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/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/cms/cmsGCStats.hpp" 30 #include "gc/cms/cmsHeap.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/cmsVMOperations.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/promotionInfo.inline.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/gcLocker.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.inline.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/owstTaskTerminator.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 56 #include "gc/shared/space.inline.hpp" 57 #include "gc/shared/strongRootsScope.hpp" 58 #include "gc/shared/taskqueue.inline.hpp" 59 #include "gc/shared/weakProcessor.hpp" 60 #include "gc/shared/workerPolicy.hpp" 61 #include "logging/log.hpp" 62 #include "logging/logStream.hpp" 63 #include "memory/allocation.hpp" 64 #include "memory/binaryTreeDictionary.inline.hpp" 65 #include "memory/iterator.inline.hpp" 66 #include "memory/padded.hpp" 67 #include "memory/resourceArea.hpp" 68 #include "memory/universe.hpp" 69 #include "oops/access.inline.hpp" 70 #include "oops/oop.inline.hpp" 71 #include "prims/jvmtiExport.hpp" 72 #include "runtime/atomic.hpp" 73 #include "runtime/flags/flagSetting.hpp" 74 #include "runtime/globals_extension.hpp" 75 #include "runtime/handles.inline.hpp" 76 #include "runtime/java.hpp" 77 #include "runtime/orderAccess.hpp" 78 #include "runtime/timer.hpp" 79 #include "runtime/vmThread.hpp" 80 #include "services/memoryService.hpp" 81 #include "services/runtimeService.hpp" 82 #include "utilities/align.hpp" 83 #include "utilities/stack.inline.hpp" 84 #if INCLUDE_JVMCI 85 #include "jvmci/jvmci.hpp" 86 #endif 87 88 // statics 89 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 90 bool CMSCollector::_full_gc_requested = false; 91 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 92 93 ////////////////////////////////////////////////////////////////// 94 // In support of CMS/VM thread synchronization 95 ////////////////////////////////////////////////////////////////// 96 // We split use of the CGC_lock into 2 "levels". 97 // The low-level locking is of the usual CGC_lock monitor. We introduce 98 // a higher level "token" (hereafter "CMS token") built on top of the 99 // low level monitor (hereafter "CGC lock"). 100 // The token-passing protocol gives priority to the VM thread. The 101 // CMS-lock doesn't provide any fairness guarantees, but clients 102 // should ensure that it is only held for very short, bounded 103 // durations. 104 // 105 // When either of the CMS thread or the VM thread is involved in 106 // collection operations during which it does not want the other 107 // thread to interfere, it obtains the CMS token. 108 // 109 // If either thread tries to get the token while the other has 110 // it, that thread waits. However, if the VM thread and CMS thread 111 // both want the token, then the VM thread gets priority while the 112 // CMS thread waits. This ensures, for instance, that the "concurrent" 113 // phases of the CMS thread's work do not block out the VM thread 114 // for long periods of time as the CMS thread continues to hog 115 // the token. (See bug 4616232). 116 // 117 // The baton-passing functions are, however, controlled by the 118 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 119 // and here the low-level CMS lock, not the high level token, 120 // ensures mutual exclusion. 121 // 122 // Two important conditions that we have to satisfy: 123 // 1. if a thread does a low-level wait on the CMS lock, then it 124 // relinquishes the CMS token if it were holding that token 125 // when it acquired the low-level CMS lock. 126 // 2. any low-level notifications on the low-level lock 127 // should only be sent when a thread has relinquished the token. 128 // 129 // In the absence of either property, we'd have potential deadlock. 130 // 131 // We protect each of the CMS (concurrent and sequential) phases 132 // with the CMS _token_, not the CMS _lock_. 133 // 134 // The only code protected by CMS lock is the token acquisition code 135 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 136 // baton-passing code. 137 // 138 // Unfortunately, i couldn't come up with a good abstraction to factor and 139 // hide the naked CGC_lock manipulation in the baton-passing code 140 // further below. That's something we should try to do. Also, the proof 141 // of correctness of this 2-level locking scheme is far from obvious, 142 // and potentially quite slippery. We have an uneasy suspicion, for instance, 143 // that there may be a theoretical possibility of delay/starvation in the 144 // low-level lock/wait/notify scheme used for the baton-passing because of 145 // potential interference with the priority scheme embodied in the 146 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 147 // invocation further below and marked with "XXX 20011219YSR". 148 // Indeed, as we note elsewhere, this may become yet more slippery 149 // in the presence of multiple CMS and/or multiple VM threads. XXX 150 151 class CMSTokenSync: public StackObj { 152 private: 153 bool _is_cms_thread; 154 public: 155 CMSTokenSync(bool is_cms_thread): 156 _is_cms_thread(is_cms_thread) { 157 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 158 "Incorrect argument to constructor"); 159 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 160 } 161 162 ~CMSTokenSync() { 163 assert(_is_cms_thread ? 164 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 165 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 166 "Incorrect state"); 167 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 168 } 169 }; 170 171 // Convenience class that does a CMSTokenSync, and then acquires 172 // upto three locks. 173 class CMSTokenSyncWithLocks: public CMSTokenSync { 174 private: 175 // Note: locks are acquired in textual declaration order 176 // and released in the opposite order 177 MutexLocker _locker1, _locker2, _locker3; 178 public: 179 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 180 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 181 CMSTokenSync(is_cms_thread), 182 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 183 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 184 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 185 { } 186 }; 187 188 189 ////////////////////////////////////////////////////////////////// 190 // Concurrent Mark-Sweep Generation ///////////////////////////// 191 ////////////////////////////////////////////////////////////////// 192 193 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 194 195 // This struct contains per-thread things necessary to support parallel 196 // young-gen collection. 197 class CMSParGCThreadState: public CHeapObj<mtGC> { 198 public: 199 CompactibleFreeListSpaceLAB lab; 200 PromotionInfo promo; 201 202 // Constructor. 203 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 204 promo.setSpace(cfls); 205 } 206 }; 207 208 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 209 ReservedSpace rs, 210 size_t initial_byte_size, 211 size_t min_byte_size, 212 size_t max_byte_size, 213 CardTableRS* ct) : 214 CardGeneration(rs, initial_byte_size, ct), 215 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 216 _did_compact(false) 217 { 218 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 219 HeapWord* end = (HeapWord*) _virtual_space.high(); 220 221 _direct_allocated_words = 0; 222 NOT_PRODUCT( 223 _numObjectsPromoted = 0; 224 _numWordsPromoted = 0; 225 _numObjectsAllocated = 0; 226 _numWordsAllocated = 0; 227 ) 228 229 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 230 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 231 _cmsSpace->_old_gen = this; 232 233 _gc_stats = new CMSGCStats(); 234 235 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 236 // offsets match. The ability to tell free chunks from objects 237 // depends on this property. 238 debug_only( 239 FreeChunk* junk = NULL; 240 assert(UseCompressedClassPointers || 241 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 242 "Offset of FreeChunk::_prev within FreeChunk must match" 243 " that of OopDesc::_klass within OopDesc"); 244 ) 245 246 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 247 for (uint i = 0; i < ParallelGCThreads; i++) { 248 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 249 } 250 251 _incremental_collection_failed = false; 252 // The "dilatation_factor" is the expansion that can occur on 253 // account of the fact that the minimum object size in the CMS 254 // generation may be larger than that in, say, a contiguous young 255 // generation. 256 // Ideally, in the calculation below, we'd compute the dilatation 257 // factor as: MinChunkSize/(promoting_gen's min object size) 258 // Since we do not have such a general query interface for the 259 // promoting generation, we'll instead just use the minimum 260 // object size (which today is a header's worth of space); 261 // note that all arithmetic is in units of HeapWords. 262 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 263 assert(_dilatation_factor >= 1.0, "from previous assert"); 264 265 initialize_performance_counters(min_byte_size, max_byte_size); 266 } 267 268 269 // The field "_initiating_occupancy" represents the occupancy percentage 270 // at which we trigger a new collection cycle. Unless explicitly specified 271 // via CMSInitiatingOccupancyFraction (argument "io" below), it 272 // is calculated by: 273 // 274 // Let "f" be MinHeapFreeRatio in 275 // 276 // _initiating_occupancy = 100-f + 277 // f * (CMSTriggerRatio/100) 278 // where CMSTriggerRatio is the argument "tr" below. 279 // 280 // That is, if we assume the heap is at its desired maximum occupancy at the 281 // end of a collection, we let CMSTriggerRatio of the (purported) free 282 // space be allocated before initiating a new collection cycle. 283 // 284 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 285 assert(io <= 100 && tr <= 100, "Check the arguments"); 286 if (io >= 0) { 287 _initiating_occupancy = (double)io / 100.0; 288 } else { 289 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 290 (double)(tr * MinHeapFreeRatio) / 100.0) 291 / 100.0; 292 } 293 } 294 295 void ConcurrentMarkSweepGeneration::ref_processor_init() { 296 assert(collector() != NULL, "no collector"); 297 collector()->ref_processor_init(); 298 } 299 300 void CMSCollector::ref_processor_init() { 301 if (_ref_processor == NULL) { 302 // Allocate and initialize a reference processor 303 _ref_processor = 304 new ReferenceProcessor(&_span_based_discoverer, 305 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 306 ParallelGCThreads, // mt processing degree 307 _cmsGen->refs_discovery_is_mt(), // mt discovery 308 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 309 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 310 &_is_alive_closure, // closure for liveness info 311 false); // disable adjusting number of processing threads 312 // Initialize the _ref_processor field of CMSGen 313 _cmsGen->set_ref_processor(_ref_processor); 314 315 } 316 } 317 318 AdaptiveSizePolicy* CMSCollector::size_policy() { 319 return CMSHeap::heap()->size_policy(); 320 } 321 322 void ConcurrentMarkSweepGeneration::initialize_performance_counters(size_t min_old_size, 323 size_t max_old_size) { 324 325 const char* gen_name = "old"; 326 // Generation Counters - generation 1, 1 subspace 327 _gen_counters = new GenerationCounters(gen_name, 1, 1, 328 min_old_size, max_old_size, &_virtual_space); 329 330 _space_counters = new GSpaceCounters(gen_name, 0, 331 _virtual_space.reserved_size(), 332 this, _gen_counters); 333 } 334 335 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 336 _cms_gen(cms_gen) 337 { 338 assert(alpha <= 100, "bad value"); 339 _saved_alpha = alpha; 340 341 // Initialize the alphas to the bootstrap value of 100. 342 _gc0_alpha = _cms_alpha = 100; 343 344 _cms_begin_time.update(); 345 _cms_end_time.update(); 346 347 _gc0_duration = 0.0; 348 _gc0_period = 0.0; 349 _gc0_promoted = 0; 350 351 _cms_duration = 0.0; 352 _cms_period = 0.0; 353 _cms_allocated = 0; 354 355 _cms_used_at_gc0_begin = 0; 356 _cms_used_at_gc0_end = 0; 357 _allow_duty_cycle_reduction = false; 358 _valid_bits = 0; 359 } 360 361 double CMSStats::cms_free_adjustment_factor(size_t free) const { 362 // TBD: CR 6909490 363 return 1.0; 364 } 365 366 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 367 } 368 369 // If promotion failure handling is on use 370 // the padded average size of the promotion for each 371 // young generation collection. 372 double CMSStats::time_until_cms_gen_full() const { 373 size_t cms_free = _cms_gen->cmsSpace()->free(); 374 CMSHeap* heap = CMSHeap::heap(); 375 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 376 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 377 if (cms_free > expected_promotion) { 378 // Start a cms collection if there isn't enough space to promote 379 // for the next young collection. Use the padded average as 380 // a safety factor. 381 cms_free -= expected_promotion; 382 383 // Adjust by the safety factor. 384 double cms_free_dbl = (double)cms_free; 385 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 386 // Apply a further correction factor which tries to adjust 387 // for recent occurance of concurrent mode failures. 388 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 389 cms_free_dbl = cms_free_dbl * cms_adjustment; 390 391 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 392 cms_free, expected_promotion); 393 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 394 // Add 1 in case the consumption rate goes to zero. 395 return cms_free_dbl / (cms_consumption_rate() + 1.0); 396 } 397 return 0.0; 398 } 399 400 // Compare the duration of the cms collection to the 401 // time remaining before the cms generation is empty. 402 // Note that the time from the start of the cms collection 403 // to the start of the cms sweep (less than the total 404 // duration of the cms collection) can be used. This 405 // has been tried and some applications experienced 406 // promotion failures early in execution. This was 407 // possibly because the averages were not accurate 408 // enough at the beginning. 409 double CMSStats::time_until_cms_start() const { 410 // We add "gc0_period" to the "work" calculation 411 // below because this query is done (mostly) at the 412 // end of a scavenge, so we need to conservatively 413 // account for that much possible delay 414 // in the query so as to avoid concurrent mode failures 415 // due to starting the collection just a wee bit too 416 // late. 417 double work = cms_duration() + gc0_period(); 418 double deadline = time_until_cms_gen_full(); 419 // If a concurrent mode failure occurred recently, we want to be 420 // more conservative and halve our expected time_until_cms_gen_full() 421 if (work > deadline) { 422 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 423 cms_duration(), gc0_period(), time_until_cms_gen_full()); 424 return 0.0; 425 } 426 return work - deadline; 427 } 428 429 #ifndef PRODUCT 430 void CMSStats::print_on(outputStream *st) const { 431 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 432 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 433 gc0_duration(), gc0_period(), gc0_promoted()); 434 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 435 cms_duration(), cms_period(), cms_allocated()); 436 st->print(",cms_since_beg=%g,cms_since_end=%g", 437 cms_time_since_begin(), cms_time_since_end()); 438 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 439 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 440 441 if (valid()) { 442 st->print(",promo_rate=%g,cms_alloc_rate=%g", 443 promotion_rate(), cms_allocation_rate()); 444 st->print(",cms_consumption_rate=%g,time_until_full=%g", 445 cms_consumption_rate(), time_until_cms_gen_full()); 446 } 447 st->cr(); 448 } 449 #endif // #ifndef PRODUCT 450 451 CMSCollector::CollectorState CMSCollector::_collectorState = 452 CMSCollector::Idling; 453 bool CMSCollector::_foregroundGCIsActive = false; 454 bool CMSCollector::_foregroundGCShouldWait = false; 455 456 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 457 CardTableRS* ct): 458 _overflow_list(NULL), 459 _conc_workers(NULL), // may be set later 460 _completed_initialization(false), 461 _collection_count_start(0), 462 _should_unload_classes(CMSClassUnloadingEnabled), 463 _concurrent_cycles_since_last_unload(0), 464 _roots_scanning_options(GenCollectedHeap::SO_None), 465 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 466 _verifying(false), 467 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 468 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 469 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 470 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 471 _cms_start_registered(false), 472 _cmsGen(cmsGen), 473 // Adjust span to cover old (cms) gen 474 _span(cmsGen->reserved()), 475 _ct(ct), 476 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 477 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 478 -1 /* lock-free */, "No_lock" /* dummy */), 479 _restart_addr(NULL), 480 _ser_pmc_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_preclean_ovflw(0), 484 _ser_kac_ovflw(0), 485 _par_kac_ovflw(0), 486 #ifndef PRODUCT 487 _num_par_pushes(0), 488 #endif 489 _span_based_discoverer(_span), 490 _ref_processor(NULL), // will be set later 491 // Construct the is_alive_closure with _span & markBitMap 492 _is_alive_closure(_span, &_markBitMap), 493 _modUnionClosurePar(&_modUnionTable), 494 _between_prologue_and_epilogue(false), 495 _abort_preclean(false), 496 _start_sampling(false), 497 _stats(cmsGen), 498 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 499 //verify that this lock should be acquired with safepoint check. 500 Monitor::_safepoint_check_never)), 501 _eden_chunk_array(NULL), // may be set in ctor body 502 _eden_chunk_index(0), // -- ditto -- 503 _eden_chunk_capacity(0), // -- ditto -- 504 _survivor_chunk_array(NULL), // -- ditto -- 505 _survivor_chunk_index(0), // -- ditto -- 506 _survivor_chunk_capacity(0), // -- ditto -- 507 _survivor_plab_array(NULL) // -- ditto -- 508 { 509 // Now expand the span and allocate the collection support structures 510 // (MUT, marking bit map etc.) to cover both generations subject to 511 // collection. 512 513 // For use by dirty card to oop closures. 514 _cmsGen->cmsSpace()->set_collector(this); 515 516 // Allocate MUT and marking bit map 517 { 518 MutexLocker x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 519 if (!_markBitMap.allocate(_span)) { 520 log_warning(gc)("Failed to allocate CMS Bit Map"); 521 return; 522 } 523 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 524 } 525 { 526 _modUnionTable.allocate(_span); 527 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 528 } 529 530 if (!_markStack.allocate(MarkStackSize)) { 531 log_warning(gc)("Failed to allocate CMS Marking Stack"); 532 return; 533 } 534 535 // Support for multi-threaded concurrent phases 536 if (CMSConcurrentMTEnabled) { 537 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 538 // just for now 539 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 540 } 541 if (ConcGCThreads > 1) { 542 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 543 ConcGCThreads, true); 544 if (_conc_workers == NULL) { 545 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 546 CMSConcurrentMTEnabled = false; 547 } else { 548 _conc_workers->initialize_workers(); 549 } 550 } else { 551 CMSConcurrentMTEnabled = false; 552 } 553 } 554 if (!CMSConcurrentMTEnabled) { 555 ConcGCThreads = 0; 556 } else { 557 // Turn off CMSCleanOnEnter optimization temporarily for 558 // the MT case where it's not fixed yet; see 6178663. 559 CMSCleanOnEnter = false; 560 } 561 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 562 "Inconsistency"); 563 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 564 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 565 566 // Parallel task queues; these are shared for the 567 // concurrent and stop-world phases of CMS, but 568 // are not shared with parallel scavenge (ParNew). 569 { 570 uint i; 571 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 572 573 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 574 || ParallelRefProcEnabled) 575 && num_queues > 0) { 576 _task_queues = new OopTaskQueueSet(num_queues); 577 if (_task_queues == NULL) { 578 log_warning(gc)("task_queues allocation failure."); 579 return; 580 } 581 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 582 for (i = 0; i < num_queues; i++) { 583 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 584 if (q == NULL) { 585 log_warning(gc)("work_queue allocation failure."); 586 return; 587 } 588 _task_queues->register_queue(i, q); 589 } 590 for (i = 0; i < num_queues; i++) { 591 _task_queues->queue(i)->initialize(); 592 } 593 } 594 } 595 596 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 597 598 // Clip CMSBootstrapOccupancy between 0 and 100. 599 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 600 601 // Now tell CMS generations the identity of their collector 602 ConcurrentMarkSweepGeneration::set_collector(this); 603 604 // Create & start a CMS thread for this CMS collector 605 _cmsThread = ConcurrentMarkSweepThread::start(this); 606 assert(cmsThread() != NULL, "CMS Thread should have been created"); 607 assert(cmsThread()->collector() == this, 608 "CMS Thread should refer to this gen"); 609 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 610 611 // Support for parallelizing young gen rescan 612 CMSHeap* heap = CMSHeap::heap(); 613 _young_gen = heap->young_gen(); 614 if (heap->supports_inline_contig_alloc()) { 615 _top_addr = heap->top_addr(); 616 _end_addr = heap->end_addr(); 617 assert(_young_gen != NULL, "no _young_gen"); 618 _eden_chunk_index = 0; 619 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 620 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 621 } 622 623 // Support for parallelizing survivor space rescan 624 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 625 const size_t max_plab_samples = 626 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 627 628 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 629 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 631 _survivor_chunk_capacity = max_plab_samples; 632 for (uint i = 0; i < ParallelGCThreads; i++) { 633 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 634 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 635 assert(cur->end() == 0, "Should be 0"); 636 assert(cur->array() == vec, "Should be vec"); 637 assert(cur->capacity() == max_plab_samples, "Error"); 638 } 639 } 640 641 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 642 _gc_counters = new CollectorCounters("CMS full collection pauses", 1); 643 _cgc_counters = new CollectorCounters("CMS concurrent cycle pauses", 2); 644 _completed_initialization = true; 645 _inter_sweep_timer.start(); // start of time 646 } 647 648 const char* ConcurrentMarkSweepGeneration::name() const { 649 return "concurrent mark-sweep generation"; 650 } 651 void ConcurrentMarkSweepGeneration::update_counters() { 652 if (UsePerfData) { 653 _space_counters->update_all(); 654 _gen_counters->update_all(); 655 } 656 } 657 658 // this is an optimized version of update_counters(). it takes the 659 // used value as a parameter rather than computing it. 660 // 661 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 662 if (UsePerfData) { 663 _space_counters->update_used(used); 664 _space_counters->update_capacity(); 665 _gen_counters->update_all(); 666 } 667 } 668 669 void ConcurrentMarkSweepGeneration::print() const { 670 Generation::print(); 671 cmsSpace()->print(); 672 } 673 674 #ifndef PRODUCT 675 void ConcurrentMarkSweepGeneration::print_statistics() { 676 cmsSpace()->printFLCensus(0); 677 } 678 #endif 679 680 size_t 681 ConcurrentMarkSweepGeneration::contiguous_available() const { 682 // dld proposes an improvement in precision here. If the committed 683 // part of the space ends in a free block we should add that to 684 // uncommitted size in the calculation below. Will make this 685 // change later, staying with the approximation below for the 686 // time being. -- ysr. 687 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 688 } 689 690 size_t 691 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 692 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 693 } 694 695 size_t ConcurrentMarkSweepGeneration::max_available() const { 696 return free() + _virtual_space.uncommitted_size(); 697 } 698 699 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 700 size_t available = max_available(); 701 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 702 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 703 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 704 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 705 return res; 706 } 707 708 // At a promotion failure dump information on block layout in heap 709 // (cms old generation). 710 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 711 Log(gc, promotion) log; 712 if (log.is_trace()) { 713 LogStream ls(log.trace()); 714 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 715 } 716 } 717 718 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 719 // Clear the promotion information. These pointers can be adjusted 720 // along with all the other pointers into the heap but 721 // compaction is expected to be a rare event with 722 // a heap using cms so don't do it without seeing the need. 723 for (uint i = 0; i < ParallelGCThreads; i++) { 724 _par_gc_thread_states[i]->promo.reset(); 725 } 726 } 727 728 void ConcurrentMarkSweepGeneration::compute_new_size() { 729 assert_locked_or_safepoint(Heap_lock); 730 731 // If incremental collection failed, we just want to expand 732 // to the limit. 733 if (incremental_collection_failed()) { 734 clear_incremental_collection_failed(); 735 grow_to_reserved(); 736 return; 737 } 738 739 // The heap has been compacted but not reset yet. 740 // Any metric such as free() or used() will be incorrect. 741 742 CardGeneration::compute_new_size(); 743 744 // Reset again after a possible resizing 745 if (did_compact()) { 746 cmsSpace()->reset_after_compaction(); 747 } 748 } 749 750 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 751 assert_locked_or_safepoint(Heap_lock); 752 753 // If incremental collection failed, we just want to expand 754 // to the limit. 755 if (incremental_collection_failed()) { 756 clear_incremental_collection_failed(); 757 grow_to_reserved(); 758 return; 759 } 760 761 double free_percentage = ((double) free()) / capacity(); 762 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 763 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 764 765 // compute expansion delta needed for reaching desired free percentage 766 if (free_percentage < desired_free_percentage) { 767 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 768 assert(desired_capacity >= capacity(), "invalid expansion size"); 769 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 770 Log(gc) log; 771 if (log.is_trace()) { 772 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 773 log.trace("From compute_new_size: "); 774 log.trace(" Free fraction %f", free_percentage); 775 log.trace(" Desired free fraction %f", desired_free_percentage); 776 log.trace(" Maximum free fraction %f", maximum_free_percentage); 777 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 778 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 779 CMSHeap* heap = CMSHeap::heap(); 780 size_t young_size = heap->young_gen()->capacity(); 781 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 782 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 783 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 784 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 785 } 786 // safe if expansion fails 787 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 788 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 789 } else { 790 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 791 assert(desired_capacity <= capacity(), "invalid expansion size"); 792 size_t shrink_bytes = capacity() - desired_capacity; 793 // Don't shrink unless the delta is greater than the minimum shrink we want 794 if (shrink_bytes >= MinHeapDeltaBytes) { 795 shrink_free_list_by(shrink_bytes); 796 } 797 } 798 } 799 800 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 801 return cmsSpace()->freelistLock(); 802 } 803 804 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 805 CMSSynchronousYieldRequest yr; 806 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 807 return have_lock_and_allocate(size, tlab); 808 } 809 810 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 811 bool tlab /* ignored */) { 812 assert_lock_strong(freelistLock()); 813 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 814 HeapWord* res = cmsSpace()->allocate(adjustedSize); 815 // Allocate the object live (grey) if the background collector has 816 // started marking. This is necessary because the marker may 817 // have passed this address and consequently this object will 818 // not otherwise be greyed and would be incorrectly swept up. 819 // Note that if this object contains references, the writing 820 // of those references will dirty the card containing this object 821 // allowing the object to be blackened (and its references scanned) 822 // either during a preclean phase or at the final checkpoint. 823 if (res != NULL) { 824 // We may block here with an uninitialized object with 825 // its mark-bit or P-bits not yet set. Such objects need 826 // to be safely navigable by block_start(). 827 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 828 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 829 collector()->direct_allocated(res, adjustedSize); 830 _direct_allocated_words += adjustedSize; 831 // allocation counters 832 NOT_PRODUCT( 833 _numObjectsAllocated++; 834 _numWordsAllocated += (int)adjustedSize; 835 ) 836 } 837 return res; 838 } 839 840 // In the case of direct allocation by mutators in a generation that 841 // is being concurrently collected, the object must be allocated 842 // live (grey) if the background collector has started marking. 843 // This is necessary because the marker may 844 // have passed this address and consequently this object will 845 // not otherwise be greyed and would be incorrectly swept up. 846 // Note that if this object contains references, the writing 847 // of those references will dirty the card containing this object 848 // allowing the object to be blackened (and its references scanned) 849 // either during a preclean phase or at the final checkpoint. 850 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 851 assert(_markBitMap.covers(start, size), "Out of bounds"); 852 if (_collectorState >= Marking) { 853 MutexLocker y(_markBitMap.lock(), 854 Mutex::_no_safepoint_check_flag); 855 // [see comments preceding SweepClosure::do_blk() below for details] 856 // 857 // Can the P-bits be deleted now? JJJ 858 // 859 // 1. need to mark the object as live so it isn't collected 860 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 861 // 3. need to mark the end of the object so marking, precleaning or sweeping 862 // can skip over uninitialized or unparsable objects. An allocated 863 // object is considered uninitialized for our purposes as long as 864 // its klass word is NULL. All old gen objects are parsable 865 // as soon as they are initialized.) 866 _markBitMap.mark(start); // object is live 867 _markBitMap.mark(start + 1); // object is potentially uninitialized? 868 _markBitMap.mark(start + size - 1); 869 // mark end of object 870 } 871 // check that oop looks uninitialized 872 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 873 } 874 875 void CMSCollector::promoted(bool par, HeapWord* start, 876 bool is_obj_array, size_t obj_size) { 877 assert(_markBitMap.covers(start), "Out of bounds"); 878 // See comment in direct_allocated() about when objects should 879 // be allocated live. 880 if (_collectorState >= Marking) { 881 // we already hold the marking bit map lock, taken in 882 // the prologue 883 if (par) { 884 _markBitMap.par_mark(start); 885 } else { 886 _markBitMap.mark(start); 887 } 888 // We don't need to mark the object as uninitialized (as 889 // in direct_allocated above) because this is being done with the 890 // world stopped and the object will be initialized by the 891 // time the marking, precleaning or sweeping get to look at it. 892 // But see the code for copying objects into the CMS generation, 893 // where we need to ensure that concurrent readers of the 894 // block offset table are able to safely navigate a block that 895 // is in flux from being free to being allocated (and in 896 // transition while being copied into) and subsequently 897 // becoming a bona-fide object when the copy/promotion is complete. 