1 /* 2 * Copyright (c) 2001, 2019, 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/cmsVMOperations.hpp" 36 #include "gc/cms/compactibleFreeListSpace.hpp" 37 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 38 #include "gc/cms/concurrentMarkSweepThread.hpp" 39 #include "gc/cms/parNewGeneration.hpp" 40 #include "gc/cms/promotionInfo.inline.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/owstTaskTerminator.hpp" 59 #include "gc/shared/referencePolicy.hpp" 60 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 61 #include "gc/shared/space.inline.hpp" 62 #include "gc/shared/strongRootsScope.hpp" 63 #include "gc/shared/taskqueue.inline.hpp" 64 #include "gc/shared/weakProcessor.hpp" 65 #include "gc/shared/workerPolicy.hpp" 66 #include "logging/log.hpp" 67 #include "logging/logStream.hpp" 68 #include "memory/allocation.hpp" 69 #include "memory/binaryTreeDictionary.inline.hpp" 70 #include "memory/iterator.inline.hpp" 71 #include "memory/padded.hpp" 72 #include "memory/resourceArea.hpp" 73 #include "oops/access.inline.hpp" 74 #include "oops/oop.inline.hpp" 75 #include "prims/jvmtiExport.hpp" 76 #include "runtime/atomic.hpp" 77 #include "runtime/flags/flagSetting.hpp" 78 #include "runtime/globals_extension.hpp" 79 #include "runtime/handles.inline.hpp" 80 #include "runtime/java.hpp" 81 #include "runtime/orderAccess.hpp" 82 #include "runtime/timer.hpp" 83 #include "runtime/vmThread.hpp" 84 #include "services/memoryService.hpp" 85 #include "services/runtimeService.hpp" 86 #include "utilities/align.hpp" 87 #include "utilities/stack.inline.hpp" 88 89 // statics 90 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 91 bool CMSCollector::_full_gc_requested = false; 92 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 93 94 ////////////////////////////////////////////////////////////////// 95 // In support of CMS/VM thread synchronization 96 ////////////////////////////////////////////////////////////////// 97 // We split use of the CGC_lock into 2 "levels". 98 // The low-level locking is of the usual CGC_lock monitor. We introduce 99 // a higher level "token" (hereafter "CMS token") built on top of the 100 // low level monitor (hereafter "CGC lock"). 101 // The token-passing protocol gives priority to the VM thread. The 102 // CMS-lock doesn't provide any fairness guarantees, but clients 103 // should ensure that it is only held for very short, bounded 104 // durations. 105 // 106 // When either of the CMS thread or the VM thread is involved in 107 // collection operations during which it does not want the other 108 // thread to interfere, it obtains the CMS token. 109 // 110 // If either thread tries to get the token while the other has 111 // it, that thread waits. However, if the VM thread and CMS thread 112 // both want the token, then the VM thread gets priority while the 113 // CMS thread waits. This ensures, for instance, that the "concurrent" 114 // phases of the CMS thread's work do not block out the VM thread 115 // for long periods of time as the CMS thread continues to hog 116 // the token. (See bug 4616232). 117 // 118 // The baton-passing functions are, however, controlled by the 119 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 120 // and here the low-level CMS lock, not the high level token, 121 // ensures mutual exclusion. 122 // 123 // Two important conditions that we have to satisfy: 124 // 1. if a thread does a low-level wait on the CMS lock, then it 125 // relinquishes the CMS token if it were holding that token 126 // when it acquired the low-level CMS lock. 127 // 2. any low-level notifications on the low-level lock 128 // should only be sent when a thread has relinquished the token. 129 // 130 // In the absence of either property, we'd have potential deadlock. 131 // 132 // We protect each of the CMS (concurrent and sequential) phases 133 // with the CMS _token_, not the CMS _lock_. 134 // 135 // The only code protected by CMS lock is the token acquisition code 136 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 137 // baton-passing code. 138 // 139 // Unfortunately, i couldn't come up with a good abstraction to factor and 140 // hide the naked CGC_lock manipulation in the baton-passing code 141 // further below. That's something we should try to do. Also, the proof 142 // of correctness of this 2-level locking scheme is far from obvious, 143 // and potentially quite slippery. We have an uneasy suspicion, for instance, 144 // that there may be a theoretical possibility of delay/starvation in the 145 // low-level lock/wait/notify scheme used for the baton-passing because of 146 // potential interference with the priority scheme embodied in the 147 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 148 // invocation further below and marked with "XXX 20011219YSR". 149 // Indeed, as we note elsewhere, this may become yet more slippery 150 // in the presence of multiple CMS and/or multiple VM threads. XXX 151 152 class CMSTokenSync: public StackObj { 153 private: 154 bool _is_cms_thread; 155 public: 156 CMSTokenSync(bool is_cms_thread): 157 _is_cms_thread(is_cms_thread) { 158 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 159 "Incorrect argument to constructor"); 160 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 161 } 162 163 ~CMSTokenSync() { 164 assert(_is_cms_thread ? 165 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 166 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 167 "Incorrect state"); 168 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 169 } 170 }; 171 172 // Convenience class that does a CMSTokenSync, and then acquires 173 // upto three locks. 174 class CMSTokenSyncWithLocks: public CMSTokenSync { 175 private: 176 // Note: locks are acquired in textual declaration order 177 // and released in the opposite order 178 MutexLockerEx _locker1, _locker2, _locker3; 179 public: 180 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 181 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 182 CMSTokenSync(is_cms_thread), 183 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 184 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 185 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 186 { } 187 }; 188 189 190 ////////////////////////////////////////////////////////////////// 191 // Concurrent Mark-Sweep Generation ///////////////////////////// 192 ////////////////////////////////////////////////////////////////// 193 194 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 195 196 // This struct contains per-thread things necessary to support parallel 197 // young-gen collection. 198 class CMSParGCThreadState: public CHeapObj<mtGC> { 199 public: 200 CompactibleFreeListSpaceLAB lab; 201 PromotionInfo promo; 202 203 // Constructor. 204 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 205 promo.setSpace(cfls); 206 } 207 }; 208 209 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 210 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 211 CardGeneration(rs, initial_byte_size, ct), 212 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 213 _did_compact(false) 214 { 215 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 216 HeapWord* end = (HeapWord*) _virtual_space.high(); 217 218 _direct_allocated_words = 0; 219 NOT_PRODUCT( 220 _numObjectsPromoted = 0; 221 _numWordsPromoted = 0; 222 _numObjectsAllocated = 0; 223 _numWordsAllocated = 0; 224 ) 225 226 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 227 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 228 _cmsSpace->_old_gen = this; 229 230 _gc_stats = new CMSGCStats(); 231 232 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 233 // offsets match. The ability to tell free chunks from objects 234 // depends on this property. 235 debug_only( 236 FreeChunk* junk = NULL; 237 assert(UseCompressedClassPointers || 238 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 239 "Offset of FreeChunk::_prev within FreeChunk must match" 240 " that of OopDesc::_klass within OopDesc"); 241 ) 242 243 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 244 for (uint i = 0; i < ParallelGCThreads; i++) { 245 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 246 } 247 248 _incremental_collection_failed = false; 249 // The "dilatation_factor" is the expansion that can occur on 250 // account of the fact that the minimum object size in the CMS 251 // generation may be larger than that in, say, a contiguous young 252 // generation. 253 // Ideally, in the calculation below, we'd compute the dilatation 254 // factor as: MinChunkSize/(promoting_gen's min object size) 255 // Since we do not have such a general query interface for the 256 // promoting generation, we'll instead just use the minimum 257 // object size (which today is a header's worth of space); 258 // note that all arithmetic is in units of HeapWords. 259 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 260 assert(_dilatation_factor >= 1.0, "from previous assert"); 261 } 262 263 264 // The field "_initiating_occupancy" represents the occupancy percentage 265 // at which we trigger a new collection cycle. Unless explicitly specified 266 // via CMSInitiatingOccupancyFraction (argument "io" below), it 267 // is calculated by: 268 // 269 // Let "f" be MinHeapFreeRatio in 270 // 271 // _initiating_occupancy = 100-f + 272 // f * (CMSTriggerRatio/100) 273 // where CMSTriggerRatio is the argument "tr" below. 274 // 275 // That is, if we assume the heap is at its desired maximum occupancy at the 276 // end of a collection, we let CMSTriggerRatio of the (purported) free 277 // space be allocated before initiating a new collection cycle. 278 // 279 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 280 assert(io <= 100 && tr <= 100, "Check the arguments"); 281 if (io >= 0) { 282 _initiating_occupancy = (double)io / 100.0; 283 } else { 284 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 285 (double)(tr * MinHeapFreeRatio) / 100.0) 286 / 100.0; 287 } 288 } 289 290 void ConcurrentMarkSweepGeneration::ref_processor_init() { 291 assert(collector() != NULL, "no collector"); 292 collector()->ref_processor_init(); 293 } 294 295 void CMSCollector::ref_processor_init() { 296 if (_ref_processor == NULL) { 297 // Allocate and initialize a reference processor 298 _ref_processor = 299 new ReferenceProcessor(&_span_based_discoverer, 300 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 301 ParallelGCThreads, // mt processing degree 302 _cmsGen->refs_discovery_is_mt(), // mt discovery 303 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 304 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 305 &_is_alive_closure, // closure for liveness info 306 false); // disable adjusting number of processing threads 307 // Initialize the _ref_processor field of CMSGen 308 _cmsGen->set_ref_processor(_ref_processor); 309 310 } 311 } 312 313 AdaptiveSizePolicy* CMSCollector::size_policy() { 314 return CMSHeap::heap()->size_policy(); 315 } 316 317 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 318 319 const char* gen_name = "old"; 320 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 321 // Generation Counters - generation 1, 1 subspace 322 _gen_counters = new GenerationCounters(gen_name, 1, 1, 323 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 324 325 _space_counters = new GSpaceCounters(gen_name, 0, 326 _virtual_space.reserved_size(), 327 this, _gen_counters); 328 } 329 330 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 331 _cms_gen(cms_gen) 332 { 333 assert(alpha <= 100, "bad value"); 334 _saved_alpha = alpha; 335 336 // Initialize the alphas to the bootstrap value of 100. 337 _gc0_alpha = _cms_alpha = 100; 338 339 _cms_begin_time.update(); 340 _cms_end_time.update(); 341 342 _gc0_duration = 0.0; 343 _gc0_period = 0.0; 344 _gc0_promoted = 0; 345 346 _cms_duration = 0.0; 347 _cms_period = 0.0; 348 _cms_allocated = 0; 349 350 _cms_used_at_gc0_begin = 0; 351 _cms_used_at_gc0_end = 0; 352 _allow_duty_cycle_reduction = false; 353 _valid_bits = 0; 354 } 355 356 double CMSStats::cms_free_adjustment_factor(size_t free) const { 357 // TBD: CR 6909490 358 return 1.0; 359 } 360 361 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 362 } 363 364 // If promotion failure handling is on use 365 // the padded average size of the promotion for each 366 // young generation collection. 367 double CMSStats::time_until_cms_gen_full() const { 368 size_t cms_free = _cms_gen->cmsSpace()->free(); 369 CMSHeap* heap = CMSHeap::heap(); 370 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 371 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 372 if (cms_free > expected_promotion) { 373 // Start a cms collection if there isn't enough space to promote 374 // for the next young collection. Use the padded average as 375 // a safety factor. 376 cms_free -= expected_promotion; 377 378 // Adjust by the safety factor. 379 double cms_free_dbl = (double)cms_free; 380 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 381 // Apply a further correction factor which tries to adjust 382 // for recent occurance of concurrent mode failures. 383 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 384 cms_free_dbl = cms_free_dbl * cms_adjustment; 385 386 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 387 cms_free, expected_promotion); 388 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 389 // Add 1 in case the consumption rate goes to zero. 390 return cms_free_dbl / (cms_consumption_rate() + 1.0); 391 } 392 return 0.0; 393 } 394 395 // Compare the duration of the cms collection to the 396 // time remaining before the cms generation is empty. 397 // Note that the time from the start of the cms collection 398 // to the start of the cms sweep (less than the total 399 // duration of the cms collection) can be used. This 400 // has been tried and some applications experienced 401 // promotion failures early in execution. This was 402 // possibly because the averages were not accurate 403 // enough at the beginning. 404 double CMSStats::time_until_cms_start() const { 405 // We add "gc0_period" to the "work" calculation 406 // below because this query is done (mostly) at the 407 // end of a scavenge, so we need to conservatively 408 // account for that much possible delay 409 // in the query so as to avoid concurrent mode failures 410 // due to starting the collection just a wee bit too 411 // late. 412 double work = cms_duration() + gc0_period(); 413 double deadline = time_until_cms_gen_full(); 414 // If a concurrent mode failure occurred recently, we want to be 415 // more conservative and halve our expected time_until_cms_gen_full() 416 if (work > deadline) { 417 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 418 cms_duration(), gc0_period(), time_until_cms_gen_full()); 419 return 0.0; 420 } 421 return work - deadline; 422 } 423 424 #ifndef PRODUCT 425 void CMSStats::print_on(outputStream *st) const { 426 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 427 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 428 gc0_duration(), gc0_period(), gc0_promoted()); 429 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 430 cms_duration(), cms_period(), cms_allocated()); 431 st->print(",cms_since_beg=%g,cms_since_end=%g", 432 cms_time_since_begin(), cms_time_since_end()); 433 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 434 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 435 436 if (valid()) { 437 st->print(",promo_rate=%g,cms_alloc_rate=%g", 438 promotion_rate(), cms_allocation_rate()); 439 st->print(",cms_consumption_rate=%g,time_until_full=%g", 440 cms_consumption_rate(), time_until_cms_gen_full()); 441 } 442 st->cr(); 443 } 444 #endif // #ifndef PRODUCT 445 446 CMSCollector::CollectorState CMSCollector::_collectorState = 447 CMSCollector::Idling; 448 bool CMSCollector::_foregroundGCIsActive = false; 449 bool CMSCollector::_foregroundGCShouldWait = false; 450 451 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 452 CardTableRS* ct, 453 ConcurrentMarkSweepPolicy* cp): 454 _overflow_list(NULL), 455 _conc_workers(NULL), // may be set later 456 _completed_initialization(false), 457 _collection_count_start(0), 458 _should_unload_classes(CMSClassUnloadingEnabled), 459 _concurrent_cycles_since_last_unload(0), 460 _roots_scanning_options(GenCollectedHeap::SO_None), 461 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 462 _verifying(false), 463 _collector_policy(cp), 464 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 465 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 466 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 467 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 468 _cms_start_registered(false), 469 _cmsGen(cmsGen), 470 // Adjust span to cover old (cms) gen 471 _span(cmsGen->reserved()), 472 _ct(ct), 473 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 474 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 475 -1 /* lock-free */, "No_lock" /* dummy */), 476 _restart_addr(NULL), 477 _ser_pmc_preclean_ovflw(0), 478 _ser_pmc_remark_ovflw(0), 479 _par_pmc_remark_ovflw(0), 480 _ser_kac_preclean_ovflw(0), 481 _ser_kac_ovflw(0), 482 _par_kac_ovflw(0), 483 #ifndef PRODUCT 484 _num_par_pushes(0), 485 #endif 486 _span_based_discoverer(_span), 487 _ref_processor(NULL), // will be set later 488 // Construct the is_alive_closure with _span & markBitMap 489 _is_alive_closure(_span, &_markBitMap), 490 _modUnionClosurePar(&_modUnionTable), 491 _between_prologue_and_epilogue(false), 492 _abort_preclean(false), 493 _start_sampling(false), 494 _stats(cmsGen), 495 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 496 //verify that this lock should be acquired with safepoint check. 497 Monitor::_safepoint_check_sometimes)), 498 _eden_chunk_array(NULL), // may be set in ctor body 499 _eden_chunk_index(0), // -- ditto -- 500 _eden_chunk_capacity(0), // -- ditto -- 501 _survivor_chunk_array(NULL), // -- ditto -- 502 _survivor_chunk_index(0), // -- ditto -- 503 _survivor_chunk_capacity(0), // -- ditto -- 504 _survivor_plab_array(NULL) // -- ditto -- 505 { 506 // Now expand the span and allocate the collection support structures 507 // (MUT, marking bit map etc.) to cover both generations subject to 508 // collection. 509 510 // For use by dirty card to oop closures. 511 _cmsGen->cmsSpace()->set_collector(this); 512 513 // Allocate MUT and marking bit map 514 { 515 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 516 if (!_markBitMap.allocate(_span)) { 517 log_warning(gc)("Failed to allocate CMS Bit Map"); 518 return; 519 } 520 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 521 } 522 { 523 _modUnionTable.allocate(_span); 524 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 525 } 526 527 if (!_markStack.allocate(MarkStackSize)) { 528 log_warning(gc)("Failed to allocate CMS Marking Stack"); 529 return; 530 } 531 532 // Support for multi-threaded concurrent phases 533 if (CMSConcurrentMTEnabled) { 534 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 535 // just for now 536 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 537 } 538 if (ConcGCThreads > 1) { 539 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 540 ConcGCThreads, true); 541 if (_conc_workers == NULL) { 542 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 543 CMSConcurrentMTEnabled = false; 544 } else { 545 _conc_workers->initialize_workers(); 546 } 547 } else { 548 CMSConcurrentMTEnabled = false; 549 } 550 } 551 if (!CMSConcurrentMTEnabled) { 552 ConcGCThreads = 0; 553 } else { 554 // Turn off CMSCleanOnEnter optimization temporarily for 555 // the MT case where it's not fixed yet; see 6178663. 556 CMSCleanOnEnter = false; 557 } 558 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 559 "Inconsistency"); 560 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 561 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 562 563 // Parallel task queues; these are shared for the 564 // concurrent and stop-world phases of CMS, but 565 // are not shared with parallel scavenge (ParNew). 566 { 567 uint i; 568 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 569 570 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 571 || ParallelRefProcEnabled) 572 && num_queues > 0) { 573 _task_queues = new OopTaskQueueSet(num_queues); 574 if (_task_queues == NULL) { 575 log_warning(gc)("task_queues allocation failure."); 576 return; 577 } 578 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 579 for (i = 0; i < num_queues; i++) { 580 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 581 if (q == NULL) { 582 log_warning(gc)("work_queue allocation failure."); 583 return; 584 } 585 _task_queues->register_queue(i, q); 586 } 587 for (i = 0; i < num_queues; i++) { 588 _task_queues->queue(i)->initialize(); 589 } 590 } 591 } 592 593 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 594 595 // Clip CMSBootstrapOccupancy between 0 and 100. 596 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 597 598 // Now tell CMS generations the identity of their collector 599 ConcurrentMarkSweepGeneration::set_collector(this); 600 601 // Create & start a CMS thread for this CMS collector 602 _cmsThread = ConcurrentMarkSweepThread::start(this); 603 assert(cmsThread() != NULL, "CMS Thread should have been created"); 604 assert(cmsThread()->collector() == this, 605 "CMS Thread should refer to this gen"); 606 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 607 608 // Support for parallelizing young gen rescan 609 CMSHeap* heap = CMSHeap::heap(); 610 _young_gen = heap->young_gen(); 611 if (heap->supports_inline_contig_alloc()) { 612 _top_addr = heap->top_addr(); 613 _end_addr = heap->end_addr(); 614 assert(_young_gen != NULL, "no _young_gen"); 615 _eden_chunk_index = 0; 616 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 617 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 618 } 619 620 // Support for parallelizing survivor space rescan 621 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 622 const size_t max_plab_samples = 623 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 624 625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 628 _survivor_chunk_capacity = max_plab_samples; 629 for (uint i = 0; i < ParallelGCThreads; i++) { 630 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 631 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 632 assert(cur->end() == 0, "Should be 0"); 633 assert(cur->array() == vec, "Should be vec"); 634 assert(cur->capacity() == max_plab_samples, "Error"); 635 } 636 } 637 638 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 639 _gc_counters = new CollectorCounters("CMS", 1); 640 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 641 _completed_initialization = true; 642 _inter_sweep_timer.start(); // start of time 643 } 644 645 const char* ConcurrentMarkSweepGeneration::name() const { 646 return "concurrent mark-sweep generation"; 647 } 648 void ConcurrentMarkSweepGeneration::update_counters() { 649 if (UsePerfData) { 650 _space_counters->update_all(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 // this is an optimized version of update_counters(). it takes the 656 // used value as a parameter rather than computing it. 657 // 658 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 659 if (UsePerfData) { 660 _space_counters->update_used(used); 661 _space_counters->update_capacity(); 662 _gen_counters->update_all(); 663 } 664 } 665 666 void ConcurrentMarkSweepGeneration::print() const { 667 Generation::print(); 668 cmsSpace()->print(); 669 } 670 671 #ifndef PRODUCT 672 void ConcurrentMarkSweepGeneration::print_statistics() { 673 cmsSpace()->printFLCensus(0); 674 } 675 #endif 676 677 size_t 678 ConcurrentMarkSweepGeneration::contiguous_available() const { 679 // dld proposes an improvement in precision here. If the committed 680 // part of the space ends in a free block we should add that to 681 // uncommitted size in the calculation below. Will make this 682 // change later, staying with the approximation below for the 683 // time being. -- ysr. 684 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 685 } 686 687 size_t 688 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 689 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 690 } 691 692 size_t ConcurrentMarkSweepGeneration::max_available() const { 693 return free() + _virtual_space.uncommitted_size(); 694 } 695 696 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 697 size_t available = max_available(); 698 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 699 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 700 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 701 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 702 return res; 703 } 704 705 // At a promotion failure dump information on block layout in heap 706 // (cms old generation). 707 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 708 Log(gc, promotion) log; 709 if (log.is_trace()) { 710 LogStream ls(log.trace()); 711 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 712 } 713 } 714 715 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 716 // Clear the promotion information. These pointers can be adjusted 717 // along with all the other pointers into the heap but 718 // compaction is expected to be a rare event with 719 // a heap using cms so don't do it without seeing the need. 720 for (uint i = 0; i < ParallelGCThreads; i++) { 721 _par_gc_thread_states[i]->promo.reset(); 722 } 723 } 724 725 void ConcurrentMarkSweepGeneration::compute_new_size() { 726 assert_locked_or_safepoint(Heap_lock); 727 728 // If incremental collection failed, we just want to expand 729 // to the limit. 730 if (incremental_collection_failed()) { 731 clear_incremental_collection_failed(); 732 grow_to_reserved(); 733 return; 734 } 735 736 // The heap has been compacted but not reset yet. 737 // Any metric such as free() or used() will be incorrect. 738 739 CardGeneration::compute_new_size(); 740 741 // Reset again after a possible resizing 742 if (did_compact()) { 743 cmsSpace()->reset_after_compaction(); 744 } 745 } 746 747 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 748 assert_locked_or_safepoint(Heap_lock); 749 750 // If incremental collection failed, we just want to expand 751 // to the limit. 752 if (incremental_collection_failed()) { 753 clear_incremental_collection_failed(); 754 grow_to_reserved(); 755 return; 756 } 757 758 double free_percentage = ((double) free()) / capacity(); 759 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 760 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 761 762 // compute expansion delta needed for reaching desired free percentage 763 if (free_percentage < desired_free_percentage) { 764 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 765 assert(desired_capacity >= capacity(), "invalid expansion size"); 766 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 767 Log(gc) log; 768 if (log.is_trace()) { 769 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 770 log.trace("From compute_new_size: "); 771 log.trace(" Free fraction %f", free_percentage); 772 log.trace(" Desired free fraction %f", desired_free_percentage); 773 log.trace(" Maximum free fraction %f", maximum_free_percentage); 774 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 775 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 776 CMSHeap* heap = CMSHeap::heap(); 777 size_t young_size = heap->young_gen()->capacity(); 778 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 779 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 780 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 781 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 782 } 783 // safe if expansion fails 784 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 785 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 786 } else { 787 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 788 assert(desired_capacity <= capacity(), "invalid expansion size"); 789 size_t shrink_bytes = capacity() - desired_capacity; 790 // Don't shrink unless the delta is greater than the minimum shrink we want 791 if (shrink_bytes >= MinHeapDeltaBytes) { 792 shrink_free_list_by(shrink_bytes); 793 } 794 } 795 } 796 797 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 798 return cmsSpace()->freelistLock(); 799 } 800 801 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 802 CMSSynchronousYieldRequest yr; 803 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 804 return have_lock_and_allocate(size, tlab); 805 } 806 807 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 808 bool tlab /* ignored */) { 809 assert_lock_strong(freelistLock()); 810 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 811 HeapWord* res = cmsSpace()->allocate(adjustedSize); 812 // Allocate the object live (grey) if the background collector has 813 // started marking. This is necessary because the marker may 814 // have passed this address and consequently this object will 815 // not otherwise be greyed and would be incorrectly swept up. 816 // Note that if this object contains references, the writing 817 // of those references will dirty the card containing this object 818 // allowing the object to be blackened (and its references scanned) 819 // either during a preclean phase or at the final checkpoint. 820 if (res != NULL) { 821 // We may block here with an uninitialized object with 822 // its mark-bit or P-bits not yet set. Such objects need 823 // to be safely navigable by block_start(). 824 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 825 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 826 collector()->direct_allocated(res, adjustedSize); 827 _direct_allocated_words += adjustedSize; 828 // allocation counters 829 NOT_PRODUCT( 830 _numObjectsAllocated++; 831 _numWordsAllocated += (int)adjustedSize; 832 ) 833 } 834 return res; 835 } 836 837 // In the case of direct allocation by mutators in a generation that 838 // is being concurrently collected, the object must be allocated 839 // live (grey) if the background collector has started marking. 840 // This is necessary because the marker may 841 // have passed this address and consequently this object will 842 // not otherwise be greyed and would be incorrectly swept up. 843 // Note that if this object contains references, the writing 844 // of those references will dirty the card containing this object 845 // allowing the object to be blackened (and its references scanned) 846 // either during a preclean phase or at the final checkpoint. 847 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 848 assert(_markBitMap.covers(start, size), "Out of bounds"); 849 if (_collectorState >= Marking) { 850 MutexLockerEx y(_markBitMap.lock(), 851 Mutex::_no_safepoint_check_flag); 852 // [see comments preceding SweepClosure::do_blk() below for details] 853 // 854 // Can the P-bits be deleted now? JJJ 855 // 856 // 1. need to mark the object as live so it isn't collected 857 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 858 // 3. need to mark the end of the object so marking, precleaning or sweeping 859 // can skip over uninitialized or unparsable objects. An allocated 860 // object is considered uninitialized for our purposes as long as 861 // its klass word is NULL. All old gen objects are parsable 862 // as soon as they are initialized.) 863 _markBitMap.mark(start); // object is live 864 _markBitMap.mark(start + 1); // object is potentially uninitialized? 865 _markBitMap.mark(start + size - 1); 866 // mark end of object 867 } 868 // check that oop looks uninitialized 869 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 870 } 871 872 void CMSCollector::promoted(bool par, HeapWord* start, 873 bool is_obj_array, size_t obj_size) { 874 assert(_markBitMap.covers(start), "Out of bounds"); 875 // See comment in direct_allocated() about when objects should 876 // be allocated live. 877 if (_collectorState >= Marking) { 878 // we already hold the marking bit map lock, taken in 879 // the prologue 880 if (par) { 881 _markBitMap.par_mark(start); 882 } else { 883 _markBitMap.mark(start); 884 } 885 // We don't need to mark the object as uninitialized (as 886 // in direct_allocated above) because this is being done with the 887 // world stopped and the object will be initialized by the 888 // time the marking, precleaning or sweeping get to look at it. 889 // But see the code for copying objects into the CMS generation, 890 // where we need to ensure that concurrent readers of the 891 // block offset table are able to safely navigate a block that 892 // is in flux from being free to being allocated (and in 893 // transition while being copied into) and subsequently 894 // becoming a bona-fide object when the copy/promotion is complete. 895 assert(SafepointSynchronize::is_at_safepoint(), 896 "expect promotion only at safepoints"); 897 898 if (_collectorState < Sweeping) { 899 // Mark the appropriate cards in the modUnionTable, so that 900 // this object gets scanned before the sweep. If this is 901 // not done, CMS generation references in the object might 902 // not get marked. 903 // For the case of arrays, which are otherwise precisely 904 // marked, we need to dirty the entire array, not just its head. 905 if (is_obj_array) { 906 // The [par_]mark_range() method expects mr.end() below to 907 // be aligned to the granularity of a bit's representation 908 // in the heap. In the case of the MUT below, that's a 909 // card size. 910 MemRegion mr(start, 911 align_up(start + obj_size, 912 CardTable::card_size /* bytes */)); 913 if (par) { 914 _modUnionTable.par_mark_range(mr); 915 } else { 916 _modUnionTable.mark_range(mr); 917 } 918 } else { // not an obj array; we can just mark the head 919 if (par) { 920 _modUnionTable.par_mark(start); 921 } else { 922 _modUnionTable.mark(start); 923 } 924 } 925 } 926 } 927 } 928 929 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 930 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 931 // allocate, copy and if necessary update promoinfo -- 932 // delegate to underlying space. 933 assert_lock_strong(freelistLock()); 934 935 #ifndef PRODUCT 936 if (CMSHeap::heap()->promotion_should_fail()) { 937 return NULL; 938 } 939 #endif // #ifndef PRODUCT 940 941 oop res = _cmsSpace->promote(obj, obj_size); 942 if (res == NULL) { 943 // expand and retry 944 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 945 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 946 // Since this is the old generation, we don't try to promote 947 // into a more senior generation. 948 res = _cmsSpace->promote(obj, obj_size); 949 } 950 if (res != NULL) { 951 // See comment in allocate() about when objects should 952 // be allocated live. 953 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 954 collector()->promoted(false, // Not parallel 955 (HeapWord*)res, obj->is_objArray(), obj_size); 956 // promotion counters 957 NOT_PRODUCT( 958 _numObjectsPromoted++; 959 _numWordsPromoted += 960 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 961 ) 962 } 963 return res; 964 } 965 966 967 // IMPORTANT: Notes on object size recognition in CMS. 968 // --------------------------------------------------- 969 // A block of storage in the CMS generation is always in 970 // one of three states. A free block (FREE), an allocated 971 // object (OBJECT) whose size() method reports the correct size, 972 // and an intermediate state (TRANSIENT) in which its size cannot 973 // be accurately determined. 974 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 975 // ----------------------------------------------------- 976 // FREE: klass_word & 1 == 1; mark_word holds block size 977 // 978 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 979 // obj->size() computes correct size 980 // 981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 982 // 983 // STATE IDENTIFICATION: (64 bit+COOPS) 984 // ------------------------------------ 985 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 986 // 987 // OBJECT: klass_word installed; klass_word != 0; 988 // obj->size() computes correct size 989 // 990 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 991 // 992 // 993 // STATE TRANSITION DIAGRAM 994 // 995 // mut / parnew mut / parnew 996 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 997 // ^ | 998 // |------------------------ DEAD <------------------------------------| 999 // sweep mut 1000 // 1001 // While a block is in TRANSIENT state its size cannot be determined 1002 // so readers will either need to come back later or stall until 1003 // the size can be determined. Note that for the case of direct 1004 // allocation, P-bits, when available, may be used to determine the 1005 // size of an object that may not yet have been initialized. 1006 1007 // Things to support parallel young-gen collection. 1008 oop 1009 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1010 oop old, markOop m, 1011 size_t word_sz) { 1012 #ifndef PRODUCT 1013 if (CMSHeap::heap()->promotion_should_fail()) { 1014 return NULL; 1015 } 1016 #endif // #ifndef PRODUCT 1017 1018 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1019 PromotionInfo* promoInfo = &ps->promo; 1020 // if we are tracking promotions, then first ensure space for 1021 // promotion (including spooling space for saving header if necessary). 1022 // then allocate and copy, then track promoted info if needed. 1023 // When tracking (see PromotionInfo::track()), the mark word may 1024 // be displaced and in this case restoration of the mark word 1025 // occurs in the (oop_since_save_marks_)iterate phase. 1026 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1027 // Out of space for allocating spooling buffers; 1028 // try expanding and allocating spooling buffers. 1029 if (!expand_and_ensure_spooling_space(promoInfo)) { 1030 return NULL; 1031 } 1032 } 1033 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1034 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1035 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1036 if (obj_ptr == NULL) { 1037 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1038 if (obj_ptr == NULL) { 1039 return NULL; 1040 } 1041 } 1042 oop obj = oop(obj_ptr); 1043 OrderAccess::storestore(); 1044 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1045 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1046 // IMPORTANT: See note on object initialization for CMS above. 1047 // Otherwise, copy the object. Here we must be careful to insert the 1048 // klass pointer last, since this marks the block as an allocated object. 1049 // Except with compressed oops it's the mark word. 1050 HeapWord* old_ptr = (HeapWord*)old; 1051 // Restore the mark word copied above. 1052 obj->set_mark_raw(m); 1053 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1054 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1055 OrderAccess::storestore(); 1056 1057 if (UseCompressedClassPointers) { 1058 // Copy gap missed by (aligned) header size calculation below 1059 obj->set_klass_gap(old->klass_gap()); 1060 } 1061 if (word_sz > (size_t)oopDesc::header_size()) { 1062 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1063 obj_ptr + oopDesc::header_size(), 1064 word_sz - oopDesc::header_size()); 1065 } 1066 1067 // Now we can track the promoted object, if necessary. We take care 1068 // to delay the transition from uninitialized to full object 1069 // (i.e., insertion of klass pointer) until after, so that it 1070 // atomically becomes a promoted object. 