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_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardGeneration.inline.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/strongRootsScope.hpp" 57 #include "memory/tenuredGeneration.hpp" 58 #include "oops/oop.inline.hpp" 59 #include "prims/jvmtiExport.hpp" 60 #include "runtime/atomic.inline.hpp" 61 #include "runtime/globals_extension.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/orderAccess.inline.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "services/memoryService.hpp" 67 #include "services/runtimeService.hpp" 68 #include "utilities/stack.inline.hpp" 69 #include "utilities/taskqueue.inline.hpp" 70 71 // statics 72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 73 bool CMSCollector::_full_gc_requested = false; 74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 75 76 ////////////////////////////////////////////////////////////////// 77 // In support of CMS/VM thread synchronization 78 ////////////////////////////////////////////////////////////////// 79 // We split use of the CGC_lock into 2 "levels". 80 // The low-level locking is of the usual CGC_lock monitor. We introduce 81 // a higher level "token" (hereafter "CMS token") built on top of the 82 // low level monitor (hereafter "CGC lock"). 83 // The token-passing protocol gives priority to the VM thread. The 84 // CMS-lock doesn't provide any fairness guarantees, but clients 85 // should ensure that it is only held for very short, bounded 86 // durations. 87 // 88 // When either of the CMS thread or the VM thread is involved in 89 // collection operations during which it does not want the other 90 // thread to interfere, it obtains the CMS token. 91 // 92 // If either thread tries to get the token while the other has 93 // it, that thread waits. However, if the VM thread and CMS thread 94 // both want the token, then the VM thread gets priority while the 95 // CMS thread waits. This ensures, for instance, that the "concurrent" 96 // phases of the CMS thread's work do not block out the VM thread 97 // for long periods of time as the CMS thread continues to hog 98 // the token. (See bug 4616232). 99 // 100 // The baton-passing functions are, however, controlled by the 101 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 102 // and here the low-level CMS lock, not the high level token, 103 // ensures mutual exclusion. 104 // 105 // Two important conditions that we have to satisfy: 106 // 1. if a thread does a low-level wait on the CMS lock, then it 107 // relinquishes the CMS token if it were holding that token 108 // when it acquired the low-level CMS lock. 109 // 2. any low-level notifications on the low-level lock 110 // should only be sent when a thread has relinquished the token. 111 // 112 // In the absence of either property, we'd have potential deadlock. 113 // 114 // We protect each of the CMS (concurrent and sequential) phases 115 // with the CMS _token_, not the CMS _lock_. 116 // 117 // The only code protected by CMS lock is the token acquisition code 118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 119 // baton-passing code. 120 // 121 // Unfortunately, i couldn't come up with a good abstraction to factor and 122 // hide the naked CGC_lock manipulation in the baton-passing code 123 // further below. That's something we should try to do. Also, the proof 124 // of correctness of this 2-level locking scheme is far from obvious, 125 // and potentially quite slippery. We have an uneasy suspicion, for instance, 126 // that there may be a theoretical possibility of delay/starvation in the 127 // low-level lock/wait/notify scheme used for the baton-passing because of 128 // potential interference with the priority scheme embodied in the 129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 130 // invocation further below and marked with "XXX 20011219YSR". 131 // Indeed, as we note elsewhere, this may become yet more slippery 132 // in the presence of multiple CMS and/or multiple VM threads. XXX 133 134 class CMSTokenSync: public StackObj { 135 private: 136 bool _is_cms_thread; 137 public: 138 CMSTokenSync(bool is_cms_thread): 139 _is_cms_thread(is_cms_thread) { 140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 141 "Incorrect argument to constructor"); 142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 143 } 144 145 ~CMSTokenSync() { 146 assert(_is_cms_thread ? 147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 149 "Incorrect state"); 150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 151 } 152 }; 153 154 // Convenience class that does a CMSTokenSync, and then acquires 155 // upto three locks. 156 class CMSTokenSyncWithLocks: public CMSTokenSync { 157 private: 158 // Note: locks are acquired in textual declaration order 159 // and released in the opposite order 160 MutexLockerEx _locker1, _locker2, _locker3; 161 public: 162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 164 CMSTokenSync(is_cms_thread), 165 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 166 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 167 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 168 { } 169 }; 170 171 172 ////////////////////////////////////////////////////////////////// 173 // Concurrent Mark-Sweep Generation ///////////////////////////// 174 ////////////////////////////////////////////////////////////////// 175 176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 177 178 // This struct contains per-thread things necessary to support parallel 179 // young-gen collection. 180 class CMSParGCThreadState: public CHeapObj<mtGC> { 181 public: 182 CFLS_LAB lab; 183 PromotionInfo promo; 184 185 // Constructor. 186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 187 promo.setSpace(cfls); 188 } 189 }; 190 191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 192 ReservedSpace rs, size_t initial_byte_size, int level, 193 CardTableRS* ct, bool use_adaptive_freelists, 194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 195 CardGeneration(rs, initial_byte_size, level, ct), 196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 (int) ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 307 308 // Generation Counters - generation 1, 1 subspace 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, 310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 311 312 _space_counters = new GSpaceCounters(gen_name, 0, 313 _virtual_space.reserved_size(), 314 this, _gen_counters); 315 } 316 317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 318 _cms_gen(cms_gen) 319 { 320 assert(alpha <= 100, "bad value"); 321 _saved_alpha = alpha; 322 323 // Initialize the alphas to the bootstrap value of 100. 324 _gc0_alpha = _cms_alpha = 100; 325 326 _cms_begin_time.update(); 327 _cms_end_time.update(); 328 329 _gc0_duration = 0.0; 330 _gc0_period = 0.0; 331 _gc0_promoted = 0; 332 333 _cms_duration = 0.0; 334 _cms_period = 0.0; 335 _cms_allocated = 0; 336 337 _cms_used_at_gc0_begin = 0; 338 _cms_used_at_gc0_end = 0; 339 _allow_duty_cycle_reduction = false; 340 _valid_bits = 0; 341 } 342 343 double CMSStats::cms_free_adjustment_factor(size_t free) const { 344 // TBD: CR 6909490 345 return 1.0; 346 } 347 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 349 } 350 351 // If promotion failure handling is on use 352 // the padded average size of the promotion for each 353 // young generation collection. 354 double CMSStats::time_until_cms_gen_full() const { 355 size_t cms_free = _cms_gen->cmsSpace()->free(); 356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 359 if (cms_free > expected_promotion) { 360 // Start a cms collection if there isn't enough space to promote 361 // for the next minor collection. Use the padded average as 362 // a safety factor. 363 cms_free -= expected_promotion; 364 365 // Adjust by the safety factor. 366 double cms_free_dbl = (double)cms_free; 367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 368 // Apply a further correction factor which tries to adjust 369 // for recent occurance of concurrent mode failures. 370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 371 cms_free_dbl = cms_free_dbl * cms_adjustment; 372 373 if (PrintGCDetails && Verbose) { 374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 376 cms_free, expected_promotion); 377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 378 cms_free_dbl, cms_consumption_rate() + 1.0); 379 } 380 // Add 1 in case the consumption rate goes to zero. 381 return cms_free_dbl / (cms_consumption_rate() + 1.0); 382 } 383 return 0.0; 384 } 385 386 // Compare the duration of the cms collection to the 387 // time remaining before the cms generation is empty. 388 // Note that the time from the start of the cms collection 389 // to the start of the cms sweep (less than the total 390 // duration of the cms collection) can be used. This 391 // has been tried and some applications experienced 392 // promotion failures early in execution. This was 393 // possibly because the averages were not accurate 394 // enough at the beginning. 395 double CMSStats::time_until_cms_start() const { 396 // We add "gc0_period" to the "work" calculation 397 // below because this query is done (mostly) at the 398 // end of a scavenge, so we need to conservatively 399 // account for that much possible delay 400 // in the query so as to avoid concurrent mode failures 401 // due to starting the collection just a wee bit too 402 // late. 403 double work = cms_duration() + gc0_period(); 404 double deadline = time_until_cms_gen_full(); 405 // If a concurrent mode failure occurred recently, we want to be 406 // more conservative and halve our expected time_until_cms_gen_full() 407 if (work > deadline) { 408 if (Verbose && PrintGCDetails) { 409 gclog_or_tty->print( 410 " CMSCollector: collect because of anticipated promotion " 411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 412 gc0_period(), time_until_cms_gen_full()); 413 } 414 return 0.0; 415 } 416 return work - deadline; 417 } 418 419 #ifndef PRODUCT 420 void CMSStats::print_on(outputStream *st) const { 421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 423 gc0_duration(), gc0_period(), gc0_promoted()); 424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 425 cms_duration(), cms_period(), cms_allocated()); 426 st->print(",cms_since_beg=%g,cms_since_end=%g", 427 cms_time_since_begin(), cms_time_since_end()); 428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 430 431 if (valid()) { 432 st->print(",promo_rate=%g,cms_alloc_rate=%g", 433 promotion_rate(), cms_allocation_rate()); 434 st->print(",cms_consumption_rate=%g,time_until_full=%g", 435 cms_consumption_rate(), time_until_cms_gen_full()); 436 } 437 st->print(" "); 438 } 439 #endif // #ifndef PRODUCT 440 441 CMSCollector::CollectorState CMSCollector::_collectorState = 442 CMSCollector::Idling; 443 bool CMSCollector::_foregroundGCIsActive = false; 444 bool CMSCollector::_foregroundGCShouldWait = false; 445 446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 447 CardTableRS* ct, 448 ConcurrentMarkSweepPolicy* cp): 449 _cmsGen(cmsGen), 450 _ct(ct), 451 _ref_processor(NULL), // will be set later 452 _conc_workers(NULL), // may be set later 453 _abort_preclean(false), 454 _start_sampling(false), 455 _between_prologue_and_epilogue(false), 456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 458 -1 /* lock-free */, "No_lock" /* dummy */), 459 _modUnionClosurePar(&_modUnionTable), 460 // Adjust my span to cover old (cms) gen 461 _span(cmsGen->reserved()), 462 // Construct the is_alive_closure with _span & markBitMap 463 _is_alive_closure(_span, &_markBitMap), 464 _restart_addr(NULL), 465 _overflow_list(NULL), 466 _stats(cmsGen), 467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 468 //verify that this lock should be acquired with safepoint check. 469 Monitor::_safepoint_check_sometimes)), 470 _eden_chunk_array(NULL), // may be set in ctor body 471 _eden_chunk_capacity(0), // -- ditto -- 472 _eden_chunk_index(0), // -- ditto -- 473 _survivor_plab_array(NULL), // -- ditto -- 474 _survivor_chunk_array(NULL), // -- ditto -- 475 _survivor_chunk_capacity(0), // -- ditto -- 476 _survivor_chunk_index(0), // -- ditto -- 477 _ser_pmc_preclean_ovflw(0), 478 _ser_kac_preclean_ovflw(0), 479 _ser_pmc_remark_ovflw(0), 480 _par_pmc_remark_ovflw(0), 481 _ser_kac_ovflw(0), 482 _par_kac_ovflw(0), 483 #ifndef PRODUCT 484 _num_par_pushes(0), 485 #endif 486 _collection_count_start(0), 487 _verifying(false), 488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 489 _completed_initialization(false), 490 _collector_policy(cp), 491 _should_unload_classes(CMSClassUnloadingEnabled), 492 _concurrent_cycles_since_last_unload(0), 493 _roots_scanning_options(GenCollectedHeap::SO_None), 494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 498 _cms_start_registered(false) 499 { 500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 501 ExplicitGCInvokesConcurrent = true; 502 } 503 // Now expand the span and allocate the collection support structures 504 // (MUT, marking bit map etc.) to cover both generations subject to 505 // collection. 506 507 // For use by dirty card to oop closures. 508 _cmsGen->cmsSpace()->set_collector(this); 509 510 // Allocate MUT and marking bit map 511 { 512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 513 if (!_markBitMap.allocate(_span)) { 514 warning("Failed to allocate CMS Bit Map"); 515 return; 516 } 517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 518 } 519 { 520 _modUnionTable.allocate(_span); 521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 522 } 523 524 if (!_markStack.allocate(MarkStackSize)) { 525 warning("Failed to allocate CMS Marking Stack"); 526 return; 527 } 528 529 // Support for multi-threaded concurrent phases 530 if (CMSConcurrentMTEnabled) { 531 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 532 // just for now 533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 534 } 535 if (ConcGCThreads > 1) { 536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 537 ConcGCThreads, true); 538 if (_conc_workers == NULL) { 539 warning("GC/CMS: _conc_workers allocation failure: " 540 "forcing -CMSConcurrentMTEnabled"); 541 CMSConcurrentMTEnabled = false; 542 } else { 543 _conc_workers->initialize_workers(); 544 } 545 } else { 546 CMSConcurrentMTEnabled = false; 547 } 548 } 549 if (!CMSConcurrentMTEnabled) { 550 ConcGCThreads = 0; 551 } else { 552 // Turn off CMSCleanOnEnter optimization temporarily for 553 // the MT case where it's not fixed yet; see 6178663. 554 CMSCleanOnEnter = false; 555 } 556 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 557 "Inconsistency"); 558 559 // Parallel task queues; these are shared for the 560 // concurrent and stop-world phases of CMS, but 561 // are not shared with parallel scavenge (ParNew). 562 { 563 uint i; 564 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 565 566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 567 || ParallelRefProcEnabled) 568 && num_queues > 0) { 569 _task_queues = new OopTaskQueueSet(num_queues); 570 if (_task_queues == NULL) { 571 warning("task_queues allocation failure."); 572 return; 573 } 574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 576 for (i = 0; i < num_queues; i++) { 577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 578 if (q == NULL) { 579 warning("work_queue allocation failure."); 580 return; 581 } 582 _task_queues->register_queue(i, q); 583 } 584 for (i = 0; i < num_queues; i++) { 585 _task_queues->queue(i)->initialize(); 586 _hash_seed[i] = 17; // copied from ParNew 587 } 588 } 589 } 590 591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 592 593 // Clip CMSBootstrapOccupancy between 0 and 100. 594 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 595 596 // Now tell CMS generations the identity of their collector 597 ConcurrentMarkSweepGeneration::set_collector(this); 598 599 // Create & start a CMS thread for this CMS collector 600 _cmsThread = ConcurrentMarkSweepThread::start(this); 601 assert(cmsThread() != NULL, "CMS Thread should have been created"); 602 assert(cmsThread()->collector() == this, 603 "CMS Thread should refer to this gen"); 604 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 605 606 // Support for parallelizing young gen rescan 607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 609 _young_gen = (ParNewGeneration*)gch->young_gen(); 610 if (gch->supports_inline_contig_alloc()) { 611 _top_addr = gch->top_addr(); 612 _end_addr = gch->end_addr(); 613 assert(_young_gen != NULL, "no _young_gen"); 614 _eden_chunk_index = 0; 615 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 617 } 618 619 // Support for parallelizing survivor space rescan 620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 621 const size_t max_plab_samples = 622 ((DefNewGeneration*)_young_gen)->max_survivor_size() / YoungPLABSize; 623 624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 627 _survivor_chunk_capacity = 2*max_plab_samples; 628 for (uint i = 0; i < ParallelGCThreads; i++) { 629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 631 assert(cur->end() == 0, "Should be 0"); 632 assert(cur->array() == vec, "Should be vec"); 633 assert(cur->capacity() == max_plab_samples, "Error"); 634 } 635 } 636 637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 638 _gc_counters = new CollectorCounters("CMS", 1); 639 _completed_initialization = true; 640 _inter_sweep_timer.start(); // start of time 641 } 642 643 const char* ConcurrentMarkSweepGeneration::name() const { 644 return "concurrent mark-sweep generation"; 645 } 646 void ConcurrentMarkSweepGeneration::update_counters() { 647 if (UsePerfData) { 648 _space_counters->update_all(); 649 _gen_counters->update_all(); 650 } 651 } 652 653 // this is an optimized version of update_counters(). it takes the 654 // used value as a parameter rather than computing it. 655 // 656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 657 if (UsePerfData) { 658 _space_counters->update_used(used); 659 _space_counters->update_capacity(); 660 _gen_counters->update_all(); 661 } 662 } 663 664 void ConcurrentMarkSweepGeneration::print() const { 665 Generation::print(); 666 cmsSpace()->print(); 667 } 668 669 #ifndef PRODUCT 670 void ConcurrentMarkSweepGeneration::print_statistics() { 671 cmsSpace()->printFLCensus(0); 672 } 673 #endif 674 675 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 676 GenCollectedHeap* gch = GenCollectedHeap::heap(); 677 if (PrintGCDetails) { 678 if (Verbose) { 679 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 680 level(), short_name(), s, used(), capacity()); 681 } else { 682 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 683 level(), short_name(), s, used() / K, capacity() / K); 684 } 685 } 686 if (Verbose) { 687 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 688 gch->used(), gch->capacity()); 689 } else { 690 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 691 gch->used() / K, gch->capacity() / K); 692 } 693 } 694 695 size_t 696 ConcurrentMarkSweepGeneration::contiguous_available() const { 697 // dld proposes an improvement in precision here. If the committed 698 // part of the space ends in a free block we should add that to 699 // uncommitted size in the calculation below. Will make this 700 // change later, staying with the approximation below for the 701 // time being. -- ysr. 702 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 703 } 704 705 size_t 706 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 707 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 708 } 709 710 size_t ConcurrentMarkSweepGeneration::max_available() const { 711 return free() + _virtual_space.uncommitted_size(); 712 } 713 714 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 715 size_t available = max_available(); 716 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 717 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 718 if (Verbose && PrintGCDetails) { 719 gclog_or_tty->print_cr( 720 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 721 "max_promo("SIZE_FORMAT")", 722 res? "":" not", available, res? ">=":"<", 723 av_promo, max_promotion_in_bytes); 724 } 725 return res; 726 } 727 728 // At a promotion failure dump information on block layout in heap 729 // (cms old generation). 730 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 731 if (CMSDumpAtPromotionFailure) { 732 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 733 } 734 } 735 736 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 737 // Clear the promotion information. These pointers can be adjusted 738 // along with all the other pointers into the heap but 739 // compaction is expected to be a rare event with 740 // a heap using cms so don't do it without seeing the need. 741 for (uint i = 0; i < ParallelGCThreads; i++) { 742 _par_gc_thread_states[i]->promo.reset(); 743 } 744 } 745 746 void ConcurrentMarkSweepGeneration::compute_new_size() { 747 assert_locked_or_safepoint(Heap_lock); 748 749 // If incremental collection failed, we just want to expand 750 // to the limit. 751 if (incremental_collection_failed()) { 752 clear_incremental_collection_failed(); 753 grow_to_reserved(); 754 return; 755 } 756 757 // The heap has been compacted but not reset yet. 758 // Any metric such as free() or used() will be incorrect. 759 760 CardGeneration::compute_new_size(); 761 762 // Reset again after a possible resizing 763 if (did_compact()) { 764 cmsSpace()->reset_after_compaction(); 765 } 766 } 767 768 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 769 assert_locked_or_safepoint(Heap_lock); 770 771 // If incremental collection failed, we just want to expand 772 // to the limit. 773 if (incremental_collection_failed()) { 774 clear_incremental_collection_failed(); 775 grow_to_reserved(); 776 return; 777 } 778 779 double free_percentage = ((double) free()) / capacity(); 780 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 781 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 782 783 // compute expansion delta needed for reaching desired free percentage 784 if (free_percentage < desired_free_percentage) { 785 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 786 assert(desired_capacity >= capacity(), "invalid expansion size"); 787 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 788 if (PrintGCDetails && Verbose) { 789 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 790 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 791 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 792 gclog_or_tty->print_cr(" Desired free fraction %f", 793 desired_free_percentage); 794 gclog_or_tty->print_cr(" Maximum free fraction %f", 795 maximum_free_percentage); 796 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 797 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 798 desired_capacity/1000); 799 int prev_level = level() - 1; 800 if (prev_level >= 0) { 801 size_t prev_size = 0; 802 GenCollectedHeap* gch = GenCollectedHeap::heap(); 803 Generation* prev_gen = gch->young_gen(); 804 prev_size = prev_gen->capacity(); 805 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 806 prev_size/1000); 807 } 808 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 809 unsafe_max_alloc_nogc()/1000); 810 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 811 contiguous_available()/1000); 812 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 813 expand_bytes); 814 } 815 // safe if expansion fails 816 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 817 if (PrintGCDetails && Verbose) { 818 gclog_or_tty->print_cr(" Expanded free fraction %f", 819 ((double) free()) / capacity()); 820 } 821 } else { 822 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 823 assert(desired_capacity <= capacity(), "invalid expansion size"); 824 size_t shrink_bytes = capacity() - desired_capacity; 825 // Don't shrink unless the delta is greater than the minimum shrink we want 826 if (shrink_bytes >= MinHeapDeltaBytes) { 827 shrink_free_list_by(shrink_bytes); 828 } 829 } 830 } 831 832 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 833 return cmsSpace()->freelistLock(); 834 } 835 836 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 837 bool tlab) { 838 CMSSynchronousYieldRequest yr; 839 MutexLockerEx x(freelistLock(), 840 Mutex::_no_safepoint_check_flag); 841 return have_lock_and_allocate(size, tlab); 842 } 843 844 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 845 bool tlab /* ignored */) { 846 assert_lock_strong(freelistLock()); 847 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 848 HeapWord* res = cmsSpace()->allocate(adjustedSize); 849 // Allocate the object live (grey) if the background collector has 850 // started marking. This is necessary because the marker may 851 // have passed this address and consequently this object will 852 // not otherwise be greyed and would be incorrectly swept up. 853 // Note that if this object contains references, the writing 854 // of those references will dirty the card containing this object 855 // allowing the object to be blackened (and its references scanned) 856 // either during a preclean phase or at the final checkpoint. 857 if (res != NULL) { 858 // We may block here with an uninitialized object with 859 // its mark-bit or P-bits not yet set. Such objects need 860 // to be safely navigable by block_start(). 861 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 862 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 863 collector()->direct_allocated(res, adjustedSize); 864 _direct_allocated_words += adjustedSize; 865 // allocation counters 866 NOT_PRODUCT( 867 _numObjectsAllocated++; 868 _numWordsAllocated += (int)adjustedSize; 869 ) 870 } 871 return res; 872 } 873 874 // In the case of direct allocation by mutators in a generation that 875 // is being concurrently collected, the object must be allocated 876 // live (grey) if the background collector has started marking. 877 // This is necessary because the marker may 878 // have passed this address and consequently this object will 879 // not otherwise be greyed and would be incorrectly swept up. 880 // Note that if this object contains references, the writing 881 // of those references will dirty the card containing this object 882 // allowing the object to be blackened (and its references scanned) 883 // either during a preclean phase or at the final checkpoint. 884 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 885 assert(_markBitMap.covers(start, size), "Out of bounds"); 886 if (_collectorState >= Marking) { 887 MutexLockerEx y(_markBitMap.lock(), 888 Mutex::_no_safepoint_check_flag); 889 // [see comments preceding SweepClosure::do_blk() below for details] 890 // 891 // Can the P-bits be deleted now? JJJ 892 // 893 // 1. need to mark the object as live so it isn't collected 894 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 895 // 3. need to mark the end of the object so marking, precleaning or sweeping 896 // can skip over uninitialized or unparsable objects. An allocated 897 // object is considered uninitialized for our purposes as long as 898 // its klass word is NULL. All old gen objects are parsable 899 // as soon as they are initialized.) 900 _markBitMap.mark(start); // object is live 901 _markBitMap.mark(start + 1); // object is potentially uninitialized? 902 _markBitMap.mark(start + size - 1); 903 // mark end of object 904 } 905 // check that oop looks uninitialized 906 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 907 } 908 909 void CMSCollector::promoted(bool par, HeapWord* start, 910 bool is_obj_array, size_t obj_size) { 911 assert(_markBitMap.covers(start), "Out of bounds"); 912 // See comment in direct_allocated() about when objects should 913 // be allocated live. 914 if (_collectorState >= Marking) { 915 // we already hold the marking bit map lock, taken in 916 // the prologue 917 if (par) { 918 _markBitMap.par_mark(start); 919 } else { 920 _markBitMap.mark(start); 921 } 922 // We don't need to mark the object as uninitialized (as 923 // in direct_allocated above) because this is being done with the 924 // world stopped and the object will be initialized by the 925 // time the marking, precleaning or sweeping get to look at it. 926 // But see the code for copying objects into the CMS generation, 927 // where we need to ensure that concurrent readers of the 928 // block offset table are able to safely navigate a block that 929 // is in flux from being free to being allocated (and in 930 // transition while being copied into) and subsequently 931 // becoming a bona-fide object when the copy/promotion is complete. 932 assert(SafepointSynchronize::is_at_safepoint(), 933 "expect promotion only at safepoints"); 934 935 if (_collectorState < Sweeping) { 936 // Mark the appropriate cards in the modUnionTable, so that 937 // this object gets scanned before the sweep. If this is 938 // not done, CMS generation references in the object might 939 // not get marked. 940 // For the case of arrays, which are otherwise precisely 941 // marked, we need to dirty the entire array, not just its head. 942 if (is_obj_array) { 943 // The [par_]mark_range() method expects mr.end() below to 944 // be aligned to the granularity of a bit's representation 945 // in the heap. In the case of the MUT below, that's a 946 // card size. 947 MemRegion mr(start, 948 (HeapWord*)round_to((intptr_t)(start + obj_size), 949 CardTableModRefBS::card_size /* bytes */)); 950 if (par) { 951 _modUnionTable.par_mark_range(mr); 952 } else { 953 _modUnionTable.mark_range(mr); 954 } 955 } else { // not an obj array; we can just mark the head 956 if (par) { 957 _modUnionTable.par_mark(start); 958 } else { 959 _modUnionTable.mark(start); 960 } 961 } 962 } 963 } 964 } 965 966 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 967 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 968 // allocate, copy and if necessary update promoinfo -- 969 // delegate to underlying space. 970 assert_lock_strong(freelistLock()); 971 972 #ifndef PRODUCT 973 if (GenCollectedHeap::heap()->promotion_should_fail()) { 974 return NULL; 975 } 976 #endif // #ifndef PRODUCT 977 978 oop res = _cmsSpace->promote(obj, obj_size); 979 if (res == NULL) { 980 // expand and retry 981 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 982 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 983 // Since this is the old generation, we don't try to promote 984 // into a more senior generation. 985 res = _cmsSpace->promote(obj, obj_size); 986 } 987 if (res != NULL) { 988 // See comment in allocate() about when objects should 989 // be allocated live. 990 assert(obj->is_oop(), "Will dereference klass pointer below"); 991 collector()->promoted(false, // Not parallel 992 (HeapWord*)res, obj->is_objArray(), obj_size); 993 // promotion counters 994 NOT_PRODUCT( 995 _numObjectsPromoted++; 996 _numWordsPromoted += 997 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 998 ) 999 } 1000 return res; 1001 } 1002 1003 1004 // IMPORTANT: Notes on object size recognition in CMS. 1005 // --------------------------------------------------- 1006 // A block of storage in the CMS generation is always in 1007 // one of three states. A free block (FREE), an allocated 1008 // object (OBJECT) whose size() method reports the correct size, 1009 // and an intermediate state (TRANSIENT) in which its size cannot 1010 // be accurately determined. 1011 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1012 // ----------------------------------------------------- 1013 // FREE: klass_word & 1 == 1; mark_word holds block size 1014 // 1015 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1016 // obj->size() computes correct size 1017 // 1018 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1019 // 1020 // STATE IDENTIFICATION: (64 bit+COOPS) 1021 // ------------------------------------ 1022 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1023 // 1024 // OBJECT: klass_word installed; klass_word != 0; 1025 // obj->size() computes correct size 1026 // 1027 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1028 // 1029 // 1030 // STATE TRANSITION DIAGRAM 1031 // 1032 // mut / parnew mut / parnew 1033 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1034 // ^ | 1035 // |------------------------ DEAD <------------------------------------| 1036 // sweep mut 1037 // 1038 // While a block is in TRANSIENT state its size cannot be determined 1039 // so readers will either need to come back later or stall until 1040 // the size can be determined. Note that for the case of direct 1041 // allocation, P-bits, when available, may be used to determine the 1042 // size of an object that may not yet have been initialized. 1043 1044 // Things to support parallel young-gen collection. 1045 oop 1046 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1047 oop old, markOop m, 1048 size_t word_sz) { 1049 #ifndef PRODUCT 1050 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1051 return NULL; 1052 } 1053 #endif // #ifndef PRODUCT 1054 1055 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1056 PromotionInfo* promoInfo = &ps->promo; 1057 // if we are tracking promotions, then first ensure space for 1058 // promotion (including spooling space for saving header if necessary). 1059 // then allocate and copy, then track promoted info if needed. 1060 // When tracking (see PromotionInfo::track()), the mark word may 1061 // be displaced and in this case restoration of the mark word 1062 // occurs in the (oop_since_save_marks_)iterate phase. 1063 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1064 // Out of space for allocating spooling buffers; 1065 // try expanding and allocating spooling buffers. 1066 if (!expand_and_ensure_spooling_space(promoInfo)) { 1067 return NULL; 1068 } 1069 } 1070 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1071 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1072 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1073 if (obj_ptr == NULL) { 1074 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1075 if (obj_ptr == NULL) { 1076 return NULL; 1077 } 1078 } 1079 oop obj = oop(obj_ptr); 1080 OrderAccess::storestore(); 1081 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1082 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1083 // IMPORTANT: See note on object initialization for CMS above. 1084 // Otherwise, copy the object. Here we must be careful to insert the 1085 // klass pointer last, since this marks the block as an allocated object. 1086 // Except with compressed oops it's the mark word. 1087 HeapWord* old_ptr = (HeapWord*)old; 1088 // Restore the mark word copied above. 1089 obj->set_mark(m); 1090 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1091 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1092 OrderAccess::storestore(); 1093 1094 if (UseCompressedClassPointers) { 1095 // Copy gap missed by (aligned) header size calculation below 1096 obj->set_klass_gap(old->klass_gap()); 1097 } 1098 if (word_sz > (size_t)oopDesc::header_size()) { 1099 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1100 obj_ptr + oopDesc::header_size(), 1101 word_sz - oopDesc::header_size()); 1102 } 1103 1104 // Now we can track the promoted object, if necessary. We take care 1105 // to delay the transition from uninitialized to full object 1106 // (i.e., insertion of klass pointer) until after, so that it 1107 // atomically becomes a promoted object. 1108 if (promoInfo->tracking()) { 1109 promoInfo->track((PromotedObject*)obj, old->klass()); 1110 } 1111 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1112 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1113 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1114 1115 // Finally, install the klass pointer (this should be volatile). 1116 OrderAccess::storestore(); 1117 obj->set_klass(old->klass()); 1118 // We should now be able to calculate the right size for this object 1119 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1120 1121 collector()->promoted(true, // parallel 1122 obj_ptr, old->is_objArray(), word_sz); 1123 1124 NOT_PRODUCT( 1125 Atomic::inc_ptr(&_numObjectsPromoted); 1126 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1127 ) 1128 1129 return obj; 1130 } 1131 1132 void 1133 ConcurrentMarkSweepGeneration:: 1134 par_promote_alloc_done(int thread_num) { 1135 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1136 ps->lab.retire(thread_num); 1137 } 1138 1139 void 1140 ConcurrentMarkSweepGeneration:: 1141 par_oop_since_save_marks_iterate_done(int thread_num) { 1142 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1143 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1144 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1145 } 1146 1147 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1148 size_t size, 1149 bool tlab) 1150 { 1151 // We allow a STW collection only if a full 1152 // collection was requested. 1153 return full || should_allocate(size, tlab); // FIX ME !!! 1154 // This and promotion failure handling are connected at the 1155 // hip and should be fixed by untying them. 1156 } 1157 1158 bool CMSCollector::shouldConcurrentCollect() { 1159 if (_full_gc_requested) { 1160 if (Verbose && PrintGCDetails) { 1161 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1162 " gc request (or gc_locker)"); 1163 } 1164 return true; 1165 } 1166 1167 FreelistLocker x(this); 1168 // ------------------------------------------------------------------ 1169 // Print out lots of information which affects the initiation of 1170 // a collection. 1171 if (PrintCMSInitiationStatistics && stats().valid()) { 1172 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1173 gclog_or_tty->stamp(); 1174 gclog_or_tty->cr(); 1175 stats().print_on(gclog_or_tty); 1176 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1177 stats().time_until_cms_gen_full()); 1178 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1179 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1180 _cmsGen->contiguous_available()); 1181 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1182 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1183 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1184 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1185 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1186 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1187 gclog_or_tty->print_cr("metadata initialized %d", 1188 MetaspaceGC::should_concurrent_collect()); 1189 } 1190 // ------------------------------------------------------------------ 1191 1192 // If the estimated time to complete a cms collection (cms_duration()) 1193 // is less than the estimated time remaining until the cms generation 1194 // is full, start a collection. 1195 if (!UseCMSInitiatingOccupancyOnly) { 1196 if (stats().valid()) { 1197 if (stats().time_until_cms_start() == 0.0) { 1198 return true; 1199 } 1200 } else { 1201 // We want to conservatively collect somewhat early in order 1202 // to try and "bootstrap" our CMS/promotion statistics; 1203 // this branch will not fire after the first successful CMS 1204 // collection because the stats should then be valid. 1205 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1206 if (Verbose && PrintGCDetails) { 1207 gclog_or_tty->print_cr( 1208 " CMSCollector: collect for bootstrapping statistics:" 1209 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1210 _bootstrap_occupancy); 1211 } 1212 return true; 1213 } 1214 } 1215 } 1216 1217 // Otherwise, we start a collection cycle if 1218 // old gen want a collection cycle started. Each may use 1219 // an appropriate criterion for making this decision. 1220 // XXX We need to make sure that the gen expansion 1221 // criterion dovetails well with this. XXX NEED TO FIX THIS 1222 if (_cmsGen->should_concurrent_collect()) { 1223 if (Verbose && PrintGCDetails) { 1224 gclog_or_tty->print_cr("CMS old gen initiated"); 1225 } 1226 return true; 1227 } 1228 1229 // We start a collection if we believe an incremental collection may fail; 1230 // this is not likely to be productive in practice because it's probably too 1231 // late anyway. 