1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardGeneration.inline.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/strongRootsScope.hpp" 57 #include "memory/tenuredGeneration.hpp" 58 #include "oops/oop.inline.hpp" 59 #include "prims/jvmtiExport.hpp" 60 #include "runtime/atomic.inline.hpp" 61 #include "runtime/globals_extension.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/orderAccess.inline.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "services/memoryService.hpp" 67 #include "services/runtimeService.hpp" 68 #include "utilities/stack.inline.hpp" 69 #include "utilities/taskqueue.inline.hpp" 70 71 // statics 72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 73 bool CMSCollector::_full_gc_requested = false; 74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 75 76 ////////////////////////////////////////////////////////////////// 77 // In support of CMS/VM thread synchronization 78 ////////////////////////////////////////////////////////////////// 79 // We split use of the CGC_lock into 2 "levels". 80 // The low-level locking is of the usual CGC_lock monitor. We introduce 81 // a higher level "token" (hereafter "CMS token") built on top of the 82 // low level monitor (hereafter "CGC lock"). 83 // The token-passing protocol gives priority to the VM thread. The 84 // CMS-lock doesn't provide any fairness guarantees, but clients 85 // should ensure that it is only held for very short, bounded 86 // durations. 87 // 88 // When either of the CMS thread or the VM thread is involved in 89 // collection operations during which it does not want the other 90 // thread to interfere, it obtains the CMS token. 91 // 92 // If either thread tries to get the token while the other has 93 // it, that thread waits. However, if the VM thread and CMS thread 94 // both want the token, then the VM thread gets priority while the 95 // CMS thread waits. This ensures, for instance, that the "concurrent" 96 // phases of the CMS thread's work do not block out the VM thread 97 // for long periods of time as the CMS thread continues to hog 98 // the token. (See bug 4616232). 99 // 100 // The baton-passing functions are, however, controlled by the 101 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 102 // and here the low-level CMS lock, not the high level token, 103 // ensures mutual exclusion. 104 // 105 // Two important conditions that we have to satisfy: 106 // 1. if a thread does a low-level wait on the CMS lock, then it 107 // relinquishes the CMS token if it were holding that token 108 // when it acquired the low-level CMS lock. 109 // 2. any low-level notifications on the low-level lock 110 // should only be sent when a thread has relinquished the token. 111 // 112 // In the absence of either property, we'd have potential deadlock. 113 // 114 // We protect each of the CMS (concurrent and sequential) phases 115 // with the CMS _token_, not the CMS _lock_. 116 // 117 // The only code protected by CMS lock is the token acquisition code 118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 119 // baton-passing code. 120 // 121 // Unfortunately, i couldn't come up with a good abstraction to factor and 122 // hide the naked CGC_lock manipulation in the baton-passing code 123 // further below. That's something we should try to do. Also, the proof 124 // of correctness of this 2-level locking scheme is far from obvious, 125 // and potentially quite slippery. We have an uneasy suspicion, for instance, 126 // that there may be a theoretical possibility of delay/starvation in the 127 // low-level lock/wait/notify scheme used for the baton-passing because of 128 // potential interference with the priority scheme embodied in the 129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 130 // invocation further below and marked with "XXX 20011219YSR". 131 // Indeed, as we note elsewhere, this may become yet more slippery 132 // in the presence of multiple CMS and/or multiple VM threads. XXX 133 134 class CMSTokenSync: public StackObj { 135 private: 136 bool _is_cms_thread; 137 public: 138 CMSTokenSync(bool is_cms_thread): 139 _is_cms_thread(is_cms_thread) { 140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 141 "Incorrect argument to constructor"); 142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 143 } 144 145 ~CMSTokenSync() { 146 assert(_is_cms_thread ? 147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 149 "Incorrect state"); 150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 151 } 152 }; 153 154 // Convenience class that does a CMSTokenSync, and then acquires 155 // upto three locks. 156 class CMSTokenSyncWithLocks: public CMSTokenSync { 157 private: 158 // Note: locks are acquired in textual declaration order 159 // and released in the opposite order 160 MutexLockerEx _locker1, _locker2, _locker3; 161 public: 162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 164 CMSTokenSync(is_cms_thread), 165 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 166 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 167 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 168 { } 169 }; 170 171 172 ////////////////////////////////////////////////////////////////// 173 // Concurrent Mark-Sweep Generation ///////////////////////////// 174 ////////////////////////////////////////////////////////////////// 175 176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 177 178 // This struct contains per-thread things necessary to support parallel 179 // young-gen collection. 180 class CMSParGCThreadState: public CHeapObj<mtGC> { 181 public: 182 CFLS_LAB lab; 183 PromotionInfo promo; 184 185 // Constructor. 186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 187 promo.setSpace(cfls); 188 } 189 }; 190 191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 192 ReservedSpace rs, size_t initial_byte_size, int level, 193 CardTableRS* ct, bool use_adaptive_freelists, 194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 195 CardGeneration(rs, initial_byte_size, level, ct), 196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 231 for (uint i = 0; i < ParallelGCThreads; i++) { 232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 233 } 234 235 _incremental_collection_failed = false; 236 // The "dilatation_factor" is the expansion that can occur on 237 // account of the fact that the minimum object size in the CMS 238 // generation may be larger than that in, say, a contiguous young 239 // generation. 240 // Ideally, in the calculation below, we'd compute the dilatation 241 // factor as: MinChunkSize/(promoting_gen's min object size) 242 // Since we do not have such a general query interface for the 243 // promoting generation, we'll instead just use the minimum 244 // object size (which today is a header's worth of space); 245 // note that all arithmetic is in units of HeapWords. 246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 247 assert(_dilatation_factor >= 1.0, "from previous assert"); 248 } 249 250 251 // The field "_initiating_occupancy" represents the occupancy percentage 252 // at which we trigger a new collection cycle. Unless explicitly specified 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it 254 // is calculated by: 255 // 256 // Let "f" be MinHeapFreeRatio in 257 // 258 // _initiating_occupancy = 100-f + 259 // f * (CMSTriggerRatio/100) 260 // where CMSTriggerRatio is the argument "tr" below. 261 // 262 // That is, if we assume the heap is at its desired maximum occupancy at the 263 // end of a collection, we let CMSTriggerRatio of the (purported) free 264 // space be allocated before initiating a new collection cycle. 265 // 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 267 assert(io <= 100 && tr <= 100, "Check the arguments"); 268 if (io >= 0) { 269 _initiating_occupancy = (double)io / 100.0; 270 } else { 271 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 272 (double)(tr * MinHeapFreeRatio) / 100.0) 273 / 100.0; 274 } 275 } 276 277 void ConcurrentMarkSweepGeneration::ref_processor_init() { 278 assert(collector() != NULL, "no collector"); 279 collector()->ref_processor_init(); 280 } 281 282 void CMSCollector::ref_processor_init() { 283 if (_ref_processor == NULL) { 284 // Allocate and initialize a reference processor 285 _ref_processor = 286 new ReferenceProcessor(_span, // span 287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 288 (int) ParallelGCThreads, // mt processing degree 289 _cmsGen->refs_discovery_is_mt(), // mt discovery 290 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 292 &_is_alive_closure); // closure for liveness info 293 // Initialize the _ref_processor field of CMSGen 294 _cmsGen->set_ref_processor(_ref_processor); 295 296 } 297 } 298 299 AdaptiveSizePolicy* CMSCollector::size_policy() { 300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 301 return gch->gen_policy()->size_policy(); 302 } 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 308 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next minor collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 if (PrintGCDetails && Verbose) { 375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 377 cms_free, expected_promotion); 378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 379 cms_free_dbl, cms_consumption_rate() + 1.0); 380 } 381 // Add 1 in case the consumption rate goes to zero. 382 return cms_free_dbl / (cms_consumption_rate() + 1.0); 383 } 384 return 0.0; 385 } 386 387 // Compare the duration of the cms collection to the 388 // time remaining before the cms generation is empty. 389 // Note that the time from the start of the cms collection 390 // to the start of the cms sweep (less than the total 391 // duration of the cms collection) can be used. This 392 // has been tried and some applications experienced 393 // promotion failures early in execution. This was 394 // possibly because the averages were not accurate 395 // enough at the beginning. 396 double CMSStats::time_until_cms_start() const { 397 // We add "gc0_period" to the "work" calculation 398 // below because this query is done (mostly) at the 399 // end of a scavenge, so we need to conservatively 400 // account for that much possible delay 401 // in the query so as to avoid concurrent mode failures 402 // due to starting the collection just a wee bit too 403 // late. 404 double work = cms_duration() + gc0_period(); 405 double deadline = time_until_cms_gen_full(); 406 // If a concurrent mode failure occurred recently, we want to be 407 // more conservative and halve our expected time_until_cms_gen_full() 408 if (work > deadline) { 409 if (Verbose && PrintGCDetails) { 410 gclog_or_tty->print( 411 " CMSCollector: collect because of anticipated promotion " 412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 413 gc0_period(), time_until_cms_gen_full()); 414 } 415 return 0.0; 416 } 417 return work - deadline; 418 } 419 420 #ifndef PRODUCT 421 void CMSStats::print_on(outputStream *st) const { 422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 424 gc0_duration(), gc0_period(), gc0_promoted()); 425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 426 cms_duration(), cms_period(), cms_allocated()); 427 st->print(",cms_since_beg=%g,cms_since_end=%g", 428 cms_time_since_begin(), cms_time_since_end()); 429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 431 432 if (valid()) { 433 st->print(",promo_rate=%g,cms_alloc_rate=%g", 434 promotion_rate(), cms_allocation_rate()); 435 st->print(",cms_consumption_rate=%g,time_until_full=%g", 436 cms_consumption_rate(), time_until_cms_gen_full()); 437 } 438 st->print(" "); 439 } 440 #endif // #ifndef PRODUCT 441 442 CMSCollector::CollectorState CMSCollector::_collectorState = 443 CMSCollector::Idling; 444 bool CMSCollector::_foregroundGCIsActive = false; 445 bool CMSCollector::_foregroundGCShouldWait = false; 446 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 448 CardTableRS* ct, 449 ConcurrentMarkSweepPolicy* cp): 450 _cmsGen(cmsGen), 451 _ct(ct), 452 _ref_processor(NULL), // will be set later 453 _conc_workers(NULL), // may be set later 454 _abort_preclean(false), 455 _start_sampling(false), 456 _between_prologue_and_epilogue(false), 457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 458 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 459 -1 /* lock-free */, "No_lock" /* dummy */), 460 _modUnionClosure(&_modUnionTable), 461 _modUnionClosurePar(&_modUnionTable), 462 // Adjust my span to cover old (cms) gen 463 _span(cmsGen->reserved()), 464 // Construct the is_alive_closure with _span & markBitMap 465 _is_alive_closure(_span, &_markBitMap), 466 _restart_addr(NULL), 467 _overflow_list(NULL), 468 _stats(cmsGen), 469 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 470 //verify that this lock should be acquired with safepoint check. 471 Monitor::_safepoint_check_sometimes)), 472 _eden_chunk_array(NULL), // may be set in ctor body 473 _eden_chunk_capacity(0), // -- ditto -- 474 _eden_chunk_index(0), // -- ditto -- 475 _survivor_plab_array(NULL), // -- ditto -- 476 _survivor_chunk_array(NULL), // -- ditto -- 477 _survivor_chunk_capacity(0), // -- ditto -- 478 _survivor_chunk_index(0), // -- ditto -- 479 _ser_pmc_preclean_ovflw(0), 480 _ser_kac_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_ovflw(0), 484 _par_kac_ovflw(0), 485 #ifndef PRODUCT 486 _num_par_pushes(0), 487 #endif 488 _collection_count_start(0), 489 _verifying(false), 490 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 491 _completed_initialization(false), 492 _collector_policy(cp), 493 _should_unload_classes(CMSClassUnloadingEnabled), 494 _concurrent_cycles_since_last_unload(0), 495 _roots_scanning_options(GenCollectedHeap::SO_None), 496 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 498 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 499 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 500 _cms_start_registered(false) 501 { 502 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 503 ExplicitGCInvokesConcurrent = true; 504 } 505 // Now expand the span and allocate the collection support structures 506 // (MUT, marking bit map etc.) to cover both generations subject to 507 // collection. 508 509 // For use by dirty card to oop closures. 510 _cmsGen->cmsSpace()->set_collector(this); 511 512 // Allocate MUT and marking bit map 513 { 514 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 515 if (!_markBitMap.allocate(_span)) { 516 warning("Failed to allocate CMS Bit Map"); 517 return; 518 } 519 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 520 } 521 { 522 _modUnionTable.allocate(_span); 523 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 524 } 525 526 if (!_markStack.allocate(MarkStackSize)) { 527 warning("Failed to allocate CMS Marking Stack"); 528 return; 529 } 530 531 // Support for multi-threaded concurrent phases 532 if (CMSConcurrentMTEnabled) { 533 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 534 // just for now 535 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 536 } 537 if (ConcGCThreads > 1) { 538 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 539 ConcGCThreads, true); 540 if (_conc_workers == NULL) { 541 warning("GC/CMS: _conc_workers allocation failure: " 542 "forcing -CMSConcurrentMTEnabled"); 543 CMSConcurrentMTEnabled = false; 544 } else { 545 _conc_workers->initialize_workers(); 546 } 547 } else { 548 CMSConcurrentMTEnabled = false; 549 } 550 } 551 if (!CMSConcurrentMTEnabled) { 552 ConcGCThreads = 0; 553 } else { 554 // Turn off CMSCleanOnEnter optimization temporarily for 555 // the MT case where it's not fixed yet; see 6178663. 556 CMSCleanOnEnter = false; 557 } 558 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 559 "Inconsistency"); 560 561 // Parallel task queues; these are shared for the 562 // concurrent and stop-world phases of CMS, but 563 // are not shared with parallel scavenge (ParNew). 564 { 565 uint i; 566 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 567 568 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 569 || ParallelRefProcEnabled) 570 && num_queues > 0) { 571 _task_queues = new OopTaskQueueSet(num_queues); 572 if (_task_queues == NULL) { 573 warning("task_queues allocation failure."); 574 return; 575 } 576 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 577 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 578 for (i = 0; i < num_queues; i++) { 579 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 580 if (q == NULL) { 581 warning("work_queue allocation failure."); 582 return; 583 } 584 _task_queues->register_queue(i, q); 585 } 586 for (i = 0; i < num_queues; i++) { 587 _task_queues->queue(i)->initialize(); 588 _hash_seed[i] = 17; // copied from ParNew 589 } 590 } 591 } 592 593 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 594 595 // Clip CMSBootstrapOccupancy between 0 and 100. 596 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 597 598 // Now tell CMS generations the identity of their collector 599 ConcurrentMarkSweepGeneration::set_collector(this); 600 601 // Create & start a CMS thread for this CMS collector 602 _cmsThread = ConcurrentMarkSweepThread::start(this); 603 assert(cmsThread() != NULL, "CMS Thread should have been created"); 604 assert(cmsThread()->collector() == this, 605 "CMS Thread should refer to this gen"); 606 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 607 608 // Support for parallelizing young gen rescan 609 GenCollectedHeap* gch = GenCollectedHeap::heap(); 610 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 611 _young_gen = (ParNewGeneration*)gch->young_gen(); 612 if (gch->supports_inline_contig_alloc()) { 613 _top_addr = gch->top_addr(); 614 _end_addr = gch->end_addr(); 615 assert(_young_gen != NULL, "no _young_gen"); 616 _eden_chunk_index = 0; 617 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 618 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 619 } 620 621 // Support for parallelizing survivor space rescan 622 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 623 const size_t max_plab_samples = 624 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 625 626 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 627 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 628 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 629 _survivor_chunk_capacity = 2*max_plab_samples; 630 for (uint i = 0; i < ParallelGCThreads; i++) { 631 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 632 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 633 assert(cur->end() == 0, "Should be 0"); 634 assert(cur->array() == vec, "Should be vec"); 635 assert(cur->capacity() == max_plab_samples, "Error"); 636 } 637 } 638 639 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 640 _gc_counters = new CollectorCounters("CMS", 1); 641 _completed_initialization = true; 642 _inter_sweep_timer.start(); // start of time 643 } 644 645 size_t CMSCollector::plab_sample_minimum_size() { 646 // The default value of MinTLABSize is 2k, but there is 647 // no way to get the default value if the flag has been overridden. 648 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 649 } 650 651 const char* ConcurrentMarkSweepGeneration::name() const { 652 return "concurrent mark-sweep generation"; 653 } 654 void ConcurrentMarkSweepGeneration::update_counters() { 655 if (UsePerfData) { 656 _space_counters->update_all(); 657 _gen_counters->update_all(); 658 } 659 } 660 661 // this is an optimized version of update_counters(). it takes the 662 // used value as a parameter rather than computing it. 663 // 664 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 665 if (UsePerfData) { 666 _space_counters->update_used(used); 667 _space_counters->update_capacity(); 668 _gen_counters->update_all(); 669 } 670 } 671 672 void ConcurrentMarkSweepGeneration::print() const { 673 Generation::print(); 674 cmsSpace()->print(); 675 } 676 677 #ifndef PRODUCT 678 void ConcurrentMarkSweepGeneration::print_statistics() { 679 cmsSpace()->printFLCensus(0); 680 } 681 #endif 682 683 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 684 GenCollectedHeap* gch = GenCollectedHeap::heap(); 685 if (PrintGCDetails) { 686 if (Verbose) { 687 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 688 level(), short_name(), s, used(), capacity()); 689 } else { 690 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 691 level(), short_name(), s, used() / K, capacity() / K); 692 } 693 } 694 if (Verbose) { 695 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 696 gch->used(), gch->capacity()); 697 } else { 698 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 699 gch->used() / K, gch->capacity() / K); 700 } 701 } 702 703 size_t 704 ConcurrentMarkSweepGeneration::contiguous_available() const { 705 // dld proposes an improvement in precision here. If the committed 706 // part of the space ends in a free block we should add that to 707 // uncommitted size in the calculation below. Will make this 708 // change later, staying with the approximation below for the 709 // time being. -- ysr. 710 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 711 } 712 713 size_t 714 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 715 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 716 } 717 718 size_t ConcurrentMarkSweepGeneration::max_available() const { 719 return free() + _virtual_space.uncommitted_size(); 720 } 721 722 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 723 size_t available = max_available(); 724 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 725 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 726 if (Verbose && PrintGCDetails) { 727 gclog_or_tty->print_cr( 728 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 729 "max_promo("SIZE_FORMAT")", 730 res? "":" not", available, res? ">=":"<", 731 av_promo, max_promotion_in_bytes); 732 } 733 return res; 734 } 735 736 // At a promotion failure dump information on block layout in heap 737 // (cms old generation). 738 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 739 if (CMSDumpAtPromotionFailure) { 740 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 741 } 742 } 743 744 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 745 // Clear the promotion information. These pointers can be adjusted 746 // along with all the other pointers into the heap but 747 // compaction is expected to be a rare event with 748 // a heap using cms so don't do it without seeing the need. 749 for (uint i = 0; i < ParallelGCThreads; i++) { 750 _par_gc_thread_states[i]->promo.reset(); 751 } 752 } 753 754 void ConcurrentMarkSweepGeneration::compute_new_size() { 755 assert_locked_or_safepoint(Heap_lock); 756 757 // If incremental collection failed, we just want to expand 758 // to the limit. 759 if (incremental_collection_failed()) { 760 clear_incremental_collection_failed(); 761 grow_to_reserved(); 762 return; 763 } 764 765 // The heap has been compacted but not reset yet. 766 // Any metric such as free() or used() will be incorrect. 767 768 CardGeneration::compute_new_size(); 769 770 // Reset again after a possible resizing 771 if (did_compact()) { 772 cmsSpace()->reset_after_compaction(); 773 } 774 } 775 776 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 777 assert_locked_or_safepoint(Heap_lock); 778 779 // If incremental collection failed, we just want to expand 780 // to the limit. 781 if (incremental_collection_failed()) { 782 clear_incremental_collection_failed(); 783 grow_to_reserved(); 784 return; 785 } 786 787 double free_percentage = ((double) free()) / capacity(); 788 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 789 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 790 791 // compute expansion delta needed for reaching desired free percentage 792 if (free_percentage < desired_free_percentage) { 793 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 794 assert(desired_capacity >= capacity(), "invalid expansion size"); 795 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 796 if (PrintGCDetails && Verbose) { 797 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 798 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 799 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 800 gclog_or_tty->print_cr(" Desired free fraction %f", 801 desired_free_percentage); 802 gclog_or_tty->print_cr(" Maximum free fraction %f", 803 maximum_free_percentage); 804 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 805 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 806 desired_capacity/1000); 807 int prev_level = level() - 1; 808 if (prev_level >= 0) { 809 size_t prev_size = 0; 810 GenCollectedHeap* gch = GenCollectedHeap::heap(); 811 Generation* prev_gen = gch->young_gen(); 812 prev_size = prev_gen->capacity(); 813 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 814 prev_size/1000); 815 } 816 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 817 unsafe_max_alloc_nogc()/1000); 818 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 819 contiguous_available()/1000); 820 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 821 expand_bytes); 822 } 823 // safe if expansion fails 824 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 825 if (PrintGCDetails && Verbose) { 826 gclog_or_tty->print_cr(" Expanded free fraction %f", 827 ((double) free()) / capacity()); 828 } 829 } else { 830 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 831 assert(desired_capacity <= capacity(), "invalid expansion size"); 832 size_t shrink_bytes = capacity() - desired_capacity; 833 // Don't shrink unless the delta is greater than the minimum shrink we want 834 if (shrink_bytes >= MinHeapDeltaBytes) { 835 shrink_free_list_by(shrink_bytes); 836 } 837 } 838 } 839 840 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 841 return cmsSpace()->freelistLock(); 842 } 843 844 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 845 bool tlab) { 846 CMSSynchronousYieldRequest yr; 847 MutexLockerEx x(freelistLock(), 848 Mutex::_no_safepoint_check_flag); 849 return have_lock_and_allocate(size, tlab); 850 } 851 852 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 853 bool tlab /* ignored */) { 854 assert_lock_strong(freelistLock()); 855 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 856 HeapWord* res = cmsSpace()->allocate(adjustedSize); 857 // Allocate the object live (grey) if the background collector has 858 // started marking. This is necessary because the marker may 859 // have passed this address and consequently this object will 860 // not otherwise be greyed and would be incorrectly swept up. 861 // Note that if this object contains references, the writing 862 // of those references will dirty the card containing this object 863 // allowing the object to be blackened (and its references scanned) 864 // either during a preclean phase or at the final checkpoint. 865 if (res != NULL) { 866 // We may block here with an uninitialized object with 867 // its mark-bit or P-bits not yet set. Such objects need 868 // to be safely navigable by block_start(). 