898 assert(SafepointSynchronize::is_at_safepoint(), 899 "expect promotion only at safepoints"); 900 901 if (_collectorState < Sweeping) { 902 // Mark the appropriate cards in the modUnionTable, so that 903 // this object gets scanned before the sweep. If this is 904 // not done, CMS generation references in the object might 905 // not get marked. 906 // For the case of arrays, which are otherwise precisely 907 // marked, we need to dirty the entire array, not just its head. 908 if (is_obj_array) { 909 // The [par_]mark_range() method expects mr.end() below to 910 // be aligned to the granularity of a bit's representation 911 // in the heap. In the case of the MUT below, that's a 912 // card size. 913 MemRegion mr(start, 914 align_up(start + obj_size, 915 CardTable::card_size /* bytes */)); 916 if (par) { 917 _modUnionTable.par_mark_range(mr); 918 } else { 919 _modUnionTable.mark_range(mr); 920 } 921 } else { // not an obj array; we can just mark the head 922 if (par) { 923 _modUnionTable.par_mark(start); 924 } else { 925 _modUnionTable.mark(start); 926 } 927 } 928 } 929 } 930 } 931 932 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 933 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 934 // allocate, copy and if necessary update promoinfo -- 935 // delegate to underlying space. 936 assert_lock_strong(freelistLock()); 937 938 #ifndef PRODUCT 939 if (CMSHeap::heap()->promotion_should_fail()) { 940 return NULL; 941 } 942 #endif // #ifndef PRODUCT 943 944 oop res = _cmsSpace->promote(obj, obj_size); 945 if (res == NULL) { 946 // expand and retry 947 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 948 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 949 // Since this is the old generation, we don't try to promote 950 // into a more senior generation. 951 res = _cmsSpace->promote(obj, obj_size); 952 } 953 if (res != NULL) { 954 // See comment in allocate() about when objects should 955 // be allocated live. 956 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 957 collector()->promoted(false, // Not parallel 958 (HeapWord*)res, obj->is_objArray(), obj_size); 959 // promotion counters 960 NOT_PRODUCT( 961 _numObjectsPromoted++; 962 _numWordsPromoted += 963 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 964 ) 965 } 966 return res; 967 } 968 969 970 // IMPORTANT: Notes on object size recognition in CMS. 971 // --------------------------------------------------- 972 // A block of storage in the CMS generation is always in 973 // one of three states. A free block (FREE), an allocated 974 // object (OBJECT) whose size() method reports the correct size, 975 // and an intermediate state (TRANSIENT) in which its size cannot 976 // be accurately determined. 977 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 978 // ----------------------------------------------------- 979 // FREE: klass_word & 1 == 1; mark_word holds block size 980 // 981 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 982 // obj->size() computes correct size 983 // 984 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 985 // 986 // STATE IDENTIFICATION: (64 bit+COOPS) 987 // ------------------------------------ 988 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 989 // 990 // OBJECT: klass_word installed; klass_word != 0; 991 // obj->size() computes correct size 992 // 993 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 994 // 995 // 996 // STATE TRANSITION DIAGRAM 997 // 998 // mut / parnew mut / parnew 999 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1000 // ^ | 1001 // |------------------------ DEAD <------------------------------------| 1002 // sweep mut 1003 // 1004 // While a block is in TRANSIENT state its size cannot be determined 1005 // so readers will either need to come back later or stall until 1006 // the size can be determined. Note that for the case of direct 1007 // allocation, P-bits, when available, may be used to determine the 1008 // size of an object that may not yet have been initialized. 1009 1010 // Things to support parallel young-gen collection. 1011 oop 1012 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1013 oop old, markOop m, 1014 size_t word_sz) { 1015 #ifndef PRODUCT 1016 if (CMSHeap::heap()->promotion_should_fail()) { 1017 return NULL; 1018 } 1019 #endif // #ifndef PRODUCT 1020 1021 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1022 PromotionInfo* promoInfo = &ps->promo; 1023 // if we are tracking promotions, then first ensure space for 1024 // promotion (including spooling space for saving header if necessary). 1025 // then allocate and copy, then track promoted info if needed. 1026 // When tracking (see PromotionInfo::track()), the mark word may 1027 // be displaced and in this case restoration of the mark word 1028 // occurs in the (oop_since_save_marks_)iterate phase. 1029 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1030 // Out of space for allocating spooling buffers; 1031 // try expanding and allocating spooling buffers. 1032 if (!expand_and_ensure_spooling_space(promoInfo)) { 1033 return NULL; 1034 } 1035 } 1036 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1037 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1038 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1039 if (obj_ptr == NULL) { 1040 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1041 if (obj_ptr == NULL) { 1042 return NULL; 1043 } 1044 } 1045 oop obj = oop(obj_ptr); 1046 OrderAccess::storestore(); 1047 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1048 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1049 // IMPORTANT: See note on object initialization for CMS above. 1050 // Otherwise, copy the object. Here we must be careful to insert the 1051 // klass pointer last, since this marks the block as an allocated object. 1052 // Except with compressed oops it's the mark word. 1053 HeapWord* old_ptr = (HeapWord*)old; 1054 // Restore the mark word copied above. 1055 obj->set_mark_raw(m); 1056 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1057 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1058 OrderAccess::storestore(); 1059 1060 if (UseCompressedClassPointers) { 1061 // Copy gap missed by (aligned) header size calculation below 1062 obj->set_klass_gap(old->klass_gap()); 1063 } 1064 if (word_sz > (size_t)oopDesc::header_size()) { 1065 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1066 obj_ptr + oopDesc::header_size(), 1067 word_sz - oopDesc::header_size()); 1068 } 1069 1070 // Now we can track the promoted object, if necessary. We take care 1071 // to delay the transition from uninitialized to full object 1072 // (i.e., insertion of klass pointer) until after, so that it 1073 // atomically becomes a promoted object. 1074 if (promoInfo->tracking()) { 1075 promoInfo->track((PromotedObject*)obj, old->klass()); 1076 } 1077 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1078 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1079 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1080 1081 // Finally, install the klass pointer (this should be volatile). 1082 OrderAccess::storestore(); 1083 obj->set_klass(old->klass()); 1084 // We should now be able to calculate the right size for this object 1085 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1086 1087 collector()->promoted(true, // parallel 1088 obj_ptr, old->is_objArray(), word_sz); 1089 1090 NOT_PRODUCT( 1091 Atomic::inc(&_numObjectsPromoted); 1092 Atomic::add(alloc_sz, &_numWordsPromoted); 1093 ) 1094 1095 return obj; 1096 } 1097 1098 void 1099 ConcurrentMarkSweepGeneration:: 1100 par_promote_alloc_done(int thread_num) { 1101 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1102 ps->lab.retire(thread_num); 1103 } 1104 1105 void 1106 ConcurrentMarkSweepGeneration:: 1107 par_oop_since_save_marks_iterate_done(int thread_num) { 1108 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1109 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1110 ps->promo.promoted_oops_iterate(dummy_cl); 1111 1112 // Because card-scanning has been completed, subsequent phases 1113 // (e.g., reference processing) will not need to recognize which 1114 // objects have been promoted during this GC. So, we can now disable 1115 // promotion tracking. 1116 ps->promo.stopTrackingPromotions(); 1117 } 1118 1119 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1120 size_t size, 1121 bool tlab) 1122 { 1123 // We allow a STW collection only if a full 1124 // collection was requested. 1125 return full || should_allocate(size, tlab); // FIX ME !!! 1126 // This and promotion failure handling are connected at the 1127 // hip and should be fixed by untying them. 1128 } 1129 1130 bool CMSCollector::shouldConcurrentCollect() { 1131 LogTarget(Trace, gc) log; 1132 1133 if (_full_gc_requested) { 1134 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1135 return true; 1136 } 1137 1138 FreelistLocker x(this); 1139 // ------------------------------------------------------------------ 1140 // Print out lots of information which affects the initiation of 1141 // a collection. 1142 if (log.is_enabled() && stats().valid()) { 1143 log.print("CMSCollector shouldConcurrentCollect: "); 1144 1145 LogStream out(log); 1146 stats().print_on(&out); 1147 1148 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1149 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1150 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1151 log.print("promotion_rate=%g", stats().promotion_rate()); 1152 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1153 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1154 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1155 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1156 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1157 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1158 } 1159 // ------------------------------------------------------------------ 1160 1161 // If the estimated time to complete a cms collection (cms_duration()) 1162 // is less than the estimated time remaining until the cms generation 1163 // is full, start a collection. 1164 if (!UseCMSInitiatingOccupancyOnly) { 1165 if (stats().valid()) { 1166 if (stats().time_until_cms_start() == 0.0) { 1167 return true; 1168 } 1169 } else { 1170 // We want to conservatively collect somewhat early in order 1171 // to try and "bootstrap" our CMS/promotion statistics; 1172 // this branch will not fire after the first successful CMS 1173 // collection because the stats should then be valid. 1174 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1175 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1176 _cmsGen->occupancy(), _bootstrap_occupancy); 1177 return true; 1178 } 1179 } 1180 } 1181 1182 // Otherwise, we start a collection cycle if 1183 // old gen want a collection cycle started. Each may use 1184 // an appropriate criterion for making this decision. 1185 // XXX We need to make sure that the gen expansion 1186 // criterion dovetails well with this. XXX NEED TO FIX THIS 1187 if (_cmsGen->should_concurrent_collect()) { 1188 log.print("CMS old gen initiated"); 1189 return true; 1190 } 1191 1192 // We start a collection if we believe an incremental collection may fail; 1193 // this is not likely to be productive in practice because it's probably too 1194 // late anyway. 1195 CMSHeap* heap = CMSHeap::heap(); 1196 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1197 log.print("CMSCollector: collect because incremental collection will fail "); 1198 return true; 1199 } 1200 1201 if (MetaspaceGC::should_concurrent_collect()) { 1202 log.print("CMSCollector: collect for metadata allocation "); 1203 return true; 1204 } 1205 1206 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1207 if (CMSTriggerInterval >= 0) { 1208 if (CMSTriggerInterval == 0) { 1209 // Trigger always 1210 return true; 1211 } 1212 1213 // Check the CMS time since begin (we do not check the stats validity 1214 // as we want to be able to trigger the first CMS cycle as well) 1215 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1216 if (stats().valid()) { 1217 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1218 stats().cms_time_since_begin()); 1219 } else { 1220 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1221 } 1222 return true; 1223 } 1224 } 1225 1226 return false; 1227 } 1228 1229 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1230 1231 // Clear _expansion_cause fields of constituent generations 1232 void CMSCollector::clear_expansion_cause() { 1233 _cmsGen->clear_expansion_cause(); 1234 } 1235 1236 // We should be conservative in starting a collection cycle. To 1237 // start too eagerly runs the risk of collecting too often in the 1238 // extreme. To collect too rarely falls back on full collections, 1239 // which works, even if not optimum in terms of concurrent work. 1240 // As a work around for too eagerly collecting, use the flag 1241 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1242 // giving the user an easily understandable way of controlling the 1243 // collections. 1244 // We want to start a new collection cycle if any of the following 1245 // conditions hold: 1246 // . our current occupancy exceeds the configured initiating occupancy 1247 // for this generation, or 1248 // . we recently needed to expand this space and have not, since that 1249 // expansion, done a collection of this generation, or 1250 // . the underlying space believes that it may be a good idea to initiate 1251 // a concurrent collection (this may be based on criteria such as the 1252 // following: the space uses linear allocation and linear allocation is 1253 // going to fail, or there is believed to be excessive fragmentation in 1254 // the generation, etc... or ... 1255 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1256 // the case of the old generation; see CR 6543076): 1257 // we may be approaching a point at which allocation requests may fail because 1258 // we will be out of sufficient free space given allocation rate estimates.] 1259 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1260 1261 assert_lock_strong(freelistLock()); 1262 if (occupancy() > initiating_occupancy()) { 1263 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1264 short_name(), occupancy(), initiating_occupancy()); 1265 return true; 1266 } 1267 if (UseCMSInitiatingOccupancyOnly) { 1268 return false; 1269 } 1270 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1271 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1272 return true; 1273 } 1274 return false; 1275 } 1276 1277 void ConcurrentMarkSweepGeneration::collect(bool full, 1278 bool clear_all_soft_refs, 1279 size_t size, 1280 bool tlab) 1281 { 1282 collector()->collect(full, clear_all_soft_refs, size, tlab); 1283 } 1284 1285 void CMSCollector::collect(bool full, 1286 bool clear_all_soft_refs, 1287 size_t size, 1288 bool tlab) 1289 { 1290 // The following "if" branch is present for defensive reasons. 1291 // In the current uses of this interface, it can be replaced with: 1292 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1293 // But I am not placing that assert here to allow future 1294 // generality in invoking this interface. 1295 if (GCLocker::is_active()) { 1296 // A consistency test for GCLocker 1297 assert(GCLocker::needs_gc(), "Should have been set already"); 1298 // Skip this foreground collection, instead 1299 // expanding the heap if necessary. 1300 // Need the free list locks for the call to free() in compute_new_size() 1301 compute_new_size(); 1302 return; 1303 } 1304 acquire_control_and_collect(full, clear_all_soft_refs); 1305 } 1306 1307 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1308 CMSHeap* heap = CMSHeap::heap(); 1309 unsigned int gc_count = heap->total_full_collections(); 1310 if (gc_count == full_gc_count) { 1311 MutexLocker y(CGC_lock, Mutex::_no_safepoint_check_flag); 1312 _full_gc_requested = true; 1313 _full_gc_cause = cause; 1314 CGC_lock->notify(); // nudge CMS thread 1315 } else { 1316 assert(gc_count > full_gc_count, "Error: causal loop"); 1317 } 1318 } 1319 1320 bool CMSCollector::is_external_interruption() { 1321 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1322 return GCCause::is_user_requested_gc(cause) || 1323 GCCause::is_serviceability_requested_gc(cause); 1324 } 1325 1326 void CMSCollector::report_concurrent_mode_interruption() { 1327 if (is_external_interruption()) { 1328 log_debug(gc)("Concurrent mode interrupted"); 1329 } else { 1330 log_debug(gc)("Concurrent mode failure"); 1331 _gc_tracer_cm->report_concurrent_mode_failure(); 1332 } 1333 } 1334 1335 1336 // The foreground and background collectors need to coordinate in order 1337 // to make sure that they do not mutually interfere with CMS collections. 1338 // When a background collection is active, 1339 // the foreground collector may need to take over (preempt) and 1340 // synchronously complete an ongoing collection. Depending on the 1341 // frequency of the background collections and the heap usage 1342 // of the application, this preemption can be seldom or frequent. 1343 // There are only certain 1344 // points in the background collection that the "collection-baton" 1345 // can be passed to the foreground collector. 1346 // 1347 // The foreground collector will wait for the baton before 1348 // starting any part of the collection. The foreground collector 1349 // will only wait at one location. 1350 // 1351 // The background collector will yield the baton before starting a new 1352 // phase of the collection (e.g., before initial marking, marking from roots, 1353 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1354 // of the loop which switches the phases. The background collector does some 1355 // of the phases (initial mark, final re-mark) with the world stopped. 1356 // Because of locking involved in stopping the world, 1357 // the foreground collector should not block waiting for the background 1358 // collector when it is doing a stop-the-world phase. The background 1359 // collector will yield the baton at an additional point just before 1360 // it enters a stop-the-world phase. Once the world is stopped, the 1361 // background collector checks the phase of the collection. If the 1362 // phase has not changed, it proceeds with the collection. If the 1363 // phase has changed, it skips that phase of the collection. See 1364 // the comments on the use of the Heap_lock in collect_in_background(). 1365 // 1366 // Variable used in baton passing. 1367 // _foregroundGCIsActive - Set to true by the foreground collector when 1368 // it wants the baton. The foreground clears it when it has finished 1369 // the collection. 1370 // _foregroundGCShouldWait - Set to true by the background collector 1371 // when it is running. The foreground collector waits while 1372 // _foregroundGCShouldWait is true. 1373 // CGC_lock - monitor used to protect access to the above variables 1374 // and to notify the foreground and background collectors. 1375 // _collectorState - current state of the CMS collection. 1376 // 1377 // The foreground collector 1378 // acquires the CGC_lock 1379 // sets _foregroundGCIsActive 1380 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1381 // various locks acquired in preparation for the collection 1382 // are released so as not to block the background collector 1383 // that is in the midst of a collection 1384 // proceeds with the collection 1385 // clears _foregroundGCIsActive 1386 // returns 1387 // 1388 // The background collector in a loop iterating on the phases of the 1389 // collection 1390 // acquires the CGC_lock 1391 // sets _foregroundGCShouldWait 1392 // if _foregroundGCIsActive is set 1393 // clears _foregroundGCShouldWait, notifies _CGC_lock 1394 // waits on _CGC_lock for _foregroundGCIsActive to become false 1395 // and exits the loop. 1396 // otherwise 1397 // proceed with that phase of the collection 1398 // if the phase is a stop-the-world phase, 1399 // yield the baton once more just before enqueueing 1400 // the stop-world CMS operation (executed by the VM thread). 1401 // returns after all phases of the collection are done 1402 // 1403 1404 void CMSCollector::acquire_control_and_collect(bool full, 1405 bool clear_all_soft_refs) { 1406 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1407 assert(!Thread::current()->is_ConcurrentGC_thread(), 1408 "shouldn't try to acquire control from self!"); 1409 1410 // Start the protocol for acquiring control of the 1411 // collection from the background collector (aka CMS thread). 1412 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1413 "VM thread should have CMS token"); 1414 // Remember the possibly interrupted state of an ongoing 1415 // concurrent collection 1416 CollectorState first_state = _collectorState; 1417 1418 // Signal to a possibly ongoing concurrent collection that 1419 // we want to do a foreground collection. 1420 _foregroundGCIsActive = true; 1421 1422 // release locks and wait for a notify from the background collector 1423 // releasing the locks in only necessary for phases which 1424 // do yields to improve the granularity of the collection. 1425 assert_lock_strong(bitMapLock()); 1426 // We need to lock the Free list lock for the space that we are 1427 // currently collecting. 1428 assert(haveFreelistLocks(), "Must be holding free list locks"); 1429 bitMapLock()->unlock(); 1430 releaseFreelistLocks(); 1431 { 1432 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1433 if (_foregroundGCShouldWait) { 1434 // We are going to be waiting for action for the CMS thread; 1435 // it had better not be gone (for instance at shutdown)! 1436 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1437 "CMS thread must be running"); 1438 // Wait here until the background collector gives us the go-ahead 1439 ConcurrentMarkSweepThread::clear_CMS_flag( 1440 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1441 // Get a possibly blocked CMS thread going: 1442 // Note that we set _foregroundGCIsActive true above, 1443 // without protection of the CGC_lock. 1444 CGC_lock->notify(); 1445 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1446 "Possible deadlock"); 1447 while (_foregroundGCShouldWait) { 1448 // wait for notification 1449 CGC_lock->wait_without_safepoint_check(); 1450 // Possibility of delay/starvation here, since CMS token does 1451 // not know to give priority to VM thread? Actually, i think 1452 // there wouldn't be any delay/starvation, but the proof of 1453 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1454 } 1455 ConcurrentMarkSweepThread::set_CMS_flag( 1456 ConcurrentMarkSweepThread::CMS_vm_has_token); 1457 } 1458 } 1459 // The CMS_token is already held. Get back the other locks. 1460 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1461 "VM thread should have CMS token"); 1462 getFreelistLocks(); 1463 bitMapLock()->lock_without_safepoint_check(); 1464 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1465 p2i(Thread::current()), first_state); 1466 log_debug(gc, state)(" gets control with state %d", _collectorState); 1467 1468 // Inform cms gen if this was due to partial collection failing. 1469 // The CMS gen may use this fact to determine its expansion policy. 1470 CMSHeap* heap = CMSHeap::heap(); 1471 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1472 assert(!_cmsGen->incremental_collection_failed(), 1473 "Should have been noticed, reacted to and cleared"); 1474 _cmsGen->set_incremental_collection_failed(); 1475 } 1476 1477 if (first_state > Idling) { 1478 report_concurrent_mode_interruption(); 1479 } 1480 1481 set_did_compact(true); 1482 1483 // If the collection is being acquired from the background 1484 // collector, there may be references on the discovered 1485 // references lists. Abandon those references, since some 1486 // of them may have become unreachable after concurrent 1487 // discovery; the STW compacting collector will redo discovery 1488 // more precisely, without being subject to floating garbage. 1489 // Leaving otherwise unreachable references in the discovered 1490 // lists would require special handling. 1491 ref_processor()->disable_discovery(); 1492 ref_processor()->abandon_partial_discovery(); 1493 ref_processor()->verify_no_references_recorded(); 1494 1495 if (first_state > Idling) { 1496 save_heap_summary(); 1497 } 1498 1499 do_compaction_work(clear_all_soft_refs); 1500 1501 // Has the GC time limit been exceeded? 1502 size_t max_eden_size = _young_gen->max_eden_size(); 1503 GCCause::Cause gc_cause = heap->gc_cause(); 1504 size_policy()->check_gc_overhead_limit(_young_gen->eden()->used(), 1505 _cmsGen->max_capacity(), 1506 max_eden_size, 1507 full, 1508 gc_cause, 1509 heap->soft_ref_policy()); 1510 1511 // Reset the expansion cause, now that we just completed 1512 // a collection cycle. 1513 clear_expansion_cause(); 1514 _foregroundGCIsActive = false; 1515 return; 1516 } 1517 1518 // Resize the tenured generation 1519 // after obtaining the free list locks for the 1520 // two generations. 1521 void CMSCollector::compute_new_size() { 1522 assert_locked_or_safepoint(Heap_lock); 1523 FreelistLocker z(this); 1524 MetaspaceGC::compute_new_size(); 1525 _cmsGen->compute_new_size_free_list(); 1526 } 1527 1528 // A work method used by the foreground collector to do 1529 // a mark-sweep-compact. 1530 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1531 CMSHeap* heap = CMSHeap::heap(); 1532 1533 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1534 gc_timer->register_gc_start(); 1535 1536 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1537 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1538 1539 heap->pre_full_gc_dump(gc_timer); 1540 1541 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1542 1543 // Temporarily widen the span of the weak reference processing to 1544 // the entire heap. 1545 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1546 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1547 // Temporarily, clear the "is_alive_non_header" field of the 1548 // reference processor. 1549 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1550 // Temporarily make reference _processing_ single threaded (non-MT). 1551 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1552 // Temporarily make refs discovery atomic 1553 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1554 // Temporarily make reference _discovery_ single threaded (non-MT) 1555 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1556 1557 ref_processor()->set_enqueuing_is_done(false); 1558 ref_processor()->enable_discovery(); 1559 ref_processor()->setup_policy(clear_all_soft_refs); 1560 // If an asynchronous collection finishes, the _modUnionTable is 1561 // all clear. If we are assuming the collection from an asynchronous 1562 // collection, clear the _modUnionTable. 1563 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1564 "_modUnionTable should be clear if the baton was not passed"); 1565 _modUnionTable.clear_all(); 1566 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1567 "mod union for klasses should be clear if the baton was passed"); 1568 _ct->cld_rem_set()->clear_mod_union(); 1569 1570 1571 // We must adjust the allocation statistics being maintained 1572 // in the free list space. We do so by reading and clearing 1573 // the sweep timer and updating the block flux rate estimates below. 1574 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1575 if (_inter_sweep_timer.is_active()) { 1576 _inter_sweep_timer.stop(); 1577 // Note that we do not use this sample to update the _inter_sweep_estimate. 1578 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1579 _inter_sweep_estimate.padded_average(), 1580 _intra_sweep_estimate.padded_average()); 1581 } 1582 1583 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1584 #ifdef ASSERT 1585 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1586 size_t free_size = cms_space->free(); 1587 assert(free_size == 1588 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1589 * HeapWordSize, 1590 "All the free space should be compacted into one chunk at top"); 1591 assert(cms_space->dictionary()->total_chunk_size( 1592 debug_only(cms_space->freelistLock())) == 0 || 1593 cms_space->totalSizeInIndexedFreeLists() == 0, 1594 "All the free space should be in a single chunk"); 1595 size_t num = cms_space->totalCount(); 1596 assert((free_size == 0 && num == 0) || 1597 (free_size > 0 && (num == 1 || num == 2)), 1598 "There should be at most 2 free chunks after compaction"); 1599 #endif // ASSERT 1600 _collectorState = Resetting; 1601 assert(_restart_addr == NULL, 1602 "Should have been NULL'd before baton was passed"); 1603 reset_stw(); 1604 _cmsGen->reset_after_compaction(); 1605 _concurrent_cycles_since_last_unload = 0; 1606 1607 // Clear any data recorded in the PLAB chunk arrays. 1608 if (_survivor_plab_array != NULL) { 1609 reset_survivor_plab_arrays(); 1610 } 1611 1612 // Adjust the per-size allocation stats for the next epoch. 1613 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1614 // Restart the "inter sweep timer" for the next epoch. 1615 _inter_sweep_timer.reset(); 1616 _inter_sweep_timer.start(); 1617 1618 // No longer a need to do a concurrent collection for Metaspace. 1619 MetaspaceGC::set_should_concurrent_collect(false); 1620 1621 heap->post_full_gc_dump(gc_timer); 1622 1623 gc_timer->register_gc_end(); 1624 1625 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1626 1627 // For a mark-sweep-compact, compute_new_size() will be called 1628 // in the heap's do_collection() method. 1629 } 1630 1631 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1632 Log(gc, heap) log; 1633 if (!log.is_trace()) { 1634 return; 1635 } 1636 1637 ContiguousSpace* eden_space = _young_gen->eden(); 1638 ContiguousSpace* from_space = _young_gen->from(); 1639 ContiguousSpace* to_space = _young_gen->to(); 1640 // Eden 1641 if (_eden_chunk_array != NULL) { 1642 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1643 p2i(eden_space->bottom()), p2i(eden_space->top()), 1644 p2i(eden_space->end()), eden_space->capacity()); 1645 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1646 _eden_chunk_index, _eden_chunk_capacity); 1647 for (size_t i = 0; i < _eden_chunk_index; i++) { 1648 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1649 } 1650 } 1651 // Survivor 1652 if (_survivor_chunk_array != NULL) { 1653 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1654 p2i(from_space->bottom()), p2i(from_space->top()), 1655 p2i(from_space->end()), from_space->capacity()); 1656 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1657 _survivor_chunk_index, _survivor_chunk_capacity); 1658 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1659 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1660 } 1661 } 1662 } 1663 1664 void CMSCollector::getFreelistLocks() const { 1665 // Get locks for all free lists in all generations that this 1666 // collector is responsible for 1667 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1668 } 1669 1670 void CMSCollector::releaseFreelistLocks() const { 1671 // Release locks for all free lists in all generations that this 1672 // collector is responsible for 1673 _cmsGen->freelistLock()->unlock(); 1674 } 1675 1676 bool CMSCollector::haveFreelistLocks() const { 1677 // Check locks for all free lists in all generations that this 1678 // collector is responsible for 1679 assert_lock_strong(_cmsGen->freelistLock()); 1680 PRODUCT_ONLY(ShouldNotReachHere()); 1681 return true; 1682 } 1683 1684 // A utility class that is used by the CMS collector to 1685 // temporarily "release" the foreground collector from its 1686 // usual obligation to wait for the background collector to 1687 // complete an ongoing phase before proceeding. 