1071 if (promoInfo->tracking()) { 1072 promoInfo->track((PromotedObject*)obj, old->klass()); 1073 } 1074 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1075 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1076 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1077 1078 // Finally, install the klass pointer (this should be volatile). 1079 OrderAccess::storestore(); 1080 obj->set_klass(old->klass()); 1081 // We should now be able to calculate the right size for this object 1082 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1083 1084 collector()->promoted(true, // parallel 1085 obj_ptr, old->is_objArray(), word_sz); 1086 1087 NOT_PRODUCT( 1088 Atomic::inc(&_numObjectsPromoted); 1089 Atomic::add(alloc_sz, &_numWordsPromoted); 1090 ) 1091 1092 return obj; 1093 } 1094 1095 void 1096 ConcurrentMarkSweepGeneration:: 1097 par_promote_alloc_done(int thread_num) { 1098 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1099 ps->lab.retire(thread_num); 1100 } 1101 1102 void 1103 ConcurrentMarkSweepGeneration:: 1104 par_oop_since_save_marks_iterate_done(int thread_num) { 1105 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1106 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1107 ps->promo.promoted_oops_iterate(dummy_cl); 1108 1109 // Because card-scanning has been completed, subsequent phases 1110 // (e.g., reference processing) will not need to recognize which 1111 // objects have been promoted during this GC. So, we can now disable 1112 // promotion tracking. 1113 ps->promo.stopTrackingPromotions(); 1114 } 1115 1116 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1117 size_t size, 1118 bool tlab) 1119 { 1120 // We allow a STW collection only if a full 1121 // collection was requested. 1122 return full || should_allocate(size, tlab); // FIX ME !!! 1123 // This and promotion failure handling are connected at the 1124 // hip and should be fixed by untying them. 1125 } 1126 1127 bool CMSCollector::shouldConcurrentCollect() { 1128 LogTarget(Trace, gc) log; 1129 1130 if (_full_gc_requested) { 1131 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1132 return true; 1133 } 1134 1135 FreelistLocker x(this); 1136 // ------------------------------------------------------------------ 1137 // Print out lots of information which affects the initiation of 1138 // a collection. 1139 if (log.is_enabled() && stats().valid()) { 1140 log.print("CMSCollector shouldConcurrentCollect: "); 1141 1142 LogStream out(log); 1143 stats().print_on(&out); 1144 1145 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1146 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1147 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1148 log.print("promotion_rate=%g", stats().promotion_rate()); 1149 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1150 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1151 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1152 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1153 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1154 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1155 } 1156 // ------------------------------------------------------------------ 1157 1158 // If the estimated time to complete a cms collection (cms_duration()) 1159 // is less than the estimated time remaining until the cms generation 1160 // is full, start a collection. 1161 if (!UseCMSInitiatingOccupancyOnly) { 1162 if (stats().valid()) { 1163 if (stats().time_until_cms_start() == 0.0) { 1164 return true; 1165 } 1166 } else { 1167 // We want to conservatively collect somewhat early in order 1168 // to try and "bootstrap" our CMS/promotion statistics; 1169 // this branch will not fire after the first successful CMS 1170 // collection because the stats should then be valid. 1171 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1172 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1173 _cmsGen->occupancy(), _bootstrap_occupancy); 1174 return true; 1175 } 1176 } 1177 } 1178 1179 // Otherwise, we start a collection cycle if 1180 // old gen want a collection cycle started. Each may use 1181 // an appropriate criterion for making this decision. 1182 // XXX We need to make sure that the gen expansion 1183 // criterion dovetails well with this. XXX NEED TO FIX THIS 1184 if (_cmsGen->should_concurrent_collect()) { 1185 log.print("CMS old gen initiated"); 1186 return true; 1187 } 1188 1189 // We start a collection if we believe an incremental collection may fail; 1190 // this is not likely to be productive in practice because it's probably too 1191 // late anyway. 1192 CMSHeap* heap = CMSHeap::heap(); 1193 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1194 log.print("CMSCollector: collect because incremental collection will fail "); 1195 return true; 1196 } 1197 1198 if (MetaspaceGC::should_concurrent_collect()) { 1199 log.print("CMSCollector: collect for metadata allocation "); 1200 return true; 1201 } 1202 1203 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1204 if (CMSTriggerInterval >= 0) { 1205 if (CMSTriggerInterval == 0) { 1206 // Trigger always 1207 return true; 1208 } 1209 1210 // Check the CMS time since begin (we do not check the stats validity 1211 // as we want to be able to trigger the first CMS cycle as well) 1212 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1213 if (stats().valid()) { 1214 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1215 stats().cms_time_since_begin()); 1216 } else { 1217 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1218 } 1219 return true; 1220 } 1221 } 1222 1223 return false; 1224 } 1225 1226 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1227 1228 // Clear _expansion_cause fields of constituent generations 1229 void CMSCollector::clear_expansion_cause() { 1230 _cmsGen->clear_expansion_cause(); 1231 } 1232 1233 // We should be conservative in starting a collection cycle. To 1234 // start too eagerly runs the risk of collecting too often in the 1235 // extreme. To collect too rarely falls back on full collections, 1236 // which works, even if not optimum in terms of concurrent work. 1237 // As a work around for too eagerly collecting, use the flag 1238 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1239 // giving the user an easily understandable way of controlling the 1240 // collections. 1241 // We want to start a new collection cycle if any of the following 1242 // conditions hold: 1243 // . our current occupancy exceeds the configured initiating occupancy 1244 // for this generation, or 1245 // . we recently needed to expand this space and have not, since that 1246 // expansion, done a collection of this generation, or 1247 // . the underlying space believes that it may be a good idea to initiate 1248 // a concurrent collection (this may be based on criteria such as the 1249 // following: the space uses linear allocation and linear allocation is 1250 // going to fail, or there is believed to be excessive fragmentation in 1251 // the generation, etc... or ... 1252 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1253 // the case of the old generation; see CR 6543076): 1254 // we may be approaching a point at which allocation requests may fail because 1255 // we will be out of sufficient free space given allocation rate estimates.] 1256 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1257 1258 assert_lock_strong(freelistLock()); 1259 if (occupancy() > initiating_occupancy()) { 1260 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1261 short_name(), occupancy(), initiating_occupancy()); 1262 return true; 1263 } 1264 if (UseCMSInitiatingOccupancyOnly) { 1265 return false; 1266 } 1267 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1268 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1269 return true; 1270 } 1271 return false; 1272 } 1273 1274 void ConcurrentMarkSweepGeneration::collect(bool full, 1275 bool clear_all_soft_refs, 1276 size_t size, 1277 bool tlab) 1278 { 1279 collector()->collect(full, clear_all_soft_refs, size, tlab); 1280 } 1281 1282 void CMSCollector::collect(bool full, 1283 bool clear_all_soft_refs, 1284 size_t size, 1285 bool tlab) 1286 { 1287 // The following "if" branch is present for defensive reasons. 1288 // In the current uses of this interface, it can be replaced with: 1289 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1290 // But I am not placing that assert here to allow future 1291 // generality in invoking this interface. 1292 if (GCLocker::is_active()) { 1293 // A consistency test for GCLocker 1294 assert(GCLocker::needs_gc(), "Should have been set already"); 1295 // Skip this foreground collection, instead 1296 // expanding the heap if necessary. 1297 // Need the free list locks for the call to free() in compute_new_size() 1298 compute_new_size(); 1299 return; 1300 } 1301 acquire_control_and_collect(full, clear_all_soft_refs); 1302 } 1303 1304 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1305 CMSHeap* heap = CMSHeap::heap(); 1306 unsigned int gc_count = heap->total_full_collections(); 1307 if (gc_count == full_gc_count) { 1308 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1309 _full_gc_requested = true; 1310 _full_gc_cause = cause; 1311 CGC_lock->notify(); // nudge CMS thread 1312 } else { 1313 assert(gc_count > full_gc_count, "Error: causal loop"); 1314 } 1315 } 1316 1317 bool CMSCollector::is_external_interruption() { 1318 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1319 return GCCause::is_user_requested_gc(cause) || 1320 GCCause::is_serviceability_requested_gc(cause); 1321 } 1322 1323 void CMSCollector::report_concurrent_mode_interruption() { 1324 if (is_external_interruption()) { 1325 log_debug(gc)("Concurrent mode interrupted"); 1326 } else { 1327 log_debug(gc)("Concurrent mode failure"); 1328 _gc_tracer_cm->report_concurrent_mode_failure(); 1329 } 1330 } 1331 1332 1333 // The foreground and background collectors need to coordinate in order 1334 // to make sure that they do not mutually interfere with CMS collections. 1335 // When a background collection is active, 1336 // the foreground collector may need to take over (preempt) and 1337 // synchronously complete an ongoing collection. Depending on the 1338 // frequency of the background collections and the heap usage 1339 // of the application, this preemption can be seldom or frequent. 1340 // There are only certain 1341 // points in the background collection that the "collection-baton" 1342 // can be passed to the foreground collector. 1343 // 1344 // The foreground collector will wait for the baton before 1345 // starting any part of the collection. The foreground collector 1346 // will only wait at one location. 1347 // 1348 // The background collector will yield the baton before starting a new 1349 // phase of the collection (e.g., before initial marking, marking from roots, 1350 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1351 // of the loop which switches the phases. The background collector does some 1352 // of the phases (initial mark, final re-mark) with the world stopped. 1353 // Because of locking involved in stopping the world, 1354 // the foreground collector should not block waiting for the background 1355 // collector when it is doing a stop-the-world phase. The background 1356 // collector will yield the baton at an additional point just before 1357 // it enters a stop-the-world phase. Once the world is stopped, the 1358 // background collector checks the phase of the collection. If the 1359 // phase has not changed, it proceeds with the collection. If the 1360 // phase has changed, it skips that phase of the collection. See 1361 // the comments on the use of the Heap_lock in collect_in_background(). 1362 // 1363 // Variable used in baton passing. 1364 // _foregroundGCIsActive - Set to true by the foreground collector when 1365 // it wants the baton. The foreground clears it when it has finished 1366 // the collection. 1367 // _foregroundGCShouldWait - Set to true by the background collector 1368 // when it is running. The foreground collector waits while 1369 // _foregroundGCShouldWait is true. 1370 // CGC_lock - monitor used to protect access to the above variables 1371 // and to notify the foreground and background collectors. 1372 // _collectorState - current state of the CMS collection. 1373 // 1374 // The foreground collector 1375 // acquires the CGC_lock 1376 // sets _foregroundGCIsActive 1377 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1378 // various locks acquired in preparation for the collection 1379 // are released so as not to block the background collector 1380 // that is in the midst of a collection 1381 // proceeds with the collection 1382 // clears _foregroundGCIsActive 1383 // returns 1384 // 1385 // The background collector in a loop iterating on the phases of the 1386 // collection 1387 // acquires the CGC_lock 1388 // sets _foregroundGCShouldWait 1389 // if _foregroundGCIsActive is set 1390 // clears _foregroundGCShouldWait, notifies _CGC_lock 1391 // waits on _CGC_lock for _foregroundGCIsActive to become false 1392 // and exits the loop. 1393 // otherwise 1394 // proceed with that phase of the collection 1395 // if the phase is a stop-the-world phase, 1396 // yield the baton once more just before enqueueing 1397 // the stop-world CMS operation (executed by the VM thread). 1398 // returns after all phases of the collection are done 1399 // 1400 1401 void CMSCollector::acquire_control_and_collect(bool full, 1402 bool clear_all_soft_refs) { 1403 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1404 assert(!Thread::current()->is_ConcurrentGC_thread(), 1405 "shouldn't try to acquire control from self!"); 1406 1407 // Start the protocol for acquiring control of the 1408 // collection from the background collector (aka CMS thread). 1409 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1410 "VM thread should have CMS token"); 1411 // Remember the possibly interrupted state of an ongoing 1412 // concurrent collection 1413 CollectorState first_state = _collectorState; 1414 1415 // Signal to a possibly ongoing concurrent collection that 1416 // we want to do a foreground collection. 1417 _foregroundGCIsActive = true; 1418 1419 // release locks and wait for a notify from the background collector 1420 // releasing the locks in only necessary for phases which 1421 // do yields to improve the granularity of the collection. 1422 assert_lock_strong(bitMapLock()); 1423 // We need to lock the Free list lock for the space that we are 1424 // currently collecting. 1425 assert(haveFreelistLocks(), "Must be holding free list locks"); 1426 bitMapLock()->unlock(); 1427 releaseFreelistLocks(); 1428 { 1429 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1430 if (_foregroundGCShouldWait) { 1431 // We are going to be waiting for action for the CMS thread; 1432 // it had better not be gone (for instance at shutdown)! 1433 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1434 "CMS thread must be running"); 1435 // Wait here until the background collector gives us the go-ahead 1436 ConcurrentMarkSweepThread::clear_CMS_flag( 1437 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1438 // Get a possibly blocked CMS thread going: 1439 // Note that we set _foregroundGCIsActive true above, 1440 // without protection of the CGC_lock. 1441 CGC_lock->notify(); 1442 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1443 "Possible deadlock"); 1444 while (_foregroundGCShouldWait) { 1445 // wait for notification 1446 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1447 // Possibility of delay/starvation here, since CMS token does 1448 // not know to give priority to VM thread? Actually, i think 1449 // there wouldn't be any delay/starvation, but the proof of 1450 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1451 } 1452 ConcurrentMarkSweepThread::set_CMS_flag( 1453 ConcurrentMarkSweepThread::CMS_vm_has_token); 1454 } 1455 } 1456 // The CMS_token is already held. Get back the other locks. 1457 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1458 "VM thread should have CMS token"); 1459 getFreelistLocks(); 1460 bitMapLock()->lock_without_safepoint_check(); 1461 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1462 p2i(Thread::current()), first_state); 1463 log_debug(gc, state)(" gets control with state %d", _collectorState); 1464 1465 // Inform cms gen if this was due to partial collection failing. 1466 // The CMS gen may use this fact to determine its expansion policy. 1467 CMSHeap* heap = CMSHeap::heap(); 1468 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1469 assert(!_cmsGen->incremental_collection_failed(), 1470 "Should have been noticed, reacted to and cleared"); 1471 _cmsGen->set_incremental_collection_failed(); 1472 } 1473 1474 if (first_state > Idling) { 1475 report_concurrent_mode_interruption(); 1476 } 1477 1478 set_did_compact(true); 1479 1480 // If the collection is being acquired from the background 1481 // collector, there may be references on the discovered 1482 // references lists. Abandon those references, since some 1483 // of them may have become unreachable after concurrent 1484 // discovery; the STW compacting collector will redo discovery 1485 // more precisely, without being subject to floating garbage. 1486 // Leaving otherwise unreachable references in the discovered 1487 // lists would require special handling. 1488 ref_processor()->disable_discovery(); 1489 ref_processor()->abandon_partial_discovery(); 1490 ref_processor()->verify_no_references_recorded(); 1491 1492 if (first_state > Idling) { 1493 save_heap_summary(); 1494 } 1495 1496 do_compaction_work(clear_all_soft_refs); 1497 1498 // Has the GC time limit been exceeded? 1499 size_t max_eden_size = _young_gen->max_eden_size(); 1500 GCCause::Cause gc_cause = heap->gc_cause(); 1501 size_policy()->check_gc_overhead_limit(_young_gen->eden()->used(), 1502 _cmsGen->max_capacity(), 1503 max_eden_size, 1504 full, 1505 gc_cause, 1506 heap->soft_ref_policy()); 1507 1508 // Reset the expansion cause, now that we just completed 1509 // a collection cycle. 1510 clear_expansion_cause(); 1511 _foregroundGCIsActive = false; 1512 return; 1513 } 1514 1515 // Resize the tenured generation 1516 // after obtaining the free list locks for the 1517 // two generations. 1518 void CMSCollector::compute_new_size() { 1519 assert_locked_or_safepoint(Heap_lock); 1520 FreelistLocker z(this); 1521 MetaspaceGC::compute_new_size(); 1522 _cmsGen->compute_new_size_free_list(); 1523 } 1524 1525 // A work method used by the foreground collector to do 1526 // a mark-sweep-compact. 1527 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1528 CMSHeap* heap = CMSHeap::heap(); 1529 1530 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1531 gc_timer->register_gc_start(); 1532 1533 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1534 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1535 1536 heap->pre_full_gc_dump(gc_timer); 1537 1538 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1539 1540 // Temporarily widen the span of the weak reference processing to 1541 // the entire heap. 1542 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1543 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1544 // Temporarily, clear the "is_alive_non_header" field of the 1545 // reference processor. 1546 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1547 // Temporarily make reference _processing_ single threaded (non-MT). 1548 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1549 // Temporarily make refs discovery atomic 1550 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1551 // Temporarily make reference _discovery_ single threaded (non-MT) 1552 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1553 1554 ref_processor()->set_enqueuing_is_done(false); 1555 ref_processor()->enable_discovery(); 1556 ref_processor()->setup_policy(clear_all_soft_refs); 1557 // If an asynchronous collection finishes, the _modUnionTable is 1558 // all clear. If we are assuming the collection from an asynchronous 1559 // collection, clear the _modUnionTable. 1560 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1561 "_modUnionTable should be clear if the baton was not passed"); 1562 _modUnionTable.clear_all(); 1563 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1564 "mod union for klasses should be clear if the baton was passed"); 1565 _ct->cld_rem_set()->clear_mod_union(); 1566 1567 1568 // We must adjust the allocation statistics being maintained 1569 // in the free list space. We do so by reading and clearing 1570 // the sweep timer and updating the block flux rate estimates below. 1571 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1572 if (_inter_sweep_timer.is_active()) { 1573 _inter_sweep_timer.stop(); 1574 // Note that we do not use this sample to update the _inter_sweep_estimate. 1575 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1576 _inter_sweep_estimate.padded_average(), 1577 _intra_sweep_estimate.padded_average()); 1578 } 1579 1580 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1581 #ifdef ASSERT 1582 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1583 size_t free_size = cms_space->free(); 1584 assert(free_size == 1585 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1586 * HeapWordSize, 1587 "All the free space should be compacted into one chunk at top"); 1588 assert(cms_space->dictionary()->total_chunk_size( 1589 debug_only(cms_space->freelistLock())) == 0 || 1590 cms_space->totalSizeInIndexedFreeLists() == 0, 1591 "All the free space should be in a single chunk"); 1592 size_t num = cms_space->totalCount(); 1593 assert((free_size == 0 && num == 0) || 1594 (free_size > 0 && (num == 1 || num == 2)), 1595 "There should be at most 2 free chunks after compaction"); 1596 #endif // ASSERT 1597 _collectorState = Resetting; 1598 assert(_restart_addr == NULL, 1599 "Should have been NULL'd before baton was passed"); 1600 reset_stw(); 1601 _cmsGen->reset_after_compaction(); 1602 _concurrent_cycles_since_last_unload = 0; 1603 1604 // Clear any data recorded in the PLAB chunk arrays. 1605 if (_survivor_plab_array != NULL) { 1606 reset_survivor_plab_arrays(); 1607 } 1608 1609 // Adjust the per-size allocation stats for the next epoch. 1610 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1611 // Restart the "inter sweep timer" for the next epoch. 1612 _inter_sweep_timer.reset(); 1613 _inter_sweep_timer.start(); 1614 1615 // No longer a need to do a concurrent collection for Metaspace. 1616 MetaspaceGC::set_should_concurrent_collect(false); 1617 1618 heap->post_full_gc_dump(gc_timer); 1619 1620 gc_timer->register_gc_end(); 1621 1622 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1623 1624 // For a mark-sweep-compact, compute_new_size() will be called 1625 // in the heap's do_collection() method. 1626 } 1627 1628 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1629 Log(gc, heap) log; 1630 if (!log.is_trace()) { 1631 return; 1632 } 1633 1634 ContiguousSpace* eden_space = _young_gen->eden(); 1635 ContiguousSpace* from_space = _young_gen->from(); 1636 ContiguousSpace* to_space = _young_gen->to(); 1637 // Eden 1638 if (_eden_chunk_array != NULL) { 1639 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1640 p2i(eden_space->bottom()), p2i(eden_space->top()), 1641 p2i(eden_space->end()), eden_space->capacity()); 1642 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1643 _eden_chunk_index, _eden_chunk_capacity); 1644 for (size_t i = 0; i < _eden_chunk_index; i++) { 1645 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1646 } 1647 } 1648 // Survivor 1649 if (_survivor_chunk_array != NULL) { 1650 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1651 p2i(from_space->bottom()), p2i(from_space->top()), 1652 p2i(from_space->end()), from_space->capacity()); 1653 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1654 _survivor_chunk_index, _survivor_chunk_capacity); 1655 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1656 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1657 } 1658 } 1659 } 1660 1661 void CMSCollector::getFreelistLocks() const { 1662 // Get locks for all free lists in all generations that this 1663 // collector is responsible for 1664 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1665 } 1666 1667 void CMSCollector::releaseFreelistLocks() const { 1668 // Release locks for all free lists in all generations that this 1669 // collector is responsible for 1670 _cmsGen->freelistLock()->unlock(); 1671 } 1672 1673 bool CMSCollector::haveFreelistLocks() const { 1674 // Check locks for all free lists in all generations that this 1675 // collector is responsible for 1676 assert_lock_strong(_cmsGen->freelistLock()); 1677 PRODUCT_ONLY(ShouldNotReachHere()); 1678 return true; 1679 } 1680 1681 // A utility class that is used by the CMS collector to 1682 // temporarily "release" the foreground collector from its 1683 // usual obligation to wait for the background collector to 1684 // complete an ongoing phase before proceeding. 1685 class ReleaseForegroundGC: public StackObj { 1686 private: 1687 CMSCollector* _c; 1688 public: 1689 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1690 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1691 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1692 // allow a potentially blocked foreground collector to proceed 1693 _c->_foregroundGCShouldWait = false; 1694 if (_c->_foregroundGCIsActive) { 1695 CGC_lock->notify(); 1696 } 1697 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1698 "Possible deadlock"); 1699 } 1700 1701 ~ReleaseForegroundGC() { 1702 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1703 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1704 _c->_foregroundGCShouldWait = true; 1705 } 1706 }; 1707 1708 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1709 assert(Thread::current()->is_ConcurrentGC_thread(), 1710 "A CMS asynchronous collection is only allowed on a CMS thread."); 1711 1712 CMSHeap* heap = CMSHeap::heap(); 1713 { 1714 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1715 MutexLockerEx hl(Heap_lock, safepoint_check); 1716 FreelistLocker fll(this); 1717 MutexLockerEx x(CGC_lock, safepoint_check); 1718 if (_foregroundGCIsActive) { 1719 // The foreground collector is. Skip this 1720 // background collection. 1721 assert(!_foregroundGCShouldWait, "Should be clear"); 1722 return; 1723 } else { 1724 assert(_collectorState == Idling, "Should be idling before start."); 1725 _collectorState = InitialMarking; 1726 register_gc_start(cause); 1727 // Reset the expansion cause, now that we are about to begin 1728 // a new cycle. 1729 clear_expansion_cause(); 1730 1731 // Clear the MetaspaceGC flag since a concurrent collection 1732 // is starting but also clear it after the collection. 1733 MetaspaceGC::set_should_concurrent_collect(false); 1734 } 1735 // Decide if we want to enable class unloading as part of the 1736 // ensuing concurrent GC cycle. 1737 update_should_unload_classes(); 1738 _full_gc_requested = false; // acks all outstanding full gc requests 1739 _full_gc_cause = GCCause::_no_gc; 1740 // Signal that we are about to start a collection 1741 heap->increment_total_full_collections(); // ... starting a collection cycle 1742 _collection_count_start = heap->total_full_collections(); 1743 } 1744 1745 size_t prev_used = _cmsGen->used(); 1746 1747 // The change of the collection state is normally done at this level; 1748 // the exceptions are phases that are executed while the world is 1749 // stopped. For those phases the change of state is done while the 1750 // world is stopped. For baton passing purposes this allows the 1751 // background collector to finish the phase and change state atomically. 1752 // The foreground collector cannot wait on a phase that is done 1753 // while the world is stopped because the foreground collector already 1754 // has the world stopped and would deadlock. 1755 while (_collectorState != Idling) { 1756 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1757 p2i(Thread::current()), _collectorState); 1758 // The foreground collector 1759 // holds the Heap_lock throughout its collection. 1760 // holds the CMS token (but not the lock) 1761 // except while it is waiting for the background collector to yield. 1762 // 1763 // The foreground collector should be blocked (not for long) 1764 // if the background collector is about to start a phase 1765 // executed with world stopped. If the background 1766 // collector has already started such a phase, the 1767 // foreground collector is blocked waiting for the 1768 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1769 // are executed in the VM thread. 1770 // 1771 // The locking order is 1772 // PendingListLock (PLL) -- if applicable (FinalMarking) 1773 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1774 // CMS token (claimed in 1775 // stop_world_and_do() --> 1776 // safepoint_synchronize() --> 1777 // CMSThread::synchronize()) 1778 1779 { 1780 // Check if the FG collector wants us to yield. 1781 CMSTokenSync x(true); // is cms thread 1782 if (waitForForegroundGC()) { 1783 // We yielded to a foreground GC, nothing more to be 1784 // done this round. 1785 assert(_foregroundGCShouldWait == false, "We set it to false in " 1786 "waitForForegroundGC()"); 1787 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1788 p2i(Thread::current()), _collectorState); 1789 return; 1790 } else { 1791 // The background collector can run but check to see if the 1792 // foreground collector has done a collection while the 1793 // background collector was waiting to get the CGC_lock 1794 // above. If yes, break so that _foregroundGCShouldWait 1795 // is cleared before returning. 1796 if (_collectorState == Idling) { 1797 break; 1798 } 1799 } 1800 } 1801 1802 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1803 "should be waiting"); 1804 1805 switch (_collectorState) { 1806 case InitialMarking: 1807 { 1808 ReleaseForegroundGC x(this); 1809 stats().record_cms_begin(); 1810 VM_CMS_Initial_Mark initial_mark_op(this); 1811 VMThread::execute(&initial_mark_op); 1812 } 1813 // The collector state may be any legal state at this point 1814 // since the background collector may have yielded to the 1815 // foreground collector. 1816 break; 1817 case Marking: 1818 // initial marking in checkpointRootsInitialWork has been completed 1819 if (markFromRoots()) { // we were successful 1820 assert(_collectorState == Precleaning, "Collector state should " 1821 "have changed"); 1822 } else { 1823 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1824 } 1825 break; 1826 case Precleaning: 1827 // marking from roots in markFromRoots has been completed 1828 preclean(); 1829 assert(_collectorState == AbortablePreclean || 1830 _collectorState == FinalMarking, 1831 "Collector state should have changed"); 1832 break; 1833 case AbortablePreclean: 1834 abortable_preclean(); 1835 assert(_collectorState == FinalMarking, "Collector state should " 1836 "have changed"); 1837 break; 1838 case FinalMarking: 1839 { 1840 ReleaseForegroundGC x(this); 1841 1842 VM_CMS_Final_Remark final_remark_op(this); 1843 VMThread::execute(&final_remark_op); 1844 } 1845 assert(_foregroundGCShouldWait, "block post-condition"); 1846 break; 1847 case Sweeping: 1848 // final marking in checkpointRootsFinal has been completed 1849 sweep(); 1850 assert(_collectorState == Resizing, "Collector state change " 1851 "to Resizing must be done under the free_list_lock"); 1852 1853 case Resizing: { 1854 // Sweeping has been completed... 1855 // At this point the background collection has completed. 1856 // Don't move the call to compute_new_size() down 1857 // into code that might be executed if the background 1858 // collection was preempted. 1859 { 1860 ReleaseForegroundGC x(this); // unblock FG collection 1861 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1862 CMSTokenSync z(true); // not strictly needed. 1863 if (_collectorState == Resizing) { 1864 compute_new_size(); 1865 save_heap_summary(); 1866 _collectorState = Resetting; 1867 } else { 1868 assert(_collectorState == Idling, "The state should only change" 1869 " because the foreground collector has finished the collection"); 1870 } 1871 } 1872 break; 1873 } 1874 case Resetting: 1875 // CMS heap resizing has been completed 1876 reset_concurrent(); 1877 assert(_collectorState == Idling, "Collector state should " 1878 "have changed"); 1879 1880 MetaspaceGC::set_should_concurrent_collect(false); 1881 1882 stats().record_cms_end(); 1883 // Don't move the concurrent_phases_end() and compute_new_size() 1884 // calls to here because a preempted background collection 1885 // has it's state set to "Resetting". 1886 break; 1887 case Idling: 1888 default: 1889 ShouldNotReachHere(); 1890 break; 1891 } 1892 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1893 p2i(Thread::current()), _collectorState); 1894 assert(_foregroundGCShouldWait, "block post-condition"); 1895 } 1896 1897 // Should this be in gc_epilogue? 1898 heap->counters()->update_counters(); 1899 1900 { 1901 // Clear _foregroundGCShouldWait and, in the event that the 1902 // foreground collector is waiting, notify it, before 1903 // returning. 1904 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1905 _foregroundGCShouldWait = false; 1906 if (_foregroundGCIsActive) { 1907 CGC_lock->notify(); 1908 } 1909 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1910 "Possible deadlock"); 1911 } 1912 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1913 p2i(Thread::current()), _collectorState); 1914 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1915 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1916 } 1917 1918 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1919 _cms_start_registered = true; 1920 _gc_timer_cm->register_gc_start(); 1921 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1922 } 1923 1924 void CMSCollector::register_gc_end() { 1925 if (_cms_start_registered) { 1926 report_heap_summary(GCWhen::AfterGC); 1927 1928 _gc_timer_cm->register_gc_end(); 1929 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1930 _cms_start_registered = false; 1931 } 1932 } 1933 1934 void CMSCollector::save_heap_summary() { 1935 CMSHeap* heap = CMSHeap::heap(); 1936 _last_heap_summary = heap->create_heap_summary(); 1937 _last_metaspace_summary = heap->create_metaspace_summary(); 1938 } 1939 1940 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1941 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1942 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1943 } 1944 1945 bool CMSCollector::waitForForegroundGC() { 1946 bool res = false; 1947 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1948 "CMS thread should have CMS token"); 1949 // Block the foreground collector until the 1950 // background collectors decides whether to 1951 // yield. 1952 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1953 _foregroundGCShouldWait = true; 1954 if (_foregroundGCIsActive) { 1955 // The background collector yields to the 1956 // foreground collector and returns a value 1957 // indicating that it has yielded. The foreground 1958 // collector can proceed. 1959 res = true; 1960 _foregroundGCShouldWait = false; 1961 ConcurrentMarkSweepThread::clear_CMS_flag( 1962 ConcurrentMarkSweepThread::CMS_cms_has_token); 1963 ConcurrentMarkSweepThread::set_CMS_flag( 1964 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1965 // Get a possibly blocked foreground thread going 1966 CGC_lock->notify(); 1967 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1968 p2i(Thread::current()), _collectorState); 1969 while (_foregroundGCIsActive) { 1970 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1971 } 1972 ConcurrentMarkSweepThread::set_CMS_flag( 1973 ConcurrentMarkSweepThread::CMS_cms_has_token); 1974 ConcurrentMarkSweepThread::clear_CMS_flag( 1975 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1976 } 1977 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1978 p2i(Thread::current()), _collectorState); 1979 return res; 1980 } 1981 1982 // Because of the need to lock the free lists and other structures in 1983 // the collector, common to all the generations that the collector is 1984 // collecting, we need the gc_prologues of individual CMS generations 1985 // delegate to their collector. It may have been simpler had the 1986 // current infrastructure allowed one to call a prologue on a 1987 // collector. In the absence of that we have the generation's 1988 // prologue delegate to the collector, which delegates back 1989 // some "local" work to a worker method in the individual generations 1990 // that it's responsible for collecting, while itself doing any 1991 // work common to all generations it's responsible for. A similar 1992 // comment applies to the gc_epilogue()'s. 1993 // The role of the variable _between_prologue_and_epilogue is to 1994 // enforce the invocation protocol. 1995 void CMSCollector::gc_prologue(bool full) { 1996 // Call gc_prologue_work() for the CMSGen 1997 // we are responsible for. 1998 1999 // The following locking discipline assumes that we are only called 2000 // when the world is stopped. 2001 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2002 2003 // The CMSCollector prologue must call the gc_prologues for the 2004 // "generations" that it's responsible 2005 // for. 2006 2007 assert( Thread::current()->is_VM_thread() 2008 || ( CMSScavengeBeforeRemark 2009 && Thread::current()->is_ConcurrentGC_thread()), 2010 "Incorrect thread type for prologue execution"); 2011 2012 if (_between_prologue_and_epilogue) { 2013 // We have already been invoked; this is a gc_prologue delegation 2014 // from yet another CMS generation that we are responsible for, just 2015 // ignore it since all relevant work has already been done. 2016 return; 2017 } 2018 2019 // set a bit saying prologue has been called; cleared in epilogue 2020 _between_prologue_and_epilogue = true; 2021 // Claim locks for common data structures, then call gc_prologue_work() 2022 // for each CMSGen. 2023 2024 getFreelistLocks(); // gets free list locks on constituent spaces 2025 bitMapLock()->lock_without_safepoint_check(); 2026 2027 // Should call gc_prologue_work() for all cms gens we are responsible for 2028 bool duringMarking = _collectorState >= Marking 2029 && _collectorState < Sweeping; 2030 2031 // The young collections clear the modified oops state, which tells if 2032 // there are any modified oops in the class. The remark phase also needs 2033 // that information. Tell the young collection to save the union of all 2034 // modified klasses. 