1232 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1233 assert(gch->collector_policy()->is_generation_policy(), 1234 "You may want to check the correctness of the following"); 1235 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1236 if (Verbose && PrintGCDetails) { 1237 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1238 } 1239 return true; 1240 } 1241 1242 if (MetaspaceGC::should_concurrent_collect()) { 1243 if (Verbose && PrintGCDetails) { 1244 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1245 } 1246 return true; 1247 } 1248 1249 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1250 if (CMSTriggerInterval >= 0) { 1251 if (CMSTriggerInterval == 0) { 1252 // Trigger always 1253 return true; 1254 } 1255 1256 // Check the CMS time since begin (we do not check the stats validity 1257 // as we want to be able to trigger the first CMS cycle as well) 1258 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1259 if (Verbose && PrintGCDetails) { 1260 if (stats().valid()) { 1261 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1262 stats().cms_time_since_begin()); 1263 } else { 1264 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1265 } 1266 } 1267 return true; 1268 } 1269 } 1270 1271 return false; 1272 } 1273 1274 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1275 1276 // Clear _expansion_cause fields of constituent generations 1277 void CMSCollector::clear_expansion_cause() { 1278 _cmsGen->clear_expansion_cause(); 1279 } 1280 1281 // We should be conservative in starting a collection cycle. To 1282 // start too eagerly runs the risk of collecting too often in the 1283 // extreme. To collect too rarely falls back on full collections, 1284 // which works, even if not optimum in terms of concurrent work. 1285 // As a work around for too eagerly collecting, use the flag 1286 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1287 // giving the user an easily understandable way of controlling the 1288 // collections. 1289 // We want to start a new collection cycle if any of the following 1290 // conditions hold: 1291 // . our current occupancy exceeds the configured initiating occupancy 1292 // for this generation, or 1293 // . we recently needed to expand this space and have not, since that 1294 // expansion, done a collection of this generation, or 1295 // . the underlying space believes that it may be a good idea to initiate 1296 // a concurrent collection (this may be based on criteria such as the 1297 // following: the space uses linear allocation and linear allocation is 1298 // going to fail, or there is believed to be excessive fragmentation in 1299 // the generation, etc... or ... 1300 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1301 // the case of the old generation; see CR 6543076): 1302 // we may be approaching a point at which allocation requests may fail because 1303 // we will be out of sufficient free space given allocation rate estimates.] 1304 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1305 1306 assert_lock_strong(freelistLock()); 1307 if (occupancy() > initiating_occupancy()) { 1308 if (PrintGCDetails && Verbose) { 1309 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1310 short_name(), occupancy(), initiating_occupancy()); 1311 } 1312 return true; 1313 } 1314 if (UseCMSInitiatingOccupancyOnly) { 1315 return false; 1316 } 1317 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1318 if (PrintGCDetails && Verbose) { 1319 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1320 short_name()); 1321 } 1322 return true; 1323 } 1324 if (_cmsSpace->should_concurrent_collect()) { 1325 if (PrintGCDetails && Verbose) { 1326 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1327 short_name()); 1328 } 1329 return true; 1330 } 1331 return false; 1332 } 1333 1334 void ConcurrentMarkSweepGeneration::collect(bool full, 1335 bool clear_all_soft_refs, 1336 size_t size, 1337 bool tlab) 1338 { 1339 collector()->collect(full, clear_all_soft_refs, size, tlab); 1340 } 1341 1342 void CMSCollector::collect(bool full, 1343 bool clear_all_soft_refs, 1344 size_t size, 1345 bool tlab) 1346 { 1347 // The following "if" branch is present for defensive reasons. 1348 // In the current uses of this interface, it can be replaced with: 1349 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1350 // But I am not placing that assert here to allow future 1351 // generality in invoking this interface. 1352 if (GC_locker::is_active()) { 1353 // A consistency test for GC_locker 1354 assert(GC_locker::needs_gc(), "Should have been set already"); 1355 // Skip this foreground collection, instead 1356 // expanding the heap if necessary. 1357 // Need the free list locks for the call to free() in compute_new_size() 1358 compute_new_size(); 1359 return; 1360 } 1361 acquire_control_and_collect(full, clear_all_soft_refs); 1362 } 1363 1364 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1365 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1366 unsigned int gc_count = gch->total_full_collections(); 1367 if (gc_count == full_gc_count) { 1368 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1369 _full_gc_requested = true; 1370 _full_gc_cause = cause; 1371 CGC_lock->notify(); // nudge CMS thread 1372 } else { 1373 assert(gc_count > full_gc_count, "Error: causal loop"); 1374 } 1375 } 1376 1377 bool CMSCollector::is_external_interruption() { 1378 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1379 return GCCause::is_user_requested_gc(cause) || 1380 GCCause::is_serviceability_requested_gc(cause); 1381 } 1382 1383 void CMSCollector::report_concurrent_mode_interruption() { 1384 if (is_external_interruption()) { 1385 if (PrintGCDetails) { 1386 gclog_or_tty->print(" (concurrent mode interrupted)"); 1387 } 1388 } else { 1389 if (PrintGCDetails) { 1390 gclog_or_tty->print(" (concurrent mode failure)"); 1391 } 1392 _gc_tracer_cm->report_concurrent_mode_failure(); 1393 } 1394 } 1395 1396 1397 // The foreground and background collectors need to coordinate in order 1398 // to make sure that they do not mutually interfere with CMS collections. 1399 // When a background collection is active, 1400 // the foreground collector may need to take over (preempt) and 1401 // synchronously complete an ongoing collection. Depending on the 1402 // frequency of the background collections and the heap usage 1403 // of the application, this preemption can be seldom or frequent. 1404 // There are only certain 1405 // points in the background collection that the "collection-baton" 1406 // can be passed to the foreground collector. 1407 // 1408 // The foreground collector will wait for the baton before 1409 // starting any part of the collection. The foreground collector 1410 // will only wait at one location. 1411 // 1412 // The background collector will yield the baton before starting a new 1413 // phase of the collection (e.g., before initial marking, marking from roots, 1414 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1415 // of the loop which switches the phases. The background collector does some 1416 // of the phases (initial mark, final re-mark) with the world stopped. 1417 // Because of locking involved in stopping the world, 1418 // the foreground collector should not block waiting for the background 1419 // collector when it is doing a stop-the-world phase. The background 1420 // collector will yield the baton at an additional point just before 1421 // it enters a stop-the-world phase. Once the world is stopped, the 1422 // background collector checks the phase of the collection. If the 1423 // phase has not changed, it proceeds with the collection. If the 1424 // phase has changed, it skips that phase of the collection. See 1425 // the comments on the use of the Heap_lock in collect_in_background(). 1426 // 1427 // Variable used in baton passing. 1428 // _foregroundGCIsActive - Set to true by the foreground collector when 1429 // it wants the baton. The foreground clears it when it has finished 1430 // the collection. 1431 // _foregroundGCShouldWait - Set to true by the background collector 1432 // when it is running. The foreground collector waits while 1433 // _foregroundGCShouldWait is true. 1434 // CGC_lock - monitor used to protect access to the above variables 1435 // and to notify the foreground and background collectors. 1436 // _collectorState - current state of the CMS collection. 1437 // 1438 // The foreground collector 1439 // acquires the CGC_lock 1440 // sets _foregroundGCIsActive 1441 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1442 // various locks acquired in preparation for the collection 1443 // are released so as not to block the background collector 1444 // that is in the midst of a collection 1445 // proceeds with the collection 1446 // clears _foregroundGCIsActive 1447 // returns 1448 // 1449 // The background collector in a loop iterating on the phases of the 1450 // collection 1451 // acquires the CGC_lock 1452 // sets _foregroundGCShouldWait 1453 // if _foregroundGCIsActive is set 1454 // clears _foregroundGCShouldWait, notifies _CGC_lock 1455 // waits on _CGC_lock for _foregroundGCIsActive to become false 1456 // and exits the loop. 1457 // otherwise 1458 // proceed with that phase of the collection 1459 // if the phase is a stop-the-world phase, 1460 // yield the baton once more just before enqueueing 1461 // the stop-world CMS operation (executed by the VM thread). 1462 // returns after all phases of the collection are done 1463 // 1464 1465 void CMSCollector::acquire_control_and_collect(bool full, 1466 bool clear_all_soft_refs) { 1467 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1468 assert(!Thread::current()->is_ConcurrentGC_thread(), 1469 "shouldn't try to acquire control from self!"); 1470 1471 // Start the protocol for acquiring control of the 1472 // collection from the background collector (aka CMS thread). 1473 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1474 "VM thread should have CMS token"); 1475 // Remember the possibly interrupted state of an ongoing 1476 // concurrent collection 1477 CollectorState first_state = _collectorState; 1478 1479 // Signal to a possibly ongoing concurrent collection that 1480 // we want to do a foreground collection. 1481 _foregroundGCIsActive = true; 1482 1483 // release locks and wait for a notify from the background collector 1484 // releasing the locks in only necessary for phases which 1485 // do yields to improve the granularity of the collection. 1486 assert_lock_strong(bitMapLock()); 1487 // We need to lock the Free list lock for the space that we are 1488 // currently collecting. 1489 assert(haveFreelistLocks(), "Must be holding free list locks"); 1490 bitMapLock()->unlock(); 1491 releaseFreelistLocks(); 1492 { 1493 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1494 if (_foregroundGCShouldWait) { 1495 // We are going to be waiting for action for the CMS thread; 1496 // it had better not be gone (for instance at shutdown)! 1497 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1498 "CMS thread must be running"); 1499 // Wait here until the background collector gives us the go-ahead 1500 ConcurrentMarkSweepThread::clear_CMS_flag( 1501 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1502 // Get a possibly blocked CMS thread going: 1503 // Note that we set _foregroundGCIsActive true above, 1504 // without protection of the CGC_lock. 1505 CGC_lock->notify(); 1506 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1507 "Possible deadlock"); 1508 while (_foregroundGCShouldWait) { 1509 // wait for notification 1510 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1511 // Possibility of delay/starvation here, since CMS token does 1512 // not know to give priority to VM thread? Actually, i think 1513 // there wouldn't be any delay/starvation, but the proof of 1514 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1515 } 1516 ConcurrentMarkSweepThread::set_CMS_flag( 1517 ConcurrentMarkSweepThread::CMS_vm_has_token); 1518 } 1519 } 1520 // The CMS_token is already held. Get back the other locks. 1521 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1522 "VM thread should have CMS token"); 1523 getFreelistLocks(); 1524 bitMapLock()->lock_without_safepoint_check(); 1525 if (TraceCMSState) { 1526 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1527 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1528 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1529 } 1530 1531 // Inform cms gen if this was due to partial collection failing. 1532 // The CMS gen may use this fact to determine its expansion policy. 1533 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1534 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1535 assert(!_cmsGen->incremental_collection_failed(), 1536 "Should have been noticed, reacted to and cleared"); 1537 _cmsGen->set_incremental_collection_failed(); 1538 } 1539 1540 if (first_state > Idling) { 1541 report_concurrent_mode_interruption(); 1542 } 1543 1544 set_did_compact(true); 1545 1546 // If the collection is being acquired from the background 1547 // collector, there may be references on the discovered 1548 // references lists. Abandon those references, since some 1549 // of them may have become unreachable after concurrent 1550 // discovery; the STW compacting collector will redo discovery 1551 // more precisely, without being subject to floating garbage. 1552 // Leaving otherwise unreachable references in the discovered 1553 // lists would require special handling. 1554 ref_processor()->disable_discovery(); 1555 ref_processor()->abandon_partial_discovery(); 1556 ref_processor()->verify_no_references_recorded(); 1557 1558 if (first_state > Idling) { 1559 save_heap_summary(); 1560 } 1561 1562 do_compaction_work(clear_all_soft_refs); 1563 1564 // Has the GC time limit been exceeded? 1565 size_t max_eden_size = _young_gen->max_capacity() - 1566 _young_gen->to()->capacity() - 1567 _young_gen->from()->capacity(); 1568 GCCause::Cause gc_cause = gch->gc_cause(); 1569 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1570 _young_gen->eden()->used(), 1571 _cmsGen->max_capacity(), 1572 max_eden_size, 1573 full, 1574 gc_cause, 1575 gch->collector_policy()); 1576 1577 // Reset the expansion cause, now that we just completed 1578 // a collection cycle. 1579 clear_expansion_cause(); 1580 _foregroundGCIsActive = false; 1581 return; 1582 } 1583 1584 // Resize the tenured generation 1585 // after obtaining the free list locks for the 1586 // two generations. 1587 void CMSCollector::compute_new_size() { 1588 assert_locked_or_safepoint(Heap_lock); 1589 FreelistLocker z(this); 1590 MetaspaceGC::compute_new_size(); 1591 _cmsGen->compute_new_size_free_list(); 1592 } 1593 1594 // A work method used by the foreground collector to do 1595 // a mark-sweep-compact. 1596 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1598 1599 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1600 gc_timer->register_gc_start(); 1601 1602 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1603 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1604 1605 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1606 1607 // Temporarily widen the span of the weak reference processing to 1608 // the entire heap. 1609 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1610 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1611 // Temporarily, clear the "is_alive_non_header" field of the 1612 // reference processor. 1613 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1614 // Temporarily make reference _processing_ single threaded (non-MT). 1615 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1616 // Temporarily make refs discovery atomic 1617 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1618 // Temporarily make reference _discovery_ single threaded (non-MT) 1619 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1620 1621 ref_processor()->set_enqueuing_is_done(false); 1622 ref_processor()->enable_discovery(); 1623 ref_processor()->setup_policy(clear_all_soft_refs); 1624 // If an asynchronous collection finishes, the _modUnionTable is 1625 // all clear. If we are assuming the collection from an asynchronous 1626 // collection, clear the _modUnionTable. 1627 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1628 "_modUnionTable should be clear if the baton was not passed"); 1629 _modUnionTable.clear_all(); 1630 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1631 "mod union for klasses should be clear if the baton was passed"); 1632 _ct->klass_rem_set()->clear_mod_union(); 1633 1634 // We must adjust the allocation statistics being maintained 1635 // in the free list space. We do so by reading and clearing 1636 // the sweep timer and updating the block flux rate estimates below. 1637 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1638 if (_inter_sweep_timer.is_active()) { 1639 _inter_sweep_timer.stop(); 1640 // Note that we do not use this sample to update the _inter_sweep_estimate. 1641 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1642 _inter_sweep_estimate.padded_average(), 1643 _intra_sweep_estimate.padded_average()); 1644 } 1645 1646 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1647 ref_processor(), clear_all_soft_refs); 1648 #ifdef ASSERT 1649 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1650 size_t free_size = cms_space->free(); 1651 assert(free_size == 1652 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1653 * HeapWordSize, 1654 "All the free space should be compacted into one chunk at top"); 1655 assert(cms_space->dictionary()->total_chunk_size( 1656 debug_only(cms_space->freelistLock())) == 0 || 1657 cms_space->totalSizeInIndexedFreeLists() == 0, 1658 "All the free space should be in a single chunk"); 1659 size_t num = cms_space->totalCount(); 1660 assert((free_size == 0 && num == 0) || 1661 (free_size > 0 && (num == 1 || num == 2)), 1662 "There should be at most 2 free chunks after compaction"); 1663 #endif // ASSERT 1664 _collectorState = Resetting; 1665 assert(_restart_addr == NULL, 1666 "Should have been NULL'd before baton was passed"); 1667 reset(false /* == !concurrent */); 1668 _cmsGen->reset_after_compaction(); 1669 _concurrent_cycles_since_last_unload = 0; 1670 1671 // Clear any data recorded in the PLAB chunk arrays. 1672 if (_survivor_plab_array != NULL) { 1673 reset_survivor_plab_arrays(); 1674 } 1675 1676 // Adjust the per-size allocation stats for the next epoch. 1677 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1678 // Restart the "inter sweep timer" for the next epoch. 1679 _inter_sweep_timer.reset(); 1680 _inter_sweep_timer.start(); 1681 1682 gc_timer->register_gc_end(); 1683 1684 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1685 1686 // For a mark-sweep-compact, compute_new_size() will be called 1687 // in the heap's do_collection() method. 1688 } 1689 1690 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1691 ContiguousSpace* eden_space = _young_gen->eden(); 1692 ContiguousSpace* from_space = _young_gen->from(); 1693 ContiguousSpace* to_space = _young_gen->to(); 1694 // Eden 1695 if (_eden_chunk_array != NULL) { 1696 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1697 p2i(eden_space->bottom()), p2i(eden_space->top()), 1698 p2i(eden_space->end()), eden_space->capacity()); 1699 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1700 "_eden_chunk_capacity=" SIZE_FORMAT, 1701 _eden_chunk_index, _eden_chunk_capacity); 1702 for (size_t i = 0; i < _eden_chunk_index; i++) { 1703 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1704 i, p2i(_eden_chunk_array[i])); 1705 } 1706 } 1707 // Survivor 1708 if (_survivor_chunk_array != NULL) { 1709 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1710 p2i(from_space->bottom()), p2i(from_space->top()), 1711 p2i(from_space->end()), from_space->capacity()); 1712 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1713 "_survivor_chunk_capacity=" SIZE_FORMAT, 1714 _survivor_chunk_index, _survivor_chunk_capacity); 1715 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1716 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1717 i, p2i(_survivor_chunk_array[i])); 1718 } 1719 } 1720 } 1721 1722 void CMSCollector::getFreelistLocks() const { 1723 // Get locks for all free lists in all generations that this 1724 // collector is responsible for 1725 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1726 } 1727 1728 void CMSCollector::releaseFreelistLocks() const { 1729 // Release locks for all free lists in all generations that this 1730 // collector is responsible for 1731 _cmsGen->freelistLock()->unlock(); 1732 } 1733 1734 bool CMSCollector::haveFreelistLocks() const { 1735 // Check locks for all free lists in all generations that this 1736 // collector is responsible for 1737 assert_lock_strong(_cmsGen->freelistLock()); 1738 PRODUCT_ONLY(ShouldNotReachHere()); 1739 return true; 1740 } 1741 1742 // A utility class that is used by the CMS collector to 1743 // temporarily "release" the foreground collector from its 1744 // usual obligation to wait for the background collector to 1745 // complete an ongoing phase before proceeding. 1746 class ReleaseForegroundGC: public StackObj { 1747 private: 1748 CMSCollector* _c; 1749 public: 1750 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1751 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1752 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1753 // allow a potentially blocked foreground collector to proceed 1754 _c->_foregroundGCShouldWait = false; 1755 if (_c->_foregroundGCIsActive) { 1756 CGC_lock->notify(); 1757 } 1758 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1759 "Possible deadlock"); 1760 } 1761 1762 ~ReleaseForegroundGC() { 1763 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1764 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1765 _c->_foregroundGCShouldWait = true; 1766 } 1767 }; 1768 1769 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1770 assert(Thread::current()->is_ConcurrentGC_thread(), 1771 "A CMS asynchronous collection is only allowed on a CMS thread."); 1772 1773 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1774 { 1775 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1776 MutexLockerEx hl(Heap_lock, safepoint_check); 1777 FreelistLocker fll(this); 1778 MutexLockerEx x(CGC_lock, safepoint_check); 1779 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1780 // The foreground collector is active or we're 1781 // not using asynchronous collections. Skip this 1782 // background collection. 1783 assert(!_foregroundGCShouldWait, "Should be clear"); 1784 return; 1785 } else { 1786 assert(_collectorState == Idling, "Should be idling before start."); 1787 _collectorState = InitialMarking; 1788 register_gc_start(cause); 1789 // Reset the expansion cause, now that we are about to begin 1790 // a new cycle. 1791 clear_expansion_cause(); 1792 1793 // Clear the MetaspaceGC flag since a concurrent collection 1794 // is starting but also clear it after the collection. 1795 MetaspaceGC::set_should_concurrent_collect(false); 1796 } 1797 // Decide if we want to enable class unloading as part of the 1798 // ensuing concurrent GC cycle. 1799 update_should_unload_classes(); 1800 _full_gc_requested = false; // acks all outstanding full gc requests 1801 _full_gc_cause = GCCause::_no_gc; 1802 // Signal that we are about to start a collection 1803 gch->increment_total_full_collections(); // ... starting a collection cycle 1804 _collection_count_start = gch->total_full_collections(); 1805 } 1806 1807 // Used for PrintGC 1808 size_t prev_used; 1809 if (PrintGC && Verbose) { 1810 prev_used = _cmsGen->used(); 1811 } 1812 1813 // The change of the collection state is normally done at this level; 1814 // the exceptions are phases that are executed while the world is 1815 // stopped. For those phases the change of state is done while the 1816 // world is stopped. For baton passing purposes this allows the 1817 // background collector to finish the phase and change state atomically. 1818 // The foreground collector cannot wait on a phase that is done 1819 // while the world is stopped because the foreground collector already 1820 // has the world stopped and would deadlock. 1821 while (_collectorState != Idling) { 1822 if (TraceCMSState) { 1823 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1824 p2i(Thread::current()), _collectorState); 1825 } 1826 // The foreground collector 1827 // holds the Heap_lock throughout its collection. 1828 // holds the CMS token (but not the lock) 1829 // except while it is waiting for the background collector to yield. 1830 // 1831 // The foreground collector should be blocked (not for long) 1832 // if the background collector is about to start a phase 1833 // executed with world stopped. If the background 1834 // collector has already started such a phase, the 1835 // foreground collector is blocked waiting for the 1836 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1837 // are executed in the VM thread. 1838 // 1839 // The locking order is 1840 // PendingListLock (PLL) -- if applicable (FinalMarking) 1841 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1842 // CMS token (claimed in 1843 // stop_world_and_do() --> 1844 // safepoint_synchronize() --> 1845 // CMSThread::synchronize()) 1846 1847 { 1848 // Check if the FG collector wants us to yield. 1849 CMSTokenSync x(true); // is cms thread 1850 if (waitForForegroundGC()) { 1851 // We yielded to a foreground GC, nothing more to be 1852 // done this round. 1853 assert(_foregroundGCShouldWait == false, "We set it to false in " 1854 "waitForForegroundGC()"); 1855 if (TraceCMSState) { 1856 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1857 " exiting collection CMS state %d", 1858 p2i(Thread::current()), _collectorState); 1859 } 1860 return; 1861 } else { 1862 // The background collector can run but check to see if the 1863 // foreground collector has done a collection while the 1864 // background collector was waiting to get the CGC_lock 1865 // above. If yes, break so that _foregroundGCShouldWait 1866 // is cleared before returning. 1867 if (_collectorState == Idling) { 1868 break; 1869 } 1870 } 1871 } 1872 1873 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1874 "should be waiting"); 1875 1876 switch (_collectorState) { 1877 case InitialMarking: 1878 { 1879 ReleaseForegroundGC x(this); 1880 stats().record_cms_begin(); 1881 VM_CMS_Initial_Mark initial_mark_op(this); 1882 VMThread::execute(&initial_mark_op); 1883 } 1884 // The collector state may be any legal state at this point 1885 // since the background collector may have yielded to the 1886 // foreground collector. 1887 break; 1888 case Marking: 1889 // initial marking in checkpointRootsInitialWork has been completed 1890 if (markFromRoots()) { // we were successful 1891 assert(_collectorState == Precleaning, "Collector state should " 1892 "have changed"); 1893 } else { 1894 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1895 } 1896 break; 1897 case Precleaning: 1898 // marking from roots in markFromRoots has been completed 1899 preclean(); 1900 assert(_collectorState == AbortablePreclean || 1901 _collectorState == FinalMarking, 1902 "Collector state should have changed"); 1903 break; 1904 case AbortablePreclean: 1905 abortable_preclean(); 1906 assert(_collectorState == FinalMarking, "Collector state should " 1907 "have changed"); 1908 break; 1909 case FinalMarking: 1910 { 1911 ReleaseForegroundGC x(this); 1912 1913 VM_CMS_Final_Remark final_remark_op(this); 1914 VMThread::execute(&final_remark_op); 1915 } 1916 assert(_foregroundGCShouldWait, "block post-condition"); 1917 break; 1918 case Sweeping: 1919 // final marking in checkpointRootsFinal has been completed 1920 sweep(); 1921 assert(_collectorState == Resizing, "Collector state change " 1922 "to Resizing must be done under the free_list_lock"); 1923 1924 case Resizing: { 1925 // Sweeping has been completed... 1926 // At this point the background collection has completed. 1927 // Don't move the call to compute_new_size() down 1928 // into code that might be executed if the background 1929 // collection was preempted. 1930 { 1931 ReleaseForegroundGC x(this); // unblock FG collection 1932 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1933 CMSTokenSync z(true); // not strictly needed. 1934 if (_collectorState == Resizing) { 1935 compute_new_size(); 1936 save_heap_summary(); 1937 _collectorState = Resetting; 1938 } else { 1939 assert(_collectorState == Idling, "The state should only change" 1940 " because the foreground collector has finished the collection"); 1941 } 1942 } 1943 break; 1944 } 1945 case Resetting: 1946 // CMS heap resizing has been completed 1947 reset(true); 1948 assert(_collectorState == Idling, "Collector state should " 1949 "have changed"); 1950 1951 MetaspaceGC::set_should_concurrent_collect(false); 1952 1953 stats().record_cms_end(); 1954 // Don't move the concurrent_phases_end() and compute_new_size() 1955 // calls to here because a preempted background collection 1956 // has it's state set to "Resetting". 1957 break; 1958 case Idling: 1959 default: 1960 ShouldNotReachHere(); 1961 break; 1962 } 1963 if (TraceCMSState) { 1964 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1965 p2i(Thread::current()), _collectorState); 1966 } 1967 assert(_foregroundGCShouldWait, "block post-condition"); 1968 } 1969 1970 // Should this be in gc_epilogue? 1971 collector_policy()->counters()->update_counters(); 1972 1973 { 1974 // Clear _foregroundGCShouldWait and, in the event that the 1975 // foreground collector is waiting, notify it, before 1976 // returning. 1977 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1978 _foregroundGCShouldWait = false; 1979 if (_foregroundGCIsActive) { 1980 CGC_lock->notify(); 1981 } 1982 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1983 "Possible deadlock"); 1984 } 1985 if (TraceCMSState) { 1986 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1987 " exiting collection CMS state %d", 1988 p2i(Thread::current()), _collectorState); 1989 } 1990 if (PrintGC && Verbose) { 1991 _cmsGen->print_heap_change(prev_used); 1992 } 1993 } 1994 1995 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1996 _cms_start_registered = true; 1997 _gc_timer_cm->register_gc_start(); 1998 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1999 } 2000 2001 void CMSCollector::register_gc_end() { 2002 if (_cms_start_registered) { 2003 report_heap_summary(GCWhen::AfterGC); 2004 2005 _gc_timer_cm->register_gc_end(); 2006 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2007 _cms_start_registered = false; 2008 } 2009 } 2010 2011 void CMSCollector::save_heap_summary() { 2012 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2013 _last_heap_summary = gch->create_heap_summary(); 2014 _last_metaspace_summary = gch->create_metaspace_summary(); 2015 } 2016 2017 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2018 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2019 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2020 } 2021 2022 bool CMSCollector::waitForForegroundGC() { 2023 bool res = false; 2024 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2025 "CMS thread should have CMS token"); 2026 // Block the foreground collector until the 2027 // background collectors decides whether to 2028 // yield. 2029 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2030 _foregroundGCShouldWait = true; 2031 if (_foregroundGCIsActive) { 2032 // The background collector yields to the 2033 // foreground collector and returns a value 2034 // indicating that it has yielded. The foreground 2035 // collector can proceed. 2036 res = true; 2037 _foregroundGCShouldWait = false; 2038 ConcurrentMarkSweepThread::clear_CMS_flag( 2039 ConcurrentMarkSweepThread::CMS_cms_has_token); 2040 ConcurrentMarkSweepThread::set_CMS_flag( 2041 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2042 // Get a possibly blocked foreground thread going 2043 CGC_lock->notify(); 2044 if (TraceCMSState) { 2045 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2046 p2i(Thread::current()), _collectorState); 2047 } 2048 while (_foregroundGCIsActive) { 2049 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2050 } 2051 ConcurrentMarkSweepThread::set_CMS_flag( 2052 ConcurrentMarkSweepThread::CMS_cms_has_token); 2053 ConcurrentMarkSweepThread::clear_CMS_flag( 2054 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2055 } 2056 if (TraceCMSState) { 2057 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2058 p2i(Thread::current()), _collectorState); 2059 } 2060 return res; 2061 } 2062 2063 // Because of the need to lock the free lists and other structures in 2064 // the collector, common to all the generations that the collector is 2065 // collecting, we need the gc_prologues of individual CMS generations 2066 // delegate to their collector. It may have been simpler had the 2067 // current infrastructure allowed one to call a prologue on a 2068 // collector. In the absence of that we have the generation's 2069 // prologue delegate to the collector, which delegates back 2070 // some "local" work to a worker method in the individual generations 2071 // that it's responsible for collecting, while itself doing any 2072 // work common to all generations it's responsible for. A similar 2073 // comment applies to the gc_epilogue()'s. 2074 // The role of the variable _between_prologue_and_epilogue is to 2075 // enforce the invocation protocol. 2076 void CMSCollector::gc_prologue(bool full) { 2077 // Call gc_prologue_work() for the CMSGen 2078 // we are responsible for. 2079 2080 // The following locking discipline assumes that we are only called 2081 // when the world is stopped. 2082 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2083 2084 // The CMSCollector prologue must call the gc_prologues for the 2085 // "generations" that it's responsible 2086 // for. 2087 2088 assert( Thread::current()->is_VM_thread() 2089 || ( CMSScavengeBeforeRemark 2090 && Thread::current()->is_ConcurrentGC_thread()), 2091 "Incorrect thread type for prologue execution"); 2092 2093 if (_between_prologue_and_epilogue) { 2094 // We have already been invoked; this is a gc_prologue delegation 2095 // from yet another CMS generation that we are responsible for, just 2096 // ignore it since all relevant work has already been done. 2097 return; 2098 } 2099 2100 // set a bit saying prologue has been called; cleared in epilogue 2101 _between_prologue_and_epilogue = true; 2102 // Claim locks for common data structures, then call gc_prologue_work() 2103 // for each CMSGen. 2104 2105 getFreelistLocks(); // gets free list locks on constituent spaces 2106 bitMapLock()->lock_without_safepoint_check(); 2107 2108 // Should call gc_prologue_work() for all cms gens we are responsible for 2109 bool duringMarking = _collectorState >= Marking 2110 && _collectorState < Sweeping; 2111 2112 // The young collections clear the modified oops state, which tells if 2113 // there are any modified oops in the class. The remark phase also needs 2114 // that information. Tell the young collection to save the union of all 2115 // modified klasses. 2116 if (duringMarking) { 2117 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2118 } 2119 2120 bool registerClosure = duringMarking; 2121 2122 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2123 2124 if (!full) { 2125 stats().record_gc0_begin(); 2126 } 2127 } 2128 2129 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2130 2131 _capacity_at_prologue = capacity(); 2132 _used_at_prologue = used(); 2133 2134 // Delegate to CMScollector which knows how to coordinate between 2135 // this and any other CMS generations that it is responsible for 2136 // collecting. 2137 collector()->gc_prologue(full); 2138 } 2139 2140 // This is a "private" interface for use by this generation's CMSCollector. 2141 // Not to be called directly by any other entity (for instance, 2142 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2143 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2144 bool registerClosure, ModUnionClosure* modUnionClosure) { 2145 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2146 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2147 "Should be NULL"); 2148 if (registerClosure) { 2149 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2150 } 2151 cmsSpace()->gc_prologue(); 2152 // Clear stat counters 2153 NOT_PRODUCT( 2154 assert(_numObjectsPromoted == 0, "check"); 2155 assert(_numWordsPromoted == 0, "check"); 2156 if (Verbose && PrintGC) { 2157 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2158 SIZE_FORMAT" bytes concurrently", 2159 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2160 } 2161 _numObjectsAllocated = 0; 2162 _numWordsAllocated = 0; 2163 ) 2164 } 2165 2166 void CMSCollector::gc_epilogue(bool full) { 2167 // The following locking discipline assumes that we are only called 2168 // when the world is stopped. 2169 assert(SafepointSynchronize::is_at_safepoint(), 2170 "world is stopped assumption"); 2171 2172 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2173 // if linear allocation blocks need to be appropriately marked to allow the 2174 // the blocks to be parsable. We also check here whether we need to nudge the 2175 // CMS collector thread to start a new cycle (if it's not already active). 2176 assert( Thread::current()->is_VM_thread() 2177 || ( CMSScavengeBeforeRemark 2178 && Thread::current()->is_ConcurrentGC_thread()), 2179 "Incorrect thread type for epilogue execution"); 2180 2181 if (!_between_prologue_and_epilogue) { 2182 // We have already been invoked; this is a gc_epilogue delegation 2183 // from yet another CMS generation that we are responsible for, just 2184 // ignore it since all relevant work has already been done. 2185 return; 2186 } 2187 assert(haveFreelistLocks(), "must have freelist locks"); 2188 assert_lock_strong(bitMapLock()); 2189 2190 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2191 2192 _cmsGen->gc_epilogue_work(full); 2193 2194 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2195 // in case sampling was not already enabled, enable it 2196 _start_sampling = true; 2197 } 2198 // reset _eden_chunk_array so sampling starts afresh 2199 _eden_chunk_index = 0; 2200 2201 size_t cms_used = _cmsGen->cmsSpace()->used(); 2202 2203 // update performance counters - this uses a special version of 2204 // update_counters() that allows the utilization to be passed as a 2205 // parameter, avoiding multiple calls to used(). 2206 // 2207 _cmsGen->update_counters(cms_used); 2208 2209 bitMapLock()->unlock(); 2210 releaseFreelistLocks(); 2211 2212 if (!CleanChunkPoolAsync) { 2213 Chunk::clean_chunk_pool(); 2214 } 2215 2216 set_did_compact(false); 2217 _between_prologue_and_epilogue = false; // ready for next cycle 2218 } 2219 2220 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2221 collector()->gc_epilogue(full); 2222 2223 // Also reset promotion tracking in par gc thread states. 2224 for (uint i = 0; i < ParallelGCThreads; i++) { 2225 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2226 } 2227 } 2228 2229 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2230 assert(!incremental_collection_failed(), "Should have been cleared"); 2231 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2232 cmsSpace()->gc_epilogue(); 2233 // Print stat counters 2234 NOT_PRODUCT( 2235 assert(_numObjectsAllocated == 0, "check"); 2236 assert(_numWordsAllocated == 0, "check"); 2237 if (Verbose && PrintGC) { 2238 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2239 SIZE_FORMAT" bytes", 2240 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2241 } 2242 _numObjectsPromoted = 0; 2243 _numWordsPromoted = 0; 2244 ) 2245 2246 if (PrintGC && Verbose) { 2247 // Call down the chain in contiguous_available needs the freelistLock 2248 // so print this out before releasing the freeListLock. 