869 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 870 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 871 collector()->direct_allocated(res, adjustedSize); 872 _direct_allocated_words += adjustedSize; 873 // allocation counters 874 NOT_PRODUCT( 875 _numObjectsAllocated++; 876 _numWordsAllocated += (int)adjustedSize; 877 ) 878 } 879 return res; 880 } 881 882 // In the case of direct allocation by mutators in a generation that 883 // is being concurrently collected, the object must be allocated 884 // live (grey) if the background collector has started marking. 885 // This is necessary because the marker may 886 // have passed this address and consequently this object will 887 // not otherwise be greyed and would be incorrectly swept up. 888 // Note that if this object contains references, the writing 889 // of those references will dirty the card containing this object 890 // allowing the object to be blackened (and its references scanned) 891 // either during a preclean phase or at the final checkpoint. 892 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 893 assert(_markBitMap.covers(start, size), "Out of bounds"); 894 if (_collectorState >= Marking) { 895 MutexLockerEx y(_markBitMap.lock(), 896 Mutex::_no_safepoint_check_flag); 897 // [see comments preceding SweepClosure::do_blk() below for details] 898 // 899 // Can the P-bits be deleted now? JJJ 900 // 901 // 1. need to mark the object as live so it isn't collected 902 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 903 // 3. need to mark the end of the object so marking, precleaning or sweeping 904 // can skip over uninitialized or unparsable objects. An allocated 905 // object is considered uninitialized for our purposes as long as 906 // its klass word is NULL. All old gen objects are parsable 907 // as soon as they are initialized.) 908 _markBitMap.mark(start); // object is live 909 _markBitMap.mark(start + 1); // object is potentially uninitialized? 910 _markBitMap.mark(start + size - 1); 911 // mark end of object 912 } 913 // check that oop looks uninitialized 914 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 915 } 916 917 void CMSCollector::promoted(bool par, HeapWord* start, 918 bool is_obj_array, size_t obj_size) { 919 assert(_markBitMap.covers(start), "Out of bounds"); 920 // See comment in direct_allocated() about when objects should 921 // be allocated live. 922 if (_collectorState >= Marking) { 923 // we already hold the marking bit map lock, taken in 924 // the prologue 925 if (par) { 926 _markBitMap.par_mark(start); 927 } else { 928 _markBitMap.mark(start); 929 } 930 // We don't need to mark the object as uninitialized (as 931 // in direct_allocated above) because this is being done with the 932 // world stopped and the object will be initialized by the 933 // time the marking, precleaning or sweeping get to look at it. 934 // But see the code for copying objects into the CMS generation, 935 // where we need to ensure that concurrent readers of the 936 // block offset table are able to safely navigate a block that 937 // is in flux from being free to being allocated (and in 938 // transition while being copied into) and subsequently 939 // becoming a bona-fide object when the copy/promotion is complete. 940 assert(SafepointSynchronize::is_at_safepoint(), 941 "expect promotion only at safepoints"); 942 943 if (_collectorState < Sweeping) { 944 // Mark the appropriate cards in the modUnionTable, so that 945 // this object gets scanned before the sweep. If this is 946 // not done, CMS generation references in the object might 947 // not get marked. 948 // For the case of arrays, which are otherwise precisely 949 // marked, we need to dirty the entire array, not just its head. 950 if (is_obj_array) { 951 // The [par_]mark_range() method expects mr.end() below to 952 // be aligned to the granularity of a bit's representation 953 // in the heap. In the case of the MUT below, that's a 954 // card size. 955 MemRegion mr(start, 956 (HeapWord*)round_to((intptr_t)(start + obj_size), 957 CardTableModRefBS::card_size /* bytes */)); 958 if (par) { 959 _modUnionTable.par_mark_range(mr); 960 } else { 961 _modUnionTable.mark_range(mr); 962 } 963 } else { // not an obj array; we can just mark the head 964 if (par) { 965 _modUnionTable.par_mark(start); 966 } else { 967 _modUnionTable.mark(start); 968 } 969 } 970 } 971 } 972 } 973 974 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 975 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 976 // allocate, copy and if necessary update promoinfo -- 977 // delegate to underlying space. 978 assert_lock_strong(freelistLock()); 979 980 #ifndef PRODUCT 981 if (GenCollectedHeap::heap()->promotion_should_fail()) { 982 return NULL; 983 } 984 #endif // #ifndef PRODUCT 985 986 oop res = _cmsSpace->promote(obj, obj_size); 987 if (res == NULL) { 988 // expand and retry 989 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 990 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 991 // Since this is the old generation, we don't try to promote 992 // into a more senior generation. 993 res = _cmsSpace->promote(obj, obj_size); 994 } 995 if (res != NULL) { 996 // See comment in allocate() about when objects should 997 // be allocated live. 998 assert(obj->is_oop(), "Will dereference klass pointer below"); 999 collector()->promoted(false, // Not parallel 1000 (HeapWord*)res, obj->is_objArray(), obj_size); 1001 // promotion counters 1002 NOT_PRODUCT( 1003 _numObjectsPromoted++; 1004 _numWordsPromoted += 1005 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1006 ) 1007 } 1008 return res; 1009 } 1010 1011 1012 // IMPORTANT: Notes on object size recognition in CMS. 1013 // --------------------------------------------------- 1014 // A block of storage in the CMS generation is always in 1015 // one of three states. A free block (FREE), an allocated 1016 // object (OBJECT) whose size() method reports the correct size, 1017 // and an intermediate state (TRANSIENT) in which its size cannot 1018 // be accurately determined. 1019 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1020 // ----------------------------------------------------- 1021 // FREE: klass_word & 1 == 1; mark_word holds block size 1022 // 1023 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1024 // obj->size() computes correct size 1025 // 1026 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1027 // 1028 // STATE IDENTIFICATION: (64 bit+COOPS) 1029 // ------------------------------------ 1030 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1031 // 1032 // OBJECT: klass_word installed; klass_word != 0; 1033 // obj->size() computes correct size 1034 // 1035 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1036 // 1037 // 1038 // STATE TRANSITION DIAGRAM 1039 // 1040 // mut / parnew mut / parnew 1041 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1042 // ^ | 1043 // |------------------------ DEAD <------------------------------------| 1044 // sweep mut 1045 // 1046 // While a block is in TRANSIENT state its size cannot be determined 1047 // so readers will either need to come back later or stall until 1048 // the size can be determined. Note that for the case of direct 1049 // allocation, P-bits, when available, may be used to determine the 1050 // size of an object that may not yet have been initialized. 1051 1052 // Things to support parallel young-gen collection. 1053 oop 1054 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1055 oop old, markOop m, 1056 size_t word_sz) { 1057 #ifndef PRODUCT 1058 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1059 return NULL; 1060 } 1061 #endif // #ifndef PRODUCT 1062 1063 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1064 PromotionInfo* promoInfo = &ps->promo; 1065 // if we are tracking promotions, then first ensure space for 1066 // promotion (including spooling space for saving header if necessary). 1067 // then allocate and copy, then track promoted info if needed. 1068 // When tracking (see PromotionInfo::track()), the mark word may 1069 // be displaced and in this case restoration of the mark word 1070 // occurs in the (oop_since_save_marks_)iterate phase. 1071 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1072 // Out of space for allocating spooling buffers; 1073 // try expanding and allocating spooling buffers. 1074 if (!expand_and_ensure_spooling_space(promoInfo)) { 1075 return NULL; 1076 } 1077 } 1078 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1079 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1080 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1081 if (obj_ptr == NULL) { 1082 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1083 if (obj_ptr == NULL) { 1084 return NULL; 1085 } 1086 } 1087 oop obj = oop(obj_ptr); 1088 OrderAccess::storestore(); 1089 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1090 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1091 // IMPORTANT: See note on object initialization for CMS above. 1092 // Otherwise, copy the object. Here we must be careful to insert the 1093 // klass pointer last, since this marks the block as an allocated object. 1094 // Except with compressed oops it's the mark word. 1095 HeapWord* old_ptr = (HeapWord*)old; 1096 // Restore the mark word copied above. 1097 obj->set_mark(m); 1098 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1099 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1100 OrderAccess::storestore(); 1101 1102 if (UseCompressedClassPointers) { 1103 // Copy gap missed by (aligned) header size calculation below 1104 obj->set_klass_gap(old->klass_gap()); 1105 } 1106 if (word_sz > (size_t)oopDesc::header_size()) { 1107 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1108 obj_ptr + oopDesc::header_size(), 1109 word_sz - oopDesc::header_size()); 1110 } 1111 1112 // Now we can track the promoted object, if necessary. We take care 1113 // to delay the transition from uninitialized to full object 1114 // (i.e., insertion of klass pointer) until after, so that it 1115 // atomically becomes a promoted object. 1116 if (promoInfo->tracking()) { 1117 promoInfo->track((PromotedObject*)obj, old->klass()); 1118 } 1119 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1120 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1121 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1122 1123 // Finally, install the klass pointer (this should be volatile). 1124 OrderAccess::storestore(); 1125 obj->set_klass(old->klass()); 1126 // We should now be able to calculate the right size for this object 1127 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1128 1129 collector()->promoted(true, // parallel 1130 obj_ptr, old->is_objArray(), word_sz); 1131 1132 NOT_PRODUCT( 1133 Atomic::inc_ptr(&_numObjectsPromoted); 1134 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1135 ) 1136 1137 return obj; 1138 } 1139 1140 void 1141 ConcurrentMarkSweepGeneration:: 1142 par_promote_alloc_done(int thread_num) { 1143 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1144 ps->lab.retire(thread_num); 1145 } 1146 1147 void 1148 ConcurrentMarkSweepGeneration:: 1149 par_oop_since_save_marks_iterate_done(int thread_num) { 1150 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1151 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1152 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1153 } 1154 1155 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1156 size_t size, 1157 bool tlab) 1158 { 1159 // We allow a STW collection only if a full 1160 // collection was requested. 1161 return full || should_allocate(size, tlab); // FIX ME !!! 1162 // This and promotion failure handling are connected at the 1163 // hip and should be fixed by untying them. 1164 } 1165 1166 bool CMSCollector::shouldConcurrentCollect() { 1167 if (_full_gc_requested) { 1168 if (Verbose && PrintGCDetails) { 1169 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1170 " gc request (or gc_locker)"); 1171 } 1172 return true; 1173 } 1174 1175 FreelistLocker x(this); 1176 // ------------------------------------------------------------------ 1177 // Print out lots of information which affects the initiation of 1178 // a collection. 1179 if (PrintCMSInitiationStatistics && stats().valid()) { 1180 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1181 gclog_or_tty->stamp(); 1182 gclog_or_tty->cr(); 1183 stats().print_on(gclog_or_tty); 1184 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1185 stats().time_until_cms_gen_full()); 1186 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1187 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1188 _cmsGen->contiguous_available()); 1189 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1190 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1191 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1192 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1193 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1194 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1195 gclog_or_tty->print_cr("metadata initialized %d", 1196 MetaspaceGC::should_concurrent_collect()); 1197 } 1198 // ------------------------------------------------------------------ 1199 1200 // If the estimated time to complete a cms collection (cms_duration()) 1201 // is less than the estimated time remaining until the cms generation 1202 // is full, start a collection. 1203 if (!UseCMSInitiatingOccupancyOnly) { 1204 if (stats().valid()) { 1205 if (stats().time_until_cms_start() == 0.0) { 1206 return true; 1207 } 1208 } else { 1209 // We want to conservatively collect somewhat early in order 1210 // to try and "bootstrap" our CMS/promotion statistics; 1211 // this branch will not fire after the first successful CMS 1212 // collection because the stats should then be valid. 1213 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1214 if (Verbose && PrintGCDetails) { 1215 gclog_or_tty->print_cr( 1216 " CMSCollector: collect for bootstrapping statistics:" 1217 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1218 _bootstrap_occupancy); 1219 } 1220 return true; 1221 } 1222 } 1223 } 1224 1225 // Otherwise, we start a collection cycle if 1226 // old gen want a collection cycle started. Each may use 1227 // an appropriate criterion for making this decision. 1228 // XXX We need to make sure that the gen expansion 1229 // criterion dovetails well with this. XXX NEED TO FIX THIS 1230 if (_cmsGen->should_concurrent_collect()) { 1231 if (Verbose && PrintGCDetails) { 1232 gclog_or_tty->print_cr("CMS old gen initiated"); 1233 } 1234 return true; 1235 } 1236 1237 // We start a collection if we believe an incremental collection may fail; 1238 // this is not likely to be productive in practice because it's probably too 1239 // late anyway. 1240 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1241 assert(gch->collector_policy()->is_generation_policy(), 1242 "You may want to check the correctness of the following"); 1243 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1244 if (Verbose && PrintGCDetails) { 1245 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1246 } 1247 return true; 1248 } 1249 1250 if (MetaspaceGC::should_concurrent_collect()) { 1251 if (Verbose && PrintGCDetails) { 1252 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1253 } 1254 return true; 1255 } 1256 1257 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1258 if (CMSTriggerInterval >= 0) { 1259 if (CMSTriggerInterval == 0) { 1260 // Trigger always 1261 return true; 1262 } 1263 1264 // Check the CMS time since begin (we do not check the stats validity 1265 // as we want to be able to trigger the first CMS cycle as well) 1266 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1267 if (Verbose && PrintGCDetails) { 1268 if (stats().valid()) { 1269 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1270 stats().cms_time_since_begin()); 1271 } else { 1272 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1273 } 1274 } 1275 return true; 1276 } 1277 } 1278 1279 return false; 1280 } 1281 1282 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1283 1284 // Clear _expansion_cause fields of constituent generations 1285 void CMSCollector::clear_expansion_cause() { 1286 _cmsGen->clear_expansion_cause(); 1287 } 1288 1289 // We should be conservative in starting a collection cycle. To 1290 // start too eagerly runs the risk of collecting too often in the 1291 // extreme. To collect too rarely falls back on full collections, 1292 // which works, even if not optimum in terms of concurrent work. 1293 // As a work around for too eagerly collecting, use the flag 1294 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1295 // giving the user an easily understandable way of controlling the 1296 // collections. 1297 // We want to start a new collection cycle if any of the following 1298 // conditions hold: 1299 // . our current occupancy exceeds the configured initiating occupancy 1300 // for this generation, or 1301 // . we recently needed to expand this space and have not, since that 1302 // expansion, done a collection of this generation, or 1303 // . the underlying space believes that it may be a good idea to initiate 1304 // a concurrent collection (this may be based on criteria such as the 1305 // following: the space uses linear allocation and linear allocation is 1306 // going to fail, or there is believed to be excessive fragmentation in 1307 // the generation, etc... or ... 1308 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1309 // the case of the old generation; see CR 6543076): 1310 // we may be approaching a point at which allocation requests may fail because 1311 // we will be out of sufficient free space given allocation rate estimates.] 1312 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1313 1314 assert_lock_strong(freelistLock()); 1315 if (occupancy() > initiating_occupancy()) { 1316 if (PrintGCDetails && Verbose) { 1317 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1318 short_name(), occupancy(), initiating_occupancy()); 1319 } 1320 return true; 1321 } 1322 if (UseCMSInitiatingOccupancyOnly) { 1323 return false; 1324 } 1325 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1326 if (PrintGCDetails && Verbose) { 1327 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1328 short_name()); 1329 } 1330 return true; 1331 } 1332 if (_cmsSpace->should_concurrent_collect()) { 1333 if (PrintGCDetails && Verbose) { 1334 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1335 short_name()); 1336 } 1337 return true; 1338 } 1339 return false; 1340 } 1341 1342 void ConcurrentMarkSweepGeneration::collect(bool full, 1343 bool clear_all_soft_refs, 1344 size_t size, 1345 bool tlab) 1346 { 1347 collector()->collect(full, clear_all_soft_refs, size, tlab); 1348 } 1349 1350 void CMSCollector::collect(bool full, 1351 bool clear_all_soft_refs, 1352 size_t size, 1353 bool tlab) 1354 { 1355 // The following "if" branch is present for defensive reasons. 1356 // In the current uses of this interface, it can be replaced with: 1357 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1358 // But I am not placing that assert here to allow future 1359 // generality in invoking this interface. 1360 if (GC_locker::is_active()) { 1361 // A consistency test for GC_locker 1362 assert(GC_locker::needs_gc(), "Should have been set already"); 1363 // Skip this foreground collection, instead 1364 // expanding the heap if necessary. 1365 // Need the free list locks for the call to free() in compute_new_size() 1366 compute_new_size(); 1367 return; 1368 } 1369 acquire_control_and_collect(full, clear_all_soft_refs); 1370 } 1371 1372 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1373 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1374 unsigned int gc_count = gch->total_full_collections(); 1375 if (gc_count == full_gc_count) { 1376 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1377 _full_gc_requested = true; 1378 _full_gc_cause = cause; 1379 CGC_lock->notify(); // nudge CMS thread 1380 } else { 1381 assert(gc_count > full_gc_count, "Error: causal loop"); 1382 } 1383 } 1384 1385 bool CMSCollector::is_external_interruption() { 1386 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1387 return GCCause::is_user_requested_gc(cause) || 1388 GCCause::is_serviceability_requested_gc(cause); 1389 } 1390 1391 void CMSCollector::report_concurrent_mode_interruption() { 1392 if (is_external_interruption()) { 1393 if (PrintGCDetails) { 1394 gclog_or_tty->print(" (concurrent mode interrupted)"); 1395 } 1396 } else { 1397 if (PrintGCDetails) { 1398 gclog_or_tty->print(" (concurrent mode failure)"); 1399 } 1400 _gc_tracer_cm->report_concurrent_mode_failure(); 1401 } 1402 } 1403 1404 1405 // The foreground and background collectors need to coordinate in order 1406 // to make sure that they do not mutually interfere with CMS collections. 1407 // When a background collection is active, 1408 // the foreground collector may need to take over (preempt) and 1409 // synchronously complete an ongoing collection. Depending on the 1410 // frequency of the background collections and the heap usage 1411 // of the application, this preemption can be seldom or frequent. 1412 // There are only certain 1413 // points in the background collection that the "collection-baton" 1414 // can be passed to the foreground collector. 1415 // 1416 // The foreground collector will wait for the baton before 1417 // starting any part of the collection. The foreground collector 1418 // will only wait at one location. 1419 // 1420 // The background collector will yield the baton before starting a new 1421 // phase of the collection (e.g., before initial marking, marking from roots, 1422 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1423 // of the loop which switches the phases. The background collector does some 1424 // of the phases (initial mark, final re-mark) with the world stopped. 1425 // Because of locking involved in stopping the world, 1426 // the foreground collector should not block waiting for the background 1427 // collector when it is doing a stop-the-world phase. The background 1428 // collector will yield the baton at an additional point just before 1429 // it enters a stop-the-world phase. Once the world is stopped, the 1430 // background collector checks the phase of the collection. If the 1431 // phase has not changed, it proceeds with the collection. If the 1432 // phase has changed, it skips that phase of the collection. See 1433 // the comments on the use of the Heap_lock in collect_in_background(). 1434 // 1435 // Variable used in baton passing. 1436 // _foregroundGCIsActive - Set to true by the foreground collector when 1437 // it wants the baton. The foreground clears it when it has finished 1438 // the collection. 1439 // _foregroundGCShouldWait - Set to true by the background collector 1440 // when it is running. The foreground collector waits while 1441 // _foregroundGCShouldWait is true. 1442 // CGC_lock - monitor used to protect access to the above variables 1443 // and to notify the foreground and background collectors. 1444 // _collectorState - current state of the CMS collection. 1445 // 1446 // The foreground collector 1447 // acquires the CGC_lock 1448 // sets _foregroundGCIsActive 1449 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1450 // various locks acquired in preparation for the collection 1451 // are released so as not to block the background collector 1452 // that is in the midst of a collection 1453 // proceeds with the collection 1454 // clears _foregroundGCIsActive 1455 // returns 1456 // 1457 // The background collector in a loop iterating on the phases of the 1458 // collection 1459 // acquires the CGC_lock 1460 // sets _foregroundGCShouldWait 1461 // if _foregroundGCIsActive is set 1462 // clears _foregroundGCShouldWait, notifies _CGC_lock 1463 // waits on _CGC_lock for _foregroundGCIsActive to become false 1464 // and exits the loop. 1465 // otherwise 1466 // proceed with that phase of the collection 1467 // if the phase is a stop-the-world phase, 1468 // yield the baton once more just before enqueueing 1469 // the stop-world CMS operation (executed by the VM thread). 1470 // returns after all phases of the collection are done 1471 // 1472 1473 void CMSCollector::acquire_control_and_collect(bool full, 1474 bool clear_all_soft_refs) { 1475 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1476 assert(!Thread::current()->is_ConcurrentGC_thread(), 1477 "shouldn't try to acquire control from self!"); 1478 1479 // Start the protocol for acquiring control of the 1480 // collection from the background collector (aka CMS thread). 1481 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1482 "VM thread should have CMS token"); 1483 // Remember the possibly interrupted state of an ongoing 1484 // concurrent collection 1485 CollectorState first_state = _collectorState; 1486 1487 // Signal to a possibly ongoing concurrent collection that 1488 // we want to do a foreground collection. 1489 _foregroundGCIsActive = true; 1490 1491 // release locks and wait for a notify from the background collector 1492 // releasing the locks in only necessary for phases which 1493 // do yields to improve the granularity of the collection. 1494 assert_lock_strong(bitMapLock()); 1495 // We need to lock the Free list lock for the space that we are 1496 // currently collecting. 1497 assert(haveFreelistLocks(), "Must be holding free list locks"); 1498 bitMapLock()->unlock(); 1499 releaseFreelistLocks(); 1500 { 1501 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1502 if (_foregroundGCShouldWait) { 1503 // We are going to be waiting for action for the CMS thread; 1504 // it had better not be gone (for instance at shutdown)! 1505 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1506 "CMS thread must be running"); 1507 // Wait here until the background collector gives us the go-ahead 1508 ConcurrentMarkSweepThread::clear_CMS_flag( 1509 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1510 // Get a possibly blocked CMS thread going: 1511 // Note that we set _foregroundGCIsActive true above, 1512 // without protection of the CGC_lock. 1513 CGC_lock->notify(); 1514 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1515 "Possible deadlock"); 1516 while (_foregroundGCShouldWait) { 1517 // wait for notification 1518 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1519 // Possibility of delay/starvation here, since CMS token does 1520 // not know to give priority to VM thread? Actually, i think 1521 // there wouldn't be any delay/starvation, but the proof of 1522 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1523 } 1524 ConcurrentMarkSweepThread::set_CMS_flag( 1525 ConcurrentMarkSweepThread::CMS_vm_has_token); 1526 } 1527 } 1528 // The CMS_token is already held. Get back the other locks. 1529 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1530 "VM thread should have CMS token"); 1531 getFreelistLocks(); 1532 bitMapLock()->lock_without_safepoint_check(); 1533 if (TraceCMSState) { 1534 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1535 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1536 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1537 } 1538 1539 // Inform cms gen if this was due to partial collection failing. 1540 // The CMS gen may use this fact to determine its expansion policy. 