1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/cms/cmsCollectorPolicy.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/compactibleFreeListSpace.hpp" 33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 34 #include "gc/cms/concurrentMarkSweepThread.hpp" 35 #include "gc/cms/parNewGeneration.hpp" 36 #include "gc/cms/vmCMSOperations.hpp" 37 #include "gc/serial/genMarkSweep.hpp" 38 #include "gc/serial/tenuredGeneration.hpp" 39 #include "gc/shared/adaptiveSizePolicy.hpp" 40 #include "gc/shared/cardGeneration.inline.hpp" 41 #include "gc/shared/cardTableRS.hpp" 42 #include "gc/shared/collectedHeap.inline.hpp" 43 #include "gc/shared/collectorCounters.hpp" 44 #include "gc/shared/collectorPolicy.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/referencePolicy.hpp" 54 #include "gc/shared/strongRootsScope.hpp" 55 #include "gc/shared/taskqueue.inline.hpp" 56 #include "memory/allocation.hpp" 57 #include "memory/iterator.inline.hpp" 58 #include "memory/padded.hpp" 59 #include "memory/resourceArea.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "prims/jvmtiExport.hpp" 62 #include "runtime/atomic.inline.hpp" 63 #include "runtime/globals_extension.hpp" 64 #include "runtime/handles.inline.hpp" 65 #include "runtime/java.hpp" 66 #include "runtime/orderAccess.inline.hpp" 67 #include "runtime/vmThread.hpp" 68 #include "services/memoryService.hpp" 69 #include "services/runtimeService.hpp" 70 #include "utilities/stack.inline.hpp" 71 72 // statics 73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 74 bool CMSCollector::_full_gc_requested = false; 75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 76 77 ////////////////////////////////////////////////////////////////// 78 // In support of CMS/VM thread synchronization 79 ////////////////////////////////////////////////////////////////// 80 // We split use of the CGC_lock into 2 "levels". 81 // The low-level locking is of the usual CGC_lock monitor. We introduce 82 // a higher level "token" (hereafter "CMS token") built on top of the 83 // low level monitor (hereafter "CGC lock"). 84 // The token-passing protocol gives priority to the VM thread. The 85 // CMS-lock doesn't provide any fairness guarantees, but clients 86 // should ensure that it is only held for very short, bounded 87 // durations. 88 // 89 // When either of the CMS thread or the VM thread is involved in 90 // collection operations during which it does not want the other 91 // thread to interfere, it obtains the CMS token. 92 // 93 // If either thread tries to get the token while the other has 94 // it, that thread waits. However, if the VM thread and CMS thread 95 // both want the token, then the VM thread gets priority while the 96 // CMS thread waits. This ensures, for instance, that the "concurrent" 97 // phases of the CMS thread's work do not block out the VM thread 98 // for long periods of time as the CMS thread continues to hog 99 // the token. (See bug 4616232). 100 // 101 // The baton-passing functions are, however, controlled by the 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 103 // and here the low-level CMS lock, not the high level token, 104 // ensures mutual exclusion. 105 // 106 // Two important conditions that we have to satisfy: 107 // 1. if a thread does a low-level wait on the CMS lock, then it 108 // relinquishes the CMS token if it were holding that token 109 // when it acquired the low-level CMS lock. 110 // 2. any low-level notifications on the low-level lock 111 // should only be sent when a thread has relinquished the token. 112 // 113 // In the absence of either property, we'd have potential deadlock. 114 // 115 // We protect each of the CMS (concurrent and sequential) phases 116 // with the CMS _token_, not the CMS _lock_. 117 // 118 // The only code protected by CMS lock is the token acquisition code 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 120 // baton-passing code. 121 // 122 // Unfortunately, i couldn't come up with a good abstraction to factor and 123 // hide the naked CGC_lock manipulation in the baton-passing code 124 // further below. That's something we should try to do. Also, the proof 125 // of correctness of this 2-level locking scheme is far from obvious, 126 // and potentially quite slippery. We have an uneasy suspicion, for instance, 127 // that there may be a theoretical possibility of delay/starvation in the 128 // low-level lock/wait/notify scheme used for the baton-passing because of 129 // potential interference with the priority scheme embodied in the 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 131 // invocation further below and marked with "XXX 20011219YSR". 132 // Indeed, as we note elsewhere, this may become yet more slippery 133 // in the presence of multiple CMS and/or multiple VM threads. XXX 134 135 class CMSTokenSync: public StackObj { 136 private: 137 bool _is_cms_thread; 138 public: 139 CMSTokenSync(bool is_cms_thread): 140 _is_cms_thread(is_cms_thread) { 141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 142 "Incorrect argument to constructor"); 143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 144 } 145 146 ~CMSTokenSync() { 147 assert(_is_cms_thread ? 148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 150 "Incorrect state"); 151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 152 } 153 }; 154 155 // Convenience class that does a CMSTokenSync, and then acquires 156 // upto three locks. 157 class CMSTokenSyncWithLocks: public CMSTokenSync { 158 private: 159 // Note: locks are acquired in textual declaration order 160 // and released in the opposite order 161 MutexLockerEx _locker1, _locker2, _locker3; 162 public: 163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 165 CMSTokenSync(is_cms_thread), 166 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 167 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 168 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 169 { } 170 }; 171 172 173 ////////////////////////////////////////////////////////////////// 174 // Concurrent Mark-Sweep Generation ///////////////////////////// 175 ////////////////////////////////////////////////////////////////// 176 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 178 179 // This struct contains per-thread things necessary to support parallel 180 // young-gen collection. 181 class CMSParGCThreadState: public CHeapObj<mtGC> { 182 public: 183 CFLS_LAB lab; 184 PromotionInfo promo; 185 186 // Constructor. 187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 188 promo.setSpace(cfls); 189 } 190 }; 191 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 193 ReservedSpace rs, size_t initial_byte_size, int level, 194 CardTableRS* ct, bool use_adaptive_freelists, 195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 196 CardGeneration(rs, initial_byte_size, level, ct), 197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 198 _did_compact(false) 199 { 200 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 201 HeapWord* end = (HeapWord*) _virtual_space.high(); 202 203 _direct_allocated_words = 0; 204 NOT_PRODUCT( 205 _numObjectsPromoted = 0; 206 _numWordsPromoted = 0; 207 _numObjectsAllocated = 0; 208 _numWordsAllocated = 0; 209 ) 210 211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 212 use_adaptive_freelists, 213 dictionaryChoice); 214 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 215 _cmsSpace->_gen = this; 216 217 _gc_stats = new CMSGCStats(); 218 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 220 // offsets match. The ability to tell free chunks from objects 221 // depends on this property. 222 debug_only( 223 FreeChunk* junk = NULL; 224 assert(UseCompressedClassPointers || 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 226 "Offset of FreeChunk::_prev within FreeChunk must match" 227 " that of OopDesc::_klass within OopDesc"); 228 ) 229 230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 231 for (uint i = 0; i < ParallelGCThreads; i++) { 232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 233 } 234 235 _incremental_collection_failed = false; 236 // The "dilatation_factor" is the expansion that can occur on 237 // account of the fact that the minimum object size in the CMS 238 // generation may be larger than that in, say, a contiguous young 239 // generation. 240 // Ideally, in the calculation below, we'd compute the dilatation 241 // factor as: MinChunkSize/(promoting_gen's min object size) 242 // Since we do not have such a general query interface for the 243 // promoting generation, we'll instead just use the minimum 244 // object size (which today is a header's worth of space); 245 // note that all arithmetic is in units of HeapWords. 246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 247 assert(_dilatation_factor >= 1.0, "from previous assert"); 248 } 249 250 251 // The field "_initiating_occupancy" represents the occupancy percentage 252 // at which we trigger a new collection cycle. Unless explicitly specified 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it 254 // is calculated by: 255 // 256 // Let "f" be MinHeapFreeRatio in 257 // 258 // _initiating_occupancy = 100-f + 259 // f * (CMSTriggerRatio/100) 260 // where CMSTriggerRatio is the argument "tr" below. 261 // 262 // That is, if we assume the heap is at its desired maximum occupancy at the 263 // end of a collection, we let CMSTriggerRatio of the (purported) free 264 // space be allocated before initiating a new collection cycle. 265 // 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 267 assert(io <= 100 && tr <= 100, "Check the arguments"); 268 if (io >= 0) { 269 _initiating_occupancy = (double)io / 100.0; 270 } else { 271 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 272 (double)(tr * MinHeapFreeRatio) / 100.0) 273 / 100.0; 274 } 275 } 276 277 void ConcurrentMarkSweepGeneration::ref_processor_init() { 278 assert(collector() != NULL, "no collector"); 279 collector()->ref_processor_init(); 280 } 281 282 void CMSCollector::ref_processor_init() { 283 if (_ref_processor == NULL) { 284 // Allocate and initialize a reference processor 285 _ref_processor = 286 new ReferenceProcessor(_span, // span 287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 288 (int) ParallelGCThreads, // mt processing degree 289 _cmsGen->refs_discovery_is_mt(), // mt discovery 290 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 292 &_is_alive_closure); // closure for liveness info 293 // Initialize the _ref_processor field of CMSGen 294 _cmsGen->set_ref_processor(_ref_processor); 295 296 } 297 } 298 299 AdaptiveSizePolicy* CMSCollector::size_policy() { 300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 301 return gch->gen_policy()->size_policy(); 302 } 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 308 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next minor collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 if (PrintGCDetails && Verbose) { 375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 377 cms_free, expected_promotion); 378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 379 cms_free_dbl, cms_consumption_rate() + 1.0); 380 } 381 // Add 1 in case the consumption rate goes to zero. 382 return cms_free_dbl / (cms_consumption_rate() + 1.0); 383 } 384 return 0.0; 385 } 386 387 // Compare the duration of the cms collection to the 388 // time remaining before the cms generation is empty. 389 // Note that the time from the start of the cms collection 390 // to the start of the cms sweep (less than the total 391 // duration of the cms collection) can be used. This 392 // has been tried and some applications experienced 393 // promotion failures early in execution. This was 394 // possibly because the averages were not accurate 395 // enough at the beginning. 396 double CMSStats::time_until_cms_start() const { 397 // We add "gc0_period" to the "work" calculation 398 // below because this query is done (mostly) at the 399 // end of a scavenge, so we need to conservatively 400 // account for that much possible delay 401 // in the query so as to avoid concurrent mode failures 402 // due to starting the collection just a wee bit too 403 // late. 404 double work = cms_duration() + gc0_period(); 405 double deadline = time_until_cms_gen_full(); 406 // If a concurrent mode failure occurred recently, we want to be 407 // more conservative and halve our expected time_until_cms_gen_full() 408 if (work > deadline) { 409 if (Verbose && PrintGCDetails) { 410 gclog_or_tty->print( 411 " CMSCollector: collect because of anticipated promotion " 412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 413 gc0_period(), time_until_cms_gen_full()); 414 } 415 return 0.0; 416 } 417 return work - deadline; 418 } 419 420 #ifndef PRODUCT 421 void CMSStats::print_on(outputStream *st) const { 422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 424 gc0_duration(), gc0_period(), gc0_promoted()); 425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 426 cms_duration(), cms_period(), cms_allocated()); 427 st->print(",cms_since_beg=%g,cms_since_end=%g", 428 cms_time_since_begin(), cms_time_since_end()); 429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 431 432 if (valid()) { 433 st->print(",promo_rate=%g,cms_alloc_rate=%g", 434 promotion_rate(), cms_allocation_rate()); 435 st->print(",cms_consumption_rate=%g,time_until_full=%g", 436 cms_consumption_rate(), time_until_cms_gen_full()); 437 } 438 st->print(" "); 439 } 440 #endif // #ifndef PRODUCT 441 442 CMSCollector::CollectorState CMSCollector::_collectorState = 443 CMSCollector::Idling; 444 bool CMSCollector::_foregroundGCIsActive = false; 445 bool CMSCollector::_foregroundGCShouldWait = false; 446 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 448 CardTableRS* ct, 449 ConcurrentMarkSweepPolicy* cp): 450 _cmsGen(cmsGen), 451 _ct(ct), 452 _ref_processor(NULL), // will be set later 453 _conc_workers(NULL), // may be set later 454 _abort_preclean(false), 455 _start_sampling(false), 456 _between_prologue_and_epilogue(false), 457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 458 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 459 -1 /* lock-free */, "No_lock" /* dummy */), 460 _modUnionClosurePar(&_modUnionTable), 461 // Adjust my span to cover old (cms) gen 462 _span(cmsGen->reserved()), 463 // Construct the is_alive_closure with _span & markBitMap 464 _is_alive_closure(_span, &_markBitMap), 465 _restart_addr(NULL), 466 _overflow_list(NULL), 467 _stats(cmsGen), 468 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 469 //verify that this lock should be acquired with safepoint check. 470 Monitor::_safepoint_check_sometimes)), 471 _eden_chunk_array(NULL), // may be set in ctor body 472 _eden_chunk_capacity(0), // -- ditto -- 473 _eden_chunk_index(0), // -- ditto -- 474 _survivor_plab_array(NULL), // -- ditto -- 475 _survivor_chunk_array(NULL), // -- ditto -- 476 _survivor_chunk_capacity(0), // -- ditto -- 477 _survivor_chunk_index(0), // -- ditto -- 478 _ser_pmc_preclean_ovflw(0), 479 _ser_kac_preclean_ovflw(0), 480 _ser_pmc_remark_ovflw(0), 481 _par_pmc_remark_ovflw(0), 482 _ser_kac_ovflw(0), 483 _par_kac_ovflw(0), 484 #ifndef PRODUCT 485 _num_par_pushes(0), 486 #endif 487 _collection_count_start(0), 488 _verifying(false), 489 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 490 _completed_initialization(false), 491 _collector_policy(cp), 492 _should_unload_classes(CMSClassUnloadingEnabled), 493 _concurrent_cycles_since_last_unload(0), 494 _roots_scanning_options(GenCollectedHeap::SO_None), 495 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 498 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 499 _cms_start_registered(false) 500 { 501 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 502 ExplicitGCInvokesConcurrent = true; 503 } 504 // Now expand the span and allocate the collection support structures 505 // (MUT, marking bit map etc.) to cover both generations subject to 506 // collection. 507 508 // For use by dirty card to oop closures. 509 _cmsGen->cmsSpace()->set_collector(this); 510 511 // Allocate MUT and marking bit map 512 { 513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 514 if (!_markBitMap.allocate(_span)) { 515 warning("Failed to allocate CMS Bit Map"); 516 return; 517 } 518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 519 } 520 { 521 _modUnionTable.allocate(_span); 522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 523 } 524 525 if (!_markStack.allocate(MarkStackSize)) { 526 warning("Failed to allocate CMS Marking Stack"); 527 return; 528 } 529 530 // Support for multi-threaded concurrent phases 531 if (CMSConcurrentMTEnabled) { 532 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 533 // just for now 534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 535 } 536 if (ConcGCThreads > 1) { 537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 538 ConcGCThreads, true); 539 if (_conc_workers == NULL) { 540 warning("GC/CMS: _conc_workers allocation failure: " 541 "forcing -CMSConcurrentMTEnabled"); 542 CMSConcurrentMTEnabled = false; 543 } else { 544 _conc_workers->initialize_workers(); 545 } 546 } else { 547 CMSConcurrentMTEnabled = false; 548 } 549 } 550 if (!CMSConcurrentMTEnabled) { 551 ConcGCThreads = 0; 552 } else { 553 // Turn off CMSCleanOnEnter optimization temporarily for 554 // the MT case where it's not fixed yet; see 6178663. 555 CMSCleanOnEnter = false; 556 } 557 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 558 "Inconsistency"); 559 560 // Parallel task queues; these are shared for the 561 // concurrent and stop-world phases of CMS, but 562 // are not shared with parallel scavenge (ParNew). 563 { 564 uint i; 565 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 566 567 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 568 || ParallelRefProcEnabled) 569 && num_queues > 0) { 570 _task_queues = new OopTaskQueueSet(num_queues); 571 if (_task_queues == NULL) { 572 warning("task_queues allocation failure."); 573 return; 574 } 575 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 576 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 577 for (i = 0; i < num_queues; i++) { 578 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 579 if (q == NULL) { 580 warning("work_queue allocation failure."); 581 return; 582 } 583 _task_queues->register_queue(i, q); 584 } 585 for (i = 0; i < num_queues; i++) { 586 _task_queues->queue(i)->initialize(); 587 _hash_seed[i] = 17; // copied from ParNew 588 } 589 } 590 } 591 592 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 593 594 // Clip CMSBootstrapOccupancy between 0 and 100. 595 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 596 597 // Now tell CMS generations the identity of their collector 598 ConcurrentMarkSweepGeneration::set_collector(this); 599 600 // Create & start a CMS thread for this CMS collector 601 _cmsThread = ConcurrentMarkSweepThread::start(this); 602 assert(cmsThread() != NULL, "CMS Thread should have been created"); 603 assert(cmsThread()->collector() == this, 604 "CMS Thread should refer to this gen"); 605 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 606 607 // Support for parallelizing young gen rescan 608 GenCollectedHeap* gch = GenCollectedHeap::heap(); 609 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 610 _young_gen = (ParNewGeneration*)gch->young_gen(); 611 if (gch->supports_inline_contig_alloc()) { 612 _top_addr = gch->top_addr(); 613 _end_addr = gch->end_addr(); 614 assert(_young_gen != NULL, "no _young_gen"); 615 _eden_chunk_index = 0; 616 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 617 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 618 } 619 620 // Support for parallelizing survivor space rescan 621 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 622 const size_t max_plab_samples = 623 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 624 625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 628 _survivor_chunk_capacity = 2*max_plab_samples; 629 for (uint i = 0; i < ParallelGCThreads; i++) { 630 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 631 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 632 assert(cur->end() == 0, "Should be 0"); 633 assert(cur->array() == vec, "Should be vec"); 634 assert(cur->capacity() == max_plab_samples, "Error"); 635 } 636 } 637 638 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 639 _gc_counters = new CollectorCounters("CMS", 1); 640 _completed_initialization = true; 641 _inter_sweep_timer.start(); // start of time 642 } 643 644 size_t CMSCollector::plab_sample_minimum_size() { 645 // The default value of MinTLABSize is 2k, but there is 646 // no way to get the default value if the flag has been overridden. 647 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 648 } 649 650 const char* ConcurrentMarkSweepGeneration::name() const { 651 return "concurrent mark-sweep generation"; 652 } 653 void ConcurrentMarkSweepGeneration::update_counters() { 654 if (UsePerfData) { 655 _space_counters->update_all(); 656 _gen_counters->update_all(); 657 } 658 } 659 660 // this is an optimized version of update_counters(). it takes the 661 // used value as a parameter rather than computing it. 662 // 663 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 664 if (UsePerfData) { 665 _space_counters->update_used(used); 666 _space_counters->update_capacity(); 667 _gen_counters->update_all(); 668 } 669 } 670 671 void ConcurrentMarkSweepGeneration::print() const { 672 Generation::print(); 673 cmsSpace()->print(); 674 } 675 676 #ifndef PRODUCT 677 void ConcurrentMarkSweepGeneration::print_statistics() { 678 cmsSpace()->printFLCensus(0); 679 } 680 #endif 681 682 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 683 GenCollectedHeap* gch = GenCollectedHeap::heap(); 684 if (PrintGCDetails) { 685 if (Verbose) { 686 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 687 level(), short_name(), s, used(), capacity()); 688 } else { 689 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 690 level(), short_name(), s, used() / K, capacity() / K); 691 } 692 } 693 if (Verbose) { 694 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 695 gch->used(), gch->capacity()); 696 } else { 697 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 698 gch->used() / K, gch->capacity() / K); 699 } 700 } 701 702 size_t 703 ConcurrentMarkSweepGeneration::contiguous_available() const { 704 // dld proposes an improvement in precision here. If the committed 705 // part of the space ends in a free block we should add that to 706 // uncommitted size in the calculation below. Will make this 707 // change later, staying with the approximation below for the 708 // time being. -- ysr. 709 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 710 } 711 712 size_t 713 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 714 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 715 } 716 717 size_t ConcurrentMarkSweepGeneration::max_available() const { 718 return free() + _virtual_space.uncommitted_size(); 719 } 720 721 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 722 size_t available = max_available(); 723 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 724 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 725 if (Verbose && PrintGCDetails) { 726 gclog_or_tty->print_cr( 727 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 728 "max_promo("SIZE_FORMAT")", 729 res? "":" not", available, res? ">=":"<", 730 av_promo, max_promotion_in_bytes); 731 } 732 return res; 733 } 734 735 // At a promotion failure dump information on block layout in heap 736 // (cms old generation). 737 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 738 if (CMSDumpAtPromotionFailure) { 739 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 740 } 741 } 742 743 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 744 // Clear the promotion information. These pointers can be adjusted 745 // along with all the other pointers into the heap but 746 // compaction is expected to be a rare event with 747 // a heap using cms so don't do it without seeing the need. 748 for (uint i = 0; i < ParallelGCThreads; i++) { 749 _par_gc_thread_states[i]->promo.reset(); 750 } 751 } 752 753 void ConcurrentMarkSweepGeneration::compute_new_size() { 754 assert_locked_or_safepoint(Heap_lock); 755 756 // If incremental collection failed, we just want to expand 757 // to the limit. 758 if (incremental_collection_failed()) { 759 clear_incremental_collection_failed(); 760 grow_to_reserved(); 761 return; 762 } 763 764 // The heap has been compacted but not reset yet. 765 // Any metric such as free() or used() will be incorrect. 766 767 CardGeneration::compute_new_size(); 768 769 // Reset again after a possible resizing 770 if (did_compact()) { 771 cmsSpace()->reset_after_compaction(); 772 } 773 } 774 775 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 776 assert_locked_or_safepoint(Heap_lock); 777 778 // If incremental collection failed, we just want to expand 779 // to the limit. 780 if (incremental_collection_failed()) { 781 clear_incremental_collection_failed(); 782 grow_to_reserved(); 783 return; 784 } 785 786 double free_percentage = ((double) free()) / capacity(); 787 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 788 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 789 790 // compute expansion delta needed for reaching desired free percentage 791 if (free_percentage < desired_free_percentage) { 792 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 793 assert(desired_capacity >= capacity(), "invalid expansion size"); 794 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 795 if (PrintGCDetails && Verbose) { 796 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 797 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 798 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 799 gclog_or_tty->print_cr(" Desired free fraction %f", 800 desired_free_percentage); 801 gclog_or_tty->print_cr(" Maximum free fraction %f", 802 maximum_free_percentage); 803 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 804 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 805 desired_capacity/1000); 806 int prev_level = level() - 1; 807 if (prev_level >= 0) { 808 size_t prev_size = 0; 809 GenCollectedHeap* gch = GenCollectedHeap::heap(); 810 Generation* prev_gen = gch->young_gen(); 811 prev_size = prev_gen->capacity(); 812 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 813 prev_size/1000); 814 } 815 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 816 unsafe_max_alloc_nogc()/1000); 817 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 818 contiguous_available()/1000); 819 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 820 expand_bytes); 821 } 822 // safe if expansion fails 823 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 824 if (PrintGCDetails && Verbose) { 825 gclog_or_tty->print_cr(" Expanded free fraction %f", 826 ((double) free()) / capacity()); 827 } 828 } else { 829 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 830 assert(desired_capacity <= capacity(), "invalid expansion size"); 831 size_t shrink_bytes = capacity() - desired_capacity; 832 // Don't shrink unless the delta is greater than the minimum shrink we want 833 if (shrink_bytes >= MinHeapDeltaBytes) { 834 shrink_free_list_by(shrink_bytes); 835 } 836 } 837 } 838 839 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 840 return cmsSpace()->freelistLock(); 841 } 842 843 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 844 bool tlab) { 845 CMSSynchronousYieldRequest yr; 846 MutexLockerEx x(freelistLock(), 847 Mutex::_no_safepoint_check_flag); 848 return have_lock_and_allocate(size, tlab); 849 } 850 851 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 852 bool tlab /* ignored */) { 853 assert_lock_strong(freelistLock()); 854 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 855 HeapWord* res = cmsSpace()->allocate(adjustedSize); 856 // Allocate the object live (grey) if the background collector has 857 // started marking. This is necessary because the marker may 858 // have passed this address and consequently this object will 859 // not otherwise be greyed and would be incorrectly swept up. 860 // Note that if this object contains references, the writing 861 // of those references will dirty the card containing this object 862 // allowing the object to be blackened (and its references scanned) 863 // either during a preclean phase or at the final checkpoint. 864 if (res != NULL) { 865 // We may block here with an uninitialized object with 866 // its mark-bit or P-bits not yet set. Such objects need 867 // to be safely navigable by block_start(). 868 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 869 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 870 collector()->direct_allocated(res, adjustedSize); 871 _direct_allocated_words += adjustedSize; 872 // allocation counters 873 NOT_PRODUCT( 874 _numObjectsAllocated++; 875 _numWordsAllocated += (int)adjustedSize; 876 ) 877 } 878 return res; 879 } 880 881 // In the case of direct allocation by mutators in a generation that 882 // is being concurrently collected, the object must be allocated 883 // live (grey) if the background collector has started marking. 884 // This is necessary because the marker may 885 // have passed this address and consequently this object will 886 // not otherwise be greyed and would be incorrectly swept up. 887 // Note that if this object contains references, the writing 888 // of those references will dirty the card containing this object 889 // allowing the object to be blackened (and its references scanned) 890 // either during a preclean phase or at the final checkpoint. 891 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 892 assert(_markBitMap.covers(start, size), "Out of bounds"); 893 if (_collectorState >= Marking) { 894 MutexLockerEx y(_markBitMap.lock(), 895 Mutex::_no_safepoint_check_flag); 896 // [see comments preceding SweepClosure::do_blk() below for details] 897 // 898 // Can the P-bits be deleted now? JJJ 899 // 900 // 1. need to mark the object as live so it isn't collected 901 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 902 // 3. need to mark the end of the object so marking, precleaning or sweeping 903 // can skip over uninitialized or unparsable objects. An allocated 904 // object is considered uninitialized for our purposes as long as 905 // its klass word is NULL. All old gen objects are parsable 906 // as soon as they are initialized.) 907 _markBitMap.mark(start); // object is live 908 _markBitMap.mark(start + 1); // object is potentially uninitialized? 909 _markBitMap.mark(start + size - 1); 910 // mark end of object 911 } 912 // check that oop looks uninitialized 913 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 914 } 915 916 void CMSCollector::promoted(bool par, HeapWord* start, 917 bool is_obj_array, size_t obj_size) { 918 assert(_markBitMap.covers(start), "Out of bounds"); 919 // See comment in direct_allocated() about when objects should 920 // be allocated live. 921 if (_collectorState >= Marking) { 922 // we already hold the marking bit map lock, taken in 923 // the prologue 924 if (par) { 925 _markBitMap.par_mark(start); 926 } else { 927 _markBitMap.mark(start); 928 } 929 // We don't need to mark the object as uninitialized (as 930 // in direct_allocated above) because this is being done with the 931 // world stopped and the object will be initialized by the 932 // time the marking, precleaning or sweeping get to look at it. 933 // But see the code for copying objects into the CMS generation, 934 // where we need to ensure that concurrent readers of the 935 // block offset table are able to safely navigate a block that 936 // is in flux from being free to being allocated (and in 937 // transition while being copied into) and subsequently 938 // becoming a bona-fide object when the copy/promotion is complete. 939 assert(SafepointSynchronize::is_at_safepoint(), 940 "expect promotion only at safepoints"); 941 942 if (_collectorState < Sweeping) { 943 // Mark the appropriate cards in the modUnionTable, so that 944 // this object gets scanned before the sweep. If this is 945 // not done, CMS generation references in the object might 946 // not get marked. 947 // For the case of arrays, which are otherwise precisely 948 // marked, we need to dirty the entire array, not just its head. 949 if (is_obj_array) { 950 // The [par_]mark_range() method expects mr.end() below to 951 // be aligned to the granularity of a bit's representation 952 // in the heap. In the case of the MUT below, that's a 953 // card size. 954 MemRegion mr(start, 955 (HeapWord*)round_to((intptr_t)(start + obj_size), 956 CardTableModRefBS::card_size /* bytes */)); 957 if (par) { 958 _modUnionTable.par_mark_range(mr); 959 } else { 960 _modUnionTable.mark_range(mr); 961 } 962 } else { // not an obj array; we can just mark the head 963 if (par) { 964 _modUnionTable.par_mark(start); 965 } else { 966 _modUnionTable.mark(start); 967 } 968 } 969 } 970 } 971 } 972 973 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 974 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 975 // allocate, copy and if necessary update promoinfo -- 976 // delegate to underlying space. 977 assert_lock_strong(freelistLock()); 978 979 #ifndef PRODUCT 980 if (GenCollectedHeap::heap()->promotion_should_fail()) { 981 return NULL; 982 } 983 #endif // #ifndef PRODUCT 984 985 oop res = _cmsSpace->promote(obj, obj_size); 986 if (res == NULL) { 987 // expand and retry 988 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 989 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 990 // Since this is the old generation, we don't try to promote 991 // into a more senior generation. 992 res = _cmsSpace->promote(obj, obj_size); 993 } 994 if (res != NULL) { 995 // See comment in allocate() about when objects should 996 // be allocated live. 997 assert(obj->is_oop(), "Will dereference klass pointer below"); 998 collector()->promoted(false, // Not parallel 999 (HeapWord*)res, obj->is_objArray(), obj_size); 1000 // promotion counters 1001 NOT_PRODUCT( 1002 _numObjectsPromoted++; 1003 _numWordsPromoted += 1004 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1005 ) 1006 } 1007 return res; 1008 } 1009 1010 1011 // IMPORTANT: Notes on object size recognition in CMS. 1012 // --------------------------------------------------- 1013 // A block of storage in the CMS generation is always in 1014 // one of three states. A free block (FREE), an allocated 1015 // object (OBJECT) whose size() method reports the correct size, 1016 // and an intermediate state (TRANSIENT) in which its size cannot 1017 // be accurately determined. 1018 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1019 // ----------------------------------------------------- 1020 // FREE: klass_word & 1 == 1; mark_word holds block size 1021 // 1022 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1023 // obj->size() computes correct size 1024 // 1025 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1026 // 1027 // STATE IDENTIFICATION: (64 bit+COOPS) 1028 // ------------------------------------ 1029 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1030 // 1031 // OBJECT: klass_word installed; klass_word != 0; 1032 // obj->size() computes correct size 1033 // 1034 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1035 // 1036 // 1037 // STATE TRANSITION DIAGRAM 1038 // 1039 // mut / parnew mut / parnew 1040 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1041 // ^ | 1042 // |------------------------ DEAD <------------------------------------| 1043 // sweep mut 1044 // 1045 // While a block is in TRANSIENT state its size cannot be determined 1046 // so readers will either need to come back later or stall until 1047 // the size can be determined. Note that for the case of direct 1048 // allocation, P-bits, when available, may be used to determine the 1049 // size of an object that may not yet have been initialized. 1050 1051 // Things to support parallel young-gen collection. 1052 oop 1053 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1054 oop old, markOop m, 1055 size_t word_sz) { 1056 #ifndef PRODUCT 1057 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1058 return NULL; 1059 } 1060 #endif // #ifndef PRODUCT 1061 1062 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1063 PromotionInfo* promoInfo = &ps->promo; 1064 // if we are tracking promotions, then first ensure space for 1065 // promotion (including spooling space for saving header if necessary). 1066 // then allocate and copy, then track promoted info if needed. 1067 // When tracking (see PromotionInfo::track()), the mark word may 1068 // be displaced and in this case restoration of the mark word 1069 // occurs in the (oop_since_save_marks_)iterate phase. 1070 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1071 // Out of space for allocating spooling buffers; 1072 // try expanding and allocating spooling buffers. 1073 if (!expand_and_ensure_spooling_space(promoInfo)) { 1074 return NULL; 1075 } 1076 } 1077 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1078 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1079 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1080 if (obj_ptr == NULL) { 1081 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1082 if (obj_ptr == NULL) { 1083 return NULL; 1084 } 1085 } 1086 oop obj = oop(obj_ptr); 1087 OrderAccess::storestore(); 1088 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1089 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1090 // IMPORTANT: See note on object initialization for CMS above. 1091 // Otherwise, copy the object. Here we must be careful to insert the 1092 // klass pointer last, since this marks the block as an allocated object. 1093 // Except with compressed oops it's the mark word. 1094 HeapWord* old_ptr = (HeapWord*)old; 1095 // Restore the mark word copied above. 1096 obj->set_mark(m); 1097 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1098 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1099 OrderAccess::storestore(); 1100 1101 if (UseCompressedClassPointers) { 1102 // Copy gap missed by (aligned) header size calculation below 1103 obj->set_klass_gap(old->klass_gap()); 1104 } 1105 if (word_sz > (size_t)oopDesc::header_size()) { 1106 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1107 obj_ptr + oopDesc::header_size(), 1108 word_sz - oopDesc::header_size()); 1109 } 1110 1111 // Now we can track the promoted object, if necessary. We take care 1112 // to delay the transition from uninitialized to full object 1113 // (i.e., insertion of klass pointer) until after, so that it 1114 // atomically becomes a promoted object. 1115 if (promoInfo->tracking()) { 1116 promoInfo->track((PromotedObject*)obj, old->klass()); 1117 } 1118 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1119 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1120 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1121 1122 // Finally, install the klass pointer (this should be volatile). 1123 OrderAccess::storestore(); 1124 obj->set_klass(old->klass()); 1125 // We should now be able to calculate the right size for this object 1126 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1127 1128 collector()->promoted(true, // parallel 1129 obj_ptr, old->is_objArray(), word_sz); 1130 1131 NOT_PRODUCT( 1132 Atomic::inc_ptr(&_numObjectsPromoted); 1133 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1134 ) 1135 1136 return obj; 1137 } 1138 1139 void 1140 ConcurrentMarkSweepGeneration:: 1141 par_promote_alloc_done(int thread_num) { 1142 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1143 ps->lab.retire(thread_num); 1144 } 1145 1146 void 1147 ConcurrentMarkSweepGeneration:: 1148 par_oop_since_save_marks_iterate_done(int thread_num) { 1149 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1150 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1151 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1152 } 1153 1154 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1155 size_t size, 1156 bool tlab) 1157 { 1158 // We allow a STW collection only if a full 1159 // collection was requested. 1160 return full || should_allocate(size, tlab); // FIX ME !!! 