2249 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2250 contiguous_available()); 2251 } 2252 } 2253 2254 #ifndef PRODUCT 2255 bool CMSCollector::have_cms_token() { 2256 Thread* thr = Thread::current(); 2257 if (thr->is_VM_thread()) { 2258 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2259 } else if (thr->is_ConcurrentGC_thread()) { 2260 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2261 } else if (thr->is_GC_task_thread()) { 2262 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2263 ParGCRareEvent_lock->owned_by_self(); 2264 } 2265 return false; 2266 } 2267 #endif 2268 2269 // Check reachability of the given heap address in CMS generation, 2270 // treating all other generations as roots. 2271 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2272 // We could "guarantee" below, rather than assert, but I'll 2273 // leave these as "asserts" so that an adventurous debugger 2274 // could try this in the product build provided some subset of 2275 // the conditions were met, provided they were interested in the 2276 // results and knew that the computation below wouldn't interfere 2277 // with other concurrent computations mutating the structures 2278 // being read or written. 2279 assert(SafepointSynchronize::is_at_safepoint(), 2280 "Else mutations in object graph will make answer suspect"); 2281 assert(have_cms_token(), "Should hold cms token"); 2282 assert(haveFreelistLocks(), "must hold free list locks"); 2283 assert_lock_strong(bitMapLock()); 2284 2285 // Clear the marking bit map array before starting, but, just 2286 // for kicks, first report if the given address is already marked 2287 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2288 _markBitMap.isMarked(addr) ? "" : " not"); 2289 2290 if (verify_after_remark()) { 2291 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2292 bool result = verification_mark_bm()->isMarked(addr); 2293 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2294 result ? "IS" : "is NOT"); 2295 return result; 2296 } else { 2297 gclog_or_tty->print_cr("Could not compute result"); 2298 return false; 2299 } 2300 } 2301 2302 2303 void 2304 CMSCollector::print_on_error(outputStream* st) { 2305 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2306 if (collector != NULL) { 2307 CMSBitMap* bitmap = &collector->_markBitMap; 2308 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2309 bitmap->print_on_error(st, " Bits: "); 2310 2311 st->cr(); 2312 2313 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2314 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2315 mut_bitmap->print_on_error(st, " Bits: "); 2316 } 2317 } 2318 2319 //////////////////////////////////////////////////////// 2320 // CMS Verification Support 2321 //////////////////////////////////////////////////////// 2322 // Following the remark phase, the following invariant 2323 // should hold -- each object in the CMS heap which is 2324 // marked in markBitMap() should be marked in the verification_mark_bm(). 2325 2326 class VerifyMarkedClosure: public BitMapClosure { 2327 CMSBitMap* _marks; 2328 bool _failed; 2329 2330 public: 2331 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2332 2333 bool do_bit(size_t offset) { 2334 HeapWord* addr = _marks->offsetToHeapWord(offset); 2335 if (!_marks->isMarked(addr)) { 2336 oop(addr)->print_on(gclog_or_tty); 2337 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2338 _failed = true; 2339 } 2340 return true; 2341 } 2342 2343 bool failed() { return _failed; } 2344 }; 2345 2346 bool CMSCollector::verify_after_remark(bool silent) { 2347 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2348 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2349 static bool init = false; 2350 2351 assert(SafepointSynchronize::is_at_safepoint(), 2352 "Else mutations in object graph will make answer suspect"); 2353 assert(have_cms_token(), 2354 "Else there may be mutual interference in use of " 2355 " verification data structures"); 2356 assert(_collectorState > Marking && _collectorState <= Sweeping, 2357 "Else marking info checked here may be obsolete"); 2358 assert(haveFreelistLocks(), "must hold free list locks"); 2359 assert_lock_strong(bitMapLock()); 2360 2361 2362 // Allocate marking bit map if not already allocated 2363 if (!init) { // first time 2364 if (!verification_mark_bm()->allocate(_span)) { 2365 return false; 2366 } 2367 init = true; 2368 } 2369 2370 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2371 2372 // Turn off refs discovery -- so we will be tracing through refs. 2373 // This is as intended, because by this time 2374 // GC must already have cleared any refs that need to be cleared, 2375 // and traced those that need to be marked; moreover, 2376 // the marking done here is not going to interfere in any 2377 // way with the marking information used by GC. 2378 NoRefDiscovery no_discovery(ref_processor()); 2379 2380 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2381 2382 // Clear any marks from a previous round 2383 verification_mark_bm()->clear_all(); 2384 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2385 verify_work_stacks_empty(); 2386 2387 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2388 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2389 // Update the saved marks which may affect the root scans. 2390 gch->save_marks(); 2391 2392 if (CMSRemarkVerifyVariant == 1) { 2393 // In this first variant of verification, we complete 2394 // all marking, then check if the new marks-vector is 2395 // a subset of the CMS marks-vector. 2396 verify_after_remark_work_1(); 2397 } else if (CMSRemarkVerifyVariant == 2) { 2398 // In this second variant of verification, we flag an error 2399 // (i.e. an object reachable in the new marks-vector not reachable 2400 // in the CMS marks-vector) immediately, also indicating the 2401 // identify of an object (A) that references the unmarked object (B) -- 2402 // presumably, a mutation to A failed to be picked up by preclean/remark? 2403 verify_after_remark_work_2(); 2404 } else { 2405 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2406 CMSRemarkVerifyVariant); 2407 } 2408 if (!silent) gclog_or_tty->print(" done] "); 2409 return true; 2410 } 2411 2412 void CMSCollector::verify_after_remark_work_1() { 2413 ResourceMark rm; 2414 HandleMark hm; 2415 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2416 2417 // Get a clear set of claim bits for the roots processing to work with. 2418 ClassLoaderDataGraph::clear_claimed_marks(); 2419 2420 // Mark from roots one level into CMS 2421 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2422 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2423 2424 gch->gen_process_roots(_cmsGen->level(), 2425 true, // younger gens are roots 2426 true, // activate StrongRootsScope 2427 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2428 should_unload_classes(), 2429 ¬Older, 2430 NULL, 2431 NULL); // SSS: Provide correct closure 2432 2433 // Now mark from the roots 2434 MarkFromRootsClosure markFromRootsClosure(this, _span, 2435 verification_mark_bm(), verification_mark_stack(), 2436 false /* don't yield */, true /* verifying */); 2437 assert(_restart_addr == NULL, "Expected pre-condition"); 2438 verification_mark_bm()->iterate(&markFromRootsClosure); 2439 while (_restart_addr != NULL) { 2440 // Deal with stack overflow: by restarting at the indicated 2441 // address. 2442 HeapWord* ra = _restart_addr; 2443 markFromRootsClosure.reset(ra); 2444 _restart_addr = NULL; 2445 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2446 } 2447 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2448 verify_work_stacks_empty(); 2449 2450 // Marking completed -- now verify that each bit marked in 2451 // verification_mark_bm() is also marked in markBitMap(); flag all 2452 // errors by printing corresponding objects. 2453 VerifyMarkedClosure vcl(markBitMap()); 2454 verification_mark_bm()->iterate(&vcl); 2455 if (vcl.failed()) { 2456 gclog_or_tty->print("Verification failed"); 2457 gch->print_on(gclog_or_tty); 2458 fatal("CMS: failed marking verification after remark"); 2459 } 2460 } 2461 2462 class VerifyKlassOopsKlassClosure : public KlassClosure { 2463 class VerifyKlassOopsClosure : public OopClosure { 2464 CMSBitMap* _bitmap; 2465 public: 2466 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2467 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2468 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2469 } _oop_closure; 2470 public: 2471 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2472 void do_klass(Klass* k) { 2473 k->oops_do(&_oop_closure); 2474 } 2475 }; 2476 2477 void CMSCollector::verify_after_remark_work_2() { 2478 ResourceMark rm; 2479 HandleMark hm; 2480 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2481 2482 // Get a clear set of claim bits for the roots processing to work with. 2483 ClassLoaderDataGraph::clear_claimed_marks(); 2484 2485 // Mark from roots one level into CMS 2486 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2487 markBitMap()); 2488 CLDToOopClosure cld_closure(¬Older, true); 2489 2490 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2491 2492 gch->gen_process_roots(_cmsGen->level(), 2493 true, // younger gens are roots 2494 true, // activate StrongRootsScope 2495 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2496 should_unload_classes(), 2497 ¬Older, 2498 NULL, 2499 &cld_closure); 2500 2501 // Now mark from the roots 2502 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2503 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2504 assert(_restart_addr == NULL, "Expected pre-condition"); 2505 verification_mark_bm()->iterate(&markFromRootsClosure); 2506 while (_restart_addr != NULL) { 2507 // Deal with stack overflow: by restarting at the indicated 2508 // address. 2509 HeapWord* ra = _restart_addr; 2510 markFromRootsClosure.reset(ra); 2511 _restart_addr = NULL; 2512 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2513 } 2514 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2515 verify_work_stacks_empty(); 2516 2517 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2518 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2519 2520 // Marking completed -- now verify that each bit marked in 2521 // verification_mark_bm() is also marked in markBitMap(); flag all 2522 // errors by printing corresponding objects. 2523 VerifyMarkedClosure vcl(markBitMap()); 2524 verification_mark_bm()->iterate(&vcl); 2525 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2526 } 2527 2528 void ConcurrentMarkSweepGeneration::save_marks() { 2529 // delegate to CMS space 2530 cmsSpace()->save_marks(); 2531 for (uint i = 0; i < ParallelGCThreads; i++) { 2532 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2533 } 2534 } 2535 2536 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2537 return cmsSpace()->no_allocs_since_save_marks(); 2538 } 2539 2540 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2541 \ 2542 void ConcurrentMarkSweepGeneration:: \ 2543 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2544 cl->set_generation(this); \ 2545 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2546 cl->reset_generation(); \ 2547 save_marks(); \ 2548 } 2549 2550 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2551 2552 void 2553 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2554 if (freelistLock()->owned_by_self()) { 2555 Generation::oop_iterate(cl); 2556 } else { 2557 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2558 Generation::oop_iterate(cl); 2559 } 2560 } 2561 2562 void 2563 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2564 if (freelistLock()->owned_by_self()) { 2565 Generation::object_iterate(cl); 2566 } else { 2567 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2568 Generation::object_iterate(cl); 2569 } 2570 } 2571 2572 void 2573 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2574 if (freelistLock()->owned_by_self()) { 2575 Generation::safe_object_iterate(cl); 2576 } else { 2577 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2578 Generation::safe_object_iterate(cl); 2579 } 2580 } 2581 2582 void 2583 ConcurrentMarkSweepGeneration::post_compact() { 2584 } 2585 2586 void 2587 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2588 // Fix the linear allocation blocks to look like free blocks. 2589 2590 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2591 // are not called when the heap is verified during universe initialization and 2592 // at vm shutdown. 2593 if (freelistLock()->owned_by_self()) { 2594 cmsSpace()->prepare_for_verify(); 2595 } else { 2596 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2597 cmsSpace()->prepare_for_verify(); 2598 } 2599 } 2600 2601 void 2602 ConcurrentMarkSweepGeneration::verify() { 2603 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2604 // are not called when the heap is verified during universe initialization and 2605 // at vm shutdown. 2606 if (freelistLock()->owned_by_self()) { 2607 cmsSpace()->verify(); 2608 } else { 2609 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2610 cmsSpace()->verify(); 2611 } 2612 } 2613 2614 void CMSCollector::verify() { 2615 _cmsGen->verify(); 2616 } 2617 2618 #ifndef PRODUCT 2619 bool CMSCollector::overflow_list_is_empty() const { 2620 assert(_num_par_pushes >= 0, "Inconsistency"); 2621 if (_overflow_list == NULL) { 2622 assert(_num_par_pushes == 0, "Inconsistency"); 2623 } 2624 return _overflow_list == NULL; 2625 } 2626 2627 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2628 // merely consolidate assertion checks that appear to occur together frequently. 2629 void CMSCollector::verify_work_stacks_empty() const { 2630 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2631 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2632 } 2633 2634 void CMSCollector::verify_overflow_empty() const { 2635 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2636 assert(no_preserved_marks(), "No preserved marks"); 2637 } 2638 #endif // PRODUCT 2639 2640 // Decide if we want to enable class unloading as part of the 2641 // ensuing concurrent GC cycle. We will collect and 2642 // unload classes if it's the case that: 2643 // (1) an explicit gc request has been made and the flag 2644 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2645 // (2) (a) class unloading is enabled at the command line, and 2646 // (b) old gen is getting really full 2647 // NOTE: Provided there is no change in the state of the heap between 2648 // calls to this method, it should have idempotent results. Moreover, 2649 // its results should be monotonically increasing (i.e. going from 0 to 1, 2650 // but not 1 to 0) between successive calls between which the heap was 2651 // not collected. For the implementation below, it must thus rely on 2652 // the property that concurrent_cycles_since_last_unload() 2653 // will not decrease unless a collection cycle happened and that 2654 // _cmsGen->is_too_full() are 2655 // themselves also monotonic in that sense. See check_monotonicity() 2656 // below. 2657 void CMSCollector::update_should_unload_classes() { 2658 _should_unload_classes = false; 2659 // Condition 1 above 2660 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2661 _should_unload_classes = true; 2662 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2663 // Disjuncts 2.b.(i,ii,iii) above 2664 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2665 CMSClassUnloadingMaxInterval) 2666 || _cmsGen->is_too_full(); 2667 } 2668 } 2669 2670 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2671 bool res = should_concurrent_collect(); 2672 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2673 return res; 2674 } 2675 2676 void CMSCollector::setup_cms_unloading_and_verification_state() { 2677 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2678 || VerifyBeforeExit; 2679 const int rso = GenCollectedHeap::SO_AllCodeCache; 2680 2681 // We set the proper root for this CMS cycle here. 2682 if (should_unload_classes()) { // Should unload classes this cycle 2683 remove_root_scanning_option(rso); // Shrink the root set appropriately 2684 set_verifying(should_verify); // Set verification state for this cycle 2685 return; // Nothing else needs to be done at this time 2686 } 2687 2688 // Not unloading classes this cycle 2689 assert(!should_unload_classes(), "Inconsistency!"); 2690 2691 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2692 // Include symbols, strings and code cache elements to prevent their resurrection. 2693 add_root_scanning_option(rso); 2694 set_verifying(true); 2695 } else if (verifying() && !should_verify) { 2696 // We were verifying, but some verification flags got disabled. 2697 set_verifying(false); 2698 // Exclude symbols, strings and code cache elements from root scanning to 2699 // reduce IM and RM pauses. 2700 remove_root_scanning_option(rso); 2701 } 2702 } 2703 2704 2705 #ifndef PRODUCT 2706 HeapWord* CMSCollector::block_start(const void* p) const { 2707 const HeapWord* addr = (HeapWord*)p; 2708 if (_span.contains(p)) { 2709 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2710 return _cmsGen->cmsSpace()->block_start(p); 2711 } 2712 } 2713 return NULL; 2714 } 2715 #endif 2716 2717 HeapWord* 2718 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2719 bool tlab, 2720 bool parallel) { 2721 CMSSynchronousYieldRequest yr; 2722 assert(!tlab, "Can't deal with TLAB allocation"); 2723 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2724 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2725 if (GCExpandToAllocateDelayMillis > 0) { 2726 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2727 } 2728 return have_lock_and_allocate(word_size, tlab); 2729 } 2730 2731 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2732 size_t bytes, 2733 size_t expand_bytes, 2734 CMSExpansionCause::Cause cause) 2735 { 2736 2737 bool success = expand(bytes, expand_bytes); 2738 2739 // remember why we expanded; this information is used 2740 // by shouldConcurrentCollect() when making decisions on whether to start 2741 // a new CMS cycle. 2742 if (success) { 2743 set_expansion_cause(cause); 2744 if (PrintGCDetails && Verbose) { 2745 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2746 CMSExpansionCause::to_string(cause)); 2747 } 2748 } 2749 } 2750 2751 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2752 HeapWord* res = NULL; 2753 MutexLocker x(ParGCRareEvent_lock); 2754 while (true) { 2755 // Expansion by some other thread might make alloc OK now: 2756 res = ps->lab.alloc(word_sz); 2757 if (res != NULL) return res; 2758 // If there's not enough expansion space available, give up. 2759 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2760 return NULL; 2761 } 2762 // Otherwise, we try expansion. 2763 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2764 // Now go around the loop and try alloc again; 2765 // A competing par_promote might beat us to the expansion space, 2766 // so we may go around the loop again if promotion fails again. 2767 if (GCExpandToAllocateDelayMillis > 0) { 2768 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2769 } 2770 } 2771 } 2772 2773 2774 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2775 PromotionInfo* promo) { 2776 MutexLocker x(ParGCRareEvent_lock); 2777 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2778 while (true) { 2779 // Expansion by some other thread might make alloc OK now: 2780 if (promo->ensure_spooling_space()) { 2781 assert(promo->has_spooling_space(), 2782 "Post-condition of successful ensure_spooling_space()"); 2783 return true; 2784 } 2785 // If there's not enough expansion space available, give up. 2786 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2787 return false; 2788 } 2789 // Otherwise, we try expansion. 2790 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2791 // Now go around the loop and try alloc again; 2792 // A competing allocation might beat us to the expansion space, 2793 // so we may go around the loop again if allocation fails again. 2794 if (GCExpandToAllocateDelayMillis > 0) { 2795 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2796 } 2797 } 2798 } 2799 2800 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2801 // Only shrink if a compaction was done so that all the free space 2802 // in the generation is in a contiguous block at the end. 2803 if (did_compact()) { 2804 CardGeneration::shrink(bytes); 2805 } 2806 } 2807 2808 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2809 assert_locked_or_safepoint(Heap_lock); 2810 } 2811 2812 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2813 assert_locked_or_safepoint(Heap_lock); 2814 assert_lock_strong(freelistLock()); 2815 if (PrintGCDetails && Verbose) { 2816 warning("Shrinking of CMS not yet implemented"); 2817 } 2818 return; 2819 } 2820 2821 2822 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2823 // phases. 2824 class CMSPhaseAccounting: public StackObj { 2825 public: 2826 CMSPhaseAccounting(CMSCollector *collector, 2827 const char *phase, 2828 const GCId gc_id, 2829 bool print_cr = true); 2830 ~CMSPhaseAccounting(); 2831 2832 private: 2833 CMSCollector *_collector; 2834 const char *_phase; 2835 elapsedTimer _wallclock; 2836 bool _print_cr; 2837 const GCId _gc_id; 2838 2839 public: 2840 // Not MT-safe; so do not pass around these StackObj's 2841 // where they may be accessed by other threads. 2842 jlong wallclock_millis() { 2843 assert(_wallclock.is_active(), "Wall clock should not stop"); 2844 _wallclock.stop(); // to record time 2845 jlong ret = _wallclock.milliseconds(); 2846 _wallclock.start(); // restart 2847 return ret; 2848 } 2849 }; 2850 2851 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2852 const char *phase, 2853 const GCId gc_id, 2854 bool print_cr) : 2855 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2856 2857 if (PrintCMSStatistics != 0) { 2858 _collector->resetYields(); 2859 } 2860 if (PrintGCDetails) { 2861 gclog_or_tty->gclog_stamp(_gc_id); 2862 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2863 _collector->cmsGen()->short_name(), _phase); 2864 } 2865 _collector->resetTimer(); 2866 _wallclock.start(); 2867 _collector->startTimer(); 2868 } 2869 2870 CMSPhaseAccounting::~CMSPhaseAccounting() { 2871 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2872 _collector->stopTimer(); 2873 _wallclock.stop(); 2874 if (PrintGCDetails) { 2875 gclog_or_tty->gclog_stamp(_gc_id); 2876 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2877 _collector->cmsGen()->short_name(), 2878 _phase, _collector->timerValue(), _wallclock.seconds()); 2879 if (_print_cr) { 2880 gclog_or_tty->cr(); 2881 } 2882 if (PrintCMSStatistics != 0) { 2883 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2884 _collector->yields()); 2885 } 2886 } 2887 } 2888 2889 // CMS work 2890 2891 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2892 class CMSParMarkTask : public AbstractGangTask { 2893 protected: 2894 CMSCollector* _collector; 2895 uint _n_workers; 2896 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2897 AbstractGangTask(name), 2898 _collector(collector), 2899 _n_workers(n_workers) {} 2900 // Work method in support of parallel rescan ... of young gen spaces 2901 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2902 ContiguousSpace* space, 2903 HeapWord** chunk_array, size_t chunk_top); 2904 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2905 }; 2906 2907 // Parallel initial mark task 2908 class CMSParInitialMarkTask: public CMSParMarkTask { 2909 public: 2910 CMSParInitialMarkTask(CMSCollector* collector, uint n_workers) : 2911 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2912 collector, n_workers) {} 2913 void work(uint worker_id); 2914 }; 2915 2916 // Checkpoint the roots into this generation from outside 2917 // this generation. [Note this initial checkpoint need only 2918 // be approximate -- we'll do a catch up phase subsequently.] 2919 void CMSCollector::checkpointRootsInitial() { 2920 assert(_collectorState == InitialMarking, "Wrong collector state"); 2921 check_correct_thread_executing(); 2922 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2923 2924 save_heap_summary(); 2925 report_heap_summary(GCWhen::BeforeGC); 2926 2927 ReferenceProcessor* rp = ref_processor(); 2928 assert(_restart_addr == NULL, "Control point invariant"); 2929 { 2930 // acquire locks for subsequent manipulations 2931 MutexLockerEx x(bitMapLock(), 2932 Mutex::_no_safepoint_check_flag); 2933 checkpointRootsInitialWork(); 2934 // enable ("weak") refs discovery 2935 rp->enable_discovery(); 2936 _collectorState = Marking; 2937 } 2938 } 2939 2940 void CMSCollector::checkpointRootsInitialWork() { 2941 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2942 assert(_collectorState == InitialMarking, "just checking"); 2943 2944 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2945 // precede our marking with a collection of all 2946 // younger generations to keep floating garbage to a minimum. 2947 // XXX: we won't do this for now -- it's an optimization to be done later. 2948 2949 // already have locks 2950 assert_lock_strong(bitMapLock()); 2951 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2952 2953 // Setup the verification and class unloading state for this 2954 // CMS collection cycle. 2955 setup_cms_unloading_and_verification_state(); 2956 2957 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2958 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2959 2960 // Reset all the PLAB chunk arrays if necessary. 2961 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2962 reset_survivor_plab_arrays(); 2963 } 2964 2965 ResourceMark rm; 2966 HandleMark hm; 2967 2968 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2969 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2970 2971 verify_work_stacks_empty(); 2972 verify_overflow_empty(); 2973 2974 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2975 // Update the saved marks which may affect the root scans. 2976 gch->save_marks(); 2977 2978 // weak reference processing has not started yet. 2979 ref_processor()->set_enqueuing_is_done(false); 2980 2981 // Need to remember all newly created CLDs, 2982 // so that we can guarantee that the remark finds them. 2983 ClassLoaderDataGraph::remember_new_clds(true); 2984 2985 // Whenever a CLD is found, it will be claimed before proceeding to mark 2986 // the klasses. The claimed marks need to be cleared before marking starts. 2987 ClassLoaderDataGraph::clear_claimed_marks(); 2988 2989 if (CMSPrintEdenSurvivorChunks) { 2990 print_eden_and_survivor_chunk_arrays(); 2991 } 2992 2993 { 2994 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2995 if (CMSParallelInitialMarkEnabled) { 2996 // The parallel version. 2997 FlexibleWorkGang* workers = gch->workers(); 2998 assert(workers != NULL, "Need parallel worker threads."); 2999 uint n_workers = workers->active_workers(); 3000 CMSParInitialMarkTask tsk(this, n_workers); 3001 gch->set_par_threads(n_workers); 3002 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3003 if (n_workers > 1) { 3004 StrongRootsScope srs; 3005 workers->run_task(&tsk); 3006 } else { 3007 StrongRootsScope srs; 3008 tsk.work(0); 3009 } 3010 gch->set_par_threads(0); 3011 } else { 3012 // The serial version. 3013 CLDToOopClosure cld_closure(¬Older, true); 3014 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3015 gch->gen_process_roots(_cmsGen->level(), 3016 true, // younger gens are roots 3017 true, // activate StrongRootsScope 3018 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3019 should_unload_classes(), 3020 ¬Older, 3021 NULL, 3022 &cld_closure); 3023 } 3024 } 3025 3026 // Clear mod-union table; it will be dirtied in the prologue of 3027 // CMS generation per each younger generation collection. 3028 3029 assert(_modUnionTable.isAllClear(), 3030 "Was cleared in most recent final checkpoint phase" 3031 " or no bits are set in the gc_prologue before the start of the next " 3032 "subsequent marking phase."); 3033 3034 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3035 3036 // Save the end of the used_region of the constituent generations 3037 // to be used to limit the extent of sweep in each generation. 3038 save_sweep_limits(); 3039 verify_overflow_empty(); 3040 } 3041 3042 bool CMSCollector::markFromRoots() { 3043 // we might be tempted to assert that: 3044 // assert(!SafepointSynchronize::is_at_safepoint(), 3045 // "inconsistent argument?"); 3046 // However that wouldn't be right, because it's possible that 3047 // a safepoint is indeed in progress as a younger generation 3048 // stop-the-world GC happens even as we mark in this generation. 3049 assert(_collectorState == Marking, "inconsistent state?"); 3050 check_correct_thread_executing(); 3051 verify_overflow_empty(); 3052 3053 // Weak ref discovery note: We may be discovering weak 3054 // refs in this generation concurrent (but interleaved) with 3055 // weak ref discovery by a younger generation collector. 3056 3057 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3058 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3059 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3060 bool res = markFromRootsWork(); 3061 if (res) { 3062 _collectorState = Precleaning; 3063 } else { // We failed and a foreground collection wants to take over 3064 assert(_foregroundGCIsActive, "internal state inconsistency"); 3065 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3066 if (PrintGCDetails) { 3067 gclog_or_tty->print_cr("bailing out to foreground collection"); 3068 } 3069 } 3070 verify_overflow_empty(); 3071 return res; 3072 } 3073 3074 bool CMSCollector::markFromRootsWork() { 3075 // iterate over marked bits in bit map, doing a full scan and mark 3076 // from these roots using the following algorithm: 3077 // . if oop is to the right of the current scan pointer, 3078 // mark corresponding bit (we'll process it later) 3079 // . else (oop is to left of current scan pointer) 3080 // push oop on marking stack 3081 // . drain the marking stack 3082 3083 // Note that when we do a marking step we need to hold the 3084 // bit map lock -- recall that direct allocation (by mutators) 3085 // and promotion (by younger generation collectors) is also 3086 // marking the bit map. [the so-called allocate live policy.] 3087 // Because the implementation of bit map marking is not 3088 // robust wrt simultaneous marking of bits in the same word, 3089 // we need to make sure that there is no such interference 3090 // between concurrent such updates. 3091 3092 // already have locks 3093 assert_lock_strong(bitMapLock()); 3094 3095 verify_work_stacks_empty(); 3096 verify_overflow_empty(); 3097 bool result = false; 3098 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3099 result = do_marking_mt(); 3100 } else { 3101 result = do_marking_st(); 3102 } 3103 return result; 3104 } 3105 3106 // Forward decl 3107 class CMSConcMarkingTask; 3108 3109 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3110 CMSCollector* _collector; 3111 CMSConcMarkingTask* _task; 3112 public: 3113 virtual void yield(); 3114 3115 // "n_threads" is the number of threads to be terminated. 3116 // "queue_set" is a set of work queues of other threads. 3117 // "collector" is the CMS collector associated with this task terminator. 3118 // "yield" indicates whether we need the gang as a whole to yield. 3119 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3120 ParallelTaskTerminator(n_threads, queue_set), 3121 _collector(collector) { } 3122 3123 void set_task(CMSConcMarkingTask* task) { 3124 _task = task; 3125 } 3126 }; 3127 3128 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3129 CMSConcMarkingTask* _task; 3130 public: 3131 bool should_exit_termination(); 3132 void set_task(CMSConcMarkingTask* task) { 3133 _task = task; 3134 } 3135 }; 3136 3137 // MT Concurrent Marking Task 3138 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3139 CMSCollector* _collector; 3140 uint _n_workers; // requested/desired # workers 3141 bool _result; 3142 CompactibleFreeListSpace* _cms_space; 3143 char _pad_front[64]; // padding to ... 3144 HeapWord* _global_finger; // ... avoid sharing cache line 3145 char _pad_back[64]; 3146 HeapWord* _restart_addr; 3147 3148 // Exposed here for yielding support 3149 Mutex* const _bit_map_lock; 3150 3151 // The per thread work queues, available here for stealing 3152 OopTaskQueueSet* _task_queues; 3153 3154 // Termination (and yielding) support 3155 CMSConcMarkingTerminator _term; 3156 CMSConcMarkingTerminatorTerminator _term_term; 3157 3158 public: 3159 CMSConcMarkingTask(CMSCollector* collector, 3160 CompactibleFreeListSpace* cms_space, 3161 YieldingFlexibleWorkGang* workers, 3162 OopTaskQueueSet* task_queues): 3163 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3164 _collector(collector), 3165 _cms_space(cms_space), 3166 _n_workers(0), _result(true), 3167 _task_queues(task_queues), 3168 _term(_n_workers, task_queues, _collector), 3169 _bit_map_lock(collector->bitMapLock()) 3170 { 3171 _requested_size = _n_workers; 3172 _term.set_task(this); 3173 _term_term.set_task(this); 3174 _restart_addr = _global_finger = _cms_space->bottom(); 3175 } 3176 3177 3178 OopTaskQueueSet* task_queues() { return _task_queues; } 3179 3180 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3181 3182 HeapWord** global_finger_addr() { return &_global_finger; } 3183 3184 CMSConcMarkingTerminator* terminator() { return &_term; } 3185 3186 virtual void set_for_termination(uint active_workers) { 3187 terminator()->reset_for_reuse(active_workers); 3188 } 3189 3190 void work(uint worker_id); 3191 bool should_yield() { 3192 return ConcurrentMarkSweepThread::should_yield() 3193 && !_collector->foregroundGCIsActive(); 3194 } 3195 3196 virtual void coordinator_yield(); // stuff done by coordinator 3197 bool result() { return _result; } 3198 3199 void reset(HeapWord* ra) { 3200 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3201 _restart_addr = _global_finger = ra; 3202 _term.reset_for_reuse(); 3203 } 3204 3205 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3206 OopTaskQueue* work_q); 3207 3208 private: 3209 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3210 void do_work_steal(int i); 3211 void bump_global_finger(HeapWord* f); 3212 }; 3213 3214 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3215 assert(_task != NULL, "Error"); 3216 return _task->yielding(); 3217 // Note that we do not need the disjunct || _task->should_yield() above 3218 // because we want terminating threads to yield only if the task 3219 // is already in the midst of yielding, which happens only after at least one 3220 // thread has yielded. 3221 } 3222 3223 void CMSConcMarkingTerminator::yield() { 3224 if (_task->should_yield()) { 3225 _task->yield(); 3226 } else { 3227 ParallelTaskTerminator::yield(); 3228 } 3229 } 3230 3231 //////////////////////////////////////////////////////////////// 3232 // Concurrent Marking Algorithm Sketch 3233 //////////////////////////////////////////////////////////////// 3234 // Until all tasks exhausted (both spaces): 3235 // -- claim next available chunk 3236 // -- bump global finger via CAS 3237 // -- find first object that starts in this chunk 3238 // and start scanning bitmap from that position 3239 // -- scan marked objects for oops 3240 // -- CAS-mark target, and if successful: 3241 // . if target oop is above global finger (volatile read) 3242 // nothing to do 3243 // . if target oop is in chunk and above local finger 3244 // then nothing to do 3245 // . else push on work-queue 3246 // -- Deal with possible overflow issues: 3247 // . local work-queue overflow causes stuff to be pushed on 3248 // global (common) overflow queue 3249 // . always first empty local work queue 3250 // . then get a batch of oops from global work queue if any 3251 // . then do work stealing 3252 // -- When all tasks claimed (both spaces) 3253 // and local work queue empty, 3254 // then in a loop do: 3255 // . check global overflow stack; steal a batch of oops and trace 3256 // . try to steal from other threads oif GOS is empty 3257 // . if neither is available, offer termination 3258 // -- Terminate and return result 3259 // 3260 void CMSConcMarkingTask::work(uint worker_id) { 3261 elapsedTimer _timer; 3262 ResourceMark rm; 3263 HandleMark hm; 3264 3265 DEBUG_ONLY(_collector->verify_overflow_empty();) 3266 3267 // Before we begin work, our work queue should be empty 3268 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3269 // Scan the bitmap covering _cms_space, tracing through grey objects. 3270 _timer.start(); 3271 do_scan_and_mark(worker_id, _cms_space); 3272 _timer.stop(); 3273 if (PrintCMSStatistics != 0) { 3274 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3275 worker_id, _timer.seconds()); 3276 // XXX: need xxx/xxx type of notation, two timers 3277 } 3278 3279 // ... do work stealing 3280 _timer.reset(); 3281 _timer.start(); 3282 do_work_steal(worker_id); 3283 _timer.stop(); 3284 if (PrintCMSStatistics != 0) { 3285 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3286 worker_id, _timer.seconds()); 3287 // XXX: need xxx/xxx type of notation, two timers 3288 } 3289 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3290 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3291 // Note that under the current task protocol, the 3292 // following assertion is true even of the spaces 3293 // expanded since the completion of the concurrent 3294 // marking. XXX This will likely change under a strict 3295 // ABORT semantics. 3296 // After perm removal the comparison was changed to 3297 // greater than or equal to from strictly greater than. 3298 // Before perm removal the highest address sweep would 3299 // have been at the end of perm gen but now is at the 3300 // end of the tenured gen. 3301 assert(_global_finger >= _cms_space->end(), 3302 "All tasks have been completed"); 3303 DEBUG_ONLY(_collector->verify_overflow_empty();) 3304 } 3305 3306 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3307 HeapWord* read = _global_finger; 3308 HeapWord* cur = read; 3309 while (f > read) { 3310 cur = read; 3311 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3312 if (cur == read) { 3313 // our cas succeeded 3314 assert(_global_finger >= f, "protocol consistency"); 3315 break; 3316 } 3317 } 3318 } 3319 3320 // This is really inefficient, and should be redone by 3321 // using (not yet available) block-read and -write interfaces to the 3322 // stack and the work_queue. XXX FIX ME !!! 3323 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3324 OopTaskQueue* work_q) { 3325 // Fast lock-free check 3326 if (ovflw_stk->length() == 0) { 3327 return false; 3328 } 3329 assert(work_q->size() == 0, "Shouldn't steal"); 3330 MutexLockerEx ml(ovflw_stk->par_lock(), 3331 Mutex::_no_safepoint_check_flag); 3332 // Grab up to 1/4 the size of the work queue 3333 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3334 (size_t)ParGCDesiredObjsFromOverflowList); 3335 num = MIN2(num, ovflw_stk->length()); 3336 for (int i = (int) num; i > 0; i--) { 3337 oop cur = ovflw_stk->pop(); 3338 assert(cur != NULL, "Counted wrong?"); 3339 work_q->push(cur); 3340 } 3341 return num > 0; 3342 } 3343 3344 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3345 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3346 int n_tasks = pst->n_tasks(); 3347 // We allow that there may be no tasks to do here because 3348 // we are restarting after a stack overflow. 3349 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3350 uint nth_task = 0; 3351 3352 HeapWord* aligned_start = sp->bottom(); 3353 if (sp->used_region().contains(_restart_addr)) { 3354 // Align down to a card boundary for the start of 0th task 3355 // for this space. 3356 aligned_start = 3357 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3358 CardTableModRefBS::card_size); 3359 } 3360 3361 size_t chunk_size = sp->marking_task_size(); 3362 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3363 // Having claimed the nth task in this space, 3364 // compute the chunk that it corresponds to: 3365 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3366 aligned_start + (nth_task+1)*chunk_size); 3367 // Try and bump the global finger via a CAS; 3368 // note that we need to do the global finger bump 3369 // _before_ taking the intersection below, because 3370 // the task corresponding to that region will be 3371 // deemed done even if the used_region() expands 3372 // because of allocation -- as it almost certainly will 3373 // during start-up while the threads yield in the 3374 // closure below. 