1541 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1542 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1543 assert(!_cmsGen->incremental_collection_failed(), 1544 "Should have been noticed, reacted to and cleared"); 1545 _cmsGen->set_incremental_collection_failed(); 1546 } 1547 1548 if (first_state > Idling) { 1549 report_concurrent_mode_interruption(); 1550 } 1551 1552 set_did_compact(true); 1553 1554 // If the collection is being acquired from the background 1555 // collector, there may be references on the discovered 1556 // references lists. Abandon those references, since some 1557 // of them may have become unreachable after concurrent 1558 // discovery; the STW compacting collector will redo discovery 1559 // more precisely, without being subject to floating garbage. 1560 // Leaving otherwise unreachable references in the discovered 1561 // lists would require special handling. 1562 ref_processor()->disable_discovery(); 1563 ref_processor()->abandon_partial_discovery(); 1564 ref_processor()->verify_no_references_recorded(); 1565 1566 if (first_state > Idling) { 1567 save_heap_summary(); 1568 } 1569 1570 do_compaction_work(clear_all_soft_refs); 1571 1572 // Has the GC time limit been exceeded? 1573 size_t max_eden_size = _young_gen->max_capacity() - 1574 _young_gen->to()->capacity() - 1575 _young_gen->from()->capacity(); 1576 GCCause::Cause gc_cause = gch->gc_cause(); 1577 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1578 _young_gen->eden()->used(), 1579 _cmsGen->max_capacity(), 1580 max_eden_size, 1581 full, 1582 gc_cause, 1583 gch->collector_policy()); 1584 1585 // Reset the expansion cause, now that we just completed 1586 // a collection cycle. 1587 clear_expansion_cause(); 1588 _foregroundGCIsActive = false; 1589 return; 1590 } 1591 1592 // Resize the tenured generation 1593 // after obtaining the free list locks for the 1594 // two generations. 1595 void CMSCollector::compute_new_size() { 1596 assert_locked_or_safepoint(Heap_lock); 1597 FreelistLocker z(this); 1598 MetaspaceGC::compute_new_size(); 1599 _cmsGen->compute_new_size_free_list(); 1600 } 1601 1602 // A work method used by the foreground collector to do 1603 // a mark-sweep-compact. 1604 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1605 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1606 1607 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1608 gc_timer->register_gc_start(); 1609 1610 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1611 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1612 1613 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1614 1615 // Temporarily widen the span of the weak reference processing to 1616 // the entire heap. 1617 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1618 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1619 // Temporarily, clear the "is_alive_non_header" field of the 1620 // reference processor. 1621 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1622 // Temporarily make reference _processing_ single threaded (non-MT). 1623 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1624 // Temporarily make refs discovery atomic 1625 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1626 // Temporarily make reference _discovery_ single threaded (non-MT) 1627 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1628 1629 ref_processor()->set_enqueuing_is_done(false); 1630 ref_processor()->enable_discovery(); 1631 ref_processor()->setup_policy(clear_all_soft_refs); 1632 // If an asynchronous collection finishes, the _modUnionTable is 1633 // all clear. If we are assuming the collection from an asynchronous 1634 // collection, clear the _modUnionTable. 1635 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1636 "_modUnionTable should be clear if the baton was not passed"); 1637 _modUnionTable.clear_all(); 1638 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1639 "mod union for klasses should be clear if the baton was passed"); 1640 _ct->klass_rem_set()->clear_mod_union(); 1641 1642 // We must adjust the allocation statistics being maintained 1643 // in the free list space. We do so by reading and clearing 1644 // the sweep timer and updating the block flux rate estimates below. 1645 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1646 if (_inter_sweep_timer.is_active()) { 1647 _inter_sweep_timer.stop(); 1648 // Note that we do not use this sample to update the _inter_sweep_estimate. 1649 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1650 _inter_sweep_estimate.padded_average(), 1651 _intra_sweep_estimate.padded_average()); 1652 } 1653 1654 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1655 ref_processor(), clear_all_soft_refs); 1656 #ifdef ASSERT 1657 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1658 size_t free_size = cms_space->free(); 1659 assert(free_size == 1660 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1661 * HeapWordSize, 1662 "All the free space should be compacted into one chunk at top"); 1663 assert(cms_space->dictionary()->total_chunk_size( 1664 debug_only(cms_space->freelistLock())) == 0 || 1665 cms_space->totalSizeInIndexedFreeLists() == 0, 1666 "All the free space should be in a single chunk"); 1667 size_t num = cms_space->totalCount(); 1668 assert((free_size == 0 && num == 0) || 1669 (free_size > 0 && (num == 1 || num == 2)), 1670 "There should be at most 2 free chunks after compaction"); 1671 #endif // ASSERT 1672 _collectorState = Resetting; 1673 assert(_restart_addr == NULL, 1674 "Should have been NULL'd before baton was passed"); 1675 reset(false /* == !concurrent */); 1676 _cmsGen->reset_after_compaction(); 1677 _concurrent_cycles_since_last_unload = 0; 1678 1679 // Clear any data recorded in the PLAB chunk arrays. 1680 if (_survivor_plab_array != NULL) { 1681 reset_survivor_plab_arrays(); 1682 } 1683 1684 // Adjust the per-size allocation stats for the next epoch. 1685 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1686 // Restart the "inter sweep timer" for the next epoch. 1687 _inter_sweep_timer.reset(); 1688 _inter_sweep_timer.start(); 1689 1690 gc_timer->register_gc_end(); 1691 1692 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1693 1694 // For a mark-sweep-compact, compute_new_size() will be called 1695 // in the heap's do_collection() method. 1696 } 1697 1698 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1699 ContiguousSpace* eden_space = _young_gen->eden(); 1700 ContiguousSpace* from_space = _young_gen->from(); 1701 ContiguousSpace* to_space = _young_gen->to(); 1702 // Eden 1703 if (_eden_chunk_array != NULL) { 1704 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1705 p2i(eden_space->bottom()), p2i(eden_space->top()), 1706 p2i(eden_space->end()), eden_space->capacity()); 1707 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1708 "_eden_chunk_capacity=" SIZE_FORMAT, 1709 _eden_chunk_index, _eden_chunk_capacity); 1710 for (size_t i = 0; i < _eden_chunk_index; i++) { 1711 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1712 i, p2i(_eden_chunk_array[i])); 1713 } 1714 } 1715 // Survivor 1716 if (_survivor_chunk_array != NULL) { 1717 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1718 p2i(from_space->bottom()), p2i(from_space->top()), 1719 p2i(from_space->end()), from_space->capacity()); 1720 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1721 "_survivor_chunk_capacity=" SIZE_FORMAT, 1722 _survivor_chunk_index, _survivor_chunk_capacity); 1723 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1724 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1725 i, p2i(_survivor_chunk_array[i])); 1726 } 1727 } 1728 } 1729 1730 void CMSCollector::getFreelistLocks() const { 1731 // Get locks for all free lists in all generations that this 1732 // collector is responsible for 1733 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1734 } 1735 1736 void CMSCollector::releaseFreelistLocks() const { 1737 // Release locks for all free lists in all generations that this 1738 // collector is responsible for 1739 _cmsGen->freelistLock()->unlock(); 1740 } 1741 1742 bool CMSCollector::haveFreelistLocks() const { 1743 // Check locks for all free lists in all generations that this 1744 // collector is responsible for 1745 assert_lock_strong(_cmsGen->freelistLock()); 1746 PRODUCT_ONLY(ShouldNotReachHere()); 1747 return true; 1748 } 1749 1750 // A utility class that is used by the CMS collector to 1751 // temporarily "release" the foreground collector from its 1752 // usual obligation to wait for the background collector to 1753 // complete an ongoing phase before proceeding. 1754 class ReleaseForegroundGC: public StackObj { 1755 private: 1756 CMSCollector* _c; 1757 public: 1758 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1759 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1760 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1761 // allow a potentially blocked foreground collector to proceed 1762 _c->_foregroundGCShouldWait = false; 1763 if (_c->_foregroundGCIsActive) { 1764 CGC_lock->notify(); 1765 } 1766 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1767 "Possible deadlock"); 1768 } 1769 1770 ~ReleaseForegroundGC() { 1771 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1772 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1773 _c->_foregroundGCShouldWait = true; 1774 } 1775 }; 1776 1777 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1778 assert(Thread::current()->is_ConcurrentGC_thread(), 1779 "A CMS asynchronous collection is only allowed on a CMS thread."); 1780 1781 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1782 { 1783 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1784 MutexLockerEx hl(Heap_lock, safepoint_check); 1785 FreelistLocker fll(this); 1786 MutexLockerEx x(CGC_lock, safepoint_check); 1787 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1788 // The foreground collector is active or we're 1789 // not using asynchronous collections. Skip this 1790 // background collection. 1791 assert(!_foregroundGCShouldWait, "Should be clear"); 1792 return; 1793 } else { 1794 assert(_collectorState == Idling, "Should be idling before start."); 1795 _collectorState = InitialMarking; 1796 register_gc_start(cause); 1797 // Reset the expansion cause, now that we are about to begin 1798 // a new cycle. 1799 clear_expansion_cause(); 1800 1801 // Clear the MetaspaceGC flag since a concurrent collection 1802 // is starting but also clear it after the collection. 1803 MetaspaceGC::set_should_concurrent_collect(false); 1804 } 1805 // Decide if we want to enable class unloading as part of the 1806 // ensuing concurrent GC cycle. 1807 update_should_unload_classes(); 1808 _full_gc_requested = false; // acks all outstanding full gc requests 1809 _full_gc_cause = GCCause::_no_gc; 1810 // Signal that we are about to start a collection 1811 gch->increment_total_full_collections(); // ... starting a collection cycle 1812 _collection_count_start = gch->total_full_collections(); 1813 } 1814 1815 // Used for PrintGC 1816 size_t prev_used; 1817 if (PrintGC && Verbose) { 1818 prev_used = _cmsGen->used(); 1819 } 1820 1821 // The change of the collection state is normally done at this level; 1822 // the exceptions are phases that are executed while the world is 1823 // stopped. For those phases the change of state is done while the 1824 // world is stopped. For baton passing purposes this allows the 1825 // background collector to finish the phase and change state atomically. 1826 // The foreground collector cannot wait on a phase that is done 1827 // while the world is stopped because the foreground collector already 1828 // has the world stopped and would deadlock. 1829 while (_collectorState != Idling) { 1830 if (TraceCMSState) { 1831 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1832 p2i(Thread::current()), _collectorState); 1833 } 1834 // The foreground collector 1835 // holds the Heap_lock throughout its collection. 1836 // holds the CMS token (but not the lock) 1837 // except while it is waiting for the background collector to yield. 1838 // 1839 // The foreground collector should be blocked (not for long) 1840 // if the background collector is about to start a phase 1841 // executed with world stopped. If the background 1842 // collector has already started such a phase, the 1843 // foreground collector is blocked waiting for the 1844 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1845 // are executed in the VM thread. 1846 // 1847 // The locking order is 1848 // PendingListLock (PLL) -- if applicable (FinalMarking) 1849 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1850 // CMS token (claimed in 1851 // stop_world_and_do() --> 1852 // safepoint_synchronize() --> 1853 // CMSThread::synchronize()) 1854 1855 { 1856 // Check if the FG collector wants us to yield. 1857 CMSTokenSync x(true); // is cms thread 1858 if (waitForForegroundGC()) { 1859 // We yielded to a foreground GC, nothing more to be 1860 // done this round. 1861 assert(_foregroundGCShouldWait == false, "We set it to false in " 1862 "waitForForegroundGC()"); 1863 if (TraceCMSState) { 1864 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1865 " exiting collection CMS state %d", 1866 p2i(Thread::current()), _collectorState); 1867 } 1868 return; 1869 } else { 1870 // The background collector can run but check to see if the 1871 // foreground collector has done a collection while the 1872 // background collector was waiting to get the CGC_lock 1873 // above. If yes, break so that _foregroundGCShouldWait 1874 // is cleared before returning. 1875 if (_collectorState == Idling) { 1876 break; 1877 } 1878 } 1879 } 1880 1881 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1882 "should be waiting"); 1883 1884 switch (_collectorState) { 1885 case InitialMarking: 1886 { 1887 ReleaseForegroundGC x(this); 1888 stats().record_cms_begin(); 1889 VM_CMS_Initial_Mark initial_mark_op(this); 1890 VMThread::execute(&initial_mark_op); 1891 } 1892 // The collector state may be any legal state at this point 1893 // since the background collector may have yielded to the 1894 // foreground collector. 1895 break; 1896 case Marking: 1897 // initial marking in checkpointRootsInitialWork has been completed 1898 if (markFromRoots()) { // we were successful 1899 assert(_collectorState == Precleaning, "Collector state should " 1900 "have changed"); 1901 } else { 1902 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1903 } 1904 break; 1905 case Precleaning: 1906 // marking from roots in markFromRoots has been completed 1907 preclean(); 1908 assert(_collectorState == AbortablePreclean || 1909 _collectorState == FinalMarking, 1910 "Collector state should have changed"); 1911 break; 1912 case AbortablePreclean: 1913 abortable_preclean(); 1914 assert(_collectorState == FinalMarking, "Collector state should " 1915 "have changed"); 1916 break; 1917 case FinalMarking: 1918 { 1919 ReleaseForegroundGC x(this); 1920 1921 VM_CMS_Final_Remark final_remark_op(this); 1922 VMThread::execute(&final_remark_op); 1923 } 1924 assert(_foregroundGCShouldWait, "block post-condition"); 1925 break; 1926 case Sweeping: 1927 // final marking in checkpointRootsFinal has been completed 1928 sweep(); 1929 assert(_collectorState == Resizing, "Collector state change " 1930 "to Resizing must be done under the free_list_lock"); 1931 1932 case Resizing: { 1933 // Sweeping has been completed... 1934 // At this point the background collection has completed. 1935 // Don't move the call to compute_new_size() down 1936 // into code that might be executed if the background 1937 // collection was preempted. 1938 { 1939 ReleaseForegroundGC x(this); // unblock FG collection 1940 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1941 CMSTokenSync z(true); // not strictly needed. 1942 if (_collectorState == Resizing) { 1943 compute_new_size(); 1944 save_heap_summary(); 1945 _collectorState = Resetting; 1946 } else { 1947 assert(_collectorState == Idling, "The state should only change" 1948 " because the foreground collector has finished the collection"); 1949 } 1950 } 1951 break; 1952 } 1953 case Resetting: 1954 // CMS heap resizing has been completed 1955 reset(true); 1956 assert(_collectorState == Idling, "Collector state should " 1957 "have changed"); 1958 1959 MetaspaceGC::set_should_concurrent_collect(false); 1960 1961 stats().record_cms_end(); 1962 // Don't move the concurrent_phases_end() and compute_new_size() 1963 // calls to here because a preempted background collection 1964 // has it's state set to "Resetting". 1965 break; 1966 case Idling: 1967 default: 1968 ShouldNotReachHere(); 1969 break; 1970 } 1971 if (TraceCMSState) { 1972 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1973 p2i(Thread::current()), _collectorState); 1974 } 1975 assert(_foregroundGCShouldWait, "block post-condition"); 1976 } 1977 1978 // Should this be in gc_epilogue? 1979 collector_policy()->counters()->update_counters(); 1980 1981 { 1982 // Clear _foregroundGCShouldWait and, in the event that the 1983 // foreground collector is waiting, notify it, before 1984 // returning. 1985 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1986 _foregroundGCShouldWait = false; 1987 if (_foregroundGCIsActive) { 1988 CGC_lock->notify(); 1989 } 1990 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1991 "Possible deadlock"); 1992 } 1993 if (TraceCMSState) { 1994 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1995 " exiting collection CMS state %d", 1996 p2i(Thread::current()), _collectorState); 1997 } 1998 if (PrintGC && Verbose) { 1999 _cmsGen->print_heap_change(prev_used); 2000 } 2001 } 2002 2003 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2004 _cms_start_registered = true; 2005 _gc_timer_cm->register_gc_start(); 2006 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2007 } 2008 2009 void CMSCollector::register_gc_end() { 2010 if (_cms_start_registered) { 2011 report_heap_summary(GCWhen::AfterGC); 2012 2013 _gc_timer_cm->register_gc_end(); 2014 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2015 _cms_start_registered = false; 2016 } 2017 } 2018 2019 void CMSCollector::save_heap_summary() { 2020 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2021 _last_heap_summary = gch->create_heap_summary(); 2022 _last_metaspace_summary = gch->create_metaspace_summary(); 2023 } 2024 2025 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2026 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2027 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2028 } 2029 2030 bool CMSCollector::waitForForegroundGC() { 2031 bool res = false; 2032 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2033 "CMS thread should have CMS token"); 2034 // Block the foreground collector until the 2035 // background collectors decides whether to 2036 // yield. 2037 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2038 _foregroundGCShouldWait = true; 2039 if (_foregroundGCIsActive) { 2040 // The background collector yields to the 2041 // foreground collector and returns a value 2042 // indicating that it has yielded. The foreground 2043 // collector can proceed. 2044 res = true; 2045 _foregroundGCShouldWait = false; 2046 ConcurrentMarkSweepThread::clear_CMS_flag( 2047 ConcurrentMarkSweepThread::CMS_cms_has_token); 2048 ConcurrentMarkSweepThread::set_CMS_flag( 2049 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2050 // Get a possibly blocked foreground thread going 2051 CGC_lock->notify(); 2052 if (TraceCMSState) { 2053 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2054 p2i(Thread::current()), _collectorState); 2055 } 2056 while (_foregroundGCIsActive) { 2057 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2058 } 2059 ConcurrentMarkSweepThread::set_CMS_flag( 2060 ConcurrentMarkSweepThread::CMS_cms_has_token); 2061 ConcurrentMarkSweepThread::clear_CMS_flag( 2062 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2063 } 2064 if (TraceCMSState) { 2065 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2066 p2i(Thread::current()), _collectorState); 2067 } 2068 return res; 2069 } 2070 2071 // Because of the need to lock the free lists and other structures in 2072 // the collector, common to all the generations that the collector is 2073 // collecting, we need the gc_prologues of individual CMS generations 2074 // delegate to their collector. It may have been simpler had the 2075 // current infrastructure allowed one to call a prologue on a 2076 // collector. In the absence of that we have the generation's 2077 // prologue delegate to the collector, which delegates back 2078 // some "local" work to a worker method in the individual generations 2079 // that it's responsible for collecting, while itself doing any 2080 // work common to all generations it's responsible for. A similar 2081 // comment applies to the gc_epilogue()'s. 2082 // The role of the variable _between_prologue_and_epilogue is to 2083 // enforce the invocation protocol. 2084 void CMSCollector::gc_prologue(bool full) { 2085 // Call gc_prologue_work() for the CMSGen 2086 // we are responsible for. 2087 2088 // The following locking discipline assumes that we are only called 2089 // when the world is stopped. 2090 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2091 2092 // The CMSCollector prologue must call the gc_prologues for the 2093 // "generations" that it's responsible 2094 // for. 2095 2096 assert( Thread::current()->is_VM_thread() 2097 || ( CMSScavengeBeforeRemark 2098 && Thread::current()->is_ConcurrentGC_thread()), 2099 "Incorrect thread type for prologue execution"); 2100 2101 if (_between_prologue_and_epilogue) { 2102 // We have already been invoked; this is a gc_prologue delegation 2103 // from yet another CMS generation that we are responsible for, just 2104 // ignore it since all relevant work has already been done. 2105 return; 2106 } 2107 2108 // set a bit saying prologue has been called; cleared in epilogue 2109 _between_prologue_and_epilogue = true; 2110 // Claim locks for common data structures, then call gc_prologue_work() 2111 // for each CMSGen. 2112 2113 getFreelistLocks(); // gets free list locks on constituent spaces 2114 bitMapLock()->lock_without_safepoint_check(); 2115 2116 // Should call gc_prologue_work() for all cms gens we are responsible for 2117 bool duringMarking = _collectorState >= Marking 2118 && _collectorState < Sweeping; 2119 2120 // The young collections clear the modified oops state, which tells if 2121 // there are any modified oops in the class. The remark phase also needs 2122 // that information. Tell the young collection to save the union of all 2123 // modified klasses. 2124 if (duringMarking) { 2125 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2126 } 2127 2128 bool registerClosure = duringMarking; 2129 2130 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2131 2132 if (!full) { 2133 stats().record_gc0_begin(); 2134 } 2135 } 2136 2137 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2138 2139 _capacity_at_prologue = capacity(); 2140 _used_at_prologue = used(); 2141 2142 // Delegate to CMScollector which knows how to coordinate between 2143 // this and any other CMS generations that it is responsible for 2144 // collecting. 2145 collector()->gc_prologue(full); 2146 } 2147 2148 // This is a "private" interface for use by this generation's CMSCollector. 2149 // Not to be called directly by any other entity (for instance, 2150 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2151 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2152 bool registerClosure, ModUnionClosure* modUnionClosure) { 2153 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2154 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2155 "Should be NULL"); 2156 if (registerClosure) { 2157 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2158 } 2159 cmsSpace()->gc_prologue(); 2160 // Clear stat counters 2161 NOT_PRODUCT( 2162 assert(_numObjectsPromoted == 0, "check"); 2163 assert(_numWordsPromoted == 0, "check"); 2164 if (Verbose && PrintGC) { 2165 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2166 SIZE_FORMAT" bytes concurrently", 2167 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2168 } 2169 _numObjectsAllocated = 0; 2170 _numWordsAllocated = 0; 2171 ) 2172 } 2173 2174 void CMSCollector::gc_epilogue(bool full) { 2175 // The following locking discipline assumes that we are only called 2176 // when the world is stopped. 2177 assert(SafepointSynchronize::is_at_safepoint(), 2178 "world is stopped assumption"); 2179 2180 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2181 // if linear allocation blocks need to be appropriately marked to allow the 2182 // the blocks to be parsable. We also check here whether we need to nudge the 2183 // CMS collector thread to start a new cycle (if it's not already active). 2184 assert( Thread::current()->is_VM_thread() 2185 || ( CMSScavengeBeforeRemark 2186 && Thread::current()->is_ConcurrentGC_thread()), 2187 "Incorrect thread type for epilogue execution"); 2188 2189 if (!_between_prologue_and_epilogue) { 2190 // We have already been invoked; this is a gc_epilogue delegation 2191 // from yet another CMS generation that we are responsible for, just 2192 // ignore it since all relevant work has already been done. 2193 return; 2194 } 2195 assert(haveFreelistLocks(), "must have freelist locks"); 2196 assert_lock_strong(bitMapLock()); 2197 2198 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2199 2200 _cmsGen->gc_epilogue_work(full); 2201 2202 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2203 // in case sampling was not already enabled, enable it 2204 _start_sampling = true; 2205 } 2206 // reset _eden_chunk_array so sampling starts afresh 2207 _eden_chunk_index = 0; 2208 2209 size_t cms_used = _cmsGen->cmsSpace()->used(); 2210 2211 // update performance counters - this uses a special version of 2212 // update_counters() that allows the utilization to be passed as a 2213 // parameter, avoiding multiple calls to used(). 2214 // 2215 _cmsGen->update_counters(cms_used); 2216 2217 bitMapLock()->unlock(); 2218 releaseFreelistLocks(); 2219 2220 if (!CleanChunkPoolAsync) { 2221 Chunk::clean_chunk_pool(); 2222 } 2223 2224 set_did_compact(false); 2225 _between_prologue_and_epilogue = false; // ready for next cycle 2226 } 2227 2228 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2229 collector()->gc_epilogue(full); 2230 2231 // Also reset promotion tracking in par gc thread states. 2232 for (uint i = 0; i < ParallelGCThreads; i++) { 2233 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2234 } 2235 } 2236 2237 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2238 assert(!incremental_collection_failed(), "Should have been cleared"); 2239 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2240 cmsSpace()->gc_epilogue(); 2241 // Print stat counters 2242 NOT_PRODUCT( 2243 assert(_numObjectsAllocated == 0, "check"); 2244 assert(_numWordsAllocated == 0, "check"); 2245 if (Verbose && PrintGC) { 2246 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2247 SIZE_FORMAT" bytes", 2248 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2249 } 2250 _numObjectsPromoted = 0; 2251 _numWordsPromoted = 0; 2252 ) 2253 2254 if (PrintGC && Verbose) { 2255 // Call down the chain in contiguous_available needs the freelistLock 2256 // so print this out before releasing the freeListLock. 