2035 if (duringMarking) { 2036 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2037 } 2038 2039 bool registerClosure = duringMarking; 2040 2041 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2042 2043 if (!full) { 2044 stats().record_gc0_begin(); 2045 } 2046 } 2047 2048 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2049 2050 _capacity_at_prologue = capacity(); 2051 _used_at_prologue = used(); 2052 2053 // We enable promotion tracking so that card-scanning can recognize 2054 // which objects have been promoted during this GC and skip them. 2055 for (uint i = 0; i < ParallelGCThreads; i++) { 2056 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2057 } 2058 2059 // Delegate to CMScollector which knows how to coordinate between 2060 // this and any other CMS generations that it is responsible for 2061 // collecting. 2062 collector()->gc_prologue(full); 2063 } 2064 2065 // This is a "private" interface for use by this generation's CMSCollector. 2066 // Not to be called directly by any other entity (for instance, 2067 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2068 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2069 bool registerClosure, ModUnionClosure* modUnionClosure) { 2070 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2071 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2072 "Should be NULL"); 2073 if (registerClosure) { 2074 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2075 } 2076 cmsSpace()->gc_prologue(); 2077 // Clear stat counters 2078 NOT_PRODUCT( 2079 assert(_numObjectsPromoted == 0, "check"); 2080 assert(_numWordsPromoted == 0, "check"); 2081 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2082 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2083 _numObjectsAllocated = 0; 2084 _numWordsAllocated = 0; 2085 ) 2086 } 2087 2088 void CMSCollector::gc_epilogue(bool full) { 2089 // The following locking discipline assumes that we are only called 2090 // when the world is stopped. 2091 assert(SafepointSynchronize::is_at_safepoint(), 2092 "world is stopped assumption"); 2093 2094 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2095 // if linear allocation blocks need to be appropriately marked to allow the 2096 // the blocks to be parsable. We also check here whether we need to nudge the 2097 // CMS collector thread to start a new cycle (if it's not already active). 2098 assert( Thread::current()->is_VM_thread() 2099 || ( CMSScavengeBeforeRemark 2100 && Thread::current()->is_ConcurrentGC_thread()), 2101 "Incorrect thread type for epilogue execution"); 2102 2103 if (!_between_prologue_and_epilogue) { 2104 // We have already been invoked; this is a gc_epilogue delegation 2105 // from yet another CMS generation that we are responsible for, just 2106 // ignore it since all relevant work has already been done. 2107 return; 2108 } 2109 assert(haveFreelistLocks(), "must have freelist locks"); 2110 assert_lock_strong(bitMapLock()); 2111 2112 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2113 2114 _cmsGen->gc_epilogue_work(full); 2115 2116 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2117 // in case sampling was not already enabled, enable it 2118 _start_sampling = true; 2119 } 2120 // reset _eden_chunk_array so sampling starts afresh 2121 _eden_chunk_index = 0; 2122 2123 size_t cms_used = _cmsGen->cmsSpace()->used(); 2124 2125 // update performance counters - this uses a special version of 2126 // update_counters() that allows the utilization to be passed as a 2127 // parameter, avoiding multiple calls to used(). 2128 // 2129 _cmsGen->update_counters(cms_used); 2130 2131 bitMapLock()->unlock(); 2132 releaseFreelistLocks(); 2133 2134 if (!CleanChunkPoolAsync) { 2135 Chunk::clean_chunk_pool(); 2136 } 2137 2138 set_did_compact(false); 2139 _between_prologue_and_epilogue = false; // ready for next cycle 2140 } 2141 2142 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2143 collector()->gc_epilogue(full); 2144 2145 // When using ParNew, promotion tracking should have already been 2146 // disabled. However, the prologue (which enables promotion 2147 // tracking) and epilogue are called irrespective of the type of 2148 // GC. So they will also be called before and after Full GCs, during 2149 // which promotion tracking will not be explicitly disabled. So, 2150 // it's safer to also disable it here too (to be symmetric with 2151 // enabling it in the prologue). 2152 for (uint i = 0; i < ParallelGCThreads; i++) { 2153 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2154 } 2155 } 2156 2157 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2158 assert(!incremental_collection_failed(), "Should have been cleared"); 2159 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2160 cmsSpace()->gc_epilogue(); 2161 // Print stat counters 2162 NOT_PRODUCT( 2163 assert(_numObjectsAllocated == 0, "check"); 2164 assert(_numWordsAllocated == 0, "check"); 2165 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2166 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2167 _numObjectsPromoted = 0; 2168 _numWordsPromoted = 0; 2169 ) 2170 2171 // Call down the chain in contiguous_available needs the freelistLock 2172 // so print this out before releasing the freeListLock. 2173 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2174 } 2175 2176 #ifndef PRODUCT 2177 bool CMSCollector::have_cms_token() { 2178 Thread* thr = Thread::current(); 2179 if (thr->is_VM_thread()) { 2180 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2181 } else if (thr->is_ConcurrentGC_thread()) { 2182 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2183 } else if (thr->is_GC_task_thread()) { 2184 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2185 ParGCRareEvent_lock->owned_by_self(); 2186 } 2187 return false; 2188 } 2189 2190 // Check reachability of the given heap address in CMS generation, 2191 // treating all other generations as roots. 2192 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2193 // We could "guarantee" below, rather than assert, but I'll 2194 // leave these as "asserts" so that an adventurous debugger 2195 // could try this in the product build provided some subset of 2196 // the conditions were met, provided they were interested in the 2197 // results and knew that the computation below wouldn't interfere 2198 // with other concurrent computations mutating the structures 2199 // being read or written. 2200 assert(SafepointSynchronize::is_at_safepoint(), 2201 "Else mutations in object graph will make answer suspect"); 2202 assert(have_cms_token(), "Should hold cms token"); 2203 assert(haveFreelistLocks(), "must hold free list locks"); 2204 assert_lock_strong(bitMapLock()); 2205 2206 // Clear the marking bit map array before starting, but, just 2207 // for kicks, first report if the given address is already marked 2208 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2209 _markBitMap.isMarked(addr) ? "" : " not"); 2210 2211 if (verify_after_remark()) { 2212 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2213 bool result = verification_mark_bm()->isMarked(addr); 2214 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2215 result ? "IS" : "is NOT"); 2216 return result; 2217 } else { 2218 tty->print_cr("Could not compute result"); 2219 return false; 2220 } 2221 } 2222 #endif 2223 2224 void 2225 CMSCollector::print_on_error(outputStream* st) { 2226 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2227 if (collector != NULL) { 2228 CMSBitMap* bitmap = &collector->_markBitMap; 2229 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2230 bitmap->print_on_error(st, " Bits: "); 2231 2232 st->cr(); 2233 2234 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2235 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2236 mut_bitmap->print_on_error(st, " Bits: "); 2237 } 2238 } 2239 2240 //////////////////////////////////////////////////////// 2241 // CMS Verification Support 2242 //////////////////////////////////////////////////////// 2243 // Following the remark phase, the following invariant 2244 // should hold -- each object in the CMS heap which is 2245 // marked in markBitMap() should be marked in the verification_mark_bm(). 2246 2247 class VerifyMarkedClosure: public BitMapClosure { 2248 CMSBitMap* _marks; 2249 bool _failed; 2250 2251 public: 2252 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2253 2254 bool do_bit(size_t offset) { 2255 HeapWord* addr = _marks->offsetToHeapWord(offset); 2256 if (!_marks->isMarked(addr)) { 2257 Log(gc, verify) log; 2258 ResourceMark rm; 2259 LogStream ls(log.error()); 2260 oop(addr)->print_on(&ls); 2261 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2262 _failed = true; 2263 } 2264 return true; 2265 } 2266 2267 bool failed() { return _failed; } 2268 }; 2269 2270 bool CMSCollector::verify_after_remark() { 2271 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2272 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2273 static bool init = false; 2274 2275 assert(SafepointSynchronize::is_at_safepoint(), 2276 "Else mutations in object graph will make answer suspect"); 2277 assert(have_cms_token(), 2278 "Else there may be mutual interference in use of " 2279 " verification data structures"); 2280 assert(_collectorState > Marking && _collectorState <= Sweeping, 2281 "Else marking info checked here may be obsolete"); 2282 assert(haveFreelistLocks(), "must hold free list locks"); 2283 assert_lock_strong(bitMapLock()); 2284 2285 2286 // Allocate marking bit map if not already allocated 2287 if (!init) { // first time 2288 if (!verification_mark_bm()->allocate(_span)) { 2289 return false; 2290 } 2291 init = true; 2292 } 2293 2294 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2295 2296 // Turn off refs discovery -- so we will be tracing through refs. 2297 // This is as intended, because by this time 2298 // GC must already have cleared any refs that need to be cleared, 2299 // and traced those that need to be marked; moreover, 2300 // the marking done here is not going to interfere in any 2301 // way with the marking information used by GC. 2302 NoRefDiscovery no_discovery(ref_processor()); 2303 2304 #if COMPILER2_OR_JVMCI 2305 DerivedPointerTableDeactivate dpt_deact; 2306 #endif 2307 2308 // Clear any marks from a previous round 2309 verification_mark_bm()->clear_all(); 2310 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2311 verify_work_stacks_empty(); 2312 2313 CMSHeap* heap = CMSHeap::heap(); 2314 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2315 // Update the saved marks which may affect the root scans. 2316 heap->save_marks(); 2317 2318 if (CMSRemarkVerifyVariant == 1) { 2319 // In this first variant of verification, we complete 2320 // all marking, then check if the new marks-vector is 2321 // a subset of the CMS marks-vector. 2322 verify_after_remark_work_1(); 2323 } else { 2324 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2325 // In this second variant of verification, we flag an error 2326 // (i.e. an object reachable in the new marks-vector not reachable 2327 // in the CMS marks-vector) immediately, also indicating the 2328 // identify of an object (A) that references the unmarked object (B) -- 2329 // presumably, a mutation to A failed to be picked up by preclean/remark? 2330 verify_after_remark_work_2(); 2331 } 2332 2333 return true; 2334 } 2335 2336 void CMSCollector::verify_after_remark_work_1() { 2337 ResourceMark rm; 2338 HandleMark hm; 2339 CMSHeap* heap = CMSHeap::heap(); 2340 2341 // Get a clear set of claim bits for the roots processing to work with. 2342 ClassLoaderDataGraph::clear_claimed_marks(); 2343 2344 // Mark from roots one level into CMS 2345 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2346 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2347 2348 { 2349 StrongRootsScope srs(1); 2350 2351 heap->cms_process_roots(&srs, 2352 true, // young gen as roots 2353 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2354 should_unload_classes(), 2355 ¬Older, 2356 NULL); 2357 } 2358 2359 // Now mark from the roots 2360 MarkFromRootsClosure markFromRootsClosure(this, _span, 2361 verification_mark_bm(), verification_mark_stack(), 2362 false /* don't yield */, true /* verifying */); 2363 assert(_restart_addr == NULL, "Expected pre-condition"); 2364 verification_mark_bm()->iterate(&markFromRootsClosure); 2365 while (_restart_addr != NULL) { 2366 // Deal with stack overflow: by restarting at the indicated 2367 // address. 2368 HeapWord* ra = _restart_addr; 2369 markFromRootsClosure.reset(ra); 2370 _restart_addr = NULL; 2371 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2372 } 2373 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2374 verify_work_stacks_empty(); 2375 2376 // Marking completed -- now verify that each bit marked in 2377 // verification_mark_bm() is also marked in markBitMap(); flag all 2378 // errors by printing corresponding objects. 2379 VerifyMarkedClosure vcl(markBitMap()); 2380 verification_mark_bm()->iterate(&vcl); 2381 if (vcl.failed()) { 2382 Log(gc, verify) log; 2383 log.error("Failed marking verification after remark"); 2384 ResourceMark rm; 2385 LogStream ls(log.error()); 2386 heap->print_on(&ls); 2387 fatal("CMS: failed marking verification after remark"); 2388 } 2389 } 2390 2391 class VerifyCLDOopsCLDClosure : public CLDClosure { 2392 class VerifyCLDOopsClosure : public OopClosure { 2393 CMSBitMap* _bitmap; 2394 public: 2395 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2396 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2397 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2398 } _oop_closure; 2399 public: 2400 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2401 void do_cld(ClassLoaderData* cld) { 2402 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2403 } 2404 }; 2405 2406 void CMSCollector::verify_after_remark_work_2() { 2407 ResourceMark rm; 2408 HandleMark hm; 2409 CMSHeap* heap = CMSHeap::heap(); 2410 2411 // Get a clear set of claim bits for the roots processing to work with. 2412 ClassLoaderDataGraph::clear_claimed_marks(); 2413 2414 // Mark from roots one level into CMS 2415 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2416 markBitMap()); 2417 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2418 2419 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2420 2421 { 2422 StrongRootsScope srs(1); 2423 2424 heap->cms_process_roots(&srs, 2425 true, // young gen as roots 2426 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2427 should_unload_classes(), 2428 ¬Older, 2429 &cld_closure); 2430 } 2431 2432 // Now mark from the roots 2433 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2434 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2435 assert(_restart_addr == NULL, "Expected pre-condition"); 2436 verification_mark_bm()->iterate(&markFromRootsClosure); 2437 while (_restart_addr != NULL) { 2438 // Deal with stack overflow: by restarting at the indicated 2439 // address. 2440 HeapWord* ra = _restart_addr; 2441 markFromRootsClosure.reset(ra); 2442 _restart_addr = NULL; 2443 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2444 } 2445 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2446 verify_work_stacks_empty(); 2447 2448 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2449 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2450 2451 // Marking completed -- now verify that each bit marked in 2452 // verification_mark_bm() is also marked in markBitMap(); flag all 2453 // errors by printing corresponding objects. 2454 VerifyMarkedClosure vcl(markBitMap()); 2455 verification_mark_bm()->iterate(&vcl); 2456 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2457 } 2458 2459 void ConcurrentMarkSweepGeneration::save_marks() { 2460 // delegate to CMS space 2461 cmsSpace()->save_marks(); 2462 } 2463 2464 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2465 return cmsSpace()->no_allocs_since_save_marks(); 2466 } 2467 2468 void 2469 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2470 if (freelistLock()->owned_by_self()) { 2471 Generation::oop_iterate(cl); 2472 } else { 2473 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2474 Generation::oop_iterate(cl); 2475 } 2476 } 2477 2478 void 2479 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2480 if (freelistLock()->owned_by_self()) { 2481 Generation::object_iterate(cl); 2482 } else { 2483 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2484 Generation::object_iterate(cl); 2485 } 2486 } 2487 2488 void 2489 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2490 if (freelistLock()->owned_by_self()) { 2491 Generation::safe_object_iterate(cl); 2492 } else { 2493 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2494 Generation::safe_object_iterate(cl); 2495 } 2496 } 2497 2498 void 2499 ConcurrentMarkSweepGeneration::post_compact() { 2500 } 2501 2502 void 2503 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2504 // Fix the linear allocation blocks to look like free blocks. 2505 2506 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2507 // are not called when the heap is verified during universe initialization and 2508 // at vm shutdown. 2509 if (freelistLock()->owned_by_self()) { 2510 cmsSpace()->prepare_for_verify(); 2511 } else { 2512 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2513 cmsSpace()->prepare_for_verify(); 2514 } 2515 } 2516 2517 void 2518 ConcurrentMarkSweepGeneration::verify() { 2519 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2520 // are not called when the heap is verified during universe initialization and 2521 // at vm shutdown. 2522 if (freelistLock()->owned_by_self()) { 2523 cmsSpace()->verify(); 2524 } else { 2525 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2526 cmsSpace()->verify(); 2527 } 2528 } 2529 2530 void CMSCollector::verify() { 2531 _cmsGen->verify(); 2532 } 2533 2534 #ifndef PRODUCT 2535 bool CMSCollector::overflow_list_is_empty() const { 2536 assert(_num_par_pushes >= 0, "Inconsistency"); 2537 if (_overflow_list == NULL) { 2538 assert(_num_par_pushes == 0, "Inconsistency"); 2539 } 2540 return _overflow_list == NULL; 2541 } 2542 2543 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2544 // merely consolidate assertion checks that appear to occur together frequently. 2545 void CMSCollector::verify_work_stacks_empty() const { 2546 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2547 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2548 } 2549 2550 void CMSCollector::verify_overflow_empty() const { 2551 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2552 assert(no_preserved_marks(), "No preserved marks"); 2553 } 2554 #endif // PRODUCT 2555 2556 // Decide if we want to enable class unloading as part of the 2557 // ensuing concurrent GC cycle. We will collect and 2558 // unload classes if it's the case that: 2559 // (a) class unloading is enabled at the command line, and 2560 // (b) old gen is getting really full 2561 // NOTE: Provided there is no change in the state of the heap between 2562 // calls to this method, it should have idempotent results. Moreover, 2563 // its results should be monotonically increasing (i.e. going from 0 to 1, 2564 // but not 1 to 0) between successive calls between which the heap was 2565 // not collected. For the implementation below, it must thus rely on 2566 // the property that concurrent_cycles_since_last_unload() 2567 // will not decrease unless a collection cycle happened and that 2568 // _cmsGen->is_too_full() are 2569 // themselves also monotonic in that sense. See check_monotonicity() 2570 // below. 2571 void CMSCollector::update_should_unload_classes() { 2572 _should_unload_classes = false; 2573 if (CMSClassUnloadingEnabled) { 2574 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2575 CMSClassUnloadingMaxInterval) 2576 || _cmsGen->is_too_full(); 2577 } 2578 } 2579 2580 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2581 bool res = should_concurrent_collect(); 2582 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2583 return res; 2584 } 2585 2586 void CMSCollector::setup_cms_unloading_and_verification_state() { 2587 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2588 || VerifyBeforeExit; 2589 const int rso = GenCollectedHeap::SO_AllCodeCache; 2590 2591 // We set the proper root for this CMS cycle here. 2592 if (should_unload_classes()) { // Should unload classes this cycle 2593 remove_root_scanning_option(rso); // Shrink the root set appropriately 2594 set_verifying(should_verify); // Set verification state for this cycle 2595 return; // Nothing else needs to be done at this time 2596 } 2597 2598 // Not unloading classes this cycle 2599 assert(!should_unload_classes(), "Inconsistency!"); 2600 2601 // If we are not unloading classes then add SO_AllCodeCache to root 2602 // scanning options. 2603 add_root_scanning_option(rso); 2604 2605 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2606 set_verifying(true); 2607 } else if (verifying() && !should_verify) { 2608 // We were verifying, but some verification flags got disabled. 2609 set_verifying(false); 2610 // Exclude symbols, strings and code cache elements from root scanning to 2611 // reduce IM and RM pauses. 2612 remove_root_scanning_option(rso); 2613 } 2614 } 2615 2616 2617 #ifndef PRODUCT 2618 HeapWord* CMSCollector::block_start(const void* p) const { 2619 const HeapWord* addr = (HeapWord*)p; 2620 if (_span.contains(p)) { 2621 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2622 return _cmsGen->cmsSpace()->block_start(p); 2623 } 2624 } 2625 return NULL; 2626 } 2627 #endif 2628 2629 HeapWord* 2630 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2631 bool tlab, 2632 bool parallel) { 2633 CMSSynchronousYieldRequest yr; 2634 assert(!tlab, "Can't deal with TLAB allocation"); 2635 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2636 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2637 if (GCExpandToAllocateDelayMillis > 0) { 2638 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2639 } 2640 return have_lock_and_allocate(word_size, tlab); 2641 } 2642 2643 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2644 size_t bytes, 2645 size_t expand_bytes, 2646 CMSExpansionCause::Cause cause) 2647 { 2648 2649 bool success = expand(bytes, expand_bytes); 2650 2651 // remember why we expanded; this information is used 2652 // by shouldConcurrentCollect() when making decisions on whether to start 2653 // a new CMS cycle. 2654 if (success) { 2655 set_expansion_cause(cause); 2656 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2657 } 2658 } 2659 2660 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2661 HeapWord* res = NULL; 2662 MutexLocker x(ParGCRareEvent_lock); 2663 while (true) { 2664 // Expansion by some other thread might make alloc OK now: 2665 res = ps->lab.alloc(word_sz); 2666 if (res != NULL) return res; 2667 // If there's not enough expansion space available, give up. 2668 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2669 return NULL; 2670 } 2671 // Otherwise, we try expansion. 2672 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2673 // Now go around the loop and try alloc again; 2674 // A competing par_promote might beat us to the expansion space, 2675 // so we may go around the loop again if promotion fails again. 2676 if (GCExpandToAllocateDelayMillis > 0) { 2677 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2678 } 2679 } 2680 } 2681 2682 2683 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2684 PromotionInfo* promo) { 2685 MutexLocker x(ParGCRareEvent_lock); 2686 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2687 while (true) { 2688 // Expansion by some other thread might make alloc OK now: 2689 if (promo->ensure_spooling_space()) { 2690 assert(promo->has_spooling_space(), 2691 "Post-condition of successful ensure_spooling_space()"); 2692 return true; 2693 } 2694 // If there's not enough expansion space available, give up. 2695 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2696 return false; 2697 } 2698 // Otherwise, we try expansion. 2699 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2700 // Now go around the loop and try alloc again; 2701 // A competing allocation might beat us to the expansion space, 2702 // so we may go around the loop again if allocation fails again. 2703 if (GCExpandToAllocateDelayMillis > 0) { 2704 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2705 } 2706 } 2707 } 2708 2709 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2710 // Only shrink if a compaction was done so that all the free space 2711 // in the generation is in a contiguous block at the end. 2712 if (did_compact()) { 2713 CardGeneration::shrink(bytes); 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2718 assert_locked_or_safepoint(Heap_lock); 2719 } 2720 2721 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2722 assert_locked_or_safepoint(Heap_lock); 2723 assert_lock_strong(freelistLock()); 2724 log_trace(gc)("Shrinking of CMS not yet implemented"); 2725 return; 2726 } 2727 2728 2729 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2730 // phases. 2731 class CMSPhaseAccounting: public StackObj { 2732 public: 2733 CMSPhaseAccounting(CMSCollector *collector, 2734 const char *title); 2735 ~CMSPhaseAccounting(); 2736 2737 private: 2738 CMSCollector *_collector; 2739 const char *_title; 2740 GCTraceConcTime(Info, gc) _trace_time; 2741 2742 public: 2743 // Not MT-safe; so do not pass around these StackObj's 2744 // where they may be accessed by other threads. 2745 double wallclock_millis() { 2746 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2747 } 2748 }; 2749 2750 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2751 const char *title) : 2752 _collector(collector), _title(title), _trace_time(title) { 2753 2754 _collector->resetYields(); 2755 _collector->resetTimer(); 2756 _collector->startTimer(); 2757 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2758 } 2759 2760 CMSPhaseAccounting::~CMSPhaseAccounting() { 2761 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2762 _collector->stopTimer(); 2763 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2764 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2765 } 2766 2767 // CMS work 2768 2769 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2770 class CMSParMarkTask : public AbstractGangTask { 2771 protected: 2772 CMSCollector* _collector; 2773 uint _n_workers; 2774 OopStorage::ParState<false, false> _par_state_string; 2775 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2776 AbstractGangTask(name), 2777 _collector(collector), 2778 _n_workers(n_workers), 2779 _par_state_string(StringTable::weak_storage()) {} 2780 // Work method in support of parallel rescan ... of young gen spaces 2781 void do_young_space_rescan(OopsInGenClosure* cl, 2782 ContiguousSpace* space, 2783 HeapWord** chunk_array, size_t chunk_top); 2784 void work_on_young_gen_roots(OopsInGenClosure* cl); 2785 }; 2786 2787 // Parallel initial mark task 2788 class CMSParInitialMarkTask: public CMSParMarkTask { 2789 StrongRootsScope* _strong_roots_scope; 2790 public: 2791 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2792 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2793 _strong_roots_scope(strong_roots_scope) {} 2794 void work(uint worker_id); 2795 }; 2796 2797 // Checkpoint the roots into this generation from outside 2798 // this generation. [Note this initial checkpoint need only 2799 // be approximate -- we'll do a catch up phase subsequently.] 2800 void CMSCollector::checkpointRootsInitial() { 2801 assert(_collectorState == InitialMarking, "Wrong collector state"); 2802 check_correct_thread_executing(); 2803 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2804 2805 save_heap_summary(); 2806 report_heap_summary(GCWhen::BeforeGC); 2807 2808 ReferenceProcessor* rp = ref_processor(); 2809 assert(_restart_addr == NULL, "Control point invariant"); 2810 { 2811 // acquire locks for subsequent manipulations 2812 MutexLockerEx x(bitMapLock(), 2813 Mutex::_no_safepoint_check_flag); 2814 checkpointRootsInitialWork(); 2815 // enable ("weak") refs discovery 2816 rp->enable_discovery(); 2817 _collectorState = Marking; 2818 } 2819 } 2820 2821 void CMSCollector::checkpointRootsInitialWork() { 2822 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2823 assert(_collectorState == InitialMarking, "just checking"); 2824 2825 // Already have locks. 2826 assert_lock_strong(bitMapLock()); 2827 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2828 2829 // Setup the verification and class unloading state for this 2830 // CMS collection cycle. 2831 setup_cms_unloading_and_verification_state(); 2832 2833 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2834 2835 // Reset all the PLAB chunk arrays if necessary. 2836 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2837 reset_survivor_plab_arrays(); 2838 } 2839 2840 ResourceMark rm; 2841 HandleMark hm; 2842 2843 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2844 CMSHeap* heap = CMSHeap::heap(); 2845 2846 verify_work_stacks_empty(); 2847 verify_overflow_empty(); 2848 2849 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2850 // Update the saved marks which may affect the root scans. 2851 heap->save_marks(); 2852 2853 // weak reference processing has not started yet. 2854 ref_processor()->set_enqueuing_is_done(false); 2855 2856 // Need to remember all newly created CLDs, 2857 // so that we can guarantee that the remark finds them. 2858 ClassLoaderDataGraph::remember_new_clds(true); 2859 2860 // Whenever a CLD is found, it will be claimed before proceeding to mark 2861 // the klasses. The claimed marks need to be cleared before marking starts. 2862 ClassLoaderDataGraph::clear_claimed_marks(); 2863 2864 print_eden_and_survivor_chunk_arrays(); 2865 2866 { 2867 #if COMPILER2_OR_JVMCI 2868 DerivedPointerTableDeactivate dpt_deact; 2869 #endif 2870 if (CMSParallelInitialMarkEnabled) { 2871 // The parallel version. 2872 WorkGang* workers = heap->workers(); 2873 assert(workers != NULL, "Need parallel worker threads."); 2874 uint n_workers = workers->active_workers(); 2875 2876 StrongRootsScope srs(n_workers); 2877 2878 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2879 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2880 // If the total workers is greater than 1, then multiple workers 2881 // may be used at some time and the initialization has been set 2882 // such that the single threaded path cannot be used. 2883 if (workers->total_workers() > 1) { 2884 workers->run_task(&tsk); 2885 } else { 2886 tsk.work(0); 2887 } 2888 } else { 2889 // The serial version. 2890 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2891 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2892 2893 StrongRootsScope srs(1); 2894 2895 heap->cms_process_roots(&srs, 2896 true, // young gen as roots 2897 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2898 should_unload_classes(), 2899 ¬Older, 2900 &cld_closure); 2901 } 2902 } 2903 2904 // Clear mod-union table; it will be dirtied in the prologue of 2905 // CMS generation per each young generation collection. 2906 2907 assert(_modUnionTable.isAllClear(), 2908 "Was cleared in most recent final checkpoint phase" 2909 " or no bits are set in the gc_prologue before the start of the next " 2910 "subsequent marking phase."); 2911 2912 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2913 2914 // Save the end of the used_region of the constituent generations 2915 // to be used to limit the extent of sweep in each generation. 2916 save_sweep_limits(); 2917 verify_overflow_empty(); 2918 } 2919 2920 bool CMSCollector::markFromRoots() { 2921 // we might be tempted to assert that: 2922 // assert(!SafepointSynchronize::is_at_safepoint(), 2923 // "inconsistent argument?"); 2924 // However that wouldn't be right, because it's possible that 2925 // a safepoint is indeed in progress as a young generation 2926 // stop-the-world GC happens even as we mark in this generation. 2927 assert(_collectorState == Marking, "inconsistent state?"); 2928 check_correct_thread_executing(); 2929 verify_overflow_empty(); 2930 2931 // Weak ref discovery note: We may be discovering weak 2932 // refs in this generation concurrent (but interleaved) with 2933 // weak ref discovery by the young generation collector. 2934 2935 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2936 GCTraceCPUTime tcpu; 2937 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2938 bool res = markFromRootsWork(); 2939 if (res) { 2940 _collectorState = Precleaning; 2941 } else { // We failed and a foreground collection wants to take over 2942 assert(_foregroundGCIsActive, "internal state inconsistency"); 2943 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2944 log_debug(gc)("bailing out to foreground collection"); 2945 } 2946 verify_overflow_empty(); 2947 return res; 2948 } 2949 2950 bool CMSCollector::markFromRootsWork() { 2951 // iterate over marked bits in bit map, doing a full scan and mark 2952 // from these roots using the following algorithm: 2953 // . if oop is to the right of the current scan pointer, 2954 // mark corresponding bit (we'll process it later) 2955 // . else (oop is to left of current scan pointer) 2956 // push oop on marking stack 2957 // . drain the marking stack 2958 2959 // Note that when we do a marking step we need to hold the 2960 // bit map lock -- recall that direct allocation (by mutators) 2961 // and promotion (by the young generation collector) is also 2962 // marking the bit map. [the so-called allocate live policy.] 2963 // Because the implementation of bit map marking is not 2964 // robust wrt simultaneous marking of bits in the same word, 2965 // we need to make sure that there is no such interference 2966 // between concurrent such updates. 2967 2968 // already have locks 2969 assert_lock_strong(bitMapLock()); 2970 2971 verify_work_stacks_empty(); 2972 verify_overflow_empty(); 2973 bool result = false; 2974 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2975 result = do_marking_mt(); 2976 } else { 2977 result = do_marking_st(); 2978 } 2979 return result; 2980 } 2981 2982 // Forward decl 2983 class CMSConcMarkingTask; 2984 2985 class CMSConcMarkingParallelTerminator: public ParallelTaskTerminator { 2986 CMSCollector* _collector; 2987 CMSConcMarkingTask* _task; 2988 public: 2989 virtual void yield(); 2990 2991 // "n_threads" is the number of threads to be terminated. 2992 // "queue_set" is a set of work queues of other threads. 2993 // "collector" is the CMS collector associated with this task terminator. 2994 // "yield" indicates whether we need the gang as a whole to yield. 2995 CMSConcMarkingParallelTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2996 ParallelTaskTerminator(n_threads, queue_set), 2997 _collector(collector) { } 2998 2999 void set_task(CMSConcMarkingTask* task) { 3000 _task = task; 3001 } 3002 }; 3003 3004 class CMSConcMarkingOWSTTerminator: public OWSTTaskTerminator { 3005 CMSCollector* _collector; 3006 CMSConcMarkingTask* _task; 3007 public: 3008 virtual void yield(); 3009 3010 // "n_threads" is the number of threads to be terminated. 3011 // "queue_set" is a set of work queues of other threads. 3012 // "collector" is the CMS collector associated with this task terminator. 3013 // "yield" indicates whether we need the gang as a whole to yield. 3014 CMSConcMarkingOWSTTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3015 OWSTTaskTerminator(n_threads, queue_set), 3016 _collector(collector) { } 3017 3018 void set_task(CMSConcMarkingTask* task) { 3019 _task = task; 3020 } 3021 }; 3022 3023 class CMSConcMarkingTaskTerminator { 3024 private: 3025 ParallelTaskTerminator* _term; 3026 public: 3027 CMSConcMarkingTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) { 3028 if (UseOWSTTaskTerminator) { 3029 _term = new CMSConcMarkingOWSTTerminator(n_threads, queue_set, collector); 3030 } else { 3031 _term = new CMSConcMarkingParallelTerminator(n_threads, queue_set, collector); 3032 } 3033 } 3034 ~CMSConcMarkingTaskTerminator() { 3035 assert(_term != NULL, "Must not be NULL"); 3036 delete _term; 3037 } 3038 3039 void set_task(CMSConcMarkingTask* task); 3040 ParallelTaskTerminator* terminator() const { return _term; } 3041 }; 3042 3043 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3044 CMSConcMarkingTask* _task; 3045 public: 3046 bool should_exit_termination(); 3047 void set_task(CMSConcMarkingTask* task) { 3048 _task = task; 3049 } 3050 }; 3051 3052 // MT Concurrent Marking Task 3053 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3054 CMSCollector* _collector; 3055 uint _n_workers; // requested/desired # workers 3056 bool _result; 3057 CompactibleFreeListSpace* _cms_space; 3058 char _pad_front[64]; // padding to ... 3059 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3060 char _pad_back[64]; 3061 HeapWord* _restart_addr; 3062 3063 // Exposed here for yielding support 3064 Mutex* const _bit_map_lock; 3065 3066 // The per thread work queues, available here for stealing 3067 OopTaskQueueSet* _task_queues; 3068 3069 // Termination (and yielding) support 3070 CMSConcMarkingTaskTerminator _term; 3071 CMSConcMarkingTerminatorTerminator _term_term; 3072 3073 public: 3074 CMSConcMarkingTask(CMSCollector* collector, 3075 CompactibleFreeListSpace* cms_space, 3076 YieldingFlexibleWorkGang* workers, 3077 OopTaskQueueSet* task_queues): 3078 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3079 _collector(collector), 3080 _n_workers(0), 3081 _result(true), 3082 _cms_space(cms_space), 3083 _bit_map_lock(collector->bitMapLock()), 3084 _task_queues(task_queues), 3085 _term(_n_workers, task_queues, _collector) 3086 { 3087 _requested_size = _n_workers; 3088 _term.set_task(this); 3089 _term_term.set_task(this); 3090 _restart_addr = _global_finger = _cms_space->bottom(); 3091 } 3092 3093 3094 OopTaskQueueSet* task_queues() { return _task_queues; } 3095 3096 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3097 3098 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3099 3100 ParallelTaskTerminator* terminator() { return _term.terminator(); } 3101 3102 virtual void set_for_termination(uint active_workers) { 3103 terminator()->reset_for_reuse(active_workers); 3104 } 3105 3106 void work(uint worker_id); 3107 bool should_yield() { 3108 return ConcurrentMarkSweepThread::should_yield() 3109 && !_collector->foregroundGCIsActive(); 3110 } 3111 3112 virtual void coordinator_yield(); // stuff done by coordinator 3113 bool result() { return _result; } 3114 3115 void reset(HeapWord* ra) { 3116 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3117 _restart_addr = _global_finger = ra; 3118 _term.terminator()->reset_for_reuse(); 3119 } 3120 3121 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3122 OopTaskQueue* work_q); 3123 3124 private: 3125 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3126 void do_work_steal(int i); 3127 void bump_global_finger(HeapWord* f); 3128 }; 3129 3130 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3131 assert(_task != NULL, "Error"); 3132 return _task->yielding(); 3133 // Note that we do not need the disjunct || _task->should_yield() above 3134 // because we want terminating threads to yield only if the task 3135 // is already in the midst of yielding, which happens only after at least one 3136 // thread has yielded. 