1161 // This and promotion failure handling are connected at the 1162 // hip and should be fixed by untying them. 1163 } 1164 1165 bool CMSCollector::shouldConcurrentCollect() { 1166 if (_full_gc_requested) { 1167 if (Verbose && PrintGCDetails) { 1168 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1169 " gc request (or gc_locker)"); 1170 } 1171 return true; 1172 } 1173 1174 FreelistLocker x(this); 1175 // ------------------------------------------------------------------ 1176 // Print out lots of information which affects the initiation of 1177 // a collection. 1178 if (PrintCMSInitiationStatistics && stats().valid()) { 1179 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1180 gclog_or_tty->stamp(); 1181 gclog_or_tty->cr(); 1182 stats().print_on(gclog_or_tty); 1183 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1184 stats().time_until_cms_gen_full()); 1185 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1186 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1187 _cmsGen->contiguous_available()); 1188 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1189 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1190 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1191 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1192 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1193 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1194 gclog_or_tty->print_cr("metadata initialized %d", 1195 MetaspaceGC::should_concurrent_collect()); 1196 } 1197 // ------------------------------------------------------------------ 1198 1199 // If the estimated time to complete a cms collection (cms_duration()) 1200 // is less than the estimated time remaining until the cms generation 1201 // is full, start a collection. 1202 if (!UseCMSInitiatingOccupancyOnly) { 1203 if (stats().valid()) { 1204 if (stats().time_until_cms_start() == 0.0) { 1205 return true; 1206 } 1207 } else { 1208 // We want to conservatively collect somewhat early in order 1209 // to try and "bootstrap" our CMS/promotion statistics; 1210 // this branch will not fire after the first successful CMS 1211 // collection because the stats should then be valid. 1212 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1213 if (Verbose && PrintGCDetails) { 1214 gclog_or_tty->print_cr( 1215 " CMSCollector: collect for bootstrapping statistics:" 1216 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1217 _bootstrap_occupancy); 1218 } 1219 return true; 1220 } 1221 } 1222 } 1223 1224 // Otherwise, we start a collection cycle if 1225 // old gen want a collection cycle started. Each may use 1226 // an appropriate criterion for making this decision. 1227 // XXX We need to make sure that the gen expansion 1228 // criterion dovetails well with this. XXX NEED TO FIX THIS 1229 if (_cmsGen->should_concurrent_collect()) { 1230 if (Verbose && PrintGCDetails) { 1231 gclog_or_tty->print_cr("CMS old gen initiated"); 1232 } 1233 return true; 1234 } 1235 1236 // We start a collection if we believe an incremental collection may fail; 1237 // this is not likely to be productive in practice because it's probably too 1238 // late anyway. 1239 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1240 assert(gch->collector_policy()->is_generation_policy(), 1241 "You may want to check the correctness of the following"); 1242 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1243 if (Verbose && PrintGCDetails) { 1244 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1245 } 1246 return true; 1247 } 1248 1249 if (MetaspaceGC::should_concurrent_collect()) { 1250 if (Verbose && PrintGCDetails) { 1251 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1252 } 1253 return true; 1254 } 1255 1256 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1257 if (CMSTriggerInterval >= 0) { 1258 if (CMSTriggerInterval == 0) { 1259 // Trigger always 1260 return true; 1261 } 1262 1263 // Check the CMS time since begin (we do not check the stats validity 1264 // as we want to be able to trigger the first CMS cycle as well) 1265 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1266 if (Verbose && PrintGCDetails) { 1267 if (stats().valid()) { 1268 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1269 stats().cms_time_since_begin()); 1270 } else { 1271 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1272 } 1273 } 1274 return true; 1275 } 1276 } 1277 1278 return false; 1279 } 1280 1281 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1282 1283 // Clear _expansion_cause fields of constituent generations 1284 void CMSCollector::clear_expansion_cause() { 1285 _cmsGen->clear_expansion_cause(); 1286 } 1287 1288 // We should be conservative in starting a collection cycle. To 1289 // start too eagerly runs the risk of collecting too often in the 1290 // extreme. To collect too rarely falls back on full collections, 1291 // which works, even if not optimum in terms of concurrent work. 1292 // As a work around for too eagerly collecting, use the flag 1293 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1294 // giving the user an easily understandable way of controlling the 1295 // collections. 1296 // We want to start a new collection cycle if any of the following 1297 // conditions hold: 1298 // . our current occupancy exceeds the configured initiating occupancy 1299 // for this generation, or 1300 // . we recently needed to expand this space and have not, since that 1301 // expansion, done a collection of this generation, or 1302 // . the underlying space believes that it may be a good idea to initiate 1303 // a concurrent collection (this may be based on criteria such as the 1304 // following: the space uses linear allocation and linear allocation is 1305 // going to fail, or there is believed to be excessive fragmentation in 1306 // the generation, etc... or ... 1307 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1308 // the case of the old generation; see CR 6543076): 1309 // we may be approaching a point at which allocation requests may fail because 1310 // we will be out of sufficient free space given allocation rate estimates.] 1311 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1312 1313 assert_lock_strong(freelistLock()); 1314 if (occupancy() > initiating_occupancy()) { 1315 if (PrintGCDetails && Verbose) { 1316 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1317 short_name(), occupancy(), initiating_occupancy()); 1318 } 1319 return true; 1320 } 1321 if (UseCMSInitiatingOccupancyOnly) { 1322 return false; 1323 } 1324 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1325 if (PrintGCDetails && Verbose) { 1326 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1327 short_name()); 1328 } 1329 return true; 1330 } 1331 if (_cmsSpace->should_concurrent_collect()) { 1332 if (PrintGCDetails && Verbose) { 1333 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1334 short_name()); 1335 } 1336 return true; 1337 } 1338 return false; 1339 } 1340 1341 void ConcurrentMarkSweepGeneration::collect(bool full, 1342 bool clear_all_soft_refs, 1343 size_t size, 1344 bool tlab) 1345 { 1346 collector()->collect(full, clear_all_soft_refs, size, tlab); 1347 } 1348 1349 void CMSCollector::collect(bool full, 1350 bool clear_all_soft_refs, 1351 size_t size, 1352 bool tlab) 1353 { 1354 // The following "if" branch is present for defensive reasons. 1355 // In the current uses of this interface, it can be replaced with: 1356 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1357 // But I am not placing that assert here to allow future 1358 // generality in invoking this interface. 1359 if (GC_locker::is_active()) { 1360 // A consistency test for GC_locker 1361 assert(GC_locker::needs_gc(), "Should have been set already"); 1362 // Skip this foreground collection, instead 1363 // expanding the heap if necessary. 1364 // Need the free list locks for the call to free() in compute_new_size() 1365 compute_new_size(); 1366 return; 1367 } 1368 acquire_control_and_collect(full, clear_all_soft_refs); 1369 } 1370 1371 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1372 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1373 unsigned int gc_count = gch->total_full_collections(); 1374 if (gc_count == full_gc_count) { 1375 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1376 _full_gc_requested = true; 1377 _full_gc_cause = cause; 1378 CGC_lock->notify(); // nudge CMS thread 1379 } else { 1380 assert(gc_count > full_gc_count, "Error: causal loop"); 1381 } 1382 } 1383 1384 bool CMSCollector::is_external_interruption() { 1385 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1386 return GCCause::is_user_requested_gc(cause) || 1387 GCCause::is_serviceability_requested_gc(cause); 1388 } 1389 1390 void CMSCollector::report_concurrent_mode_interruption() { 1391 if (is_external_interruption()) { 1392 if (PrintGCDetails) { 1393 gclog_or_tty->print(" (concurrent mode interrupted)"); 1394 } 1395 } else { 1396 if (PrintGCDetails) { 1397 gclog_or_tty->print(" (concurrent mode failure)"); 1398 } 1399 _gc_tracer_cm->report_concurrent_mode_failure(); 1400 } 1401 } 1402 1403 1404 // The foreground and background collectors need to coordinate in order 1405 // to make sure that they do not mutually interfere with CMS collections. 1406 // When a background collection is active, 1407 // the foreground collector may need to take over (preempt) and 1408 // synchronously complete an ongoing collection. Depending on the 1409 // frequency of the background collections and the heap usage 1410 // of the application, this preemption can be seldom or frequent. 1411 // There are only certain 1412 // points in the background collection that the "collection-baton" 1413 // can be passed to the foreground collector. 1414 // 1415 // The foreground collector will wait for the baton before 1416 // starting any part of the collection. The foreground collector 1417 // will only wait at one location. 1418 // 1419 // The background collector will yield the baton before starting a new 1420 // phase of the collection (e.g., before initial marking, marking from roots, 1421 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1422 // of the loop which switches the phases. The background collector does some 1423 // of the phases (initial mark, final re-mark) with the world stopped. 1424 // Because of locking involved in stopping the world, 1425 // the foreground collector should not block waiting for the background 1426 // collector when it is doing a stop-the-world phase. The background 1427 // collector will yield the baton at an additional point just before 1428 // it enters a stop-the-world phase. Once the world is stopped, the 1429 // background collector checks the phase of the collection. If the 1430 // phase has not changed, it proceeds with the collection. If the 1431 // phase has changed, it skips that phase of the collection. See 1432 // the comments on the use of the Heap_lock in collect_in_background(). 1433 // 1434 // Variable used in baton passing. 1435 // _foregroundGCIsActive - Set to true by the foreground collector when 1436 // it wants the baton. The foreground clears it when it has finished 1437 // the collection. 1438 // _foregroundGCShouldWait - Set to true by the background collector 1439 // when it is running. The foreground collector waits while 1440 // _foregroundGCShouldWait is true. 1441 // CGC_lock - monitor used to protect access to the above variables 1442 // and to notify the foreground and background collectors. 1443 // _collectorState - current state of the CMS collection. 1444 // 1445 // The foreground collector 1446 // acquires the CGC_lock 1447 // sets _foregroundGCIsActive 1448 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1449 // various locks acquired in preparation for the collection 1450 // are released so as not to block the background collector 1451 // that is in the midst of a collection 1452 // proceeds with the collection 1453 // clears _foregroundGCIsActive 1454 // returns 1455 // 1456 // The background collector in a loop iterating on the phases of the 1457 // collection 1458 // acquires the CGC_lock 1459 // sets _foregroundGCShouldWait 1460 // if _foregroundGCIsActive is set 1461 // clears _foregroundGCShouldWait, notifies _CGC_lock 1462 // waits on _CGC_lock for _foregroundGCIsActive to become false 1463 // and exits the loop. 1464 // otherwise 1465 // proceed with that phase of the collection 1466 // if the phase is a stop-the-world phase, 1467 // yield the baton once more just before enqueueing 1468 // the stop-world CMS operation (executed by the VM thread). 1469 // returns after all phases of the collection are done 1470 // 1471 1472 void CMSCollector::acquire_control_and_collect(bool full, 1473 bool clear_all_soft_refs) { 1474 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1475 assert(!Thread::current()->is_ConcurrentGC_thread(), 1476 "shouldn't try to acquire control from self!"); 1477 1478 // Start the protocol for acquiring control of the 1479 // collection from the background collector (aka CMS thread). 1480 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1481 "VM thread should have CMS token"); 1482 // Remember the possibly interrupted state of an ongoing 1483 // concurrent collection 1484 CollectorState first_state = _collectorState; 1485 1486 // Signal to a possibly ongoing concurrent collection that 1487 // we want to do a foreground collection. 1488 _foregroundGCIsActive = true; 1489 1490 // release locks and wait for a notify from the background collector 1491 // releasing the locks in only necessary for phases which 1492 // do yields to improve the granularity of the collection. 1493 assert_lock_strong(bitMapLock()); 1494 // We need to lock the Free list lock for the space that we are 1495 // currently collecting. 1496 assert(haveFreelistLocks(), "Must be holding free list locks"); 1497 bitMapLock()->unlock(); 1498 releaseFreelistLocks(); 1499 { 1500 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1501 if (_foregroundGCShouldWait) { 1502 // We are going to be waiting for action for the CMS thread; 1503 // it had better not be gone (for instance at shutdown)! 1504 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1505 "CMS thread must be running"); 1506 // Wait here until the background collector gives us the go-ahead 1507 ConcurrentMarkSweepThread::clear_CMS_flag( 1508 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1509 // Get a possibly blocked CMS thread going: 1510 // Note that we set _foregroundGCIsActive true above, 1511 // without protection of the CGC_lock. 1512 CGC_lock->notify(); 1513 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1514 "Possible deadlock"); 1515 while (_foregroundGCShouldWait) { 1516 // wait for notification 1517 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1518 // Possibility of delay/starvation here, since CMS token does 1519 // not know to give priority to VM thread? Actually, i think 1520 // there wouldn't be any delay/starvation, but the proof of 1521 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1522 } 1523 ConcurrentMarkSweepThread::set_CMS_flag( 1524 ConcurrentMarkSweepThread::CMS_vm_has_token); 1525 } 1526 } 1527 // The CMS_token is already held. Get back the other locks. 1528 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1529 "VM thread should have CMS token"); 1530 getFreelistLocks(); 1531 bitMapLock()->lock_without_safepoint_check(); 1532 if (TraceCMSState) { 1533 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1534 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1535 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1536 } 1537 1538 // Inform cms gen if this was due to partial collection failing. 1539 // The CMS gen may use this fact to determine its expansion policy. 1540 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1541 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1542 assert(!_cmsGen->incremental_collection_failed(), 1543 "Should have been noticed, reacted to and cleared"); 1544 _cmsGen->set_incremental_collection_failed(); 1545 } 1546 1547 if (first_state > Idling) { 1548 report_concurrent_mode_interruption(); 1549 } 1550 1551 set_did_compact(true); 1552 1553 // If the collection is being acquired from the background 1554 // collector, there may be references on the discovered 1555 // references lists. Abandon those references, since some 1556 // of them may have become unreachable after concurrent 1557 // discovery; the STW compacting collector will redo discovery 1558 // more precisely, without being subject to floating garbage. 1559 // Leaving otherwise unreachable references in the discovered 1560 // lists would require special handling. 1561 ref_processor()->disable_discovery(); 1562 ref_processor()->abandon_partial_discovery(); 1563 ref_processor()->verify_no_references_recorded(); 1564 1565 if (first_state > Idling) { 1566 save_heap_summary(); 1567 } 1568 1569 do_compaction_work(clear_all_soft_refs); 1570 1571 // Has the GC time limit been exceeded? 1572 size_t max_eden_size = _young_gen->max_capacity() - 1573 _young_gen->to()->capacity() - 1574 _young_gen->from()->capacity(); 1575 GCCause::Cause gc_cause = gch->gc_cause(); 1576 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1577 _young_gen->eden()->used(), 1578 _cmsGen->max_capacity(), 1579 max_eden_size, 1580 full, 1581 gc_cause, 1582 gch->collector_policy()); 1583 1584 // Reset the expansion cause, now that we just completed 1585 // a collection cycle. 1586 clear_expansion_cause(); 1587 _foregroundGCIsActive = false; 1588 return; 1589 } 1590 1591 // Resize the tenured generation 1592 // after obtaining the free list locks for the 1593 // two generations. 1594 void CMSCollector::compute_new_size() { 1595 assert_locked_or_safepoint(Heap_lock); 1596 FreelistLocker z(this); 1597 MetaspaceGC::compute_new_size(); 1598 _cmsGen->compute_new_size_free_list(); 1599 } 1600 1601 // A work method used by the foreground collector to do 1602 // a mark-sweep-compact. 1603 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1604 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1605 1606 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1607 gc_timer->register_gc_start(); 1608 1609 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1610 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1611 1612 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1613 1614 // Temporarily widen the span of the weak reference processing to 1615 // the entire heap. 1616 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1617 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1618 // Temporarily, clear the "is_alive_non_header" field of the 1619 // reference processor. 1620 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1621 // Temporarily make reference _processing_ single threaded (non-MT). 1622 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1623 // Temporarily make refs discovery atomic 1624 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1625 // Temporarily make reference _discovery_ single threaded (non-MT) 1626 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1627 1628 ref_processor()->set_enqueuing_is_done(false); 1629 ref_processor()->enable_discovery(); 1630 ref_processor()->setup_policy(clear_all_soft_refs); 1631 // If an asynchronous collection finishes, the _modUnionTable is 1632 // all clear. If we are assuming the collection from an asynchronous 1633 // collection, clear the _modUnionTable. 1634 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1635 "_modUnionTable should be clear if the baton was not passed"); 1636 _modUnionTable.clear_all(); 1637 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1638 "mod union for klasses should be clear if the baton was passed"); 1639 _ct->klass_rem_set()->clear_mod_union(); 1640 1641 // We must adjust the allocation statistics being maintained 1642 // in the free list space. We do so by reading and clearing 1643 // the sweep timer and updating the block flux rate estimates below. 1644 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1645 if (_inter_sweep_timer.is_active()) { 1646 _inter_sweep_timer.stop(); 1647 // Note that we do not use this sample to update the _inter_sweep_estimate. 1648 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1649 _inter_sweep_estimate.padded_average(), 1650 _intra_sweep_estimate.padded_average()); 1651 } 1652 1653 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1654 ref_processor(), clear_all_soft_refs); 1655 #ifdef ASSERT 1656 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1657 size_t free_size = cms_space->free(); 1658 assert(free_size == 1659 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1660 * HeapWordSize, 1661 "All the free space should be compacted into one chunk at top"); 1662 assert(cms_space->dictionary()->total_chunk_size( 1663 debug_only(cms_space->freelistLock())) == 0 || 1664 cms_space->totalSizeInIndexedFreeLists() == 0, 1665 "All the free space should be in a single chunk"); 1666 size_t num = cms_space->totalCount(); 1667 assert((free_size == 0 && num == 0) || 1668 (free_size > 0 && (num == 1 || num == 2)), 1669 "There should be at most 2 free chunks after compaction"); 1670 #endif // ASSERT 1671 _collectorState = Resetting; 1672 assert(_restart_addr == NULL, 1673 "Should have been NULL'd before baton was passed"); 1674 reset(false /* == !concurrent */); 1675 _cmsGen->reset_after_compaction(); 1676 _concurrent_cycles_since_last_unload = 0; 1677 1678 // Clear any data recorded in the PLAB chunk arrays. 1679 if (_survivor_plab_array != NULL) { 1680 reset_survivor_plab_arrays(); 1681 } 1682 1683 // Adjust the per-size allocation stats for the next epoch. 1684 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1685 // Restart the "inter sweep timer" for the next epoch. 1686 _inter_sweep_timer.reset(); 1687 _inter_sweep_timer.start(); 1688 1689 gc_timer->register_gc_end(); 1690 1691 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1692 1693 // For a mark-sweep-compact, compute_new_size() will be called 1694 // in the heap's do_collection() method. 1695 } 1696 1697 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1698 ContiguousSpace* eden_space = _young_gen->eden(); 1699 ContiguousSpace* from_space = _young_gen->from(); 1700 ContiguousSpace* to_space = _young_gen->to(); 1701 // Eden 1702 if (_eden_chunk_array != NULL) { 1703 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1704 p2i(eden_space->bottom()), p2i(eden_space->top()), 1705 p2i(eden_space->end()), eden_space->capacity()); 1706 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1707 "_eden_chunk_capacity=" SIZE_FORMAT, 1708 _eden_chunk_index, _eden_chunk_capacity); 1709 for (size_t i = 0; i < _eden_chunk_index; i++) { 1710 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1711 i, p2i(_eden_chunk_array[i])); 1712 } 1713 } 1714 // Survivor 1715 if (_survivor_chunk_array != NULL) { 1716 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1717 p2i(from_space->bottom()), p2i(from_space->top()), 1718 p2i(from_space->end()), from_space->capacity()); 1719 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1720 "_survivor_chunk_capacity=" SIZE_FORMAT, 1721 _survivor_chunk_index, _survivor_chunk_capacity); 1722 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1723 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1724 i, p2i(_survivor_chunk_array[i])); 1725 } 1726 } 1727 } 1728 1729 void CMSCollector::getFreelistLocks() const { 1730 // Get locks for all free lists in all generations that this 1731 // collector is responsible for 1732 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1733 } 1734 1735 void CMSCollector::releaseFreelistLocks() const { 1736 // Release locks for all free lists in all generations that this 1737 // collector is responsible for 1738 _cmsGen->freelistLock()->unlock(); 1739 } 1740 1741 bool CMSCollector::haveFreelistLocks() const { 1742 // Check locks for all free lists in all generations that this 1743 // collector is responsible for 1744 assert_lock_strong(_cmsGen->freelistLock()); 1745 PRODUCT_ONLY(ShouldNotReachHere()); 1746 return true; 1747 } 1748 1749 // A utility class that is used by the CMS collector to 1750 // temporarily "release" the foreground collector from its 1751 // usual obligation to wait for the background collector to 1752 // complete an ongoing phase before proceeding. 1753 class ReleaseForegroundGC: public StackObj { 1754 private: 1755 CMSCollector* _c; 1756 public: 1757 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1758 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1759 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1760 // allow a potentially blocked foreground collector to proceed 1761 _c->_foregroundGCShouldWait = false; 1762 if (_c->_foregroundGCIsActive) { 1763 CGC_lock->notify(); 1764 } 1765 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1766 "Possible deadlock"); 1767 } 1768 1769 ~ReleaseForegroundGC() { 1770 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1771 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1772 _c->_foregroundGCShouldWait = true; 1773 } 1774 }; 1775 1776 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1777 assert(Thread::current()->is_ConcurrentGC_thread(), 1778 "A CMS asynchronous collection is only allowed on a CMS thread."); 1779 1780 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1781 { 1782 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1783 MutexLockerEx hl(Heap_lock, safepoint_check); 1784 FreelistLocker fll(this); 1785 MutexLockerEx x(CGC_lock, safepoint_check); 1786 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1787 // The foreground collector is active or we're 1788 // not using asynchronous collections. Skip this 1789 // background collection. 1790 assert(!_foregroundGCShouldWait, "Should be clear"); 1791 return; 1792 } else { 1793 assert(_collectorState == Idling, "Should be idling before start."); 1794 _collectorState = InitialMarking; 1795 register_gc_start(cause); 1796 // Reset the expansion cause, now that we are about to begin 1797 // a new cycle. 1798 clear_expansion_cause(); 1799 1800 // Clear the MetaspaceGC flag since a concurrent collection 1801 // is starting but also clear it after the collection. 1802 MetaspaceGC::set_should_concurrent_collect(false); 1803 } 1804 // Decide if we want to enable class unloading as part of the 1805 // ensuing concurrent GC cycle. 1806 update_should_unload_classes(); 1807 _full_gc_requested = false; // acks all outstanding full gc requests 1808 _full_gc_cause = GCCause::_no_gc; 1809 // Signal that we are about to start a collection 1810 gch->increment_total_full_collections(); // ... starting a collection cycle 1811 _collection_count_start = gch->total_full_collections(); 1812 } 1813 1814 // Used for PrintGC 1815 size_t prev_used; 1816 if (PrintGC && Verbose) { 1817 prev_used = _cmsGen->used(); 1818 } 1819 1820 // The change of the collection state is normally done at this level; 1821 // the exceptions are phases that are executed while the world is 1822 // stopped. For those phases the change of state is done while the 1823 // world is stopped. For baton passing purposes this allows the 1824 // background collector to finish the phase and change state atomically. 1825 // The foreground collector cannot wait on a phase that is done 1826 // while the world is stopped because the foreground collector already 1827 // has the world stopped and would deadlock. 1828 while (_collectorState != Idling) { 1829 if (TraceCMSState) { 1830 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1831 p2i(Thread::current()), _collectorState); 1832 } 1833 // The foreground collector 1834 // holds the Heap_lock throughout its collection. 1835 // holds the CMS token (but not the lock) 1836 // except while it is waiting for the background collector to yield. 1837 // 1838 // The foreground collector should be blocked (not for long) 1839 // if the background collector is about to start a phase 1840 // executed with world stopped. If the background 1841 // collector has already started such a phase, the 1842 // foreground collector is blocked waiting for the 1843 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1844 // are executed in the VM thread. 1845 // 1846 // The locking order is 1847 // PendingListLock (PLL) -- if applicable (FinalMarking) 1848 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1849 // CMS token (claimed in 1850 // stop_world_and_do() --> 1851 // safepoint_synchronize() --> 1852 // CMSThread::synchronize()) 1853 1854 { 1855 // Check if the FG collector wants us to yield. 1856 CMSTokenSync x(true); // is cms thread 1857 if (waitForForegroundGC()) { 1858 // We yielded to a foreground GC, nothing more to be 1859 // done this round. 1860 assert(_foregroundGCShouldWait == false, "We set it to false in " 1861 "waitForForegroundGC()"); 1862 if (TraceCMSState) { 1863 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1864 " exiting collection CMS state %d", 1865 p2i(Thread::current()), _collectorState); 1866 } 1867 return; 1868 } else { 1869 // The background collector can run but check to see if the 1870 // foreground collector has done a collection while the 1871 // background collector was waiting to get the CGC_lock 1872 // above. If yes, break so that _foregroundGCShouldWait 1873 // is cleared before returning. 1874 if (_collectorState == Idling) { 1875 break; 1876 } 1877 } 1878 } 1879 1880 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1881 "should be waiting"); 1882 1883 switch (_collectorState) { 1884 case InitialMarking: 1885 { 1886 ReleaseForegroundGC x(this); 1887 stats().record_cms_begin(); 1888 VM_CMS_Initial_Mark initial_mark_op(this); 1889 VMThread::execute(&initial_mark_op); 1890 } 1891 // The collector state may be any legal state at this point 1892 // since the background collector may have yielded to the 1893 // foreground collector. 1894 break; 1895 case Marking: 1896 // initial marking in checkpointRootsInitialWork has been completed 1897 if (markFromRoots()) { // we were successful 1898 assert(_collectorState == Precleaning, "Collector state should " 1899 "have changed"); 1900 } else { 1901 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1902 } 1903 break; 1904 case Precleaning: 1905 // marking from roots in markFromRoots has been completed 1906 preclean(); 1907 assert(_collectorState == AbortablePreclean || 1908 _collectorState == FinalMarking, 1909 "Collector state should have changed"); 1910 break; 1911 case AbortablePreclean: 1912 abortable_preclean(); 1913 assert(_collectorState == FinalMarking, "Collector state should " 1914 "have changed"); 1915 break; 1916 case FinalMarking: 1917 { 1918 ReleaseForegroundGC x(this); 1919 1920 VM_CMS_Final_Remark final_remark_op(this); 1921 VMThread::execute(&final_remark_op); 1922 } 1923 assert(_foregroundGCShouldWait, "block post-condition"); 1924 break; 1925 case Sweeping: 1926 // final marking in checkpointRootsFinal has been completed 1927 sweep(); 1928 assert(_collectorState == Resizing, "Collector state change " 1929 "to Resizing must be done under the free_list_lock"); 1930 1931 case Resizing: { 1932 // Sweeping has been completed... 1933 // At this point the background collection has completed. 1934 // Don't move the call to compute_new_size() down 1935 // into code that might be executed if the background 1936 // collection was preempted. 1937 { 1938 ReleaseForegroundGC x(this); // unblock FG collection 1939 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1940 CMSTokenSync z(true); // not strictly needed. 1941 if (_collectorState == Resizing) { 1942 compute_new_size(); 1943 save_heap_summary(); 1944 _collectorState = Resetting; 1945 } else { 1946 assert(_collectorState == Idling, "The state should only change" 1947 " because the foreground collector has finished the collection"); 1948 } 1949 } 1950 break; 1951 } 1952 case Resetting: 1953 // CMS heap resizing has been completed 1954 reset(true); 1955 assert(_collectorState == Idling, "Collector state should " 1956 "have changed"); 1957 1958 MetaspaceGC::set_should_concurrent_collect(false); 1959 1960 stats().record_cms_end(); 1961 // Don't move the concurrent_phases_end() and compute_new_size() 1962 // calls to here because a preempted background collection 1963 // has it's state set to "Resetting". 1964 break; 1965 case Idling: 1966 default: 1967 ShouldNotReachHere(); 1968 break; 1969 } 1970 if (TraceCMSState) { 1971 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1972 p2i(Thread::current()), _collectorState); 1973 } 1974 assert(_foregroundGCShouldWait, "block post-condition"); 1975 } 1976 1977 // Should this be in gc_epilogue? 1978 collector_policy()->counters()->update_counters(); 1979 1980 { 1981 // Clear _foregroundGCShouldWait and, in the event that the 1982 // foreground collector is waiting, notify it, before 1983 // returning. 1984 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1985 _foregroundGCShouldWait = false; 1986 if (_foregroundGCIsActive) { 1987 CGC_lock->notify(); 1988 } 1989 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1990 "Possible deadlock"); 1991 } 1992 if (TraceCMSState) { 1993 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1994 " exiting collection CMS state %d", 1995 p2i(Thread::current()), _collectorState); 1996 } 1997 if (PrintGC && Verbose) { 1998 _cmsGen->print_heap_change(prev_used); 1999 } 2000 } 2001 2002 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2003 _cms_start_registered = true; 2004 _gc_timer_cm->register_gc_start(); 2005 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2006 } 2007 2008 void CMSCollector::register_gc_end() { 2009 if (_cms_start_registered) { 2010 report_heap_summary(GCWhen::AfterGC); 2011 2012 _gc_timer_cm->register_gc_end(); 2013 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2014 _cms_start_registered = false; 2015 } 2016 } 2017 2018 void CMSCollector::save_heap_summary() { 2019 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2020 _last_heap_summary = gch->create_heap_summary(); 2021 _last_metaspace_summary = gch->create_metaspace_summary(); 2022 } 2023 2024 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2025 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2026 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2027 } 2028 2029 bool CMSCollector::waitForForegroundGC() { 2030 bool res = false; 2031 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2032 "CMS thread should have CMS token"); 2033 // Block the foreground collector until the 2034 // background collectors decides whether to 2035 // yield. 2036 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2037 _foregroundGCShouldWait = true; 2038 if (_foregroundGCIsActive) { 2039 // The background collector yields to the 2040 // foreground collector and returns a value 2041 // indicating that it has yielded. The foreground 2042 // collector can proceed. 2043 res = true; 2044 _foregroundGCShouldWait = false; 2045 ConcurrentMarkSweepThread::clear_CMS_flag( 2046 ConcurrentMarkSweepThread::CMS_cms_has_token); 2047 ConcurrentMarkSweepThread::set_CMS_flag( 2048 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2049 // Get a possibly blocked foreground thread going 2050 CGC_lock->notify(); 2051 if (TraceCMSState) { 2052 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2053 p2i(Thread::current()), _collectorState); 2054 } 2055 while (_foregroundGCIsActive) { 2056 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2057 } 2058 ConcurrentMarkSweepThread::set_CMS_flag( 2059 ConcurrentMarkSweepThread::CMS_cms_has_token); 2060 ConcurrentMarkSweepThread::clear_CMS_flag( 2061 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2062 } 2063 if (TraceCMSState) { 2064 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2065 p2i(Thread::current()), _collectorState); 2066 } 2067 return res; 2068 } 2069 2070 // Because of the need to lock the free lists and other structures in 2071 // the collector, common to all the generations that the collector is 2072 // collecting, we need the gc_prologues of individual CMS generations 2073 // delegate to their collector. It may have been simpler had the 2074 // current infrastructure allowed one to call a prologue on a 2075 // collector. In the absence of that we have the generation's 2076 // prologue delegate to the collector, which delegates back 2077 // some "local" work to a worker method in the individual generations 2078 // that it's responsible for collecting, while itself doing any 2079 // work common to all generations it's responsible for. A similar 2080 // comment applies to the gc_epilogue()'s. 2081 // The role of the variable _between_prologue_and_epilogue is to 2082 // enforce the invocation protocol. 2083 void CMSCollector::gc_prologue(bool full) { 2084 // Call gc_prologue_work() for the CMSGen 2085 // we are responsible for. 2086 2087 // The following locking discipline assumes that we are only called 2088 // when the world is stopped. 2089 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2090 2091 // The CMSCollector prologue must call the gc_prologues for the 2092 // "generations" that it's responsible 2093 // for. 2094 2095 assert( Thread::current()->is_VM_thread() 2096 || ( CMSScavengeBeforeRemark 2097 && Thread::current()->is_ConcurrentGC_thread()), 2098 "Incorrect thread type for prologue execution"); 2099 2100 if (_between_prologue_and_epilogue) { 2101 // We have already been invoked; this is a gc_prologue delegation 2102 // from yet another CMS generation that we are responsible for, just 2103 // ignore it since all relevant work has already been done. 2104 return; 2105 } 2106 2107 // set a bit saying prologue has been called; cleared in epilogue 2108 _between_prologue_and_epilogue = true; 2109 // Claim locks for common data structures, then call gc_prologue_work() 2110 // for each CMSGen. 2111 2112 getFreelistLocks(); // gets free list locks on constituent spaces 2113 bitMapLock()->lock_without_safepoint_check(); 2114 2115 // Should call gc_prologue_work() for all cms gens we are responsible for 2116 bool duringMarking = _collectorState >= Marking 2117 && _collectorState < Sweeping; 2118 2119 // The young collections clear the modified oops state, which tells if 2120 // there are any modified oops in the class. The remark phase also needs 2121 // that information. Tell the young collection to save the union of all 2122 // modified klasses. 2123 if (duringMarking) { 2124 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2125 } 2126 2127 bool registerClosure = duringMarking; 2128 2129 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2130 2131 if (!full) { 2132 stats().record_gc0_begin(); 2133 } 2134 } 2135 2136 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2137 2138 _capacity_at_prologue = capacity(); 2139 _used_at_prologue = used(); 2140 2141 // Delegate to CMScollector which knows how to coordinate between 2142 // this and any other CMS generations that it is responsible for 2143 // collecting. 2144 collector()->gc_prologue(full); 2145 } 2146 2147 // This is a "private" interface for use by this generation's CMSCollector. 2148 // Not to be called directly by any other entity (for instance, 2149 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2150 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2151 bool registerClosure, ModUnionClosure* modUnionClosure) { 2152 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2153 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2154 "Should be NULL"); 2155 if (registerClosure) { 2156 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2157 } 2158 cmsSpace()->gc_prologue(); 2159 // Clear stat counters 2160 NOT_PRODUCT( 2161 assert(_numObjectsPromoted == 0, "check"); 2162 assert(_numWordsPromoted == 0, "check"); 2163 if (Verbose && PrintGC) { 2164 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2165 SIZE_FORMAT" bytes concurrently", 2166 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2167 } 2168 _numObjectsAllocated = 0; 2169 _numWordsAllocated = 0; 2170 ) 2171 } 2172 2173 void CMSCollector::gc_epilogue(bool full) { 2174 // The following locking discipline assumes that we are only called 2175 // when the world is stopped. 2176 assert(SafepointSynchronize::is_at_safepoint(), 2177 "world is stopped assumption"); 2178 2179 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2180 // if linear allocation blocks need to be appropriately marked to allow the 2181 // the blocks to be parsable. We also check here whether we need to nudge the 2182 // CMS collector thread to start a new cycle (if it's not already active). 