2257 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2258 contiguous_available()); 2259 } 2260 } 2261 2262 #ifndef PRODUCT 2263 bool CMSCollector::have_cms_token() { 2264 Thread* thr = Thread::current(); 2265 if (thr->is_VM_thread()) { 2266 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2267 } else if (thr->is_ConcurrentGC_thread()) { 2268 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2269 } else if (thr->is_GC_task_thread()) { 2270 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2271 ParGCRareEvent_lock->owned_by_self(); 2272 } 2273 return false; 2274 } 2275 #endif 2276 2277 // Check reachability of the given heap address in CMS generation, 2278 // treating all other generations as roots. 2279 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2280 // We could "guarantee" below, rather than assert, but I'll 2281 // leave these as "asserts" so that an adventurous debugger 2282 // could try this in the product build provided some subset of 2283 // the conditions were met, provided they were interested in the 2284 // results and knew that the computation below wouldn't interfere 2285 // with other concurrent computations mutating the structures 2286 // being read or written. 2287 assert(SafepointSynchronize::is_at_safepoint(), 2288 "Else mutations in object graph will make answer suspect"); 2289 assert(have_cms_token(), "Should hold cms token"); 2290 assert(haveFreelistLocks(), "must hold free list locks"); 2291 assert_lock_strong(bitMapLock()); 2292 2293 // Clear the marking bit map array before starting, but, just 2294 // for kicks, first report if the given address is already marked 2295 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2296 _markBitMap.isMarked(addr) ? "" : " not"); 2297 2298 if (verify_after_remark()) { 2299 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2300 bool result = verification_mark_bm()->isMarked(addr); 2301 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2302 result ? "IS" : "is NOT"); 2303 return result; 2304 } else { 2305 gclog_or_tty->print_cr("Could not compute result"); 2306 return false; 2307 } 2308 } 2309 2310 2311 void 2312 CMSCollector::print_on_error(outputStream* st) { 2313 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2314 if (collector != NULL) { 2315 CMSBitMap* bitmap = &collector->_markBitMap; 2316 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2317 bitmap->print_on_error(st, " Bits: "); 2318 2319 st->cr(); 2320 2321 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2322 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2323 mut_bitmap->print_on_error(st, " Bits: "); 2324 } 2325 } 2326 2327 //////////////////////////////////////////////////////// 2328 // CMS Verification Support 2329 //////////////////////////////////////////////////////// 2330 // Following the remark phase, the following invariant 2331 // should hold -- each object in the CMS heap which is 2332 // marked in markBitMap() should be marked in the verification_mark_bm(). 2333 2334 class VerifyMarkedClosure: public BitMapClosure { 2335 CMSBitMap* _marks; 2336 bool _failed; 2337 2338 public: 2339 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2340 2341 bool do_bit(size_t offset) { 2342 HeapWord* addr = _marks->offsetToHeapWord(offset); 2343 if (!_marks->isMarked(addr)) { 2344 oop(addr)->print_on(gclog_or_tty); 2345 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2346 _failed = true; 2347 } 2348 return true; 2349 } 2350 2351 bool failed() { return _failed; } 2352 }; 2353 2354 bool CMSCollector::verify_after_remark(bool silent) { 2355 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2356 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2357 static bool init = false; 2358 2359 assert(SafepointSynchronize::is_at_safepoint(), 2360 "Else mutations in object graph will make answer suspect"); 2361 assert(have_cms_token(), 2362 "Else there may be mutual interference in use of " 2363 " verification data structures"); 2364 assert(_collectorState > Marking && _collectorState <= Sweeping, 2365 "Else marking info checked here may be obsolete"); 2366 assert(haveFreelistLocks(), "must hold free list locks"); 2367 assert_lock_strong(bitMapLock()); 2368 2369 2370 // Allocate marking bit map if not already allocated 2371 if (!init) { // first time 2372 if (!verification_mark_bm()->allocate(_span)) { 2373 return false; 2374 } 2375 init = true; 2376 } 2377 2378 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2379 2380 // Turn off refs discovery -- so we will be tracing through refs. 2381 // This is as intended, because by this time 2382 // GC must already have cleared any refs that need to be cleared, 2383 // and traced those that need to be marked; moreover, 2384 // the marking done here is not going to interfere in any 2385 // way with the marking information used by GC. 2386 NoRefDiscovery no_discovery(ref_processor()); 2387 2388 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2389 2390 // Clear any marks from a previous round 2391 verification_mark_bm()->clear_all(); 2392 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2393 verify_work_stacks_empty(); 2394 2395 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2396 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2397 // Update the saved marks which may affect the root scans. 2398 gch->save_marks(); 2399 2400 if (CMSRemarkVerifyVariant == 1) { 2401 // In this first variant of verification, we complete 2402 // all marking, then check if the new marks-vector is 2403 // a subset of the CMS marks-vector. 2404 verify_after_remark_work_1(); 2405 } else if (CMSRemarkVerifyVariant == 2) { 2406 // In this second variant of verification, we flag an error 2407 // (i.e. an object reachable in the new marks-vector not reachable 2408 // in the CMS marks-vector) immediately, also indicating the 2409 // identify of an object (A) that references the unmarked object (B) -- 2410 // presumably, a mutation to A failed to be picked up by preclean/remark? 2411 verify_after_remark_work_2(); 2412 } else { 2413 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2414 CMSRemarkVerifyVariant); 2415 } 2416 if (!silent) gclog_or_tty->print(" done] "); 2417 return true; 2418 } 2419 2420 void CMSCollector::verify_after_remark_work_1() { 2421 ResourceMark rm; 2422 HandleMark hm; 2423 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2424 2425 // Get a clear set of claim bits for the roots processing to work with. 2426 ClassLoaderDataGraph::clear_claimed_marks(); 2427 2428 // Mark from roots one level into CMS 2429 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2430 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2431 2432 gch->gen_process_roots(_cmsGen->level(), 2433 true, // younger gens are roots 2434 true, // activate StrongRootsScope 2435 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2436 should_unload_classes(), 2437 ¬Older, 2438 NULL, 2439 NULL); // SSS: Provide correct closure 2440 2441 // Now mark from the roots 2442 MarkFromRootsClosure markFromRootsClosure(this, _span, 2443 verification_mark_bm(), verification_mark_stack(), 2444 false /* don't yield */, true /* verifying */); 2445 assert(_restart_addr == NULL, "Expected pre-condition"); 2446 verification_mark_bm()->iterate(&markFromRootsClosure); 2447 while (_restart_addr != NULL) { 2448 // Deal with stack overflow: by restarting at the indicated 2449 // address. 2450 HeapWord* ra = _restart_addr; 2451 markFromRootsClosure.reset(ra); 2452 _restart_addr = NULL; 2453 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2454 } 2455 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2456 verify_work_stacks_empty(); 2457 2458 // Marking completed -- now verify that each bit marked in 2459 // verification_mark_bm() is also marked in markBitMap(); flag all 2460 // errors by printing corresponding objects. 2461 VerifyMarkedClosure vcl(markBitMap()); 2462 verification_mark_bm()->iterate(&vcl); 2463 if (vcl.failed()) { 2464 gclog_or_tty->print("Verification failed"); 2465 gch->print_on(gclog_or_tty); 2466 fatal("CMS: failed marking verification after remark"); 2467 } 2468 } 2469 2470 class VerifyKlassOopsKlassClosure : public KlassClosure { 2471 class VerifyKlassOopsClosure : public OopClosure { 2472 CMSBitMap* _bitmap; 2473 public: 2474 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2475 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2476 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2477 } _oop_closure; 2478 public: 2479 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2480 void do_klass(Klass* k) { 2481 k->oops_do(&_oop_closure); 2482 } 2483 }; 2484 2485 void CMSCollector::verify_after_remark_work_2() { 2486 ResourceMark rm; 2487 HandleMark hm; 2488 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2489 2490 // Get a clear set of claim bits for the roots processing to work with. 2491 ClassLoaderDataGraph::clear_claimed_marks(); 2492 2493 // Mark from roots one level into CMS 2494 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2495 markBitMap()); 2496 CLDToOopClosure cld_closure(¬Older, true); 2497 2498 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2499 2500 gch->gen_process_roots(_cmsGen->level(), 2501 true, // younger gens are roots 2502 true, // activate StrongRootsScope 2503 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2504 should_unload_classes(), 2505 ¬Older, 2506 NULL, 2507 &cld_closure); 2508 2509 // Now mark from the roots 2510 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2511 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2512 assert(_restart_addr == NULL, "Expected pre-condition"); 2513 verification_mark_bm()->iterate(&markFromRootsClosure); 2514 while (_restart_addr != NULL) { 2515 // Deal with stack overflow: by restarting at the indicated 2516 // address. 2517 HeapWord* ra = _restart_addr; 2518 markFromRootsClosure.reset(ra); 2519 _restart_addr = NULL; 2520 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2521 } 2522 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2523 verify_work_stacks_empty(); 2524 2525 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2526 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2527 2528 // Marking completed -- now verify that each bit marked in 2529 // verification_mark_bm() is also marked in markBitMap(); flag all 2530 // errors by printing corresponding objects. 2531 VerifyMarkedClosure vcl(markBitMap()); 2532 verification_mark_bm()->iterate(&vcl); 2533 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2534 } 2535 2536 void ConcurrentMarkSweepGeneration::save_marks() { 2537 // delegate to CMS space 2538 cmsSpace()->save_marks(); 2539 for (uint i = 0; i < ParallelGCThreads; i++) { 2540 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2541 } 2542 } 2543 2544 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2545 return cmsSpace()->no_allocs_since_save_marks(); 2546 } 2547 2548 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2549 \ 2550 void ConcurrentMarkSweepGeneration:: \ 2551 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2552 cl->set_generation(this); \ 2553 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2554 cl->reset_generation(); \ 2555 save_marks(); \ 2556 } 2557 2558 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2559 2560 void 2561 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2562 if (freelistLock()->owned_by_self()) { 2563 Generation::oop_iterate(cl); 2564 } else { 2565 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2566 Generation::oop_iterate(cl); 2567 } 2568 } 2569 2570 void 2571 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2572 if (freelistLock()->owned_by_self()) { 2573 Generation::object_iterate(cl); 2574 } else { 2575 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2576 Generation::object_iterate(cl); 2577 } 2578 } 2579 2580 void 2581 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2582 if (freelistLock()->owned_by_self()) { 2583 Generation::safe_object_iterate(cl); 2584 } else { 2585 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2586 Generation::safe_object_iterate(cl); 2587 } 2588 } 2589 2590 void 2591 ConcurrentMarkSweepGeneration::post_compact() { 2592 } 2593 2594 void 2595 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2596 // Fix the linear allocation blocks to look like free blocks. 2597 2598 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2599 // are not called when the heap is verified during universe initialization and 2600 // at vm shutdown. 2601 if (freelistLock()->owned_by_self()) { 2602 cmsSpace()->prepare_for_verify(); 2603 } else { 2604 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2605 cmsSpace()->prepare_for_verify(); 2606 } 2607 } 2608 2609 void 2610 ConcurrentMarkSweepGeneration::verify() { 2611 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2612 // are not called when the heap is verified during universe initialization and 2613 // at vm shutdown. 2614 if (freelistLock()->owned_by_self()) { 2615 cmsSpace()->verify(); 2616 } else { 2617 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2618 cmsSpace()->verify(); 2619 } 2620 } 2621 2622 void CMSCollector::verify() { 2623 _cmsGen->verify(); 2624 } 2625 2626 #ifndef PRODUCT 2627 bool CMSCollector::overflow_list_is_empty() const { 2628 assert(_num_par_pushes >= 0, "Inconsistency"); 2629 if (_overflow_list == NULL) { 2630 assert(_num_par_pushes == 0, "Inconsistency"); 2631 } 2632 return _overflow_list == NULL; 2633 } 2634 2635 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2636 // merely consolidate assertion checks that appear to occur together frequently. 2637 void CMSCollector::verify_work_stacks_empty() const { 2638 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2639 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2640 } 2641 2642 void CMSCollector::verify_overflow_empty() const { 2643 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2644 assert(no_preserved_marks(), "No preserved marks"); 2645 } 2646 #endif // PRODUCT 2647 2648 // Decide if we want to enable class unloading as part of the 2649 // ensuing concurrent GC cycle. We will collect and 2650 // unload classes if it's the case that: 2651 // (1) an explicit gc request has been made and the flag 2652 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2653 // (2) (a) class unloading is enabled at the command line, and 2654 // (b) old gen is getting really full 2655 // NOTE: Provided there is no change in the state of the heap between 2656 // calls to this method, it should have idempotent results. Moreover, 2657 // its results should be monotonically increasing (i.e. going from 0 to 1, 2658 // but not 1 to 0) between successive calls between which the heap was 2659 // not collected. For the implementation below, it must thus rely on 2660 // the property that concurrent_cycles_since_last_unload() 2661 // will not decrease unless a collection cycle happened and that 2662 // _cmsGen->is_too_full() are 2663 // themselves also monotonic in that sense. See check_monotonicity() 2664 // below. 2665 void CMSCollector::update_should_unload_classes() { 2666 _should_unload_classes = false; 2667 // Condition 1 above 2668 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2669 _should_unload_classes = true; 2670 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2671 // Disjuncts 2.b.(i,ii,iii) above 2672 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2673 CMSClassUnloadingMaxInterval) 2674 || _cmsGen->is_too_full(); 2675 } 2676 } 2677 2678 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2679 bool res = should_concurrent_collect(); 2680 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2681 return res; 2682 } 2683 2684 void CMSCollector::setup_cms_unloading_and_verification_state() { 2685 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2686 || VerifyBeforeExit; 2687 const int rso = GenCollectedHeap::SO_AllCodeCache; 2688 2689 // We set the proper root for this CMS cycle here. 2690 if (should_unload_classes()) { // Should unload classes this cycle 2691 remove_root_scanning_option(rso); // Shrink the root set appropriately 2692 set_verifying(should_verify); // Set verification state for this cycle 2693 return; // Nothing else needs to be done at this time 2694 } 2695 2696 // Not unloading classes this cycle 2697 assert(!should_unload_classes(), "Inconsistency!"); 2698 2699 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2700 // Include symbols, strings and code cache elements to prevent their resurrection. 2701 add_root_scanning_option(rso); 2702 set_verifying(true); 2703 } else if (verifying() && !should_verify) { 2704 // We were verifying, but some verification flags got disabled. 2705 set_verifying(false); 2706 // Exclude symbols, strings and code cache elements from root scanning to 2707 // reduce IM and RM pauses. 2708 remove_root_scanning_option(rso); 2709 } 2710 } 2711 2712 2713 #ifndef PRODUCT 2714 HeapWord* CMSCollector::block_start(const void* p) const { 2715 const HeapWord* addr = (HeapWord*)p; 2716 if (_span.contains(p)) { 2717 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2718 return _cmsGen->cmsSpace()->block_start(p); 2719 } 2720 } 2721 return NULL; 2722 } 2723 #endif 2724 2725 HeapWord* 2726 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2727 bool tlab, 2728 bool parallel) { 2729 CMSSynchronousYieldRequest yr; 2730 assert(!tlab, "Can't deal with TLAB allocation"); 2731 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2732 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2733 if (GCExpandToAllocateDelayMillis > 0) { 2734 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2735 } 2736 return have_lock_and_allocate(word_size, tlab); 2737 } 2738 2739 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2740 size_t bytes, 2741 size_t expand_bytes, 2742 CMSExpansionCause::Cause cause) 2743 { 2744 2745 bool success = expand(bytes, expand_bytes); 2746 2747 // remember why we expanded; this information is used 2748 // by shouldConcurrentCollect() when making decisions on whether to start 2749 // a new CMS cycle. 2750 if (success) { 2751 set_expansion_cause(cause); 2752 if (PrintGCDetails && Verbose) { 2753 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2754 CMSExpansionCause::to_string(cause)); 2755 } 2756 } 2757 } 2758 2759 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2760 HeapWord* res = NULL; 2761 MutexLocker x(ParGCRareEvent_lock); 2762 while (true) { 2763 // Expansion by some other thread might make alloc OK now: 2764 res = ps->lab.alloc(word_sz); 2765 if (res != NULL) return res; 2766 // If there's not enough expansion space available, give up. 2767 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2768 return NULL; 2769 } 2770 // Otherwise, we try expansion. 2771 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2772 // Now go around the loop and try alloc again; 2773 // A competing par_promote might beat us to the expansion space, 2774 // so we may go around the loop again if promotion fails again. 2775 if (GCExpandToAllocateDelayMillis > 0) { 2776 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2777 } 2778 } 2779 } 2780 2781 2782 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2783 PromotionInfo* promo) { 2784 MutexLocker x(ParGCRareEvent_lock); 2785 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2786 while (true) { 2787 // Expansion by some other thread might make alloc OK now: 2788 if (promo->ensure_spooling_space()) { 2789 assert(promo->has_spooling_space(), 2790 "Post-condition of successful ensure_spooling_space()"); 2791 return true; 2792 } 2793 // If there's not enough expansion space available, give up. 2794 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2795 return false; 2796 } 2797 // Otherwise, we try expansion. 2798 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2799 // Now go around the loop and try alloc again; 2800 // A competing allocation might beat us to the expansion space, 2801 // so we may go around the loop again if allocation fails again. 2802 if (GCExpandToAllocateDelayMillis > 0) { 2803 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2804 } 2805 } 2806 } 2807 2808 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2809 // Only shrink if a compaction was done so that all the free space 2810 // in the generation is in a contiguous block at the end. 2811 if (did_compact()) { 2812 CardGeneration::shrink(bytes); 2813 } 2814 } 2815 2816 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2817 assert_locked_or_safepoint(Heap_lock); 2818 } 2819 2820 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2821 assert_locked_or_safepoint(Heap_lock); 2822 assert_lock_strong(freelistLock()); 2823 if (PrintGCDetails && Verbose) { 2824 warning("Shrinking of CMS not yet implemented"); 2825 } 2826 return; 2827 } 2828 2829 2830 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2831 // phases. 2832 class CMSPhaseAccounting: public StackObj { 2833 public: 2834 CMSPhaseAccounting(CMSCollector *collector, 2835 const char *phase, 2836 const GCId gc_id, 2837 bool print_cr = true); 2838 ~CMSPhaseAccounting(); 2839 2840 private: 2841 CMSCollector *_collector; 2842 const char *_phase; 2843 elapsedTimer _wallclock; 2844 bool _print_cr; 2845 const GCId _gc_id; 2846 2847 public: 2848 // Not MT-safe; so do not pass around these StackObj's 2849 // where they may be accessed by other threads. 2850 jlong wallclock_millis() { 2851 assert(_wallclock.is_active(), "Wall clock should not stop"); 2852 _wallclock.stop(); // to record time 2853 jlong ret = _wallclock.milliseconds(); 2854 _wallclock.start(); // restart 2855 return ret; 2856 } 2857 }; 2858 2859 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2860 const char *phase, 2861 const GCId gc_id, 2862 bool print_cr) : 2863 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2864 2865 if (PrintCMSStatistics != 0) { 2866 _collector->resetYields(); 2867 } 2868 if (PrintGCDetails) { 2869 gclog_or_tty->gclog_stamp(_gc_id); 2870 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2871 _collector->cmsGen()->short_name(), _phase); 2872 } 2873 _collector->resetTimer(); 2874 _wallclock.start(); 2875 _collector->startTimer(); 2876 } 2877 2878 CMSPhaseAccounting::~CMSPhaseAccounting() { 2879 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2880 _collector->stopTimer(); 2881 _wallclock.stop(); 2882 if (PrintGCDetails) { 2883 gclog_or_tty->gclog_stamp(_gc_id); 2884 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2885 _collector->cmsGen()->short_name(), 2886 _phase, _collector->timerValue(), _wallclock.seconds()); 2887 if (_print_cr) { 2888 gclog_or_tty->cr(); 2889 } 2890 if (PrintCMSStatistics != 0) { 2891 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2892 _collector->yields()); 2893 } 2894 } 2895 } 2896 2897 // CMS work 2898 2899 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2900 class CMSParMarkTask : public AbstractGangTask { 2901 protected: 2902 CMSCollector* _collector; 2903 int _n_workers; 2904 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 2905 AbstractGangTask(name), 2906 _collector(collector), 2907 _n_workers(n_workers) {} 2908 // Work method in support of parallel rescan ... of young gen spaces 2909 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2910 ContiguousSpace* space, 2911 HeapWord** chunk_array, size_t chunk_top); 2912 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2913 }; 2914 2915 // Parallel initial mark task 2916 class CMSParInitialMarkTask: public CMSParMarkTask { 2917 public: 2918 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 2919 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2920 collector, n_workers) {} 2921 void work(uint worker_id); 2922 }; 2923 2924 // Checkpoint the roots into this generation from outside 2925 // this generation. [Note this initial checkpoint need only 2926 // be approximate -- we'll do a catch up phase subsequently.] 2927 void CMSCollector::checkpointRootsInitial() { 2928 assert(_collectorState == InitialMarking, "Wrong collector state"); 2929 check_correct_thread_executing(); 2930 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2931 2932 save_heap_summary(); 2933 report_heap_summary(GCWhen::BeforeGC); 2934 2935 ReferenceProcessor* rp = ref_processor(); 2936 assert(_restart_addr == NULL, "Control point invariant"); 2937 { 2938 // acquire locks for subsequent manipulations 2939 MutexLockerEx x(bitMapLock(), 2940 Mutex::_no_safepoint_check_flag); 2941 checkpointRootsInitialWork(); 2942 // enable ("weak") refs discovery 2943 rp->enable_discovery(); 2944 _collectorState = Marking; 2945 } 2946 } 2947 2948 void CMSCollector::checkpointRootsInitialWork() { 2949 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2950 assert(_collectorState == InitialMarking, "just checking"); 2951 2952 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2953 // precede our marking with a collection of all 2954 // younger generations to keep floating garbage to a minimum. 2955 // XXX: we won't do this for now -- it's an optimization to be done later. 2956 2957 // already have locks 2958 assert_lock_strong(bitMapLock()); 2959 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2960 2961 // Setup the verification and class unloading state for this 2962 // CMS collection cycle. 2963 setup_cms_unloading_and_verification_state(); 2964 2965 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2966 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2967 2968 // Reset all the PLAB chunk arrays if necessary. 2969 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2970 reset_survivor_plab_arrays(); 2971 } 2972 2973 ResourceMark rm; 2974 HandleMark hm; 2975 2976 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2977 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2978 2979 verify_work_stacks_empty(); 2980 verify_overflow_empty(); 2981 2982 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2983 // Update the saved marks which may affect the root scans. 2984 gch->save_marks(); 2985 2986 // weak reference processing has not started yet. 2987 ref_processor()->set_enqueuing_is_done(false); 2988 2989 // Need to remember all newly created CLDs, 2990 // so that we can guarantee that the remark finds them. 2991 ClassLoaderDataGraph::remember_new_clds(true); 2992 2993 // Whenever a CLD is found, it will be claimed before proceeding to mark 2994 // the klasses. The claimed marks need to be cleared before marking starts. 2995 ClassLoaderDataGraph::clear_claimed_marks(); 2996 2997 if (CMSPrintEdenSurvivorChunks) { 2998 print_eden_and_survivor_chunk_arrays(); 2999 } 3000 3001 { 3002 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3003 if (CMSParallelInitialMarkEnabled) { 3004 // The parallel version. 3005 FlexibleWorkGang* workers = gch->workers(); 3006 assert(workers != NULL, "Need parallel worker threads."); 3007 int n_workers = workers->active_workers(); 3008 CMSParInitialMarkTask tsk(this, n_workers); 3009 gch->set_par_threads(n_workers); 3010 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3011 if (n_workers > 1) { 3012 StrongRootsScope srs; 3013 workers->run_task(&tsk); 3014 } else { 3015 StrongRootsScope srs; 3016 tsk.work(0); 3017 } 3018 gch->set_par_threads(0); 3019 } else { 3020 // The serial version. 