3137 } 3138 3139 void CMSConcMarkingParallelTerminator::yield() { 3140 if (_task->should_yield()) { 3141 _task->yield(); 3142 } else { 3143 ParallelTaskTerminator::yield(); 3144 } 3145 } 3146 3147 void CMSConcMarkingOWSTTerminator::yield() { 3148 if (_task->should_yield()) { 3149 _task->yield(); 3150 } else { 3151 OWSTTaskTerminator::yield(); 3152 } 3153 } 3154 3155 void CMSConcMarkingTaskTerminator::set_task(CMSConcMarkingTask* task) { 3156 if (UseOWSTTaskTerminator) { 3157 ((CMSConcMarkingOWSTTerminator*)_term)->set_task(task); 3158 } else { 3159 ((CMSConcMarkingParallelTerminator*)_term)->set_task(task); 3160 } 3161 } 3162 3163 //////////////////////////////////////////////////////////////// 3164 // Concurrent Marking Algorithm Sketch 3165 //////////////////////////////////////////////////////////////// 3166 // Until all tasks exhausted (both spaces): 3167 // -- claim next available chunk 3168 // -- bump global finger via CAS 3169 // -- find first object that starts in this chunk 3170 // and start scanning bitmap from that position 3171 // -- scan marked objects for oops 3172 // -- CAS-mark target, and if successful: 3173 // . if target oop is above global finger (volatile read) 3174 // nothing to do 3175 // . if target oop is in chunk and above local finger 3176 // then nothing to do 3177 // . else push on work-queue 3178 // -- Deal with possible overflow issues: 3179 // . local work-queue overflow causes stuff to be pushed on 3180 // global (common) overflow queue 3181 // . always first empty local work queue 3182 // . then get a batch of oops from global work queue if any 3183 // . then do work stealing 3184 // -- When all tasks claimed (both spaces) 3185 // and local work queue empty, 3186 // then in a loop do: 3187 // . check global overflow stack; steal a batch of oops and trace 3188 // . try to steal from other threads oif GOS is empty 3189 // . if neither is available, offer termination 3190 // -- Terminate and return result 3191 // 3192 void CMSConcMarkingTask::work(uint worker_id) { 3193 elapsedTimer _timer; 3194 ResourceMark rm; 3195 HandleMark hm; 3196 3197 DEBUG_ONLY(_collector->verify_overflow_empty();) 3198 3199 // Before we begin work, our work queue should be empty 3200 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3201 // Scan the bitmap covering _cms_space, tracing through grey objects. 3202 _timer.start(); 3203 do_scan_and_mark(worker_id, _cms_space); 3204 _timer.stop(); 3205 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3206 3207 // ... do work stealing 3208 _timer.reset(); 3209 _timer.start(); 3210 do_work_steal(worker_id); 3211 _timer.stop(); 3212 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3213 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3214 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3215 // Note that under the current task protocol, the 3216 // following assertion is true even of the spaces 3217 // expanded since the completion of the concurrent 3218 // marking. XXX This will likely change under a strict 3219 // ABORT semantics. 3220 // After perm removal the comparison was changed to 3221 // greater than or equal to from strictly greater than. 3222 // Before perm removal the highest address sweep would 3223 // have been at the end of perm gen but now is at the 3224 // end of the tenured gen. 3225 assert(_global_finger >= _cms_space->end(), 3226 "All tasks have been completed"); 3227 DEBUG_ONLY(_collector->verify_overflow_empty();) 3228 } 3229 3230 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3231 HeapWord* read = _global_finger; 3232 HeapWord* cur = read; 3233 while (f > read) { 3234 cur = read; 3235 read = Atomic::cmpxchg(f, &_global_finger, cur); 3236 if (cur == read) { 3237 // our cas succeeded 3238 assert(_global_finger >= f, "protocol consistency"); 3239 break; 3240 } 3241 } 3242 } 3243 3244 // This is really inefficient, and should be redone by 3245 // using (not yet available) block-read and -write interfaces to the 3246 // stack and the work_queue. XXX FIX ME !!! 3247 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3248 OopTaskQueue* work_q) { 3249 // Fast lock-free check 3250 if (ovflw_stk->length() == 0) { 3251 return false; 3252 } 3253 assert(work_q->size() == 0, "Shouldn't steal"); 3254 MutexLockerEx ml(ovflw_stk->par_lock(), 3255 Mutex::_no_safepoint_check_flag); 3256 // Grab up to 1/4 the size of the work queue 3257 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3258 (size_t)ParGCDesiredObjsFromOverflowList); 3259 num = MIN2(num, ovflw_stk->length()); 3260 for (int i = (int) num; i > 0; i--) { 3261 oop cur = ovflw_stk->pop(); 3262 assert(cur != NULL, "Counted wrong?"); 3263 work_q->push(cur); 3264 } 3265 return num > 0; 3266 } 3267 3268 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3269 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3270 int n_tasks = pst->n_tasks(); 3271 // We allow that there may be no tasks to do here because 3272 // we are restarting after a stack overflow. 3273 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3274 uint nth_task = 0; 3275 3276 HeapWord* aligned_start = sp->bottom(); 3277 if (sp->used_region().contains(_restart_addr)) { 3278 // Align down to a card boundary for the start of 0th task 3279 // for this space. 3280 aligned_start = align_down(_restart_addr, CardTable::card_size); 3281 } 3282 3283 size_t chunk_size = sp->marking_task_size(); 3284 while (pst->try_claim_task(/* reference */ nth_task)) { 3285 // Having claimed the nth task in this space, 3286 // compute the chunk that it corresponds to: 3287 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3288 aligned_start + (nth_task+1)*chunk_size); 3289 // Try and bump the global finger via a CAS; 3290 // note that we need to do the global finger bump 3291 // _before_ taking the intersection below, because 3292 // the task corresponding to that region will be 3293 // deemed done even if the used_region() expands 3294 // because of allocation -- as it almost certainly will 3295 // during start-up while the threads yield in the 3296 // closure below. 3297 HeapWord* finger = span.end(); 3298 bump_global_finger(finger); // atomically 3299 // There are null tasks here corresponding to chunks 3300 // beyond the "top" address of the space. 3301 span = span.intersection(sp->used_region()); 3302 if (!span.is_empty()) { // Non-null task 3303 HeapWord* prev_obj; 3304 assert(!span.contains(_restart_addr) || nth_task == 0, 3305 "Inconsistency"); 3306 if (nth_task == 0) { 3307 // For the 0th task, we'll not need to compute a block_start. 3308 if (span.contains(_restart_addr)) { 3309 // In the case of a restart because of stack overflow, 3310 // we might additionally skip a chunk prefix. 3311 prev_obj = _restart_addr; 3312 } else { 3313 prev_obj = span.start(); 3314 } 3315 } else { 3316 // We want to skip the first object because 3317 // the protocol is to scan any object in its entirety 3318 // that _starts_ in this span; a fortiori, any 3319 // object starting in an earlier span is scanned 3320 // as part of an earlier claimed task. 3321 // Below we use the "careful" version of block_start 3322 // so we do not try to navigate uninitialized objects. 3323 prev_obj = sp->block_start_careful(span.start()); 3324 // Below we use a variant of block_size that uses the 3325 // Printezis bits to avoid waiting for allocated 3326 // objects to become initialized/parsable. 3327 while (prev_obj < span.start()) { 3328 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3329 if (sz > 0) { 3330 prev_obj += sz; 3331 } else { 3332 // In this case we may end up doing a bit of redundant 3333 // scanning, but that appears unavoidable, short of 3334 // locking the free list locks; see bug 6324141. 3335 break; 3336 } 3337 } 3338 } 3339 if (prev_obj < span.end()) { 3340 MemRegion my_span = MemRegion(prev_obj, span.end()); 3341 // Do the marking work within a non-empty span -- 3342 // the last argument to the constructor indicates whether the 3343 // iteration should be incremental with periodic yields. 3344 ParMarkFromRootsClosure cl(this, _collector, my_span, 3345 &_collector->_markBitMap, 3346 work_queue(i), 3347 &_collector->_markStack); 3348 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3349 } // else nothing to do for this task 3350 } // else nothing to do for this task 3351 } 3352 // We'd be tempted to assert here that since there are no 3353 // more tasks left to claim in this space, the global_finger 3354 // must exceed space->top() and a fortiori space->end(). However, 3355 // that would not quite be correct because the bumping of 3356 // global_finger occurs strictly after the claiming of a task, 3357 // so by the time we reach here the global finger may not yet 3358 // have been bumped up by the thread that claimed the last 3359 // task. 3360 pst->all_tasks_completed(); 3361 } 3362 3363 class ParConcMarkingClosure: public MetadataVisitingOopIterateClosure { 3364 private: 3365 CMSCollector* _collector; 3366 CMSConcMarkingTask* _task; 3367 MemRegion _span; 3368 CMSBitMap* _bit_map; 3369 CMSMarkStack* _overflow_stack; 3370 OopTaskQueue* _work_queue; 3371 protected: 3372 DO_OOP_WORK_DEFN 3373 public: 3374 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3375 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3376 MetadataVisitingOopIterateClosure(collector->ref_processor()), 3377 _collector(collector), 3378 _task(task), 3379 _span(collector->_span), 3380 _bit_map(bit_map), 3381 _overflow_stack(overflow_stack), 3382 _work_queue(work_queue) 3383 { } 3384 virtual void do_oop(oop* p); 3385 virtual void do_oop(narrowOop* p); 3386 3387 void trim_queue(size_t max); 3388 void handle_stack_overflow(HeapWord* lost); 3389 void do_yield_check() { 3390 if (_task->should_yield()) { 3391 _task->yield(); 3392 } 3393 } 3394 }; 3395 3396 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3397 3398 // Grey object scanning during work stealing phase -- 3399 // the salient assumption here is that any references 3400 // that are in these stolen objects being scanned must 3401 // already have been initialized (else they would not have 3402 // been published), so we do not need to check for 3403 // uninitialized objects before pushing here. 3404 void ParConcMarkingClosure::do_oop(oop obj) { 3405 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3406 HeapWord* addr = (HeapWord*)obj; 3407 // Check if oop points into the CMS generation 3408 // and is not marked 3409 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3410 // a white object ... 3411 // If we manage to "claim" the object, by being the 3412 // first thread to mark it, then we push it on our 3413 // marking stack 3414 if (_bit_map->par_mark(addr)) { // ... now grey 3415 // push on work queue (grey set) 3416 bool simulate_overflow = false; 3417 NOT_PRODUCT( 3418 if (CMSMarkStackOverflowALot && 3419 _collector->simulate_overflow()) { 3420 // simulate a stack overflow 3421 simulate_overflow = true; 3422 } 3423 ) 3424 if (simulate_overflow || 3425 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3426 // stack overflow 3427 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3428 // We cannot assert that the overflow stack is full because 3429 // it may have been emptied since. 3430 assert(simulate_overflow || 3431 _work_queue->size() == _work_queue->max_elems(), 3432 "Else push should have succeeded"); 3433 handle_stack_overflow(addr); 3434 } 3435 } // Else, some other thread got there first 3436 do_yield_check(); 3437 } 3438 } 3439 3440 void ParConcMarkingClosure::trim_queue(size_t max) { 3441 while (_work_queue->size() > max) { 3442 oop new_oop; 3443 if (_work_queue->pop_local(new_oop)) { 3444 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3445 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3446 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3447 new_oop->oop_iterate(this); // do_oop() above 3448 do_yield_check(); 3449 } 3450 } 3451 } 3452 3453 // Upon stack overflow, we discard (part of) the stack, 3454 // remembering the least address amongst those discarded 3455 // in CMSCollector's _restart_address. 3456 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3457 // We need to do this under a mutex to prevent other 3458 // workers from interfering with the work done below. 3459 MutexLockerEx ml(_overflow_stack->par_lock(), 3460 Mutex::_no_safepoint_check_flag); 3461 // Remember the least grey address discarded 3462 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3463 _collector->lower_restart_addr(ra); 3464 _overflow_stack->reset(); // discard stack contents 3465 _overflow_stack->expand(); // expand the stack if possible 3466 } 3467 3468 3469 void CMSConcMarkingTask::do_work_steal(int i) { 3470 OopTaskQueue* work_q = work_queue(i); 3471 oop obj_to_scan; 3472 CMSBitMap* bm = &(_collector->_markBitMap); 3473 CMSMarkStack* ovflw = &(_collector->_markStack); 3474 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3475 while (true) { 3476 cl.trim_queue(0); 3477 assert(work_q->size() == 0, "Should have been emptied above"); 3478 if (get_work_from_overflow_stack(ovflw, work_q)) { 3479 // Can't assert below because the work obtained from the 3480 // overflow stack may already have been stolen from us. 3481 // assert(work_q->size() > 0, "Work from overflow stack"); 3482 continue; 3483 } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 3484 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3485 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3486 obj_to_scan->oop_iterate(&cl); 3487 } else if (terminator()->offer_termination(&_term_term)) { 3488 assert(work_q->size() == 0, "Impossible!"); 3489 break; 3490 } else if (yielding() || should_yield()) { 3491 yield(); 3492 } 3493 } 3494 } 3495 3496 // This is run by the CMS (coordinator) thread. 3497 void CMSConcMarkingTask::coordinator_yield() { 3498 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3499 "CMS thread should hold CMS token"); 3500 // First give up the locks, then yield, then re-lock 3501 // We should probably use a constructor/destructor idiom to 3502 // do this unlock/lock or modify the MutexUnlocker class to 3503 // serve our purpose. XXX 3504 assert_lock_strong(_bit_map_lock); 3505 _bit_map_lock->unlock(); 3506 ConcurrentMarkSweepThread::desynchronize(true); 3507 _collector->stopTimer(); 3508 _collector->incrementYields(); 3509 3510 // It is possible for whichever thread initiated the yield request 3511 // not to get a chance to wake up and take the bitmap lock between 3512 // this thread releasing it and reacquiring it. So, while the 3513 // should_yield() flag is on, let's sleep for a bit to give the 3514 // other thread a chance to wake up. The limit imposed on the number 3515 // of iterations is defensive, to avoid any unforseen circumstances 3516 // putting us into an infinite loop. Since it's always been this 3517 // (coordinator_yield()) method that was observed to cause the 3518 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3519 // which is by default non-zero. For the other seven methods that 3520 // also perform the yield operation, as are using a different 3521 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3522 // can enable the sleeping for those methods too, if necessary. 3523 // See 6442774. 3524 // 3525 // We really need to reconsider the synchronization between the GC 3526 // thread and the yield-requesting threads in the future and we 3527 // should really use wait/notify, which is the recommended 3528 // way of doing this type of interaction. Additionally, we should 3529 // consolidate the eight methods that do the yield operation and they 3530 // are almost identical into one for better maintainability and 3531 // readability. See 6445193. 3532 // 3533 // Tony 2006.06.29 3534 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3535 ConcurrentMarkSweepThread::should_yield() && 3536 !CMSCollector::foregroundGCIsActive(); ++i) { 3537 os::sleep(Thread::current(), 1, false); 3538 } 3539 3540 ConcurrentMarkSweepThread::synchronize(true); 3541 _bit_map_lock->lock_without_safepoint_check(); 3542 _collector->startTimer(); 3543 } 3544 3545 bool CMSCollector::do_marking_mt() { 3546 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3547 uint num_workers = WorkerPolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3548 conc_workers()->active_workers(), 3549 Threads::number_of_non_daemon_threads()); 3550 num_workers = conc_workers()->update_active_workers(num_workers); 3551 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3552 3553 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3554 3555 CMSConcMarkingTask tsk(this, 3556 cms_space, 3557 conc_workers(), 3558 task_queues()); 3559 3560 // Since the actual number of workers we get may be different 3561 // from the number we requested above, do we need to do anything different 3562 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3563 // class?? XXX 3564 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3565 3566 // Refs discovery is already non-atomic. 3567 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3568 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3569 conc_workers()->start_task(&tsk); 3570 while (tsk.yielded()) { 3571 tsk.coordinator_yield(); 3572 conc_workers()->continue_task(&tsk); 3573 } 3574 // If the task was aborted, _restart_addr will be non-NULL 3575 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3576 while (_restart_addr != NULL) { 3577 // XXX For now we do not make use of ABORTED state and have not 3578 // yet implemented the right abort semantics (even in the original 3579 // single-threaded CMS case). That needs some more investigation 3580 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3581 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3582 // If _restart_addr is non-NULL, a marking stack overflow 3583 // occurred; we need to do a fresh marking iteration from the 3584 // indicated restart address. 3585 if (_foregroundGCIsActive) { 3586 // We may be running into repeated stack overflows, having 3587 // reached the limit of the stack size, while making very 3588 // slow forward progress. It may be best to bail out and 3589 // let the foreground collector do its job. 3590 // Clear _restart_addr, so that foreground GC 3591 // works from scratch. This avoids the headache of 3592 // a "rescan" which would otherwise be needed because 3593 // of the dirty mod union table & card table. 3594 _restart_addr = NULL; 3595 return false; 3596 } 3597 // Adjust the task to restart from _restart_addr 3598 tsk.reset(_restart_addr); 3599 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3600 _restart_addr); 3601 _restart_addr = NULL; 3602 // Get the workers going again 3603 conc_workers()->start_task(&tsk); 3604 while (tsk.yielded()) { 3605 tsk.coordinator_yield(); 3606 conc_workers()->continue_task(&tsk); 3607 } 3608 } 3609 assert(tsk.completed(), "Inconsistency"); 3610 assert(tsk.result() == true, "Inconsistency"); 3611 return true; 3612 } 3613 3614 bool CMSCollector::do_marking_st() { 3615 ResourceMark rm; 3616 HandleMark hm; 3617 3618 // Temporarily make refs discovery single threaded (non-MT) 3619 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3620 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3621 &_markStack, CMSYield); 3622 // the last argument to iterate indicates whether the iteration 3623 // should be incremental with periodic yields. 3624 _markBitMap.iterate(&markFromRootsClosure); 3625 // If _restart_addr is non-NULL, a marking stack overflow 3626 // occurred; we need to do a fresh iteration from the 3627 // indicated restart address. 3628 while (_restart_addr != NULL) { 3629 if (_foregroundGCIsActive) { 3630 // We may be running into repeated stack overflows, having 3631 // reached the limit of the stack size, while making very 3632 // slow forward progress. It may be best to bail out and 3633 // let the foreground collector do its job. 3634 // Clear _restart_addr, so that foreground GC 3635 // works from scratch. This avoids the headache of 3636 // a "rescan" which would otherwise be needed because 3637 // of the dirty mod union table & card table. 3638 _restart_addr = NULL; 3639 return false; // indicating failure to complete marking 3640 } 3641 // Deal with stack overflow: 3642 // we restart marking from _restart_addr 3643 HeapWord* ra = _restart_addr; 3644 markFromRootsClosure.reset(ra); 3645 _restart_addr = NULL; 3646 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3647 } 3648 return true; 3649 } 3650 3651 void CMSCollector::preclean() { 3652 check_correct_thread_executing(); 3653 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3654 verify_work_stacks_empty(); 3655 verify_overflow_empty(); 3656 _abort_preclean = false; 3657 if (CMSPrecleaningEnabled) { 3658 if (!CMSEdenChunksRecordAlways) { 3659 _eden_chunk_index = 0; 3660 } 3661 size_t used = get_eden_used(); 3662 size_t capacity = get_eden_capacity(); 3663 // Don't start sampling unless we will get sufficiently 3664 // many samples. 3665 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3666 * CMSScheduleRemarkEdenPenetration)) { 3667 _start_sampling = true; 3668 } else { 3669 _start_sampling = false; 3670 } 3671 GCTraceCPUTime tcpu; 3672 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3673 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3674 } 3675 CMSTokenSync x(true); // is cms thread 3676 if (CMSPrecleaningEnabled) { 3677 sample_eden(); 3678 _collectorState = AbortablePreclean; 3679 } else { 3680 _collectorState = FinalMarking; 3681 } 3682 verify_work_stacks_empty(); 3683 verify_overflow_empty(); 3684 } 3685 3686 // Try and schedule the remark such that young gen 3687 // occupancy is CMSScheduleRemarkEdenPenetration %. 3688 void CMSCollector::abortable_preclean() { 3689 check_correct_thread_executing(); 3690 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3691 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3692 3693 // If Eden's current occupancy is below this threshold, 3694 // immediately schedule the remark; else preclean 3695 // past the next scavenge in an effort to 3696 // schedule the pause as described above. By choosing 3697 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3698 // we will never do an actual abortable preclean cycle. 3699 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3700 GCTraceCPUTime tcpu; 3701 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3702 // We need more smarts in the abortable preclean 3703 // loop below to deal with cases where allocation 3704 // in young gen is very very slow, and our precleaning 3705 // is running a losing race against a horde of 3706 // mutators intent on flooding us with CMS updates 3707 // (dirty cards). 3708 // One, admittedly dumb, strategy is to give up 3709 // after a certain number of abortable precleaning loops 3710 // or after a certain maximum time. We want to make 3711 // this smarter in the next iteration. 3712 // XXX FIX ME!!! YSR 3713 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3714 while (!(should_abort_preclean() || 3715 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3716 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3717 cumworkdone += workdone; 3718 loops++; 3719 // Voluntarily terminate abortable preclean phase if we have 3720 // been at it for too long. 3721 if ((CMSMaxAbortablePrecleanLoops != 0) && 3722 loops >= CMSMaxAbortablePrecleanLoops) { 3723 log_debug(gc)(" CMS: abort preclean due to loops "); 3724 break; 3725 } 3726 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3727 log_debug(gc)(" CMS: abort preclean due to time "); 3728 break; 3729 } 3730 // If we are doing little work each iteration, we should 3731 // take a short break. 3732 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3733 // Sleep for some time, waiting for work to accumulate 3734 stopTimer(); 3735 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3736 startTimer(); 3737 waited++; 3738 } 3739 } 3740 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3741 loops, waited, cumworkdone); 3742 } 3743 CMSTokenSync x(true); // is cms thread 3744 if (_collectorState != Idling) { 3745 assert(_collectorState == AbortablePreclean, 3746 "Spontaneous state transition?"); 3747 _collectorState = FinalMarking; 3748 } // Else, a foreground collection completed this CMS cycle. 3749 return; 3750 } 3751 3752 // Respond to an Eden sampling opportunity 3753 void CMSCollector::sample_eden() { 3754 // Make sure a young gc cannot sneak in between our 3755 // reading and recording of a sample. 3756 assert(Thread::current()->is_ConcurrentGC_thread(), 3757 "Only the cms thread may collect Eden samples"); 3758 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3759 "Should collect samples while holding CMS token"); 3760 if (!_start_sampling) { 3761 return; 3762 } 3763 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3764 // is populated by the young generation. 3765 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3766 if (_eden_chunk_index < _eden_chunk_capacity) { 3767 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3768 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3769 "Unexpected state of Eden"); 3770 // We'd like to check that what we just sampled is an oop-start address; 3771 // however, we cannot do that here since the object may not yet have been 3772 // initialized. So we'll instead do the check when we _use_ this sample 3773 // later. 3774 if (_eden_chunk_index == 0 || 3775 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3776 _eden_chunk_array[_eden_chunk_index-1]) 3777 >= CMSSamplingGrain)) { 3778 _eden_chunk_index++; // commit sample 3779 } 3780 } 3781 } 3782 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3783 size_t used = get_eden_used(); 3784 size_t capacity = get_eden_capacity(); 3785 assert(used <= capacity, "Unexpected state of Eden"); 3786 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3787 _abort_preclean = true; 3788 } 3789 } 3790 } 3791 3792 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3793 assert(_collectorState == Precleaning || 3794 _collectorState == AbortablePreclean, "incorrect state"); 3795 ResourceMark rm; 3796 HandleMark hm; 3797 3798 // Precleaning is currently not MT but the reference processor 3799 // may be set for MT. Disable it temporarily here. 3800 ReferenceProcessor* rp = ref_processor(); 3801 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3802 3803 // Do one pass of scrubbing the discovered reference lists 3804 // to remove any reference objects with strongly-reachable 3805 // referents. 3806 if (clean_refs) { 3807 CMSPrecleanRefsYieldClosure yield_cl(this); 3808 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3809 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3810 &_markStack, true /* preclean */); 3811 CMSDrainMarkingStackClosure complete_trace(this, 3812 _span, &_markBitMap, &_markStack, 3813 &keep_alive, true /* preclean */); 3814 3815 // We don't want this step to interfere with a young 3816 // collection because we don't want to take CPU 3817 // or memory bandwidth away from the young GC threads 3818 // (which may be as many as there are CPUs). 3819 // Note that we don't need to protect ourselves from 3820 // interference with mutators because they can't 3821 // manipulate the discovered reference lists nor affect 3822 // the computed reachability of the referents, the 3823 // only properties manipulated by the precleaning 3824 // of these reference lists. 3825 stopTimer(); 3826 CMSTokenSyncWithLocks x(true /* is cms thread */, 3827 bitMapLock()); 3828 startTimer(); 3829 sample_eden(); 3830 3831 // The following will yield to allow foreground 3832 // collection to proceed promptly. XXX YSR: 3833 // The code in this method may need further 3834 // tweaking for better performance and some restructuring 3835 // for cleaner interfaces. 3836 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3837 rp->preclean_discovered_references( 3838 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3839 gc_timer); 3840 } 3841 3842 if (clean_survivor) { // preclean the active survivor space(s) 3843 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3844 &_markBitMap, &_modUnionTable, 3845 &_markStack, true /* precleaning phase */); 3846 stopTimer(); 3847 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3848 bitMapLock()); 3849 startTimer(); 3850 unsigned int before_count = 3851 CMSHeap::heap()->total_collections(); 3852 SurvivorSpacePrecleanClosure 3853 sss_cl(this, _span, &_markBitMap, &_markStack, 3854 &pam_cl, before_count, CMSYield); 3855 _young_gen->from()->object_iterate_careful(&sss_cl); 3856 _young_gen->to()->object_iterate_careful(&sss_cl); 3857 } 3858 MarkRefsIntoAndScanClosure 3859 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3860 &_markStack, this, CMSYield, 3861 true /* precleaning phase */); 3862 // CAUTION: The following closure has persistent state that may need to 3863 // be reset upon a decrease in the sequence of addresses it 3864 // processes. 3865 ScanMarkedObjectsAgainCarefullyClosure 3866 smoac_cl(this, _span, 3867 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3868 3869 // Preclean dirty cards in ModUnionTable and CardTable using 3870 // appropriate convergence criterion; 3871 // repeat CMSPrecleanIter times unless we find that 3872 // we are losing. 3873 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3874 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3875 "Bad convergence multiplier"); 3876 assert(CMSPrecleanThreshold >= 100, 3877 "Unreasonably low CMSPrecleanThreshold"); 3878 3879 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3880 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3881 numIter < CMSPrecleanIter; 3882 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3883 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3884 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3885 // Either there are very few dirty cards, so re-mark 3886 // pause will be small anyway, or our pre-cleaning isn't 3887 // that much faster than the rate at which cards are being 3888 // dirtied, so we might as well stop and re-mark since 3889 // precleaning won't improve our re-mark time by much. 3890 if (curNumCards <= CMSPrecleanThreshold || 3891 (numIter > 0 && 3892 (curNumCards * CMSPrecleanDenominator > 3893 lastNumCards * CMSPrecleanNumerator))) { 3894 numIter++; 3895 cumNumCards += curNumCards; 3896 break; 3897 } 3898 } 3899 3900 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3901 3902 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3903 cumNumCards += curNumCards; 3904 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3905 curNumCards, cumNumCards, numIter); 3906 return cumNumCards; // as a measure of useful work done 3907 } 3908 3909 // PRECLEANING NOTES: 3910 // Precleaning involves: 3911 // . reading the bits of the modUnionTable and clearing the set bits. 3912 // . For the cards corresponding to the set bits, we scan the 3913 // objects on those cards. This means we need the free_list_lock 3914 // so that we can safely iterate over the CMS space when scanning 3915 // for oops. 3916 // . When we scan the objects, we'll be both reading and setting 3917 // marks in the marking bit map, so we'll need the marking bit map. 3918 // . For protecting _collector_state transitions, we take the CGC_lock. 3919 // Note that any races in the reading of of card table entries by the 3920 // CMS thread on the one hand and the clearing of those entries by the 3921 // VM thread or the setting of those entries by the mutator threads on the 3922 // other are quite benign. However, for efficiency it makes sense to keep 3923 // the VM thread from racing with the CMS thread while the latter is 3924 // dirty card info to the modUnionTable. We therefore also use the 3925 // CGC_lock to protect the reading of the card table and the mod union 3926 // table by the CM thread. 3927 // . We run concurrently with mutator updates, so scanning 3928 // needs to be done carefully -- we should not try to scan 3929 // potentially uninitialized objects. 3930 // 3931 // Locking strategy: While holding the CGC_lock, we scan over and 3932 // reset a maximal dirty range of the mod union / card tables, then lock 3933 // the free_list_lock and bitmap lock to do a full marking, then 3934 // release these locks; and repeat the cycle. This allows for a 3935 // certain amount of fairness in the sharing of these locks between 3936 // the CMS collector on the one hand, and the VM thread and the 3937 // mutators on the other. 3938 3939 // NOTE: preclean_mod_union_table() and preclean_card_table() 3940 // further below are largely identical; if you need to modify 3941 // one of these methods, please check the other method too. 3942 3943 size_t CMSCollector::preclean_mod_union_table( 3944 ConcurrentMarkSweepGeneration* old_gen, 3945 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3946 verify_work_stacks_empty(); 3947 verify_overflow_empty(); 3948 3949 // strategy: starting with the first card, accumulate contiguous 3950 // ranges of dirty cards; clear these cards, then scan the region 3951 // covered by these cards. 3952 3953 // Since all of the MUT is committed ahead, we can just use 3954 // that, in case the generations expand while we are precleaning. 3955 // It might also be fine to just use the committed part of the 3956 // generation, but we might potentially miss cards when the 3957 // generation is rapidly expanding while we are in the midst 3958 // of precleaning. 3959 HeapWord* startAddr = old_gen->reserved().start(); 3960 HeapWord* endAddr = old_gen->reserved().end(); 3961 3962 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3963 3964 size_t numDirtyCards, cumNumDirtyCards; 3965 HeapWord *nextAddr, *lastAddr; 3966 for (cumNumDirtyCards = numDirtyCards = 0, 3967 nextAddr = lastAddr = startAddr; 3968 nextAddr < endAddr; 3969 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3970 3971 ResourceMark rm; 3972 HandleMark hm; 3973 3974 MemRegion dirtyRegion; 3975 { 3976 stopTimer(); 3977 // Potential yield point 3978 CMSTokenSync ts(true); 3979 startTimer(); 3980 sample_eden(); 3981 // Get dirty region starting at nextOffset (inclusive), 3982 // simultaneously clearing it. 3983 dirtyRegion = 3984 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3985 assert(dirtyRegion.start() >= nextAddr, 3986 "returned region inconsistent?"); 3987 } 3988 // Remember where the next search should begin. 3989 // The returned region (if non-empty) is a right open interval, 3990 // so lastOffset is obtained from the right end of that 3991 // interval. 3992 lastAddr = dirtyRegion.end(); 3993 // Should do something more transparent and less hacky XXX 3994 numDirtyCards = 3995 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3996 3997 // We'll scan the cards in the dirty region (with periodic 3998 // yields for foreground GC as needed). 3999 if (!dirtyRegion.is_empty()) { 4000 assert(numDirtyCards > 0, "consistency check"); 4001 HeapWord* stop_point = NULL; 4002 stopTimer(); 4003 // Potential yield point 4004 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4005 bitMapLock()); 4006 startTimer(); 4007 { 4008 verify_work_stacks_empty(); 4009 verify_overflow_empty(); 4010 sample_eden(); 4011 stop_point = 4012 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4013 } 4014 if (stop_point != NULL) { 4015 // The careful iteration stopped early either because it found an 4016 // uninitialized object, or because we were in the midst of an 4017 // "abortable preclean", which should now be aborted. Redirty 4018 // the bits corresponding to the partially-scanned or unscanned 4019 // cards. We'll either restart at the next block boundary or 4020 // abort the preclean. 4021 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4022 "Should only be AbortablePreclean."); 4023 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4024 if (should_abort_preclean()) { 4025 break; // out of preclean loop 4026 } else { 4027 // Compute the next address at which preclean should pick up; 4028 // might need bitMapLock in order to read P-bits. 4029 lastAddr = next_card_start_after_block(stop_point); 4030 } 4031 } 4032 } else { 4033 assert(lastAddr == endAddr, "consistency check"); 4034 assert(numDirtyCards == 0, "consistency check"); 4035 break; 4036 } 4037 } 4038 verify_work_stacks_empty(); 4039 verify_overflow_empty(); 4040 return cumNumDirtyCards; 4041 } 4042 4043 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4044 // below are largely identical; if you need to modify 4045 // one of these methods, please check the other method too. 4046 4047 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4048 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4049 // strategy: it's similar to precleamModUnionTable above, in that 4050 // we accumulate contiguous ranges of dirty cards, mark these cards 4051 // precleaned, then scan the region covered by these cards. 4052 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4053 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4054 4055 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4056 4057 size_t numDirtyCards, cumNumDirtyCards; 4058 HeapWord *lastAddr, *nextAddr; 4059 4060 for (cumNumDirtyCards = numDirtyCards = 0, 4061 nextAddr = lastAddr = startAddr; 4062 nextAddr < endAddr; 4063 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4064 4065 ResourceMark rm; 4066 HandleMark hm; 4067 4068 MemRegion dirtyRegion; 4069 { 4070 // See comments in "Precleaning notes" above on why we 4071 // do this locking. XXX Could the locking overheads be 4072 // too high when dirty cards are sparse? [I don't think so.] 4073 stopTimer(); 4074 CMSTokenSync x(true); // is cms thread 4075 startTimer(); 4076 sample_eden(); 4077 // Get and clear dirty region from card table 4078 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4079 true, 4080 CardTable::precleaned_card_val()); 4081 4082 assert(dirtyRegion.start() >= nextAddr, 4083 "returned region inconsistent?"); 4084 } 4085 lastAddr = dirtyRegion.end(); 4086 numDirtyCards = 4087 dirtyRegion.word_size()/CardTable::card_size_in_words; 4088 4089 if (!dirtyRegion.is_empty()) { 4090 stopTimer(); 4091 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4092 startTimer(); 4093 sample_eden(); 4094 verify_work_stacks_empty(); 4095 verify_overflow_empty(); 4096 HeapWord* stop_point = 4097 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4098 if (stop_point != NULL) { 4099 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4100 "Should only be AbortablePreclean."); 4101 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4102 if (should_abort_preclean()) { 4103 break; // out of preclean loop 4104 } else { 4105 // Compute the next address at which preclean should pick up. 4106 lastAddr = next_card_start_after_block(stop_point); 4107 } 4108 } 4109 } else { 4110 break; 4111 } 4112 } 4113 verify_work_stacks_empty(); 4114 verify_overflow_empty(); 4115 return cumNumDirtyCards; 4116 } 4117 4118 class PrecleanCLDClosure : public CLDClosure { 4119 MetadataVisitingOopsInGenClosure* _cm_closure; 4120 public: 4121 PrecleanCLDClosure(MetadataVisitingOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4122 void do_cld(ClassLoaderData* cld) { 4123 if (cld->has_accumulated_modified_oops()) { 4124 cld->clear_accumulated_modified_oops(); 4125 4126 _cm_closure->do_cld(cld); 4127 } 4128 } 4129 }; 4130 4131 // The freelist lock is needed to prevent asserts, is it really needed? 