2183 assert( Thread::current()->is_VM_thread() 2184 || ( CMSScavengeBeforeRemark 2185 && Thread::current()->is_ConcurrentGC_thread()), 2186 "Incorrect thread type for epilogue execution"); 2187 2188 if (!_between_prologue_and_epilogue) { 2189 // We have already been invoked; this is a gc_epilogue delegation 2190 // from yet another CMS generation that we are responsible for, just 2191 // ignore it since all relevant work has already been done. 2192 return; 2193 } 2194 assert(haveFreelistLocks(), "must have freelist locks"); 2195 assert_lock_strong(bitMapLock()); 2196 2197 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2198 2199 _cmsGen->gc_epilogue_work(full); 2200 2201 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2202 // in case sampling was not already enabled, enable it 2203 _start_sampling = true; 2204 } 2205 // reset _eden_chunk_array so sampling starts afresh 2206 _eden_chunk_index = 0; 2207 2208 size_t cms_used = _cmsGen->cmsSpace()->used(); 2209 2210 // update performance counters - this uses a special version of 2211 // update_counters() that allows the utilization to be passed as a 2212 // parameter, avoiding multiple calls to used(). 2213 // 2214 _cmsGen->update_counters(cms_used); 2215 2216 bitMapLock()->unlock(); 2217 releaseFreelistLocks(); 2218 2219 if (!CleanChunkPoolAsync) { 2220 Chunk::clean_chunk_pool(); 2221 } 2222 2223 set_did_compact(false); 2224 _between_prologue_and_epilogue = false; // ready for next cycle 2225 } 2226 2227 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2228 collector()->gc_epilogue(full); 2229 2230 // Also reset promotion tracking in par gc thread states. 2231 for (uint i = 0; i < ParallelGCThreads; i++) { 2232 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2233 } 2234 } 2235 2236 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2237 assert(!incremental_collection_failed(), "Should have been cleared"); 2238 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2239 cmsSpace()->gc_epilogue(); 2240 // Print stat counters 2241 NOT_PRODUCT( 2242 assert(_numObjectsAllocated == 0, "check"); 2243 assert(_numWordsAllocated == 0, "check"); 2244 if (Verbose && PrintGC) { 2245 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2246 SIZE_FORMAT" bytes", 2247 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2248 } 2249 _numObjectsPromoted = 0; 2250 _numWordsPromoted = 0; 2251 ) 2252 2253 if (PrintGC && Verbose) { 2254 // Call down the chain in contiguous_available needs the freelistLock 2255 // so print this out before releasing the freeListLock. 2256 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2257 contiguous_available()); 2258 } 2259 } 2260 2261 #ifndef PRODUCT 2262 bool CMSCollector::have_cms_token() { 2263 Thread* thr = Thread::current(); 2264 if (thr->is_VM_thread()) { 2265 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2266 } else if (thr->is_ConcurrentGC_thread()) { 2267 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2268 } else if (thr->is_GC_task_thread()) { 2269 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2270 ParGCRareEvent_lock->owned_by_self(); 2271 } 2272 return false; 2273 } 2274 #endif 2275 2276 // Check reachability of the given heap address in CMS generation, 2277 // treating all other generations as roots. 2278 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2279 // We could "guarantee" below, rather than assert, but I'll 2280 // leave these as "asserts" so that an adventurous debugger 2281 // could try this in the product build provided some subset of 2282 // the conditions were met, provided they were interested in the 2283 // results and knew that the computation below wouldn't interfere 2284 // with other concurrent computations mutating the structures 2285 // being read or written. 2286 assert(SafepointSynchronize::is_at_safepoint(), 2287 "Else mutations in object graph will make answer suspect"); 2288 assert(have_cms_token(), "Should hold cms token"); 2289 assert(haveFreelistLocks(), "must hold free list locks"); 2290 assert_lock_strong(bitMapLock()); 2291 2292 // Clear the marking bit map array before starting, but, just 2293 // for kicks, first report if the given address is already marked 2294 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2295 _markBitMap.isMarked(addr) ? "" : " not"); 2296 2297 if (verify_after_remark()) { 2298 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2299 bool result = verification_mark_bm()->isMarked(addr); 2300 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2301 result ? "IS" : "is NOT"); 2302 return result; 2303 } else { 2304 gclog_or_tty->print_cr("Could not compute result"); 2305 return false; 2306 } 2307 } 2308 2309 2310 void 2311 CMSCollector::print_on_error(outputStream* st) { 2312 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2313 if (collector != NULL) { 2314 CMSBitMap* bitmap = &collector->_markBitMap; 2315 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2316 bitmap->print_on_error(st, " Bits: "); 2317 2318 st->cr(); 2319 2320 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2321 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2322 mut_bitmap->print_on_error(st, " Bits: "); 2323 } 2324 } 2325 2326 //////////////////////////////////////////////////////// 2327 // CMS Verification Support 2328 //////////////////////////////////////////////////////// 2329 // Following the remark phase, the following invariant 2330 // should hold -- each object in the CMS heap which is 2331 // marked in markBitMap() should be marked in the verification_mark_bm(). 2332 2333 class VerifyMarkedClosure: public BitMapClosure { 2334 CMSBitMap* _marks; 2335 bool _failed; 2336 2337 public: 2338 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2339 2340 bool do_bit(size_t offset) { 2341 HeapWord* addr = _marks->offsetToHeapWord(offset); 2342 if (!_marks->isMarked(addr)) { 2343 oop(addr)->print_on(gclog_or_tty); 2344 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2345 _failed = true; 2346 } 2347 return true; 2348 } 2349 2350 bool failed() { return _failed; } 2351 }; 2352 2353 bool CMSCollector::verify_after_remark(bool silent) { 2354 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2355 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2356 static bool init = false; 2357 2358 assert(SafepointSynchronize::is_at_safepoint(), 2359 "Else mutations in object graph will make answer suspect"); 2360 assert(have_cms_token(), 2361 "Else there may be mutual interference in use of " 2362 " verification data structures"); 2363 assert(_collectorState > Marking && _collectorState <= Sweeping, 2364 "Else marking info checked here may be obsolete"); 2365 assert(haveFreelistLocks(), "must hold free list locks"); 2366 assert_lock_strong(bitMapLock()); 2367 2368 2369 // Allocate marking bit map if not already allocated 2370 if (!init) { // first time 2371 if (!verification_mark_bm()->allocate(_span)) { 2372 return false; 2373 } 2374 init = true; 2375 } 2376 2377 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2378 2379 // Turn off refs discovery -- so we will be tracing through refs. 2380 // This is as intended, because by this time 2381 // GC must already have cleared any refs that need to be cleared, 2382 // and traced those that need to be marked; moreover, 2383 // the marking done here is not going to interfere in any 2384 // way with the marking information used by GC. 2385 NoRefDiscovery no_discovery(ref_processor()); 2386 2387 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2388 2389 // Clear any marks from a previous round 2390 verification_mark_bm()->clear_all(); 2391 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2392 verify_work_stacks_empty(); 2393 2394 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2395 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2396 // Update the saved marks which may affect the root scans. 2397 gch->save_marks(); 2398 2399 if (CMSRemarkVerifyVariant == 1) { 2400 // In this first variant of verification, we complete 2401 // all marking, then check if the new marks-vector is 2402 // a subset of the CMS marks-vector. 2403 verify_after_remark_work_1(); 2404 } else if (CMSRemarkVerifyVariant == 2) { 2405 // In this second variant of verification, we flag an error 2406 // (i.e. an object reachable in the new marks-vector not reachable 2407 // in the CMS marks-vector) immediately, also indicating the 2408 // identify of an object (A) that references the unmarked object (B) -- 2409 // presumably, a mutation to A failed to be picked up by preclean/remark? 2410 verify_after_remark_work_2(); 2411 } else { 2412 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2413 CMSRemarkVerifyVariant); 2414 } 2415 if (!silent) gclog_or_tty->print(" done] "); 2416 return true; 2417 } 2418 2419 void CMSCollector::verify_after_remark_work_1() { 2420 ResourceMark rm; 2421 HandleMark hm; 2422 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2423 2424 // Get a clear set of claim bits for the roots processing to work with. 2425 ClassLoaderDataGraph::clear_claimed_marks(); 2426 2427 // Mark from roots one level into CMS 2428 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2429 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2430 2431 { 2432 StrongRootsScope srs(1); 2433 2434 gch->gen_process_roots(&srs, 2435 _cmsGen->level(), 2436 true, // younger gens are roots 2437 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2438 should_unload_classes(), 2439 ¬Older, 2440 NULL, 2441 NULL); 2442 } 2443 2444 // Now mark from the roots 2445 MarkFromRootsClosure markFromRootsClosure(this, _span, 2446 verification_mark_bm(), verification_mark_stack(), 2447 false /* don't yield */, true /* verifying */); 2448 assert(_restart_addr == NULL, "Expected pre-condition"); 2449 verification_mark_bm()->iterate(&markFromRootsClosure); 2450 while (_restart_addr != NULL) { 2451 // Deal with stack overflow: by restarting at the indicated 2452 // address. 2453 HeapWord* ra = _restart_addr; 2454 markFromRootsClosure.reset(ra); 2455 _restart_addr = NULL; 2456 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2457 } 2458 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2459 verify_work_stacks_empty(); 2460 2461 // Marking completed -- now verify that each bit marked in 2462 // verification_mark_bm() is also marked in markBitMap(); flag all 2463 // errors by printing corresponding objects. 2464 VerifyMarkedClosure vcl(markBitMap()); 2465 verification_mark_bm()->iterate(&vcl); 2466 if (vcl.failed()) { 2467 gclog_or_tty->print("Verification failed"); 2468 gch->print_on(gclog_or_tty); 2469 fatal("CMS: failed marking verification after remark"); 2470 } 2471 } 2472 2473 class VerifyKlassOopsKlassClosure : public KlassClosure { 2474 class VerifyKlassOopsClosure : public OopClosure { 2475 CMSBitMap* _bitmap; 2476 public: 2477 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2478 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2479 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2480 } _oop_closure; 2481 public: 2482 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2483 void do_klass(Klass* k) { 2484 k->oops_do(&_oop_closure); 2485 } 2486 }; 2487 2488 void CMSCollector::verify_after_remark_work_2() { 2489 ResourceMark rm; 2490 HandleMark hm; 2491 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2492 2493 // Get a clear set of claim bits for the roots processing to work with. 2494 ClassLoaderDataGraph::clear_claimed_marks(); 2495 2496 // Mark from roots one level into CMS 2497 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2498 markBitMap()); 2499 CLDToOopClosure cld_closure(¬Older, true); 2500 2501 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2502 2503 { 2504 StrongRootsScope srs(1); 2505 2506 gch->gen_process_roots(&srs, 2507 _cmsGen->level(), 2508 true, // younger gens are roots 2509 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2510 should_unload_classes(), 2511 ¬Older, 2512 NULL, 2513 &cld_closure); 2514 } 2515 2516 // Now mark from the roots 2517 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2518 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2519 assert(_restart_addr == NULL, "Expected pre-condition"); 2520 verification_mark_bm()->iterate(&markFromRootsClosure); 2521 while (_restart_addr != NULL) { 2522 // Deal with stack overflow: by restarting at the indicated 2523 // address. 2524 HeapWord* ra = _restart_addr; 2525 markFromRootsClosure.reset(ra); 2526 _restart_addr = NULL; 2527 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2528 } 2529 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2530 verify_work_stacks_empty(); 2531 2532 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2533 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2534 2535 // Marking completed -- now verify that each bit marked in 2536 // verification_mark_bm() is also marked in markBitMap(); flag all 2537 // errors by printing corresponding objects. 2538 VerifyMarkedClosure vcl(markBitMap()); 2539 verification_mark_bm()->iterate(&vcl); 2540 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2541 } 2542 2543 void ConcurrentMarkSweepGeneration::save_marks() { 2544 // delegate to CMS space 2545 cmsSpace()->save_marks(); 2546 for (uint i = 0; i < ParallelGCThreads; i++) { 2547 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2548 } 2549 } 2550 2551 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2552 return cmsSpace()->no_allocs_since_save_marks(); 2553 } 2554 2555 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2556 \ 2557 void ConcurrentMarkSweepGeneration:: \ 2558 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2559 cl->set_generation(this); \ 2560 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2561 cl->reset_generation(); \ 2562 save_marks(); \ 2563 } 2564 2565 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2566 2567 void 2568 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2569 if (freelistLock()->owned_by_self()) { 2570 Generation::oop_iterate(cl); 2571 } else { 2572 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2573 Generation::oop_iterate(cl); 2574 } 2575 } 2576 2577 void 2578 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2579 if (freelistLock()->owned_by_self()) { 2580 Generation::object_iterate(cl); 2581 } else { 2582 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2583 Generation::object_iterate(cl); 2584 } 2585 } 2586 2587 void 2588 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2589 if (freelistLock()->owned_by_self()) { 2590 Generation::safe_object_iterate(cl); 2591 } else { 2592 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2593 Generation::safe_object_iterate(cl); 2594 } 2595 } 2596 2597 void 2598 ConcurrentMarkSweepGeneration::post_compact() { 2599 } 2600 2601 void 2602 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2603 // Fix the linear allocation blocks to look like free blocks. 2604 2605 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2606 // are not called when the heap is verified during universe initialization and 2607 // at vm shutdown. 2608 if (freelistLock()->owned_by_self()) { 2609 cmsSpace()->prepare_for_verify(); 2610 } else { 2611 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2612 cmsSpace()->prepare_for_verify(); 2613 } 2614 } 2615 2616 void 2617 ConcurrentMarkSweepGeneration::verify() { 2618 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2619 // are not called when the heap is verified during universe initialization and 2620 // at vm shutdown. 2621 if (freelistLock()->owned_by_self()) { 2622 cmsSpace()->verify(); 2623 } else { 2624 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2625 cmsSpace()->verify(); 2626 } 2627 } 2628 2629 void CMSCollector::verify() { 2630 _cmsGen->verify(); 2631 } 2632 2633 #ifndef PRODUCT 2634 bool CMSCollector::overflow_list_is_empty() const { 2635 assert(_num_par_pushes >= 0, "Inconsistency"); 2636 if (_overflow_list == NULL) { 2637 assert(_num_par_pushes == 0, "Inconsistency"); 2638 } 2639 return _overflow_list == NULL; 2640 } 2641 2642 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2643 // merely consolidate assertion checks that appear to occur together frequently. 2644 void CMSCollector::verify_work_stacks_empty() const { 2645 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2646 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2647 } 2648 2649 void CMSCollector::verify_overflow_empty() const { 2650 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2651 assert(no_preserved_marks(), "No preserved marks"); 2652 } 2653 #endif // PRODUCT 2654 2655 // Decide if we want to enable class unloading as part of the 2656 // ensuing concurrent GC cycle. We will collect and 2657 // unload classes if it's the case that: 2658 // (1) an explicit gc request has been made and the flag 2659 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2660 // (2) (a) class unloading is enabled at the command line, and 2661 // (b) old gen is getting really full 2662 // NOTE: Provided there is no change in the state of the heap between 2663 // calls to this method, it should have idempotent results. Moreover, 2664 // its results should be monotonically increasing (i.e. going from 0 to 1, 2665 // but not 1 to 0) between successive calls between which the heap was 2666 // not collected. For the implementation below, it must thus rely on 2667 // the property that concurrent_cycles_since_last_unload() 2668 // will not decrease unless a collection cycle happened and that 2669 // _cmsGen->is_too_full() are 2670 // themselves also monotonic in that sense. See check_monotonicity() 2671 // below. 2672 void CMSCollector::update_should_unload_classes() { 2673 _should_unload_classes = false; 2674 // Condition 1 above 2675 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2676 _should_unload_classes = true; 2677 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2678 // Disjuncts 2.b.(i,ii,iii) above 2679 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2680 CMSClassUnloadingMaxInterval) 2681 || _cmsGen->is_too_full(); 2682 } 2683 } 2684 2685 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2686 bool res = should_concurrent_collect(); 2687 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2688 return res; 2689 } 2690 2691 void CMSCollector::setup_cms_unloading_and_verification_state() { 2692 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2693 || VerifyBeforeExit; 2694 const int rso = GenCollectedHeap::SO_AllCodeCache; 2695 2696 // We set the proper root for this CMS cycle here. 2697 if (should_unload_classes()) { // Should unload classes this cycle 2698 remove_root_scanning_option(rso); // Shrink the root set appropriately 2699 set_verifying(should_verify); // Set verification state for this cycle 2700 return; // Nothing else needs to be done at this time 2701 } 2702 2703 // Not unloading classes this cycle 2704 assert(!should_unload_classes(), "Inconsistency!"); 2705 2706 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2707 // Include symbols, strings and code cache elements to prevent their resurrection. 2708 add_root_scanning_option(rso); 2709 set_verifying(true); 2710 } else if (verifying() && !should_verify) { 2711 // We were verifying, but some verification flags got disabled. 2712 set_verifying(false); 2713 // Exclude symbols, strings and code cache elements from root scanning to 2714 // reduce IM and RM pauses. 2715 remove_root_scanning_option(rso); 2716 } 2717 } 2718 2719 2720 #ifndef PRODUCT 2721 HeapWord* CMSCollector::block_start(const void* p) const { 2722 const HeapWord* addr = (HeapWord*)p; 2723 if (_span.contains(p)) { 2724 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2725 return _cmsGen->cmsSpace()->block_start(p); 2726 } 2727 } 2728 return NULL; 2729 } 2730 #endif 2731 2732 HeapWord* 2733 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2734 bool tlab, 2735 bool parallel) { 2736 CMSSynchronousYieldRequest yr; 2737 assert(!tlab, "Can't deal with TLAB allocation"); 2738 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2739 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2740 if (GCExpandToAllocateDelayMillis > 0) { 2741 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2742 } 2743 return have_lock_and_allocate(word_size, tlab); 2744 } 2745 2746 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2747 size_t bytes, 2748 size_t expand_bytes, 2749 CMSExpansionCause::Cause cause) 2750 { 2751 2752 bool success = expand(bytes, expand_bytes); 2753 2754 // remember why we expanded; this information is used 2755 // by shouldConcurrentCollect() when making decisions on whether to start 2756 // a new CMS cycle. 2757 if (success) { 2758 set_expansion_cause(cause); 2759 if (PrintGCDetails && Verbose) { 2760 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2761 CMSExpansionCause::to_string(cause)); 2762 } 2763 } 2764 } 2765 2766 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2767 HeapWord* res = NULL; 2768 MutexLocker x(ParGCRareEvent_lock); 2769 while (true) { 2770 // Expansion by some other thread might make alloc OK now: 2771 res = ps->lab.alloc(word_sz); 2772 if (res != NULL) return res; 2773 // If there's not enough expansion space available, give up. 2774 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2775 return NULL; 2776 } 2777 // Otherwise, we try expansion. 2778 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2779 // Now go around the loop and try alloc again; 2780 // A competing par_promote might beat us to the expansion space, 2781 // so we may go around the loop again if promotion fails again. 2782 if (GCExpandToAllocateDelayMillis > 0) { 2783 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2784 } 2785 } 2786 } 2787 2788 2789 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2790 PromotionInfo* promo) { 2791 MutexLocker x(ParGCRareEvent_lock); 2792 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2793 while (true) { 2794 // Expansion by some other thread might make alloc OK now: 2795 if (promo->ensure_spooling_space()) { 2796 assert(promo->has_spooling_space(), 2797 "Post-condition of successful ensure_spooling_space()"); 2798 return true; 2799 } 2800 // If there's not enough expansion space available, give up. 2801 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2802 return false; 2803 } 2804 // Otherwise, we try expansion. 2805 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2806 // Now go around the loop and try alloc again; 2807 // A competing allocation might beat us to the expansion space, 2808 // so we may go around the loop again if allocation fails again. 2809 if (GCExpandToAllocateDelayMillis > 0) { 2810 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2811 } 2812 } 2813 } 2814 2815 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2816 // Only shrink if a compaction was done so that all the free space 2817 // in the generation is in a contiguous block at the end. 2818 if (did_compact()) { 2819 CardGeneration::shrink(bytes); 2820 } 2821 } 2822 2823 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2824 assert_locked_or_safepoint(Heap_lock); 2825 } 2826 2827 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2828 assert_locked_or_safepoint(Heap_lock); 2829 assert_lock_strong(freelistLock()); 2830 if (PrintGCDetails && Verbose) { 2831 warning("Shrinking of CMS not yet implemented"); 2832 } 2833 return; 2834 } 2835 2836 2837 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2838 // phases. 2839 class CMSPhaseAccounting: public StackObj { 2840 public: 2841 CMSPhaseAccounting(CMSCollector *collector, 2842 const char *phase, 2843 const GCId gc_id, 2844 bool print_cr = true); 2845 ~CMSPhaseAccounting(); 2846 2847 private: 2848 CMSCollector *_collector; 2849 const char *_phase; 2850 elapsedTimer _wallclock; 2851 bool _print_cr; 2852 const GCId _gc_id; 2853 2854 public: 2855 // Not MT-safe; so do not pass around these StackObj's 2856 // where they may be accessed by other threads. 2857 jlong wallclock_millis() { 2858 assert(_wallclock.is_active(), "Wall clock should not stop"); 2859 _wallclock.stop(); // to record time 2860 jlong ret = _wallclock.milliseconds(); 2861 _wallclock.start(); // restart 2862 return ret; 2863 } 2864 }; 2865 2866 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2867 const char *phase, 2868 const GCId gc_id, 2869 bool print_cr) : 2870 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2871 2872 if (PrintCMSStatistics != 0) { 2873 _collector->resetYields(); 2874 } 2875 if (PrintGCDetails) { 2876 gclog_or_tty->gclog_stamp(_gc_id); 2877 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2878 _collector->cmsGen()->short_name(), _phase); 2879 } 2880 _collector->resetTimer(); 2881 _wallclock.start(); 2882 _collector->startTimer(); 2883 } 2884 2885 CMSPhaseAccounting::~CMSPhaseAccounting() { 2886 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2887 _collector->stopTimer(); 2888 _wallclock.stop(); 2889 if (PrintGCDetails) { 2890 gclog_or_tty->gclog_stamp(_gc_id); 2891 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2892 _collector->cmsGen()->short_name(), 2893 _phase, _collector->timerValue(), _wallclock.seconds()); 2894 if (_print_cr) { 2895 gclog_or_tty->cr(); 2896 } 2897 if (PrintCMSStatistics != 0) { 2898 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2899 _collector->yields()); 2900 } 2901 } 2902 } 2903 2904 // CMS work 2905 2906 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2907 class CMSParMarkTask : public AbstractGangTask { 2908 protected: 2909 CMSCollector* _collector; 2910 uint _n_workers; 2911 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2912 AbstractGangTask(name), 2913 _collector(collector), 2914 _n_workers(n_workers) {} 2915 // Work method in support of parallel rescan ... of young gen spaces 2916 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2917 ContiguousSpace* space, 2918 HeapWord** chunk_array, size_t chunk_top); 2919 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2920 }; 2921 2922 // Parallel initial mark task 2923 class CMSParInitialMarkTask: public CMSParMarkTask { 2924 StrongRootsScope* _strong_roots_scope; 2925 public: 2926 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2927 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2928 _strong_roots_scope(strong_roots_scope) {} 2929 void work(uint worker_id); 2930 }; 2931 2932 // Checkpoint the roots into this generation from outside 2933 // this generation. [Note this initial checkpoint need only 2934 // be approximate -- we'll do a catch up phase subsequently.] 2935 void CMSCollector::checkpointRootsInitial() { 2936 assert(_collectorState == InitialMarking, "Wrong collector state"); 2937 check_correct_thread_executing(); 2938 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2939 2940 save_heap_summary(); 2941 report_heap_summary(GCWhen::BeforeGC); 2942 2943 ReferenceProcessor* rp = ref_processor(); 2944 assert(_restart_addr == NULL, "Control point invariant"); 2945 { 2946 // acquire locks for subsequent manipulations 2947 MutexLockerEx x(bitMapLock(), 2948 Mutex::_no_safepoint_check_flag); 2949 checkpointRootsInitialWork(); 2950 // enable ("weak") refs discovery 2951 rp->enable_discovery(); 2952 _collectorState = Marking; 2953 } 2954 } 2955 2956 void CMSCollector::checkpointRootsInitialWork() { 2957 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2958 assert(_collectorState == InitialMarking, "just checking"); 2959 2960 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2961 // precede our marking with a collection of all 2962 // younger generations to keep floating garbage to a minimum. 2963 // XXX: we won't do this for now -- it's an optimization to be done later. 2964 2965 // already have locks 2966 assert_lock_strong(bitMapLock()); 2967 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2968 2969 // Setup the verification and class unloading state for this 2970 // CMS collection cycle. 2971 setup_cms_unloading_and_verification_state(); 2972 2973 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2974 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2975 2976 // Reset all the PLAB chunk arrays if necessary. 2977 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2978 reset_survivor_plab_arrays(); 2979 } 2980 2981 ResourceMark rm; 2982 HandleMark hm; 2983 2984 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2985 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2986 2987 verify_work_stacks_empty(); 2988 verify_overflow_empty(); 2989 2990 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2991 // Update the saved marks which may affect the root scans. 2992 gch->save_marks(); 2993 2994 // weak reference processing has not started yet. 2995 ref_processor()->set_enqueuing_is_done(false); 2996 2997 // Need to remember all newly created CLDs, 2998 // so that we can guarantee that the remark finds them. 2999 ClassLoaderDataGraph::remember_new_clds(true); 3000 3001 // Whenever a CLD is found, it will be claimed before proceeding to mark 3002 // the klasses. The claimed marks need to be cleared before marking starts. 3003 ClassLoaderDataGraph::clear_claimed_marks(); 3004 3005 if (CMSPrintEdenSurvivorChunks) { 3006 print_eden_and_survivor_chunk_arrays(); 3007 } 3008 3009 { 3010 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3011 if (CMSParallelInitialMarkEnabled) { 3012 // The parallel version. 3013 FlexibleWorkGang* workers = gch->workers(); 3014 assert(workers != NULL, "Need parallel worker threads."); 3015 uint n_workers = workers->active_workers(); 3016 3017 StrongRootsScope srs(n_workers); 3018 3019 CMSParInitialMarkTask tsk(this, &srs, n_workers); 3020 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3021 if (n_workers > 1) { 3022 workers->run_task(&tsk); 3023 } else { 3024 tsk.work(0); 3025 } 3026 } else { 3027 // The serial version. 3028 CLDToOopClosure cld_closure(¬Older, true); 3029 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3030 3031 StrongRootsScope srs(1); 3032 3033 gch->gen_process_roots(&srs, 3034 _cmsGen->level(), 3035 true, // younger gens are roots 3036 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3037 should_unload_classes(), 3038 ¬Older, 3039 NULL, 3040 &cld_closure); 3041 } 3042 } 3043 3044 // Clear mod-union table; it will be dirtied in the prologue of 3045 // CMS generation per each younger generation collection. 3046 3047 assert(_modUnionTable.isAllClear(), 3048 "Was cleared in most recent final checkpoint phase" 3049 " or no bits are set in the gc_prologue before the start of the next " 3050 "subsequent marking phase."); 3051 3052 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3053 3054 // Save the end of the used_region of the constituent generations 3055 // to be used to limit the extent of sweep in each generation. 3056 save_sweep_limits(); 3057 verify_overflow_empty(); 3058 } 3059 3060 bool CMSCollector::markFromRoots() { 3061 // we might be tempted to assert that: 3062 // assert(!SafepointSynchronize::is_at_safepoint(), 3063 // "inconsistent argument?"); 3064 // However that wouldn't be right, because it's possible that 3065 // a safepoint is indeed in progress as a younger generation 3066 // stop-the-world GC happens even as we mark in this generation. 3067 assert(_collectorState == Marking, "inconsistent state?"); 3068 check_correct_thread_executing(); 3069 verify_overflow_empty(); 3070 3071 // Weak ref discovery note: We may be discovering weak 3072 // refs in this generation concurrent (but interleaved) with 3073 // weak ref discovery by a younger generation collector. 3074 3075 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3076 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3077 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3078 bool res = markFromRootsWork(); 3079 if (res) { 3080 _collectorState = Precleaning; 3081 } else { // We failed and a foreground collection wants to take over 3082 assert(_foregroundGCIsActive, "internal state inconsistency"); 3083 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3084 if (PrintGCDetails) { 3085 gclog_or_tty->print_cr("bailing out to foreground collection"); 3086 } 3087 } 3088 verify_overflow_empty(); 3089 return res; 3090 } 3091 3092 bool CMSCollector::markFromRootsWork() { 3093 // iterate over marked bits in bit map, doing a full scan and mark 3094 // from these roots using the following algorithm: 3095 // . if oop is to the right of the current scan pointer, 3096 // mark corresponding bit (we'll process it later) 3097 // . else (oop is to left of current scan pointer) 3098 // push oop on marking stack 3099 // . drain the marking stack 3100 3101 // Note that when we do a marking step we need to hold the 3102 // bit map lock -- recall that direct allocation (by mutators) 3103 // and promotion (by younger generation collectors) is also 3104 // marking the bit map. [the so-called allocate live policy.] 3105 // Because the implementation of bit map marking is not 3106 // robust wrt simultaneous marking of bits in the same word, 3107 // we need to make sure that there is no such interference 3108 // between concurrent such updates. 3109 3110 // already have locks 3111 assert_lock_strong(bitMapLock()); 3112 3113 verify_work_stacks_empty(); 3114 verify_overflow_empty(); 3115 bool result = false; 3116 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3117 result = do_marking_mt(); 3118 } else { 3119 result = do_marking_st(); 3120 } 3121 return result; 3122 } 3123 3124 // Forward decl 3125 class CMSConcMarkingTask; 3126 3127 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3128 CMSCollector* _collector; 3129 CMSConcMarkingTask* _task; 3130 public: 3131 virtual void yield(); 3132 3133 // "n_threads" is the number of threads to be terminated. 3134 // "queue_set" is a set of work queues of other threads. 3135 // "collector" is the CMS collector associated with this task terminator. 3136 // "yield" indicates whether we need the gang as a whole to yield. 3137 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3138 ParallelTaskTerminator(n_threads, queue_set), 3139 _collector(collector) { } 3140 3141 void set_task(CMSConcMarkingTask* task) { 3142 _task = task; 3143 } 3144 }; 3145 3146 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3147 CMSConcMarkingTask* _task; 3148 public: 3149 bool should_exit_termination(); 3150 void set_task(CMSConcMarkingTask* task) { 3151 _task = task; 3152 } 3153 }; 3154 3155 // MT Concurrent Marking Task 3156 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3157 CMSCollector* _collector; 3158 uint _n_workers; // requested/desired # workers 3159 bool _result; 3160 CompactibleFreeListSpace* _cms_space; 3161 char _pad_front[64]; // padding to ... 3162 HeapWord* _global_finger; // ... avoid sharing cache line 3163 char _pad_back[64]; 3164 HeapWord* _restart_addr; 3165 3166 // Exposed here for yielding support 3167 Mutex* const _bit_map_lock; 3168 3169 // The per thread work queues, available here for stealing 3170 OopTaskQueueSet* _task_queues; 3171 3172 // Termination (and yielding) support 3173 CMSConcMarkingTerminator _term; 3174 CMSConcMarkingTerminatorTerminator _term_term; 3175 3176 public: 3177 CMSConcMarkingTask(CMSCollector* collector, 3178 CompactibleFreeListSpace* cms_space, 3179 YieldingFlexibleWorkGang* workers, 3180 OopTaskQueueSet* task_queues): 3181 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3182 _collector(collector), 3183 _cms_space(cms_space), 3184 _n_workers(0), _result(true), 3185 _task_queues(task_queues), 3186 _term(_n_workers, task_queues, _collector), 3187 _bit_map_lock(collector->bitMapLock()) 3188 { 3189 _requested_size = _n_workers; 3190 _term.set_task(this); 3191 _term_term.set_task(this); 3192 _restart_addr = _global_finger = _cms_space->bottom(); 3193 } 3194 3195 3196 OopTaskQueueSet* task_queues() { return _task_queues; } 3197 3198 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3199 3200 HeapWord** global_finger_addr() { return &_global_finger; } 3201 3202 CMSConcMarkingTerminator* terminator() { return &_term; } 3203 3204 virtual void set_for_termination(uint active_workers) { 3205 terminator()->reset_for_reuse(active_workers); 3206 } 3207 3208 void work(uint worker_id); 3209 bool should_yield() { 3210 return ConcurrentMarkSweepThread::should_yield() 3211 && !_collector->foregroundGCIsActive(); 3212 } 3213 3214 virtual void coordinator_yield(); // stuff done by coordinator 3215 bool result() { return _result; } 3216 3217 void reset(HeapWord* ra) { 3218 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3219 _restart_addr = _global_finger = ra; 3220 _term.reset_for_reuse(); 3221 } 3222 3223 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3224 OopTaskQueue* work_q); 3225 3226 private: 3227 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3228 void do_work_steal(int i); 3229 void bump_global_finger(HeapWord* f); 3230 }; 3231 3232 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3233 assert(_task != NULL, "Error"); 3234 return _task->yielding(); 3235 // Note that we do not need the disjunct || _task->should_yield() above 3236 // because we want terminating threads to yield only if the task 3237 // is already in the midst of yielding, which happens only after at least one 3238 // thread has yielded. 3239 } 3240 3241 void CMSConcMarkingTerminator::yield() { 3242 if (_task->should_yield()) { 3243 _task->yield(); 3244 } else { 3245 ParallelTaskTerminator::yield(); 3246 } 3247 } 3248 3249 //////////////////////////////////////////////////////////////// 3250 // Concurrent Marking Algorithm Sketch 3251 //////////////////////////////////////////////////////////////// 3252 // Until all tasks exhausted (both spaces): 3253 // -- claim next available chunk 3254 // -- bump global finger via CAS 3255 // -- find first object that starts in this chunk 3256 // and start scanning bitmap from that position 3257 // -- scan marked objects for oops 3258 // -- CAS-mark target, and if successful: 3259 // . if target oop is above global finger (volatile read) 3260 // nothing to do 3261 // . if target oop is in chunk and above local finger 3262 // then nothing to do 3263 // . else push on work-queue 3264 // -- Deal with possible overflow issues: 3265 // . local work-queue overflow causes stuff to be pushed on 3266 // global (common) overflow queue 3267 // . always first empty local work queue 3268 // . then get a batch of oops from global work queue if any 3269 // . then do work stealing 3270 // -- When all tasks claimed (both spaces) 3271 // and local work queue empty, 3272 // then in a loop do: 3273 // . check global overflow stack; steal a batch of oops and trace 3274 // . try to steal from other threads oif GOS is empty 3275 // . if neither is available, offer termination 3276 // -- Terminate and return result 3277 // 3278 void CMSConcMarkingTask::work(uint worker_id) { 3279 elapsedTimer _timer; 3280 ResourceMark rm; 3281 HandleMark hm; 3282 3283 DEBUG_ONLY(_collector->verify_overflow_empty();) 3284 3285 // Before we begin work, our work queue should be empty 3286 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3287 // Scan the bitmap covering _cms_space, tracing through grey objects. 3288 _timer.start(); 3289 do_scan_and_mark(worker_id, _cms_space); 3290 _timer.stop(); 3291 if (PrintCMSStatistics != 0) { 3292 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3293 worker_id, _timer.seconds()); 3294 // XXX: need xxx/xxx type of notation, two timers 3295 } 3296 3297 // ... do work stealing 3298 _timer.reset(); 3299 _timer.start(); 3300 do_work_steal(worker_id); 3301 _timer.stop(); 3302 if (PrintCMSStatistics != 0) { 3303 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3304 worker_id, _timer.seconds()); 3305 // XXX: need xxx/xxx type of notation, two timers 3306 } 3307 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3308 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3309 // Note that under the current task protocol, the 3310 // following assertion is true even of the spaces 3311 // expanded since the completion of the concurrent 3312 // marking. XXX This will likely change under a strict 3313 // ABORT semantics. 3314 // After perm removal the comparison was changed to 3315 // greater than or equal to from strictly greater than. 3316 // Before perm removal the highest address sweep would 3317 // have been at the end of perm gen but now is at the 3318 // end of the tenured gen. 3319 assert(_global_finger >= _cms_space->end(), 3320 "All tasks have been completed"); 3321 DEBUG_ONLY(_collector->verify_overflow_empty();) 3322 } 3323 3324 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3325 HeapWord* read = _global_finger; 3326 HeapWord* cur = read; 3327 while (f > read) { 3328 cur = read; 3329 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3330 if (cur == read) { 3331 // our cas succeeded 3332 assert(_global_finger >= f, "protocol consistency"); 3333 break; 3334 } 3335 } 3336 } 3337 3338 // This is really inefficient, and should be redone by 3339 // using (not yet available) block-read and -write interfaces to the 3340 // stack and the work_queue. XXX FIX ME !!! 