3021 CLDToOopClosure cld_closure(¬Older, true); 3022 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3023 gch->gen_process_roots(_cmsGen->level(), 3024 true, // younger gens are roots 3025 true, // activate StrongRootsScope 3026 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3027 should_unload_classes(), 3028 ¬Older, 3029 NULL, 3030 &cld_closure); 3031 } 3032 } 3033 3034 // Clear mod-union table; it will be dirtied in the prologue of 3035 // CMS generation per each younger generation collection. 3036 3037 assert(_modUnionTable.isAllClear(), 3038 "Was cleared in most recent final checkpoint phase" 3039 " or no bits are set in the gc_prologue before the start of the next " 3040 "subsequent marking phase."); 3041 3042 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3043 3044 // Save the end of the used_region of the constituent generations 3045 // to be used to limit the extent of sweep in each generation. 3046 save_sweep_limits(); 3047 verify_overflow_empty(); 3048 } 3049 3050 bool CMSCollector::markFromRoots() { 3051 // we might be tempted to assert that: 3052 // assert(!SafepointSynchronize::is_at_safepoint(), 3053 // "inconsistent argument?"); 3054 // However that wouldn't be right, because it's possible that 3055 // a safepoint is indeed in progress as a younger generation 3056 // stop-the-world GC happens even as we mark in this generation. 3057 assert(_collectorState == Marking, "inconsistent state?"); 3058 check_correct_thread_executing(); 3059 verify_overflow_empty(); 3060 3061 // Weak ref discovery note: We may be discovering weak 3062 // refs in this generation concurrent (but interleaved) with 3063 // weak ref discovery by a younger generation collector. 3064 3065 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3066 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3067 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3068 bool res = markFromRootsWork(); 3069 if (res) { 3070 _collectorState = Precleaning; 3071 } else { // We failed and a foreground collection wants to take over 3072 assert(_foregroundGCIsActive, "internal state inconsistency"); 3073 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3074 if (PrintGCDetails) { 3075 gclog_or_tty->print_cr("bailing out to foreground collection"); 3076 } 3077 } 3078 verify_overflow_empty(); 3079 return res; 3080 } 3081 3082 bool CMSCollector::markFromRootsWork() { 3083 // iterate over marked bits in bit map, doing a full scan and mark 3084 // from these roots using the following algorithm: 3085 // . if oop is to the right of the current scan pointer, 3086 // mark corresponding bit (we'll process it later) 3087 // . else (oop is to left of current scan pointer) 3088 // push oop on marking stack 3089 // . drain the marking stack 3090 3091 // Note that when we do a marking step we need to hold the 3092 // bit map lock -- recall that direct allocation (by mutators) 3093 // and promotion (by younger generation collectors) is also 3094 // marking the bit map. [the so-called allocate live policy.] 3095 // Because the implementation of bit map marking is not 3096 // robust wrt simultaneous marking of bits in the same word, 3097 // we need to make sure that there is no such interference 3098 // between concurrent such updates. 3099 3100 // already have locks 3101 assert_lock_strong(bitMapLock()); 3102 3103 verify_work_stacks_empty(); 3104 verify_overflow_empty(); 3105 bool result = false; 3106 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3107 result = do_marking_mt(); 3108 } else { 3109 result = do_marking_st(); 3110 } 3111 return result; 3112 } 3113 3114 // Forward decl 3115 class CMSConcMarkingTask; 3116 3117 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3118 CMSCollector* _collector; 3119 CMSConcMarkingTask* _task; 3120 public: 3121 virtual void yield(); 3122 3123 // "n_threads" is the number of threads to be terminated. 3124 // "queue_set" is a set of work queues of other threads. 3125 // "collector" is the CMS collector associated with this task terminator. 3126 // "yield" indicates whether we need the gang as a whole to yield. 3127 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3128 ParallelTaskTerminator(n_threads, queue_set), 3129 _collector(collector) { } 3130 3131 void set_task(CMSConcMarkingTask* task) { 3132 _task = task; 3133 } 3134 }; 3135 3136 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3137 CMSConcMarkingTask* _task; 3138 public: 3139 bool should_exit_termination(); 3140 void set_task(CMSConcMarkingTask* task) { 3141 _task = task; 3142 } 3143 }; 3144 3145 // MT Concurrent Marking Task 3146 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3147 CMSCollector* _collector; 3148 int _n_workers; // requested/desired # workers 3149 bool _result; 3150 CompactibleFreeListSpace* _cms_space; 3151 char _pad_front[64]; // padding to ... 3152 HeapWord* _global_finger; // ... avoid sharing cache line 3153 char _pad_back[64]; 3154 HeapWord* _restart_addr; 3155 3156 // Exposed here for yielding support 3157 Mutex* const _bit_map_lock; 3158 3159 // The per thread work queues, available here for stealing 3160 OopTaskQueueSet* _task_queues; 3161 3162 // Termination (and yielding) support 3163 CMSConcMarkingTerminator _term; 3164 CMSConcMarkingTerminatorTerminator _term_term; 3165 3166 public: 3167 CMSConcMarkingTask(CMSCollector* collector, 3168 CompactibleFreeListSpace* cms_space, 3169 YieldingFlexibleWorkGang* workers, 3170 OopTaskQueueSet* task_queues): 3171 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3172 _collector(collector), 3173 _cms_space(cms_space), 3174 _n_workers(0), _result(true), 3175 _task_queues(task_queues), 3176 _term(_n_workers, task_queues, _collector), 3177 _bit_map_lock(collector->bitMapLock()) 3178 { 3179 _requested_size = _n_workers; 3180 _term.set_task(this); 3181 _term_term.set_task(this); 3182 _restart_addr = _global_finger = _cms_space->bottom(); 3183 } 3184 3185 3186 OopTaskQueueSet* task_queues() { return _task_queues; } 3187 3188 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3189 3190 HeapWord** global_finger_addr() { return &_global_finger; } 3191 3192 CMSConcMarkingTerminator* terminator() { return &_term; } 3193 3194 virtual void set_for_termination(int active_workers) { 3195 terminator()->reset_for_reuse(active_workers); 3196 } 3197 3198 void work(uint worker_id); 3199 bool should_yield() { 3200 return ConcurrentMarkSweepThread::should_yield() 3201 && !_collector->foregroundGCIsActive(); 3202 } 3203 3204 virtual void coordinator_yield(); // stuff done by coordinator 3205 bool result() { return _result; } 3206 3207 void reset(HeapWord* ra) { 3208 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3209 _restart_addr = _global_finger = ra; 3210 _term.reset_for_reuse(); 3211 } 3212 3213 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3214 OopTaskQueue* work_q); 3215 3216 private: 3217 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3218 void do_work_steal(int i); 3219 void bump_global_finger(HeapWord* f); 3220 }; 3221 3222 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3223 assert(_task != NULL, "Error"); 3224 return _task->yielding(); 3225 // Note that we do not need the disjunct || _task->should_yield() above 3226 // because we want terminating threads to yield only if the task 3227 // is already in the midst of yielding, which happens only after at least one 3228 // thread has yielded. 3229 } 3230 3231 void CMSConcMarkingTerminator::yield() { 3232 if (_task->should_yield()) { 3233 _task->yield(); 3234 } else { 3235 ParallelTaskTerminator::yield(); 3236 } 3237 } 3238 3239 //////////////////////////////////////////////////////////////// 3240 // Concurrent Marking Algorithm Sketch 3241 //////////////////////////////////////////////////////////////// 3242 // Until all tasks exhausted (both spaces): 3243 // -- claim next available chunk 3244 // -- bump global finger via CAS 3245 // -- find first object that starts in this chunk 3246 // and start scanning bitmap from that position 3247 // -- scan marked objects for oops 3248 // -- CAS-mark target, and if successful: 3249 // . if target oop is above global finger (volatile read) 3250 // nothing to do 3251 // . if target oop is in chunk and above local finger 3252 // then nothing to do 3253 // . else push on work-queue 3254 // -- Deal with possible overflow issues: 3255 // . local work-queue overflow causes stuff to be pushed on 3256 // global (common) overflow queue 3257 // . always first empty local work queue 3258 // . then get a batch of oops from global work queue if any 3259 // . then do work stealing 3260 // -- When all tasks claimed (both spaces) 3261 // and local work queue empty, 3262 // then in a loop do: 3263 // . check global overflow stack; steal a batch of oops and trace 3264 // . try to steal from other threads oif GOS is empty 3265 // . if neither is available, offer termination 3266 // -- Terminate and return result 3267 // 3268 void CMSConcMarkingTask::work(uint worker_id) { 3269 elapsedTimer _timer; 3270 ResourceMark rm; 3271 HandleMark hm; 3272 3273 DEBUG_ONLY(_collector->verify_overflow_empty();) 3274 3275 // Before we begin work, our work queue should be empty 3276 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3277 // Scan the bitmap covering _cms_space, tracing through grey objects. 3278 _timer.start(); 3279 do_scan_and_mark(worker_id, _cms_space); 3280 _timer.stop(); 3281 if (PrintCMSStatistics != 0) { 3282 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3283 worker_id, _timer.seconds()); 3284 // XXX: need xxx/xxx type of notation, two timers 3285 } 3286 3287 // ... do work stealing 3288 _timer.reset(); 3289 _timer.start(); 3290 do_work_steal(worker_id); 3291 _timer.stop(); 3292 if (PrintCMSStatistics != 0) { 3293 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3294 worker_id, _timer.seconds()); 3295 // XXX: need xxx/xxx type of notation, two timers 3296 } 3297 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3298 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3299 // Note that under the current task protocol, the 3300 // following assertion is true even of the spaces 3301 // expanded since the completion of the concurrent 3302 // marking. XXX This will likely change under a strict 3303 // ABORT semantics. 3304 // After perm removal the comparison was changed to 3305 // greater than or equal to from strictly greater than. 3306 // Before perm removal the highest address sweep would 3307 // have been at the end of perm gen but now is at the 3308 // end of the tenured gen. 3309 assert(_global_finger >= _cms_space->end(), 3310 "All tasks have been completed"); 3311 DEBUG_ONLY(_collector->verify_overflow_empty();) 3312 } 3313 3314 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3315 HeapWord* read = _global_finger; 3316 HeapWord* cur = read; 3317 while (f > read) { 3318 cur = read; 3319 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3320 if (cur == read) { 3321 // our cas succeeded 3322 assert(_global_finger >= f, "protocol consistency"); 3323 break; 3324 } 3325 } 3326 } 3327 3328 // This is really inefficient, and should be redone by 3329 // using (not yet available) block-read and -write interfaces to the 3330 // stack and the work_queue. XXX FIX ME !!! 3331 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3332 OopTaskQueue* work_q) { 3333 // Fast lock-free check 3334 if (ovflw_stk->length() == 0) { 3335 return false; 3336 } 3337 assert(work_q->size() == 0, "Shouldn't steal"); 3338 MutexLockerEx ml(ovflw_stk->par_lock(), 3339 Mutex::_no_safepoint_check_flag); 3340 // Grab up to 1/4 the size of the work queue 3341 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3342 (size_t)ParGCDesiredObjsFromOverflowList); 3343 num = MIN2(num, ovflw_stk->length()); 3344 for (int i = (int) num; i > 0; i--) { 3345 oop cur = ovflw_stk->pop(); 3346 assert(cur != NULL, "Counted wrong?"); 3347 work_q->push(cur); 3348 } 3349 return num > 0; 3350 } 3351 3352 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3353 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3354 int n_tasks = pst->n_tasks(); 3355 // We allow that there may be no tasks to do here because 3356 // we are restarting after a stack overflow. 3357 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3358 uint nth_task = 0; 3359 3360 HeapWord* aligned_start = sp->bottom(); 3361 if (sp->used_region().contains(_restart_addr)) { 3362 // Align down to a card boundary for the start of 0th task 3363 // for this space. 3364 aligned_start = 3365 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3366 CardTableModRefBS::card_size); 3367 } 3368 3369 size_t chunk_size = sp->marking_task_size(); 3370 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3371 // Having claimed the nth task in this space, 3372 // compute the chunk that it corresponds to: 3373 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3374 aligned_start + (nth_task+1)*chunk_size); 3375 // Try and bump the global finger via a CAS; 3376 // note that we need to do the global finger bump 3377 // _before_ taking the intersection below, because 3378 // the task corresponding to that region will be 3379 // deemed done even if the used_region() expands 3380 // because of allocation -- as it almost certainly will 3381 // during start-up while the threads yield in the 3382 // closure below. 3383 HeapWord* finger = span.end(); 3384 bump_global_finger(finger); // atomically 3385 // There are null tasks here corresponding to chunks 3386 // beyond the "top" address of the space. 3387 span = span.intersection(sp->used_region()); 3388 if (!span.is_empty()) { // Non-null task 3389 HeapWord* prev_obj; 3390 assert(!span.contains(_restart_addr) || nth_task == 0, 3391 "Inconsistency"); 3392 if (nth_task == 0) { 3393 // For the 0th task, we'll not need to compute a block_start. 3394 if (span.contains(_restart_addr)) { 3395 // In the case of a restart because of stack overflow, 3396 // we might additionally skip a chunk prefix. 3397 prev_obj = _restart_addr; 3398 } else { 3399 prev_obj = span.start(); 3400 } 3401 } else { 3402 // We want to skip the first object because 3403 // the protocol is to scan any object in its entirety 3404 // that _starts_ in this span; a fortiori, any 3405 // object starting in an earlier span is scanned 3406 // as part of an earlier claimed task. 3407 // Below we use the "careful" version of block_start 3408 // so we do not try to navigate uninitialized objects. 3409 prev_obj = sp->block_start_careful(span.start()); 3410 // Below we use a variant of block_size that uses the 3411 // Printezis bits to avoid waiting for allocated 3412 // objects to become initialized/parsable. 3413 while (prev_obj < span.start()) { 3414 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3415 if (sz > 0) { 3416 prev_obj += sz; 3417 } else { 3418 // In this case we may end up doing a bit of redundant 3419 // scanning, but that appears unavoidable, short of 3420 // locking the free list locks; see bug 6324141. 3421 break; 3422 } 3423 } 3424 } 3425 if (prev_obj < span.end()) { 3426 MemRegion my_span = MemRegion(prev_obj, span.end()); 3427 // Do the marking work within a non-empty span -- 3428 // the last argument to the constructor indicates whether the 3429 // iteration should be incremental with periodic yields. 3430 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3431 &_collector->_markBitMap, 3432 work_queue(i), 3433 &_collector->_markStack); 3434 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3435 } // else nothing to do for this task 3436 } // else nothing to do for this task 3437 } 3438 // We'd be tempted to assert here that since there are no 3439 // more tasks left to claim in this space, the global_finger 3440 // must exceed space->top() and a fortiori space->end(). However, 3441 // that would not quite be correct because the bumping of 3442 // global_finger occurs strictly after the claiming of a task, 3443 // so by the time we reach here the global finger may not yet 3444 // have been bumped up by the thread that claimed the last 3445 // task. 3446 pst->all_tasks_completed(); 3447 } 3448 3449 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3450 private: 3451 CMSCollector* _collector; 3452 CMSConcMarkingTask* _task; 3453 MemRegion _span; 3454 CMSBitMap* _bit_map; 3455 CMSMarkStack* _overflow_stack; 3456 OopTaskQueue* _work_queue; 3457 protected: 3458 DO_OOP_WORK_DEFN 3459 public: 3460 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3461 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3462 MetadataAwareOopClosure(collector->ref_processor()), 3463 _collector(collector), 3464 _task(task), 3465 _span(collector->_span), 3466 _work_queue(work_queue), 3467 _bit_map(bit_map), 3468 _overflow_stack(overflow_stack) 3469 { } 3470 virtual void do_oop(oop* p); 3471 virtual void do_oop(narrowOop* p); 3472 3473 void trim_queue(size_t max); 3474 void handle_stack_overflow(HeapWord* lost); 3475 void do_yield_check() { 3476 if (_task->should_yield()) { 3477 _task->yield(); 3478 } 3479 } 3480 }; 3481 3482 // Grey object scanning during work stealing phase -- 3483 // the salient assumption here is that any references 3484 // that are in these stolen objects being scanned must 3485 // already have been initialized (else they would not have 3486 // been published), so we do not need to check for 3487 // uninitialized objects before pushing here. 3488 void Par_ConcMarkingClosure::do_oop(oop obj) { 3489 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3490 HeapWord* addr = (HeapWord*)obj; 3491 // Check if oop points into the CMS generation 3492 // and is not marked 3493 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3494 // a white object ... 3495 // If we manage to "claim" the object, by being the 3496 // first thread to mark it, then we push it on our 3497 // marking stack 3498 if (_bit_map->par_mark(addr)) { // ... now grey 3499 // push on work queue (grey set) 3500 bool simulate_overflow = false; 3501 NOT_PRODUCT( 3502 if (CMSMarkStackOverflowALot && 3503 _collector->simulate_overflow()) { 3504 // simulate a stack overflow 3505 simulate_overflow = true; 3506 } 3507 ) 3508 if (simulate_overflow || 3509 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3510 // stack overflow 3511 if (PrintCMSStatistics != 0) { 3512 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3513 SIZE_FORMAT, _overflow_stack->capacity()); 3514 } 3515 // We cannot assert that the overflow stack is full because 3516 // it may have been emptied since. 3517 assert(simulate_overflow || 3518 _work_queue->size() == _work_queue->max_elems(), 3519 "Else push should have succeeded"); 3520 handle_stack_overflow(addr); 3521 } 3522 } // Else, some other thread got there first 3523 do_yield_check(); 3524 } 3525 } 3526 3527 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3528 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3529 3530 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3531 while (_work_queue->size() > max) { 3532 oop new_oop; 3533 if (_work_queue->pop_local(new_oop)) { 3534 assert(new_oop->is_oop(), "Should be an oop"); 3535 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3536 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3537 new_oop->oop_iterate(this); // do_oop() above 3538 do_yield_check(); 3539 } 3540 } 3541 } 3542 3543 // Upon stack overflow, we discard (part of) the stack, 3544 // remembering the least address amongst those discarded 3545 // in CMSCollector's _restart_address. 3546 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3547 // We need to do this under a mutex to prevent other 3548 // workers from interfering with the work done below. 3549 MutexLockerEx ml(_overflow_stack->par_lock(), 3550 Mutex::_no_safepoint_check_flag); 3551 // Remember the least grey address discarded 3552 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3553 _collector->lower_restart_addr(ra); 3554 _overflow_stack->reset(); // discard stack contents 3555 _overflow_stack->expand(); // expand the stack if possible 3556 } 3557 3558 3559 void CMSConcMarkingTask::do_work_steal(int i) { 3560 OopTaskQueue* work_q = work_queue(i); 3561 oop obj_to_scan; 3562 CMSBitMap* bm = &(_collector->_markBitMap); 3563 CMSMarkStack* ovflw = &(_collector->_markStack); 3564 int* seed = _collector->hash_seed(i); 3565 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3566 while (true) { 3567 cl.trim_queue(0); 3568 assert(work_q->size() == 0, "Should have been emptied above"); 3569 if (get_work_from_overflow_stack(ovflw, work_q)) { 3570 // Can't assert below because the work obtained from the 3571 // overflow stack may already have been stolen from us. 3572 // assert(work_q->size() > 0, "Work from overflow stack"); 3573 continue; 3574 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3575 assert(obj_to_scan->is_oop(), "Should be an oop"); 3576 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3577 obj_to_scan->oop_iterate(&cl); 3578 } else if (terminator()->offer_termination(&_term_term)) { 3579 assert(work_q->size() == 0, "Impossible!"); 3580 break; 3581 } else if (yielding() || should_yield()) { 3582 yield(); 3583 } 3584 } 3585 } 3586 3587 // This is run by the CMS (coordinator) thread. 3588 void CMSConcMarkingTask::coordinator_yield() { 3589 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3590 "CMS thread should hold CMS token"); 3591 // First give up the locks, then yield, then re-lock 3592 // We should probably use a constructor/destructor idiom to 3593 // do this unlock/lock or modify the MutexUnlocker class to 3594 // serve our purpose. XXX 3595 assert_lock_strong(_bit_map_lock); 3596 _bit_map_lock->unlock(); 3597 ConcurrentMarkSweepThread::desynchronize(true); 3598 _collector->stopTimer(); 3599 if (PrintCMSStatistics != 0) { 3600 _collector->incrementYields(); 3601 } 3602 3603 // It is possible for whichever thread initiated the yield request 3604 // not to get a chance to wake up and take the bitmap lock between 3605 // this thread releasing it and reacquiring it. So, while the 3606 // should_yield() flag is on, let's sleep for a bit to give the 3607 // other thread a chance to wake up. The limit imposed on the number 3608 // of iterations is defensive, to avoid any unforseen circumstances 3609 // putting us into an infinite loop. Since it's always been this 3610 // (coordinator_yield()) method that was observed to cause the 3611 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3612 // which is by default non-zero. For the other seven methods that 3613 // also perform the yield operation, as are using a different 3614 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3615 // can enable the sleeping for those methods too, if necessary. 3616 // See 6442774. 3617 // 3618 // We really need to reconsider the synchronization between the GC 3619 // thread and the yield-requesting threads in the future and we 3620 // should really use wait/notify, which is the recommended 3621 // way of doing this type of interaction. Additionally, we should 3622 // consolidate the eight methods that do the yield operation and they 3623 // are almost identical into one for better maintainability and 3624 // readability. See 6445193. 3625 // 3626 // Tony 2006.06.29 3627 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3628 ConcurrentMarkSweepThread::should_yield() && 3629 !CMSCollector::foregroundGCIsActive(); ++i) { 3630 os::sleep(Thread::current(), 1, false); 3631 } 3632 3633 ConcurrentMarkSweepThread::synchronize(true); 3634 _bit_map_lock->lock_without_safepoint_check(); 3635 _collector->startTimer(); 3636 } 3637 3638 bool CMSCollector::do_marking_mt() { 3639 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3640 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 3641 conc_workers()->total_workers(), 3642 conc_workers()->active_workers(), 3643 Threads::number_of_non_daemon_threads()); 3644 conc_workers()->set_active_workers(num_workers); 3645 3646 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3647 3648 CMSConcMarkingTask tsk(this, 3649 cms_space, 3650 conc_workers(), 3651 task_queues()); 3652 3653 // Since the actual number of workers we get may be different 3654 // from the number we requested above, do we need to do anything different 3655 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3656 // class?? XXX 3657 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3658 3659 // Refs discovery is already non-atomic. 3660 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3661 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3662 conc_workers()->start_task(&tsk); 3663 while (tsk.yielded()) { 3664 tsk.coordinator_yield(); 3665 conc_workers()->continue_task(&tsk); 3666 } 3667 // If the task was aborted, _restart_addr will be non-NULL 3668 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3669 while (_restart_addr != NULL) { 3670 // XXX For now we do not make use of ABORTED state and have not 3671 // yet implemented the right abort semantics (even in the original 3672 // single-threaded CMS case). That needs some more investigation 3673 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3674 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3675 // If _restart_addr is non-NULL, a marking stack overflow 3676 // occurred; we need to do a fresh marking iteration from the 3677 // indicated restart address. 3678 if (_foregroundGCIsActive) { 3679 // We may be running into repeated stack overflows, having 3680 // reached the limit of the stack size, while making very 3681 // slow forward progress. It may be best to bail out and 3682 // let the foreground collector do its job. 3683 // Clear _restart_addr, so that foreground GC 3684 // works from scratch. This avoids the headache of 3685 // a "rescan" which would otherwise be needed because 3686 // of the dirty mod union table & card table. 3687 _restart_addr = NULL; 3688 return false; 3689 } 3690 // Adjust the task to restart from _restart_addr 3691 tsk.reset(_restart_addr); 3692 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3693 _restart_addr); 3694 _restart_addr = NULL; 3695 // Get the workers going again 3696 conc_workers()->start_task(&tsk); 3697 while (tsk.yielded()) { 3698 tsk.coordinator_yield(); 3699 conc_workers()->continue_task(&tsk); 3700 } 3701 } 3702 assert(tsk.completed(), "Inconsistency"); 3703 assert(tsk.result() == true, "Inconsistency"); 3704 return true; 3705 } 3706 3707 bool CMSCollector::do_marking_st() { 3708 ResourceMark rm; 3709 HandleMark hm; 3710 3711 // Temporarily make refs discovery single threaded (non-MT) 3712 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3713 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3714 &_markStack, CMSYield); 3715 // the last argument to iterate indicates whether the iteration 3716 // should be incremental with periodic yields. 3717 _markBitMap.iterate(&markFromRootsClosure); 3718 // If _restart_addr is non-NULL, a marking stack overflow 3719 // occurred; we need to do a fresh iteration from the 3720 // indicated restart address. 