1688 class ReleaseForegroundGC: public StackObj { 1689 private: 1690 CMSCollector* _c; 1691 public: 1692 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1693 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1694 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1695 // allow a potentially blocked foreground collector to proceed 1696 _c->_foregroundGCShouldWait = false; 1697 if (_c->_foregroundGCIsActive) { 1698 CGC_lock->notify(); 1699 } 1700 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1701 "Possible deadlock"); 1702 } 1703 1704 ~ReleaseForegroundGC() { 1705 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1706 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1707 _c->_foregroundGCShouldWait = true; 1708 } 1709 }; 1710 1711 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1712 assert(Thread::current()->is_ConcurrentGC_thread(), 1713 "A CMS asynchronous collection is only allowed on a CMS thread."); 1714 1715 CMSHeap* heap = CMSHeap::heap(); 1716 { 1717 MutexLocker hl(Heap_lock, Mutex::_no_safepoint_check_flag); 1718 FreelistLocker fll(this); 1719 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1720 if (_foregroundGCIsActive) { 1721 // The foreground collector is. Skip this 1722 // background collection. 1723 assert(!_foregroundGCShouldWait, "Should be clear"); 1724 return; 1725 } else { 1726 assert(_collectorState == Idling, "Should be idling before start."); 1727 _collectorState = InitialMarking; 1728 register_gc_start(cause); 1729 // Reset the expansion cause, now that we are about to begin 1730 // a new cycle. 1731 clear_expansion_cause(); 1732 1733 // Clear the MetaspaceGC flag since a concurrent collection 1734 // is starting but also clear it after the collection. 1735 MetaspaceGC::set_should_concurrent_collect(false); 1736 } 1737 // Decide if we want to enable class unloading as part of the 1738 // ensuing concurrent GC cycle. 1739 update_should_unload_classes(); 1740 _full_gc_requested = false; // acks all outstanding full gc requests 1741 _full_gc_cause = GCCause::_no_gc; 1742 // Signal that we are about to start a collection 1743 heap->increment_total_full_collections(); // ... starting a collection cycle 1744 _collection_count_start = heap->total_full_collections(); 1745 } 1746 1747 size_t prev_used = _cmsGen->used(); 1748 1749 // The change of the collection state is normally done at this level; 1750 // the exceptions are phases that are executed while the world is 1751 // stopped. For those phases the change of state is done while the 1752 // world is stopped. For baton passing purposes this allows the 1753 // background collector to finish the phase and change state atomically. 1754 // The foreground collector cannot wait on a phase that is done 1755 // while the world is stopped because the foreground collector already 1756 // has the world stopped and would deadlock. 1757 while (_collectorState != Idling) { 1758 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1759 p2i(Thread::current()), _collectorState); 1760 // The foreground collector 1761 // holds the Heap_lock throughout its collection. 1762 // holds the CMS token (but not the lock) 1763 // except while it is waiting for the background collector to yield. 1764 // 1765 // The foreground collector should be blocked (not for long) 1766 // if the background collector is about to start a phase 1767 // executed with world stopped. If the background 1768 // collector has already started such a phase, the 1769 // foreground collector is blocked waiting for the 1770 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1771 // are executed in the VM thread. 1772 // 1773 // The locking order is 1774 // PendingListLock (PLL) -- if applicable (FinalMarking) 1775 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1776 // CMS token (claimed in 1777 // stop_world_and_do() --> 1778 // safepoint_synchronize() --> 1779 // CMSThread::synchronize()) 1780 1781 { 1782 // Check if the FG collector wants us to yield. 1783 CMSTokenSync x(true); // is cms thread 1784 if (waitForForegroundGC()) { 1785 // We yielded to a foreground GC, nothing more to be 1786 // done this round. 1787 assert(_foregroundGCShouldWait == false, "We set it to false in " 1788 "waitForForegroundGC()"); 1789 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1790 p2i(Thread::current()), _collectorState); 1791 return; 1792 } else { 1793 // The background collector can run but check to see if the 1794 // foreground collector has done a collection while the 1795 // background collector was waiting to get the CGC_lock 1796 // above. If yes, break so that _foregroundGCShouldWait 1797 // is cleared before returning. 1798 if (_collectorState == Idling) { 1799 break; 1800 } 1801 } 1802 } 1803 1804 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1805 "should be waiting"); 1806 1807 switch (_collectorState) { 1808 case InitialMarking: 1809 { 1810 ReleaseForegroundGC x(this); 1811 stats().record_cms_begin(); 1812 VM_CMS_Initial_Mark initial_mark_op(this); 1813 VMThread::execute(&initial_mark_op); 1814 } 1815 // The collector state may be any legal state at this point 1816 // since the background collector may have yielded to the 1817 // foreground collector. 1818 break; 1819 case Marking: 1820 // initial marking in checkpointRootsInitialWork has been completed 1821 if (markFromRoots()) { // we were successful 1822 assert(_collectorState == Precleaning, "Collector state should " 1823 "have changed"); 1824 } else { 1825 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1826 } 1827 break; 1828 case Precleaning: 1829 // marking from roots in markFromRoots has been completed 1830 preclean(); 1831 assert(_collectorState == AbortablePreclean || 1832 _collectorState == FinalMarking, 1833 "Collector state should have changed"); 1834 break; 1835 case AbortablePreclean: 1836 abortable_preclean(); 1837 assert(_collectorState == FinalMarking, "Collector state should " 1838 "have changed"); 1839 break; 1840 case FinalMarking: 1841 { 1842 ReleaseForegroundGC x(this); 1843 1844 VM_CMS_Final_Remark final_remark_op(this); 1845 VMThread::execute(&final_remark_op); 1846 } 1847 assert(_foregroundGCShouldWait, "block post-condition"); 1848 break; 1849 case Sweeping: 1850 // final marking in checkpointRootsFinal has been completed 1851 sweep(); 1852 assert(_collectorState == Resizing, "Collector state change " 1853 "to Resizing must be done under the free_list_lock"); 1854 1855 case Resizing: { 1856 // Sweeping has been completed... 1857 // At this point the background collection has completed. 1858 // Don't move the call to compute_new_size() down 1859 // into code that might be executed if the background 1860 // collection was preempted. 1861 { 1862 ReleaseForegroundGC x(this); // unblock FG collection 1863 MutexLocker y(Heap_lock, Mutex::_no_safepoint_check_flag); 1864 CMSTokenSync z(true); // not strictly needed. 1865 if (_collectorState == Resizing) { 1866 compute_new_size(); 1867 save_heap_summary(); 1868 _collectorState = Resetting; 1869 } else { 1870 assert(_collectorState == Idling, "The state should only change" 1871 " because the foreground collector has finished the collection"); 1872 } 1873 } 1874 break; 1875 } 1876 case Resetting: 1877 // CMS heap resizing has been completed 1878 reset_concurrent(); 1879 assert(_collectorState == Idling, "Collector state should " 1880 "have changed"); 1881 1882 MetaspaceGC::set_should_concurrent_collect(false); 1883 1884 stats().record_cms_end(); 1885 // Don't move the concurrent_phases_end() and compute_new_size() 1886 // calls to here because a preempted background collection 1887 // has it's state set to "Resetting". 1888 break; 1889 case Idling: 1890 default: 1891 ShouldNotReachHere(); 1892 break; 1893 } 1894 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1895 p2i(Thread::current()), _collectorState); 1896 assert(_foregroundGCShouldWait, "block post-condition"); 1897 } 1898 1899 // Should this be in gc_epilogue? 1900 heap->counters()->update_counters(); 1901 1902 { 1903 // Clear _foregroundGCShouldWait and, in the event that the 1904 // foreground collector is waiting, notify it, before 1905 // returning. 1906 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1907 _foregroundGCShouldWait = false; 1908 if (_foregroundGCIsActive) { 1909 CGC_lock->notify(); 1910 } 1911 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1912 "Possible deadlock"); 1913 } 1914 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1915 p2i(Thread::current()), _collectorState); 1916 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1917 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1918 } 1919 1920 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1921 _cms_start_registered = true; 1922 _gc_timer_cm->register_gc_start(); 1923 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1924 } 1925 1926 void CMSCollector::register_gc_end() { 1927 if (_cms_start_registered) { 1928 report_heap_summary(GCWhen::AfterGC); 1929 1930 _gc_timer_cm->register_gc_end(); 1931 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1932 _cms_start_registered = false; 1933 } 1934 } 1935 1936 void CMSCollector::save_heap_summary() { 1937 CMSHeap* heap = CMSHeap::heap(); 1938 _last_heap_summary = heap->create_heap_summary(); 1939 _last_metaspace_summary = heap->create_metaspace_summary(); 1940 } 1941 1942 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1943 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1944 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1945 } 1946 1947 bool CMSCollector::waitForForegroundGC() { 1948 bool res = false; 1949 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1950 "CMS thread should have CMS token"); 1951 // Block the foreground collector until the 1952 // background collectors decides whether to 1953 // yield. 1954 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1955 _foregroundGCShouldWait = true; 1956 if (_foregroundGCIsActive) { 1957 // The background collector yields to the 1958 // foreground collector and returns a value 1959 // indicating that it has yielded. The foreground 1960 // collector can proceed. 1961 res = true; 1962 _foregroundGCShouldWait = false; 1963 ConcurrentMarkSweepThread::clear_CMS_flag( 1964 ConcurrentMarkSweepThread::CMS_cms_has_token); 1965 ConcurrentMarkSweepThread::set_CMS_flag( 1966 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1967 // Get a possibly blocked foreground thread going 1968 CGC_lock->notify(); 1969 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1970 p2i(Thread::current()), _collectorState); 1971 while (_foregroundGCIsActive) { 1972 CGC_lock->wait_without_safepoint_check(); 1973 } 1974 ConcurrentMarkSweepThread::set_CMS_flag( 1975 ConcurrentMarkSweepThread::CMS_cms_has_token); 1976 ConcurrentMarkSweepThread::clear_CMS_flag( 1977 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1978 } 1979 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1980 p2i(Thread::current()), _collectorState); 1981 return res; 1982 } 1983 1984 // Because of the need to lock the free lists and other structures in 1985 // the collector, common to all the generations that the collector is 1986 // collecting, we need the gc_prologues of individual CMS generations 1987 // delegate to their collector. It may have been simpler had the 1988 // current infrastructure allowed one to call a prologue on a 1989 // collector. In the absence of that we have the generation's 1990 // prologue delegate to the collector, which delegates back 1991 // some "local" work to a worker method in the individual generations 1992 // that it's responsible for collecting, while itself doing any 1993 // work common to all generations it's responsible for. A similar 1994 // comment applies to the gc_epilogue()'s. 1995 // The role of the variable _between_prologue_and_epilogue is to 1996 // enforce the invocation protocol. 1997 void CMSCollector::gc_prologue(bool full) { 1998 // Call gc_prologue_work() for the CMSGen 1999 // we are responsible for. 2000 2001 // The following locking discipline assumes that we are only called 2002 // when the world is stopped. 2003 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2004 2005 // The CMSCollector prologue must call the gc_prologues for the 2006 // "generations" that it's responsible 2007 // for. 2008 2009 assert( Thread::current()->is_VM_thread() 2010 || ( CMSScavengeBeforeRemark 2011 && Thread::current()->is_ConcurrentGC_thread()), 2012 "Incorrect thread type for prologue execution"); 2013 2014 if (_between_prologue_and_epilogue) { 2015 // We have already been invoked; this is a gc_prologue delegation 2016 // from yet another CMS generation that we are responsible for, just 2017 // ignore it since all relevant work has already been done. 2018 return; 2019 } 2020 2021 // set a bit saying prologue has been called; cleared in epilogue 2022 _between_prologue_and_epilogue = true; 2023 // Claim locks for common data structures, then call gc_prologue_work() 2024 // for each CMSGen. 2025 2026 getFreelistLocks(); // gets free list locks on constituent spaces 2027 bitMapLock()->lock_without_safepoint_check(); 2028 2029 // Should call gc_prologue_work() for all cms gens we are responsible for 2030 bool duringMarking = _collectorState >= Marking 2031 && _collectorState < Sweeping; 2032 2033 // The young collections clear the modified oops state, which tells if 2034 // there are any modified oops in the class. The remark phase also needs 2035 // that information. Tell the young collection to save the union of all 2036 // modified klasses. 2037 if (duringMarking) { 2038 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2039 } 2040 2041 bool registerClosure = duringMarking; 2042 2043 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2044 2045 if (!full) { 2046 stats().record_gc0_begin(); 2047 } 2048 } 2049 2050 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2051 2052 _capacity_at_prologue = capacity(); 2053 _used_at_prologue = used(); 2054 2055 // We enable promotion tracking so that card-scanning can recognize 2056 // which objects have been promoted during this GC and skip them. 2057 for (uint i = 0; i < ParallelGCThreads; i++) { 2058 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2059 } 2060 2061 // Delegate to CMScollector which knows how to coordinate between 2062 // this and any other CMS generations that it is responsible for 2063 // collecting. 2064 collector()->gc_prologue(full); 2065 } 2066 2067 // This is a "private" interface for use by this generation's CMSCollector. 2068 // Not to be called directly by any other entity (for instance, 2069 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2070 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2071 bool registerClosure, ModUnionClosure* modUnionClosure) { 2072 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2073 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2074 "Should be NULL"); 2075 if (registerClosure) { 2076 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2077 } 2078 cmsSpace()->gc_prologue(); 2079 // Clear stat counters 2080 NOT_PRODUCT( 2081 assert(_numObjectsPromoted == 0, "check"); 2082 assert(_numWordsPromoted == 0, "check"); 2083 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2084 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2085 _numObjectsAllocated = 0; 2086 _numWordsAllocated = 0; 2087 ) 2088 } 2089 2090 void CMSCollector::gc_epilogue(bool full) { 2091 // The following locking discipline assumes that we are only called 2092 // when the world is stopped. 2093 assert(SafepointSynchronize::is_at_safepoint(), 2094 "world is stopped assumption"); 2095 2096 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2097 // if linear allocation blocks need to be appropriately marked to allow the 2098 // the blocks to be parsable. We also check here whether we need to nudge the 2099 // CMS collector thread to start a new cycle (if it's not already active). 2100 assert( Thread::current()->is_VM_thread() 2101 || ( CMSScavengeBeforeRemark 2102 && Thread::current()->is_ConcurrentGC_thread()), 2103 "Incorrect thread type for epilogue execution"); 2104 2105 if (!_between_prologue_and_epilogue) { 2106 // We have already been invoked; this is a gc_epilogue delegation 2107 // from yet another CMS generation that we are responsible for, just 2108 // ignore it since all relevant work has already been done. 2109 return; 2110 } 2111 assert(haveFreelistLocks(), "must have freelist locks"); 2112 assert_lock_strong(bitMapLock()); 2113 2114 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2115 2116 _cmsGen->gc_epilogue_work(full); 2117 2118 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2119 // in case sampling was not already enabled, enable it 2120 _start_sampling = true; 2121 } 2122 // reset _eden_chunk_array so sampling starts afresh 2123 _eden_chunk_index = 0; 2124 2125 size_t cms_used = _cmsGen->cmsSpace()->used(); 2126 2127 // update performance counters - this uses a special version of 2128 // update_counters() that allows the utilization to be passed as a 2129 // parameter, avoiding multiple calls to used(). 2130 // 2131 _cmsGen->update_counters(cms_used); 2132 2133 bitMapLock()->unlock(); 2134 releaseFreelistLocks(); 2135 2136 if (!CleanChunkPoolAsync) { 2137 Chunk::clean_chunk_pool(); 2138 } 2139 2140 set_did_compact(false); 2141 _between_prologue_and_epilogue = false; // ready for next cycle 2142 } 2143 2144 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2145 collector()->gc_epilogue(full); 2146 2147 // When using ParNew, promotion tracking should have already been 2148 // disabled. However, the prologue (which enables promotion 2149 // tracking) and epilogue are called irrespective of the type of 2150 // GC. So they will also be called before and after Full GCs, during 2151 // which promotion tracking will not be explicitly disabled. So, 2152 // it's safer to also disable it here too (to be symmetric with 2153 // enabling it in the prologue). 2154 for (uint i = 0; i < ParallelGCThreads; i++) { 2155 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2156 } 2157 } 2158 2159 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2160 assert(!incremental_collection_failed(), "Should have been cleared"); 2161 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2162 cmsSpace()->gc_epilogue(); 2163 // Print stat counters 2164 NOT_PRODUCT( 2165 assert(_numObjectsAllocated == 0, "check"); 2166 assert(_numWordsAllocated == 0, "check"); 2167 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2168 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2169 _numObjectsPromoted = 0; 2170 _numWordsPromoted = 0; 2171 ) 2172 2173 // Call down the chain in contiguous_available needs the freelistLock 2174 // so print this out before releasing the freeListLock. 2175 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2176 } 2177 2178 #ifndef PRODUCT 2179 bool CMSCollector::have_cms_token() { 2180 Thread* thr = Thread::current(); 2181 if (thr->is_VM_thread()) { 2182 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2183 } else if (thr->is_ConcurrentGC_thread()) { 2184 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2185 } else if (thr->is_GC_task_thread()) { 2186 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2187 ParGCRareEvent_lock->owned_by_self(); 2188 } 2189 return false; 2190 } 2191 2192 // Check reachability of the given heap address in CMS generation, 2193 // treating all other generations as roots. 2194 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2195 // We could "guarantee" below, rather than assert, but I'll 2196 // leave these as "asserts" so that an adventurous debugger 2197 // could try this in the product build provided some subset of 2198 // the conditions were met, provided they were interested in the 2199 // results and knew that the computation below wouldn't interfere 2200 // with other concurrent computations mutating the structures 2201 // being read or written. 2202 assert(SafepointSynchronize::is_at_safepoint(), 2203 "Else mutations in object graph will make answer suspect"); 2204 assert(have_cms_token(), "Should hold cms token"); 2205 assert(haveFreelistLocks(), "must hold free list locks"); 2206 assert_lock_strong(bitMapLock()); 2207 2208 // Clear the marking bit map array before starting, but, just 2209 // for kicks, first report if the given address is already marked 2210 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2211 _markBitMap.isMarked(addr) ? "" : " not"); 2212 2213 if (verify_after_remark()) { 2214 MutexLocker x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2215 bool result = verification_mark_bm()->isMarked(addr); 2216 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2217 result ? "IS" : "is NOT"); 2218 return result; 2219 } else { 2220 tty->print_cr("Could not compute result"); 2221 return false; 2222 } 2223 } 2224 #endif 2225 2226 void 2227 CMSCollector::print_on_error(outputStream* st) { 2228 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2229 if (collector != NULL) { 2230 CMSBitMap* bitmap = &collector->_markBitMap; 2231 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2232 bitmap->print_on_error(st, " Bits: "); 2233 2234 st->cr(); 2235 2236 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2237 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2238 mut_bitmap->print_on_error(st, " Bits: "); 2239 } 2240 } 2241 2242 //////////////////////////////////////////////////////// 2243 // CMS Verification Support 2244 //////////////////////////////////////////////////////// 2245 // Following the remark phase, the following invariant 2246 // should hold -- each object in the CMS heap which is 2247 // marked in markBitMap() should be marked in the verification_mark_bm(). 2248 2249 class VerifyMarkedClosure: public BitMapClosure { 2250 CMSBitMap* _marks; 2251 bool _failed; 2252 2253 public: 2254 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2255 2256 bool do_bit(size_t offset) { 2257 HeapWord* addr = _marks->offsetToHeapWord(offset); 2258 if (!_marks->isMarked(addr)) { 2259 Log(gc, verify) log; 2260 ResourceMark rm; 2261 LogStream ls(log.error()); 2262 oop(addr)->print_on(&ls); 2263 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2264 _failed = true; 2265 } 2266 return true; 2267 } 2268 2269 bool failed() { return _failed; } 2270 }; 2271 2272 bool CMSCollector::verify_after_remark() { 2273 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2274 MutexLocker ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2275 static bool init = false; 2276 2277 assert(SafepointSynchronize::is_at_safepoint(), 2278 "Else mutations in object graph will make answer suspect"); 2279 assert(have_cms_token(), 2280 "Else there may be mutual interference in use of " 2281 " verification data structures"); 2282 assert(_collectorState > Marking && _collectorState <= Sweeping, 2283 "Else marking info checked here may be obsolete"); 2284 assert(haveFreelistLocks(), "must hold free list locks"); 2285 assert_lock_strong(bitMapLock()); 2286 2287 2288 // Allocate marking bit map if not already allocated 2289 if (!init) { // first time 2290 if (!verification_mark_bm()->allocate(_span)) { 2291 return false; 2292 } 2293 init = true; 2294 } 2295 2296 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2297 2298 // Turn off refs discovery -- so we will be tracing through refs. 2299 // This is as intended, because by this time 2300 // GC must already have cleared any refs that need to be cleared, 2301 // and traced those that need to be marked; moreover, 2302 // the marking done here is not going to interfere in any 2303 // way with the marking information used by GC. 2304 NoRefDiscovery no_discovery(ref_processor()); 2305 2306 #if COMPILER2_OR_JVMCI 2307 DerivedPointerTableDeactivate dpt_deact; 2308 #endif 2309 2310 // Clear any marks from a previous round 2311 verification_mark_bm()->clear_all(); 2312 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2313 verify_work_stacks_empty(); 2314 2315 CMSHeap* heap = CMSHeap::heap(); 2316 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2317 // Update the saved marks which may affect the root scans. 2318 heap->save_marks(); 2319 2320 if (CMSRemarkVerifyVariant == 1) { 2321 // In this first variant of verification, we complete 2322 // all marking, then check if the new marks-vector is 2323 // a subset of the CMS marks-vector. 2324 verify_after_remark_work_1(); 2325 } else { 2326 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2327 // In this second variant of verification, we flag an error 2328 // (i.e. an object reachable in the new marks-vector not reachable 2329 // in the CMS marks-vector) immediately, also indicating the 2330 // identify of an object (A) that references the unmarked object (B) -- 2331 // presumably, a mutation to A failed to be picked up by preclean/remark? 2332 verify_after_remark_work_2(); 2333 } 2334 2335 return true; 2336 } 2337 2338 void CMSCollector::verify_after_remark_work_1() { 2339 ResourceMark rm; 2340 HandleMark hm; 2341 CMSHeap* heap = CMSHeap::heap(); 2342 2343 // Get a clear set of claim bits for the roots processing to work with. 2344 ClassLoaderDataGraph::clear_claimed_marks(); 2345 2346 // Mark from roots one level into CMS 2347 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2348 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2349 2350 { 2351 StrongRootsScope srs(1); 2352 2353 heap->cms_process_roots(&srs, 2354 true, // young gen as roots 2355 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2356 should_unload_classes(), 2357 ¬Older, 2358 NULL); 2359 } 2360 2361 // Now mark from the roots 2362 MarkFromRootsClosure markFromRootsClosure(this, _span, 2363 verification_mark_bm(), verification_mark_stack(), 2364 false /* don't yield */, true /* verifying */); 2365 assert(_restart_addr == NULL, "Expected pre-condition"); 2366 verification_mark_bm()->iterate(&markFromRootsClosure); 2367 while (_restart_addr != NULL) { 2368 // Deal with stack overflow: by restarting at the indicated 2369 // address. 2370 HeapWord* ra = _restart_addr; 2371 markFromRootsClosure.reset(ra); 2372 _restart_addr = NULL; 2373 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2374 } 2375 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2376 verify_work_stacks_empty(); 2377 2378 // Marking completed -- now verify that each bit marked in 2379 // verification_mark_bm() is also marked in markBitMap(); flag all 2380 // errors by printing corresponding objects. 2381 VerifyMarkedClosure vcl(markBitMap()); 2382 verification_mark_bm()->iterate(&vcl); 2383 if (vcl.failed()) { 2384 Log(gc, verify) log; 2385 log.error("Failed marking verification after remark"); 2386 ResourceMark rm; 2387 LogStream ls(log.error()); 2388 heap->print_on(&ls); 2389 fatal("CMS: failed marking verification after remark"); 2390 } 2391 } 2392 2393 class VerifyCLDOopsCLDClosure : public CLDClosure { 2394 class VerifyCLDOopsClosure : public OopClosure { 2395 CMSBitMap* _bitmap; 2396 public: 2397 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2398 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2399 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2400 } _oop_closure; 2401 public: 2402 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2403 void do_cld(ClassLoaderData* cld) { 2404 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2405 } 2406 }; 2407 2408 void CMSCollector::verify_after_remark_work_2() { 2409 ResourceMark rm; 2410 HandleMark hm; 2411 CMSHeap* heap = CMSHeap::heap(); 2412 2413 // Get a clear set of claim bits for the roots processing to work with. 2414 ClassLoaderDataGraph::clear_claimed_marks(); 2415 2416 // Mark from roots one level into CMS 2417 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2418 markBitMap()); 2419 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2420 2421 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2422 2423 { 2424 StrongRootsScope srs(1); 2425 2426 heap->cms_process_roots(&srs, 2427 true, // young gen as roots 2428 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2429 should_unload_classes(), 2430 ¬Older, 2431 &cld_closure); 2432 } 2433 2434 // Now mark from the roots 2435 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2436 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2437 assert(_restart_addr == NULL, "Expected pre-condition"); 2438 verification_mark_bm()->iterate(&markFromRootsClosure); 2439 while (_restart_addr != NULL) { 2440 // Deal with stack overflow: by restarting at the indicated 2441 // address. 2442 HeapWord* ra = _restart_addr; 2443 markFromRootsClosure.reset(ra); 2444 _restart_addr = NULL; 2445 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2446 } 2447 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2448 verify_work_stacks_empty(); 2449 2450 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2451 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2452 2453 // Marking completed -- now verify that each bit marked in 2454 // verification_mark_bm() is also marked in markBitMap(); flag all 2455 // errors by printing corresponding objects. 2456 VerifyMarkedClosure vcl(markBitMap()); 2457 verification_mark_bm()->iterate(&vcl); 2458 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2459 } 2460 2461 void ConcurrentMarkSweepGeneration::save_marks() { 2462 // delegate to CMS space 2463 cmsSpace()->save_marks(); 2464 } 2465 2466 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2467 return cmsSpace()->no_allocs_since_save_marks(); 2468 } 2469 2470 void 2471 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2472 if (freelistLock()->owned_by_self()) { 2473 Generation::oop_iterate(cl); 2474 } else { 2475 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2476 Generation::oop_iterate(cl); 2477 } 2478 } 2479 2480 void 2481 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2482 if (freelistLock()->owned_by_self()) { 2483 Generation::object_iterate(cl); 2484 } else { 2485 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2486 Generation::object_iterate(cl); 2487 } 2488 } 2489 2490 void 2491 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2492 if (freelistLock()->owned_by_self()) { 2493 Generation::safe_object_iterate(cl); 2494 } else { 2495 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2496 Generation::safe_object_iterate(cl); 2497 } 2498 } 2499 2500 void 2501 ConcurrentMarkSweepGeneration::post_compact() { 2502 } 2503 2504 void 2505 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2506 // Fix the linear allocation blocks to look like free blocks. 2507 2508 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2509 // are not called when the heap is verified during universe initialization and 2510 // at vm shutdown. 2511 if (freelistLock()->owned_by_self()) { 2512 cmsSpace()->prepare_for_verify(); 2513 } else { 2514 MutexLocker fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2515 cmsSpace()->prepare_for_verify(); 2516 } 2517 } 2518 2519 void 2520 ConcurrentMarkSweepGeneration::verify() { 2521 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2522 // are not called when the heap is verified during universe initialization and 2523 // at vm shutdown. 2524 if (freelistLock()->owned_by_self()) { 2525 cmsSpace()->verify(); 2526 } else { 2527 MutexLocker fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2528 cmsSpace()->verify(); 2529 } 2530 } 2531 2532 void CMSCollector::verify() { 2533 _cmsGen->verify(); 2534 } 2535 2536 #ifndef PRODUCT 2537 bool CMSCollector::overflow_list_is_empty() const { 2538 assert(_num_par_pushes >= 0, "Inconsistency"); 2539 if (_overflow_list == NULL) { 2540 assert(_num_par_pushes == 0, "Inconsistency"); 2541 } 2542 return _overflow_list == NULL; 2543 } 2544 2545 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2546 // merely consolidate assertion checks that appear to occur together frequently. 2547 void CMSCollector::verify_work_stacks_empty() const { 2548 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2549 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2550 } 2551 2552 void CMSCollector::verify_overflow_empty() const { 2553 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2554 assert(no_preserved_marks(), "No preserved marks"); 2555 } 2556 #endif // PRODUCT 2557 2558 // Decide if we want to enable class unloading as part of the 2559 // ensuing concurrent GC cycle. We will collect and 2560 // unload classes if it's the case that: 2561 // (a) class unloading is enabled at the command line, and 2562 // (b) old gen is getting really full 2563 // NOTE: Provided there is no change in the state of the heap between 2564 // calls to this method, it should have idempotent results. Moreover, 2565 // its results should be monotonically increasing (i.e. going from 0 to 1, 2566 // but not 1 to 0) between successive calls between which the heap was 2567 // not collected. For the implementation below, it must thus rely on 2568 // the property that concurrent_cycles_since_last_unload() 2569 // will not decrease unless a collection cycle happened and that 2570 // _cmsGen->is_too_full() are 2571 // themselves also monotonic in that sense. See check_monotonicity() 2572 // below. 