3341 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3342 OopTaskQueue* work_q) { 3343 // Fast lock-free check 3344 if (ovflw_stk->length() == 0) { 3345 return false; 3346 } 3347 assert(work_q->size() == 0, "Shouldn't steal"); 3348 MutexLockerEx ml(ovflw_stk->par_lock(), 3349 Mutex::_no_safepoint_check_flag); 3350 // Grab up to 1/4 the size of the work queue 3351 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3352 (size_t)ParGCDesiredObjsFromOverflowList); 3353 num = MIN2(num, ovflw_stk->length()); 3354 for (int i = (int) num; i > 0; i--) { 3355 oop cur = ovflw_stk->pop(); 3356 assert(cur != NULL, "Counted wrong?"); 3357 work_q->push(cur); 3358 } 3359 return num > 0; 3360 } 3361 3362 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3363 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3364 int n_tasks = pst->n_tasks(); 3365 // We allow that there may be no tasks to do here because 3366 // we are restarting after a stack overflow. 3367 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3368 uint nth_task = 0; 3369 3370 HeapWord* aligned_start = sp->bottom(); 3371 if (sp->used_region().contains(_restart_addr)) { 3372 // Align down to a card boundary for the start of 0th task 3373 // for this space. 3374 aligned_start = 3375 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3376 CardTableModRefBS::card_size); 3377 } 3378 3379 size_t chunk_size = sp->marking_task_size(); 3380 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3381 // Having claimed the nth task in this space, 3382 // compute the chunk that it corresponds to: 3383 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3384 aligned_start + (nth_task+1)*chunk_size); 3385 // Try and bump the global finger via a CAS; 3386 // note that we need to do the global finger bump 3387 // _before_ taking the intersection below, because 3388 // the task corresponding to that region will be 3389 // deemed done even if the used_region() expands 3390 // because of allocation -- as it almost certainly will 3391 // during start-up while the threads yield in the 3392 // closure below. 3393 HeapWord* finger = span.end(); 3394 bump_global_finger(finger); // atomically 3395 // There are null tasks here corresponding to chunks 3396 // beyond the "top" address of the space. 3397 span = span.intersection(sp->used_region()); 3398 if (!span.is_empty()) { // Non-null task 3399 HeapWord* prev_obj; 3400 assert(!span.contains(_restart_addr) || nth_task == 0, 3401 "Inconsistency"); 3402 if (nth_task == 0) { 3403 // For the 0th task, we'll not need to compute a block_start. 3404 if (span.contains(_restart_addr)) { 3405 // In the case of a restart because of stack overflow, 3406 // we might additionally skip a chunk prefix. 3407 prev_obj = _restart_addr; 3408 } else { 3409 prev_obj = span.start(); 3410 } 3411 } else { 3412 // We want to skip the first object because 3413 // the protocol is to scan any object in its entirety 3414 // that _starts_ in this span; a fortiori, any 3415 // object starting in an earlier span is scanned 3416 // as part of an earlier claimed task. 3417 // Below we use the "careful" version of block_start 3418 // so we do not try to navigate uninitialized objects. 3419 prev_obj = sp->block_start_careful(span.start()); 3420 // Below we use a variant of block_size that uses the 3421 // Printezis bits to avoid waiting for allocated 3422 // objects to become initialized/parsable. 3423 while (prev_obj < span.start()) { 3424 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3425 if (sz > 0) { 3426 prev_obj += sz; 3427 } else { 3428 // In this case we may end up doing a bit of redundant 3429 // scanning, but that appears unavoidable, short of 3430 // locking the free list locks; see bug 6324141. 3431 break; 3432 } 3433 } 3434 } 3435 if (prev_obj < span.end()) { 3436 MemRegion my_span = MemRegion(prev_obj, span.end()); 3437 // Do the marking work within a non-empty span -- 3438 // the last argument to the constructor indicates whether the 3439 // iteration should be incremental with periodic yields. 3440 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3441 &_collector->_markBitMap, 3442 work_queue(i), 3443 &_collector->_markStack); 3444 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3445 } // else nothing to do for this task 3446 } // else nothing to do for this task 3447 } 3448 // We'd be tempted to assert here that since there are no 3449 // more tasks left to claim in this space, the global_finger 3450 // must exceed space->top() and a fortiori space->end(). However, 3451 // that would not quite be correct because the bumping of 3452 // global_finger occurs strictly after the claiming of a task, 3453 // so by the time we reach here the global finger may not yet 3454 // have been bumped up by the thread that claimed the last 3455 // task. 3456 pst->all_tasks_completed(); 3457 } 3458 3459 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3460 private: 3461 CMSCollector* _collector; 3462 CMSConcMarkingTask* _task; 3463 MemRegion _span; 3464 CMSBitMap* _bit_map; 3465 CMSMarkStack* _overflow_stack; 3466 OopTaskQueue* _work_queue; 3467 protected: 3468 DO_OOP_WORK_DEFN 3469 public: 3470 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3471 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3472 MetadataAwareOopClosure(collector->ref_processor()), 3473 _collector(collector), 3474 _task(task), 3475 _span(collector->_span), 3476 _work_queue(work_queue), 3477 _bit_map(bit_map), 3478 _overflow_stack(overflow_stack) 3479 { } 3480 virtual void do_oop(oop* p); 3481 virtual void do_oop(narrowOop* p); 3482 3483 void trim_queue(size_t max); 3484 void handle_stack_overflow(HeapWord* lost); 3485 void do_yield_check() { 3486 if (_task->should_yield()) { 3487 _task->yield(); 3488 } 3489 } 3490 }; 3491 3492 // Grey object scanning during work stealing phase -- 3493 // the salient assumption here is that any references 3494 // that are in these stolen objects being scanned must 3495 // already have been initialized (else they would not have 3496 // been published), so we do not need to check for 3497 // uninitialized objects before pushing here. 3498 void Par_ConcMarkingClosure::do_oop(oop obj) { 3499 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3500 HeapWord* addr = (HeapWord*)obj; 3501 // Check if oop points into the CMS generation 3502 // and is not marked 3503 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3504 // a white object ... 3505 // If we manage to "claim" the object, by being the 3506 // first thread to mark it, then we push it on our 3507 // marking stack 3508 if (_bit_map->par_mark(addr)) { // ... now grey 3509 // push on work queue (grey set) 3510 bool simulate_overflow = false; 3511 NOT_PRODUCT( 3512 if (CMSMarkStackOverflowALot && 3513 _collector->simulate_overflow()) { 3514 // simulate a stack overflow 3515 simulate_overflow = true; 3516 } 3517 ) 3518 if (simulate_overflow || 3519 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3520 // stack overflow 3521 if (PrintCMSStatistics != 0) { 3522 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3523 SIZE_FORMAT, _overflow_stack->capacity()); 3524 } 3525 // We cannot assert that the overflow stack is full because 3526 // it may have been emptied since. 3527 assert(simulate_overflow || 3528 _work_queue->size() == _work_queue->max_elems(), 3529 "Else push should have succeeded"); 3530 handle_stack_overflow(addr); 3531 } 3532 } // Else, some other thread got there first 3533 do_yield_check(); 3534 } 3535 } 3536 3537 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3538 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3539 3540 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3541 while (_work_queue->size() > max) { 3542 oop new_oop; 3543 if (_work_queue->pop_local(new_oop)) { 3544 assert(new_oop->is_oop(), "Should be an oop"); 3545 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3546 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3547 new_oop->oop_iterate(this); // do_oop() above 3548 do_yield_check(); 3549 } 3550 } 3551 } 3552 3553 // Upon stack overflow, we discard (part of) the stack, 3554 // remembering the least address amongst those discarded 3555 // in CMSCollector's _restart_address. 3556 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3557 // We need to do this under a mutex to prevent other 3558 // workers from interfering with the work done below. 3559 MutexLockerEx ml(_overflow_stack->par_lock(), 3560 Mutex::_no_safepoint_check_flag); 3561 // Remember the least grey address discarded 3562 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3563 _collector->lower_restart_addr(ra); 3564 _overflow_stack->reset(); // discard stack contents 3565 _overflow_stack->expand(); // expand the stack if possible 3566 } 3567 3568 3569 void CMSConcMarkingTask::do_work_steal(int i) { 3570 OopTaskQueue* work_q = work_queue(i); 3571 oop obj_to_scan; 3572 CMSBitMap* bm = &(_collector->_markBitMap); 3573 CMSMarkStack* ovflw = &(_collector->_markStack); 3574 int* seed = _collector->hash_seed(i); 3575 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3576 while (true) { 3577 cl.trim_queue(0); 3578 assert(work_q->size() == 0, "Should have been emptied above"); 3579 if (get_work_from_overflow_stack(ovflw, work_q)) { 3580 // Can't assert below because the work obtained from the 3581 // overflow stack may already have been stolen from us. 3582 // assert(work_q->size() > 0, "Work from overflow stack"); 3583 continue; 3584 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3585 assert(obj_to_scan->is_oop(), "Should be an oop"); 3586 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3587 obj_to_scan->oop_iterate(&cl); 3588 } else if (terminator()->offer_termination(&_term_term)) { 3589 assert(work_q->size() == 0, "Impossible!"); 3590 break; 3591 } else if (yielding() || should_yield()) { 3592 yield(); 3593 } 3594 } 3595 } 3596 3597 // This is run by the CMS (coordinator) thread. 3598 void CMSConcMarkingTask::coordinator_yield() { 3599 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3600 "CMS thread should hold CMS token"); 3601 // First give up the locks, then yield, then re-lock 3602 // We should probably use a constructor/destructor idiom to 3603 // do this unlock/lock or modify the MutexUnlocker class to 3604 // serve our purpose. XXX 3605 assert_lock_strong(_bit_map_lock); 3606 _bit_map_lock->unlock(); 3607 ConcurrentMarkSweepThread::desynchronize(true); 3608 _collector->stopTimer(); 3609 if (PrintCMSStatistics != 0) { 3610 _collector->incrementYields(); 3611 } 3612 3613 // It is possible for whichever thread initiated the yield request 3614 // not to get a chance to wake up and take the bitmap lock between 3615 // this thread releasing it and reacquiring it. So, while the 3616 // should_yield() flag is on, let's sleep for a bit to give the 3617 // other thread a chance to wake up. The limit imposed on the number 3618 // of iterations is defensive, to avoid any unforseen circumstances 3619 // putting us into an infinite loop. Since it's always been this 3620 // (coordinator_yield()) method that was observed to cause the 3621 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3622 // which is by default non-zero. For the other seven methods that 3623 // also perform the yield operation, as are using a different 3624 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3625 // can enable the sleeping for those methods too, if necessary. 3626 // See 6442774. 3627 // 3628 // We really need to reconsider the synchronization between the GC 3629 // thread and the yield-requesting threads in the future and we 3630 // should really use wait/notify, which is the recommended 3631 // way of doing this type of interaction. Additionally, we should 3632 // consolidate the eight methods that do the yield operation and they 3633 // are almost identical into one for better maintainability and 3634 // readability. See 6445193. 3635 // 3636 // Tony 2006.06.29 3637 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3638 ConcurrentMarkSweepThread::should_yield() && 3639 !CMSCollector::foregroundGCIsActive(); ++i) { 3640 os::sleep(Thread::current(), 1, false); 3641 } 3642 3643 ConcurrentMarkSweepThread::synchronize(true); 3644 _bit_map_lock->lock_without_safepoint_check(); 3645 _collector->startTimer(); 3646 } 3647 3648 bool CMSCollector::do_marking_mt() { 3649 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3650 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3651 conc_workers()->active_workers(), 3652 Threads::number_of_non_daemon_threads()); 3653 conc_workers()->set_active_workers(num_workers); 3654 3655 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3656 3657 CMSConcMarkingTask tsk(this, 3658 cms_space, 3659 conc_workers(), 3660 task_queues()); 3661 3662 // Since the actual number of workers we get may be different 3663 // from the number we requested above, do we need to do anything different 3664 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3665 // class?? XXX 3666 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3667 3668 // Refs discovery is already non-atomic. 3669 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3670 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3671 conc_workers()->start_task(&tsk); 3672 while (tsk.yielded()) { 3673 tsk.coordinator_yield(); 3674 conc_workers()->continue_task(&tsk); 3675 } 3676 // If the task was aborted, _restart_addr will be non-NULL 3677 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3678 while (_restart_addr != NULL) { 3679 // XXX For now we do not make use of ABORTED state and have not 3680 // yet implemented the right abort semantics (even in the original 3681 // single-threaded CMS case). That needs some more investigation 3682 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3683 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3684 // If _restart_addr is non-NULL, a marking stack overflow 3685 // occurred; we need to do a fresh marking iteration from the 3686 // indicated restart address. 3687 if (_foregroundGCIsActive) { 3688 // We may be running into repeated stack overflows, having 3689 // reached the limit of the stack size, while making very 3690 // slow forward progress. It may be best to bail out and 3691 // let the foreground collector do its job. 3692 // Clear _restart_addr, so that foreground GC 3693 // works from scratch. This avoids the headache of 3694 // a "rescan" which would otherwise be needed because 3695 // of the dirty mod union table & card table. 3696 _restart_addr = NULL; 3697 return false; 3698 } 3699 // Adjust the task to restart from _restart_addr 3700 tsk.reset(_restart_addr); 3701 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3702 _restart_addr); 3703 _restart_addr = NULL; 3704 // Get the workers going again 3705 conc_workers()->start_task(&tsk); 3706 while (tsk.yielded()) { 3707 tsk.coordinator_yield(); 3708 conc_workers()->continue_task(&tsk); 3709 } 3710 } 3711 assert(tsk.completed(), "Inconsistency"); 3712 assert(tsk.result() == true, "Inconsistency"); 3713 return true; 3714 } 3715 3716 bool CMSCollector::do_marking_st() { 3717 ResourceMark rm; 3718 HandleMark hm; 3719 3720 // Temporarily make refs discovery single threaded (non-MT) 3721 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3722 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3723 &_markStack, CMSYield); 3724 // the last argument to iterate indicates whether the iteration 3725 // should be incremental with periodic yields. 3726 _markBitMap.iterate(&markFromRootsClosure); 3727 // If _restart_addr is non-NULL, a marking stack overflow 3728 // occurred; we need to do a fresh iteration from the 3729 // indicated restart address. 3730 while (_restart_addr != NULL) { 3731 if (_foregroundGCIsActive) { 3732 // We may be running into repeated stack overflows, having 3733 // reached the limit of the stack size, while making very 3734 // slow forward progress. It may be best to bail out and 3735 // let the foreground collector do its job. 3736 // Clear _restart_addr, so that foreground GC 3737 // works from scratch. This avoids the headache of 3738 // a "rescan" which would otherwise be needed because 3739 // of the dirty mod union table & card table. 3740 _restart_addr = NULL; 3741 return false; // indicating failure to complete marking 3742 } 3743 // Deal with stack overflow: 3744 // we restart marking from _restart_addr 3745 HeapWord* ra = _restart_addr; 3746 markFromRootsClosure.reset(ra); 3747 _restart_addr = NULL; 3748 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3749 } 3750 return true; 3751 } 3752 3753 void CMSCollector::preclean() { 3754 check_correct_thread_executing(); 3755 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3756 verify_work_stacks_empty(); 3757 verify_overflow_empty(); 3758 _abort_preclean = false; 3759 if (CMSPrecleaningEnabled) { 3760 if (!CMSEdenChunksRecordAlways) { 3761 _eden_chunk_index = 0; 3762 } 3763 size_t used = get_eden_used(); 3764 size_t capacity = get_eden_capacity(); 3765 // Don't start sampling unless we will get sufficiently 3766 // many samples. 3767 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3768 * CMSScheduleRemarkEdenPenetration)) { 3769 _start_sampling = true; 3770 } else { 3771 _start_sampling = false; 3772 } 3773 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3774 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3775 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3776 } 3777 CMSTokenSync x(true); // is cms thread 3778 if (CMSPrecleaningEnabled) { 3779 sample_eden(); 3780 _collectorState = AbortablePreclean; 3781 } else { 3782 _collectorState = FinalMarking; 3783 } 3784 verify_work_stacks_empty(); 3785 verify_overflow_empty(); 3786 } 3787 3788 // Try and schedule the remark such that young gen 3789 // occupancy is CMSScheduleRemarkEdenPenetration %. 3790 void CMSCollector::abortable_preclean() { 3791 check_correct_thread_executing(); 3792 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3793 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3794 3795 // If Eden's current occupancy is below this threshold, 3796 // immediately schedule the remark; else preclean 3797 // past the next scavenge in an effort to 3798 // schedule the pause as described above. By choosing 3799 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3800 // we will never do an actual abortable preclean cycle. 3801 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3802 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3803 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3804 // We need more smarts in the abortable preclean 3805 // loop below to deal with cases where allocation 3806 // in young gen is very very slow, and our precleaning 3807 // is running a losing race against a horde of 3808 // mutators intent on flooding us with CMS updates 3809 // (dirty cards). 3810 // One, admittedly dumb, strategy is to give up 3811 // after a certain number of abortable precleaning loops 3812 // or after a certain maximum time. We want to make 3813 // this smarter in the next iteration. 3814 // XXX FIX ME!!! YSR 3815 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3816 while (!(should_abort_preclean() || 3817 ConcurrentMarkSweepThread::should_terminate())) { 3818 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3819 cumworkdone += workdone; 3820 loops++; 3821 // Voluntarily terminate abortable preclean phase if we have 3822 // been at it for too long. 3823 if ((CMSMaxAbortablePrecleanLoops != 0) && 3824 loops >= CMSMaxAbortablePrecleanLoops) { 3825 if (PrintGCDetails) { 3826 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3827 } 3828 break; 3829 } 3830 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3831 if (PrintGCDetails) { 3832 gclog_or_tty->print(" CMS: abort preclean due to time "); 3833 } 3834 break; 3835 } 3836 // If we are doing little work each iteration, we should 3837 // take a short break. 3838 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3839 // Sleep for some time, waiting for work to accumulate 3840 stopTimer(); 3841 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3842 startTimer(); 3843 waited++; 3844 } 3845 } 3846 if (PrintCMSStatistics > 0) { 3847 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3848 loops, waited, cumworkdone); 3849 } 3850 } 3851 CMSTokenSync x(true); // is cms thread 3852 if (_collectorState != Idling) { 3853 assert(_collectorState == AbortablePreclean, 3854 "Spontaneous state transition?"); 3855 _collectorState = FinalMarking; 3856 } // Else, a foreground collection completed this CMS cycle. 3857 return; 3858 } 3859 3860 // Respond to an Eden sampling opportunity 3861 void CMSCollector::sample_eden() { 3862 // Make sure a young gc cannot sneak in between our 3863 // reading and recording of a sample. 3864 assert(Thread::current()->is_ConcurrentGC_thread(), 3865 "Only the cms thread may collect Eden samples"); 3866 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3867 "Should collect samples while holding CMS token"); 3868 if (!_start_sampling) { 3869 return; 3870 } 3871 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3872 // is populated by the young generation. 3873 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3874 if (_eden_chunk_index < _eden_chunk_capacity) { 3875 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3876 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3877 "Unexpected state of Eden"); 3878 // We'd like to check that what we just sampled is an oop-start address; 3879 // however, we cannot do that here since the object may not yet have been 3880 // initialized. So we'll instead do the check when we _use_ this sample 3881 // later. 3882 if (_eden_chunk_index == 0 || 3883 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3884 _eden_chunk_array[_eden_chunk_index-1]) 3885 >= CMSSamplingGrain)) { 3886 _eden_chunk_index++; // commit sample 3887 } 3888 } 3889 } 3890 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3891 size_t used = get_eden_used(); 3892 size_t capacity = get_eden_capacity(); 3893 assert(used <= capacity, "Unexpected state of Eden"); 3894 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3895 _abort_preclean = true; 3896 } 3897 } 3898 } 3899 3900 3901 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3902 assert(_collectorState == Precleaning || 3903 _collectorState == AbortablePreclean, "incorrect state"); 3904 ResourceMark rm; 3905 HandleMark hm; 3906 3907 // Precleaning is currently not MT but the reference processor 3908 // may be set for MT. Disable it temporarily here. 3909 ReferenceProcessor* rp = ref_processor(); 3910 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3911 3912 // Do one pass of scrubbing the discovered reference lists 3913 // to remove any reference objects with strongly-reachable 3914 // referents. 3915 if (clean_refs) { 3916 CMSPrecleanRefsYieldClosure yield_cl(this); 3917 assert(rp->span().equals(_span), "Spans should be equal"); 3918 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3919 &_markStack, true /* preclean */); 3920 CMSDrainMarkingStackClosure complete_trace(this, 3921 _span, &_markBitMap, &_markStack, 3922 &keep_alive, true /* preclean */); 3923 3924 // We don't want this step to interfere with a young 3925 // collection because we don't want to take CPU 3926 // or memory bandwidth away from the young GC threads 3927 // (which may be as many as there are CPUs). 3928 // Note that we don't need to protect ourselves from 3929 // interference with mutators because they can't 3930 // manipulate the discovered reference lists nor affect 3931 // the computed reachability of the referents, the 3932 // only properties manipulated by the precleaning 3933 // of these reference lists. 3934 stopTimer(); 3935 CMSTokenSyncWithLocks x(true /* is cms thread */, 3936 bitMapLock()); 3937 startTimer(); 3938 sample_eden(); 3939 3940 // The following will yield to allow foreground 3941 // collection to proceed promptly. XXX YSR: 3942 // The code in this method may need further 3943 // tweaking for better performance and some restructuring 3944 // for cleaner interfaces. 3945 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3946 rp->preclean_discovered_references( 3947 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3948 gc_timer, _gc_tracer_cm->gc_id()); 3949 } 3950 3951 if (clean_survivor) { // preclean the active survivor space(s) 3952 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3953 &_markBitMap, &_modUnionTable, 3954 &_markStack, true /* precleaning phase */); 3955 stopTimer(); 3956 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3957 bitMapLock()); 3958 startTimer(); 3959 unsigned int before_count = 3960 GenCollectedHeap::heap()->total_collections(); 3961 SurvivorSpacePrecleanClosure 3962 sss_cl(this, _span, &_markBitMap, &_markStack, 3963 &pam_cl, before_count, CMSYield); 3964 _young_gen->from()->object_iterate_careful(&sss_cl); 3965 _young_gen->to()->object_iterate_careful(&sss_cl); 3966 } 3967 MarkRefsIntoAndScanClosure 3968 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3969 &_markStack, this, CMSYield, 3970 true /* precleaning phase */); 3971 // CAUTION: The following closure has persistent state that may need to 3972 // be reset upon a decrease in the sequence of addresses it 3973 // processes. 3974 ScanMarkedObjectsAgainCarefullyClosure 3975 smoac_cl(this, _span, 3976 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3977 3978 // Preclean dirty cards in ModUnionTable and CardTable using 3979 // appropriate convergence criterion; 3980 // repeat CMSPrecleanIter times unless we find that 3981 // we are losing. 3982 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3983 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3984 "Bad convergence multiplier"); 3985 assert(CMSPrecleanThreshold >= 100, 3986 "Unreasonably low CMSPrecleanThreshold"); 3987 3988 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3989 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3990 numIter < CMSPrecleanIter; 3991 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3992 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3993 if (Verbose && PrintGCDetails) { 3994 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3995 } 3996 // Either there are very few dirty cards, so re-mark 3997 // pause will be small anyway, or our pre-cleaning isn't 3998 // that much faster than the rate at which cards are being 3999 // dirtied, so we might as well stop and re-mark since 4000 // precleaning won't improve our re-mark time by much. 4001 if (curNumCards <= CMSPrecleanThreshold || 4002 (numIter > 0 && 4003 (curNumCards * CMSPrecleanDenominator > 4004 lastNumCards * CMSPrecleanNumerator))) { 4005 numIter++; 4006 cumNumCards += curNumCards; 4007 break; 4008 } 4009 } 4010 4011 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4012 4013 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4014 cumNumCards += curNumCards; 4015 if (PrintGCDetails && PrintCMSStatistics != 0) { 4016 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4017 curNumCards, cumNumCards, numIter); 4018 } 4019 return cumNumCards; // as a measure of useful work done 4020 } 4021 4022 // PRECLEANING NOTES: 4023 // Precleaning involves: 4024 // . reading the bits of the modUnionTable and clearing the set bits. 4025 // . For the cards corresponding to the set bits, we scan the 4026 // objects on those cards. This means we need the free_list_lock 4027 // so that we can safely iterate over the CMS space when scanning 4028 // for oops. 4029 // . When we scan the objects, we'll be both reading and setting 4030 // marks in the marking bit map, so we'll need the marking bit map. 4031 // . For protecting _collector_state transitions, we take the CGC_lock. 4032 // Note that any races in the reading of of card table entries by the 4033 // CMS thread on the one hand and the clearing of those entries by the 4034 // VM thread or the setting of those entries by the mutator threads on the 4035 // other are quite benign. However, for efficiency it makes sense to keep 4036 // the VM thread from racing with the CMS thread while the latter is 4037 // dirty card info to the modUnionTable. We therefore also use the 4038 // CGC_lock to protect the reading of the card table and the mod union 4039 // table by the CM thread. 4040 // . We run concurrently with mutator updates, so scanning 4041 // needs to be done carefully -- we should not try to scan 4042 // potentially uninitialized objects. 4043 // 4044 // Locking strategy: While holding the CGC_lock, we scan over and 4045 // reset a maximal dirty range of the mod union / card tables, then lock 4046 // the free_list_lock and bitmap lock to do a full marking, then 4047 // release these locks; and repeat the cycle. This allows for a 4048 // certain amount of fairness in the sharing of these locks between 4049 // the CMS collector on the one hand, and the VM thread and the 4050 // mutators on the other. 4051 4052 // NOTE: preclean_mod_union_table() and preclean_card_table() 4053 // further below are largely identical; if you need to modify 4054 // one of these methods, please check the other method too. 4055 4056 size_t CMSCollector::preclean_mod_union_table( 4057 ConcurrentMarkSweepGeneration* gen, 4058 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4059 verify_work_stacks_empty(); 4060 verify_overflow_empty(); 4061 4062 // strategy: starting with the first card, accumulate contiguous 4063 // ranges of dirty cards; clear these cards, then scan the region 4064 // covered by these cards. 4065 4066 // Since all of the MUT is committed ahead, we can just use 4067 // that, in case the generations expand while we are precleaning. 4068 // It might also be fine to just use the committed part of the 4069 // generation, but we might potentially miss cards when the 4070 // generation is rapidly expanding while we are in the midst 4071 // of precleaning. 4072 HeapWord* startAddr = gen->reserved().start(); 4073 HeapWord* endAddr = gen->reserved().end(); 4074 4075 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4076 4077 size_t numDirtyCards, cumNumDirtyCards; 4078 HeapWord *nextAddr, *lastAddr; 4079 for (cumNumDirtyCards = numDirtyCards = 0, 4080 nextAddr = lastAddr = startAddr; 4081 nextAddr < endAddr; 4082 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4083 4084 ResourceMark rm; 4085 HandleMark hm; 4086 4087 MemRegion dirtyRegion; 4088 { 4089 stopTimer(); 4090 // Potential yield point 4091 CMSTokenSync ts(true); 4092 startTimer(); 4093 sample_eden(); 4094 // Get dirty region starting at nextOffset (inclusive), 4095 // simultaneously clearing it. 4096 dirtyRegion = 4097 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4098 assert(dirtyRegion.start() >= nextAddr, 4099 "returned region inconsistent?"); 4100 } 4101 // Remember where the next search should begin. 4102 // The returned region (if non-empty) is a right open interval, 4103 // so lastOffset is obtained from the right end of that 4104 // interval. 4105 lastAddr = dirtyRegion.end(); 4106 // Should do something more transparent and less hacky XXX 4107 numDirtyCards = 4108 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4109 4110 // We'll scan the cards in the dirty region (with periodic 4111 // yields for foreground GC as needed). 4112 if (!dirtyRegion.is_empty()) { 4113 assert(numDirtyCards > 0, "consistency check"); 4114 HeapWord* stop_point = NULL; 4115 stopTimer(); 4116 // Potential yield point 4117 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4118 bitMapLock()); 4119 startTimer(); 4120 { 4121 verify_work_stacks_empty(); 4122 verify_overflow_empty(); 4123 sample_eden(); 4124 stop_point = 4125 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4126 } 4127 if (stop_point != NULL) { 4128 // The careful iteration stopped early either because it found an 4129 // uninitialized object, or because we were in the midst of an 4130 // "abortable preclean", which should now be aborted. Redirty 4131 // the bits corresponding to the partially-scanned or unscanned 4132 // cards. We'll either restart at the next block boundary or 4133 // abort the preclean. 4134 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4135 "Should only be AbortablePreclean."); 4136 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4137 if (should_abort_preclean()) { 4138 break; // out of preclean loop 4139 } else { 4140 // Compute the next address at which preclean should pick up; 4141 // might need bitMapLock in order to read P-bits. 4142 lastAddr = next_card_start_after_block(stop_point); 4143 } 4144 } 4145 } else { 4146 assert(lastAddr == endAddr, "consistency check"); 4147 assert(numDirtyCards == 0, "consistency check"); 4148 break; 4149 } 4150 } 4151 verify_work_stacks_empty(); 4152 verify_overflow_empty(); 4153 return cumNumDirtyCards; 4154 } 4155 4156 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4157 // below are largely identical; if you need to modify 4158 // one of these methods, please check the other method too. 4159 4160 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4161 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4162 // strategy: it's similar to precleamModUnionTable above, in that 4163 // we accumulate contiguous ranges of dirty cards, mark these cards 4164 // precleaned, then scan the region covered by these cards. 4165 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4166 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4167 4168 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4169 4170 size_t numDirtyCards, cumNumDirtyCards; 4171 HeapWord *lastAddr, *nextAddr; 4172 4173 for (cumNumDirtyCards = numDirtyCards = 0, 4174 nextAddr = lastAddr = startAddr; 4175 nextAddr < endAddr; 4176 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4177 4178 ResourceMark rm; 4179 HandleMark hm; 4180 4181 MemRegion dirtyRegion; 4182 { 4183 // See comments in "Precleaning notes" above on why we 4184 // do this locking. XXX Could the locking overheads be 4185 // too high when dirty cards are sparse? [I don't think so.] 4186 stopTimer(); 4187 CMSTokenSync x(true); // is cms thread 4188 startTimer(); 4189 sample_eden(); 4190 // Get and clear dirty region from card table 4191 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4192 MemRegion(nextAddr, endAddr), 4193 true, 4194 CardTableModRefBS::precleaned_card_val()); 4195 4196 assert(dirtyRegion.start() >= nextAddr, 4197 "returned region inconsistent?"); 4198 } 4199 lastAddr = dirtyRegion.end(); 4200 numDirtyCards = 4201 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4202 4203 if (!dirtyRegion.is_empty()) { 4204 stopTimer(); 4205 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4206 startTimer(); 4207 sample_eden(); 4208 verify_work_stacks_empty(); 4209 verify_overflow_empty(); 4210 HeapWord* stop_point = 4211 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4212 if (stop_point != NULL) { 4213 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4214 "Should only be AbortablePreclean."); 4215 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4216 if (should_abort_preclean()) { 4217 break; // out of preclean loop 4218 } else { 4219 // Compute the next address at which preclean should pick up. 4220 lastAddr = next_card_start_after_block(stop_point); 4221 } 4222 } 4223 } else { 4224 break; 4225 } 4226 } 4227 verify_work_stacks_empty(); 4228 verify_overflow_empty(); 4229 return cumNumDirtyCards; 4230 } 4231 4232 class PrecleanKlassClosure : public KlassClosure { 4233 KlassToOopClosure _cm_klass_closure; 4234 public: 4235 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4236 void do_klass(Klass* k) { 4237 if (k->has_accumulated_modified_oops()) { 4238 k->clear_accumulated_modified_oops(); 4239 4240 _cm_klass_closure.do_klass(k); 4241 } 4242 } 4243 }; 4244 4245 // The freelist lock is needed to prevent asserts, is it really needed? 4246 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4247 4248 cl->set_freelistLock(freelistLock); 4249 4250 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4251 4252 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4253 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4254 PrecleanKlassClosure preclean_klass_closure(cl); 4255 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4256 4257 verify_work_stacks_empty(); 4258 verify_overflow_empty(); 4259 } 4260 4261 void CMSCollector::checkpointRootsFinal() { 4262 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4263 check_correct_thread_executing(); 4264 // world is stopped at this checkpoint 4265 assert(SafepointSynchronize::is_at_safepoint(), 4266 "world should be stopped"); 4267 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4268 4269 verify_work_stacks_empty(); 4270 verify_overflow_empty(); 4271 4272 if (PrintGCDetails) { 4273 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4274 _young_gen->used() / K, 4275 _young_gen->capacity() / K); 4276 } 4277 { 4278 if (CMSScavengeBeforeRemark) { 4279 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4280 // Temporarily set flag to false, GCH->do_collection will 4281 // expect it to be false and set to true 4282 FlagSetting fl(gch->_is_gc_active, false); 4283 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4284 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4285 int level = _cmsGen->level() - 1; 4286 if (level >= 0) { 4287 gch->do_collection(true, // full (i.e. force, see below) 4288 false, // !clear_all_soft_refs 4289 0, // size 4290 false, // is_tlab 4291 level // max_level 4292 ); 4293 } 4294 } 4295 FreelistLocker x(this); 4296 MutexLockerEx y(bitMapLock(), 4297 Mutex::_no_safepoint_check_flag); 4298 checkpointRootsFinalWork(); 4299 } 4300 verify_work_stacks_empty(); 4301 verify_overflow_empty(); 4302 } 4303 4304 void CMSCollector::checkpointRootsFinalWork() { 4305 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4306 4307 assert(haveFreelistLocks(), "must have free list locks"); 4308 assert_lock_strong(bitMapLock()); 4309 4310 ResourceMark rm; 4311 HandleMark hm; 4312 4313 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4314 4315 if (should_unload_classes()) { 4316 CodeCache::gc_prologue(); 4317 } 4318 assert(haveFreelistLocks(), "must have free list locks"); 4319 assert_lock_strong(bitMapLock()); 4320 4321 // We might assume that we need not fill TLAB's when 4322 // CMSScavengeBeforeRemark is set, because we may have just done 4323 // a scavenge which would have filled all TLAB's -- and besides 4324 // Eden would be empty. This however may not always be the case -- 4325 // for instance although we asked for a scavenge, it may not have 4326 // happened because of a JNI critical section. We probably need 4327 // a policy for deciding whether we can in that case wait until 4328 // the critical section releases and then do the remark following 4329 // the scavenge, and skip it here. In the absence of that policy, 4330 // or of an indication of whether the scavenge did indeed occur, 4331 // we cannot rely on TLAB's having been filled and must do 4332 // so here just in case a scavenge did not happen. 4333 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4334 // Update the saved marks which may affect the root scans. 4335 gch->save_marks(); 4336 4337 if (CMSPrintEdenSurvivorChunks) { 4338 print_eden_and_survivor_chunk_arrays(); 4339 } 4340 4341 { 4342 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4343 4344 // Note on the role of the mod union table: 4345 // Since the marker in "markFromRoots" marks concurrently with 4346 // mutators, it is possible for some reachable objects not to have been 4347 // scanned. For instance, an only reference to an object A was 4348 // placed in object B after the marker scanned B. Unless B is rescanned, 4349 // A would be collected. Such updates to references in marked objects 4350 // are detected via the mod union table which is the set of all cards 4351 // dirtied since the first checkpoint in this GC cycle and prior to 4352 // the most recent young generation GC, minus those cleaned up by the 4353 // concurrent precleaning. 