3721 while (_restart_addr != NULL) { 3722 if (_foregroundGCIsActive) { 3723 // We may be running into repeated stack overflows, having 3724 // reached the limit of the stack size, while making very 3725 // slow forward progress. It may be best to bail out and 3726 // let the foreground collector do its job. 3727 // Clear _restart_addr, so that foreground GC 3728 // works from scratch. This avoids the headache of 3729 // a "rescan" which would otherwise be needed because 3730 // of the dirty mod union table & card table. 3731 _restart_addr = NULL; 3732 return false; // indicating failure to complete marking 3733 } 3734 // Deal with stack overflow: 3735 // we restart marking from _restart_addr 3736 HeapWord* ra = _restart_addr; 3737 markFromRootsClosure.reset(ra); 3738 _restart_addr = NULL; 3739 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3740 } 3741 return true; 3742 } 3743 3744 void CMSCollector::preclean() { 3745 check_correct_thread_executing(); 3746 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3747 verify_work_stacks_empty(); 3748 verify_overflow_empty(); 3749 _abort_preclean = false; 3750 if (CMSPrecleaningEnabled) { 3751 if (!CMSEdenChunksRecordAlways) { 3752 _eden_chunk_index = 0; 3753 } 3754 size_t used = get_eden_used(); 3755 size_t capacity = get_eden_capacity(); 3756 // Don't start sampling unless we will get sufficiently 3757 // many samples. 3758 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3759 * CMSScheduleRemarkEdenPenetration)) { 3760 _start_sampling = true; 3761 } else { 3762 _start_sampling = false; 3763 } 3764 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3765 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3766 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3767 } 3768 CMSTokenSync x(true); // is cms thread 3769 if (CMSPrecleaningEnabled) { 3770 sample_eden(); 3771 _collectorState = AbortablePreclean; 3772 } else { 3773 _collectorState = FinalMarking; 3774 } 3775 verify_work_stacks_empty(); 3776 verify_overflow_empty(); 3777 } 3778 3779 // Try and schedule the remark such that young gen 3780 // occupancy is CMSScheduleRemarkEdenPenetration %. 3781 void CMSCollector::abortable_preclean() { 3782 check_correct_thread_executing(); 3783 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3784 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3785 3786 // If Eden's current occupancy is below this threshold, 3787 // immediately schedule the remark; else preclean 3788 // past the next scavenge in an effort to 3789 // schedule the pause as described above. By choosing 3790 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3791 // we will never do an actual abortable preclean cycle. 3792 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3793 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3794 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3795 // We need more smarts in the abortable preclean 3796 // loop below to deal with cases where allocation 3797 // in young gen is very very slow, and our precleaning 3798 // is running a losing race against a horde of 3799 // mutators intent on flooding us with CMS updates 3800 // (dirty cards). 3801 // One, admittedly dumb, strategy is to give up 3802 // after a certain number of abortable precleaning loops 3803 // or after a certain maximum time. We want to make 3804 // this smarter in the next iteration. 3805 // XXX FIX ME!!! YSR 3806 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3807 while (!(should_abort_preclean() || 3808 ConcurrentMarkSweepThread::should_terminate())) { 3809 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3810 cumworkdone += workdone; 3811 loops++; 3812 // Voluntarily terminate abortable preclean phase if we have 3813 // been at it for too long. 3814 if ((CMSMaxAbortablePrecleanLoops != 0) && 3815 loops >= CMSMaxAbortablePrecleanLoops) { 3816 if (PrintGCDetails) { 3817 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3818 } 3819 break; 3820 } 3821 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3822 if (PrintGCDetails) { 3823 gclog_or_tty->print(" CMS: abort preclean due to time "); 3824 } 3825 break; 3826 } 3827 // If we are doing little work each iteration, we should 3828 // take a short break. 3829 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3830 // Sleep for some time, waiting for work to accumulate 3831 stopTimer(); 3832 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3833 startTimer(); 3834 waited++; 3835 } 3836 } 3837 if (PrintCMSStatistics > 0) { 3838 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3839 loops, waited, cumworkdone); 3840 } 3841 } 3842 CMSTokenSync x(true); // is cms thread 3843 if (_collectorState != Idling) { 3844 assert(_collectorState == AbortablePreclean, 3845 "Spontaneous state transition?"); 3846 _collectorState = FinalMarking; 3847 } // Else, a foreground collection completed this CMS cycle. 3848 return; 3849 } 3850 3851 // Respond to an Eden sampling opportunity 3852 void CMSCollector::sample_eden() { 3853 // Make sure a young gc cannot sneak in between our 3854 // reading and recording of a sample. 3855 assert(Thread::current()->is_ConcurrentGC_thread(), 3856 "Only the cms thread may collect Eden samples"); 3857 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3858 "Should collect samples while holding CMS token"); 3859 if (!_start_sampling) { 3860 return; 3861 } 3862 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3863 // is populated by the young generation. 3864 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3865 if (_eden_chunk_index < _eden_chunk_capacity) { 3866 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3867 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3868 "Unexpected state of Eden"); 3869 // We'd like to check that what we just sampled is an oop-start address; 3870 // however, we cannot do that here since the object may not yet have been 3871 // initialized. So we'll instead do the check when we _use_ this sample 3872 // later. 3873 if (_eden_chunk_index == 0 || 3874 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3875 _eden_chunk_array[_eden_chunk_index-1]) 3876 >= CMSSamplingGrain)) { 3877 _eden_chunk_index++; // commit sample 3878 } 3879 } 3880 } 3881 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3882 size_t used = get_eden_used(); 3883 size_t capacity = get_eden_capacity(); 3884 assert(used <= capacity, "Unexpected state of Eden"); 3885 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3886 _abort_preclean = true; 3887 } 3888 } 3889 } 3890 3891 3892 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3893 assert(_collectorState == Precleaning || 3894 _collectorState == AbortablePreclean, "incorrect state"); 3895 ResourceMark rm; 3896 HandleMark hm; 3897 3898 // Precleaning is currently not MT but the reference processor 3899 // may be set for MT. Disable it temporarily here. 3900 ReferenceProcessor* rp = ref_processor(); 3901 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3902 3903 // Do one pass of scrubbing the discovered reference lists 3904 // to remove any reference objects with strongly-reachable 3905 // referents. 3906 if (clean_refs) { 3907 CMSPrecleanRefsYieldClosure yield_cl(this); 3908 assert(rp->span().equals(_span), "Spans should be equal"); 3909 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3910 &_markStack, true /* preclean */); 3911 CMSDrainMarkingStackClosure complete_trace(this, 3912 _span, &_markBitMap, &_markStack, 3913 &keep_alive, true /* preclean */); 3914 3915 // We don't want this step to interfere with a young 3916 // collection because we don't want to take CPU 3917 // or memory bandwidth away from the young GC threads 3918 // (which may be as many as there are CPUs). 3919 // Note that we don't need to protect ourselves from 3920 // interference with mutators because they can't 3921 // manipulate the discovered reference lists nor affect 3922 // the computed reachability of the referents, the 3923 // only properties manipulated by the precleaning 3924 // of these reference lists. 3925 stopTimer(); 3926 CMSTokenSyncWithLocks x(true /* is cms thread */, 3927 bitMapLock()); 3928 startTimer(); 3929 sample_eden(); 3930 3931 // The following will yield to allow foreground 3932 // collection to proceed promptly. XXX YSR: 3933 // The code in this method may need further 3934 // tweaking for better performance and some restructuring 3935 // for cleaner interfaces. 3936 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3937 rp->preclean_discovered_references( 3938 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3939 gc_timer, _gc_tracer_cm->gc_id()); 3940 } 3941 3942 if (clean_survivor) { // preclean the active survivor space(s) 3943 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3944 &_markBitMap, &_modUnionTable, 3945 &_markStack, true /* precleaning phase */); 3946 stopTimer(); 3947 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3948 bitMapLock()); 3949 startTimer(); 3950 unsigned int before_count = 3951 GenCollectedHeap::heap()->total_collections(); 3952 SurvivorSpacePrecleanClosure 3953 sss_cl(this, _span, &_markBitMap, &_markStack, 3954 &pam_cl, before_count, CMSYield); 3955 _young_gen->from()->object_iterate_careful(&sss_cl); 3956 _young_gen->to()->object_iterate_careful(&sss_cl); 3957 } 3958 MarkRefsIntoAndScanClosure 3959 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3960 &_markStack, this, CMSYield, 3961 true /* precleaning phase */); 3962 // CAUTION: The following closure has persistent state that may need to 3963 // be reset upon a decrease in the sequence of addresses it 3964 // processes. 3965 ScanMarkedObjectsAgainCarefullyClosure 3966 smoac_cl(this, _span, 3967 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3968 3969 // Preclean dirty cards in ModUnionTable and CardTable using 3970 // appropriate convergence criterion; 3971 // repeat CMSPrecleanIter times unless we find that 3972 // we are losing. 3973 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3974 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3975 "Bad convergence multiplier"); 3976 assert(CMSPrecleanThreshold >= 100, 3977 "Unreasonably low CMSPrecleanThreshold"); 3978 3979 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3980 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3981 numIter < CMSPrecleanIter; 3982 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3983 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3984 if (Verbose && PrintGCDetails) { 3985 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3986 } 3987 // Either there are very few dirty cards, so re-mark 3988 // pause will be small anyway, or our pre-cleaning isn't 3989 // that much faster than the rate at which cards are being 3990 // dirtied, so we might as well stop and re-mark since 3991 // precleaning won't improve our re-mark time by much. 3992 if (curNumCards <= CMSPrecleanThreshold || 3993 (numIter > 0 && 3994 (curNumCards * CMSPrecleanDenominator > 3995 lastNumCards * CMSPrecleanNumerator))) { 3996 numIter++; 3997 cumNumCards += curNumCards; 3998 break; 3999 } 4000 } 4001 4002 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4003 4004 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4005 cumNumCards += curNumCards; 4006 if (PrintGCDetails && PrintCMSStatistics != 0) { 4007 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4008 curNumCards, cumNumCards, numIter); 4009 } 4010 return cumNumCards; // as a measure of useful work done 4011 } 4012 4013 // PRECLEANING NOTES: 4014 // Precleaning involves: 4015 // . reading the bits of the modUnionTable and clearing the set bits. 4016 // . For the cards corresponding to the set bits, we scan the 4017 // objects on those cards. This means we need the free_list_lock 4018 // so that we can safely iterate over the CMS space when scanning 4019 // for oops. 4020 // . When we scan the objects, we'll be both reading and setting 4021 // marks in the marking bit map, so we'll need the marking bit map. 4022 // . For protecting _collector_state transitions, we take the CGC_lock. 4023 // Note that any races in the reading of of card table entries by the 4024 // CMS thread on the one hand and the clearing of those entries by the 4025 // VM thread or the setting of those entries by the mutator threads on the 4026 // other are quite benign. However, for efficiency it makes sense to keep 4027 // the VM thread from racing with the CMS thread while the latter is 4028 // dirty card info to the modUnionTable. We therefore also use the 4029 // CGC_lock to protect the reading of the card table and the mod union 4030 // table by the CM thread. 4031 // . We run concurrently with mutator updates, so scanning 4032 // needs to be done carefully -- we should not try to scan 4033 // potentially uninitialized objects. 4034 // 4035 // Locking strategy: While holding the CGC_lock, we scan over and 4036 // reset a maximal dirty range of the mod union / card tables, then lock 4037 // the free_list_lock and bitmap lock to do a full marking, then 4038 // release these locks; and repeat the cycle. This allows for a 4039 // certain amount of fairness in the sharing of these locks between 4040 // the CMS collector on the one hand, and the VM thread and the 4041 // mutators on the other. 4042 4043 // NOTE: preclean_mod_union_table() and preclean_card_table() 4044 // further below are largely identical; if you need to modify 4045 // one of these methods, please check the other method too. 4046 4047 size_t CMSCollector::preclean_mod_union_table( 4048 ConcurrentMarkSweepGeneration* gen, 4049 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4050 verify_work_stacks_empty(); 4051 verify_overflow_empty(); 4052 4053 // strategy: starting with the first card, accumulate contiguous 4054 // ranges of dirty cards; clear these cards, then scan the region 4055 // covered by these cards. 4056 4057 // Since all of the MUT is committed ahead, we can just use 4058 // that, in case the generations expand while we are precleaning. 4059 // It might also be fine to just use the committed part of the 4060 // generation, but we might potentially miss cards when the 4061 // generation is rapidly expanding while we are in the midst 4062 // of precleaning. 4063 HeapWord* startAddr = gen->reserved().start(); 4064 HeapWord* endAddr = gen->reserved().end(); 4065 4066 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4067 4068 size_t numDirtyCards, cumNumDirtyCards; 4069 HeapWord *nextAddr, *lastAddr; 4070 for (cumNumDirtyCards = numDirtyCards = 0, 4071 nextAddr = lastAddr = startAddr; 4072 nextAddr < endAddr; 4073 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4074 4075 ResourceMark rm; 4076 HandleMark hm; 4077 4078 MemRegion dirtyRegion; 4079 { 4080 stopTimer(); 4081 // Potential yield point 4082 CMSTokenSync ts(true); 4083 startTimer(); 4084 sample_eden(); 4085 // Get dirty region starting at nextOffset (inclusive), 4086 // simultaneously clearing it. 4087 dirtyRegion = 4088 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4089 assert(dirtyRegion.start() >= nextAddr, 4090 "returned region inconsistent?"); 4091 } 4092 // Remember where the next search should begin. 4093 // The returned region (if non-empty) is a right open interval, 4094 // so lastOffset is obtained from the right end of that 4095 // interval. 4096 lastAddr = dirtyRegion.end(); 4097 // Should do something more transparent and less hacky XXX 4098 numDirtyCards = 4099 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4100 4101 // We'll scan the cards in the dirty region (with periodic 4102 // yields for foreground GC as needed). 4103 if (!dirtyRegion.is_empty()) { 4104 assert(numDirtyCards > 0, "consistency check"); 4105 HeapWord* stop_point = NULL; 4106 stopTimer(); 4107 // Potential yield point 4108 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4109 bitMapLock()); 4110 startTimer(); 4111 { 4112 verify_work_stacks_empty(); 4113 verify_overflow_empty(); 4114 sample_eden(); 4115 stop_point = 4116 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4117 } 4118 if (stop_point != NULL) { 4119 // The careful iteration stopped early either because it found an 4120 // uninitialized object, or because we were in the midst of an 4121 // "abortable preclean", which should now be aborted. Redirty 4122 // the bits corresponding to the partially-scanned or unscanned 4123 // cards. We'll either restart at the next block boundary or 4124 // abort the preclean. 4125 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4126 "Should only be AbortablePreclean."); 4127 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4128 if (should_abort_preclean()) { 4129 break; // out of preclean loop 4130 } else { 4131 // Compute the next address at which preclean should pick up; 4132 // might need bitMapLock in order to read P-bits. 4133 lastAddr = next_card_start_after_block(stop_point); 4134 } 4135 } 4136 } else { 4137 assert(lastAddr == endAddr, "consistency check"); 4138 assert(numDirtyCards == 0, "consistency check"); 4139 break; 4140 } 4141 } 4142 verify_work_stacks_empty(); 4143 verify_overflow_empty(); 4144 return cumNumDirtyCards; 4145 } 4146 4147 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4148 // below are largely identical; if you need to modify 4149 // one of these methods, please check the other method too. 4150 4151 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4152 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4153 // strategy: it's similar to precleamModUnionTable above, in that 4154 // we accumulate contiguous ranges of dirty cards, mark these cards 4155 // precleaned, then scan the region covered by these cards. 4156 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4157 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4158 4159 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4160 4161 size_t numDirtyCards, cumNumDirtyCards; 4162 HeapWord *lastAddr, *nextAddr; 4163 4164 for (cumNumDirtyCards = numDirtyCards = 0, 4165 nextAddr = lastAddr = startAddr; 4166 nextAddr < endAddr; 4167 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4168 4169 ResourceMark rm; 4170 HandleMark hm; 4171 4172 MemRegion dirtyRegion; 4173 { 4174 // See comments in "Precleaning notes" above on why we 4175 // do this locking. XXX Could the locking overheads be 4176 // too high when dirty cards are sparse? [I don't think so.] 4177 stopTimer(); 4178 CMSTokenSync x(true); // is cms thread 4179 startTimer(); 4180 sample_eden(); 4181 // Get and clear dirty region from card table 4182 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4183 MemRegion(nextAddr, endAddr), 4184 true, 4185 CardTableModRefBS::precleaned_card_val()); 4186 4187 assert(dirtyRegion.start() >= nextAddr, 4188 "returned region inconsistent?"); 4189 } 4190 lastAddr = dirtyRegion.end(); 4191 numDirtyCards = 4192 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4193 4194 if (!dirtyRegion.is_empty()) { 4195 stopTimer(); 4196 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4197 startTimer(); 4198 sample_eden(); 4199 verify_work_stacks_empty(); 4200 verify_overflow_empty(); 4201 HeapWord* stop_point = 4202 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4203 if (stop_point != NULL) { 4204 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4205 "Should only be AbortablePreclean."); 4206 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4207 if (should_abort_preclean()) { 4208 break; // out of preclean loop 4209 } else { 4210 // Compute the next address at which preclean should pick up. 4211 lastAddr = next_card_start_after_block(stop_point); 4212 } 4213 } 4214 } else { 4215 break; 4216 } 4217 } 4218 verify_work_stacks_empty(); 4219 verify_overflow_empty(); 4220 return cumNumDirtyCards; 4221 } 4222 4223 class PrecleanKlassClosure : public KlassClosure { 4224 KlassToOopClosure _cm_klass_closure; 4225 public: 4226 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4227 void do_klass(Klass* k) { 4228 if (k->has_accumulated_modified_oops()) { 4229 k->clear_accumulated_modified_oops(); 4230 4231 _cm_klass_closure.do_klass(k); 4232 } 4233 } 4234 }; 4235 4236 // The freelist lock is needed to prevent asserts, is it really needed? 4237 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4238 4239 cl->set_freelistLock(freelistLock); 4240 4241 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4242 4243 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4244 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4245 PrecleanKlassClosure preclean_klass_closure(cl); 4246 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4247 4248 verify_work_stacks_empty(); 4249 verify_overflow_empty(); 4250 } 4251 4252 void CMSCollector::checkpointRootsFinal() { 4253 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4254 check_correct_thread_executing(); 4255 // world is stopped at this checkpoint 4256 assert(SafepointSynchronize::is_at_safepoint(), 4257 "world should be stopped"); 4258 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4259 4260 verify_work_stacks_empty(); 4261 verify_overflow_empty(); 4262 4263 if (PrintGCDetails) { 4264 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4265 _young_gen->used() / K, 4266 _young_gen->capacity() / K); 4267 } 4268 { 4269 if (CMSScavengeBeforeRemark) { 4270 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4271 // Temporarily set flag to false, GCH->do_collection will 4272 // expect it to be false and set to true 4273 FlagSetting fl(gch->_is_gc_active, false); 4274 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4275 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4276 int level = _cmsGen->level() - 1; 4277 if (level >= 0) { 4278 gch->do_collection(true, // full (i.e. force, see below) 4279 false, // !clear_all_soft_refs 4280 0, // size 4281 false, // is_tlab 4282 level // max_level 4283 ); 4284 } 4285 } 4286 FreelistLocker x(this); 4287 MutexLockerEx y(bitMapLock(), 4288 Mutex::_no_safepoint_check_flag); 4289 checkpointRootsFinalWork(); 4290 } 4291 verify_work_stacks_empty(); 4292 verify_overflow_empty(); 4293 } 4294 4295 void CMSCollector::checkpointRootsFinalWork() { 4296 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4297 4298 assert(haveFreelistLocks(), "must have free list locks"); 4299 assert_lock_strong(bitMapLock()); 4300 4301 ResourceMark rm; 4302 HandleMark hm; 4303 4304 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4305 4306 if (should_unload_classes()) { 4307 CodeCache::gc_prologue(); 4308 } 4309 assert(haveFreelistLocks(), "must have free list locks"); 4310 assert_lock_strong(bitMapLock()); 4311 4312 // We might assume that we need not fill TLAB's when 4313 // CMSScavengeBeforeRemark is set, because we may have just done 4314 // a scavenge which would have filled all TLAB's -- and besides 4315 // Eden would be empty. This however may not always be the case -- 4316 // for instance although we asked for a scavenge, it may not have 4317 // happened because of a JNI critical section. We probably need 4318 // a policy for deciding whether we can in that case wait until 4319 // the critical section releases and then do the remark following 4320 // the scavenge, and skip it here. In the absence of that policy, 4321 // or of an indication of whether the scavenge did indeed occur, 4322 // we cannot rely on TLAB's having been filled and must do 4323 // so here just in case a scavenge did not happen. 4324 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4325 // Update the saved marks which may affect the root scans. 4326 gch->save_marks(); 4327 4328 if (CMSPrintEdenSurvivorChunks) { 4329 print_eden_and_survivor_chunk_arrays(); 4330 } 4331 4332 { 4333 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4334 4335 // Note on the role of the mod union table: 4336 // Since the marker in "markFromRoots" marks concurrently with 4337 // mutators, it is possible for some reachable objects not to have been 4338 // scanned. For instance, an only reference to an object A was 4339 // placed in object B after the marker scanned B. Unless B is rescanned, 4340 // A would be collected. Such updates to references in marked objects 4341 // are detected via the mod union table which is the set of all cards 4342 // dirtied since the first checkpoint in this GC cycle and prior to 4343 // the most recent young generation GC, minus those cleaned up by the 4344 // concurrent precleaning. 4345 if (CMSParallelRemarkEnabled) { 4346 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4347 do_remark_parallel(); 4348 } else { 4349 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4350 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4351 do_remark_non_parallel(); 4352 } 4353 } 4354 verify_work_stacks_empty(); 4355 verify_overflow_empty(); 4356 4357 { 4358 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4359 refProcessingWork(); 4360 } 4361 verify_work_stacks_empty(); 4362 verify_overflow_empty(); 4363 4364 if (should_unload_classes()) { 4365 CodeCache::gc_epilogue(); 4366 } 4367 JvmtiExport::gc_epilogue(); 4368 4369 // If we encountered any (marking stack / work queue) overflow 4370 // events during the current CMS cycle, take appropriate 4371 // remedial measures, where possible, so as to try and avoid 4372 // recurrence of that condition. 4373 assert(_markStack.isEmpty(), "No grey objects"); 4374 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4375 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4376 if (ser_ovflw > 0) { 4377 if (PrintCMSStatistics != 0) { 4378 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4379 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4380 ", kac_preclean="SIZE_FORMAT")", 4381 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4382 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4383 } 4384 _markStack.expand(); 4385 _ser_pmc_remark_ovflw = 0; 4386 _ser_pmc_preclean_ovflw = 0; 4387 _ser_kac_preclean_ovflw = 0; 4388 _ser_kac_ovflw = 0; 4389 } 4390 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4391 if (PrintCMSStatistics != 0) { 4392 gclog_or_tty->print_cr("Work queue overflow (benign) " 4393 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4394 _par_pmc_remark_ovflw, _par_kac_ovflw); 4395 } 4396 _par_pmc_remark_ovflw = 0; 4397 _par_kac_ovflw = 0; 4398 } 4399 if (PrintCMSStatistics != 0) { 4400 if (_markStack._hit_limit > 0) { 4401 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4402 _markStack._hit_limit); 4403 } 4404 if (_markStack._failed_double > 0) { 4405 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4406 " current capacity "SIZE_FORMAT, 4407 _markStack._failed_double, 4408 _markStack.capacity()); 4409 } 4410 } 4411 _markStack._hit_limit = 0; 4412 _markStack._failed_double = 0; 4413 4414 if ((VerifyAfterGC || VerifyDuringGC) && 4415 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4416 verify_after_remark(); 4417 } 4418 4419 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4420 4421 // Change under the freelistLocks. 