4132 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4133 // Needed to walk CLDG 4134 MutexLocker ml(ClassLoaderDataGraph_lock); 4135 4136 cl->set_freelistLock(freelistLock); 4137 4138 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4139 4140 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4141 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4142 PrecleanCLDClosure preclean_closure(cl); 4143 ClassLoaderDataGraph::cld_do(&preclean_closure); 4144 4145 verify_work_stacks_empty(); 4146 verify_overflow_empty(); 4147 } 4148 4149 void CMSCollector::checkpointRootsFinal() { 4150 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4151 check_correct_thread_executing(); 4152 // world is stopped at this checkpoint 4153 assert(SafepointSynchronize::is_at_safepoint(), 4154 "world should be stopped"); 4155 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4156 4157 verify_work_stacks_empty(); 4158 verify_overflow_empty(); 4159 4160 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4161 _young_gen->used() / K, _young_gen->capacity() / K); 4162 { 4163 if (CMSScavengeBeforeRemark) { 4164 CMSHeap* heap = CMSHeap::heap(); 4165 // Temporarily set flag to false, GCH->do_collection will 4166 // expect it to be false and set to true 4167 FlagSetting fl(heap->_is_gc_active, false); 4168 4169 heap->do_collection(true, // full (i.e. force, see below) 4170 false, // !clear_all_soft_refs 4171 0, // size 4172 false, // is_tlab 4173 GenCollectedHeap::YoungGen // type 4174 ); 4175 } 4176 FreelistLocker x(this); 4177 MutexLockerEx y(bitMapLock(), 4178 Mutex::_no_safepoint_check_flag); 4179 checkpointRootsFinalWork(); 4180 } 4181 verify_work_stacks_empty(); 4182 verify_overflow_empty(); 4183 } 4184 4185 void CMSCollector::checkpointRootsFinalWork() { 4186 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4187 4188 assert(haveFreelistLocks(), "must have free list locks"); 4189 assert_lock_strong(bitMapLock()); 4190 4191 ResourceMark rm; 4192 HandleMark hm; 4193 4194 CMSHeap* heap = CMSHeap::heap(); 4195 4196 if (should_unload_classes()) { 4197 CodeCache::gc_prologue(); 4198 } 4199 assert(haveFreelistLocks(), "must have free list locks"); 4200 assert_lock_strong(bitMapLock()); 4201 4202 // We might assume that we need not fill TLAB's when 4203 // CMSScavengeBeforeRemark is set, because we may have just done 4204 // a scavenge which would have filled all TLAB's -- and besides 4205 // Eden would be empty. This however may not always be the case -- 4206 // for instance although we asked for a scavenge, it may not have 4207 // happened because of a JNI critical section. We probably need 4208 // a policy for deciding whether we can in that case wait until 4209 // the critical section releases and then do the remark following 4210 // the scavenge, and skip it here. In the absence of that policy, 4211 // or of an indication of whether the scavenge did indeed occur, 4212 // we cannot rely on TLAB's having been filled and must do 4213 // so here just in case a scavenge did not happen. 4214 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4215 // Update the saved marks which may affect the root scans. 4216 heap->save_marks(); 4217 4218 print_eden_and_survivor_chunk_arrays(); 4219 4220 { 4221 #if COMPILER2_OR_JVMCI 4222 DerivedPointerTableDeactivate dpt_deact; 4223 #endif 4224 4225 // Note on the role of the mod union table: 4226 // Since the marker in "markFromRoots" marks concurrently with 4227 // mutators, it is possible for some reachable objects not to have been 4228 // scanned. For instance, an only reference to an object A was 4229 // placed in object B after the marker scanned B. Unless B is rescanned, 4230 // A would be collected. Such updates to references in marked objects 4231 // are detected via the mod union table which is the set of all cards 4232 // dirtied since the first checkpoint in this GC cycle and prior to 4233 // the most recent young generation GC, minus those cleaned up by the 4234 // concurrent precleaning. 4235 if (CMSParallelRemarkEnabled) { 4236 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4237 do_remark_parallel(); 4238 } else { 4239 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4240 do_remark_non_parallel(); 4241 } 4242 } 4243 verify_work_stacks_empty(); 4244 verify_overflow_empty(); 4245 4246 { 4247 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4248 refProcessingWork(); 4249 } 4250 verify_work_stacks_empty(); 4251 verify_overflow_empty(); 4252 4253 if (should_unload_classes()) { 4254 CodeCache::gc_epilogue(); 4255 } 4256 JvmtiExport::gc_epilogue(); 4257 4258 // If we encountered any (marking stack / work queue) overflow 4259 // events during the current CMS cycle, take appropriate 4260 // remedial measures, where possible, so as to try and avoid 4261 // recurrence of that condition. 4262 assert(_markStack.isEmpty(), "No grey objects"); 4263 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4264 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4265 if (ser_ovflw > 0) { 4266 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4267 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4268 _markStack.expand(); 4269 _ser_pmc_remark_ovflw = 0; 4270 _ser_pmc_preclean_ovflw = 0; 4271 _ser_kac_preclean_ovflw = 0; 4272 _ser_kac_ovflw = 0; 4273 } 4274 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4275 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4276 _par_pmc_remark_ovflw, _par_kac_ovflw); 4277 _par_pmc_remark_ovflw = 0; 4278 _par_kac_ovflw = 0; 4279 } 4280 if (_markStack._hit_limit > 0) { 4281 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4282 _markStack._hit_limit); 4283 } 4284 if (_markStack._failed_double > 0) { 4285 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4286 _markStack._failed_double, _markStack.capacity()); 4287 } 4288 _markStack._hit_limit = 0; 4289 _markStack._failed_double = 0; 4290 4291 if ((VerifyAfterGC || VerifyDuringGC) && 4292 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4293 verify_after_remark(); 4294 } 4295 4296 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4297 4298 // Change under the freelistLocks. 4299 _collectorState = Sweeping; 4300 // Call isAllClear() under bitMapLock 4301 assert(_modUnionTable.isAllClear(), 4302 "Should be clear by end of the final marking"); 4303 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4304 "Should be clear by end of the final marking"); 4305 } 4306 4307 void CMSParInitialMarkTask::work(uint worker_id) { 4308 elapsedTimer _timer; 4309 ResourceMark rm; 4310 HandleMark hm; 4311 4312 // ---------- scan from roots -------------- 4313 _timer.start(); 4314 CMSHeap* heap = CMSHeap::heap(); 4315 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4316 4317 // ---------- young gen roots -------------- 4318 { 4319 work_on_young_gen_roots(&par_mri_cl); 4320 _timer.stop(); 4321 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4322 } 4323 4324 // ---------- remaining roots -------------- 4325 _timer.reset(); 4326 _timer.start(); 4327 4328 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4329 4330 heap->cms_process_roots(_strong_roots_scope, 4331 false, // yg was scanned above 4332 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4333 _collector->should_unload_classes(), 4334 &par_mri_cl, 4335 &cld_closure, 4336 &_par_state_string); 4337 4338 assert(_collector->should_unload_classes() 4339 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4340 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4341 _timer.stop(); 4342 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4343 } 4344 4345 // Parallel remark task 4346 class CMSParRemarkTask: public CMSParMarkTask { 4347 CompactibleFreeListSpace* _cms_space; 4348 4349 // The per-thread work queues, available here for stealing. 4350 OopTaskQueueSet* _task_queues; 4351 TaskTerminator _term; 4352 StrongRootsScope* _strong_roots_scope; 4353 4354 public: 4355 // A value of 0 passed to n_workers will cause the number of 4356 // workers to be taken from the active workers in the work gang. 4357 CMSParRemarkTask(CMSCollector* collector, 4358 CompactibleFreeListSpace* cms_space, 4359 uint n_workers, WorkGang* workers, 4360 OopTaskQueueSet* task_queues, 4361 StrongRootsScope* strong_roots_scope): 4362 CMSParMarkTask("Rescan roots and grey objects in parallel", 4363 collector, n_workers), 4364 _cms_space(cms_space), 4365 _task_queues(task_queues), 4366 _term(n_workers, task_queues), 4367 _strong_roots_scope(strong_roots_scope) { } 4368 4369 OopTaskQueueSet* task_queues() { return _task_queues; } 4370 4371 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4372 4373 ParallelTaskTerminator* terminator() { return _term.terminator(); } 4374 uint n_workers() { return _n_workers; } 4375 4376 void work(uint worker_id); 4377 4378 private: 4379 // ... of dirty cards in old space 4380 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4381 ParMarkRefsIntoAndScanClosure* cl); 4382 4383 // ... work stealing for the above 4384 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4385 }; 4386 4387 class RemarkCLDClosure : public CLDClosure { 4388 CLDToOopClosure _cm_closure; 4389 public: 4390 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4391 void do_cld(ClassLoaderData* cld) { 4392 // Check if we have modified any oops in the CLD during the concurrent marking. 4393 if (cld->has_accumulated_modified_oops()) { 4394 cld->clear_accumulated_modified_oops(); 4395 4396 // We could have transfered the current modified marks to the accumulated marks, 4397 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4398 } else if (cld->has_modified_oops()) { 4399 // Don't clear anything, this info is needed by the next young collection. 4400 } else { 4401 // No modified oops in the ClassLoaderData. 4402 return; 4403 } 4404 4405 // The klass has modified fields, need to scan the klass. 4406 _cm_closure.do_cld(cld); 4407 } 4408 }; 4409 4410 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4411 ParNewGeneration* young_gen = _collector->_young_gen; 4412 ContiguousSpace* eden_space = young_gen->eden(); 4413 ContiguousSpace* from_space = young_gen->from(); 4414 ContiguousSpace* to_space = young_gen->to(); 4415 4416 HeapWord** eca = _collector->_eden_chunk_array; 4417 size_t ect = _collector->_eden_chunk_index; 4418 HeapWord** sca = _collector->_survivor_chunk_array; 4419 size_t sct = _collector->_survivor_chunk_index; 4420 4421 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4422 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4423 4424 do_young_space_rescan(cl, to_space, NULL, 0); 4425 do_young_space_rescan(cl, from_space, sca, sct); 4426 do_young_space_rescan(cl, eden_space, eca, ect); 4427 } 4428 4429 // work_queue(i) is passed to the closure 4430 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4431 // also is passed to do_dirty_card_rescan_tasks() and to 4432 // do_work_steal() to select the i-th task_queue. 4433 4434 void CMSParRemarkTask::work(uint worker_id) { 4435 elapsedTimer _timer; 4436 ResourceMark rm; 4437 HandleMark hm; 4438 4439 // ---------- rescan from roots -------------- 4440 _timer.start(); 4441 CMSHeap* heap = CMSHeap::heap(); 4442 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4443 _collector->_span, _collector->ref_processor(), 4444 &(_collector->_markBitMap), 4445 work_queue(worker_id)); 4446 4447 // Rescan young gen roots first since these are likely 4448 // coarsely partitioned and may, on that account, constitute 4449 // the critical path; thus, it's best to start off that 4450 // work first. 4451 // ---------- young gen roots -------------- 4452 { 4453 work_on_young_gen_roots(&par_mrias_cl); 4454 _timer.stop(); 4455 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4456 } 4457 4458 // ---------- remaining roots -------------- 4459 _timer.reset(); 4460 _timer.start(); 4461 heap->cms_process_roots(_strong_roots_scope, 4462 false, // yg was scanned above 4463 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4464 _collector->should_unload_classes(), 4465 &par_mrias_cl, 4466 NULL, // The dirty klasses will be handled below 4467 &_par_state_string); 4468 4469 assert(_collector->should_unload_classes() 4470 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4471 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4472 _timer.stop(); 4473 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4474 4475 // ---------- unhandled CLD scanning ---------- 4476 if (worker_id == 0) { // Single threaded at the moment. 4477 _timer.reset(); 4478 _timer.start(); 4479 4480 // Scan all new class loader data objects and new dependencies that were 4481 // introduced during concurrent marking. 4482 ResourceMark rm; 4483 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4484 for (int i = 0; i < array->length(); i++) { 4485 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4486 } 4487 4488 // We don't need to keep track of new CLDs anymore. 4489 ClassLoaderDataGraph::remember_new_clds(false); 4490 4491 _timer.stop(); 4492 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4493 } 4494 4495 // We might have added oops to ClassLoaderData::_handles during the 4496 // concurrent marking phase. These oops do not always point to newly allocated objects 4497 // that are guaranteed to be kept alive. Hence, 4498 // we do have to revisit the _handles block during the remark phase. 4499 4500 // ---------- dirty CLD scanning ---------- 4501 if (worker_id == 0) { // Single threaded at the moment. 4502 _timer.reset(); 4503 _timer.start(); 4504 4505 // Scan all classes that was dirtied during the concurrent marking phase. 4506 RemarkCLDClosure remark_closure(&par_mrias_cl); 4507 ClassLoaderDataGraph::cld_do(&remark_closure); 4508 4509 _timer.stop(); 4510 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4511 } 4512 4513 4514 // ---------- rescan dirty cards ------------ 4515 _timer.reset(); 4516 _timer.start(); 4517 4518 // Do the rescan tasks for each of the two spaces 4519 // (cms_space) in turn. 4520 // "worker_id" is passed to select the task_queue for "worker_id" 4521 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4522 _timer.stop(); 4523 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4524 4525 // ---------- steal work from other threads ... 4526 // ---------- ... and drain overflow list. 4527 _timer.reset(); 4528 _timer.start(); 4529 do_work_steal(worker_id, &par_mrias_cl); 4530 _timer.stop(); 4531 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4532 } 4533 4534 void 4535 CMSParMarkTask::do_young_space_rescan( 4536 OopsInGenClosure* cl, ContiguousSpace* space, 4537 HeapWord** chunk_array, size_t chunk_top) { 4538 // Until all tasks completed: 4539 // . claim an unclaimed task 4540 // . compute region boundaries corresponding to task claimed 4541 // using chunk_array 4542 // . par_oop_iterate(cl) over that region 4543 4544 ResourceMark rm; 4545 HandleMark hm; 4546 4547 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4548 4549 uint nth_task = 0; 4550 uint n_tasks = pst->n_tasks(); 4551 4552 if (n_tasks > 0) { 4553 assert(pst->valid(), "Uninitialized use?"); 4554 HeapWord *start, *end; 4555 while (pst->try_claim_task(/* reference */ nth_task)) { 4556 // We claimed task # nth_task; compute its boundaries. 4557 if (chunk_top == 0) { // no samples were taken 4558 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4559 start = space->bottom(); 4560 end = space->top(); 4561 } else if (nth_task == 0) { 4562 start = space->bottom(); 4563 end = chunk_array[nth_task]; 4564 } else if (nth_task < (uint)chunk_top) { 4565 assert(nth_task >= 1, "Control point invariant"); 4566 start = chunk_array[nth_task - 1]; 4567 end = chunk_array[nth_task]; 4568 } else { 4569 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4570 start = chunk_array[chunk_top - 1]; 4571 end = space->top(); 4572 } 4573 MemRegion mr(start, end); 4574 // Verify that mr is in space 4575 assert(mr.is_empty() || space->used_region().contains(mr), 4576 "Should be in space"); 4577 // Verify that "start" is an object boundary 4578 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4579 "Should be an oop"); 4580 space->par_oop_iterate(mr, cl); 4581 } 4582 pst->all_tasks_completed(); 4583 } 4584 } 4585 4586 void 4587 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4588 CompactibleFreeListSpace* sp, int i, 4589 ParMarkRefsIntoAndScanClosure* cl) { 4590 // Until all tasks completed: 4591 // . claim an unclaimed task 4592 // . compute region boundaries corresponding to task claimed 4593 // . transfer dirty bits ct->mut for that region 4594 // . apply rescanclosure to dirty mut bits for that region 4595 4596 ResourceMark rm; 4597 HandleMark hm; 4598 4599 OopTaskQueue* work_q = work_queue(i); 4600 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4601 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4602 // CAUTION: This closure has state that persists across calls to 4603 // the work method dirty_range_iterate_clear() in that it has 4604 // embedded in it a (subtype of) UpwardsObjectClosure. The 4605 // use of that state in the embedded UpwardsObjectClosure instance 4606 // assumes that the cards are always iterated (even if in parallel 4607 // by several threads) in monotonically increasing order per each 4608 // thread. This is true of the implementation below which picks 4609 // card ranges (chunks) in monotonically increasing order globally 4610 // and, a-fortiori, in monotonically increasing order per thread 4611 // (the latter order being a subsequence of the former). 4612 // If the work code below is ever reorganized into a more chaotic 4613 // work-partitioning form than the current "sequential tasks" 4614 // paradigm, the use of that persistent state will have to be 4615 // revisited and modified appropriately. See also related 4616 // bug 4756801 work on which should examine this code to make 4617 // sure that the changes there do not run counter to the 4618 // assumptions made here and necessary for correctness and 4619 // efficiency. Note also that this code might yield inefficient 4620 // behavior in the case of very large objects that span one or 4621 // more work chunks. Such objects would potentially be scanned 4622 // several times redundantly. Work on 4756801 should try and 4623 // address that performance anomaly if at all possible. XXX 4624 MemRegion full_span = _collector->_span; 4625 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4626 MarkFromDirtyCardsClosure 4627 greyRescanClosure(_collector, full_span, // entire span of interest 4628 sp, bm, work_q, cl); 4629 4630 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4631 assert(pst->valid(), "Uninitialized use?"); 4632 uint nth_task = 0; 4633 const int alignment = CardTable::card_size * BitsPerWord; 4634 MemRegion span = sp->used_region(); 4635 HeapWord* start_addr = span.start(); 4636 HeapWord* end_addr = align_up(span.end(), alignment); 4637 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4638 assert(is_aligned(start_addr, alignment), "Check alignment"); 4639 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4640 4641 while (pst->try_claim_task(/* reference */ nth_task)) { 4642 // Having claimed the nth_task, compute corresponding mem-region, 4643 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4644 // The alignment restriction ensures that we do not need any 4645 // synchronization with other gang-workers while setting or 4646 // clearing bits in thus chunk of the MUT. 4647 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4648 start_addr + (nth_task+1)*chunk_size); 4649 // The last chunk's end might be way beyond end of the 4650 // used region. In that case pull back appropriately. 4651 if (this_span.end() > end_addr) { 4652 this_span.set_end(end_addr); 4653 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4654 } 4655 // Iterate over the dirty cards covering this chunk, marking them 4656 // precleaned, and setting the corresponding bits in the mod union 4657 // table. Since we have been careful to partition at Card and MUT-word 4658 // boundaries no synchronization is needed between parallel threads. 4659 _collector->_ct->dirty_card_iterate(this_span, 4660 &modUnionClosure); 4661 4662 // Having transferred these marks into the modUnionTable, 4663 // rescan the marked objects on the dirty cards in the modUnionTable. 4664 // Even if this is at a synchronous collection, the initial marking 4665 // may have been done during an asynchronous collection so there 4666 // may be dirty bits in the mod-union table. 4667 _collector->_modUnionTable.dirty_range_iterate_clear( 4668 this_span, &greyRescanClosure); 4669 _collector->_modUnionTable.verifyNoOneBitsInRange( 4670 this_span.start(), 4671 this_span.end()); 4672 } 4673 pst->all_tasks_completed(); // declare that i am done 4674 } 4675 4676 // . see if we can share work_queues with ParNew? XXX 4677 void 4678 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4679 OopTaskQueue* work_q = work_queue(i); 4680 NOT_PRODUCT(int num_steals = 0;) 4681 oop obj_to_scan; 4682 CMSBitMap* bm = &(_collector->_markBitMap); 4683 4684 while (true) { 4685 // Completely finish any left over work from (an) earlier round(s) 4686 cl->trim_queue(0); 4687 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4688 (size_t)ParGCDesiredObjsFromOverflowList); 4689 // Now check if there's any work in the overflow list 4690 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4691 // only affects the number of attempts made to get work from the 4692 // overflow list and does not affect the number of workers. Just 4693 // pass ParallelGCThreads so this behavior is unchanged. 4694 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4695 work_q, 4696 ParallelGCThreads)) { 4697 // found something in global overflow list; 4698 // not yet ready to go stealing work from others. 4699 // We'd like to assert(work_q->size() != 0, ...) 4700 // because we just took work from the overflow list, 4701 // but of course we can't since all of that could have 4702 // been already stolen from us. 4703 // "He giveth and He taketh away." 4704 continue; 4705 } 4706 // Verify that we have no work before we resort to stealing 4707 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4708 // Try to steal from other queues that have work 4709 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4710 NOT_PRODUCT(num_steals++;) 4711 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4712 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4713 // Do scanning work 4714 obj_to_scan->oop_iterate(cl); 4715 // Loop around, finish this work, and try to steal some more 4716 } else if (terminator()->offer_termination()) { 4717 break; // nirvana from the infinite cycle 4718 } 4719 } 4720 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4721 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4722 "Else our work is not yet done"); 4723 } 4724 4725 // Record object boundaries in _eden_chunk_array by sampling the eden 4726 // top in the slow-path eden object allocation code path and record 4727 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4728 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4729 // sampling in sample_eden() that activates during the part of the 4730 // preclean phase. 4731 void CMSCollector::sample_eden_chunk() { 4732 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4733 if (_eden_chunk_lock->try_lock()) { 4734 // Record a sample. This is the critical section. The contents 4735 // of the _eden_chunk_array have to be non-decreasing in the 4736 // address order. 4737 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4738 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4739 "Unexpected state of Eden"); 4740 if (_eden_chunk_index == 0 || 4741 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4742 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4743 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4744 _eden_chunk_index++; // commit sample 4745 } 4746 _eden_chunk_lock->unlock(); 4747 } 4748 } 4749 } 4750 4751 // Return a thread-local PLAB recording array, as appropriate. 4752 void* CMSCollector::get_data_recorder(int thr_num) { 4753 if (_survivor_plab_array != NULL && 4754 (CMSPLABRecordAlways || 4755 (_collectorState > Marking && _collectorState < FinalMarking))) { 4756 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4757 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4758 ca->reset(); // clear it so that fresh data is recorded 4759 return (void*) ca; 4760 } else { 4761 return NULL; 4762 } 4763 } 4764 4765 // Reset all the thread-local PLAB recording arrays 4766 void CMSCollector::reset_survivor_plab_arrays() { 4767 for (uint i = 0; i < ParallelGCThreads; i++) { 4768 _survivor_plab_array[i].reset(); 4769 } 4770 } 4771 4772 // Merge the per-thread plab arrays into the global survivor chunk 4773 // array which will provide the partitioning of the survivor space 4774 // for CMS initial scan and rescan. 4775 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4776 int no_of_gc_threads) { 4777 assert(_survivor_plab_array != NULL, "Error"); 4778 assert(_survivor_chunk_array != NULL, "Error"); 4779 assert(_collectorState == FinalMarking || 4780 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4781 for (int j = 0; j < no_of_gc_threads; j++) { 4782 _cursor[j] = 0; 4783 } 4784 HeapWord* top = surv->top(); 4785 size_t i; 4786 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4787 HeapWord* min_val = top; // Higher than any PLAB address 4788 uint min_tid = 0; // position of min_val this round 4789 for (int j = 0; j < no_of_gc_threads; j++) { 4790 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4791 if (_cursor[j] == cur_sca->end()) { 4792 continue; 4793 } 4794 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4795 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4796 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4797 if (cur_val < min_val) { 4798 min_tid = j; 4799 min_val = cur_val; 4800 } else { 4801 assert(cur_val < top, "All recorded addresses should be less"); 4802 } 4803 } 4804 // At this point min_val and min_tid are respectively 4805 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4806 // and the thread (j) that witnesses that address. 4807 // We record this address in the _survivor_chunk_array[i] 4808 // and increment _cursor[min_tid] prior to the next round i. 4809 if (min_val == top) { 4810 break; 4811 } 4812 _survivor_chunk_array[i] = min_val; 4813 _cursor[min_tid]++; 4814 } 4815 // We are all done; record the size of the _survivor_chunk_array 4816 _survivor_chunk_index = i; // exclusive: [0, i) 4817 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4818 // Verify that we used up all the recorded entries 4819 #ifdef ASSERT 4820 size_t total = 0; 4821 for (int j = 0; j < no_of_gc_threads; j++) { 4822 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4823 total += _cursor[j]; 4824 } 4825 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4826 // Check that the merged array is in sorted order 4827 if (total > 0) { 4828 for (size_t i = 0; i < total - 1; i++) { 4829 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4830 i, p2i(_survivor_chunk_array[i])); 4831 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4832 "Not sorted"); 4833 } 4834 } 4835 #endif // ASSERT 4836 } 4837 4838 // Set up the space's par_seq_tasks structure for work claiming 4839 // for parallel initial scan and rescan of young gen. 4840 // See ParRescanTask where this is currently used. 4841 void 4842 CMSCollector:: 4843 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4844 assert(n_threads > 0, "Unexpected n_threads argument"); 4845 4846 // Eden space 4847 if (!_young_gen->eden()->is_empty()) { 4848 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4849 assert(!pst->valid(), "Clobbering existing data?"); 4850 // Each valid entry in [0, _eden_chunk_index) represents a task. 4851 size_t n_tasks = _eden_chunk_index + 1; 4852 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4853 // Sets the condition for completion of the subtask (how many threads 4854 // need to finish in order to be done). 4855 pst->set_n_threads(n_threads); 4856 pst->set_n_tasks((int)n_tasks); 4857 } 4858 4859 // Merge the survivor plab arrays into _survivor_chunk_array 4860 if (_survivor_plab_array != NULL) { 4861 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4862 } else { 4863 assert(_survivor_chunk_index == 0, "Error"); 4864 } 4865 4866 // To space 4867 { 4868 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4869 assert(!pst->valid(), "Clobbering existing data?"); 4870 // Sets the condition for completion of the subtask (how many threads 4871 // need to finish in order to be done). 4872 pst->set_n_threads(n_threads); 4873 pst->set_n_tasks(1); 4874 assert(pst->valid(), "Error"); 4875 } 4876 4877 // From space 4878 { 4879 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4880 assert(!pst->valid(), "Clobbering existing data?"); 4881 size_t n_tasks = _survivor_chunk_index + 1; 4882 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4883 // Sets the condition for completion of the subtask (how many threads 4884 // need to finish in order to be done). 4885 pst->set_n_threads(n_threads); 4886 pst->set_n_tasks((int)n_tasks); 4887 assert(pst->valid(), "Error"); 4888 } 4889 } 4890 4891 // Parallel version of remark 4892 void CMSCollector::do_remark_parallel() { 4893 CMSHeap* heap = CMSHeap::heap(); 4894 WorkGang* workers = heap->workers(); 4895 assert(workers != NULL, "Need parallel worker threads."); 4896 // Choose to use the number of GC workers most recently set 4897 // into "active_workers". 4898 uint n_workers = workers->active_workers(); 4899 4900 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4901 4902 StrongRootsScope srs(n_workers); 4903 4904 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4905 4906 // We won't be iterating over the cards in the card table updating 4907 // the younger_gen cards, so we shouldn't call the following else 4908 // the verification code as well as subsequent younger_refs_iterate 4909 // code would get confused. XXX 4910 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4911 4912 // The young gen rescan work will not be done as part of 4913 // process_roots (which currently doesn't know how to 4914 // parallelize such a scan), but rather will be broken up into 4915 // a set of parallel tasks (via the sampling that the [abortable] 4916 // preclean phase did of eden, plus the [two] tasks of 4917 // scanning the [two] survivor spaces. Further fine-grain 4918 // parallelization of the scanning of the survivor spaces 4919 // themselves, and of precleaning of the young gen itself 4920 // is deferred to the future. 4921 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4922 4923 // The dirty card rescan work is broken up into a "sequence" 4924 // of parallel tasks (per constituent space) that are dynamically 4925 // claimed by the parallel threads. 4926 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4927 4928 // It turns out that even when we're using 1 thread, doing the work in a 4929 // separate thread causes wide variance in run times. We can't help this 4930 // in the multi-threaded case, but we special-case n=1 here to get 4931 // repeatable measurements of the 1-thread overhead of the parallel code. 4932 if (n_workers > 1) { 4933 // Make refs discovery MT-safe, if it isn't already: it may not 4934 // necessarily be so, since it's possible that we are doing 4935 // ST marking. 4936 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4937 workers->run_task(&tsk); 4938 } else { 4939 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4940 tsk.work(0); 4941 } 4942 4943 // restore, single-threaded for now, any preserved marks 4944 // as a result of work_q overflow 4945 restore_preserved_marks_if_any(); 4946 } 4947 4948 // Non-parallel version of remark 4949 void CMSCollector::do_remark_non_parallel() { 4950 ResourceMark rm; 4951 HandleMark hm; 4952 CMSHeap* heap = CMSHeap::heap(); 4953 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4954 4955 MarkRefsIntoAndScanClosure 4956 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4957 &_markStack, this, 4958 false /* should_yield */, false /* not precleaning */); 4959 MarkFromDirtyCardsClosure 4960 markFromDirtyCardsClosure(this, _span, 4961 NULL, // space is set further below 4962 &_markBitMap, &_markStack, &mrias_cl); 4963 { 4964 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4965 // Iterate over the dirty cards, setting the corresponding bits in the 4966 // mod union table. 4967 { 4968 ModUnionClosure modUnionClosure(&_modUnionTable); 4969 _ct->dirty_card_iterate(_cmsGen->used_region(), 4970 &modUnionClosure); 4971 } 4972 // Having transferred these marks into the modUnionTable, we just need 4973 // to rescan the marked objects on the dirty cards in the modUnionTable. 4974 // The initial marking may have been done during an asynchronous 4975 // collection so there may be dirty bits in the mod-union table. 4976 const int alignment = CardTable::card_size * BitsPerWord; 4977 { 4978 // ... First handle dirty cards in CMS gen 4979 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4980 MemRegion ur = _cmsGen->used_region(); 4981 HeapWord* lb = ur.start(); 4982 HeapWord* ub = align_up(ur.end(), alignment); 4983 MemRegion cms_span(lb, ub); 4984 _modUnionTable.dirty_range_iterate_clear(cms_span, 4985 &markFromDirtyCardsClosure); 4986 verify_work_stacks_empty(); 4987 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4988 } 4989 } 4990 if (VerifyDuringGC && 4991 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4992 HandleMark hm; // Discard invalid handles created during verification 4993 Universe::verify(); 4994 } 4995 { 4996 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4997 4998 verify_work_stacks_empty(); 4999 5000 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5001 StrongRootsScope srs(1); 5002 5003 heap->cms_process_roots(&srs, 5004 true, // young gen as roots 5005 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5006 should_unload_classes(), 5007 &mrias_cl, 5008 NULL); // The dirty klasses will be handled below 5009 5010 assert(should_unload_classes() 5011 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5012 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5013 } 5014 5015 { 5016 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 5017 5018 verify_work_stacks_empty(); 5019 5020 // Scan all class loader data objects that might have been introduced 5021 // during concurrent marking. 5022 ResourceMark rm; 5023 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5024 for (int i = 0; i < array->length(); i++) { 5025 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 5026 } 5027 5028 // We don't need to keep track of new CLDs anymore. 5029 ClassLoaderDataGraph::remember_new_clds(false); 5030 5031 verify_work_stacks_empty(); 5032 } 5033 5034 // We might have added oops to ClassLoaderData::_handles during the 5035 // concurrent marking phase. These oops do not point to newly allocated objects 5036 // that are guaranteed to be kept alive. Hence, 5037 // we do have to revisit the _handles block during the remark phase. 5038 { 5039 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 5040 5041 verify_work_stacks_empty(); 5042 5043 RemarkCLDClosure remark_closure(&mrias_cl); 5044 ClassLoaderDataGraph::cld_do(&remark_closure); 5045 5046 verify_work_stacks_empty(); 5047 } 5048 5049 verify_work_stacks_empty(); 5050 // Restore evacuated mark words, if any, used for overflow list links 5051 restore_preserved_marks_if_any(); 5052 5053 verify_overflow_empty(); 5054 } 5055 5056 //////////////////////////////////////////////////////// 5057 // Parallel Reference Processing Task Proxy Class 5058 //////////////////////////////////////////////////////// 5059 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5060 OopTaskQueueSet* _queues; 5061 TaskTerminator _terminator; 5062 public: 5063 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5064 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5065 ParallelTaskTerminator* terminator() { return _terminator.terminator(); } 5066 OopTaskQueueSet* queues() { return _queues; } 5067 }; 5068 5069 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5070 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5071 CMSCollector* _collector; 5072 CMSBitMap* _mark_bit_map; 5073 const MemRegion _span; 5074 ProcessTask& _task; 5075 5076 public: 5077 CMSRefProcTaskProxy(ProcessTask& task, 5078 CMSCollector* collector, 5079 const MemRegion& span, 5080 CMSBitMap* mark_bit_map, 5081 AbstractWorkGang* workers, 5082 OopTaskQueueSet* task_queues): 5083 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5084 task_queues, 5085 workers->active_workers()), 5086 _collector(collector), 5087 _mark_bit_map(mark_bit_map), 5088 _span(span), 5089 _task(task) 5090 { 5091 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5092 "Inconsistency in _span"); 5093 } 5094 5095 OopTaskQueueSet* task_queues() { return queues(); } 5096 5097 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5098 5099 void do_work_steal(int i, 5100 CMSParDrainMarkingStackClosure* drain, 5101 CMSParKeepAliveClosure* keep_alive); 5102 5103 virtual void work(uint worker_id); 5104 }; 5105 5106 void CMSRefProcTaskProxy::work(uint worker_id) { 5107 ResourceMark rm; 5108 HandleMark hm; 5109 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5110 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5111 _mark_bit_map, 5112 work_queue(worker_id)); 5113 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5114 _mark_bit_map, 5115 work_queue(worker_id)); 5116 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5117 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5118 if (_task.marks_oops_alive()) { 5119 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5120 } 5121 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5122 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5123 } 5124 5125 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5126 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5127 _span(span), 5128 _work_queue(work_queue), 5129 _bit_map(bit_map), 5130 _mark_and_push(collector, span, bit_map, work_queue), 5131 _low_water_mark(MIN2((work_queue->max_elems()/4), 5132 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5133 { } 5134 5135 // . see if we can share work_queues with ParNew? XXX 5136 void CMSRefProcTaskProxy::do_work_steal(int i, 5137 CMSParDrainMarkingStackClosure* drain, 5138 CMSParKeepAliveClosure* keep_alive) { 5139 OopTaskQueue* work_q = work_queue(i); 5140 NOT_PRODUCT(int num_steals = 0;) 5141 oop obj_to_scan; 5142 5143 while (true) { 5144 // Completely finish any left over work from (an) earlier round(s) 5145 drain->trim_queue(0); 5146 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5147 (size_t)ParGCDesiredObjsFromOverflowList); 5148 // Now check if there's any work in the overflow list 5149 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5150 // only affects the number of attempts made to get work from the 5151 // overflow list and does not affect the number of workers. Just 5152 // pass ParallelGCThreads so this behavior is unchanged. 5153 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5154 work_q, 5155 ParallelGCThreads)) { 5156 // Found something in global overflow list; 5157 // not yet ready to go stealing work from others. 5158 // We'd like to assert(work_q->size() != 0, ...) 