4354 if (CMSParallelRemarkEnabled) { 4355 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4356 do_remark_parallel(); 4357 } else { 4358 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4359 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4360 do_remark_non_parallel(); 4361 } 4362 } 4363 verify_work_stacks_empty(); 4364 verify_overflow_empty(); 4365 4366 { 4367 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4368 refProcessingWork(); 4369 } 4370 verify_work_stacks_empty(); 4371 verify_overflow_empty(); 4372 4373 if (should_unload_classes()) { 4374 CodeCache::gc_epilogue(); 4375 } 4376 JvmtiExport::gc_epilogue(); 4377 4378 // If we encountered any (marking stack / work queue) overflow 4379 // events during the current CMS cycle, take appropriate 4380 // remedial measures, where possible, so as to try and avoid 4381 // recurrence of that condition. 4382 assert(_markStack.isEmpty(), "No grey objects"); 4383 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4384 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4385 if (ser_ovflw > 0) { 4386 if (PrintCMSStatistics != 0) { 4387 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4388 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4389 ", kac_preclean="SIZE_FORMAT")", 4390 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4391 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4392 } 4393 _markStack.expand(); 4394 _ser_pmc_remark_ovflw = 0; 4395 _ser_pmc_preclean_ovflw = 0; 4396 _ser_kac_preclean_ovflw = 0; 4397 _ser_kac_ovflw = 0; 4398 } 4399 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4400 if (PrintCMSStatistics != 0) { 4401 gclog_or_tty->print_cr("Work queue overflow (benign) " 4402 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4403 _par_pmc_remark_ovflw, _par_kac_ovflw); 4404 } 4405 _par_pmc_remark_ovflw = 0; 4406 _par_kac_ovflw = 0; 4407 } 4408 if (PrintCMSStatistics != 0) { 4409 if (_markStack._hit_limit > 0) { 4410 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4411 _markStack._hit_limit); 4412 } 4413 if (_markStack._failed_double > 0) { 4414 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4415 " current capacity "SIZE_FORMAT, 4416 _markStack._failed_double, 4417 _markStack.capacity()); 4418 } 4419 } 4420 _markStack._hit_limit = 0; 4421 _markStack._failed_double = 0; 4422 4423 if ((VerifyAfterGC || VerifyDuringGC) && 4424 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4425 verify_after_remark(); 4426 } 4427 4428 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4429 4430 // Change under the freelistLocks. 4431 _collectorState = Sweeping; 4432 // Call isAllClear() under bitMapLock 4433 assert(_modUnionTable.isAllClear(), 4434 "Should be clear by end of the final marking"); 4435 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4436 "Should be clear by end of the final marking"); 4437 } 4438 4439 void CMSParInitialMarkTask::work(uint worker_id) { 4440 elapsedTimer _timer; 4441 ResourceMark rm; 4442 HandleMark hm; 4443 4444 // ---------- scan from roots -------------- 4445 _timer.start(); 4446 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4447 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4448 4449 // ---------- young gen roots -------------- 4450 { 4451 work_on_young_gen_roots(worker_id, &par_mri_cl); 4452 _timer.stop(); 4453 if (PrintCMSStatistics != 0) { 4454 gclog_or_tty->print_cr( 4455 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4456 worker_id, _timer.seconds()); 4457 } 4458 } 4459 4460 // ---------- remaining roots -------------- 4461 _timer.reset(); 4462 _timer.start(); 4463 4464 CLDToOopClosure cld_closure(&par_mri_cl, true); 4465 4466 gch->gen_process_roots(_strong_roots_scope, 4467 _collector->_cmsGen->level(), 4468 false, // yg was scanned above 4469 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4470 _collector->should_unload_classes(), 4471 &par_mri_cl, 4472 NULL, 4473 &cld_closure); 4474 assert(_collector->should_unload_classes() 4475 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4476 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4477 _timer.stop(); 4478 if (PrintCMSStatistics != 0) { 4479 gclog_or_tty->print_cr( 4480 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4481 worker_id, _timer.seconds()); 4482 } 4483 } 4484 4485 // Parallel remark task 4486 class CMSParRemarkTask: public CMSParMarkTask { 4487 CompactibleFreeListSpace* _cms_space; 4488 4489 // The per-thread work queues, available here for stealing. 4490 OopTaskQueueSet* _task_queues; 4491 ParallelTaskTerminator _term; 4492 StrongRootsScope* _strong_roots_scope; 4493 4494 public: 4495 // A value of 0 passed to n_workers will cause the number of 4496 // workers to be taken from the active workers in the work gang. 4497 CMSParRemarkTask(CMSCollector* collector, 4498 CompactibleFreeListSpace* cms_space, 4499 uint n_workers, FlexibleWorkGang* workers, 4500 OopTaskQueueSet* task_queues, 4501 StrongRootsScope* strong_roots_scope): 4502 CMSParMarkTask("Rescan roots and grey objects in parallel", 4503 collector, n_workers), 4504 _cms_space(cms_space), 4505 _task_queues(task_queues), 4506 _term(n_workers, task_queues), 4507 _strong_roots_scope(strong_roots_scope) { } 4508 4509 OopTaskQueueSet* task_queues() { return _task_queues; } 4510 4511 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4512 4513 ParallelTaskTerminator* terminator() { return &_term; } 4514 uint n_workers() { return _n_workers; } 4515 4516 void work(uint worker_id); 4517 4518 private: 4519 // ... of dirty cards in old space 4520 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4521 Par_MarkRefsIntoAndScanClosure* cl); 4522 4523 // ... work stealing for the above 4524 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4525 }; 4526 4527 class RemarkKlassClosure : public KlassClosure { 4528 KlassToOopClosure _cm_klass_closure; 4529 public: 4530 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4531 void do_klass(Klass* k) { 4532 // Check if we have modified any oops in the Klass during the concurrent marking. 4533 if (k->has_accumulated_modified_oops()) { 4534 k->clear_accumulated_modified_oops(); 4535 4536 // We could have transfered the current modified marks to the accumulated marks, 4537 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4538 } else if (k->has_modified_oops()) { 4539 // Don't clear anything, this info is needed by the next young collection. 4540 } else { 4541 // No modified oops in the Klass. 4542 return; 4543 } 4544 4545 // The klass has modified fields, need to scan the klass. 4546 _cm_klass_closure.do_klass(k); 4547 } 4548 }; 4549 4550 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4551 ParNewGeneration* young_gen = _collector->_young_gen; 4552 ContiguousSpace* eden_space = young_gen->eden(); 4553 ContiguousSpace* from_space = young_gen->from(); 4554 ContiguousSpace* to_space = young_gen->to(); 4555 4556 HeapWord** eca = _collector->_eden_chunk_array; 4557 size_t ect = _collector->_eden_chunk_index; 4558 HeapWord** sca = _collector->_survivor_chunk_array; 4559 size_t sct = _collector->_survivor_chunk_index; 4560 4561 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4562 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4563 4564 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4565 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4566 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4567 } 4568 4569 // work_queue(i) is passed to the closure 4570 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4571 // also is passed to do_dirty_card_rescan_tasks() and to 4572 // do_work_steal() to select the i-th task_queue. 4573 4574 void CMSParRemarkTask::work(uint worker_id) { 4575 elapsedTimer _timer; 4576 ResourceMark rm; 4577 HandleMark hm; 4578 4579 // ---------- rescan from roots -------------- 4580 _timer.start(); 4581 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4582 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4583 _collector->_span, _collector->ref_processor(), 4584 &(_collector->_markBitMap), 4585 work_queue(worker_id)); 4586 4587 // Rescan young gen roots first since these are likely 4588 // coarsely partitioned and may, on that account, constitute 4589 // the critical path; thus, it's best to start off that 4590 // work first. 4591 // ---------- young gen roots -------------- 4592 { 4593 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4594 _timer.stop(); 4595 if (PrintCMSStatistics != 0) { 4596 gclog_or_tty->print_cr( 4597 "Finished young gen rescan work in %dth thread: %3.3f sec", 4598 worker_id, _timer.seconds()); 4599 } 4600 } 4601 4602 // ---------- remaining roots -------------- 4603 _timer.reset(); 4604 _timer.start(); 4605 gch->gen_process_roots(_strong_roots_scope, 4606 _collector->_cmsGen->level(), 4607 false, // yg was scanned above 4608 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4609 _collector->should_unload_classes(), 4610 &par_mrias_cl, 4611 NULL, 4612 NULL); // The dirty klasses will be handled below 4613 4614 assert(_collector->should_unload_classes() 4615 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4616 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4617 _timer.stop(); 4618 if (PrintCMSStatistics != 0) { 4619 gclog_or_tty->print_cr( 4620 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4621 worker_id, _timer.seconds()); 4622 } 4623 4624 // ---------- unhandled CLD scanning ---------- 4625 if (worker_id == 0) { // Single threaded at the moment. 4626 _timer.reset(); 4627 _timer.start(); 4628 4629 // Scan all new class loader data objects and new dependencies that were 4630 // introduced during concurrent marking. 4631 ResourceMark rm; 4632 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4633 for (int i = 0; i < array->length(); i++) { 4634 par_mrias_cl.do_class_loader_data(array->at(i)); 4635 } 4636 4637 // We don't need to keep track of new CLDs anymore. 4638 ClassLoaderDataGraph::remember_new_clds(false); 4639 4640 _timer.stop(); 4641 if (PrintCMSStatistics != 0) { 4642 gclog_or_tty->print_cr( 4643 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4644 worker_id, _timer.seconds()); 4645 } 4646 } 4647 4648 // ---------- dirty klass scanning ---------- 4649 if (worker_id == 0) { // Single threaded at the moment. 4650 _timer.reset(); 4651 _timer.start(); 4652 4653 // Scan all classes that was dirtied during the concurrent marking phase. 4654 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4655 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4656 4657 _timer.stop(); 4658 if (PrintCMSStatistics != 0) { 4659 gclog_or_tty->print_cr( 4660 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4661 worker_id, _timer.seconds()); 4662 } 4663 } 4664 4665 // We might have added oops to ClassLoaderData::_handles during the 4666 // concurrent marking phase. These oops point to newly allocated objects 4667 // that are guaranteed to be kept alive. Either by the direct allocation 4668 // code, or when the young collector processes the roots. Hence, 4669 // we don't have to revisit the _handles block during the remark phase. 4670 4671 // ---------- rescan dirty cards ------------ 4672 _timer.reset(); 4673 _timer.start(); 4674 4675 // Do the rescan tasks for each of the two spaces 4676 // (cms_space) in turn. 4677 // "worker_id" is passed to select the task_queue for "worker_id" 4678 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4679 _timer.stop(); 4680 if (PrintCMSStatistics != 0) { 4681 gclog_or_tty->print_cr( 4682 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4683 worker_id, _timer.seconds()); 4684 } 4685 4686 // ---------- steal work from other threads ... 4687 // ---------- ... and drain overflow list. 4688 _timer.reset(); 4689 _timer.start(); 4690 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4691 _timer.stop(); 4692 if (PrintCMSStatistics != 0) { 4693 gclog_or_tty->print_cr( 4694 "Finished work stealing in %dth thread: %3.3f sec", 4695 worker_id, _timer.seconds()); 4696 } 4697 } 4698 4699 // Note that parameter "i" is not used. 4700 void 4701 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4702 OopsInGenClosure* cl, ContiguousSpace* space, 4703 HeapWord** chunk_array, size_t chunk_top) { 4704 // Until all tasks completed: 4705 // . claim an unclaimed task 4706 // . compute region boundaries corresponding to task claimed 4707 // using chunk_array 4708 // . par_oop_iterate(cl) over that region 4709 4710 ResourceMark rm; 4711 HandleMark hm; 4712 4713 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4714 4715 uint nth_task = 0; 4716 uint n_tasks = pst->n_tasks(); 4717 4718 if (n_tasks > 0) { 4719 assert(pst->valid(), "Uninitialized use?"); 4720 HeapWord *start, *end; 4721 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4722 // We claimed task # nth_task; compute its boundaries. 4723 if (chunk_top == 0) { // no samples were taken 4724 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4725 start = space->bottom(); 4726 end = space->top(); 4727 } else if (nth_task == 0) { 4728 start = space->bottom(); 4729 end = chunk_array[nth_task]; 4730 } else if (nth_task < (uint)chunk_top) { 4731 assert(nth_task >= 1, "Control point invariant"); 4732 start = chunk_array[nth_task - 1]; 4733 end = chunk_array[nth_task]; 4734 } else { 4735 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4736 start = chunk_array[chunk_top - 1]; 4737 end = space->top(); 4738 } 4739 MemRegion mr(start, end); 4740 // Verify that mr is in space 4741 assert(mr.is_empty() || space->used_region().contains(mr), 4742 "Should be in space"); 4743 // Verify that "start" is an object boundary 4744 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4745 "Should be an oop"); 4746 space->par_oop_iterate(mr, cl); 4747 } 4748 pst->all_tasks_completed(); 4749 } 4750 } 4751 4752 void 4753 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4754 CompactibleFreeListSpace* sp, int i, 4755 Par_MarkRefsIntoAndScanClosure* cl) { 4756 // Until all tasks completed: 4757 // . claim an unclaimed task 4758 // . compute region boundaries corresponding to task claimed 4759 // . transfer dirty bits ct->mut for that region 4760 // . apply rescanclosure to dirty mut bits for that region 4761 4762 ResourceMark rm; 4763 HandleMark hm; 4764 4765 OopTaskQueue* work_q = work_queue(i); 4766 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4767 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4768 // CAUTION: This closure has state that persists across calls to 4769 // the work method dirty_range_iterate_clear() in that it has 4770 // embedded in it a (subtype of) UpwardsObjectClosure. The 4771 // use of that state in the embedded UpwardsObjectClosure instance 4772 // assumes that the cards are always iterated (even if in parallel 4773 // by several threads) in monotonically increasing order per each 4774 // thread. This is true of the implementation below which picks 4775 // card ranges (chunks) in monotonically increasing order globally 4776 // and, a-fortiori, in monotonically increasing order per thread 4777 // (the latter order being a subsequence of the former). 4778 // If the work code below is ever reorganized into a more chaotic 4779 // work-partitioning form than the current "sequential tasks" 4780 // paradigm, the use of that persistent state will have to be 4781 // revisited and modified appropriately. See also related 4782 // bug 4756801 work on which should examine this code to make 4783 // sure that the changes there do not run counter to the 4784 // assumptions made here and necessary for correctness and 4785 // efficiency. Note also that this code might yield inefficient 4786 // behavior in the case of very large objects that span one or 4787 // more work chunks. Such objects would potentially be scanned 4788 // several times redundantly. Work on 4756801 should try and 4789 // address that performance anomaly if at all possible. XXX 4790 MemRegion full_span = _collector->_span; 4791 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4792 MarkFromDirtyCardsClosure 4793 greyRescanClosure(_collector, full_span, // entire span of interest 4794 sp, bm, work_q, cl); 4795 4796 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4797 assert(pst->valid(), "Uninitialized use?"); 4798 uint nth_task = 0; 4799 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4800 MemRegion span = sp->used_region(); 4801 HeapWord* start_addr = span.start(); 4802 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4803 alignment); 4804 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4805 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4806 start_addr, "Check alignment"); 4807 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4808 chunk_size, "Check alignment"); 4809 4810 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4811 // Having claimed the nth_task, compute corresponding mem-region, 4812 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4813 // The alignment restriction ensures that we do not need any 4814 // synchronization with other gang-workers while setting or 4815 // clearing bits in thus chunk of the MUT. 4816 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4817 start_addr + (nth_task+1)*chunk_size); 4818 // The last chunk's end might be way beyond end of the 4819 // used region. In that case pull back appropriately. 4820 if (this_span.end() > end_addr) { 4821 this_span.set_end(end_addr); 4822 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4823 } 4824 // Iterate over the dirty cards covering this chunk, marking them 4825 // precleaned, and setting the corresponding bits in the mod union 4826 // table. Since we have been careful to partition at Card and MUT-word 4827 // boundaries no synchronization is needed between parallel threads. 4828 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4829 &modUnionClosure); 4830 4831 // Having transferred these marks into the modUnionTable, 4832 // rescan the marked objects on the dirty cards in the modUnionTable. 4833 // Even if this is at a synchronous collection, the initial marking 4834 // may have been done during an asynchronous collection so there 4835 // may be dirty bits in the mod-union table. 4836 _collector->_modUnionTable.dirty_range_iterate_clear( 4837 this_span, &greyRescanClosure); 4838 _collector->_modUnionTable.verifyNoOneBitsInRange( 4839 this_span.start(), 4840 this_span.end()); 4841 } 4842 pst->all_tasks_completed(); // declare that i am done 4843 } 4844 4845 // . see if we can share work_queues with ParNew? XXX 4846 void 4847 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4848 int* seed) { 4849 OopTaskQueue* work_q = work_queue(i); 4850 NOT_PRODUCT(int num_steals = 0;) 4851 oop obj_to_scan; 4852 CMSBitMap* bm = &(_collector->_markBitMap); 4853 4854 while (true) { 4855 // Completely finish any left over work from (an) earlier round(s) 4856 cl->trim_queue(0); 4857 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4858 (size_t)ParGCDesiredObjsFromOverflowList); 4859 // Now check if there's any work in the overflow list 4860 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4861 // only affects the number of attempts made to get work from the 4862 // overflow list and does not affect the number of workers. Just 4863 // pass ParallelGCThreads so this behavior is unchanged. 4864 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4865 work_q, 4866 ParallelGCThreads)) { 4867 // found something in global overflow list; 4868 // not yet ready to go stealing work from others. 4869 // We'd like to assert(work_q->size() != 0, ...) 4870 // because we just took work from the overflow list, 4871 // but of course we can't since all of that could have 4872 // been already stolen from us. 4873 // "He giveth and He taketh away." 4874 continue; 4875 } 4876 // Verify that we have no work before we resort to stealing 4877 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4878 // Try to steal from other queues that have work 4879 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4880 NOT_PRODUCT(num_steals++;) 4881 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4882 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4883 // Do scanning work 4884 obj_to_scan->oop_iterate(cl); 4885 // Loop around, finish this work, and try to steal some more 4886 } else if (terminator()->offer_termination()) { 4887 break; // nirvana from the infinite cycle 4888 } 4889 } 4890 NOT_PRODUCT( 4891 if (PrintCMSStatistics != 0) { 4892 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4893 } 4894 ) 4895 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4896 "Else our work is not yet done"); 4897 } 4898 4899 // Record object boundaries in _eden_chunk_array by sampling the eden 4900 // top in the slow-path eden object allocation code path and record 4901 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4902 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4903 // sampling in sample_eden() that activates during the part of the 4904 // preclean phase. 4905 void CMSCollector::sample_eden_chunk() { 4906 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4907 if (_eden_chunk_lock->try_lock()) { 4908 // Record a sample. This is the critical section. The contents 4909 // of the _eden_chunk_array have to be non-decreasing in the 4910 // address order. 4911 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4912 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4913 "Unexpected state of Eden"); 4914 if (_eden_chunk_index == 0 || 4915 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4916 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4917 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4918 _eden_chunk_index++; // commit sample 4919 } 4920 _eden_chunk_lock->unlock(); 4921 } 4922 } 4923 } 4924 4925 // Return a thread-local PLAB recording array, as appropriate. 4926 void* CMSCollector::get_data_recorder(int thr_num) { 4927 if (_survivor_plab_array != NULL && 4928 (CMSPLABRecordAlways || 4929 (_collectorState > Marking && _collectorState < FinalMarking))) { 4930 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4931 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4932 ca->reset(); // clear it so that fresh data is recorded 4933 return (void*) ca; 4934 } else { 4935 return NULL; 4936 } 4937 } 4938 4939 // Reset all the thread-local PLAB recording arrays 4940 void CMSCollector::reset_survivor_plab_arrays() { 4941 for (uint i = 0; i < ParallelGCThreads; i++) { 4942 _survivor_plab_array[i].reset(); 4943 } 4944 } 4945 4946 // Merge the per-thread plab arrays into the global survivor chunk 4947 // array which will provide the partitioning of the survivor space 4948 // for CMS initial scan and rescan. 4949 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4950 int no_of_gc_threads) { 4951 assert(_survivor_plab_array != NULL, "Error"); 4952 assert(_survivor_chunk_array != NULL, "Error"); 4953 assert(_collectorState == FinalMarking || 4954 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4955 for (int j = 0; j < no_of_gc_threads; j++) { 4956 _cursor[j] = 0; 4957 } 4958 HeapWord* top = surv->top(); 4959 size_t i; 4960 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4961 HeapWord* min_val = top; // Higher than any PLAB address 4962 uint min_tid = 0; // position of min_val this round 4963 for (int j = 0; j < no_of_gc_threads; j++) { 4964 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4965 if (_cursor[j] == cur_sca->end()) { 4966 continue; 4967 } 4968 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4969 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4970 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4971 if (cur_val < min_val) { 4972 min_tid = j; 4973 min_val = cur_val; 4974 } else { 4975 assert(cur_val < top, "All recorded addresses should be less"); 4976 } 4977 } 4978 // At this point min_val and min_tid are respectively 4979 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4980 // and the thread (j) that witnesses that address. 4981 // We record this address in the _survivor_chunk_array[i] 4982 // and increment _cursor[min_tid] prior to the next round i. 4983 if (min_val == top) { 4984 break; 4985 } 4986 _survivor_chunk_array[i] = min_val; 4987 _cursor[min_tid]++; 4988 } 4989 // We are all done; record the size of the _survivor_chunk_array 4990 _survivor_chunk_index = i; // exclusive: [0, i) 4991 if (PrintCMSStatistics > 0) { 4992 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4993 } 4994 // Verify that we used up all the recorded entries 4995 #ifdef ASSERT 4996 size_t total = 0; 4997 for (int j = 0; j < no_of_gc_threads; j++) { 4998 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4999 total += _cursor[j]; 5000 } 5001 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5002 // Check that the merged array is in sorted order 5003 if (total > 0) { 5004 for (size_t i = 0; i < total - 1; i++) { 5005 if (PrintCMSStatistics > 0) { 5006 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5007 i, p2i(_survivor_chunk_array[i])); 5008 } 5009 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5010 "Not sorted"); 5011 } 5012 } 5013 #endif // ASSERT 5014 } 5015 5016 // Set up the space's par_seq_tasks structure for work claiming 5017 // for parallel initial scan and rescan of young gen. 5018 // See ParRescanTask where this is currently used. 5019 void 5020 CMSCollector:: 5021 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5022 assert(n_threads > 0, "Unexpected n_threads argument"); 5023 5024 // Eden space 5025 if (!_young_gen->eden()->is_empty()) { 5026 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5027 assert(!pst->valid(), "Clobbering existing data?"); 5028 // Each valid entry in [0, _eden_chunk_index) represents a task. 5029 size_t n_tasks = _eden_chunk_index + 1; 5030 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5031 // Sets the condition for completion of the subtask (how many threads 5032 // need to finish in order to be done). 5033 pst->set_n_threads(n_threads); 5034 pst->set_n_tasks((int)n_tasks); 5035 } 5036 5037 // Merge the survivor plab arrays into _survivor_chunk_array 5038 if (_survivor_plab_array != NULL) { 5039 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5040 } else { 5041 assert(_survivor_chunk_index == 0, "Error"); 5042 } 5043 5044 // To space 5045 { 5046 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5047 assert(!pst->valid(), "Clobbering existing data?"); 5048 // Sets the condition for completion of the subtask (how many threads 5049 // need to finish in order to be done). 5050 pst->set_n_threads(n_threads); 5051 pst->set_n_tasks(1); 5052 assert(pst->valid(), "Error"); 5053 } 5054 5055 // From space 5056 { 5057 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5058 assert(!pst->valid(), "Clobbering existing data?"); 5059 size_t n_tasks = _survivor_chunk_index + 1; 5060 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5061 // Sets the condition for completion of the subtask (how many threads 5062 // need to finish in order to be done). 5063 pst->set_n_threads(n_threads); 5064 pst->set_n_tasks((int)n_tasks); 5065 assert(pst->valid(), "Error"); 5066 } 5067 } 5068 5069 // Parallel version of remark 5070 void CMSCollector::do_remark_parallel() { 5071 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5072 FlexibleWorkGang* workers = gch->workers(); 5073 assert(workers != NULL, "Need parallel worker threads."); 5074 // Choose to use the number of GC workers most recently set 5075 // into "active_workers". 5076 uint n_workers = workers->active_workers(); 5077 5078 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5079 5080 StrongRootsScope srs(n_workers); 5081 5082 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5083 5084 // We won't be iterating over the cards in the card table updating 5085 // the younger_gen cards, so we shouldn't call the following else 5086 // the verification code as well as subsequent younger_refs_iterate 5087 // code would get confused. XXX 5088 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5089 5090 // The young gen rescan work will not be done as part of 5091 // process_roots (which currently doesn't know how to 5092 // parallelize such a scan), but rather will be broken up into 5093 // a set of parallel tasks (via the sampling that the [abortable] 5094 // preclean phase did of eden, plus the [two] tasks of 5095 // scanning the [two] survivor spaces. Further fine-grain 5096 // parallelization of the scanning of the survivor spaces 5097 // themselves, and of precleaning of the younger gen itself 5098 // is deferred to the future. 5099 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5100 5101 // The dirty card rescan work is broken up into a "sequence" 5102 // of parallel tasks (per constituent space) that are dynamically 5103 // claimed by the parallel threads. 5104 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5105 5106 // It turns out that even when we're using 1 thread, doing the work in a 5107 // separate thread causes wide variance in run times. We can't help this 5108 // in the multi-threaded case, but we special-case n=1 here to get 5109 // repeatable measurements of the 1-thread overhead of the parallel code. 5110 if (n_workers > 1) { 5111 // Make refs discovery MT-safe, if it isn't already: it may not 5112 // necessarily be so, since it's possible that we are doing 5113 // ST marking. 5114 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5115 workers->run_task(&tsk); 5116 } else { 5117 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5118 tsk.work(0); 5119 } 5120 5121 // restore, single-threaded for now, any preserved marks 5122 // as a result of work_q overflow 5123 restore_preserved_marks_if_any(); 5124 } 5125 5126 // Non-parallel version of remark 5127 void CMSCollector::do_remark_non_parallel() { 5128 ResourceMark rm; 5129 HandleMark hm; 5130 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5131 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5132 5133 MarkRefsIntoAndScanClosure 5134 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5135 &_markStack, this, 5136 false /* should_yield */, false /* not precleaning */); 5137 MarkFromDirtyCardsClosure 5138 markFromDirtyCardsClosure(this, _span, 5139 NULL, // space is set further below 5140 &_markBitMap, &_markStack, &mrias_cl); 5141 { 5142 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5143 // Iterate over the dirty cards, setting the corresponding bits in the 5144 // mod union table. 5145 { 5146 ModUnionClosure modUnionClosure(&_modUnionTable); 5147 _ct->ct_bs()->dirty_card_iterate( 5148 _cmsGen->used_region(), 5149 &modUnionClosure); 5150 } 5151 // Having transferred these marks into the modUnionTable, we just need 5152 // to rescan the marked objects on the dirty cards in the modUnionTable. 5153 // The initial marking may have been done during an asynchronous 5154 // collection so there may be dirty bits in the mod-union table. 5155 const int alignment = 5156 CardTableModRefBS::card_size * BitsPerWord; 5157 { 5158 // ... First handle dirty cards in CMS gen 5159 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5160 MemRegion ur = _cmsGen->used_region(); 5161 HeapWord* lb = ur.start(); 5162 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5163 MemRegion cms_span(lb, ub); 5164 _modUnionTable.dirty_range_iterate_clear(cms_span, 5165 &markFromDirtyCardsClosure); 5166 verify_work_stacks_empty(); 5167 if (PrintCMSStatistics != 0) { 5168 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5169 markFromDirtyCardsClosure.num_dirty_cards()); 5170 } 5171 } 5172 } 5173 if (VerifyDuringGC && 5174 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5175 HandleMark hm; // Discard invalid handles created during verification 5176 Universe::verify(); 5177 } 5178 { 5179 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5180 5181 verify_work_stacks_empty(); 5182 5183 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5184 StrongRootsScope srs(1); 5185 5186 gch->gen_process_roots(&srs, 5187 _cmsGen->level(), 5188 true, // younger gens as roots 5189 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5190 should_unload_classes(), 5191 &mrias_cl, 5192 NULL, 5193 NULL); // The dirty klasses will be handled below 5194 5195 assert(should_unload_classes() 5196 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5197 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5198 } 5199 5200 { 5201 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5202 5203 verify_work_stacks_empty(); 5204 5205 // Scan all class loader data objects that might have been introduced 5206 // during concurrent marking. 5207 ResourceMark rm; 5208 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5209 for (int i = 0; i < array->length(); i++) { 5210 mrias_cl.do_class_loader_data(array->at(i)); 5211 } 5212 5213 // We don't need to keep track of new CLDs anymore. 5214 ClassLoaderDataGraph::remember_new_clds(false); 5215 5216 verify_work_stacks_empty(); 5217 } 5218 5219 { 5220 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5221 5222 verify_work_stacks_empty(); 5223 5224 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5225 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5226 5227 verify_work_stacks_empty(); 5228 } 5229 5230 // We might have added oops to ClassLoaderData::_handles during the 5231 // concurrent marking phase. These oops point to newly allocated objects 5232 // that are guaranteed to be kept alive. Either by the direct allocation 5233 // code, or when the young collector processes the roots. Hence, 5234 // we don't have to revisit the _handles block during the remark phase. 5235 5236 verify_work_stacks_empty(); 5237 // Restore evacuated mark words, if any, used for overflow list links 5238 if (!CMSOverflowEarlyRestoration) { 5239 restore_preserved_marks_if_any(); 5240 } 5241 verify_overflow_empty(); 5242 } 5243 5244 //////////////////////////////////////////////////////// 5245 // Parallel Reference Processing Task Proxy Class 5246 //////////////////////////////////////////////////////// 5247 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5248 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5249 CMSCollector* _collector; 5250 CMSBitMap* _mark_bit_map; 5251 const MemRegion _span; 5252 ProcessTask& _task; 5253 5254 public: 5255 CMSRefProcTaskProxy(ProcessTask& task, 5256 CMSCollector* collector, 5257 const MemRegion& span, 5258 CMSBitMap* mark_bit_map, 5259 AbstractWorkGang* workers, 5260 OopTaskQueueSet* task_queues): 5261 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5262 task_queues, 5263 workers->active_workers()), 5264 _task(task), 5265 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5266 { 5267 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5268 "Inconsistency in _span"); 5269 } 5270 5271 OopTaskQueueSet* task_queues() { return queues(); } 5272 5273 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5274 5275 void do_work_steal(int i, 5276 CMSParDrainMarkingStackClosure* drain, 5277 CMSParKeepAliveClosure* keep_alive, 5278 int* seed); 5279 5280 virtual void work(uint worker_id); 5281 }; 5282 5283 void CMSRefProcTaskProxy::work(uint worker_id) { 5284 ResourceMark rm; 5285 HandleMark hm; 5286 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5287 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5288 _mark_bit_map, 5289 work_queue(worker_id)); 5290 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5291 _mark_bit_map, 5292 work_queue(worker_id)); 5293 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5294 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5295 if (_task.marks_oops_alive()) { 5296 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5297 _collector->hash_seed(worker_id)); 5298 } 5299 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5300 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5301 } 5302 5303 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5304 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5305 EnqueueTask& _task; 5306 5307 public: 5308 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5309 : AbstractGangTask("Enqueue reference objects in parallel"), 5310 _task(task) 5311 { } 5312 5313 virtual void work(uint worker_id) 5314 { 5315 _task.work(worker_id); 5316 } 5317 }; 5318 5319 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5320 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5321 _span(span), 5322 _bit_map(bit_map), 5323 _work_queue(work_queue), 5324 _mark_and_push(collector, span, bit_map, work_queue), 5325 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5326 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5327 { } 5328 5329 // . see if we can share work_queues with ParNew? XXX 5330 void CMSRefProcTaskProxy::do_work_steal(int i, 5331 CMSParDrainMarkingStackClosure* drain, 5332 CMSParKeepAliveClosure* keep_alive, 5333 int* seed) { 5334 OopTaskQueue* work_q = work_queue(i); 5335 NOT_PRODUCT(int num_steals = 0;) 5336 oop obj_to_scan; 5337 5338 while (true) { 5339 // Completely finish any left over work from (an) earlier round(s) 5340 drain->trim_queue(0); 5341 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5342 (size_t)ParGCDesiredObjsFromOverflowList); 5343 // Now check if there's any work in the overflow list 5344 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5345 // only affects the number of attempts made to get work from the 5346 // overflow list and does not affect the number of workers. Just 5347 // pass ParallelGCThreads so this behavior is unchanged. 5348 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5349 work_q, 5350 ParallelGCThreads)) { 5351 // Found something in global overflow list; 5352 // not yet ready to go stealing work from others. 5353 // We'd like to assert(work_q->size() != 0, ...) 5354 // because we just took work from the overflow list, 5355 // but of course we can't, since all of that might have 5356 // been already stolen from us. 5357 continue; 5358 } 5359 // Verify that we have no work before we resort to stealing 5360 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5361 // Try to steal from other queues that have work 5362 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5363 NOT_PRODUCT(num_steals++;) 5364 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5365 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5366 // Do scanning work 5367 obj_to_scan->oop_iterate(keep_alive); 5368 // Loop around, finish this work, and try to steal some more 5369 } else if (terminator()->offer_termination()) { 5370 break; // nirvana from the infinite cycle 5371 } 5372 } 5373 NOT_PRODUCT( 5374 if (PrintCMSStatistics != 0) { 5375 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5376 } 5377 ) 5378 } 5379 5380 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5381 { 5382 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5383 FlexibleWorkGang* workers = gch->workers(); 5384 assert(workers != NULL, "Need parallel worker threads."); 5385 CMSRefProcTaskProxy rp_task(task, &_collector, 5386 _collector.ref_processor()->span(), 5387 _collector.markBitMap(), 5388 workers, _collector.task_queues()); 5389 workers->run_task(&rp_task); 5390 } 5391 5392 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5393 { 5394 5395 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5396 FlexibleWorkGang* workers = gch->workers(); 5397 assert(workers != NULL, "Need parallel worker threads."); 5398 CMSRefEnqueueTaskProxy enq_task(task); 5399 workers->run_task(&enq_task); 5400 } 5401 5402 void CMSCollector::refProcessingWork() { 5403 ResourceMark rm; 5404 HandleMark hm; 5405 5406 ReferenceProcessor* rp = ref_processor(); 5407 assert(rp->span().equals(_span), "Spans should be equal"); 5408 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5409 // Process weak references. 5410 rp->setup_policy(false); 5411 verify_work_stacks_empty(); 5412 5413 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5414 &_markStack, false /* !preclean */); 5415 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5416 _span, &_markBitMap, &_markStack, 5417 &cmsKeepAliveClosure, false /* !preclean */); 5418 { 5419 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5420 5421 ReferenceProcessorStats stats; 5422 if (rp->processing_is_mt()) { 5423 // Set the degree of MT here. If the discovery is done MT, there 5424 // may have been a different number of threads doing the discovery 5425 // and a different number of discovered lists may have Ref objects. 5426 // That is OK as long as the Reference lists are balanced (see 5427 // balance_all_queues() and balance_queues()). 