2573 void CMSCollector::update_should_unload_classes() { 2574 _should_unload_classes = false; 2575 if (CMSClassUnloadingEnabled) { 2576 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2577 CMSClassUnloadingMaxInterval) 2578 || _cmsGen->is_too_full(); 2579 } 2580 } 2581 2582 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2583 bool res = should_concurrent_collect(); 2584 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2585 return res; 2586 } 2587 2588 void CMSCollector::setup_cms_unloading_and_verification_state() { 2589 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2590 || VerifyBeforeExit; 2591 const int rso = GenCollectedHeap::SO_AllCodeCache; 2592 2593 // We set the proper root for this CMS cycle here. 2594 if (should_unload_classes()) { // Should unload classes this cycle 2595 remove_root_scanning_option(rso); // Shrink the root set appropriately 2596 set_verifying(should_verify); // Set verification state for this cycle 2597 return; // Nothing else needs to be done at this time 2598 } 2599 2600 // Not unloading classes this cycle 2601 assert(!should_unload_classes(), "Inconsistency!"); 2602 2603 // If we are not unloading classes then add SO_AllCodeCache to root 2604 // scanning options. 2605 add_root_scanning_option(rso); 2606 2607 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2608 set_verifying(true); 2609 } else if (verifying() && !should_verify) { 2610 // We were verifying, but some verification flags got disabled. 2611 set_verifying(false); 2612 // Exclude symbols, strings and code cache elements from root scanning to 2613 // reduce IM and RM pauses. 2614 remove_root_scanning_option(rso); 2615 } 2616 } 2617 2618 2619 #ifndef PRODUCT 2620 HeapWord* CMSCollector::block_start(const void* p) const { 2621 const HeapWord* addr = (HeapWord*)p; 2622 if (_span.contains(p)) { 2623 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2624 return _cmsGen->cmsSpace()->block_start(p); 2625 } 2626 } 2627 return NULL; 2628 } 2629 #endif 2630 2631 HeapWord* 2632 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2633 bool tlab, 2634 bool parallel) { 2635 CMSSynchronousYieldRequest yr; 2636 assert(!tlab, "Can't deal with TLAB allocation"); 2637 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2638 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2639 if (GCExpandToAllocateDelayMillis > 0) { 2640 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2641 } 2642 return have_lock_and_allocate(word_size, tlab); 2643 } 2644 2645 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2646 size_t bytes, 2647 size_t expand_bytes, 2648 CMSExpansionCause::Cause cause) 2649 { 2650 2651 bool success = expand(bytes, expand_bytes); 2652 2653 // remember why we expanded; this information is used 2654 // by shouldConcurrentCollect() when making decisions on whether to start 2655 // a new CMS cycle. 2656 if (success) { 2657 set_expansion_cause(cause); 2658 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2659 } 2660 } 2661 2662 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2663 HeapWord* res = NULL; 2664 MutexLocker x(ParGCRareEvent_lock); 2665 while (true) { 2666 // Expansion by some other thread might make alloc OK now: 2667 res = ps->lab.alloc(word_sz); 2668 if (res != NULL) return res; 2669 // If there's not enough expansion space available, give up. 2670 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2671 return NULL; 2672 } 2673 // Otherwise, we try expansion. 2674 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2675 // Now go around the loop and try alloc again; 2676 // A competing par_promote might beat us to the expansion space, 2677 // so we may go around the loop again if promotion fails again. 2678 if (GCExpandToAllocateDelayMillis > 0) { 2679 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2680 } 2681 } 2682 } 2683 2684 2685 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2686 PromotionInfo* promo) { 2687 MutexLocker x(ParGCRareEvent_lock); 2688 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2689 while (true) { 2690 // Expansion by some other thread might make alloc OK now: 2691 if (promo->ensure_spooling_space()) { 2692 assert(promo->has_spooling_space(), 2693 "Post-condition of successful ensure_spooling_space()"); 2694 return true; 2695 } 2696 // If there's not enough expansion space available, give up. 2697 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2698 return false; 2699 } 2700 // Otherwise, we try expansion. 2701 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2702 // Now go around the loop and try alloc again; 2703 // A competing allocation might beat us to the expansion space, 2704 // so we may go around the loop again if allocation fails again. 2705 if (GCExpandToAllocateDelayMillis > 0) { 2706 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2707 } 2708 } 2709 } 2710 2711 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2712 // Only shrink if a compaction was done so that all the free space 2713 // in the generation is in a contiguous block at the end. 2714 if (did_compact()) { 2715 CardGeneration::shrink(bytes); 2716 } 2717 } 2718 2719 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2720 assert_locked_or_safepoint(Heap_lock); 2721 } 2722 2723 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2724 assert_locked_or_safepoint(Heap_lock); 2725 assert_lock_strong(freelistLock()); 2726 log_trace(gc)("Shrinking of CMS not yet implemented"); 2727 return; 2728 } 2729 2730 2731 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2732 // phases. 2733 class CMSPhaseAccounting: public StackObj { 2734 public: 2735 CMSPhaseAccounting(CMSCollector *collector, 2736 const char *title); 2737 ~CMSPhaseAccounting(); 2738 2739 private: 2740 CMSCollector *_collector; 2741 const char *_title; 2742 GCTraceConcTime(Info, gc) _trace_time; 2743 2744 public: 2745 // Not MT-safe; so do not pass around these StackObj's 2746 // where they may be accessed by other threads. 2747 double wallclock_millis() { 2748 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2749 } 2750 }; 2751 2752 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2753 const char *title) : 2754 _collector(collector), _title(title), _trace_time(title) { 2755 2756 _collector->resetYields(); 2757 _collector->resetTimer(); 2758 _collector->startTimer(); 2759 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2760 } 2761 2762 CMSPhaseAccounting::~CMSPhaseAccounting() { 2763 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2764 _collector->stopTimer(); 2765 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_millis(_collector->timerTicks())); 2766 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2767 } 2768 2769 // CMS work 2770 2771 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2772 class CMSParMarkTask : public AbstractGangTask { 2773 protected: 2774 CMSCollector* _collector; 2775 uint _n_workers; 2776 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2777 AbstractGangTask(name), 2778 _collector(collector), 2779 _n_workers(n_workers) {} 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 MutexLocker 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 MutexLocker 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 MutexLocker 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 MutexLocker 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 assert(haveFreelistLocks(), "must have free list locks"); 4197 assert_lock_strong(bitMapLock()); 4198 4199 // We might assume that we need not fill TLAB's when 4200 // CMSScavengeBeforeRemark is set, because we may have just done 4201 // a scavenge which would have filled all TLAB's -- and besides 4202 // Eden would be empty. This however may not always be the case -- 4203 // for instance although we asked for a scavenge, it may not have 4204 // happened because of a JNI critical section. We probably need 4205 // a policy for deciding whether we can in that case wait until 4206 // the critical section releases and then do the remark following 4207 // the scavenge, and skip it here. In the absence of that policy, 4208 // or of an indication of whether the scavenge did indeed occur, 4209 // we cannot rely on TLAB's having been filled and must do 4210 // so here just in case a scavenge did not happen. 4211 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4212 // Update the saved marks which may affect the root scans. 4213 heap->save_marks(); 4214 4215 print_eden_and_survivor_chunk_arrays(); 4216 4217 { 4218 #if COMPILER2_OR_JVMCI 4219 DerivedPointerTableDeactivate dpt_deact; 4220 #endif 4221 4222 // Note on the role of the mod union table: 4223 // Since the marker in "markFromRoots" marks concurrently with 4224 // mutators, it is possible for some reachable objects not to have been 4225 // scanned. For instance, an only reference to an object A was 4226 // placed in object B after the marker scanned B. Unless B is rescanned, 4227 // A would be collected. Such updates to references in marked objects 4228 // are detected via the mod union table which is the set of all cards 4229 // dirtied since the first checkpoint in this GC cycle and prior to 4230 // the most recent young generation GC, minus those cleaned up by the 4231 // concurrent precleaning. 4232 if (CMSParallelRemarkEnabled) { 4233 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4234 do_remark_parallel(); 4235 } else { 4236 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4237 do_remark_non_parallel(); 4238 } 4239 } 4240 verify_work_stacks_empty(); 4241 verify_overflow_empty(); 4242 4243 { 4244 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4245 refProcessingWork(); 4246 } 4247 verify_work_stacks_empty(); 4248 verify_overflow_empty(); 4249 4250 if (should_unload_classes()) { 4251 heap->prune_scavengable_nmethods(); 4252 } 4253 JvmtiExport::gc_epilogue(); 4254 4255 // If we encountered any (marking stack / work queue) overflow 4256 // events during the current CMS cycle, take appropriate 4257 // remedial measures, where possible, so as to try and avoid 4258 // recurrence of that condition. 4259 assert(_markStack.isEmpty(), "No grey objects"); 4260 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4261 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4262 if (ser_ovflw > 0) { 4263 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4264 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4265 _markStack.expand(); 4266 _ser_pmc_remark_ovflw = 0; 4267 _ser_pmc_preclean_ovflw = 0; 4268 _ser_kac_preclean_ovflw = 0; 4269 _ser_kac_ovflw = 0; 4270 } 4271 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4272 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4273 _par_pmc_remark_ovflw, _par_kac_ovflw); 4274 _par_pmc_remark_ovflw = 0; 4275 _par_kac_ovflw = 0; 4276 } 4277 if (_markStack._hit_limit > 0) { 4278 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4279 _markStack._hit_limit); 4280 } 4281 if (_markStack._failed_double > 0) { 4282 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4283 _markStack._failed_double, _markStack.capacity()); 4284 } 4285 _markStack._hit_limit = 0; 4286 _markStack._failed_double = 0; 4287 4288 if ((VerifyAfterGC || VerifyDuringGC) && 4289 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4290 verify_after_remark(); 4291 } 4292 4293 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4294 4295 // Change under the freelistLocks. 4296 _collectorState = Sweeping; 4297 // Call isAllClear() under bitMapLock 4298 assert(_modUnionTable.isAllClear(), 4299 "Should be clear by end of the final marking"); 4300 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4301 "Should be clear by end of the final marking"); 4302 } 4303 4304 void CMSParInitialMarkTask::work(uint worker_id) { 4305 elapsedTimer _timer; 4306 ResourceMark rm; 4307 HandleMark hm; 4308 4309 // ---------- scan from roots -------------- 4310 _timer.start(); 4311 CMSHeap* heap = CMSHeap::heap(); 4312 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4313 4314 // ---------- young gen roots -------------- 4315 { 4316 work_on_young_gen_roots(&par_mri_cl); 4317 _timer.stop(); 4318 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4319 } 4320 4321 // ---------- remaining roots -------------- 4322 _timer.reset(); 4323 _timer.start(); 4324 4325 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4326 4327 heap->cms_process_roots(_strong_roots_scope, 4328 false, // yg was scanned above 4329 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4330 _collector->should_unload_classes(), 4331 &par_mri_cl, 4332 &cld_closure); 4333 4334 assert(_collector->should_unload_classes() 4335 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4336 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4337 _timer.stop(); 4338 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4339 } 4340 4341 // Parallel remark task 4342 class CMSParRemarkTask: public CMSParMarkTask { 4343 CompactibleFreeListSpace* _cms_space; 4344 4345 // The per-thread work queues, available here for stealing. 4346 OopTaskQueueSet* _task_queues; 4347 TaskTerminator _term; 4348 StrongRootsScope* _strong_roots_scope; 4349 4350 public: 4351 // A value of 0 passed to n_workers will cause the number of 4352 // workers to be taken from the active workers in the work gang. 4353 CMSParRemarkTask(CMSCollector* collector, 4354 CompactibleFreeListSpace* cms_space, 4355 uint n_workers, WorkGang* workers, 4356 OopTaskQueueSet* task_queues, 4357 StrongRootsScope* strong_roots_scope): 4358 CMSParMarkTask("Rescan roots and grey objects in parallel", 4359 collector, n_workers), 4360 _cms_space(cms_space), 4361 _task_queues(task_queues), 4362 _term(n_workers, task_queues), 4363 _strong_roots_scope(strong_roots_scope) { } 4364 4365 OopTaskQueueSet* task_queues() { return _task_queues; } 4366 4367 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4368 4369 ParallelTaskTerminator* terminator() { return _term.terminator(); } 4370 uint n_workers() { return _n_workers; } 4371 4372 void work(uint worker_id); 4373 4374 private: 4375 // ... of dirty cards in old space 4376 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4377 ParMarkRefsIntoAndScanClosure* cl); 4378 4379 // ... work stealing for the above 4380 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4381 }; 4382 4383 class RemarkCLDClosure : public CLDClosure { 4384 CLDToOopClosure _cm_closure; 4385 public: 4386 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4387 void do_cld(ClassLoaderData* cld) { 4388 // Check if we have modified any oops in the CLD during the concurrent marking. 4389 if (cld->has_accumulated_modified_oops()) { 4390 cld->clear_accumulated_modified_oops(); 4391 4392 // We could have transfered the current modified marks to the accumulated marks, 4393 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4394 } else if (cld->has_modified_oops()) { 4395 // Don't clear anything, this info is needed by the next young collection. 4396 } else { 4397 // No modified oops in the ClassLoaderData. 4398 return; 4399 } 4400 4401 // The klass has modified fields, need to scan the klass. 4402 _cm_closure.do_cld(cld); 4403 } 4404 }; 4405 4406 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4407 ParNewGeneration* young_gen = _collector->_young_gen; 4408 ContiguousSpace* eden_space = young_gen->eden(); 4409 ContiguousSpace* from_space = young_gen->from(); 4410 ContiguousSpace* to_space = young_gen->to(); 4411 4412 HeapWord** eca = _collector->_eden_chunk_array; 4413 size_t ect = _collector->_eden_chunk_index; 4414 HeapWord** sca = _collector->_survivor_chunk_array; 4415 size_t sct = _collector->_survivor_chunk_index; 4416 4417 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4418 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4419 4420 do_young_space_rescan(cl, to_space, NULL, 0); 4421 do_young_space_rescan(cl, from_space, sca, sct); 4422 do_young_space_rescan(cl, eden_space, eca, ect); 4423 } 4424 4425 // work_queue(i) is passed to the closure 4426 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4427 // also is passed to do_dirty_card_rescan_tasks() and to 4428 // do_work_steal() to select the i-th task_queue. 4429 4430 void CMSParRemarkTask::work(uint worker_id) { 4431 elapsedTimer _timer; 4432 ResourceMark rm; 4433 HandleMark hm; 4434 4435 // ---------- rescan from roots -------------- 4436 _timer.start(); 4437 CMSHeap* heap = CMSHeap::heap(); 4438 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4439 _collector->_span, _collector->ref_processor(), 4440 &(_collector->_markBitMap), 4441 work_queue(worker_id)); 4442 4443 // Rescan young gen roots first since these are likely 4444 // coarsely partitioned and may, on that account, constitute 4445 // the critical path; thus, it's best to start off that 4446 // work first. 4447 // ---------- young gen roots -------------- 4448 { 4449 work_on_young_gen_roots(&par_mrias_cl); 4450 _timer.stop(); 4451 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4452 } 4453 4454 // ---------- remaining roots -------------- 4455 _timer.reset(); 4456 _timer.start(); 4457 heap->cms_process_roots(_strong_roots_scope, 4458 false, // yg was scanned above 4459 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4460 _collector->should_unload_classes(), 4461 &par_mrias_cl, 4462 NULL); // The dirty klasses will be handled below 4463 4464 assert(_collector->should_unload_classes() 4465 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4466 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4467 _timer.stop(); 4468 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4469 4470 // ---------- unhandled CLD scanning ---------- 4471 if (worker_id == 0) { // Single threaded at the moment. 4472 _timer.reset(); 4473 _timer.start(); 4474 4475 // Scan all new class loader data objects and new dependencies that were 4476 // introduced during concurrent marking. 4477 ResourceMark rm; 4478 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4479 for (int i = 0; i < array->length(); i++) { 4480 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4481 } 4482 4483 // We don't need to keep track of new CLDs anymore. 4484 ClassLoaderDataGraph::remember_new_clds(false); 4485 4486 _timer.stop(); 4487 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4488 } 4489 4490 // We might have added oops to ClassLoaderData::_handles during the 4491 // concurrent marking phase. These oops do not always point to newly allocated objects 4492 // that are guaranteed to be kept alive. Hence, 4493 // we do have to revisit the _handles block during the remark phase. 4494 4495 // ---------- dirty CLD scanning ---------- 4496 if (worker_id == 0) { // Single threaded at the moment. 4497 _timer.reset(); 4498 _timer.start(); 4499 4500 // Scan all classes that was dirtied during the concurrent marking phase. 4501 RemarkCLDClosure remark_closure(&par_mrias_cl); 4502 ClassLoaderDataGraph::cld_do(&remark_closure); 4503 4504 _timer.stop(); 4505 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4506 } 4507 4508 4509 // ---------- rescan dirty cards ------------ 4510 _timer.reset(); 4511 _timer.start(); 4512 4513 // Do the rescan tasks for each of the two spaces 4514 // (cms_space) in turn. 4515 // "worker_id" is passed to select the task_queue for "worker_id" 4516 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4517 _timer.stop(); 4518 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4519 4520 // ---------- steal work from other threads ... 4521 // ---------- ... and drain overflow list. 4522 _timer.reset(); 4523 _timer.start(); 4524 do_work_steal(worker_id, &par_mrias_cl); 4525 _timer.stop(); 4526 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4527 } 4528 4529 void 4530 CMSParMarkTask::do_young_space_rescan( 4531 OopsInGenClosure* cl, ContiguousSpace* space, 4532 HeapWord** chunk_array, size_t chunk_top) { 4533 // Until all tasks completed: 4534 // . claim an unclaimed task 4535 // . compute region boundaries corresponding to task claimed 4536 // using chunk_array 4537 // . par_oop_iterate(cl) over that region 4538 4539 ResourceMark rm; 4540 HandleMark hm; 4541 4542 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4543 4544 uint nth_task = 0; 4545 uint n_tasks = pst->n_tasks(); 4546 4547 if (n_tasks > 0) { 4548 assert(pst->valid(), "Uninitialized use?"); 4549 HeapWord *start, *end; 4550 while (pst->try_claim_task(/* reference */ nth_task)) { 4551 // We claimed task # nth_task; compute its boundaries. 4552 if (chunk_top == 0) { // no samples were taken 4553 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4554 start = space->bottom(); 4555 end = space->top(); 4556 } else if (nth_task == 0) { 4557 start = space->bottom(); 4558 end = chunk_array[nth_task]; 4559 } else if (nth_task < (uint)chunk_top) { 4560 assert(nth_task >= 1, "Control point invariant"); 4561 start = chunk_array[nth_task - 1]; 4562 end = chunk_array[nth_task]; 4563 } else { 4564 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4565 start = chunk_array[chunk_top - 1]; 4566 end = space->top(); 4567 } 4568 MemRegion mr(start, end); 4569 // Verify that mr is in space 4570 assert(mr.is_empty() || space->used_region().contains(mr), 4571 "Should be in space"); 4572 // Verify that "start" is an object boundary 4573 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4574 "Should be an oop"); 4575 space->par_oop_iterate(mr, cl); 4576 } 4577 pst->all_tasks_completed(); 4578 } 4579 } 4580 4581 void 4582 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4583 CompactibleFreeListSpace* sp, int i, 4584 ParMarkRefsIntoAndScanClosure* cl) { 4585 // Until all tasks completed: 4586 // . claim an unclaimed task 4587 // . compute region boundaries corresponding to task claimed 4588 // . transfer dirty bits ct->mut for that region 4589 // . apply rescanclosure to dirty mut bits for that region 4590 4591 ResourceMark rm; 4592 HandleMark hm; 4593 4594 OopTaskQueue* work_q = work_queue(i); 4595 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4596 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4597 // CAUTION: This closure has state that persists across calls to 4598 // the work method dirty_range_iterate_clear() in that it has 4599 // embedded in it a (subtype of) UpwardsObjectClosure. The 4600 // use of that state in the embedded UpwardsObjectClosure instance 4601 // assumes that the cards are always iterated (even if in parallel 4602 // by several threads) in monotonically increasing order per each 4603 // thread. This is true of the implementation below which picks 4604 // card ranges (chunks) in monotonically increasing order globally 4605 // and, a-fortiori, in monotonically increasing order per thread 4606 // (the latter order being a subsequence of the former). 4607 // If the work code below is ever reorganized into a more chaotic 4608 // work-partitioning form than the current "sequential tasks" 4609 // paradigm, the use of that persistent state will have to be 4610 // revisited and modified appropriately. See also related 4611 // bug 4756801 work on which should examine this code to make 4612 // sure that the changes there do not run counter to the 4613 // assumptions made here and necessary for correctness and 4614 // efficiency. Note also that this code might yield inefficient 4615 // behavior in the case of very large objects that span one or 4616 // more work chunks. Such objects would potentially be scanned 4617 // several times redundantly. Work on 4756801 should try and 4618 // address that performance anomaly if at all possible. XXX 4619 MemRegion full_span = _collector->_span; 4620 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4621 MarkFromDirtyCardsClosure 4622 greyRescanClosure(_collector, full_span, // entire span of interest 4623 sp, bm, work_q, cl); 4624 4625 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4626 assert(pst->valid(), "Uninitialized use?"); 4627 uint nth_task = 0; 4628 const int alignment = CardTable::card_size * BitsPerWord; 4629 MemRegion span = sp->used_region(); 4630 HeapWord* start_addr = span.start(); 4631 HeapWord* end_addr = align_up(span.end(), alignment); 4632 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4633 assert(is_aligned(start_addr, alignment), "Check alignment"); 4634 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4635 4636 while (pst->try_claim_task(/* reference */ nth_task)) { 4637 // Having claimed the nth_task, compute corresponding mem-region, 4638 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4639 // The alignment restriction ensures that we do not need any 4640 // synchronization with other gang-workers while setting or 4641 // clearing bits in thus chunk of the MUT. 4642 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4643 start_addr + (nth_task+1)*chunk_size); 4644 // The last chunk's end might be way beyond end of the 4645 // used region. In that case pull back appropriately. 4646 if (this_span.end() > end_addr) { 4647 this_span.set_end(end_addr); 4648 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4649 } 4650 // Iterate over the dirty cards covering this chunk, marking them 4651 // precleaned, and setting the corresponding bits in the mod union 4652 // table. Since we have been careful to partition at Card and MUT-word 4653 // boundaries no synchronization is needed between parallel threads. 4654 _collector->_ct->dirty_card_iterate(this_span, 4655 &modUnionClosure); 4656 4657 // Having transferred these marks into the modUnionTable, 4658 // rescan the marked objects on the dirty cards in the modUnionTable. 4659 // Even if this is at a synchronous collection, the initial marking 4660 // may have been done during an asynchronous collection so there 4661 // may be dirty bits in the mod-union table. 4662 _collector->_modUnionTable.dirty_range_iterate_clear( 4663 this_span, &greyRescanClosure); 4664 _collector->_modUnionTable.verifyNoOneBitsInRange( 4665 this_span.start(), 4666 this_span.end()); 4667 } 4668 pst->all_tasks_completed(); // declare that i am done 4669 } 4670 4671 // . see if we can share work_queues with ParNew? XXX 4672 void 4673 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4674 OopTaskQueue* work_q = work_queue(i); 4675 NOT_PRODUCT(int num_steals = 0;) 4676 oop obj_to_scan; 4677 CMSBitMap* bm = &(_collector->_markBitMap); 4678 4679 while (true) { 4680 // Completely finish any left over work from (an) earlier round(s) 4681 cl->trim_queue(0); 4682 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4683 (size_t)ParGCDesiredObjsFromOverflowList); 4684 // Now check if there's any work in the overflow list 4685 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4686 // only affects the number of attempts made to get work from the 4687 // overflow list and does not affect the number of workers. Just 4688 // pass ParallelGCThreads so this behavior is unchanged. 4689 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4690 work_q, 4691 ParallelGCThreads)) { 4692 // found something in global overflow list; 4693 // not yet ready to go stealing work from others. 4694 // We'd like to assert(work_q->size() != 0, ...) 4695 // because we just took work from the overflow list, 4696 // but of course we can't since all of that could have 4697 // been already stolen from us. 4698 // "He giveth and He taketh away." 4699 continue; 4700 } 4701 // Verify that we have no work before we resort to stealing 4702 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4703 // Try to steal from other queues that have work 4704 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4705 NOT_PRODUCT(num_steals++;) 4706 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4707 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4708 // Do scanning work 4709 obj_to_scan->oop_iterate(cl); 4710 // Loop around, finish this work, and try to steal some more 4711 } else if (terminator()->offer_termination()) { 4712 break; // nirvana from the infinite cycle 4713 } 4714 } 4715 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4716 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4717 "Else our work is not yet done"); 4718 } 4719 4720 // Record object boundaries in _eden_chunk_array by sampling the eden 4721 // top in the slow-path eden object allocation code path and record 4722 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4723 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4724 // sampling in sample_eden() that activates during the part of the 4725 // preclean phase. 4726 void CMSCollector::sample_eden_chunk() { 4727 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4728 if (_eden_chunk_lock->try_lock()) { 4729 // Record a sample. This is the critical section. The contents 4730 // of the _eden_chunk_array have to be non-decreasing in the 4731 // address order. 4732 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4733 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4734 "Unexpected state of Eden"); 4735 if (_eden_chunk_index == 0 || 4736 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4737 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4738 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4739 _eden_chunk_index++; // commit sample 4740 } 4741 _eden_chunk_lock->unlock(); 4742 } 4743 } 4744 } 4745 4746 // Return a thread-local PLAB recording array, as appropriate. 4747 void* CMSCollector::get_data_recorder(int thr_num) { 4748 if (_survivor_plab_array != NULL && 4749 (CMSPLABRecordAlways || 4750 (_collectorState > Marking && _collectorState < FinalMarking))) { 4751 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4752 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4753 ca->reset(); // clear it so that fresh data is recorded 4754 return (void*) ca; 4755 } else { 4756 return NULL; 4757 } 4758 } 4759 4760 // Reset all the thread-local PLAB recording arrays 4761 void CMSCollector::reset_survivor_plab_arrays() { 4762 for (uint i = 0; i < ParallelGCThreads; i++) { 4763 _survivor_plab_array[i].reset(); 4764 } 4765 } 4766 4767 // Merge the per-thread plab arrays into the global survivor chunk 4768 // array which will provide the partitioning of the survivor space 4769 // for CMS initial scan and rescan. 4770 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4771 int no_of_gc_threads) { 4772 assert(_survivor_plab_array != NULL, "Error"); 4773 assert(_survivor_chunk_array != NULL, "Error"); 4774 assert(_collectorState == FinalMarking || 4775 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4776 for (int j = 0; j < no_of_gc_threads; j++) { 4777 _cursor[j] = 0; 4778 } 4779 HeapWord* top = surv->top(); 4780 size_t i; 4781 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4782 HeapWord* min_val = top; // Higher than any PLAB address 4783 uint min_tid = 0; // position of min_val this round 4784 for (int j = 0; j < no_of_gc_threads; j++) { 4785 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4786 if (_cursor[j] == cur_sca->end()) { 4787 continue; 4788 } 4789 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4790 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4791 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4792 if (cur_val < min_val) { 4793 min_tid = j; 4794 min_val = cur_val; 4795 } else { 4796 assert(cur_val < top, "All recorded addresses should be less"); 4797 } 4798 } 4799 // At this point min_val and min_tid are respectively 4800 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4801 // and the thread (j) that witnesses that address. 4802 // We record this address in the _survivor_chunk_array[i] 4803 // and increment _cursor[min_tid] prior to the next round i. 4804 if (min_val == top) { 4805 break; 4806 } 4807 _survivor_chunk_array[i] = min_val; 4808 _cursor[min_tid]++; 4809 } 4810 // We are all done; record the size of the _survivor_chunk_array 4811 _survivor_chunk_index = i; // exclusive: [0, i) 4812 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4813 // Verify that we used up all the recorded entries 4814 #ifdef ASSERT 4815 size_t total = 0; 4816 for (int j = 0; j < no_of_gc_threads; j++) { 4817 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4818 total += _cursor[j]; 4819 } 4820 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4821 // Check that the merged array is in sorted order 4822 if (total > 0) { 4823 for (size_t i = 0; i < total - 1; i++) { 4824 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4825 i, p2i(_survivor_chunk_array[i])); 4826 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4827 "Not sorted"); 4828 } 4829 } 4830 #endif // ASSERT 4831 } 4832 4833 // Set up the space's par_seq_tasks structure for work claiming 4834 // for parallel initial scan and rescan of young gen. 4835 // See ParRescanTask where this is currently used. 4836 void 4837 CMSCollector:: 4838 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4839 assert(n_threads > 0, "Unexpected n_threads argument"); 4840 4841 // Eden space 4842 if (!