3375 HeapWord* finger = span.end(); 3376 bump_global_finger(finger); // atomically 3377 // There are null tasks here corresponding to chunks 3378 // beyond the "top" address of the space. 3379 span = span.intersection(sp->used_region()); 3380 if (!span.is_empty()) { // Non-null task 3381 HeapWord* prev_obj; 3382 assert(!span.contains(_restart_addr) || nth_task == 0, 3383 "Inconsistency"); 3384 if (nth_task == 0) { 3385 // For the 0th task, we'll not need to compute a block_start. 3386 if (span.contains(_restart_addr)) { 3387 // In the case of a restart because of stack overflow, 3388 // we might additionally skip a chunk prefix. 3389 prev_obj = _restart_addr; 3390 } else { 3391 prev_obj = span.start(); 3392 } 3393 } else { 3394 // We want to skip the first object because 3395 // the protocol is to scan any object in its entirety 3396 // that _starts_ in this span; a fortiori, any 3397 // object starting in an earlier span is scanned 3398 // as part of an earlier claimed task. 3399 // Below we use the "careful" version of block_start 3400 // so we do not try to navigate uninitialized objects. 3401 prev_obj = sp->block_start_careful(span.start()); 3402 // Below we use a variant of block_size that uses the 3403 // Printezis bits to avoid waiting for allocated 3404 // objects to become initialized/parsable. 3405 while (prev_obj < span.start()) { 3406 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3407 if (sz > 0) { 3408 prev_obj += sz; 3409 } else { 3410 // In this case we may end up doing a bit of redundant 3411 // scanning, but that appears unavoidable, short of 3412 // locking the free list locks; see bug 6324141. 3413 break; 3414 } 3415 } 3416 } 3417 if (prev_obj < span.end()) { 3418 MemRegion my_span = MemRegion(prev_obj, span.end()); 3419 // Do the marking work within a non-empty span -- 3420 // the last argument to the constructor indicates whether the 3421 // iteration should be incremental with periodic yields. 3422 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3423 &_collector->_markBitMap, 3424 work_queue(i), 3425 &_collector->_markStack); 3426 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3427 } // else nothing to do for this task 3428 } // else nothing to do for this task 3429 } 3430 // We'd be tempted to assert here that since there are no 3431 // more tasks left to claim in this space, the global_finger 3432 // must exceed space->top() and a fortiori space->end(). However, 3433 // that would not quite be correct because the bumping of 3434 // global_finger occurs strictly after the claiming of a task, 3435 // so by the time we reach here the global finger may not yet 3436 // have been bumped up by the thread that claimed the last 3437 // task. 3438 pst->all_tasks_completed(); 3439 } 3440 3441 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3442 private: 3443 CMSCollector* _collector; 3444 CMSConcMarkingTask* _task; 3445 MemRegion _span; 3446 CMSBitMap* _bit_map; 3447 CMSMarkStack* _overflow_stack; 3448 OopTaskQueue* _work_queue; 3449 protected: 3450 DO_OOP_WORK_DEFN 3451 public: 3452 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3453 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3454 MetadataAwareOopClosure(collector->ref_processor()), 3455 _collector(collector), 3456 _task(task), 3457 _span(collector->_span), 3458 _work_queue(work_queue), 3459 _bit_map(bit_map), 3460 _overflow_stack(overflow_stack) 3461 { } 3462 virtual void do_oop(oop* p); 3463 virtual void do_oop(narrowOop* p); 3464 3465 void trim_queue(size_t max); 3466 void handle_stack_overflow(HeapWord* lost); 3467 void do_yield_check() { 3468 if (_task->should_yield()) { 3469 _task->yield(); 3470 } 3471 } 3472 }; 3473 3474 // Grey object scanning during work stealing phase -- 3475 // the salient assumption here is that any references 3476 // that are in these stolen objects being scanned must 3477 // already have been initialized (else they would not have 3478 // been published), so we do not need to check for 3479 // uninitialized objects before pushing here. 3480 void Par_ConcMarkingClosure::do_oop(oop obj) { 3481 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3482 HeapWord* addr = (HeapWord*)obj; 3483 // Check if oop points into the CMS generation 3484 // and is not marked 3485 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3486 // a white object ... 3487 // If we manage to "claim" the object, by being the 3488 // first thread to mark it, then we push it on our 3489 // marking stack 3490 if (_bit_map->par_mark(addr)) { // ... now grey 3491 // push on work queue (grey set) 3492 bool simulate_overflow = false; 3493 NOT_PRODUCT( 3494 if (CMSMarkStackOverflowALot && 3495 _collector->simulate_overflow()) { 3496 // simulate a stack overflow 3497 simulate_overflow = true; 3498 } 3499 ) 3500 if (simulate_overflow || 3501 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3502 // stack overflow 3503 if (PrintCMSStatistics != 0) { 3504 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3505 SIZE_FORMAT, _overflow_stack->capacity()); 3506 } 3507 // We cannot assert that the overflow stack is full because 3508 // it may have been emptied since. 3509 assert(simulate_overflow || 3510 _work_queue->size() == _work_queue->max_elems(), 3511 "Else push should have succeeded"); 3512 handle_stack_overflow(addr); 3513 } 3514 } // Else, some other thread got there first 3515 do_yield_check(); 3516 } 3517 } 3518 3519 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3520 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3521 3522 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3523 while (_work_queue->size() > max) { 3524 oop new_oop; 3525 if (_work_queue->pop_local(new_oop)) { 3526 assert(new_oop->is_oop(), "Should be an oop"); 3527 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3528 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3529 new_oop->oop_iterate(this); // do_oop() above 3530 do_yield_check(); 3531 } 3532 } 3533 } 3534 3535 // Upon stack overflow, we discard (part of) the stack, 3536 // remembering the least address amongst those discarded 3537 // in CMSCollector's _restart_address. 3538 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3539 // We need to do this under a mutex to prevent other 3540 // workers from interfering with the work done below. 3541 MutexLockerEx ml(_overflow_stack->par_lock(), 3542 Mutex::_no_safepoint_check_flag); 3543 // Remember the least grey address discarded 3544 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3545 _collector->lower_restart_addr(ra); 3546 _overflow_stack->reset(); // discard stack contents 3547 _overflow_stack->expand(); // expand the stack if possible 3548 } 3549 3550 3551 void CMSConcMarkingTask::do_work_steal(int i) { 3552 OopTaskQueue* work_q = work_queue(i); 3553 oop obj_to_scan; 3554 CMSBitMap* bm = &(_collector->_markBitMap); 3555 CMSMarkStack* ovflw = &(_collector->_markStack); 3556 int* seed = _collector->hash_seed(i); 3557 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3558 while (true) { 3559 cl.trim_queue(0); 3560 assert(work_q->size() == 0, "Should have been emptied above"); 3561 if (get_work_from_overflow_stack(ovflw, work_q)) { 3562 // Can't assert below because the work obtained from the 3563 // overflow stack may already have been stolen from us. 3564 // assert(work_q->size() > 0, "Work from overflow stack"); 3565 continue; 3566 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3567 assert(obj_to_scan->is_oop(), "Should be an oop"); 3568 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3569 obj_to_scan->oop_iterate(&cl); 3570 } else if (terminator()->offer_termination(&_term_term)) { 3571 assert(work_q->size() == 0, "Impossible!"); 3572 break; 3573 } else if (yielding() || should_yield()) { 3574 yield(); 3575 } 3576 } 3577 } 3578 3579 // This is run by the CMS (coordinator) thread. 3580 void CMSConcMarkingTask::coordinator_yield() { 3581 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3582 "CMS thread should hold CMS token"); 3583 // First give up the locks, then yield, then re-lock 3584 // We should probably use a constructor/destructor idiom to 3585 // do this unlock/lock or modify the MutexUnlocker class to 3586 // serve our purpose. XXX 3587 assert_lock_strong(_bit_map_lock); 3588 _bit_map_lock->unlock(); 3589 ConcurrentMarkSweepThread::desynchronize(true); 3590 _collector->stopTimer(); 3591 if (PrintCMSStatistics != 0) { 3592 _collector->incrementYields(); 3593 } 3594 3595 // It is possible for whichever thread initiated the yield request 3596 // not to get a chance to wake up and take the bitmap lock between 3597 // this thread releasing it and reacquiring it. So, while the 3598 // should_yield() flag is on, let's sleep for a bit to give the 3599 // other thread a chance to wake up. The limit imposed on the number 3600 // of iterations is defensive, to avoid any unforseen circumstances 3601 // putting us into an infinite loop. Since it's always been this 3602 // (coordinator_yield()) method that was observed to cause the 3603 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3604 // which is by default non-zero. For the other seven methods that 3605 // also perform the yield operation, as are using a different 3606 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3607 // can enable the sleeping for those methods too, if necessary. 3608 // See 6442774. 3609 // 3610 // We really need to reconsider the synchronization between the GC 3611 // thread and the yield-requesting threads in the future and we 3612 // should really use wait/notify, which is the recommended 3613 // way of doing this type of interaction. Additionally, we should 3614 // consolidate the eight methods that do the yield operation and they 3615 // are almost identical into one for better maintainability and 3616 // readability. See 6445193. 3617 // 3618 // Tony 2006.06.29 3619 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3620 ConcurrentMarkSweepThread::should_yield() && 3621 !CMSCollector::foregroundGCIsActive(); ++i) { 3622 os::sleep(Thread::current(), 1, false); 3623 } 3624 3625 ConcurrentMarkSweepThread::synchronize(true); 3626 _bit_map_lock->lock_without_safepoint_check(); 3627 _collector->startTimer(); 3628 } 3629 3630 bool CMSCollector::do_marking_mt() { 3631 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3632 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3633 conc_workers()->active_workers(), 3634 Threads::number_of_non_daemon_threads()); 3635 conc_workers()->set_active_workers(num_workers); 3636 3637 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3638 3639 CMSConcMarkingTask tsk(this, 3640 cms_space, 3641 conc_workers(), 3642 task_queues()); 3643 3644 // Since the actual number of workers we get may be different 3645 // from the number we requested above, do we need to do anything different 3646 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3647 // class?? XXX 3648 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3649 3650 // Refs discovery is already non-atomic. 3651 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3652 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3653 conc_workers()->start_task(&tsk); 3654 while (tsk.yielded()) { 3655 tsk.coordinator_yield(); 3656 conc_workers()->continue_task(&tsk); 3657 } 3658 // If the task was aborted, _restart_addr will be non-NULL 3659 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3660 while (_restart_addr != NULL) { 3661 // XXX For now we do not make use of ABORTED state and have not 3662 // yet implemented the right abort semantics (even in the original 3663 // single-threaded CMS case). That needs some more investigation 3664 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3665 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3666 // If _restart_addr is non-NULL, a marking stack overflow 3667 // occurred; we need to do a fresh marking iteration from the 3668 // indicated restart address. 3669 if (_foregroundGCIsActive) { 3670 // We may be running into repeated stack overflows, having 3671 // reached the limit of the stack size, while making very 3672 // slow forward progress. It may be best to bail out and 3673 // let the foreground collector do its job. 3674 // Clear _restart_addr, so that foreground GC 3675 // works from scratch. This avoids the headache of 3676 // a "rescan" which would otherwise be needed because 3677 // of the dirty mod union table & card table. 3678 _restart_addr = NULL; 3679 return false; 3680 } 3681 // Adjust the task to restart from _restart_addr 3682 tsk.reset(_restart_addr); 3683 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3684 _restart_addr); 3685 _restart_addr = NULL; 3686 // Get the workers going again 3687 conc_workers()->start_task(&tsk); 3688 while (tsk.yielded()) { 3689 tsk.coordinator_yield(); 3690 conc_workers()->continue_task(&tsk); 3691 } 3692 } 3693 assert(tsk.completed(), "Inconsistency"); 3694 assert(tsk.result() == true, "Inconsistency"); 3695 return true; 3696 } 3697 3698 bool CMSCollector::do_marking_st() { 3699 ResourceMark rm; 3700 HandleMark hm; 3701 3702 // Temporarily make refs discovery single threaded (non-MT) 3703 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3704 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3705 &_markStack, CMSYield); 3706 // the last argument to iterate indicates whether the iteration 3707 // should be incremental with periodic yields. 3708 _markBitMap.iterate(&markFromRootsClosure); 3709 // If _restart_addr is non-NULL, a marking stack overflow 3710 // occurred; we need to do a fresh iteration from the 3711 // indicated restart address. 3712 while (_restart_addr != NULL) { 3713 if (_foregroundGCIsActive) { 3714 // We may be running into repeated stack overflows, having 3715 // reached the limit of the stack size, while making very 3716 // slow forward progress. It may be best to bail out and 3717 // let the foreground collector do its job. 3718 // Clear _restart_addr, so that foreground GC 3719 // works from scratch. This avoids the headache of 3720 // a "rescan" which would otherwise be needed because 3721 // of the dirty mod union table & card table. 3722 _restart_addr = NULL; 3723 return false; // indicating failure to complete marking 3724 } 3725 // Deal with stack overflow: 3726 // we restart marking from _restart_addr 3727 HeapWord* ra = _restart_addr; 3728 markFromRootsClosure.reset(ra); 3729 _restart_addr = NULL; 3730 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3731 } 3732 return true; 3733 } 3734 3735 void CMSCollector::preclean() { 3736 check_correct_thread_executing(); 3737 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3738 verify_work_stacks_empty(); 3739 verify_overflow_empty(); 3740 _abort_preclean = false; 3741 if (CMSPrecleaningEnabled) { 3742 if (!CMSEdenChunksRecordAlways) { 3743 _eden_chunk_index = 0; 3744 } 3745 size_t used = get_eden_used(); 3746 size_t capacity = get_eden_capacity(); 3747 // Don't start sampling unless we will get sufficiently 3748 // many samples. 3749 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3750 * CMSScheduleRemarkEdenPenetration)) { 3751 _start_sampling = true; 3752 } else { 3753 _start_sampling = false; 3754 } 3755 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3756 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3757 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3758 } 3759 CMSTokenSync x(true); // is cms thread 3760 if (CMSPrecleaningEnabled) { 3761 sample_eden(); 3762 _collectorState = AbortablePreclean; 3763 } else { 3764 _collectorState = FinalMarking; 3765 } 3766 verify_work_stacks_empty(); 3767 verify_overflow_empty(); 3768 } 3769 3770 // Try and schedule the remark such that young gen 3771 // occupancy is CMSScheduleRemarkEdenPenetration %. 3772 void CMSCollector::abortable_preclean() { 3773 check_correct_thread_executing(); 3774 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3775 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3776 3777 // If Eden's current occupancy is below this threshold, 3778 // immediately schedule the remark; else preclean 3779 // past the next scavenge in an effort to 3780 // schedule the pause as described above. By choosing 3781 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3782 // we will never do an actual abortable preclean cycle. 3783 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3784 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3785 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3786 // We need more smarts in the abortable preclean 3787 // loop below to deal with cases where allocation 3788 // in young gen is very very slow, and our precleaning 3789 // is running a losing race against a horde of 3790 // mutators intent on flooding us with CMS updates 3791 // (dirty cards). 3792 // One, admittedly dumb, strategy is to give up 3793 // after a certain number of abortable precleaning loops 3794 // or after a certain maximum time. We want to make 3795 // this smarter in the next iteration. 3796 // XXX FIX ME!!! YSR 3797 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3798 while (!(should_abort_preclean() || 3799 ConcurrentMarkSweepThread::should_terminate())) { 3800 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3801 cumworkdone += workdone; 3802 loops++; 3803 // Voluntarily terminate abortable preclean phase if we have 3804 // been at it for too long. 3805 if ((CMSMaxAbortablePrecleanLoops != 0) && 3806 loops >= CMSMaxAbortablePrecleanLoops) { 3807 if (PrintGCDetails) { 3808 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3809 } 3810 break; 3811 } 3812 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3813 if (PrintGCDetails) { 3814 gclog_or_tty->print(" CMS: abort preclean due to time "); 3815 } 3816 break; 3817 } 3818 // If we are doing little work each iteration, we should 3819 // take a short break. 3820 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3821 // Sleep for some time, waiting for work to accumulate 3822 stopTimer(); 3823 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3824 startTimer(); 3825 waited++; 3826 } 3827 } 3828 if (PrintCMSStatistics > 0) { 3829 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3830 loops, waited, cumworkdone); 3831 } 3832 } 3833 CMSTokenSync x(true); // is cms thread 3834 if (_collectorState != Idling) { 3835 assert(_collectorState == AbortablePreclean, 3836 "Spontaneous state transition?"); 3837 _collectorState = FinalMarking; 3838 } // Else, a foreground collection completed this CMS cycle. 3839 return; 3840 } 3841 3842 // Respond to an Eden sampling opportunity 3843 void CMSCollector::sample_eden() { 3844 // Make sure a young gc cannot sneak in between our 3845 // reading and recording of a sample. 3846 assert(Thread::current()->is_ConcurrentGC_thread(), 3847 "Only the cms thread may collect Eden samples"); 3848 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3849 "Should collect samples while holding CMS token"); 3850 if (!_start_sampling) { 3851 return; 3852 } 3853 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3854 // is populated by the young generation. 3855 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3856 if (_eden_chunk_index < _eden_chunk_capacity) { 3857 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3858 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3859 "Unexpected state of Eden"); 3860 // We'd like to check that what we just sampled is an oop-start address; 3861 // however, we cannot do that here since the object may not yet have been 3862 // initialized. So we'll instead do the check when we _use_ this sample 3863 // later. 3864 if (_eden_chunk_index == 0 || 3865 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3866 _eden_chunk_array[_eden_chunk_index-1]) 3867 >= CMSSamplingGrain)) { 3868 _eden_chunk_index++; // commit sample 3869 } 3870 } 3871 } 3872 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3873 size_t used = get_eden_used(); 3874 size_t capacity = get_eden_capacity(); 3875 assert(used <= capacity, "Unexpected state of Eden"); 3876 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3877 _abort_preclean = true; 3878 } 3879 } 3880 } 3881 3882 3883 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3884 assert(_collectorState == Precleaning || 3885 _collectorState == AbortablePreclean, "incorrect state"); 3886 ResourceMark rm; 3887 HandleMark hm; 3888 3889 // Precleaning is currently not MT but the reference processor 3890 // may be set for MT. Disable it temporarily here. 3891 ReferenceProcessor* rp = ref_processor(); 3892 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3893 3894 // Do one pass of scrubbing the discovered reference lists 3895 // to remove any reference objects with strongly-reachable 3896 // referents. 3897 if (clean_refs) { 3898 CMSPrecleanRefsYieldClosure yield_cl(this); 3899 assert(rp->span().equals(_span), "Spans should be equal"); 3900 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3901 &_markStack, true /* preclean */); 3902 CMSDrainMarkingStackClosure complete_trace(this, 3903 _span, &_markBitMap, &_markStack, 3904 &keep_alive, true /* preclean */); 3905 3906 // We don't want this step to interfere with a young 3907 // collection because we don't want to take CPU 3908 // or memory bandwidth away from the young GC threads 3909 // (which may be as many as there are CPUs). 3910 // Note that we don't need to protect ourselves from 3911 // interference with mutators because they can't 3912 // manipulate the discovered reference lists nor affect 3913 // the computed reachability of the referents, the 3914 // only properties manipulated by the precleaning 3915 // of these reference lists. 3916 stopTimer(); 3917 CMSTokenSyncWithLocks x(true /* is cms thread */, 3918 bitMapLock()); 3919 startTimer(); 3920 sample_eden(); 3921 3922 // The following will yield to allow foreground 3923 // collection to proceed promptly. XXX YSR: 3924 // The code in this method may need further 3925 // tweaking for better performance and some restructuring 3926 // for cleaner interfaces. 3927 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3928 rp->preclean_discovered_references( 3929 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3930 gc_timer, _gc_tracer_cm->gc_id()); 3931 } 3932 3933 if (clean_survivor) { // preclean the active survivor space(s) 3934 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3935 &_markBitMap, &_modUnionTable, 3936 &_markStack, true /* precleaning phase */); 3937 stopTimer(); 3938 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3939 bitMapLock()); 3940 startTimer(); 3941 unsigned int before_count = 3942 GenCollectedHeap::heap()->total_collections(); 3943 SurvivorSpacePrecleanClosure 3944 sss_cl(this, _span, &_markBitMap, &_markStack, 3945 &pam_cl, before_count, CMSYield); 3946 _young_gen->from()->object_iterate_careful(&sss_cl); 3947 _young_gen->to()->object_iterate_careful(&sss_cl); 3948 } 3949 MarkRefsIntoAndScanClosure 3950 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3951 &_markStack, this, CMSYield, 3952 true /* precleaning phase */); 3953 // CAUTION: The following closure has persistent state that may need to 3954 // be reset upon a decrease in the sequence of addresses it 3955 // processes. 3956 ScanMarkedObjectsAgainCarefullyClosure 3957 smoac_cl(this, _span, 3958 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3959 3960 // Preclean dirty cards in ModUnionTable and CardTable using 3961 // appropriate convergence criterion; 3962 // repeat CMSPrecleanIter times unless we find that 3963 // we are losing. 3964 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3965 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3966 "Bad convergence multiplier"); 3967 assert(CMSPrecleanThreshold >= 100, 3968 "Unreasonably low CMSPrecleanThreshold"); 3969 3970 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3971 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3972 numIter < CMSPrecleanIter; 3973 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3974 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3975 if (Verbose && PrintGCDetails) { 3976 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3977 } 3978 // Either there are very few dirty cards, so re-mark 3979 // pause will be small anyway, or our pre-cleaning isn't 3980 // that much faster than the rate at which cards are being 3981 // dirtied, so we might as well stop and re-mark since 3982 // precleaning won't improve our re-mark time by much. 3983 if (curNumCards <= CMSPrecleanThreshold || 3984 (numIter > 0 && 3985 (curNumCards * CMSPrecleanDenominator > 3986 lastNumCards * CMSPrecleanNumerator))) { 3987 numIter++; 3988 cumNumCards += curNumCards; 3989 break; 3990 } 3991 } 3992 3993 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3994 3995 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3996 cumNumCards += curNumCards; 3997 if (PrintGCDetails && PrintCMSStatistics != 0) { 3998 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3999 curNumCards, cumNumCards, numIter); 4000 } 4001 return cumNumCards; // as a measure of useful work done 4002 } 4003 4004 // PRECLEANING NOTES: 4005 // Precleaning involves: 4006 // . reading the bits of the modUnionTable and clearing the set bits. 4007 // . For the cards corresponding to the set bits, we scan the 4008 // objects on those cards. This means we need the free_list_lock 4009 // so that we can safely iterate over the CMS space when scanning 4010 // for oops. 4011 // . When we scan the objects, we'll be both reading and setting 4012 // marks in the marking bit map, so we'll need the marking bit map. 4013 // . For protecting _collector_state transitions, we take the CGC_lock. 4014 // Note that any races in the reading of of card table entries by the 4015 // CMS thread on the one hand and the clearing of those entries by the 4016 // VM thread or the setting of those entries by the mutator threads on the 4017 // other are quite benign. However, for efficiency it makes sense to keep 4018 // the VM thread from racing with the CMS thread while the latter is 4019 // dirty card info to the modUnionTable. We therefore also use the 4020 // CGC_lock to protect the reading of the card table and the mod union 4021 // table by the CM thread. 4022 // . We run concurrently with mutator updates, so scanning 4023 // needs to be done carefully -- we should not try to scan 4024 // potentially uninitialized objects. 4025 // 4026 // Locking strategy: While holding the CGC_lock, we scan over and 4027 // reset a maximal dirty range of the mod union / card tables, then lock 4028 // the free_list_lock and bitmap lock to do a full marking, then 4029 // release these locks; and repeat the cycle. This allows for a 4030 // certain amount of fairness in the sharing of these locks between 4031 // the CMS collector on the one hand, and the VM thread and the 4032 // mutators on the other. 4033 4034 // NOTE: preclean_mod_union_table() and preclean_card_table() 4035 // further below are largely identical; if you need to modify 4036 // one of these methods, please check the other method too. 4037 4038 size_t CMSCollector::preclean_mod_union_table( 4039 ConcurrentMarkSweepGeneration* gen, 4040 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4041 verify_work_stacks_empty(); 4042 verify_overflow_empty(); 4043 4044 // strategy: starting with the first card, accumulate contiguous 4045 // ranges of dirty cards; clear these cards, then scan the region 4046 // covered by these cards. 4047 4048 // Since all of the MUT is committed ahead, we can just use 4049 // that, in case the generations expand while we are precleaning. 4050 // It might also be fine to just use the committed part of the 4051 // generation, but we might potentially miss cards when the 4052 // generation is rapidly expanding while we are in the midst 4053 // of precleaning. 4054 HeapWord* startAddr = gen->reserved().start(); 4055 HeapWord* endAddr = gen->reserved().end(); 4056 4057 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4058 4059 size_t numDirtyCards, cumNumDirtyCards; 4060 HeapWord *nextAddr, *lastAddr; 4061 for (cumNumDirtyCards = numDirtyCards = 0, 4062 nextAddr = lastAddr = startAddr; 4063 nextAddr < endAddr; 4064 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4065 4066 ResourceMark rm; 4067 HandleMark hm; 4068 4069 MemRegion dirtyRegion; 4070 { 4071 stopTimer(); 4072 // Potential yield point 4073 CMSTokenSync ts(true); 4074 startTimer(); 4075 sample_eden(); 4076 // Get dirty region starting at nextOffset (inclusive), 4077 // simultaneously clearing it. 4078 dirtyRegion = 4079 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4080 assert(dirtyRegion.start() >= nextAddr, 4081 "returned region inconsistent?"); 4082 } 4083 // Remember where the next search should begin. 4084 // The returned region (if non-empty) is a right open interval, 4085 // so lastOffset is obtained from the right end of that 4086 // interval. 4087 lastAddr = dirtyRegion.end(); 4088 // Should do something more transparent and less hacky XXX 4089 numDirtyCards = 4090 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4091 4092 // We'll scan the cards in the dirty region (with periodic 4093 // yields for foreground GC as needed). 4094 if (!dirtyRegion.is_empty()) { 4095 assert(numDirtyCards > 0, "consistency check"); 4096 HeapWord* stop_point = NULL; 4097 stopTimer(); 4098 // Potential yield point 4099 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4100 bitMapLock()); 4101 startTimer(); 4102 { 4103 verify_work_stacks_empty(); 4104 verify_overflow_empty(); 4105 sample_eden(); 4106 stop_point = 4107 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4108 } 4109 if (stop_point != NULL) { 4110 // The careful iteration stopped early either because it found an 4111 // uninitialized object, or because we were in the midst of an 4112 // "abortable preclean", which should now be aborted. Redirty 4113 // the bits corresponding to the partially-scanned or unscanned 4114 // cards. We'll either restart at the next block boundary or 4115 // abort the preclean. 4116 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4117 "Should only be AbortablePreclean."); 4118 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4119 if (should_abort_preclean()) { 4120 break; // out of preclean loop 4121 } else { 4122 // Compute the next address at which preclean should pick up; 4123 // might need bitMapLock in order to read P-bits. 4124 lastAddr = next_card_start_after_block(stop_point); 4125 } 4126 } 4127 } else { 4128 assert(lastAddr == endAddr, "consistency check"); 4129 assert(numDirtyCards == 0, "consistency check"); 4130 break; 4131 } 4132 } 4133 verify_work_stacks_empty(); 4134 verify_overflow_empty(); 4135 return cumNumDirtyCards; 4136 } 4137 4138 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4139 // below are largely identical; if you need to modify 4140 // one of these methods, please check the other method too. 4141 4142 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4143 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4144 // strategy: it's similar to precleamModUnionTable above, in that 4145 // we accumulate contiguous ranges of dirty cards, mark these cards 4146 // precleaned, then scan the region covered by these cards. 4147 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4148 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4149 4150 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4151 4152 size_t numDirtyCards, cumNumDirtyCards; 4153 HeapWord *lastAddr, *nextAddr; 4154 4155 for (cumNumDirtyCards = numDirtyCards = 0, 4156 nextAddr = lastAddr = startAddr; 4157 nextAddr < endAddr; 4158 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4159 4160 ResourceMark rm; 4161 HandleMark hm; 4162 4163 MemRegion dirtyRegion; 4164 { 4165 // See comments in "Precleaning notes" above on why we 4166 // do this locking. XXX Could the locking overheads be 4167 // too high when dirty cards are sparse? [I don't think so.] 4168 stopTimer(); 4169 CMSTokenSync x(true); // is cms thread 4170 startTimer(); 4171 sample_eden(); 4172 // Get and clear dirty region from card table 4173 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4174 MemRegion(nextAddr, endAddr), 4175 true, 4176 CardTableModRefBS::precleaned_card_val()); 4177 4178 assert(dirtyRegion.start() >= nextAddr, 4179 "returned region inconsistent?"); 4180 } 4181 lastAddr = dirtyRegion.end(); 4182 numDirtyCards = 4183 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4184 4185 if (!dirtyRegion.is_empty()) { 4186 stopTimer(); 4187 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4188 startTimer(); 4189 sample_eden(); 4190 verify_work_stacks_empty(); 4191 verify_overflow_empty(); 4192 HeapWord* stop_point = 4193 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4194 if (stop_point != NULL) { 4195 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4196 "Should only be AbortablePreclean."); 4197 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4198 if (should_abort_preclean()) { 4199 break; // out of preclean loop 4200 } else { 4201 // Compute the next address at which preclean should pick up. 4202 lastAddr = next_card_start_after_block(stop_point); 4203 } 4204 } 4205 } else { 4206 break; 4207 } 4208 } 4209 verify_work_stacks_empty(); 4210 verify_overflow_empty(); 4211 return cumNumDirtyCards; 4212 } 4213 4214 class PrecleanKlassClosure : public KlassClosure { 4215 KlassToOopClosure _cm_klass_closure; 4216 public: 4217 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4218 void do_klass(Klass* k) { 4219 if (k->has_accumulated_modified_oops()) { 4220 k->clear_accumulated_modified_oops(); 4221 4222 _cm_klass_closure.do_klass(k); 4223 } 4224 } 4225 }; 4226 4227 // The freelist lock is needed to prevent asserts, is it really needed? 4228 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4229 4230 cl->set_freelistLock(freelistLock); 4231 4232 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4233 4234 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4235 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4236 PrecleanKlassClosure preclean_klass_closure(cl); 4237 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4238 4239 verify_work_stacks_empty(); 4240 verify_overflow_empty(); 4241 } 4242 4243 void CMSCollector::checkpointRootsFinal() { 4244 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4245 check_correct_thread_executing(); 4246 // world is stopped at this checkpoint 4247 assert(SafepointSynchronize::is_at_safepoint(), 4248 "world should be stopped"); 4249 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4250 4251 verify_work_stacks_empty(); 4252 verify_overflow_empty(); 4253 4254 if (PrintGCDetails) { 4255 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4256 _young_gen->used() / K, 4257 _young_gen->capacity() / K); 4258 } 4259 { 4260 if (CMSScavengeBeforeRemark) { 4261 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4262 // Temporarily set flag to false, GCH->do_collection will 4263 // expect it to be false and set to true 4264 FlagSetting fl(gch->_is_gc_active, false); 4265 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4266 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4267 int level = _cmsGen->level() - 1; 4268 if (level >= 0) { 4269 gch->do_collection(true, // full (i.e. force, see below) 4270 false, // !clear_all_soft_refs 4271 0, // size 4272 false, // is_tlab 4273 level // max_level 4274 ); 4275 } 4276 } 4277 FreelistLocker x(this); 4278 MutexLockerEx y(bitMapLock(), 4279 Mutex::_no_safepoint_check_flag); 4280 checkpointRootsFinalWork(); 4281 } 4282 verify_work_stacks_empty(); 4283 verify_overflow_empty(); 4284 } 4285 4286 void CMSCollector::checkpointRootsFinalWork() { 4287 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4288 4289 assert(haveFreelistLocks(), "must have free list locks"); 4290 assert_lock_strong(bitMapLock()); 4291 4292 ResourceMark rm; 4293 HandleMark hm; 4294 4295 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4296 4297 if (should_unload_classes()) { 4298 CodeCache::gc_prologue(); 4299 } 4300 assert(haveFreelistLocks(), "must have free list locks"); 4301 assert_lock_strong(bitMapLock()); 4302 4303 // We might assume that we need not fill TLAB's when 4304 // CMSScavengeBeforeRemark is set, because we may have just done 4305 // a scavenge which would have filled all TLAB's -- and besides 4306 // Eden would be empty. This however may not always be the case -- 4307 // for instance although we asked for a scavenge, it may not have 4308 // happened because of a JNI critical section. We probably need 4309 // a policy for deciding whether we can in that case wait until 4310 // the critical section releases and then do the remark following 4311 // the scavenge, and skip it here. In the absence of that policy, 4312 // or of an indication of whether the scavenge did indeed occur, 4313 // we cannot rely on TLAB's having been filled and must do 4314 // so here just in case a scavenge did not happen. 4315 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4316 // Update the saved marks which may affect the root scans. 4317 gch->save_marks(); 4318 4319 if (CMSPrintEdenSurvivorChunks) { 4320 print_eden_and_survivor_chunk_arrays(); 4321 } 4322 4323 { 4324 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4325 4326 // Note on the role of the mod union table: 4327 // Since the marker in "markFromRoots" marks concurrently with 4328 // mutators, it is possible for some reachable objects not to have been 4329 // scanned. For instance, an only reference to an object A was 4330 // placed in object B after the marker scanned B. Unless B is rescanned, 4331 // A would be collected. Such updates to references in marked objects 4332 // are detected via the mod union table which is the set of all cards 4333 // dirtied since the first checkpoint in this GC cycle and prior to 4334 // the most recent young generation GC, minus those cleaned up by the 4335 // concurrent precleaning. 4336 if (CMSParallelRemarkEnabled) { 4337 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4338 do_remark_parallel(); 4339 } else { 4340 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4341 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4342 do_remark_non_parallel(); 4343 } 4344 } 4345 verify_work_stacks_empty(); 4346 verify_overflow_empty(); 4347 4348 { 4349 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4350 refProcessingWork(); 4351 } 4352 verify_work_stacks_empty(); 4353 verify_overflow_empty(); 4354 4355 if (should_unload_classes()) { 4356 CodeCache::gc_epilogue(); 4357 } 4358 JvmtiExport::gc_epilogue(); 4359 4360 // If we encountered any (marking stack / work queue) overflow 4361 // events during the current CMS cycle, take appropriate 4362 // remedial measures, where possible, so as to try and avoid 4363 // recurrence of that condition. 4364 assert(_markStack.isEmpty(), "No grey objects"); 4365 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4366 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4367 if (ser_ovflw > 0) { 4368 if (PrintCMSStatistics != 0) { 4369 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4370 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4371 ", kac_preclean="SIZE_FORMAT")", 4372 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4373 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4374 } 4375 _markStack.expand(); 4376 _ser_pmc_remark_ovflw = 0; 4377 _ser_pmc_preclean_ovflw = 0; 4378 _ser_kac_preclean_ovflw = 0; 4379 _ser_kac_ovflw = 0; 4380 } 4381 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4382 if (PrintCMSStatistics != 0) { 4383 gclog_or_tty->print_cr("Work queue overflow (benign) " 4384 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4385 _par_pmc_remark_ovflw, _par_kac_ovflw); 4386 } 4387 _par_pmc_remark_ovflw = 0; 4388 _par_kac_ovflw = 0; 4389 } 4390 if (PrintCMSStatistics != 0) { 4391 if (_markStack._hit_limit > 0) { 4392 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4393 _markStack._hit_limit); 4394 } 4395 if (_markStack._failed_double > 0) { 4396 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4397 " current capacity "SIZE_FORMAT, 4398 _markStack._failed_double, 4399 _markStack.capacity()); 4400 } 4401 } 4402 _markStack._hit_limit = 0; 4403 _markStack._failed_double = 0; 4404 4405 if ((VerifyAfterGC || VerifyDuringGC) && 4406 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4407 verify_after_remark(); 4408 } 4409 4410 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4411 4412 // Change under the freelistLocks. 