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