_young_gen->eden()->is_empty()) { 4843 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4844 assert(!pst->valid(), "Clobbering existing data?"); 4845 // Each valid entry in [0, _eden_chunk_index) represents a task. 4846 size_t n_tasks = _eden_chunk_index + 1; 4847 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4848 // Sets the condition for completion of the subtask (how many threads 4849 // need to finish in order to be done). 4850 pst->set_n_threads(n_threads); 4851 pst->set_n_tasks((int)n_tasks); 4852 } 4853 4854 // Merge the survivor plab arrays into _survivor_chunk_array 4855 if (_survivor_plab_array != NULL) { 4856 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4857 } else { 4858 assert(_survivor_chunk_index == 0, "Error"); 4859 } 4860 4861 // To space 4862 { 4863 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4864 assert(!pst->valid(), "Clobbering existing data?"); 4865 // Sets the condition for completion of the subtask (how many threads 4866 // need to finish in order to be done). 4867 pst->set_n_threads(n_threads); 4868 pst->set_n_tasks(1); 4869 assert(pst->valid(), "Error"); 4870 } 4871 4872 // From space 4873 { 4874 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4875 assert(!pst->valid(), "Clobbering existing data?"); 4876 size_t n_tasks = _survivor_chunk_index + 1; 4877 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4878 // Sets the condition for completion of the subtask (how many threads 4879 // need to finish in order to be done). 4880 pst->set_n_threads(n_threads); 4881 pst->set_n_tasks((int)n_tasks); 4882 assert(pst->valid(), "Error"); 4883 } 4884 } 4885 4886 // Parallel version of remark 4887 void CMSCollector::do_remark_parallel() { 4888 CMSHeap* heap = CMSHeap::heap(); 4889 WorkGang* workers = heap->workers(); 4890 assert(workers != NULL, "Need parallel worker threads."); 4891 // Choose to use the number of GC workers most recently set 4892 // into "active_workers". 4893 uint n_workers = workers->active_workers(); 4894 4895 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4896 4897 StrongRootsScope srs(n_workers); 4898 4899 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4900 4901 // We won't be iterating over the cards in the card table updating 4902 // the younger_gen cards, so we shouldn't call the following else 4903 // the verification code as well as subsequent younger_refs_iterate 4904 // code would get confused. XXX 4905 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4906 4907 // The young gen rescan work will not be done as part of 4908 // process_roots (which currently doesn't know how to 4909 // parallelize such a scan), but rather will be broken up into 4910 // a set of parallel tasks (via the sampling that the [abortable] 4911 // preclean phase did of eden, plus the [two] tasks of 4912 // scanning the [two] survivor spaces. Further fine-grain 4913 // parallelization of the scanning of the survivor spaces 4914 // themselves, and of precleaning of the young gen itself 4915 // is deferred to the future. 4916 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4917 4918 // The dirty card rescan work is broken up into a "sequence" 4919 // of parallel tasks (per constituent space) that are dynamically 4920 // claimed by the parallel threads. 4921 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4922 4923 // It turns out that even when we're using 1 thread, doing the work in a 4924 // separate thread causes wide variance in run times. We can't help this 4925 // in the multi-threaded case, but we special-case n=1 here to get 4926 // repeatable measurements of the 1-thread overhead of the parallel code. 4927 if (n_workers > 1) { 4928 // Make refs discovery MT-safe, if it isn't already: it may not 4929 // necessarily be so, since it's possible that we are doing 4930 // ST marking. 4931 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4932 workers->run_task(&tsk); 4933 } else { 4934 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4935 tsk.work(0); 4936 } 4937 4938 // restore, single-threaded for now, any preserved marks 4939 // as a result of work_q overflow 4940 restore_preserved_marks_if_any(); 4941 } 4942 4943 // Non-parallel version of remark 4944 void CMSCollector::do_remark_non_parallel() { 4945 ResourceMark rm; 4946 HandleMark hm; 4947 CMSHeap* heap = CMSHeap::heap(); 4948 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4949 4950 MarkRefsIntoAndScanClosure 4951 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4952 &_markStack, this, 4953 false /* should_yield */, false /* not precleaning */); 4954 MarkFromDirtyCardsClosure 4955 markFromDirtyCardsClosure(this, _span, 4956 NULL, // space is set further below 4957 &_markBitMap, &_markStack, &mrias_cl); 4958 { 4959 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4960 // Iterate over the dirty cards, setting the corresponding bits in the 4961 // mod union table. 4962 { 4963 ModUnionClosure modUnionClosure(&_modUnionTable); 4964 _ct->dirty_card_iterate(_cmsGen->used_region(), 4965 &modUnionClosure); 4966 } 4967 // Having transferred these marks into the modUnionTable, we just need 4968 // to rescan the marked objects on the dirty cards in the modUnionTable. 4969 // The initial marking may have been done during an asynchronous 4970 // collection so there may be dirty bits in the mod-union table. 4971 const int alignment = CardTable::card_size * BitsPerWord; 4972 { 4973 // ... First handle dirty cards in CMS gen 4974 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4975 MemRegion ur = _cmsGen->used_region(); 4976 HeapWord* lb = ur.start(); 4977 HeapWord* ub = align_up(ur.end(), alignment); 4978 MemRegion cms_span(lb, ub); 4979 _modUnionTable.dirty_range_iterate_clear(cms_span, 4980 &markFromDirtyCardsClosure); 4981 verify_work_stacks_empty(); 4982 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4983 } 4984 } 4985 if (VerifyDuringGC && 4986 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4987 HandleMark hm; // Discard invalid handles created during verification 4988 Universe::verify(); 4989 } 4990 { 4991 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4992 4993 verify_work_stacks_empty(); 4994 4995 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4996 StrongRootsScope srs(1); 4997 4998 heap->cms_process_roots(&srs, 4999 true, // young gen as roots 5000 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5001 should_unload_classes(), 5002 &mrias_cl, 5003 NULL); // The dirty klasses will be handled below 5004 5005 assert(should_unload_classes() 5006 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5007 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5008 } 5009 5010 { 5011 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 5012 5013 verify_work_stacks_empty(); 5014 5015 // Scan all class loader data objects that might have been introduced 5016 // during concurrent marking. 5017 ResourceMark rm; 5018 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5019 for (int i = 0; i < array->length(); i++) { 5020 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 5021 } 5022 5023 // We don't need to keep track of new CLDs anymore. 5024 ClassLoaderDataGraph::remember_new_clds(false); 5025 5026 verify_work_stacks_empty(); 5027 } 5028 5029 // We might have added oops to ClassLoaderData::_handles during the 5030 // concurrent marking phase. These oops do not point to newly allocated objects 5031 // that are guaranteed to be kept alive. Hence, 5032 // we do have to revisit the _handles block during the remark phase. 5033 { 5034 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 5035 5036 verify_work_stacks_empty(); 5037 5038 RemarkCLDClosure remark_closure(&mrias_cl); 5039 ClassLoaderDataGraph::cld_do(&remark_closure); 5040 5041 verify_work_stacks_empty(); 5042 } 5043 5044 verify_work_stacks_empty(); 5045 // Restore evacuated mark words, if any, used for overflow list links 5046 restore_preserved_marks_if_any(); 5047 5048 verify_overflow_empty(); 5049 } 5050 5051 //////////////////////////////////////////////////////// 5052 // Parallel Reference Processing Task Proxy Class 5053 //////////////////////////////////////////////////////// 5054 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5055 OopTaskQueueSet* _queues; 5056 TaskTerminator _terminator; 5057 public: 5058 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5059 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5060 ParallelTaskTerminator* terminator() { return _terminator.terminator(); } 5061 OopTaskQueueSet* queues() { return _queues; } 5062 }; 5063 5064 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5065 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5066 CMSCollector* _collector; 5067 CMSBitMap* _mark_bit_map; 5068 const MemRegion _span; 5069 ProcessTask& _task; 5070 5071 public: 5072 CMSRefProcTaskProxy(ProcessTask& task, 5073 CMSCollector* collector, 5074 const MemRegion& span, 5075 CMSBitMap* mark_bit_map, 5076 AbstractWorkGang* workers, 5077 OopTaskQueueSet* task_queues): 5078 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5079 task_queues, 5080 workers->active_workers()), 5081 _collector(collector), 5082 _mark_bit_map(mark_bit_map), 5083 _span(span), 5084 _task(task) 5085 { 5086 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5087 "Inconsistency in _span"); 5088 } 5089 5090 OopTaskQueueSet* task_queues() { return queues(); } 5091 5092 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5093 5094 void do_work_steal(int i, 5095 CMSParDrainMarkingStackClosure* drain, 5096 CMSParKeepAliveClosure* keep_alive); 5097 5098 virtual void work(uint worker_id); 5099 }; 5100 5101 void CMSRefProcTaskProxy::work(uint worker_id) { 5102 ResourceMark rm; 5103 HandleMark hm; 5104 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5105 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5106 _mark_bit_map, 5107 work_queue(worker_id)); 5108 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5109 _mark_bit_map, 5110 work_queue(worker_id)); 5111 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5112 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5113 if (_task.marks_oops_alive()) { 5114 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5115 } 5116 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5117 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5118 } 5119 5120 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5121 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5122 _span(span), 5123 _work_queue(work_queue), 5124 _bit_map(bit_map), 5125 _mark_and_push(collector, span, bit_map, work_queue), 5126 _low_water_mark(MIN2((work_queue->max_elems()/4), 5127 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5128 { } 5129 5130 // . see if we can share work_queues with ParNew? XXX 5131 void CMSRefProcTaskProxy::do_work_steal(int i, 5132 CMSParDrainMarkingStackClosure* drain, 5133 CMSParKeepAliveClosure* keep_alive) { 5134 OopTaskQueue* work_q = work_queue(i); 5135 NOT_PRODUCT(int num_steals = 0;) 5136 oop obj_to_scan; 5137 5138 while (true) { 5139 // Completely finish any left over work from (an) earlier round(s) 5140 drain->trim_queue(0); 5141 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5142 (size_t)ParGCDesiredObjsFromOverflowList); 5143 // Now check if there's any work in the overflow list 5144 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5145 // only affects the number of attempts made to get work from the 5146 // overflow list and does not affect the number of workers. Just 5147 // pass ParallelGCThreads so this behavior is unchanged. 5148 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5149 work_q, 5150 ParallelGCThreads)) { 5151 // Found something in global overflow list; 5152 // not yet ready to go stealing work from others. 5153 // We'd like to assert(work_q->size() != 0, ...) 5154 // because we just took work from the overflow list, 5155 // but of course we can't, since all of that might have 5156 // been already stolen from us. 5157 continue; 5158 } 5159 // Verify that we have no work before we resort to stealing 5160 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5161 // Try to steal from other queues that have work 5162 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5163 NOT_PRODUCT(num_steals++;) 5164 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5165 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5166 // Do scanning work 5167 obj_to_scan->oop_iterate(keep_alive); 5168 // Loop around, finish this work, and try to steal some more 5169 } else if (terminator()->offer_termination()) { 5170 break; // nirvana from the infinite cycle 5171 } 5172 } 5173 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5174 } 5175 5176 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5177 CMSHeap* heap = CMSHeap::heap(); 5178 WorkGang* workers = heap->workers(); 5179 assert(workers != NULL, "Need parallel worker threads."); 5180 assert(workers->active_workers() == ergo_workers, 5181 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5182 ergo_workers, workers->active_workers()); 5183 CMSRefProcTaskProxy rp_task(task, &_collector, 5184 _collector.ref_processor_span(), 5185 _collector.markBitMap(), 5186 workers, _collector.task_queues()); 5187 workers->run_task(&rp_task, workers->active_workers()); 5188 } 5189 5190 void CMSCollector::refProcessingWork() { 5191 ResourceMark rm; 5192 HandleMark hm; 5193 5194 ReferenceProcessor* rp = ref_processor(); 5195 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5196 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5197 // Process weak references. 5198 rp->setup_policy(false); 5199 verify_work_stacks_empty(); 5200 5201 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5202 { 5203 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5204 5205 // Setup keep_alive and complete closures. 5206 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5207 &_markStack, false /* !preclean */); 5208 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5209 _span, &_markBitMap, &_markStack, 5210 &cmsKeepAliveClosure, false /* !preclean */); 5211 5212 ReferenceProcessorStats stats; 5213 if (rp->processing_is_mt()) { 5214 // Set the degree of MT here. If the discovery is done MT, there 5215 // may have been a different number of threads doing the discovery 5216 // and a different number of discovered lists may have Ref objects. 5217 // That is OK as long as the Reference lists are balanced (see 5218 // balance_all_queues() and balance_queues()). 5219 CMSHeap* heap = CMSHeap::heap(); 5220 uint active_workers = ParallelGCThreads; 5221 WorkGang* workers = heap->workers(); 5222 if (workers != NULL) { 5223 active_workers = workers->active_workers(); 5224 // The expectation is that active_workers will have already 5225 // been set to a reasonable value. If it has not been set, 5226 // investigate. 5227 assert(active_workers > 0, "Should have been set during scavenge"); 5228 } 5229 rp->set_active_mt_degree(active_workers); 5230 CMSRefProcTaskExecutor task_executor(*this); 5231 stats = rp->process_discovered_references(&_is_alive_closure, 5232 &cmsKeepAliveClosure, 5233 &cmsDrainMarkingStackClosure, 5234 &task_executor, 5235 &pt); 5236 } else { 5237 stats = rp->process_discovered_references(&_is_alive_closure, 5238 &cmsKeepAliveClosure, 5239 &cmsDrainMarkingStackClosure, 5240 NULL, 5241 &pt); 5242 } 5243 _gc_tracer_cm->report_gc_reference_stats(stats); 5244 pt.print_all_references(); 5245 } 5246 5247 // This is the point where the entire marking should have completed. 5248 verify_work_stacks_empty(); 5249 5250 { 5251 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5252 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5253 } 5254 5255 if (should_unload_classes()) { 5256 { 5257 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5258 5259 // Unload classes and purge the SystemDictionary. 5260 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5261 5262 // Unload nmethods. 5263 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5264 5265 // Prune dead klasses from subklass/sibling/implementor lists. 5266 Klass::clean_weak_klass_links(purged_class); 5267 5268 // Clean JVMCI metadata handles. 5269 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 5270 } 5271 } 5272 5273 // Restore any preserved marks as a result of mark stack or 5274 // work queue overflow 5275 restore_preserved_marks_if_any(); // done single-threaded for now 5276 5277 rp->set_enqueuing_is_done(true); 5278 rp->verify_no_references_recorded(); 5279 } 5280 5281 #ifndef PRODUCT 5282 void CMSCollector::check_correct_thread_executing() { 5283 Thread* t = Thread::current(); 5284 // Only the VM thread or the CMS thread should be here. 5285 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5286 "Unexpected thread type"); 5287 // If this is the vm thread, the foreground process 5288 // should not be waiting. Note that _foregroundGCIsActive is 5289 // true while the foreground collector is waiting. 5290 if (_foregroundGCShouldWait) { 5291 // We cannot be the VM thread 5292 assert(t->is_ConcurrentGC_thread(), 5293 "Should be CMS thread"); 5294 } else { 5295 // We can be the CMS thread only if we are in a stop-world 5296 // phase of CMS collection. 5297 if (t->is_ConcurrentGC_thread()) { 5298 assert(_collectorState == InitialMarking || 5299 _collectorState == FinalMarking, 5300 "Should be a stop-world phase"); 5301 // The CMS thread should be holding the CMS_token. 5302 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5303 "Potential interference with concurrently " 5304 "executing VM thread"); 5305 } 5306 } 5307 } 5308 #endif 5309 5310 void CMSCollector::sweep() { 5311 assert(_collectorState == Sweeping, "just checking"); 5312 check_correct_thread_executing(); 5313 verify_work_stacks_empty(); 5314 verify_overflow_empty(); 5315 increment_sweep_count(); 5316 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5317 5318 _inter_sweep_timer.stop(); 5319 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5320 5321 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5322 _intra_sweep_timer.reset(); 5323 _intra_sweep_timer.start(); 5324 { 5325 GCTraceCPUTime tcpu; 5326 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5327 // First sweep the old gen 5328 { 5329 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5330 bitMapLock()); 5331 sweepWork(_cmsGen); 5332 } 5333 5334 // Update Universe::_heap_*_at_gc figures. 5335 // We need all the free list locks to make the abstract state 5336 // transition from Sweeping to Resetting. See detailed note 5337 // further below. 5338 { 5339 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5340 // Update heap occupancy information which is used as 5341 // input to soft ref clearing policy at the next gc. 5342 Universe::update_heap_info_at_gc(); 5343 _collectorState = Resizing; 5344 } 5345 } 5346 verify_work_stacks_empty(); 5347 verify_overflow_empty(); 5348 5349 if (should_unload_classes()) { 5350 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5351 // requires that the virtual spaces are stable and not deleted. 5352 ClassLoaderDataGraph::set_should_purge(true); 5353 } 5354 5355 _intra_sweep_timer.stop(); 5356 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5357 5358 _inter_sweep_timer.reset(); 5359 _inter_sweep_timer.start(); 5360 5361 // We need to use a monotonically non-decreasing time in ms 5362 // or we will see time-warp warnings and os::javaTimeMillis() 5363 // does not guarantee monotonicity. 5364 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5365 update_time_of_last_gc(now); 5366 5367 // NOTE on abstract state transitions: 5368 // Mutators allocate-live and/or mark the mod-union table dirty 5369 // based on the state of the collection. The former is done in 5370 // the interval [Marking, Sweeping] and the latter in the interval 5371 // [Marking, Sweeping). Thus the transitions into the Marking state 5372 // and out of the Sweeping state must be synchronously visible 5373 // globally to the mutators. 5374 // The transition into the Marking state happens with the world 5375 // stopped so the mutators will globally see it. Sweeping is 5376 // done asynchronously by the background collector so the transition 5377 // from the Sweeping state to the Resizing state must be done 5378 // under the freelistLock (as is the check for whether to 5379 // allocate-live and whether to dirty the mod-union table). 5380 assert(_collectorState == Resizing, "Change of collector state to" 5381 " Resizing must be done under the freelistLocks (plural)"); 5382 5383 // Now that sweeping has been completed, we clear 5384 // the incremental_collection_failed flag, 5385 // thus inviting a younger gen collection to promote into 5386 // this generation. If such a promotion may still fail, 5387 // the flag will be set again when a young collection is 5388 // attempted. 5389 CMSHeap* heap = CMSHeap::heap(); 5390 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5391 heap->update_full_collections_completed(_collection_count_start); 5392 } 5393 5394 // FIX ME!!! Looks like this belongs in CFLSpace, with 5395 // CMSGen merely delegating to it. 5396 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5397 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5398 HeapWord* minAddr = _cmsSpace->bottom(); 5399 HeapWord* largestAddr = 5400 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5401 if (largestAddr == NULL) { 5402 // The dictionary appears to be empty. In this case 5403 // try to coalesce at the end of the heap. 5404 largestAddr = _cmsSpace->end(); 5405 } 5406 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5407 size_t nearLargestOffset = 5408 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5409 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5410 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5411 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5412 } 5413 5414 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5415 return addr >= _cmsSpace->nearLargestChunk(); 5416 } 5417 5418 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5419 return _cmsSpace->find_chunk_at_end(); 5420 } 5421 5422 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5423 bool full) { 5424 // If the young generation has been collected, gather any statistics 5425 // that are of interest at this point. 5426 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5427 if (!full && current_is_young) { 5428 // Gather statistics on the young generation collection. 5429 collector()->stats().record_gc0_end(used()); 5430 } 5431 } 5432 5433 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5434 // We iterate over the space(s) underlying this generation, 5435 // checking the mark bit map to see if the bits corresponding 5436 // to specific blocks are marked or not. Blocks that are 5437 // marked are live and are not swept up. All remaining blocks 5438 // are swept up, with coalescing on-the-fly as we sweep up 5439 // contiguous free and/or garbage blocks: 5440 // We need to ensure that the sweeper synchronizes with allocators 5441 // and stop-the-world collectors. In particular, the following 5442 // locks are used: 5443 // . CMS token: if this is held, a stop the world collection cannot occur 5444 // . freelistLock: if this is held no allocation can occur from this 5445 // generation by another thread 5446 // . bitMapLock: if this is held, no other thread can access or update 5447 // 5448 5449 // Note that we need to hold the freelistLock if we use 5450 // block iterate below; else the iterator might go awry if 5451 // a mutator (or promotion) causes block contents to change 5452 // (for instance if the allocator divvies up a block). 5453 // If we hold the free list lock, for all practical purposes 5454 // young generation GC's can't occur (they'll usually need to 5455 // promote), so we might as well prevent all young generation 5456 // GC's while we do a sweeping step. For the same reason, we might 5457 // as well take the bit map lock for the entire duration 5458 5459 // check that we hold the requisite locks 5460 assert(have_cms_token(), "Should hold cms token"); 5461 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5462 assert_lock_strong(old_gen->freelistLock()); 5463 assert_lock_strong(bitMapLock()); 5464 5465 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5466 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5467 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5468 _inter_sweep_estimate.padded_average(), 5469 _intra_sweep_estimate.padded_average()); 5470 old_gen->setNearLargestChunk(); 5471 5472 { 5473 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5474 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5475 // We need to free-up/coalesce garbage/blocks from a 5476 // co-terminal free run. This is done in the SweepClosure 5477 // destructor; so, do not remove this scope, else the 5478 // end-of-sweep-census below will be off by a little bit. 5479 } 5480 old_gen->cmsSpace()->sweep_completed(); 5481 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5482 if (should_unload_classes()) { // unloaded classes this cycle, 5483 _concurrent_cycles_since_last_unload = 0; // ... reset count 5484 } else { // did not unload classes, 5485 _concurrent_cycles_since_last_unload++; // ... increment count 5486 } 5487 } 5488 5489 // Reset CMS data structures (for now just the marking bit map) 5490 // preparatory for the next cycle. 5491 void CMSCollector::reset_concurrent() { 5492 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5493 5494 // If the state is not "Resetting", the foreground thread 5495 // has done a collection and the resetting. 5496 if (_collectorState != Resetting) { 5497 assert(_collectorState == Idling, "The state should only change" 5498 " because the foreground collector has finished the collection"); 5499 return; 5500 } 5501 5502 { 5503 // Clear the mark bitmap (no grey objects to start with) 5504 // for the next cycle. 5505 GCTraceCPUTime tcpu; 5506 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5507 5508 HeapWord* curAddr = _markBitMap.startWord(); 5509 while (curAddr < _markBitMap.endWord()) { 5510 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5511 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5512 _markBitMap.clear_large_range(chunk); 5513 if (ConcurrentMarkSweepThread::should_yield() && 5514 !foregroundGCIsActive() && 5515 CMSYield) { 5516 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5517 "CMS thread should hold CMS token"); 5518 assert_lock_strong(bitMapLock()); 5519 bitMapLock()->unlock(); 5520 ConcurrentMarkSweepThread::desynchronize(true); 5521 stopTimer(); 5522 incrementYields(); 5523 5524 // See the comment in coordinator_yield() 5525 for (unsigned i = 0; i < CMSYieldSleepCount && 5526 ConcurrentMarkSweepThread::should_yield() && 5527 !CMSCollector::foregroundGCIsActive(); ++i) { 5528 os::sleep(Thread::current(), 1, false); 5529 } 5530 5531 ConcurrentMarkSweepThread::synchronize(true); 5532 bitMapLock()->lock_without_safepoint_check(); 5533 startTimer(); 5534 } 5535 curAddr = chunk.end(); 5536 } 5537 // A successful mostly concurrent collection has been done. 5538 // Because only the full (i.e., concurrent mode failure) collections 5539 // are being measured for gc overhead limits, clean the "near" flag 5540 // and count. 5541 size_policy()->reset_gc_overhead_limit_count(); 5542 _collectorState = Idling; 5543 } 5544 5545 register_gc_end(); 5546 } 5547 5548 // Same as above but for STW paths 5549 void CMSCollector::reset_stw() { 5550 // already have the lock 5551 assert(_collectorState == Resetting, "just checking"); 5552 assert_lock_strong(bitMapLock()); 5553 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5554 _markBitMap.clear_all(); 5555 _collectorState = Idling; 5556 register_gc_end(); 5557 } 5558 5559 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5560 GCTraceCPUTime tcpu; 5561 TraceCollectorStats tcs_cgc(cgc_counters()); 5562 5563 switch (op) { 5564 case CMS_op_checkpointRootsInitial: { 5565 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5566 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5567 checkpointRootsInitial(); 5568 break; 5569 } 5570 case CMS_op_checkpointRootsFinal: { 5571 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5572 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5573 checkpointRootsFinal(); 5574 break; 5575 } 5576 default: 5577 fatal("No such CMS_op"); 5578 } 5579 } 5580 5581 #ifndef PRODUCT 5582 size_t const CMSCollector::skip_header_HeapWords() { 5583 return FreeChunk::header_size(); 5584 } 5585 5586 // Try and collect here conditions that should hold when 5587 // CMS thread is exiting. The idea is that the foreground GC 5588 // thread should not be blocked if it wants to terminate 5589 // the CMS thread and yet continue to run the VM for a while 5590 // after that. 5591 void CMSCollector::verify_ok_to_terminate() const { 5592 assert(Thread::current()->is_ConcurrentGC_thread(), 5593 "should be called by CMS thread"); 5594 assert(!_foregroundGCShouldWait, "should be false"); 5595 // We could check here that all the various low-level locks 5596 // are not held by the CMS thread, but that is overkill; see 5597 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5598 // is checked. 5599 } 5600 #endif 5601 5602 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5603 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5604 "missing Printezis mark?"); 5605 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5606 size_t size = pointer_delta(nextOneAddr + 1, addr); 5607 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5608 "alignment problem"); 5609 assert(size >= 3, "Necessary for Printezis marks to work"); 5610 return size; 5611 } 5612 5613 // A variant of the above (block_size_using_printezis_bits()) except 5614 // that we return 0 if the P-bits are not yet set. 5615 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5616 if (_markBitMap.isMarked(addr + 1)) { 5617 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5618 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5619 size_t size = pointer_delta(nextOneAddr + 1, addr); 5620 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5621 "alignment problem"); 5622 assert(size >= 3, "Necessary for Printezis marks to work"); 5623 return size; 5624 } 5625 return 0; 5626 } 5627 5628 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5629 size_t sz = 0; 5630 oop p = (oop)addr; 5631 if (p->klass_or_null_acquire() != NULL) { 5632 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5633 } else { 5634 sz = block_size_using_printezis_bits(addr); 5635 } 5636 assert(sz > 0, "size must be nonzero"); 5637 HeapWord* next_block = addr + sz; 5638 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5639 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5640 align_down((uintptr_t)next_card, CardTable::card_size), 5641 "must be different cards"); 5642 return next_card; 5643 } 5644 5645 5646 // CMS Bit Map Wrapper ///////////////////////////////////////// 5647 5648 // Construct a CMS bit map infrastructure, but don't create the 5649 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5650 // further below. 5651 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5652 _shifter(shifter), 5653 _bm(), 5654 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5655 Monitor::_safepoint_check_never) : NULL) 5656 { 5657 _bmStartWord = 0; 5658 _bmWordSize = 0; 5659 } 5660 5661 bool CMSBitMap::allocate(MemRegion mr) { 5662 _bmStartWord = mr.start(); 5663 _bmWordSize = mr.word_size(); 5664 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5665 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5666 if (!brs.is_reserved()) { 5667 log_warning(gc)("CMS bit map allocation failure"); 5668 return false; 5669 } 5670 // For now we'll just commit all of the bit map up front. 