4413 _collectorState = Sweeping; 4414 // Call isAllClear() under bitMapLock 4415 assert(_modUnionTable.isAllClear(), 4416 "Should be clear by end of the final marking"); 4417 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4418 "Should be clear by end of the final marking"); 4419 } 4420 4421 void CMSParInitialMarkTask::work(uint worker_id) { 4422 elapsedTimer _timer; 4423 ResourceMark rm; 4424 HandleMark hm; 4425 4426 // ---------- scan from roots -------------- 4427 _timer.start(); 4428 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4429 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4430 4431 // ---------- young gen roots -------------- 4432 { 4433 work_on_young_gen_roots(worker_id, &par_mri_cl); 4434 _timer.stop(); 4435 if (PrintCMSStatistics != 0) { 4436 gclog_or_tty->print_cr( 4437 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4438 worker_id, _timer.seconds()); 4439 } 4440 } 4441 4442 // ---------- remaining roots -------------- 4443 _timer.reset(); 4444 _timer.start(); 4445 4446 CLDToOopClosure cld_closure(&par_mri_cl, true); 4447 4448 gch->gen_process_roots(_collector->_cmsGen->level(), 4449 false, // yg was scanned above 4450 false, // this is parallel code 4451 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4452 _collector->should_unload_classes(), 4453 &par_mri_cl, 4454 NULL, 4455 &cld_closure); 4456 assert(_collector->should_unload_classes() 4457 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4458 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4459 _timer.stop(); 4460 if (PrintCMSStatistics != 0) { 4461 gclog_or_tty->print_cr( 4462 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4463 worker_id, _timer.seconds()); 4464 } 4465 } 4466 4467 // Parallel remark task 4468 class CMSParRemarkTask: public CMSParMarkTask { 4469 CompactibleFreeListSpace* _cms_space; 4470 4471 // The per-thread work queues, available here for stealing. 4472 OopTaskQueueSet* _task_queues; 4473 ParallelTaskTerminator _term; 4474 4475 public: 4476 // A value of 0 passed to n_workers will cause the number of 4477 // workers to be taken from the active workers in the work gang. 4478 CMSParRemarkTask(CMSCollector* collector, 4479 CompactibleFreeListSpace* cms_space, 4480 uint n_workers, FlexibleWorkGang* workers, 4481 OopTaskQueueSet* task_queues): 4482 CMSParMarkTask("Rescan roots and grey objects in parallel", 4483 collector, n_workers), 4484 _cms_space(cms_space), 4485 _task_queues(task_queues), 4486 _term(n_workers, task_queues) { } 4487 4488 OopTaskQueueSet* task_queues() { return _task_queues; } 4489 4490 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4491 4492 ParallelTaskTerminator* terminator() { return &_term; } 4493 uint n_workers() { return _n_workers; } 4494 4495 void work(uint worker_id); 4496 4497 private: 4498 // ... of dirty cards in old space 4499 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4500 Par_MarkRefsIntoAndScanClosure* cl); 4501 4502 // ... work stealing for the above 4503 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4504 }; 4505 4506 class RemarkKlassClosure : public KlassClosure { 4507 KlassToOopClosure _cm_klass_closure; 4508 public: 4509 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4510 void do_klass(Klass* k) { 4511 // Check if we have modified any oops in the Klass during the concurrent marking. 4512 if (k->has_accumulated_modified_oops()) { 4513 k->clear_accumulated_modified_oops(); 4514 4515 // We could have transfered the current modified marks to the accumulated marks, 4516 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4517 } else if (k->has_modified_oops()) { 4518 // Don't clear anything, this info is needed by the next young collection. 4519 } else { 4520 // No modified oops in the Klass. 4521 return; 4522 } 4523 4524 // The klass has modified fields, need to scan the klass. 4525 _cm_klass_closure.do_klass(k); 4526 } 4527 }; 4528 4529 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4530 ParNewGeneration* young_gen = _collector->_young_gen; 4531 ContiguousSpace* eden_space = young_gen->eden(); 4532 ContiguousSpace* from_space = young_gen->from(); 4533 ContiguousSpace* to_space = young_gen->to(); 4534 4535 HeapWord** eca = _collector->_eden_chunk_array; 4536 size_t ect = _collector->_eden_chunk_index; 4537 HeapWord** sca = _collector->_survivor_chunk_array; 4538 size_t sct = _collector->_survivor_chunk_index; 4539 4540 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4541 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4542 4543 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4544 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4545 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4546 } 4547 4548 // work_queue(i) is passed to the closure 4549 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4550 // also is passed to do_dirty_card_rescan_tasks() and to 4551 // do_work_steal() to select the i-th task_queue. 4552 4553 void CMSParRemarkTask::work(uint worker_id) { 4554 elapsedTimer _timer; 4555 ResourceMark rm; 4556 HandleMark hm; 4557 4558 // ---------- rescan from roots -------------- 4559 _timer.start(); 4560 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4561 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4562 _collector->_span, _collector->ref_processor(), 4563 &(_collector->_markBitMap), 4564 work_queue(worker_id)); 4565 4566 // Rescan young gen roots first since these are likely 4567 // coarsely partitioned and may, on that account, constitute 4568 // the critical path; thus, it's best to start off that 4569 // work first. 4570 // ---------- young gen roots -------------- 4571 { 4572 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4573 _timer.stop(); 4574 if (PrintCMSStatistics != 0) { 4575 gclog_or_tty->print_cr( 4576 "Finished young gen rescan work in %dth thread: %3.3f sec", 4577 worker_id, _timer.seconds()); 4578 } 4579 } 4580 4581 // ---------- remaining roots -------------- 4582 _timer.reset(); 4583 _timer.start(); 4584 gch->gen_process_roots(_collector->_cmsGen->level(), 4585 false, // yg was scanned above 4586 false, // this is parallel code 4587 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4588 _collector->should_unload_classes(), 4589 &par_mrias_cl, 4590 NULL, 4591 NULL); // The dirty klasses will be handled below 4592 4593 assert(_collector->should_unload_classes() 4594 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4595 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4596 _timer.stop(); 4597 if (PrintCMSStatistics != 0) { 4598 gclog_or_tty->print_cr( 4599 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4600 worker_id, _timer.seconds()); 4601 } 4602 4603 // ---------- unhandled CLD scanning ---------- 4604 if (worker_id == 0) { // Single threaded at the moment. 4605 _timer.reset(); 4606 _timer.start(); 4607 4608 // Scan all new class loader data objects and new dependencies that were 4609 // introduced during concurrent marking. 4610 ResourceMark rm; 4611 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4612 for (int i = 0; i < array->length(); i++) { 4613 par_mrias_cl.do_class_loader_data(array->at(i)); 4614 } 4615 4616 // We don't need to keep track of new CLDs anymore. 4617 ClassLoaderDataGraph::remember_new_clds(false); 4618 4619 _timer.stop(); 4620 if (PrintCMSStatistics != 0) { 4621 gclog_or_tty->print_cr( 4622 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4623 worker_id, _timer.seconds()); 4624 } 4625 } 4626 4627 // ---------- dirty klass scanning ---------- 4628 if (worker_id == 0) { // Single threaded at the moment. 4629 _timer.reset(); 4630 _timer.start(); 4631 4632 // Scan all classes that was dirtied during the concurrent marking phase. 4633 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4634 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4635 4636 _timer.stop(); 4637 if (PrintCMSStatistics != 0) { 4638 gclog_or_tty->print_cr( 4639 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4640 worker_id, _timer.seconds()); 4641 } 4642 } 4643 4644 // We might have added oops to ClassLoaderData::_handles during the 4645 // concurrent marking phase. These oops point to newly allocated objects 4646 // that are guaranteed to be kept alive. Either by the direct allocation 4647 // code, or when the young collector processes the roots. Hence, 4648 // we don't have to revisit the _handles block during the remark phase. 4649 4650 // ---------- rescan dirty cards ------------ 4651 _timer.reset(); 4652 _timer.start(); 4653 4654 // Do the rescan tasks for each of the two spaces 4655 // (cms_space) in turn. 4656 // "worker_id" is passed to select the task_queue for "worker_id" 4657 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4658 _timer.stop(); 4659 if (PrintCMSStatistics != 0) { 4660 gclog_or_tty->print_cr( 4661 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4662 worker_id, _timer.seconds()); 4663 } 4664 4665 // ---------- steal work from other threads ... 4666 // ---------- ... and drain overflow list. 4667 _timer.reset(); 4668 _timer.start(); 4669 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4670 _timer.stop(); 4671 if (PrintCMSStatistics != 0) { 4672 gclog_or_tty->print_cr( 4673 "Finished work stealing in %dth thread: %3.3f sec", 4674 worker_id, _timer.seconds()); 4675 } 4676 } 4677 4678 // Note that parameter "i" is not used. 4679 void 4680 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4681 OopsInGenClosure* cl, ContiguousSpace* space, 4682 HeapWord** chunk_array, size_t chunk_top) { 4683 // Until all tasks completed: 4684 // . claim an unclaimed task 4685 // . compute region boundaries corresponding to task claimed 4686 // using chunk_array 4687 // . par_oop_iterate(cl) over that region 4688 4689 ResourceMark rm; 4690 HandleMark hm; 4691 4692 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4693 4694 uint nth_task = 0; 4695 uint n_tasks = pst->n_tasks(); 4696 4697 if (n_tasks > 0) { 4698 assert(pst->valid(), "Uninitialized use?"); 4699 HeapWord *start, *end; 4700 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4701 // We claimed task # nth_task; compute its boundaries. 4702 if (chunk_top == 0) { // no samples were taken 4703 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4704 start = space->bottom(); 4705 end = space->top(); 4706 } else if (nth_task == 0) { 4707 start = space->bottom(); 4708 end = chunk_array[nth_task]; 4709 } else if (nth_task < (uint)chunk_top) { 4710 assert(nth_task >= 1, "Control point invariant"); 4711 start = chunk_array[nth_task - 1]; 4712 end = chunk_array[nth_task]; 4713 } else { 4714 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4715 start = chunk_array[chunk_top - 1]; 4716 end = space->top(); 4717 } 4718 MemRegion mr(start, end); 4719 // Verify that mr is in space 4720 assert(mr.is_empty() || space->used_region().contains(mr), 4721 "Should be in space"); 4722 // Verify that "start" is an object boundary 4723 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4724 "Should be an oop"); 4725 space->par_oop_iterate(mr, cl); 4726 } 4727 pst->all_tasks_completed(); 4728 } 4729 } 4730 4731 void 4732 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4733 CompactibleFreeListSpace* sp, int i, 4734 Par_MarkRefsIntoAndScanClosure* cl) { 4735 // Until all tasks completed: 4736 // . claim an unclaimed task 4737 // . compute region boundaries corresponding to task claimed 4738 // . transfer dirty bits ct->mut for that region 4739 // . apply rescanclosure to dirty mut bits for that region 4740 4741 ResourceMark rm; 4742 HandleMark hm; 4743 4744 OopTaskQueue* work_q = work_queue(i); 4745 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4746 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4747 // CAUTION: This closure has state that persists across calls to 4748 // the work method dirty_range_iterate_clear() in that it has 4749 // embedded in it a (subtype of) UpwardsObjectClosure. The 4750 // use of that state in the embedded UpwardsObjectClosure instance 4751 // assumes that the cards are always iterated (even if in parallel 4752 // by several threads) in monotonically increasing order per each 4753 // thread. This is true of the implementation below which picks 4754 // card ranges (chunks) in monotonically increasing order globally 4755 // and, a-fortiori, in monotonically increasing order per thread 4756 // (the latter order being a subsequence of the former). 4757 // If the work code below is ever reorganized into a more chaotic 4758 // work-partitioning form than the current "sequential tasks" 4759 // paradigm, the use of that persistent state will have to be 4760 // revisited and modified appropriately. See also related 4761 // bug 4756801 work on which should examine this code to make 4762 // sure that the changes there do not run counter to the 4763 // assumptions made here and necessary for correctness and 4764 // efficiency. Note also that this code might yield inefficient 4765 // behavior in the case of very large objects that span one or 4766 // more work chunks. Such objects would potentially be scanned 4767 // several times redundantly. Work on 4756801 should try and 4768 // address that performance anomaly if at all possible. XXX 4769 MemRegion full_span = _collector->_span; 4770 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4771 MarkFromDirtyCardsClosure 4772 greyRescanClosure(_collector, full_span, // entire span of interest 4773 sp, bm, work_q, cl); 4774 4775 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4776 assert(pst->valid(), "Uninitialized use?"); 4777 uint nth_task = 0; 4778 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4779 MemRegion span = sp->used_region(); 4780 HeapWord* start_addr = span.start(); 4781 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4782 alignment); 4783 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4784 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4785 start_addr, "Check alignment"); 4786 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4787 chunk_size, "Check alignment"); 4788 4789 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4790 // Having claimed the nth_task, compute corresponding mem-region, 4791 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4792 // The alignment restriction ensures that we do not need any 4793 // synchronization with other gang-workers while setting or 4794 // clearing bits in thus chunk of the MUT. 4795 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4796 start_addr + (nth_task+1)*chunk_size); 4797 // The last chunk's end might be way beyond end of the 4798 // used region. In that case pull back appropriately. 4799 if (this_span.end() > end_addr) { 4800 this_span.set_end(end_addr); 4801 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4802 } 4803 // Iterate over the dirty cards covering this chunk, marking them 4804 // precleaned, and setting the corresponding bits in the mod union 4805 // table. Since we have been careful to partition at Card and MUT-word 4806 // boundaries no synchronization is needed between parallel threads. 4807 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4808 &modUnionClosure); 4809 4810 // Having transferred these marks into the modUnionTable, 4811 // rescan the marked objects on the dirty cards in the modUnionTable. 4812 // Even if this is at a synchronous collection, the initial marking 4813 // may have been done during an asynchronous collection so there 4814 // may be dirty bits in the mod-union table. 4815 _collector->_modUnionTable.dirty_range_iterate_clear( 4816 this_span, &greyRescanClosure); 4817 _collector->_modUnionTable.verifyNoOneBitsInRange( 4818 this_span.start(), 4819 this_span.end()); 4820 } 4821 pst->all_tasks_completed(); // declare that i am done 4822 } 4823 4824 // . see if we can share work_queues with ParNew? XXX 4825 void 4826 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4827 int* seed) { 4828 OopTaskQueue* work_q = work_queue(i); 4829 NOT_PRODUCT(int num_steals = 0;) 4830 oop obj_to_scan; 4831 CMSBitMap* bm = &(_collector->_markBitMap); 4832 4833 while (true) { 4834 // Completely finish any left over work from (an) earlier round(s) 4835 cl->trim_queue(0); 4836 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4837 (size_t)ParGCDesiredObjsFromOverflowList); 4838 // Now check if there's any work in the overflow list 4839 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4840 // only affects the number of attempts made to get work from the 4841 // overflow list and does not affect the number of workers. Just 4842 // pass ParallelGCThreads so this behavior is unchanged. 4843 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4844 work_q, 4845 ParallelGCThreads)) { 4846 // found something in global overflow list; 4847 // not yet ready to go stealing work from others. 4848 // We'd like to assert(work_q->size() != 0, ...) 4849 // because we just took work from the overflow list, 4850 // but of course we can't since all of that could have 4851 // been already stolen from us. 4852 // "He giveth and He taketh away." 4853 continue; 4854 } 4855 // Verify that we have no work before we resort to stealing 4856 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4857 // Try to steal from other queues that have work 4858 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4859 NOT_PRODUCT(num_steals++;) 4860 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4861 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4862 // Do scanning work 4863 obj_to_scan->oop_iterate(cl); 4864 // Loop around, finish this work, and try to steal some more 4865 } else if (terminator()->offer_termination()) { 4866 break; // nirvana from the infinite cycle 4867 } 4868 } 4869 NOT_PRODUCT( 4870 if (PrintCMSStatistics != 0) { 4871 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4872 } 4873 ) 4874 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4875 "Else our work is not yet done"); 4876 } 4877 4878 // Record object boundaries in _eden_chunk_array by sampling the eden 4879 // top in the slow-path eden object allocation code path and record 4880 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4881 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4882 // sampling in sample_eden() that activates during the part of the 4883 // preclean phase. 4884 void CMSCollector::sample_eden_chunk() { 4885 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4886 if (_eden_chunk_lock->try_lock()) { 4887 // Record a sample. This is the critical section. The contents 4888 // of the _eden_chunk_array have to be non-decreasing in the 4889 // address order. 4890 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4891 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4892 "Unexpected state of Eden"); 4893 if (_eden_chunk_index == 0 || 4894 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4895 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4896 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4897 _eden_chunk_index++; // commit sample 4898 } 4899 _eden_chunk_lock->unlock(); 4900 } 4901 } 4902 } 4903 4904 // Return a thread-local PLAB recording array, as appropriate. 4905 void* CMSCollector::get_data_recorder(int thr_num) { 4906 if (_survivor_plab_array != NULL && 4907 (CMSPLABRecordAlways || 4908 (_collectorState > Marking && _collectorState < FinalMarking))) { 4909 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4910 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4911 ca->reset(); // clear it so that fresh data is recorded 4912 return (void*) ca; 4913 } else { 4914 return NULL; 4915 } 4916 } 4917 4918 // Reset all the thread-local PLAB recording arrays 4919 void CMSCollector::reset_survivor_plab_arrays() { 4920 for (uint i = 0; i < ParallelGCThreads; i++) { 4921 _survivor_plab_array[i].reset(); 4922 } 4923 } 4924 4925 // Merge the per-thread plab arrays into the global survivor chunk 4926 // array which will provide the partitioning of the survivor space 4927 // for CMS initial scan and rescan. 4928 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4929 int no_of_gc_threads) { 4930 assert(_survivor_plab_array != NULL, "Error"); 4931 assert(_survivor_chunk_array != NULL, "Error"); 4932 assert(_collectorState == FinalMarking || 4933 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4934 for (int j = 0; j < no_of_gc_threads; j++) { 4935 _cursor[j] = 0; 4936 } 4937 HeapWord* top = surv->top(); 4938 size_t i; 4939 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4940 HeapWord* min_val = top; // Higher than any PLAB address 4941 uint min_tid = 0; // position of min_val this round 4942 for (int j = 0; j < no_of_gc_threads; j++) { 4943 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4944 if (_cursor[j] == cur_sca->end()) { 4945 continue; 4946 } 4947 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4948 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4949 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4950 if (cur_val < min_val) { 4951 min_tid = j; 4952 min_val = cur_val; 4953 } else { 4954 assert(cur_val < top, "All recorded addresses should be less"); 4955 } 4956 } 4957 // At this point min_val and min_tid are respectively 4958 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4959 // and the thread (j) that witnesses that address. 4960 // We record this address in the _survivor_chunk_array[i] 4961 // and increment _cursor[min_tid] prior to the next round i. 4962 if (min_val == top) { 4963 break; 4964 } 4965 _survivor_chunk_array[i] = min_val; 4966 _cursor[min_tid]++; 4967 } 4968 // We are all done; record the size of the _survivor_chunk_array 4969 _survivor_chunk_index = i; // exclusive: [0, i) 4970 if (PrintCMSStatistics > 0) { 4971 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4972 } 4973 // Verify that we used up all the recorded entries 4974 #ifdef ASSERT 4975 size_t total = 0; 4976 for (int j = 0; j < no_of_gc_threads; j++) { 4977 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4978 total += _cursor[j]; 4979 } 4980 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4981 // Check that the merged array is in sorted order 4982 if (total > 0) { 4983 for (size_t i = 0; i < total - 1; i++) { 4984 if (PrintCMSStatistics > 0) { 4985 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4986 i, p2i(_survivor_chunk_array[i])); 4987 } 4988 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4989 "Not sorted"); 4990 } 4991 } 4992 #endif // ASSERT 4993 } 4994 4995 // Set up the space's par_seq_tasks structure for work claiming 4996 // for parallel initial scan and rescan of young gen. 4997 // See ParRescanTask where this is currently used. 4998 void 4999 CMSCollector:: 5000 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5001 assert(n_threads > 0, "Unexpected n_threads argument"); 5002 5003 // Eden space 5004 if (!_young_gen->eden()->is_empty()) { 5005 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5006 assert(!pst->valid(), "Clobbering existing data?"); 5007 // Each valid entry in [0, _eden_chunk_index) represents a task. 5008 size_t n_tasks = _eden_chunk_index + 1; 5009 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5010 // Sets the condition for completion of the subtask (how many threads 5011 // need to finish in order to be done). 5012 pst->set_n_threads(n_threads); 5013 pst->set_n_tasks((int)n_tasks); 5014 } 5015 5016 // Merge the survivor plab arrays into _survivor_chunk_array 5017 if (_survivor_plab_array != NULL) { 5018 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5019 } else { 5020 assert(_survivor_chunk_index == 0, "Error"); 5021 } 5022 5023 // To space 5024 { 5025 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5026 assert(!pst->valid(), "Clobbering existing data?"); 5027 // Sets the condition for completion of the subtask (how many threads 5028 // need to finish in order to be done). 5029 pst->set_n_threads(n_threads); 5030 pst->set_n_tasks(1); 5031 assert(pst->valid(), "Error"); 5032 } 5033 5034 // From space 5035 { 5036 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5037 assert(!pst->valid(), "Clobbering existing data?"); 5038 size_t n_tasks = _survivor_chunk_index + 1; 5039 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5040 // Sets the condition for completion of the subtask (how many threads 5041 // need to finish in order to be done). 5042 pst->set_n_threads(n_threads); 5043 pst->set_n_tasks((int)n_tasks); 5044 assert(pst->valid(), "Error"); 5045 } 5046 } 5047 5048 // Parallel version of remark 5049 void CMSCollector::do_remark_parallel() { 5050 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5051 FlexibleWorkGang* workers = gch->workers(); 5052 assert(workers != NULL, "Need parallel worker threads."); 5053 // Choose to use the number of GC workers most recently set 5054 // into "active_workers". If active_workers is not set, set it 5055 // to ParallelGCThreads. 5056 uint n_workers = workers->active_workers(); 5057 if (n_workers == 0) { 5058 assert(n_workers > 0, "Should have been set during scavenge"); 5059 n_workers = ParallelGCThreads; 5060 workers->set_active_workers(n_workers); 5061 } 5062 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5063 5064 CMSParRemarkTask tsk(this, 5065 cms_space, 5066 n_workers, workers, task_queues()); 5067 5068 // Set up for parallel process_roots work. 5069 gch->set_par_threads(n_workers); 5070 // We won't be iterating over the cards in the card table updating 5071 // the younger_gen cards, so we shouldn't call the following else 5072 // the verification code as well as subsequent younger_refs_iterate 5073 // code would get confused. XXX 5074 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5075 5076 // The young gen rescan work will not be done as part of 5077 // process_roots (which currently doesn't know how to 5078 // parallelize such a scan), but rather will be broken up into 5079 // a set of parallel tasks (via the sampling that the [abortable] 5080 // preclean phase did of eden, plus the [two] tasks of 5081 // scanning the [two] survivor spaces. Further fine-grain 5082 // parallelization of the scanning of the survivor spaces 5083 // themselves, and of precleaning of the younger gen itself 5084 // is deferred to the future. 5085 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5086 5087 // The dirty card rescan work is broken up into a "sequence" 5088 // of parallel tasks (per constituent space) that are dynamically 5089 // claimed by the parallel threads. 5090 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5091 5092 // It turns out that even when we're using 1 thread, doing the work in a 5093 // separate thread causes wide variance in run times. We can't help this 5094 // in the multi-threaded case, but we special-case n=1 here to get 5095 // repeatable measurements of the 1-thread overhead of the parallel code. 5096 if (n_workers > 1) { 5097 // Make refs discovery MT-safe, if it isn't already: it may not 5098 // necessarily be so, since it's possible that we are doing 5099 // ST marking. 5100 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5101 StrongRootsScope srs; 5102 workers->run_task(&tsk); 5103 } else { 5104 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5105 StrongRootsScope srs; 5106 tsk.work(0); 5107 } 5108 5109 gch->set_par_threads(0); // 0 ==> non-parallel. 5110 // restore, single-threaded for now, any preserved marks 5111 // as a result of work_q overflow 5112 restore_preserved_marks_if_any(); 5113 } 5114 5115 // Non-parallel version of remark 5116 void CMSCollector::do_remark_non_parallel() { 5117 ResourceMark rm; 5118 HandleMark hm; 5119 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5120 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5121 5122 MarkRefsIntoAndScanClosure 5123 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5124 &_markStack, this, 5125 false /* should_yield */, false /* not precleaning */); 5126 MarkFromDirtyCardsClosure 5127 markFromDirtyCardsClosure(this, _span, 5128 NULL, // space is set further below 5129 &_markBitMap, &_markStack, &mrias_cl); 5130 { 5131 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5132 // Iterate over the dirty cards, setting the corresponding bits in the 5133 // mod union table. 5134 { 5135 ModUnionClosure modUnionClosure(&_modUnionTable); 5136 _ct->ct_bs()->dirty_card_iterate( 5137 _cmsGen->used_region(), 5138 &modUnionClosure); 5139 } 5140 // Having transferred these marks into the modUnionTable, we just need 5141 // to rescan the marked objects on the dirty cards in the modUnionTable. 5142 // The initial marking may have been done during an asynchronous 5143 // collection so there may be dirty bits in the mod-union table. 5144 const int alignment = 5145 CardTableModRefBS::card_size * BitsPerWord; 5146 { 5147 // ... First handle dirty cards in CMS gen 5148 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5149 MemRegion ur = _cmsGen->used_region(); 5150 HeapWord* lb = ur.start(); 5151 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5152 MemRegion cms_span(lb, ub); 5153 _modUnionTable.dirty_range_iterate_clear(cms_span, 5154 &markFromDirtyCardsClosure); 5155 verify_work_stacks_empty(); 5156 if (PrintCMSStatistics != 0) { 5157 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5158 markFromDirtyCardsClosure.num_dirty_cards()); 5159 } 5160 } 5161 } 5162 if (VerifyDuringGC && 5163 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5164 HandleMark hm; // Discard invalid handles created during verification 5165 Universe::verify(); 5166 } 5167 { 5168 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5169 5170 verify_work_stacks_empty(); 5171 5172 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5173 StrongRootsScope srs; 5174 5175 gch->gen_process_roots(_cmsGen->level(), 5176 true, // younger gens as roots 5177 false, // use the local StrongRootsScope 5178 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5179 should_unload_classes(), 5180 &mrias_cl, 5181 NULL, 5182 NULL); // The dirty klasses will be handled below 5183 5184 assert(should_unload_classes() 5185 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5186 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5187 } 5188 5189 { 5190 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5191 5192 verify_work_stacks_empty(); 5193 5194 // Scan all class loader data objects that might have been introduced 5195 // during concurrent marking. 5196 ResourceMark rm; 5197 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5198 for (int i = 0; i < array->length(); i++) { 5199 mrias_cl.do_class_loader_data(array->at(i)); 5200 } 5201 5202 // We don't need to keep track of new CLDs anymore. 5203 ClassLoaderDataGraph::remember_new_clds(false); 5204 5205 verify_work_stacks_empty(); 5206 } 5207 5208 { 5209 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5210 5211 verify_work_stacks_empty(); 5212 5213 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5214 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5215 5216 verify_work_stacks_empty(); 5217 } 5218 5219 // We might have added oops to ClassLoaderData::_handles during the 5220 // concurrent marking phase. These oops point to newly allocated objects 5221 // that are guaranteed to be kept alive. Either by the direct allocation 5222 // code, or when the young collector processes the roots. Hence, 5223 // we don't have to revisit the _handles block during the remark phase. 5224 5225 verify_work_stacks_empty(); 5226 // Restore evacuated mark words, if any, used for overflow list links 5227 if (!CMSOverflowEarlyRestoration) { 5228 restore_preserved_marks_if_any(); 5229 } 5230 verify_overflow_empty(); 5231 } 5232 5233 //////////////////////////////////////////////////////// 5234 // Parallel Reference Processing Task Proxy Class 5235 //////////////////////////////////////////////////////// 5236 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5237 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5238 CMSCollector* _collector; 5239 CMSBitMap* _mark_bit_map; 5240 const MemRegion _span; 5241 ProcessTask& _task; 5242 5243 public: 5244 CMSRefProcTaskProxy(ProcessTask& task, 5245 CMSCollector* collector, 5246 const MemRegion& span, 5247 CMSBitMap* mark_bit_map, 5248 AbstractWorkGang* workers, 5249 OopTaskQueueSet* task_queues): 5250 // XXX Should superclass AGTWOQ also know about AWG since it knows 5251 // about the task_queues used by the AWG? Then it could initialize 5252 // the terminator() object. See 6984287. The set_for_termination() 5253 // below is a temporary band-aid for the regression in 6984287. 5254 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5255 task_queues), 5256 _task(task), 5257 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5258 { 5259 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5260 "Inconsistency in _span"); 5261 set_for_termination(workers->active_workers()); 5262 } 5263 5264 OopTaskQueueSet* task_queues() { return queues(); } 5265 5266 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5267 5268 void do_work_steal(int i, 5269 CMSParDrainMarkingStackClosure* drain, 5270 CMSParKeepAliveClosure* keep_alive, 5271 int* seed); 5272 5273 virtual void work(uint worker_id); 5274 }; 5275 5276 void CMSRefProcTaskProxy::work(uint worker_id) { 5277 ResourceMark rm; 5278 HandleMark hm; 5279 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5280 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5281 _mark_bit_map, 5282 work_queue(worker_id)); 5283 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5284 _mark_bit_map, 5285 work_queue(worker_id)); 5286 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5287 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5288 if (_task.marks_oops_alive()) { 5289 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5290 _collector->hash_seed(worker_id)); 5291 } 5292 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5293 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5294 } 5295 5296 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5297 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5298 EnqueueTask& _task; 5299 5300 public: 5301 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5302 : AbstractGangTask("Enqueue reference objects in parallel"), 5303 _task(task) 5304 { } 5305 5306 virtual void work(uint worker_id) 5307 { 5308 _task.work(worker_id); 5309 } 5310 }; 5311 5312 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5313 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5314 _span(span), 5315 _bit_map(bit_map), 5316 _work_queue(work_queue), 5317 _mark_and_push(collector, span, bit_map, work_queue), 5318 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5319 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5320 { } 5321 5322 // . see if we can share work_queues with ParNew? XXX 5323 void CMSRefProcTaskProxy::do_work_steal(int i, 5324 CMSParDrainMarkingStackClosure* drain, 5325 CMSParKeepAliveClosure* keep_alive, 5326 int* seed) { 5327 OopTaskQueue* work_q = work_queue(i); 5328 NOT_PRODUCT(int num_steals = 0;) 5329 oop obj_to_scan; 5330 5331 while (true) { 5332 // Completely finish any left over work from (an) earlier round(s) 5333 drain->trim_queue(0); 5334 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5335 (size_t)ParGCDesiredObjsFromOverflowList); 5336 // Now check if there's any work in the overflow list 5337 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5338 // only affects the number of attempts made to get work from the 5339 // overflow list and does not affect the number of workers. Just 5340 // pass ParallelGCThreads so this behavior is unchanged. 5341 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5342 work_q, 5343 ParallelGCThreads)) { 5344 // Found something in global overflow list; 5345 // not yet ready to go stealing work from others. 5346 // We'd like to assert(work_q->size() != 0, ...) 5347 // because we just took work from the overflow list, 5348 // but of course we can't, since all of that might have 5349 // been already stolen from us. 5350 continue; 5351 } 5352 // Verify that we have no work before we resort to stealing 5353 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5354 // Try to steal from other queues that have work 5355 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5356 NOT_PRODUCT(num_steals++;) 5357 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5358 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5359 // Do scanning work 5360 obj_to_scan->oop_iterate(keep_alive); 5361 // Loop around, finish this work, and try to steal some more 5362 } else if (terminator()->offer_termination()) { 5363 break; // nirvana from the infinite cycle 5364 } 5365 } 5366 NOT_PRODUCT( 5367 if (PrintCMSStatistics != 0) { 5368 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5369 } 5370 ) 5371 } 5372 5373 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5374 { 5375 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5376 FlexibleWorkGang* workers = gch->workers(); 5377 assert(workers != NULL, "Need parallel worker threads."); 5378 CMSRefProcTaskProxy rp_task(task, &_collector, 5379 _collector.ref_processor()->span(), 5380 _collector.markBitMap(), 5381 workers, _collector.task_queues()); 5382 workers->run_task(&rp_task); 5383 } 5384 5385 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5386 { 5387 5388 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5389 FlexibleWorkGang* workers = gch->workers(); 5390 assert(workers != NULL, "Need parallel worker threads."); 5391 CMSRefEnqueueTaskProxy enq_task(task); 5392 workers->run_task(&enq_task); 5393 } 5394 5395 void CMSCollector::refProcessingWork() { 5396 ResourceMark rm; 5397 HandleMark hm; 5398 5399 ReferenceProcessor* rp = ref_processor(); 5400 assert(rp->span().equals(_span), "Spans should be equal"); 5401 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5402 // Process weak references. 5403 rp->setup_policy(false); 5404 verify_work_stacks_empty(); 5405 5406 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5407 &_markStack, false /* !preclean */); 5408 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5409 _span, &_markBitMap, &_markStack, 5410 &cmsKeepAliveClosure, false /* !preclean */); 5411 { 5412 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5413 5414 ReferenceProcessorStats stats; 5415 if (rp->processing_is_mt()) { 5416 // Set the degree of MT here. If the discovery is done MT, there 5417 // may have been a different number of threads doing the discovery 5418 // and a different number of discovered lists may have Ref objects. 5419 // That is OK as long as the Reference lists are balanced (see 5420 // balance_all_queues() and balance_queues()). 5421 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5422 uint active_workers = ParallelGCThreads; 5423 FlexibleWorkGang* workers = gch->workers(); 5424 if (workers != NULL) { 5425 active_workers = workers->active_workers(); 5426 // The expectation is that active_workers will have already 5427 // been set to a reasonable value. If it has not been set, 5428 // investigate. 5429 assert(active_workers > 0, "Should have been set during scavenge"); 5430 } 5431 rp->set_active_mt_degree(active_workers); 5432 CMSRefProcTaskExecutor task_executor(*this); 5433 stats = rp->process_discovered_references(&_is_alive_closure, 5434 &cmsKeepAliveClosure, 5435 &cmsDrainMarkingStackClosure, 5436 &task_executor, 5437 _gc_timer_cm, 5438 _gc_tracer_cm->gc_id()); 5439 } else { 5440 stats = rp->process_discovered_references(&_is_alive_closure, 5441 &cmsKeepAliveClosure, 5442 &cmsDrainMarkingStackClosure, 5443 NULL, 5444 _gc_timer_cm, 5445 _gc_tracer_cm->gc_id()); 5446 } 5447 _gc_tracer_cm->report_gc_reference_stats(stats); 5448 5449 } 5450 5451 // This is the point where the entire marking should have completed. 