1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardGeneration.inline.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/strongRootsScope.hpp" 57 #include "memory/tenuredGeneration.hpp" 58 #include "oops/oop.inline.hpp" 59 #include "prims/jvmtiExport.hpp" 60 #include "runtime/atomic.inline.hpp" 61 #include "runtime/globals_extension.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/orderAccess.inline.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "services/memoryService.hpp" 67 #include "services/runtimeService.hpp" 68 #include "utilities/stack.inline.hpp" 69 #include "utilities/taskqueue.inline.hpp" 70 71 // statics 72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 73 bool CMSCollector::_full_gc_requested = false; 74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 75 76 ////////////////////////////////////////////////////////////////// 77 // In support of CMS/VM thread synchronization 78 ////////////////////////////////////////////////////////////////// 79 // We split use of the CGC_lock into 2 "levels". 80 // The low-level locking is of the usual CGC_lock monitor. We introduce 81 // a higher level "token" (hereafter "CMS token") built on top of the 82 // low level monitor (hereafter "CGC lock"). 83 // The token-passing protocol gives priority to the VM thread. The 84 // CMS-lock doesn't provide any fairness guarantees, but clients 85 // should ensure that it is only held for very short, bounded 86 // durations. 87 // 88 // When either of the CMS thread or the VM thread is involved in 89 // collection operations during which it does not want the other 90 // thread to interfere, it obtains the CMS token. 91 // 92 // If either thread tries to get the token while the other has 93 // it, that thread waits. However, if the VM thread and CMS thread 94 // both want the token, then the VM thread gets priority while the 95 // CMS thread waits. This ensures, for instance, that the "concurrent" 96 // phases of the CMS thread's work do not block out the VM thread 97 // for long periods of time as the CMS thread continues to hog 98 // the token. (See bug 4616232). 99 // 100 // The baton-passing functions are, however, controlled by the 101 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 102 // and here the low-level CMS lock, not the high level token, 103 // ensures mutual exclusion. 104 // 105 // Two important conditions that we have to satisfy: 106 // 1. if a thread does a low-level wait on the CMS lock, then it 107 // relinquishes the CMS token if it were holding that token 108 // when it acquired the low-level CMS lock. 109 // 2. any low-level notifications on the low-level lock 110 // should only be sent when a thread has relinquished the token. 111 // 112 // In the absence of either property, we'd have potential deadlock. 113 // 114 // We protect each of the CMS (concurrent and sequential) phases 115 // with the CMS _token_, not the CMS _lock_. 116 // 117 // The only code protected by CMS lock is the token acquisition code 118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 119 // baton-passing code. 120 // 121 // Unfortunately, i couldn't come up with a good abstraction to factor and 122 // hide the naked CGC_lock manipulation in the baton-passing code 123 // further below. That's something we should try to do. Also, the proof 124 // of correctness of this 2-level locking scheme is far from obvious, 125 // and potentially quite slippery. We have an uneasy suspicion, for instance, 126 // that there may be a theoretical possibility of delay/starvation in the 127 // low-level lock/wait/notify scheme used for the baton-passing because of 128 // potential interference with the priority scheme embodied in the 129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 130 // invocation further below and marked with "XXX 20011219YSR". 131 // Indeed, as we note elsewhere, this may become yet more slippery 132 // in the presence of multiple CMS and/or multiple VM threads. XXX 133 134 class CMSTokenSync: public StackObj { 135 private: 136 bool _is_cms_thread; 137 public: 138 CMSTokenSync(bool is_cms_thread): 139 _is_cms_thread(is_cms_thread) { 140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 141 "Incorrect argument to constructor"); 142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 143 } 144 145 ~CMSTokenSync() { 146 assert(_is_cms_thread ? 147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 149 "Incorrect state"); 150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 151 } 152 }; 153 154 // Convenience class that does a CMSTokenSync, and then acquires 155 // upto three locks. 156 class CMSTokenSyncWithLocks: public CMSTokenSync { 157 private: 158 // Note: locks are acquired in textual declaration order 159 // and released in the opposite order 160 MutexLockerEx _locker1, _locker2, _locker3; 161 public: 162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 164 CMSTokenSync(is_cms_thread), 165 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 166 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 167 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 168 { } 169 }; 170 171 172 ////////////////////////////////////////////////////////////////// 173 // Concurrent Mark-Sweep Generation ///////////////////////////// 174 ////////////////////////////////////////////////////////////////// 175 176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 177 178 // This struct contains per-thread things necessary to support parallel 179 // young-gen collection. 180 class CMSParGCThreadState: public CHeapObj<mtGC> { 181 public: 182 CFLS_LAB lab; 183 PromotionInfo promo; 184 185 // Constructor. 186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 187 promo.setSpace(cfls); 188 } 189 }; 190 191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 192 ReservedSpace rs, size_t initial_byte_size, 193 CardTableRS* ct, bool use_adaptive_freelists, 194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 195 CardGeneration(rs, initial_byte_size, ct), 196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 (int) ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 307 308 // Generation Counters - generation 1, 1 subspace 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, 310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 311 312 _space_counters = new GSpaceCounters(gen_name, 0, 313 _virtual_space.reserved_size(), 314 this, _gen_counters); 315 } 316 317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 318 _cms_gen(cms_gen) 319 { 320 assert(alpha <= 100, "bad value"); 321 _saved_alpha = alpha; 322 323 // Initialize the alphas to the bootstrap value of 100. 324 _gc0_alpha = _cms_alpha = 100; 325 326 _cms_begin_time.update(); 327 _cms_end_time.update(); 328 329 _gc0_duration = 0.0; 330 _gc0_period = 0.0; 331 _gc0_promoted = 0; 332 333 _cms_duration = 0.0; 334 _cms_period = 0.0; 335 _cms_allocated = 0; 336 337 _cms_used_at_gc0_begin = 0; 338 _cms_used_at_gc0_end = 0; 339 _allow_duty_cycle_reduction = false; 340 _valid_bits = 0; 341 } 342 343 double CMSStats::cms_free_adjustment_factor(size_t free) const { 344 // TBD: CR 6909490 345 return 1.0; 346 } 347 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 349 } 350 351 // If promotion failure handling is on use 352 // the padded average size of the promotion for each 353 // young generation collection. 354 double CMSStats::time_until_cms_gen_full() const { 355 size_t cms_free = _cms_gen->cmsSpace()->free(); 356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 359 if (cms_free > expected_promotion) { 360 // Start a cms collection if there isn't enough space to promote 361 // for the next minor collection. Use the padded average as 362 // a safety factor. 363 cms_free -= expected_promotion; 364 365 // Adjust by the safety factor. 366 double cms_free_dbl = (double)cms_free; 367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 368 // Apply a further correction factor which tries to adjust 369 // for recent occurance of concurrent mode failures. 370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 371 cms_free_dbl = cms_free_dbl * cms_adjustment; 372 373 if (PrintGCDetails && Verbose) { 374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 376 cms_free, expected_promotion); 377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 378 cms_free_dbl, cms_consumption_rate() + 1.0); 379 } 380 // Add 1 in case the consumption rate goes to zero. 381 return cms_free_dbl / (cms_consumption_rate() + 1.0); 382 } 383 return 0.0; 384 } 385 386 // Compare the duration of the cms collection to the 387 // time remaining before the cms generation is empty. 388 // Note that the time from the start of the cms collection 389 // to the start of the cms sweep (less than the total 390 // duration of the cms collection) can be used. This 391 // has been tried and some applications experienced 392 // promotion failures early in execution. This was 393 // possibly because the averages were not accurate 394 // enough at the beginning. 395 double CMSStats::time_until_cms_start() const { 396 // We add "gc0_period" to the "work" calculation 397 // below because this query is done (mostly) at the 398 // end of a scavenge, so we need to conservatively 399 // account for that much possible delay 400 // in the query so as to avoid concurrent mode failures 401 // due to starting the collection just a wee bit too 402 // late. 403 double work = cms_duration() + gc0_period(); 404 double deadline = time_until_cms_gen_full(); 405 // If a concurrent mode failure occurred recently, we want to be 406 // more conservative and halve our expected time_until_cms_gen_full() 407 if (work > deadline) { 408 if (Verbose && PrintGCDetails) { 409 gclog_or_tty->print( 410 " CMSCollector: collect because of anticipated promotion " 411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 412 gc0_period(), time_until_cms_gen_full()); 413 } 414 return 0.0; 415 } 416 return work - deadline; 417 } 418 419 #ifndef PRODUCT 420 void CMSStats::print_on(outputStream *st) const { 421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 423 gc0_duration(), gc0_period(), gc0_promoted()); 424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 425 cms_duration(), cms_period(), cms_allocated()); 426 st->print(",cms_since_beg=%g,cms_since_end=%g", 427 cms_time_since_begin(), cms_time_since_end()); 428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 430 431 if (valid()) { 432 st->print(",promo_rate=%g,cms_alloc_rate=%g", 433 promotion_rate(), cms_allocation_rate()); 434 st->print(",cms_consumption_rate=%g,time_until_full=%g", 435 cms_consumption_rate(), time_until_cms_gen_full()); 436 } 437 st->print(" "); 438 } 439 #endif // #ifndef PRODUCT 440 441 CMSCollector::CollectorState CMSCollector::_collectorState = 442 CMSCollector::Idling; 443 bool CMSCollector::_foregroundGCIsActive = false; 444 bool CMSCollector::_foregroundGCShouldWait = false; 445 446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 447 CardTableRS* ct, 448 ConcurrentMarkSweepPolicy* cp): 449 _cmsGen(cmsGen), 450 _ct(ct), 451 _ref_processor(NULL), // will be set later 452 _conc_workers(NULL), // may be set later 453 _abort_preclean(false), 454 _start_sampling(false), 455 _between_prologue_and_epilogue(false), 456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 458 -1 /* lock-free */, "No_lock" /* dummy */), 459 _modUnionClosurePar(&_modUnionTable), 460 // Adjust my span to cover old (cms) gen 461 _span(cmsGen->reserved()), 462 // Construct the is_alive_closure with _span & markBitMap 463 _is_alive_closure(_span, &_markBitMap), 464 _restart_addr(NULL), 465 _overflow_list(NULL), 466 _stats(cmsGen), 467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 468 //verify that this lock should be acquired with safepoint check. 469 Monitor::_safepoint_check_sometimes)), 470 _eden_chunk_array(NULL), // may be set in ctor body 471 _eden_chunk_capacity(0), // -- ditto -- 472 _eden_chunk_index(0), // -- ditto -- 473 _survivor_plab_array(NULL), // -- ditto -- 474 _survivor_chunk_array(NULL), // -- ditto -- 475 _survivor_chunk_capacity(0), // -- ditto -- 476 _survivor_chunk_index(0), // -- ditto -- 477 _ser_pmc_preclean_ovflw(0), 478 _ser_kac_preclean_ovflw(0), 479 _ser_pmc_remark_ovflw(0), 480 _par_pmc_remark_ovflw(0), 481 _ser_kac_ovflw(0), 482 _par_kac_ovflw(0), 483 #ifndef PRODUCT 484 _num_par_pushes(0), 485 #endif 486 _collection_count_start(0), 487 _verifying(false), 488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 489 _completed_initialization(false), 490 _collector_policy(cp), 491 _should_unload_classes(CMSClassUnloadingEnabled), 492 _concurrent_cycles_since_last_unload(0), 493 _roots_scanning_options(GenCollectedHeap::SO_None), 494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 498 _cms_start_registered(false) 499 { 500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 501 ExplicitGCInvokesConcurrent = true; 502 } 503 // Now expand the span and allocate the collection support structures 504 // (MUT, marking bit map etc.) to cover both generations subject to 505 // collection. 506 507 // For use by dirty card to oop closures. 508 _cmsGen->cmsSpace()->set_collector(this); 509 510 // Allocate MUT and marking bit map 511 { 512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 513 if (!_markBitMap.allocate(_span)) { 514 warning("Failed to allocate CMS Bit Map"); 515 return; 516 } 517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 518 } 519 { 520 _modUnionTable.allocate(_span); 521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 522 } 523 524 if (!_markStack.allocate(MarkStackSize)) { 525 warning("Failed to allocate CMS Marking Stack"); 526 return; 527 } 528 529 // Support for multi-threaded concurrent phases 530 if (CMSConcurrentMTEnabled) { 531 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 532 // just for now 533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 534 } 535 if (ConcGCThreads > 1) { 536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 537 ConcGCThreads, true); 538 if (_conc_workers == NULL) { 539 warning("GC/CMS: _conc_workers allocation failure: " 540 "forcing -CMSConcurrentMTEnabled"); 541 CMSConcurrentMTEnabled = false; 542 } else { 543 _conc_workers->initialize_workers(); 544 } 545 } else { 546 CMSConcurrentMTEnabled = false; 547 } 548 } 549 if (!CMSConcurrentMTEnabled) { 550 ConcGCThreads = 0; 551 } else { 552 // Turn off CMSCleanOnEnter optimization temporarily for 553 // the MT case where it's not fixed yet; see 6178663. 554 CMSCleanOnEnter = false; 555 } 556 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 557 "Inconsistency"); 558 559 // Parallel task queues; these are shared for the 560 // concurrent and stop-world phases of CMS, but 561 // are not shared with parallel scavenge (ParNew). 562 { 563 uint i; 564 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 565 566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 567 || ParallelRefProcEnabled) 568 && num_queues > 0) { 569 _task_queues = new OopTaskQueueSet(num_queues); 570 if (_task_queues == NULL) { 571 warning("task_queues allocation failure."); 572 return; 573 } 574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 576 for (i = 0; i < num_queues; i++) { 577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 578 if (q == NULL) { 579 warning("work_queue allocation failure."); 580 return; 581 } 582 _task_queues->register_queue(i, q); 583 } 584 for (i = 0; i < num_queues; i++) { 585 _task_queues->queue(i)->initialize(); 586 _hash_seed[i] = 17; // copied from ParNew 587 } 588 } 589 } 590 591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 592 593 // Clip CMSBootstrapOccupancy between 0 and 100. 594 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 595 596 // Now tell CMS generations the identity of their collector 597 ConcurrentMarkSweepGeneration::set_collector(this); 598 599 // Create & start a CMS thread for this CMS collector 600 _cmsThread = ConcurrentMarkSweepThread::start(this); 601 assert(cmsThread() != NULL, "CMS Thread should have been created"); 602 assert(cmsThread()->collector() == this, 603 "CMS Thread should refer to this gen"); 604 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 605 606 // Support for parallelizing young gen rescan 607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 609 _young_gen = (ParNewGeneration*)gch->young_gen(); 610 if (gch->supports_inline_contig_alloc()) { 611 _top_addr = gch->top_addr(); 612 _end_addr = gch->end_addr(); 613 assert(_young_gen != NULL, "no _young_gen"); 614 _eden_chunk_index = 0; 615 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 617 } 618 619 // Support for parallelizing survivor space rescan 620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 621 const size_t max_plab_samples = 622 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 623 624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 627 _survivor_chunk_capacity = 2*max_plab_samples; 628 for (uint i = 0; i < ParallelGCThreads; i++) { 629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 631 assert(cur->end() == 0, "Should be 0"); 632 assert(cur->array() == vec, "Should be vec"); 633 assert(cur->capacity() == max_plab_samples, "Error"); 634 } 635 } 636 637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 638 _gc_counters = new CollectorCounters("CMS", 1); 639 _completed_initialization = true; 640 _inter_sweep_timer.start(); // start of time 641 } 642 643 size_t CMSCollector::plab_sample_minimum_size() { 644 // The default value of MinTLABSize is 2k, but there is 645 // no way to get the default value if the flag has been overridden. 646 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 647 } 648 649 const char* ConcurrentMarkSweepGeneration::name() const { 650 return "concurrent mark-sweep generation"; 651 } 652 void ConcurrentMarkSweepGeneration::update_counters() { 653 if (UsePerfData) { 654 _space_counters->update_all(); 655 _gen_counters->update_all(); 656 } 657 } 658 659 // this is an optimized version of update_counters(). it takes the 660 // used value as a parameter rather than computing it. 661 // 662 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 663 if (UsePerfData) { 664 _space_counters->update_used(used); 665 _space_counters->update_capacity(); 666 _gen_counters->update_all(); 667 } 668 } 669 670 void ConcurrentMarkSweepGeneration::print() const { 671 Generation::print(); 672 cmsSpace()->print(); 673 } 674 675 #ifndef PRODUCT 676 void ConcurrentMarkSweepGeneration::print_statistics() { 677 cmsSpace()->printFLCensus(0); 678 } 679 #endif 680 681 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 682 GenCollectedHeap* gch = GenCollectedHeap::heap(); 683 if (PrintGCDetails) { 684 // I didn't want to change the logging when removing the level concept, 685 // but I guess this logging could say "old" or something instead of "1". 686 assert(this == gch->old_gen(), 687 "The CMS generation should be the old generation"); 688 uint level = 1; 689 if (Verbose) { 690 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 691 level, short_name(), s, used(), capacity()); 692 } else { 693 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 694 level, short_name(), s, used() / K, capacity() / K); 695 } 696 } 697 if (Verbose) { 698 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 699 gch->used(), gch->capacity()); 700 } else { 701 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 702 gch->used() / K, gch->capacity() / K); 703 } 704 } 705 706 size_t 707 ConcurrentMarkSweepGeneration::contiguous_available() const { 708 // dld proposes an improvement in precision here. If the committed 709 // part of the space ends in a free block we should add that to 710 // uncommitted size in the calculation below. Will make this 711 // change later, staying with the approximation below for the 712 // time being. -- ysr. 713 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 714 } 715 716 size_t 717 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 718 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 719 } 720 721 size_t ConcurrentMarkSweepGeneration::max_available() const { 722 return free() + _virtual_space.uncommitted_size(); 723 } 724 725 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 726 size_t available = max_available(); 727 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 728 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 729 if (Verbose && PrintGCDetails) { 730 gclog_or_tty->print_cr( 731 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 732 "max_promo("SIZE_FORMAT")", 733 res? "":" not", available, res? ">=":"<", 734 av_promo, max_promotion_in_bytes); 735 } 736 return res; 737 } 738 739 // At a promotion failure dump information on block layout in heap 740 // (cms old generation). 741 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 742 if (CMSDumpAtPromotionFailure) { 743 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 744 } 745 } 746 747 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 748 // Clear the promotion information. These pointers can be adjusted 749 // along with all the other pointers into the heap but 750 // compaction is expected to be a rare event with 751 // a heap using cms so don't do it without seeing the need. 752 for (uint i = 0; i < ParallelGCThreads; i++) { 753 _par_gc_thread_states[i]->promo.reset(); 754 } 755 } 756 757 void ConcurrentMarkSweepGeneration::compute_new_size() { 758 assert_locked_or_safepoint(Heap_lock); 759 760 // If incremental collection failed, we just want to expand 761 // to the limit. 762 if (incremental_collection_failed()) { 763 clear_incremental_collection_failed(); 764 grow_to_reserved(); 765 return; 766 } 767 768 // The heap has been compacted but not reset yet. 769 // Any metric such as free() or used() will be incorrect. 770 771 CardGeneration::compute_new_size(); 772 773 // Reset again after a possible resizing 774 if (did_compact()) { 775 cmsSpace()->reset_after_compaction(); 776 } 777 } 778 779 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 780 assert_locked_or_safepoint(Heap_lock); 781 782 // If incremental collection failed, we just want to expand 783 // to the limit. 784 if (incremental_collection_failed()) { 785 clear_incremental_collection_failed(); 786 grow_to_reserved(); 787 return; 788 } 789 790 double free_percentage = ((double) free()) / capacity(); 791 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 792 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 793 794 // compute expansion delta needed for reaching desired free percentage 795 if (free_percentage < desired_free_percentage) { 796 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 797 assert(desired_capacity >= capacity(), "invalid expansion size"); 798 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 799 if (PrintGCDetails && Verbose) { 800 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 801 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 802 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 803 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage); 804 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage); 805 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity() / 1000); 806 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, desired_capacity / 1000); 807 GenCollectedHeap* gch = GenCollectedHeap::heap(); 808 assert(this == gch->_old_gen, "The CMS generation should always be the old generation"); 809 size_t young_size = gch->_young_gen->capacity(); 810 gclog_or_tty->print_cr(" Young gen size "SIZE_FORMAT, young_size / 1000); 811 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 812 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, contiguous_available() / 1000); 813 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", expand_bytes); 814 } 815 // safe if expansion fails 816 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 817 if (PrintGCDetails && Verbose) { 818 gclog_or_tty->print_cr(" Expanded free fraction %f", 819 ((double) free()) / capacity()); 820 } 821 } else { 822 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 823 assert(desired_capacity <= capacity(), "invalid expansion size"); 824 size_t shrink_bytes = capacity() - desired_capacity; 825 // Don't shrink unless the delta is greater than the minimum shrink we want 826 if (shrink_bytes >= MinHeapDeltaBytes) { 827 shrink_free_list_by(shrink_bytes); 828 } 829 } 830 } 831 832 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 833 return cmsSpace()->freelistLock(); 834 } 835 836 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 837 bool tlab) { 838 CMSSynchronousYieldRequest yr; 839 MutexLockerEx x(freelistLock(), 840 Mutex::_no_safepoint_check_flag); 841 return have_lock_and_allocate(size, tlab); 842 } 843 844 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 845 bool tlab /* ignored */) { 846 assert_lock_strong(freelistLock()); 847 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 848 HeapWord* res = cmsSpace()->allocate(adjustedSize); 849 // Allocate the object live (grey) if the background collector has 850 // started marking. This is necessary because the marker may 851 // have passed this address and consequently this object will 852 // not otherwise be greyed and would be incorrectly swept up. 853 // Note that if this object contains references, the writing 854 // of those references will dirty the card containing this object 855 // allowing the object to be blackened (and its references scanned) 856 // either during a preclean phase or at the final checkpoint. 857 if (res != NULL) { 858 // We may block here with an uninitialized object with 859 // its mark-bit or P-bits not yet set. Such objects need 860 // to be safely navigable by block_start(). 861 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 862 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 863 collector()->direct_allocated(res, adjustedSize); 864 _direct_allocated_words += adjustedSize; 865 // allocation counters 866 NOT_PRODUCT( 867 _numObjectsAllocated++; 868 _numWordsAllocated += (int)adjustedSize; 869 ) 870 } 871 return res; 872 } 873 874 // In the case of direct allocation by mutators in a generation that 875 // is being concurrently collected, the object must be allocated 876 // live (grey) if the background collector has started marking. 877 // This is necessary because the marker may 878 // have passed this address and consequently this object will 879 // not otherwise be greyed and would be incorrectly swept up. 880 // Note that if this object contains references, the writing 881 // of those references will dirty the card containing this object 882 // allowing the object to be blackened (and its references scanned) 883 // either during a preclean phase or at the final checkpoint. 884 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 885 assert(_markBitMap.covers(start, size), "Out of bounds"); 886 if (_collectorState >= Marking) { 887 MutexLockerEx y(_markBitMap.lock(), 888 Mutex::_no_safepoint_check_flag); 889 // [see comments preceding SweepClosure::do_blk() below for details] 890 // 891 // Can the P-bits be deleted now? JJJ 892 // 893 // 1. need to mark the object as live so it isn't collected 894 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 895 // 3. need to mark the end of the object so marking, precleaning or sweeping 896 // can skip over uninitialized or unparsable objects. An allocated 897 // object is considered uninitialized for our purposes as long as 898 // its klass word is NULL. All old gen objects are parsable 899 // as soon as they are initialized.) 900 _markBitMap.mark(start); // object is live 901 _markBitMap.mark(start + 1); // object is potentially uninitialized? 902 _markBitMap.mark(start + size - 1); 903 // mark end of object 904 } 905 // check that oop looks uninitialized 906 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 907 } 908 909 void CMSCollector::promoted(bool par, HeapWord* start, 910 bool is_obj_array, size_t obj_size) { 911 assert(_markBitMap.covers(start), "Out of bounds"); 912 // See comment in direct_allocated() about when objects should 913 // be allocated live. 914 if (_collectorState >= Marking) { 915 // we already hold the marking bit map lock, taken in 916 // the prologue 917 if (par) { 918 _markBitMap.par_mark(start); 919 } else { 920 _markBitMap.mark(start); 921 } 922 // We don't need to mark the object as uninitialized (as 923 // in direct_allocated above) because this is being done with the 924 // world stopped and the object will be initialized by the 925 // time the marking, precleaning or sweeping get to look at it. 926 // But see the code for copying objects into the CMS generation, 927 // where we need to ensure that concurrent readers of the 928 // block offset table are able to safely navigate a block that 929 // is in flux from being free to being allocated (and in 930 // transition while being copied into) and subsequently 931 // becoming a bona-fide object when the copy/promotion is complete. 932 assert(SafepointSynchronize::is_at_safepoint(), 933 "expect promotion only at safepoints"); 934 935 if (_collectorState < Sweeping) { 936 // Mark the appropriate cards in the modUnionTable, so that 937 // this object gets scanned before the sweep. If this is 938 // not done, CMS generation references in the object might 939 // not get marked. 940 // For the case of arrays, which are otherwise precisely 941 // marked, we need to dirty the entire array, not just its head. 942 if (is_obj_array) { 943 // The [par_]mark_range() method expects mr.end() below to 944 // be aligned to the granularity of a bit's representation 945 // in the heap. In the case of the MUT below, that's a 946 // card size. 947 MemRegion mr(start, 948 (HeapWord*)round_to((intptr_t)(start + obj_size), 949 CardTableModRefBS::card_size /* bytes */)); 950 if (par) { 951 _modUnionTable.par_mark_range(mr); 952 } else { 953 _modUnionTable.mark_range(mr); 954 } 955 } else { // not an obj array; we can just mark the head 956 if (par) { 957 _modUnionTable.par_mark(start); 958 } else { 959 _modUnionTable.mark(start); 960 } 961 } 962 } 963 } 964 } 965 966 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 967 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 968 // allocate, copy and if necessary update promoinfo -- 969 // delegate to underlying space. 970 assert_lock_strong(freelistLock()); 971 972 #ifndef PRODUCT 973 if (GenCollectedHeap::heap()->promotion_should_fail()) { 974 return NULL; 975 } 976 #endif // #ifndef PRODUCT 977 978 oop res = _cmsSpace->promote(obj, obj_size); 979 if (res == NULL) { 980 // expand and retry 981 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 982 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 983 // Since this is the old generation, we don't try to promote 984 // into a more senior generation. 985 res = _cmsSpace->promote(obj, obj_size); 986 } 987 if (res != NULL) { 988 // See comment in allocate() about when objects should 989 // be allocated live. 990 assert(obj->is_oop(), "Will dereference klass pointer below"); 991 collector()->promoted(false, // Not parallel 992 (HeapWord*)res, obj->is_objArray(), obj_size); 993 // promotion counters 994 NOT_PRODUCT( 995 _numObjectsPromoted++; 996 _numWordsPromoted += 997 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 998 ) 999 } 1000 return res; 1001 } 1002 1003 1004 // IMPORTANT: Notes on object size recognition in CMS. 1005 // --------------------------------------------------- 1006 // A block of storage in the CMS generation is always in 1007 // one of three states. A free block (FREE), an allocated 1008 // object (OBJECT) whose size() method reports the correct size, 1009 // and an intermediate state (TRANSIENT) in which its size cannot 1010 // be accurately determined. 1011 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1012 // ----------------------------------------------------- 1013 // FREE: klass_word & 1 == 1; mark_word holds block size 1014 // 1015 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1016 // obj->size() computes correct size 1017 // 1018 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1019 // 1020 // STATE IDENTIFICATION: (64 bit+COOPS) 1021 // ------------------------------------ 1022 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1023 // 1024 // OBJECT: klass_word installed; klass_word != 0; 1025 // obj->size() computes correct size 1026 // 1027 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1028 // 1029 // 1030 // STATE TRANSITION DIAGRAM 1031 // 1032 // mut / parnew mut / parnew 1033 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1034 // ^ | 1035 // |------------------------ DEAD <------------------------------------| 1036 // sweep mut 1037 // 1038 // While a block is in TRANSIENT state its size cannot be determined 1039 // so readers will either need to come back later or stall until 1040 // the size can be determined. Note that for the case of direct 1041 // allocation, P-bits, when available, may be used to determine the 1042 // size of an object that may not yet have been initialized. 1043 1044 // Things to support parallel young-gen collection. 1045 oop 1046 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1047 oop old, markOop m, 1048 size_t word_sz) { 1049 #ifndef PRODUCT 1050 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1051 return NULL; 1052 } 1053 #endif // #ifndef PRODUCT 1054 1055 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1056 PromotionInfo* promoInfo = &ps->promo; 1057 // if we are tracking promotions, then first ensure space for 1058 // promotion (including spooling space for saving header if necessary). 1059 // then allocate and copy, then track promoted info if needed. 1060 // When tracking (see PromotionInfo::track()), the mark word may 1061 // be displaced and in this case restoration of the mark word 1062 // occurs in the (oop_since_save_marks_)iterate phase. 1063 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1064 // Out of space for allocating spooling buffers; 1065 // try expanding and allocating spooling buffers. 1066 if (!expand_and_ensure_spooling_space(promoInfo)) { 1067 return NULL; 1068 } 1069 } 1070 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1071 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1072 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1073 if (obj_ptr == NULL) { 1074 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1075 if (obj_ptr == NULL) { 1076 return NULL; 1077 } 1078 } 1079 oop obj = oop(obj_ptr); 1080 OrderAccess::storestore(); 1081 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1082 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1083 // IMPORTANT: See note on object initialization for CMS above. 1084 // Otherwise, copy the object. Here we must be careful to insert the 1085 // klass pointer last, since this marks the block as an allocated object. 1086 // Except with compressed oops it's the mark word. 1087 HeapWord* old_ptr = (HeapWord*)old; 1088 // Restore the mark word copied above. 1089 obj->set_mark(m); 1090 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1091 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1092 OrderAccess::storestore(); 1093 1094 if (UseCompressedClassPointers) { 1095 // Copy gap missed by (aligned) header size calculation below 1096 obj->set_klass_gap(old->klass_gap()); 1097 } 1098 if (word_sz > (size_t)oopDesc::header_size()) { 1099 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1100 obj_ptr + oopDesc::header_size(), 1101 word_sz - oopDesc::header_size()); 1102 } 1103 1104 // Now we can track the promoted object, if necessary. We take care 1105 // to delay the transition from uninitialized to full object 1106 // (i.e., insertion of klass pointer) until after, so that it 1107 // atomically becomes a promoted object. 1108 if (promoInfo->tracking()) { 1109 promoInfo->track((PromotedObject*)obj, old->klass()); 1110 } 1111 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1112 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1113 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1114 1115 // Finally, install the klass pointer (this should be volatile). 1116 OrderAccess::storestore(); 1117 obj->set_klass(old->klass()); 1118 // We should now be able to calculate the right size for this object 1119 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1120 1121 collector()->promoted(true, // parallel 1122 obj_ptr, old->is_objArray(), word_sz); 1123 1124 NOT_PRODUCT( 1125 Atomic::inc_ptr(&_numObjectsPromoted); 1126 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1127 ) 1128 1129 return obj; 1130 } 1131 1132 void 1133 ConcurrentMarkSweepGeneration:: 1134 par_promote_alloc_done(int thread_num) { 1135 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1136 ps->lab.retire(thread_num); 1137 } 1138 1139 void 1140 ConcurrentMarkSweepGeneration:: 1141 par_oop_since_save_marks_iterate_done(int thread_num) { 1142 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1143 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1144 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1145 } 1146 1147 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1148 size_t size, 1149 bool tlab) 1150 { 1151 // We allow a STW collection only if a full 1152 // collection was requested. 1153 return full || should_allocate(size, tlab); // FIX ME !!! 1154 // This and promotion failure handling are connected at the 1155 // hip and should be fixed by untying them. 1156 } 1157 1158 bool CMSCollector::shouldConcurrentCollect() { 1159 if (_full_gc_requested) { 1160 if (Verbose && PrintGCDetails) { 1161 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1162 " gc request (or gc_locker)"); 1163 } 1164 return true; 1165 } 1166 1167 FreelistLocker x(this); 1168 // ------------------------------------------------------------------ 1169 // Print out lots of information which affects the initiation of 1170 // a collection. 1171 if (PrintCMSInitiationStatistics && stats().valid()) { 1172 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1173 gclog_or_tty->stamp(); 1174 gclog_or_tty->cr(); 1175 stats().print_on(gclog_or_tty); 1176 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1177 stats().time_until_cms_gen_full()); 1178 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1179 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1180 _cmsGen->contiguous_available()); 1181 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1182 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1183 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1184 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1185 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1186 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1187 gclog_or_tty->print_cr("metadata initialized %d", 1188 MetaspaceGC::should_concurrent_collect()); 1189 } 1190 // ------------------------------------------------------------------ 1191 1192 // If the estimated time to complete a cms collection (cms_duration()) 1193 // is less than the estimated time remaining until the cms generation 1194 // is full, start a collection. 