5159 // because we just took work from the overflow list, 5160 // but of course we can't, since all of that might have 5161 // been already stolen from us. 5162 continue; 5163 } 5164 // Verify that we have no work before we resort to stealing 5165 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5166 // Try to steal from other queues that have work 5167 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5168 NOT_PRODUCT(num_steals++;) 5169 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5170 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5171 // Do scanning work 5172 obj_to_scan->oop_iterate(keep_alive); 5173 // Loop around, finish this work, and try to steal some more 5174 } else if (terminator()->offer_termination()) { 5175 break; // nirvana from the infinite cycle 5176 } 5177 } 5178 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5179 } 5180 5181 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5182 CMSHeap* heap = CMSHeap::heap(); 5183 WorkGang* workers = heap->workers(); 5184 assert(workers != NULL, "Need parallel worker threads."); 5185 assert(workers->active_workers() == ergo_workers, 5186 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5187 ergo_workers, workers->active_workers()); 5188 CMSRefProcTaskProxy rp_task(task, &_collector, 5189 _collector.ref_processor_span(), 5190 _collector.markBitMap(), 5191 workers, _collector.task_queues()); 5192 workers->run_task(&rp_task, workers->active_workers()); 5193 } 5194 5195 void CMSCollector::refProcessingWork() { 5196 ResourceMark rm; 5197 HandleMark hm; 5198 5199 ReferenceProcessor* rp = ref_processor(); 5200 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5201 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5202 // Process weak references. 5203 rp->setup_policy(false); 5204 verify_work_stacks_empty(); 5205 5206 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5207 { 5208 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5209 5210 // Setup keep_alive and complete closures. 5211 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5212 &_markStack, false /* !preclean */); 5213 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5214 _span, &_markBitMap, &_markStack, 5215 &cmsKeepAliveClosure, false /* !preclean */); 5216 5217 ReferenceProcessorStats stats; 5218 if (rp->processing_is_mt()) { 5219 // Set the degree of MT here. If the discovery is done MT, there 5220 // may have been a different number of threads doing the discovery 5221 // and a different number of discovered lists may have Ref objects. 5222 // That is OK as long as the Reference lists are balanced (see 5223 // balance_all_queues() and balance_queues()). 5224 CMSHeap* heap = CMSHeap::heap(); 5225 uint active_workers = ParallelGCThreads; 5226 WorkGang* workers = heap->workers(); 5227 if (workers != NULL) { 5228 active_workers = workers->active_workers(); 5229 // The expectation is that active_workers will have already 5230 // been set to a reasonable value. If it has not been set, 5231 // investigate. 5232 assert(active_workers > 0, "Should have been set during scavenge"); 5233 } 5234 rp->set_active_mt_degree(active_workers); 5235 CMSRefProcTaskExecutor task_executor(*this); 5236 stats = rp->process_discovered_references(&_is_alive_closure, 5237 &cmsKeepAliveClosure, 5238 &cmsDrainMarkingStackClosure, 5239 &task_executor, 5240 &pt); 5241 } else { 5242 stats = rp->process_discovered_references(&_is_alive_closure, 5243 &cmsKeepAliveClosure, 5244 &cmsDrainMarkingStackClosure, 5245 NULL, 5246 &pt); 5247 } 5248 _gc_tracer_cm->report_gc_reference_stats(stats); 5249 pt.print_all_references(); 5250 } 5251 5252 // This is the point where the entire marking should have completed. 5253 verify_work_stacks_empty(); 5254 5255 { 5256 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5257 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5258 } 5259 5260 if (should_unload_classes()) { 5261 { 5262 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5263 5264 // Unload classes and purge the SystemDictionary. 5265 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5266 5267 // Unload nmethods. 5268 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5269 5270 // Prune dead klasses from subklass/sibling/implementor lists. 5271 Klass::clean_weak_klass_links(purged_class); 5272 } 5273 5274 { 5275 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5276 // Clean up unreferenced symbols in symbol table. 5277 SymbolTable::unlink(); 5278 } 5279 5280 { 5281 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5282 // Delete entries for dead interned strings. 5283 StringTable::unlink(&_is_alive_closure); 5284 } 5285 } 5286 5287 // Restore any preserved marks as a result of mark stack or 5288 // work queue overflow 5289 restore_preserved_marks_if_any(); // done single-threaded for now 5290 5291 rp->set_enqueuing_is_done(true); 5292 rp->verify_no_references_recorded(); 5293 } 5294 5295 #ifndef PRODUCT 5296 void CMSCollector::check_correct_thread_executing() { 5297 Thread* t = Thread::current(); 5298 // Only the VM thread or the CMS thread should be here. 5299 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5300 "Unexpected thread type"); 5301 // If this is the vm thread, the foreground process 5302 // should not be waiting. Note that _foregroundGCIsActive is 5303 // true while the foreground collector is waiting. 5304 if (_foregroundGCShouldWait) { 5305 // We cannot be the VM thread 5306 assert(t->is_ConcurrentGC_thread(), 5307 "Should be CMS thread"); 5308 } else { 5309 // We can be the CMS thread only if we are in a stop-world 5310 // phase of CMS collection. 5311 if (t->is_ConcurrentGC_thread()) { 5312 assert(_collectorState == InitialMarking || 5313 _collectorState == FinalMarking, 5314 "Should be a stop-world phase"); 5315 // The CMS thread should be holding the CMS_token. 5316 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5317 "Potential interference with concurrently " 5318 "executing VM thread"); 5319 } 5320 } 5321 } 5322 #endif 5323 5324 void CMSCollector::sweep() { 5325 assert(_collectorState == Sweeping, "just checking"); 5326 check_correct_thread_executing(); 5327 verify_work_stacks_empty(); 5328 verify_overflow_empty(); 5329 increment_sweep_count(); 5330 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5331 5332 _inter_sweep_timer.stop(); 5333 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5334 5335 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5336 _intra_sweep_timer.reset(); 5337 _intra_sweep_timer.start(); 5338 { 5339 GCTraceCPUTime tcpu; 5340 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5341 // First sweep the old gen 5342 { 5343 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5344 bitMapLock()); 5345 sweepWork(_cmsGen); 5346 } 5347 5348 // Update Universe::_heap_*_at_gc figures. 5349 // We need all the free list locks to make the abstract state 5350 // transition from Sweeping to Resetting. See detailed note 5351 // further below. 5352 { 5353 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5354 // Update heap occupancy information which is used as 5355 // input to soft ref clearing policy at the next gc. 5356 Universe::update_heap_info_at_gc(); 5357 _collectorState = Resizing; 5358 } 5359 } 5360 verify_work_stacks_empty(); 5361 verify_overflow_empty(); 5362 5363 if (should_unload_classes()) { 5364 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5365 // requires that the virtual spaces are stable and not deleted. 5366 ClassLoaderDataGraph::set_should_purge(true); 5367 } 5368 5369 _intra_sweep_timer.stop(); 5370 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5371 5372 _inter_sweep_timer.reset(); 5373 _inter_sweep_timer.start(); 5374 5375 // We need to use a monotonically non-decreasing time in ms 5376 // or we will see time-warp warnings and os::javaTimeMillis() 5377 // does not guarantee monotonicity. 5378 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5379 update_time_of_last_gc(now); 5380 5381 // NOTE on abstract state transitions: 5382 // Mutators allocate-live and/or mark the mod-union table dirty 5383 // based on the state of the collection. The former is done in 5384 // the interval [Marking, Sweeping] and the latter in the interval 5385 // [Marking, Sweeping). Thus the transitions into the Marking state 5386 // and out of the Sweeping state must be synchronously visible 5387 // globally to the mutators. 5388 // The transition into the Marking state happens with the world 5389 // stopped so the mutators will globally see it. Sweeping is 5390 // done asynchronously by the background collector so the transition 5391 // from the Sweeping state to the Resizing state must be done 5392 // under the freelistLock (as is the check for whether to 5393 // allocate-live and whether to dirty the mod-union table). 5394 assert(_collectorState == Resizing, "Change of collector state to" 5395 " Resizing must be done under the freelistLocks (plural)"); 5396 5397 // Now that sweeping has been completed, we clear 5398 // the incremental_collection_failed flag, 5399 // thus inviting a younger gen collection to promote into 5400 // this generation. If such a promotion may still fail, 5401 // the flag will be set again when a young collection is 5402 // attempted. 5403 CMSHeap* heap = CMSHeap::heap(); 5404 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5405 heap->update_full_collections_completed(_collection_count_start); 5406 } 5407 5408 // FIX ME!!! Looks like this belongs in CFLSpace, with 5409 // CMSGen merely delegating to it. 5410 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5411 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5412 HeapWord* minAddr = _cmsSpace->bottom(); 5413 HeapWord* largestAddr = 5414 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5415 if (largestAddr == NULL) { 5416 // The dictionary appears to be empty. In this case 5417 // try to coalesce at the end of the heap. 5418 largestAddr = _cmsSpace->end(); 5419 } 5420 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5421 size_t nearLargestOffset = 5422 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5423 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5424 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5425 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5426 } 5427 5428 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5429 return addr >= _cmsSpace->nearLargestChunk(); 5430 } 5431 5432 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5433 return _cmsSpace->find_chunk_at_end(); 5434 } 5435 5436 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5437 bool full) { 5438 // If the young generation has been collected, gather any statistics 5439 // that are of interest at this point. 5440 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5441 if (!full && current_is_young) { 5442 // Gather statistics on the young generation collection. 5443 collector()->stats().record_gc0_end(used()); 5444 } 5445 } 5446 5447 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5448 // We iterate over the space(s) underlying this generation, 5449 // checking the mark bit map to see if the bits corresponding 5450 // to specific blocks are marked or not. Blocks that are 5451 // marked are live and are not swept up. All remaining blocks 5452 // are swept up, with coalescing on-the-fly as we sweep up 5453 // contiguous free and/or garbage blocks: 5454 // We need to ensure that the sweeper synchronizes with allocators 5455 // and stop-the-world collectors. In particular, the following 5456 // locks are used: 5457 // . CMS token: if this is held, a stop the world collection cannot occur 5458 // . freelistLock: if this is held no allocation can occur from this 5459 // generation by another thread 5460 // . bitMapLock: if this is held, no other thread can access or update 5461 // 5462 5463 // Note that we need to hold the freelistLock if we use 5464 // block iterate below; else the iterator might go awry if 5465 // a mutator (or promotion) causes block contents to change 5466 // (for instance if the allocator divvies up a block). 5467 // If we hold the free list lock, for all practical purposes 5468 // young generation GC's can't occur (they'll usually need to 5469 // promote), so we might as well prevent all young generation 5470 // GC's while we do a sweeping step. For the same reason, we might 5471 // as well take the bit map lock for the entire duration 5472 5473 // check that we hold the requisite locks 5474 assert(have_cms_token(), "Should hold cms token"); 5475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5476 assert_lock_strong(old_gen->freelistLock()); 5477 assert_lock_strong(bitMapLock()); 5478 5479 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5480 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5481 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5482 _inter_sweep_estimate.padded_average(), 5483 _intra_sweep_estimate.padded_average()); 5484 old_gen->setNearLargestChunk(); 5485 5486 { 5487 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5488 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5489 // We need to free-up/coalesce garbage/blocks from a 5490 // co-terminal free run. This is done in the SweepClosure 5491 // destructor; so, do not remove this scope, else the 5492 // end-of-sweep-census below will be off by a little bit. 5493 } 5494 old_gen->cmsSpace()->sweep_completed(); 5495 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5496 if (should_unload_classes()) { // unloaded classes this cycle, 5497 _concurrent_cycles_since_last_unload = 0; // ... reset count 5498 } else { // did not unload classes, 5499 _concurrent_cycles_since_last_unload++; // ... increment count 5500 } 5501 } 5502 5503 // Reset CMS data structures (for now just the marking bit map) 5504 // preparatory for the next cycle. 5505 void CMSCollector::reset_concurrent() { 5506 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5507 5508 // If the state is not "Resetting", the foreground thread 5509 // has done a collection and the resetting. 5510 if (_collectorState != Resetting) { 5511 assert(_collectorState == Idling, "The state should only change" 5512 " because the foreground collector has finished the collection"); 5513 return; 5514 } 5515 5516 { 5517 // Clear the mark bitmap (no grey objects to start with) 5518 // for the next cycle. 5519 GCTraceCPUTime tcpu; 5520 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5521 5522 HeapWord* curAddr = _markBitMap.startWord(); 5523 while (curAddr < _markBitMap.endWord()) { 5524 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5525 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5526 _markBitMap.clear_large_range(chunk); 5527 if (ConcurrentMarkSweepThread::should_yield() && 5528 !foregroundGCIsActive() && 5529 CMSYield) { 5530 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5531 "CMS thread should hold CMS token"); 5532 assert_lock_strong(bitMapLock()); 5533 bitMapLock()->unlock(); 5534 ConcurrentMarkSweepThread::desynchronize(true); 5535 stopTimer(); 5536 incrementYields(); 5537 5538 // See the comment in coordinator_yield() 5539 for (unsigned i = 0; i < CMSYieldSleepCount && 5540 ConcurrentMarkSweepThread::should_yield() && 5541 !CMSCollector::foregroundGCIsActive(); ++i) { 5542 os::sleep(Thread::current(), 1, false); 5543 } 5544 5545 ConcurrentMarkSweepThread::synchronize(true); 5546 bitMapLock()->lock_without_safepoint_check(); 5547 startTimer(); 5548 } 5549 curAddr = chunk.end(); 5550 } 5551 // A successful mostly concurrent collection has been done. 5552 // Because only the full (i.e., concurrent mode failure) collections 5553 // are being measured for gc overhead limits, clean the "near" flag 5554 // and count. 5555 size_policy()->reset_gc_overhead_limit_count(); 5556 _collectorState = Idling; 5557 } 5558 5559 register_gc_end(); 5560 } 5561 5562 // Same as above but for STW paths 5563 void CMSCollector::reset_stw() { 5564 // already have the lock 5565 assert(_collectorState == Resetting, "just checking"); 5566 assert_lock_strong(bitMapLock()); 5567 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5568 _markBitMap.clear_all(); 5569 _collectorState = Idling; 5570 register_gc_end(); 5571 } 5572 5573 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5574 GCTraceCPUTime tcpu; 5575 TraceCollectorStats tcs_cgc(cgc_counters()); 5576 5577 switch (op) { 5578 case CMS_op_checkpointRootsInitial: { 5579 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5580 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5581 checkpointRootsInitial(); 5582 break; 5583 } 5584 case CMS_op_checkpointRootsFinal: { 5585 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5586 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5587 checkpointRootsFinal(); 5588 break; 5589 } 5590 default: 5591 fatal("No such CMS_op"); 5592 } 5593 } 5594 5595 #ifndef PRODUCT 5596 size_t const CMSCollector::skip_header_HeapWords() { 5597 return FreeChunk::header_size(); 5598 } 5599 5600 // Try and collect here conditions that should hold when 5601 // CMS thread is exiting. The idea is that the foreground GC 5602 // thread should not be blocked if it wants to terminate 5603 // the CMS thread and yet continue to run the VM for a while 5604 // after that. 5605 void CMSCollector::verify_ok_to_terminate() const { 5606 assert(Thread::current()->is_ConcurrentGC_thread(), 5607 "should be called by CMS thread"); 5608 assert(!_foregroundGCShouldWait, "should be false"); 5609 // We could check here that all the various low-level locks 5610 // are not held by the CMS thread, but that is overkill; see 5611 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5612 // is checked. 5613 } 5614 #endif 5615 5616 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5617 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5618 "missing Printezis mark?"); 5619 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5620 size_t size = pointer_delta(nextOneAddr + 1, addr); 5621 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5622 "alignment problem"); 5623 assert(size >= 3, "Necessary for Printezis marks to work"); 5624 return size; 5625 } 5626 5627 // A variant of the above (block_size_using_printezis_bits()) except 5628 // that we return 0 if the P-bits are not yet set. 5629 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5630 if (_markBitMap.isMarked(addr + 1)) { 5631 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5632 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5633 size_t size = pointer_delta(nextOneAddr + 1, addr); 5634 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5635 "alignment problem"); 5636 assert(size >= 3, "Necessary for Printezis marks to work"); 5637 return size; 5638 } 5639 return 0; 5640 } 5641 5642 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5643 size_t sz = 0; 5644 oop p = (oop)addr; 5645 if (p->klass_or_null_acquire() != NULL) { 5646 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5647 } else { 5648 sz = block_size_using_printezis_bits(addr); 5649 } 5650 assert(sz > 0, "size must be nonzero"); 5651 HeapWord* next_block = addr + sz; 5652 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5653 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5654 align_down((uintptr_t)next_card, CardTable::card_size), 5655 "must be different cards"); 5656 return next_card; 5657 } 5658 5659 5660 // CMS Bit Map Wrapper ///////////////////////////////////////// 5661 5662 // Construct a CMS bit map infrastructure, but don't create the 5663 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5664 // further below. 5665 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5666 _shifter(shifter), 5667 _bm(), 5668 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5669 Monitor::_safepoint_check_sometimes) : NULL) 5670 { 5671 _bmStartWord = 0; 5672 _bmWordSize = 0; 5673 } 5674 5675 bool CMSBitMap::allocate(MemRegion mr) { 5676 _bmStartWord = mr.start(); 5677 _bmWordSize = mr.word_size(); 5678 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5679 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5680 if (!brs.is_reserved()) { 5681 log_warning(gc)("CMS bit map allocation failure"); 5682 return false; 5683 } 5684 // For now we'll just commit all of the bit map up front. 5685 // Later on we'll try to be more parsimonious with swap. 5686 if (!_virtual_space.initialize(brs, brs.size())) { 5687 log_warning(gc)("CMS bit map backing store failure"); 5688 return false; 5689 } 5690 assert(_virtual_space.committed_size() == brs.size(), 5691 "didn't reserve backing store for all of CMS bit map?"); 5692 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5693 _bmWordSize, "inconsistency in bit map sizing"); 5694 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5695 5696 // bm.clear(); // can we rely on getting zero'd memory? verify below 5697 assert(isAllClear(), 5698 "Expected zero'd memory from ReservedSpace constructor"); 5699 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5700 "consistency check"); 5701 return true; 5702 } 5703 5704 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5705 HeapWord *next_addr, *end_addr, *last_addr; 5706 assert_locked(); 5707 assert(covers(mr), "out-of-range error"); 5708 // XXX assert that start and end are appropriately aligned 5709 for (next_addr = mr.start(), end_addr = mr.end(); 5710 next_addr < end_addr; next_addr = last_addr) { 5711 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5712 last_addr = dirty_region.end(); 5713 if (!dirty_region.is_empty()) { 5714 cl->do_MemRegion(dirty_region); 5715 } else { 5716 assert(last_addr == end_addr, "program logic"); 5717 return; 5718 } 5719 } 5720 } 5721 5722 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5723 _bm.print_on_error(st, prefix); 5724 } 5725 5726 #ifndef PRODUCT 5727 void CMSBitMap::assert_locked() const { 5728 CMSLockVerifier::assert_locked(lock()); 5729 } 5730 5731 bool CMSBitMap::covers(MemRegion mr) const { 5732 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5733 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5734 "size inconsistency"); 5735 return (mr.start() >= _bmStartWord) && 5736 (mr.end() <= endWord()); 5737 } 5738 5739 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5740 return (start >= _bmStartWord && (start + size) <= endWord()); 5741 } 5742 5743 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5744 // verify that there are no 1 bits in the interval [left, right) 5745 FalseBitMapClosure falseBitMapClosure; 5746 iterate(&falseBitMapClosure, left, right); 5747 } 5748 5749 void CMSBitMap::region_invariant(MemRegion mr) 5750 { 5751 assert_locked(); 5752 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5753 assert(!mr.is_empty(), "unexpected empty region"); 5754 assert(covers(mr), "mr should be covered by bit map"); 5755 // convert address range into offset range 5756 size_t start_ofs = heapWordToOffset(mr.start()); 5757 // Make sure that end() is appropriately aligned 5758 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5759 "Misaligned mr.end()"); 5760 size_t end_ofs = heapWordToOffset(mr.end()); 5761 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5762 } 5763 5764 #endif 5765 5766 bool CMSMarkStack::allocate(size_t size) { 5767 // allocate a stack of the requisite depth 5768 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5769 size * sizeof(oop))); 5770 if (!rs.is_reserved()) { 5771 log_warning(gc)("CMSMarkStack allocation failure"); 5772 return false; 5773 } 5774 if (!_virtual_space.initialize(rs, rs.size())) { 5775 log_warning(gc)("CMSMarkStack backing store failure"); 5776 return false; 5777 } 5778 assert(_virtual_space.committed_size() == rs.size(), 5779 "didn't reserve backing store for all of CMS stack?"); 5780 _base = (oop*)(_virtual_space.low()); 5781 _index = 0; 5782 _capacity = size; 5783 NOT_PRODUCT(_max_depth = 0); 5784 return true; 5785 } 5786 5787 // XXX FIX ME !!! In the MT case we come in here holding a 5788 // leaf lock. For printing we need to take a further lock 5789 // which has lower rank. We need to recalibrate the two 5790 // lock-ranks involved in order to be able to print the 5791 // messages below. (Or defer the printing to the caller. 5792 // For now we take the expedient path of just disabling the 5793 // messages for the problematic case.) 5794 void CMSMarkStack::expand() { 5795 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5796 if (_capacity == MarkStackSizeMax) { 5797 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5798 // We print a warning message only once per CMS cycle. 5799 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5800 } 5801 return; 5802 } 5803 // Double capacity if possible 5804 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5805 // Do not give up existing stack until we have managed to 5806 // get the double capacity that we desired. 5807 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5808 new_capacity * sizeof(oop))); 5809 if (rs.is_reserved()) { 5810 // Release the backing store associated with old stack 5811 _virtual_space.release(); 5812 // Reinitialize virtual space for new stack 5813 if (!_virtual_space.initialize(rs, rs.size())) { 5814 fatal("Not enough swap for expanded marking stack"); 5815 } 5816 _base = (oop*)(_virtual_space.low()); 5817 _index = 0; 5818 _capacity = new_capacity; 5819 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5820 // Failed to double capacity, continue; 5821 // we print a detail message only once per CMS cycle. 5822 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5823 _capacity / K, new_capacity / K); 5824 } 5825 } 5826 5827 5828 // Closures 5829 // XXX: there seems to be a lot of code duplication here; 5830 // should refactor and consolidate common code. 5831 5832 // This closure is used to mark refs into the CMS generation in 5833 // the CMS bit map. Called at the first checkpoint. This closure 5834 // assumes that we do not need to re-mark dirty cards; if the CMS 5835 // generation on which this is used is not an oldest 5836 // generation then this will lose younger_gen cards! 5837 5838 MarkRefsIntoClosure::MarkRefsIntoClosure( 5839 MemRegion span, CMSBitMap* bitMap): 5840 _span(span), 5841 _bitMap(bitMap) 5842 { 5843 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5844 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5845 } 5846 5847 void MarkRefsIntoClosure::do_oop(oop obj) { 5848 // if p points into _span, then mark corresponding bit in _markBitMap 5849 assert(oopDesc::is_oop(obj), "expected an oop"); 5850 HeapWord* addr = (HeapWord*)obj; 5851 if (_span.contains(addr)) { 5852 // this should be made more efficient 5853 _bitMap->mark(addr); 5854 } 5855 } 5856 5857 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5858 MemRegion span, CMSBitMap* bitMap): 5859 _span(span), 5860 _bitMap(bitMap) 5861 { 5862 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5863 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5864 } 5865 5866 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5867 // if p points into _span, then mark corresponding bit in _markBitMap 5868 assert(oopDesc::is_oop(obj), "expected an oop"); 5869 HeapWord* addr = (HeapWord*)obj; 5870 if (_span.contains(addr)) { 5871 // this should be made more efficient 5872 _bitMap->par_mark(addr); 5873 } 5874 } 5875 5876 // A variant of the above, used for CMS marking verification. 5877 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5878 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5879 _span(span), 5880 _verification_bm(verification_bm), 5881 _cms_bm(cms_bm) 5882 { 5883 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5884 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5885 } 5886 5887 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5888 // if p points into _span, then mark corresponding bit in _markBitMap 5889 assert(oopDesc::is_oop(obj), "expected an oop"); 5890 HeapWord* addr = (HeapWord*)obj; 5891 if (_span.contains(addr)) { 5892 _verification_bm->mark(addr); 5893 if (!_cms_bm->isMarked(addr)) { 5894 Log(gc, verify) log; 5895 ResourceMark rm; 5896 LogStream ls(log.error()); 5897 oop(addr)->print_on(&ls); 5898 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5899 fatal("... aborting"); 5900 } 5901 } 5902 } 5903 5904 ////////////////////////////////////////////////// 5905 // MarkRefsIntoAndScanClosure 5906 ////////////////////////////////////////////////// 5907 5908 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5909 ReferenceDiscoverer* rd, 5910 CMSBitMap* bit_map, 5911 CMSBitMap* mod_union_table, 5912 CMSMarkStack* mark_stack, 5913 CMSCollector* collector, 5914 bool should_yield, 5915 bool concurrent_precleaning): 5916 _span(span), 5917 _bit_map(bit_map), 5918 _mark_stack(mark_stack), 5919 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5920 mark_stack, concurrent_precleaning), 5921 _collector(collector), 5922 _freelistLock(NULL), 5923 _yield(should_yield), 5924 _concurrent_precleaning(concurrent_precleaning) 5925 { 5926 // FIXME: Should initialize in base class constructor. 5927 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5928 set_ref_discoverer_internal(rd); 5929 } 5930 5931 // This closure is used to mark refs into the CMS generation at the 5932 // second (final) checkpoint, and to scan and transitively follow 5933 // the unmarked oops. It is also used during the concurrent precleaning 5934 // phase while scanning objects on dirty cards in the CMS generation. 5935 // The marks are made in the marking bit map and the marking stack is 5936 // used for keeping the (newly) grey objects during the scan. 5937 // The parallel version (Par_...) appears further below. 5938 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5939 if (obj != NULL) { 5940 assert(oopDesc::is_oop(obj), "expected an oop"); 5941 HeapWord* addr = (HeapWord*)obj; 5942 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5943 assert(_collector->overflow_list_is_empty(), 5944 "overflow list should be empty"); 5945 if (_span.contains(addr) && 5946 !_bit_map->isMarked(addr)) { 5947 // mark bit map (object is now grey) 5948 _bit_map->mark(addr); 5949 // push on marking stack (stack should be empty), and drain the 5950 // stack by applying this closure to the oops in the oops popped 5951 // from the stack (i.e. blacken the grey objects) 5952 bool res = _mark_stack->push(obj); 5953 assert(res, "Should have space to push on empty stack"); 5954 do { 5955 oop new_oop = _mark_stack->pop(); 5956 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5957 assert(_bit_map->isMarked((HeapWord*)new_oop), 5958 "only grey objects on this stack"); 5959 // iterate over the oops in this oop, marking and pushing 5960 // the ones in CMS heap (i.e. in _span). 5961 new_oop->oop_iterate(&_pushAndMarkClosure); 5962 // check if it's time to yield 5963 do_yield_check(); 5964 } while (!_mark_stack->isEmpty() || 5965 (!_concurrent_precleaning && take_from_overflow_list())); 5966 // if marking stack is empty, and we are not doing this 5967 // during precleaning, then check the overflow list 5968 } 5969 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5970 assert(_collector->overflow_list_is_empty(), 5971 "overflow list was drained above"); 5972 5973 assert(_collector->no_preserved_marks(), 5974 "All preserved marks should have been restored above"); 5975 } 5976 } 5977 5978 void MarkRefsIntoAndScanClosure::do_yield_work() { 5979 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5980 "CMS thread should hold CMS token"); 5981 assert_lock_strong(_freelistLock); 5982 assert_lock_strong(_bit_map->lock()); 5983 // relinquish the free_list_lock and bitMaplock() 5984 _bit_map->lock()->unlock(); 5985 _freelistLock->unlock(); 5986 ConcurrentMarkSweepThread::desynchronize(true); 5987 _collector->stopTimer(); 5988 _collector->incrementYields(); 5989 5990 // See the comment in coordinator_yield() 5991 for (unsigned i = 0; 5992 i < CMSYieldSleepCount && 5993 ConcurrentMarkSweepThread::should_yield() && 5994 !CMSCollector::foregroundGCIsActive(); 5995 ++i) { 5996 os::sleep(Thread::current(), 1, false); 5997 } 5998 5999 ConcurrentMarkSweepThread::synchronize(true); 6000 _freelistLock->lock_without_safepoint_check(); 6001 _bit_map->lock()->lock_without_safepoint_check(); 6002 _collector->startTimer(); 6003 } 6004 6005 /////////////////////////////////////////////////////////// 6006 // ParMarkRefsIntoAndScanClosure: a parallel version of 6007 // MarkRefsIntoAndScanClosure 6008 /////////////////////////////////////////////////////////// 6009 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6010 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 6011 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6012 _span(span), 6013 _bit_map(bit_map), 6014 _work_queue(work_queue), 6015 _low_water_mark(MIN2((work_queue->max_elems()/4), 6016 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6017 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 6018 { 6019 // FIXME: Should initialize in base class constructor. 6020 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 6021 set_ref_discoverer_internal(rd); 6022 } 6023 6024 // This closure is used to mark refs into the CMS generation at the 6025 // second (final) checkpoint, and to scan and transitively follow 6026 // the unmarked oops. The marks are made in the marking bit map and 6027 // the work_queue is used for keeping the (newly) grey objects during 6028 // the scan phase whence they are also available for stealing by parallel 6029 // threads. Since the marking bit map is shared, updates are 6030 // synchronized (via CAS). 6031 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6032 if (obj != NULL) { 6033 // Ignore mark word because this could be an already marked oop 6034 // that may be chained at the end of the overflow list. 6035 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6036 HeapWord* addr = (HeapWord*)obj; 6037 if (_span.contains(addr) && 6038 !_bit_map->isMarked(addr)) { 6039 // mark bit map (object will become grey): 6040 // It is possible for several threads to be 6041 // trying to "claim" this object concurrently; 6042 // the unique thread that succeeds in marking the 6043 // object first will do the subsequent push on 6044 // to the work queue (or overflow list). 6045 if (_bit_map->par_mark(addr)) { 6046 // push on work_queue (which may not be empty), and trim the 6047 // queue to an appropriate length by applying this closure to 6048 // the oops in the oops popped from the stack (i.e. blacken the 6049 // grey objects) 6050 bool res = _work_queue->push(obj); 6051 assert(res, "Low water mark should be less than capacity?"); 6052 trim_queue(_low_water_mark); 6053 } // Else, another thread claimed the object 6054 } 6055 } 6056 } 6057 6058 // This closure is used to rescan the marked objects on the dirty cards 6059 // in the mod union table and the card table proper. 6060 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6061 oop p, MemRegion mr) { 6062 6063 size_t size = 0; 6064 HeapWord* addr = (HeapWord*)p; 6065 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6066 assert(_span.contains(addr), "we are scanning the CMS generation"); 6067 // check if it's time to yield 6068 if (do_yield_check()) { 6069 // We yielded for some foreground stop-world work, 6070 // and we have been asked to abort this ongoing preclean cycle. 6071 return 0; 6072 } 6073 if (_bitMap->isMarked(addr)) { 6074 // it's marked; is it potentially uninitialized? 6075 if (p->klass_or_null_acquire() != NULL) { 6076 // an initialized object; ignore mark word in verification below 6077 // since we are running concurrent with mutators 6078 assert(oopDesc::is_oop(p, true), "should be an oop"); 6079 if (p->is_objArray()) { 6080 // objArrays are precisely marked; restrict scanning 6081 // to dirty cards only. 6082 size = CompactibleFreeListSpace::adjustObjectSize( 6083 p->oop_iterate_size(_scanningClosure, mr)); 6084 } else { 6085 // A non-array may have been imprecisely marked; we need 6086 // to scan object in its entirety. 6087 size = CompactibleFreeListSpace::adjustObjectSize( 6088 p->oop_iterate_size(_scanningClosure)); 6089 } 6090 #ifdef ASSERT 6091 size_t direct_size = 6092 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6093 assert(size == direct_size, "Inconsistency in size"); 6094 assert(size >= 3, "Necessary for Printezis marks to work"); 6095 HeapWord* start_pbit = addr + 1; 6096 HeapWord* end_pbit = addr + size - 1; 6097 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6098 "inconsistent Printezis mark"); 6099 // Verify inner mark bits (between Printezis bits) are clear, 6100 // but don't repeat if there are multiple dirty regions for 6101 // the same object, to avoid potential O(N^2) performance. 6102 if (addr != _last_scanned_object) { 6103 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6104 _last_scanned_object = addr; 6105 } 6106 #endif // ASSERT 6107 } else { 6108 // An uninitialized object. 6109 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6110 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6111 size = pointer_delta(nextOneAddr + 1, addr); 6112 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6113 "alignment problem"); 6114 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6115 // will dirty the card when the klass pointer is installed in the 6116 // object (signaling the completion of initialization). 6117 } 6118 } else { 6119 // Either a not yet marked object or an uninitialized object 6120 if (p->klass_or_null_acquire() == NULL) { 6121 // An uninitialized object, skip to the next card, since 6122 // we may not be able to read its P-bits yet. 6123 assert(size == 0, "Initial value"); 6124 } else { 6125 // An object not (yet) reached by marking: we merely need to 6126 // compute its size so as to go look at the next block. 6127 assert(oopDesc::is_oop(p, true), "should be an oop"); 6128 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6129 } 6130 } 6131 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6132 return size; 6133 } 6134 6135 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6136 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6137 "CMS thread should hold CMS token"); 6138 assert_lock_strong(_freelistLock); 6139 assert_lock_strong(_bitMap->lock()); 6140 // relinquish the free_list_lock and bitMaplock() 6141 _bitMap->lock()->unlock(); 6142 _freelistLock->unlock(); 6143 ConcurrentMarkSweepThread::desynchronize(true); 6144 _collector->stopTimer(); 6145 _collector->incrementYields(); 6146 6147 // See the comment in coordinator_yield() 6148 for (unsigned i = 0; i < CMSYieldSleepCount && 6149 ConcurrentMarkSweepThread::should_yield() && 6150 !CMSCollector::foregroundGCIsActive(); ++i) { 6151 os::sleep(Thread::current(), 1, false); 6152 } 6153 6154 ConcurrentMarkSweepThread::synchronize(true); 6155 _freelistLock->lock_without_safepoint_check(); 6156 _bitMap->lock()->lock_without_safepoint_check(); 6157 _collector->startTimer(); 6158 } 6159 6160 6161 ////////////////////////////////////////////////////////////////// 6162 // SurvivorSpacePrecleanClosure 6163 ////////////////////////////////////////////////////////////////// 6164 // This (single-threaded) closure is used to preclean the oops in 6165 // the survivor spaces. 