5452 verify_work_stacks_empty(); 5453 5454 if (should_unload_classes()) { 5455 { 5456 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5457 5458 // Unload classes and purge the SystemDictionary. 5459 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5460 5461 // Unload nmethods. 5462 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5463 5464 // Prune dead klasses from subklass/sibling/implementor lists. 5465 Klass::clean_weak_klass_links(&_is_alive_closure); 5466 } 5467 5468 { 5469 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5470 // Clean up unreferenced symbols in symbol table. 5471 SymbolTable::unlink(); 5472 } 5473 5474 { 5475 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5476 // Delete entries for dead interned strings. 5477 StringTable::unlink(&_is_alive_closure); 5478 } 5479 } 5480 5481 5482 // Restore any preserved marks as a result of mark stack or 5483 // work queue overflow 5484 restore_preserved_marks_if_any(); // done single-threaded for now 5485 5486 rp->set_enqueuing_is_done(true); 5487 if (rp->processing_is_mt()) { 5488 rp->balance_all_queues(); 5489 CMSRefProcTaskExecutor task_executor(*this); 5490 rp->enqueue_discovered_references(&task_executor); 5491 } else { 5492 rp->enqueue_discovered_references(NULL); 5493 } 5494 rp->verify_no_references_recorded(); 5495 assert(!rp->discovery_enabled(), "should have been disabled"); 5496 } 5497 5498 #ifndef PRODUCT 5499 void CMSCollector::check_correct_thread_executing() { 5500 Thread* t = Thread::current(); 5501 // Only the VM thread or the CMS thread should be here. 5502 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5503 "Unexpected thread type"); 5504 // If this is the vm thread, the foreground process 5505 // should not be waiting. Note that _foregroundGCIsActive is 5506 // true while the foreground collector is waiting. 5507 if (_foregroundGCShouldWait) { 5508 // We cannot be the VM thread 5509 assert(t->is_ConcurrentGC_thread(), 5510 "Should be CMS thread"); 5511 } else { 5512 // We can be the CMS thread only if we are in a stop-world 5513 // phase of CMS collection. 5514 if (t->is_ConcurrentGC_thread()) { 5515 assert(_collectorState == InitialMarking || 5516 _collectorState == FinalMarking, 5517 "Should be a stop-world phase"); 5518 // The CMS thread should be holding the CMS_token. 5519 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5520 "Potential interference with concurrently " 5521 "executing VM thread"); 5522 } 5523 } 5524 } 5525 #endif 5526 5527 void CMSCollector::sweep() { 5528 assert(_collectorState == Sweeping, "just checking"); 5529 check_correct_thread_executing(); 5530 verify_work_stacks_empty(); 5531 verify_overflow_empty(); 5532 increment_sweep_count(); 5533 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5534 5535 _inter_sweep_timer.stop(); 5536 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5537 5538 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5539 _intra_sweep_timer.reset(); 5540 _intra_sweep_timer.start(); 5541 { 5542 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5543 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5544 // First sweep the old gen 5545 { 5546 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5547 bitMapLock()); 5548 sweepWork(_cmsGen); 5549 } 5550 5551 // Update Universe::_heap_*_at_gc figures. 5552 // We need all the free list locks to make the abstract state 5553 // transition from Sweeping to Resetting. See detailed note 5554 // further below. 5555 { 5556 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5557 // Update heap occupancy information which is used as 5558 // input to soft ref clearing policy at the next gc. 5559 Universe::update_heap_info_at_gc(); 5560 _collectorState = Resizing; 5561 } 5562 } 5563 verify_work_stacks_empty(); 5564 verify_overflow_empty(); 5565 5566 if (should_unload_classes()) { 5567 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5568 // requires that the virtual spaces are stable and not deleted. 5569 ClassLoaderDataGraph::set_should_purge(true); 5570 } 5571 5572 _intra_sweep_timer.stop(); 5573 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5574 5575 _inter_sweep_timer.reset(); 5576 _inter_sweep_timer.start(); 5577 5578 // We need to use a monotonically non-decreasing time in ms 5579 // or we will see time-warp warnings and os::javaTimeMillis() 5580 // does not guarantee monotonicity. 5581 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5582 update_time_of_last_gc(now); 5583 5584 // NOTE on abstract state transitions: 5585 // Mutators allocate-live and/or mark the mod-union table dirty 5586 // based on the state of the collection. The former is done in 5587 // the interval [Marking, Sweeping] and the latter in the interval 5588 // [Marking, Sweeping). Thus the transitions into the Marking state 5589 // and out of the Sweeping state must be synchronously visible 5590 // globally to the mutators. 5591 // The transition into the Marking state happens with the world 5592 // stopped so the mutators will globally see it. Sweeping is 5593 // done asynchronously by the background collector so the transition 5594 // from the Sweeping state to the Resizing state must be done 5595 // under the freelistLock (as is the check for whether to 5596 // allocate-live and whether to dirty the mod-union table). 5597 assert(_collectorState == Resizing, "Change of collector state to" 5598 " Resizing must be done under the freelistLocks (plural)"); 5599 5600 // Now that sweeping has been completed, we clear 5601 // the incremental_collection_failed flag, 5602 // thus inviting a younger gen collection to promote into 5603 // this generation. If such a promotion may still fail, 5604 // the flag will be set again when a young collection is 5605 // attempted. 5606 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5607 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5608 gch->update_full_collections_completed(_collection_count_start); 5609 } 5610 5611 // FIX ME!!! Looks like this belongs in CFLSpace, with 5612 // CMSGen merely delegating to it. 5613 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5614 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5615 HeapWord* minAddr = _cmsSpace->bottom(); 5616 HeapWord* largestAddr = 5617 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5618 if (largestAddr == NULL) { 5619 // The dictionary appears to be empty. In this case 5620 // try to coalesce at the end of the heap. 5621 largestAddr = _cmsSpace->end(); 5622 } 5623 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5624 size_t nearLargestOffset = 5625 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5626 if (PrintFLSStatistics != 0) { 5627 gclog_or_tty->print_cr( 5628 "CMS: Large Block: " PTR_FORMAT ";" 5629 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5630 p2i(largestAddr), 5631 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5632 } 5633 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5634 } 5635 5636 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5637 return addr >= _cmsSpace->nearLargestChunk(); 5638 } 5639 5640 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5641 return _cmsSpace->find_chunk_at_end(); 5642 } 5643 5644 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5645 bool full) { 5646 // The next lower level has been collected. Gather any statistics 5647 // that are of interest at this point. 5648 if (!full && (current_level + 1) == level()) { 5649 // Gather statistics on the young generation collection. 5650 collector()->stats().record_gc0_end(used()); 5651 } 5652 } 5653 5654 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5655 // We iterate over the space(s) underlying this generation, 5656 // checking the mark bit map to see if the bits corresponding 5657 // to specific blocks are marked or not. Blocks that are 5658 // marked are live and are not swept up. All remaining blocks 5659 // are swept up, with coalescing on-the-fly as we sweep up 5660 // contiguous free and/or garbage blocks: 5661 // We need to ensure that the sweeper synchronizes with allocators 5662 // and stop-the-world collectors. In particular, the following 5663 // locks are used: 5664 // . CMS token: if this is held, a stop the world collection cannot occur 5665 // . freelistLock: if this is held no allocation can occur from this 5666 // generation by another thread 5667 // . bitMapLock: if this is held, no other thread can access or update 5668 // 5669 5670 // Note that we need to hold the freelistLock if we use 5671 // block iterate below; else the iterator might go awry if 5672 // a mutator (or promotion) causes block contents to change 5673 // (for instance if the allocator divvies up a block). 5674 // If we hold the free list lock, for all practical purposes 5675 // young generation GC's can't occur (they'll usually need to 5676 // promote), so we might as well prevent all young generation 5677 // GC's while we do a sweeping step. For the same reason, we might 5678 // as well take the bit map lock for the entire duration 5679 5680 // check that we hold the requisite locks 5681 assert(have_cms_token(), "Should hold cms token"); 5682 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5683 assert_lock_strong(gen->freelistLock()); 5684 assert_lock_strong(bitMapLock()); 5685 5686 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5687 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5688 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5689 _inter_sweep_estimate.padded_average(), 5690 _intra_sweep_estimate.padded_average()); 5691 gen->setNearLargestChunk(); 5692 5693 { 5694 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5695 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5696 // We need to free-up/coalesce garbage/blocks from a 5697 // co-terminal free run. This is done in the SweepClosure 5698 // destructor; so, do not remove this scope, else the 5699 // end-of-sweep-census below will be off by a little bit. 5700 } 5701 gen->cmsSpace()->sweep_completed(); 5702 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5703 if (should_unload_classes()) { // unloaded classes this cycle, 5704 _concurrent_cycles_since_last_unload = 0; // ... reset count 5705 } else { // did not unload classes, 5706 _concurrent_cycles_since_last_unload++; // ... increment count 5707 } 5708 } 5709 5710 // Reset CMS data structures (for now just the marking bit map) 5711 // preparatory for the next cycle. 5712 void CMSCollector::reset(bool concurrent) { 5713 if (concurrent) { 5714 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5715 5716 // If the state is not "Resetting", the foreground thread 5717 // has done a collection and the resetting. 5718 if (_collectorState != Resetting) { 5719 assert(_collectorState == Idling, "The state should only change" 5720 " because the foreground collector has finished the collection"); 5721 return; 5722 } 5723 5724 // Clear the mark bitmap (no grey objects to start with) 5725 // for the next cycle. 5726 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5727 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5728 5729 HeapWord* curAddr = _markBitMap.startWord(); 5730 while (curAddr < _markBitMap.endWord()) { 5731 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5732 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5733 _markBitMap.clear_large_range(chunk); 5734 if (ConcurrentMarkSweepThread::should_yield() && 5735 !foregroundGCIsActive() && 5736 CMSYield) { 5737 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5738 "CMS thread should hold CMS token"); 5739 assert_lock_strong(bitMapLock()); 5740 bitMapLock()->unlock(); 5741 ConcurrentMarkSweepThread::desynchronize(true); 5742 stopTimer(); 5743 if (PrintCMSStatistics != 0) { 5744 incrementYields(); 5745 } 5746 5747 // See the comment in coordinator_yield() 5748 for (unsigned i = 0; i < CMSYieldSleepCount && 5749 ConcurrentMarkSweepThread::should_yield() && 5750 !CMSCollector::foregroundGCIsActive(); ++i) { 5751 os::sleep(Thread::current(), 1, false); 5752 } 5753 5754 ConcurrentMarkSweepThread::synchronize(true); 5755 bitMapLock()->lock_without_safepoint_check(); 5756 startTimer(); 5757 } 5758 curAddr = chunk.end(); 5759 } 5760 // A successful mostly concurrent collection has been done. 5761 // Because only the full (i.e., concurrent mode failure) collections 5762 // are being measured for gc overhead limits, clean the "near" flag 5763 // and count. 5764 size_policy()->reset_gc_overhead_limit_count(); 5765 _collectorState = Idling; 5766 } else { 5767 // already have the lock 5768 assert(_collectorState == Resetting, "just checking"); 5769 assert_lock_strong(bitMapLock()); 5770 _markBitMap.clear_all(); 5771 _collectorState = Idling; 5772 } 5773 5774 register_gc_end(); 5775 } 5776 5777 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5778 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5779 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5780 TraceCollectorStats tcs(counters()); 5781 5782 switch (op) { 5783 case CMS_op_checkpointRootsInitial: { 5784 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5785 checkpointRootsInitial(); 5786 if (PrintGC) { 5787 _cmsGen->printOccupancy("initial-mark"); 5788 } 5789 break; 5790 } 5791 case CMS_op_checkpointRootsFinal: { 5792 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5793 checkpointRootsFinal(); 5794 if (PrintGC) { 5795 _cmsGen->printOccupancy("remark"); 5796 } 5797 break; 5798 } 5799 default: 5800 fatal("No such CMS_op"); 5801 } 5802 } 5803 5804 #ifndef PRODUCT 5805 size_t const CMSCollector::skip_header_HeapWords() { 5806 return FreeChunk::header_size(); 5807 } 5808 5809 // Try and collect here conditions that should hold when 5810 // CMS thread is exiting. The idea is that the foreground GC 5811 // thread should not be blocked if it wants to terminate 5812 // the CMS thread and yet continue to run the VM for a while 5813 // after that. 5814 void CMSCollector::verify_ok_to_terminate() const { 5815 assert(Thread::current()->is_ConcurrentGC_thread(), 5816 "should be called by CMS thread"); 5817 assert(!_foregroundGCShouldWait, "should be false"); 5818 // We could check here that all the various low-level locks 5819 // are not held by the CMS thread, but that is overkill; see 5820 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5821 // is checked. 5822 } 5823 #endif 5824 5825 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5826 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5827 "missing Printezis mark?"); 5828 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5829 size_t size = pointer_delta(nextOneAddr + 1, addr); 5830 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5831 "alignment problem"); 5832 assert(size >= 3, "Necessary for Printezis marks to work"); 5833 return size; 5834 } 5835 5836 // A variant of the above (block_size_using_printezis_bits()) except 5837 // that we return 0 if the P-bits are not yet set. 5838 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5839 if (_markBitMap.isMarked(addr + 1)) { 5840 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5841 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5842 size_t size = pointer_delta(nextOneAddr + 1, addr); 5843 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5844 "alignment problem"); 5845 assert(size >= 3, "Necessary for Printezis marks to work"); 5846 return size; 5847 } 5848 return 0; 5849 } 5850 5851 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5852 size_t sz = 0; 5853 oop p = (oop)addr; 5854 if (p->klass_or_null() != NULL) { 5855 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5856 } else { 5857 sz = block_size_using_printezis_bits(addr); 5858 } 5859 assert(sz > 0, "size must be nonzero"); 5860 HeapWord* next_block = addr + sz; 5861 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5862 CardTableModRefBS::card_size); 5863 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5864 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5865 "must be different cards"); 5866 return next_card; 5867 } 5868 5869 5870 // CMS Bit Map Wrapper ///////////////////////////////////////// 5871 5872 // Construct a CMS bit map infrastructure, but don't create the 5873 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5874 // further below. 5875 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5876 _bm(), 5877 _shifter(shifter), 5878 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5879 Monitor::_safepoint_check_sometimes) : NULL) 5880 { 5881 _bmStartWord = 0; 5882 _bmWordSize = 0; 5883 } 5884 5885 bool CMSBitMap::allocate(MemRegion mr) { 5886 _bmStartWord = mr.start(); 5887 _bmWordSize = mr.word_size(); 5888 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5889 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5890 if (!brs.is_reserved()) { 5891 warning("CMS bit map allocation failure"); 5892 return false; 5893 } 5894 // For now we'll just commit all of the bit map up front. 5895 // Later on we'll try to be more parsimonious with swap. 5896 if (!_virtual_space.initialize(brs, brs.size())) { 5897 warning("CMS bit map backing store failure"); 5898 return false; 5899 } 5900 assert(_virtual_space.committed_size() == brs.size(), 5901 "didn't reserve backing store for all of CMS bit map?"); 5902 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5903 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5904 _bmWordSize, "inconsistency in bit map sizing"); 5905 _bm.set_size(_bmWordSize >> _shifter); 5906 5907 // bm.clear(); // can we rely on getting zero'd memory? verify below 5908 assert(isAllClear(), 5909 "Expected zero'd memory from ReservedSpace constructor"); 5910 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5911 "consistency check"); 5912 return true; 5913 } 5914 5915 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5916 HeapWord *next_addr, *end_addr, *last_addr; 5917 assert_locked(); 5918 assert(covers(mr), "out-of-range error"); 5919 // XXX assert that start and end are appropriately aligned 5920 for (next_addr = mr.start(), end_addr = mr.end(); 5921 next_addr < end_addr; next_addr = last_addr) { 5922 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5923 last_addr = dirty_region.end(); 5924 if (!dirty_region.is_empty()) { 5925 cl->do_MemRegion(dirty_region); 5926 } else { 5927 assert(last_addr == end_addr, "program logic"); 5928 return; 5929 } 5930 } 5931 } 5932 5933 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5934 _bm.print_on_error(st, prefix); 5935 } 5936 5937 #ifndef PRODUCT 5938 void CMSBitMap::assert_locked() const { 5939 CMSLockVerifier::assert_locked(lock()); 5940 } 5941 5942 bool CMSBitMap::covers(MemRegion mr) const { 5943 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5944 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5945 "size inconsistency"); 5946 return (mr.start() >= _bmStartWord) && 5947 (mr.end() <= endWord()); 5948 } 5949 5950 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5951 return (start >= _bmStartWord && (start + size) <= endWord()); 5952 } 5953 5954 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5955 // verify that there are no 1 bits in the interval [left, right) 5956 FalseBitMapClosure falseBitMapClosure; 5957 iterate(&falseBitMapClosure, left, right); 5958 } 5959 5960 void CMSBitMap::region_invariant(MemRegion mr) 5961 { 5962 assert_locked(); 5963 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5964 assert(!mr.is_empty(), "unexpected empty region"); 5965 assert(covers(mr), "mr should be covered by bit map"); 5966 // convert address range into offset range 5967 size_t start_ofs = heapWordToOffset(mr.start()); 5968 // Make sure that end() is appropriately aligned 5969 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5970 (1 << (_shifter+LogHeapWordSize))), 5971 "Misaligned mr.end()"); 5972 size_t end_ofs = heapWordToOffset(mr.end()); 5973 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5974 } 5975 5976 #endif 5977 5978 bool CMSMarkStack::allocate(size_t size) { 5979 // allocate a stack of the requisite depth 5980 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5981 size * sizeof(oop))); 5982 if (!rs.is_reserved()) { 5983 warning("CMSMarkStack allocation failure"); 5984 return false; 5985 } 5986 if (!_virtual_space.initialize(rs, rs.size())) { 5987 warning("CMSMarkStack backing store failure"); 5988 return false; 5989 } 5990 assert(_virtual_space.committed_size() == rs.size(), 5991 "didn't reserve backing store for all of CMS stack?"); 5992 _base = (oop*)(_virtual_space.low()); 5993 _index = 0; 5994 _capacity = size; 5995 NOT_PRODUCT(_max_depth = 0); 5996 return true; 5997 } 5998 5999 // XXX FIX ME !!! In the MT case we come in here holding a 6000 // leaf lock. For printing we need to take a further lock 6001 // which has lower rank. We need to recalibrate the two 6002 // lock-ranks involved in order to be able to print the 6003 // messages below. (Or defer the printing to the caller. 6004 // For now we take the expedient path of just disabling the 6005 // messages for the problematic case.) 6006 void CMSMarkStack::expand() { 6007 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6008 if (_capacity == MarkStackSizeMax) { 6009 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6010 // We print a warning message only once per CMS cycle. 6011 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6012 } 6013 return; 6014 } 6015 // Double capacity if possible 6016 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6017 // Do not give up existing stack until we have managed to 6018 // get the double capacity that we desired. 6019 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6020 new_capacity * sizeof(oop))); 6021 if (rs.is_reserved()) { 6022 // Release the backing store associated with old stack 6023 _virtual_space.release(); 6024 // Reinitialize virtual space for new stack 6025 if (!_virtual_space.initialize(rs, rs.size())) { 6026 fatal("Not enough swap for expanded marking stack"); 6027 } 6028 _base = (oop*)(_virtual_space.low()); 6029 _index = 0; 6030 _capacity = new_capacity; 6031 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6032 // Failed to double capacity, continue; 6033 // we print a detail message only once per CMS cycle. 6034 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6035 SIZE_FORMAT"K", 6036 _capacity / K, new_capacity / K); 6037 } 6038 } 6039 6040 6041 // Closures 6042 // XXX: there seems to be a lot of code duplication here; 6043 // should refactor and consolidate common code. 6044 6045 // This closure is used to mark refs into the CMS generation in 6046 // the CMS bit map. Called at the first checkpoint. This closure 6047 // assumes that we do not need to re-mark dirty cards; if the CMS 6048 // generation on which this is used is not an oldest 6049 // generation then this will lose younger_gen cards! 6050 6051 MarkRefsIntoClosure::MarkRefsIntoClosure( 6052 MemRegion span, CMSBitMap* bitMap): 6053 _span(span), 6054 _bitMap(bitMap) 6055 { 6056 assert(_ref_processor == NULL, "deliberately left NULL"); 6057 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6058 } 6059 6060 void MarkRefsIntoClosure::do_oop(oop obj) { 6061 // if p points into _span, then mark corresponding bit in _markBitMap 6062 assert(obj->is_oop(), "expected an oop"); 6063 HeapWord* addr = (HeapWord*)obj; 6064 if (_span.contains(addr)) { 6065 // this should be made more efficient 6066 _bitMap->mark(addr); 6067 } 6068 } 6069 6070 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6071 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6072 6073 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6074 MemRegion span, CMSBitMap* bitMap): 6075 _span(span), 6076 _bitMap(bitMap) 6077 { 6078 assert(_ref_processor == NULL, "deliberately left NULL"); 6079 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6080 } 6081 6082 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6083 // if p points into _span, then mark corresponding bit in _markBitMap 6084 assert(obj->is_oop(), "expected an oop"); 6085 HeapWord* addr = (HeapWord*)obj; 6086 if (_span.contains(addr)) { 6087 // this should be made more efficient 6088 _bitMap->par_mark(addr); 6089 } 6090 } 6091 6092 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6093 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6094 6095 // A variant of the above, used for CMS marking verification. 6096 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6097 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6098 _span(span), 6099 _verification_bm(verification_bm), 6100 _cms_bm(cms_bm) 6101 { 6102 assert(_ref_processor == NULL, "deliberately left NULL"); 6103 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6104 } 6105 6106 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6107 // if p points into _span, then mark corresponding bit in _markBitMap 6108 assert(obj->is_oop(), "expected an oop"); 6109 HeapWord* addr = (HeapWord*)obj; 6110 if (_span.contains(addr)) { 6111 _verification_bm->mark(addr); 6112 if (!_cms_bm->isMarked(addr)) { 6113 oop(addr)->print(); 6114 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6115 fatal("... aborting"); 6116 } 6117 } 6118 } 6119 6120 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6121 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6122 6123 ////////////////////////////////////////////////// 6124 // MarkRefsIntoAndScanClosure 6125 ////////////////////////////////////////////////// 6126 6127 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6128 ReferenceProcessor* rp, 6129 CMSBitMap* bit_map, 6130 CMSBitMap* mod_union_table, 6131 CMSMarkStack* mark_stack, 6132 CMSCollector* collector, 6133 bool should_yield, 6134 bool concurrent_precleaning): 6135 _collector(collector), 6136 _span(span), 6137 _bit_map(bit_map), 6138 _mark_stack(mark_stack), 6139 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6140 mark_stack, concurrent_precleaning), 6141 _yield(should_yield), 6142 _concurrent_precleaning(concurrent_precleaning), 6143 _freelistLock(NULL) 6144 { 6145 _ref_processor = rp; 6146 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6147 } 6148 6149 // This closure is used to mark refs into the CMS generation at the 6150 // second (final) checkpoint, and to scan and transitively follow 6151 // the unmarked oops. It is also used during the concurrent precleaning 6152 // phase while scanning objects on dirty cards in the CMS generation. 6153 // The marks are made in the marking bit map and the marking stack is 6154 // used for keeping the (newly) grey objects during the scan. 6155 // The parallel version (Par_...) appears further below. 6156 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6157 if (obj != NULL) { 6158 assert(obj->is_oop(), "expected an oop"); 6159 HeapWord* addr = (HeapWord*)obj; 6160 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6161 assert(_collector->overflow_list_is_empty(), 6162 "overflow list should be empty"); 6163 if (_span.contains(addr) && 6164 !_bit_map->isMarked(addr)) { 6165 // mark bit map (object is now grey) 6166 _bit_map->mark(addr); 6167 // push on marking stack (stack should be empty), and drain the 6168 // stack by applying this closure to the oops in the oops popped 6169 // from the stack (i.e. blacken the grey objects) 6170 bool res = _mark_stack->push(obj); 6171 assert(res, "Should have space to push on empty stack"); 6172 do { 6173 oop new_oop = _mark_stack->pop(); 6174 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6175 assert(_bit_map->isMarked((HeapWord*)new_oop), 6176 "only grey objects on this stack"); 6177 // iterate over the oops in this oop, marking and pushing 6178 // the ones in CMS heap (i.e. in _span). 6179 new_oop->oop_iterate(&_pushAndMarkClosure); 6180 // check if it's time to yield 6181 do_yield_check(); 6182 } while (!_mark_stack->isEmpty() || 6183 (!_concurrent_precleaning && take_from_overflow_list())); 6184 // if marking stack is empty, and we are not doing this 6185 // during precleaning, then check the overflow list 6186 } 6187 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6188 assert(_collector->overflow_list_is_empty(), 6189 "overflow list was drained above"); 6190 // We could restore evacuated mark words, if any, used for 6191 // overflow list links here because the overflow list is 6192 // provably empty here. That would reduce the maximum 6193 // size requirements for preserved_{oop,mark}_stack. 6194 // But we'll just postpone it until we are all done 6195 // so we can just stream through. 6196 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6197 _collector->restore_preserved_marks_if_any(); 6198 assert(_collector->no_preserved_marks(), "No preserved marks"); 6199 } 6200 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6201 "All preserved marks should have been restored above"); 6202 } 6203 } 6204 6205 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6206 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6207 6208 void MarkRefsIntoAndScanClosure::do_yield_work() { 6209 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6210 "CMS thread should hold CMS token"); 6211 assert_lock_strong(_freelistLock); 6212 assert_lock_strong(_bit_map->lock()); 6213 // relinquish the free_list_lock and bitMaplock() 6214 _bit_map->lock()->unlock(); 6215 _freelistLock->unlock(); 6216 ConcurrentMarkSweepThread::desynchronize(true); 6217 _collector->stopTimer(); 6218 if (PrintCMSStatistics != 0) { 6219 _collector->incrementYields(); 6220 } 6221 6222 // See the comment in coordinator_yield() 6223 for (unsigned i = 0; 6224 i < CMSYieldSleepCount && 6225 ConcurrentMarkSweepThread::should_yield() && 6226 !CMSCollector::foregroundGCIsActive(); 6227 ++i) { 6228 os::sleep(Thread::current(), 1, false); 6229 } 6230 6231 ConcurrentMarkSweepThread::synchronize(true); 6232 _freelistLock->lock_without_safepoint_check(); 6233 _bit_map->lock()->lock_without_safepoint_check(); 6234 _collector->startTimer(); 6235 } 6236 6237 /////////////////////////////////////////////////////////// 6238 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6239 // MarkRefsIntoAndScanClosure 6240 /////////////////////////////////////////////////////////// 6241 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6242 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6243 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6244 _span(span), 6245 _bit_map(bit_map), 6246 _work_queue(work_queue), 6247 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6248 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6249 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6250 { 6251 _ref_processor = rp; 6252 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6253 } 6254 6255 // This closure is used to mark refs into the CMS generation at the 6256 // second (final) checkpoint, and to scan and transitively follow 6257 // the unmarked oops. The marks are made in the marking bit map and 6258 // the work_queue is used for keeping the (newly) grey objects during 6259 // the scan phase whence they are also available for stealing by parallel 6260 // threads. Since the marking bit map is shared, updates are 6261 // synchronized (via CAS). 6262 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6263 if (obj != NULL) { 6264 // Ignore mark word because this could be an already marked oop 6265 // that may be chained at the end of the overflow list. 6266 assert(obj->is_oop(true), "expected an oop"); 6267 HeapWord* addr = (HeapWord*)obj; 6268 if (_span.contains(addr) && 6269 !_bit_map->isMarked(addr)) { 6270 // mark bit map (object will become grey): 6271 // It is possible for several threads to be 6272 // trying to "claim" this object concurrently; 6273 // the unique thread that succeeds in marking the 6274 // object first will do the subsequent push on 6275 // to the work queue (or overflow list). 6276 if (_bit_map->par_mark(addr)) { 6277 // push on work_queue (which may not be empty), and trim the 6278 // queue to an appropriate length by applying this closure to 6279 // the oops in the oops popped from the stack (i.e. blacken the 6280 // grey objects) 6281 bool res = _work_queue->push(obj); 6282 assert(res, "Low water mark should be less than capacity?"); 6283 trim_queue(_low_water_mark); 6284 } // Else, another thread claimed the object 6285 } 6286 } 6287 } 6288 6289 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6290 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6291 6292 // This closure is used to rescan the marked objects on the dirty cards 6293 // in the mod union table and the card table proper. 6294 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6295 oop p, MemRegion mr) { 6296 6297 size_t size = 0; 6298 HeapWord* addr = (HeapWord*)p; 6299 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6300 assert(_span.contains(addr), "we are scanning the CMS generation"); 6301 // check if it's time to yield 6302 if (do_yield_check()) { 6303 // We yielded for some foreground stop-world work, 6304 // and we have been asked to abort this ongoing preclean cycle. 6305 return 0; 6306 } 6307 if (_bitMap->isMarked(addr)) { 6308 // it's marked; is it potentially uninitialized? 6309 if (p->klass_or_null() != NULL) { 6310 // an initialized object; ignore mark word in verification below 6311 // since we are running concurrent with mutators 6312 assert(p->is_oop(true), "should be an oop"); 6313 if (p->is_objArray()) { 6314 // objArrays are precisely marked; restrict scanning 6315 // to dirty cards only. 6316 size = CompactibleFreeListSpace::adjustObjectSize( 6317 p->oop_iterate(_scanningClosure, mr)); 6318 } else { 6319 // A non-array may have been imprecisely marked; we need 6320 // to scan object in its entirety. 6321 size = CompactibleFreeListSpace::adjustObjectSize( 6322 p->oop_iterate(_scanningClosure)); 6323 } 6324 #ifdef ASSERT 6325 size_t direct_size = 6326 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6327 assert(size == direct_size, "Inconsistency in size"); 6328 assert(size >= 3, "Necessary for Printezis marks to work"); 6329 if (!_bitMap->isMarked(addr+1)) { 6330 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6331 } else { 6332 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6333 assert(_bitMap->isMarked(addr+size-1), 6334 "inconsistent Printezis mark"); 6335 } 6336 #endif // ASSERT 6337 } else { 6338 // An uninitialized object. 6339 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6340 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6341 size = pointer_delta(nextOneAddr + 1, addr); 6342 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6343 "alignment problem"); 6344 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6345 // will dirty the card when the klass pointer is installed in the 6346 // object (signaling the completion of initialization). 6347 } 6348 } else { 6349 // Either a not yet marked object or an uninitialized object 6350 if (p->klass_or_null() == NULL) { 6351 // An uninitialized object, skip to the next card, since 6352 // we may not be able to read its P-bits yet. 6353 assert(size == 0, "Initial value"); 6354 } else { 6355 // An object not (yet) reached by marking: we merely need to 6356 // compute its size so as to go look at the next block. 6357 assert(p->is_oop(true), "should be an oop"); 6358 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6359 } 6360 } 6361 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6362 return size; 6363 } 6364 6365 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6366 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6367 "CMS thread should hold CMS token"); 6368 assert_lock_strong(_freelistLock); 6369 assert_lock_strong(_bitMap->lock()); 6370 // relinquish the free_list_lock and bitMaplock() 6371 _bitMap->lock()->unlock(); 6372 _freelistLock->unlock(); 6373 ConcurrentMarkSweepThread::desynchronize(true); 6374 _collector->stopTimer(); 6375 if (PrintCMSStatistics != 0) { 6376 _collector->incrementYields(); 6377 } 6378 6379 // See the comment in coordinator_yield() 6380 for (unsigned i = 0; i < CMSYieldSleepCount && 6381 ConcurrentMarkSweepThread::should_yield() && 6382 !CMSCollector::foregroundGCIsActive(); ++i) { 6383 os::sleep(Thread::current(), 1, false); 6384 } 6385 6386 ConcurrentMarkSweepThread::synchronize(true); 6387 _freelistLock->lock_without_safepoint_check(); 6388 _bitMap->lock()->lock_without_safepoint_check(); 6389 _collector->startTimer(); 6390 } 6391 6392 6393 ////////////////////////////////////////////////////////////////// 6394 // SurvivorSpacePrecleanClosure 6395 ////////////////////////////////////////////////////////////////// 6396 // This (single-threaded) closure is used to preclean the oops in 6397 // the survivor spaces. 6398 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6399 6400 HeapWord* addr = (HeapWord*)p; 6401 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6402 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6403 assert(p->klass_or_null() != NULL, "object should be initialized"); 6404 // an initialized object; ignore mark word in verification below 6405 // since we are running concurrent with mutators 6406 assert(p->is_oop(true), "should be an oop"); 6407 // Note that we do not yield while we iterate over 6408 // the interior oops of p, pushing the relevant ones 6409 // on our marking stack. 6410 size_t size = p->oop_iterate(_scanning_closure); 6411 do_yield_check(); 6412 // Observe that below, we do not abandon the preclean 6413 // phase as soon as we should; rather we empty the 6414 // marking stack before returning. This is to satisfy 6415 // some existing assertions. In general, it may be a 6416 // good idea to abort immediately and complete the marking 6417 // from the grey objects at a later time. 6418 while (!_mark_stack->isEmpty()) { 6419 oop new_oop = _mark_stack->pop(); 6420 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6421 assert(_bit_map->isMarked((HeapWord*)new_oop), 6422 "only grey objects on this stack"); 6423 // iterate over the oops in this oop, marking and pushing 6424 // the ones in CMS heap (i.e. in _span). 6425 new_oop->oop_iterate(_scanning_closure); 6426 // check if it's time to yield 6427 do_yield_check(); 6428 } 6429 unsigned int after_count = 6430 GenCollectedHeap::heap()->total_collections(); 6431 bool abort = (_before_count != after_count) || 6432 _collector->should_abort_preclean(); 6433 return abort ? 0 : size; 6434 } 6435 6436 void SurvivorSpacePrecleanClosure::do_yield_work() { 6437 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6438 "CMS thread should hold CMS token"); 6439 assert_lock_strong(_bit_map->lock()); 6440 // Relinquish the bit map lock 6441 _bit_map->lock()->unlock(); 6442 ConcurrentMarkSweepThread::desynchronize(true); 6443 _collector->stopTimer(); 6444 if (PrintCMSStatistics != 0) { 6445 _collector->incrementYields(); 6446 } 6447 6448 // See the comment in coordinator_yield() 6449 for (unsigned i = 0; i < CMSYieldSleepCount && 6450 ConcurrentMarkSweepThread::should_yield() && 6451 !CMSCollector::foregroundGCIsActive(); ++i) { 6452 os::sleep(Thread::current(), 1, false); 6453 } 6454 6455 ConcurrentMarkSweepThread::synchronize(true); 6456 _bit_map->lock()->lock_without_safepoint_check(); 6457 _collector->startTimer(); 6458 } 6459 6460 // This closure is used to rescan the marked objects on the dirty cards 6461 // in the mod union table and the card table proper. In the parallel 6462 // case, although the bitMap is shared, we do a single read so the 6463 // isMarked() query is "safe". 6464 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6465 // Ignore mark word because we are running concurrent with mutators 6466 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6467 HeapWord* addr = (HeapWord*)p; 6468 assert(_span.contains(addr), "we are scanning the CMS generation"); 6469 bool is_obj_array = false; 6470 #ifdef ASSERT 6471 if (!_parallel) { 6472 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6473 assert(_collector->overflow_list_is_empty(), 6474 "overflow list should be empty"); 6475 6476 } 6477 #endif // ASSERT 6478 if (_bit_map->isMarked(addr)) { 6479 // Obj arrays are precisely marked, non-arrays are not; 6480 // so we scan objArrays precisely and non-arrays in their 6481 // entirety. 