1195 if (!UseCMSInitiatingOccupancyOnly) { 1196 if (stats().valid()) { 1197 if (stats().time_until_cms_start() == 0.0) { 1198 return true; 1199 } 1200 } else { 1201 // We want to conservatively collect somewhat early in order 1202 // to try and "bootstrap" our CMS/promotion statistics; 1203 // this branch will not fire after the first successful CMS 1204 // collection because the stats should then be valid. 1205 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1206 if (Verbose && PrintGCDetails) { 1207 gclog_or_tty->print_cr( 1208 " CMSCollector: collect for bootstrapping statistics:" 1209 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1210 _bootstrap_occupancy); 1211 } 1212 return true; 1213 } 1214 } 1215 } 1216 1217 // Otherwise, we start a collection cycle if 1218 // old gen want a collection cycle started. Each may use 1219 // an appropriate criterion for making this decision. 1220 // XXX We need to make sure that the gen expansion 1221 // criterion dovetails well with this. XXX NEED TO FIX THIS 1222 if (_cmsGen->should_concurrent_collect()) { 1223 if (Verbose && PrintGCDetails) { 1224 gclog_or_tty->print_cr("CMS old gen initiated"); 1225 } 1226 return true; 1227 } 1228 1229 // We start a collection if we believe an incremental collection may fail; 1230 // this is not likely to be productive in practice because it's probably too 1231 // late anyway. 1232 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1233 assert(gch->collector_policy()->is_generation_policy(), 1234 "You may want to check the correctness of the following"); 1235 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1236 if (Verbose && PrintGCDetails) { 1237 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1238 } 1239 return true; 1240 } 1241 1242 if (MetaspaceGC::should_concurrent_collect()) { 1243 if (Verbose && PrintGCDetails) { 1244 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1245 } 1246 return true; 1247 } 1248 1249 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1250 if (CMSTriggerInterval >= 0) { 1251 if (CMSTriggerInterval == 0) { 1252 // Trigger always 1253 return true; 1254 } 1255 1256 // Check the CMS time since begin (we do not check the stats validity 1257 // as we want to be able to trigger the first CMS cycle as well) 1258 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1259 if (Verbose && PrintGCDetails) { 1260 if (stats().valid()) { 1261 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1262 stats().cms_time_since_begin()); 1263 } else { 1264 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1265 } 1266 } 1267 return true; 1268 } 1269 } 1270 1271 return false; 1272 } 1273 1274 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1275 1276 // Clear _expansion_cause fields of constituent generations 1277 void CMSCollector::clear_expansion_cause() { 1278 _cmsGen->clear_expansion_cause(); 1279 } 1280 1281 // We should be conservative in starting a collection cycle. To 1282 // start too eagerly runs the risk of collecting too often in the 1283 // extreme. To collect too rarely falls back on full collections, 1284 // which works, even if not optimum in terms of concurrent work. 1285 // As a work around for too eagerly collecting, use the flag 1286 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1287 // giving the user an easily understandable way of controlling the 1288 // collections. 1289 // We want to start a new collection cycle if any of the following 1290 // conditions hold: 1291 // . our current occupancy exceeds the configured initiating occupancy 1292 // for this generation, or 1293 // . we recently needed to expand this space and have not, since that 1294 // expansion, done a collection of this generation, or 1295 // . the underlying space believes that it may be a good idea to initiate 1296 // a concurrent collection (this may be based on criteria such as the 1297 // following: the space uses linear allocation and linear allocation is 1298 // going to fail, or there is believed to be excessive fragmentation in 1299 // the generation, etc... or ... 1300 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1301 // the case of the old generation; see CR 6543076): 1302 // we may be approaching a point at which allocation requests may fail because 1303 // we will be out of sufficient free space given allocation rate estimates.] 1304 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1305 1306 assert_lock_strong(freelistLock()); 1307 if (occupancy() > initiating_occupancy()) { 1308 if (PrintGCDetails && Verbose) { 1309 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1310 short_name(), occupancy(), initiating_occupancy()); 1311 } 1312 return true; 1313 } 1314 if (UseCMSInitiatingOccupancyOnly) { 1315 return false; 1316 } 1317 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1318 if (PrintGCDetails && Verbose) { 1319 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1320 short_name()); 1321 } 1322 return true; 1323 } 1324 if (_cmsSpace->should_concurrent_collect()) { 1325 if (PrintGCDetails && Verbose) { 1326 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1327 short_name()); 1328 } 1329 return true; 1330 } 1331 return false; 1332 } 1333 1334 void ConcurrentMarkSweepGeneration::collect(bool full, 1335 bool clear_all_soft_refs, 1336 size_t size, 1337 bool tlab) 1338 { 1339 collector()->collect(full, clear_all_soft_refs, size, tlab); 1340 } 1341 1342 void CMSCollector::collect(bool full, 1343 bool clear_all_soft_refs, 1344 size_t size, 1345 bool tlab) 1346 { 1347 // The following "if" branch is present for defensive reasons. 1348 // In the current uses of this interface, it can be replaced with: 1349 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1350 // But I am not placing that assert here to allow future 1351 // generality in invoking this interface. 1352 if (GC_locker::is_active()) { 1353 // A consistency test for GC_locker 1354 assert(GC_locker::needs_gc(), "Should have been set already"); 1355 // Skip this foreground collection, instead 1356 // expanding the heap if necessary. 1357 // Need the free list locks for the call to free() in compute_new_size() 1358 compute_new_size(); 1359 return; 1360 } 1361 acquire_control_and_collect(full, clear_all_soft_refs); 1362 } 1363 1364 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1365 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1366 unsigned int gc_count = gch->total_full_collections(); 1367 if (gc_count == full_gc_count) { 1368 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1369 _full_gc_requested = true; 1370 _full_gc_cause = cause; 1371 CGC_lock->notify(); // nudge CMS thread 1372 } else { 1373 assert(gc_count > full_gc_count, "Error: causal loop"); 1374 } 1375 } 1376 1377 bool CMSCollector::is_external_interruption() { 1378 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1379 return GCCause::is_user_requested_gc(cause) || 1380 GCCause::is_serviceability_requested_gc(cause); 1381 } 1382 1383 void CMSCollector::report_concurrent_mode_interruption() { 1384 if (is_external_interruption()) { 1385 if (PrintGCDetails) { 1386 gclog_or_tty->print(" (concurrent mode interrupted)"); 1387 } 1388 } else { 1389 if (PrintGCDetails) { 1390 gclog_or_tty->print(" (concurrent mode failure)"); 1391 } 1392 _gc_tracer_cm->report_concurrent_mode_failure(); 1393 } 1394 } 1395 1396 1397 // The foreground and background collectors need to coordinate in order 1398 // to make sure that they do not mutually interfere with CMS collections. 1399 // When a background collection is active, 1400 // the foreground collector may need to take over (preempt) and 1401 // synchronously complete an ongoing collection. Depending on the 1402 // frequency of the background collections and the heap usage 1403 // of the application, this preemption can be seldom or frequent. 1404 // There are only certain 1405 // points in the background collection that the "collection-baton" 1406 // can be passed to the foreground collector. 1407 // 1408 // The foreground collector will wait for the baton before 1409 // starting any part of the collection. The foreground collector 1410 // will only wait at one location. 1411 // 1412 // The background collector will yield the baton before starting a new 1413 // phase of the collection (e.g., before initial marking, marking from roots, 1414 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1415 // of the loop which switches the phases. The background collector does some 1416 // of the phases (initial mark, final re-mark) with the world stopped. 1417 // Because of locking involved in stopping the world, 1418 // the foreground collector should not block waiting for the background 1419 // collector when it is doing a stop-the-world phase. The background 1420 // collector will yield the baton at an additional point just before 1421 // it enters a stop-the-world phase. Once the world is stopped, the 1422 // background collector checks the phase of the collection. If the 1423 // phase has not changed, it proceeds with the collection. If the 1424 // phase has changed, it skips that phase of the collection. See 1425 // the comments on the use of the Heap_lock in collect_in_background(). 1426 // 1427 // Variable used in baton passing. 1428 // _foregroundGCIsActive - Set to true by the foreground collector when 1429 // it wants the baton. The foreground clears it when it has finished 1430 // the collection. 1431 // _foregroundGCShouldWait - Set to true by the background collector 1432 // when it is running. The foreground collector waits while 1433 // _foregroundGCShouldWait is true. 1434 // CGC_lock - monitor used to protect access to the above variables 1435 // and to notify the foreground and background collectors. 1436 // _collectorState - current state of the CMS collection. 1437 // 1438 // The foreground collector 1439 // acquires the CGC_lock 1440 // sets _foregroundGCIsActive 1441 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1442 // various locks acquired in preparation for the collection 1443 // are released so as not to block the background collector 1444 // that is in the midst of a collection 1445 // proceeds with the collection 1446 // clears _foregroundGCIsActive 1447 // returns 1448 // 1449 // The background collector in a loop iterating on the phases of the 1450 // collection 1451 // acquires the CGC_lock 1452 // sets _foregroundGCShouldWait 1453 // if _foregroundGCIsActive is set 1454 // clears _foregroundGCShouldWait, notifies _CGC_lock 1455 // waits on _CGC_lock for _foregroundGCIsActive to become false 1456 // and exits the loop. 1457 // otherwise 1458 // proceed with that phase of the collection 1459 // if the phase is a stop-the-world phase, 1460 // yield the baton once more just before enqueueing 1461 // the stop-world CMS operation (executed by the VM thread). 1462 // returns after all phases of the collection are done 1463 // 1464 1465 void CMSCollector::acquire_control_and_collect(bool full, 1466 bool clear_all_soft_refs) { 1467 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1468 assert(!Thread::current()->is_ConcurrentGC_thread(), 1469 "shouldn't try to acquire control from self!"); 1470 1471 // Start the protocol for acquiring control of the 1472 // collection from the background collector (aka CMS thread). 1473 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1474 "VM thread should have CMS token"); 1475 // Remember the possibly interrupted state of an ongoing 1476 // concurrent collection 1477 CollectorState first_state = _collectorState; 1478 1479 // Signal to a possibly ongoing concurrent collection that 1480 // we want to do a foreground collection. 1481 _foregroundGCIsActive = true; 1482 1483 // release locks and wait for a notify from the background collector 1484 // releasing the locks in only necessary for phases which 1485 // do yields to improve the granularity of the collection. 1486 assert_lock_strong(bitMapLock()); 1487 // We need to lock the Free list lock for the space that we are 1488 // currently collecting. 1489 assert(haveFreelistLocks(), "Must be holding free list locks"); 1490 bitMapLock()->unlock(); 1491 releaseFreelistLocks(); 1492 { 1493 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1494 if (_foregroundGCShouldWait) { 1495 // We are going to be waiting for action for the CMS thread; 1496 // it had better not be gone (for instance at shutdown)! 1497 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1498 "CMS thread must be running"); 1499 // Wait here until the background collector gives us the go-ahead 1500 ConcurrentMarkSweepThread::clear_CMS_flag( 1501 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1502 // Get a possibly blocked CMS thread going: 1503 // Note that we set _foregroundGCIsActive true above, 1504 // without protection of the CGC_lock. 1505 CGC_lock->notify(); 1506 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1507 "Possible deadlock"); 1508 while (_foregroundGCShouldWait) { 1509 // wait for notification 1510 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1511 // Possibility of delay/starvation here, since CMS token does 1512 // not know to give priority to VM thread? Actually, i think 1513 // there wouldn't be any delay/starvation, but the proof of 1514 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1515 } 1516 ConcurrentMarkSweepThread::set_CMS_flag( 1517 ConcurrentMarkSweepThread::CMS_vm_has_token); 1518 } 1519 } 1520 // The CMS_token is already held. Get back the other locks. 1521 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1522 "VM thread should have CMS token"); 1523 getFreelistLocks(); 1524 bitMapLock()->lock_without_safepoint_check(); 1525 if (TraceCMSState) { 1526 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1527 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1528 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1529 } 1530 1531 // Inform cms gen if this was due to partial collection failing. 1532 // The CMS gen may use this fact to determine its expansion policy. 1533 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1534 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1535 assert(!_cmsGen->incremental_collection_failed(), 1536 "Should have been noticed, reacted to and cleared"); 1537 _cmsGen->set_incremental_collection_failed(); 1538 } 1539 1540 if (first_state > Idling) { 1541 report_concurrent_mode_interruption(); 1542 } 1543 1544 set_did_compact(true); 1545 1546 // If the collection is being acquired from the background 1547 // collector, there may be references on the discovered 1548 // references lists. Abandon those references, since some 1549 // of them may have become unreachable after concurrent 1550 // discovery; the STW compacting collector will redo discovery 1551 // more precisely, without being subject to floating garbage. 1552 // Leaving otherwise unreachable references in the discovered 1553 // lists would require special handling. 1554 ref_processor()->disable_discovery(); 1555 ref_processor()->abandon_partial_discovery(); 1556 ref_processor()->verify_no_references_recorded(); 1557 1558 if (first_state > Idling) { 1559 save_heap_summary(); 1560 } 1561 1562 do_compaction_work(clear_all_soft_refs); 1563 1564 // Has the GC time limit been exceeded? 1565 size_t max_eden_size = _young_gen->max_capacity() - 1566 _young_gen->to()->capacity() - 1567 _young_gen->from()->capacity(); 1568 GCCause::Cause gc_cause = gch->gc_cause(); 1569 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1570 _young_gen->eden()->used(), 1571 _cmsGen->max_capacity(), 1572 max_eden_size, 1573 full, 1574 gc_cause, 1575 gch->collector_policy()); 1576 1577 // Reset the expansion cause, now that we just completed 1578 // a collection cycle. 1579 clear_expansion_cause(); 1580 _foregroundGCIsActive = false; 1581 return; 1582 } 1583 1584 // Resize the tenured generation 1585 // after obtaining the free list locks for the 1586 // two generations. 1587 void CMSCollector::compute_new_size() { 1588 assert_locked_or_safepoint(Heap_lock); 1589 FreelistLocker z(this); 1590 MetaspaceGC::compute_new_size(); 1591 _cmsGen->compute_new_size_free_list(); 1592 } 1593 1594 // A work method used by the foreground collector to do 1595 // a mark-sweep-compact. 1596 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1598 1599 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1600 gc_timer->register_gc_start(); 1601 1602 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1603 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1604 1605 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1606 1607 // Temporarily widen the span of the weak reference processing to 1608 // the entire heap. 1609 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1610 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1611 // Temporarily, clear the "is_alive_non_header" field of the 1612 // reference processor. 1613 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1614 // Temporarily make reference _processing_ single threaded (non-MT). 1615 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1616 // Temporarily make refs discovery atomic 1617 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1618 // Temporarily make reference _discovery_ single threaded (non-MT) 1619 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1620 1621 ref_processor()->set_enqueuing_is_done(false); 1622 ref_processor()->enable_discovery(); 1623 ref_processor()->setup_policy(clear_all_soft_refs); 1624 // If an asynchronous collection finishes, the _modUnionTable is 1625 // all clear. If we are assuming the collection from an asynchronous 1626 // collection, clear the _modUnionTable. 1627 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1628 "_modUnionTable should be clear if the baton was not passed"); 1629 _modUnionTable.clear_all(); 1630 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1631 "mod union for klasses should be clear if the baton was passed"); 1632 _ct->klass_rem_set()->clear_mod_union(); 1633 1634 // We must adjust the allocation statistics being maintained 1635 // in the free list space. We do so by reading and clearing 1636 // the sweep timer and updating the block flux rate estimates below. 1637 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1638 if (_inter_sweep_timer.is_active()) { 1639 _inter_sweep_timer.stop(); 1640 // Note that we do not use this sample to update the _inter_sweep_estimate. 1641 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1642 _inter_sweep_estimate.padded_average(), 1643 _intra_sweep_estimate.padded_average()); 1644 } 1645 1646 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1647 #ifdef ASSERT 1648 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1649 size_t free_size = cms_space->free(); 1650 assert(free_size == 1651 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1652 * HeapWordSize, 1653 "All the free space should be compacted into one chunk at top"); 1654 assert(cms_space->dictionary()->total_chunk_size( 1655 debug_only(cms_space->freelistLock())) == 0 || 1656 cms_space->totalSizeInIndexedFreeLists() == 0, 1657 "All the free space should be in a single chunk"); 1658 size_t num = cms_space->totalCount(); 1659 assert((free_size == 0 && num == 0) || 1660 (free_size > 0 && (num == 1 || num == 2)), 1661 "There should be at most 2 free chunks after compaction"); 1662 #endif // ASSERT 1663 _collectorState = Resetting; 1664 assert(_restart_addr == NULL, 1665 "Should have been NULL'd before baton was passed"); 1666 reset(false /* == !concurrent */); 1667 _cmsGen->reset_after_compaction(); 1668 _concurrent_cycles_since_last_unload = 0; 1669 1670 // Clear any data recorded in the PLAB chunk arrays. 1671 if (_survivor_plab_array != NULL) { 1672 reset_survivor_plab_arrays(); 1673 } 1674 1675 // Adjust the per-size allocation stats for the next epoch. 1676 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1677 // Restart the "inter sweep timer" for the next epoch. 1678 _inter_sweep_timer.reset(); 1679 _inter_sweep_timer.start(); 1680 1681 gc_timer->register_gc_end(); 1682 1683 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1684 1685 // For a mark-sweep-compact, compute_new_size() will be called 1686 // in the heap's do_collection() method. 1687 } 1688 1689 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1690 ContiguousSpace* eden_space = _young_gen->eden(); 1691 ContiguousSpace* from_space = _young_gen->from(); 1692 ContiguousSpace* to_space = _young_gen->to(); 1693 // Eden 1694 if (_eden_chunk_array != NULL) { 1695 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1696 p2i(eden_space->bottom()), p2i(eden_space->top()), 1697 p2i(eden_space->end()), eden_space->capacity()); 1698 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1699 "_eden_chunk_capacity=" SIZE_FORMAT, 1700 _eden_chunk_index, _eden_chunk_capacity); 1701 for (size_t i = 0; i < _eden_chunk_index; i++) { 1702 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1703 i, p2i(_eden_chunk_array[i])); 1704 } 1705 } 1706 // Survivor 1707 if (_survivor_chunk_array != NULL) { 1708 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1709 p2i(from_space->bottom()), p2i(from_space->top()), 1710 p2i(from_space->end()), from_space->capacity()); 1711 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1712 "_survivor_chunk_capacity=" SIZE_FORMAT, 1713 _survivor_chunk_index, _survivor_chunk_capacity); 1714 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1715 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1716 i, p2i(_survivor_chunk_array[i])); 1717 } 1718 } 1719 } 1720 1721 void CMSCollector::getFreelistLocks() const { 1722 // Get locks for all free lists in all generations that this 1723 // collector is responsible for 1724 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1725 } 1726 1727 void CMSCollector::releaseFreelistLocks() const { 1728 // Release locks for all free lists in all generations that this 1729 // collector is responsible for 1730 _cmsGen->freelistLock()->unlock(); 1731 } 1732 1733 bool CMSCollector::haveFreelistLocks() const { 1734 // Check locks for all free lists in all generations that this 1735 // collector is responsible for 1736 assert_lock_strong(_cmsGen->freelistLock()); 1737 PRODUCT_ONLY(ShouldNotReachHere()); 1738 return true; 1739 } 1740 1741 // A utility class that is used by the CMS collector to 1742 // temporarily "release" the foreground collector from its 1743 // usual obligation to wait for the background collector to 1744 // complete an ongoing phase before proceeding. 1745 class ReleaseForegroundGC: public StackObj { 1746 private: 1747 CMSCollector* _c; 1748 public: 1749 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1750 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1751 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1752 // allow a potentially blocked foreground collector to proceed 1753 _c->_foregroundGCShouldWait = false; 1754 if (_c->_foregroundGCIsActive) { 1755 CGC_lock->notify(); 1756 } 1757 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1758 "Possible deadlock"); 1759 } 1760 1761 ~ReleaseForegroundGC() { 1762 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1763 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1764 _c->_foregroundGCShouldWait = true; 1765 } 1766 }; 1767 1768 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1769 assert(Thread::current()->is_ConcurrentGC_thread(), 1770 "A CMS asynchronous collection is only allowed on a CMS thread."); 1771 1772 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1773 { 1774 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1775 MutexLockerEx hl(Heap_lock, safepoint_check); 1776 FreelistLocker fll(this); 1777 MutexLockerEx x(CGC_lock, safepoint_check); 1778 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1779 // The foreground collector is active or we're 1780 // not using asynchronous collections. Skip this 1781 // background collection. 1782 assert(!_foregroundGCShouldWait, "Should be clear"); 1783 return; 1784 } else { 1785 assert(_collectorState == Idling, "Should be idling before start."); 1786 _collectorState = InitialMarking; 1787 register_gc_start(cause); 1788 // Reset the expansion cause, now that we are about to begin 1789 // a new cycle. 1790 clear_expansion_cause(); 1791 1792 // Clear the MetaspaceGC flag since a concurrent collection 1793 // is starting but also clear it after the collection. 1794 MetaspaceGC::set_should_concurrent_collect(false); 1795 } 1796 // Decide if we want to enable class unloading as part of the 1797 // ensuing concurrent GC cycle. 1798 update_should_unload_classes(); 1799 _full_gc_requested = false; // acks all outstanding full gc requests 1800 _full_gc_cause = GCCause::_no_gc; 1801 // Signal that we are about to start a collection 1802 gch->increment_total_full_collections(); // ... starting a collection cycle 1803 _collection_count_start = gch->total_full_collections(); 1804 } 1805 1806 // Used for PrintGC 1807 size_t prev_used; 1808 if (PrintGC && Verbose) { 1809 prev_used = _cmsGen->used(); 1810 } 1811 1812 // The change of the collection state is normally done at this level; 1813 // the exceptions are phases that are executed while the world is 1814 // stopped. For those phases the change of state is done while the 1815 // world is stopped. For baton passing purposes this allows the 1816 // background collector to finish the phase and change state atomically. 1817 // The foreground collector cannot wait on a phase that is done 1818 // while the world is stopped because the foreground collector already 1819 // has the world stopped and would deadlock. 1820 while (_collectorState != Idling) { 1821 if (TraceCMSState) { 1822 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1823 p2i(Thread::current()), _collectorState); 1824 } 1825 // The foreground collector 1826 // holds the Heap_lock throughout its collection. 1827 // holds the CMS token (but not the lock) 1828 // except while it is waiting for the background collector to yield. 1829 // 1830 // The foreground collector should be blocked (not for long) 1831 // if the background collector is about to start a phase 1832 // executed with world stopped. If the background 1833 // collector has already started such a phase, the 1834 // foreground collector is blocked waiting for the 1835 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1836 // are executed in the VM thread. 1837 // 1838 // The locking order is 1839 // PendingListLock (PLL) -- if applicable (FinalMarking) 1840 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1841 // CMS token (claimed in 1842 // stop_world_and_do() --> 1843 // safepoint_synchronize() --> 1844 // CMSThread::synchronize()) 1845 1846 { 1847 // Check if the FG collector wants us to yield. 1848 CMSTokenSync x(true); // is cms thread 1849 if (waitForForegroundGC()) { 1850 // We yielded to a foreground GC, nothing more to be 1851 // done this round. 1852 assert(_foregroundGCShouldWait == false, "We set it to false in " 1853 "waitForForegroundGC()"); 1854 if (TraceCMSState) { 1855 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1856 " exiting collection CMS state %d", 1857 p2i(Thread::current()), _collectorState); 1858 } 1859 return; 1860 } else { 1861 // The background collector can run but check to see if the 1862 // foreground collector has done a collection while the 1863 // background collector was waiting to get the CGC_lock 1864 // above. If yes, break so that _foregroundGCShouldWait 1865 // is cleared before returning. 1866 if (_collectorState == Idling) { 1867 break; 1868 } 1869 } 1870 } 1871 1872 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1873 "should be waiting"); 1874 1875 switch (_collectorState) { 1876 case InitialMarking: 1877 { 1878 ReleaseForegroundGC x(this); 1879 stats().record_cms_begin(); 1880 VM_CMS_Initial_Mark initial_mark_op(this); 1881 VMThread::execute(&initial_mark_op); 1882 } 1883 // The collector state may be any legal state at this point 1884 // since the background collector may have yielded to the 1885 // foreground collector. 1886 break; 1887 case Marking: 1888 // initial marking in checkpointRootsInitialWork has been completed 1889 if (markFromRoots()) { // we were successful 1890 assert(_collectorState == Precleaning, "Collector state should " 1891 "have changed"); 1892 } else { 1893 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1894 } 1895 break; 1896 case Precleaning: 1897 // marking from roots in markFromRoots has been completed 1898 preclean(); 1899 assert(_collectorState == AbortablePreclean || 1900 _collectorState == FinalMarking, 1901 "Collector state should have changed"); 1902 break; 1903 case AbortablePreclean: 1904 abortable_preclean(); 1905 assert(_collectorState == FinalMarking, "Collector state should " 1906 "have changed"); 1907 break; 1908 case FinalMarking: 1909 { 1910 ReleaseForegroundGC x(this); 1911 1912 VM_CMS_Final_Remark final_remark_op(this); 1913 VMThread::execute(&final_remark_op); 1914 } 1915 assert(_foregroundGCShouldWait, "block post-condition"); 1916 break; 1917 case Sweeping: 1918 // final marking in checkpointRootsFinal has been completed 1919 sweep(); 1920 assert(_collectorState == Resizing, "Collector state change " 1921 "to Resizing must be done under the free_list_lock"); 1922 1923 case Resizing: { 1924 // Sweeping has been completed... 1925 // At this point the background collection has completed. 1926 // Don't move the call to compute_new_size() down 1927 // into code that might be executed if the background 1928 // collection was preempted. 1929 { 1930 ReleaseForegroundGC x(this); // unblock FG collection 1931 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1932 CMSTokenSync z(true); // not strictly needed. 1933 if (_collectorState == Resizing) { 1934 compute_new_size(); 1935 save_heap_summary(); 1936 _collectorState = Resetting; 1937 } else { 1938 assert(_collectorState == Idling, "The state should only change" 1939 " because the foreground collector has finished the collection"); 1940 } 1941 } 1942 break; 1943 } 1944 case Resetting: 1945 // CMS heap resizing has been completed 1946 reset(true); 1947 assert(_collectorState == Idling, "Collector state should " 1948 "have changed"); 1949 1950 MetaspaceGC::set_should_concurrent_collect(false); 1951 1952 stats().record_cms_end(); 1953 // Don't move the concurrent_phases_end() and compute_new_size() 1954 // calls to here because a preempted background collection 1955 // has it's state set to "Resetting". 1956 break; 1957 case Idling: 1958 default: 1959 ShouldNotReachHere(); 1960 break; 1961 } 1962 if (TraceCMSState) { 1963 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1964 p2i(Thread::current()), _collectorState); 1965 } 1966 assert(_foregroundGCShouldWait, "block post-condition"); 1967 } 1968 1969 // Should this be in gc_epilogue? 1970 collector_policy()->counters()->update_counters(); 1971 1972 { 1973 // Clear _foregroundGCShouldWait and, in the event that the 1974 // foreground collector is waiting, notify it, before 1975 // returning. 1976 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1977 _foregroundGCShouldWait = false; 1978 if (_foregroundGCIsActive) { 1979 CGC_lock->notify(); 1980 } 1981 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1982 "Possible deadlock"); 1983 } 1984 if (TraceCMSState) { 1985 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1986 " exiting collection CMS state %d", 1987 p2i(Thread::current()), _collectorState); 1988 } 1989 if (PrintGC && Verbose) { 1990 _cmsGen->print_heap_change(prev_used); 1991 } 1992 } 1993 1994 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1995 _cms_start_registered = true; 1996 _gc_timer_cm->register_gc_start(); 1997 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1998 } 1999 2000 void CMSCollector::register_gc_end() { 2001 if (_cms_start_registered) { 2002 report_heap_summary(GCWhen::AfterGC); 2003 2004 _gc_timer_cm->register_gc_end(); 2005 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2006 _cms_start_registered = false; 2007 } 2008 } 2009 2010 void CMSCollector::save_heap_summary() { 2011 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2012 _last_heap_summary = gch->create_heap_summary(); 2013 _last_metaspace_summary = gch->create_metaspace_summary(); 2014 } 2015 2016 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2017 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2018 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2019 } 2020 2021 bool CMSCollector::waitForForegroundGC() { 2022 bool res = false; 2023 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2024 "CMS thread should have CMS token"); 2025 // Block the foreground collector until the 2026 // background collectors decides whether to 2027 // yield. 2028 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2029 _foregroundGCShouldWait = true; 2030 if (_foregroundGCIsActive) { 2031 // The background collector yields to the 2032 // foreground collector and returns a value 2033 // indicating that it has yielded. The foreground 2034 // collector can proceed. 2035 res = true; 2036 _foregroundGCShouldWait = false; 2037 ConcurrentMarkSweepThread::clear_CMS_flag( 2038 ConcurrentMarkSweepThread::CMS_cms_has_token); 2039 ConcurrentMarkSweepThread::set_CMS_flag( 2040 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2041 // Get a possibly blocked foreground thread going 2042 CGC_lock->notify(); 2043 if (TraceCMSState) { 2044 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2045 p2i(Thread::current()), _collectorState); 2046 } 2047 while (_foregroundGCIsActive) { 2048 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2049 } 2050 ConcurrentMarkSweepThread::set_CMS_flag( 2051 ConcurrentMarkSweepThread::CMS_cms_has_token); 2052 ConcurrentMarkSweepThread::clear_CMS_flag( 2053 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2054 } 2055 if (TraceCMSState) { 2056 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2057 p2i(Thread::current()), _collectorState); 2058 } 2059 return res; 2060 } 2061 2062 // Because of the need to lock the free lists and other structures in 2063 // the collector, common to all the generations that the collector is 2064 // collecting, we need the gc_prologues of individual CMS generations 2065 // delegate to their collector. It may have been simpler had the 2066 // current infrastructure allowed one to call a prologue on a 2067 // collector. In the absence of that we have the generation's 2068 // prologue delegate to the collector, which delegates back 2069 // some "local" work to a worker method in the individual generations 2070 // that it's responsible for collecting, while itself doing any 2071 // work common to all generations it's responsible for. A similar 2072 // comment applies to the gc_epilogue()'s. 2073 // The role of the variable _between_prologue_and_epilogue is to 2074 // enforce the invocation protocol. 2075 void CMSCollector::gc_prologue(bool full) { 2076 // Call gc_prologue_work() for the CMSGen 2077 // we are responsible for. 2078 2079 // The following locking discipline assumes that we are only called 2080 // when the world is stopped. 2081 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2082 2083 // The CMSCollector prologue must call the gc_prologues for the 2084 // "generations" that it's responsible 2085 // for. 2086 2087 assert( Thread::current()->is_VM_thread() 2088 || ( CMSScavengeBeforeRemark 2089 && Thread::current()->is_ConcurrentGC_thread()), 2090 "Incorrect thread type for prologue execution"); 2091 2092 if (_between_prologue_and_epilogue) { 2093 // We have already been invoked; this is a gc_prologue delegation 2094 // from yet another CMS generation that we are responsible for, just 2095 // ignore it since all relevant work has already been done. 2096 return; 2097 } 2098 2099 // set a bit saying prologue has been called; cleared in epilogue 2100 _between_prologue_and_epilogue = true; 2101 // Claim locks for common data structures, then call gc_prologue_work() 2102 // for each CMSGen. 2103 2104 getFreelistLocks(); // gets free list locks on constituent spaces 2105 bitMapLock()->lock_without_safepoint_check(); 2106 2107 // Should call gc_prologue_work() for all cms gens we are responsible for 2108 bool duringMarking = _collectorState >= Marking 2109 && _collectorState < Sweeping; 2110 2111 // The young collections clear the modified oops state, which tells if 2112 // there are any modified oops in the class. The remark phase also needs 2113 // that information. Tell the young collection to save the union of all 2114 // modified klasses. 2115 if (duringMarking) { 2116 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2117 } 2118 2119 bool registerClosure = duringMarking; 2120 2121 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2122 2123 if (!full) { 2124 stats().record_gc0_begin(); 2125 } 2126 } 2127 2128 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2129 2130 _capacity_at_prologue = capacity(); 2131 _used_at_prologue = used(); 2132 2133 // Delegate to CMScollector which knows how to coordinate between 2134 // this and any other CMS generations that it is responsible for 2135 // collecting. 2136 collector()->gc_prologue(full); 2137 } 2138 2139 // This is a "private" interface for use by this generation's CMSCollector. 2140 // Not to be called directly by any other entity (for instance, 2141 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2142 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2143 bool registerClosure, ModUnionClosure* modUnionClosure) { 2144 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2145 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2146 "Should be NULL"); 2147 if (registerClosure) { 2148 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2149 } 2150 cmsSpace()->gc_prologue(); 2151 // Clear stat counters 2152 NOT_PRODUCT( 2153 assert(_numObjectsPromoted == 0, "check"); 2154 assert(_numWordsPromoted == 0, "check"); 2155 if (Verbose && PrintGC) { 2156 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2157 SIZE_FORMAT" bytes concurrently", 2158 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2159 } 2160 _numObjectsAllocated = 0; 2161 _numWordsAllocated = 0; 2162 ) 2163 } 2164 2165 void CMSCollector::gc_epilogue(bool full) { 2166 // The following locking discipline assumes that we are only called 2167 // when the world is stopped. 2168 assert(SafepointSynchronize::is_at_safepoint(), 2169 "world is stopped assumption"); 2170 2171 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2172 // if linear allocation blocks need to be appropriately marked to allow the 2173 // the blocks to be parsable. We also check here whether we need to nudge the 2174 // CMS collector thread to start a new cycle (if it's not already active). 2175 assert( Thread::current()->is_VM_thread() 2176 || ( CMSScavengeBeforeRemark 2177 && Thread::current()->is_ConcurrentGC_thread()), 2178 "Incorrect thread type for epilogue execution"); 2179 2180 if (!