5428 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5429 uint active_workers = ParallelGCThreads; 5430 FlexibleWorkGang* workers = gch->workers(); 5431 if (workers != NULL) { 5432 active_workers = workers->active_workers(); 5433 // The expectation is that active_workers will have already 5434 // been set to a reasonable value. If it has not been set, 5435 // investigate. 5436 assert(active_workers > 0, "Should have been set during scavenge"); 5437 } 5438 rp->set_active_mt_degree(active_workers); 5439 CMSRefProcTaskExecutor task_executor(*this); 5440 stats = rp->process_discovered_references(&_is_alive_closure, 5441 &cmsKeepAliveClosure, 5442 &cmsDrainMarkingStackClosure, 5443 &task_executor, 5444 _gc_timer_cm, 5445 _gc_tracer_cm->gc_id()); 5446 } else { 5447 stats = rp->process_discovered_references(&_is_alive_closure, 5448 &cmsKeepAliveClosure, 5449 &cmsDrainMarkingStackClosure, 5450 NULL, 5451 _gc_timer_cm, 5452 _gc_tracer_cm->gc_id()); 5453 } 5454 _gc_tracer_cm->report_gc_reference_stats(stats); 5455 5456 } 5457 5458 // This is the point where the entire marking should have completed. 5459 verify_work_stacks_empty(); 5460 5461 if (should_unload_classes()) { 5462 { 5463 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5464 5465 // Unload classes and purge the SystemDictionary. 5466 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5467 5468 // Unload nmethods. 5469 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5470 5471 // Prune dead klasses from subklass/sibling/implementor lists. 5472 Klass::clean_weak_klass_links(&_is_alive_closure); 5473 } 5474 5475 { 5476 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5477 // Clean up unreferenced symbols in symbol table. 5478 SymbolTable::unlink(); 5479 } 5480 5481 { 5482 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5483 // Delete entries for dead interned strings. 5484 StringTable::unlink(&_is_alive_closure); 5485 } 5486 } 5487 5488 5489 // Restore any preserved marks as a result of mark stack or 5490 // work queue overflow 5491 restore_preserved_marks_if_any(); // done single-threaded for now 5492 5493 rp->set_enqueuing_is_done(true); 5494 if (rp->processing_is_mt()) { 5495 rp->balance_all_queues(); 5496 CMSRefProcTaskExecutor task_executor(*this); 5497 rp->enqueue_discovered_references(&task_executor); 5498 } else { 5499 rp->enqueue_discovered_references(NULL); 5500 } 5501 rp->verify_no_references_recorded(); 5502 assert(!rp->discovery_enabled(), "should have been disabled"); 5503 } 5504 5505 #ifndef PRODUCT 5506 void CMSCollector::check_correct_thread_executing() { 5507 Thread* t = Thread::current(); 5508 // Only the VM thread or the CMS thread should be here. 5509 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5510 "Unexpected thread type"); 5511 // If this is the vm thread, the foreground process 5512 // should not be waiting. Note that _foregroundGCIsActive is 5513 // true while the foreground collector is waiting. 5514 if (_foregroundGCShouldWait) { 5515 // We cannot be the VM thread 5516 assert(t->is_ConcurrentGC_thread(), 5517 "Should be CMS thread"); 5518 } else { 5519 // We can be the CMS thread only if we are in a stop-world 5520 // phase of CMS collection. 5521 if (t->is_ConcurrentGC_thread()) { 5522 assert(_collectorState == InitialMarking || 5523 _collectorState == FinalMarking, 5524 "Should be a stop-world phase"); 5525 // The CMS thread should be holding the CMS_token. 5526 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5527 "Potential interference with concurrently " 5528 "executing VM thread"); 5529 } 5530 } 5531 } 5532 #endif 5533 5534 void CMSCollector::sweep() { 5535 assert(_collectorState == Sweeping, "just checking"); 5536 check_correct_thread_executing(); 5537 verify_work_stacks_empty(); 5538 verify_overflow_empty(); 5539 increment_sweep_count(); 5540 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5541 5542 _inter_sweep_timer.stop(); 5543 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5544 5545 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5546 _intra_sweep_timer.reset(); 5547 _intra_sweep_timer.start(); 5548 { 5549 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5550 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5551 // First sweep the old gen 5552 { 5553 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5554 bitMapLock()); 5555 sweepWork(_cmsGen); 5556 } 5557 5558 // Update Universe::_heap_*_at_gc figures. 5559 // We need all the free list locks to make the abstract state 5560 // transition from Sweeping to Resetting. See detailed note 5561 // further below. 5562 { 5563 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5564 // Update heap occupancy information which is used as 5565 // input to soft ref clearing policy at the next gc. 5566 Universe::update_heap_info_at_gc(); 5567 _collectorState = Resizing; 5568 } 5569 } 5570 verify_work_stacks_empty(); 5571 verify_overflow_empty(); 5572 5573 if (should_unload_classes()) { 5574 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5575 // requires that the virtual spaces are stable and not deleted. 5576 ClassLoaderDataGraph::set_should_purge(true); 5577 } 5578 5579 _intra_sweep_timer.stop(); 5580 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5581 5582 _inter_sweep_timer.reset(); 5583 _inter_sweep_timer.start(); 5584 5585 // We need to use a monotonically non-decreasing time in ms 5586 // or we will see time-warp warnings and os::javaTimeMillis() 5587 // does not guarantee monotonicity. 5588 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5589 update_time_of_last_gc(now); 5590 5591 // NOTE on abstract state transitions: 5592 // Mutators allocate-live and/or mark the mod-union table dirty 5593 // based on the state of the collection. The former is done in 5594 // the interval [Marking, Sweeping] and the latter in the interval 5595 // [Marking, Sweeping). Thus the transitions into the Marking state 5596 // and out of the Sweeping state must be synchronously visible 5597 // globally to the mutators. 5598 // The transition into the Marking state happens with the world 5599 // stopped so the mutators will globally see it. Sweeping is 5600 // done asynchronously by the background collector so the transition 5601 // from the Sweeping state to the Resizing state must be done 5602 // under the freelistLock (as is the check for whether to 5603 // allocate-live and whether to dirty the mod-union table). 5604 assert(_collectorState == Resizing, "Change of collector state to" 5605 " Resizing must be done under the freelistLocks (plural)"); 5606 5607 // Now that sweeping has been completed, we clear 5608 // the incremental_collection_failed flag, 5609 // thus inviting a younger gen collection to promote into 5610 // this generation. If such a promotion may still fail, 5611 // the flag will be set again when a young collection is 5612 // attempted. 5613 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5614 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5615 gch->update_full_collections_completed(_collection_count_start); 5616 } 5617 5618 // FIX ME!!! Looks like this belongs in CFLSpace, with 5619 // CMSGen merely delegating to it. 5620 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5621 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5622 HeapWord* minAddr = _cmsSpace->bottom(); 5623 HeapWord* largestAddr = 5624 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5625 if (largestAddr == NULL) { 5626 // The dictionary appears to be empty. In this case 5627 // try to coalesce at the end of the heap. 5628 largestAddr = _cmsSpace->end(); 5629 } 5630 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5631 size_t nearLargestOffset = 5632 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5633 if (PrintFLSStatistics != 0) { 5634 gclog_or_tty->print_cr( 5635 "CMS: Large Block: " PTR_FORMAT ";" 5636 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5637 p2i(largestAddr), 5638 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5639 } 5640 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5641 } 5642 5643 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5644 return addr >= _cmsSpace->nearLargestChunk(); 5645 } 5646 5647 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5648 return _cmsSpace->find_chunk_at_end(); 5649 } 5650 5651 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5652 bool full) { 5653 // The next lower level has been collected. Gather any statistics 5654 // that are of interest at this point. 5655 if (!full && (current_level + 1) == level()) { 5656 // Gather statistics on the young generation collection. 5657 collector()->stats().record_gc0_end(used()); 5658 } 5659 } 5660 5661 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5662 // We iterate over the space(s) underlying this generation, 5663 // checking the mark bit map to see if the bits corresponding 5664 // to specific blocks are marked or not. Blocks that are 5665 // marked are live and are not swept up. All remaining blocks 5666 // are swept up, with coalescing on-the-fly as we sweep up 5667 // contiguous free and/or garbage blocks: 5668 // We need to ensure that the sweeper synchronizes with allocators 5669 // and stop-the-world collectors. In particular, the following 5670 // locks are used: 5671 // . CMS token: if this is held, a stop the world collection cannot occur 5672 // . freelistLock: if this is held no allocation can occur from this 5673 // generation by another thread 5674 // . bitMapLock: if this is held, no other thread can access or update 5675 // 5676 5677 // Note that we need to hold the freelistLock if we use 5678 // block iterate below; else the iterator might go awry if 5679 // a mutator (or promotion) causes block contents to change 5680 // (for instance if the allocator divvies up a block). 5681 // If we hold the free list lock, for all practical purposes 5682 // young generation GC's can't occur (they'll usually need to 5683 // promote), so we might as well prevent all young generation 5684 // GC's while we do a sweeping step. For the same reason, we might 5685 // as well take the bit map lock for the entire duration 5686 5687 // check that we hold the requisite locks 5688 assert(have_cms_token(), "Should hold cms token"); 5689 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5690 assert_lock_strong(gen->freelistLock()); 5691 assert_lock_strong(bitMapLock()); 5692 5693 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5694 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5695 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5696 _inter_sweep_estimate.padded_average(), 5697 _intra_sweep_estimate.padded_average()); 5698 gen->setNearLargestChunk(); 5699 5700 { 5701 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5702 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5703 // We need to free-up/coalesce garbage/blocks from a 5704 // co-terminal free run. This is done in the SweepClosure 5705 // destructor; so, do not remove this scope, else the 5706 // end-of-sweep-census below will be off by a little bit. 5707 } 5708 gen->cmsSpace()->sweep_completed(); 5709 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5710 if (should_unload_classes()) { // unloaded classes this cycle, 5711 _concurrent_cycles_since_last_unload = 0; // ... reset count 5712 } else { // did not unload classes, 5713 _concurrent_cycles_since_last_unload++; // ... increment count 5714 } 5715 } 5716 5717 // Reset CMS data structures (for now just the marking bit map) 5718 // preparatory for the next cycle. 5719 void CMSCollector::reset(bool concurrent) { 5720 if (concurrent) { 5721 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5722 5723 // If the state is not "Resetting", the foreground thread 5724 // has done a collection and the resetting. 5725 if (_collectorState != Resetting) { 5726 assert(_collectorState == Idling, "The state should only change" 5727 " because the foreground collector has finished the collection"); 5728 return; 5729 } 5730 5731 // Clear the mark bitmap (no grey objects to start with) 5732 // for the next cycle. 5733 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5734 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5735 5736 HeapWord* curAddr = _markBitMap.startWord(); 5737 while (curAddr < _markBitMap.endWord()) { 5738 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5739 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5740 _markBitMap.clear_large_range(chunk); 5741 if (ConcurrentMarkSweepThread::should_yield() && 5742 !foregroundGCIsActive() && 5743 CMSYield) { 5744 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5745 "CMS thread should hold CMS token"); 5746 assert_lock_strong(bitMapLock()); 5747 bitMapLock()->unlock(); 5748 ConcurrentMarkSweepThread::desynchronize(true); 5749 stopTimer(); 5750 if (PrintCMSStatistics != 0) { 5751 incrementYields(); 5752 } 5753 5754 // See the comment in coordinator_yield() 5755 for (unsigned i = 0; i < CMSYieldSleepCount && 5756 ConcurrentMarkSweepThread::should_yield() && 5757 !CMSCollector::foregroundGCIsActive(); ++i) { 5758 os::sleep(Thread::current(), 1, false); 5759 } 5760 5761 ConcurrentMarkSweepThread::synchronize(true); 5762 bitMapLock()->lock_without_safepoint_check(); 5763 startTimer(); 5764 } 5765 curAddr = chunk.end(); 5766 } 5767 // A successful mostly concurrent collection has been done. 5768 // Because only the full (i.e., concurrent mode failure) collections 5769 // are being measured for gc overhead limits, clean the "near" flag 5770 // and count. 5771 size_policy()->reset_gc_overhead_limit_count(); 5772 _collectorState = Idling; 5773 } else { 5774 // already have the lock 5775 assert(_collectorState == Resetting, "just checking"); 5776 assert_lock_strong(bitMapLock()); 5777 _markBitMap.clear_all(); 5778 _collectorState = Idling; 5779 } 5780 5781 register_gc_end(); 5782 } 5783 5784 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5785 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5786 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5787 TraceCollectorStats tcs(counters()); 5788 5789 switch (op) { 5790 case CMS_op_checkpointRootsInitial: { 5791 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5792 checkpointRootsInitial(); 5793 if (PrintGC) { 5794 _cmsGen->printOccupancy("initial-mark"); 5795 } 5796 break; 5797 } 5798 case CMS_op_checkpointRootsFinal: { 5799 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5800 checkpointRootsFinal(); 5801 if (PrintGC) { 5802 _cmsGen->printOccupancy("remark"); 5803 } 5804 break; 5805 } 5806 default: 5807 fatal("No such CMS_op"); 5808 } 5809 } 5810 5811 #ifndef PRODUCT 5812 size_t const CMSCollector::skip_header_HeapWords() { 5813 return FreeChunk::header_size(); 5814 } 5815 5816 // Try and collect here conditions that should hold when 5817 // CMS thread is exiting. The idea is that the foreground GC 5818 // thread should not be blocked if it wants to terminate 5819 // the CMS thread and yet continue to run the VM for a while 5820 // after that. 5821 void CMSCollector::verify_ok_to_terminate() const { 5822 assert(Thread::current()->is_ConcurrentGC_thread(), 5823 "should be called by CMS thread"); 5824 assert(!_foregroundGCShouldWait, "should be false"); 5825 // We could check here that all the various low-level locks 5826 // are not held by the CMS thread, but that is overkill; see 5827 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5828 // is checked. 5829 } 5830 #endif 5831 5832 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5833 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5834 "missing Printezis mark?"); 5835 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5836 size_t size = pointer_delta(nextOneAddr + 1, addr); 5837 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5838 "alignment problem"); 5839 assert(size >= 3, "Necessary for Printezis marks to work"); 5840 return size; 5841 } 5842 5843 // A variant of the above (block_size_using_printezis_bits()) except 5844 // that we return 0 if the P-bits are not yet set. 5845 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5846 if (_markBitMap.isMarked(addr + 1)) { 5847 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5848 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5849 size_t size = pointer_delta(nextOneAddr + 1, addr); 5850 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5851 "alignment problem"); 5852 assert(size >= 3, "Necessary for Printezis marks to work"); 5853 return size; 5854 } 5855 return 0; 5856 } 5857 5858 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5859 size_t sz = 0; 5860 oop p = (oop)addr; 5861 if (p->klass_or_null() != NULL) { 5862 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5863 } else { 5864 sz = block_size_using_printezis_bits(addr); 5865 } 5866 assert(sz > 0, "size must be nonzero"); 5867 HeapWord* next_block = addr + sz; 5868 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5869 CardTableModRefBS::card_size); 5870 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5871 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5872 "must be different cards"); 5873 return next_card; 5874 } 5875 5876 5877 // CMS Bit Map Wrapper ///////////////////////////////////////// 5878 5879 // Construct a CMS bit map infrastructure, but don't create the 5880 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5881 // further below. 5882 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5883 _bm(), 5884 _shifter(shifter), 5885 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5886 Monitor::_safepoint_check_sometimes) : NULL) 5887 { 5888 _bmStartWord = 0; 5889 _bmWordSize = 0; 5890 } 5891 5892 bool CMSBitMap::allocate(MemRegion mr) { 5893 _bmStartWord = mr.start(); 5894 _bmWordSize = mr.word_size(); 5895 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5896 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5897 if (!brs.is_reserved()) { 5898 warning("CMS bit map allocation failure"); 5899 return false; 5900 } 5901 // For now we'll just commit all of the bit map up front. 5902 // Later on we'll try to be more parsimonious with swap. 5903 if (!_virtual_space.initialize(brs, brs.size())) { 5904 warning("CMS bit map backing store failure"); 5905 return false; 5906 } 5907 assert(_virtual_space.committed_size() == brs.size(), 5908 "didn't reserve backing store for all of CMS bit map?"); 5909 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5910 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5911 _bmWordSize, "inconsistency in bit map sizing"); 5912 _bm.set_size(_bmWordSize >> _shifter); 5913 5914 // bm.clear(); // can we rely on getting zero'd memory? verify below 5915 assert(isAllClear(), 5916 "Expected zero'd memory from ReservedSpace constructor"); 5917 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5918 "consistency check"); 5919 return true; 5920 } 5921 5922 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5923 HeapWord *next_addr, *end_addr, *last_addr; 5924 assert_locked(); 5925 assert(covers(mr), "out-of-range error"); 5926 // XXX assert that start and end are appropriately aligned 5927 for (next_addr = mr.start(), end_addr = mr.end(); 5928 next_addr < end_addr; next_addr = last_addr) { 5929 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5930 last_addr = dirty_region.end(); 5931 if (!dirty_region.is_empty()) { 5932 cl->do_MemRegion(dirty_region); 5933 } else { 5934 assert(last_addr == end_addr, "program logic"); 5935 return; 5936 } 5937 } 5938 } 5939 5940 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5941 _bm.print_on_error(st, prefix); 5942 } 5943 5944 #ifndef PRODUCT 5945 void CMSBitMap::assert_locked() const { 5946 CMSLockVerifier::assert_locked(lock()); 5947 } 5948 5949 bool CMSBitMap::covers(MemRegion mr) const { 5950 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5951 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5952 "size inconsistency"); 5953 return (mr.start() >= _bmStartWord) && 5954 (mr.end() <= endWord()); 5955 } 5956 5957 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5958 return (start >= _bmStartWord && (start + size) <= endWord()); 5959 } 5960 5961 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5962 // verify that there are no 1 bits in the interval [left, right) 5963 FalseBitMapClosure falseBitMapClosure; 5964 iterate(&falseBitMapClosure, left, right); 5965 } 5966 5967 void CMSBitMap::region_invariant(MemRegion mr) 5968 { 5969 assert_locked(); 5970 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5971 assert(!mr.is_empty(), "unexpected empty region"); 5972 assert(covers(mr), "mr should be covered by bit map"); 5973 // convert address range into offset range 5974 size_t start_ofs = heapWordToOffset(mr.start()); 5975 // Make sure that end() is appropriately aligned 5976 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5977 (1 << (_shifter+LogHeapWordSize))), 5978 "Misaligned mr.end()"); 5979 size_t end_ofs = heapWordToOffset(mr.end()); 5980 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5981 } 5982 5983 #endif 5984 5985 bool CMSMarkStack::allocate(size_t size) { 5986 // allocate a stack of the requisite depth 5987 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5988 size * sizeof(oop))); 5989 if (!rs.is_reserved()) { 5990 warning("CMSMarkStack allocation failure"); 5991 return false; 5992 } 5993 if (!_virtual_space.initialize(rs, rs.size())) { 5994 warning("CMSMarkStack backing store failure"); 5995 return false; 5996 } 5997 assert(_virtual_space.committed_size() == rs.size(), 5998 "didn't reserve backing store for all of CMS stack?"); 5999 _base = (oop*)(_virtual_space.low()); 6000 _index = 0; 6001 _capacity = size; 6002 NOT_PRODUCT(_max_depth = 0); 6003 return true; 6004 } 6005 6006 // XXX FIX ME !!! In the MT case we come in here holding a 6007 // leaf lock. For printing we need to take a further lock 6008 // which has lower rank. We need to recalibrate the two 6009 // lock-ranks involved in order to be able to print the 6010 // messages below. (Or defer the printing to the caller. 6011 // For now we take the expedient path of just disabling the 6012 // messages for the problematic case.) 6013 void CMSMarkStack::expand() { 6014 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6015 if (_capacity == MarkStackSizeMax) { 6016 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6017 // We print a warning message only once per CMS cycle. 6018 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6019 } 6020 return; 6021 } 6022 // Double capacity if possible 6023 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6024 // Do not give up existing stack until we have managed to 6025 // get the double capacity that we desired. 6026 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6027 new_capacity * sizeof(oop))); 6028 if (rs.is_reserved()) { 6029 // Release the backing store associated with old stack 6030 _virtual_space.release(); 6031 // Reinitialize virtual space for new stack 6032 if (!_virtual_space.initialize(rs, rs.size())) { 6033 fatal("Not enough swap for expanded marking stack"); 6034 } 6035 _base = (oop*)(_virtual_space.low()); 6036 _index = 0; 6037 _capacity = new_capacity; 6038 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6039 // Failed to double capacity, continue; 6040 // we print a detail message only once per CMS cycle. 6041 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6042 SIZE_FORMAT"K", 6043 _capacity / K, new_capacity / K); 6044 } 6045 } 6046 6047 6048 // Closures 6049 // XXX: there seems to be a lot of code duplication here; 6050 // should refactor and consolidate common code. 6051 6052 // This closure is used to mark refs into the CMS generation in 6053 // the CMS bit map. Called at the first checkpoint. This closure 6054 // assumes that we do not need to re-mark dirty cards; if the CMS 6055 // generation on which this is used is not an oldest 6056 // generation then this will lose younger_gen cards! 6057 6058 MarkRefsIntoClosure::MarkRefsIntoClosure( 6059 MemRegion span, CMSBitMap* bitMap): 6060 _span(span), 6061 _bitMap(bitMap) 6062 { 6063 assert(_ref_processor == NULL, "deliberately left NULL"); 6064 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6065 } 6066 6067 void MarkRefsIntoClosure::do_oop(oop obj) { 6068 // if p points into _span, then mark corresponding bit in _markBitMap 6069 assert(obj->is_oop(), "expected an oop"); 6070 HeapWord* addr = (HeapWord*)obj; 6071 if (_span.contains(addr)) { 6072 // this should be made more efficient 6073 _bitMap->mark(addr); 6074 } 6075 } 6076 6077 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6078 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6079 6080 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6081 MemRegion span, CMSBitMap* bitMap): 6082 _span(span), 6083 _bitMap(bitMap) 6084 { 6085 assert(_ref_processor == NULL, "deliberately left NULL"); 6086 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6087 } 6088 6089 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6090 // if p points into _span, then mark corresponding bit in _markBitMap 6091 assert(obj->is_oop(), "expected an oop"); 6092 HeapWord* addr = (HeapWord*)obj; 6093 if (_span.contains(addr)) { 6094 // this should be made more efficient 6095 _bitMap->par_mark(addr); 6096 } 6097 } 6098 6099 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6100 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6101 6102 // A variant of the above, used for CMS marking verification. 6103 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6104 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6105 _span(span), 6106 _verification_bm(verification_bm), 6107 _cms_bm(cms_bm) 6108 { 6109 assert(_ref_processor == NULL, "deliberately left NULL"); 6110 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6111 } 6112 6113 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6114 // if p points into _span, then mark corresponding bit in _markBitMap 6115 assert(obj->is_oop(), "expected an oop"); 6116 HeapWord* addr = (HeapWord*)obj; 6117 if (_span.contains(addr)) { 6118 _verification_bm->mark(addr); 6119 if (!_cms_bm->isMarked(addr)) { 6120 oop(addr)->print(); 6121 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6122 fatal("... aborting"); 6123 } 6124 } 6125 } 6126 6127 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6128 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6129 6130 ////////////////////////////////////////////////// 6131 // MarkRefsIntoAndScanClosure 6132 ////////////////////////////////////////////////// 6133 6134 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6135 ReferenceProcessor* rp, 6136 CMSBitMap* bit_map, 6137 CMSBitMap* mod_union_table, 6138 CMSMarkStack* mark_stack, 6139 CMSCollector* collector, 6140 bool should_yield, 6141 bool concurrent_precleaning): 6142 _collector(collector), 6143 _span(span), 6144 _bit_map(bit_map), 6145 _mark_stack(mark_stack), 6146 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6147 mark_stack, concurrent_precleaning), 6148 _yield(should_yield), 6149 _concurrent_precleaning(concurrent_precleaning), 6150 _freelistLock(NULL) 6151 { 6152 _ref_processor = rp; 6153 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6154 } 6155 6156 // This closure is used to mark refs into the CMS generation at the 6157 // second (final) checkpoint, and to scan and transitively follow 6158 // the unmarked oops. It is also used during the concurrent precleaning 6159 // phase while scanning objects on dirty cards in the CMS generation. 6160 // The marks are made in the marking bit map and the marking stack is 6161 // used for keeping the (newly) grey objects during the scan. 6162 // The parallel version (Par_...) appears further below. 6163 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6164 if (obj != NULL) { 6165 assert(obj->is_oop(), "expected an oop"); 6166 HeapWord* addr = (HeapWord*)obj; 6167 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6168 assert(_collector->overflow_list_is_empty(), 6169 "overflow list should be empty"); 6170 if (_span.contains(addr) && 6171 !_bit_map->isMarked(addr)) { 6172 // mark bit map (object is now grey) 6173 _bit_map->mark(addr); 6174 // push on marking stack (stack should be empty), and drain the 6175 // stack by applying this closure to the oops in the oops popped 6176 // from the stack (i.e. blacken the grey objects) 6177 bool res = _mark_stack->push(obj); 6178 assert(res, "Should have space to push on empty stack"); 6179 do { 6180 oop new_oop = _mark_stack->pop(); 6181 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6182 assert(_bit_map->isMarked((HeapWord*)new_oop), 6183 "only grey objects on this stack"); 6184 // iterate over the oops in this oop, marking and pushing 6185 // the ones in CMS heap (i.e. in _span). 6186 new_oop->oop_iterate(&_pushAndMarkClosure); 6187 // check if it's time to yield 6188 do_yield_check(); 6189 } while (!_mark_stack->isEmpty() || 6190 (!_concurrent_precleaning && take_from_overflow_list())); 6191 // if marking stack is empty, and we are not doing this 6192 // during precleaning, then check the overflow list 6193 } 6194 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6195 assert(_collector->overflow_list_is_empty(), 6196 "overflow list was drained above"); 6197 // We could restore evacuated mark words, if any, used for 6198 // overflow list links here because the overflow list is 6199 // provably empty here. That would reduce the maximum 6200 // size requirements for preserved_{oop,mark}_stack. 6201 // But we'll just postpone it until we are all done 6202 // so we can just stream through. 6203 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6204 _collector->restore_preserved_marks_if_any(); 6205 assert(_collector->no_preserved_marks(), "No preserved marks"); 6206 } 6207 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6208 "All preserved marks should have been restored above"); 6209 } 6210 } 6211 6212 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6213 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6214 6215 void MarkRefsIntoAndScanClosure::do_yield_work() { 6216 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6217 "CMS thread should hold CMS token"); 6218 assert_lock_strong(_freelistLock); 6219 assert_lock_strong(_bit_map->lock()); 6220 // relinquish the free_list_lock and bitMaplock() 6221 _bit_map->lock()->unlock(); 6222 _freelistLock->unlock(); 6223 ConcurrentMarkSweepThread::desynchronize(true); 6224 _collector->stopTimer(); 6225 if (PrintCMSStatistics != 0) { 6226 _collector->incrementYields(); 6227 } 6228 6229 // See the comment in coordinator_yield() 6230 for (unsigned i = 0; 6231 i < CMSYieldSleepCount && 6232 ConcurrentMarkSweepThread::should_yield() && 6233 !CMSCollector::foregroundGCIsActive(); 6234 ++i) { 6235 os::sleep(Thread::current(), 1, false); 6236 } 6237 6238 ConcurrentMarkSweepThread::synchronize(true); 6239 _freelistLock->lock_without_safepoint_check(); 6240 _bit_map->lock()->lock_without_safepoint_check(); 6241 _collector->startTimer(); 6242 } 6243 6244 /////////////////////////////////////////////////////////// 6245 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6246 // MarkRefsIntoAndScanClosure 6247 /////////////////////////////////////////////////////////// 6248 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6249 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6250 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6251 _span(span), 6252 _bit_map(bit_map), 6253 _work_queue(work_queue), 6254 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6255 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6256 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6257 { 6258 _ref_processor = rp; 6259 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6260 } 6261 6262 // This closure is used to mark refs into the CMS generation at the 6263 // second (final) checkpoint, and to scan and transitively follow 6264 // the unmarked oops. The marks are made in the marking bit map and 6265 // the work_queue is used for keeping the (newly) grey objects during 6266 // the scan phase whence they are also available for stealing by parallel 6267 // threads. Since the marking bit map is shared, updates are 6268 // synchronized (via CAS). 6269 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6270 if (obj != NULL) { 6271 // Ignore mark word because this could be an already marked oop 6272 // that may be chained at the end of the overflow list. 6273 assert(obj->is_oop(true), "expected an oop"); 6274 HeapWord* addr = (HeapWord*)obj; 6275 if (_span.contains(addr) && 6276 !_bit_map->isMarked(addr)) { 6277 // mark bit map (object will become grey): 6278 // It is possible for several threads to be 6279 // trying to "claim" this object concurrently; 6280 // the unique thread that succeeds in marking the 6281 // object first will do the subsequent push on 6282 // to the work queue (or overflow list). 6283 if (_bit_map->par_mark(addr)) { 6284 // push on work_queue (which may not be empty), and trim the 6285 // queue to an appropriate length by applying this closure to 6286 // the oops in the oops popped from the stack (i.e. blacken the 6287 // grey objects) 6288 bool res = _work_queue->push(obj); 6289 assert(res, "Low water mark should be less than capacity?"); 6290 trim_queue(_low_water_mark); 6291 } // Else, another thread claimed the object 6292 } 6293 } 6294 } 6295 6296 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6297 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6298 6299 // This closure is used to rescan the marked objects on the dirty cards 6300 // in the mod union table and the card table proper. 6301 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6302 oop p, MemRegion mr) { 6303 6304 size_t size = 0; 6305 HeapWord* addr = (HeapWord*)p; 6306 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6307 assert(_span.contains(addr), "we are scanning the CMS generation"); 6308 // check if it's time to yield 6309 if (do_yield_check()) { 6310 // We yielded for some foreground stop-world work, 6311 // and we have been asked to abort this ongoing preclean cycle. 6312 return 0; 6313 } 6314 if (_bitMap->isMarked(addr)) { 6315 // it's marked; is it potentially uninitialized? 6316 if (p->klass_or_null() != NULL) { 6317 // an initialized object; ignore mark word in verification below 6318 // since we are running concurrent with mutators 6319 assert(p->is_oop(true), "should be an oop"); 6320 if (p->is_objArray()) { 6321 // objArrays are precisely marked; restrict scanning 6322 // to dirty cards only. 6323 size = CompactibleFreeListSpace::adjustObjectSize( 6324 p->oop_iterate(_scanningClosure, mr)); 6325 } else { 6326 // A non-array may have been imprecisely marked; we need 6327 // to scan object in its entirety. 6328 size = CompactibleFreeListSpace::adjustObjectSize( 6329 p->oop_iterate(_scanningClosure)); 6330 } 6331 #ifdef ASSERT 6332 size_t direct_size = 6333 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6334 assert(size == direct_size, "Inconsistency in size"); 6335 assert(size >= 3, "Necessary for Printezis marks to work"); 6336 if (!_bitMap->isMarked(addr+1)) { 6337 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6338 } else { 6339 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6340 assert(_bitMap->isMarked(addr+size-1), 6341 "inconsistent Printezis mark"); 6342 } 6343 #endif // ASSERT 6344 } else { 6345 // An uninitialized object. 6346 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6347 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6348 size = pointer_delta(nextOneAddr + 1, addr); 6349 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6350 "alignment problem"); 6351 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6352 // will dirty the card when the klass pointer is installed in the 6353 // object (signaling the completion of initialization). 6354 } 6355 } else { 6356 // Either a not yet marked object or an uninitialized object 6357 if (p->klass_or_null() == NULL) { 6358 // An uninitialized object, skip to the next card, since 6359 // we may not be able to read its P-bits yet. 6360 assert(size == 0, "Initial value"); 6361 } else { 6362 // An object not (yet) reached by marking: we merely need to 6363 // compute its size so as to go look at the next block. 6364 assert(p->is_oop(true), "should be an oop"); 6365 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6366 } 6367 } 6368 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6369 return size; 6370 } 6371 6372 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6373 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6374 "CMS thread should hold CMS token"); 6375 assert_lock_strong(_freelistLock); 6376 assert_lock_strong(_bitMap->lock()); 6377 // relinquish the free_list_lock and bitMaplock() 6378 _bitMap->lock()->unlock(); 6379 _freelistLock->unlock(); 6380 ConcurrentMarkSweepThread::desynchronize(true); 6381 _collector->stopTimer(); 6382 if (PrintCMSStatistics != 0) { 6383 _collector->incrementYields(); 6384 } 6385 6386 // See the comment in coordinator_yield() 6387 for (unsigned i = 0; i < CMSYieldSleepCount && 6388 ConcurrentMarkSweepThread::should_yield() && 6389 !CMSCollector::foregroundGCIsActive(); ++i) { 6390 os::sleep(Thread::current(), 1, false); 6391 } 6392 6393 ConcurrentMarkSweepThread::synchronize(true); 6394 _freelistLock->lock_without_safepoint_check(); 6395 _bitMap->lock()->lock_without_safepoint_check(); 6396 _collector->startTimer(); 6397 } 6398 6399 6400 ////////////////////////////////////////////////////////////////// 6401 // SurvivorSpacePrecleanClosure 6402 ////////////////////////////////////////////////////////////////// 6403 // This (single-threaded) closure is used to preclean the oops in 6404 // the survivor spaces. 6405 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6406 6407 HeapWord* addr = (HeapWord*)p; 6408 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6409 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6410 assert(p->klass_or_null() != NULL, "object should be initialized"); 6411 // an initialized object; ignore mark word in verification below 6412 // since we are running concurrent with mutators 6413 assert(p->is_oop(true), "should be an oop"); 6414 // Note that we do not yield while we iterate over 6415 // the interior oops of p, pushing the relevant ones 6416 // on our marking stack. 6417 size_t size = p->oop_iterate(_scanning_closure); 6418 do_yield_check(); 6419 // Observe that below, we do not abandon the preclean 6420 // phase as soon as we should; rather we empty the 6421 // marking stack before returning. This is to satisfy 6422 // some existing assertions. In general, it may be a 6423 // good idea to abort immediately and complete the marking 6424 // from the grey objects at a later time. 6425 while (!_mark_stack->isEmpty()) { 6426 oop new_oop = _mark_stack->pop(); 6427 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6428 assert(_bit_map->isMarked((HeapWord*)new_oop), 6429 "only grey objects on this stack"); 6430 // iterate over the oops in this oop, marking and pushing 6431 // the ones in CMS heap (i.e. in _span). 6432 new_oop->oop_iterate(_scanning_closure); 6433 // check if it's time to yield 6434 do_yield_check(); 6435 } 6436 unsigned int after_count = 6437 GenCollectedHeap::heap()->total_collections(); 6438 bool abort = (_before_count != after_count) || 6439 _collector->should_abort_preclean(); 6440 return abort ? 0 : size; 6441 } 6442 6443 void SurvivorSpacePrecleanClosure::do_yield_work() { 6444 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6445 "CMS thread should hold CMS token"); 6446 assert_lock_strong(_bit_map->lock()); 6447 // Relinquish the bit map lock 6448 _bit_map->lock()->unlock(); 6449 ConcurrentMarkSweepThread::desynchronize(true); 6450 _collector->stopTimer(); 6451 if (PrintCMSStatistics != 0) { 6452 _collector->incrementYields(); 6453 } 6454 6455 // See the comment in coordinator_yield() 6456 for (unsigned i = 0; i < CMSYieldSleepCount && 6457 ConcurrentMarkSweepThread::should_yield() && 6458 !CMSCollector::foregroundGCIsActive(); ++i) { 6459 os::sleep(Thread::current(), 1, false); 6460 } 6461 6462 ConcurrentMarkSweepThread::synchronize(true); 6463 _bit_map->lock()->lock_without_safepoint_check(); 6464 _collector->startTimer(); 6465 } 6466 6467 // This closure is used to rescan the marked objects on the dirty cards 6468 // in the mod union table and the card table proper. In the parallel 6469 // case, although the bitMap is shared, we do a single read so the 6470 // isMarked() query is "safe". 