5671 // Later on we'll try to be more parsimonious with swap. 5672 if (!_virtual_space.initialize(brs, brs.size())) { 5673 log_warning(gc)("CMS bit map backing store failure"); 5674 return false; 5675 } 5676 assert(_virtual_space.committed_size() == brs.size(), 5677 "didn't reserve backing store for all of CMS bit map?"); 5678 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5679 _bmWordSize, "inconsistency in bit map sizing"); 5680 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5681 5682 // bm.clear(); // can we rely on getting zero'd memory? verify below 5683 assert(isAllClear(), 5684 "Expected zero'd memory from ReservedSpace constructor"); 5685 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5686 "consistency check"); 5687 return true; 5688 } 5689 5690 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5691 HeapWord *next_addr, *end_addr, *last_addr; 5692 assert_locked(); 5693 assert(covers(mr), "out-of-range error"); 5694 // XXX assert that start and end are appropriately aligned 5695 for (next_addr = mr.start(), end_addr = mr.end(); 5696 next_addr < end_addr; next_addr = last_addr) { 5697 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5698 last_addr = dirty_region.end(); 5699 if (!dirty_region.is_empty()) { 5700 cl->do_MemRegion(dirty_region); 5701 } else { 5702 assert(last_addr == end_addr, "program logic"); 5703 return; 5704 } 5705 } 5706 } 5707 5708 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5709 _bm.print_on_error(st, prefix); 5710 } 5711 5712 #ifndef PRODUCT 5713 void CMSBitMap::assert_locked() const { 5714 CMSLockVerifier::assert_locked(lock()); 5715 } 5716 5717 bool CMSBitMap::covers(MemRegion mr) const { 5718 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5719 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5720 "size inconsistency"); 5721 return (mr.start() >= _bmStartWord) && 5722 (mr.end() <= endWord()); 5723 } 5724 5725 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5726 return (start >= _bmStartWord && (start + size) <= endWord()); 5727 } 5728 5729 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5730 // verify that there are no 1 bits in the interval [left, right) 5731 FalseBitMapClosure falseBitMapClosure; 5732 iterate(&falseBitMapClosure, left, right); 5733 } 5734 5735 void CMSBitMap::region_invariant(MemRegion mr) 5736 { 5737 assert_locked(); 5738 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5739 assert(!mr.is_empty(), "unexpected empty region"); 5740 assert(covers(mr), "mr should be covered by bit map"); 5741 // convert address range into offset range 5742 size_t start_ofs = heapWordToOffset(mr.start()); 5743 // Make sure that end() is appropriately aligned 5744 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5745 "Misaligned mr.end()"); 5746 size_t end_ofs = heapWordToOffset(mr.end()); 5747 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5748 } 5749 5750 #endif 5751 5752 bool CMSMarkStack::allocate(size_t size) { 5753 // allocate a stack of the requisite depth 5754 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5755 size * sizeof(oop))); 5756 if (!rs.is_reserved()) { 5757 log_warning(gc)("CMSMarkStack allocation failure"); 5758 return false; 5759 } 5760 if (!_virtual_space.initialize(rs, rs.size())) { 5761 log_warning(gc)("CMSMarkStack backing store failure"); 5762 return false; 5763 } 5764 assert(_virtual_space.committed_size() == rs.size(), 5765 "didn't reserve backing store for all of CMS stack?"); 5766 _base = (oop*)(_virtual_space.low()); 5767 _index = 0; 5768 _capacity = size; 5769 NOT_PRODUCT(_max_depth = 0); 5770 return true; 5771 } 5772 5773 // XXX FIX ME !!! In the MT case we come in here holding a 5774 // leaf lock. For printing we need to take a further lock 5775 // which has lower rank. We need to recalibrate the two 5776 // lock-ranks involved in order to be able to print the 5777 // messages below. (Or defer the printing to the caller. 5778 // For now we take the expedient path of just disabling the 5779 // messages for the problematic case.) 5780 void CMSMarkStack::expand() { 5781 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5782 if (_capacity == MarkStackSizeMax) { 5783 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5784 // We print a warning message only once per CMS cycle. 5785 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5786 } 5787 return; 5788 } 5789 // Double capacity if possible 5790 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5791 // Do not give up existing stack until we have managed to 5792 // get the double capacity that we desired. 5793 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5794 new_capacity * sizeof(oop))); 5795 if (rs.is_reserved()) { 5796 // Release the backing store associated with old stack 5797 _virtual_space.release(); 5798 // Reinitialize virtual space for new stack 5799 if (!_virtual_space.initialize(rs, rs.size())) { 5800 fatal("Not enough swap for expanded marking stack"); 5801 } 5802 _base = (oop*)(_virtual_space.low()); 5803 _index = 0; 5804 _capacity = new_capacity; 5805 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5806 // Failed to double capacity, continue; 5807 // we print a detail message only once per CMS cycle. 5808 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5809 _capacity / K, new_capacity / K); 5810 } 5811 } 5812 5813 5814 // Closures 5815 // XXX: there seems to be a lot of code duplication here; 5816 // should refactor and consolidate common code. 5817 5818 // This closure is used to mark refs into the CMS generation in 5819 // the CMS bit map. Called at the first checkpoint. This closure 5820 // assumes that we do not need to re-mark dirty cards; if the CMS 5821 // generation on which this is used is not an oldest 5822 // generation then this will lose younger_gen cards! 5823 5824 MarkRefsIntoClosure::MarkRefsIntoClosure( 5825 MemRegion span, CMSBitMap* bitMap): 5826 _span(span), 5827 _bitMap(bitMap) 5828 { 5829 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5830 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5831 } 5832 5833 void MarkRefsIntoClosure::do_oop(oop obj) { 5834 // if p points into _span, then mark corresponding bit in _markBitMap 5835 assert(oopDesc::is_oop(obj), "expected an oop"); 5836 HeapWord* addr = (HeapWord*)obj; 5837 if (_span.contains(addr)) { 5838 // this should be made more efficient 5839 _bitMap->mark(addr); 5840 } 5841 } 5842 5843 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5844 MemRegion span, CMSBitMap* bitMap): 5845 _span(span), 5846 _bitMap(bitMap) 5847 { 5848 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5849 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5850 } 5851 5852 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5853 // if p points into _span, then mark corresponding bit in _markBitMap 5854 assert(oopDesc::is_oop(obj), "expected an oop"); 5855 HeapWord* addr = (HeapWord*)obj; 5856 if (_span.contains(addr)) { 5857 // this should be made more efficient 5858 _bitMap->par_mark(addr); 5859 } 5860 } 5861 5862 // A variant of the above, used for CMS marking verification. 5863 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5864 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5865 _span(span), 5866 _verification_bm(verification_bm), 5867 _cms_bm(cms_bm) 5868 { 5869 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5870 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5871 } 5872 5873 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5874 // if p points into _span, then mark corresponding bit in _markBitMap 5875 assert(oopDesc::is_oop(obj), "expected an oop"); 5876 HeapWord* addr = (HeapWord*)obj; 5877 if (_span.contains(addr)) { 5878 _verification_bm->mark(addr); 5879 if (!_cms_bm->isMarked(addr)) { 5880 Log(gc, verify) log; 5881 ResourceMark rm; 5882 LogStream ls(log.error()); 5883 oop(addr)->print_on(&ls); 5884 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5885 fatal("... aborting"); 5886 } 5887 } 5888 } 5889 5890 ////////////////////////////////////////////////// 5891 // MarkRefsIntoAndScanClosure 5892 ////////////////////////////////////////////////// 5893 5894 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5895 ReferenceDiscoverer* rd, 5896 CMSBitMap* bit_map, 5897 CMSBitMap* mod_union_table, 5898 CMSMarkStack* mark_stack, 5899 CMSCollector* collector, 5900 bool should_yield, 5901 bool concurrent_precleaning): 5902 _span(span), 5903 _bit_map(bit_map), 5904 _mark_stack(mark_stack), 5905 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5906 mark_stack, concurrent_precleaning), 5907 _collector(collector), 5908 _freelistLock(NULL), 5909 _yield(should_yield), 5910 _concurrent_precleaning(concurrent_precleaning) 5911 { 5912 // FIXME: Should initialize in base class constructor. 5913 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5914 set_ref_discoverer_internal(rd); 5915 } 5916 5917 // This closure is used to mark refs into the CMS generation at the 5918 // second (final) checkpoint, and to scan and transitively follow 5919 // the unmarked oops. It is also used during the concurrent precleaning 5920 // phase while scanning objects on dirty cards in the CMS generation. 5921 // The marks are made in the marking bit map and the marking stack is 5922 // used for keeping the (newly) grey objects during the scan. 5923 // The parallel version (Par_...) appears further below. 5924 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5925 if (obj != NULL) { 5926 assert(oopDesc::is_oop(obj), "expected an oop"); 5927 HeapWord* addr = (HeapWord*)obj; 5928 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5929 assert(_collector->overflow_list_is_empty(), 5930 "overflow list should be empty"); 5931 if (_span.contains(addr) && 5932 !_bit_map->isMarked(addr)) { 5933 // mark bit map (object is now grey) 5934 _bit_map->mark(addr); 5935 // push on marking stack (stack should be empty), and drain the 5936 // stack by applying this closure to the oops in the oops popped 5937 // from the stack (i.e. blacken the grey objects) 5938 bool res = _mark_stack->push(obj); 5939 assert(res, "Should have space to push on empty stack"); 5940 do { 5941 oop new_oop = _mark_stack->pop(); 5942 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5943 assert(_bit_map->isMarked((HeapWord*)new_oop), 5944 "only grey objects on this stack"); 5945 // iterate over the oops in this oop, marking and pushing 5946 // the ones in CMS heap (i.e. in _span). 5947 new_oop->oop_iterate(&_pushAndMarkClosure); 5948 // check if it's time to yield 5949 do_yield_check(); 5950 } while (!_mark_stack->isEmpty() || 5951 (!_concurrent_precleaning && take_from_overflow_list())); 5952 // if marking stack is empty, and we are not doing this 5953 // during precleaning, then check the overflow list 5954 } 5955 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5956 assert(_collector->overflow_list_is_empty(), 5957 "overflow list was drained above"); 5958 5959 assert(_collector->no_preserved_marks(), 5960 "All preserved marks should have been restored above"); 5961 } 5962 } 5963 5964 void MarkRefsIntoAndScanClosure::do_yield_work() { 5965 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5966 "CMS thread should hold CMS token"); 5967 assert_lock_strong(_freelistLock); 5968 assert_lock_strong(_bit_map->lock()); 5969 // relinquish the free_list_lock and bitMaplock() 5970 _bit_map->lock()->unlock(); 5971 _freelistLock->unlock(); 5972 ConcurrentMarkSweepThread::desynchronize(true); 5973 _collector->stopTimer(); 5974 _collector->incrementYields(); 5975 5976 // See the comment in coordinator_yield() 5977 for (unsigned i = 0; 5978 i < CMSYieldSleepCount && 5979 ConcurrentMarkSweepThread::should_yield() && 5980 !CMSCollector::foregroundGCIsActive(); 5981 ++i) { 5982 os::sleep(Thread::current(), 1, false); 5983 } 5984 5985 ConcurrentMarkSweepThread::synchronize(true); 5986 _freelistLock->lock_without_safepoint_check(); 5987 _bit_map->lock()->lock_without_safepoint_check(); 5988 _collector->startTimer(); 5989 } 5990 5991 /////////////////////////////////////////////////////////// 5992 // ParMarkRefsIntoAndScanClosure: a parallel version of 5993 // MarkRefsIntoAndScanClosure 5994 /////////////////////////////////////////////////////////// 5995 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5996 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5997 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5998 _span(span), 5999 _bit_map(bit_map), 6000 _work_queue(work_queue), 6001 _low_water_mark(MIN2((work_queue->max_elems()/4), 6002 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6003 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 6004 { 6005 // FIXME: Should initialize in base class constructor. 6006 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 6007 set_ref_discoverer_internal(rd); 6008 } 6009 6010 // This closure is used to mark refs into the CMS generation at the 6011 // second (final) checkpoint, and to scan and transitively follow 6012 // the unmarked oops. The marks are made in the marking bit map and 6013 // the work_queue is used for keeping the (newly) grey objects during 6014 // the scan phase whence they are also available for stealing by parallel 6015 // threads. Since the marking bit map is shared, updates are 6016 // synchronized (via CAS). 6017 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6018 if (obj != NULL) { 6019 // Ignore mark word because this could be an already marked oop 6020 // that may be chained at the end of the overflow list. 6021 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6022 HeapWord* addr = (HeapWord*)obj; 6023 if (_span.contains(addr) && 6024 !_bit_map->isMarked(addr)) { 6025 // mark bit map (object will become grey): 6026 // It is possible for several threads to be 6027 // trying to "claim" this object concurrently; 6028 // the unique thread that succeeds in marking the 6029 // object first will do the subsequent push on 6030 // to the work queue (or overflow list). 6031 if (_bit_map->par_mark(addr)) { 6032 // push on work_queue (which may not be empty), and trim the 6033 // queue to an appropriate length by applying this closure to 6034 // the oops in the oops popped from the stack (i.e. blacken the 6035 // grey objects) 6036 bool res = _work_queue->push(obj); 6037 assert(res, "Low water mark should be less than capacity?"); 6038 trim_queue(_low_water_mark); 6039 } // Else, another thread claimed the object 6040 } 6041 } 6042 } 6043 6044 // This closure is used to rescan the marked objects on the dirty cards 6045 // in the mod union table and the card table proper. 6046 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6047 oop p, MemRegion mr) { 6048 6049 size_t size = 0; 6050 HeapWord* addr = (HeapWord*)p; 6051 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6052 assert(_span.contains(addr), "we are scanning the CMS generation"); 6053 // check if it's time to yield 6054 if (do_yield_check()) { 6055 // We yielded for some foreground stop-world work, 6056 // and we have been asked to abort this ongoing preclean cycle. 6057 return 0; 6058 } 6059 if (_bitMap->isMarked(addr)) { 6060 // it's marked; is it potentially uninitialized? 6061 if (p->klass_or_null_acquire() != NULL) { 6062 // an initialized object; ignore mark word in verification below 6063 // since we are running concurrent with mutators 6064 assert(oopDesc::is_oop(p, true), "should be an oop"); 6065 if (p->is_objArray()) { 6066 // objArrays are precisely marked; restrict scanning 6067 // to dirty cards only. 6068 size = CompactibleFreeListSpace::adjustObjectSize( 6069 p->oop_iterate_size(_scanningClosure, mr)); 6070 } else { 6071 // A non-array may have been imprecisely marked; we need 6072 // to scan object in its entirety. 6073 size = CompactibleFreeListSpace::adjustObjectSize( 6074 p->oop_iterate_size(_scanningClosure)); 6075 } 6076 #ifdef ASSERT 6077 size_t direct_size = 6078 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6079 assert(size == direct_size, "Inconsistency in size"); 6080 assert(size >= 3, "Necessary for Printezis marks to work"); 6081 HeapWord* start_pbit = addr + 1; 6082 HeapWord* end_pbit = addr + size - 1; 6083 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6084 "inconsistent Printezis mark"); 6085 // Verify inner mark bits (between Printezis bits) are clear, 6086 // but don't repeat if there are multiple dirty regions for 6087 // the same object, to avoid potential O(N^2) performance. 6088 if (addr != _last_scanned_object) { 6089 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6090 _last_scanned_object = addr; 6091 } 6092 #endif // ASSERT 6093 } else { 6094 // An uninitialized object. 6095 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6096 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6097 size = pointer_delta(nextOneAddr + 1, addr); 6098 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6099 "alignment problem"); 6100 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6101 // will dirty the card when the klass pointer is installed in the 6102 // object (signaling the completion of initialization). 6103 } 6104 } else { 6105 // Either a not yet marked object or an uninitialized object 6106 if (p->klass_or_null_acquire() == NULL) { 6107 // An uninitialized object, skip to the next card, since 6108 // we may not be able to read its P-bits yet. 6109 assert(size == 0, "Initial value"); 6110 } else { 6111 // An object not (yet) reached by marking: we merely need to 6112 // compute its size so as to go look at the next block. 6113 assert(oopDesc::is_oop(p, true), "should be an oop"); 6114 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6115 } 6116 } 6117 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6118 return size; 6119 } 6120 6121 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6122 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6123 "CMS thread should hold CMS token"); 6124 assert_lock_strong(_freelistLock); 6125 assert_lock_strong(_bitMap->lock()); 6126 // relinquish the free_list_lock and bitMaplock() 6127 _bitMap->lock()->unlock(); 6128 _freelistLock->unlock(); 6129 ConcurrentMarkSweepThread::desynchronize(true); 6130 _collector->stopTimer(); 6131 _collector->incrementYields(); 6132 6133 // See the comment in coordinator_yield() 6134 for (unsigned i = 0; i < CMSYieldSleepCount && 6135 ConcurrentMarkSweepThread::should_yield() && 6136 !CMSCollector::foregroundGCIsActive(); ++i) { 6137 os::sleep(Thread::current(), 1, false); 6138 } 6139 6140 ConcurrentMarkSweepThread::synchronize(true); 6141 _freelistLock->lock_without_safepoint_check(); 6142 _bitMap->lock()->lock_without_safepoint_check(); 6143 _collector->startTimer(); 6144 } 6145 6146 6147 ////////////////////////////////////////////////////////////////// 6148 // SurvivorSpacePrecleanClosure 6149 ////////////////////////////////////////////////////////////////// 6150 // This (single-threaded) closure is used to preclean the oops in 6151 // the survivor spaces. 6152 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6153 6154 HeapWord* addr = (HeapWord*)p; 6155 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6156 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6157 assert(p->klass_or_null() != NULL, "object should be initialized"); 6158 // an initialized object; ignore mark word in verification below 6159 // since we are running concurrent with mutators 6160 assert(oopDesc::is_oop(p, true), "should be an oop"); 6161 // Note that we do not yield while we iterate over 6162 // the interior oops of p, pushing the relevant ones 6163 // on our marking stack. 6164 size_t size = p->oop_iterate_size(_scanning_closure); 6165 do_yield_check(); 6166 // Observe that below, we do not abandon the preclean 6167 // phase as soon as we should; rather we empty the 6168 // marking stack before returning. This is to satisfy 6169 // some existing assertions. In general, it may be a 6170 // good idea to abort immediately and complete the marking 6171 // from the grey objects at a later time. 6172 while (!_mark_stack->isEmpty()) { 6173 oop new_oop = _mark_stack->pop(); 6174 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6175 assert(_bit_map->isMarked((HeapWord*)new_oop), 6176 "only grey objects on this stack"); 6177 // iterate over the oops in this oop, marking and pushing 6178 // the ones in CMS heap (i.e. in _span). 6179 new_oop->oop_iterate(_scanning_closure); 6180 // check if it's time to yield 6181 do_yield_check(); 6182 } 6183 unsigned int after_count = 6184 CMSHeap::heap()->total_collections(); 6185 bool abort = (_before_count != after_count) || 6186 _collector->should_abort_preclean(); 6187 return abort ? 0 : size; 6188 } 6189 6190 void SurvivorSpacePrecleanClosure::do_yield_work() { 6191 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6192 "CMS thread should hold CMS token"); 6193 assert_lock_strong(_bit_map->lock()); 6194 // Relinquish the bit map lock 6195 _bit_map->lock()->unlock(); 6196 ConcurrentMarkSweepThread::desynchronize(true); 6197 _collector->stopTimer(); 6198 _collector->incrementYields(); 6199 6200 // See the comment in coordinator_yield() 6201 for (unsigned i = 0; i < CMSYieldSleepCount && 6202 ConcurrentMarkSweepThread::should_yield() && 6203 !CMSCollector::foregroundGCIsActive(); ++i) { 6204 os::sleep(Thread::current(), 1, false); 6205 } 6206 6207 ConcurrentMarkSweepThread::synchronize(true); 6208 _bit_map->lock()->lock_without_safepoint_check(); 6209 _collector->startTimer(); 6210 } 6211 6212 // This closure is used to rescan the marked objects on the dirty cards 6213 // in the mod union table and the card table proper. In the parallel 6214 // case, although the bitMap is shared, we do a single read so the 6215 // isMarked() query is "safe". 6216 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6217 // Ignore mark word because we are running concurrent with mutators 6218 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6219 HeapWord* addr = (HeapWord*)p; 6220 assert(_span.contains(addr), "we are scanning the CMS generation"); 6221 bool is_obj_array = false; 6222 #ifdef ASSERT 6223 if (!_parallel) { 6224 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6225 assert(_collector->overflow_list_is_empty(), 6226 "overflow list should be empty"); 6227 6228 } 6229 #endif // ASSERT 6230 if (_bit_map->isMarked(addr)) { 6231 // Obj arrays are precisely marked, non-arrays are not; 6232 // so we scan objArrays precisely and non-arrays in their 6233 // entirety. 6234 if (p->is_objArray()) { 6235 is_obj_array = true; 6236 if (_parallel) { 6237 p->oop_iterate(_par_scan_closure, mr); 6238 } else { 6239 p->oop_iterate(_scan_closure, mr); 6240 } 6241 } else { 6242 if (_parallel) { 6243 p->oop_iterate(_par_scan_closure); 6244 } else { 6245 p->oop_iterate(_scan_closure); 6246 } 6247 } 6248 } 6249 #ifdef ASSERT 6250 if (!_parallel) { 6251 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6252 assert(_collector->overflow_list_is_empty(), 6253 "overflow list should be empty"); 6254 6255 } 6256 #endif // ASSERT 6257 return is_obj_array; 6258 } 6259 6260 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6261 MemRegion span, 6262 CMSBitMap* bitMap, CMSMarkStack* markStack, 6263 bool should_yield, bool verifying): 6264 _collector(collector), 6265 _span(span), 6266 _bitMap(bitMap), 6267 _mut(&collector->_modUnionTable), 6268 _markStack(markStack), 6269 _yield(should_yield), 6270 _skipBits(0) 6271 { 6272 assert(_markStack->isEmpty(), "stack should be empty"); 6273 _finger = _bitMap->startWord(); 6274 _threshold = _finger; 6275 assert(_collector->_restart_addr == NULL, "Sanity check"); 6276 assert(_span.contains(_finger), "Out of bounds _finger?"); 6277 DEBUG_ONLY(_verifying = verifying;) 6278 } 6279 6280 void MarkFromRootsClosure::reset(HeapWord* addr) { 6281 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6282 assert(_span.contains(addr), "Out of bounds _finger?"); 6283 _finger = addr; 6284 _threshold = align_up(_finger, CardTable::card_size); 6285 } 6286 6287 // Should revisit to see if this should be restructured for 6288 // greater efficiency. 6289 bool MarkFromRootsClosure::do_bit(size_t offset) { 6290 if (_skipBits > 0) { 6291 _skipBits--; 6292 return true; 6293 } 6294 // convert offset into a HeapWord* 6295 HeapWord* addr = _bitMap->startWord() + offset; 6296 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6297 "address out of range"); 6298 assert(_bitMap->isMarked(addr), "tautology"); 6299 if (_bitMap->isMarked(addr+1)) { 6300 // this is an allocated but not yet initialized object 6301 assert(_skipBits == 0, "tautology"); 6302 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6303 oop p = oop(addr); 6304 if (p->klass_or_null_acquire() == NULL) { 6305 DEBUG_ONLY(if (!_verifying) {) 6306 // We re-dirty the cards on which this object lies and increase 6307 // the _threshold so that we'll come back to scan this object 6308 // during the preclean or remark phase. (CMSCleanOnEnter) 6309 if (CMSCleanOnEnter) { 6310 size_t sz = _collector->block_size_using_printezis_bits(addr); 6311 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6312 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6313 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6314 // Bump _threshold to end_card_addr; note that 6315 // _threshold cannot possibly exceed end_card_addr, anyhow. 6316 // This prevents future clearing of the card as the scan proceeds 6317 // to the right. 6318 assert(_threshold <= end_card_addr, 6319 "Because we are just scanning into this object"); 6320 if (_threshold < end_card_addr) { 6321 _threshold = end_card_addr; 6322 } 6323 if (p->klass_or_null_acquire() != NULL) { 6324 // Redirty the range of cards... 6325 _mut->mark_range(redirty_range); 6326 } // ...else the setting of klass will dirty the card anyway. 6327 } 6328 DEBUG_ONLY(}) 6329 return true; 6330 } 6331 } 6332 scanOopsInOop(addr); 6333 return true; 6334 } 6335 6336 // We take a break if we've been at this for a while, 6337 // so as to avoid monopolizing the locks involved. 6338 void MarkFromRootsClosure::do_yield_work() { 6339 // First give up the locks, then yield, then re-lock 6340 // We should probably use a constructor/destructor idiom to 6341 // do this unlock/lock or modify the MutexUnlocker class to 6342 // serve our purpose. XXX 6343 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6344 "CMS thread should hold CMS token"); 6345 assert_lock_strong(_bitMap->lock()); 6346 _bitMap->lock()->unlock(); 6347 ConcurrentMarkSweepThread::desynchronize(true); 6348 _collector->stopTimer(); 6349 _collector->incrementYields(); 6350 6351 // See the comment in coordinator_yield() 6352 for (unsigned i = 0; i < CMSYieldSleepCount && 6353 ConcurrentMarkSweepThread::should_yield() && 6354 !CMSCollector::foregroundGCIsActive(); ++i) { 6355 os::sleep(Thread::current(), 1, false); 6356 } 6357 6358 ConcurrentMarkSweepThread::synchronize(true); 6359 _bitMap->lock()->lock_without_safepoint_check(); 6360 _collector->startTimer(); 6361 } 6362 6363 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6364 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6365 assert(_markStack->isEmpty(), 6366 "should drain stack to limit stack usage"); 6367 // convert ptr to an oop preparatory to scanning 6368 oop obj = oop(ptr); 6369 // Ignore mark word in verification below, since we 6370 // may be running concurrent with mutators. 6371 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6372 assert(_finger <= ptr, "_finger runneth ahead"); 6373 // advance the finger to right end of this object 6374 _finger = ptr + obj->size(); 6375 assert(_finger > ptr, "we just incremented it above"); 6376 // On large heaps, it may take us some time to get through 6377 // the marking phase. During 6378 // this time it's possible that a lot of mutations have 6379 // accumulated in the card table and the mod union table -- 6380 // these mutation records are redundant until we have 6381 // actually traced into the corresponding card. 6382 // Here, we check whether advancing the finger would make 6383 // us cross into a new card, and if so clear corresponding 6384 // cards in the MUT (preclean them in the card-table in the 6385 // future). 6386 6387 DEBUG_ONLY(if (!_verifying) {) 6388 // The clean-on-enter optimization is disabled by default, 6389 // until we fix 6178663. 6390 if (CMSCleanOnEnter && (_finger > _threshold)) { 6391 // [_threshold, _finger) represents the interval 6392 // of cards to be cleared in MUT (or precleaned in card table). 6393 // The set of cards to be cleared is all those that overlap 6394 // with the interval [_threshold, _finger); note that 6395 // _threshold is always kept card-aligned but _finger isn't 6396 // always card-aligned. 6397 HeapWord* old_threshold = _threshold; 6398 assert(is_aligned(old_threshold, CardTable::card_size), 6399 "_threshold should always be card-aligned"); 6400 _threshold = align_up(_finger, CardTable::card_size); 6401 MemRegion mr(old_threshold, _threshold); 6402 assert(!mr.is_empty(), "Control point invariant"); 6403 assert(_span.contains(mr), "Should clear within span"); 6404 _mut->clear_range(mr); 6405 } 6406 DEBUG_ONLY(}) 6407 // Note: the finger doesn't advance while we drain 6408 // the stack below. 6409 PushOrMarkClosure pushOrMarkClosure(_collector, 6410 _span, _bitMap, _markStack, 6411 _finger, this); 6412 bool res = _markStack->push(obj); 6413 assert(res, "Empty non-zero size stack should have space for single push"); 6414 while (!_markStack->isEmpty()) { 6415 oop new_oop = _markStack->pop(); 6416 // Skip verifying header mark word below because we are 6417 // running concurrent with mutators. 6418 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6419 // now scan this oop's oops 6420 new_oop->oop_iterate(&pushOrMarkClosure); 6421 do_yield_check(); 6422 } 6423 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6424 } 6425 6426 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6427 CMSCollector* collector, MemRegion span, 6428 CMSBitMap* bit_map, 6429 OopTaskQueue* work_queue, 6430 CMSMarkStack* overflow_stack): 6431 _collector(collector), 6432 _whole_span(collector->_span), 6433 _span(span), 6434 _bit_map(bit_map), 6435 _mut(&collector->_modUnionTable), 6436 _work_queue(work_queue), 6437 _overflow_stack(overflow_stack), 6438 _skip_bits(0), 6439 _task(task) 6440 { 6441 assert(_work_queue->size() == 0, "work_queue should be empty"); 6442 _finger = span.start(); 6443 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6444 assert(_span.contains(_finger), "Out of bounds _finger?"); 6445 } 6446 6447 // Should revisit to see if this should be restructured for 6448 // greater efficiency. 6449 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6450 if (_skip_bits > 0) { 6451 _skip_bits--; 6452 return true; 6453 } 6454 // convert offset into a HeapWord* 6455 HeapWord* addr = _bit_map->startWord() + offset; 6456 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6457 "address out of range"); 6458 assert(_bit_map->isMarked(addr), "tautology"); 6459 if (_bit_map->isMarked(addr+1)) { 6460 // this is an allocated object that might not yet be initialized 6461 assert(_skip_bits == 0, "tautology"); 6462 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6463 oop p = oop(addr); 6464 if (p->klass_or_null_acquire() == NULL) { 6465 // in the case of Clean-on-Enter optimization, redirty card 6466 // and avoid clearing card by increasing the threshold. 6467 return true; 6468 } 6469 } 6470 scan_oops_in_oop(addr); 6471 return true; 6472 } 6473 6474 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6475 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6476 // Should we assert that our work queue is empty or 6477 // below some drain limit? 6478 assert(_work_queue->size() == 0, 6479 "should drain stack to limit stack usage"); 6480 // convert ptr to an oop preparatory to scanning 6481 oop obj = oop(ptr); 6482 // Ignore mark word in verification below, since we 6483 // may be running concurrent with mutators. 6484 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6485 assert(_finger <= ptr, "_finger runneth ahead"); 6486 // advance the finger to right end of this object 6487 _finger = ptr + obj->size(); 6488 assert(_finger > ptr, "we just incremented it above"); 6489 // On large heaps, it may take us some time to get through 6490 // the marking phase. During 6491 // this time it's possible that a lot of mutations have 6492 // accumulated in the card table and the mod union table -- 6493 // these mutation records are redundant until we have 6494 // actually traced into the corresponding card. 6495 // Here, we check whether advancing the finger would make 6496 // us cross into a new card, and if so clear corresponding 6497 // cards in the MUT (preclean them in the card-table in the 6498 // future). 