_between_prologue_and_epilogue) { 2181 // We have already been invoked; this is a gc_epilogue delegation 2182 // from yet another CMS generation that we are responsible for, just 2183 // ignore it since all relevant work has already been done. 2184 return; 2185 } 2186 assert(haveFreelistLocks(), "must have freelist locks"); 2187 assert_lock_strong(bitMapLock()); 2188 2189 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2190 2191 _cmsGen->gc_epilogue_work(full); 2192 2193 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2194 // in case sampling was not already enabled, enable it 2195 _start_sampling = true; 2196 } 2197 // reset _eden_chunk_array so sampling starts afresh 2198 _eden_chunk_index = 0; 2199 2200 size_t cms_used = _cmsGen->cmsSpace()->used(); 2201 2202 // update performance counters - this uses a special version of 2203 // update_counters() that allows the utilization to be passed as a 2204 // parameter, avoiding multiple calls to used(). 2205 // 2206 _cmsGen->update_counters(cms_used); 2207 2208 bitMapLock()->unlock(); 2209 releaseFreelistLocks(); 2210 2211 if (!CleanChunkPoolAsync) { 2212 Chunk::clean_chunk_pool(); 2213 } 2214 2215 set_did_compact(false); 2216 _between_prologue_and_epilogue = false; // ready for next cycle 2217 } 2218 2219 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2220 collector()->gc_epilogue(full); 2221 2222 // Also reset promotion tracking in par gc thread states. 2223 for (uint i = 0; i < ParallelGCThreads; i++) { 2224 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2225 } 2226 } 2227 2228 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2229 assert(!incremental_collection_failed(), "Should have been cleared"); 2230 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2231 cmsSpace()->gc_epilogue(); 2232 // Print stat counters 2233 NOT_PRODUCT( 2234 assert(_numObjectsAllocated == 0, "check"); 2235 assert(_numWordsAllocated == 0, "check"); 2236 if (Verbose && PrintGC) { 2237 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2238 SIZE_FORMAT" bytes", 2239 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2240 } 2241 _numObjectsPromoted = 0; 2242 _numWordsPromoted = 0; 2243 ) 2244 2245 if (PrintGC && Verbose) { 2246 // Call down the chain in contiguous_available needs the freelistLock 2247 // so print this out before releasing the freeListLock. 2248 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2249 contiguous_available()); 2250 } 2251 } 2252 2253 #ifndef PRODUCT 2254 bool CMSCollector::have_cms_token() { 2255 Thread* thr = Thread::current(); 2256 if (thr->is_VM_thread()) { 2257 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2258 } else if (thr->is_ConcurrentGC_thread()) { 2259 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2260 } else if (thr->is_GC_task_thread()) { 2261 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2262 ParGCRareEvent_lock->owned_by_self(); 2263 } 2264 return false; 2265 } 2266 #endif 2267 2268 // Check reachability of the given heap address in CMS generation, 2269 // treating all other generations as roots. 2270 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2271 // We could "guarantee" below, rather than assert, but I'll 2272 // leave these as "asserts" so that an adventurous debugger 2273 // could try this in the product build provided some subset of 2274 // the conditions were met, provided they were interested in the 2275 // results and knew that the computation below wouldn't interfere 2276 // with other concurrent computations mutating the structures 2277 // being read or written. 2278 assert(SafepointSynchronize::is_at_safepoint(), 2279 "Else mutations in object graph will make answer suspect"); 2280 assert(have_cms_token(), "Should hold cms token"); 2281 assert(haveFreelistLocks(), "must hold free list locks"); 2282 assert_lock_strong(bitMapLock()); 2283 2284 // Clear the marking bit map array before starting, but, just 2285 // for kicks, first report if the given address is already marked 2286 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2287 _markBitMap.isMarked(addr) ? "" : " not"); 2288 2289 if (verify_after_remark()) { 2290 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2291 bool result = verification_mark_bm()->isMarked(addr); 2292 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2293 result ? "IS" : "is NOT"); 2294 return result; 2295 } else { 2296 gclog_or_tty->print_cr("Could not compute result"); 2297 return false; 2298 } 2299 } 2300 2301 2302 void 2303 CMSCollector::print_on_error(outputStream* st) { 2304 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2305 if (collector != NULL) { 2306 CMSBitMap* bitmap = &collector->_markBitMap; 2307 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2308 bitmap->print_on_error(st, " Bits: "); 2309 2310 st->cr(); 2311 2312 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2313 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2314 mut_bitmap->print_on_error(st, " Bits: "); 2315 } 2316 } 2317 2318 //////////////////////////////////////////////////////// 2319 // CMS Verification Support 2320 //////////////////////////////////////////////////////// 2321 // Following the remark phase, the following invariant 2322 // should hold -- each object in the CMS heap which is 2323 // marked in markBitMap() should be marked in the verification_mark_bm(). 2324 2325 class VerifyMarkedClosure: public BitMapClosure { 2326 CMSBitMap* _marks; 2327 bool _failed; 2328 2329 public: 2330 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2331 2332 bool do_bit(size_t offset) { 2333 HeapWord* addr = _marks->offsetToHeapWord(offset); 2334 if (!_marks->isMarked(addr)) { 2335 oop(addr)->print_on(gclog_or_tty); 2336 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2337 _failed = true; 2338 } 2339 return true; 2340 } 2341 2342 bool failed() { return _failed; } 2343 }; 2344 2345 bool CMSCollector::verify_after_remark(bool silent) { 2346 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2347 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2348 static bool init = false; 2349 2350 assert(SafepointSynchronize::is_at_safepoint(), 2351 "Else mutations in object graph will make answer suspect"); 2352 assert(have_cms_token(), 2353 "Else there may be mutual interference in use of " 2354 " verification data structures"); 2355 assert(_collectorState > Marking && _collectorState <= Sweeping, 2356 "Else marking info checked here may be obsolete"); 2357 assert(haveFreelistLocks(), "must hold free list locks"); 2358 assert_lock_strong(bitMapLock()); 2359 2360 2361 // Allocate marking bit map if not already allocated 2362 if (!init) { // first time 2363 if (!verification_mark_bm()->allocate(_span)) { 2364 return false; 2365 } 2366 init = true; 2367 } 2368 2369 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2370 2371 // Turn off refs discovery -- so we will be tracing through refs. 2372 // This is as intended, because by this time 2373 // GC must already have cleared any refs that need to be cleared, 2374 // and traced those that need to be marked; moreover, 2375 // the marking done here is not going to interfere in any 2376 // way with the marking information used by GC. 2377 NoRefDiscovery no_discovery(ref_processor()); 2378 2379 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2380 2381 // Clear any marks from a previous round 2382 verification_mark_bm()->clear_all(); 2383 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2384 verify_work_stacks_empty(); 2385 2386 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2387 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2388 // Update the saved marks which may affect the root scans. 2389 gch->save_marks(); 2390 2391 if (CMSRemarkVerifyVariant == 1) { 2392 // In this first variant of verification, we complete 2393 // all marking, then check if the new marks-vector is 2394 // a subset of the CMS marks-vector. 2395 verify_after_remark_work_1(); 2396 } else if (CMSRemarkVerifyVariant == 2) { 2397 // In this second variant of verification, we flag an error 2398 // (i.e. an object reachable in the new marks-vector not reachable 2399 // in the CMS marks-vector) immediately, also indicating the 2400 // identify of an object (A) that references the unmarked object (B) -- 2401 // presumably, a mutation to A failed to be picked up by preclean/remark? 2402 verify_after_remark_work_2(); 2403 } else { 2404 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2405 CMSRemarkVerifyVariant); 2406 } 2407 if (!silent) gclog_or_tty->print(" done] "); 2408 return true; 2409 } 2410 2411 void CMSCollector::verify_after_remark_work_1() { 2412 ResourceMark rm; 2413 HandleMark hm; 2414 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2415 2416 // Get a clear set of claim bits for the roots processing to work with. 2417 ClassLoaderDataGraph::clear_claimed_marks(); 2418 2419 // Mark from roots one level into CMS 2420 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2421 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2422 2423 gch->gen_process_roots(Generation::Old, 2424 true, // younger gens are roots 2425 true, // activate StrongRootsScope 2426 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2427 should_unload_classes(), 2428 ¬Older, 2429 NULL, 2430 NULL); // SSS: Provide correct closure 2431 2432 // Now mark from the roots 2433 MarkFromRootsClosure markFromRootsClosure(this, _span, 2434 verification_mark_bm(), verification_mark_stack(), 2435 false /* don't yield */, true /* verifying */); 2436 assert(_restart_addr == NULL, "Expected pre-condition"); 2437 verification_mark_bm()->iterate(&markFromRootsClosure); 2438 while (_restart_addr != NULL) { 2439 // Deal with stack overflow: by restarting at the indicated 2440 // address. 2441 HeapWord* ra = _restart_addr; 2442 markFromRootsClosure.reset(ra); 2443 _restart_addr = NULL; 2444 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2445 } 2446 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2447 verify_work_stacks_empty(); 2448 2449 // Marking completed -- now verify that each bit marked in 2450 // verification_mark_bm() is also marked in markBitMap(); flag all 2451 // errors by printing corresponding objects. 2452 VerifyMarkedClosure vcl(markBitMap()); 2453 verification_mark_bm()->iterate(&vcl); 2454 if (vcl.failed()) { 2455 gclog_or_tty->print("Verification failed"); 2456 gch->print_on(gclog_or_tty); 2457 fatal("CMS: failed marking verification after remark"); 2458 } 2459 } 2460 2461 class VerifyKlassOopsKlassClosure : public KlassClosure { 2462 class VerifyKlassOopsClosure : public OopClosure { 2463 CMSBitMap* _bitmap; 2464 public: 2465 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2466 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2467 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2468 } _oop_closure; 2469 public: 2470 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2471 void do_klass(Klass* k) { 2472 k->oops_do(&_oop_closure); 2473 } 2474 }; 2475 2476 void CMSCollector::verify_after_remark_work_2() { 2477 ResourceMark rm; 2478 HandleMark hm; 2479 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2480 2481 // Get a clear set of claim bits for the roots processing to work with. 2482 ClassLoaderDataGraph::clear_claimed_marks(); 2483 2484 // Mark from roots one level into CMS 2485 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2486 markBitMap()); 2487 CLDToOopClosure cld_closure(¬Older, true); 2488 2489 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2490 2491 gch->gen_process_roots(Generation::Old, 2492 true, // younger gens are roots 2493 true, // activate StrongRootsScope 2494 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2495 should_unload_classes(), 2496 ¬Older, 2497 NULL, 2498 &cld_closure); 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 ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2691 // Include symbols, strings and code cache elements to prevent their resurrection. 2692 add_root_scanning_option(rso); 2693 set_verifying(true); 2694 } else if (verifying() && !should_verify) { 2695 // We were verifying, but some verification flags got disabled. 2696 set_verifying(false); 2697 // Exclude symbols, strings and code cache elements from root scanning to 2698 // reduce IM and RM pauses. 2699 remove_root_scanning_option(rso); 2700 } 2701 } 2702 2703 2704 #ifndef PRODUCT 2705 HeapWord* CMSCollector::block_start(const void* p) const { 2706 const HeapWord* addr = (HeapWord*)p; 2707 if (_span.contains(p)) { 2708 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2709 return _cmsGen->cmsSpace()->block_start(p); 2710 } 2711 } 2712 return NULL; 2713 } 2714 #endif 2715 2716 HeapWord* 2717 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2718 bool tlab, 2719 bool parallel) { 2720 CMSSynchronousYieldRequest yr; 2721 assert(!tlab, "Can't deal with TLAB allocation"); 2722 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2723 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2724 if (GCExpandToAllocateDelayMillis > 0) { 2725 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2726 } 2727 return have_lock_and_allocate(word_size, tlab); 2728 } 2729 2730 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2731 size_t bytes, 2732 size_t expand_bytes, 2733 CMSExpansionCause::Cause cause) 2734 { 2735 2736 bool success = expand(bytes, expand_bytes); 2737 2738 // remember why we expanded; this information is used 2739 // by shouldConcurrentCollect() when making decisions on whether to start 2740 // a new CMS cycle. 2741 if (success) { 2742 set_expansion_cause(cause); 2743 if (PrintGCDetails && Verbose) { 2744 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2745 CMSExpansionCause::to_string(cause)); 2746 } 2747 } 2748 } 2749 2750 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2751 HeapWord* res = NULL; 2752 MutexLocker x(ParGCRareEvent_lock); 2753 while (true) { 2754 // Expansion by some other thread might make alloc OK now: 2755 res = ps->lab.alloc(word_sz); 2756 if (res != NULL) return res; 2757 // If there's not enough expansion space available, give up. 2758 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2759 return NULL; 2760 } 2761 // Otherwise, we try expansion. 2762 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2763 // Now go around the loop and try alloc again; 2764 // A competing par_promote might beat us to the expansion space, 2765 // so we may go around the loop again if promotion fails again. 2766 if (GCExpandToAllocateDelayMillis > 0) { 2767 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2768 } 2769 } 2770 } 2771 2772 2773 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2774 PromotionInfo* promo) { 2775 MutexLocker x(ParGCRareEvent_lock); 2776 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2777 while (true) { 2778 // Expansion by some other thread might make alloc OK now: 2779 if (promo->ensure_spooling_space()) { 2780 assert(promo->has_spooling_space(), 2781 "Post-condition of successful ensure_spooling_space()"); 2782 return true; 2783 } 2784 // If there's not enough expansion space available, give up. 2785 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2786 return false; 2787 } 2788 // Otherwise, we try expansion. 2789 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2790 // Now go around the loop and try alloc again; 2791 // A competing allocation might beat us to the expansion space, 2792 // so we may go around the loop again if allocation fails again. 2793 if (GCExpandToAllocateDelayMillis > 0) { 2794 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2795 } 2796 } 2797 } 2798 2799 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2800 // Only shrink if a compaction was done so that all the free space 2801 // in the generation is in a contiguous block at the end. 2802 if (did_compact()) { 2803 CardGeneration::shrink(bytes); 2804 } 2805 } 2806 2807 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2808 assert_locked_or_safepoint(Heap_lock); 2809 } 2810 2811 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2812 assert_locked_or_safepoint(Heap_lock); 2813 assert_lock_strong(freelistLock()); 2814 if (PrintGCDetails && Verbose) { 2815 warning("Shrinking of CMS not yet implemented"); 2816 } 2817 return; 2818 } 2819 2820 2821 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2822 // phases. 2823 class CMSPhaseAccounting: public StackObj { 2824 public: 2825 CMSPhaseAccounting(CMSCollector *collector, 2826 const char *phase, 2827 const GCId gc_id, 2828 bool print_cr = true); 2829 ~CMSPhaseAccounting(); 2830 2831 private: 2832 CMSCollector *_collector; 2833 const char *_phase; 2834 elapsedTimer _wallclock; 2835 bool _print_cr; 2836 const GCId _gc_id; 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 const GCId gc_id, 2853 bool print_cr) : 2854 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2855 2856 if (PrintCMSStatistics != 0) { 2857 _collector->resetYields(); 2858 } 2859 if (PrintGCDetails) { 2860 gclog_or_tty->gclog_stamp(_gc_id); 2861 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2862 _collector->cmsGen()->short_name(), _phase); 2863 } 2864 _collector->resetTimer(); 2865 _wallclock.start(); 2866 _collector->startTimer(); 2867 } 2868 2869 CMSPhaseAccounting::~CMSPhaseAccounting() { 2870 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2871 _collector->stopTimer(); 2872 _wallclock.stop(); 2873 if (PrintGCDetails) { 2874 gclog_or_tty->gclog_stamp(_gc_id); 2875 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2876 _collector->cmsGen()->short_name(), 2877 _phase, _collector->timerValue(), _wallclock.seconds()); 2878 if (_print_cr) { 2879 gclog_or_tty->cr(); 2880 } 2881 if (PrintCMSStatistics != 0) { 2882 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2883 _collector->yields()); 2884 } 2885 } 2886 } 2887 2888 // CMS work 2889 2890 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2891 class CMSParMarkTask : public AbstractGangTask { 2892 protected: 2893 CMSCollector* _collector; 2894 uint _n_workers; 2895 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2896 AbstractGangTask(name), 2897 _collector(collector), 2898 _n_workers(n_workers) {} 2899 // Work method in support of parallel rescan ... of young gen spaces 2900 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2901 ContiguousSpace* space, 2902 HeapWord** chunk_array, size_t chunk_top); 2903 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2904 }; 2905 2906 // Parallel initial mark task 2907 class CMSParInitialMarkTask: public CMSParMarkTask { 2908 public: 2909 CMSParInitialMarkTask(CMSCollector* collector, uint n_workers) : 2910 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2911 collector, n_workers) {} 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 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2944 // precede our marking with a collection of all 2945 // younger generations to keep floating garbage to a minimum. 2946 // XXX: we won't do this for now -- it's an optimization to be done later. 2947 2948 // already have locks 2949 assert_lock_strong(bitMapLock()); 2950 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2951 2952 // Setup the verification and class unloading state for this 2953 // CMS collection cycle. 2954 setup_cms_unloading_and_verification_state(); 2955 2956 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2957 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2958 2959 // Reset all the PLAB chunk arrays if necessary. 2960 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2961 reset_survivor_plab_arrays(); 2962 } 2963 2964 ResourceMark rm; 2965 HandleMark hm; 2966 2967 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2968 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2969 2970 verify_work_stacks_empty(); 2971 verify_overflow_empty(); 2972 2973 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2974 // Update the saved marks which may affect the root scans. 2975 gch->save_marks(); 2976 2977 // weak reference processing has not started yet. 2978 ref_processor()->set_enqueuing_is_done(false); 2979 2980 // Need to remember all newly created CLDs, 2981 // so that we can guarantee that the remark finds them. 2982 ClassLoaderDataGraph::remember_new_clds(true); 2983 2984 // Whenever a CLD is found, it will be claimed before proceeding to mark 2985 // the klasses. The claimed marks need to be cleared before marking starts. 2986 ClassLoaderDataGraph::clear_claimed_marks(); 2987 2988 if (CMSPrintEdenSurvivorChunks) { 2989 print_eden_and_survivor_chunk_arrays(); 2990 } 2991 2992 { 2993 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2994 if (CMSParallelInitialMarkEnabled) { 2995 // The parallel version. 2996 FlexibleWorkGang* workers = gch->workers(); 2997 assert(workers != NULL, "Need parallel worker threads."); 2998 uint n_workers = workers->active_workers(); 2999 CMSParInitialMarkTask tsk(this, n_workers); 3000 gch->set_par_threads(n_workers); 3001 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3002 if (n_workers > 1) { 3003 StrongRootsScope srs; 3004 workers->run_task(&tsk); 3005 } else { 3006 StrongRootsScope srs; 3007 tsk.work(0); 3008 } 3009 gch->set_par_threads(0); 3010 } else { 3011 // The serial version. 3012 CLDToOopClosure cld_closure(¬Older, true); 3013 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3014 gch->gen_process_roots(Generation::Old, 3015 true, // younger gens are roots 3016 true, // activate StrongRootsScope 3017 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3018 should_unload_classes(), 3019 ¬Older, 3020 NULL, 3021 &cld_closure); 3022 } 3023 } 3024 3025 // Clear mod-union table; it will be dirtied in the prologue of 3026 // CMS generation per each younger generation collection. 3027 3028 assert(_modUnionTable.isAllClear(), 3029 "Was cleared in most recent final checkpoint phase" 3030 " or no bits are set in the gc_prologue before the start of the next " 3031 "subsequent marking phase."); 3032 3033 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3034 3035 // Save the end of the used_region of the constituent generations 3036 // to be used to limit the extent of sweep in each generation. 3037 save_sweep_limits(); 3038 verify_overflow_empty(); 3039 } 3040 3041 bool CMSCollector::markFromRoots() { 3042 // we might be tempted to assert that: 3043 // assert(!SafepointSynchronize::is_at_safepoint(), 3044 // "inconsistent argument?"); 3045 // However that wouldn't be right, because it's possible that 3046 // a safepoint is indeed in progress as a younger generation 3047 // stop-the-world GC happens even as we mark in this generation. 3048 assert(_collectorState == Marking, "inconsistent state?"); 3049 check_correct_thread_executing(); 3050 verify_overflow_empty(); 3051 3052 // Weak ref discovery note: We may be discovering weak 3053 // refs in this generation concurrent (but interleaved) with 3054 // weak ref discovery by a younger generation collector. 3055 3056 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3057 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3058 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3059 bool res = markFromRootsWork(); 3060 if (res) { 3061 _collectorState = Precleaning; 3062 } else { // We failed and a foreground collection wants to take over 3063 assert(_foregroundGCIsActive, "internal state inconsistency"); 3064 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3065 if (PrintGCDetails) { 3066 gclog_or_tty->print_cr("bailing out to foreground collection"); 3067 } 3068 } 3069 verify_overflow_empty(); 3070 return res; 3071 } 3072 3073 bool CMSCollector::markFromRootsWork() { 3074 // iterate over marked bits in bit map, doing a full scan and mark 3075 // from these roots using the following algorithm: 3076 // . if oop is to the right of the current scan pointer, 3077 // mark corresponding bit (we'll process it later) 3078 // . else (oop is to left of current scan pointer) 3079 // push oop on marking stack 3080 // . drain the marking stack 3081 3082 // Note that when we do a marking step we need to hold the 3083 // bit map lock -- recall that direct allocation (by mutators) 3084 // and promotion (by younger generation collectors) is also 3085 // marking the bit map. [the so-called allocate live policy.] 3086 // Because the implementation of bit map marking is not 3087 // robust wrt simultaneous marking of bits in the same word, 3088 // we need to make sure that there is no such interference 3089 // between concurrent such updates. 3090 3091 // already have locks 3092 assert_lock_strong(bitMapLock()); 3093 3094 verify_work_stacks_empty(); 3095 verify_overflow_empty(); 3096 bool result = false; 3097 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3098 result = do_marking_mt(); 3099 } else { 3100 result = do_marking_st(); 3101 } 3102 return result; 3103 } 3104 3105 // Forward decl 3106 class CMSConcMarkingTask; 3107 3108 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3109 CMSCollector* _collector; 3110 CMSConcMarkingTask* _task; 3111 public: 3112 virtual void yield(); 3113 3114 // "n_threads" is the number of threads to be terminated. 3115 // "queue_set" is a set of work queues of other threads. 3116 // "collector" is the CMS collector associated with this task terminator. 3117 // "yield" indicates whether we need the gang as a whole to yield. 3118 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3119 ParallelTaskTerminator(n_threads, queue_set), 3120 _collector(collector) { } 3121 3122 void set_task(CMSConcMarkingTask* task) { 3123 _task = task; 3124 } 3125 }; 3126 3127 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3128 CMSConcMarkingTask* _task; 3129 public: 3130 bool should_exit_termination(); 3131 void set_task(CMSConcMarkingTask* task) { 3132 _task = task; 3133 } 3134 }; 3135 3136 // MT Concurrent Marking Task 3137 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3138 CMSCollector* _collector; 3139 uint _n_workers; // requested/desired # workers 3140 bool _result; 3141 CompactibleFreeListSpace* _cms_space; 3142 char _pad_front[64]; // padding to ... 3143 HeapWord* _global_finger; // ... avoid sharing cache line 3144 char _pad_back[64]; 3145 HeapWord* _restart_addr; 3146 3147 // Exposed here for yielding support 3148 Mutex* const _bit_map_lock; 3149 3150 // The per thread work queues, available here for stealing 3151 OopTaskQueueSet* _task_queues; 3152 3153 // Termination (and yielding) support 3154 CMSConcMarkingTerminator _term; 3155 CMSConcMarkingTerminatorTerminator _term_term; 3156 3157 public: 3158 CMSConcMarkingTask(CMSCollector* collector, 3159 CompactibleFreeListSpace* cms_space, 3160 YieldingFlexibleWorkGang* workers, 3161 OopTaskQueueSet* task_queues): 3162 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3163 _collector(collector), 3164 _cms_space(cms_space), 3165 _n_workers(0), _result(true), 3166 _task_queues(task_queues), 3167 _term(_n_workers, task_queues, _collector), 3168 _bit_map_lock(collector->bitMapLock()) 3169 { 3170 _requested_size = _n_workers; 3171 _term.set_task(this); 3172 _term_term.set_task(this); 3173 _restart_addr = _global_finger = _cms_space->bottom(); 3174 } 3175 3176 3177 OopTaskQueueSet* task_queues() { return _task_queues; } 3178 3179 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3180 3181 HeapWord** global_finger_addr() { return &_global_finger; } 3182 3183 CMSConcMarkingTerminator* terminator() { return &_term; } 3184 3185 virtual void set_for_termination(uint active_workers) { 3186 terminator()->reset_for_reuse(active_workers); 3187 } 3188 3189 void work(uint worker_id); 3190 bool should_yield() { 3191 return ConcurrentMarkSweepThread::should_yield() 3192 && !_collector->foregroundGCIsActive(); 3193 } 3194 3195 virtual void coordinator_yield(); // stuff done by coordinator 3196 bool result() { return _result; } 3197 3198 void reset(HeapWord* ra) { 3199 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3200 _restart_addr = _global_finger = ra; 3201 _term.reset_for_reuse(); 3202 } 3203 3204 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3205 OopTaskQueue* work_q); 3206 3207 private: 3208 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3209 void do_work_steal(int i); 3210 void bump_global_finger(HeapWord* f); 3211 }; 3212 3213 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3214 assert(_task != NULL, "Error"); 3215 return _task->yielding(); 3216 // Note that we do not need the disjunct || _task->should_yield() above 3217 // because we want terminating threads to yield only if the task 3218 // is already in the midst of yielding, which happens only after at least one 3219 // thread has yielded. 3220 } 3221 3222 void CMSConcMarkingTerminator::yield() { 3223 if (_task->should_yield()) { 3224 _task->yield(); 3225 } else { 3226 ParallelTaskTerminator::yield(); 3227 } 3228 } 3229 3230 //////////////////////////////////////////////////////////////// 3231 // Concurrent Marking Algorithm Sketch 3232 //////////////////////////////////////////////////////////////// 3233 // Until all tasks exhausted (both spaces): 3234 // -- claim next available chunk 3235 // -- bump global finger via CAS 3236 // -- find first object that starts in this chunk 3237 // and start scanning bitmap from that position 3238 // -- scan marked objects for oops 3239 // -- CAS-mark target, and if successful: 3240 // . if target oop is above global finger (volatile read) 3241 // nothing to do 3242 // . if target oop is in chunk and above local finger 3243 // then nothing to do 3244 // . else push on work-queue 3245 // -- Deal with possible overflow issues: 3246 // . local work-queue overflow causes stuff to be pushed on 3247 // global (common) overflow queue 3248 // . always first empty local work queue 3249 // . then get a batch of oops from global work queue if any 3250 // . then do work stealing 3251 // -- When all tasks claimed (both spaces) 3252 // and local work queue empty, 3253 // then in a loop do: 3254 // . check global overflow stack; steal a batch of oops and trace 3255 // . try to steal from other threads oif GOS is empty 3256 // . if neither is available, offer termination 3257 // -- Terminate and return result 3258 // 3259 void CMSConcMarkingTask::work(uint worker_id) { 3260 elapsedTimer _timer; 3261 ResourceMark rm; 3262 HandleMark hm; 3263 3264 DEBUG_ONLY(_collector->verify_overflow_empty();) 3265 3266 // Before we begin work, our work queue should be empty 3267 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3268 // Scan the bitmap covering _cms_space, tracing through grey objects. 3269 _timer.start(); 3270 do_scan_and_mark(worker_id, _cms_space); 3271 _timer.stop(); 3272 if (PrintCMSStatistics != 0) { 3273 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3274 worker_id, _timer.seconds()); 3275 // XXX: need xxx/xxx type of notation, two timers 3276 } 3277 3278 // ... do work stealing 3279 _timer.reset(); 3280 _timer.start(); 3281 do_work_steal(worker_id); 3282 _timer.stop(); 3283 if (PrintCMSStatistics != 0) { 3284 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3285 worker_id, _timer.seconds()); 3286 // XXX: need xxx/xxx type of notation, two timers 3287 } 3288 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3289 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3290 // Note that under the current task protocol, the 3291 // following assertion is true even of the spaces 3292 // expanded since the completion of the concurrent 3293 // marking. XXX This will likely change under a strict 3294 // ABORT semantics. 3295 // After perm removal the comparison was changed to 3296 // greater than or equal to from strictly greater than. 3297 // Before perm removal the highest address sweep would 3298 // have been at the end of perm gen but now is at the 3299 // end of the tenured gen. 3300 assert(_global_finger >= _cms_space->end(), 3301 "All tasks have been completed"); 3302 DEBUG_ONLY(_collector->verify_overflow_empty();) 3303 } 3304 3305 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3306 HeapWord* read = _global_finger; 3307 HeapWord* cur = read; 3308 while (f > read) { 3309 cur = read; 3310 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3311 if (cur == read) { 3312 // our cas succeeded 3313 assert(_global_finger >= f, "protocol consistency"); 3314 break; 3315 } 3316 } 3317 } 3318 3319 // This is really inefficient, and should be redone by 3320 // using (not yet available) block-read and -write interfaces to the 3321 // stack and the work_queue. XXX FIX ME !!! 3322 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3323 OopTaskQueue* work_q) { 3324 // Fast lock-free check 3325 if (ovflw_stk->length() == 0) { 3326 return false; 3327 } 3328 assert(work_q->size() == 0, "Shouldn't steal"); 3329 MutexLockerEx ml(ovflw_stk->par_lock(), 3330 Mutex::_no_safepoint_check_flag); 3331 // Grab up to 1/4 the size of the work queue 3332 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3333 (size_t)ParGCDesiredObjsFromOverflowList); 3334 num = MIN2(num, ovflw_stk->length()); 3335 for (int i = (int) num; i > 0; i--) { 3336 oop cur = ovflw_stk->pop(); 3337 assert(cur != NULL, "Counted wrong?"); 3338 work_q->push(cur); 3339 } 3340 return num > 0; 3341 } 3342 3343 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3344 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3345 int n_tasks = pst->n_tasks(); 3346 // We allow that there may be no tasks to do here because 3347 // we are restarting after a stack overflow. 3348 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3349 uint nth_task = 0; 3350 3351 HeapWord* aligned_start = sp->bottom(); 3352 if (sp->used_region().contains(_restart_addr)) { 3353 // Align down to a card boundary for the start of 0th task 3354 // for this space. 3355 aligned_start = 3356 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3357 CardTableModRefBS::card_size); 3358 } 3359 3360 size_t chunk_size = sp->marking_task_size(); 3361 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3362 // Having claimed the nth task in this space, 3363 // compute the chunk that it corresponds to: 3364 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3365 aligned_start + (nth_task+1)*chunk_size); 3366 // Try and bump the global finger via a CAS; 3367 // note that we need to do the global finger bump 3368 // _before_ taking the intersection below, because 3369 // the task corresponding to that region will be 3370 // deemed done even if the used_region() expands 3371 // because of allocation -- as it almost certainly will 3372 // during start-up while the threads yield in the 3373 // closure below. 3374 HeapWord* finger = span.end(); 3375 bump_global_finger(finger); // atomically 3376 // There are null tasks here corresponding to chunks 3377 // beyond the "top" address of the space. 3378 span = span.intersection(sp->used_region()); 3379 if (!span.is_empty()) { // Non-null task 3380 HeapWord* prev_obj; 3381 assert(!span.contains(_restart_addr) || nth_task == 0, 3382 "Inconsistency"); 3383 if (nth_task == 0) { 3384 // For the 0th task, we'll not need to compute a block_start. 3385 if (span.contains(_restart_addr)) { 3386 // In the case of a restart because of stack overflow, 3387 // we might additionally skip a chunk prefix. 3388 prev_obj = _restart_addr; 3389 } else { 3390 prev_obj = span.start(); 3391 } 3392 } else { 3393 // We want to skip the first object because 3394 // the protocol is to scan any object in its entirety 3395 // that _starts_ in this span; a fortiori, any 3396 // object starting in an earlier span is scanned 3397 // as part of an earlier claimed task. 3398 // Below we use the "careful" version of block_start 3399 // so we do not try to navigate uninitialized objects. 3400 prev_obj = sp->block_start_careful(span.start()); 3401 // Below we use a variant of block_size that uses the 3402 // Printezis bits to avoid waiting for allocated 3403 // objects to become initialized/parsable. 3404 while (prev_obj < span.start()) { 3405 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3406 if (sz > 0) { 3407 prev_obj += sz; 3408 } else { 3409 // In this case we may end up doing a bit of redundant 3410 // scanning, but that appears unavoidable, short of 3411 // locking the free list locks; see bug 6324141. 3412 break; 3413 } 3414 } 3415 } 3416 if (prev_obj < span.end()) { 3417 MemRegion my_span = MemRegion(prev_obj, span.end()); 3418 // Do the marking work within a non-empty span -- 3419 // the last argument to the constructor indicates whether the 3420 // iteration should be incremental with periodic yields. 3421 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3422 &_collector->_markBitMap, 3423 work_queue(i), 3424 &_collector->_markStack); 3425 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3426 } // else nothing to do for this task 3427 } // else nothing to do for this task 3428 } 3429 // We'd be tempted to assert here that since there are no 3430 // more tasks left to claim in this space, the global_finger 3431 // must exceed space->top() and a fortiori space->end(). However, 3432 // that would not quite be correct because the bumping of 3433 // global_finger occurs strictly after the claiming of a task, 3434 // so by the time we reach here the global finger may not yet 3435 // have been bumped up by the thread that claimed the last 3436 // task. 3437 pst->all_tasks_completed(); 3438 } 3439 3440 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3441 private: 3442 CMSCollector* _collector; 3443 CMSConcMarkingTask* _task; 3444 MemRegion _span; 3445 CMSBitMap* _bit_map; 3446 CMSMarkStack* _overflow_stack; 3447 OopTaskQueue* _work_queue; 3448 protected: 3449 DO_OOP_WORK_DEFN 3450 public: 3451 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3452 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3453 MetadataAwareOopClosure(collector->ref_processor()), 3454 _collector(collector), 3455 _task(task), 3456 _span(collector->_span), 3457 _work_queue(work_queue), 3458 _bit_map(bit_map), 3459 _overflow_stack(overflow_stack) 3460 { } 3461 virtual void do_oop(oop* p); 3462 virtual void do_oop(narrowOop* p); 3463 3464 void trim_queue(size_t max); 3465 void handle_stack_overflow(HeapWord* lost); 3466 void do_yield_check() { 3467 if (_task->should_yield()) { 3468 _task->yield(); 3469 } 3470 } 3471 }; 3472 3473 // Grey object scanning during work stealing phase -- 3474 // the salient assumption here is that any references 3475 // that are in these stolen objects being scanned must 3476 // already have been initialized (else they would not have 3477 // been published), so we do not need to check for 3478 // uninitialized objects before pushing here. 3479 void Par_ConcMarkingClosure::do_oop(oop obj) { 3480 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3481 HeapWord* addr = (HeapWord*)obj; 3482 // Check if oop points into the CMS generation 3483 // and is not marked 3484 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3485 // a white object ... 3486 // If we manage to "claim" the object, by being the 3487 // first thread to mark it, then we push it on our 3488 // marking stack 3489 if (_bit_map->par_mark(addr)) { // ... now grey 3490 // push on work queue (grey set) 3491 bool simulate_overflow = false; 3492 NOT_PRODUCT( 3493 if (CMSMarkStackOverflowALot && 3494 _collector->simulate_overflow()) { 3495 // simulate a stack overflow 3496 simulate_overflow = true; 3497 } 3498 ) 3499 if (simulate_overflow || 3500 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3501 // stack overflow 3502 if (PrintCMSStatistics != 0) { 3503 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3504 SIZE_FORMAT, _overflow_stack->capacity()); 3505 } 3506 // We cannot assert that the overflow stack is full because 3507 // it may have been emptied since. 