6482 if (p->is_objArray()) { 6483 is_obj_array = true; 6484 if (_parallel) { 6485 p->oop_iterate(_par_scan_closure, mr); 6486 } else { 6487 p->oop_iterate(_scan_closure, mr); 6488 } 6489 } else { 6490 if (_parallel) { 6491 p->oop_iterate(_par_scan_closure); 6492 } else { 6493 p->oop_iterate(_scan_closure); 6494 } 6495 } 6496 } 6497 #ifdef ASSERT 6498 if (!_parallel) { 6499 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6500 assert(_collector->overflow_list_is_empty(), 6501 "overflow list should be empty"); 6502 6503 } 6504 #endif // ASSERT 6505 return is_obj_array; 6506 } 6507 6508 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6509 MemRegion span, 6510 CMSBitMap* bitMap, CMSMarkStack* markStack, 6511 bool should_yield, bool verifying): 6512 _collector(collector), 6513 _span(span), 6514 _bitMap(bitMap), 6515 _mut(&collector->_modUnionTable), 6516 _markStack(markStack), 6517 _yield(should_yield), 6518 _skipBits(0) 6519 { 6520 assert(_markStack->isEmpty(), "stack should be empty"); 6521 _finger = _bitMap->startWord(); 6522 _threshold = _finger; 6523 assert(_collector->_restart_addr == NULL, "Sanity check"); 6524 assert(_span.contains(_finger), "Out of bounds _finger?"); 6525 DEBUG_ONLY(_verifying = verifying;) 6526 } 6527 6528 void MarkFromRootsClosure::reset(HeapWord* addr) { 6529 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6530 assert(_span.contains(addr), "Out of bounds _finger?"); 6531 _finger = addr; 6532 _threshold = (HeapWord*)round_to( 6533 (intptr_t)_finger, CardTableModRefBS::card_size); 6534 } 6535 6536 // Should revisit to see if this should be restructured for 6537 // greater efficiency. 6538 bool MarkFromRootsClosure::do_bit(size_t offset) { 6539 if (_skipBits > 0) { 6540 _skipBits--; 6541 return true; 6542 } 6543 // convert offset into a HeapWord* 6544 HeapWord* addr = _bitMap->startWord() + offset; 6545 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6546 "address out of range"); 6547 assert(_bitMap->isMarked(addr), "tautology"); 6548 if (_bitMap->isMarked(addr+1)) { 6549 // this is an allocated but not yet initialized object 6550 assert(_skipBits == 0, "tautology"); 6551 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6552 oop p = oop(addr); 6553 if (p->klass_or_null() == NULL) { 6554 DEBUG_ONLY(if (!_verifying) {) 6555 // We re-dirty the cards on which this object lies and increase 6556 // the _threshold so that we'll come back to scan this object 6557 // during the preclean or remark phase. (CMSCleanOnEnter) 6558 if (CMSCleanOnEnter) { 6559 size_t sz = _collector->block_size_using_printezis_bits(addr); 6560 HeapWord* end_card_addr = (HeapWord*)round_to( 6561 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6562 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6563 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6564 // Bump _threshold to end_card_addr; note that 6565 // _threshold cannot possibly exceed end_card_addr, anyhow. 6566 // This prevents future clearing of the card as the scan proceeds 6567 // to the right. 6568 assert(_threshold <= end_card_addr, 6569 "Because we are just scanning into this object"); 6570 if (_threshold < end_card_addr) { 6571 _threshold = end_card_addr; 6572 } 6573 if (p->klass_or_null() != NULL) { 6574 // Redirty the range of cards... 6575 _mut->mark_range(redirty_range); 6576 } // ...else the setting of klass will dirty the card anyway. 6577 } 6578 DEBUG_ONLY(}) 6579 return true; 6580 } 6581 } 6582 scanOopsInOop(addr); 6583 return true; 6584 } 6585 6586 // We take a break if we've been at this for a while, 6587 // so as to avoid monopolizing the locks involved. 6588 void MarkFromRootsClosure::do_yield_work() { 6589 // First give up the locks, then yield, then re-lock 6590 // We should probably use a constructor/destructor idiom to 6591 // do this unlock/lock or modify the MutexUnlocker class to 6592 // serve our purpose. XXX 6593 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6594 "CMS thread should hold CMS token"); 6595 assert_lock_strong(_bitMap->lock()); 6596 _bitMap->lock()->unlock(); 6597 ConcurrentMarkSweepThread::desynchronize(true); 6598 _collector->stopTimer(); 6599 if (PrintCMSStatistics != 0) { 6600 _collector->incrementYields(); 6601 } 6602 6603 // See the comment in coordinator_yield() 6604 for (unsigned i = 0; i < CMSYieldSleepCount && 6605 ConcurrentMarkSweepThread::should_yield() && 6606 !CMSCollector::foregroundGCIsActive(); ++i) { 6607 os::sleep(Thread::current(), 1, false); 6608 } 6609 6610 ConcurrentMarkSweepThread::synchronize(true); 6611 _bitMap->lock()->lock_without_safepoint_check(); 6612 _collector->startTimer(); 6613 } 6614 6615 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6616 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6617 assert(_markStack->isEmpty(), 6618 "should drain stack to limit stack usage"); 6619 // convert ptr to an oop preparatory to scanning 6620 oop obj = oop(ptr); 6621 // Ignore mark word in verification below, since we 6622 // may be running concurrent with mutators. 6623 assert(obj->is_oop(true), "should be an oop"); 6624 assert(_finger <= ptr, "_finger runneth ahead"); 6625 // advance the finger to right end of this object 6626 _finger = ptr + obj->size(); 6627 assert(_finger > ptr, "we just incremented it above"); 6628 // On large heaps, it may take us some time to get through 6629 // the marking phase. During 6630 // this time it's possible that a lot of mutations have 6631 // accumulated in the card table and the mod union table -- 6632 // these mutation records are redundant until we have 6633 // actually traced into the corresponding card. 6634 // Here, we check whether advancing the finger would make 6635 // us cross into a new card, and if so clear corresponding 6636 // cards in the MUT (preclean them in the card-table in the 6637 // future). 6638 6639 DEBUG_ONLY(if (!_verifying) {) 6640 // The clean-on-enter optimization is disabled by default, 6641 // until we fix 6178663. 6642 if (CMSCleanOnEnter && (_finger > _threshold)) { 6643 // [_threshold, _finger) represents the interval 6644 // of cards to be cleared in MUT (or precleaned in card table). 6645 // The set of cards to be cleared is all those that overlap 6646 // with the interval [_threshold, _finger); note that 6647 // _threshold is always kept card-aligned but _finger isn't 6648 // always card-aligned. 6649 HeapWord* old_threshold = _threshold; 6650 assert(old_threshold == (HeapWord*)round_to( 6651 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6652 "_threshold should always be card-aligned"); 6653 _threshold = (HeapWord*)round_to( 6654 (intptr_t)_finger, CardTableModRefBS::card_size); 6655 MemRegion mr(old_threshold, _threshold); 6656 assert(!mr.is_empty(), "Control point invariant"); 6657 assert(_span.contains(mr), "Should clear within span"); 6658 _mut->clear_range(mr); 6659 } 6660 DEBUG_ONLY(}) 6661 // Note: the finger doesn't advance while we drain 6662 // the stack below. 6663 PushOrMarkClosure pushOrMarkClosure(_collector, 6664 _span, _bitMap, _markStack, 6665 _finger, this); 6666 bool res = _markStack->push(obj); 6667 assert(res, "Empty non-zero size stack should have space for single push"); 6668 while (!_markStack->isEmpty()) { 6669 oop new_oop = _markStack->pop(); 6670 // Skip verifying header mark word below because we are 6671 // running concurrent with mutators. 6672 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6673 // now scan this oop's oops 6674 new_oop->oop_iterate(&pushOrMarkClosure); 6675 do_yield_check(); 6676 } 6677 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6678 } 6679 6680 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6681 CMSCollector* collector, MemRegion span, 6682 CMSBitMap* bit_map, 6683 OopTaskQueue* work_queue, 6684 CMSMarkStack* overflow_stack): 6685 _collector(collector), 6686 _whole_span(collector->_span), 6687 _span(span), 6688 _bit_map(bit_map), 6689 _mut(&collector->_modUnionTable), 6690 _work_queue(work_queue), 6691 _overflow_stack(overflow_stack), 6692 _skip_bits(0), 6693 _task(task) 6694 { 6695 assert(_work_queue->size() == 0, "work_queue should be empty"); 6696 _finger = span.start(); 6697 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6698 assert(_span.contains(_finger), "Out of bounds _finger?"); 6699 } 6700 6701 // Should revisit to see if this should be restructured for 6702 // greater efficiency. 6703 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6704 if (_skip_bits > 0) { 6705 _skip_bits--; 6706 return true; 6707 } 6708 // convert offset into a HeapWord* 6709 HeapWord* addr = _bit_map->startWord() + offset; 6710 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6711 "address out of range"); 6712 assert(_bit_map->isMarked(addr), "tautology"); 6713 if (_bit_map->isMarked(addr+1)) { 6714 // this is an allocated object that might not yet be initialized 6715 assert(_skip_bits == 0, "tautology"); 6716 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6717 oop p = oop(addr); 6718 if (p->klass_or_null() == NULL) { 6719 // in the case of Clean-on-Enter optimization, redirty card 6720 // and avoid clearing card by increasing the threshold. 6721 return true; 6722 } 6723 } 6724 scan_oops_in_oop(addr); 6725 return true; 6726 } 6727 6728 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6729 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6730 // Should we assert that our work queue is empty or 6731 // below some drain limit? 6732 assert(_work_queue->size() == 0, 6733 "should drain stack to limit stack usage"); 6734 // convert ptr to an oop preparatory to scanning 6735 oop obj = oop(ptr); 6736 // Ignore mark word in verification below, since we 6737 // may be running concurrent with mutators. 6738 assert(obj->is_oop(true), "should be an oop"); 6739 assert(_finger <= ptr, "_finger runneth ahead"); 6740 // advance the finger to right end of this object 6741 _finger = ptr + obj->size(); 6742 assert(_finger > ptr, "we just incremented it above"); 6743 // On large heaps, it may take us some time to get through 6744 // the marking phase. During 6745 // this time it's possible that a lot of mutations have 6746 // accumulated in the card table and the mod union table -- 6747 // these mutation records are redundant until we have 6748 // actually traced into the corresponding card. 6749 // Here, we check whether advancing the finger would make 6750 // us cross into a new card, and if so clear corresponding 6751 // cards in the MUT (preclean them in the card-table in the 6752 // future). 6753 6754 // The clean-on-enter optimization is disabled by default, 6755 // until we fix 6178663. 6756 if (CMSCleanOnEnter && (_finger > _threshold)) { 6757 // [_threshold, _finger) represents the interval 6758 // of cards to be cleared in MUT (or precleaned in card table). 6759 // The set of cards to be cleared is all those that overlap 6760 // with the interval [_threshold, _finger); note that 6761 // _threshold is always kept card-aligned but _finger isn't 6762 // always card-aligned. 6763 HeapWord* old_threshold = _threshold; 6764 assert(old_threshold == (HeapWord*)round_to( 6765 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6766 "_threshold should always be card-aligned"); 6767 _threshold = (HeapWord*)round_to( 6768 (intptr_t)_finger, CardTableModRefBS::card_size); 6769 MemRegion mr(old_threshold, _threshold); 6770 assert(!mr.is_empty(), "Control point invariant"); 6771 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6772 _mut->clear_range(mr); 6773 } 6774 6775 // Note: the local finger doesn't advance while we drain 6776 // the stack below, but the global finger sure can and will. 6777 HeapWord** gfa = _task->global_finger_addr(); 6778 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6779 _span, _bit_map, 6780 _work_queue, 6781 _overflow_stack, 6782 _finger, 6783 gfa, this); 6784 bool res = _work_queue->push(obj); // overflow could occur here 6785 assert(res, "Will hold once we use workqueues"); 6786 while (true) { 6787 oop new_oop; 6788 if (!_work_queue->pop_local(new_oop)) { 6789 // We emptied our work_queue; check if there's stuff that can 6790 // be gotten from the overflow stack. 6791 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6792 _overflow_stack, _work_queue)) { 6793 do_yield_check(); 6794 continue; 6795 } else { // done 6796 break; 6797 } 6798 } 6799 // Skip verifying header mark word below because we are 6800 // running concurrent with mutators. 6801 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6802 // now scan this oop's oops 6803 new_oop->oop_iterate(&pushOrMarkClosure); 6804 do_yield_check(); 6805 } 6806 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6807 } 6808 6809 // Yield in response to a request from VM Thread or 6810 // from mutators. 6811 void Par_MarkFromRootsClosure::do_yield_work() { 6812 assert(_task != NULL, "sanity"); 6813 _task->yield(); 6814 } 6815 6816 // A variant of the above used for verifying CMS marking work. 6817 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6818 MemRegion span, 6819 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6820 CMSMarkStack* mark_stack): 6821 _collector(collector), 6822 _span(span), 6823 _verification_bm(verification_bm), 6824 _cms_bm(cms_bm), 6825 _mark_stack(mark_stack), 6826 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6827 mark_stack) 6828 { 6829 assert(_mark_stack->isEmpty(), "stack should be empty"); 6830 _finger = _verification_bm->startWord(); 6831 assert(_collector->_restart_addr == NULL, "Sanity check"); 6832 assert(_span.contains(_finger), "Out of bounds _finger?"); 6833 } 6834 6835 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6836 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6837 assert(_span.contains(addr), "Out of bounds _finger?"); 6838 _finger = addr; 6839 } 6840 6841 // Should revisit to see if this should be restructured for 6842 // greater efficiency. 6843 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6844 // convert offset into a HeapWord* 6845 HeapWord* addr = _verification_bm->startWord() + offset; 6846 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6847 "address out of range"); 6848 assert(_verification_bm->isMarked(addr), "tautology"); 6849 assert(_cms_bm->isMarked(addr), "tautology"); 6850 6851 assert(_mark_stack->isEmpty(), 6852 "should drain stack to limit stack usage"); 6853 // convert addr to an oop preparatory to scanning 6854 oop obj = oop(addr); 6855 assert(obj->is_oop(), "should be an oop"); 6856 assert(_finger <= addr, "_finger runneth ahead"); 6857 // advance the finger to right end of this object 6858 _finger = addr + obj->size(); 6859 assert(_finger > addr, "we just incremented it above"); 6860 // Note: the finger doesn't advance while we drain 6861 // the stack below. 6862 bool res = _mark_stack->push(obj); 6863 assert(res, "Empty non-zero size stack should have space for single push"); 6864 while (!_mark_stack->isEmpty()) { 6865 oop new_oop = _mark_stack->pop(); 6866 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6867 // now scan this oop's oops 6868 new_oop->oop_iterate(&_pam_verify_closure); 6869 } 6870 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6871 return true; 6872 } 6873 6874 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6875 CMSCollector* collector, MemRegion span, 6876 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6877 CMSMarkStack* mark_stack): 6878 MetadataAwareOopClosure(collector->ref_processor()), 6879 _collector(collector), 6880 _span(span), 6881 _verification_bm(verification_bm), 6882 _cms_bm(cms_bm), 6883 _mark_stack(mark_stack) 6884 { } 6885 6886 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6887 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6888 6889 // Upon stack overflow, we discard (part of) the stack, 6890 // remembering the least address amongst those discarded 6891 // in CMSCollector's _restart_address. 6892 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6893 // Remember the least grey address discarded 6894 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6895 _collector->lower_restart_addr(ra); 6896 _mark_stack->reset(); // discard stack contents 6897 _mark_stack->expand(); // expand the stack if possible 6898 } 6899 6900 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6901 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6902 HeapWord* addr = (HeapWord*)obj; 6903 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6904 // Oop lies in _span and isn't yet grey or black 6905 _verification_bm->mark(addr); // now grey 6906 if (!_cms_bm->isMarked(addr)) { 6907 oop(addr)->print(); 6908 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6909 p2i(addr)); 6910 fatal("... aborting"); 6911 } 6912 6913 if (!_mark_stack->push(obj)) { // stack overflow 6914 if (PrintCMSStatistics != 0) { 6915 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6916 SIZE_FORMAT, _mark_stack->capacity()); 6917 } 6918 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6919 handle_stack_overflow(addr); 6920 } 6921 // anything including and to the right of _finger 6922 // will be scanned as we iterate over the remainder of the 6923 // bit map 6924 } 6925 } 6926 6927 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6928 MemRegion span, 6929 CMSBitMap* bitMap, CMSMarkStack* markStack, 6930 HeapWord* finger, MarkFromRootsClosure* parent) : 6931 MetadataAwareOopClosure(collector->ref_processor()), 6932 _collector(collector), 6933 _span(span), 6934 _bitMap(bitMap), 6935 _markStack(markStack), 6936 _finger(finger), 6937 _parent(parent) 6938 { } 6939 6940 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6941 MemRegion span, 6942 CMSBitMap* bit_map, 6943 OopTaskQueue* work_queue, 6944 CMSMarkStack* overflow_stack, 6945 HeapWord* finger, 6946 HeapWord** global_finger_addr, 6947 Par_MarkFromRootsClosure* parent) : 6948 MetadataAwareOopClosure(collector->ref_processor()), 6949 _collector(collector), 6950 _whole_span(collector->_span), 6951 _span(span), 6952 _bit_map(bit_map), 6953 _work_queue(work_queue), 6954 _overflow_stack(overflow_stack), 6955 _finger(finger), 6956 _global_finger_addr(global_finger_addr), 6957 _parent(parent) 6958 { } 6959 6960 // Assumes thread-safe access by callers, who are 6961 // responsible for mutual exclusion. 6962 void CMSCollector::lower_restart_addr(HeapWord* low) { 6963 assert(_span.contains(low), "Out of bounds addr"); 6964 if (_restart_addr == NULL) { 6965 _restart_addr = low; 6966 } else { 6967 _restart_addr = MIN2(_restart_addr, low); 6968 } 6969 } 6970 6971 // Upon stack overflow, we discard (part of) the stack, 6972 // remembering the least address amongst those discarded 6973 // in CMSCollector's _restart_address. 6974 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6975 // Remember the least grey address discarded 6976 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6977 _collector->lower_restart_addr(ra); 6978 _markStack->reset(); // discard stack contents 6979 _markStack->expand(); // expand the stack if possible 6980 } 6981 6982 // Upon stack overflow, we discard (part of) the stack, 6983 // remembering the least address amongst those discarded 6984 // in CMSCollector's _restart_address. 6985 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6986 // We need to do this under a mutex to prevent other 6987 // workers from interfering with the work done below. 6988 MutexLockerEx ml(_overflow_stack->par_lock(), 6989 Mutex::_no_safepoint_check_flag); 6990 // Remember the least grey address discarded 6991 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6992 _collector->lower_restart_addr(ra); 6993 _overflow_stack->reset(); // discard stack contents 6994 _overflow_stack->expand(); // expand the stack if possible 6995 } 6996 6997 void PushOrMarkClosure::do_oop(oop obj) { 6998 // Ignore mark word because we are running concurrent with mutators. 6999 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7000 HeapWord* addr = (HeapWord*)obj; 7001 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7002 // Oop lies in _span and isn't yet grey or black 7003 _bitMap->mark(addr); // now grey 7004 if (addr < _finger) { 7005 // the bit map iteration has already either passed, or 7006 // sampled, this bit in the bit map; we'll need to 7007 // use the marking stack to scan this oop's oops. 7008 bool simulate_overflow = false; 7009 NOT_PRODUCT( 7010 if (CMSMarkStackOverflowALot && 7011 _collector->simulate_overflow()) { 7012 // simulate a stack overflow 7013 simulate_overflow = true; 7014 } 7015 ) 7016 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7017 if (PrintCMSStatistics != 0) { 7018 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7019 SIZE_FORMAT, _markStack->capacity()); 7020 } 7021 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7022 handle_stack_overflow(addr); 7023 } 7024 } 7025 // anything including and to the right of _finger 7026 // will be scanned as we iterate over the remainder of the 7027 // bit map 7028 do_yield_check(); 7029 } 7030 } 7031 7032 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7033 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7034 7035 void Par_PushOrMarkClosure::do_oop(oop obj) { 7036 // Ignore mark word because we are running concurrent with mutators. 7037 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7038 HeapWord* addr = (HeapWord*)obj; 7039 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7040 // Oop lies in _span and isn't yet grey or black 7041 // We read the global_finger (volatile read) strictly after marking oop 7042 bool res = _bit_map->par_mark(addr); // now grey 7043 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7044 // Should we push this marked oop on our stack? 7045 // -- if someone else marked it, nothing to do 7046 // -- if target oop is above global finger nothing to do 7047 // -- if target oop is in chunk and above local finger 7048 // then nothing to do 7049 // -- else push on work queue 7050 if ( !res // someone else marked it, they will deal with it 7051 || (addr >= *gfa) // will be scanned in a later task 7052 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7053 return; 7054 } 7055 // the bit map iteration has already either passed, or 7056 // sampled, this bit in the bit map; we'll need to 7057 // use the marking stack to scan this oop's oops. 7058 bool simulate_overflow = false; 7059 NOT_PRODUCT( 7060 if (CMSMarkStackOverflowALot && 7061 _collector->simulate_overflow()) { 7062 // simulate a stack overflow 7063 simulate_overflow = true; 7064 } 7065 ) 7066 if (simulate_overflow || 7067 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7068 // stack overflow 7069 if (PrintCMSStatistics != 0) { 7070 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7071 SIZE_FORMAT, _overflow_stack->capacity()); 7072 } 7073 // We cannot assert that the overflow stack is full because 7074 // it may have been emptied since. 7075 assert(simulate_overflow || 7076 _work_queue->size() == _work_queue->max_elems(), 7077 "Else push should have succeeded"); 7078 handle_stack_overflow(addr); 7079 } 7080 do_yield_check(); 7081 } 7082 } 7083 7084 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7085 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7086 7087 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7088 MemRegion span, 7089 ReferenceProcessor* rp, 7090 CMSBitMap* bit_map, 7091 CMSBitMap* mod_union_table, 7092 CMSMarkStack* mark_stack, 7093 bool concurrent_precleaning): 7094 MetadataAwareOopClosure(rp), 7095 _collector(collector), 7096 _span(span), 7097 _bit_map(bit_map), 7098 _mod_union_table(mod_union_table), 7099 _mark_stack(mark_stack), 7100 _concurrent_precleaning(concurrent_precleaning) 7101 { 7102 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7103 } 7104 7105 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7106 // the non-parallel version (the parallel version appears further below.) 7107 void PushAndMarkClosure::do_oop(oop obj) { 7108 // Ignore mark word verification. If during concurrent precleaning, 7109 // the object monitor may be locked. If during the checkpoint 7110 // phases, the object may already have been reached by a different 7111 // path and may be at the end of the global overflow list (so 7112 // the mark word may be NULL). 7113 assert(obj->is_oop_or_null(true /* ignore mark word */), 7114 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7115 HeapWord* addr = (HeapWord*)obj; 7116 // Check if oop points into the CMS generation 7117 // and is not marked 7118 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7119 // a white object ... 7120 _bit_map->mark(addr); // ... now grey 7121 // push on the marking stack (grey set) 7122 bool simulate_overflow = false; 7123 NOT_PRODUCT( 7124 if (CMSMarkStackOverflowALot && 7125 _collector->simulate_overflow()) { 7126 // simulate a stack overflow 7127 simulate_overflow = true; 7128 } 7129 ) 7130 if (simulate_overflow || !_mark_stack->push(obj)) { 7131 if (_concurrent_precleaning) { 7132 // During precleaning we can just dirty the appropriate card(s) 7133 // in the mod union table, thus ensuring that the object remains 7134 // in the grey set and continue. In the case of object arrays 7135 // we need to dirty all of the cards that the object spans, 7136 // since the rescan of object arrays will be limited to the 7137 // dirty cards. 7138 // Note that no one can be interfering with us in this action 7139 // of dirtying the mod union table, so no locking or atomics 7140 // are required. 7141 if (obj->is_objArray()) { 7142 size_t sz = obj->size(); 7143 HeapWord* end_card_addr = (HeapWord*)round_to( 7144 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7145 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7146 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7147 _mod_union_table->mark_range(redirty_range); 7148 } else { 7149 _mod_union_table->mark(addr); 7150 } 7151 _collector->_ser_pmc_preclean_ovflw++; 7152 } else { 7153 // During the remark phase, we need to remember this oop 7154 // in the overflow list. 7155 _collector->push_on_overflow_list(obj); 7156 _collector->_ser_pmc_remark_ovflw++; 7157 } 7158 } 7159 } 7160 } 7161 7162 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7163 MemRegion span, 7164 ReferenceProcessor* rp, 7165 CMSBitMap* bit_map, 7166 OopTaskQueue* work_queue): 7167 MetadataAwareOopClosure(rp), 7168 _collector(collector), 7169 _span(span), 7170 _bit_map(bit_map), 7171 _work_queue(work_queue) 7172 { 7173 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7174 } 7175 7176 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7177 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7178 7179 // Grey object rescan during second checkpoint phase -- 7180 // the parallel version. 7181 void Par_PushAndMarkClosure::do_oop(oop obj) { 7182 // In the assert below, we ignore the mark word because 7183 // this oop may point to an already visited object that is 7184 // on the overflow stack (in which case the mark word has 7185 // been hijacked for chaining into the overflow stack -- 7186 // if this is the last object in the overflow stack then 7187 // its mark word will be NULL). Because this object may 7188 // have been subsequently popped off the global overflow 7189 // stack, and the mark word possibly restored to the prototypical 7190 // value, by the time we get to examined this failing assert in 7191 // the debugger, is_oop_or_null(false) may subsequently start 7192 // to hold. 7193 assert(obj->is_oop_or_null(true), 7194 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7195 HeapWord* addr = (HeapWord*)obj; 7196 // Check if oop points into the CMS generation 7197 // and is not marked 7198 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7199 // a white object ... 7200 // If we manage to "claim" the object, by being the 7201 // first thread to mark it, then we push it on our 7202 // marking stack 7203 if (_bit_map->par_mark(addr)) { // ... now grey 7204 // push on work queue (grey set) 7205 bool simulate_overflow = false; 7206 NOT_PRODUCT( 7207 if (CMSMarkStackOverflowALot && 7208 _collector->par_simulate_overflow()) { 7209 // simulate a stack overflow 7210 simulate_overflow = true; 7211 } 7212 ) 7213 if (simulate_overflow || !_work_queue->push(obj)) { 7214 _collector->par_push_on_overflow_list(obj); 7215 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7216 } 7217 } // Else, some other thread got there first 7218 } 7219 } 7220 7221 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7222 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7223 7224 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7225 Mutex* bml = _collector->bitMapLock(); 7226 assert_lock_strong(bml); 7227 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7228 "CMS thread should hold CMS token"); 7229 7230 bml->unlock(); 7231 ConcurrentMarkSweepThread::desynchronize(true); 7232 7233 _collector->stopTimer(); 7234 if (PrintCMSStatistics != 0) { 7235 _collector->incrementYields(); 7236 } 7237 7238 // See the comment in coordinator_yield() 7239 for (unsigned i = 0; i < CMSYieldSleepCount && 7240 ConcurrentMarkSweepThread::should_yield() && 7241 !CMSCollector::foregroundGCIsActive(); ++i) { 7242 os::sleep(Thread::current(), 1, false); 7243 } 7244 7245 ConcurrentMarkSweepThread::synchronize(true); 7246 bml->lock(); 7247 7248 _collector->startTimer(); 7249 } 7250 7251 bool CMSPrecleanRefsYieldClosure::should_return() { 7252 if (ConcurrentMarkSweepThread::should_yield()) { 7253 do_yield_work(); 7254 } 7255 return _collector->foregroundGCIsActive(); 7256 } 7257 7258 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7259 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7260 "mr should be aligned to start at a card boundary"); 7261 // We'd like to assert: 7262 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7263 // "mr should be a range of cards"); 7264 // However, that would be too strong in one case -- the last 7265 // partition ends at _unallocated_block which, in general, can be 7266 // an arbitrary boundary, not necessarily card aligned. 7267 if (PrintCMSStatistics != 0) { 7268 _num_dirty_cards += 7269 mr.word_size()/CardTableModRefBS::card_size_in_words; 7270 } 7271 _space->object_iterate_mem(mr, &_scan_cl); 7272 } 7273 7274 SweepClosure::SweepClosure(CMSCollector* collector, 7275 ConcurrentMarkSweepGeneration* g, 7276 CMSBitMap* bitMap, bool should_yield) : 7277 _collector(collector), 7278 _g(g), 7279 _sp(g->cmsSpace()), 7280 _limit(_sp->sweep_limit()), 7281 _freelistLock(_sp->freelistLock()), 7282 _bitMap(bitMap), 7283 _yield(should_yield), 7284 _inFreeRange(false), // No free range at beginning of sweep 7285 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7286 _lastFreeRangeCoalesced(false), 7287 _freeFinger(g->used_region().start()) 7288 { 7289 NOT_PRODUCT( 7290 _numObjectsFreed = 0; 7291 _numWordsFreed = 0; 7292 _numObjectsLive = 0; 7293 _numWordsLive = 0; 7294 _numObjectsAlreadyFree = 0; 7295 _numWordsAlreadyFree = 0; 7296 _last_fc = NULL; 7297 7298 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7299 _sp->dictionary()->initialize_dict_returned_bytes(); 7300 ) 7301 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7302 "sweep _limit out of bounds"); 7303 if (CMSTraceSweeper) { 7304 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7305 p2i(_limit)); 7306 } 7307 } 7308 7309 void SweepClosure::print_on(outputStream* st) const { 7310 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7311 p2i(_sp->bottom()), p2i(_sp->end())); 7312 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7313 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7314 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7315 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7316 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7317 } 7318 7319 #ifndef PRODUCT 7320 // Assertion checking only: no useful work in product mode -- 7321 // however, if any of the flags below become product flags, 7322 // you may need to review this code to see if it needs to be 7323 // enabled in product mode. 7324 SweepClosure::~SweepClosure() { 7325 assert_lock_strong(_freelistLock); 7326 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7327 "sweep _limit out of bounds"); 7328 if (inFreeRange()) { 7329 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7330 print(); 7331 ShouldNotReachHere(); 7332 } 7333 if (Verbose && PrintGC) { 7334 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7335 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7336 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7337 SIZE_FORMAT" bytes " 7338 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7339 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7340 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7341 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7342 * sizeof(HeapWord); 7343 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7344 7345 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7346 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7347 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7348 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7349 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7350 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7351 indexListReturnedBytes); 7352 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7353 dict_returned_bytes); 7354 } 7355 } 7356 if (CMSTraceSweeper) { 7357 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7358 p2i(_limit)); 7359 } 7360 } 7361 #endif // PRODUCT 7362 7363 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7364 bool freeRangeInFreeLists) { 7365 if (CMSTraceSweeper) { 7366 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7367 p2i(freeFinger), freeRangeInFreeLists); 7368 } 7369 assert(!inFreeRange(), "Trampling existing free range"); 7370 set_inFreeRange(true); 7371 set_lastFreeRangeCoalesced(false); 7372 7373 set_freeFinger(freeFinger); 7374 set_freeRangeInFreeLists(freeRangeInFreeLists); 7375 if (CMSTestInFreeList) { 7376 if (freeRangeInFreeLists) { 7377 FreeChunk* fc = (FreeChunk*) freeFinger; 7378 assert(fc->is_free(), "A chunk on the free list should be free."); 7379 assert(fc->size() > 0, "Free range should have a size"); 7380 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7381 } 7382 } 7383 } 7384 7385 // Note that the sweeper runs concurrently with mutators. Thus, 7386 // it is possible for direct allocation in this generation to happen 7387 // in the middle of the sweep. Note that the sweeper also coalesces 7388 // contiguous free blocks. Thus, unless the sweeper and the allocator 7389 // synchronize appropriately freshly allocated blocks may get swept up. 7390 // This is accomplished by the sweeper locking the free lists while 7391 // it is sweeping. Thus blocks that are determined to be free are 7392 // indeed free. There is however one additional complication: 7393 // blocks that have been allocated since the final checkpoint and 7394 // mark, will not have been marked and so would be treated as 7395 // unreachable and swept up. To prevent this, the allocator marks 7396 // the bit map when allocating during the sweep phase. This leads, 7397 // however, to a further complication -- objects may have been allocated 7398 // but not yet initialized -- in the sense that the header isn't yet 7399 // installed. The sweeper can not then determine the size of the block 7400 // in order to skip over it. To deal with this case, we use a technique 7401 // (due to Printezis) to encode such uninitialized block sizes in the 7402 // bit map. Since the bit map uses a bit per every HeapWord, but the 7403 // CMS generation has a minimum object size of 3 HeapWords, it follows 7404 // that "normal marks" won't be adjacent in the bit map (there will 7405 // always be at least two 0 bits between successive 1 bits). We make use 7406 // of these "unused" bits to represent uninitialized blocks -- the bit 7407 // corresponding to the start of the uninitialized object and the next 7408 // bit are both set. Finally, a 1 bit marks the end of the object that 7409 // started with the two consecutive 1 bits to indicate its potentially 7410 // uninitialized state. 7411 7412 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7413 FreeChunk* fc = (FreeChunk*)addr; 7414 size_t res; 7415 7416 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7417 // than "addr == _limit" because although _limit was a block boundary when 7418 // we started the sweep, it may no longer be one because heap expansion 7419 // may have caused us to coalesce the block ending at the address _limit 7420 // with a newly expanded chunk (this happens when _limit was set to the 7421 // previous _end of the space), so we may have stepped past _limit: 7422 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7423 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7424 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7425 "sweep _limit out of bounds"); 7426 assert(addr < _sp->end(), "addr out of bounds"); 7427 // Flush any free range we might be holding as a single 7428 // coalesced chunk to the appropriate free list. 7429 if (inFreeRange()) { 7430 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7431 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7432 flush_cur_free_chunk(freeFinger(), 7433 pointer_delta(addr, freeFinger())); 7434 if (CMSTraceSweeper) { 7435 gclog_or_tty->print("Sweep: last chunk: "); 7436 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7437 "[coalesced:%d]\n", 7438 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7439 lastFreeRangeCoalesced() ? 1 : 0); 7440 } 7441 } 7442 7443 // help the iterator loop finish 7444 return pointer_delta(_sp->end(), addr); 7445 } 7446 7447 assert(addr < _limit, "sweep invariant"); 7448 // check if we should yield 7449 do_yield_check(addr); 7450 if (fc->is_free()) { 7451 // Chunk that is already free 7452 res = fc->size(); 7453 do_already_free_chunk(fc); 7454 debug_only(_sp->verifyFreeLists()); 7455 // If we flush the chunk at hand in lookahead_and_flush() 7456 // and it's coalesced with a preceding chunk, then the 7457 // process of "mangling" the payload of the coalesced block 7458 // will cause erasure of the size information from the 7459 // (erstwhile) header of all the coalesced blocks but the 7460 // first, so the first disjunct in the assert will not hold 7461 // in that specific case (in which case the second disjunct 7462 // will hold). 7463 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7464 "Otherwise the size info doesn't change at this step"); 7465 NOT_PRODUCT( 7466 _numObjectsAlreadyFree++; 7467 _numWordsAlreadyFree += res; 7468 ) 7469 NOT_PRODUCT(_last_fc = fc;) 7470 } else if (!_bitMap->isMarked(addr)) { 7471 // Chunk is fresh garbage 7472 res = do_garbage_chunk(fc); 7473 debug_only(_sp->verifyFreeLists()); 7474 NOT_PRODUCT( 7475 _numObjectsFreed++; 7476 _numWordsFreed += res; 7477 ) 7478 } else { 7479 // Chunk that is alive. 7480 res = do_live_chunk(fc); 7481 debug_only(_sp->verifyFreeLists()); 7482 NOT_PRODUCT( 7483 _numObjectsLive++; 7484 _numWordsLive += res; 7485 ) 7486 } 7487 return res; 7488 } 7489 7490 // For the smart allocation, record following 7491 // split deaths - a free chunk is removed from its free list because 7492 // it is being split into two or more chunks. 7493 // split birth - a free chunk is being added to its free list because 7494 // a larger free chunk has been split and resulted in this free chunk. 7495 // coal death - a free chunk is being removed from its free list because 7496 // it is being coalesced into a large free chunk. 7497 // coal birth - a free chunk is being added to its free list because 7498 // it was created when two or more free chunks where coalesced into 7499 // this free chunk. 