6166 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6167 6168 HeapWord* addr = (HeapWord*)p; 6169 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6170 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6171 assert(p->klass_or_null() != NULL, "object should be initialized"); 6172 // an initialized object; ignore mark word in verification below 6173 // since we are running concurrent with mutators 6174 assert(oopDesc::is_oop(p, true), "should be an oop"); 6175 // Note that we do not yield while we iterate over 6176 // the interior oops of p, pushing the relevant ones 6177 // on our marking stack. 6178 size_t size = p->oop_iterate_size(_scanning_closure); 6179 do_yield_check(); 6180 // Observe that below, we do not abandon the preclean 6181 // phase as soon as we should; rather we empty the 6182 // marking stack before returning. This is to satisfy 6183 // some existing assertions. In general, it may be a 6184 // good idea to abort immediately and complete the marking 6185 // from the grey objects at a later time. 6186 while (!_mark_stack->isEmpty()) { 6187 oop new_oop = _mark_stack->pop(); 6188 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6189 assert(_bit_map->isMarked((HeapWord*)new_oop), 6190 "only grey objects on this stack"); 6191 // iterate over the oops in this oop, marking and pushing 6192 // the ones in CMS heap (i.e. in _span). 6193 new_oop->oop_iterate(_scanning_closure); 6194 // check if it's time to yield 6195 do_yield_check(); 6196 } 6197 unsigned int after_count = 6198 CMSHeap::heap()->total_collections(); 6199 bool abort = (_before_count != after_count) || 6200 _collector->should_abort_preclean(); 6201 return abort ? 0 : size; 6202 } 6203 6204 void SurvivorSpacePrecleanClosure::do_yield_work() { 6205 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6206 "CMS thread should hold CMS token"); 6207 assert_lock_strong(_bit_map->lock()); 6208 // Relinquish the bit map lock 6209 _bit_map->lock()->unlock(); 6210 ConcurrentMarkSweepThread::desynchronize(true); 6211 _collector->stopTimer(); 6212 _collector->incrementYields(); 6213 6214 // See the comment in coordinator_yield() 6215 for (unsigned i = 0; i < CMSYieldSleepCount && 6216 ConcurrentMarkSweepThread::should_yield() && 6217 !CMSCollector::foregroundGCIsActive(); ++i) { 6218 os::sleep(Thread::current(), 1, false); 6219 } 6220 6221 ConcurrentMarkSweepThread::synchronize(true); 6222 _bit_map->lock()->lock_without_safepoint_check(); 6223 _collector->startTimer(); 6224 } 6225 6226 // This closure is used to rescan the marked objects on the dirty cards 6227 // in the mod union table and the card table proper. In the parallel 6228 // case, although the bitMap is shared, we do a single read so the 6229 // isMarked() query is "safe". 6230 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6231 // Ignore mark word because we are running concurrent with mutators 6232 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6233 HeapWord* addr = (HeapWord*)p; 6234 assert(_span.contains(addr), "we are scanning the CMS generation"); 6235 bool is_obj_array = false; 6236 #ifdef ASSERT 6237 if (!_parallel) { 6238 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6239 assert(_collector->overflow_list_is_empty(), 6240 "overflow list should be empty"); 6241 6242 } 6243 #endif // ASSERT 6244 if (_bit_map->isMarked(addr)) { 6245 // Obj arrays are precisely marked, non-arrays are not; 6246 // so we scan objArrays precisely and non-arrays in their 6247 // entirety. 6248 if (p->is_objArray()) { 6249 is_obj_array = true; 6250 if (_parallel) { 6251 p->oop_iterate(_par_scan_closure, mr); 6252 } else { 6253 p->oop_iterate(_scan_closure, mr); 6254 } 6255 } else { 6256 if (_parallel) { 6257 p->oop_iterate(_par_scan_closure); 6258 } else { 6259 p->oop_iterate(_scan_closure); 6260 } 6261 } 6262 } 6263 #ifdef ASSERT 6264 if (!_parallel) { 6265 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6266 assert(_collector->overflow_list_is_empty(), 6267 "overflow list should be empty"); 6268 6269 } 6270 #endif // ASSERT 6271 return is_obj_array; 6272 } 6273 6274 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6275 MemRegion span, 6276 CMSBitMap* bitMap, CMSMarkStack* markStack, 6277 bool should_yield, bool verifying): 6278 _collector(collector), 6279 _span(span), 6280 _bitMap(bitMap), 6281 _mut(&collector->_modUnionTable), 6282 _markStack(markStack), 6283 _yield(should_yield), 6284 _skipBits(0) 6285 { 6286 assert(_markStack->isEmpty(), "stack should be empty"); 6287 _finger = _bitMap->startWord(); 6288 _threshold = _finger; 6289 assert(_collector->_restart_addr == NULL, "Sanity check"); 6290 assert(_span.contains(_finger), "Out of bounds _finger?"); 6291 DEBUG_ONLY(_verifying = verifying;) 6292 } 6293 6294 void MarkFromRootsClosure::reset(HeapWord* addr) { 6295 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6296 assert(_span.contains(addr), "Out of bounds _finger?"); 6297 _finger = addr; 6298 _threshold = align_up(_finger, CardTable::card_size); 6299 } 6300 6301 // Should revisit to see if this should be restructured for 6302 // greater efficiency. 6303 bool MarkFromRootsClosure::do_bit(size_t offset) { 6304 if (_skipBits > 0) { 6305 _skipBits--; 6306 return true; 6307 } 6308 // convert offset into a HeapWord* 6309 HeapWord* addr = _bitMap->startWord() + offset; 6310 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6311 "address out of range"); 6312 assert(_bitMap->isMarked(addr), "tautology"); 6313 if (_bitMap->isMarked(addr+1)) { 6314 // this is an allocated but not yet initialized object 6315 assert(_skipBits == 0, "tautology"); 6316 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6317 oop p = oop(addr); 6318 if (p->klass_or_null_acquire() == NULL) { 6319 DEBUG_ONLY(if (!_verifying) {) 6320 // We re-dirty the cards on which this object lies and increase 6321 // the _threshold so that we'll come back to scan this object 6322 // during the preclean or remark phase. (CMSCleanOnEnter) 6323 if (CMSCleanOnEnter) { 6324 size_t sz = _collector->block_size_using_printezis_bits(addr); 6325 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6326 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6327 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6328 // Bump _threshold to end_card_addr; note that 6329 // _threshold cannot possibly exceed end_card_addr, anyhow. 6330 // This prevents future clearing of the card as the scan proceeds 6331 // to the right. 6332 assert(_threshold <= end_card_addr, 6333 "Because we are just scanning into this object"); 6334 if (_threshold < end_card_addr) { 6335 _threshold = end_card_addr; 6336 } 6337 if (p->klass_or_null_acquire() != NULL) { 6338 // Redirty the range of cards... 6339 _mut->mark_range(redirty_range); 6340 } // ...else the setting of klass will dirty the card anyway. 6341 } 6342 DEBUG_ONLY(}) 6343 return true; 6344 } 6345 } 6346 scanOopsInOop(addr); 6347 return true; 6348 } 6349 6350 // We take a break if we've been at this for a while, 6351 // so as to avoid monopolizing the locks involved. 6352 void MarkFromRootsClosure::do_yield_work() { 6353 // First give up the locks, then yield, then re-lock 6354 // We should probably use a constructor/destructor idiom to 6355 // do this unlock/lock or modify the MutexUnlocker class to 6356 // serve our purpose. XXX 6357 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6358 "CMS thread should hold CMS token"); 6359 assert_lock_strong(_bitMap->lock()); 6360 _bitMap->lock()->unlock(); 6361 ConcurrentMarkSweepThread::desynchronize(true); 6362 _collector->stopTimer(); 6363 _collector->incrementYields(); 6364 6365 // See the comment in coordinator_yield() 6366 for (unsigned i = 0; i < CMSYieldSleepCount && 6367 ConcurrentMarkSweepThread::should_yield() && 6368 !CMSCollector::foregroundGCIsActive(); ++i) { 6369 os::sleep(Thread::current(), 1, false); 6370 } 6371 6372 ConcurrentMarkSweepThread::synchronize(true); 6373 _bitMap->lock()->lock_without_safepoint_check(); 6374 _collector->startTimer(); 6375 } 6376 6377 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6378 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6379 assert(_markStack->isEmpty(), 6380 "should drain stack to limit stack usage"); 6381 // convert ptr to an oop preparatory to scanning 6382 oop obj = oop(ptr); 6383 // Ignore mark word in verification below, since we 6384 // may be running concurrent with mutators. 6385 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6386 assert(_finger <= ptr, "_finger runneth ahead"); 6387 // advance the finger to right end of this object 6388 _finger = ptr + obj->size(); 6389 assert(_finger > ptr, "we just incremented it above"); 6390 // On large heaps, it may take us some time to get through 6391 // the marking phase. During 6392 // this time it's possible that a lot of mutations have 6393 // accumulated in the card table and the mod union table -- 6394 // these mutation records are redundant until we have 6395 // actually traced into the corresponding card. 6396 // Here, we check whether advancing the finger would make 6397 // us cross into a new card, and if so clear corresponding 6398 // cards in the MUT (preclean them in the card-table in the 6399 // future). 6400 6401 DEBUG_ONLY(if (!_verifying) {) 6402 // The clean-on-enter optimization is disabled by default, 6403 // until we fix 6178663. 6404 if (CMSCleanOnEnter && (_finger > _threshold)) { 6405 // [_threshold, _finger) represents the interval 6406 // of cards to be cleared in MUT (or precleaned in card table). 6407 // The set of cards to be cleared is all those that overlap 6408 // with the interval [_threshold, _finger); note that 6409 // _threshold is always kept card-aligned but _finger isn't 6410 // always card-aligned. 6411 HeapWord* old_threshold = _threshold; 6412 assert(is_aligned(old_threshold, CardTable::card_size), 6413 "_threshold should always be card-aligned"); 6414 _threshold = align_up(_finger, CardTable::card_size); 6415 MemRegion mr(old_threshold, _threshold); 6416 assert(!mr.is_empty(), "Control point invariant"); 6417 assert(_span.contains(mr), "Should clear within span"); 6418 _mut->clear_range(mr); 6419 } 6420 DEBUG_ONLY(}) 6421 // Note: the finger doesn't advance while we drain 6422 // the stack below. 6423 PushOrMarkClosure pushOrMarkClosure(_collector, 6424 _span, _bitMap, _markStack, 6425 _finger, this); 6426 bool res = _markStack->push(obj); 6427 assert(res, "Empty non-zero size stack should have space for single push"); 6428 while (!_markStack->isEmpty()) { 6429 oop new_oop = _markStack->pop(); 6430 // Skip verifying header mark word below because we are 6431 // running concurrent with mutators. 6432 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6433 // now scan this oop's oops 6434 new_oop->oop_iterate(&pushOrMarkClosure); 6435 do_yield_check(); 6436 } 6437 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6438 } 6439 6440 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6441 CMSCollector* collector, MemRegion span, 6442 CMSBitMap* bit_map, 6443 OopTaskQueue* work_queue, 6444 CMSMarkStack* overflow_stack): 6445 _collector(collector), 6446 _whole_span(collector->_span), 6447 _span(span), 6448 _bit_map(bit_map), 6449 _mut(&collector->_modUnionTable), 6450 _work_queue(work_queue), 6451 _overflow_stack(overflow_stack), 6452 _skip_bits(0), 6453 _task(task) 6454 { 6455 assert(_work_queue->size() == 0, "work_queue should be empty"); 6456 _finger = span.start(); 6457 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6458 assert(_span.contains(_finger), "Out of bounds _finger?"); 6459 } 6460 6461 // Should revisit to see if this should be restructured for 6462 // greater efficiency. 6463 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6464 if (_skip_bits > 0) { 6465 _skip_bits--; 6466 return true; 6467 } 6468 // convert offset into a HeapWord* 6469 HeapWord* addr = _bit_map->startWord() + offset; 6470 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6471 "address out of range"); 6472 assert(_bit_map->isMarked(addr), "tautology"); 6473 if (_bit_map->isMarked(addr+1)) { 6474 // this is an allocated object that might not yet be initialized 6475 assert(_skip_bits == 0, "tautology"); 6476 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6477 oop p = oop(addr); 6478 if (p->klass_or_null_acquire() == NULL) { 6479 // in the case of Clean-on-Enter optimization, redirty card 6480 // and avoid clearing card by increasing the threshold. 6481 return true; 6482 } 6483 } 6484 scan_oops_in_oop(addr); 6485 return true; 6486 } 6487 6488 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6489 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6490 // Should we assert that our work queue is empty or 6491 // below some drain limit? 6492 assert(_work_queue->size() == 0, 6493 "should drain stack to limit stack usage"); 6494 // convert ptr to an oop preparatory to scanning 6495 oop obj = oop(ptr); 6496 // Ignore mark word in verification below, since we 6497 // may be running concurrent with mutators. 6498 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6499 assert(_finger <= ptr, "_finger runneth ahead"); 6500 // advance the finger to right end of this object 6501 _finger = ptr + obj->size(); 6502 assert(_finger > ptr, "we just incremented it above"); 6503 // On large heaps, it may take us some time to get through 6504 // the marking phase. During 6505 // this time it's possible that a lot of mutations have 6506 // accumulated in the card table and the mod union table -- 6507 // these mutation records are redundant until we have 6508 // actually traced into the corresponding card. 6509 // Here, we check whether advancing the finger would make 6510 // us cross into a new card, and if so clear corresponding 6511 // cards in the MUT (preclean them in the card-table in the 6512 // future). 6513 6514 // The clean-on-enter optimization is disabled by default, 6515 // until we fix 6178663. 6516 if (CMSCleanOnEnter && (_finger > _threshold)) { 6517 // [_threshold, _finger) represents the interval 6518 // of cards to be cleared in MUT (or precleaned in card table). 6519 // The set of cards to be cleared is all those that overlap 6520 // with the interval [_threshold, _finger); note that 6521 // _threshold is always kept card-aligned but _finger isn't 6522 // always card-aligned. 6523 HeapWord* old_threshold = _threshold; 6524 assert(is_aligned(old_threshold, CardTable::card_size), 6525 "_threshold should always be card-aligned"); 6526 _threshold = align_up(_finger, CardTable::card_size); 6527 MemRegion mr(old_threshold, _threshold); 6528 assert(!mr.is_empty(), "Control point invariant"); 6529 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6530 _mut->clear_range(mr); 6531 } 6532 6533 // Note: the local finger doesn't advance while we drain 6534 // the stack below, but the global finger sure can and will. 6535 HeapWord* volatile* gfa = _task->global_finger_addr(); 6536 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6537 _span, _bit_map, 6538 _work_queue, 6539 _overflow_stack, 6540 _finger, 6541 gfa, this); 6542 bool res = _work_queue->push(obj); // overflow could occur here 6543 assert(res, "Will hold once we use workqueues"); 6544 while (true) { 6545 oop new_oop; 6546 if (!_work_queue->pop_local(new_oop)) { 6547 // We emptied our work_queue; check if there's stuff that can 6548 // be gotten from the overflow stack. 6549 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6550 _overflow_stack, _work_queue)) { 6551 do_yield_check(); 6552 continue; 6553 } else { // done 6554 break; 6555 } 6556 } 6557 // Skip verifying header mark word below because we are 6558 // running concurrent with mutators. 6559 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6560 // now scan this oop's oops 6561 new_oop->oop_iterate(&pushOrMarkClosure); 6562 do_yield_check(); 6563 } 6564 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6565 } 6566 6567 // Yield in response to a request from VM Thread or 6568 // from mutators. 6569 void ParMarkFromRootsClosure::do_yield_work() { 6570 assert(_task != NULL, "sanity"); 6571 _task->yield(); 6572 } 6573 6574 // A variant of the above used for verifying CMS marking work. 6575 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6576 MemRegion span, 6577 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6578 CMSMarkStack* mark_stack): 6579 _collector(collector), 6580 _span(span), 6581 _verification_bm(verification_bm), 6582 _cms_bm(cms_bm), 6583 _mark_stack(mark_stack), 6584 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6585 mark_stack) 6586 { 6587 assert(_mark_stack->isEmpty(), "stack should be empty"); 6588 _finger = _verification_bm->startWord(); 6589 assert(_collector->_restart_addr == NULL, "Sanity check"); 6590 assert(_span.contains(_finger), "Out of bounds _finger?"); 6591 } 6592 6593 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6594 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6595 assert(_span.contains(addr), "Out of bounds _finger?"); 6596 _finger = addr; 6597 } 6598 6599 // Should revisit to see if this should be restructured for 6600 // greater efficiency. 6601 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6602 // convert offset into a HeapWord* 6603 HeapWord* addr = _verification_bm->startWord() + offset; 6604 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6605 "address out of range"); 6606 assert(_verification_bm->isMarked(addr), "tautology"); 6607 assert(_cms_bm->isMarked(addr), "tautology"); 6608 6609 assert(_mark_stack->isEmpty(), 6610 "should drain stack to limit stack usage"); 6611 // convert addr to an oop preparatory to scanning 6612 oop obj = oop(addr); 6613 assert(oopDesc::is_oop(obj), "should be an oop"); 6614 assert(_finger <= addr, "_finger runneth ahead"); 6615 // advance the finger to right end of this object 6616 _finger = addr + obj->size(); 6617 assert(_finger > addr, "we just incremented it above"); 6618 // Note: the finger doesn't advance while we drain 6619 // the stack below. 6620 bool res = _mark_stack->push(obj); 6621 assert(res, "Empty non-zero size stack should have space for single push"); 6622 while (!_mark_stack->isEmpty()) { 6623 oop new_oop = _mark_stack->pop(); 6624 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6625 // now scan this oop's oops 6626 new_oop->oop_iterate(&_pam_verify_closure); 6627 } 6628 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6629 return true; 6630 } 6631 6632 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6633 CMSCollector* collector, MemRegion span, 6634 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6635 CMSMarkStack* mark_stack): 6636 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6637 _collector(collector), 6638 _span(span), 6639 _verification_bm(verification_bm), 6640 _cms_bm(cms_bm), 6641 _mark_stack(mark_stack) 6642 { } 6643 6644 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6645 oop obj = RawAccess<>::oop_load(p); 6646 do_oop(obj); 6647 } 6648 6649 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6650 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6651 6652 // Upon stack overflow, we discard (part of) the stack, 6653 // remembering the least address amongst those discarded 6654 // in CMSCollector's _restart_address. 6655 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6656 // Remember the least grey address discarded 6657 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6658 _collector->lower_restart_addr(ra); 6659 _mark_stack->reset(); // discard stack contents 6660 _mark_stack->expand(); // expand the stack if possible 6661 } 6662 6663 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6664 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6665 HeapWord* addr = (HeapWord*)obj; 6666 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6667 // Oop lies in _span and isn't yet grey or black 6668 _verification_bm->mark(addr); // now grey 6669 if (!_cms_bm->isMarked(addr)) { 6670 Log(gc, verify) log; 6671 ResourceMark rm; 6672 LogStream ls(log.error()); 6673 oop(addr)->print_on(&ls); 6674 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6675 fatal("... aborting"); 6676 } 6677 6678 if (!_mark_stack->push(obj)) { // stack overflow 6679 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6680 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6681 handle_stack_overflow(addr); 6682 } 6683 // anything including and to the right of _finger 6684 // will be scanned as we iterate over the remainder of the 6685 // bit map 6686 } 6687 } 6688 6689 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6690 MemRegion span, 6691 CMSBitMap* bitMap, CMSMarkStack* markStack, 6692 HeapWord* finger, MarkFromRootsClosure* parent) : 6693 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6694 _collector(collector), 6695 _span(span), 6696 _bitMap(bitMap), 6697 _markStack(markStack), 6698 _finger(finger), 6699 _parent(parent) 6700 { } 6701 6702 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6703 MemRegion span, 6704 CMSBitMap* bit_map, 6705 OopTaskQueue* work_queue, 6706 CMSMarkStack* overflow_stack, 6707 HeapWord* finger, 6708 HeapWord* volatile* global_finger_addr, 6709 ParMarkFromRootsClosure* parent) : 6710 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6711 _collector(collector), 6712 _whole_span(collector->_span), 6713 _span(span), 6714 _bit_map(bit_map), 6715 _work_queue(work_queue), 6716 _overflow_stack(overflow_stack), 6717 _finger(finger), 6718 _global_finger_addr(global_finger_addr), 6719 _parent(parent) 6720 { } 6721 6722 // Assumes thread-safe access by callers, who are 6723 // responsible for mutual exclusion. 6724 void CMSCollector::lower_restart_addr(HeapWord* low) { 6725 assert(_span.contains(low), "Out of bounds addr"); 6726 if (_restart_addr == NULL) { 6727 _restart_addr = low; 6728 } else { 6729 _restart_addr = MIN2(_restart_addr, low); 6730 } 6731 } 6732 6733 // Upon stack overflow, we discard (part of) the stack, 6734 // remembering the least address amongst those discarded 6735 // in CMSCollector's _restart_address. 6736 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6737 // Remember the least grey address discarded 6738 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6739 _collector->lower_restart_addr(ra); 6740 _markStack->reset(); // discard stack contents 6741 _markStack->expand(); // expand the stack if possible 6742 } 6743 6744 // Upon stack overflow, we discard (part of) the stack, 6745 // remembering the least address amongst those discarded 6746 // in CMSCollector's _restart_address. 6747 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6748 // We need to do this under a mutex to prevent other 6749 // workers from interfering with the work done below. 6750 MutexLockerEx ml(_overflow_stack->par_lock(), 6751 Mutex::_no_safepoint_check_flag); 6752 // Remember the least grey address discarded 6753 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6754 _collector->lower_restart_addr(ra); 6755 _overflow_stack->reset(); // discard stack contents 6756 _overflow_stack->expand(); // expand the stack if possible 6757 } 6758 6759 void PushOrMarkClosure::do_oop(oop obj) { 6760 // Ignore mark word because we are running concurrent with mutators. 6761 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6762 HeapWord* addr = (HeapWord*)obj; 6763 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6764 // Oop lies in _span and isn't yet grey or black 6765 _bitMap->mark(addr); // now grey 6766 if (addr < _finger) { 6767 // the bit map iteration has already either passed, or 6768 // sampled, this bit in the bit map; we'll need to 6769 // use the marking stack to scan this oop's oops. 6770 bool simulate_overflow = false; 6771 NOT_PRODUCT( 6772 if (CMSMarkStackOverflowALot && 6773 _collector->simulate_overflow()) { 6774 // simulate a stack overflow 6775 simulate_overflow = true; 6776 } 6777 ) 6778 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6779 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6780 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6781 handle_stack_overflow(addr); 6782 } 6783 } 6784 // anything including and to the right of _finger 6785 // will be scanned as we iterate over the remainder of the 6786 // bit map 6787 do_yield_check(); 6788 } 6789 } 6790 6791 void ParPushOrMarkClosure::do_oop(oop obj) { 6792 // Ignore mark word because we are running concurrent with mutators. 6793 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6794 HeapWord* addr = (HeapWord*)obj; 6795 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6796 // Oop lies in _span and isn't yet grey or black 6797 // We read the global_finger (volatile read) strictly after marking oop 6798 bool res = _bit_map->par_mark(addr); // now grey 6799 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6800 // Should we push this marked oop on our stack? 6801 // -- if someone else marked it, nothing to do 6802 // -- if target oop is above global finger nothing to do 6803 // -- if target oop is in chunk and above local finger 6804 // then nothing to do 6805 // -- else push on work queue 6806 if ( !res // someone else marked it, they will deal with it 6807 || (addr >= *gfa) // will be scanned in a later task 6808 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6809 return; 6810 } 6811 // the bit map iteration has already either passed, or 6812 // sampled, this bit in the bit map; we'll need to 6813 // use the marking stack to scan this oop's oops. 6814 bool simulate_overflow = false; 6815 NOT_PRODUCT( 6816 if (CMSMarkStackOverflowALot && 6817 _collector->simulate_overflow()) { 6818 // simulate a stack overflow 6819 simulate_overflow = true; 6820 } 6821 ) 6822 if (simulate_overflow || 6823 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6824 // stack overflow 6825 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6826 // We cannot assert that the overflow stack is full because 6827 // it may have been emptied since. 6828 assert(simulate_overflow || 6829 _work_queue->size() == _work_queue->max_elems(), 6830 "Else push should have succeeded"); 6831 handle_stack_overflow(addr); 6832 } 6833 do_yield_check(); 6834 } 6835 } 6836 6837 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6838 MemRegion span, 6839 ReferenceDiscoverer* rd, 6840 CMSBitMap* bit_map, 6841 CMSBitMap* mod_union_table, 6842 CMSMarkStack* mark_stack, 6843 bool concurrent_precleaning): 6844 MetadataVisitingOopIterateClosure(rd), 6845 _collector(collector), 6846 _span(span), 6847 _bit_map(bit_map), 6848 _mod_union_table(mod_union_table), 6849 _mark_stack(mark_stack), 6850 _concurrent_precleaning(concurrent_precleaning) 6851 { 6852 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6853 } 6854 6855 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6856 // the non-parallel version (the parallel version appears further below.) 6857 void PushAndMarkClosure::do_oop(oop obj) { 6858 // Ignore mark word verification. If during concurrent precleaning, 6859 // the object monitor may be locked. If during the checkpoint 6860 // phases, the object may already have been reached by a different 6861 // path and may be at the end of the global overflow list (so 6862 // the mark word may be NULL). 6863 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6864 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6865 HeapWord* addr = (HeapWord*)obj; 6866 // Check if oop points into the CMS generation 6867 // and is not marked 6868 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6869 // a white object ... 6870 _bit_map->mark(addr); // ... now grey 6871 // push on the marking stack (grey set) 6872 bool simulate_overflow = false; 6873 NOT_PRODUCT( 6874 if (CMSMarkStackOverflowALot && 6875 _collector->simulate_overflow()) { 6876 // simulate a stack overflow 6877 simulate_overflow = true; 6878 } 6879 ) 6880 if (simulate_overflow || !_mark_stack->push(obj)) { 6881 if (_concurrent_precleaning) { 6882 // During precleaning we can just dirty the appropriate card(s) 6883 // in the mod union table, thus ensuring that the object remains 6884 // in the grey set and continue. In the case of object arrays 6885 // we need to dirty all of the cards that the object spans, 6886 // since the rescan of object arrays will be limited to the 6887 // dirty cards. 6888 // Note that no one can be interfering with us in this action 6889 // of dirtying the mod union table, so no locking or atomics 6890 // are required. 6891 if (obj->is_objArray()) { 6892 size_t sz = obj->size(); 6893 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6894 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6895 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6896 _mod_union_table->mark_range(redirty_range); 6897 } else { 6898 _mod_union_table->mark(addr); 6899 } 6900 _collector->_ser_pmc_preclean_ovflw++; 6901 } else { 6902 // During the remark phase, we need to remember this oop 6903 // in the overflow list. 6904 _collector->push_on_overflow_list(obj); 6905 _collector->_ser_pmc_remark_ovflw++; 6906 } 6907 } 6908 } 6909 } 6910 6911 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6912 MemRegion span, 6913 ReferenceDiscoverer* rd, 6914 CMSBitMap* bit_map, 6915 OopTaskQueue* work_queue): 6916 MetadataVisitingOopIterateClosure(rd), 6917 _collector(collector), 6918 _span(span), 6919 _bit_map(bit_map), 6920 _work_queue(work_queue) 6921 { 6922 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6923 } 6924 6925 // Grey object rescan during second checkpoint phase -- 6926 // the parallel version. 6927 void ParPushAndMarkClosure::do_oop(oop obj) { 6928 // In the assert below, we ignore the mark word because 6929 // this oop may point to an already visited object that is 6930 // on the overflow stack (in which case the mark word has 6931 // been hijacked for chaining into the overflow stack -- 6932 // if this is the last object in the overflow stack then 6933 // its mark word will be NULL). Because this object may 6934 // have been subsequently popped off the global overflow 6935 // stack, and the mark word possibly restored to the prototypical 6936 // value, by the time we get to examined this failing assert in 6937 // the debugger, is_oop_or_null(false) may subsequently start 6938 // to hold. 6939 assert(oopDesc::is_oop_or_null(obj, true), 6940 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6941 HeapWord* addr = (HeapWord*)obj; 6942 // Check if oop points into the CMS generation 6943 // and is not marked 6944 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6945 // a white object ... 6946 // If we manage to "claim" the object, by being the 6947 // first thread to mark it, then we push it on our 6948 // marking stack 6949 if (_bit_map->par_mark(addr)) { // ... now grey 6950 // push on work queue (grey set) 6951 bool simulate_overflow = false; 6952 NOT_PRODUCT( 6953 if (CMSMarkStackOverflowALot && 6954 _collector->par_simulate_overflow()) { 6955 // simulate a stack overflow 6956 simulate_overflow = true; 6957 } 6958 ) 6959 if (simulate_overflow || !_work_queue->push(obj)) { 6960 _collector->par_push_on_overflow_list(obj); 6961 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6962 } 6963 } // Else, some other thread got there first 6964 } 6965 } 6966 6967 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6968 Mutex* bml = _collector->bitMapLock(); 6969 assert_lock_strong(bml); 6970 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6971 "CMS thread should hold CMS token"); 6972 6973 bml->unlock(); 6974 ConcurrentMarkSweepThread::desynchronize(true); 6975 6976 _collector->stopTimer(); 6977 _collector->incrementYields(); 6978 6979 // See the comment in coordinator_yield() 6980 for (unsigned i = 0; i < CMSYieldSleepCount && 6981 ConcurrentMarkSweepThread::should_yield() && 6982 !CMSCollector::foregroundGCIsActive(); ++i) { 6983 os::sleep(Thread::current(), 1, false); 6984 } 6985 6986 ConcurrentMarkSweepThread::synchronize(true); 6987 bml->lock(); 6988 6989 _collector->startTimer(); 6990 } 6991 6992 bool CMSPrecleanRefsYieldClosure::should_return() { 6993 if (ConcurrentMarkSweepThread::should_yield()) { 6994 do_yield_work(); 6995 } 6996 return _collector->foregroundGCIsActive(); 6997 } 6998 6999 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7000 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7001 "mr should be aligned to start at a card boundary"); 7002 // We'd like to assert: 7003 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7004 // "mr should be a range of cards"); 7005 // However, that would be too strong in one case -- the last 7006 // partition ends at _unallocated_block which, in general, can be 7007 // an arbitrary boundary, not necessarily card aligned. 7008 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7009 _space->object_iterate_mem(mr, &_scan_cl); 7010 } 7011 7012 SweepClosure::SweepClosure(CMSCollector* collector, 7013 ConcurrentMarkSweepGeneration* g, 7014 CMSBitMap* bitMap, bool should_yield) : 7015 _collector(collector), 7016 _g(g), 7017 _sp(g->cmsSpace()), 7018 _limit(_sp->sweep_limit()), 7019 _freelistLock(_sp->freelistLock()), 7020 _bitMap(bitMap), 7021 _inFreeRange(false), // No free range at beginning of sweep 7022 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7023 _lastFreeRangeCoalesced(false), 7024 _yield(should_yield), 7025 _freeFinger(g->used_region().start()) 7026 { 7027 NOT_PRODUCT( 7028 _numObjectsFreed = 0; 7029 _numWordsFreed = 0; 7030 _numObjectsLive = 0; 7031 _numWordsLive = 0; 7032 _numObjectsAlreadyFree = 0; 7033 _numWordsAlreadyFree = 0; 7034 _last_fc = NULL; 7035 7036 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7037 _sp->dictionary()->initialize_dict_returned_bytes(); 7038 ) 7039 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7040 "sweep _limit out of bounds"); 7041 log_develop_trace(gc, sweep)("===================="); 7042 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7043 } 7044 7045 void SweepClosure::print_on(outputStream* st) const { 7046 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7047 p2i(_sp->bottom()), p2i(_sp->end())); 7048 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7049 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7050 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7051 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7052 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7053 } 7054 7055 #ifndef PRODUCT 7056 // Assertion checking only: no useful work in product mode -- 7057 // however, if any of the flags below become product flags, 7058 // you may need to review this code to see if it needs to be 7059 // enabled in product mode. 7060 SweepClosure::~SweepClosure() { 7061 assert_lock_strong(_freelistLock); 7062 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7063 "sweep _limit out of bounds"); 7064 if (inFreeRange()) { 7065 Log(gc, sweep) log; 7066 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7067 ResourceMark rm; 7068 LogStream ls(log.error()); 7069 print_on(&ls); 7070 ShouldNotReachHere(); 7071 } 7072 7073 if (log_is_enabled(Debug, gc, sweep)) { 7074 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7075 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7076 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7077 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7078 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7079 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7080 } 7081 7082 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7083 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7084 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7085 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7086 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7087 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7088 } 7089 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7090 log_develop_trace(gc, sweep)("================"); 7091 } 7092 #endif // PRODUCT 7093 7094 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7095 bool freeRangeInFreeLists) { 7096 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7097 p2i(freeFinger), freeRangeInFreeLists); 7098 assert(!inFreeRange(), "Trampling existing free range"); 7099 set_inFreeRange(true); 7100 set_lastFreeRangeCoalesced(false); 7101 7102 set_freeFinger(freeFinger); 7103 set_freeRangeInFreeLists(freeRangeInFreeLists); 7104 if (CMSTestInFreeList) { 7105 if (freeRangeInFreeLists) { 7106 FreeChunk* fc = (FreeChunk*) freeFinger; 7107 assert(fc->is_free(), "A chunk on the free list should be free."); 7108 assert(fc->size() > 0, "Free range should have a size"); 7109 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7110 } 7111 } 7112 } 7113 7114 // Note that the sweeper runs concurrently with mutators. Thus, 7115 // it is possible for direct allocation in this generation to happen 7116 // in the middle of the sweep. Note that the sweeper also coalesces 7117 // contiguous free blocks. Thus, unless the sweeper and the allocator 7118 // synchronize appropriately freshly allocated blocks may get swept up. 7119 // This is accomplished by the sweeper locking the free lists while 7120 // it is sweeping. Thus blocks that are determined to be free are 7121 // indeed free. There is however one additional complication: 7122 // blocks that have been allocated since the final checkpoint and 7123 // mark, will not have been marked and so would be treated as 7124 // unreachable and swept up. To prevent this, the allocator marks 7125 // the bit map when allocating during the sweep phase. This leads, 7126 // however, to a further complication -- objects may have been allocated 7127 // but not yet initialized -- in the sense that the header isn't yet 7128 // installed. The sweeper can not then determine the size of the block 7129 // in order to skip over it. To deal with this case, we use a technique 7130 // (due to Printezis) to encode such uninitialized block sizes in the 7131 // bit map. Since the bit map uses a bit per every HeapWord, but the 7132 // CMS generation has a minimum object size of 3 HeapWords, it follows 7133 // that "normal marks" won't be adjacent in the bit map (there will 7134 // always be at least two 0 bits between successive 1 bits). We make use 7135 // of these "unused" bits to represent uninitialized blocks -- the bit 7136 // corresponding to the start of the uninitialized object and the next 7137 // bit are both set. Finally, a 1 bit marks the end of the object that 7138 // started with the two consecutive 1 bits to indicate its potentially 7139 // uninitialized state. 7140 7141 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7142 FreeChunk* fc = (FreeChunk*)addr; 7143 size_t res; 7144 7145 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7146 // than "addr == _limit" because although _limit was a block boundary when 7147 // we started the sweep, it may no longer be one because heap expansion 7148 // may have caused us to coalesce the block ending at the address _limit 7149 // with a newly expanded chunk (this happens when _limit was set to the 7150 // previous _end of the space), so we may have stepped past _limit: 7151 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7152 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7153 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7154 "sweep _limit out of bounds"); 7155 assert(addr < _sp->end(), "addr out of bounds"); 7156 // Flush any free range we might be holding as a single 7157 // coalesced chunk to the appropriate free list. 7158 if (inFreeRange()) { 7159 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7160 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7161 flush_cur_free_chunk(freeFinger(), 7162 pointer_delta(addr, freeFinger())); 7163 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7164 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7165 lastFreeRangeCoalesced() ? 