4422 _collectorState = Sweeping; 4423 // Call isAllClear() under bitMapLock 4424 assert(_modUnionTable.isAllClear(), 4425 "Should be clear by end of the final marking"); 4426 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4427 "Should be clear by end of the final marking"); 4428 } 4429 4430 void CMSParInitialMarkTask::work(uint worker_id) { 4431 elapsedTimer _timer; 4432 ResourceMark rm; 4433 HandleMark hm; 4434 4435 // ---------- scan from roots -------------- 4436 _timer.start(); 4437 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4438 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4439 4440 // ---------- young gen roots -------------- 4441 { 4442 work_on_young_gen_roots(worker_id, &par_mri_cl); 4443 _timer.stop(); 4444 if (PrintCMSStatistics != 0) { 4445 gclog_or_tty->print_cr( 4446 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4447 worker_id, _timer.seconds()); 4448 } 4449 } 4450 4451 // ---------- remaining roots -------------- 4452 _timer.reset(); 4453 _timer.start(); 4454 4455 CLDToOopClosure cld_closure(&par_mri_cl, true); 4456 4457 gch->gen_process_roots(_collector->_cmsGen->level(), 4458 false, // yg was scanned above 4459 false, // this is parallel code 4460 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4461 _collector->should_unload_classes(), 4462 &par_mri_cl, 4463 NULL, 4464 &cld_closure); 4465 assert(_collector->should_unload_classes() 4466 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4467 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4468 _timer.stop(); 4469 if (PrintCMSStatistics != 0) { 4470 gclog_or_tty->print_cr( 4471 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4472 worker_id, _timer.seconds()); 4473 } 4474 } 4475 4476 // Parallel remark task 4477 class CMSParRemarkTask: public CMSParMarkTask { 4478 CompactibleFreeListSpace* _cms_space; 4479 4480 // The per-thread work queues, available here for stealing. 4481 OopTaskQueueSet* _task_queues; 4482 ParallelTaskTerminator _term; 4483 4484 public: 4485 // A value of 0 passed to n_workers will cause the number of 4486 // workers to be taken from the active workers in the work gang. 4487 CMSParRemarkTask(CMSCollector* collector, 4488 CompactibleFreeListSpace* cms_space, 4489 int n_workers, FlexibleWorkGang* workers, 4490 OopTaskQueueSet* task_queues): 4491 CMSParMarkTask("Rescan roots and grey objects in parallel", 4492 collector, n_workers), 4493 _cms_space(cms_space), 4494 _task_queues(task_queues), 4495 _term(n_workers, task_queues) { } 4496 4497 OopTaskQueueSet* task_queues() { return _task_queues; } 4498 4499 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4500 4501 ParallelTaskTerminator* terminator() { return &_term; } 4502 int n_workers() { return _n_workers; } 4503 4504 void work(uint worker_id); 4505 4506 private: 4507 // ... of dirty cards in old space 4508 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4509 Par_MarkRefsIntoAndScanClosure* cl); 4510 4511 // ... work stealing for the above 4512 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4513 }; 4514 4515 class RemarkKlassClosure : public KlassClosure { 4516 KlassToOopClosure _cm_klass_closure; 4517 public: 4518 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4519 void do_klass(Klass* k) { 4520 // Check if we have modified any oops in the Klass during the concurrent marking. 4521 if (k->has_accumulated_modified_oops()) { 4522 k->clear_accumulated_modified_oops(); 4523 4524 // We could have transfered the current modified marks to the accumulated marks, 4525 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4526 } else if (k->has_modified_oops()) { 4527 // Don't clear anything, this info is needed by the next young collection. 4528 } else { 4529 // No modified oops in the Klass. 4530 return; 4531 } 4532 4533 // The klass has modified fields, need to scan the klass. 4534 _cm_klass_closure.do_klass(k); 4535 } 4536 }; 4537 4538 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4539 ParNewGeneration* young_gen = _collector->_young_gen; 4540 ContiguousSpace* eden_space = young_gen->eden(); 4541 ContiguousSpace* from_space = young_gen->from(); 4542 ContiguousSpace* to_space = young_gen->to(); 4543 4544 HeapWord** eca = _collector->_eden_chunk_array; 4545 size_t ect = _collector->_eden_chunk_index; 4546 HeapWord** sca = _collector->_survivor_chunk_array; 4547 size_t sct = _collector->_survivor_chunk_index; 4548 4549 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4550 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4551 4552 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4553 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4554 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4555 } 4556 4557 // work_queue(i) is passed to the closure 4558 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4559 // also is passed to do_dirty_card_rescan_tasks() and to 4560 // do_work_steal() to select the i-th task_queue. 4561 4562 void CMSParRemarkTask::work(uint worker_id) { 4563 elapsedTimer _timer; 4564 ResourceMark rm; 4565 HandleMark hm; 4566 4567 // ---------- rescan from roots -------------- 4568 _timer.start(); 4569 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4570 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4571 _collector->_span, _collector->ref_processor(), 4572 &(_collector->_markBitMap), 4573 work_queue(worker_id)); 4574 4575 // Rescan young gen roots first since these are likely 4576 // coarsely partitioned and may, on that account, constitute 4577 // the critical path; thus, it's best to start off that 4578 // work first. 4579 // ---------- young gen roots -------------- 4580 { 4581 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4582 _timer.stop(); 4583 if (PrintCMSStatistics != 0) { 4584 gclog_or_tty->print_cr( 4585 "Finished young gen rescan work in %dth thread: %3.3f sec", 4586 worker_id, _timer.seconds()); 4587 } 4588 } 4589 4590 // ---------- remaining roots -------------- 4591 _timer.reset(); 4592 _timer.start(); 4593 gch->gen_process_roots(_collector->_cmsGen->level(), 4594 false, // yg was scanned above 4595 false, // this is parallel code 4596 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4597 _collector->should_unload_classes(), 4598 &par_mrias_cl, 4599 NULL, 4600 NULL); // The dirty klasses will be handled below 4601 4602 assert(_collector->should_unload_classes() 4603 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4604 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4605 _timer.stop(); 4606 if (PrintCMSStatistics != 0) { 4607 gclog_or_tty->print_cr( 4608 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4609 worker_id, _timer.seconds()); 4610 } 4611 4612 // ---------- unhandled CLD scanning ---------- 4613 if (worker_id == 0) { // Single threaded at the moment. 4614 _timer.reset(); 4615 _timer.start(); 4616 4617 // Scan all new class loader data objects and new dependencies that were 4618 // introduced during concurrent marking. 4619 ResourceMark rm; 4620 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4621 for (int i = 0; i < array->length(); i++) { 4622 par_mrias_cl.do_class_loader_data(array->at(i)); 4623 } 4624 4625 // We don't need to keep track of new CLDs anymore. 4626 ClassLoaderDataGraph::remember_new_clds(false); 4627 4628 _timer.stop(); 4629 if (PrintCMSStatistics != 0) { 4630 gclog_or_tty->print_cr( 4631 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4632 worker_id, _timer.seconds()); 4633 } 4634 } 4635 4636 // ---------- dirty klass scanning ---------- 4637 if (worker_id == 0) { // Single threaded at the moment. 4638 _timer.reset(); 4639 _timer.start(); 4640 4641 // Scan all classes that was dirtied during the concurrent marking phase. 4642 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4643 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4644 4645 _timer.stop(); 4646 if (PrintCMSStatistics != 0) { 4647 gclog_or_tty->print_cr( 4648 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4649 worker_id, _timer.seconds()); 4650 } 4651 } 4652 4653 // We might have added oops to ClassLoaderData::_handles during the 4654 // concurrent marking phase. These oops point to newly allocated objects 4655 // that are guaranteed to be kept alive. Either by the direct allocation 4656 // code, or when the young collector processes the roots. Hence, 4657 // we don't have to revisit the _handles block during the remark phase. 4658 4659 // ---------- rescan dirty cards ------------ 4660 _timer.reset(); 4661 _timer.start(); 4662 4663 // Do the rescan tasks for each of the two spaces 4664 // (cms_space) in turn. 4665 // "worker_id" is passed to select the task_queue for "worker_id" 4666 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4667 _timer.stop(); 4668 if (PrintCMSStatistics != 0) { 4669 gclog_or_tty->print_cr( 4670 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4671 worker_id, _timer.seconds()); 4672 } 4673 4674 // ---------- steal work from other threads ... 4675 // ---------- ... and drain overflow list. 4676 _timer.reset(); 4677 _timer.start(); 4678 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4679 _timer.stop(); 4680 if (PrintCMSStatistics != 0) { 4681 gclog_or_tty->print_cr( 4682 "Finished work stealing in %dth thread: %3.3f sec", 4683 worker_id, _timer.seconds()); 4684 } 4685 } 4686 4687 // Note that parameter "i" is not used. 4688 void 4689 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4690 OopsInGenClosure* cl, ContiguousSpace* space, 4691 HeapWord** chunk_array, size_t chunk_top) { 4692 // Until all tasks completed: 4693 // . claim an unclaimed task 4694 // . compute region boundaries corresponding to task claimed 4695 // using chunk_array 4696 // . par_oop_iterate(cl) over that region 4697 4698 ResourceMark rm; 4699 HandleMark hm; 4700 4701 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4702 4703 uint nth_task = 0; 4704 uint n_tasks = pst->n_tasks(); 4705 4706 if (n_tasks > 0) { 4707 assert(pst->valid(), "Uninitialized use?"); 4708 HeapWord *start, *end; 4709 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4710 // We claimed task # nth_task; compute its boundaries. 4711 if (chunk_top == 0) { // no samples were taken 4712 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4713 start = space->bottom(); 4714 end = space->top(); 4715 } else if (nth_task == 0) { 4716 start = space->bottom(); 4717 end = chunk_array[nth_task]; 4718 } else if (nth_task < (uint)chunk_top) { 4719 assert(nth_task >= 1, "Control point invariant"); 4720 start = chunk_array[nth_task - 1]; 4721 end = chunk_array[nth_task]; 4722 } else { 4723 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4724 start = chunk_array[chunk_top - 1]; 4725 end = space->top(); 4726 } 4727 MemRegion mr(start, end); 4728 // Verify that mr is in space 4729 assert(mr.is_empty() || space->used_region().contains(mr), 4730 "Should be in space"); 4731 // Verify that "start" is an object boundary 4732 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4733 "Should be an oop"); 4734 space->par_oop_iterate(mr, cl); 4735 } 4736 pst->all_tasks_completed(); 4737 } 4738 } 4739 4740 void 4741 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4742 CompactibleFreeListSpace* sp, int i, 4743 Par_MarkRefsIntoAndScanClosure* cl) { 4744 // Until all tasks completed: 4745 // . claim an unclaimed task 4746 // . compute region boundaries corresponding to task claimed 4747 // . transfer dirty bits ct->mut for that region 4748 // . apply rescanclosure to dirty mut bits for that region 4749 4750 ResourceMark rm; 4751 HandleMark hm; 4752 4753 OopTaskQueue* work_q = work_queue(i); 4754 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4755 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4756 // CAUTION: This closure has state that persists across calls to 4757 // the work method dirty_range_iterate_clear() in that it has 4758 // embedded in it a (subtype of) UpwardsObjectClosure. The 4759 // use of that state in the embedded UpwardsObjectClosure instance 4760 // assumes that the cards are always iterated (even if in parallel 4761 // by several threads) in monotonically increasing order per each 4762 // thread. This is true of the implementation below which picks 4763 // card ranges (chunks) in monotonically increasing order globally 4764 // and, a-fortiori, in monotonically increasing order per thread 4765 // (the latter order being a subsequence of the former). 4766 // If the work code below is ever reorganized into a more chaotic 4767 // work-partitioning form than the current "sequential tasks" 4768 // paradigm, the use of that persistent state will have to be 4769 // revisited and modified appropriately. See also related 4770 // bug 4756801 work on which should examine this code to make 4771 // sure that the changes there do not run counter to the 4772 // assumptions made here and necessary for correctness and 4773 // efficiency. Note also that this code might yield inefficient 4774 // behavior in the case of very large objects that span one or 4775 // more work chunks. Such objects would potentially be scanned 4776 // several times redundantly. Work on 4756801 should try and 4777 // address that performance anomaly if at all possible. XXX 4778 MemRegion full_span = _collector->_span; 4779 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4780 MarkFromDirtyCardsClosure 4781 greyRescanClosure(_collector, full_span, // entire span of interest 4782 sp, bm, work_q, cl); 4783 4784 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4785 assert(pst->valid(), "Uninitialized use?"); 4786 uint nth_task = 0; 4787 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4788 MemRegion span = sp->used_region(); 4789 HeapWord* start_addr = span.start(); 4790 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4791 alignment); 4792 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4793 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4794 start_addr, "Check alignment"); 4795 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4796 chunk_size, "Check alignment"); 4797 4798 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4799 // Having claimed the nth_task, compute corresponding mem-region, 4800 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4801 // The alignment restriction ensures that we do not need any 4802 // synchronization with other gang-workers while setting or 4803 // clearing bits in thus chunk of the MUT. 4804 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4805 start_addr + (nth_task+1)*chunk_size); 4806 // The last chunk's end might be way beyond end of the 4807 // used region. In that case pull back appropriately. 4808 if (this_span.end() > end_addr) { 4809 this_span.set_end(end_addr); 4810 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4811 } 4812 // Iterate over the dirty cards covering this chunk, marking them 4813 // precleaned, and setting the corresponding bits in the mod union 4814 // table. Since we have been careful to partition at Card and MUT-word 4815 // boundaries no synchronization is needed between parallel threads. 4816 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4817 &modUnionClosure); 4818 4819 // Having transferred these marks into the modUnionTable, 4820 // rescan the marked objects on the dirty cards in the modUnionTable. 4821 // Even if this is at a synchronous collection, the initial marking 4822 // may have been done during an asynchronous collection so there 4823 // may be dirty bits in the mod-union table. 4824 _collector->_modUnionTable.dirty_range_iterate_clear( 4825 this_span, &greyRescanClosure); 4826 _collector->_modUnionTable.verifyNoOneBitsInRange( 4827 this_span.start(), 4828 this_span.end()); 4829 } 4830 pst->all_tasks_completed(); // declare that i am done 4831 } 4832 4833 // . see if we can share work_queues with ParNew? XXX 4834 void 4835 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4836 int* seed) { 4837 OopTaskQueue* work_q = work_queue(i); 4838 NOT_PRODUCT(int num_steals = 0;) 4839 oop obj_to_scan; 4840 CMSBitMap* bm = &(_collector->_markBitMap); 4841 4842 while (true) { 4843 // Completely finish any left over work from (an) earlier round(s) 4844 cl->trim_queue(0); 4845 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4846 (size_t)ParGCDesiredObjsFromOverflowList); 4847 // Now check if there's any work in the overflow list 4848 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4849 // only affects the number of attempts made to get work from the 4850 // overflow list and does not affect the number of workers. Just 4851 // pass ParallelGCThreads so this behavior is unchanged. 4852 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4853 work_q, 4854 ParallelGCThreads)) { 4855 // found something in global overflow list; 4856 // not yet ready to go stealing work from others. 4857 // We'd like to assert(work_q->size() != 0, ...) 4858 // because we just took work from the overflow list, 4859 // but of course we can't since all of that could have 4860 // been already stolen from us. 4861 // "He giveth and He taketh away." 4862 continue; 4863 } 4864 // Verify that we have no work before we resort to stealing 4865 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4866 // Try to steal from other queues that have work 4867 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4868 NOT_PRODUCT(num_steals++;) 4869 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4870 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4871 // Do scanning work 4872 obj_to_scan->oop_iterate(cl); 4873 // Loop around, finish this work, and try to steal some more 4874 } else if (terminator()->offer_termination()) { 4875 break; // nirvana from the infinite cycle 4876 } 4877 } 4878 NOT_PRODUCT( 4879 if (PrintCMSStatistics != 0) { 4880 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4881 } 4882 ) 4883 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4884 "Else our work is not yet done"); 4885 } 4886 4887 // Record object boundaries in _eden_chunk_array by sampling the eden 4888 // top in the slow-path eden object allocation code path and record 4889 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4890 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4891 // sampling in sample_eden() that activates during the part of the 4892 // preclean phase. 4893 void CMSCollector::sample_eden_chunk() { 4894 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4895 if (_eden_chunk_lock->try_lock()) { 4896 // Record a sample. This is the critical section. The contents 4897 // of the _eden_chunk_array have to be non-decreasing in the 4898 // address order. 4899 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4900 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4901 "Unexpected state of Eden"); 4902 if (_eden_chunk_index == 0 || 4903 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4904 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4905 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4906 _eden_chunk_index++; // commit sample 4907 } 4908 _eden_chunk_lock->unlock(); 4909 } 4910 } 4911 } 4912 4913 // Return a thread-local PLAB recording array, as appropriate. 4914 void* CMSCollector::get_data_recorder(int thr_num) { 4915 if (_survivor_plab_array != NULL && 4916 (CMSPLABRecordAlways || 4917 (_collectorState > Marking && _collectorState < FinalMarking))) { 4918 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4919 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4920 ca->reset(); // clear it so that fresh data is recorded 4921 return (void*) ca; 4922 } else { 4923 return NULL; 4924 } 4925 } 4926 4927 // Reset all the thread-local PLAB recording arrays 4928 void CMSCollector::reset_survivor_plab_arrays() { 4929 for (uint i = 0; i < ParallelGCThreads; i++) { 4930 _survivor_plab_array[i].reset(); 4931 } 4932 } 4933 4934 // Merge the per-thread plab arrays into the global survivor chunk 4935 // array which will provide the partitioning of the survivor space 4936 // for CMS initial scan and rescan. 4937 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4938 int no_of_gc_threads) { 4939 assert(_survivor_plab_array != NULL, "Error"); 4940 assert(_survivor_chunk_array != NULL, "Error"); 4941 assert(_collectorState == FinalMarking || 4942 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4943 for (int j = 0; j < no_of_gc_threads; j++) { 4944 _cursor[j] = 0; 4945 } 4946 HeapWord* top = surv->top(); 4947 size_t i; 4948 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4949 HeapWord* min_val = top; // Higher than any PLAB address 4950 uint min_tid = 0; // position of min_val this round 4951 for (int j = 0; j < no_of_gc_threads; j++) { 4952 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4953 if (_cursor[j] == cur_sca->end()) { 4954 continue; 4955 } 4956 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4957 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4958 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4959 if (cur_val < min_val) { 4960 min_tid = j; 4961 min_val = cur_val; 4962 } else { 4963 assert(cur_val < top, "All recorded addresses should be less"); 4964 } 4965 } 4966 // At this point min_val and min_tid are respectively 4967 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4968 // and the thread (j) that witnesses that address. 4969 // We record this address in the _survivor_chunk_array[i] 4970 // and increment _cursor[min_tid] prior to the next round i. 4971 if (min_val == top) { 4972 break; 4973 } 4974 _survivor_chunk_array[i] = min_val; 4975 _cursor[min_tid]++; 4976 } 4977 // We are all done; record the size of the _survivor_chunk_array 4978 _survivor_chunk_index = i; // exclusive: [0, i) 4979 if (PrintCMSStatistics > 0) { 4980 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4981 } 4982 // Verify that we used up all the recorded entries 4983 #ifdef ASSERT 4984 size_t total = 0; 4985 for (int j = 0; j < no_of_gc_threads; j++) { 4986 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4987 total += _cursor[j]; 4988 } 4989 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4990 // Check that the merged array is in sorted order 4991 if (total > 0) { 4992 for (size_t i = 0; i < total - 1; i++) { 4993 if (PrintCMSStatistics > 0) { 4994 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4995 i, p2i(_survivor_chunk_array[i])); 4996 } 4997 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4998 "Not sorted"); 4999 } 5000 } 5001 #endif // ASSERT 5002 } 5003 5004 // Set up the space's par_seq_tasks structure for work claiming 5005 // for parallel initial scan and rescan of young gen. 5006 // See ParRescanTask where this is currently used. 5007 void 5008 CMSCollector:: 5009 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5010 assert(n_threads > 0, "Unexpected n_threads argument"); 5011 5012 // Eden space 5013 if (!_young_gen->eden()->is_empty()) { 5014 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5015 assert(!pst->valid(), "Clobbering existing data?"); 5016 // Each valid entry in [0, _eden_chunk_index) represents a task. 5017 size_t n_tasks = _eden_chunk_index + 1; 5018 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5019 // Sets the condition for completion of the subtask (how many threads 5020 // need to finish in order to be done). 5021 pst->set_n_threads(n_threads); 5022 pst->set_n_tasks((int)n_tasks); 5023 } 5024 5025 // Merge the survivor plab arrays into _survivor_chunk_array 5026 if (_survivor_plab_array != NULL) { 5027 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5028 } else { 5029 assert(_survivor_chunk_index == 0, "Error"); 5030 } 5031 5032 // To space 5033 { 5034 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5035 assert(!pst->valid(), "Clobbering existing data?"); 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(1); 5040 assert(pst->valid(), "Error"); 5041 } 5042 5043 // From space 5044 { 5045 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5046 assert(!pst->valid(), "Clobbering existing data?"); 5047 size_t n_tasks = _survivor_chunk_index + 1; 5048 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5049 // Sets the condition for completion of the subtask (how many threads 5050 // need to finish in order to be done). 5051 pst->set_n_threads(n_threads); 5052 pst->set_n_tasks((int)n_tasks); 5053 assert(pst->valid(), "Error"); 5054 } 5055 } 5056 5057 // Parallel version of remark 5058 void CMSCollector::do_remark_parallel() { 5059 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5060 FlexibleWorkGang* workers = gch->workers(); 5061 assert(workers != NULL, "Need parallel worker threads."); 5062 // Choose to use the number of GC workers most recently set 5063 // into "active_workers". If active_workers is not set, set it 5064 // to ParallelGCThreads. 5065 int n_workers = workers->active_workers(); 5066 if (n_workers == 0) { 5067 assert(n_workers > 0, "Should have been set during scavenge"); 5068 n_workers = ParallelGCThreads; 5069 workers->set_active_workers(n_workers); 5070 } 5071 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5072 5073 CMSParRemarkTask tsk(this, 5074 cms_space, 5075 n_workers, workers, task_queues()); 5076 5077 // Set up for parallel process_roots work. 5078 gch->set_par_threads(n_workers); 5079 // We won't be iterating over the cards in the card table updating 5080 // the younger_gen cards, so we shouldn't call the following else 5081 // the verification code as well as subsequent younger_refs_iterate 5082 // code would get confused. XXX 5083 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5084 5085 // The young gen rescan work will not be done as part of 5086 // process_roots (which currently doesn't know how to 5087 // parallelize such a scan), but rather will be broken up into 5088 // a set of parallel tasks (via the sampling that the [abortable] 5089 // preclean phase did of eden, plus the [two] tasks of 5090 // scanning the [two] survivor spaces. Further fine-grain 5091 // parallelization of the scanning of the survivor spaces 5092 // themselves, and of precleaning of the younger gen itself 5093 // is deferred to the future. 5094 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5095 5096 // The dirty card rescan work is broken up into a "sequence" 5097 // of parallel tasks (per constituent space) that are dynamically 5098 // claimed by the parallel threads. 5099 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5100 5101 // It turns out that even when we're using 1 thread, doing the work in a 5102 // separate thread causes wide variance in run times. We can't help this 5103 // in the multi-threaded case, but we special-case n=1 here to get 5104 // repeatable measurements of the 1-thread overhead of the parallel code. 5105 if (n_workers > 1) { 5106 // Make refs discovery MT-safe, if it isn't already: it may not 5107 // necessarily be so, since it's possible that we are doing 5108 // ST marking. 5109 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5110 StrongRootsScope srs; 5111 workers->run_task(&tsk); 5112 } else { 5113 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5114 StrongRootsScope srs; 5115 tsk.work(0); 5116 } 5117 5118 gch->set_par_threads(0); // 0 ==> non-parallel. 5119 // restore, single-threaded for now, any preserved marks 5120 // as a result of work_q overflow 5121 restore_preserved_marks_if_any(); 5122 } 5123 5124 // Non-parallel version of remark 5125 void CMSCollector::do_remark_non_parallel() { 5126 ResourceMark rm; 5127 HandleMark hm; 5128 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5129 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5130 5131 MarkRefsIntoAndScanClosure 5132 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5133 &_markStack, this, 5134 false /* should_yield */, false /* not precleaning */); 5135 MarkFromDirtyCardsClosure 5136 markFromDirtyCardsClosure(this, _span, 5137 NULL, // space is set further below 5138 &_markBitMap, &_markStack, &mrias_cl); 5139 { 5140 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5141 // Iterate over the dirty cards, setting the corresponding bits in the 5142 // mod union table. 