3508 assert(simulate_overflow || 3509 _work_queue->size() == _work_queue->max_elems(), 3510 "Else push should have succeeded"); 3511 handle_stack_overflow(addr); 3512 } 3513 } // Else, some other thread got there first 3514 do_yield_check(); 3515 } 3516 } 3517 3518 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3519 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3520 3521 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3522 while (_work_queue->size() > max) { 3523 oop new_oop; 3524 if (_work_queue->pop_local(new_oop)) { 3525 assert(new_oop->is_oop(), "Should be an oop"); 3526 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3527 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3528 new_oop->oop_iterate(this); // do_oop() above 3529 do_yield_check(); 3530 } 3531 } 3532 } 3533 3534 // Upon stack overflow, we discard (part of) the stack, 3535 // remembering the least address amongst those discarded 3536 // in CMSCollector's _restart_address. 3537 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3538 // We need to do this under a mutex to prevent other 3539 // workers from interfering with the work done below. 3540 MutexLockerEx ml(_overflow_stack->par_lock(), 3541 Mutex::_no_safepoint_check_flag); 3542 // Remember the least grey address discarded 3543 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3544 _collector->lower_restart_addr(ra); 3545 _overflow_stack->reset(); // discard stack contents 3546 _overflow_stack->expand(); // expand the stack if possible 3547 } 3548 3549 3550 void CMSConcMarkingTask::do_work_steal(int i) { 3551 OopTaskQueue* work_q = work_queue(i); 3552 oop obj_to_scan; 3553 CMSBitMap* bm = &(_collector->_markBitMap); 3554 CMSMarkStack* ovflw = &(_collector->_markStack); 3555 int* seed = _collector->hash_seed(i); 3556 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3557 while (true) { 3558 cl.trim_queue(0); 3559 assert(work_q->size() == 0, "Should have been emptied above"); 3560 if (get_work_from_overflow_stack(ovflw, work_q)) { 3561 // Can't assert below because the work obtained from the 3562 // overflow stack may already have been stolen from us. 3563 // assert(work_q->size() > 0, "Work from overflow stack"); 3564 continue; 3565 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3566 assert(obj_to_scan->is_oop(), "Should be an oop"); 3567 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3568 obj_to_scan->oop_iterate(&cl); 3569 } else if (terminator()->offer_termination(&_term_term)) { 3570 assert(work_q->size() == 0, "Impossible!"); 3571 break; 3572 } else if (yielding() || should_yield()) { 3573 yield(); 3574 } 3575 } 3576 } 3577 3578 // This is run by the CMS (coordinator) thread. 3579 void CMSConcMarkingTask::coordinator_yield() { 3580 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3581 "CMS thread should hold CMS token"); 3582 // First give up the locks, then yield, then re-lock 3583 // We should probably use a constructor/destructor idiom to 3584 // do this unlock/lock or modify the MutexUnlocker class to 3585 // serve our purpose. XXX 3586 assert_lock_strong(_bit_map_lock); 3587 _bit_map_lock->unlock(); 3588 ConcurrentMarkSweepThread::desynchronize(true); 3589 _collector->stopTimer(); 3590 if (PrintCMSStatistics != 0) { 3591 _collector->incrementYields(); 3592 } 3593 3594 // It is possible for whichever thread initiated the yield request 3595 // not to get a chance to wake up and take the bitmap lock between 3596 // this thread releasing it and reacquiring it. So, while the 3597 // should_yield() flag is on, let's sleep for a bit to give the 3598 // other thread a chance to wake up. The limit imposed on the number 3599 // of iterations is defensive, to avoid any unforseen circumstances 3600 // putting us into an infinite loop. Since it's always been this 3601 // (coordinator_yield()) method that was observed to cause the 3602 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3603 // which is by default non-zero. For the other seven methods that 3604 // also perform the yield operation, as are using a different 3605 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3606 // can enable the sleeping for those methods too, if necessary. 3607 // See 6442774. 3608 // 3609 // We really need to reconsider the synchronization between the GC 3610 // thread and the yield-requesting threads in the future and we 3611 // should really use wait/notify, which is the recommended 3612 // way of doing this type of interaction. Additionally, we should 3613 // consolidate the eight methods that do the yield operation and they 3614 // are almost identical into one for better maintainability and 3615 // readability. See 6445193. 3616 // 3617 // Tony 2006.06.29 3618 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3619 ConcurrentMarkSweepThread::should_yield() && 3620 !CMSCollector::foregroundGCIsActive(); ++i) { 3621 os::sleep(Thread::current(), 1, false); 3622 } 3623 3624 ConcurrentMarkSweepThread::synchronize(true); 3625 _bit_map_lock->lock_without_safepoint_check(); 3626 _collector->startTimer(); 3627 } 3628 3629 bool CMSCollector::do_marking_mt() { 3630 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3631 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3632 conc_workers()->active_workers(), 3633 Threads::number_of_non_daemon_threads()); 3634 conc_workers()->set_active_workers(num_workers); 3635 3636 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3637 3638 CMSConcMarkingTask tsk(this, 3639 cms_space, 3640 conc_workers(), 3641 task_queues()); 3642 3643 // Since the actual number of workers we get may be different 3644 // from the number we requested above, do we need to do anything different 3645 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3646 // class?? XXX 3647 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3648 3649 // Refs discovery is already non-atomic. 3650 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3651 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3652 conc_workers()->start_task(&tsk); 3653 while (tsk.yielded()) { 3654 tsk.coordinator_yield(); 3655 conc_workers()->continue_task(&tsk); 3656 } 3657 // If the task was aborted, _restart_addr will be non-NULL 3658 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3659 while (_restart_addr != NULL) { 3660 // XXX For now we do not make use of ABORTED state and have not 3661 // yet implemented the right abort semantics (even in the original 3662 // single-threaded CMS case). That needs some more investigation 3663 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3664 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3665 // If _restart_addr is non-NULL, a marking stack overflow 3666 // occurred; we need to do a fresh marking iteration from the 3667 // indicated restart address. 3668 if (_foregroundGCIsActive) { 3669 // We may be running into repeated stack overflows, having 3670 // reached the limit of the stack size, while making very 3671 // slow forward progress. It may be best to bail out and 3672 // let the foreground collector do its job. 3673 // Clear _restart_addr, so that foreground GC 3674 // works from scratch. This avoids the headache of 3675 // a "rescan" which would otherwise be needed because 3676 // of the dirty mod union table & card table. 3677 _restart_addr = NULL; 3678 return false; 3679 } 3680 // Adjust the task to restart from _restart_addr 3681 tsk.reset(_restart_addr); 3682 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3683 _restart_addr); 3684 _restart_addr = NULL; 3685 // Get the workers going again 3686 conc_workers()->start_task(&tsk); 3687 while (tsk.yielded()) { 3688 tsk.coordinator_yield(); 3689 conc_workers()->continue_task(&tsk); 3690 } 3691 } 3692 assert(tsk.completed(), "Inconsistency"); 3693 assert(tsk.result() == true, "Inconsistency"); 3694 return true; 3695 } 3696 3697 bool CMSCollector::do_marking_st() { 3698 ResourceMark rm; 3699 HandleMark hm; 3700 3701 // Temporarily make refs discovery single threaded (non-MT) 3702 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3703 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3704 &_markStack, CMSYield); 3705 // the last argument to iterate indicates whether the iteration 3706 // should be incremental with periodic yields. 3707 _markBitMap.iterate(&markFromRootsClosure); 3708 // If _restart_addr is non-NULL, a marking stack overflow 3709 // occurred; we need to do a fresh iteration from the 3710 // indicated restart address. 3711 while (_restart_addr != NULL) { 3712 if (_foregroundGCIsActive) { 3713 // We may be running into repeated stack overflows, having 3714 // reached the limit of the stack size, while making very 3715 // slow forward progress. It may be best to bail out and 3716 // let the foreground collector do its job. 3717 // Clear _restart_addr, so that foreground GC 3718 // works from scratch. This avoids the headache of 3719 // a "rescan" which would otherwise be needed because 3720 // of the dirty mod union table & card table. 3721 _restart_addr = NULL; 3722 return false; // indicating failure to complete marking 3723 } 3724 // Deal with stack overflow: 3725 // we restart marking from _restart_addr 3726 HeapWord* ra = _restart_addr; 3727 markFromRootsClosure.reset(ra); 3728 _restart_addr = NULL; 3729 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3730 } 3731 return true; 3732 } 3733 3734 void CMSCollector::preclean() { 3735 check_correct_thread_executing(); 3736 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3737 verify_work_stacks_empty(); 3738 verify_overflow_empty(); 3739 _abort_preclean = false; 3740 if (CMSPrecleaningEnabled) { 3741 if (!CMSEdenChunksRecordAlways) { 3742 _eden_chunk_index = 0; 3743 } 3744 size_t used = get_eden_used(); 3745 size_t capacity = get_eden_capacity(); 3746 // Don't start sampling unless we will get sufficiently 3747 // many samples. 3748 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3749 * CMSScheduleRemarkEdenPenetration)) { 3750 _start_sampling = true; 3751 } else { 3752 _start_sampling = false; 3753 } 3754 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3755 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3756 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3757 } 3758 CMSTokenSync x(true); // is cms thread 3759 if (CMSPrecleaningEnabled) { 3760 sample_eden(); 3761 _collectorState = AbortablePreclean; 3762 } else { 3763 _collectorState = FinalMarking; 3764 } 3765 verify_work_stacks_empty(); 3766 verify_overflow_empty(); 3767 } 3768 3769 // Try and schedule the remark such that young gen 3770 // occupancy is CMSScheduleRemarkEdenPenetration %. 3771 void CMSCollector::abortable_preclean() { 3772 check_correct_thread_executing(); 3773 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3774 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3775 3776 // If Eden's current occupancy is below this threshold, 3777 // immediately schedule the remark; else preclean 3778 // past the next scavenge in an effort to 3779 // schedule the pause as described above. By choosing 3780 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3781 // we will never do an actual abortable preclean cycle. 3782 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3783 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3784 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3785 // We need more smarts in the abortable preclean 3786 // loop below to deal with cases where allocation 3787 // in young gen is very very slow, and our precleaning 3788 // is running a losing race against a horde of 3789 // mutators intent on flooding us with CMS updates 3790 // (dirty cards). 3791 // One, admittedly dumb, strategy is to give up 3792 // after a certain number of abortable precleaning loops 3793 // or after a certain maximum time. We want to make 3794 // this smarter in the next iteration. 3795 // XXX FIX ME!!! YSR 3796 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3797 while (!(should_abort_preclean() || 3798 ConcurrentMarkSweepThread::should_terminate())) { 3799 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3800 cumworkdone += workdone; 3801 loops++; 3802 // Voluntarily terminate abortable preclean phase if we have 3803 // been at it for too long. 3804 if ((CMSMaxAbortablePrecleanLoops != 0) && 3805 loops >= CMSMaxAbortablePrecleanLoops) { 3806 if (PrintGCDetails) { 3807 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3808 } 3809 break; 3810 } 3811 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3812 if (PrintGCDetails) { 3813 gclog_or_tty->print(" CMS: abort preclean due to time "); 3814 } 3815 break; 3816 } 3817 // If we are doing little work each iteration, we should 3818 // take a short break. 3819 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3820 // Sleep for some time, waiting for work to accumulate 3821 stopTimer(); 3822 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3823 startTimer(); 3824 waited++; 3825 } 3826 } 3827 if (PrintCMSStatistics > 0) { 3828 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3829 loops, waited, cumworkdone); 3830 } 3831 } 3832 CMSTokenSync x(true); // is cms thread 3833 if (_collectorState != Idling) { 3834 assert(_collectorState == AbortablePreclean, 3835 "Spontaneous state transition?"); 3836 _collectorState = FinalMarking; 3837 } // Else, a foreground collection completed this CMS cycle. 3838 return; 3839 } 3840 3841 // Respond to an Eden sampling opportunity 3842 void CMSCollector::sample_eden() { 3843 // Make sure a young gc cannot sneak in between our 3844 // reading and recording of a sample. 3845 assert(Thread::current()->is_ConcurrentGC_thread(), 3846 "Only the cms thread may collect Eden samples"); 3847 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3848 "Should collect samples while holding CMS token"); 3849 if (!_start_sampling) { 3850 return; 3851 } 3852 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3853 // is populated by the young generation. 3854 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3855 if (_eden_chunk_index < _eden_chunk_capacity) { 3856 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3857 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3858 "Unexpected state of Eden"); 3859 // We'd like to check that what we just sampled is an oop-start address; 3860 // however, we cannot do that here since the object may not yet have been 3861 // initialized. So we'll instead do the check when we _use_ this sample 3862 // later. 3863 if (_eden_chunk_index == 0 || 3864 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3865 _eden_chunk_array[_eden_chunk_index-1]) 3866 >= CMSSamplingGrain)) { 3867 _eden_chunk_index++; // commit sample 3868 } 3869 } 3870 } 3871 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3872 size_t used = get_eden_used(); 3873 size_t capacity = get_eden_capacity(); 3874 assert(used <= capacity, "Unexpected state of Eden"); 3875 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3876 _abort_preclean = true; 3877 } 3878 } 3879 } 3880 3881 3882 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3883 assert(_collectorState == Precleaning || 3884 _collectorState == AbortablePreclean, "incorrect state"); 3885 ResourceMark rm; 3886 HandleMark hm; 3887 3888 // Precleaning is currently not MT but the reference processor 3889 // may be set for MT. Disable it temporarily here. 3890 ReferenceProcessor* rp = ref_processor(); 3891 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3892 3893 // Do one pass of scrubbing the discovered reference lists 3894 // to remove any reference objects with strongly-reachable 3895 // referents. 3896 if (clean_refs) { 3897 CMSPrecleanRefsYieldClosure yield_cl(this); 3898 assert(rp->span().equals(_span), "Spans should be equal"); 3899 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3900 &_markStack, true /* preclean */); 3901 CMSDrainMarkingStackClosure complete_trace(this, 3902 _span, &_markBitMap, &_markStack, 3903 &keep_alive, true /* preclean */); 3904 3905 // We don't want this step to interfere with a young 3906 // collection because we don't want to take CPU 3907 // or memory bandwidth away from the young GC threads 3908 // (which may be as many as there are CPUs). 3909 // Note that we don't need to protect ourselves from 3910 // interference with mutators because they can't 3911 // manipulate the discovered reference lists nor affect 3912 // the computed reachability of the referents, the 3913 // only properties manipulated by the precleaning 3914 // of these reference lists. 3915 stopTimer(); 3916 CMSTokenSyncWithLocks x(true /* is cms thread */, 3917 bitMapLock()); 3918 startTimer(); 3919 sample_eden(); 3920 3921 // The following will yield to allow foreground 3922 // collection to proceed promptly. XXX YSR: 3923 // The code in this method may need further 3924 // tweaking for better performance and some restructuring 3925 // for cleaner interfaces. 3926 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3927 rp->preclean_discovered_references( 3928 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3929 gc_timer, _gc_tracer_cm->gc_id()); 3930 } 3931 3932 if (clean_survivor) { // preclean the active survivor space(s) 3933 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3934 &_markBitMap, &_modUnionTable, 3935 &_markStack, true /* precleaning phase */); 3936 stopTimer(); 3937 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3938 bitMapLock()); 3939 startTimer(); 3940 unsigned int before_count = 3941 GenCollectedHeap::heap()->total_collections(); 3942 SurvivorSpacePrecleanClosure 3943 sss_cl(this, _span, &_markBitMap, &_markStack, 3944 &pam_cl, before_count, CMSYield); 3945 _young_gen->from()->object_iterate_careful(&sss_cl); 3946 _young_gen->to()->object_iterate_careful(&sss_cl); 3947 } 3948 MarkRefsIntoAndScanClosure 3949 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3950 &_markStack, this, CMSYield, 3951 true /* precleaning phase */); 3952 // CAUTION: The following closure has persistent state that may need to 3953 // be reset upon a decrease in the sequence of addresses it 3954 // processes. 3955 ScanMarkedObjectsAgainCarefullyClosure 3956 smoac_cl(this, _span, 3957 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3958 3959 // Preclean dirty cards in ModUnionTable and CardTable using 3960 // appropriate convergence criterion; 3961 // repeat CMSPrecleanIter times unless we find that 3962 // we are losing. 3963 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3964 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3965 "Bad convergence multiplier"); 3966 assert(CMSPrecleanThreshold >= 100, 3967 "Unreasonably low CMSPrecleanThreshold"); 3968 3969 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3970 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3971 numIter < CMSPrecleanIter; 3972 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3973 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3974 if (Verbose && PrintGCDetails) { 3975 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3976 } 3977 // Either there are very few dirty cards, so re-mark 3978 // pause will be small anyway, or our pre-cleaning isn't 3979 // that much faster than the rate at which cards are being 3980 // dirtied, so we might as well stop and re-mark since 3981 // precleaning won't improve our re-mark time by much. 3982 if (curNumCards <= CMSPrecleanThreshold || 3983 (numIter > 0 && 3984 (curNumCards * CMSPrecleanDenominator > 3985 lastNumCards * CMSPrecleanNumerator))) { 3986 numIter++; 3987 cumNumCards += curNumCards; 3988 break; 3989 } 3990 } 3991 3992 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3993 3994 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3995 cumNumCards += curNumCards; 3996 if (PrintGCDetails && PrintCMSStatistics != 0) { 3997 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3998 curNumCards, cumNumCards, numIter); 3999 } 4000 return cumNumCards; // as a measure of useful work done 4001 } 4002 4003 // PRECLEANING NOTES: 4004 // Precleaning involves: 4005 // . reading the bits of the modUnionTable and clearing the set bits. 4006 // . For the cards corresponding to the set bits, we scan the 4007 // objects on those cards. This means we need the free_list_lock 4008 // so that we can safely iterate over the CMS space when scanning 4009 // for oops. 4010 // . When we scan the objects, we'll be both reading and setting 4011 // marks in the marking bit map, so we'll need the marking bit map. 4012 // . For protecting _collector_state transitions, we take the CGC_lock. 4013 // Note that any races in the reading of of card table entries by the 4014 // CMS thread on the one hand and the clearing of those entries by the 4015 // VM thread or the setting of those entries by the mutator threads on the 4016 // other are quite benign. However, for efficiency it makes sense to keep 4017 // the VM thread from racing with the CMS thread while the latter is 4018 // dirty card info to the modUnionTable. We therefore also use the 4019 // CGC_lock to protect the reading of the card table and the mod union 4020 // table by the CM thread. 4021 // . We run concurrently with mutator updates, so scanning 4022 // needs to be done carefully -- we should not try to scan 4023 // potentially uninitialized objects. 4024 // 4025 // Locking strategy: While holding the CGC_lock, we scan over and 4026 // reset a maximal dirty range of the mod union / card tables, then lock 4027 // the free_list_lock and bitmap lock to do a full marking, then 4028 // release these locks; and repeat the cycle. This allows for a 4029 // certain amount of fairness in the sharing of these locks between 4030 // the CMS collector on the one hand, and the VM thread and the 4031 // mutators on the other. 4032 4033 // NOTE: preclean_mod_union_table() and preclean_card_table() 4034 // further below are largely identical; if you need to modify 4035 // one of these methods, please check the other method too. 4036 4037 size_t CMSCollector::preclean_mod_union_table( 4038 ConcurrentMarkSweepGeneration* gen, 4039 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4040 verify_work_stacks_empty(); 4041 verify_overflow_empty(); 4042 4043 // strategy: starting with the first card, accumulate contiguous 4044 // ranges of dirty cards; clear these cards, then scan the region 4045 // covered by these cards. 4046 4047 // Since all of the MUT is committed ahead, we can just use 4048 // that, in case the generations expand while we are precleaning. 4049 // It might also be fine to just use the committed part of the 4050 // generation, but we might potentially miss cards when the 4051 // generation is rapidly expanding while we are in the midst 4052 // of precleaning. 4053 HeapWord* startAddr = gen->reserved().start(); 4054 HeapWord* endAddr = gen->reserved().end(); 4055 4056 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4057 4058 size_t numDirtyCards, cumNumDirtyCards; 4059 HeapWord *nextAddr, *lastAddr; 4060 for (cumNumDirtyCards = numDirtyCards = 0, 4061 nextAddr = lastAddr = startAddr; 4062 nextAddr < endAddr; 4063 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4064 4065 ResourceMark rm; 4066 HandleMark hm; 4067 4068 MemRegion dirtyRegion; 4069 { 4070 stopTimer(); 4071 // Potential yield point 4072 CMSTokenSync ts(true); 4073 startTimer(); 4074 sample_eden(); 4075 // Get dirty region starting at nextOffset (inclusive), 4076 // simultaneously clearing it. 4077 dirtyRegion = 4078 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4079 assert(dirtyRegion.start() >= nextAddr, 4080 "returned region inconsistent?"); 4081 } 4082 // Remember where the next search should begin. 4083 // The returned region (if non-empty) is a right open interval, 4084 // so lastOffset is obtained from the right end of that 4085 // interval. 4086 lastAddr = dirtyRegion.end(); 4087 // Should do something more transparent and less hacky XXX 4088 numDirtyCards = 4089 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4090 4091 // We'll scan the cards in the dirty region (with periodic 4092 // yields for foreground GC as needed). 4093 if (!dirtyRegion.is_empty()) { 4094 assert(numDirtyCards > 0, "consistency check"); 4095 HeapWord* stop_point = NULL; 4096 stopTimer(); 4097 // Potential yield point 4098 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4099 bitMapLock()); 4100 startTimer(); 4101 { 4102 verify_work_stacks_empty(); 4103 verify_overflow_empty(); 4104 sample_eden(); 4105 stop_point = 4106 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4107 } 4108 if (stop_point != NULL) { 4109 // The careful iteration stopped early either because it found an 4110 // uninitialized object, or because we were in the midst of an 4111 // "abortable preclean", which should now be aborted. Redirty 4112 // the bits corresponding to the partially-scanned or unscanned 4113 // cards. We'll either restart at the next block boundary or 4114 // abort the preclean. 4115 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4116 "Should only be AbortablePreclean."); 4117 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4118 if (should_abort_preclean()) { 4119 break; // out of preclean loop 4120 } else { 4121 // Compute the next address at which preclean should pick up; 4122 // might need bitMapLock in order to read P-bits. 4123 lastAddr = next_card_start_after_block(stop_point); 4124 } 4125 } 4126 } else { 4127 assert(lastAddr == endAddr, "consistency check"); 4128 assert(numDirtyCards == 0, "consistency check"); 4129 break; 4130 } 4131 } 4132 verify_work_stacks_empty(); 4133 verify_overflow_empty(); 4134 return cumNumDirtyCards; 4135 } 4136 4137 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4138 // below are largely identical; if you need to modify 4139 // one of these methods, please check the other method too. 4140 4141 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4142 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4143 // strategy: it's similar to precleamModUnionTable above, in that 4144 // we accumulate contiguous ranges of dirty cards, mark these cards 4145 // precleaned, then scan the region covered by these cards. 4146 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4147 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4148 4149 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4150 4151 size_t numDirtyCards, cumNumDirtyCards; 4152 HeapWord *lastAddr, *nextAddr; 4153 4154 for (cumNumDirtyCards = numDirtyCards = 0, 4155 nextAddr = lastAddr = startAddr; 4156 nextAddr < endAddr; 4157 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4158 4159 ResourceMark rm; 4160 HandleMark hm; 4161 4162 MemRegion dirtyRegion; 4163 { 4164 // See comments in "Precleaning notes" above on why we 4165 // do this locking. XXX Could the locking overheads be 4166 // too high when dirty cards are sparse? [I don't think so.] 4167 stopTimer(); 4168 CMSTokenSync x(true); // is cms thread 4169 startTimer(); 4170 sample_eden(); 4171 // Get and clear dirty region from card table 4172 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4173 MemRegion(nextAddr, endAddr), 4174 true, 4175 CardTableModRefBS::precleaned_card_val()); 4176 4177 assert(dirtyRegion.start() >= nextAddr, 4178 "returned region inconsistent?"); 4179 } 4180 lastAddr = dirtyRegion.end(); 4181 numDirtyCards = 4182 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4183 4184 if (!dirtyRegion.is_empty()) { 4185 stopTimer(); 4186 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4187 startTimer(); 4188 sample_eden(); 4189 verify_work_stacks_empty(); 4190 verify_overflow_empty(); 4191 HeapWord* stop_point = 4192 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4193 if (stop_point != NULL) { 4194 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4195 "Should only be AbortablePreclean."); 4196 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4197 if (should_abort_preclean()) { 4198 break; // out of preclean loop 4199 } else { 4200 // Compute the next address at which preclean should pick up. 4201 lastAddr = next_card_start_after_block(stop_point); 4202 } 4203 } 4204 } else { 4205 break; 4206 } 4207 } 4208 verify_work_stacks_empty(); 4209 verify_overflow_empty(); 4210 return cumNumDirtyCards; 4211 } 4212 4213 class PrecleanKlassClosure : public KlassClosure { 4214 KlassToOopClosure _cm_klass_closure; 4215 public: 4216 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4217 void do_klass(Klass* k) { 4218 if (k->has_accumulated_modified_oops()) { 4219 k->clear_accumulated_modified_oops(); 4220 4221 _cm_klass_closure.do_klass(k); 4222 } 4223 } 4224 }; 4225 4226 // The freelist lock is needed to prevent asserts, is it really needed? 4227 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4228 4229 cl->set_freelistLock(freelistLock); 4230 4231 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4232 4233 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4234 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4235 PrecleanKlassClosure preclean_klass_closure(cl); 4236 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4237 4238 verify_work_stacks_empty(); 4239 verify_overflow_empty(); 4240 } 4241 4242 void CMSCollector::checkpointRootsFinal() { 4243 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4244 check_correct_thread_executing(); 4245 // world is stopped at this checkpoint 4246 assert(SafepointSynchronize::is_at_safepoint(), 4247 "world should be stopped"); 4248 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4249 4250 verify_work_stacks_empty(); 4251 verify_overflow_empty(); 4252 4253 if (PrintGCDetails) { 4254 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4255 _young_gen->used() / K, 4256 _young_gen->capacity() / K); 4257 } 4258 { 4259 if (CMSScavengeBeforeRemark) { 4260 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4261 // Temporarily set flag to false, GCH->do_collection will 4262 // expect it to be false and set to true 4263 FlagSetting fl(gch->_is_gc_active, false); 4264 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4265 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4266 gch->do_collection(true, // full (i.e. force, see below) 4267 false, // !clear_all_soft_refs 4268 0, // size 4269 false, // is_tlab 4270 Generation::Young // type 4271 ); 4272 } 4273 FreelistLocker x(this); 4274 MutexLockerEx y(bitMapLock(), 4275 Mutex::_no_safepoint_check_flag); 4276 checkpointRootsFinalWork(); 4277 } 4278 verify_work_stacks_empty(); 4279 verify_overflow_empty(); 4280 } 4281 4282 void CMSCollector::checkpointRootsFinalWork() { 4283 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4284 4285 assert(haveFreelistLocks(), "must have free list locks"); 4286 assert_lock_strong(bitMapLock()); 4287 4288 ResourceMark rm; 4289 HandleMark hm; 4290 4291 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4292 4293 if (should_unload_classes()) { 4294 CodeCache::gc_prologue(); 4295 } 4296 assert(haveFreelistLocks(), "must have free list locks"); 4297 assert_lock_strong(bitMapLock()); 4298 4299 // We might assume that we need not fill TLAB's when 4300 // CMSScavengeBeforeRemark is set, because we may have just done 4301 // a scavenge which would have filled all TLAB's -- and besides 4302 // Eden would be empty. This however may not always be the case -- 4303 // for instance although we asked for a scavenge, it may not have 4304 // happened because of a JNI critical section. We probably need 4305 // a policy for deciding whether we can in that case wait until 4306 // the critical section releases and then do the remark following 4307 // the scavenge, and skip it here. In the absence of that policy, 4308 // or of an indication of whether the scavenge did indeed occur, 4309 // we cannot rely on TLAB's having been filled and must do 4310 // so here just in case a scavenge did not happen. 4311 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4312 // Update the saved marks which may affect the root scans. 4313 gch->save_marks(); 4314 4315 if (CMSPrintEdenSurvivorChunks) { 4316 print_eden_and_survivor_chunk_arrays(); 4317 } 4318 4319 { 4320 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4321 4322 // Note on the role of the mod union table: 4323 // Since the marker in "markFromRoots" marks concurrently with 4324 // mutators, it is possible for some reachable objects not to have been 4325 // scanned. For instance, an only reference to an object A was 4326 // placed in object B after the marker scanned B. Unless B is rescanned, 4327 // A would be collected. Such updates to references in marked objects 4328 // are detected via the mod union table which is the set of all cards 4329 // dirtied since the first checkpoint in this GC cycle and prior to 4330 // the most recent young generation GC, minus those cleaned up by the 4331 // concurrent precleaning. 4332 if (CMSParallelRemarkEnabled) { 4333 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4334 do_remark_parallel(); 4335 } else { 4336 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4337 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4338 do_remark_non_parallel(); 4339 } 4340 } 4341 verify_work_stacks_empty(); 4342 verify_overflow_empty(); 4343 4344 { 4345 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4346 refProcessingWork(); 4347 } 4348 verify_work_stacks_empty(); 4349 verify_overflow_empty(); 4350 4351 if (should_unload_classes()) { 4352 CodeCache::gc_epilogue(); 4353 } 4354 JvmtiExport::gc_epilogue(); 4355 4356 // If we encountered any (marking stack / work queue) overflow 4357 // events during the current CMS cycle, take appropriate 4358 // remedial measures, where possible, so as to try and avoid 4359 // recurrence of that condition. 4360 assert(_markStack.isEmpty(), "No grey objects"); 4361 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4362 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4363 if (ser_ovflw > 0) { 4364 if (PrintCMSStatistics != 0) { 4365 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4366 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4367 ", kac_preclean="SIZE_FORMAT")", 4368 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4369 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4370 } 4371 _markStack.expand(); 4372 _ser_pmc_remark_ovflw = 0; 4373 _ser_pmc_preclean_ovflw = 0; 4374 _ser_kac_preclean_ovflw = 0; 4375 _ser_kac_ovflw = 0; 4376 } 4377 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4378 if (PrintCMSStatistics != 0) { 4379 gclog_or_tty->print_cr("Work queue overflow (benign) " 4380 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4381 _par_pmc_remark_ovflw, _par_kac_ovflw); 4382 } 4383 _par_pmc_remark_ovflw = 0; 4384 _par_kac_ovflw = 0; 4385 } 4386 if (PrintCMSStatistics != 0) { 4387 if (_markStack._hit_limit > 0) { 4388 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4389 _markStack._hit_limit); 4390 } 4391 if (_markStack._failed_double > 0) { 4392 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4393 " current capacity "SIZE_FORMAT, 4394 _markStack._failed_double, 4395 _markStack.capacity()); 4396 } 4397 } 4398 _markStack._hit_limit = 0; 4399 _markStack._failed_double = 0; 4400 4401 if ((VerifyAfterGC || VerifyDuringGC) && 4402 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4403 verify_after_remark(); 4404 } 4405 4406 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4407 4408 // Change under the freelistLocks. 4409 _collectorState = Sweeping; 4410 // Call isAllClear() under bitMapLock 4411 assert(_modUnionTable.isAllClear(), 4412 "Should be clear by end of the final marking"); 4413 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4414 "Should be clear by end of the final marking"); 4415 } 4416 4417 void CMSParInitialMarkTask::work(uint worker_id) { 4418 elapsedTimer _timer; 4419 ResourceMark rm; 4420 HandleMark hm; 4421 4422 // ---------- scan from roots -------------- 4423 _timer.start(); 4424 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4425 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4426 4427 // ---------- young gen roots -------------- 4428 { 4429 work_on_young_gen_roots(worker_id, &par_mri_cl); 4430 _timer.stop(); 4431 if (PrintCMSStatistics != 0) { 4432 gclog_or_tty->print_cr( 4433 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4434 worker_id, _timer.seconds()); 4435 } 4436 } 4437 4438 // ---------- remaining roots -------------- 4439 _timer.reset(); 4440 _timer.start(); 4441 4442 CLDToOopClosure cld_closure(&par_mri_cl, true); 4443 4444 gch->gen_process_roots(Generation::Old, 4445 false, // yg was scanned above 4446 false, // this is parallel code 4447 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4448 _collector->should_unload_classes(), 4449 &par_mri_cl, 4450 NULL, 4451 &cld_closure); 4452 assert(_collector->should_unload_classes() 4453 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4454 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4455 _timer.stop(); 4456 if (PrintCMSStatistics != 0) { 4457 gclog_or_tty->print_cr( 4458 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4459 worker_id, _timer.seconds()); 4460 } 4461 } 4462 4463 // Parallel remark task 4464 class CMSParRemarkTask: public CMSParMarkTask { 4465 CompactibleFreeListSpace* _cms_space; 4466 4467 // The per-thread work queues, available here for stealing. 4468 OopTaskQueueSet* _task_queues; 4469 ParallelTaskTerminator _term; 4470 4471 public: 4472 // A value of 0 passed to n_workers will cause the number of 4473 // workers to be taken from the active workers in the work gang. 4474 CMSParRemarkTask(CMSCollector* collector, 4475 CompactibleFreeListSpace* cms_space, 4476 uint n_workers, FlexibleWorkGang* workers, 4477 OopTaskQueueSet* task_queues): 4478 CMSParMarkTask("Rescan roots and grey objects in parallel", 4479 collector, n_workers), 4480 _cms_space(cms_space), 4481 _task_queues(task_queues), 4482 _term(n_workers, task_queues) { } 4483 4484 OopTaskQueueSet* task_queues() { return _task_queues; } 4485 4486 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4487 4488 ParallelTaskTerminator* terminator() { return &_term; } 4489 uint n_workers() { return _n_workers; } 4490 4491 void work(uint worker_id); 4492 4493 private: 4494 // ... of dirty cards in old space 4495 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4496 Par_MarkRefsIntoAndScanClosure* cl); 4497 4498 // ... work stealing for the above 4499 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4500 }; 4501 4502 class RemarkKlassClosure : public KlassClosure { 4503 KlassToOopClosure _cm_klass_closure; 4504 public: 4505 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4506 void do_klass(Klass* k) { 4507 // Check if we have modified any oops in the Klass during the concurrent marking. 4508 if (k->has_accumulated_modified_oops()) { 4509 k->clear_accumulated_modified_oops(); 4510 4511 // We could have transfered the current modified marks to the accumulated marks, 4512 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4513 } else if (k->has_modified_oops()) { 4514 // Don't clear anything, this info is needed by the next young collection. 