1 : 0); 7166 } 7167 7168 // help the iterator loop finish 7169 return pointer_delta(_sp->end(), addr); 7170 } 7171 7172 assert(addr < _limit, "sweep invariant"); 7173 // check if we should yield 7174 do_yield_check(addr); 7175 if (fc->is_free()) { 7176 // Chunk that is already free 7177 res = fc->size(); 7178 do_already_free_chunk(fc); 7179 debug_only(_sp->verifyFreeLists()); 7180 // If we flush the chunk at hand in lookahead_and_flush() 7181 // and it's coalesced with a preceding chunk, then the 7182 // process of "mangling" the payload of the coalesced block 7183 // will cause erasure of the size information from the 7184 // (erstwhile) header of all the coalesced blocks but the 7185 // first, so the first disjunct in the assert will not hold 7186 // in that specific case (in which case the second disjunct 7187 // will hold). 7188 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7189 "Otherwise the size info doesn't change at this step"); 7190 NOT_PRODUCT( 7191 _numObjectsAlreadyFree++; 7192 _numWordsAlreadyFree += res; 7193 ) 7194 NOT_PRODUCT(_last_fc = fc;) 7195 } else if (!_bitMap->isMarked(addr)) { 7196 // Chunk is fresh garbage 7197 res = do_garbage_chunk(fc); 7198 debug_only(_sp->verifyFreeLists()); 7199 NOT_PRODUCT( 7200 _numObjectsFreed++; 7201 _numWordsFreed += res; 7202 ) 7203 } else { 7204 // Chunk that is alive. 7205 res = do_live_chunk(fc); 7206 debug_only(_sp->verifyFreeLists()); 7207 NOT_PRODUCT( 7208 _numObjectsLive++; 7209 _numWordsLive += res; 7210 ) 7211 } 7212 return res; 7213 } 7214 7215 // For the smart allocation, record following 7216 // split deaths - a free chunk is removed from its free list because 7217 // it is being split into two or more chunks. 7218 // split birth - a free chunk is being added to its free list because 7219 // a larger free chunk has been split and resulted in this free chunk. 7220 // coal death - a free chunk is being removed from its free list because 7221 // it is being coalesced into a large free chunk. 7222 // coal birth - a free chunk is being added to its free list because 7223 // it was created when two or more free chunks where coalesced into 7224 // this free chunk. 7225 // 7226 // These statistics are used to determine the desired number of free 7227 // chunks of a given size. The desired number is chosen to be relative 7228 // to the end of a CMS sweep. The desired number at the end of a sweep 7229 // is the 7230 // count-at-end-of-previous-sweep (an amount that was enough) 7231 // - count-at-beginning-of-current-sweep (the excess) 7232 // + split-births (gains in this size during interval) 7233 // - split-deaths (demands on this size during interval) 7234 // where the interval is from the end of one sweep to the end of the 7235 // next. 7236 // 7237 // When sweeping the sweeper maintains an accumulated chunk which is 7238 // the chunk that is made up of chunks that have been coalesced. That 7239 // will be termed the left-hand chunk. A new chunk of garbage that 7240 // is being considered for coalescing will be referred to as the 7241 // right-hand chunk. 7242 // 7243 // When making a decision on whether to coalesce a right-hand chunk with 7244 // the current left-hand chunk, the current count vs. the desired count 7245 // of the left-hand chunk is considered. Also if the right-hand chunk 7246 // is near the large chunk at the end of the heap (see 7247 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7248 // left-hand chunk is coalesced. 7249 // 7250 // When making a decision about whether to split a chunk, the desired count 7251 // vs. the current count of the candidate to be split is also considered. 7252 // If the candidate is underpopulated (currently fewer chunks than desired) 7253 // a chunk of an overpopulated (currently more chunks than desired) size may 7254 // be chosen. The "hint" associated with a free list, if non-null, points 7255 // to a free list which may be overpopulated. 7256 // 7257 7258 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7259 const size_t size = fc->size(); 7260 // Chunks that cannot be coalesced are not in the 7261 // free lists. 7262 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7263 assert(_sp->verify_chunk_in_free_list(fc), 7264 "free chunk should be in free lists"); 7265 } 7266 // a chunk that is already free, should not have been 7267 // marked in the bit map 7268 HeapWord* const addr = (HeapWord*) fc; 7269 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7270 // Verify that the bit map has no bits marked between 7271 // addr and purported end of this block. 7272 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7273 7274 // Some chunks cannot be coalesced under any circumstances. 7275 // See the definition of cantCoalesce(). 7276 if (!fc->cantCoalesce()) { 7277 // This chunk can potentially be coalesced. 7278 // All the work is done in 7279 do_post_free_or_garbage_chunk(fc, size); 7280 // Note that if the chunk is not coalescable (the else arm 7281 // below), we unconditionally flush, without needing to do 7282 // a "lookahead," as we do below. 7283 if (inFreeRange()) lookahead_and_flush(fc, size); 7284 } else { 7285 // Code path common to both original and adaptive free lists. 7286 7287 // cant coalesce with previous block; this should be treated 7288 // as the end of a free run if any 7289 if (inFreeRange()) { 7290 // we kicked some butt; time to pick up the garbage 7291 assert(freeFinger() < addr, "freeFinger points too high"); 7292 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7293 } 7294 // else, nothing to do, just continue 7295 } 7296 } 7297 7298 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7299 // This is a chunk of garbage. It is not in any free list. 7300 // Add it to a free list or let it possibly be coalesced into 7301 // a larger chunk. 7302 HeapWord* const addr = (HeapWord*) fc; 7303 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7304 7305 // Verify that the bit map has no bits marked between 7306 // addr and purported end of just dead object. 7307 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7308 do_post_free_or_garbage_chunk(fc, size); 7309 7310 assert(_limit >= addr + size, 7311 "A freshly garbage chunk can't possibly straddle over _limit"); 7312 if (inFreeRange()) lookahead_and_flush(fc, size); 7313 return size; 7314 } 7315 7316 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7317 HeapWord* addr = (HeapWord*) fc; 7318 // The sweeper has just found a live object. Return any accumulated 7319 // left hand chunk to the free lists. 7320 if (inFreeRange()) { 7321 assert(freeFinger() < addr, "freeFinger points too high"); 7322 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7323 } 7324 7325 // This object is live: we'd normally expect this to be 7326 // an oop, and like to assert the following: 7327 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7328 // However, as we commented above, this may be an object whose 7329 // header hasn't yet been initialized. 7330 size_t size; 7331 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7332 if (_bitMap->isMarked(addr + 1)) { 7333 // Determine the size from the bit map, rather than trying to 7334 // compute it from the object header. 7335 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7336 size = pointer_delta(nextOneAddr + 1, addr); 7337 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7338 "alignment problem"); 7339 7340 #ifdef ASSERT 7341 if (oop(addr)->klass_or_null_acquire() != NULL) { 7342 // Ignore mark word because we are running concurrent with mutators 7343 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7344 assert(size == 7345 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7346 "P-mark and computed size do not agree"); 7347 } 7348 #endif 7349 7350 } else { 7351 // This should be an initialized object that's alive. 7352 assert(oop(addr)->klass_or_null_acquire() != NULL, 7353 "Should be an initialized object"); 7354 // Ignore mark word because we are running concurrent with mutators 7355 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7356 // Verify that the bit map has no bits marked between 7357 // addr and purported end of this block. 7358 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7359 assert(size >= 3, "Necessary for Printezis marks to work"); 7360 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7361 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7362 } 7363 return size; 7364 } 7365 7366 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7367 size_t chunkSize) { 7368 // do_post_free_or_garbage_chunk() should only be called in the case 7369 // of the adaptive free list allocator. 7370 const bool fcInFreeLists = fc->is_free(); 7371 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7372 if (CMSTestInFreeList && fcInFreeLists) { 7373 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7374 } 7375 7376 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7377 7378 HeapWord* const fc_addr = (HeapWord*) fc; 7379 7380 bool coalesce = false; 7381 const size_t left = pointer_delta(fc_addr, freeFinger()); 7382 const size_t right = chunkSize; 7383 switch (FLSCoalescePolicy) { 7384 // numeric value forms a coalition aggressiveness metric 7385 case 0: { // never coalesce 7386 coalesce = false; 7387 break; 7388 } 7389 case 1: { // coalesce if left & right chunks on overpopulated lists 7390 coalesce = _sp->coalOverPopulated(left) && 7391 _sp->coalOverPopulated(right); 7392 break; 7393 } 7394 case 2: { // coalesce if left chunk on overpopulated list (default) 7395 coalesce = _sp->coalOverPopulated(left); 7396 break; 7397 } 7398 case 3: { // coalesce if left OR right chunk on overpopulated list 7399 coalesce = _sp->coalOverPopulated(left) || 7400 _sp->coalOverPopulated(right); 7401 break; 7402 } 7403 case 4: { // always coalesce 7404 coalesce = true; 7405 break; 7406 } 7407 default: 7408 ShouldNotReachHere(); 7409 } 7410 7411 // Should the current free range be coalesced? 7412 // If the chunk is in a free range and either we decided to coalesce above 7413 // or the chunk is near the large block at the end of the heap 7414 // (isNearLargestChunk() returns true), then coalesce this chunk. 7415 const bool doCoalesce = inFreeRange() 7416 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7417 if (doCoalesce) { 7418 // Coalesce the current free range on the left with the new 7419 // chunk on the right. If either is on a free list, 7420 // it must be removed from the list and stashed in the closure. 7421 if (freeRangeInFreeLists()) { 7422 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7423 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7424 "Size of free range is inconsistent with chunk size."); 7425 if (CMSTestInFreeList) { 7426 assert(_sp->verify_chunk_in_free_list(ffc), 7427 "Chunk is not in free lists"); 7428 } 7429 _sp->coalDeath(ffc->size()); 7430 _sp->removeFreeChunkFromFreeLists(ffc); 7431 set_freeRangeInFreeLists(false); 7432 } 7433 if (fcInFreeLists) { 7434 _sp->coalDeath(chunkSize); 7435 assert(fc->size() == chunkSize, 7436 "The chunk has the wrong size or is not in the free lists"); 7437 _sp->removeFreeChunkFromFreeLists(fc); 7438 } 7439 set_lastFreeRangeCoalesced(true); 7440 print_free_block_coalesced(fc); 7441 } else { // not in a free range and/or should not coalesce 7442 // Return the current free range and start a new one. 7443 if (inFreeRange()) { 7444 // In a free range but cannot coalesce with the right hand chunk. 7445 // Put the current free range into the free lists. 7446 flush_cur_free_chunk(freeFinger(), 7447 pointer_delta(fc_addr, freeFinger())); 7448 } 7449 // Set up for new free range. Pass along whether the right hand 7450 // chunk is in the free lists. 7451 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7452 } 7453 } 7454 7455 // Lookahead flush: 7456 // If we are tracking a free range, and this is the last chunk that 7457 // we'll look at because its end crosses past _limit, we'll preemptively 7458 // flush it along with any free range we may be holding on to. Note that 7459 // this can be the case only for an already free or freshly garbage 7460 // chunk. If this block is an object, it can never straddle 7461 // over _limit. The "straddling" occurs when _limit is set at 7462 // the previous end of the space when this cycle started, and 7463 // a subsequent heap expansion caused the previously co-terminal 7464 // free block to be coalesced with the newly expanded portion, 7465 // thus rendering _limit a non-block-boundary making it dangerous 7466 // for the sweeper to step over and examine. 7467 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7468 assert(inFreeRange(), "Should only be called if currently in a free range."); 7469 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7470 assert(_sp->used_region().contains(eob - 1), 7471 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7472 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7473 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7474 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7475 if (eob >= _limit) { 7476 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7477 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7478 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7479 "[" PTR_FORMAT "," PTR_FORMAT ")", 7480 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7481 // Return the storage we are tracking back into the free lists. 7482 log_develop_trace(gc, sweep)("Flushing ... "); 7483 assert(freeFinger() < eob, "Error"); 7484 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7485 } 7486 } 7487 7488 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7489 assert(inFreeRange(), "Should only be called if currently in a free range."); 7490 assert(size > 0, 7491 "A zero sized chunk cannot be added to the free lists."); 7492 if (!freeRangeInFreeLists()) { 7493 if (CMSTestInFreeList) { 7494 FreeChunk* fc = (FreeChunk*) chunk; 7495 fc->set_size(size); 7496 assert(!_sp->verify_chunk_in_free_list(fc), 7497 "chunk should not be in free lists yet"); 7498 } 7499 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7500 // A new free range is going to be starting. The current 7501 // free range has not been added to the free lists yet or 7502 // was removed so add it back. 7503 // If the current free range was coalesced, then the death 7504 // of the free range was recorded. Record a birth now. 7505 if (lastFreeRangeCoalesced()) { 7506 _sp->coalBirth(size); 7507 } 7508 _sp->addChunkAndRepairOffsetTable(chunk, size, 7509 lastFreeRangeCoalesced()); 7510 } else { 7511 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7512 } 7513 set_inFreeRange(false); 7514 set_freeRangeInFreeLists(false); 7515 } 7516 7517 // We take a break if we've been at this for a while, 7518 // so as to avoid monopolizing the locks involved. 7519 void SweepClosure::do_yield_work(HeapWord* addr) { 7520 // Return current free chunk being used for coalescing (if any) 7521 // to the appropriate freelist. After yielding, the next 7522 // free block encountered will start a coalescing range of 7523 // free blocks. If the next free block is adjacent to the 7524 // chunk just flushed, they will need to wait for the next 7525 // sweep to be coalesced. 7526 if (inFreeRange()) { 7527 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7528 } 7529 7530 // First give up the locks, then yield, then re-lock. 7531 // We should probably use a constructor/destructor idiom to 7532 // do this unlock/lock or modify the MutexUnlocker class to 7533 // serve our purpose. XXX 7534 assert_lock_strong(_bitMap->lock()); 7535 assert_lock_strong(_freelistLock); 7536 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7537 "CMS thread should hold CMS token"); 7538 _bitMap->lock()->unlock(); 7539 _freelistLock->unlock(); 7540 ConcurrentMarkSweepThread::desynchronize(true); 7541 _collector->stopTimer(); 7542 _collector->incrementYields(); 7543 7544 // See the comment in coordinator_yield() 7545 for (unsigned i = 0; i < CMSYieldSleepCount && 7546 ConcurrentMarkSweepThread::should_yield() && 7547 !CMSCollector::foregroundGCIsActive(); ++i) { 7548 os::sleep(Thread::current(), 1, false); 7549 } 7550 7551 ConcurrentMarkSweepThread::synchronize(true); 7552 _freelistLock->lock(); 7553 _bitMap->lock()->lock_without_safepoint_check(); 7554 _collector->startTimer(); 7555 } 7556 7557 #ifndef PRODUCT 7558 // This is actually very useful in a product build if it can 7559 // be called from the debugger. Compile it into the product 7560 // as needed. 7561 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7562 return debug_cms_space->verify_chunk_in_free_list(fc); 7563 } 7564 #endif 7565 7566 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7567 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7568 p2i(fc), fc->size()); 7569 } 7570 7571 // CMSIsAliveClosure 7572 bool CMSIsAliveClosure::do_object_b(oop obj) { 7573 HeapWord* addr = (HeapWord*)obj; 7574 return addr != NULL && 7575 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7576 } 7577 7578 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7579 MemRegion span, 7580 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7581 bool cpc): 7582 _collector(collector), 7583 _span(span), 7584 _mark_stack(mark_stack), 7585 _bit_map(bit_map), 7586 _concurrent_precleaning(cpc) { 7587 assert(!_span.is_empty(), "Empty span could spell trouble"); 7588 } 7589 7590 7591 // CMSKeepAliveClosure: the serial version 7592 void CMSKeepAliveClosure::do_oop(oop obj) { 7593 HeapWord* addr = (HeapWord*)obj; 7594 if (_span.contains(addr) && 7595 !_bit_map->isMarked(addr)) { 7596 _bit_map->mark(addr); 7597 bool simulate_overflow = false; 7598 NOT_PRODUCT( 7599 if (CMSMarkStackOverflowALot && 7600 _collector->simulate_overflow()) { 7601 // simulate a stack overflow 7602 simulate_overflow = true; 7603 } 7604 ) 7605 if (simulate_overflow || !_mark_stack->push(obj)) { 7606 if (_concurrent_precleaning) { 7607 // We dirty the overflown object and let the remark 7608 // phase deal with it. 7609 assert(_collector->overflow_list_is_empty(), "Error"); 7610 // In the case of object arrays, we need to dirty all of 7611 // the cards that the object spans. No locking or atomics 7612 // are needed since no one else can be mutating the mod union 7613 // table. 7614 if (obj->is_objArray()) { 7615 size_t sz = obj->size(); 7616 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7617 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7618 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7619 _collector->_modUnionTable.mark_range(redirty_range); 7620 } else { 7621 _collector->_modUnionTable.mark(addr); 7622 } 7623 _collector->_ser_kac_preclean_ovflw++; 7624 } else { 7625 _collector->push_on_overflow_list(obj); 7626 _collector->_ser_kac_ovflw++; 7627 } 7628 } 7629 } 7630 } 7631 7632 // CMSParKeepAliveClosure: a parallel version of the above. 7633 // The work queues are private to each closure (thread), 7634 // but (may be) available for stealing by other threads. 7635 void CMSParKeepAliveClosure::do_oop(oop obj) { 7636 HeapWord* addr = (HeapWord*)obj; 7637 if (_span.contains(addr) && 7638 !_bit_map->isMarked(addr)) { 7639 // In general, during recursive tracing, several threads 7640 // may be concurrently getting here; the first one to 7641 // "tag" it, claims it. 7642 if (_bit_map->par_mark(addr)) { 7643 bool res = _work_queue->push(obj); 7644 assert(res, "Low water mark should be much less than capacity"); 7645 // Do a recursive trim in the hope that this will keep 7646 // stack usage lower, but leave some oops for potential stealers 7647 trim_queue(_low_water_mark); 7648 } // Else, another thread got there first 7649 } 7650 } 7651 7652 void CMSParKeepAliveClosure::trim_queue(uint max) { 7653 while (_work_queue->size() > max) { 7654 oop new_oop; 7655 if (_work_queue->pop_local(new_oop)) { 7656 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7657 assert(_bit_map->isMarked((HeapWord*)new_oop), 7658 "no white objects on this stack!"); 7659 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7660 // iterate over the oops in this oop, marking and pushing 7661 // the ones in CMS heap (i.e. in _span). 7662 new_oop->oop_iterate(&_mark_and_push); 7663 } 7664 } 7665 } 7666 7667 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7668 CMSCollector* collector, 7669 MemRegion span, CMSBitMap* bit_map, 7670 OopTaskQueue* work_queue): 7671 _collector(collector), 7672 _span(span), 7673 _work_queue(work_queue), 7674 _bit_map(bit_map) { } 7675 7676 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7677 HeapWord* addr = (HeapWord*)obj; 7678 if (_span.contains(addr) && 7679 !_bit_map->isMarked(addr)) { 7680 if (_bit_map->par_mark(addr)) { 7681 bool simulate_overflow = false; 7682 NOT_PRODUCT( 7683 if (CMSMarkStackOverflowALot && 7684 _collector->par_simulate_overflow()) { 7685 // simulate a stack overflow 7686 simulate_overflow = true; 7687 } 7688 ) 7689 if (simulate_overflow || !_work_queue->push(obj)) { 7690 _collector->par_push_on_overflow_list(obj); 7691 _collector->_par_kac_ovflw++; 7692 } 7693 } // Else another thread got there already 7694 } 7695 } 7696 7697 ////////////////////////////////////////////////////////////////// 7698 // CMSExpansionCause ///////////////////////////// 7699 ////////////////////////////////////////////////////////////////// 7700 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7701 switch (cause) { 7702 case _no_expansion: 7703 return "No expansion"; 7704 case _satisfy_free_ratio: 7705 return "Free ratio"; 7706 case _satisfy_promotion: 7707 return "Satisfy promotion"; 7708 case _satisfy_allocation: 7709 return "allocation"; 7710 case _allocate_par_lab: 7711 return "Par LAB"; 7712 case _allocate_par_spooling_space: 7713 return "Par Spooling Space"; 7714 case _adaptive_size_policy: 7715 return "Ergonomics"; 7716 default: 7717 return "unknown"; 7718 } 7719 } 7720 7721 void CMSDrainMarkingStackClosure::do_void() { 7722 // the max number to take from overflow list at a time 7723 const size_t num = _mark_stack->capacity()/4; 7724 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7725 "Overflow list should be NULL during concurrent phases"); 7726 while (!_mark_stack->isEmpty() || 7727 // if stack is empty, check the overflow list 7728 _collector->take_from_overflow_list(num, _mark_stack)) { 7729 oop obj = _mark_stack->pop(); 7730 HeapWord* addr = (HeapWord*)obj; 7731 assert(_span.contains(addr), "Should be within span"); 7732 assert(_bit_map->isMarked(addr), "Should be marked"); 7733 assert(oopDesc::is_oop(obj), "Should be an oop"); 7734 obj->oop_iterate(_keep_alive); 7735 } 7736 } 7737 7738 void CMSParDrainMarkingStackClosure::do_void() { 7739 // drain queue 7740 trim_queue(0); 7741 } 7742 7743 // Trim our work_queue so its length is below max at return 7744 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7745 while (_work_queue->size() > max) { 7746 oop new_oop; 7747 if (_work_queue->pop_local(new_oop)) { 7748 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7749 assert(_bit_map->isMarked((HeapWord*)new_oop), 7750 "no white objects on this stack!"); 7751 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7752 // iterate over the oops in this oop, marking and pushing 7753 // the ones in CMS heap (i.e. in _span). 7754 new_oop->oop_iterate(&_mark_and_push); 7755 } 7756 } 7757 } 7758 7759 //////////////////////////////////////////////////////////////////// 7760 // Support for Marking Stack Overflow list handling and related code 7761 //////////////////////////////////////////////////////////////////// 7762 // Much of the following code is similar in shape and spirit to the 7763 // code used in ParNewGC. We should try and share that code 7764 // as much as possible in the future. 7765 7766 #ifndef PRODUCT 7767 // Debugging support for CMSStackOverflowALot 7768 7769 // It's OK to call this multi-threaded; the worst thing 7770 // that can happen is that we'll get a bunch of closely 7771 // spaced simulated overflows, but that's OK, in fact 7772 // probably good as it would exercise the overflow code 7773 // under contention. 7774 bool CMSCollector::simulate_overflow() { 7775 if (_overflow_counter-- <= 0) { // just being defensive 7776 _overflow_counter = CMSMarkStackOverflowInterval; 7777 return true; 7778 } else { 7779 return false; 7780 } 7781 } 7782 7783 bool CMSCollector::par_simulate_overflow() { 7784 return simulate_overflow(); 7785 } 7786 #endif 7787 7788 // Single-threaded 7789 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7790 assert(stack->isEmpty(), "Expected precondition"); 7791 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7792 size_t i = num; 7793 oop cur = _overflow_list; 7794 const markOop proto = markOopDesc::prototype(); 7795 NOT_PRODUCT(ssize_t n = 0;) 7796 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7797 next = oop(cur->mark_raw()); 7798 cur->set_mark_raw(proto); // until proven otherwise 7799 assert(oopDesc::is_oop(cur), "Should be an oop"); 7800 bool res = stack->push(cur); 7801 assert(res, "Bit off more than can chew?"); 7802 NOT_PRODUCT(n++;) 7803 } 7804 _overflow_list = cur; 7805 #ifndef PRODUCT 7806 assert(_num_par_pushes >= n, "Too many pops?"); 7807 _num_par_pushes -=n; 7808 #endif 7809 return !stack->isEmpty(); 7810 } 7811 7812 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7813 // (MT-safe) Get a prefix of at most "num" from the list. 7814 // The overflow list is chained through the mark word of 7815 // each object in the list. We fetch the entire list, 7816 // break off a prefix of the right size and return the 7817 // remainder. If other threads try to take objects from 7818 // the overflow list at that time, they will wait for 7819 // some time to see if data becomes available. If (and 7820 // only if) another thread places one or more object(s) 7821 // on the global list before we have returned the suffix 7822 // to the global list, we will walk down our local list 7823 // to find its end and append the global list to 7824 // our suffix before returning it. This suffix walk can 7825 // prove to be expensive (quadratic in the amount of traffic) 7826 // when there are many objects in the overflow list and 7827 // there is much producer-consumer contention on the list. 7828 // *NOTE*: The overflow list manipulation code here and 7829 // in ParNewGeneration:: are very similar in shape, 7830 // except that in the ParNew case we use the old (from/eden) 7831 // copy of the object to thread the list via its klass word. 7832 // Because of the common code, if you make any changes in 7833 // the code below, please check the ParNew version to see if 7834 // similar changes might be needed. 7835 // CR 6797058 has been filed to consolidate the common code. 7836 bool CMSCollector::par_take_from_overflow_list(size_t num, 7837 OopTaskQueue* work_q, 7838 int no_of_gc_threads) { 7839 assert(work_q->size() == 0, "First empty local work queue"); 7840 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7841 if (_overflow_list == NULL) { 7842 return false; 7843 } 7844 // Grab the entire list; we'll put back a suffix 7845 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7846 Thread* tid = Thread::current(); 7847 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7848 // set to ParallelGCThreads. 7849 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7850 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7851 // If the list is busy, we spin for a short while, 7852 // sleeping between attempts to get the list. 7853 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7854 os::sleep(tid, sleep_time_millis, false); 7855 if (_overflow_list == NULL) { 7856 // Nothing left to take 7857 return false; 7858 } else if (_overflow_list != BUSY) { 7859 // Try and grab the prefix 7860 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7861 } 7862 } 7863 // If the list was found to be empty, or we spun long 7864 // enough, we give up and return empty-handed. If we leave 7865 // the list in the BUSY state below, it must be the case that 7866 // some other thread holds the overflow list and will set it 7867 // to a non-BUSY state in the future. 7868 if (prefix == NULL || prefix == BUSY) { 7869 // Nothing to take or waited long enough 7870 if (prefix == NULL) { 7871 // Write back the NULL in case we overwrote it with BUSY above 7872 // and it is still the same value. 7873 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7874 } 7875 return false; 7876 } 7877 assert(prefix != NULL && prefix != BUSY, "Error"); 7878 size_t i = num; 7879 oop cur = prefix; 7880 // Walk down the first "num" objects, unless we reach the end. 7881 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7882 if (cur->mark_raw() == NULL) { 7883 // We have "num" or fewer elements in the list, so there 7884 // is nothing to return to the global list. 7885 // Write back the NULL in lieu of the BUSY we wrote 7886 // above, if it is still the same value. 7887 if (_overflow_list == BUSY) { 7888 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7889 } 7890 } else { 7891 // Chop off the suffix and return it to the global list. 7892 assert(cur->mark_raw() != BUSY, "Error"); 7893 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7894 cur->set_mark_raw(NULL); // break off suffix 7895 // It's possible that the list is still in the empty(busy) state 7896 // we left it in a short while ago; in that case we may be 7897 // able to place back the suffix without incurring the cost 7898 // of a walk down the list. 7899 oop observed_overflow_list = _overflow_list; 7900 oop cur_overflow_list = observed_overflow_list; 7901 bool attached = false; 7902 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7903 observed_overflow_list = 7904 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7905 if (cur_overflow_list == observed_overflow_list) { 7906 attached = true; 7907 break; 7908 } else cur_overflow_list = observed_overflow_list; 7909 } 7910 if (!attached) { 7911 // Too bad, someone else sneaked in (at least) an element; we'll need 7912 // to do a splice. Find tail of suffix so we can prepend suffix to global 7913 // list. 7914 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7915 oop suffix_tail = cur; 7916 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7917 "Tautology"); 7918 observed_overflow_list = _overflow_list; 7919 do { 7920 cur_overflow_list = observed_overflow_list; 7921 if (cur_overflow_list != BUSY) { 7922 // Do the splice ... 7923 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7924 } else { // cur_overflow_list == BUSY 7925 suffix_tail->set_mark_raw(NULL); 7926 } 7927 // ... and try to place spliced list back on overflow_list ... 7928 observed_overflow_list = 7929 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7930 } while (cur_overflow_list != observed_overflow_list); 7931 // ... until we have succeeded in doing so. 7932 } 7933 } 7934 7935 // Push the prefix elements on work_q 7936 assert(prefix != NULL, "control point invariant"); 7937 const markOop proto = markOopDesc::prototype(); 7938 oop next; 7939 NOT_PRODUCT(ssize_t n = 0;) 7940 for (cur = prefix; cur != NULL; cur = next) { 7941 next = oop(cur->mark_raw()); 7942 cur->set_mark_raw(proto); // until proven otherwise 7943 assert(oopDesc::is_oop(cur), "Should be an oop"); 7944 bool res = work_q->push(cur); 7945 assert(res, "Bit off more than we can chew?"); 7946 NOT_PRODUCT(n++;) 7947 } 7948 #ifndef PRODUCT 7949 assert(_num_par_pushes >= n, "Too many pops?"); 7950 Atomic::sub(n, &_num_par_pushes); 7951 #endif 7952 return true; 7953 } 7954 7955 // Single-threaded 7956 void CMSCollector::push_on_overflow_list(oop p) { 7957 NOT_PRODUCT(_num_par_pushes++;) 7958 assert(oopDesc::is_oop(p), "Not an oop"); 7959 preserve_mark_if_necessary(p); 7960 p->set_mark_raw((markOop)_overflow_list); 7961 _overflow_list = p; 7962 } 7963 7964 // Multi-threaded; use CAS to prepend to overflow list 7965 void CMSCollector::par_push_on_overflow_list(oop p) { 7966 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7967 assert(oopDesc::is_oop(p), "Not an oop"); 7968 par_preserve_mark_if_necessary(p); 7969 oop observed_overflow_list = _overflow_list; 7970 oop cur_overflow_list; 7971 do { 7972 cur_overflow_list = observed_overflow_list; 7973 if (cur_overflow_list != BUSY) { 7974 p->set_mark_raw(markOop(cur_overflow_list)); 7975 } else { 7976 p->set_mark_raw(NULL); 7977 } 7978 observed_overflow_list = 7979 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7980 } while (cur_overflow_list != observed_overflow_list); 7981 } 7982 #undef BUSY 7983 7984 // Single threaded 7985 // General Note on GrowableArray: pushes may silently fail 7986 // because we are (temporarily) out of C-heap for expanding 7987 // the stack. The problem is quite ubiquitous and affects 7988 // a lot of code in the JVM. The prudent thing for GrowableArray 7989 // to do (for now) is to exit with an error. However, that may 7990 // be too draconian in some cases because the caller may be 7991 // able to recover without much harm. For such cases, we 7992 // should probably introduce a "soft_push" method which returns 7993 // an indication of success or failure with the assumption that 7994 // the caller may be able to recover from a failure; code in 7995 // the VM can then be changed, incrementally, to deal with such 7996 // failures where possible, thus, incrementally hardening the VM 7997 // in such low resource situations. 7998 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7999 _preserved_oop_stack.push(p); 8000 _preserved_mark_stack.push(m); 8001 assert(m == p->mark_raw(), "Mark word changed"); 8002 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8003 "bijection"); 8004 } 8005 8006 // Single threaded 8007 void CMSCollector::preserve_mark_if_necessary(oop p) { 8008 markOop m = p->mark_raw(); 8009 if (m->must_be_preserved(p)) { 8010 preserve_mark_work(p, m); 8011 } 8012 } 8013 8014 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8015 markOop m = p->mark_raw(); 8016 if (m->must_be_preserved(p)) { 8017 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8018 // Even though we read the mark word without holding 8019 // the lock, we are assured that it will not change 8020 // because we "own" this oop, so no other thread can 8021 // be trying to push it on the overflow list; see 8022 // the assertion in preserve_mark_work() that checks 8023 // that m == p->mark_raw(). 8024 preserve_mark_work(p, m); 8025 } 8026 } 8027 8028 // We should be able to do this multi-threaded, 8029 // a chunk of stack being a task (this is 8030 // correct because each oop only ever appears 8031 // once in the overflow list. However, it's 8032 // not very easy to completely overlap this with 8033 // other operations, so will generally not be done 8034 // until all work's been completed. Because we 8035 // expect the preserved oop stack (set) to be small, 8036 // it's probably fine to do this single-threaded. 8037 // We can explore cleverer concurrent/overlapped/parallel 8038 // processing of preserved marks if we feel the 8039 // need for this in the future. Stack overflow should 8040 // be so rare in practice and, when it happens, its 8041 // effect on performance so great that this will 8042 // likely just be in the noise anyway. 8043 void CMSCollector::restore_preserved_marks_if_any() { 8044 assert(SafepointSynchronize::is_at_safepoint(), 8045 "world should be stopped"); 8046 assert(Thread::current()->is_ConcurrentGC_thread() || 8047 Thread::current()->is_VM_thread(), 8048 "should be single-threaded"); 8049 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8050 "bijection"); 8051 8052 while (!_preserved_oop_stack.is_empty()) { 8053 oop p = _preserved_oop_stack.pop(); 8054 assert(oopDesc::is_oop(p), "Should be an oop"); 8055 assert(_span.contains(p), "oop should be in _span"); 8056 assert(p->mark_raw() == markOopDesc::prototype(), 8057 "Set when taken from overflow list"); 8058 markOop m = _preserved_mark_stack.pop(); 8059 p->set_mark_raw(m); 8060 } 8061 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8062 "stacks were cleared above"); 8063 } 8064 8065 #ifndef PRODUCT 8066 bool CMSCollector::no_preserved_marks() const { 8067 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8068 } 8069 #endif 8070 8071 // Transfer some number of overflown objects to usual marking 8072 // stack. Return true if some objects were transferred. 8073 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8074 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8075 (size_t)ParGCDesiredObjsFromOverflowList); 8076 8077 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8078 assert(_collector->overflow_list_is_empty() || res, 8079 "If list is not empty, we should have taken something"); 8080 assert(!res || !_mark_stack->isEmpty(), 8081 "If we took something, it should now be on our stack"); 8082 return res; 8083 } 8084 8085 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8086 size_t res = _sp->block_size_no_stall(addr, _collector); 8087 if (_sp->block_is_obj(addr)) { 8088 if (_live_bit_map->isMarked(addr)) { 8089 // It can't have been dead in a previous cycle 8090 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8091 } else { 8092 _dead_bit_map->mark(addr); // mark the dead object 8093 } 8094 } 8095 // Could be 0, if the block size could not be computed without stalling. 8096 return res; 8097 } 8098 8099 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8100 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8101 switch (phase) { 8102 case CMSCollector::InitialMarking: 8103 initialize(manager /* GC manager */ , 8104 cause /* cause of the GC */, 8105 true /* allMemoryPoolsAffected */, 8106 true /* recordGCBeginTime */, 8107 true /* recordPreGCUsage */, 8108 false /* recordPeakUsage */, 8109 false /* recordPostGCusage */, 8110 true /* recordAccumulatedGCTime */, 8111 false /* recordGCEndTime */, 8112 false /* countCollection */ ); 8113 break; 8114 8115 case CMSCollector::FinalMarking: 8116 initialize(manager /* GC manager */ , 8117 cause /* cause of the GC */, 8118 true /* allMemoryPoolsAffected */, 8119 false /* recordGCBeginTime */, 8120 false /* recordPreGCUsage */, 8121 false /* recordPeakUsage */, 8122 false /* recordPostGCusage */, 8123 true /* recordAccumulatedGCTime */, 8124 false /* recordGCEndTime */, 8125 false /* countCollection */ ); 8126 break; 8127 8128 case CMSCollector::Sweeping: 8129 initialize(manager /* GC manager */ , 8130 cause /* cause of the GC */, 8131 true /* allMemoryPoolsAffected */, 8132 false /* recordGCBeginTime */, 8133 false /* recordPreGCUsage */, 8134 true /* recordPeakUsage */, 8135 true /* recordPostGCusage */, 8136 false /* recordAccumulatedGCTime */, 8137 true /* recordGCEndTime */, 8138 true /* countCollection */ ); 8139 break; 8140 8141 default: 8142 ShouldNotReachHere(); 8143 } 8144 }