6471 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6472 // Ignore mark word because we are running concurrent with mutators 6473 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6474 HeapWord* addr = (HeapWord*)p; 6475 assert(_span.contains(addr), "we are scanning the CMS generation"); 6476 bool is_obj_array = false; 6477 #ifdef ASSERT 6478 if (!_parallel) { 6479 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6480 assert(_collector->overflow_list_is_empty(), 6481 "overflow list should be empty"); 6482 6483 } 6484 #endif // ASSERT 6485 if (_bit_map->isMarked(addr)) { 6486 // Obj arrays are precisely marked, non-arrays are not; 6487 // so we scan objArrays precisely and non-arrays in their 6488 // entirety. 6489 if (p->is_objArray()) { 6490 is_obj_array = true; 6491 if (_parallel) { 6492 p->oop_iterate(_par_scan_closure, mr); 6493 } else { 6494 p->oop_iterate(_scan_closure, mr); 6495 } 6496 } else { 6497 if (_parallel) { 6498 p->oop_iterate(_par_scan_closure); 6499 } else { 6500 p->oop_iterate(_scan_closure); 6501 } 6502 } 6503 } 6504 #ifdef ASSERT 6505 if (!_parallel) { 6506 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6507 assert(_collector->overflow_list_is_empty(), 6508 "overflow list should be empty"); 6509 6510 } 6511 #endif // ASSERT 6512 return is_obj_array; 6513 } 6514 6515 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6516 MemRegion span, 6517 CMSBitMap* bitMap, CMSMarkStack* markStack, 6518 bool should_yield, bool verifying): 6519 _collector(collector), 6520 _span(span), 6521 _bitMap(bitMap), 6522 _mut(&collector->_modUnionTable), 6523 _markStack(markStack), 6524 _yield(should_yield), 6525 _skipBits(0) 6526 { 6527 assert(_markStack->isEmpty(), "stack should be empty"); 6528 _finger = _bitMap->startWord(); 6529 _threshold = _finger; 6530 assert(_collector->_restart_addr == NULL, "Sanity check"); 6531 assert(_span.contains(_finger), "Out of bounds _finger?"); 6532 DEBUG_ONLY(_verifying = verifying;) 6533 } 6534 6535 void MarkFromRootsClosure::reset(HeapWord* addr) { 6536 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6537 assert(_span.contains(addr), "Out of bounds _finger?"); 6538 _finger = addr; 6539 _threshold = (HeapWord*)round_to( 6540 (intptr_t)_finger, CardTableModRefBS::card_size); 6541 } 6542 6543 // Should revisit to see if this should be restructured for 6544 // greater efficiency. 6545 bool MarkFromRootsClosure::do_bit(size_t offset) { 6546 if (_skipBits > 0) { 6547 _skipBits--; 6548 return true; 6549 } 6550 // convert offset into a HeapWord* 6551 HeapWord* addr = _bitMap->startWord() + offset; 6552 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6553 "address out of range"); 6554 assert(_bitMap->isMarked(addr), "tautology"); 6555 if (_bitMap->isMarked(addr+1)) { 6556 // this is an allocated but not yet initialized object 6557 assert(_skipBits == 0, "tautology"); 6558 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6559 oop p = oop(addr); 6560 if (p->klass_or_null() == NULL) { 6561 DEBUG_ONLY(if (!_verifying) {) 6562 // We re-dirty the cards on which this object lies and increase 6563 // the _threshold so that we'll come back to scan this object 6564 // during the preclean or remark phase. (CMSCleanOnEnter) 6565 if (CMSCleanOnEnter) { 6566 size_t sz = _collector->block_size_using_printezis_bits(addr); 6567 HeapWord* end_card_addr = (HeapWord*)round_to( 6568 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6569 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6570 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6571 // Bump _threshold to end_card_addr; note that 6572 // _threshold cannot possibly exceed end_card_addr, anyhow. 6573 // This prevents future clearing of the card as the scan proceeds 6574 // to the right. 6575 assert(_threshold <= end_card_addr, 6576 "Because we are just scanning into this object"); 6577 if (_threshold < end_card_addr) { 6578 _threshold = end_card_addr; 6579 } 6580 if (p->klass_or_null() != NULL) { 6581 // Redirty the range of cards... 6582 _mut->mark_range(redirty_range); 6583 } // ...else the setting of klass will dirty the card anyway. 6584 } 6585 DEBUG_ONLY(}) 6586 return true; 6587 } 6588 } 6589 scanOopsInOop(addr); 6590 return true; 6591 } 6592 6593 // We take a break if we've been at this for a while, 6594 // so as to avoid monopolizing the locks involved. 6595 void MarkFromRootsClosure::do_yield_work() { 6596 // First give up the locks, then yield, then re-lock 6597 // We should probably use a constructor/destructor idiom to 6598 // do this unlock/lock or modify the MutexUnlocker class to 6599 // serve our purpose. XXX 6600 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6601 "CMS thread should hold CMS token"); 6602 assert_lock_strong(_bitMap->lock()); 6603 _bitMap->lock()->unlock(); 6604 ConcurrentMarkSweepThread::desynchronize(true); 6605 _collector->stopTimer(); 6606 if (PrintCMSStatistics != 0) { 6607 _collector->incrementYields(); 6608 } 6609 6610 // See the comment in coordinator_yield() 6611 for (unsigned i = 0; i < CMSYieldSleepCount && 6612 ConcurrentMarkSweepThread::should_yield() && 6613 !CMSCollector::foregroundGCIsActive(); ++i) { 6614 os::sleep(Thread::current(), 1, false); 6615 } 6616 6617 ConcurrentMarkSweepThread::synchronize(true); 6618 _bitMap->lock()->lock_without_safepoint_check(); 6619 _collector->startTimer(); 6620 } 6621 6622 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6623 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6624 assert(_markStack->isEmpty(), 6625 "should drain stack to limit stack usage"); 6626 // convert ptr to an oop preparatory to scanning 6627 oop obj = oop(ptr); 6628 // Ignore mark word in verification below, since we 6629 // may be running concurrent with mutators. 6630 assert(obj->is_oop(true), "should be an oop"); 6631 assert(_finger <= ptr, "_finger runneth ahead"); 6632 // advance the finger to right end of this object 6633 _finger = ptr + obj->size(); 6634 assert(_finger > ptr, "we just incremented it above"); 6635 // On large heaps, it may take us some time to get through 6636 // the marking phase. During 6637 // this time it's possible that a lot of mutations have 6638 // accumulated in the card table and the mod union table -- 6639 // these mutation records are redundant until we have 6640 // actually traced into the corresponding card. 6641 // Here, we check whether advancing the finger would make 6642 // us cross into a new card, and if so clear corresponding 6643 // cards in the MUT (preclean them in the card-table in the 6644 // future). 6645 6646 DEBUG_ONLY(if (!_verifying) {) 6647 // The clean-on-enter optimization is disabled by default, 6648 // until we fix 6178663. 6649 if (CMSCleanOnEnter && (_finger > _threshold)) { 6650 // [_threshold, _finger) represents the interval 6651 // of cards to be cleared in MUT (or precleaned in card table). 6652 // The set of cards to be cleared is all those that overlap 6653 // with the interval [_threshold, _finger); note that 6654 // _threshold is always kept card-aligned but _finger isn't 6655 // always card-aligned. 6656 HeapWord* old_threshold = _threshold; 6657 assert(old_threshold == (HeapWord*)round_to( 6658 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6659 "_threshold should always be card-aligned"); 6660 _threshold = (HeapWord*)round_to( 6661 (intptr_t)_finger, CardTableModRefBS::card_size); 6662 MemRegion mr(old_threshold, _threshold); 6663 assert(!mr.is_empty(), "Control point invariant"); 6664 assert(_span.contains(mr), "Should clear within span"); 6665 _mut->clear_range(mr); 6666 } 6667 DEBUG_ONLY(}) 6668 // Note: the finger doesn't advance while we drain 6669 // the stack below. 6670 PushOrMarkClosure pushOrMarkClosure(_collector, 6671 _span, _bitMap, _markStack, 6672 _finger, this); 6673 bool res = _markStack->push(obj); 6674 assert(res, "Empty non-zero size stack should have space for single push"); 6675 while (!_markStack->isEmpty()) { 6676 oop new_oop = _markStack->pop(); 6677 // Skip verifying header mark word below because we are 6678 // running concurrent with mutators. 6679 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6680 // now scan this oop's oops 6681 new_oop->oop_iterate(&pushOrMarkClosure); 6682 do_yield_check(); 6683 } 6684 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6685 } 6686 6687 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6688 CMSCollector* collector, MemRegion span, 6689 CMSBitMap* bit_map, 6690 OopTaskQueue* work_queue, 6691 CMSMarkStack* overflow_stack): 6692 _collector(collector), 6693 _whole_span(collector->_span), 6694 _span(span), 6695 _bit_map(bit_map), 6696 _mut(&collector->_modUnionTable), 6697 _work_queue(work_queue), 6698 _overflow_stack(overflow_stack), 6699 _skip_bits(0), 6700 _task(task) 6701 { 6702 assert(_work_queue->size() == 0, "work_queue should be empty"); 6703 _finger = span.start(); 6704 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6705 assert(_span.contains(_finger), "Out of bounds _finger?"); 6706 } 6707 6708 // Should revisit to see if this should be restructured for 6709 // greater efficiency. 6710 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6711 if (_skip_bits > 0) { 6712 _skip_bits--; 6713 return true; 6714 } 6715 // convert offset into a HeapWord* 6716 HeapWord* addr = _bit_map->startWord() + offset; 6717 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6718 "address out of range"); 6719 assert(_bit_map->isMarked(addr), "tautology"); 6720 if (_bit_map->isMarked(addr+1)) { 6721 // this is an allocated object that might not yet be initialized 6722 assert(_skip_bits == 0, "tautology"); 6723 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6724 oop p = oop(addr); 6725 if (p->klass_or_null() == NULL) { 6726 // in the case of Clean-on-Enter optimization, redirty card 6727 // and avoid clearing card by increasing the threshold. 6728 return true; 6729 } 6730 } 6731 scan_oops_in_oop(addr); 6732 return true; 6733 } 6734 6735 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6736 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6737 // Should we assert that our work queue is empty or 6738 // below some drain limit? 6739 assert(_work_queue->size() == 0, 6740 "should drain stack to limit stack usage"); 6741 // convert ptr to an oop preparatory to scanning 6742 oop obj = oop(ptr); 6743 // Ignore mark word in verification below, since we 6744 // may be running concurrent with mutators. 6745 assert(obj->is_oop(true), "should be an oop"); 6746 assert(_finger <= ptr, "_finger runneth ahead"); 6747 // advance the finger to right end of this object 6748 _finger = ptr + obj->size(); 6749 assert(_finger > ptr, "we just incremented it above"); 6750 // On large heaps, it may take us some time to get through 6751 // the marking phase. During 6752 // this time it's possible that a lot of mutations have 6753 // accumulated in the card table and the mod union table -- 6754 // these mutation records are redundant until we have 6755 // actually traced into the corresponding card. 6756 // Here, we check whether advancing the finger would make 6757 // us cross into a new card, and if so clear corresponding 6758 // cards in the MUT (preclean them in the card-table in the 6759 // future). 6760 6761 // The clean-on-enter optimization is disabled by default, 6762 // until we fix 6178663. 6763 if (CMSCleanOnEnter && (_finger > _threshold)) { 6764 // [_threshold, _finger) represents the interval 6765 // of cards to be cleared in MUT (or precleaned in card table). 6766 // The set of cards to be cleared is all those that overlap 6767 // with the interval [_threshold, _finger); note that 6768 // _threshold is always kept card-aligned but _finger isn't 6769 // always card-aligned. 6770 HeapWord* old_threshold = _threshold; 6771 assert(old_threshold == (HeapWord*)round_to( 6772 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6773 "_threshold should always be card-aligned"); 6774 _threshold = (HeapWord*)round_to( 6775 (intptr_t)_finger, CardTableModRefBS::card_size); 6776 MemRegion mr(old_threshold, _threshold); 6777 assert(!mr.is_empty(), "Control point invariant"); 6778 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6779 _mut->clear_range(mr); 6780 } 6781 6782 // Note: the local finger doesn't advance while we drain 6783 // the stack below, but the global finger sure can and will. 6784 HeapWord** gfa = _task->global_finger_addr(); 6785 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6786 _span, _bit_map, 6787 _work_queue, 6788 _overflow_stack, 6789 _finger, 6790 gfa, this); 6791 bool res = _work_queue->push(obj); // overflow could occur here 6792 assert(res, "Will hold once we use workqueues"); 6793 while (true) { 6794 oop new_oop; 6795 if (!_work_queue->pop_local(new_oop)) { 6796 // We emptied our work_queue; check if there's stuff that can 6797 // be gotten from the overflow stack. 6798 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6799 _overflow_stack, _work_queue)) { 6800 do_yield_check(); 6801 continue; 6802 } else { // done 6803 break; 6804 } 6805 } 6806 // Skip verifying header mark word below because we are 6807 // running concurrent with mutators. 6808 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6809 // now scan this oop's oops 6810 new_oop->oop_iterate(&pushOrMarkClosure); 6811 do_yield_check(); 6812 } 6813 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6814 } 6815 6816 // Yield in response to a request from VM Thread or 6817 // from mutators. 6818 void Par_MarkFromRootsClosure::do_yield_work() { 6819 assert(_task != NULL, "sanity"); 6820 _task->yield(); 6821 } 6822 6823 // A variant of the above used for verifying CMS marking work. 6824 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6825 MemRegion span, 6826 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6827 CMSMarkStack* mark_stack): 6828 _collector(collector), 6829 _span(span), 6830 _verification_bm(verification_bm), 6831 _cms_bm(cms_bm), 6832 _mark_stack(mark_stack), 6833 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6834 mark_stack) 6835 { 6836 assert(_mark_stack->isEmpty(), "stack should be empty"); 6837 _finger = _verification_bm->startWord(); 6838 assert(_collector->_restart_addr == NULL, "Sanity check"); 6839 assert(_span.contains(_finger), "Out of bounds _finger?"); 6840 } 6841 6842 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6843 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6844 assert(_span.contains(addr), "Out of bounds _finger?"); 6845 _finger = addr; 6846 } 6847 6848 // Should revisit to see if this should be restructured for 6849 // greater efficiency. 6850 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6851 // convert offset into a HeapWord* 6852 HeapWord* addr = _verification_bm->startWord() + offset; 6853 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6854 "address out of range"); 6855 assert(_verification_bm->isMarked(addr), "tautology"); 6856 assert(_cms_bm->isMarked(addr), "tautology"); 6857 6858 assert(_mark_stack->isEmpty(), 6859 "should drain stack to limit stack usage"); 6860 // convert addr to an oop preparatory to scanning 6861 oop obj = oop(addr); 6862 assert(obj->is_oop(), "should be an oop"); 6863 assert(_finger <= addr, "_finger runneth ahead"); 6864 // advance the finger to right end of this object 6865 _finger = addr + obj->size(); 6866 assert(_finger > addr, "we just incremented it above"); 6867 // Note: the finger doesn't advance while we drain 6868 // the stack below. 6869 bool res = _mark_stack->push(obj); 6870 assert(res, "Empty non-zero size stack should have space for single push"); 6871 while (!_mark_stack->isEmpty()) { 6872 oop new_oop = _mark_stack->pop(); 6873 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6874 // now scan this oop's oops 6875 new_oop->oop_iterate(&_pam_verify_closure); 6876 } 6877 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6878 return true; 6879 } 6880 6881 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6882 CMSCollector* collector, MemRegion span, 6883 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6884 CMSMarkStack* mark_stack): 6885 MetadataAwareOopClosure(collector->ref_processor()), 6886 _collector(collector), 6887 _span(span), 6888 _verification_bm(verification_bm), 6889 _cms_bm(cms_bm), 6890 _mark_stack(mark_stack) 6891 { } 6892 6893 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6894 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6895 6896 // Upon stack overflow, we discard (part of) the stack, 6897 // remembering the least address amongst those discarded 6898 // in CMSCollector's _restart_address. 6899 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6900 // Remember the least grey address discarded 6901 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6902 _collector->lower_restart_addr(ra); 6903 _mark_stack->reset(); // discard stack contents 6904 _mark_stack->expand(); // expand the stack if possible 6905 } 6906 6907 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6908 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6909 HeapWord* addr = (HeapWord*)obj; 6910 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6911 // Oop lies in _span and isn't yet grey or black 6912 _verification_bm->mark(addr); // now grey 6913 if (!_cms_bm->isMarked(addr)) { 6914 oop(addr)->print(); 6915 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6916 p2i(addr)); 6917 fatal("... aborting"); 6918 } 6919 6920 if (!_mark_stack->push(obj)) { // stack overflow 6921 if (PrintCMSStatistics != 0) { 6922 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6923 SIZE_FORMAT, _mark_stack->capacity()); 6924 } 6925 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6926 handle_stack_overflow(addr); 6927 } 6928 // anything including and to the right of _finger 6929 // will be scanned as we iterate over the remainder of the 6930 // bit map 6931 } 6932 } 6933 6934 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6935 MemRegion span, 6936 CMSBitMap* bitMap, CMSMarkStack* markStack, 6937 HeapWord* finger, MarkFromRootsClosure* parent) : 6938 MetadataAwareOopClosure(collector->ref_processor()), 6939 _collector(collector), 6940 _span(span), 6941 _bitMap(bitMap), 6942 _markStack(markStack), 6943 _finger(finger), 6944 _parent(parent) 6945 { } 6946 6947 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6948 MemRegion span, 6949 CMSBitMap* bit_map, 6950 OopTaskQueue* work_queue, 6951 CMSMarkStack* overflow_stack, 6952 HeapWord* finger, 6953 HeapWord** global_finger_addr, 6954 Par_MarkFromRootsClosure* parent) : 6955 MetadataAwareOopClosure(collector->ref_processor()), 6956 _collector(collector), 6957 _whole_span(collector->_span), 6958 _span(span), 6959 _bit_map(bit_map), 6960 _work_queue(work_queue), 6961 _overflow_stack(overflow_stack), 6962 _finger(finger), 6963 _global_finger_addr(global_finger_addr), 6964 _parent(parent) 6965 { } 6966 6967 // Assumes thread-safe access by callers, who are 6968 // responsible for mutual exclusion. 6969 void CMSCollector::lower_restart_addr(HeapWord* low) { 6970 assert(_span.contains(low), "Out of bounds addr"); 6971 if (_restart_addr == NULL) { 6972 _restart_addr = low; 6973 } else { 6974 _restart_addr = MIN2(_restart_addr, low); 6975 } 6976 } 6977 6978 // Upon stack overflow, we discard (part of) the stack, 6979 // remembering the least address amongst those discarded 6980 // in CMSCollector's _restart_address. 6981 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6982 // Remember the least grey address discarded 6983 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6984 _collector->lower_restart_addr(ra); 6985 _markStack->reset(); // discard stack contents 6986 _markStack->expand(); // expand the stack if possible 6987 } 6988 6989 // Upon stack overflow, we discard (part of) the stack, 6990 // remembering the least address amongst those discarded 6991 // in CMSCollector's _restart_address. 6992 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6993 // We need to do this under a mutex to prevent other 6994 // workers from interfering with the work done below. 6995 MutexLockerEx ml(_overflow_stack->par_lock(), 6996 Mutex::_no_safepoint_check_flag); 6997 // Remember the least grey address discarded 6998 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6999 _collector->lower_restart_addr(ra); 7000 _overflow_stack->reset(); // discard stack contents 7001 _overflow_stack->expand(); // expand the stack if possible 7002 } 7003 7004 void PushOrMarkClosure::do_oop(oop obj) { 7005 // Ignore mark word because we are running concurrent with mutators. 7006 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7007 HeapWord* addr = (HeapWord*)obj; 7008 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7009 // Oop lies in _span and isn't yet grey or black 7010 _bitMap->mark(addr); // now grey 7011 if (addr < _finger) { 7012 // the bit map iteration has already either passed, or 7013 // sampled, this bit in the bit map; we'll need to 7014 // use the marking stack to scan this oop's oops. 7015 bool simulate_overflow = false; 7016 NOT_PRODUCT( 7017 if (CMSMarkStackOverflowALot && 7018 _collector->simulate_overflow()) { 7019 // simulate a stack overflow 7020 simulate_overflow = true; 7021 } 7022 ) 7023 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7024 if (PrintCMSStatistics != 0) { 7025 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7026 SIZE_FORMAT, _markStack->capacity()); 7027 } 7028 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7029 handle_stack_overflow(addr); 7030 } 7031 } 7032 // anything including and to the right of _finger 7033 // will be scanned as we iterate over the remainder of the 7034 // bit map 7035 do_yield_check(); 7036 } 7037 } 7038 7039 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7040 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7041 7042 void Par_PushOrMarkClosure::do_oop(oop obj) { 7043 // Ignore mark word because we are running concurrent with mutators. 7044 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7045 HeapWord* addr = (HeapWord*)obj; 7046 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7047 // Oop lies in _span and isn't yet grey or black 7048 // We read the global_finger (volatile read) strictly after marking oop 7049 bool res = _bit_map->par_mark(addr); // now grey 7050 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7051 // Should we push this marked oop on our stack? 7052 // -- if someone else marked it, nothing to do 7053 // -- if target oop is above global finger nothing to do 7054 // -- if target oop is in chunk and above local finger 7055 // then nothing to do 7056 // -- else push on work queue 7057 if ( !res // someone else marked it, they will deal with it 7058 || (addr >= *gfa) // will be scanned in a later task 7059 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7060 return; 7061 } 7062 // the bit map iteration has already either passed, or 7063 // sampled, this bit in the bit map; we'll need to 7064 // use the marking stack to scan this oop's oops. 7065 bool simulate_overflow = false; 7066 NOT_PRODUCT( 7067 if (CMSMarkStackOverflowALot && 7068 _collector->simulate_overflow()) { 7069 // simulate a stack overflow 7070 simulate_overflow = true; 7071 } 7072 ) 7073 if (simulate_overflow || 7074 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7075 // stack overflow 7076 if (PrintCMSStatistics != 0) { 7077 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7078 SIZE_FORMAT, _overflow_stack->capacity()); 7079 } 7080 // We cannot assert that the overflow stack is full because 7081 // it may have been emptied since. 7082 assert(simulate_overflow || 7083 _work_queue->size() == _work_queue->max_elems(), 7084 "Else push should have succeeded"); 7085 handle_stack_overflow(addr); 7086 } 7087 do_yield_check(); 7088 } 7089 } 7090 7091 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7092 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7093 7094 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7095 MemRegion span, 7096 ReferenceProcessor* rp, 7097 CMSBitMap* bit_map, 7098 CMSBitMap* mod_union_table, 7099 CMSMarkStack* mark_stack, 7100 bool concurrent_precleaning): 7101 MetadataAwareOopClosure(rp), 7102 _collector(collector), 7103 _span(span), 7104 _bit_map(bit_map), 7105 _mod_union_table(mod_union_table), 7106 _mark_stack(mark_stack), 7107 _concurrent_precleaning(concurrent_precleaning) 7108 { 7109 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7110 } 7111 7112 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7113 // the non-parallel version (the parallel version appears further below.) 7114 void PushAndMarkClosure::do_oop(oop obj) { 7115 // Ignore mark word verification. If during concurrent precleaning, 7116 // the object monitor may be locked. If during the checkpoint 7117 // phases, the object may already have been reached by a different 7118 // path and may be at the end of the global overflow list (so 7119 // the mark word may be NULL). 7120 assert(obj->is_oop_or_null(true /* ignore mark word */), 7121 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7122 HeapWord* addr = (HeapWord*)obj; 7123 // Check if oop points into the CMS generation 7124 // and is not marked 7125 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7126 // a white object ... 7127 _bit_map->mark(addr); // ... now grey 7128 // push on the marking stack (grey set) 7129 bool simulate_overflow = false; 7130 NOT_PRODUCT( 7131 if (CMSMarkStackOverflowALot && 7132 _collector->simulate_overflow()) { 7133 // simulate a stack overflow 7134 simulate_overflow = true; 7135 } 7136 ) 7137 if (simulate_overflow || !_mark_stack->push(obj)) { 7138 if (_concurrent_precleaning) { 7139 // During precleaning we can just dirty the appropriate card(s) 7140 // in the mod union table, thus ensuring that the object remains 7141 // in the grey set and continue. In the case of object arrays 7142 // we need to dirty all of the cards that the object spans, 7143 // since the rescan of object arrays will be limited to the 7144 // dirty cards. 7145 // Note that no one can be interfering with us in this action 7146 // of dirtying the mod union table, so no locking or atomics 7147 // are required. 7148 if (obj->is_objArray()) { 7149 size_t sz = obj->size(); 7150 HeapWord* end_card_addr = (HeapWord*)round_to( 7151 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7152 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7153 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7154 _mod_union_table->mark_range(redirty_range); 7155 } else { 7156 _mod_union_table->mark(addr); 7157 } 7158 _collector->_ser_pmc_preclean_ovflw++; 7159 } else { 7160 // During the remark phase, we need to remember this oop 7161 // in the overflow list. 7162 _collector->push_on_overflow_list(obj); 7163 _collector->_ser_pmc_remark_ovflw++; 7164 } 7165 } 7166 } 7167 } 7168 7169 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7170 MemRegion span, 7171 ReferenceProcessor* rp, 7172 CMSBitMap* bit_map, 7173 OopTaskQueue* work_queue): 7174 MetadataAwareOopClosure(rp), 7175 _collector(collector), 7176 _span(span), 7177 _bit_map(bit_map), 7178 _work_queue(work_queue) 7179 { 7180 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7181 } 7182 7183 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7184 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7185 7186 // Grey object rescan during second checkpoint phase -- 7187 // the parallel version. 7188 void Par_PushAndMarkClosure::do_oop(oop obj) { 7189 // In the assert below, we ignore the mark word because 7190 // this oop may point to an already visited object that is 7191 // on the overflow stack (in which case the mark word has 7192 // been hijacked for chaining into the overflow stack -- 7193 // if this is the last object in the overflow stack then 7194 // its mark word will be NULL). Because this object may 7195 // have been subsequently popped off the global overflow 7196 // stack, and the mark word possibly restored to the prototypical 7197 // value, by the time we get to examined this failing assert in 7198 // the debugger, is_oop_or_null(false) may subsequently start 7199 // to hold. 7200 assert(obj->is_oop_or_null(true), 7201 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7202 HeapWord* addr = (HeapWord*)obj; 7203 // Check if oop points into the CMS generation 7204 // and is not marked 7205 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7206 // a white object ... 7207 // If we manage to "claim" the object, by being the 7208 // first thread to mark it, then we push it on our 7209 // marking stack 7210 if (_bit_map->par_mark(addr)) { // ... now grey 7211 // push on work queue (grey set) 7212 bool simulate_overflow = false; 7213 NOT_PRODUCT( 7214 if (CMSMarkStackOverflowALot && 7215 _collector->par_simulate_overflow()) { 7216 // simulate a stack overflow 7217 simulate_overflow = true; 7218 } 7219 ) 7220 if (simulate_overflow || !_work_queue->push(obj)) { 7221 _collector->par_push_on_overflow_list(obj); 7222 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7223 } 7224 } // Else, some other thread got there first 7225 } 7226 } 7227 7228 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7229 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7230 7231 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7232 Mutex* bml = _collector->bitMapLock(); 7233 assert_lock_strong(bml); 7234 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7235 "CMS thread should hold CMS token"); 7236 7237 bml->unlock(); 7238 ConcurrentMarkSweepThread::desynchronize(true); 7239 7240 _collector->stopTimer(); 7241 if (PrintCMSStatistics != 0) { 7242 _collector->incrementYields(); 7243 } 7244 7245 // See the comment in coordinator_yield() 7246 for (unsigned i = 0; i < CMSYieldSleepCount && 7247 ConcurrentMarkSweepThread::should_yield() && 7248 !CMSCollector::foregroundGCIsActive(); ++i) { 7249 os::sleep(Thread::current(), 1, false); 7250 } 7251 7252 ConcurrentMarkSweepThread::synchronize(true); 7253 bml->lock(); 7254 7255 _collector->startTimer(); 7256 } 7257 7258 bool CMSPrecleanRefsYieldClosure::should_return() { 7259 if (ConcurrentMarkSweepThread::should_yield()) { 7260 do_yield_work(); 7261 } 7262 return _collector->foregroundGCIsActive(); 7263 } 7264 7265 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7266 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7267 "mr should be aligned to start at a card boundary"); 7268 // We'd like to assert: 7269 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7270 // "mr should be a range of cards"); 7271 // However, that would be too strong in one case -- the last 7272 // partition ends at _unallocated_block which, in general, can be 7273 // an arbitrary boundary, not necessarily card aligned. 7274 if (PrintCMSStatistics != 0) { 7275 _num_dirty_cards += 7276 mr.word_size()/CardTableModRefBS::card_size_in_words; 7277 } 7278 _space->object_iterate_mem(mr, &_scan_cl); 7279 } 7280 7281 SweepClosure::SweepClosure(CMSCollector* collector, 7282 ConcurrentMarkSweepGeneration* g, 7283 CMSBitMap* bitMap, bool should_yield) : 7284 _collector(collector), 7285 _g(g), 7286 _sp(g->cmsSpace()), 7287 _limit(_sp->sweep_limit()), 7288 _freelistLock(_sp->freelistLock()), 7289 _bitMap(bitMap), 7290 _yield(should_yield), 7291 _inFreeRange(false), // No free range at beginning of sweep 7292 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7293 _lastFreeRangeCoalesced(false), 7294 _freeFinger(g->used_region().start()) 7295 { 7296 NOT_PRODUCT( 7297 _numObjectsFreed = 0; 7298 _numWordsFreed = 0; 7299 _numObjectsLive = 0; 7300 _numWordsLive = 0; 7301 _numObjectsAlreadyFree = 0; 7302 _numWordsAlreadyFree = 0; 7303 _last_fc = NULL; 7304 7305 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7306 _sp->dictionary()->initialize_dict_returned_bytes(); 7307 ) 7308 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7309 "sweep _limit out of bounds"); 7310 if (CMSTraceSweeper) { 7311 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7312 p2i(_limit)); 7313 } 7314 } 7315 7316 void SweepClosure::print_on(outputStream* st) const { 7317 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7318 p2i(_sp->bottom()), p2i(_sp->end())); 7319 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7320 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7321 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7322 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7323 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7324 } 7325 7326 #ifndef PRODUCT 7327 // Assertion checking only: no useful work in product mode -- 7328 // however, if any of the flags below become product flags, 7329 // you may need to review this code to see if it needs to be 7330 // enabled in product mode. 7331 SweepClosure::~SweepClosure() { 7332 assert_lock_strong(_freelistLock); 7333 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7334 "sweep _limit out of bounds"); 7335 if (inFreeRange()) { 7336 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7337 print(); 7338 ShouldNotReachHere(); 7339 } 7340 if (Verbose && PrintGC) { 7341 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7342 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7343 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7344 SIZE_FORMAT" bytes " 7345 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7346 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7347 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7348 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7349 * sizeof(HeapWord); 7350 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7351 7352 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7353 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7354 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7355 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7356 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7357 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7358 indexListReturnedBytes); 7359 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7360 dict_returned_bytes); 7361 } 7362 } 7363 if (CMSTraceSweeper) { 7364 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7365 p2i(_limit)); 7366 } 7367 } 7368 #endif // PRODUCT 7369 7370 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7371 bool freeRangeInFreeLists) { 7372 if (CMSTraceSweeper) { 7373 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7374 p2i(freeFinger), freeRangeInFreeLists); 7375 } 7376 assert(!inFreeRange(), "Trampling existing free range"); 7377 set_inFreeRange(true); 7378 set_lastFreeRangeCoalesced(false); 7379 7380 set_freeFinger(freeFinger); 7381 set_freeRangeInFreeLists(freeRangeInFreeLists); 7382 if (CMSTestInFreeList) { 7383 if (freeRangeInFreeLists) { 7384 FreeChunk* fc = (FreeChunk*) freeFinger; 7385 assert(fc->is_free(), "A chunk on the free list should be free."); 7386 assert(fc->size() > 0, "Free range should have a size"); 7387 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7388 } 7389 } 7390 } 7391 7392 // Note that the sweeper runs concurrently with mutators. Thus, 7393 // it is possible for direct allocation in this generation to happen 7394 // in the middle of the sweep. Note that the sweeper also coalesces 7395 // contiguous free blocks. Thus, unless the sweeper and the allocator 7396 // synchronize appropriately freshly allocated blocks may get swept up. 7397 // This is accomplished by the sweeper locking the free lists while 7398 // it is sweeping. Thus blocks that are determined to be free are 7399 // indeed free. There is however one additional complication: 7400 // blocks that have been allocated since the final checkpoint and 7401 // mark, will not have been marked and so would be treated as 7402 // unreachable and swept up. To prevent this, the allocator marks 7403 // the bit map when allocating during the sweep phase. This leads, 7404 // however, to a further complication -- objects may have been allocated 7405 // but not yet initialized -- in the sense that the header isn't yet 7406 // installed. The sweeper can not then determine the size of the block 7407 // in order to skip over it. To deal with this case, we use a technique 7408 // (due to Printezis) to encode such uninitialized block sizes in the 7409 // bit map. Since the bit map uses a bit per every HeapWord, but the 7410 // CMS generation has a minimum object size of 3 HeapWords, it follows 7411 // that "normal marks" won't be adjacent in the bit map (there will 7412 // always be at least two 0 bits between successive 1 bits). We make use 7413 // of these "unused" bits to represent uninitialized blocks -- the bit 7414 // corresponding to the start of the uninitialized object and the next 7415 // bit are both set. Finally, a 1 bit marks the end of the object that 7416 // started with the two consecutive 1 bits to indicate its potentially 7417 // uninitialized state. 7418 7419 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7420 FreeChunk* fc = (FreeChunk*)addr; 7421 size_t res; 7422 7423 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7424 // than "addr == _limit" because although _limit was a block boundary when 7425 // we started the sweep, it may no longer be one because heap expansion 7426 // may have caused us to coalesce the block ending at the address _limit 7427 // with a newly expanded chunk (this happens when _limit was set to the 7428 // previous _end of the space), so we may have stepped past _limit: 7429 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7430 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7431 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7432 "sweep _limit out of bounds"); 7433 assert(addr < _sp->end(), "addr out of bounds"); 7434 // Flush any free range we might be holding as a single 7435 // coalesced chunk to the appropriate free list. 7436 if (inFreeRange()) { 7437 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7438 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7439 flush_cur_free_chunk(freeFinger(), 7440 pointer_delta(addr, freeFinger())); 7441 if (CMSTraceSweeper) { 7442 gclog_or_tty->print("Sweep: last chunk: "); 7443 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7444 "[coalesced:%d]\n", 7445 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7446 lastFreeRangeCoalesced() ? 1 : 0); 7447 } 7448 } 7449 7450 // help the iterator loop finish 7451 return pointer_delta(_sp->end(), addr); 7452 } 7453 7454 assert(addr < _limit, "sweep invariant"); 7455 // check if we should yield 7456 do_yield_check(addr); 7457 if (fc->is_free()) { 7458 // Chunk that is already free 7459 res = fc->size(); 7460 do_already_free_chunk(fc); 7461 debug_only(_sp->verifyFreeLists()); 7462 // If we flush the chunk at hand in lookahead_and_flush() 7463 // and it's coalesced with a preceding chunk, then the 7464 // process of "mangling" the payload of the coalesced block 7465 // will cause erasure of the size information from the 7466 // (erstwhile) header of all the coalesced blocks but the 7467 // first, so the first disjunct in the assert will not hold 7468 // in that specific case (in which case the second disjunct 7469 // will hold). 7470 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7471 "Otherwise the size info doesn't change at this step"); 7472 NOT_PRODUCT( 7473 _numObjectsAlreadyFree++; 7474 _numWordsAlreadyFree += res; 7475 ) 7476 NOT_PRODUCT(_last_fc = fc;) 7477 } else if (!_bitMap->isMarked(addr)) { 7478 // Chunk is fresh garbage 7479 res = do_garbage_chunk(fc); 7480 debug_only(_sp->verifyFreeLists()); 7481 NOT_PRODUCT( 7482 _numObjectsFreed++; 7483 _numWordsFreed += res; 7484 ) 7485 } else { 7486 // Chunk that is alive. 7487 res = do_live_chunk(fc); 7488 debug_only(_sp->verifyFreeLists()); 7489 NOT_PRODUCT( 7490 _numObjectsLive++; 7491 _numWordsLive += res; 7492 ) 7493 } 7494 return res; 7495 } 7496 7497 // For the smart allocation, record following 7498 // split deaths - a free chunk is removed from its free list because 7499 // it is being split into two or more chunks. 7500 // split birth - a free chunk is being added to its free list because 7501 // a larger free chunk has been split and resulted in this free chunk. 7502 // coal death - a free chunk is being removed from its free list because 7503 // it is being coalesced into a large free chunk. 7504 // coal birth - a free chunk is being added to its free list because 7505 // it was created when two or more free chunks where coalesced into 7506 // this free chunk. 