4515 } else { 4516 // No modified oops in the Klass. 4517 return; 4518 } 4519 4520 // The klass has modified fields, need to scan the klass. 4521 _cm_klass_closure.do_klass(k); 4522 } 4523 }; 4524 4525 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4526 ParNewGeneration* young_gen = _collector->_young_gen; 4527 ContiguousSpace* eden_space = young_gen->eden(); 4528 ContiguousSpace* from_space = young_gen->from(); 4529 ContiguousSpace* to_space = young_gen->to(); 4530 4531 HeapWord** eca = _collector->_eden_chunk_array; 4532 size_t ect = _collector->_eden_chunk_index; 4533 HeapWord** sca = _collector->_survivor_chunk_array; 4534 size_t sct = _collector->_survivor_chunk_index; 4535 4536 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4537 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4538 4539 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4540 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4541 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4542 } 4543 4544 // work_queue(i) is passed to the closure 4545 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4546 // also is passed to do_dirty_card_rescan_tasks() and to 4547 // do_work_steal() to select the i-th task_queue. 4548 4549 void CMSParRemarkTask::work(uint worker_id) { 4550 elapsedTimer _timer; 4551 ResourceMark rm; 4552 HandleMark hm; 4553 4554 // ---------- rescan from roots -------------- 4555 _timer.start(); 4556 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4557 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4558 _collector->_span, _collector->ref_processor(), 4559 &(_collector->_markBitMap), 4560 work_queue(worker_id)); 4561 4562 // Rescan young gen roots first since these are likely 4563 // coarsely partitioned and may, on that account, constitute 4564 // the critical path; thus, it's best to start off that 4565 // work first. 4566 // ---------- young gen roots -------------- 4567 { 4568 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4569 _timer.stop(); 4570 if (PrintCMSStatistics != 0) { 4571 gclog_or_tty->print_cr( 4572 "Finished young gen rescan work in %dth thread: %3.3f sec", 4573 worker_id, _timer.seconds()); 4574 } 4575 } 4576 4577 // ---------- remaining roots -------------- 4578 _timer.reset(); 4579 _timer.start(); 4580 gch->gen_process_roots(Generation::Old, 4581 false, // yg was scanned above 4582 false, // this is parallel code 4583 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4584 _collector->should_unload_classes(), 4585 &par_mrias_cl, 4586 NULL, 4587 NULL); // The dirty klasses will be handled below 4588 4589 assert(_collector->should_unload_classes() 4590 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4591 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4592 _timer.stop(); 4593 if (PrintCMSStatistics != 0) { 4594 gclog_or_tty->print_cr( 4595 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4596 worker_id, _timer.seconds()); 4597 } 4598 4599 // ---------- unhandled CLD scanning ---------- 4600 if (worker_id == 0) { // Single threaded at the moment. 4601 _timer.reset(); 4602 _timer.start(); 4603 4604 // Scan all new class loader data objects and new dependencies that were 4605 // introduced during concurrent marking. 4606 ResourceMark rm; 4607 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4608 for (int i = 0; i < array->length(); i++) { 4609 par_mrias_cl.do_class_loader_data(array->at(i)); 4610 } 4611 4612 // We don't need to keep track of new CLDs anymore. 4613 ClassLoaderDataGraph::remember_new_clds(false); 4614 4615 _timer.stop(); 4616 if (PrintCMSStatistics != 0) { 4617 gclog_or_tty->print_cr( 4618 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4619 worker_id, _timer.seconds()); 4620 } 4621 } 4622 4623 // ---------- dirty klass scanning ---------- 4624 if (worker_id == 0) { // Single threaded at the moment. 4625 _timer.reset(); 4626 _timer.start(); 4627 4628 // Scan all classes that was dirtied during the concurrent marking phase. 4629 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4630 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4631 4632 _timer.stop(); 4633 if (PrintCMSStatistics != 0) { 4634 gclog_or_tty->print_cr( 4635 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4636 worker_id, _timer.seconds()); 4637 } 4638 } 4639 4640 // We might have added oops to ClassLoaderData::_handles during the 4641 // concurrent marking phase. These oops point to newly allocated objects 4642 // that are guaranteed to be kept alive. Either by the direct allocation 4643 // code, or when the young collector processes the roots. Hence, 4644 // we don't have to revisit the _handles block during the remark phase. 4645 4646 // ---------- rescan dirty cards ------------ 4647 _timer.reset(); 4648 _timer.start(); 4649 4650 // Do the rescan tasks for each of the two spaces 4651 // (cms_space) in turn. 4652 // "worker_id" is passed to select the task_queue for "worker_id" 4653 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4654 _timer.stop(); 4655 if (PrintCMSStatistics != 0) { 4656 gclog_or_tty->print_cr( 4657 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4658 worker_id, _timer.seconds()); 4659 } 4660 4661 // ---------- steal work from other threads ... 4662 // ---------- ... and drain overflow list. 4663 _timer.reset(); 4664 _timer.start(); 4665 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4666 _timer.stop(); 4667 if (PrintCMSStatistics != 0) { 4668 gclog_or_tty->print_cr( 4669 "Finished work stealing in %dth thread: %3.3f sec", 4670 worker_id, _timer.seconds()); 4671 } 4672 } 4673 4674 // Note that parameter "i" is not used. 4675 void 4676 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4677 OopsInGenClosure* cl, ContiguousSpace* space, 4678 HeapWord** chunk_array, size_t chunk_top) { 4679 // Until all tasks completed: 4680 // . claim an unclaimed task 4681 // . compute region boundaries corresponding to task claimed 4682 // using chunk_array 4683 // . par_oop_iterate(cl) over that region 4684 4685 ResourceMark rm; 4686 HandleMark hm; 4687 4688 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4689 4690 uint nth_task = 0; 4691 uint n_tasks = pst->n_tasks(); 4692 4693 if (n_tasks > 0) { 4694 assert(pst->valid(), "Uninitialized use?"); 4695 HeapWord *start, *end; 4696 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4697 // We claimed task # nth_task; compute its boundaries. 4698 if (chunk_top == 0) { // no samples were taken 4699 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4700 start = space->bottom(); 4701 end = space->top(); 4702 } else if (nth_task == 0) { 4703 start = space->bottom(); 4704 end = chunk_array[nth_task]; 4705 } else if (nth_task < (uint)chunk_top) { 4706 assert(nth_task >= 1, "Control point invariant"); 4707 start = chunk_array[nth_task - 1]; 4708 end = chunk_array[nth_task]; 4709 } else { 4710 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4711 start = chunk_array[chunk_top - 1]; 4712 end = space->top(); 4713 } 4714 MemRegion mr(start, end); 4715 // Verify that mr is in space 4716 assert(mr.is_empty() || space->used_region().contains(mr), 4717 "Should be in space"); 4718 // Verify that "start" is an object boundary 4719 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4720 "Should be an oop"); 4721 space->par_oop_iterate(mr, cl); 4722 } 4723 pst->all_tasks_completed(); 4724 } 4725 } 4726 4727 void 4728 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4729 CompactibleFreeListSpace* sp, int i, 4730 Par_MarkRefsIntoAndScanClosure* cl) { 4731 // Until all tasks completed: 4732 // . claim an unclaimed task 4733 // . compute region boundaries corresponding to task claimed 4734 // . transfer dirty bits ct->mut for that region 4735 // . apply rescanclosure to dirty mut bits for that region 4736 4737 ResourceMark rm; 4738 HandleMark hm; 4739 4740 OopTaskQueue* work_q = work_queue(i); 4741 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4742 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4743 // CAUTION: This closure has state that persists across calls to 4744 // the work method dirty_range_iterate_clear() in that it has 4745 // embedded in it a (subtype of) UpwardsObjectClosure. The 4746 // use of that state in the embedded UpwardsObjectClosure instance 4747 // assumes that the cards are always iterated (even if in parallel 4748 // by several threads) in monotonically increasing order per each 4749 // thread. This is true of the implementation below which picks 4750 // card ranges (chunks) in monotonically increasing order globally 4751 // and, a-fortiori, in monotonically increasing order per thread 4752 // (the latter order being a subsequence of the former). 4753 // If the work code below is ever reorganized into a more chaotic 4754 // work-partitioning form than the current "sequential tasks" 4755 // paradigm, the use of that persistent state will have to be 4756 // revisited and modified appropriately. See also related 4757 // bug 4756801 work on which should examine this code to make 4758 // sure that the changes there do not run counter to the 4759 // assumptions made here and necessary for correctness and 4760 // efficiency. Note also that this code might yield inefficient 4761 // behavior in the case of very large objects that span one or 4762 // more work chunks. Such objects would potentially be scanned 4763 // several times redundantly. Work on 4756801 should try and 4764 // address that performance anomaly if at all possible. XXX 4765 MemRegion full_span = _collector->_span; 4766 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4767 MarkFromDirtyCardsClosure 4768 greyRescanClosure(_collector, full_span, // entire span of interest 4769 sp, bm, work_q, cl); 4770 4771 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4772 assert(pst->valid(), "Uninitialized use?"); 4773 uint nth_task = 0; 4774 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4775 MemRegion span = sp->used_region(); 4776 HeapWord* start_addr = span.start(); 4777 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4778 alignment); 4779 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4780 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4781 start_addr, "Check alignment"); 4782 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4783 chunk_size, "Check alignment"); 4784 4785 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4786 // Having claimed the nth_task, compute corresponding mem-region, 4787 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4788 // The alignment restriction ensures that we do not need any 4789 // synchronization with other gang-workers while setting or 4790 // clearing bits in thus chunk of the MUT. 4791 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4792 start_addr + (nth_task+1)*chunk_size); 4793 // The last chunk's end might be way beyond end of the 4794 // used region. In that case pull back appropriately. 4795 if (this_span.end() > end_addr) { 4796 this_span.set_end(end_addr); 4797 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4798 } 4799 // Iterate over the dirty cards covering this chunk, marking them 4800 // precleaned, and setting the corresponding bits in the mod union 4801 // table. Since we have been careful to partition at Card and MUT-word 4802 // boundaries no synchronization is needed between parallel threads. 4803 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4804 &modUnionClosure); 4805 4806 // Having transferred these marks into the modUnionTable, 4807 // rescan the marked objects on the dirty cards in the modUnionTable. 4808 // Even if this is at a synchronous collection, the initial marking 4809 // may have been done during an asynchronous collection so there 4810 // may be dirty bits in the mod-union table. 4811 _collector->_modUnionTable.dirty_range_iterate_clear( 4812 this_span, &greyRescanClosure); 4813 _collector->_modUnionTable.verifyNoOneBitsInRange( 4814 this_span.start(), 4815 this_span.end()); 4816 } 4817 pst->all_tasks_completed(); // declare that i am done 4818 } 4819 4820 // . see if we can share work_queues with ParNew? XXX 4821 void 4822 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4823 int* seed) { 4824 OopTaskQueue* work_q = work_queue(i); 4825 NOT_PRODUCT(int num_steals = 0;) 4826 oop obj_to_scan; 4827 CMSBitMap* bm = &(_collector->_markBitMap); 4828 4829 while (true) { 4830 // Completely finish any left over work from (an) earlier round(s) 4831 cl->trim_queue(0); 4832 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4833 (size_t)ParGCDesiredObjsFromOverflowList); 4834 // Now check if there's any work in the overflow list 4835 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4836 // only affects the number of attempts made to get work from the 4837 // overflow list and does not affect the number of workers. Just 4838 // pass ParallelGCThreads so this behavior is unchanged. 4839 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4840 work_q, 4841 ParallelGCThreads)) { 4842 // found something in global overflow list; 4843 // not yet ready to go stealing work from others. 4844 // We'd like to assert(work_q->size() != 0, ...) 4845 // because we just took work from the overflow list, 4846 // but of course we can't since all of that could have 4847 // been already stolen from us. 4848 // "He giveth and He taketh away." 4849 continue; 4850 } 4851 // Verify that we have no work before we resort to stealing 4852 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4853 // Try to steal from other queues that have work 4854 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4855 NOT_PRODUCT(num_steals++;) 4856 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4857 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4858 // Do scanning work 4859 obj_to_scan->oop_iterate(cl); 4860 // Loop around, finish this work, and try to steal some more 4861 } else if (terminator()->offer_termination()) { 4862 break; // nirvana from the infinite cycle 4863 } 4864 } 4865 NOT_PRODUCT( 4866 if (PrintCMSStatistics != 0) { 4867 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4868 } 4869 ) 4870 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4871 "Else our work is not yet done"); 4872 } 4873 4874 // Record object boundaries in _eden_chunk_array by sampling the eden 4875 // top in the slow-path eden object allocation code path and record 4876 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4877 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4878 // sampling in sample_eden() that activates during the part of the 4879 // preclean phase. 4880 void CMSCollector::sample_eden_chunk() { 4881 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4882 if (_eden_chunk_lock->try_lock()) { 4883 // Record a sample. This is the critical section. The contents 4884 // of the _eden_chunk_array have to be non-decreasing in the 4885 // address order. 4886 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4887 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4888 "Unexpected state of Eden"); 4889 if (_eden_chunk_index == 0 || 4890 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4891 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4892 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4893 _eden_chunk_index++; // commit sample 4894 } 4895 _eden_chunk_lock->unlock(); 4896 } 4897 } 4898 } 4899 4900 // Return a thread-local PLAB recording array, as appropriate. 4901 void* CMSCollector::get_data_recorder(int thr_num) { 4902 if (_survivor_plab_array != NULL && 4903 (CMSPLABRecordAlways || 4904 (_collectorState > Marking && _collectorState < FinalMarking))) { 4905 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4906 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4907 ca->reset(); // clear it so that fresh data is recorded 4908 return (void*) ca; 4909 } else { 4910 return NULL; 4911 } 4912 } 4913 4914 // Reset all the thread-local PLAB recording arrays 4915 void CMSCollector::reset_survivor_plab_arrays() { 4916 for (uint i = 0; i < ParallelGCThreads; i++) { 4917 _survivor_plab_array[i].reset(); 4918 } 4919 } 4920 4921 // Merge the per-thread plab arrays into the global survivor chunk 4922 // array which will provide the partitioning of the survivor space 4923 // for CMS initial scan and rescan. 4924 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4925 int no_of_gc_threads) { 4926 assert(_survivor_plab_array != NULL, "Error"); 4927 assert(_survivor_chunk_array != NULL, "Error"); 4928 assert(_collectorState == FinalMarking || 4929 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4930 for (int j = 0; j < no_of_gc_threads; j++) { 4931 _cursor[j] = 0; 4932 } 4933 HeapWord* top = surv->top(); 4934 size_t i; 4935 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4936 HeapWord* min_val = top; // Higher than any PLAB address 4937 uint min_tid = 0; // position of min_val this round 4938 for (int j = 0; j < no_of_gc_threads; j++) { 4939 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4940 if (_cursor[j] == cur_sca->end()) { 4941 continue; 4942 } 4943 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4944 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4945 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4946 if (cur_val < min_val) { 4947 min_tid = j; 4948 min_val = cur_val; 4949 } else { 4950 assert(cur_val < top, "All recorded addresses should be less"); 4951 } 4952 } 4953 // At this point min_val and min_tid are respectively 4954 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4955 // and the thread (j) that witnesses that address. 4956 // We record this address in the _survivor_chunk_array[i] 4957 // and increment _cursor[min_tid] prior to the next round i. 4958 if (min_val == top) { 4959 break; 4960 } 4961 _survivor_chunk_array[i] = min_val; 4962 _cursor[min_tid]++; 4963 } 4964 // We are all done; record the size of the _survivor_chunk_array 4965 _survivor_chunk_index = i; // exclusive: [0, i) 4966 if (PrintCMSStatistics > 0) { 4967 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4968 } 4969 // Verify that we used up all the recorded entries 4970 #ifdef ASSERT 4971 size_t total = 0; 4972 for (int j = 0; j < no_of_gc_threads; j++) { 4973 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4974 total += _cursor[j]; 4975 } 4976 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4977 // Check that the merged array is in sorted order 4978 if (total > 0) { 4979 for (size_t i = 0; i < total - 1; i++) { 4980 if (PrintCMSStatistics > 0) { 4981 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4982 i, p2i(_survivor_chunk_array[i])); 4983 } 4984 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4985 "Not sorted"); 4986 } 4987 } 4988 #endif // ASSERT 4989 } 4990 4991 // Set up the space's par_seq_tasks structure for work claiming 4992 // for parallel initial scan and rescan of young gen. 4993 // See ParRescanTask where this is currently used. 4994 void 4995 CMSCollector:: 4996 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4997 assert(n_threads > 0, "Unexpected n_threads argument"); 4998 4999 // Eden space 5000 if (!_young_gen->eden()->is_empty()) { 5001 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5002 assert(!pst->valid(), "Clobbering existing data?"); 5003 // Each valid entry in [0, _eden_chunk_index) represents a task. 5004 size_t n_tasks = _eden_chunk_index + 1; 5005 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5006 // Sets the condition for completion of the subtask (how many threads 5007 // need to finish in order to be done). 5008 pst->set_n_threads(n_threads); 5009 pst->set_n_tasks((int)n_tasks); 5010 } 5011 5012 // Merge the survivor plab arrays into _survivor_chunk_array 5013 if (_survivor_plab_array != NULL) { 5014 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5015 } else { 5016 assert(_survivor_chunk_index == 0, "Error"); 5017 } 5018 5019 // To space 5020 { 5021 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5022 assert(!pst->valid(), "Clobbering existing data?"); 5023 // Sets the condition for completion of the subtask (how many threads 5024 // need to finish in order to be done). 5025 pst->set_n_threads(n_threads); 5026 pst->set_n_tasks(1); 5027 assert(pst->valid(), "Error"); 5028 } 5029 5030 // From space 5031 { 5032 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5033 assert(!pst->valid(), "Clobbering existing data?"); 5034 size_t n_tasks = _survivor_chunk_index + 1; 5035 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5036 // Sets the condition for completion of the subtask (how many threads 5037 // need to finish in order to be done). 5038 pst->set_n_threads(n_threads); 5039 pst->set_n_tasks((int)n_tasks); 5040 assert(pst->valid(), "Error"); 5041 } 5042 } 5043 5044 // Parallel version of remark 5045 void CMSCollector::do_remark_parallel() { 5046 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5047 FlexibleWorkGang* workers = gch->workers(); 5048 assert(workers != NULL, "Need parallel worker threads."); 5049 // Choose to use the number of GC workers most recently set 5050 // into "active_workers". If active_workers is not set, set it 5051 // to ParallelGCThreads. 5052 uint n_workers = workers->active_workers(); 5053 if (n_workers == 0) { 5054 assert(n_workers > 0, "Should have been set during scavenge"); 5055 n_workers = ParallelGCThreads; 5056 workers->set_active_workers(n_workers); 5057 } 5058 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5059 5060 CMSParRemarkTask tsk(this, 5061 cms_space, 5062 n_workers, workers, task_queues()); 5063 5064 // Set up for parallel process_roots work. 5065 gch->set_par_threads(n_workers); 5066 // We won't be iterating over the cards in the card table updating 5067 // the younger_gen cards, so we shouldn't call the following else 5068 // the verification code as well as subsequent younger_refs_iterate 5069 // code would get confused. XXX 5070 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5071 5072 // The young gen rescan work will not be done as part of 5073 // process_roots (which currently doesn't know how to 5074 // parallelize such a scan), but rather will be broken up into 5075 // a set of parallel tasks (via the sampling that the [abortable] 5076 // preclean phase did of eden, plus the [two] tasks of 5077 // scanning the [two] survivor spaces. Further fine-grain 5078 // parallelization of the scanning of the survivor spaces 5079 // themselves, and of precleaning of the younger gen itself 5080 // is deferred to the future. 5081 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5082 5083 // The dirty card rescan work is broken up into a "sequence" 5084 // of parallel tasks (per constituent space) that are dynamically 5085 // claimed by the parallel threads. 5086 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5087 5088 // It turns out that even when we're using 1 thread, doing the work in a 5089 // separate thread causes wide variance in run times. We can't help this 5090 // in the multi-threaded case, but we special-case n=1 here to get 5091 // repeatable measurements of the 1-thread overhead of the parallel code. 5092 if (n_workers > 1) { 5093 // Make refs discovery MT-safe, if it isn't already: it may not 5094 // necessarily be so, since it's possible that we are doing 5095 // ST marking. 5096 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5097 StrongRootsScope srs; 5098 workers->run_task(&tsk); 5099 } else { 5100 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5101 StrongRootsScope srs; 5102 tsk.work(0); 5103 } 5104 5105 gch->set_par_threads(0); // 0 ==> non-parallel. 5106 // restore, single-threaded for now, any preserved marks 5107 // as a result of work_q overflow 5108 restore_preserved_marks_if_any(); 5109 } 5110 5111 // Non-parallel version of remark 5112 void CMSCollector::do_remark_non_parallel() { 5113 ResourceMark rm; 5114 HandleMark hm; 5115 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5116 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5117 5118 MarkRefsIntoAndScanClosure 5119 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5120 &_markStack, this, 5121 false /* should_yield */, false /* not precleaning */); 5122 MarkFromDirtyCardsClosure 5123 markFromDirtyCardsClosure(this, _span, 5124 NULL, // space is set further below 5125 &_markBitMap, &_markStack, &mrias_cl); 5126 { 5127 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5128 // Iterate over the dirty cards, setting the corresponding bits in the 5129 // mod union table. 5130 { 5131 ModUnionClosure modUnionClosure(&_modUnionTable); 5132 _ct->ct_bs()->dirty_card_iterate( 5133 _cmsGen->used_region(), 5134 &modUnionClosure); 5135 } 5136 // Having transferred these marks into the modUnionTable, we just need 5137 // to rescan the marked objects on the dirty cards in the modUnionTable. 5138 // The initial marking may have been done during an asynchronous 5139 // collection so there may be dirty bits in the mod-union table. 5140 const int alignment = 5141 CardTableModRefBS::card_size * BitsPerWord; 5142 { 5143 // ... First handle dirty cards in CMS gen 5144 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5145 MemRegion ur = _cmsGen->used_region(); 5146 HeapWord* lb = ur.start(); 5147 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5148 MemRegion cms_span(lb, ub); 5149 _modUnionTable.dirty_range_iterate_clear(cms_span, 5150 &markFromDirtyCardsClosure); 5151 verify_work_stacks_empty(); 5152 if (PrintCMSStatistics != 0) { 5153 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5154 markFromDirtyCardsClosure.num_dirty_cards()); 5155 } 5156 } 5157 } 5158 if (VerifyDuringGC && 5159 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5160 HandleMark hm; // Discard invalid handles created during verification 5161 Universe::verify(); 5162 } 5163 { 5164 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5165 5166 verify_work_stacks_empty(); 5167 5168 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5169 StrongRootsScope srs; 5170 5171 gch->gen_process_roots(Generation::Old, 5172 true, // younger gens as roots 5173 false, // use the local StrongRootsScope 5174 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5175 should_unload_classes(), 5176 &mrias_cl, 5177 NULL, 5178 NULL); // The dirty klasses will be handled below 5179 5180 assert(should_unload_classes() 5181 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5182 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5183 } 5184 5185 { 5186 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5187 5188 verify_work_stacks_empty(); 5189 5190 // Scan all class loader data objects that might have been introduced 5191 // during concurrent marking. 5192 ResourceMark rm; 5193 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5194 for (int i = 0; i < array->length(); i++) { 5195 mrias_cl.do_class_loader_data(array->at(i)); 5196 } 5197 5198 // We don't need to keep track of new CLDs anymore. 5199 ClassLoaderDataGraph::remember_new_clds(false); 5200 5201 verify_work_stacks_empty(); 5202 } 5203 5204 { 5205 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5206 5207 verify_work_stacks_empty(); 5208 5209 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5210 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5211 5212 verify_work_stacks_empty(); 5213 } 5214 5215 // We might have added oops to ClassLoaderData::_handles during the 5216 // concurrent marking phase. These oops point to newly allocated objects 5217 // that are guaranteed to be kept alive. Either by the direct allocation 5218 // code, or when the young collector processes the roots. Hence, 5219 // we don't have to revisit the _handles block during the remark phase. 5220 5221 verify_work_stacks_empty(); 5222 // Restore evacuated mark words, if any, used for overflow list links 5223 if (!CMSOverflowEarlyRestoration) { 5224 restore_preserved_marks_if_any(); 5225 } 5226 verify_overflow_empty(); 5227 } 5228 5229 //////////////////////////////////////////////////////// 5230 // Parallel Reference Processing Task Proxy Class 5231 //////////////////////////////////////////////////////// 5232 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5233 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5234 CMSCollector* _collector; 5235 CMSBitMap* _mark_bit_map; 5236 const MemRegion _span; 5237 ProcessTask& _task; 5238 5239 public: 5240 CMSRefProcTaskProxy(ProcessTask& task, 5241 CMSCollector* collector, 5242 const MemRegion& span, 5243 CMSBitMap* mark_bit_map, 5244 AbstractWorkGang* workers, 5245 OopTaskQueueSet* task_queues): 5246 // XXX Should superclass AGTWOQ also know about AWG since it knows 5247 // about the task_queues used by the AWG? Then it could initialize 5248 // the terminator() object. See 6984287. The set_for_termination() 5249 // below is a temporary band-aid for the regression in 6984287. 5250 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5251 task_queues), 5252 _task(task), 5253 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5254 { 5255 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5256 "Inconsistency in _span"); 5257 set_for_termination(workers->active_workers()); 5258 } 5259 5260 OopTaskQueueSet* task_queues() { return queues(); } 5261 5262 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5263 5264 void do_work_steal(int i, 5265 CMSParDrainMarkingStackClosure* drain, 5266 CMSParKeepAliveClosure* keep_alive, 5267 int* seed); 5268 5269 virtual void work(uint worker_id); 5270 }; 5271 5272 void CMSRefProcTaskProxy::work(uint worker_id) { 5273 ResourceMark rm; 5274 HandleMark hm; 5275 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5276 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5277 _mark_bit_map, 5278 work_queue(worker_id)); 5279 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5280 _mark_bit_map, 5281 work_queue(worker_id)); 5282 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5283 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5284 if (_task.marks_oops_alive()) { 5285 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5286 _collector->hash_seed(worker_id)); 5287 } 5288 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5289 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5290 } 5291 5292 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5293 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5294 EnqueueTask& _task; 5295 5296 public: 5297 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5298 : AbstractGangTask("Enqueue reference objects in parallel"), 5299 _task(task) 5300 { } 5301 5302 virtual void work(uint worker_id) 5303 { 5304 _task.work(worker_id); 5305 } 5306 }; 5307 5308 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5309 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5310 _span(span), 5311 _bit_map(bit_map), 5312 _work_queue(work_queue), 5313 _mark_and_push(collector, span, bit_map, work_queue), 5314 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5315 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5316 { } 5317 5318 // . see if we can share work_queues with ParNew? XXX 5319 void CMSRefProcTaskProxy::do_work_steal(int i, 5320 CMSParDrainMarkingStackClosure* drain, 5321 CMSParKeepAliveClosure* keep_alive, 5322 int* seed) { 5323 OopTaskQueue* work_q = work_queue(i); 5324 NOT_PRODUCT(int num_steals = 0;) 5325 oop obj_to_scan; 5326 5327 while (true) { 5328 // Completely finish any left over work from (an) earlier round(s) 5329 drain->trim_queue(0); 5330 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5331 (size_t)ParGCDesiredObjsFromOverflowList); 5332 // Now check if there's any work in the overflow list 5333 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5334 // only affects the number of attempts made to get work from the 5335 // overflow list and does not affect the number of workers. Just 5336 // pass ParallelGCThreads so this behavior is unchanged. 5337 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5338 work_q, 5339 ParallelGCThreads)) { 5340 // Found something in global overflow list; 5341 // not yet ready to go stealing work from others. 5342 // We'd like to assert(work_q->size() != 0, ...) 5343 // because we just took work from the overflow list, 5344 // but of course we can't, since all of that might have 5345 // been already stolen from us. 5346 continue; 5347 } 5348 // Verify that we have no work before we resort to stealing 5349 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5350 // Try to steal from other queues that have work 5351 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5352 NOT_PRODUCT(num_steals++;) 5353 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5354 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5355 // Do scanning work 5356 obj_to_scan->oop_iterate(keep_alive); 5357 // Loop around, finish this work, and try to steal some more 5358 } else if (terminator()->offer_termination()) { 5359 break; // nirvana from the infinite cycle 5360 } 5361 } 5362 NOT_PRODUCT( 5363 if (PrintCMSStatistics != 0) { 5364 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5365 } 5366 ) 5367 } 5368 5369 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5370 { 5371 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5372 FlexibleWorkGang* workers = gch->workers(); 5373 assert(workers != NULL, "Need parallel worker threads."); 5374 CMSRefProcTaskProxy rp_task(task, &_collector, 5375 _collector.ref_processor()->span(), 5376 _collector.markBitMap(), 5377 workers, _collector.task_queues()); 5378 workers->run_task(&rp_task); 5379 } 5380 5381 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5382 { 5383 5384 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5385 FlexibleWorkGang* workers = gch->workers(); 5386 assert(workers != NULL, "Need parallel worker threads."); 5387 CMSRefEnqueueTaskProxy enq_task(task); 5388 workers->run_task(&enq_task); 5389 } 5390 5391 void CMSCollector::refProcessingWork() { 5392 ResourceMark rm; 5393 HandleMark hm; 5394 5395 ReferenceProcessor* rp = ref_processor(); 5396 assert(rp->span().equals(_span), "Spans should be equal"); 5397 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5398 // Process weak references. 5399 rp->setup_policy(false); 5400 verify_work_stacks_empty(); 5401 5402 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5403 &_markStack, false /* !preclean */); 5404 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5405 _span, &_markBitMap, &_markStack, 5406 &cmsKeepAliveClosure, false /* !preclean */); 5407 { 5408 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5409 5410 ReferenceProcessorStats stats; 5411 if (rp->processing_is_mt()) { 5412 // Set the degree of MT here. If the discovery is done MT, there 5413 // may have been a different number of threads doing the discovery 5414 // and a different number of discovered lists may have Ref objects. 5415 // That is OK as long as the Reference lists are balanced (see 5416 // balance_all_queues() and balance_queues()). 5417 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5418 uint active_workers = ParallelGCThreads; 5419 FlexibleWorkGang* workers = gch->workers(); 5420 if (workers != NULL) { 5421 active_workers = workers->active_workers(); 5422 // The expectation is that active_workers will have already 5423 // been set to a reasonable value. If it has not been set, 5424 // investigate. 5425 assert(active_workers > 0, "Should have been set during scavenge"); 5426 } 5427 rp->set_active_mt_degree(active_workers); 5428 CMSRefProcTaskExecutor task_executor(*this); 5429 stats = rp->process_discovered_references(&_is_alive_closure, 5430 &cmsKeepAliveClosure, 5431 &cmsDrainMarkingStackClosure, 5432 &task_executor, 5433 _gc_timer_cm, 5434 _gc_tracer_cm->gc_id()); 5435 } else { 5436 stats = rp->process_discovered_references(&_is_alive_closure, 5437 &cmsKeepAliveClosure, 5438 &cmsDrainMarkingStackClosure, 5439 NULL, 5440 _gc_timer_cm, 5441 _gc_tracer_cm->gc_id()); 5442 } 5443 _gc_tracer_cm->report_gc_reference_stats(stats); 5444 5445 } 5446 5447 // This is the point where the entire marking should have completed. 5448 verify_work_stacks_empty(); 5449 5450 if (should_unload_classes()) { 5451 { 5452 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5453 5454 // Unload classes and purge the SystemDictionary. 5455 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5456 5457 // Unload nmethods. 5458 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5459 5460 // Prune dead klasses from subklass/sibling/implementor lists. 5461 Klass::clean_weak_klass_links(&_is_alive_closure); 5462 } 5463 5464 { 5465 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5466 // Clean up unreferenced symbols in symbol table. 5467 SymbolTable::unlink(); 5468 } 5469 5470 { 5471 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5472 // Delete entries for dead interned strings. 5473 StringTable::unlink(&_is_alive_closure); 5474 } 5475 } 5476 5477 5478 // Restore any preserved marks as a result of mark stack or 5479 // work queue overflow 5480 restore_preserved_marks_if_any(); // done single-threaded for now 5481 5482 rp->set_enqueuing_is_done(true); 5483 if (rp->processing_is_mt()) { 5484 rp->balance_all_queues(); 5485 CMSRefProcTaskExecutor task_executor(*this); 5486 rp->enqueue_discovered_references(&task_executor); 5487 } else { 5488 rp->enqueue_discovered_references(NULL); 5489 } 5490 rp->verify_no_references_recorded(); 5491 assert(!rp->discovery_enabled(), "should have been disabled"); 5492 } 5493 5494 #ifndef PRODUCT 5495 void CMSCollector::check_correct_thread_executing() { 5496 Thread* t = Thread::current(); 5497 // Only the VM thread or the CMS thread should be here. 5498 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5499 "Unexpected thread type"); 5500 // If this is the vm thread, the foreground process 5501 // should not be waiting. Note that _foregroundGCIsActive is 5502 // true while the foreground collector is waiting. 5503 if (_foregroundGCShouldWait) { 5504 // We cannot be the VM thread 5505 assert(t->is_ConcurrentGC_thread(), 5506 "Should be CMS thread"); 5507 } else { 5508 // We can be the CMS thread only if we are in a stop-world 5509 // phase of CMS collection. 5510 if (t->is_ConcurrentGC_thread()) { 5511 assert(_collectorState == InitialMarking || 5512 _collectorState == FinalMarking, 5513 "Should be a stop-world phase"); 5514 // The CMS thread should be holding the CMS_token. 5515 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5516 "Potential interference with concurrently " 5517 "executing VM thread"); 5518 } 5519 } 5520 } 5521 #endif 5522 5523 void CMSCollector::sweep() { 5524 assert(_collectorState == Sweeping, "just checking"); 5525 check_correct_thread_executing(); 5526 verify_work_stacks_empty(); 5527 verify_overflow_empty(); 5528 increment_sweep_count(); 5529 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5530 5531 _inter_sweep_timer.stop(); 5532 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5533 5534 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5535 _intra_sweep_timer.reset(); 5536 _intra_sweep_timer.start(); 5537 { 5538 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5539 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5540 // First sweep the old gen 5541 { 5542 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5543 bitMapLock()); 5544 sweepWork(_cmsGen); 5545 } 5546 5547 // Update Universe::_heap_*_at_gc figures. 