6499 6500 // The clean-on-enter optimization is disabled by default, 6501 // until we fix 6178663. 6502 if (CMSCleanOnEnter && (_finger > _threshold)) { 6503 // [_threshold, _finger) represents the interval 6504 // of cards to be cleared in MUT (or precleaned in card table). 6505 // The set of cards to be cleared is all those that overlap 6506 // with the interval [_threshold, _finger); note that 6507 // _threshold is always kept card-aligned but _finger isn't 6508 // always card-aligned. 6509 HeapWord* old_threshold = _threshold; 6510 assert(is_aligned(old_threshold, CardTable::card_size), 6511 "_threshold should always be card-aligned"); 6512 _threshold = align_up(_finger, CardTable::card_size); 6513 MemRegion mr(old_threshold, _threshold); 6514 assert(!mr.is_empty(), "Control point invariant"); 6515 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6516 _mut->clear_range(mr); 6517 } 6518 6519 // Note: the local finger doesn't advance while we drain 6520 // the stack below, but the global finger sure can and will. 6521 HeapWord* volatile* gfa = _task->global_finger_addr(); 6522 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6523 _span, _bit_map, 6524 _work_queue, 6525 _overflow_stack, 6526 _finger, 6527 gfa, this); 6528 bool res = _work_queue->push(obj); // overflow could occur here 6529 assert(res, "Will hold once we use workqueues"); 6530 while (true) { 6531 oop new_oop; 6532 if (!_work_queue->pop_local(new_oop)) { 6533 // We emptied our work_queue; check if there's stuff that can 6534 // be gotten from the overflow stack. 6535 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6536 _overflow_stack, _work_queue)) { 6537 do_yield_check(); 6538 continue; 6539 } else { // done 6540 break; 6541 } 6542 } 6543 // Skip verifying header mark word below because we are 6544 // running concurrent with mutators. 6545 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6546 // now scan this oop's oops 6547 new_oop->oop_iterate(&pushOrMarkClosure); 6548 do_yield_check(); 6549 } 6550 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6551 } 6552 6553 // Yield in response to a request from VM Thread or 6554 // from mutators. 6555 void ParMarkFromRootsClosure::do_yield_work() { 6556 assert(_task != NULL, "sanity"); 6557 _task->yield(); 6558 } 6559 6560 // A variant of the above used for verifying CMS marking work. 6561 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6562 MemRegion span, 6563 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6564 CMSMarkStack* mark_stack): 6565 _collector(collector), 6566 _span(span), 6567 _verification_bm(verification_bm), 6568 _cms_bm(cms_bm), 6569 _mark_stack(mark_stack), 6570 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6571 mark_stack) 6572 { 6573 assert(_mark_stack->isEmpty(), "stack should be empty"); 6574 _finger = _verification_bm->startWord(); 6575 assert(_collector->_restart_addr == NULL, "Sanity check"); 6576 assert(_span.contains(_finger), "Out of bounds _finger?"); 6577 } 6578 6579 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6580 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6581 assert(_span.contains(addr), "Out of bounds _finger?"); 6582 _finger = addr; 6583 } 6584 6585 // Should revisit to see if this should be restructured for 6586 // greater efficiency. 6587 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6588 // convert offset into a HeapWord* 6589 HeapWord* addr = _verification_bm->startWord() + offset; 6590 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6591 "address out of range"); 6592 assert(_verification_bm->isMarked(addr), "tautology"); 6593 assert(_cms_bm->isMarked(addr), "tautology"); 6594 6595 assert(_mark_stack->isEmpty(), 6596 "should drain stack to limit stack usage"); 6597 // convert addr to an oop preparatory to scanning 6598 oop obj = oop(addr); 6599 assert(oopDesc::is_oop(obj), "should be an oop"); 6600 assert(_finger <= addr, "_finger runneth ahead"); 6601 // advance the finger to right end of this object 6602 _finger = addr + obj->size(); 6603 assert(_finger > addr, "we just incremented it above"); 6604 // Note: the finger doesn't advance while we drain 6605 // the stack below. 6606 bool res = _mark_stack->push(obj); 6607 assert(res, "Empty non-zero size stack should have space for single push"); 6608 while (!_mark_stack->isEmpty()) { 6609 oop new_oop = _mark_stack->pop(); 6610 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6611 // now scan this oop's oops 6612 new_oop->oop_iterate(&_pam_verify_closure); 6613 } 6614 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6615 return true; 6616 } 6617 6618 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6619 CMSCollector* collector, MemRegion span, 6620 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6621 CMSMarkStack* mark_stack): 6622 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6623 _collector(collector), 6624 _span(span), 6625 _verification_bm(verification_bm), 6626 _cms_bm(cms_bm), 6627 _mark_stack(mark_stack) 6628 { } 6629 6630 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6631 oop obj = RawAccess<>::oop_load(p); 6632 do_oop(obj); 6633 } 6634 6635 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6636 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6637 6638 // Upon stack overflow, we discard (part of) the stack, 6639 // remembering the least address amongst those discarded 6640 // in CMSCollector's _restart_address. 6641 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6642 // Remember the least grey address discarded 6643 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6644 _collector->lower_restart_addr(ra); 6645 _mark_stack->reset(); // discard stack contents 6646 _mark_stack->expand(); // expand the stack if possible 6647 } 6648 6649 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6650 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6651 HeapWord* addr = (HeapWord*)obj; 6652 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6653 // Oop lies in _span and isn't yet grey or black 6654 _verification_bm->mark(addr); // now grey 6655 if (!_cms_bm->isMarked(addr)) { 6656 Log(gc, verify) log; 6657 ResourceMark rm; 6658 LogStream ls(log.error()); 6659 oop(addr)->print_on(&ls); 6660 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6661 fatal("... aborting"); 6662 } 6663 6664 if (!_mark_stack->push(obj)) { // stack overflow 6665 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6666 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6667 handle_stack_overflow(addr); 6668 } 6669 // anything including and to the right of _finger 6670 // will be scanned as we iterate over the remainder of the 6671 // bit map 6672 } 6673 } 6674 6675 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6676 MemRegion span, 6677 CMSBitMap* bitMap, CMSMarkStack* markStack, 6678 HeapWord* finger, MarkFromRootsClosure* parent) : 6679 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6680 _collector(collector), 6681 _span(span), 6682 _bitMap(bitMap), 6683 _markStack(markStack), 6684 _finger(finger), 6685 _parent(parent) 6686 { } 6687 6688 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6689 MemRegion span, 6690 CMSBitMap* bit_map, 6691 OopTaskQueue* work_queue, 6692 CMSMarkStack* overflow_stack, 6693 HeapWord* finger, 6694 HeapWord* volatile* global_finger_addr, 6695 ParMarkFromRootsClosure* parent) : 6696 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6697 _collector(collector), 6698 _whole_span(collector->_span), 6699 _span(span), 6700 _bit_map(bit_map), 6701 _work_queue(work_queue), 6702 _overflow_stack(overflow_stack), 6703 _finger(finger), 6704 _global_finger_addr(global_finger_addr), 6705 _parent(parent) 6706 { } 6707 6708 // Assumes thread-safe access by callers, who are 6709 // responsible for mutual exclusion. 6710 void CMSCollector::lower_restart_addr(HeapWord* low) { 6711 assert(_span.contains(low), "Out of bounds addr"); 6712 if (_restart_addr == NULL) { 6713 _restart_addr = low; 6714 } else { 6715 _restart_addr = MIN2(_restart_addr, low); 6716 } 6717 } 6718 6719 // Upon stack overflow, we discard (part of) the stack, 6720 // remembering the least address amongst those discarded 6721 // in CMSCollector's _restart_address. 6722 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6723 // Remember the least grey address discarded 6724 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6725 _collector->lower_restart_addr(ra); 6726 _markStack->reset(); // discard stack contents 6727 _markStack->expand(); // expand the stack if possible 6728 } 6729 6730 // Upon stack overflow, we discard (part of) the stack, 6731 // remembering the least address amongst those discarded 6732 // in CMSCollector's _restart_address. 6733 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6734 // We need to do this under a mutex to prevent other 6735 // workers from interfering with the work done below. 6736 MutexLocker ml(_overflow_stack->par_lock(), 6737 Mutex::_no_safepoint_check_flag); 6738 // Remember the least grey address discarded 6739 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6740 _collector->lower_restart_addr(ra); 6741 _overflow_stack->reset(); // discard stack contents 6742 _overflow_stack->expand(); // expand the stack if possible 6743 } 6744 6745 void PushOrMarkClosure::do_oop(oop obj) { 6746 // Ignore mark word because we are running concurrent with mutators. 6747 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6748 HeapWord* addr = (HeapWord*)obj; 6749 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6750 // Oop lies in _span and isn't yet grey or black 6751 _bitMap->mark(addr); // now grey 6752 if (addr < _finger) { 6753 // the bit map iteration has already either passed, or 6754 // sampled, this bit in the bit map; we'll need to 6755 // use the marking stack to scan this oop's oops. 6756 bool simulate_overflow = false; 6757 NOT_PRODUCT( 6758 if (CMSMarkStackOverflowALot && 6759 _collector->simulate_overflow()) { 6760 // simulate a stack overflow 6761 simulate_overflow = true; 6762 } 6763 ) 6764 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6765 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6766 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6767 handle_stack_overflow(addr); 6768 } 6769 } 6770 // anything including and to the right of _finger 6771 // will be scanned as we iterate over the remainder of the 6772 // bit map 6773 do_yield_check(); 6774 } 6775 } 6776 6777 void ParPushOrMarkClosure::do_oop(oop obj) { 6778 // Ignore mark word because we are running concurrent with mutators. 6779 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6780 HeapWord* addr = (HeapWord*)obj; 6781 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6782 // Oop lies in _span and isn't yet grey or black 6783 // We read the global_finger (volatile read) strictly after marking oop 6784 bool res = _bit_map->par_mark(addr); // now grey 6785 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6786 // Should we push this marked oop on our stack? 6787 // -- if someone else marked it, nothing to do 6788 // -- if target oop is above global finger nothing to do 6789 // -- if target oop is in chunk and above local finger 6790 // then nothing to do 6791 // -- else push on work queue 6792 if ( !res // someone else marked it, they will deal with it 6793 || (addr >= *gfa) // will be scanned in a later task 6794 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6795 return; 6796 } 6797 // the bit map iteration has already either passed, or 6798 // sampled, this bit in the bit map; we'll need to 6799 // use the marking stack to scan this oop's oops. 6800 bool simulate_overflow = false; 6801 NOT_PRODUCT( 6802 if (CMSMarkStackOverflowALot && 6803 _collector->simulate_overflow()) { 6804 // simulate a stack overflow 6805 simulate_overflow = true; 6806 } 6807 ) 6808 if (simulate_overflow || 6809 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6810 // stack overflow 6811 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6812 // We cannot assert that the overflow stack is full because 6813 // it may have been emptied since. 6814 assert(simulate_overflow || 6815 _work_queue->size() == _work_queue->max_elems(), 6816 "Else push should have succeeded"); 6817 handle_stack_overflow(addr); 6818 } 6819 do_yield_check(); 6820 } 6821 } 6822 6823 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6824 MemRegion span, 6825 ReferenceDiscoverer* rd, 6826 CMSBitMap* bit_map, 6827 CMSBitMap* mod_union_table, 6828 CMSMarkStack* mark_stack, 6829 bool concurrent_precleaning): 6830 MetadataVisitingOopIterateClosure(rd), 6831 _collector(collector), 6832 _span(span), 6833 _bit_map(bit_map), 6834 _mod_union_table(mod_union_table), 6835 _mark_stack(mark_stack), 6836 _concurrent_precleaning(concurrent_precleaning) 6837 { 6838 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6839 } 6840 6841 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6842 // the non-parallel version (the parallel version appears further below.) 6843 void PushAndMarkClosure::do_oop(oop obj) { 6844 // Ignore mark word verification. If during concurrent precleaning, 6845 // the object monitor may be locked. If during the checkpoint 6846 // phases, the object may already have been reached by a different 6847 // path and may be at the end of the global overflow list (so 6848 // the mark word may be NULL). 6849 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6850 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6851 HeapWord* addr = (HeapWord*)obj; 6852 // Check if oop points into the CMS generation 6853 // and is not marked 6854 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6855 // a white object ... 6856 _bit_map->mark(addr); // ... now grey 6857 // push on the marking stack (grey set) 6858 bool simulate_overflow = false; 6859 NOT_PRODUCT( 6860 if (CMSMarkStackOverflowALot && 6861 _collector->simulate_overflow()) { 6862 // simulate a stack overflow 6863 simulate_overflow = true; 6864 } 6865 ) 6866 if (simulate_overflow || !_mark_stack->push(obj)) { 6867 if (_concurrent_precleaning) { 6868 // During precleaning we can just dirty the appropriate card(s) 6869 // in the mod union table, thus ensuring that the object remains 6870 // in the grey set and continue. In the case of object arrays 6871 // we need to dirty all of the cards that the object spans, 6872 // since the rescan of object arrays will be limited to the 6873 // dirty cards. 6874 // Note that no one can be interfering with us in this action 6875 // of dirtying the mod union table, so no locking or atomics 6876 // are required. 6877 if (obj->is_objArray()) { 6878 size_t sz = obj->size(); 6879 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6880 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6881 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6882 _mod_union_table->mark_range(redirty_range); 6883 } else { 6884 _mod_union_table->mark(addr); 6885 } 6886 _collector->_ser_pmc_preclean_ovflw++; 6887 } else { 6888 // During the remark phase, we need to remember this oop 6889 // in the overflow list. 6890 _collector->push_on_overflow_list(obj); 6891 _collector->_ser_pmc_remark_ovflw++; 6892 } 6893 } 6894 } 6895 } 6896 6897 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6898 MemRegion span, 6899 ReferenceDiscoverer* rd, 6900 CMSBitMap* bit_map, 6901 OopTaskQueue* work_queue): 6902 MetadataVisitingOopIterateClosure(rd), 6903 _collector(collector), 6904 _span(span), 6905 _bit_map(bit_map), 6906 _work_queue(work_queue) 6907 { 6908 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6909 } 6910 6911 // Grey object rescan during second checkpoint phase -- 6912 // the parallel version. 6913 void ParPushAndMarkClosure::do_oop(oop obj) { 6914 // In the assert below, we ignore the mark word because 6915 // this oop may point to an already visited object that is 6916 // on the overflow stack (in which case the mark word has 6917 // been hijacked for chaining into the overflow stack -- 6918 // if this is the last object in the overflow stack then 6919 // its mark word will be NULL). Because this object may 6920 // have been subsequently popped off the global overflow 6921 // stack, and the mark word possibly restored to the prototypical 6922 // value, by the time we get to examined this failing assert in 6923 // the debugger, is_oop_or_null(false) may subsequently start 6924 // to hold. 6925 assert(oopDesc::is_oop_or_null(obj, true), 6926 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6927 HeapWord* addr = (HeapWord*)obj; 6928 // Check if oop points into the CMS generation 6929 // and is not marked 6930 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6931 // a white object ... 6932 // If we manage to "claim" the object, by being the 6933 // first thread to mark it, then we push it on our 6934 // marking stack 6935 if (_bit_map->par_mark(addr)) { // ... now grey 6936 // push on work queue (grey set) 6937 bool simulate_overflow = false; 6938 NOT_PRODUCT( 6939 if (CMSMarkStackOverflowALot && 6940 _collector->par_simulate_overflow()) { 6941 // simulate a stack overflow 6942 simulate_overflow = true; 6943 } 6944 ) 6945 if (simulate_overflow || !_work_queue->push(obj)) { 6946 _collector->par_push_on_overflow_list(obj); 6947 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6948 } 6949 } // Else, some other thread got there first 6950 } 6951 } 6952 6953 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6954 Mutex* bml = _collector->bitMapLock(); 6955 assert_lock_strong(bml); 6956 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6957 "CMS thread should hold CMS token"); 6958 6959 bml->unlock(); 6960 ConcurrentMarkSweepThread::desynchronize(true); 6961 6962 _collector->stopTimer(); 6963 _collector->incrementYields(); 6964 6965 // See the comment in coordinator_yield() 6966 for (unsigned i = 0; i < CMSYieldSleepCount && 6967 ConcurrentMarkSweepThread::should_yield() && 6968 !CMSCollector::foregroundGCIsActive(); ++i) { 6969 os::sleep(Thread::current(), 1, false); 6970 } 6971 6972 ConcurrentMarkSweepThread::synchronize(true); 6973 bml->lock(); 6974 6975 _collector->startTimer(); 6976 } 6977 6978 bool CMSPrecleanRefsYieldClosure::should_return() { 6979 if (ConcurrentMarkSweepThread::should_yield()) { 6980 do_yield_work(); 6981 } 6982 return _collector->foregroundGCIsActive(); 6983 } 6984 6985 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6986 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6987 "mr should be aligned to start at a card boundary"); 6988 // We'd like to assert: 6989 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6990 // "mr should be a range of cards"); 6991 // However, that would be too strong in one case -- the last 6992 // partition ends at _unallocated_block which, in general, can be 6993 // an arbitrary boundary, not necessarily card aligned. 6994 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 6995 _space->object_iterate_mem(mr, &_scan_cl); 6996 } 6997 6998 SweepClosure::SweepClosure(CMSCollector* collector, 6999 ConcurrentMarkSweepGeneration* g, 7000 CMSBitMap* bitMap, bool should_yield) : 7001 _collector(collector), 7002 _g(g), 7003 _sp(g->cmsSpace()), 7004 _limit(_sp->sweep_limit()), 7005 _freelistLock(_sp->freelistLock()), 7006 _bitMap(bitMap), 7007 _inFreeRange(false), // No free range at beginning of sweep 7008 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7009 _lastFreeRangeCoalesced(false), 7010 _yield(should_yield), 7011 _freeFinger(g->used_region().start()) 7012 { 7013 NOT_PRODUCT( 7014 _numObjectsFreed = 0; 7015 _numWordsFreed = 0; 7016 _numObjectsLive = 0; 7017 _numWordsLive = 0; 7018 _numObjectsAlreadyFree = 0; 7019 _numWordsAlreadyFree = 0; 7020 _last_fc = NULL; 7021 7022 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7023 _sp->dictionary()->initialize_dict_returned_bytes(); 7024 ) 7025 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7026 "sweep _limit out of bounds"); 7027 log_develop_trace(gc, sweep)("===================="); 7028 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7029 } 7030 7031 void SweepClosure::print_on(outputStream* st) const { 7032 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7033 p2i(_sp->bottom()), p2i(_sp->end())); 7034 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7035 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7036 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7037 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7038 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7039 } 7040 7041 #ifndef PRODUCT 7042 // Assertion checking only: no useful work in product mode -- 7043 // however, if any of the flags below become product flags, 7044 // you may need to review this code to see if it needs to be 7045 // enabled in product mode. 7046 SweepClosure::~SweepClosure() { 7047 assert_lock_strong(_freelistLock); 7048 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7049 "sweep _limit out of bounds"); 7050 if (inFreeRange()) { 7051 Log(gc, sweep) log; 7052 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7053 ResourceMark rm; 7054 LogStream ls(log.error()); 7055 print_on(&ls); 7056 ShouldNotReachHere(); 7057 } 7058 7059 if (log_is_enabled(Debug, gc, sweep)) { 7060 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7061 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7062 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7063 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7064 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7065 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7066 } 7067 7068 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7069 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7070 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7071 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7072 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7073 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7074 } 7075 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7076 log_develop_trace(gc, sweep)("================"); 7077 } 7078 #endif // PRODUCT 7079 7080 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7081 bool freeRangeInFreeLists) { 7082 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7083 p2i(freeFinger), freeRangeInFreeLists); 7084 assert(!inFreeRange(), "Trampling existing free range"); 7085 set_inFreeRange(true); 7086 set_lastFreeRangeCoalesced(false); 7087 7088 set_freeFinger(freeFinger); 7089 set_freeRangeInFreeLists(freeRangeInFreeLists); 7090 if (CMSTestInFreeList) { 7091 if (freeRangeInFreeLists) { 7092 FreeChunk* fc = (FreeChunk*) freeFinger; 7093 assert(fc->is_free(), "A chunk on the free list should be free."); 7094 assert(fc->size() > 0, "Free range should have a size"); 7095 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7096 } 7097 } 7098 } 7099 7100 // Note that the sweeper runs concurrently with mutators. Thus, 7101 // it is possible for direct allocation in this generation to happen 7102 // in the middle of the sweep. Note that the sweeper also coalesces 7103 // contiguous free blocks. Thus, unless the sweeper and the allocator 7104 // synchronize appropriately freshly allocated blocks may get swept up. 7105 // This is accomplished by the sweeper locking the free lists while 7106 // it is sweeping. Thus blocks that are determined to be free are 7107 // indeed free. There is however one additional complication: 7108 // blocks that have been allocated since the final checkpoint and 7109 // mark, will not have been marked and so would be treated as 7110 // unreachable and swept up. To prevent this, the allocator marks 7111 // the bit map when allocating during the sweep phase. This leads, 7112 // however, to a further complication -- objects may have been allocated 7113 // but not yet initialized -- in the sense that the header isn't yet 7114 // installed. The sweeper can not then determine the size of the block 7115 // in order to skip over it. To deal with this case, we use a technique 7116 // (due to Printezis) to encode such uninitialized block sizes in the 7117 // bit map. Since the bit map uses a bit per every HeapWord, but the 7118 // CMS generation has a minimum object size of 3 HeapWords, it follows 7119 // that "normal marks" won't be adjacent in the bit map (there will 7120 // always be at least two 0 bits between successive 1 bits). We make use 7121 // of these "unused" bits to represent uninitialized blocks -- the bit 7122 // corresponding to the start of the uninitialized object and the next 7123 // bit are both set. Finally, a 1 bit marks the end of the object that 7124 // started with the two consecutive 1 bits to indicate its potentially 7125 // uninitialized state. 7126 7127 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7128 FreeChunk* fc = (FreeChunk*)addr; 7129 size_t res; 7130 7131 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7132 // than "addr == _limit" because although _limit was a block boundary when 7133 // we started the sweep, it may no longer be one because heap expansion 7134 // may have caused us to coalesce the block ending at the address _limit 7135 // with a newly expanded chunk (this happens when _limit was set to the 7136 // previous _end of the space), so we may have stepped past _limit: 7137 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7138 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7139 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7140 "sweep _limit out of bounds"); 7141 assert(addr < _sp->end(), "addr out of bounds"); 7142 // Flush any free range we might be holding as a single 7143 // coalesced chunk to the appropriate free list. 7144 if (inFreeRange()) { 7145 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7146 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7147 flush_cur_free_chunk(freeFinger(), 7148 pointer_delta(addr, freeFinger())); 7149 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7150 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7151 lastFreeRangeCoalesced() ? 1 : 0); 7152 } 7153 7154 // help the iterator loop finish 7155 return pointer_delta(_sp->end(), addr); 7156 } 7157 7158 assert(addr < _limit, "sweep invariant"); 7159 // check if we should yield 7160 do_yield_check(addr); 7161 if (fc->is_free()) { 7162 // Chunk that is already free 7163 res = fc->size(); 7164 do_already_free_chunk(fc); 7165 debug_only(_sp->verifyFreeLists()); 7166 // If we flush the chunk at hand in lookahead_and_flush() 7167 // and it's coalesced with a preceding chunk, then the 7168 // process of "mangling" the payload of the coalesced block 7169 // will cause erasure of the size information from the 7170 // (erstwhile) header of all the coalesced blocks but the 7171 // first, so the first disjunct in the assert will not hold 7172 // in that specific case (in which case the second disjunct 7173 // will hold). 7174 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7175 "Otherwise the size info doesn't change at this step"); 7176 NOT_PRODUCT( 7177 _numObjectsAlreadyFree++; 7178 _numWordsAlreadyFree += res; 7179 ) 7180 NOT_PRODUCT(_last_fc = fc;) 7181 } else if (!_bitMap->isMarked(addr)) { 7182 // Chunk is fresh garbage 7183 res = do_garbage_chunk(fc); 7184 debug_only(_sp->verifyFreeLists()); 7185 NOT_PRODUCT( 7186 _numObjectsFreed++; 7187 _numWordsFreed += res; 7188 ) 7189 } else { 7190 // Chunk that is alive. 7191 res = do_live_chunk(fc); 7192 debug_only(_sp->verifyFreeLists()); 7193 NOT_PRODUCT( 7194 _numObjectsLive++; 7195 _numWordsLive += res; 7196 ) 7197 } 7198 return res; 7199 } 7200 7201 // For the smart allocation, record following 7202 // split deaths - a free chunk is removed from its free list because 7203 // it is being split into two or more chunks. 7204 // split birth - a free chunk is being added to its free list because 7205 // a larger free chunk has been split and resulted in this free chunk. 7206 // coal death - a free chunk is being removed from its free list because 7207 // it is being coalesced into a large free chunk. 7208 // coal birth - a free chunk is being added to its free list because 7209 // it was created when two or more free chunks where coalesced into 7210 // this free chunk. 7211 // 7212 // These statistics are used to determine the desired number of free 7213 // chunks of a given size. The desired number is chosen to be relative 7214 // to the end of a CMS sweep. The desired number at the end of a sweep 7215 // is the 7216 // count-at-end-of-previous-sweep (an amount that was enough) 7217 // - count-at-beginning-of-current-sweep (the excess) 7218 // + split-births (gains in this size during interval) 7219 // - split-deaths (demands on this size during interval) 7220 // where the interval is from the end of one sweep to the end of the 7221 // next. 7222 // 7223 // When sweeping the sweeper maintains an accumulated chunk which is 7224 // the chunk that is made up of chunks that have been coalesced. That 7225 // will be termed the left-hand chunk. A new chunk of garbage that 7226 // is being considered for coalescing will be referred to as the 7227 // right-hand chunk. 7228 // 7229 // When making a decision on whether to coalesce a right-hand chunk with 7230 // the current left-hand chunk, the current count vs. the desired count 7231 // of the left-hand chunk is considered. Also if the right-hand chunk 7232 // is near the large chunk at the end of the heap (see 7233 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7234 // left-hand chunk is coalesced. 7235 // 7236 // When making a decision about whether to split a chunk, the desired count 7237 // vs. the current count of the candidate to be split is also considered. 7238 // If the candidate is underpopulated (currently fewer chunks than desired) 7239 // a chunk of an overpopulated (currently more chunks than desired) size may 7240 // be chosen. The "hint" associated with a free list, if non-null, points 7241 // to a free list which may be overpopulated. 7242 // 7243 7244 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7245 const size_t size = fc->size(); 7246 // Chunks that cannot be coalesced are not in the 7247 // free lists. 7248 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7249 assert(_sp->verify_chunk_in_free_list(fc), 7250 "free chunk should be in free lists"); 7251 } 7252 // a chunk that is already free, should not have been 7253 // marked in the bit map 7254 HeapWord* const addr = (HeapWord*) fc; 7255 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7256 // Verify that the bit map has no bits marked between 7257 // addr and purported end of this block. 7258 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7259 7260 // Some chunks cannot be coalesced under any circumstances. 7261 // See the definition of cantCoalesce(). 7262 if (!fc->cantCoalesce()) { 7263 // This chunk can potentially be coalesced. 7264 // All the work is done in 7265 do_post_free_or_garbage_chunk(fc, size); 7266 // Note that if the chunk is not coalescable (the else arm 7267 // below), we unconditionally flush, without needing to do 7268 // a "lookahead," as we do below. 7269 if (inFreeRange()) lookahead_and_flush(fc, size); 7270 } else { 7271 // Code path common to both original and adaptive free lists. 7272 7273 // cant coalesce with previous block; this should be treated 7274 // as the end of a free run if any 7275 if (inFreeRange()) { 7276 // we kicked some butt; time to pick up the garbage 7277 assert(freeFinger() < addr, "freeFinger points too high"); 7278 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7279 } 7280 // else, nothing to do, just continue 7281 } 7282 } 7283 7284 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7285 // This is a chunk of garbage. It is not in any free list. 7286 // Add it to a free list or let it possibly be coalesced into 7287 // a larger chunk. 7288 HeapWord* const addr = (HeapWord*) fc; 7289 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7290 7291 // Verify that the bit map has no bits marked between 7292 // addr and purported end of just dead object. 7293 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7294 do_post_free_or_garbage_chunk(fc, size); 7295 7296 assert(_limit >= addr + size, 7297 "A freshly garbage chunk can't possibly straddle over _limit"); 7298 if (inFreeRange()) lookahead_and_flush(fc, size); 7299 return size; 7300 } 7301 7302 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7303 HeapWord* addr = (HeapWord*) fc; 7304 // The sweeper has just found a live object. Return any accumulated 7305 // left hand chunk to the free lists. 