7507 // 7508 // These statistics are used to determine the desired number of free 7509 // chunks of a given size. The desired number is chosen to be relative 7510 // to the end of a CMS sweep. The desired number at the end of a sweep 7511 // is the 7512 // count-at-end-of-previous-sweep (an amount that was enough) 7513 // - count-at-beginning-of-current-sweep (the excess) 7514 // + split-births (gains in this size during interval) 7515 // - split-deaths (demands on this size during interval) 7516 // where the interval is from the end of one sweep to the end of the 7517 // next. 7518 // 7519 // When sweeping the sweeper maintains an accumulated chunk which is 7520 // the chunk that is made up of chunks that have been coalesced. That 7521 // will be termed the left-hand chunk. A new chunk of garbage that 7522 // is being considered for coalescing will be referred to as the 7523 // right-hand chunk. 7524 // 7525 // When making a decision on whether to coalesce a right-hand chunk with 7526 // the current left-hand chunk, the current count vs. the desired count 7527 // of the left-hand chunk is considered. Also if the right-hand chunk 7528 // is near the large chunk at the end of the heap (see 7529 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7530 // left-hand chunk is coalesced. 7531 // 7532 // When making a decision about whether to split a chunk, the desired count 7533 // vs. the current count of the candidate to be split is also considered. 7534 // If the candidate is underpopulated (currently fewer chunks than desired) 7535 // a chunk of an overpopulated (currently more chunks than desired) size may 7536 // be chosen. The "hint" associated with a free list, if non-null, points 7537 // to a free list which may be overpopulated. 7538 // 7539 7540 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7541 const size_t size = fc->size(); 7542 // Chunks that cannot be coalesced are not in the 7543 // free lists. 7544 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7545 assert(_sp->verify_chunk_in_free_list(fc), 7546 "free chunk should be in free lists"); 7547 } 7548 // a chunk that is already free, should not have been 7549 // marked in the bit map 7550 HeapWord* const addr = (HeapWord*) fc; 7551 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7552 // Verify that the bit map has no bits marked between 7553 // addr and purported end of this block. 7554 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7555 7556 // Some chunks cannot be coalesced under any circumstances. 7557 // See the definition of cantCoalesce(). 7558 if (!fc->cantCoalesce()) { 7559 // This chunk can potentially be coalesced. 7560 if (_sp->adaptive_freelists()) { 7561 // All the work is done in 7562 do_post_free_or_garbage_chunk(fc, size); 7563 } else { // Not adaptive free lists 7564 // this is a free chunk that can potentially be coalesced by the sweeper; 7565 if (!inFreeRange()) { 7566 // if the next chunk is a free block that can't be coalesced 7567 // it doesn't make sense to remove this chunk from the free lists 7568 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7569 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7570 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7571 nextChunk->is_free() && // ... which is free... 7572 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7573 // nothing to do 7574 } else { 7575 // Potentially the start of a new free range: 7576 // Don't eagerly remove it from the free lists. 7577 // No need to remove it if it will just be put 7578 // back again. (Also from a pragmatic point of view 7579 // if it is a free block in a region that is beyond 7580 // any allocated blocks, an assertion will fail) 7581 // Remember the start of a free run. 7582 initialize_free_range(addr, true); 7583 // end - can coalesce with next chunk 7584 } 7585 } else { 7586 // the midst of a free range, we are coalescing 7587 print_free_block_coalesced(fc); 7588 if (CMSTraceSweeper) { 7589 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7590 } 7591 // remove it from the free lists 7592 _sp->removeFreeChunkFromFreeLists(fc); 7593 set_lastFreeRangeCoalesced(true); 7594 // If the chunk is being coalesced and the current free range is 7595 // in the free lists, remove the current free range so that it 7596 // will be returned to the free lists in its entirety - all 7597 // the coalesced pieces included. 7598 if (freeRangeInFreeLists()) { 7599 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7600 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7601 "Size of free range is inconsistent with chunk size."); 7602 if (CMSTestInFreeList) { 7603 assert(_sp->verify_chunk_in_free_list(ffc), 7604 "free range is not in free lists"); 7605 } 7606 _sp->removeFreeChunkFromFreeLists(ffc); 7607 set_freeRangeInFreeLists(false); 7608 } 7609 } 7610 } 7611 // Note that if the chunk is not coalescable (the else arm 7612 // below), we unconditionally flush, without needing to do 7613 // a "lookahead," as we do below. 7614 if (inFreeRange()) lookahead_and_flush(fc, size); 7615 } else { 7616 // Code path common to both original and adaptive free lists. 7617 7618 // cant coalesce with previous block; this should be treated 7619 // as the end of a free run if any 7620 if (inFreeRange()) { 7621 // we kicked some butt; time to pick up the garbage 7622 assert(freeFinger() < addr, "freeFinger points too high"); 7623 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7624 } 7625 // else, nothing to do, just continue 7626 } 7627 } 7628 7629 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7630 // This is a chunk of garbage. It is not in any free list. 7631 // Add it to a free list or let it possibly be coalesced into 7632 // a larger chunk. 7633 HeapWord* const addr = (HeapWord*) fc; 7634 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7635 7636 if (_sp->adaptive_freelists()) { 7637 // Verify that the bit map has no bits marked between 7638 // addr and purported end of just dead object. 7639 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7640 7641 do_post_free_or_garbage_chunk(fc, size); 7642 } else { 7643 if (!inFreeRange()) { 7644 // start of a new free range 7645 assert(size > 0, "A free range should have a size"); 7646 initialize_free_range(addr, false); 7647 } else { 7648 // this will be swept up when we hit the end of the 7649 // free range 7650 if (CMSTraceSweeper) { 7651 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7652 } 7653 // If the chunk is being coalesced and the current free range is 7654 // in the free lists, remove the current free range so that it 7655 // will be returned to the free lists in its entirety - all 7656 // the coalesced pieces included. 7657 if (freeRangeInFreeLists()) { 7658 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7659 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7660 "Size of free range is inconsistent with chunk size."); 7661 if (CMSTestInFreeList) { 7662 assert(_sp->verify_chunk_in_free_list(ffc), 7663 "free range is not in free lists"); 7664 } 7665 _sp->removeFreeChunkFromFreeLists(ffc); 7666 set_freeRangeInFreeLists(false); 7667 } 7668 set_lastFreeRangeCoalesced(true); 7669 } 7670 // this will be swept up when we hit the end of the free range 7671 7672 // Verify that the bit map has no bits marked between 7673 // addr and purported end of just dead object. 7674 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7675 } 7676 assert(_limit >= addr + size, 7677 "A freshly garbage chunk can't possibly straddle over _limit"); 7678 if (inFreeRange()) lookahead_and_flush(fc, size); 7679 return size; 7680 } 7681 7682 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7683 HeapWord* addr = (HeapWord*) fc; 7684 // The sweeper has just found a live object. Return any accumulated 7685 // left hand chunk to the free lists. 7686 if (inFreeRange()) { 7687 assert(freeFinger() < addr, "freeFinger points too high"); 7688 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7689 } 7690 7691 // This object is live: we'd normally expect this to be 7692 // an oop, and like to assert the following: 7693 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7694 // However, as we commented above, this may be an object whose 7695 // header hasn't yet been initialized. 7696 size_t size; 7697 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7698 if (_bitMap->isMarked(addr + 1)) { 7699 // Determine the size from the bit map, rather than trying to 7700 // compute it from the object header. 7701 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7702 size = pointer_delta(nextOneAddr + 1, addr); 7703 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7704 "alignment problem"); 7705 7706 #ifdef ASSERT 7707 if (oop(addr)->klass_or_null() != NULL) { 7708 // Ignore mark word because we are running concurrent with mutators 7709 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7710 assert(size == 7711 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7712 "P-mark and computed size do not agree"); 7713 } 7714 #endif 7715 7716 } else { 7717 // This should be an initialized object that's alive. 7718 assert(oop(addr)->klass_or_null() != NULL, 7719 "Should be an initialized object"); 7720 // Ignore mark word because we are running concurrent with mutators 7721 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7722 // Verify that the bit map has no bits marked between 7723 // addr and purported end of this block. 7724 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7725 assert(size >= 3, "Necessary for Printezis marks to work"); 7726 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7727 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7728 } 7729 return size; 7730 } 7731 7732 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7733 size_t chunkSize) { 7734 // do_post_free_or_garbage_chunk() should only be called in the case 7735 // of the adaptive free list allocator. 7736 const bool fcInFreeLists = fc->is_free(); 7737 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7738 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7739 if (CMSTestInFreeList && fcInFreeLists) { 7740 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7741 } 7742 7743 if (CMSTraceSweeper) { 7744 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7745 } 7746 7747 HeapWord* const fc_addr = (HeapWord*) fc; 7748 7749 bool coalesce; 7750 const size_t left = pointer_delta(fc_addr, freeFinger()); 7751 const size_t right = chunkSize; 7752 switch (FLSCoalescePolicy) { 7753 // numeric value forms a coalition aggressiveness metric 7754 case 0: { // never coalesce 7755 coalesce = false; 7756 break; 7757 } 7758 case 1: { // coalesce if left & right chunks on overpopulated lists 7759 coalesce = _sp->coalOverPopulated(left) && 7760 _sp->coalOverPopulated(right); 7761 break; 7762 } 7763 case 2: { // coalesce if left chunk on overpopulated list (default) 7764 coalesce = _sp->coalOverPopulated(left); 7765 break; 7766 } 7767 case 3: { // coalesce if left OR right chunk on overpopulated list 7768 coalesce = _sp->coalOverPopulated(left) || 7769 _sp->coalOverPopulated(right); 7770 break; 7771 } 7772 case 4: { // always coalesce 7773 coalesce = true; 7774 break; 7775 } 7776 default: 7777 ShouldNotReachHere(); 7778 } 7779 7780 // Should the current free range be coalesced? 7781 // If the chunk is in a free range and either we decided to coalesce above 7782 // or the chunk is near the large block at the end of the heap 7783 // (isNearLargestChunk() returns true), then coalesce this chunk. 7784 const bool doCoalesce = inFreeRange() 7785 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7786 if (doCoalesce) { 7787 // Coalesce the current free range on the left with the new 7788 // chunk on the right. If either is on a free list, 7789 // it must be removed from the list and stashed in the closure. 7790 if (freeRangeInFreeLists()) { 7791 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7792 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7793 "Size of free range is inconsistent with chunk size."); 7794 if (CMSTestInFreeList) { 7795 assert(_sp->verify_chunk_in_free_list(ffc), 7796 "Chunk is not in free lists"); 7797 } 7798 _sp->coalDeath(ffc->size()); 7799 _sp->removeFreeChunkFromFreeLists(ffc); 7800 set_freeRangeInFreeLists(false); 7801 } 7802 if (fcInFreeLists) { 7803 _sp->coalDeath(chunkSize); 7804 assert(fc->size() == chunkSize, 7805 "The chunk has the wrong size or is not in the free lists"); 7806 _sp->removeFreeChunkFromFreeLists(fc); 7807 } 7808 set_lastFreeRangeCoalesced(true); 7809 print_free_block_coalesced(fc); 7810 } else { // not in a free range and/or should not coalesce 7811 // Return the current free range and start a new one. 7812 if (inFreeRange()) { 7813 // In a free range but cannot coalesce with the right hand chunk. 7814 // Put the current free range into the free lists. 7815 flush_cur_free_chunk(freeFinger(), 7816 pointer_delta(fc_addr, freeFinger())); 7817 } 7818 // Set up for new free range. Pass along whether the right hand 7819 // chunk is in the free lists. 7820 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7821 } 7822 } 7823 7824 // Lookahead flush: 7825 // If we are tracking a free range, and this is the last chunk that 7826 // we'll look at because its end crosses past _limit, we'll preemptively 7827 // flush it along with any free range we may be holding on to. Note that 7828 // this can be the case only for an already free or freshly garbage 7829 // chunk. If this block is an object, it can never straddle 7830 // over _limit. The "straddling" occurs when _limit is set at 7831 // the previous end of the space when this cycle started, and 7832 // a subsequent heap expansion caused the previously co-terminal 7833 // free block to be coalesced with the newly expanded portion, 7834 // thus rendering _limit a non-block-boundary making it dangerous 7835 // for the sweeper to step over and examine. 7836 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7837 assert(inFreeRange(), "Should only be called if currently in a free range."); 7838 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7839 assert(_sp->used_region().contains(eob - 1), 7840 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7841 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7842 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7843 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7844 if (eob >= _limit) { 7845 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7846 if (CMSTraceSweeper) { 7847 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7848 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7849 "[" PTR_FORMAT "," PTR_FORMAT ")", 7850 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7851 } 7852 // Return the storage we are tracking back into the free lists. 7853 if (CMSTraceSweeper) { 7854 gclog_or_tty->print_cr("Flushing ... "); 7855 } 7856 assert(freeFinger() < eob, "Error"); 7857 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7858 } 7859 } 7860 7861 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7862 assert(inFreeRange(), "Should only be called if currently in a free range."); 7863 assert(size > 0, 7864 "A zero sized chunk cannot be added to the free lists."); 7865 if (!freeRangeInFreeLists()) { 7866 if (CMSTestInFreeList) { 7867 FreeChunk* fc = (FreeChunk*) chunk; 7868 fc->set_size(size); 7869 assert(!_sp->verify_chunk_in_free_list(fc), 7870 "chunk should not be in free lists yet"); 7871 } 7872 if (CMSTraceSweeper) { 7873 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7874 p2i(chunk), size); 7875 } 7876 // A new free range is going to be starting. The current 7877 // free range has not been added to the free lists yet or 7878 // was removed so add it back. 7879 // If the current free range was coalesced, then the death 7880 // of the free range was recorded. Record a birth now. 7881 if (lastFreeRangeCoalesced()) { 7882 _sp->coalBirth(size); 7883 } 7884 _sp->addChunkAndRepairOffsetTable(chunk, size, 7885 lastFreeRangeCoalesced()); 7886 } else if (CMSTraceSweeper) { 7887 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7888 } 7889 set_inFreeRange(false); 7890 set_freeRangeInFreeLists(false); 7891 } 7892 7893 // We take a break if we've been at this for a while, 7894 // so as to avoid monopolizing the locks involved. 7895 void SweepClosure::do_yield_work(HeapWord* addr) { 7896 // Return current free chunk being used for coalescing (if any) 7897 // to the appropriate freelist. After yielding, the next 7898 // free block encountered will start a coalescing range of 7899 // free blocks. If the next free block is adjacent to the 7900 // chunk just flushed, they will need to wait for the next 7901 // sweep to be coalesced. 7902 if (inFreeRange()) { 7903 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7904 } 7905 7906 // First give up the locks, then yield, then re-lock. 7907 // We should probably use a constructor/destructor idiom to 7908 // do this unlock/lock or modify the MutexUnlocker class to 7909 // serve our purpose. XXX 7910 assert_lock_strong(_bitMap->lock()); 7911 assert_lock_strong(_freelistLock); 7912 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7913 "CMS thread should hold CMS token"); 7914 _bitMap->lock()->unlock(); 7915 _freelistLock->unlock(); 7916 ConcurrentMarkSweepThread::desynchronize(true); 7917 _collector->stopTimer(); 7918 if (PrintCMSStatistics != 0) { 7919 _collector->incrementYields(); 7920 } 7921 7922 // See the comment in coordinator_yield() 7923 for (unsigned i = 0; i < CMSYieldSleepCount && 7924 ConcurrentMarkSweepThread::should_yield() && 7925 !CMSCollector::foregroundGCIsActive(); ++i) { 7926 os::sleep(Thread::current(), 1, false); 7927 } 7928 7929 ConcurrentMarkSweepThread::synchronize(true); 7930 _freelistLock->lock(); 7931 _bitMap->lock()->lock_without_safepoint_check(); 7932 _collector->startTimer(); 7933 } 7934 7935 #ifndef PRODUCT 7936 // This is actually very useful in a product build if it can 7937 // be called from the debugger. Compile it into the product 7938 // as needed. 7939 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7940 return debug_cms_space->verify_chunk_in_free_list(fc); 7941 } 7942 #endif 7943 7944 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7945 if (CMSTraceSweeper) { 7946 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7947 p2i(fc), fc->size()); 7948 } 7949 } 7950 7951 // CMSIsAliveClosure 7952 bool CMSIsAliveClosure::do_object_b(oop obj) { 7953 HeapWord* addr = (HeapWord*)obj; 7954 return addr != NULL && 7955 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7956 } 7957 7958 7959 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7960 MemRegion span, 7961 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7962 bool cpc): 7963 _collector(collector), 7964 _span(span), 7965 _bit_map(bit_map), 7966 _mark_stack(mark_stack), 7967 _concurrent_precleaning(cpc) { 7968 assert(!_span.is_empty(), "Empty span could spell trouble"); 7969 } 7970 7971 7972 // CMSKeepAliveClosure: the serial version 7973 void CMSKeepAliveClosure::do_oop(oop obj) { 7974 HeapWord* addr = (HeapWord*)obj; 7975 if (_span.contains(addr) && 7976 !_bit_map->isMarked(addr)) { 7977 _bit_map->mark(addr); 7978 bool simulate_overflow = false; 7979 NOT_PRODUCT( 7980 if (CMSMarkStackOverflowALot && 7981 _collector->simulate_overflow()) { 7982 // simulate a stack overflow 7983 simulate_overflow = true; 7984 } 7985 ) 7986 if (simulate_overflow || !_mark_stack->push(obj)) { 7987 if (_concurrent_precleaning) { 7988 // We dirty the overflown object and let the remark 7989 // phase deal with it. 7990 assert(_collector->overflow_list_is_empty(), "Error"); 7991 // In the case of object arrays, we need to dirty all of 7992 // the cards that the object spans. No locking or atomics 7993 // are needed since no one else can be mutating the mod union 7994 // table. 7995 if (obj->is_objArray()) { 7996 size_t sz = obj->size(); 7997 HeapWord* end_card_addr = 7998 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7999 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8000 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8001 _collector->_modUnionTable.mark_range(redirty_range); 8002 } else { 8003 _collector->_modUnionTable.mark(addr); 8004 } 8005 _collector->_ser_kac_preclean_ovflw++; 8006 } else { 8007 _collector->push_on_overflow_list(obj); 8008 _collector->_ser_kac_ovflw++; 8009 } 8010 } 8011 } 8012 } 8013 8014 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8015 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8016 8017 // CMSParKeepAliveClosure: a parallel version of the above. 8018 // The work queues are private to each closure (thread), 8019 // but (may be) available for stealing by other threads. 8020 void CMSParKeepAliveClosure::do_oop(oop obj) { 8021 HeapWord* addr = (HeapWord*)obj; 8022 if (_span.contains(addr) && 8023 !_bit_map->isMarked(addr)) { 8024 // In general, during recursive tracing, several threads 8025 // may be concurrently getting here; the first one to 8026 // "tag" it, claims it. 8027 if (_bit_map->par_mark(addr)) { 8028 bool res = _work_queue->push(obj); 8029 assert(res, "Low water mark should be much less than capacity"); 8030 // Do a recursive trim in the hope that this will keep 8031 // stack usage lower, but leave some oops for potential stealers 8032 trim_queue(_low_water_mark); 8033 } // Else, another thread got there first 8034 } 8035 } 8036 8037 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8038 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8039 8040 void CMSParKeepAliveClosure::trim_queue(uint max) { 8041 while (_work_queue->size() > max) { 8042 oop new_oop; 8043 if (_work_queue->pop_local(new_oop)) { 8044 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8045 assert(_bit_map->isMarked((HeapWord*)new_oop), 8046 "no white objects on this stack!"); 8047 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8048 // iterate over the oops in this oop, marking and pushing 8049 // the ones in CMS heap (i.e. in _span). 8050 new_oop->oop_iterate(&_mark_and_push); 8051 } 8052 } 8053 } 8054 8055 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8056 CMSCollector* collector, 8057 MemRegion span, CMSBitMap* bit_map, 8058 OopTaskQueue* work_queue): 8059 _collector(collector), 8060 _span(span), 8061 _bit_map(bit_map), 8062 _work_queue(work_queue) { } 8063 8064 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8065 HeapWord* addr = (HeapWord*)obj; 8066 if (_span.contains(addr) && 8067 !_bit_map->isMarked(addr)) { 8068 if (_bit_map->par_mark(addr)) { 8069 bool simulate_overflow = false; 8070 NOT_PRODUCT( 8071 if (CMSMarkStackOverflowALot && 8072 _collector->par_simulate_overflow()) { 8073 // simulate a stack overflow 8074 simulate_overflow = true; 8075 } 8076 ) 8077 if (simulate_overflow || !_work_queue->push(obj)) { 8078 _collector->par_push_on_overflow_list(obj); 8079 _collector->_par_kac_ovflw++; 8080 } 8081 } // Else another thread got there already 8082 } 8083 } 8084 8085 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8086 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8087 8088 ////////////////////////////////////////////////////////////////// 8089 // CMSExpansionCause ///////////////////////////// 8090 ////////////////////////////////////////////////////////////////// 8091 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8092 switch (cause) { 8093 case _no_expansion: 8094 return "No expansion"; 8095 case _satisfy_free_ratio: 8096 return "Free ratio"; 8097 case _satisfy_promotion: 8098 return "Satisfy promotion"; 8099 case _satisfy_allocation: 8100 return "allocation"; 8101 case _allocate_par_lab: 8102 return "Par LAB"; 8103 case _allocate_par_spooling_space: 8104 return "Par Spooling Space"; 8105 case _adaptive_size_policy: 8106 return "Ergonomics"; 8107 default: 8108 return "unknown"; 8109 } 8110 } 8111 8112 void CMSDrainMarkingStackClosure::do_void() { 8113 // the max number to take from overflow list at a time 8114 const size_t num = _mark_stack->capacity()/4; 8115 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8116 "Overflow list should be NULL during concurrent phases"); 8117 while (!_mark_stack->isEmpty() || 8118 // if stack is empty, check the overflow list 8119 _collector->take_from_overflow_list(num, _mark_stack)) { 8120 oop obj = _mark_stack->pop(); 8121 HeapWord* addr = (HeapWord*)obj; 8122 assert(_span.contains(addr), "Should be within span"); 8123 assert(_bit_map->isMarked(addr), "Should be marked"); 8124 assert(obj->is_oop(), "Should be an oop"); 8125 obj->oop_iterate(_keep_alive); 8126 } 8127 } 8128 8129 void CMSParDrainMarkingStackClosure::do_void() { 8130 // drain queue 8131 trim_queue(0); 8132 } 8133 8134 // Trim our work_queue so its length is below max at return 8135 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8136 while (_work_queue->size() > max) { 8137 oop new_oop; 8138 if (_work_queue->pop_local(new_oop)) { 8139 assert(new_oop->is_oop(), "Expected an oop"); 8140 assert(_bit_map->isMarked((HeapWord*)new_oop), 8141 "no white objects on this stack!"); 8142 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8143 // iterate over the oops in this oop, marking and pushing 8144 // the ones in CMS heap (i.e. in _span). 8145 new_oop->oop_iterate(&_mark_and_push); 8146 } 8147 } 8148 } 8149 8150 //////////////////////////////////////////////////////////////////// 8151 // Support for Marking Stack Overflow list handling and related code 8152 //////////////////////////////////////////////////////////////////// 8153 // Much of the following code is similar in shape and spirit to the 8154 // code used in ParNewGC. We should try and share that code 8155 // as much as possible in the future. 8156 8157 #ifndef PRODUCT 8158 // Debugging support for CMSStackOverflowALot 8159 8160 // It's OK to call this multi-threaded; the worst thing 8161 // that can happen is that we'll get a bunch of closely 8162 // spaced simulated overflows, but that's OK, in fact 8163 // probably good as it would exercise the overflow code 8164 // under contention. 8165 bool CMSCollector::simulate_overflow() { 8166 if (_overflow_counter-- <= 0) { // just being defensive 8167 _overflow_counter = CMSMarkStackOverflowInterval; 8168 return true; 8169 } else { 8170 return false; 8171 } 8172 } 8173 8174 bool CMSCollector::par_simulate_overflow() { 8175 return simulate_overflow(); 8176 } 8177 #endif 8178 8179 // Single-threaded 8180 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8181 assert(stack->isEmpty(), "Expected precondition"); 8182 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8183 size_t i = num; 8184 oop cur = _overflow_list; 8185 const markOop proto = markOopDesc::prototype(); 8186 NOT_PRODUCT(ssize_t n = 0;) 8187 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8188 next = oop(cur->mark()); 8189 cur->set_mark(proto); // until proven otherwise 8190 assert(cur->is_oop(), "Should be an oop"); 8191 bool res = stack->push(cur); 8192 assert(res, "Bit off more than can chew?"); 8193 NOT_PRODUCT(n++;) 8194 } 8195 _overflow_list = cur; 8196 #ifndef PRODUCT 8197 assert(_num_par_pushes >= n, "Too many pops?"); 8198 _num_par_pushes -=n; 8199 #endif 8200 return !stack->isEmpty(); 8201 } 8202 8203 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8204 // (MT-safe) Get a prefix of at most "num" from the list. 8205 // The overflow list is chained through the mark word of 8206 // each object in the list. We fetch the entire list, 8207 // break off a prefix of the right size and return the 8208 // remainder. If other threads try to take objects from 8209 // the overflow list at that time, they will wait for 8210 // some time to see if data becomes available. If (and 8211 // only if) another thread places one or more object(s) 8212 // on the global list before we have returned the suffix 8213 // to the global list, we will walk down our local list 8214 // to find its end and append the global list to 8215 // our suffix before returning it. This suffix walk can 8216 // prove to be expensive (quadratic in the amount of traffic) 8217 // when there are many objects in the overflow list and 8218 // there is much producer-consumer contention on the list. 8219 // *NOTE*: The overflow list manipulation code here and 8220 // in ParNewGeneration:: are very similar in shape, 8221 // except that in the ParNew case we use the old (from/eden) 8222 // copy of the object to thread the list via its klass word. 8223 // Because of the common code, if you make any changes in 8224 // the code below, please check the ParNew version to see if 8225 // similar changes might be needed. 8226 // CR 6797058 has been filed to consolidate the common code. 8227 bool CMSCollector::par_take_from_overflow_list(size_t num, 8228 OopTaskQueue* work_q, 8229 int no_of_gc_threads) { 8230 assert(work_q->size() == 0, "First empty local work queue"); 8231 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8232 if (_overflow_list == NULL) { 8233 return false; 8234 } 8235 // Grab the entire list; we'll put back a suffix 8236 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8237 Thread* tid = Thread::current(); 8238 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8239 // set to ParallelGCThreads. 8240 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8241 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8242 // If the list is busy, we spin for a short while, 8243 // sleeping between attempts to get the list. 8244 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8245 os::sleep(tid, sleep_time_millis, false); 8246 if (_overflow_list == NULL) { 8247 // Nothing left to take 8248 return false; 8249 } else if (_overflow_list != BUSY) { 8250 // Try and grab the prefix 8251 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8252 } 8253 } 8254 // If the list was found to be empty, or we spun long 8255 // enough, we give up and return empty-handed. If we leave 8256 // the list in the BUSY state below, it must be the case that 8257 // some other thread holds the overflow list and will set it 8258 // to a non-BUSY state in the future. 8259 if (prefix == NULL || prefix == BUSY) { 8260 // Nothing to take or waited long enough 8261 if (prefix == NULL) { 8262 // Write back the NULL in case we overwrote it with BUSY above 8263 // and it is still the same value. 8264 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8265 } 8266 return false; 8267 } 8268 assert(prefix != NULL && prefix != BUSY, "Error"); 8269 size_t i = num; 8270 oop cur = prefix; 8271 // Walk down the first "num" objects, unless we reach the end. 8272 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8273 if (cur->mark() == NULL) { 8274 // We have "num" or fewer elements in the list, so there 8275 // is nothing to return to the global list. 8276 // Write back the NULL in lieu of the BUSY we wrote 8277 // above, if it is still the same value. 8278 if (_overflow_list == BUSY) { 8279 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8280 } 8281 } else { 8282 // Chop off the suffix and return it to the global list. 8283 assert(cur->mark() != BUSY, "Error"); 8284 oop suffix_head = cur->mark(); // suffix will be put back on global list 8285 cur->set_mark(NULL); // break off suffix 8286 // It's possible that the list is still in the empty(busy) state 8287 // we left it in a short while ago; in that case we may be 8288 // able to place back the suffix without incurring the cost 8289 // of a walk down the list. 8290 oop observed_overflow_list = _overflow_list; 8291 oop cur_overflow_list = observed_overflow_list; 8292 bool attached = false; 8293 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8294 observed_overflow_list = 8295 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8296 if (cur_overflow_list == observed_overflow_list) { 8297 attached = true; 8298 break; 8299 } else cur_overflow_list = observed_overflow_list; 8300 } 8301 if (!attached) { 8302 // Too bad, someone else sneaked in (at least) an element; we'll need 8303 // to do a splice. Find tail of suffix so we can prepend suffix to global 8304 // list. 8305 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8306 oop suffix_tail = cur; 8307 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8308 "Tautology"); 8309 observed_overflow_list = _overflow_list; 8310 do { 8311 cur_overflow_list = observed_overflow_list; 8312 if (cur_overflow_list != BUSY) { 8313 // Do the splice ... 8314 suffix_tail->set_mark(markOop(cur_overflow_list)); 8315 } else { // cur_overflow_list == BUSY 8316 suffix_tail->set_mark(NULL); 8317 } 8318 // ... and try to place spliced list back on overflow_list ... 8319 observed_overflow_list = 8320 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8321 } while (cur_overflow_list != observed_overflow_list); 8322 // ... until we have succeeded in doing so. 8323 } 8324 } 8325 8326 // Push the prefix elements on work_q 8327 assert(prefix != NULL, "control point invariant"); 8328 const markOop proto = markOopDesc::prototype(); 8329 oop next; 8330 NOT_PRODUCT(ssize_t n = 0;) 8331 for (cur = prefix; cur != NULL; cur = next) { 8332 next = oop(cur->mark()); 8333 cur->set_mark(proto); // until proven otherwise 8334 assert(cur->is_oop(), "Should be an oop"); 8335 bool res = work_q->push(cur); 8336 assert(res, "Bit off more than we can chew?"); 8337 NOT_PRODUCT(n++;) 8338 } 8339 #ifndef PRODUCT 8340 assert(_num_par_pushes >= n, "Too many pops?"); 8341 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8342 #endif 8343 return true; 8344 } 8345 8346 // Single-threaded 8347 void CMSCollector::push_on_overflow_list(oop p) { 8348 NOT_PRODUCT(_num_par_pushes++;) 8349 assert(p->is_oop(), "Not an oop"); 8350 preserve_mark_if_necessary(p); 8351 p->set_mark((markOop)_overflow_list); 8352 _overflow_list = p; 8353 } 8354 8355 // Multi-threaded; use CAS to prepend to overflow list 8356 void CMSCollector::par_push_on_overflow_list(oop p) { 8357 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8358 assert(p->is_oop(), "Not an oop"); 8359 par_preserve_mark_if_necessary(p); 8360 oop observed_overflow_list = _overflow_list; 8361 oop cur_overflow_list; 8362 do { 8363 cur_overflow_list = observed_overflow_list; 8364 if (cur_overflow_list != BUSY) { 8365 p->set_mark(markOop(cur_overflow_list)); 8366 } else { 8367 p->set_mark(NULL); 8368 } 8369 observed_overflow_list = 8370 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8371 } while (cur_overflow_list != observed_overflow_list); 8372 } 8373 #undef BUSY 8374 8375 // Single threaded 8376 // General Note on GrowableArray: pushes may silently fail 8377 // because we are (temporarily) out of C-heap for expanding 8378 // the stack. The problem is quite ubiquitous and affects 8379 // a lot of code in the JVM. The prudent thing for GrowableArray 8380 // to do (for now) is to exit with an error. However, that may 8381 // be too draconian in some cases because the caller may be 8382 // able to recover without much harm. For such cases, we 8383 // should probably introduce a "soft_push" method which returns 8384 // an indication of success or failure with the assumption that 8385 // the caller may be able to recover from a failure; code in 8386 // the VM can then be changed, incrementally, to deal with such 8387 // failures where possible, thus, incrementally hardening the VM 8388 // in such low resource situations. 8389 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8390 _preserved_oop_stack.push(p); 8391 _preserved_mark_stack.push(m); 8392 assert(m == p->mark(), "Mark word changed"); 8393 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8394 "bijection"); 8395 } 8396 8397 // Single threaded 8398 void CMSCollector::preserve_mark_if_necessary(oop p) { 8399 markOop m = p->mark(); 8400 if (m->must_be_preserved(p)) { 8401 preserve_mark_work(p, m); 8402 } 8403 } 8404 8405 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8406 markOop m = p->mark(); 8407 if (m->must_be_preserved(p)) { 8408 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8409 // Even though we read the mark word without holding 8410 // the lock, we are assured that it will not change 8411 // because we "own" this oop, so no other thread can 8412 // be trying to push it on the overflow list; see 8413 // the assertion in preserve_mark_work() that checks 8414 // that m == p->mark(). 8415 preserve_mark_work(p, m); 8416 } 8417 } 8418 8419 // We should be able to do this multi-threaded, 8420 // a chunk of stack being a task (this is 8421 // correct because each oop only ever appears 8422 // once in the overflow list. However, it's 8423 // not very easy to completely overlap this with 8424 // other operations, so will generally not be done 8425 // until all work's been completed. Because we 8426 // expect the preserved oop stack (set) to be small, 8427 // it's probably fine to do this single-threaded. 8428 // We can explore cleverer concurrent/overlapped/parallel 8429 // processing of preserved marks if we feel the 8430 // need for this in the future. Stack overflow should 8431 // be so rare in practice and, when it happens, its 8432 // effect on performance so great that this will 8433 // likely just be in the noise anyway. 8434 void CMSCollector::restore_preserved_marks_if_any() { 8435 assert(SafepointSynchronize::is_at_safepoint(), 8436 "world should be stopped"); 8437 assert(Thread::current()->is_ConcurrentGC_thread() || 8438 Thread::current()->is_VM_thread(), 8439 "should be single-threaded"); 8440 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8441 "bijection"); 8442 8443 while (!_preserved_oop_stack.is_empty()) { 8444 oop p = _preserved_oop_stack.pop(); 8445 assert(p->is_oop(), "Should be an oop"); 8446 assert(_span.contains(p), "oop should be in _span"); 8447 assert(p->mark() == markOopDesc::prototype(), 8448 "Set when taken from overflow list"); 8449 markOop m = _preserved_mark_stack.pop(); 8450 p->set_mark(m); 8451 } 8452 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8453 "stacks were cleared above"); 8454 } 8455 8456 #ifndef PRODUCT 8457 bool CMSCollector::no_preserved_marks() const { 8458 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8459 } 8460 #endif 8461 8462 // Transfer some number of overflown objects to usual marking 8463 // stack. Return true if some objects were transferred. 8464 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8465 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8466 (size_t)ParGCDesiredObjsFromOverflowList); 8467 8468 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8469 assert(_collector->overflow_list_is_empty() || res, 8470 "If list is not empty, we should have taken something"); 8471 assert(!res || !_mark_stack->isEmpty(), 8472 "If we took something, it should now be on our stack"); 8473 return res; 8474 } 8475 8476 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8477 size_t res = _sp->block_size_no_stall(addr, _collector); 8478 if (_sp->block_is_obj(addr)) { 8479 if (_live_bit_map->isMarked(addr)) { 8480 // It can't have been dead in a previous cycle 8481 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8482 } else { 8483 _dead_bit_map->mark(addr); // mark the dead object 8484 } 8485 } 8486 // Could be 0, if the block size could not be computed without stalling. 8487 return res; 8488 } 8489 8490 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8491 8492 switch (phase) { 8493 case CMSCollector::InitialMarking: 8494 initialize(true /* fullGC */ , 8495 cause /* cause of the GC */, 8496 true /* recordGCBeginTime */, 8497 true /* recordPreGCUsage */, 8498 false /* recordPeakUsage */, 8499 false /* recordPostGCusage */, 8500 true /* recordAccumulatedGCTime */, 8501 false /* recordGCEndTime */, 8502 false /* countCollection */ ); 8503 break; 8504 8505 case CMSCollector::FinalMarking: 8506 initialize(true /* fullGC */ , 8507 cause /* cause of the GC */, 8508 false /* recordGCBeginTime */, 8509 false /* recordPreGCUsage */, 8510 false /* recordPeakUsage */, 8511 false /* recordPostGCusage */, 8512 true /* recordAccumulatedGCTime */, 8513 false /* recordGCEndTime */, 8514 false /* countCollection */ ); 8515 break; 8516 8517 case CMSCollector::Sweeping: 8518 initialize(true /* fullGC */ , 8519 cause /* cause of the GC */, 8520 false /* recordGCBeginTime */, 8521 false /* recordPreGCUsage */, 8522 true /* recordPeakUsage */, 8523 true /* recordPostGCusage */, 8524 false /* recordAccumulatedGCTime */, 8525 true /* recordGCEndTime */, 8526 true /* countCollection */ ); 8527 break; 8528 8529 default: 8530 ShouldNotReachHere(); 8531 } 8532 }