5548 // We need all the free list locks to make the abstract state 5549 // transition from Sweeping to Resetting. See detailed note 5550 // further below. 5551 { 5552 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5553 // Update heap occupancy information which is used as 5554 // input to soft ref clearing policy at the next gc. 5555 Universe::update_heap_info_at_gc(); 5556 _collectorState = Resizing; 5557 } 5558 } 5559 verify_work_stacks_empty(); 5560 verify_overflow_empty(); 5561 5562 if (should_unload_classes()) { 5563 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5564 // requires that the virtual spaces are stable and not deleted. 5565 ClassLoaderDataGraph::set_should_purge(true); 5566 } 5567 5568 _intra_sweep_timer.stop(); 5569 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5570 5571 _inter_sweep_timer.reset(); 5572 _inter_sweep_timer.start(); 5573 5574 // We need to use a monotonically non-decreasing time in ms 5575 // or we will see time-warp warnings and os::javaTimeMillis() 5576 // does not guarantee monotonicity. 5577 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5578 update_time_of_last_gc(now); 5579 5580 // NOTE on abstract state transitions: 5581 // Mutators allocate-live and/or mark the mod-union table dirty 5582 // based on the state of the collection. The former is done in 5583 // the interval [Marking, Sweeping] and the latter in the interval 5584 // [Marking, Sweeping). Thus the transitions into the Marking state 5585 // and out of the Sweeping state must be synchronously visible 5586 // globally to the mutators. 5587 // The transition into the Marking state happens with the world 5588 // stopped so the mutators will globally see it. Sweeping is 5589 // done asynchronously by the background collector so the transition 5590 // from the Sweeping state to the Resizing state must be done 5591 // under the freelistLock (as is the check for whether to 5592 // allocate-live and whether to dirty the mod-union table). 5593 assert(_collectorState == Resizing, "Change of collector state to" 5594 " Resizing must be done under the freelistLocks (plural)"); 5595 5596 // Now that sweeping has been completed, we clear 5597 // the incremental_collection_failed flag, 5598 // thus inviting a younger gen collection to promote into 5599 // this generation. If such a promotion may still fail, 5600 // the flag will be set again when a young collection is 5601 // attempted. 5602 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5603 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5604 gch->update_full_collections_completed(_collection_count_start); 5605 } 5606 5607 // FIX ME!!! Looks like this belongs in CFLSpace, with 5608 // CMSGen merely delegating to it. 5609 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5610 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5611 HeapWord* minAddr = _cmsSpace->bottom(); 5612 HeapWord* largestAddr = 5613 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5614 if (largestAddr == NULL) { 5615 // The dictionary appears to be empty. In this case 5616 // try to coalesce at the end of the heap. 5617 largestAddr = _cmsSpace->end(); 5618 } 5619 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5620 size_t nearLargestOffset = 5621 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5622 if (PrintFLSStatistics != 0) { 5623 gclog_or_tty->print_cr( 5624 "CMS: Large Block: " PTR_FORMAT ";" 5625 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5626 p2i(largestAddr), 5627 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5628 } 5629 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5630 } 5631 5632 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5633 return addr >= _cmsSpace->nearLargestChunk(); 5634 } 5635 5636 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5637 return _cmsSpace->find_chunk_at_end(); 5638 } 5639 5640 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5641 bool full) { 5642 // If the young generation has been collected, gather any statistics 5643 // that are of interest at this point. 5644 bool current_is_young = (current_generation == GenCollectedHeap::heap()->young_gen()); 5645 if (!full && current_is_young) { 5646 // Gather statistics on the young generation collection. 5647 collector()->stats().record_gc0_end(used()); 5648 } 5649 } 5650 5651 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5652 // We iterate over the space(s) underlying this generation, 5653 // checking the mark bit map to see if the bits corresponding 5654 // to specific blocks are marked or not. Blocks that are 5655 // marked are live and are not swept up. All remaining blocks 5656 // are swept up, with coalescing on-the-fly as we sweep up 5657 // contiguous free and/or garbage blocks: 5658 // We need to ensure that the sweeper synchronizes with allocators 5659 // and stop-the-world collectors. In particular, the following 5660 // locks are used: 5661 // . CMS token: if this is held, a stop the world collection cannot occur 5662 // . freelistLock: if this is held no allocation can occur from this 5663 // generation by another thread 5664 // . bitMapLock: if this is held, no other thread can access or update 5665 // 5666 5667 // Note that we need to hold the freelistLock if we use 5668 // block iterate below; else the iterator might go awry if 5669 // a mutator (or promotion) causes block contents to change 5670 // (for instance if the allocator divvies up a block). 5671 // If we hold the free list lock, for all practical purposes 5672 // young generation GC's can't occur (they'll usually need to 5673 // promote), so we might as well prevent all young generation 5674 // GC's while we do a sweeping step. For the same reason, we might 5675 // as well take the bit map lock for the entire duration 5676 5677 // check that we hold the requisite locks 5678 assert(have_cms_token(), "Should hold cms token"); 5679 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5680 assert_lock_strong(gen->freelistLock()); 5681 assert_lock_strong(bitMapLock()); 5682 5683 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5684 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5685 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5686 _inter_sweep_estimate.padded_average(), 5687 _intra_sweep_estimate.padded_average()); 5688 gen->setNearLargestChunk(); 5689 5690 { 5691 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5692 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5693 // We need to free-up/coalesce garbage/blocks from a 5694 // co-terminal free run. This is done in the SweepClosure 5695 // destructor; so, do not remove this scope, else the 5696 // end-of-sweep-census below will be off by a little bit. 5697 } 5698 gen->cmsSpace()->sweep_completed(); 5699 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5700 if (should_unload_classes()) { // unloaded classes this cycle, 5701 _concurrent_cycles_since_last_unload = 0; // ... reset count 5702 } else { // did not unload classes, 5703 _concurrent_cycles_since_last_unload++; // ... increment count 5704 } 5705 } 5706 5707 // Reset CMS data structures (for now just the marking bit map) 5708 // preparatory for the next cycle. 5709 void CMSCollector::reset(bool concurrent) { 5710 if (concurrent) { 5711 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5712 5713 // If the state is not "Resetting", the foreground thread 5714 // has done a collection and the resetting. 5715 if (_collectorState != Resetting) { 5716 assert(_collectorState == Idling, "The state should only change" 5717 " because the foreground collector has finished the collection"); 5718 return; 5719 } 5720 5721 // Clear the mark bitmap (no grey objects to start with) 5722 // for the next cycle. 5723 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5724 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5725 5726 HeapWord* curAddr = _markBitMap.startWord(); 5727 while (curAddr < _markBitMap.endWord()) { 5728 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5729 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5730 _markBitMap.clear_large_range(chunk); 5731 if (ConcurrentMarkSweepThread::should_yield() && 5732 !foregroundGCIsActive() && 5733 CMSYield) { 5734 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5735 "CMS thread should hold CMS token"); 5736 assert_lock_strong(bitMapLock()); 5737 bitMapLock()->unlock(); 5738 ConcurrentMarkSweepThread::desynchronize(true); 5739 stopTimer(); 5740 if (PrintCMSStatistics != 0) { 5741 incrementYields(); 5742 } 5743 5744 // See the comment in coordinator_yield() 5745 for (unsigned i = 0; i < CMSYieldSleepCount && 5746 ConcurrentMarkSweepThread::should_yield() && 5747 !CMSCollector::foregroundGCIsActive(); ++i) { 5748 os::sleep(Thread::current(), 1, false); 5749 } 5750 5751 ConcurrentMarkSweepThread::synchronize(true); 5752 bitMapLock()->lock_without_safepoint_check(); 5753 startTimer(); 5754 } 5755 curAddr = chunk.end(); 5756 } 5757 // A successful mostly concurrent collection has been done. 5758 // Because only the full (i.e., concurrent mode failure) collections 5759 // are being measured for gc overhead limits, clean the "near" flag 5760 // and count. 5761 size_policy()->reset_gc_overhead_limit_count(); 5762 _collectorState = Idling; 5763 } else { 5764 // already have the lock 5765 assert(_collectorState == Resetting, "just checking"); 5766 assert_lock_strong(bitMapLock()); 5767 _markBitMap.clear_all(); 5768 _collectorState = Idling; 5769 } 5770 5771 register_gc_end(); 5772 } 5773 5774 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5775 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5776 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5777 TraceCollectorStats tcs(counters()); 5778 5779 switch (op) { 5780 case CMS_op_checkpointRootsInitial: { 5781 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5782 checkpointRootsInitial(); 5783 if (PrintGC) { 5784 _cmsGen->printOccupancy("initial-mark"); 5785 } 5786 break; 5787 } 5788 case CMS_op_checkpointRootsFinal: { 5789 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5790 checkpointRootsFinal(); 5791 if (PrintGC) { 5792 _cmsGen->printOccupancy("remark"); 5793 } 5794 break; 5795 } 5796 default: 5797 fatal("No such CMS_op"); 5798 } 5799 } 5800 5801 #ifndef PRODUCT 5802 size_t const CMSCollector::skip_header_HeapWords() { 5803 return FreeChunk::header_size(); 5804 } 5805 5806 // Try and collect here conditions that should hold when 5807 // CMS thread is exiting. The idea is that the foreground GC 5808 // thread should not be blocked if it wants to terminate 5809 // the CMS thread and yet continue to run the VM for a while 5810 // after that. 5811 void CMSCollector::verify_ok_to_terminate() const { 5812 assert(Thread::current()->is_ConcurrentGC_thread(), 5813 "should be called by CMS thread"); 5814 assert(!_foregroundGCShouldWait, "should be false"); 5815 // We could check here that all the various low-level locks 5816 // are not held by the CMS thread, but that is overkill; see 5817 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5818 // is checked. 5819 } 5820 #endif 5821 5822 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5823 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5824 "missing Printezis mark?"); 5825 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5826 size_t size = pointer_delta(nextOneAddr + 1, addr); 5827 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5828 "alignment problem"); 5829 assert(size >= 3, "Necessary for Printezis marks to work"); 5830 return size; 5831 } 5832 5833 // A variant of the above (block_size_using_printezis_bits()) except 5834 // that we return 0 if the P-bits are not yet set. 5835 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5836 if (_markBitMap.isMarked(addr + 1)) { 5837 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5838 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5839 size_t size = pointer_delta(nextOneAddr + 1, addr); 5840 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5841 "alignment problem"); 5842 assert(size >= 3, "Necessary for Printezis marks to work"); 5843 return size; 5844 } 5845 return 0; 5846 } 5847 5848 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5849 size_t sz = 0; 5850 oop p = (oop)addr; 5851 if (p->klass_or_null() != NULL) { 5852 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5853 } else { 5854 sz = block_size_using_printezis_bits(addr); 5855 } 5856 assert(sz > 0, "size must be nonzero"); 5857 HeapWord* next_block = addr + sz; 5858 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5859 CardTableModRefBS::card_size); 5860 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5861 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5862 "must be different cards"); 5863 return next_card; 5864 } 5865 5866 5867 // CMS Bit Map Wrapper ///////////////////////////////////////// 5868 5869 // Construct a CMS bit map infrastructure, but don't create the 5870 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5871 // further below. 5872 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5873 _bm(), 5874 _shifter(shifter), 5875 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5876 Monitor::_safepoint_check_sometimes) : NULL) 5877 { 5878 _bmStartWord = 0; 5879 _bmWordSize = 0; 5880 } 5881 5882 bool CMSBitMap::allocate(MemRegion mr) { 5883 _bmStartWord = mr.start(); 5884 _bmWordSize = mr.word_size(); 5885 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5886 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5887 if (!brs.is_reserved()) { 5888 warning("CMS bit map allocation failure"); 5889 return false; 5890 } 5891 // For now we'll just commit all of the bit map up front. 5892 // Later on we'll try to be more parsimonious with swap. 5893 if (!_virtual_space.initialize(brs, brs.size())) { 5894 warning("CMS bit map backing store failure"); 5895 return false; 5896 } 5897 assert(_virtual_space.committed_size() == brs.size(), 5898 "didn't reserve backing store for all of CMS bit map?"); 5899 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5900 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5901 _bmWordSize, "inconsistency in bit map sizing"); 5902 _bm.set_size(_bmWordSize >> _shifter); 5903 5904 // bm.clear(); // can we rely on getting zero'd memory? verify below 5905 assert(isAllClear(), 5906 "Expected zero'd memory from ReservedSpace constructor"); 5907 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5908 "consistency check"); 5909 return true; 5910 } 5911 5912 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5913 HeapWord *next_addr, *end_addr, *last_addr; 5914 assert_locked(); 5915 assert(covers(mr), "out-of-range error"); 5916 // XXX assert that start and end are appropriately aligned 5917 for (next_addr = mr.start(), end_addr = mr.end(); 5918 next_addr < end_addr; next_addr = last_addr) { 5919 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5920 last_addr = dirty_region.end(); 5921 if (!dirty_region.is_empty()) { 5922 cl->do_MemRegion(dirty_region); 5923 } else { 5924 assert(last_addr == end_addr, "program logic"); 5925 return; 5926 } 5927 } 5928 } 5929 5930 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5931 _bm.print_on_error(st, prefix); 5932 } 5933 5934 #ifndef PRODUCT 5935 void CMSBitMap::assert_locked() const { 5936 CMSLockVerifier::assert_locked(lock()); 5937 } 5938 5939 bool CMSBitMap::covers(MemRegion mr) const { 5940 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5941 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5942 "size inconsistency"); 5943 return (mr.start() >= _bmStartWord) && 5944 (mr.end() <= endWord()); 5945 } 5946 5947 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5948 return (start >= _bmStartWord && (start + size) <= endWord()); 5949 } 5950 5951 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5952 // verify that there are no 1 bits in the interval [left, right) 5953 FalseBitMapClosure falseBitMapClosure; 5954 iterate(&falseBitMapClosure, left, right); 5955 } 5956 5957 void CMSBitMap::region_invariant(MemRegion mr) 5958 { 5959 assert_locked(); 5960 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5961 assert(!mr.is_empty(), "unexpected empty region"); 5962 assert(covers(mr), "mr should be covered by bit map"); 5963 // convert address range into offset range 5964 size_t start_ofs = heapWordToOffset(mr.start()); 5965 // Make sure that end() is appropriately aligned 5966 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5967 (1 << (_shifter+LogHeapWordSize))), 5968 "Misaligned mr.end()"); 5969 size_t end_ofs = heapWordToOffset(mr.end()); 5970 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5971 } 5972 5973 #endif 5974 5975 bool CMSMarkStack::allocate(size_t size) { 5976 // allocate a stack of the requisite depth 5977 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5978 size * sizeof(oop))); 5979 if (!rs.is_reserved()) { 5980 warning("CMSMarkStack allocation failure"); 5981 return false; 5982 } 5983 if (!_virtual_space.initialize(rs, rs.size())) { 5984 warning("CMSMarkStack backing store failure"); 5985 return false; 5986 } 5987 assert(_virtual_space.committed_size() == rs.size(), 5988 "didn't reserve backing store for all of CMS stack?"); 5989 _base = (oop*)(_virtual_space.low()); 5990 _index = 0; 5991 _capacity = size; 5992 NOT_PRODUCT(_max_depth = 0); 5993 return true; 5994 } 5995 5996 // XXX FIX ME !!! In the MT case we come in here holding a 5997 // leaf lock. For printing we need to take a further lock 5998 // which has lower rank. We need to recalibrate the two 5999 // lock-ranks involved in order to be able to print the 6000 // messages below. (Or defer the printing to the caller. 6001 // For now we take the expedient path of just disabling the 6002 // messages for the problematic case.) 6003 void CMSMarkStack::expand() { 6004 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6005 if (_capacity == MarkStackSizeMax) { 6006 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6007 // We print a warning message only once per CMS cycle. 6008 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6009 } 6010 return; 6011 } 6012 // Double capacity if possible 6013 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6014 // Do not give up existing stack until we have managed to 6015 // get the double capacity that we desired. 6016 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6017 new_capacity * sizeof(oop))); 6018 if (rs.is_reserved()) { 6019 // Release the backing store associated with old stack 6020 _virtual_space.release(); 6021 // Reinitialize virtual space for new stack 6022 if (!_virtual_space.initialize(rs, rs.size())) { 6023 fatal("Not enough swap for expanded marking stack"); 6024 } 6025 _base = (oop*)(_virtual_space.low()); 6026 _index = 0; 6027 _capacity = new_capacity; 6028 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6029 // Failed to double capacity, continue; 6030 // we print a detail message only once per CMS cycle. 6031 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6032 SIZE_FORMAT"K", 6033 _capacity / K, new_capacity / K); 6034 } 6035 } 6036 6037 6038 // Closures 6039 // XXX: there seems to be a lot of code duplication here; 6040 // should refactor and consolidate common code. 6041 6042 // This closure is used to mark refs into the CMS generation in 6043 // the CMS bit map. Called at the first checkpoint. This closure 6044 // assumes that we do not need to re-mark dirty cards; if the CMS 6045 // generation on which this is used is not an oldest 6046 // generation then this will lose younger_gen cards! 6047 6048 MarkRefsIntoClosure::MarkRefsIntoClosure( 6049 MemRegion span, CMSBitMap* bitMap): 6050 _span(span), 6051 _bitMap(bitMap) 6052 { 6053 assert(_ref_processor == NULL, "deliberately left NULL"); 6054 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6055 } 6056 6057 void MarkRefsIntoClosure::do_oop(oop obj) { 6058 // if p points into _span, then mark corresponding bit in _markBitMap 6059 assert(obj->is_oop(), "expected an oop"); 6060 HeapWord* addr = (HeapWord*)obj; 6061 if (_span.contains(addr)) { 6062 // this should be made more efficient 6063 _bitMap->mark(addr); 6064 } 6065 } 6066 6067 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6068 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6069 6070 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6071 MemRegion span, CMSBitMap* bitMap): 6072 _span(span), 6073 _bitMap(bitMap) 6074 { 6075 assert(_ref_processor == NULL, "deliberately left NULL"); 6076 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6077 } 6078 6079 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6080 // if p points into _span, then mark corresponding bit in _markBitMap 6081 assert(obj->is_oop(), "expected an oop"); 6082 HeapWord* addr = (HeapWord*)obj; 6083 if (_span.contains(addr)) { 6084 // this should be made more efficient 6085 _bitMap->par_mark(addr); 6086 } 6087 } 6088 6089 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6090 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6091 6092 // A variant of the above, used for CMS marking verification. 6093 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6094 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6095 _span(span), 6096 _verification_bm(verification_bm), 6097 _cms_bm(cms_bm) 6098 { 6099 assert(_ref_processor == NULL, "deliberately left NULL"); 6100 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6101 } 6102 6103 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6104 // if p points into _span, then mark corresponding bit in _markBitMap 6105 assert(obj->is_oop(), "expected an oop"); 6106 HeapWord* addr = (HeapWord*)obj; 6107 if (_span.contains(addr)) { 6108 _verification_bm->mark(addr); 6109 if (!_cms_bm->isMarked(addr)) { 6110 oop(addr)->print(); 6111 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6112 fatal("... aborting"); 6113 } 6114 } 6115 } 6116 6117 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6118 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6119 6120 ////////////////////////////////////////////////// 6121 // MarkRefsIntoAndScanClosure 6122 ////////////////////////////////////////////////// 6123 6124 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6125 ReferenceProcessor* rp, 6126 CMSBitMap* bit_map, 6127 CMSBitMap* mod_union_table, 6128 CMSMarkStack* mark_stack, 6129 CMSCollector* collector, 6130 bool should_yield, 6131 bool concurrent_precleaning): 6132 _collector(collector), 6133 _span(span), 6134 _bit_map(bit_map), 6135 _mark_stack(mark_stack), 6136 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6137 mark_stack, concurrent_precleaning), 6138 _yield(should_yield), 6139 _concurrent_precleaning(concurrent_precleaning), 6140 _freelistLock(NULL) 6141 { 6142 _ref_processor = rp; 6143 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6144 } 6145 6146 // This closure is used to mark refs into the CMS generation at the 6147 // second (final) checkpoint, and to scan and transitively follow 6148 // the unmarked oops. It is also used during the concurrent precleaning 6149 // phase while scanning objects on dirty cards in the CMS generation. 6150 // The marks are made in the marking bit map and the marking stack is 6151 // used for keeping the (newly) grey objects during the scan. 6152 // The parallel version (Par_...) appears further below. 6153 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6154 if (obj != NULL) { 6155 assert(obj->is_oop(), "expected an oop"); 6156 HeapWord* addr = (HeapWord*)obj; 6157 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6158 assert(_collector->overflow_list_is_empty(), 6159 "overflow list should be empty"); 6160 if (_span.contains(addr) && 6161 !_bit_map->isMarked(addr)) { 6162 // mark bit map (object is now grey) 6163 _bit_map->mark(addr); 6164 // push on marking stack (stack should be empty), and drain the 6165 // stack by applying this closure to the oops in the oops popped 6166 // from the stack (i.e. blacken the grey objects) 6167 bool res = _mark_stack->push(obj); 6168 assert(res, "Should have space to push on empty stack"); 6169 do { 6170 oop new_oop = _mark_stack->pop(); 6171 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6172 assert(_bit_map->isMarked((HeapWord*)new_oop), 6173 "only grey objects on this stack"); 6174 // iterate over the oops in this oop, marking and pushing 6175 // the ones in CMS heap (i.e. in _span). 6176 new_oop->oop_iterate(&_pushAndMarkClosure); 6177 // check if it's time to yield 6178 do_yield_check(); 6179 } while (!_mark_stack->isEmpty() || 6180 (!_concurrent_precleaning && take_from_overflow_list())); 6181 // if marking stack is empty, and we are not doing this 6182 // during precleaning, then check the overflow list 6183 } 6184 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6185 assert(_collector->overflow_list_is_empty(), 6186 "overflow list was drained above"); 6187 // We could restore evacuated mark words, if any, used for 6188 // overflow list links here because the overflow list is 6189 // provably empty here. That would reduce the maximum 6190 // size requirements for preserved_{oop,mark}_stack. 6191 // But we'll just postpone it until we are all done 6192 // so we can just stream through. 6193 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6194 _collector->restore_preserved_marks_if_any(); 6195 assert(_collector->no_preserved_marks(), "No preserved marks"); 6196 } 6197 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6198 "All preserved marks should have been restored above"); 6199 } 6200 } 6201 6202 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6203 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6204 6205 void MarkRefsIntoAndScanClosure::do_yield_work() { 6206 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6207 "CMS thread should hold CMS token"); 6208 assert_lock_strong(_freelistLock); 6209 assert_lock_strong(_bit_map->lock()); 6210 // relinquish the free_list_lock and bitMaplock() 6211 _bit_map->lock()->unlock(); 6212 _freelistLock->unlock(); 6213 ConcurrentMarkSweepThread::desynchronize(true); 6214 _collector->stopTimer(); 6215 if (PrintCMSStatistics != 0) { 6216 _collector->incrementYields(); 6217 } 6218 6219 // See the comment in coordinator_yield() 6220 for (unsigned i = 0; 6221 i < CMSYieldSleepCount && 6222 ConcurrentMarkSweepThread::should_yield() && 6223 !CMSCollector::foregroundGCIsActive(); 6224 ++i) { 6225 os::sleep(Thread::current(), 1, false); 6226 } 6227 6228 ConcurrentMarkSweepThread::synchronize(true); 6229 _freelistLock->lock_without_safepoint_check(); 6230 _bit_map->lock()->lock_without_safepoint_check(); 6231 _collector->startTimer(); 6232 } 6233 6234 /////////////////////////////////////////////////////////// 6235 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6236 // MarkRefsIntoAndScanClosure 6237 /////////////////////////////////////////////////////////// 6238 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6239 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6240 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6241 _span(span), 6242 _bit_map(bit_map), 6243 _work_queue(work_queue), 6244 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6245 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6246 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6247 { 6248 _ref_processor = rp; 6249 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6250 } 6251 6252 // This closure is used to mark refs into the CMS generation at the 6253 // second (final) checkpoint, and to scan and transitively follow 6254 // the unmarked oops. The marks are made in the marking bit map and 6255 // the work_queue is used for keeping the (newly) grey objects during 6256 // the scan phase whence they are also available for stealing by parallel 6257 // threads. Since the marking bit map is shared, updates are 6258 // synchronized (via CAS). 6259 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6260 if (obj != NULL) { 6261 // Ignore mark word because this could be an already marked oop 6262 // that may be chained at the end of the overflow list. 6263 assert(obj->is_oop(true), "expected an oop"); 6264 HeapWord* addr = (HeapWord*)obj; 6265 if (_span.contains(addr) && 6266 !_bit_map->isMarked(addr)) { 6267 // mark bit map (object will become grey): 6268 // It is possible for several threads to be 6269 // trying to "claim" this object concurrently; 6270 // the unique thread that succeeds in marking the 6271 // object first will do the subsequent push on 6272 // to the work queue (or overflow list). 6273 if (_bit_map->par_mark(addr)) { 6274 // push on work_queue (which may not be empty), and trim the 6275 // queue to an appropriate length by applying this closure to 6276 // the oops in the oops popped from the stack (i.e. blacken the 6277 // grey objects) 6278 bool res = _work_queue->push(obj); 6279 assert(res, "Low water mark should be less than capacity?"); 6280 trim_queue(_low_water_mark); 6281 } // Else, another thread claimed the object 6282 } 6283 } 6284 } 6285 6286 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6287 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6288 6289 // This closure is used to rescan the marked objects on the dirty cards 6290 // in the mod union table and the card table proper. 6291 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6292 oop p, MemRegion mr) { 6293 6294 size_t size = 0; 6295 HeapWord* addr = (HeapWord*)p; 6296 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6297 assert(_span.contains(addr), "we are scanning the CMS generation"); 6298 // check if it's time to yield 6299 if (do_yield_check()) { 6300 // We yielded for some foreground stop-world work, 6301 // and we have been asked to abort this ongoing preclean cycle. 6302 return 0; 6303 } 6304 if (_bitMap->isMarked(addr)) { 6305 // it's marked; is it potentially uninitialized? 6306 if (p->klass_or_null() != NULL) { 6307 // an initialized object; ignore mark word in verification below 6308 // since we are running concurrent with mutators 6309 assert(p->is_oop(true), "should be an oop"); 6310 if (p->is_objArray()) { 6311 // objArrays are precisely marked; restrict scanning 6312 // to dirty cards only. 6313 size = CompactibleFreeListSpace::adjustObjectSize( 6314 p->oop_iterate(_scanningClosure, mr)); 6315 } else { 6316 // A non-array may have been imprecisely marked; we need 6317 // to scan object in its entirety. 6318 size = CompactibleFreeListSpace::adjustObjectSize( 6319 p->oop_iterate(_scanningClosure)); 6320 } 6321 #ifdef ASSERT 6322 size_t direct_size = 6323 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6324 assert(size == direct_size, "Inconsistency in size"); 6325 assert(size >= 3, "Necessary for Printezis marks to work"); 6326 if (!_bitMap->isMarked(addr+1)) { 6327 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6328 } else { 6329 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6330 assert(_bitMap->isMarked(addr+size-1), 6331 "inconsistent Printezis mark"); 6332 } 6333 #endif // ASSERT 6334 } else { 6335 // An uninitialized object. 6336 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6337 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6338 size = pointer_delta(nextOneAddr + 1, addr); 6339 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6340 "alignment problem"); 6341 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6342 // will dirty the card when the klass pointer is installed in the 6343 // object (signaling the completion of initialization). 6344 } 6345 } else { 6346 // Either a not yet marked object or an uninitialized object 6347 if (p->klass_or_null() == NULL) { 6348 // An uninitialized object, skip to the next card, since 6349 // we may not be able to read its P-bits yet. 6350 assert(size == 0, "Initial value"); 6351 } else { 6352 // An object not (yet) reached by marking: we merely need to 6353 // compute its size so as to go look at the next block. 6354 assert(p->is_oop(true), "should be an oop"); 6355 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6356 } 6357 } 6358 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6359 return size; 6360 } 6361 6362 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6363 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6364 "CMS thread should hold CMS token"); 6365 assert_lock_strong(_freelistLock); 6366 assert_lock_strong(_bitMap->lock()); 6367 // relinquish the free_list_lock and bitMaplock() 6368 _bitMap->lock()->unlock(); 6369 _freelistLock->unlock(); 6370 ConcurrentMarkSweepThread::desynchronize(true); 6371 _collector->stopTimer(); 6372 if (PrintCMSStatistics != 0) { 6373 _collector->incrementYields(); 6374 } 6375 6376 // See the comment in coordinator_yield() 6377 for (unsigned i = 0; i < CMSYieldSleepCount && 6378 ConcurrentMarkSweepThread::should_yield() && 6379 !CMSCollector::foregroundGCIsActive(); ++i) { 6380 os::sleep(Thread::current(), 1, false); 6381 } 6382 6383 ConcurrentMarkSweepThread::synchronize(true); 6384 _freelistLock->lock_without_safepoint_check(); 6385 _bitMap->lock()->lock_without_safepoint_check(); 6386 _collector->startTimer(); 6387 } 6388 6389 6390 ////////////////////////////////////////////////////////////////// 6391 // SurvivorSpacePrecleanClosure 6392 ////////////////////////////////////////////////////////////////// 6393 // This (single-threaded) closure is used to preclean the oops in 6394 // the survivor spaces. 6395 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6396 6397 HeapWord* addr = (HeapWord*)p; 6398 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6399 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6400 assert(p->klass_or_null() != NULL, "object should be initialized"); 6401 // an initialized object; ignore mark word in verification below 6402 // since we are running concurrent with mutators 6403 assert(p->is_oop(true), "should be an oop"); 6404 // Note that we do not yield while we iterate over 6405 // the interior oops of p, pushing the relevant ones 6406 // on our marking stack. 6407 size_t size = p->oop_iterate(_scanning_closure); 6408 do_yield_check(); 6409 // Observe that below, we do not abandon the preclean 6410 // phase as soon as we should; rather we empty the 6411 // marking stack before returning. This is to satisfy 6412 // some existing assertions. In general, it may be a 6413 // good idea to abort immediately and complete the marking 6414 // from the grey objects at a later time. 6415 while (!_mark_stack->isEmpty()) { 6416 oop new_oop = _mark_stack->pop(); 6417 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6418 assert(_bit_map->isMarked((HeapWord*)new_oop), 6419 "only grey objects on this stack"); 6420 // iterate over the oops in this oop, marking and pushing 6421 // the ones in CMS heap (i.e. in _span). 6422 new_oop->oop_iterate(_scanning_closure); 6423 // check if it's time to yield 6424 do_yield_check(); 6425 } 6426 unsigned int after_count = 6427 GenCollectedHeap::heap()->total_collections(); 6428 bool abort = (_before_count != after_count) || 6429 _collector->should_abort_preclean(); 6430 return abort ? 0 : size; 6431 } 6432 6433 void SurvivorSpacePrecleanClosure::do_yield_work() { 6434 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6435 "CMS thread should hold CMS token"); 6436 assert_lock_strong(_bit_map->lock()); 6437 // Relinquish the bit map lock 6438 _bit_map->lock()->unlock(); 6439 ConcurrentMarkSweepThread::desynchronize(true); 6440 _collector->stopTimer(); 6441 if (PrintCMSStatistics != 0) { 6442 _collector->incrementYields(); 6443 } 6444 6445 // See the comment in coordinator_yield() 6446 for (unsigned i = 0; i < CMSYieldSleepCount && 6447 ConcurrentMarkSweepThread::should_yield() && 6448 !CMSCollector::foregroundGCIsActive(); ++i) { 6449 os::sleep(Thread::current(), 1, false); 6450 } 6451 6452 ConcurrentMarkSweepThread::synchronize(true); 6453 _bit_map->lock()->lock_without_safepoint_check(); 6454 _collector->startTimer(); 6455 } 6456 6457 // This closure is used to rescan the marked objects on the dirty cards 6458 // in the mod union table and the card table proper. In the parallel 6459 // case, although the bitMap is shared, we do a single read so the 6460 // isMarked() query is "safe". 6461 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6462 // Ignore mark word because we are running concurrent with mutators 6463 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6464 HeapWord* addr = (HeapWord*)p; 6465 assert(_span.contains(addr), "we are scanning the CMS generation"); 6466 bool is_obj_array = false; 6467 #ifdef ASSERT 6468 if (!_parallel) { 6469 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6470 assert(_collector->overflow_list_is_empty(), 6471 "overflow list should be empty"); 6472 6473 } 6474 #endif // ASSERT 6475 if (_bit_map->isMarked(addr)) { 6476 // Obj arrays are precisely marked, non-arrays are not; 6477 // so we scan objArrays precisely and non-arrays in their 6478 // entirety. 6479 if (p->is_objArray()) { 6480 is_obj_array = true; 6481 if (_parallel) { 6482 p->oop_iterate(_par_scan_closure, mr); 6483 } else { 6484 p->oop_iterate(_scan_closure, mr); 6485 } 6486 } else { 6487 if (_parallel) { 6488 p->oop_iterate(_par_scan_closure); 6489 } else { 6490 p->oop_iterate(_scan_closure); 6491 } 6492 } 6493 } 6494 #ifdef ASSERT 6495 if (!_parallel) { 6496 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6497 assert(_collector->overflow_list_is_empty(), 6498 "overflow list should be empty"); 6499 6500 } 6501 #endif // ASSERT 6502 return is_obj_array; 6503 } 6504 6505 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6506 MemRegion span, 6507 CMSBitMap* bitMap, CMSMarkStack* markStack, 6508 bool should_yield, bool verifying): 6509 _collector(collector), 6510 _span(span), 6511 _bitMap(bitMap), 6512 _mut(&collector->_modUnionTable), 6513 _markStack(markStack), 6514 _yield(should_yield), 6515 _skipBits(0) 6516 { 6517 assert(_markStack->isEmpty(), "stack should be empty"); 6518 _finger = _bitMap->startWord(); 6519 _threshold = _finger; 6520 assert(_collector->_restart_addr == NULL, "Sanity check"); 6521 assert(_span.contains(_finger), "Out of bounds _finger?"); 6522 DEBUG_ONLY(_verifying = verifying;) 6523 } 6524 6525 void MarkFromRootsClosure::reset(HeapWord* addr) { 6526 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6527 assert(_span.contains(addr), "Out of bounds _finger?"); 6528 _finger = addr; 6529 _threshold = (HeapWord*)round_to( 6530 (intptr_t)_finger, CardTableModRefBS::card_size); 6531 } 6532 6533 // Should revisit to see if this should be restructured for 6534 // greater efficiency. 