7306 if (inFreeRange()) { 7307 assert(freeFinger() < addr, "freeFinger points too high"); 7308 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7309 } 7310 7311 // This object is live: we'd normally expect this to be 7312 // an oop, and like to assert the following: 7313 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7314 // However, as we commented above, this may be an object whose 7315 // header hasn't yet been initialized. 7316 size_t size; 7317 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7318 if (_bitMap->isMarked(addr + 1)) { 7319 // Determine the size from the bit map, rather than trying to 7320 // compute it from the object header. 7321 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7322 size = pointer_delta(nextOneAddr + 1, addr); 7323 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7324 "alignment problem"); 7325 7326 #ifdef ASSERT 7327 if (oop(addr)->klass_or_null_acquire() != NULL) { 7328 // Ignore mark word because we are running concurrent with mutators 7329 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7330 assert(size == 7331 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7332 "P-mark and computed size do not agree"); 7333 } 7334 #endif 7335 7336 } else { 7337 // This should be an initialized object that's alive. 7338 assert(oop(addr)->klass_or_null_acquire() != NULL, 7339 "Should be an initialized object"); 7340 // Ignore mark word because we are running concurrent with mutators 7341 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7342 // Verify that the bit map has no bits marked between 7343 // addr and purported end of this block. 7344 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7345 assert(size >= 3, "Necessary for Printezis marks to work"); 7346 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7347 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7348 } 7349 return size; 7350 } 7351 7352 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7353 size_t chunkSize) { 7354 // do_post_free_or_garbage_chunk() should only be called in the case 7355 // of the adaptive free list allocator. 7356 const bool fcInFreeLists = fc->is_free(); 7357 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7358 if (CMSTestInFreeList && fcInFreeLists) { 7359 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7360 } 7361 7362 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7363 7364 HeapWord* const fc_addr = (HeapWord*) fc; 7365 7366 bool coalesce = false; 7367 const size_t left = pointer_delta(fc_addr, freeFinger()); 7368 const size_t right = chunkSize; 7369 switch (FLSCoalescePolicy) { 7370 // numeric value forms a coalition aggressiveness metric 7371 case 0: { // never coalesce 7372 coalesce = false; 7373 break; 7374 } 7375 case 1: { // coalesce if left & right chunks on overpopulated lists 7376 coalesce = _sp->coalOverPopulated(left) && 7377 _sp->coalOverPopulated(right); 7378 break; 7379 } 7380 case 2: { // coalesce if left chunk on overpopulated list (default) 7381 coalesce = _sp->coalOverPopulated(left); 7382 break; 7383 } 7384 case 3: { // coalesce if left OR right chunk on overpopulated list 7385 coalesce = _sp->coalOverPopulated(left) || 7386 _sp->coalOverPopulated(right); 7387 break; 7388 } 7389 case 4: { // always coalesce 7390 coalesce = true; 7391 break; 7392 } 7393 default: 7394 ShouldNotReachHere(); 7395 } 7396 7397 // Should the current free range be coalesced? 7398 // If the chunk is in a free range and either we decided to coalesce above 7399 // or the chunk is near the large block at the end of the heap 7400 // (isNearLargestChunk() returns true), then coalesce this chunk. 7401 const bool doCoalesce = inFreeRange() 7402 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7403 if (doCoalesce) { 7404 // Coalesce the current free range on the left with the new 7405 // chunk on the right. If either is on a free list, 7406 // it must be removed from the list and stashed in the closure. 7407 if (freeRangeInFreeLists()) { 7408 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7409 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7410 "Size of free range is inconsistent with chunk size."); 7411 if (CMSTestInFreeList) { 7412 assert(_sp->verify_chunk_in_free_list(ffc), 7413 "Chunk is not in free lists"); 7414 } 7415 _sp->coalDeath(ffc->size()); 7416 _sp->removeFreeChunkFromFreeLists(ffc); 7417 set_freeRangeInFreeLists(false); 7418 } 7419 if (fcInFreeLists) { 7420 _sp->coalDeath(chunkSize); 7421 assert(fc->size() == chunkSize, 7422 "The chunk has the wrong size or is not in the free lists"); 7423 _sp->removeFreeChunkFromFreeLists(fc); 7424 } 7425 set_lastFreeRangeCoalesced(true); 7426 print_free_block_coalesced(fc); 7427 } else { // not in a free range and/or should not coalesce 7428 // Return the current free range and start a new one. 7429 if (inFreeRange()) { 7430 // In a free range but cannot coalesce with the right hand chunk. 7431 // Put the current free range into the free lists. 7432 flush_cur_free_chunk(freeFinger(), 7433 pointer_delta(fc_addr, freeFinger())); 7434 } 7435 // Set up for new free range. Pass along whether the right hand 7436 // chunk is in the free lists. 7437 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7438 } 7439 } 7440 7441 // Lookahead flush: 7442 // If we are tracking a free range, and this is the last chunk that 7443 // we'll look at because its end crosses past _limit, we'll preemptively 7444 // flush it along with any free range we may be holding on to. Note that 7445 // this can be the case only for an already free or freshly garbage 7446 // chunk. If this block is an object, it can never straddle 7447 // over _limit. The "straddling" occurs when _limit is set at 7448 // the previous end of the space when this cycle started, and 7449 // a subsequent heap expansion caused the previously co-terminal 7450 // free block to be coalesced with the newly expanded portion, 7451 // thus rendering _limit a non-block-boundary making it dangerous 7452 // for the sweeper to step over and examine. 7453 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7454 assert(inFreeRange(), "Should only be called if currently in a free range."); 7455 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7456 assert(_sp->used_region().contains(eob - 1), 7457 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7458 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7459 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7460 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7461 if (eob >= _limit) { 7462 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7463 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7464 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7465 "[" PTR_FORMAT "," PTR_FORMAT ")", 7466 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7467 // Return the storage we are tracking back into the free lists. 7468 log_develop_trace(gc, sweep)("Flushing ... "); 7469 assert(freeFinger() < eob, "Error"); 7470 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7471 } 7472 } 7473 7474 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7475 assert(inFreeRange(), "Should only be called if currently in a free range."); 7476 assert(size > 0, 7477 "A zero sized chunk cannot be added to the free lists."); 7478 if (!freeRangeInFreeLists()) { 7479 if (CMSTestInFreeList) { 7480 FreeChunk* fc = (FreeChunk*) chunk; 7481 fc->set_size(size); 7482 assert(!_sp->verify_chunk_in_free_list(fc), 7483 "chunk should not be in free lists yet"); 7484 } 7485 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7486 // A new free range is going to be starting. The current 7487 // free range has not been added to the free lists yet or 7488 // was removed so add it back. 7489 // If the current free range was coalesced, then the death 7490 // of the free range was recorded. Record a birth now. 7491 if (lastFreeRangeCoalesced()) { 7492 _sp->coalBirth(size); 7493 } 7494 _sp->addChunkAndRepairOffsetTable(chunk, size, 7495 lastFreeRangeCoalesced()); 7496 } else { 7497 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7498 } 7499 set_inFreeRange(false); 7500 set_freeRangeInFreeLists(false); 7501 } 7502 7503 // We take a break if we've been at this for a while, 7504 // so as to avoid monopolizing the locks involved. 7505 void SweepClosure::do_yield_work(HeapWord* addr) { 7506 // Return current free chunk being used for coalescing (if any) 7507 // to the appropriate freelist. After yielding, the next 7508 // free block encountered will start a coalescing range of 7509 // free blocks. If the next free block is adjacent to the 7510 // chunk just flushed, they will need to wait for the next 7511 // sweep to be coalesced. 7512 if (inFreeRange()) { 7513 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7514 } 7515 7516 // First give up the locks, then yield, then re-lock. 7517 // We should probably use a constructor/destructor idiom to 7518 // do this unlock/lock or modify the MutexUnlocker class to 7519 // serve our purpose. XXX 7520 assert_lock_strong(_bitMap->lock()); 7521 assert_lock_strong(_freelistLock); 7522 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7523 "CMS thread should hold CMS token"); 7524 _bitMap->lock()->unlock(); 7525 _freelistLock->unlock(); 7526 ConcurrentMarkSweepThread::desynchronize(true); 7527 _collector->stopTimer(); 7528 _collector->incrementYields(); 7529 7530 // See the comment in coordinator_yield() 7531 for (unsigned i = 0; i < CMSYieldSleepCount && 7532 ConcurrentMarkSweepThread::should_yield() && 7533 !CMSCollector::foregroundGCIsActive(); ++i) { 7534 os::sleep(Thread::current(), 1, false); 7535 } 7536 7537 ConcurrentMarkSweepThread::synchronize(true); 7538 _freelistLock->lock_without_safepoint_check(); 7539 _bitMap->lock()->lock_without_safepoint_check(); 7540 _collector->startTimer(); 7541 } 7542 7543 #ifndef PRODUCT 7544 // This is actually very useful in a product build if it can 7545 // be called from the debugger. Compile it into the product 7546 // as needed. 7547 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7548 return debug_cms_space->verify_chunk_in_free_list(fc); 7549 } 7550 #endif 7551 7552 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7553 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7554 p2i(fc), fc->size()); 7555 } 7556 7557 // CMSIsAliveClosure 7558 bool CMSIsAliveClosure::do_object_b(oop obj) { 7559 HeapWord* addr = (HeapWord*)obj; 7560 return addr != NULL && 7561 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7562 } 7563 7564 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7565 MemRegion span, 7566 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7567 bool cpc): 7568 _collector(collector), 7569 _span(span), 7570 _mark_stack(mark_stack), 7571 _bit_map(bit_map), 7572 _concurrent_precleaning(cpc) { 7573 assert(!_span.is_empty(), "Empty span could spell trouble"); 7574 } 7575 7576 7577 // CMSKeepAliveClosure: the serial version 7578 void CMSKeepAliveClosure::do_oop(oop obj) { 7579 HeapWord* addr = (HeapWord*)obj; 7580 if (_span.contains(addr) && 7581 !_bit_map->isMarked(addr)) { 7582 _bit_map->mark(addr); 7583 bool simulate_overflow = false; 7584 NOT_PRODUCT( 7585 if (CMSMarkStackOverflowALot && 7586 _collector->simulate_overflow()) { 7587 // simulate a stack overflow 7588 simulate_overflow = true; 7589 } 7590 ) 7591 if (simulate_overflow || !_mark_stack->push(obj)) { 7592 if (_concurrent_precleaning) { 7593 // We dirty the overflown object and let the remark 7594 // phase deal with it. 7595 assert(_collector->overflow_list_is_empty(), "Error"); 7596 // In the case of object arrays, we need to dirty all of 7597 // the cards that the object spans. No locking or atomics 7598 // are needed since no one else can be mutating the mod union 7599 // table. 7600 if (obj->is_objArray()) { 7601 size_t sz = obj->size(); 7602 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7603 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7604 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7605 _collector->_modUnionTable.mark_range(redirty_range); 7606 } else { 7607 _collector->_modUnionTable.mark(addr); 7608 } 7609 _collector->_ser_kac_preclean_ovflw++; 7610 } else { 7611 _collector->push_on_overflow_list(obj); 7612 _collector->_ser_kac_ovflw++; 7613 } 7614 } 7615 } 7616 } 7617 7618 // CMSParKeepAliveClosure: a parallel version of the above. 7619 // The work queues are private to each closure (thread), 7620 // but (may be) available for stealing by other threads. 7621 void CMSParKeepAliveClosure::do_oop(oop obj) { 7622 HeapWord* addr = (HeapWord*)obj; 7623 if (_span.contains(addr) && 7624 !_bit_map->isMarked(addr)) { 7625 // In general, during recursive tracing, several threads 7626 // may be concurrently getting here; the first one to 7627 // "tag" it, claims it. 7628 if (_bit_map->par_mark(addr)) { 7629 bool res = _work_queue->push(obj); 7630 assert(res, "Low water mark should be much less than capacity"); 7631 // Do a recursive trim in the hope that this will keep 7632 // stack usage lower, but leave some oops for potential stealers 7633 trim_queue(_low_water_mark); 7634 } // Else, another thread got there first 7635 } 7636 } 7637 7638 void CMSParKeepAliveClosure::trim_queue(uint max) { 7639 while (_work_queue->size() > max) { 7640 oop new_oop; 7641 if (_work_queue->pop_local(new_oop)) { 7642 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7643 assert(_bit_map->isMarked((HeapWord*)new_oop), 7644 "no white objects on this stack!"); 7645 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7646 // iterate over the oops in this oop, marking and pushing 7647 // the ones in CMS heap (i.e. in _span). 7648 new_oop->oop_iterate(&_mark_and_push); 7649 } 7650 } 7651 } 7652 7653 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7654 CMSCollector* collector, 7655 MemRegion span, CMSBitMap* bit_map, 7656 OopTaskQueue* work_queue): 7657 _collector(collector), 7658 _span(span), 7659 _work_queue(work_queue), 7660 _bit_map(bit_map) { } 7661 7662 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7663 HeapWord* addr = (HeapWord*)obj; 7664 if (_span.contains(addr) && 7665 !_bit_map->isMarked(addr)) { 7666 if (_bit_map->par_mark(addr)) { 7667 bool simulate_overflow = false; 7668 NOT_PRODUCT( 7669 if (CMSMarkStackOverflowALot && 7670 _collector->par_simulate_overflow()) { 7671 // simulate a stack overflow 7672 simulate_overflow = true; 7673 } 7674 ) 7675 if (simulate_overflow || !_work_queue->push(obj)) { 7676 _collector->par_push_on_overflow_list(obj); 7677 _collector->_par_kac_ovflw++; 7678 } 7679 } // Else another thread got there already 7680 } 7681 } 7682 7683 ////////////////////////////////////////////////////////////////// 7684 // CMSExpansionCause ///////////////////////////// 7685 ////////////////////////////////////////////////////////////////// 7686 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7687 switch (cause) { 7688 case _no_expansion: 7689 return "No expansion"; 7690 case _satisfy_free_ratio: 7691 return "Free ratio"; 7692 case _satisfy_promotion: 7693 return "Satisfy promotion"; 7694 case _satisfy_allocation: 7695 return "allocation"; 7696 case _allocate_par_lab: 7697 return "Par LAB"; 7698 case _allocate_par_spooling_space: 7699 return "Par Spooling Space"; 7700 case _adaptive_size_policy: 7701 return "Ergonomics"; 7702 default: 7703 return "unknown"; 7704 } 7705 } 7706 7707 void CMSDrainMarkingStackClosure::do_void() { 7708 // the max number to take from overflow list at a time 7709 const size_t num = _mark_stack->capacity()/4; 7710 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7711 "Overflow list should be NULL during concurrent phases"); 7712 while (!_mark_stack->isEmpty() || 7713 // if stack is empty, check the overflow list 7714 _collector->take_from_overflow_list(num, _mark_stack)) { 7715 oop obj = _mark_stack->pop(); 7716 HeapWord* addr = (HeapWord*)obj; 7717 assert(_span.contains(addr), "Should be within span"); 7718 assert(_bit_map->isMarked(addr), "Should be marked"); 7719 assert(oopDesc::is_oop(obj), "Should be an oop"); 7720 obj->oop_iterate(_keep_alive); 7721 } 7722 } 7723 7724 void CMSParDrainMarkingStackClosure::do_void() { 7725 // drain queue 7726 trim_queue(0); 7727 } 7728 7729 // Trim our work_queue so its length is below max at return 7730 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7731 while (_work_queue->size() > max) { 7732 oop new_oop; 7733 if (_work_queue->pop_local(new_oop)) { 7734 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7735 assert(_bit_map->isMarked((HeapWord*)new_oop), 7736 "no white objects on this stack!"); 7737 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7738 // iterate over the oops in this oop, marking and pushing 7739 // the ones in CMS heap (i.e. in _span). 7740 new_oop->oop_iterate(&_mark_and_push); 7741 } 7742 } 7743 } 7744 7745 //////////////////////////////////////////////////////////////////// 7746 // Support for Marking Stack Overflow list handling and related code 7747 //////////////////////////////////////////////////////////////////// 7748 // Much of the following code is similar in shape and spirit to the 7749 // code used in ParNewGC. We should try and share that code 7750 // as much as possible in the future. 7751 7752 #ifndef PRODUCT 7753 // Debugging support for CMSStackOverflowALot 7754 7755 // It's OK to call this multi-threaded; the worst thing 7756 // that can happen is that we'll get a bunch of closely 7757 // spaced simulated overflows, but that's OK, in fact 7758 // probably good as it would exercise the overflow code 7759 // under contention. 7760 bool CMSCollector::simulate_overflow() { 7761 if (_overflow_counter-- <= 0) { // just being defensive 7762 _overflow_counter = CMSMarkStackOverflowInterval; 7763 return true; 7764 } else { 7765 return false; 7766 } 7767 } 7768 7769 bool CMSCollector::par_simulate_overflow() { 7770 return simulate_overflow(); 7771 } 7772 #endif 7773 7774 // Single-threaded 7775 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7776 assert(stack->isEmpty(), "Expected precondition"); 7777 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7778 size_t i = num; 7779 oop cur = _overflow_list; 7780 const markOop proto = markOopDesc::prototype(); 7781 NOT_PRODUCT(ssize_t n = 0;) 7782 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7783 next = oop(cur->mark_raw()); 7784 cur->set_mark_raw(proto); // until proven otherwise 7785 assert(oopDesc::is_oop(cur), "Should be an oop"); 7786 bool res = stack->push(cur); 7787 assert(res, "Bit off more than can chew?"); 7788 NOT_PRODUCT(n++;) 7789 } 7790 _overflow_list = cur; 7791 #ifndef PRODUCT 7792 assert(_num_par_pushes >= n, "Too many pops?"); 7793 _num_par_pushes -=n; 7794 #endif 7795 return !stack->isEmpty(); 7796 } 7797 7798 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7799 // (MT-safe) Get a prefix of at most "num" from the list. 7800 // The overflow list is chained through the mark word of 7801 // each object in the list. We fetch the entire list, 7802 // break off a prefix of the right size and return the 7803 // remainder. If other threads try to take objects from 7804 // the overflow list at that time, they will wait for 7805 // some time to see if data becomes available. If (and 7806 // only if) another thread places one or more object(s) 7807 // on the global list before we have returned the suffix 7808 // to the global list, we will walk down our local list 7809 // to find its end and append the global list to 7810 // our suffix before returning it. This suffix walk can 7811 // prove to be expensive (quadratic in the amount of traffic) 7812 // when there are many objects in the overflow list and 7813 // there is much producer-consumer contention on the list. 7814 // *NOTE*: The overflow list manipulation code here and 7815 // in ParNewGeneration:: are very similar in shape, 7816 // except that in the ParNew case we use the old (from/eden) 7817 // copy of the object to thread the list via its klass word. 7818 // Because of the common code, if you make any changes in 7819 // the code below, please check the ParNew version to see if 7820 // similar changes might be needed. 7821 // CR 6797058 has been filed to consolidate the common code. 7822 bool CMSCollector::par_take_from_overflow_list(size_t num, 7823 OopTaskQueue* work_q, 7824 int no_of_gc_threads) { 7825 assert(work_q->size() == 0, "First empty local work queue"); 7826 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7827 if (_overflow_list == NULL) { 7828 return false; 7829 } 7830 // Grab the entire list; we'll put back a suffix 7831 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7832 Thread* tid = Thread::current(); 7833 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7834 // set to ParallelGCThreads. 7835 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7836 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7837 // If the list is busy, we spin for a short while, 7838 // sleeping between attempts to get the list. 7839 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7840 os::sleep(tid, sleep_time_millis, false); 7841 if (_overflow_list == NULL) { 7842 // Nothing left to take 7843 return false; 7844 } else if (_overflow_list != BUSY) { 7845 // Try and grab the prefix 7846 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7847 } 7848 } 7849 // If the list was found to be empty, or we spun long 7850 // enough, we give up and return empty-handed. If we leave 7851 // the list in the BUSY state below, it must be the case that 7852 // some other thread holds the overflow list and will set it 7853 // to a non-BUSY state in the future. 7854 if (prefix == NULL || prefix == BUSY) { 7855 // Nothing to take or waited long enough 7856 if (prefix == NULL) { 7857 // Write back the NULL in case we overwrote it with BUSY above 7858 // and it is still the same value. 7859 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7860 } 7861 return false; 7862 } 7863 assert(prefix != NULL && prefix != BUSY, "Error"); 7864 size_t i = num; 7865 oop cur = prefix; 7866 // Walk down the first "num" objects, unless we reach the end. 7867 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7868 if (cur->mark_raw() == NULL) { 7869 // We have "num" or fewer elements in the list, so there 7870 // is nothing to return to the global list. 7871 // Write back the NULL in lieu of the BUSY we wrote 7872 // above, if it is still the same value. 7873 if (_overflow_list == BUSY) { 7874 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7875 } 7876 } else { 7877 // Chop off the suffix and return it to the global list. 7878 assert(cur->mark_raw() != BUSY, "Error"); 7879 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7880 cur->set_mark_raw(NULL); // break off suffix 7881 // It's possible that the list is still in the empty(busy) state 7882 // we left it in a short while ago; in that case we may be 7883 // able to place back the suffix without incurring the cost 7884 // of a walk down the list. 7885 oop observed_overflow_list = _overflow_list; 7886 oop cur_overflow_list = observed_overflow_list; 7887 bool attached = false; 7888 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7889 observed_overflow_list = 7890 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7891 if (cur_overflow_list == observed_overflow_list) { 7892 attached = true; 7893 break; 7894 } else cur_overflow_list = observed_overflow_list; 7895 } 7896 if (!attached) { 7897 // Too bad, someone else sneaked in (at least) an element; we'll need 7898 // to do a splice. Find tail of suffix so we can prepend suffix to global 7899 // list. 7900 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7901 oop suffix_tail = cur; 7902 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7903 "Tautology"); 7904 observed_overflow_list = _overflow_list; 7905 do { 7906 cur_overflow_list = observed_overflow_list; 7907 if (cur_overflow_list != BUSY) { 7908 // Do the splice ... 7909 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7910 } else { // cur_overflow_list == BUSY 7911 suffix_tail->set_mark_raw(NULL); 7912 } 7913 // ... and try to place spliced list back on overflow_list ... 7914 observed_overflow_list = 7915 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7916 } while (cur_overflow_list != observed_overflow_list); 7917 // ... until we have succeeded in doing so. 7918 } 7919 } 7920 7921 // Push the prefix elements on work_q 7922 assert(prefix != NULL, "control point invariant"); 7923 const markOop proto = markOopDesc::prototype(); 7924 oop next; 7925 NOT_PRODUCT(ssize_t n = 0;) 7926 for (cur = prefix; cur != NULL; cur = next) { 7927 next = oop(cur->mark_raw()); 7928 cur->set_mark_raw(proto); // until proven otherwise 7929 assert(oopDesc::is_oop(cur), "Should be an oop"); 7930 bool res = work_q->push(cur); 7931 assert(res, "Bit off more than we can chew?"); 7932 NOT_PRODUCT(n++;) 7933 } 7934 #ifndef PRODUCT 7935 assert(_num_par_pushes >= n, "Too many pops?"); 7936 Atomic::sub(n, &_num_par_pushes); 7937 #endif 7938 return true; 7939 } 7940 7941 // Single-threaded 7942 void CMSCollector::push_on_overflow_list(oop p) { 7943 NOT_PRODUCT(_num_par_pushes++;) 7944 assert(oopDesc::is_oop(p), "Not an oop"); 7945 preserve_mark_if_necessary(p); 7946 p->set_mark_raw((markOop)_overflow_list); 7947 _overflow_list = p; 7948 } 7949 7950 // Multi-threaded; use CAS to prepend to overflow list 7951 void CMSCollector::par_push_on_overflow_list(oop p) { 7952 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7953 assert(oopDesc::is_oop(p), "Not an oop"); 7954 par_preserve_mark_if_necessary(p); 7955 oop observed_overflow_list = _overflow_list; 7956 oop cur_overflow_list; 7957 do { 7958 cur_overflow_list = observed_overflow_list; 7959 if (cur_overflow_list != BUSY) { 7960 p->set_mark_raw(markOop(cur_overflow_list)); 7961 } else { 7962 p->set_mark_raw(NULL); 7963 } 7964 observed_overflow_list = 7965 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7966 } while (cur_overflow_list != observed_overflow_list); 7967 } 7968 #undef BUSY 7969 7970 // Single threaded 7971 // General Note on GrowableArray: pushes may silently fail 7972 // because we are (temporarily) out of C-heap for expanding 7973 // the stack. The problem is quite ubiquitous and affects 7974 // a lot of code in the JVM. The prudent thing for GrowableArray 7975 // to do (for now) is to exit with an error. However, that may 7976 // be too draconian in some cases because the caller may be 7977 // able to recover without much harm. For such cases, we 7978 // should probably introduce a "soft_push" method which returns 7979 // an indication of success or failure with the assumption that 7980 // the caller may be able to recover from a failure; code in 7981 // the VM can then be changed, incrementally, to deal with such 7982 // failures where possible, thus, incrementally hardening the VM 7983 // in such low resource situations. 7984 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7985 _preserved_oop_stack.push(p); 7986 _preserved_mark_stack.push(m); 7987 assert(m == p->mark_raw(), "Mark word changed"); 7988 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7989 "bijection"); 7990 } 7991 7992 // Single threaded 7993 void CMSCollector::preserve_mark_if_necessary(oop p) { 7994 markOop m = p->mark_raw(); 7995 if (m->must_be_preserved(p)) { 7996 preserve_mark_work(p, m); 7997 } 7998 } 7999 8000 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8001 markOop m = p->mark_raw(); 8002 if (m->must_be_preserved(p)) { 8003 MutexLocker x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8004 // Even though we read the mark word without holding 8005 // the lock, we are assured that it will not change 8006 // because we "own" this oop, so no other thread can 8007 // be trying to push it on the overflow list; see 8008 // the assertion in preserve_mark_work() that checks 8009 // that m == p->mark_raw(). 8010 preserve_mark_work(p, m); 8011 } 8012 } 8013 8014 // We should be able to do this multi-threaded, 8015 // a chunk of stack being a task (this is 8016 // correct because each oop only ever appears 8017 // once in the overflow list. However, it's 8018 // not very easy to completely overlap this with 8019 // other operations, so will generally not be done 8020 // until all work's been completed. Because we 8021 // expect the preserved oop stack (set) to be small, 8022 // it's probably fine to do this single-threaded. 8023 // We can explore cleverer concurrent/overlapped/parallel 8024 // processing of preserved marks if we feel the 8025 // need for this in the future. Stack overflow should 8026 // be so rare in practice and, when it happens, its 8027 // effect on performance so great that this will 8028 // likely just be in the noise anyway. 8029 void CMSCollector::restore_preserved_marks_if_any() { 8030 assert(SafepointSynchronize::is_at_safepoint(), 8031 "world should be stopped"); 8032 assert(Thread::current()->is_ConcurrentGC_thread() || 8033 Thread::current()->is_VM_thread(), 8034 "should be single-threaded"); 8035 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8036 "bijection"); 8037 8038 while (!_preserved_oop_stack.is_empty()) { 8039 oop p = _preserved_oop_stack.pop(); 8040 assert(oopDesc::is_oop(p), "Should be an oop"); 8041 assert(_span.contains(p), "oop should be in _span"); 8042 assert(p->mark_raw() == markOopDesc::prototype(), 8043 "Set when taken from overflow list"); 8044 markOop m = _preserved_mark_stack.pop(); 8045 p->set_mark_raw(m); 8046 } 8047 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8048 "stacks were cleared above"); 8049 } 8050 8051 #ifndef PRODUCT 8052 bool CMSCollector::no_preserved_marks() const { 8053 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8054 } 8055 #endif 8056 8057 // Transfer some number of overflown objects to usual marking 8058 // stack. Return true if some objects were transferred. 8059 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8060 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8061 (size_t)ParGCDesiredObjsFromOverflowList); 8062 8063 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8064 assert(_collector->overflow_list_is_empty() || res, 8065 "If list is not empty, we should have taken something"); 8066 assert(!res || !_mark_stack->isEmpty(), 8067 "If we took something, it should now be on our stack"); 8068 return res; 8069 } 8070 8071 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8072 size_t res = _sp->block_size_no_stall(addr, _collector); 8073 if (_sp->block_is_obj(addr)) { 8074 if (_live_bit_map->isMarked(addr)) { 8075 // It can't have been dead in a previous cycle 8076 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8077 } else { 8078 _dead_bit_map->mark(addr); // mark the dead object 8079 } 8080 } 8081 // Could be 0, if the block size could not be computed without stalling. 8082 return res; 8083 } 8084 8085 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8086 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8087 switch (phase) { 8088 case CMSCollector::InitialMarking: 8089 initialize(manager /* GC manager */ , 8090 cause /* cause of the GC */, 8091 true /* allMemoryPoolsAffected */, 8092 true /* recordGCBeginTime */, 8093 true /* recordPreGCUsage */, 8094 false /* recordPeakUsage */, 8095 false /* recordPostGCusage */, 8096 true /* recordAccumulatedGCTime */, 8097 false /* recordGCEndTime */, 8098 false /* countCollection */ ); 8099 break; 8100 8101 case CMSCollector::FinalMarking: 8102 initialize(manager /* GC manager */ , 8103 cause /* cause of the GC */, 8104 true /* allMemoryPoolsAffected */, 8105 false /* recordGCBeginTime */, 8106 false /* recordPreGCUsage */, 8107 false /* recordPeakUsage */, 8108 false /* recordPostGCusage */, 8109 true /* recordAccumulatedGCTime */, 8110 false /* recordGCEndTime */, 8111 false /* countCollection */ ); 8112 break; 8113 8114 case CMSCollector::Sweeping: 8115 initialize(manager /* GC manager */ , 8116 cause /* cause of the GC */, 8117 true /* allMemoryPoolsAffected */, 8118 false /* recordGCBeginTime */, 8119 false /* recordPreGCUsage */, 8120 true /* recordPeakUsage */, 8121 true /* recordPostGCusage */, 8122 false /* recordAccumulatedGCTime */, 8123 true /* recordGCEndTime */, 8124 true /* countCollection */ ); 8125 break; 8126 8127 default: 8128 ShouldNotReachHere(); 8129 } 8130 }