7500 // 7501 // These statistics are used to determine the desired number of free 7502 // chunks of a given size. The desired number is chosen to be relative 7503 // to the end of a CMS sweep. The desired number at the end of a sweep 7504 // is the 7505 // count-at-end-of-previous-sweep (an amount that was enough) 7506 // - count-at-beginning-of-current-sweep (the excess) 7507 // + split-births (gains in this size during interval) 7508 // - split-deaths (demands on this size during interval) 7509 // where the interval is from the end of one sweep to the end of the 7510 // next. 7511 // 7512 // When sweeping the sweeper maintains an accumulated chunk which is 7513 // the chunk that is made up of chunks that have been coalesced. That 7514 // will be termed the left-hand chunk. A new chunk of garbage that 7515 // is being considered for coalescing will be referred to as the 7516 // right-hand chunk. 7517 // 7518 // When making a decision on whether to coalesce a right-hand chunk with 7519 // the current left-hand chunk, the current count vs. the desired count 7520 // of the left-hand chunk is considered. Also if the right-hand chunk 7521 // is near the large chunk at the end of the heap (see 7522 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7523 // left-hand chunk is coalesced. 7524 // 7525 // When making a decision about whether to split a chunk, the desired count 7526 // vs. the current count of the candidate to be split is also considered. 7527 // If the candidate is underpopulated (currently fewer chunks than desired) 7528 // a chunk of an overpopulated (currently more chunks than desired) size may 7529 // be chosen. The "hint" associated with a free list, if non-null, points 7530 // to a free list which may be overpopulated. 7531 // 7532 7533 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7534 const size_t size = fc->size(); 7535 // Chunks that cannot be coalesced are not in the 7536 // free lists. 7537 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7538 assert(_sp->verify_chunk_in_free_list(fc), 7539 "free chunk should be in free lists"); 7540 } 7541 // a chunk that is already free, should not have been 7542 // marked in the bit map 7543 HeapWord* const addr = (HeapWord*) fc; 7544 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7545 // Verify that the bit map has no bits marked between 7546 // addr and purported end of this block. 7547 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7548 7549 // Some chunks cannot be coalesced under any circumstances. 7550 // See the definition of cantCoalesce(). 7551 if (!fc->cantCoalesce()) { 7552 // This chunk can potentially be coalesced. 7553 if (_sp->adaptive_freelists()) { 7554 // All the work is done in 7555 do_post_free_or_garbage_chunk(fc, size); 7556 } else { // Not adaptive free lists 7557 // this is a free chunk that can potentially be coalesced by the sweeper; 7558 if (!inFreeRange()) { 7559 // if the next chunk is a free block that can't be coalesced 7560 // it doesn't make sense to remove this chunk from the free lists 7561 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7562 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7563 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7564 nextChunk->is_free() && // ... which is free... 7565 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7566 // nothing to do 7567 } else { 7568 // Potentially the start of a new free range: 7569 // Don't eagerly remove it from the free lists. 7570 // No need to remove it if it will just be put 7571 // back again. (Also from a pragmatic point of view 7572 // if it is a free block in a region that is beyond 7573 // any allocated blocks, an assertion will fail) 7574 // Remember the start of a free run. 7575 initialize_free_range(addr, true); 7576 // end - can coalesce with next chunk 7577 } 7578 } else { 7579 // the midst of a free range, we are coalescing 7580 print_free_block_coalesced(fc); 7581 if (CMSTraceSweeper) { 7582 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7583 } 7584 // remove it from the free lists 7585 _sp->removeFreeChunkFromFreeLists(fc); 7586 set_lastFreeRangeCoalesced(true); 7587 // If the chunk is being coalesced and the current free range is 7588 // in the free lists, remove the current free range so that it 7589 // will be returned to the free lists in its entirety - all 7590 // the coalesced pieces included. 7591 if (freeRangeInFreeLists()) { 7592 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7593 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7594 "Size of free range is inconsistent with chunk size."); 7595 if (CMSTestInFreeList) { 7596 assert(_sp->verify_chunk_in_free_list(ffc), 7597 "free range is not in free lists"); 7598 } 7599 _sp->removeFreeChunkFromFreeLists(ffc); 7600 set_freeRangeInFreeLists(false); 7601 } 7602 } 7603 } 7604 // Note that if the chunk is not coalescable (the else arm 7605 // below), we unconditionally flush, without needing to do 7606 // a "lookahead," as we do below. 7607 if (inFreeRange()) lookahead_and_flush(fc, size); 7608 } else { 7609 // Code path common to both original and adaptive free lists. 7610 7611 // cant coalesce with previous block; this should be treated 7612 // as the end of a free run if any 7613 if (inFreeRange()) { 7614 // we kicked some butt; time to pick up the garbage 7615 assert(freeFinger() < addr, "freeFinger points too high"); 7616 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7617 } 7618 // else, nothing to do, just continue 7619 } 7620 } 7621 7622 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7623 // This is a chunk of garbage. It is not in any free list. 7624 // Add it to a free list or let it possibly be coalesced into 7625 // a larger chunk. 7626 HeapWord* const addr = (HeapWord*) fc; 7627 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7628 7629 if (_sp->adaptive_freelists()) { 7630 // Verify that the bit map has no bits marked between 7631 // addr and purported end of just dead object. 7632 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7633 7634 do_post_free_or_garbage_chunk(fc, size); 7635 } else { 7636 if (!inFreeRange()) { 7637 // start of a new free range 7638 assert(size > 0, "A free range should have a size"); 7639 initialize_free_range(addr, false); 7640 } else { 7641 // this will be swept up when we hit the end of the 7642 // free range 7643 if (CMSTraceSweeper) { 7644 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7645 } 7646 // If the chunk is being coalesced and the current free range is 7647 // in the free lists, remove the current free range so that it 7648 // will be returned to the free lists in its entirety - all 7649 // the coalesced pieces included. 7650 if (freeRangeInFreeLists()) { 7651 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7652 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7653 "Size of free range is inconsistent with chunk size."); 7654 if (CMSTestInFreeList) { 7655 assert(_sp->verify_chunk_in_free_list(ffc), 7656 "free range is not in free lists"); 7657 } 7658 _sp->removeFreeChunkFromFreeLists(ffc); 7659 set_freeRangeInFreeLists(false); 7660 } 7661 set_lastFreeRangeCoalesced(true); 7662 } 7663 // this will be swept up when we hit the end of the free range 7664 7665 // Verify that the bit map has no bits marked between 7666 // addr and purported end of just dead object. 7667 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7668 } 7669 assert(_limit >= addr + size, 7670 "A freshly garbage chunk can't possibly straddle over _limit"); 7671 if (inFreeRange()) lookahead_and_flush(fc, size); 7672 return size; 7673 } 7674 7675 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7676 HeapWord* addr = (HeapWord*) fc; 7677 // The sweeper has just found a live object. Return any accumulated 7678 // left hand chunk to the free lists. 7679 if (inFreeRange()) { 7680 assert(freeFinger() < addr, "freeFinger points too high"); 7681 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7682 } 7683 7684 // This object is live: we'd normally expect this to be 7685 // an oop, and like to assert the following: 7686 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7687 // However, as we commented above, this may be an object whose 7688 // header hasn't yet been initialized. 7689 size_t size; 7690 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7691 if (_bitMap->isMarked(addr + 1)) { 7692 // Determine the size from the bit map, rather than trying to 7693 // compute it from the object header. 7694 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7695 size = pointer_delta(nextOneAddr + 1, addr); 7696 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7697 "alignment problem"); 7698 7699 #ifdef ASSERT 7700 if (oop(addr)->klass_or_null() != NULL) { 7701 // Ignore mark word because we are running concurrent with mutators 7702 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7703 assert(size == 7704 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7705 "P-mark and computed size do not agree"); 7706 } 7707 #endif 7708 7709 } else { 7710 // This should be an initialized object that's alive. 7711 assert(oop(addr)->klass_or_null() != NULL, 7712 "Should be an initialized object"); 7713 // Ignore mark word because we are running concurrent with mutators 7714 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7715 // Verify that the bit map has no bits marked between 7716 // addr and purported end of this block. 7717 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7718 assert(size >= 3, "Necessary for Printezis marks to work"); 7719 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7720 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7721 } 7722 return size; 7723 } 7724 7725 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7726 size_t chunkSize) { 7727 // do_post_free_or_garbage_chunk() should only be called in the case 7728 // of the adaptive free list allocator. 7729 const bool fcInFreeLists = fc->is_free(); 7730 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7731 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7732 if (CMSTestInFreeList && fcInFreeLists) { 7733 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7734 } 7735 7736 if (CMSTraceSweeper) { 7737 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7738 } 7739 7740 HeapWord* const fc_addr = (HeapWord*) fc; 7741 7742 bool coalesce; 7743 const size_t left = pointer_delta(fc_addr, freeFinger()); 7744 const size_t right = chunkSize; 7745 switch (FLSCoalescePolicy) { 7746 // numeric value forms a coalition aggressiveness metric 7747 case 0: { // never coalesce 7748 coalesce = false; 7749 break; 7750 } 7751 case 1: { // coalesce if left & right chunks on overpopulated lists 7752 coalesce = _sp->coalOverPopulated(left) && 7753 _sp->coalOverPopulated(right); 7754 break; 7755 } 7756 case 2: { // coalesce if left chunk on overpopulated list (default) 7757 coalesce = _sp->coalOverPopulated(left); 7758 break; 7759 } 7760 case 3: { // coalesce if left OR right chunk on overpopulated list 7761 coalesce = _sp->coalOverPopulated(left) || 7762 _sp->coalOverPopulated(right); 7763 break; 7764 } 7765 case 4: { // always coalesce 7766 coalesce = true; 7767 break; 7768 } 7769 default: 7770 ShouldNotReachHere(); 7771 } 7772 7773 // Should the current free range be coalesced? 7774 // If the chunk is in a free range and either we decided to coalesce above 7775 // or the chunk is near the large block at the end of the heap 7776 // (isNearLargestChunk() returns true), then coalesce this chunk. 7777 const bool doCoalesce = inFreeRange() 7778 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7779 if (doCoalesce) { 7780 // Coalesce the current free range on the left with the new 7781 // chunk on the right. If either is on a free list, 7782 // it must be removed from the list and stashed in the closure. 7783 if (freeRangeInFreeLists()) { 7784 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7785 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7786 "Size of free range is inconsistent with chunk size."); 7787 if (CMSTestInFreeList) { 7788 assert(_sp->verify_chunk_in_free_list(ffc), 7789 "Chunk is not in free lists"); 7790 } 7791 _sp->coalDeath(ffc->size()); 7792 _sp->removeFreeChunkFromFreeLists(ffc); 7793 set_freeRangeInFreeLists(false); 7794 } 7795 if (fcInFreeLists) { 7796 _sp->coalDeath(chunkSize); 7797 assert(fc->size() == chunkSize, 7798 "The chunk has the wrong size or is not in the free lists"); 7799 _sp->removeFreeChunkFromFreeLists(fc); 7800 } 7801 set_lastFreeRangeCoalesced(true); 7802 print_free_block_coalesced(fc); 7803 } else { // not in a free range and/or should not coalesce 7804 // Return the current free range and start a new one. 7805 if (inFreeRange()) { 7806 // In a free range but cannot coalesce with the right hand chunk. 7807 // Put the current free range into the free lists. 7808 flush_cur_free_chunk(freeFinger(), 7809 pointer_delta(fc_addr, freeFinger())); 7810 } 7811 // Set up for new free range. Pass along whether the right hand 7812 // chunk is in the free lists. 7813 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7814 } 7815 } 7816 7817 // Lookahead flush: 7818 // If we are tracking a free range, and this is the last chunk that 7819 // we'll look at because its end crosses past _limit, we'll preemptively 7820 // flush it along with any free range we may be holding on to. Note that 7821 // this can be the case only for an already free or freshly garbage 7822 // chunk. If this block is an object, it can never straddle 7823 // over _limit. The "straddling" occurs when _limit is set at 7824 // the previous end of the space when this cycle started, and 7825 // a subsequent heap expansion caused the previously co-terminal 7826 // free block to be coalesced with the newly expanded portion, 7827 // thus rendering _limit a non-block-boundary making it dangerous 7828 // for the sweeper to step over and examine. 7829 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7830 assert(inFreeRange(), "Should only be called if currently in a free range."); 7831 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7832 assert(_sp->used_region().contains(eob - 1), 7833 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7834 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7835 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7836 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7837 if (eob >= _limit) { 7838 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7839 if (CMSTraceSweeper) { 7840 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7841 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7842 "[" PTR_FORMAT "," PTR_FORMAT ")", 7843 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7844 } 7845 // Return the storage we are tracking back into the free lists. 7846 if (CMSTraceSweeper) { 7847 gclog_or_tty->print_cr("Flushing ... "); 7848 } 7849 assert(freeFinger() < eob, "Error"); 7850 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7851 } 7852 } 7853 7854 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7855 assert(inFreeRange(), "Should only be called if currently in a free range."); 7856 assert(size > 0, 7857 "A zero sized chunk cannot be added to the free lists."); 7858 if (!freeRangeInFreeLists()) { 7859 if (CMSTestInFreeList) { 7860 FreeChunk* fc = (FreeChunk*) chunk; 7861 fc->set_size(size); 7862 assert(!_sp->verify_chunk_in_free_list(fc), 7863 "chunk should not be in free lists yet"); 7864 } 7865 if (CMSTraceSweeper) { 7866 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7867 p2i(chunk), size); 7868 } 7869 // A new free range is going to be starting. The current 7870 // free range has not been added to the free lists yet or 7871 // was removed so add it back. 7872 // If the current free range was coalesced, then the death 7873 // of the free range was recorded. Record a birth now. 7874 if (lastFreeRangeCoalesced()) { 7875 _sp->coalBirth(size); 7876 } 7877 _sp->addChunkAndRepairOffsetTable(chunk, size, 7878 lastFreeRangeCoalesced()); 7879 } else if (CMSTraceSweeper) { 7880 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7881 } 7882 set_inFreeRange(false); 7883 set_freeRangeInFreeLists(false); 7884 } 7885 7886 // We take a break if we've been at this for a while, 7887 // so as to avoid monopolizing the locks involved. 7888 void SweepClosure::do_yield_work(HeapWord* addr) { 7889 // Return current free chunk being used for coalescing (if any) 7890 // to the appropriate freelist. After yielding, the next 7891 // free block encountered will start a coalescing range of 7892 // free blocks. If the next free block is adjacent to the 7893 // chunk just flushed, they will need to wait for the next 7894 // sweep to be coalesced. 7895 if (inFreeRange()) { 7896 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7897 } 7898 7899 // First give up the locks, then yield, then re-lock. 7900 // We should probably use a constructor/destructor idiom to 7901 // do this unlock/lock or modify the MutexUnlocker class to 7902 // serve our purpose. XXX 7903 assert_lock_strong(_bitMap->lock()); 7904 assert_lock_strong(_freelistLock); 7905 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7906 "CMS thread should hold CMS token"); 7907 _bitMap->lock()->unlock(); 7908 _freelistLock->unlock(); 7909 ConcurrentMarkSweepThread::desynchronize(true); 7910 _collector->stopTimer(); 7911 if (PrintCMSStatistics != 0) { 7912 _collector->incrementYields(); 7913 } 7914 7915 // See the comment in coordinator_yield() 7916 for (unsigned i = 0; i < CMSYieldSleepCount && 7917 ConcurrentMarkSweepThread::should_yield() && 7918 !CMSCollector::foregroundGCIsActive(); ++i) { 7919 os::sleep(Thread::current(), 1, false); 7920 } 7921 7922 ConcurrentMarkSweepThread::synchronize(true); 7923 _freelistLock->lock(); 7924 _bitMap->lock()->lock_without_safepoint_check(); 7925 _collector->startTimer(); 7926 } 7927 7928 #ifndef PRODUCT 7929 // This is actually very useful in a product build if it can 7930 // be called from the debugger. Compile it into the product 7931 // as needed. 7932 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7933 return debug_cms_space->verify_chunk_in_free_list(fc); 7934 } 7935 #endif 7936 7937 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7938 if (CMSTraceSweeper) { 7939 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7940 p2i(fc), fc->size()); 7941 } 7942 } 7943 7944 // CMSIsAliveClosure 7945 bool CMSIsAliveClosure::do_object_b(oop obj) { 7946 HeapWord* addr = (HeapWord*)obj; 7947 return addr != NULL && 7948 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7949 } 7950 7951 7952 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7953 MemRegion span, 7954 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7955 bool cpc): 7956 _collector(collector), 7957 _span(span), 7958 _bit_map(bit_map), 7959 _mark_stack(mark_stack), 7960 _concurrent_precleaning(cpc) { 7961 assert(!_span.is_empty(), "Empty span could spell trouble"); 7962 } 7963 7964 7965 // CMSKeepAliveClosure: the serial version 7966 void CMSKeepAliveClosure::do_oop(oop obj) { 7967 HeapWord* addr = (HeapWord*)obj; 7968 if (_span.contains(addr) && 7969 !_bit_map->isMarked(addr)) { 7970 _bit_map->mark(addr); 7971 bool simulate_overflow = false; 7972 NOT_PRODUCT( 7973 if (CMSMarkStackOverflowALot && 7974 _collector->simulate_overflow()) { 7975 // simulate a stack overflow 7976 simulate_overflow = true; 7977 } 7978 ) 7979 if (simulate_overflow || !_mark_stack->push(obj)) { 7980 if (_concurrent_precleaning) { 7981 // We dirty the overflown object and let the remark 7982 // phase deal with it. 7983 assert(_collector->overflow_list_is_empty(), "Error"); 7984 // In the case of object arrays, we need to dirty all of 7985 // the cards that the object spans. No locking or atomics 7986 // are needed since no one else can be mutating the mod union 7987 // table. 7988 if (obj->is_objArray()) { 7989 size_t sz = obj->size(); 7990 HeapWord* end_card_addr = 7991 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7992 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7993 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7994 _collector->_modUnionTable.mark_range(redirty_range); 7995 } else { 7996 _collector->_modUnionTable.mark(addr); 7997 } 7998 _collector->_ser_kac_preclean_ovflw++; 7999 } else { 8000 _collector->push_on_overflow_list(obj); 8001 _collector->_ser_kac_ovflw++; 8002 } 8003 } 8004 } 8005 } 8006 8007 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8008 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8009 8010 // CMSParKeepAliveClosure: a parallel version of the above. 8011 // The work queues are private to each closure (thread), 8012 // but (may be) available for stealing by other threads. 8013 void CMSParKeepAliveClosure::do_oop(oop obj) { 8014 HeapWord* addr = (HeapWord*)obj; 8015 if (_span.contains(addr) && 8016 !_bit_map->isMarked(addr)) { 8017 // In general, during recursive tracing, several threads 8018 // may be concurrently getting here; the first one to 8019 // "tag" it, claims it. 8020 if (_bit_map->par_mark(addr)) { 8021 bool res = _work_queue->push(obj); 8022 assert(res, "Low water mark should be much less than capacity"); 8023 // Do a recursive trim in the hope that this will keep 8024 // stack usage lower, but leave some oops for potential stealers 8025 trim_queue(_low_water_mark); 8026 } // Else, another thread got there first 8027 } 8028 } 8029 8030 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8031 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8032 8033 void CMSParKeepAliveClosure::trim_queue(uint max) { 8034 while (_work_queue->size() > max) { 8035 oop new_oop; 8036 if (_work_queue->pop_local(new_oop)) { 8037 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8038 assert(_bit_map->isMarked((HeapWord*)new_oop), 8039 "no white objects on this stack!"); 8040 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8041 // iterate over the oops in this oop, marking and pushing 8042 // the ones in CMS heap (i.e. in _span). 8043 new_oop->oop_iterate(&_mark_and_push); 8044 } 8045 } 8046 } 8047 8048 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8049 CMSCollector* collector, 8050 MemRegion span, CMSBitMap* bit_map, 8051 OopTaskQueue* work_queue): 8052 _collector(collector), 8053 _span(span), 8054 _bit_map(bit_map), 8055 _work_queue(work_queue) { } 8056 8057 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8058 HeapWord* addr = (HeapWord*)obj; 8059 if (_span.contains(addr) && 8060 !_bit_map->isMarked(addr)) { 8061 if (_bit_map->par_mark(addr)) { 8062 bool simulate_overflow = false; 8063 NOT_PRODUCT( 8064 if (CMSMarkStackOverflowALot && 8065 _collector->par_simulate_overflow()) { 8066 // simulate a stack overflow 8067 simulate_overflow = true; 8068 } 8069 ) 8070 if (simulate_overflow || !_work_queue->push(obj)) { 8071 _collector->par_push_on_overflow_list(obj); 8072 _collector->_par_kac_ovflw++; 8073 } 8074 } // Else another thread got there already 8075 } 8076 } 8077 8078 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8079 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8080 8081 ////////////////////////////////////////////////////////////////// 8082 // CMSExpansionCause ///////////////////////////// 8083 ////////////////////////////////////////////////////////////////// 8084 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8085 switch (cause) { 8086 case _no_expansion: 8087 return "No expansion"; 8088 case _satisfy_free_ratio: 8089 return "Free ratio"; 8090 case _satisfy_promotion: 8091 return "Satisfy promotion"; 8092 case _satisfy_allocation: 8093 return "allocation"; 8094 case _allocate_par_lab: 8095 return "Par LAB"; 8096 case _allocate_par_spooling_space: 8097 return "Par Spooling Space"; 8098 case _adaptive_size_policy: 8099 return "Ergonomics"; 8100 default: 8101 return "unknown"; 8102 } 8103 } 8104 8105 void CMSDrainMarkingStackClosure::do_void() { 8106 // the max number to take from overflow list at a time 8107 const size_t num = _mark_stack->capacity()/4; 8108 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8109 "Overflow list should be NULL during concurrent phases"); 8110 while (!_mark_stack->isEmpty() || 8111 // if stack is empty, check the overflow list 8112 _collector->take_from_overflow_list(num, _mark_stack)) { 8113 oop obj = _mark_stack->pop(); 8114 HeapWord* addr = (HeapWord*)obj; 8115 assert(_span.contains(addr), "Should be within span"); 8116 assert(_bit_map->isMarked(addr), "Should be marked"); 8117 assert(obj->is_oop(), "Should be an oop"); 8118 obj->oop_iterate(_keep_alive); 8119 } 8120 } 8121 8122 void CMSParDrainMarkingStackClosure::do_void() { 8123 // drain queue 8124 trim_queue(0); 8125 } 8126 8127 // Trim our work_queue so its length is below max at return 8128 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8129 while (_work_queue->size() > max) { 8130 oop new_oop; 8131 if (_work_queue->pop_local(new_oop)) { 8132 assert(new_oop->is_oop(), "Expected an oop"); 8133 assert(_bit_map->isMarked((HeapWord*)new_oop), 8134 "no white objects on this stack!"); 8135 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8136 // iterate over the oops in this oop, marking and pushing 8137 // the ones in CMS heap (i.e. in _span). 8138 new_oop->oop_iterate(&_mark_and_push); 8139 } 8140 } 8141 } 8142 8143 //////////////////////////////////////////////////////////////////// 8144 // Support for Marking Stack Overflow list handling and related code 8145 //////////////////////////////////////////////////////////////////// 8146 // Much of the following code is similar in shape and spirit to the 8147 // code used in ParNewGC. We should try and share that code 8148 // as much as possible in the future. 8149 8150 #ifndef PRODUCT 8151 // Debugging support for CMSStackOverflowALot 8152 8153 // It's OK to call this multi-threaded; the worst thing 8154 // that can happen is that we'll get a bunch of closely 8155 // spaced simulated overflows, but that's OK, in fact 8156 // probably good as it would exercise the overflow code 8157 // under contention. 8158 bool CMSCollector::simulate_overflow() { 8159 if (_overflow_counter-- <= 0) { // just being defensive 8160 _overflow_counter = CMSMarkStackOverflowInterval; 8161 return true; 8162 } else { 8163 return false; 8164 } 8165 } 8166 8167 bool CMSCollector::par_simulate_overflow() { 8168 return simulate_overflow(); 8169 } 8170 #endif 8171 8172 // Single-threaded 8173 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8174 assert(stack->isEmpty(), "Expected precondition"); 8175 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8176 size_t i = num; 8177 oop cur = _overflow_list; 8178 const markOop proto = markOopDesc::prototype(); 8179 NOT_PRODUCT(ssize_t n = 0;) 8180 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8181 next = oop(cur->mark()); 8182 cur->set_mark(proto); // until proven otherwise 8183 assert(cur->is_oop(), "Should be an oop"); 8184 bool res = stack->push(cur); 8185 assert(res, "Bit off more than can chew?"); 8186 NOT_PRODUCT(n++;) 8187 } 8188 _overflow_list = cur; 8189 #ifndef PRODUCT 8190 assert(_num_par_pushes >= n, "Too many pops?"); 8191 _num_par_pushes -=n; 8192 #endif 8193 return !stack->isEmpty(); 8194 } 8195 8196 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8197 // (MT-safe) Get a prefix of at most "num" from the list. 8198 // The overflow list is chained through the mark word of 8199 // each object in the list. We fetch the entire list, 8200 // break off a prefix of the right size and return the 8201 // remainder. If other threads try to take objects from 8202 // the overflow list at that time, they will wait for 8203 // some time to see if data becomes available. If (and 8204 // only if) another thread places one or more object(s) 8205 // on the global list before we have returned the suffix 8206 // to the global list, we will walk down our local list 8207 // to find its end and append the global list to 8208 // our suffix before returning it. This suffix walk can 8209 // prove to be expensive (quadratic in the amount of traffic) 8210 // when there are many objects in the overflow list and 8211 // there is much producer-consumer contention on the list. 8212 // *NOTE*: The overflow list manipulation code here and 8213 // in ParNewGeneration:: are very similar in shape, 8214 // except that in the ParNew case we use the old (from/eden) 8215 // copy of the object to thread the list via its klass word. 8216 // Because of the common code, if you make any changes in 8217 // the code below, please check the ParNew version to see if 8218 // similar changes might be needed. 8219 // CR 6797058 has been filed to consolidate the common code. 8220 bool CMSCollector::par_take_from_overflow_list(size_t num, 8221 OopTaskQueue* work_q, 8222 int no_of_gc_threads) { 8223 assert(work_q->size() == 0, "First empty local work queue"); 8224 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8225 if (_overflow_list == NULL) { 8226 return false; 8227 } 8228 // Grab the entire list; we'll put back a suffix 8229 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8230 Thread* tid = Thread::current(); 8231 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8232 // set to ParallelGCThreads. 8233 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8234 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8235 // If the list is busy, we spin for a short while, 8236 // sleeping between attempts to get the list. 8237 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8238 os::sleep(tid, sleep_time_millis, false); 8239 if (_overflow_list == NULL) { 8240 // Nothing left to take 8241 return false; 8242 } else if (_overflow_list != BUSY) { 8243 // Try and grab the prefix 8244 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8245 } 8246 } 8247 // If the list was found to be empty, or we spun long 8248 // enough, we give up and return empty-handed. If we leave 8249 // the list in the BUSY state below, it must be the case that 8250 // some other thread holds the overflow list and will set it 8251 // to a non-BUSY state in the future. 8252 if (prefix == NULL || prefix == BUSY) { 8253 // Nothing to take or waited long enough 8254 if (prefix == NULL) { 8255 // Write back the NULL in case we overwrote it with BUSY above 8256 // and it is still the same value. 8257 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8258 } 8259 return false; 8260 } 8261 assert(prefix != NULL && prefix != BUSY, "Error"); 8262 size_t i = num; 8263 oop cur = prefix; 8264 // Walk down the first "num" objects, unless we reach the end. 8265 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8266 if (cur->mark() == NULL) { 8267 // We have "num" or fewer elements in the list, so there 8268 // is nothing to return to the global list. 8269 // Write back the NULL in lieu of the BUSY we wrote 8270 // above, if it is still the same value. 8271 if (_overflow_list == BUSY) { 8272 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8273 } 8274 } else { 8275 // Chop off the suffix and return it to the global list. 8276 assert(cur->mark() != BUSY, "Error"); 8277 oop suffix_head = cur->mark(); // suffix will be put back on global list 8278 cur->set_mark(NULL); // break off suffix 8279 // It's possible that the list is still in the empty(busy) state 8280 // we left it in a short while ago; in that case we may be 8281 // able to place back the suffix without incurring the cost 8282 // of a walk down the list. 8283 oop observed_overflow_list = _overflow_list; 8284 oop cur_overflow_list = observed_overflow_list; 8285 bool attached = false; 8286 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8287 observed_overflow_list = 8288 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8289 if (cur_overflow_list == observed_overflow_list) { 8290 attached = true; 8291 break; 8292 } else cur_overflow_list = observed_overflow_list; 8293 } 8294 if (!attached) { 8295 // Too bad, someone else sneaked in (at least) an element; we'll need 8296 // to do a splice. Find tail of suffix so we can prepend suffix to global 8297 // list. 8298 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8299 oop suffix_tail = cur; 8300 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8301 "Tautology"); 8302 observed_overflow_list = _overflow_list; 8303 do { 8304 cur_overflow_list = observed_overflow_list; 8305 if (cur_overflow_list != BUSY) { 8306 // Do the splice ... 8307 suffix_tail->set_mark(markOop(cur_overflow_list)); 8308 } else { // cur_overflow_list == BUSY 8309 suffix_tail->set_mark(NULL); 8310 } 8311 // ... and try to place spliced list back on overflow_list ... 8312 observed_overflow_list = 8313 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8314 } while (cur_overflow_list != observed_overflow_list); 8315 // ... until we have succeeded in doing so. 8316 } 8317 } 8318 8319 // Push the prefix elements on work_q 8320 assert(prefix != NULL, "control point invariant"); 8321 const markOop proto = markOopDesc::prototype(); 8322 oop next; 8323 NOT_PRODUCT(ssize_t n = 0;) 8324 for (cur = prefix; cur != NULL; cur = next) { 8325 next = oop(cur->mark()); 8326 cur->set_mark(proto); // until proven otherwise 8327 assert(cur->is_oop(), "Should be an oop"); 8328 bool res = work_q->push(cur); 8329 assert(res, "Bit off more than we can chew?"); 8330 NOT_PRODUCT(n++;) 8331 } 8332 #ifndef PRODUCT 8333 assert(_num_par_pushes >= n, "Too many pops?"); 8334 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8335 #endif 8336 return true; 8337 } 8338 8339 // Single-threaded 8340 void CMSCollector::push_on_overflow_list(oop p) { 8341 NOT_PRODUCT(_num_par_pushes++;) 8342 assert(p->is_oop(), "Not an oop"); 8343 preserve_mark_if_necessary(p); 8344 p->set_mark((markOop)_overflow_list); 8345 _overflow_list = p; 8346 } 8347 8348 // Multi-threaded; use CAS to prepend to overflow list 8349 void CMSCollector::par_push_on_overflow_list(oop p) { 8350 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8351 assert(p->is_oop(), "Not an oop"); 8352 par_preserve_mark_if_necessary(p); 8353 oop observed_overflow_list = _overflow_list; 8354 oop cur_overflow_list; 8355 do { 8356 cur_overflow_list = observed_overflow_list; 8357 if (cur_overflow_list != BUSY) { 8358 p->set_mark(markOop(cur_overflow_list)); 8359 } else { 8360 p->set_mark(NULL); 8361 } 8362 observed_overflow_list = 8363 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8364 } while (cur_overflow_list != observed_overflow_list); 8365 } 8366 #undef BUSY 8367 8368 // Single threaded 8369 // General Note on GrowableArray: pushes may silently fail 8370 // because we are (temporarily) out of C-heap for expanding 8371 // the stack. The problem is quite ubiquitous and affects 8372 // a lot of code in the JVM. The prudent thing for GrowableArray 8373 // to do (for now) is to exit with an error. However, that may 8374 // be too draconian in some cases because the caller may be 8375 // able to recover without much harm. For such cases, we 8376 // should probably introduce a "soft_push" method which returns 8377 // an indication of success or failure with the assumption that 8378 // the caller may be able to recover from a failure; code in 8379 // the VM can then be changed, incrementally, to deal with such 8380 // failures where possible, thus, incrementally hardening the VM 8381 // in such low resource situations. 8382 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8383 _preserved_oop_stack.push(p); 8384 _preserved_mark_stack.push(m); 8385 assert(m == p->mark(), "Mark word changed"); 8386 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8387 "bijection"); 8388 } 8389 8390 // Single threaded 8391 void CMSCollector::preserve_mark_if_necessary(oop p) { 8392 markOop m = p->mark(); 8393 if (m->must_be_preserved(p)) { 8394 preserve_mark_work(p, m); 8395 } 8396 } 8397 8398 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8399 markOop m = p->mark(); 8400 if (m->must_be_preserved(p)) { 8401 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8402 // Even though we read the mark word without holding 8403 // the lock, we are assured that it will not change 8404 // because we "own" this oop, so no other thread can 8405 // be trying to push it on the overflow list; see 8406 // the assertion in preserve_mark_work() that checks 8407 // that m == p->mark(). 8408 preserve_mark_work(p, m); 8409 } 8410 } 8411 8412 // We should be able to do this multi-threaded, 8413 // a chunk of stack being a task (this is 8414 // correct because each oop only ever appears 8415 // once in the overflow list. However, it's 8416 // not very easy to completely overlap this with 8417 // other operations, so will generally not be done 8418 // until all work's been completed. Because we 8419 // expect the preserved oop stack (set) to be small, 8420 // it's probably fine to do this single-threaded. 8421 // We can explore cleverer concurrent/overlapped/parallel 8422 // processing of preserved marks if we feel the 8423 // need for this in the future. Stack overflow should 8424 // be so rare in practice and, when it happens, its 8425 // effect on performance so great that this will 8426 // likely just be in the noise anyway. 8427 void CMSCollector::restore_preserved_marks_if_any() { 8428 assert(SafepointSynchronize::is_at_safepoint(), 8429 "world should be stopped"); 8430 assert(Thread::current()->is_ConcurrentGC_thread() || 8431 Thread::current()->is_VM_thread(), 8432 "should be single-threaded"); 8433 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8434 "bijection"); 8435 8436 while (!_preserved_oop_stack.is_empty()) { 8437 oop p = _preserved_oop_stack.pop(); 8438 assert(p->is_oop(), "Should be an oop"); 8439 assert(_span.contains(p), "oop should be in _span"); 8440 assert(p->mark() == markOopDesc::prototype(), 8441 "Set when taken from overflow list"); 8442 markOop m = _preserved_mark_stack.pop(); 8443 p->set_mark(m); 8444 } 8445 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8446 "stacks were cleared above"); 8447 } 8448 8449 #ifndef PRODUCT 8450 bool CMSCollector::no_preserved_marks() const { 8451 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8452 } 8453 #endif 8454 8455 // Transfer some number of overflown objects to usual marking 8456 // stack. Return true if some objects were transferred. 8457 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8458 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8459 (size_t)ParGCDesiredObjsFromOverflowList); 8460 8461 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8462 assert(_collector->overflow_list_is_empty() || res, 8463 "If list is not empty, we should have taken something"); 8464 assert(!res || !_mark_stack->isEmpty(), 8465 "If we took something, it should now be on our stack"); 8466 return res; 8467 } 8468 8469 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8470 size_t res = _sp->block_size_no_stall(addr, _collector); 8471 if (_sp->block_is_obj(addr)) { 8472 if (_live_bit_map->isMarked(addr)) { 8473 // It can't have been dead in a previous cycle 8474 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8475 } else { 8476 _dead_bit_map->mark(addr); // mark the dead object 8477 } 8478 } 8479 // Could be 0, if the block size could not be computed without stalling. 8480 return res; 8481 } 8482 8483 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8484 8485 switch (phase) { 8486 case CMSCollector::InitialMarking: 8487 initialize(true /* fullGC */ , 8488 cause /* cause of the GC */, 8489 true /* recordGCBeginTime */, 8490 true /* recordPreGCUsage */, 8491 false /* recordPeakUsage */, 8492 false /* recordPostGCusage */, 8493 true /* recordAccumulatedGCTime */, 8494 false /* recordGCEndTime */, 8495 false /* countCollection */ ); 8496 break; 8497 8498 case CMSCollector::FinalMarking: 8499 initialize(true /* fullGC */ , 8500 cause /* cause of the GC */, 8501 false /* recordGCBeginTime */, 8502 false /* recordPreGCUsage */, 8503 false /* recordPeakUsage */, 8504 false /* recordPostGCusage */, 8505 true /* recordAccumulatedGCTime */, 8506 false /* recordGCEndTime */, 8507 false /* countCollection */ ); 8508 break; 8509 8510 case CMSCollector::Sweeping: 8511 initialize(true /* fullGC */ , 8512 cause /* cause of the GC */, 8513 false /* recordGCBeginTime */, 8514 false /* recordPreGCUsage */, 8515 true /* recordPeakUsage */, 8516 true /* recordPostGCusage */, 8517 false /* recordAccumulatedGCTime */, 8518 true /* recordGCEndTime */, 8519 true /* countCollection */ ); 8520 break; 8521 8522 default: 8523 ShouldNotReachHere(); 8524 } 8525 }