6535 bool MarkFromRootsClosure::do_bit(size_t offset) { 6536 if (_skipBits > 0) { 6537 _skipBits--; 6538 return true; 6539 } 6540 // convert offset into a HeapWord* 6541 HeapWord* addr = _bitMap->startWord() + offset; 6542 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6543 "address out of range"); 6544 assert(_bitMap->isMarked(addr), "tautology"); 6545 if (_bitMap->isMarked(addr+1)) { 6546 // this is an allocated but not yet initialized object 6547 assert(_skipBits == 0, "tautology"); 6548 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6549 oop p = oop(addr); 6550 if (p->klass_or_null() == NULL) { 6551 DEBUG_ONLY(if (!_verifying) {) 6552 // We re-dirty the cards on which this object lies and increase 6553 // the _threshold so that we'll come back to scan this object 6554 // during the preclean or remark phase. (CMSCleanOnEnter) 6555 if (CMSCleanOnEnter) { 6556 size_t sz = _collector->block_size_using_printezis_bits(addr); 6557 HeapWord* end_card_addr = (HeapWord*)round_to( 6558 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6559 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6560 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6561 // Bump _threshold to end_card_addr; note that 6562 // _threshold cannot possibly exceed end_card_addr, anyhow. 6563 // This prevents future clearing of the card as the scan proceeds 6564 // to the right. 6565 assert(_threshold <= end_card_addr, 6566 "Because we are just scanning into this object"); 6567 if (_threshold < end_card_addr) { 6568 _threshold = end_card_addr; 6569 } 6570 if (p->klass_or_null() != NULL) { 6571 // Redirty the range of cards... 6572 _mut->mark_range(redirty_range); 6573 } // ...else the setting of klass will dirty the card anyway. 6574 } 6575 DEBUG_ONLY(}) 6576 return true; 6577 } 6578 } 6579 scanOopsInOop(addr); 6580 return true; 6581 } 6582 6583 // We take a break if we've been at this for a while, 6584 // so as to avoid monopolizing the locks involved. 6585 void MarkFromRootsClosure::do_yield_work() { 6586 // First give up the locks, then yield, then re-lock 6587 // We should probably use a constructor/destructor idiom to 6588 // do this unlock/lock or modify the MutexUnlocker class to 6589 // serve our purpose. XXX 6590 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6591 "CMS thread should hold CMS token"); 6592 assert_lock_strong(_bitMap->lock()); 6593 _bitMap->lock()->unlock(); 6594 ConcurrentMarkSweepThread::desynchronize(true); 6595 _collector->stopTimer(); 6596 if (PrintCMSStatistics != 0) { 6597 _collector->incrementYields(); 6598 } 6599 6600 // See the comment in coordinator_yield() 6601 for (unsigned i = 0; i < CMSYieldSleepCount && 6602 ConcurrentMarkSweepThread::should_yield() && 6603 !CMSCollector::foregroundGCIsActive(); ++i) { 6604 os::sleep(Thread::current(), 1, false); 6605 } 6606 6607 ConcurrentMarkSweepThread::synchronize(true); 6608 _bitMap->lock()->lock_without_safepoint_check(); 6609 _collector->startTimer(); 6610 } 6611 6612 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6613 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6614 assert(_markStack->isEmpty(), 6615 "should drain stack to limit stack usage"); 6616 // convert ptr to an oop preparatory to scanning 6617 oop obj = oop(ptr); 6618 // Ignore mark word in verification below, since we 6619 // may be running concurrent with mutators. 6620 assert(obj->is_oop(true), "should be an oop"); 6621 assert(_finger <= ptr, "_finger runneth ahead"); 6622 // advance the finger to right end of this object 6623 _finger = ptr + obj->size(); 6624 assert(_finger > ptr, "we just incremented it above"); 6625 // On large heaps, it may take us some time to get through 6626 // the marking phase. During 6627 // this time it's possible that a lot of mutations have 6628 // accumulated in the card table and the mod union table -- 6629 // these mutation records are redundant until we have 6630 // actually traced into the corresponding card. 6631 // Here, we check whether advancing the finger would make 6632 // us cross into a new card, and if so clear corresponding 6633 // cards in the MUT (preclean them in the card-table in the 6634 // future). 6635 6636 DEBUG_ONLY(if (!_verifying) {) 6637 // The clean-on-enter optimization is disabled by default, 6638 // until we fix 6178663. 6639 if (CMSCleanOnEnter && (_finger > _threshold)) { 6640 // [_threshold, _finger) represents the interval 6641 // of cards to be cleared in MUT (or precleaned in card table). 6642 // The set of cards to be cleared is all those that overlap 6643 // with the interval [_threshold, _finger); note that 6644 // _threshold is always kept card-aligned but _finger isn't 6645 // always card-aligned. 6646 HeapWord* old_threshold = _threshold; 6647 assert(old_threshold == (HeapWord*)round_to( 6648 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6649 "_threshold should always be card-aligned"); 6650 _threshold = (HeapWord*)round_to( 6651 (intptr_t)_finger, CardTableModRefBS::card_size); 6652 MemRegion mr(old_threshold, _threshold); 6653 assert(!mr.is_empty(), "Control point invariant"); 6654 assert(_span.contains(mr), "Should clear within span"); 6655 _mut->clear_range(mr); 6656 } 6657 DEBUG_ONLY(}) 6658 // Note: the finger doesn't advance while we drain 6659 // the stack below. 6660 PushOrMarkClosure pushOrMarkClosure(_collector, 6661 _span, _bitMap, _markStack, 6662 _finger, this); 6663 bool res = _markStack->push(obj); 6664 assert(res, "Empty non-zero size stack should have space for single push"); 6665 while (!_markStack->isEmpty()) { 6666 oop new_oop = _markStack->pop(); 6667 // Skip verifying header mark word below because we are 6668 // running concurrent with mutators. 6669 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6670 // now scan this oop's oops 6671 new_oop->oop_iterate(&pushOrMarkClosure); 6672 do_yield_check(); 6673 } 6674 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6675 } 6676 6677 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6678 CMSCollector* collector, MemRegion span, 6679 CMSBitMap* bit_map, 6680 OopTaskQueue* work_queue, 6681 CMSMarkStack* overflow_stack): 6682 _collector(collector), 6683 _whole_span(collector->_span), 6684 _span(span), 6685 _bit_map(bit_map), 6686 _mut(&collector->_modUnionTable), 6687 _work_queue(work_queue), 6688 _overflow_stack(overflow_stack), 6689 _skip_bits(0), 6690 _task(task) 6691 { 6692 assert(_work_queue->size() == 0, "work_queue should be empty"); 6693 _finger = span.start(); 6694 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6695 assert(_span.contains(_finger), "Out of bounds _finger?"); 6696 } 6697 6698 // Should revisit to see if this should be restructured for 6699 // greater efficiency. 6700 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6701 if (_skip_bits > 0) { 6702 _skip_bits--; 6703 return true; 6704 } 6705 // convert offset into a HeapWord* 6706 HeapWord* addr = _bit_map->startWord() + offset; 6707 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6708 "address out of range"); 6709 assert(_bit_map->isMarked(addr), "tautology"); 6710 if (_bit_map->isMarked(addr+1)) { 6711 // this is an allocated object that might not yet be initialized 6712 assert(_skip_bits == 0, "tautology"); 6713 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6714 oop p = oop(addr); 6715 if (p->klass_or_null() == NULL) { 6716 // in the case of Clean-on-Enter optimization, redirty card 6717 // and avoid clearing card by increasing the threshold. 6718 return true; 6719 } 6720 } 6721 scan_oops_in_oop(addr); 6722 return true; 6723 } 6724 6725 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6726 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6727 // Should we assert that our work queue is empty or 6728 // below some drain limit? 6729 assert(_work_queue->size() == 0, 6730 "should drain stack to limit stack usage"); 6731 // convert ptr to an oop preparatory to scanning 6732 oop obj = oop(ptr); 6733 // Ignore mark word in verification below, since we 6734 // may be running concurrent with mutators. 6735 assert(obj->is_oop(true), "should be an oop"); 6736 assert(_finger <= ptr, "_finger runneth ahead"); 6737 // advance the finger to right end of this object 6738 _finger = ptr + obj->size(); 6739 assert(_finger > ptr, "we just incremented it above"); 6740 // On large heaps, it may take us some time to get through 6741 // the marking phase. During 6742 // this time it's possible that a lot of mutations have 6743 // accumulated in the card table and the mod union table -- 6744 // these mutation records are redundant until we have 6745 // actually traced into the corresponding card. 6746 // Here, we check whether advancing the finger would make 6747 // us cross into a new card, and if so clear corresponding 6748 // cards in the MUT (preclean them in the card-table in the 6749 // future). 6750 6751 // The clean-on-enter optimization is disabled by default, 6752 // until we fix 6178663. 6753 if (CMSCleanOnEnter && (_finger > _threshold)) { 6754 // [_threshold, _finger) represents the interval 6755 // of cards to be cleared in MUT (or precleaned in card table). 6756 // The set of cards to be cleared is all those that overlap 6757 // with the interval [_threshold, _finger); note that 6758 // _threshold is always kept card-aligned but _finger isn't 6759 // always card-aligned. 6760 HeapWord* old_threshold = _threshold; 6761 assert(old_threshold == (HeapWord*)round_to( 6762 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6763 "_threshold should always be card-aligned"); 6764 _threshold = (HeapWord*)round_to( 6765 (intptr_t)_finger, CardTableModRefBS::card_size); 6766 MemRegion mr(old_threshold, _threshold); 6767 assert(!mr.is_empty(), "Control point invariant"); 6768 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6769 _mut->clear_range(mr); 6770 } 6771 6772 // Note: the local finger doesn't advance while we drain 6773 // the stack below, but the global finger sure can and will. 6774 HeapWord** gfa = _task->global_finger_addr(); 6775 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6776 _span, _bit_map, 6777 _work_queue, 6778 _overflow_stack, 6779 _finger, 6780 gfa, this); 6781 bool res = _work_queue->push(obj); // overflow could occur here 6782 assert(res, "Will hold once we use workqueues"); 6783 while (true) { 6784 oop new_oop; 6785 if (!_work_queue->pop_local(new_oop)) { 6786 // We emptied our work_queue; check if there's stuff that can 6787 // be gotten from the overflow stack. 6788 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6789 _overflow_stack, _work_queue)) { 6790 do_yield_check(); 6791 continue; 6792 } else { // done 6793 break; 6794 } 6795 } 6796 // Skip verifying header mark word below because we are 6797 // running concurrent with mutators. 6798 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6799 // now scan this oop's oops 6800 new_oop->oop_iterate(&pushOrMarkClosure); 6801 do_yield_check(); 6802 } 6803 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6804 } 6805 6806 // Yield in response to a request from VM Thread or 6807 // from mutators. 6808 void Par_MarkFromRootsClosure::do_yield_work() { 6809 assert(_task != NULL, "sanity"); 6810 _task->yield(); 6811 } 6812 6813 // A variant of the above used for verifying CMS marking work. 6814 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6815 MemRegion span, 6816 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6817 CMSMarkStack* mark_stack): 6818 _collector(collector), 6819 _span(span), 6820 _verification_bm(verification_bm), 6821 _cms_bm(cms_bm), 6822 _mark_stack(mark_stack), 6823 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6824 mark_stack) 6825 { 6826 assert(_mark_stack->isEmpty(), "stack should be empty"); 6827 _finger = _verification_bm->startWord(); 6828 assert(_collector->_restart_addr == NULL, "Sanity check"); 6829 assert(_span.contains(_finger), "Out of bounds _finger?"); 6830 } 6831 6832 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6833 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6834 assert(_span.contains(addr), "Out of bounds _finger?"); 6835 _finger = addr; 6836 } 6837 6838 // Should revisit to see if this should be restructured for 6839 // greater efficiency. 6840 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6841 // convert offset into a HeapWord* 6842 HeapWord* addr = _verification_bm->startWord() + offset; 6843 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6844 "address out of range"); 6845 assert(_verification_bm->isMarked(addr), "tautology"); 6846 assert(_cms_bm->isMarked(addr), "tautology"); 6847 6848 assert(_mark_stack->isEmpty(), 6849 "should drain stack to limit stack usage"); 6850 // convert addr to an oop preparatory to scanning 6851 oop obj = oop(addr); 6852 assert(obj->is_oop(), "should be an oop"); 6853 assert(_finger <= addr, "_finger runneth ahead"); 6854 // advance the finger to right end of this object 6855 _finger = addr + obj->size(); 6856 assert(_finger > addr, "we just incremented it above"); 6857 // Note: the finger doesn't advance while we drain 6858 // the stack below. 6859 bool res = _mark_stack->push(obj); 6860 assert(res, "Empty non-zero size stack should have space for single push"); 6861 while (!_mark_stack->isEmpty()) { 6862 oop new_oop = _mark_stack->pop(); 6863 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6864 // now scan this oop's oops 6865 new_oop->oop_iterate(&_pam_verify_closure); 6866 } 6867 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6868 return true; 6869 } 6870 6871 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6872 CMSCollector* collector, MemRegion span, 6873 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6874 CMSMarkStack* mark_stack): 6875 MetadataAwareOopClosure(collector->ref_processor()), 6876 _collector(collector), 6877 _span(span), 6878 _verification_bm(verification_bm), 6879 _cms_bm(cms_bm), 6880 _mark_stack(mark_stack) 6881 { } 6882 6883 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6884 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6885 6886 // Upon stack overflow, we discard (part of) the stack, 6887 // remembering the least address amongst those discarded 6888 // in CMSCollector's _restart_address. 6889 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6890 // Remember the least grey address discarded 6891 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6892 _collector->lower_restart_addr(ra); 6893 _mark_stack->reset(); // discard stack contents 6894 _mark_stack->expand(); // expand the stack if possible 6895 } 6896 6897 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6898 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6899 HeapWord* addr = (HeapWord*)obj; 6900 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6901 // Oop lies in _span and isn't yet grey or black 6902 _verification_bm->mark(addr); // now grey 6903 if (!_cms_bm->isMarked(addr)) { 6904 oop(addr)->print(); 6905 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6906 p2i(addr)); 6907 fatal("... aborting"); 6908 } 6909 6910 if (!_mark_stack->push(obj)) { // stack overflow 6911 if (PrintCMSStatistics != 0) { 6912 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6913 SIZE_FORMAT, _mark_stack->capacity()); 6914 } 6915 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6916 handle_stack_overflow(addr); 6917 } 6918 // anything including and to the right of _finger 6919 // will be scanned as we iterate over the remainder of the 6920 // bit map 6921 } 6922 } 6923 6924 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6925 MemRegion span, 6926 CMSBitMap* bitMap, CMSMarkStack* markStack, 6927 HeapWord* finger, MarkFromRootsClosure* parent) : 6928 MetadataAwareOopClosure(collector->ref_processor()), 6929 _collector(collector), 6930 _span(span), 6931 _bitMap(bitMap), 6932 _markStack(markStack), 6933 _finger(finger), 6934 _parent(parent) 6935 { } 6936 6937 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6938 MemRegion span, 6939 CMSBitMap* bit_map, 6940 OopTaskQueue* work_queue, 6941 CMSMarkStack* overflow_stack, 6942 HeapWord* finger, 6943 HeapWord** global_finger_addr, 6944 Par_MarkFromRootsClosure* parent) : 6945 MetadataAwareOopClosure(collector->ref_processor()), 6946 _collector(collector), 6947 _whole_span(collector->_span), 6948 _span(span), 6949 _bit_map(bit_map), 6950 _work_queue(work_queue), 6951 _overflow_stack(overflow_stack), 6952 _finger(finger), 6953 _global_finger_addr(global_finger_addr), 6954 _parent(parent) 6955 { } 6956 6957 // Assumes thread-safe access by callers, who are 6958 // responsible for mutual exclusion. 6959 void CMSCollector::lower_restart_addr(HeapWord* low) { 6960 assert(_span.contains(low), "Out of bounds addr"); 6961 if (_restart_addr == NULL) { 6962 _restart_addr = low; 6963 } else { 6964 _restart_addr = MIN2(_restart_addr, low); 6965 } 6966 } 6967 6968 // Upon stack overflow, we discard (part of) the stack, 6969 // remembering the least address amongst those discarded 6970 // in CMSCollector's _restart_address. 6971 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6972 // Remember the least grey address discarded 6973 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6974 _collector->lower_restart_addr(ra); 6975 _markStack->reset(); // discard stack contents 6976 _markStack->expand(); // expand the stack if possible 6977 } 6978 6979 // Upon stack overflow, we discard (part of) the stack, 6980 // remembering the least address amongst those discarded 6981 // in CMSCollector's _restart_address. 6982 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6983 // We need to do this under a mutex to prevent other 6984 // workers from interfering with the work done below. 6985 MutexLockerEx ml(_overflow_stack->par_lock(), 6986 Mutex::_no_safepoint_check_flag); 6987 // Remember the least grey address discarded 6988 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6989 _collector->lower_restart_addr(ra); 6990 _overflow_stack->reset(); // discard stack contents 6991 _overflow_stack->expand(); // expand the stack if possible 6992 } 6993 6994 void PushOrMarkClosure::do_oop(oop obj) { 6995 // Ignore mark word because we are running concurrent with mutators. 6996 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6997 HeapWord* addr = (HeapWord*)obj; 6998 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6999 // Oop lies in _span and isn't yet grey or black 7000 _bitMap->mark(addr); // now grey 7001 if (addr < _finger) { 7002 // the bit map iteration has already either passed, or 7003 // sampled, this bit in the bit map; we'll need to 7004 // use the marking stack to scan this oop's oops. 7005 bool simulate_overflow = false; 7006 NOT_PRODUCT( 7007 if (CMSMarkStackOverflowALot && 7008 _collector->simulate_overflow()) { 7009 // simulate a stack overflow 7010 simulate_overflow = true; 7011 } 7012 ) 7013 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7014 if (PrintCMSStatistics != 0) { 7015 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7016 SIZE_FORMAT, _markStack->capacity()); 7017 } 7018 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7019 handle_stack_overflow(addr); 7020 } 7021 } 7022 // anything including and to the right of _finger 7023 // will be scanned as we iterate over the remainder of the 7024 // bit map 7025 do_yield_check(); 7026 } 7027 } 7028 7029 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7030 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7031 7032 void Par_PushOrMarkClosure::do_oop(oop obj) { 7033 // Ignore mark word because we are running concurrent with mutators. 7034 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7035 HeapWord* addr = (HeapWord*)obj; 7036 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7037 // Oop lies in _span and isn't yet grey or black 7038 // We read the global_finger (volatile read) strictly after marking oop 7039 bool res = _bit_map->par_mark(addr); // now grey 7040 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7041 // Should we push this marked oop on our stack? 7042 // -- if someone else marked it, nothing to do 7043 // -- if target oop is above global finger nothing to do 7044 // -- if target oop is in chunk and above local finger 7045 // then nothing to do 7046 // -- else push on work queue 7047 if ( !res // someone else marked it, they will deal with it 7048 || (addr >= *gfa) // will be scanned in a later task 7049 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7050 return; 7051 } 7052 // the bit map iteration has already either passed, or 7053 // sampled, this bit in the bit map; we'll need to 7054 // use the marking stack to scan this oop's oops. 7055 bool simulate_overflow = false; 7056 NOT_PRODUCT( 7057 if (CMSMarkStackOverflowALot && 7058 _collector->simulate_overflow()) { 7059 // simulate a stack overflow 7060 simulate_overflow = true; 7061 } 7062 ) 7063 if (simulate_overflow || 7064 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7065 // stack overflow 7066 if (PrintCMSStatistics != 0) { 7067 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7068 SIZE_FORMAT, _overflow_stack->capacity()); 7069 } 7070 // We cannot assert that the overflow stack is full because 7071 // it may have been emptied since. 7072 assert(simulate_overflow || 7073 _work_queue->size() == _work_queue->max_elems(), 7074 "Else push should have succeeded"); 7075 handle_stack_overflow(addr); 7076 } 7077 do_yield_check(); 7078 } 7079 } 7080 7081 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7082 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7083 7084 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7085 MemRegion span, 7086 ReferenceProcessor* rp, 7087 CMSBitMap* bit_map, 7088 CMSBitMap* mod_union_table, 7089 CMSMarkStack* mark_stack, 7090 bool concurrent_precleaning): 7091 MetadataAwareOopClosure(rp), 7092 _collector(collector), 7093 _span(span), 7094 _bit_map(bit_map), 7095 _mod_union_table(mod_union_table), 7096 _mark_stack(mark_stack), 7097 _concurrent_precleaning(concurrent_precleaning) 7098 { 7099 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7100 } 7101 7102 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7103 // the non-parallel version (the parallel version appears further below.) 7104 void PushAndMarkClosure::do_oop(oop obj) { 7105 // Ignore mark word verification. If during concurrent precleaning, 7106 // the object monitor may be locked. If during the checkpoint 7107 // phases, the object may already have been reached by a different 7108 // path and may be at the end of the global overflow list (so 7109 // the mark word may be NULL). 7110 assert(obj->is_oop_or_null(true /* ignore mark word */), 7111 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7112 HeapWord* addr = (HeapWord*)obj; 7113 // Check if oop points into the CMS generation 7114 // and is not marked 7115 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7116 // a white object ... 7117 _bit_map->mark(addr); // ... now grey 7118 // push on the marking stack (grey set) 7119 bool simulate_overflow = false; 7120 NOT_PRODUCT( 7121 if (CMSMarkStackOverflowALot && 7122 _collector->simulate_overflow()) { 7123 // simulate a stack overflow 7124 simulate_overflow = true; 7125 } 7126 ) 7127 if (simulate_overflow || !_mark_stack->push(obj)) { 7128 if (_concurrent_precleaning) { 7129 // During precleaning we can just dirty the appropriate card(s) 7130 // in the mod union table, thus ensuring that the object remains 7131 // in the grey set and continue. In the case of object arrays 7132 // we need to dirty all of the cards that the object spans, 7133 // since the rescan of object arrays will be limited to the 7134 // dirty cards. 7135 // Note that no one can be interfering with us in this action 7136 // of dirtying the mod union table, so no locking or atomics 7137 // are required. 7138 if (obj->is_objArray()) { 7139 size_t sz = obj->size(); 7140 HeapWord* end_card_addr = (HeapWord*)round_to( 7141 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7142 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7143 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7144 _mod_union_table->mark_range(redirty_range); 7145 } else { 7146 _mod_union_table->mark(addr); 7147 } 7148 _collector->_ser_pmc_preclean_ovflw++; 7149 } else { 7150 // During the remark phase, we need to remember this oop 7151 // in the overflow list. 7152 _collector->push_on_overflow_list(obj); 7153 _collector->_ser_pmc_remark_ovflw++; 7154 } 7155 } 7156 } 7157 } 7158 7159 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7160 MemRegion span, 7161 ReferenceProcessor* rp, 7162 CMSBitMap* bit_map, 7163 OopTaskQueue* work_queue): 7164 MetadataAwareOopClosure(rp), 7165 _collector(collector), 7166 _span(span), 7167 _bit_map(bit_map), 7168 _work_queue(work_queue) 7169 { 7170 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7171 } 7172 7173 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7174 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7175 7176 // Grey object rescan during second checkpoint phase -- 7177 // the parallel version. 7178 void Par_PushAndMarkClosure::do_oop(oop obj) { 7179 // In the assert below, we ignore the mark word because 7180 // this oop may point to an already visited object that is 7181 // on the overflow stack (in which case the mark word has 7182 // been hijacked for chaining into the overflow stack -- 7183 // if this is the last object in the overflow stack then 7184 // its mark word will be NULL). Because this object may 7185 // have been subsequently popped off the global overflow 7186 // stack, and the mark word possibly restored to the prototypical 7187 // value, by the time we get to examined this failing assert in 7188 // the debugger, is_oop_or_null(false) may subsequently start 7189 // to hold. 7190 assert(obj->is_oop_or_null(true), 7191 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7192 HeapWord* addr = (HeapWord*)obj; 7193 // Check if oop points into the CMS generation 7194 // and is not marked 7195 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7196 // a white object ... 7197 // If we manage to "claim" the object, by being the 7198 // first thread to mark it, then we push it on our 7199 // marking stack 7200 if (_bit_map->par_mark(addr)) { // ... now grey 7201 // push on work queue (grey set) 7202 bool simulate_overflow = false; 7203 NOT_PRODUCT( 7204 if (CMSMarkStackOverflowALot && 7205 _collector->par_simulate_overflow()) { 7206 // simulate a stack overflow 7207 simulate_overflow = true; 7208 } 7209 ) 7210 if (simulate_overflow || !_work_queue->push(obj)) { 7211 _collector->par_push_on_overflow_list(obj); 7212 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7213 } 7214 } // Else, some other thread got there first 7215 } 7216 } 7217 7218 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7219 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7220 7221 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7222 Mutex* bml = _collector->bitMapLock(); 7223 assert_lock_strong(bml); 7224 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7225 "CMS thread should hold CMS token"); 7226 7227 bml->unlock(); 7228 ConcurrentMarkSweepThread::desynchronize(true); 7229 7230 _collector->stopTimer(); 7231 if (PrintCMSStatistics != 0) { 7232 _collector->incrementYields(); 7233 } 7234 7235 // See the comment in coordinator_yield() 7236 for (unsigned i = 0; i < CMSYieldSleepCount && 7237 ConcurrentMarkSweepThread::should_yield() && 7238 !CMSCollector::foregroundGCIsActive(); ++i) { 7239 os::sleep(Thread::current(), 1, false); 7240 } 7241 7242 ConcurrentMarkSweepThread::synchronize(true); 7243 bml->lock(); 7244 7245 _collector->startTimer(); 7246 } 7247 7248 bool CMSPrecleanRefsYieldClosure::should_return() { 7249 if (ConcurrentMarkSweepThread::should_yield()) { 7250 do_yield_work(); 7251 } 7252 return _collector->foregroundGCIsActive(); 7253 } 7254 7255 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7256 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7257 "mr should be aligned to start at a card boundary"); 7258 // We'd like to assert: 7259 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7260 // "mr should be a range of cards"); 7261 // However, that would be too strong in one case -- the last 7262 // partition ends at _unallocated_block which, in general, can be 7263 // an arbitrary boundary, not necessarily card aligned. 7264 if (PrintCMSStatistics != 0) { 7265 _num_dirty_cards += 7266 mr.word_size()/CardTableModRefBS::card_size_in_words; 7267 } 7268 _space->object_iterate_mem(mr, &_scan_cl); 7269 } 7270 7271 SweepClosure::SweepClosure(CMSCollector* collector, 7272 ConcurrentMarkSweepGeneration* g, 7273 CMSBitMap* bitMap, bool should_yield) : 7274 _collector(collector), 7275 _g(g), 7276 _sp(g->cmsSpace()), 7277 _limit(_sp->sweep_limit()), 7278 _freelistLock(_sp->freelistLock()), 7279 _bitMap(bitMap), 7280 _yield(should_yield), 7281 _inFreeRange(false), // No free range at beginning of sweep 7282 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7283 _lastFreeRangeCoalesced(false), 7284 _freeFinger(g->used_region().start()) 7285 { 7286 NOT_PRODUCT( 7287 _numObjectsFreed = 0; 7288 _numWordsFreed = 0; 7289 _numObjectsLive = 0; 7290 _numWordsLive = 0; 7291 _numObjectsAlreadyFree = 0; 7292 _numWordsAlreadyFree = 0; 7293 _last_fc = NULL; 7294 7295 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7296 _sp->dictionary()->initialize_dict_returned_bytes(); 7297 ) 7298 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7299 "sweep _limit out of bounds"); 7300 if (CMSTraceSweeper) { 7301 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7302 p2i(_limit)); 7303 } 7304 } 7305 7306 void SweepClosure::print_on(outputStream* st) const { 7307 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7308 p2i(_sp->bottom()), p2i(_sp->end())); 7309 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7310 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7311 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7312 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7313 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7314 } 7315 7316 #ifndef PRODUCT 7317 // Assertion checking only: no useful work in product mode -- 7318 // however, if any of the flags below become product flags, 7319 // you may need to review this code to see if it needs to be 7320 // enabled in product mode. 7321 SweepClosure::~SweepClosure() { 7322 assert_lock_strong(_freelistLock); 7323 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7324 "sweep _limit out of bounds"); 7325 if (inFreeRange()) { 7326 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7327 print(); 7328 ShouldNotReachHere(); 7329 } 7330 if (Verbose && PrintGC) { 7331 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7332 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7333 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7334 SIZE_FORMAT" bytes " 7335 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7336 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7337 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7338 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7339 * sizeof(HeapWord); 7340 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7341 7342 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7343 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7344 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7345 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7346 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7347 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7348 indexListReturnedBytes); 7349 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7350 dict_returned_bytes); 7351 } 7352 } 7353 if (CMSTraceSweeper) { 7354 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7355 p2i(_limit)); 7356 } 7357 } 7358 #endif // PRODUCT 7359 7360 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7361 bool freeRangeInFreeLists) { 7362 if (CMSTraceSweeper) { 7363 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7364 p2i(freeFinger), freeRangeInFreeLists); 7365 } 7366 assert(!inFreeRange(), "Trampling existing free range"); 7367 set_inFreeRange(true); 7368 set_lastFreeRangeCoalesced(false); 7369 7370 set_freeFinger(freeFinger); 7371 set_freeRangeInFreeLists(freeRangeInFreeLists); 7372 if (CMSTestInFreeList) { 7373 if (freeRangeInFreeLists) { 7374 FreeChunk* fc = (FreeChunk*) freeFinger; 7375 assert(fc->is_free(), "A chunk on the free list should be free."); 7376 assert(fc->size() > 0, "Free range should have a size"); 7377 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7378 } 7379 } 7380 } 7381 7382 // Note that the sweeper runs concurrently with mutators. Thus, 7383 // it is possible for direct allocation in this generation to happen 7384 // in the middle of the sweep. Note that the sweeper also coalesces 7385 // contiguous free blocks. Thus, unless the sweeper and the allocator 7386 // synchronize appropriately freshly allocated blocks may get swept up. 7387 // This is accomplished by the sweeper locking the free lists while 7388 // it is sweeping. Thus blocks that are determined to be free are 7389 // indeed free. There is however one additional complication: 7390 // blocks that have been allocated since the final checkpoint and 7391 // mark, will not have been marked and so would be treated as 7392 // unreachable and swept up. To prevent this, the allocator marks 7393 // the bit map when allocating during the sweep phase. This leads, 7394 // however, to a further complication -- objects may have been allocated 7395 // but not yet initialized -- in the sense that the header isn't yet 7396 // installed. The sweeper can not then determine the size of the block 7397 // in order to skip over it. To deal with this case, we use a technique 7398 // (due to Printezis) to encode such uninitialized block sizes in the 7399 // bit map. Since the bit map uses a bit per every HeapWord, but the 7400 // CMS generation has a minimum object size of 3 HeapWords, it follows 7401 // that "normal marks" won't be adjacent in the bit map (there will 7402 // always be at least two 0 bits between successive 1 bits). We make use 7403 // of these "unused" bits to represent uninitialized blocks -- the bit 7404 // corresponding to the start of the uninitialized object and the next 7405 // bit are both set. Finally, a 1 bit marks the end of the object that 7406 // started with the two consecutive 1 bits to indicate its potentially 7407 // uninitialized state. 7408 7409 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7410 FreeChunk* fc = (FreeChunk*)addr; 7411 size_t res; 7412 7413 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7414 // than "addr == _limit" because although _limit was a block boundary when 7415 // we started the sweep, it may no longer be one because heap expansion 7416 // may have caused us to coalesce the block ending at the address _limit 7417 // with a newly expanded chunk (this happens when _limit was set to the 7418 // previous _end of the space), so we may have stepped past _limit: 7419 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7420 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7421 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7422 "sweep _limit out of bounds"); 7423 assert(addr < _sp->end(), "addr out of bounds"); 7424 // Flush any free range we might be holding as a single 7425 // coalesced chunk to the appropriate free list. 7426 if (inFreeRange()) { 7427 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7428 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7429 flush_cur_free_chunk(freeFinger(), 7430 pointer_delta(addr, freeFinger())); 7431 if (CMSTraceSweeper) { 7432 gclog_or_tty->print("Sweep: last chunk: "); 7433 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7434 "[coalesced:%d]\n", 7435 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7436 lastFreeRangeCoalesced() ? 1 : 0); 7437 } 7438 } 7439 7440 // help the iterator loop finish 7441 return pointer_delta(_sp->end(), addr); 7442 } 7443 7444 assert(addr < _limit, "sweep invariant"); 7445 // check if we should yield 7446 do_yield_check(addr); 7447 if (fc->is_free()) { 7448 // Chunk that is already free 7449 res = fc->size(); 7450 do_already_free_chunk(fc); 7451 debug_only(_sp->verifyFreeLists()); 7452 // If we flush the chunk at hand in lookahead_and_flush() 7453 // and it's coalesced with a preceding chunk, then the 7454 // process of "mangling" the payload of the coalesced block 7455 // will cause erasure of the size information from the 7456 // (erstwhile) header of all the coalesced blocks but the 7457 // first, so the first disjunct in the assert will not hold 7458 // in that specific case (in which case the second disjunct 7459 // will hold). 7460 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7461 "Otherwise the size info doesn't change at this step"); 7462 NOT_PRODUCT( 7463 _numObjectsAlreadyFree++; 7464 _numWordsAlreadyFree += res; 7465 ) 7466 NOT_PRODUCT(_last_fc = fc;) 7467 } else if (!_bitMap->isMarked(addr)) { 7468 // Chunk is fresh garbage 7469 res = do_garbage_chunk(fc); 7470 debug_only(_sp->verifyFreeLists()); 7471 NOT_PRODUCT( 7472 _numObjectsFreed++; 7473 _numWordsFreed += res; 7474 ) 7475 } else { 7476 // Chunk that is alive. 7477 res = do_live_chunk(fc); 7478 debug_only(_sp->verifyFreeLists()); 7479 NOT_PRODUCT( 7480 _numObjectsLive++; 7481 _numWordsLive += res; 7482 ) 7483 } 7484 return res; 7485 } 7486 7487 // For the smart allocation, record following 7488 // split deaths - a free chunk is removed from its free list because 7489 // it is being split into two or more chunks. 7490 // split birth - a free chunk is being added to its free list because 7491 // a larger free chunk has been split and resulted in this free chunk. 7492 // coal death - a free chunk is being removed from its free list because 7493 // it is being coalesced into a large free chunk. 7494 // coal birth - a free chunk is being added to its free list because 7495 // it was created when two or more free chunks where coalesced into 7496 // this free chunk. 7497 // 7498 // These statistics are used to determine the desired number of free 7499 // chunks of a given size. The desired number is chosen to be relative 7500 // to the end of a CMS sweep. The desired number at the end of a sweep 7501 // is the 7502 // count-at-end-of-previous-sweep (an amount that was enough) 7503 // - count-at-beginning-of-current-sweep (the excess) 7504 // + split-births (gains in this size during interval) 7505 // - split-deaths (demands on this size during interval) 7506 // where the interval is from the end of one sweep to the end of the 7507 // next. 7508 // 7509 // When sweeping the sweeper maintains an accumulated chunk which is 7510 // the chunk that is made up of chunks that have been coalesced. That 7511 // will be termed the left-hand chunk. A new chunk of garbage that 7512 // is being considered for coalescing will be referred to as the 7513 // right-hand chunk. 7514 // 7515 // When making a decision on whether to coalesce a right-hand chunk with 7516 // the current left-hand chunk, the current count vs. the desired count 7517 // of the left-hand chunk is considered. Also if the right-hand chunk 7518 // is near the large chunk at the end of the heap (see 7519 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7520 // left-hand chunk is coalesced. 7521 // 7522 // When making a decision about whether to split a chunk, the desired count 7523 // vs. the current count of the candidate to be split is also considered. 7524 // If the candidate is underpopulated (currently fewer chunks than desired) 7525 // a chunk of an overpopulated (currently more chunks than desired) size may 7526 // be chosen. The "hint" associated with a free list, if non-null, points 7527 // to a free list which may be overpopulated. 7528 // 7529 7530 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7531 const size_t size = fc->size(); 7532 // Chunks that cannot be coalesced are not in the 7533 // free lists. 7534 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7535 assert(_sp->verify_chunk_in_free_list(fc), 7536 "free chunk should be in free lists"); 7537 } 7538 // a chunk that is already free, should not have been 7539 // marked in the bit map 7540 HeapWord* const addr = (HeapWord*) fc; 7541 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7542 // Verify that the bit map has no bits marked between 7543 // addr and purported end of this block. 7544 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7545 7546 // Some chunks cannot be coalesced under any circumstances. 7547 // See the definition of cantCoalesce(). 7548 if (!fc->cantCoalesce()) { 7549 // This chunk can potentially be coalesced. 7550 if (_sp->adaptive_freelists()) { 7551 // All the work is done in 7552 do_post_free_or_garbage_chunk(fc, size); 7553 } else { // Not adaptive free lists 7554 // this is a free chunk that can potentially be coalesced by the sweeper; 7555 if (!inFreeRange()) { 7556 // if the next chunk is a free block that can't be coalesced 7557 // it doesn't make sense to remove this chunk from the free lists 7558 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7559 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7560 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7561 nextChunk->is_free() && // ... which is free... 7562 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7563 // nothing to do 7564 } else { 7565 // Potentially the start of a new free range: 7566 // Don't eagerly remove it from the free lists. 7567 // No need to remove it if it will just be put 7568 // back again. (Also from a pragmatic point of view 7569 // if it is a free block in a region that is beyond 7570 // any allocated blocks, an assertion will fail) 7571 // Remember the start of a free run. 7572 initialize_free_range(addr, true); 7573 // end - can coalesce with next chunk 7574 } 7575 } else { 7576 // the midst of a free range, we are coalescing 7577 print_free_block_coalesced(fc); 7578 if (CMSTraceSweeper) { 7579 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7580 } 7581 // remove it from the free lists 7582 _sp->removeFreeChunkFromFreeLists(fc); 7583 set_lastFreeRangeCoalesced(true); 7584 // If the chunk is being coalesced and the current free range is 7585 // in the free lists, remove the current free range so that it 7586 // will be returned to the free lists in its entirety - all 7587 // the coalesced pieces included. 7588 if (freeRangeInFreeLists()) { 7589 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7590 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7591 "Size of free range is inconsistent with chunk size."); 7592 if (CMSTestInFreeList) { 7593 assert(_sp->verify_chunk_in_free_list(ffc), 7594 "free range is not in free lists"); 7595 } 7596 _sp->removeFreeChunkFromFreeLists(ffc); 7597 set_freeRangeInFreeLists(false); 7598 } 7599 } 7600 } 7601 // Note that if the chunk is not coalescable (the else arm 7602 // below), we unconditionally flush, without needing to do 7603 // a "lookahead," as we do below. 7604 if (inFreeRange()) lookahead_and_flush(fc, size); 7605 } else { 7606 // Code path common to both original and adaptive free lists. 7607 7608 // cant coalesce with previous block; this should be treated 7609 // as the end of a free run if any 7610 if (inFreeRange()) { 7611 // we kicked some butt; time to pick up the garbage 7612 assert(freeFinger() < addr, "freeFinger points too high"); 7613 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7614 } 7615 // else, nothing to do, just continue 7616 } 7617 } 7618 7619 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7620 // This is a chunk of garbage. It is not in any free list. 7621 // Add it to a free list or let it possibly be coalesced into 7622 // a larger chunk. 7623 HeapWord* const addr = (HeapWord*) fc; 7624 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7625 7626 if (_sp->adaptive_freelists()) { 7627 // Verify that the bit map has no bits marked between 7628 // addr and purported end of just dead object. 7629 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7630 7631 do_post_free_or_garbage_chunk(fc, size); 7632 } else { 7633 if (!inFreeRange()) { 7634 // start of a new free range 7635 assert(size > 0, "A free range should have a size"); 7636 initialize_free_range(addr, false); 7637 } else { 7638 // this will be swept up when we hit the end of the 7639 // free range 7640 if (CMSTraceSweeper) { 7641 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7642 } 7643 // If the chunk is being coalesced and the current free range is 7644 // in the free lists, remove the current free range so that it 7645 // will be returned to the free lists in its entirety - all 7646 // the coalesced pieces included. 7647 if (freeRangeInFreeLists()) { 7648 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7649 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7650 "Size of free range is inconsistent with chunk size."); 7651 if (CMSTestInFreeList) { 7652 assert(_sp->verify_chunk_in_free_list(ffc), 7653 "free range is not in free lists"); 7654 } 7655 _sp->removeFreeChunkFromFreeLists(ffc); 7656 set_freeRangeInFreeLists(false); 7657 } 7658 set_lastFreeRangeCoalesced(true); 7659 } 7660 // this will be swept up when we hit the end of the free range 7661 7662 // Verify that the bit map has no bits marked between 7663 // addr and purported end of just dead object. 7664 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7665 } 7666 assert(_limit >= addr + size, 7667 "A freshly garbage chunk can't possibly straddle over _limit"); 7668 if (inFreeRange()) lookahead_and_flush(fc, size); 7669 return size; 7670 } 7671 7672 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7673 HeapWord* addr = (HeapWord*) fc; 7674 // The sweeper has just found a live object. Return any accumulated 7675 // left hand chunk to the free lists. 7676 if (inFreeRange()) { 7677 assert(freeFinger() < addr, "freeFinger points too high"); 7678 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7679 } 7680 7681 // This object is live: we'd normally expect this to be 7682 // an oop, and like to assert the following: 7683 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7684 // However, as we commented above, this may be an object whose 7685 // header hasn't yet been initialized. 7686 size_t size; 7687 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7688 if (_bitMap->isMarked(addr + 1)) { 7689 // Determine the size from the bit map, rather than trying to 7690 // compute it from the object header. 7691 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7692 size = pointer_delta(nextOneAddr + 1, addr); 7693 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7694 "alignment problem"); 7695 7696 #ifdef ASSERT 7697 if (oop(addr)->klass_or_null() != NULL) { 7698 // Ignore mark word because we are running concurrent with mutators 7699 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7700 assert(size == 7701 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7702 "P-mark and computed size do not agree"); 7703 } 7704 #endif 7705 7706 } else { 7707 // This should be an initialized object that's alive. 7708 assert(oop(addr)->klass_or_null() != NULL, 7709 "Should be an initialized object"); 7710 // Ignore mark word because we are running concurrent with mutators 7711 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7712 // Verify that the bit map has no bits marked between 7713 // addr and purported end of this block. 7714 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7715 assert(size >= 3, "Necessary for Printezis marks to work"); 7716 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7717 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7718 } 7719 return size; 7720 } 7721 7722 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7723 size_t chunkSize) { 7724 // do_post_free_or_garbage_chunk() should only be called in the case 7725 // of the adaptive free list allocator. 7726 const bool fcInFreeLists = fc->is_free(); 7727 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7728 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7729 if (CMSTestInFreeList && fcInFreeLists) { 7730 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7731 } 7732 7733 if (CMSTraceSweeper) { 7734 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7735 } 7736 7737 HeapWord* const fc_addr = (HeapWord*) fc; 7738 7739 bool coalesce; 7740 const size_t left = pointer_delta(fc_addr, freeFinger()); 7741 const size_t right = chunkSize; 7742 switch (FLSCoalescePolicy) { 7743 // numeric value forms a coalition aggressiveness metric 7744 case 0: { // never coalesce 7745 coalesce = false; 7746 break; 7747 } 7748 case 1: { // coalesce if left & right chunks on overpopulated lists 7749 coalesce = _sp->coalOverPopulated(left) && 7750 _sp->coalOverPopulated(right); 7751 break; 7752 } 7753 case 2: { // coalesce if left chunk on overpopulated list (default) 7754 coalesce = _sp->coalOverPopulated(left); 7755 break; 7756 } 7757 case 3: { // coalesce if left OR right chunk on overpopulated list 7758 coalesce = _sp->coalOverPopulated(left) || 7759 _sp->coalOverPopulated(right); 7760 break; 7761 } 7762 case 4: { // always coalesce 7763 coalesce = true; 7764 break; 7765 } 7766 default: 7767 ShouldNotReachHere(); 7768 } 7769 7770 // Should the current free range be coalesced? 7771 // If the chunk is in a free range and either we decided to coalesce above 7772 // or the chunk is near the large block at the end of the heap 7773 // (isNearLargestChunk() returns true), then coalesce this chunk. 7774 const bool doCoalesce = inFreeRange() 7775 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7776 if (doCoalesce) { 7777 // Coalesce the current free range on the left with the new 7778 // chunk on the right. If either is on a free list, 7779 // it must be removed from the list and stashed in the closure. 7780 if (freeRangeInFreeLists()) { 7781 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7782 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7783 "Size of free range is inconsistent with chunk size."); 7784 if (CMSTestInFreeList) { 7785 assert(_sp->verify_chunk_in_free_list(ffc), 7786 "Chunk is not in free lists"); 7787 } 7788 _sp->coalDeath(ffc->size()); 7789 _sp->removeFreeChunkFromFreeLists(ffc); 7790 set_freeRangeInFreeLists(false); 7791 } 7792 if (fcInFreeLists) { 7793 _sp->coalDeath(chunkSize); 7794 assert(fc->size() == chunkSize, 7795 "The chunk has the wrong size or is not in the free lists"); 7796 _sp->removeFreeChunkFromFreeLists(fc); 7797 } 7798 set_lastFreeRangeCoalesced(true); 7799 print_free_block_coalesced(fc); 7800 } else { // not in a free range and/or should not coalesce 7801 // Return the current free range and start a new one. 7802 if (inFreeRange()) { 7803 // In a free range but cannot coalesce with the right hand chunk. 7804 // Put the current free range into the free lists. 7805 flush_cur_free_chunk(freeFinger(), 7806 pointer_delta(fc_addr, freeFinger())); 7807 } 7808 // Set up for new free range. Pass along whether the right hand 7809 // chunk is in the free lists. 7810 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7811 } 7812 } 7813 7814 // Lookahead flush: 7815 // If we are tracking a free range, and this is the last chunk that 7816 // we'll look at because its end crosses past _limit, we'll preemptively 7817 // flush it along with any free range we may be holding on to. Note that 7818 // this can be the case only for an already free or freshly garbage 7819 // chunk. If this block is an object, it can never straddle 7820 // over _limit. The "straddling" occurs when _limit is set at 7821 // the previous end of the space when this cycle started, and 7822 // a subsequent heap expansion caused the previously co-terminal 7823 // free block to be coalesced with the newly expanded portion, 7824 // thus rendering _limit a non-block-boundary making it dangerous 7825 // for the sweeper to step over and examine. 7826 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7827 assert(inFreeRange(), "Should only be called if currently in a free range."); 7828 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7829 assert(_sp->used_region().contains(eob - 1), 7830 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7831 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7832 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7833 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7834 if (eob >= _limit) { 7835 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7836 if (CMSTraceSweeper) { 7837 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7838 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7839 "[" PTR_FORMAT "," PTR_FORMAT ")", 7840 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7841 } 7842 // Return the storage we are tracking back into the free lists. 7843 if (CMSTraceSweeper) { 7844 gclog_or_tty->print_cr("Flushing ... "); 7845 } 7846 assert(freeFinger() < eob, "Error"); 7847 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7848 } 7849 } 7850 7851 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7852 assert(inFreeRange(), "Should only be called if currently in a free range."); 7853 assert(size > 0, 7854 "A zero sized chunk cannot be added to the free lists."); 7855 if (!freeRangeInFreeLists()) { 7856 if (CMSTestInFreeList) { 7857 FreeChunk* fc = (FreeChunk*) chunk; 7858 fc->set_size(size); 7859 assert(!_sp->verify_chunk_in_free_list(fc), 7860 "chunk should not be in free lists yet"); 7861 } 7862 if (CMSTraceSweeper) { 7863 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7864 p2i(chunk), size); 7865 } 7866 // A new free range is going to be starting. The current 7867 // free range has not been added to the free lists yet or 7868 // was removed so add it back. 7869 // If the current free range was coalesced, then the death 7870 // of the free range was recorded. Record a birth now. 7871 if (lastFreeRangeCoalesced()) { 7872 _sp->coalBirth(size); 7873 } 7874 _sp->addChunkAndRepairOffsetTable(chunk, size, 7875 lastFreeRangeCoalesced()); 7876 } else if (CMSTraceSweeper) { 7877 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7878 } 7879 set_inFreeRange(false); 7880 set_freeRangeInFreeLists(false); 7881 } 7882 7883 // We take a break if we've been at this for a while, 7884 // so as to avoid monopolizing the locks involved. 7885 void SweepClosure::do_yield_work(HeapWord* addr) { 7886 // Return current free chunk being used for coalescing (if any) 7887 // to the appropriate freelist. After yielding, the next 7888 // free block encountered will start a coalescing range of 7889 // free blocks. If the next free block is adjacent to the 7890 // chunk just flushed, they will need to wait for the next 7891 // sweep to be coalesced. 7892 if (inFreeRange()) { 7893 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7894 } 7895 7896 // First give up the locks, then yield, then re-lock. 7897 // We should probably use a constructor/destructor idiom to 7898 // do this unlock/lock or modify the MutexUnlocker class to 7899 // serve our purpose. XXX 7900 assert_lock_strong(_bitMap->lock()); 7901 assert_lock_strong(_freelistLock); 7902 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7903 "CMS thread should hold CMS token"); 7904 _bitMap->lock()->unlock(); 7905 _freelistLock->unlock(); 7906 ConcurrentMarkSweepThread::desynchronize(true); 7907 _collector->stopTimer(); 7908 if (PrintCMSStatistics != 0) { 7909 _collector->incrementYields(); 7910 } 7911 7912 // See the comment in coordinator_yield() 7913 for (unsigned i = 0; i < CMSYieldSleepCount && 7914 ConcurrentMarkSweepThread::should_yield() && 7915 !CMSCollector::foregroundGCIsActive(); ++i) { 7916 os::sleep(Thread::current(), 1, false); 7917 } 7918 7919 ConcurrentMarkSweepThread::synchronize(true); 7920 _freelistLock->lock(); 7921 _bitMap->lock()->lock_without_safepoint_check(); 7922 _collector->startTimer(); 7923 } 7924 7925 #ifndef PRODUCT 7926 // This is actually very useful in a product build if it can 7927 // be called from the debugger. Compile it into the product 7928 // as needed. 7929 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7930 return debug_cms_space->verify_chunk_in_free_list(fc); 7931 } 7932 #endif 7933 7934 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7935 if (CMSTraceSweeper) { 7936 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7937 p2i(fc), fc->size()); 7938 } 7939 } 7940 7941 // CMSIsAliveClosure 7942 bool CMSIsAliveClosure::do_object_b(oop obj) { 7943 HeapWord* addr = (HeapWord*)obj; 7944 return addr != NULL && 7945 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7946 } 7947 7948 7949 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7950 MemRegion span, 7951 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7952 bool cpc): 7953 _collector(collector), 7954 _span(span), 7955 _bit_map(bit_map), 7956 _mark_stack(mark_stack), 7957 _concurrent_precleaning(cpc) { 7958 assert(!_span.is_empty(), "Empty span could spell trouble"); 7959 } 7960 7961 7962 // CMSKeepAliveClosure: the serial version 7963 void CMSKeepAliveClosure::do_oop(oop obj) { 7964 HeapWord* addr = (HeapWord*)obj; 7965 if (_span.contains(addr) && 7966 !_bit_map->isMarked(addr)) { 7967 _bit_map->mark(addr); 7968 bool simulate_overflow = false; 7969 NOT_PRODUCT( 7970 if (CMSMarkStackOverflowALot && 7971 _collector->simulate_overflow()) { 7972 // simulate a stack overflow 7973 simulate_overflow = true; 7974 } 7975 ) 7976 if (simulate_overflow || !_mark_stack->push(obj)) { 7977 if (_concurrent_precleaning) { 7978 // We dirty the overflown object and let the remark 7979 // phase deal with it. 7980 assert(_collector->overflow_list_is_empty(), "Error"); 7981 // In the case of object arrays, we need to dirty all of 7982 // the cards that the object spans. No locking or atomics 7983 // are needed since no one else can be mutating the mod union 7984 // table. 7985 if (obj->is_objArray()) { 7986 size_t sz = obj->size(); 7987 HeapWord* end_card_addr = 7988 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7989 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7990 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7991 _collector->_modUnionTable.mark_range(redirty_range); 7992 } else { 7993 _collector->_modUnionTable.mark(addr); 7994 } 7995 _collector->_ser_kac_preclean_ovflw++; 7996 } else { 7997 _collector->push_on_overflow_list(obj); 7998 _collector->_ser_kac_ovflw++; 7999 } 8000 } 8001 } 8002 } 8003 8004 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8005 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8006 8007 // CMSParKeepAliveClosure: a parallel version of the above. 8008 // The work queues are private to each closure (thread), 8009 // but (may be) available for stealing by other threads. 8010 void CMSParKeepAliveClosure::do_oop(oop obj) { 8011 HeapWord* addr = (HeapWord*)obj; 8012 if (_span.contains(addr) && 8013 !_bit_map->isMarked(addr)) { 8014 // In general, during recursive tracing, several threads 8015 // may be concurrently getting here; the first one to 8016 // "tag" it, claims it. 8017 if (_bit_map->par_mark(addr)) { 8018 bool res = _work_queue->push(obj); 8019 assert(res, "Low water mark should be much less than capacity"); 8020 // Do a recursive trim in the hope that this will keep 8021 // stack usage lower, but leave some oops for potential stealers 8022 trim_queue(_low_water_mark); 8023 } // Else, another thread got there first 8024 } 8025 } 8026 8027 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8028 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8029 8030 void CMSParKeepAliveClosure::trim_queue(uint max) { 8031 while (_work_queue->size() > max) { 8032 oop new_oop; 8033 if (_work_queue->pop_local(new_oop)) { 8034 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8035 assert(_bit_map->isMarked((HeapWord*)new_oop), 8036 "no white objects on this stack!"); 8037 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8038 // iterate over the oops in this oop, marking and pushing 8039 // the ones in CMS heap (i.e. in _span). 8040 new_oop->oop_iterate(&_mark_and_push); 8041 } 8042 } 8043 } 8044 8045 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8046 CMSCollector* collector, 8047 MemRegion span, CMSBitMap* bit_map, 8048 OopTaskQueue* work_queue): 8049 _collector(collector), 8050 _span(span), 8051 _bit_map(bit_map), 8052 _work_queue(work_queue) { } 8053 8054 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8055 HeapWord* addr = (HeapWord*)obj; 8056 if (_span.contains(addr) && 8057 !_bit_map->isMarked(addr)) { 8058 if (_bit_map->par_mark(addr)) { 8059 bool simulate_overflow = false; 8060 NOT_PRODUCT( 8061 if (CMSMarkStackOverflowALot && 8062 _collector->par_simulate_overflow()) { 8063 // simulate a stack overflow 8064 simulate_overflow = true; 8065 } 8066 ) 8067 if (simulate_overflow || !_work_queue->push(obj)) { 8068 _collector->par_push_on_overflow_list(obj); 8069 _collector->_par_kac_ovflw++; 8070 } 8071 } // Else another thread got there already 8072 } 8073 } 8074 8075 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8076 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8077 8078 ////////////////////////////////////////////////////////////////// 8079 // CMSExpansionCause ///////////////////////////// 8080 ////////////////////////////////////////////////////////////////// 8081 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8082 switch (cause) { 8083 case _no_expansion: 8084 return "No expansion"; 8085 case _satisfy_free_ratio: 8086 return "Free ratio"; 8087 case _satisfy_promotion: 8088 return "Satisfy promotion"; 8089 case _satisfy_allocation: 8090 return "allocation"; 8091 case _allocate_par_lab: 8092 return "Par LAB"; 8093 case _allocate_par_spooling_space: 8094 return "Par Spooling Space"; 8095 case _adaptive_size_policy: 8096 return "Ergonomics"; 8097 default: 8098 return "unknown"; 8099 } 8100 } 8101 8102 void CMSDrainMarkingStackClosure::do_void() { 8103 // the max number to take from overflow list at a time 8104 const size_t num = _mark_stack->capacity()/4; 8105 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8106 "Overflow list should be NULL during concurrent phases"); 8107 while (!_mark_stack->isEmpty() || 8108 // if stack is empty, check the overflow list 8109 _collector->take_from_overflow_list(num, _mark_stack)) { 8110 oop obj = _mark_stack->pop(); 8111 HeapWord* addr = (HeapWord*)obj; 8112 assert(_span.contains(addr), "Should be within span"); 8113 assert(_bit_map->isMarked(addr), "Should be marked"); 8114 assert(obj->is_oop(), "Should be an oop"); 8115 obj->oop_iterate(_keep_alive); 8116 } 8117 } 8118 8119 void CMSParDrainMarkingStackClosure::do_void() { 8120 // drain queue 8121 trim_queue(0); 8122 } 8123 8124 // Trim our work_queue so its length is below max at return 8125 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8126 while (_work_queue->size() > max) { 8127 oop new_oop; 8128 if (_work_queue->pop_local(new_oop)) { 8129 assert(new_oop->is_oop(), "Expected an oop"); 8130 assert(_bit_map->isMarked((HeapWord*)new_oop), 8131 "no white objects on this stack!"); 8132 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8133 // iterate over the oops in this oop, marking and pushing 8134 // the ones in CMS heap (i.e. in _span). 8135 new_oop->oop_iterate(&_mark_and_push); 8136 } 8137 } 8138 } 8139 8140 //////////////////////////////////////////////////////////////////// 8141 // Support for Marking Stack Overflow list handling and related code 8142 //////////////////////////////////////////////////////////////////// 8143 // Much of the following code is similar in shape and spirit to the 8144 // code used in ParNewGC. We should try and share that code 8145 // as much as possible in the future. 8146 8147 #ifndef PRODUCT 8148 // Debugging support for CMSStackOverflowALot 8149 8150 // It's OK to call this multi-threaded; the worst thing 8151 // that can happen is that we'll get a bunch of closely 8152 // spaced simulated overflows, but that's OK, in fact 8153 // probably good as it would exercise the overflow code 8154 // under contention. 8155 bool CMSCollector::simulate_overflow() { 8156 if (_overflow_counter-- <= 0) { // just being defensive 8157 _overflow_counter = CMSMarkStackOverflowInterval; 8158 return true; 8159 } else { 8160 return false; 8161 } 8162 } 8163 8164 bool CMSCollector::par_simulate_overflow() { 8165 return simulate_overflow(); 8166 } 8167 #endif 8168 8169 // Single-threaded 8170 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8171 assert(stack->isEmpty(), "Expected precondition"); 8172 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8173 size_t i = num; 8174 oop cur = _overflow_list; 8175 const markOop proto = markOopDesc::prototype(); 8176 NOT_PRODUCT(ssize_t n = 0;) 8177 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8178 next = oop(cur->mark()); 8179 cur->set_mark(proto); // until proven otherwise 8180 assert(cur->is_oop(), "Should be an oop"); 8181 bool res = stack->push(cur); 8182 assert(res, "Bit off more than can chew?"); 8183 NOT_PRODUCT(n++;) 8184 } 8185 _overflow_list = cur; 8186 #ifndef PRODUCT 8187 assert(_num_par_pushes >= n, "Too many pops?"); 8188 _num_par_pushes -=n; 8189 #endif 8190 return !stack->isEmpty(); 8191 } 8192 8193 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8194 // (MT-safe) Get a prefix of at most "num" from the list. 8195 // The overflow list is chained through the mark word of 8196 // each object in the list. We fetch the entire list, 8197 // break off a prefix of the right size and return the 8198 // remainder. If other threads try to take objects from 8199 // the overflow list at that time, they will wait for 8200 // some time to see if data becomes available. If (and 8201 // only if) another thread places one or more object(s) 8202 // on the global list before we have returned the suffix 8203 // to the global list, we will walk down our local list 8204 // to find its end and append the global list to 8205 // our suffix before returning it. This suffix walk can 8206 // prove to be expensive (quadratic in the amount of traffic) 8207 // when there are many objects in the overflow list and 8208 // there is much producer-consumer contention on the list. 8209 // *NOTE*: The overflow list manipulation code here and 8210 // in ParNewGeneration:: are very similar in shape, 8211 // except that in the ParNew case we use the old (from/eden) 8212 // copy of the object to thread the list via its klass word. 8213 // Because of the common code, if you make any changes in 8214 // the code below, please check the ParNew version to see if 8215 // similar changes might be needed. 8216 // CR 6797058 has been filed to consolidate the common code. 8217 bool CMSCollector::par_take_from_overflow_list(size_t num, 8218 OopTaskQueue* work_q, 8219 int no_of_gc_threads) { 8220 assert(work_q->size() == 0, "First empty local work queue"); 8221 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8222 if (_overflow_list == NULL) { 8223 return false; 8224 } 8225 // Grab the entire list; we'll put back a suffix 8226 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8227 Thread* tid = Thread::current(); 8228 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8229 // set to ParallelGCThreads. 8230 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8231 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8232 // If the list is busy, we spin for a short while, 8233 // sleeping between attempts to get the list. 8234 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8235 os::sleep(tid, sleep_time_millis, false); 8236 if (_overflow_list == NULL) { 8237 // Nothing left to take 8238 return false; 8239 } else if (_overflow_list != BUSY) { 8240 // Try and grab the prefix 8241 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8242 } 8243 } 8244 // If the list was found to be empty, or we spun long 8245 // enough, we give up and return empty-handed. If we leave 8246 // the list in the BUSY state below, it must be the case that 8247 // some other thread holds the overflow list and will set it 8248 // to a non-BUSY state in the future. 8249 if (prefix == NULL || prefix == BUSY) { 8250 // Nothing to take or waited long enough 8251 if (prefix == NULL) { 8252 // Write back the NULL in case we overwrote it with BUSY above 8253 // and it is still the same value. 8254 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8255 } 8256 return false; 8257 } 8258 assert(prefix != NULL && prefix != BUSY, "Error"); 8259 size_t i = num; 8260 oop cur = prefix; 8261 // Walk down the first "num" objects, unless we reach the end. 8262 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8263 if (cur->mark() == NULL) { 8264 // We have "num" or fewer elements in the list, so there 8265 // is nothing to return to the global list. 8266 // Write back the NULL in lieu of the BUSY we wrote 8267 // above, if it is still the same value. 8268 if (_overflow_list == BUSY) { 8269 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8270 } 8271 } else { 8272 // Chop off the suffix and return it to the global list. 8273 assert(cur->mark() != BUSY, "Error"); 8274 oop suffix_head = cur->mark(); // suffix will be put back on global list 8275 cur->set_mark(NULL); // break off suffix 8276 // It's possible that the list is still in the empty(busy) state 8277 // we left it in a short while ago; in that case we may be 8278 // able to place back the suffix without incurring the cost 8279 // of a walk down the list. 8280 oop observed_overflow_list = _overflow_list; 8281 oop cur_overflow_list = observed_overflow_list; 8282 bool attached = false; 8283 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8284 observed_overflow_list = 8285 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8286 if (cur_overflow_list == observed_overflow_list) { 8287 attached = true; 8288 break; 8289 } else cur_overflow_list = observed_overflow_list; 8290 } 8291 if (!attached) { 8292 // Too bad, someone else sneaked in (at least) an element; we'll need 8293 // to do a splice. Find tail of suffix so we can prepend suffix to global 8294 // list. 8295 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8296 oop suffix_tail = cur; 8297 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8298 "Tautology"); 8299 observed_overflow_list = _overflow_list; 8300 do { 8301 cur_overflow_list = observed_overflow_list; 8302 if (cur_overflow_list != BUSY) { 8303 // Do the splice ... 8304 suffix_tail->set_mark(markOop(cur_overflow_list)); 8305 } else { // cur_overflow_list == BUSY 8306 suffix_tail->set_mark(NULL); 8307 } 8308 // ... and try to place spliced list back on overflow_list ... 8309 observed_overflow_list = 8310 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8311 } while (cur_overflow_list != observed_overflow_list); 8312 // ... until we have succeeded in doing so. 8313 } 8314 } 8315 8316 // Push the prefix elements on work_q 8317 assert(prefix != NULL, "control point invariant"); 8318 const markOop proto = markOopDesc::prototype(); 8319 oop next; 8320 NOT_PRODUCT(ssize_t n = 0;) 8321 for (cur = prefix; cur != NULL; cur = next) { 8322 next = oop(cur->mark()); 8323 cur->set_mark(proto); // until proven otherwise 8324 assert(cur->is_oop(), "Should be an oop"); 8325 bool res = work_q->push(cur); 8326 assert(res, "Bit off more than we can chew?"); 8327 NOT_PRODUCT(n++;) 8328 } 8329 #ifndef PRODUCT 8330 assert(_num_par_pushes >= n, "Too many pops?"); 8331 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8332 #endif 8333 return true; 8334 } 8335 8336 // Single-threaded 8337 void CMSCollector::push_on_overflow_list(oop p) { 8338 NOT_PRODUCT(_num_par_pushes++;) 8339 assert(p->is_oop(), "Not an oop"); 8340 preserve_mark_if_necessary(p); 8341 p->set_mark((markOop)_overflow_list); 8342 _overflow_list = p; 8343 } 8344 8345 // Multi-threaded; use CAS to prepend to overflow list 8346 void CMSCollector::par_push_on_overflow_list(oop p) { 8347 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8348 assert(p->is_oop(), "Not an oop"); 8349 par_preserve_mark_if_necessary(p); 8350 oop observed_overflow_list = _overflow_list; 8351 oop cur_overflow_list; 8352 do { 8353 cur_overflow_list = observed_overflow_list; 8354 if (cur_overflow_list != BUSY) { 8355 p->set_mark(markOop(cur_overflow_list)); 8356 } else { 8357 p->set_mark(NULL); 8358 } 8359 observed_overflow_list = 8360 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8361 } while (cur_overflow_list != observed_overflow_list); 8362 } 8363 #undef BUSY 8364 8365 // Single threaded 8366 // General Note on GrowableArray: pushes may silently fail 8367 // because we are (temporarily) out of C-heap for expanding 8368 // the stack. The problem is quite ubiquitous and affects 8369 // a lot of code in the JVM. The prudent thing for GrowableArray 8370 // to do (for now) is to exit with an error. However, that may 8371 // be too draconian in some cases because the caller may be 8372 // able to recover without much harm. For such cases, we 8373 // should probably introduce a "soft_push" method which returns 8374 // an indication of success or failure with the assumption that 8375 // the caller may be able to recover from a failure; code in 8376 // the VM can then be changed, incrementally, to deal with such 8377 // failures where possible, thus, incrementally hardening the VM 8378 // in such low resource situations. 8379 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8380 _preserved_oop_stack.push(p); 8381 _preserved_mark_stack.push(m); 8382 assert(m == p->mark(), "Mark word changed"); 8383 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8384 "bijection"); 8385 } 8386 8387 // Single threaded 8388 void CMSCollector::preserve_mark_if_necessary(oop p) { 8389 markOop m = p->mark(); 8390 if (m->must_be_preserved(p)) { 8391 preserve_mark_work(p, m); 8392 } 8393 } 8394 8395 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8396 markOop m = p->mark(); 8397 if (m->must_be_preserved(p)) { 8398 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8399 // Even though we read the mark word without holding 8400 // the lock, we are assured that it will not change 8401 // because we "own" this oop, so no other thread can 8402 // be trying to push it on the overflow list; see 8403 // the assertion in preserve_mark_work() that checks 8404 // that m == p->mark(). 8405 preserve_mark_work(p, m); 8406 } 8407 } 8408 8409 // We should be able to do this multi-threaded, 8410 // a chunk of stack being a task (this is 8411 // correct because each oop only ever appears 8412 // once in the overflow list. However, it's 8413 // not very easy to completely overlap this with 8414 // other operations, so will generally not be done 8415 // until all work's been completed. Because we 8416 // expect the preserved oop stack (set) to be small, 8417 // it's probably fine to do this single-threaded. 8418 // We can explore cleverer concurrent/overlapped/parallel 8419 // processing of preserved marks if we feel the 8420 // need for this in the future. Stack overflow should 8421 // be so rare in practice and, when it happens, its 8422 // effect on performance so great that this will 8423 // likely just be in the noise anyway. 8424 void CMSCollector::restore_preserved_marks_if_any() { 8425 assert(SafepointSynchronize::is_at_safepoint(), 8426 "world should be stopped"); 8427 assert(Thread::current()->is_ConcurrentGC_thread() || 8428 Thread::current()->is_VM_thread(), 8429 "should be single-threaded"); 8430 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8431 "bijection"); 8432 8433 while (!_preserved_oop_stack.is_empty()) { 8434 oop p = _preserved_oop_stack.pop(); 8435 assert(p->is_oop(), "Should be an oop"); 8436 assert(_span.contains(p), "oop should be in _span"); 8437 assert(p->mark() == markOopDesc::prototype(), 8438 "Set when taken from overflow list"); 8439 markOop m = _preserved_mark_stack.pop(); 8440 p->set_mark(m); 8441 } 8442 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8443 "stacks were cleared above"); 8444 } 8445 8446 #ifndef PRODUCT 8447 bool CMSCollector::no_preserved_marks() const { 8448 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8449 } 8450 #endif 8451 8452 // Transfer some number of overflown objects to usual marking 8453 // stack. Return true if some objects were transferred. 8454 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8455 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8456 (size_t)ParGCDesiredObjsFromOverflowList); 8457 8458 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8459 assert(_collector->overflow_list_is_empty() || res, 8460 "If list is not empty, we should have taken something"); 8461 assert(!res || !_mark_stack->isEmpty(), 8462 "If we took something, it should now be on our stack"); 8463 return res; 8464 } 8465 8466 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8467 size_t res = _sp->block_size_no_stall(addr, _collector); 8468 if (_sp->block_is_obj(addr)) { 8469 if (_live_bit_map->isMarked(addr)) { 8470 // It can't have been dead in a previous cycle 8471 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8472 } else { 8473 _dead_bit_map->mark(addr); // mark the dead object 8474 } 8475 } 8476 // Could be 0, if the block size could not be computed without stalling. 8477 return res; 8478 } 8479 8480 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8481 8482 switch (phase) { 8483 case CMSCollector::InitialMarking: 8484 initialize(true /* fullGC */ , 8485 cause /* cause of the GC */, 8486 true /* recordGCBeginTime */, 8487 true /* recordPreGCUsage */, 8488 false /* recordPeakUsage */, 8489 false /* recordPostGCusage */, 8490 true /* recordAccumulatedGCTime */, 8491 false /* recordGCEndTime */, 8492 false /* countCollection */ ); 8493 break; 8494 8495 case CMSCollector::FinalMarking: 8496 initialize(true /* fullGC */ , 8497 cause /* cause of the GC */, 8498 false /* recordGCBeginTime */, 8499 false /* recordPreGCUsage */, 8500 false /* recordPeakUsage */, 8501 false /* recordPostGCusage */, 8502 true /* recordAccumulatedGCTime */, 8503 false /* recordGCEndTime */, 8504 false /* countCollection */ ); 8505 break; 8506 8507 case CMSCollector::Sweeping: 8508 initialize(true /* fullGC */ , 8509 cause /* cause of the GC */, 8510 false /* recordGCBeginTime */, 8511 false /* recordPreGCUsage */, 8512 true /* recordPeakUsage */, 8513 true /* recordPostGCusage */, 8514 false /* recordAccumulatedGCTime */, 8515 true /* recordGCEndTime */, 8516 true /* countCollection */ ); 8517 break; 8518 8519 default: 8520 ShouldNotReachHere(); 8521 } 8522 }