5143 { 5144 ModUnionClosure modUnionClosure(&_modUnionTable); 5145 _ct->ct_bs()->dirty_card_iterate( 5146 _cmsGen->used_region(), 5147 &modUnionClosure); 5148 } 5149 // Having transferred these marks into the modUnionTable, we just need 5150 // to rescan the marked objects on the dirty cards in the modUnionTable. 5151 // The initial marking may have been done during an asynchronous 5152 // collection so there may be dirty bits in the mod-union table. 5153 const int alignment = 5154 CardTableModRefBS::card_size * BitsPerWord; 5155 { 5156 // ... First handle dirty cards in CMS gen 5157 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5158 MemRegion ur = _cmsGen->used_region(); 5159 HeapWord* lb = ur.start(); 5160 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5161 MemRegion cms_span(lb, ub); 5162 _modUnionTable.dirty_range_iterate_clear(cms_span, 5163 &markFromDirtyCardsClosure); 5164 verify_work_stacks_empty(); 5165 if (PrintCMSStatistics != 0) { 5166 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5167 markFromDirtyCardsClosure.num_dirty_cards()); 5168 } 5169 } 5170 } 5171 if (VerifyDuringGC && 5172 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5173 HandleMark hm; // Discard invalid handles created during verification 5174 Universe::verify(); 5175 } 5176 { 5177 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5178 5179 verify_work_stacks_empty(); 5180 5181 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5182 StrongRootsScope srs; 5183 5184 gch->gen_process_roots(_cmsGen->level(), 5185 true, // younger gens as roots 5186 false, // use the local StrongRootsScope 5187 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5188 should_unload_classes(), 5189 &mrias_cl, 5190 NULL, 5191 NULL); // The dirty klasses will be handled below 5192 5193 assert(should_unload_classes() 5194 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5195 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5196 } 5197 5198 { 5199 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5200 5201 verify_work_stacks_empty(); 5202 5203 // Scan all class loader data objects that might have been introduced 5204 // during concurrent marking. 5205 ResourceMark rm; 5206 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5207 for (int i = 0; i < array->length(); i++) { 5208 mrias_cl.do_class_loader_data(array->at(i)); 5209 } 5210 5211 // We don't need to keep track of new CLDs anymore. 5212 ClassLoaderDataGraph::remember_new_clds(false); 5213 5214 verify_work_stacks_empty(); 5215 } 5216 5217 { 5218 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5219 5220 verify_work_stacks_empty(); 5221 5222 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5223 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5224 5225 verify_work_stacks_empty(); 5226 } 5227 5228 // We might have added oops to ClassLoaderData::_handles during the 5229 // concurrent marking phase. These oops point to newly allocated objects 5230 // that are guaranteed to be kept alive. Either by the direct allocation 5231 // code, or when the young collector processes the roots. Hence, 5232 // we don't have to revisit the _handles block during the remark phase. 5233 5234 verify_work_stacks_empty(); 5235 // Restore evacuated mark words, if any, used for overflow list links 5236 if (!CMSOverflowEarlyRestoration) { 5237 restore_preserved_marks_if_any(); 5238 } 5239 verify_overflow_empty(); 5240 } 5241 5242 //////////////////////////////////////////////////////// 5243 // Parallel Reference Processing Task Proxy Class 5244 //////////////////////////////////////////////////////// 5245 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5246 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5247 CMSCollector* _collector; 5248 CMSBitMap* _mark_bit_map; 5249 const MemRegion _span; 5250 ProcessTask& _task; 5251 5252 public: 5253 CMSRefProcTaskProxy(ProcessTask& task, 5254 CMSCollector* collector, 5255 const MemRegion& span, 5256 CMSBitMap* mark_bit_map, 5257 AbstractWorkGang* workers, 5258 OopTaskQueueSet* task_queues): 5259 // XXX Should superclass AGTWOQ also know about AWG since it knows 5260 // about the task_queues used by the AWG? Then it could initialize 5261 // the terminator() object. See 6984287. The set_for_termination() 5262 // below is a temporary band-aid for the regression in 6984287. 5263 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5264 task_queues), 5265 _task(task), 5266 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5267 { 5268 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5269 "Inconsistency in _span"); 5270 set_for_termination(workers->active_workers()); 5271 } 5272 5273 OopTaskQueueSet* task_queues() { return queues(); } 5274 5275 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5276 5277 void do_work_steal(int i, 5278 CMSParDrainMarkingStackClosure* drain, 5279 CMSParKeepAliveClosure* keep_alive, 5280 int* seed); 5281 5282 virtual void work(uint worker_id); 5283 }; 5284 5285 void CMSRefProcTaskProxy::work(uint worker_id) { 5286 ResourceMark rm; 5287 HandleMark hm; 5288 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5289 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5290 _mark_bit_map, 5291 work_queue(worker_id)); 5292 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5293 _mark_bit_map, 5294 work_queue(worker_id)); 5295 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5296 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5297 if (_task.marks_oops_alive()) { 5298 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5299 _collector->hash_seed(worker_id)); 5300 } 5301 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5302 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5303 } 5304 5305 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5306 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5307 EnqueueTask& _task; 5308 5309 public: 5310 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5311 : AbstractGangTask("Enqueue reference objects in parallel"), 5312 _task(task) 5313 { } 5314 5315 virtual void work(uint worker_id) 5316 { 5317 _task.work(worker_id); 5318 } 5319 }; 5320 5321 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5322 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5323 _span(span), 5324 _bit_map(bit_map), 5325 _work_queue(work_queue), 5326 _mark_and_push(collector, span, bit_map, work_queue), 5327 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5328 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5329 { } 5330 5331 // . see if we can share work_queues with ParNew? XXX 5332 void CMSRefProcTaskProxy::do_work_steal(int i, 5333 CMSParDrainMarkingStackClosure* drain, 5334 CMSParKeepAliveClosure* keep_alive, 5335 int* seed) { 5336 OopTaskQueue* work_q = work_queue(i); 5337 NOT_PRODUCT(int num_steals = 0;) 5338 oop obj_to_scan; 5339 5340 while (true) { 5341 // Completely finish any left over work from (an) earlier round(s) 5342 drain->trim_queue(0); 5343 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5344 (size_t)ParGCDesiredObjsFromOverflowList); 5345 // Now check if there's any work in the overflow list 5346 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5347 // only affects the number of attempts made to get work from the 5348 // overflow list and does not affect the number of workers. Just 5349 // pass ParallelGCThreads so this behavior is unchanged. 5350 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5351 work_q, 5352 ParallelGCThreads)) { 5353 // Found something in global overflow list; 5354 // not yet ready to go stealing work from others. 5355 // We'd like to assert(work_q->size() != 0, ...) 5356 // because we just took work from the overflow list, 5357 // but of course we can't, since all of that might have 5358 // been already stolen from us. 5359 continue; 5360 } 5361 // Verify that we have no work before we resort to stealing 5362 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5363 // Try to steal from other queues that have work 5364 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5365 NOT_PRODUCT(num_steals++;) 5366 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5367 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5368 // Do scanning work 5369 obj_to_scan->oop_iterate(keep_alive); 5370 // Loop around, finish this work, and try to steal some more 5371 } else if (terminator()->offer_termination()) { 5372 break; // nirvana from the infinite cycle 5373 } 5374 } 5375 NOT_PRODUCT( 5376 if (PrintCMSStatistics != 0) { 5377 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5378 } 5379 ) 5380 } 5381 5382 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5383 { 5384 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5385 FlexibleWorkGang* workers = gch->workers(); 5386 assert(workers != NULL, "Need parallel worker threads."); 5387 CMSRefProcTaskProxy rp_task(task, &_collector, 5388 _collector.ref_processor()->span(), 5389 _collector.markBitMap(), 5390 workers, _collector.task_queues()); 5391 workers->run_task(&rp_task); 5392 } 5393 5394 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5395 { 5396 5397 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5398 FlexibleWorkGang* workers = gch->workers(); 5399 assert(workers != NULL, "Need parallel worker threads."); 5400 CMSRefEnqueueTaskProxy enq_task(task); 5401 workers->run_task(&enq_task); 5402 } 5403 5404 void CMSCollector::refProcessingWork() { 5405 ResourceMark rm; 5406 HandleMark hm; 5407 5408 ReferenceProcessor* rp = ref_processor(); 5409 assert(rp->span().equals(_span), "Spans should be equal"); 5410 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5411 // Process weak references. 5412 rp->setup_policy(false); 5413 verify_work_stacks_empty(); 5414 5415 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5416 &_markStack, false /* !preclean */); 5417 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5418 _span, &_markBitMap, &_markStack, 5419 &cmsKeepAliveClosure, false /* !preclean */); 5420 { 5421 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5422 5423 ReferenceProcessorStats stats; 5424 if (rp->processing_is_mt()) { 5425 // Set the degree of MT here. If the discovery is done MT, there 5426 // may have been a different number of threads doing the discovery 5427 // and a different number of discovered lists may have Ref objects. 5428 // That is OK as long as the Reference lists are balanced (see 5429 // balance_all_queues() and balance_queues()). 5430 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5431 int active_workers = ParallelGCThreads; 5432 FlexibleWorkGang* workers = gch->workers(); 5433 if (workers != NULL) { 5434 active_workers = workers->active_workers(); 5435 // The expectation is that active_workers will have already 5436 // been set to a reasonable value. If it has not been set, 5437 // investigate. 5438 assert(active_workers > 0, "Should have been set during scavenge"); 5439 } 5440 rp->set_active_mt_degree(active_workers); 5441 CMSRefProcTaskExecutor task_executor(*this); 5442 stats = rp->process_discovered_references(&_is_alive_closure, 5443 &cmsKeepAliveClosure, 5444 &cmsDrainMarkingStackClosure, 5445 &task_executor, 5446 _gc_timer_cm, 5447 _gc_tracer_cm->gc_id()); 5448 } else { 5449 stats = rp->process_discovered_references(&_is_alive_closure, 5450 &cmsKeepAliveClosure, 5451 &cmsDrainMarkingStackClosure, 5452 NULL, 5453 _gc_timer_cm, 5454 _gc_tracer_cm->gc_id()); 5455 } 5456 _gc_tracer_cm->report_gc_reference_stats(stats); 5457 5458 } 5459 5460 // This is the point where the entire marking should have completed. 5461 verify_work_stacks_empty(); 5462 5463 if (should_unload_classes()) { 5464 { 5465 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5466 5467 // Unload classes and purge the SystemDictionary. 5468 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5469 5470 // Unload nmethods. 5471 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5472 5473 // Prune dead klasses from subklass/sibling/implementor lists. 5474 Klass::clean_weak_klass_links(&_is_alive_closure); 5475 } 5476 5477 { 5478 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5479 // Clean up unreferenced symbols in symbol table. 5480 SymbolTable::unlink(); 5481 } 5482 5483 { 5484 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5485 // Delete entries for dead interned strings. 5486 StringTable::unlink(&_is_alive_closure); 5487 } 5488 } 5489 5490 5491 // Restore any preserved marks as a result of mark stack or 5492 // work queue overflow 5493 restore_preserved_marks_if_any(); // done single-threaded for now 5494 5495 rp->set_enqueuing_is_done(true); 5496 if (rp->processing_is_mt()) { 5497 rp->balance_all_queues(); 5498 CMSRefProcTaskExecutor task_executor(*this); 5499 rp->enqueue_discovered_references(&task_executor); 5500 } else { 5501 rp->enqueue_discovered_references(NULL); 5502 } 5503 rp->verify_no_references_recorded(); 5504 assert(!rp->discovery_enabled(), "should have been disabled"); 5505 } 5506 5507 #ifndef PRODUCT 5508 void CMSCollector::check_correct_thread_executing() { 5509 Thread* t = Thread::current(); 5510 // Only the VM thread or the CMS thread should be here. 5511 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5512 "Unexpected thread type"); 5513 // If this is the vm thread, the foreground process 5514 // should not be waiting. Note that _foregroundGCIsActive is 5515 // true while the foreground collector is waiting. 5516 if (_foregroundGCShouldWait) { 5517 // We cannot be the VM thread 5518 assert(t->is_ConcurrentGC_thread(), 5519 "Should be CMS thread"); 5520 } else { 5521 // We can be the CMS thread only if we are in a stop-world 5522 // phase of CMS collection. 5523 if (t->is_ConcurrentGC_thread()) { 5524 assert(_collectorState == InitialMarking || 5525 _collectorState == FinalMarking, 5526 "Should be a stop-world phase"); 5527 // The CMS thread should be holding the CMS_token. 5528 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5529 "Potential interference with concurrently " 5530 "executing VM thread"); 5531 } 5532 } 5533 } 5534 #endif 5535 5536 void CMSCollector::sweep() { 5537 assert(_collectorState == Sweeping, "just checking"); 5538 check_correct_thread_executing(); 5539 verify_work_stacks_empty(); 5540 verify_overflow_empty(); 5541 increment_sweep_count(); 5542 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5543 5544 _inter_sweep_timer.stop(); 5545 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5546 5547 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5548 _intra_sweep_timer.reset(); 5549 _intra_sweep_timer.start(); 5550 { 5551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5552 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5553 // First sweep the old gen 5554 { 5555 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5556 bitMapLock()); 5557 sweepWork(_cmsGen); 5558 } 5559 5560 // Update Universe::_heap_*_at_gc figures. 5561 // We need all the free list locks to make the abstract state 5562 // transition from Sweeping to Resetting. See detailed note 5563 // further below. 5564 { 5565 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5566 // Update heap occupancy information which is used as 5567 // input to soft ref clearing policy at the next gc. 5568 Universe::update_heap_info_at_gc(); 5569 _collectorState = Resizing; 5570 } 5571 } 5572 verify_work_stacks_empty(); 5573 verify_overflow_empty(); 5574 5575 if (should_unload_classes()) { 5576 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5577 // requires that the virtual spaces are stable and not deleted. 5578 ClassLoaderDataGraph::set_should_purge(true); 5579 } 5580 5581 _intra_sweep_timer.stop(); 5582 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5583 5584 _inter_sweep_timer.reset(); 5585 _inter_sweep_timer.start(); 5586 5587 // We need to use a monotonically non-decreasing time in ms 5588 // or we will see time-warp warnings and os::javaTimeMillis() 5589 // does not guarantee monotonicity. 5590 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5591 update_time_of_last_gc(now); 5592 5593 // NOTE on abstract state transitions: 5594 // Mutators allocate-live and/or mark the mod-union table dirty 5595 // based on the state of the collection. The former is done in 5596 // the interval [Marking, Sweeping] and the latter in the interval 5597 // [Marking, Sweeping). Thus the transitions into the Marking state 5598 // and out of the Sweeping state must be synchronously visible 5599 // globally to the mutators. 5600 // The transition into the Marking state happens with the world 5601 // stopped so the mutators will globally see it. Sweeping is 5602 // done asynchronously by the background collector so the transition 5603 // from the Sweeping state to the Resizing state must be done 5604 // under the freelistLock (as is the check for whether to 5605 // allocate-live and whether to dirty the mod-union table). 5606 assert(_collectorState == Resizing, "Change of collector state to" 5607 " Resizing must be done under the freelistLocks (plural)"); 5608 5609 // Now that sweeping has been completed, we clear 5610 // the incremental_collection_failed flag, 5611 // thus inviting a younger gen collection to promote into 5612 // this generation. If such a promotion may still fail, 5613 // the flag will be set again when a young collection is 5614 // attempted. 5615 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5616 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5617 gch->update_full_collections_completed(_collection_count_start); 5618 } 5619 5620 // FIX ME!!! Looks like this belongs in CFLSpace, with 5621 // CMSGen merely delegating to it. 5622 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5623 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5624 HeapWord* minAddr = _cmsSpace->bottom(); 5625 HeapWord* largestAddr = 5626 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5627 if (largestAddr == NULL) { 5628 // The dictionary appears to be empty. In this case 5629 // try to coalesce at the end of the heap. 5630 largestAddr = _cmsSpace->end(); 5631 } 5632 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5633 size_t nearLargestOffset = 5634 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5635 if (PrintFLSStatistics != 0) { 5636 gclog_or_tty->print_cr( 5637 "CMS: Large Block: " PTR_FORMAT ";" 5638 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5639 p2i(largestAddr), 5640 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5641 } 5642 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5643 } 5644 5645 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5646 return addr >= _cmsSpace->nearLargestChunk(); 5647 } 5648 5649 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5650 return _cmsSpace->find_chunk_at_end(); 5651 } 5652 5653 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5654 bool full) { 5655 // The next lower level has been collected. Gather any statistics 5656 // that are of interest at this point. 5657 if (!full && (current_level + 1) == level()) { 5658 // Gather statistics on the young generation collection. 5659 collector()->stats().record_gc0_end(used()); 5660 } 5661 } 5662 5663 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5664 // We iterate over the space(s) underlying this generation, 5665 // checking the mark bit map to see if the bits corresponding 5666 // to specific blocks are marked or not. Blocks that are 5667 // marked are live and are not swept up. All remaining blocks 5668 // are swept up, with coalescing on-the-fly as we sweep up 5669 // contiguous free and/or garbage blocks: 5670 // We need to ensure that the sweeper synchronizes with allocators 5671 // and stop-the-world collectors. In particular, the following 5672 // locks are used: 5673 // . CMS token: if this is held, a stop the world collection cannot occur 5674 // . freelistLock: if this is held no allocation can occur from this 5675 // generation by another thread 5676 // . bitMapLock: if this is held, no other thread can access or update 5677 // 5678 5679 // Note that we need to hold the freelistLock if we use 5680 // block iterate below; else the iterator might go awry if 5681 // a mutator (or promotion) causes block contents to change 5682 // (for instance if the allocator divvies up a block). 5683 // If we hold the free list lock, for all practical purposes 5684 // young generation GC's can't occur (they'll usually need to 5685 // promote), so we might as well prevent all young generation 5686 // GC's while we do a sweeping step. For the same reason, we might 5687 // as well take the bit map lock for the entire duration 5688 5689 // check that we hold the requisite locks 5690 assert(have_cms_token(), "Should hold cms token"); 5691 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5692 assert_lock_strong(gen->freelistLock()); 5693 assert_lock_strong(bitMapLock()); 5694 5695 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5696 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5697 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5698 _inter_sweep_estimate.padded_average(), 5699 _intra_sweep_estimate.padded_average()); 5700 gen->setNearLargestChunk(); 5701 5702 { 5703 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5704 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5705 // We need to free-up/coalesce garbage/blocks from a 5706 // co-terminal free run. This is done in the SweepClosure 5707 // destructor; so, do not remove this scope, else the 5708 // end-of-sweep-census below will be off by a little bit. 5709 } 5710 gen->cmsSpace()->sweep_completed(); 5711 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5712 if (should_unload_classes()) { // unloaded classes this cycle, 5713 _concurrent_cycles_since_last_unload = 0; // ... reset count 5714 } else { // did not unload classes, 5715 _concurrent_cycles_since_last_unload++; // ... increment count 5716 } 5717 } 5718 5719 // Reset CMS data structures (for now just the marking bit map) 5720 // preparatory for the next cycle. 5721 void CMSCollector::reset(bool concurrent) { 5722 if (concurrent) { 5723 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5724 5725 // If the state is not "Resetting", the foreground thread 5726 // has done a collection and the resetting. 5727 if (_collectorState != Resetting) { 5728 assert(_collectorState == Idling, "The state should only change" 5729 " because the foreground collector has finished the collection"); 5730 return; 5731 } 5732 5733 // Clear the mark bitmap (no grey objects to start with) 5734 // for the next cycle. 5735 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5736 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5737 5738 HeapWord* curAddr = _markBitMap.startWord(); 5739 while (curAddr < _markBitMap.endWord()) { 5740 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5741 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5742 _markBitMap.clear_large_range(chunk); 5743 if (ConcurrentMarkSweepThread::should_yield() && 5744 !foregroundGCIsActive() && 5745 CMSYield) { 5746 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5747 "CMS thread should hold CMS token"); 5748 assert_lock_strong(bitMapLock()); 5749 bitMapLock()->unlock(); 5750 ConcurrentMarkSweepThread::desynchronize(true); 5751 stopTimer(); 5752 if (PrintCMSStatistics != 0) { 5753 incrementYields(); 5754 } 5755 5756 // See the comment in coordinator_yield() 5757 for (unsigned i = 0; i < CMSYieldSleepCount && 5758 ConcurrentMarkSweepThread::should_yield() && 5759 !CMSCollector::foregroundGCIsActive(); ++i) { 5760 os::sleep(Thread::current(), 1, false); 5761 } 5762 5763 ConcurrentMarkSweepThread::synchronize(true); 5764 bitMapLock()->lock_without_safepoint_check(); 5765 startTimer(); 5766 } 5767 curAddr = chunk.end(); 5768 } 5769 // A successful mostly concurrent collection has been done. 5770 // Because only the full (i.e., concurrent mode failure) collections 5771 // are being measured for gc overhead limits, clean the "near" flag 5772 // and count. 5773 size_policy()->reset_gc_overhead_limit_count(); 5774 _collectorState = Idling; 5775 } else { 5776 // already have the lock 5777 assert(_collectorState == Resetting, "just checking"); 5778 assert_lock_strong(bitMapLock()); 5779 _markBitMap.clear_all(); 5780 _collectorState = Idling; 5781 } 5782 5783 register_gc_end(); 5784 } 5785 5786 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5787 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5788 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5789 TraceCollectorStats tcs(counters()); 5790 5791 switch (op) { 5792 case CMS_op_checkpointRootsInitial: { 5793 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5794 checkpointRootsInitial(); 5795 if (PrintGC) { 5796 _cmsGen->printOccupancy("initial-mark"); 5797 } 5798 break; 5799 } 5800 case CMS_op_checkpointRootsFinal: { 5801 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5802 checkpointRootsFinal(); 5803 if (PrintGC) { 5804 _cmsGen->printOccupancy("remark"); 5805 } 5806 break; 5807 } 5808 default: 5809 fatal("No such CMS_op"); 5810 } 5811 } 5812 5813 #ifndef PRODUCT 5814 size_t const CMSCollector::skip_header_HeapWords() { 5815 return FreeChunk::header_size(); 5816 } 5817 5818 // Try and collect here conditions that should hold when 5819 // CMS thread is exiting. The idea is that the foreground GC 5820 // thread should not be blocked if it wants to terminate 5821 // the CMS thread and yet continue to run the VM for a while 5822 // after that. 5823 void CMSCollector::verify_ok_to_terminate() const { 5824 assert(Thread::current()->is_ConcurrentGC_thread(), 5825 "should be called by CMS thread"); 5826 assert(!_foregroundGCShouldWait, "should be false"); 5827 // We could check here that all the various low-level locks 5828 // are not held by the CMS thread, but that is overkill; see 5829 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5830 // is checked. 5831 } 5832 #endif 5833 5834 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5835 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5836 "missing Printezis mark?"); 5837 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5838 size_t size = pointer_delta(nextOneAddr + 1, addr); 5839 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5840 "alignment problem"); 5841 assert(size >= 3, "Necessary for Printezis marks to work"); 5842 return size; 5843 } 5844 5845 // A variant of the above (block_size_using_printezis_bits()) except 5846 // that we return 0 if the P-bits are not yet set. 5847 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5848 if (_markBitMap.isMarked(addr + 1)) { 5849 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5850 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5851 size_t size = pointer_delta(nextOneAddr + 1, addr); 5852 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5853 "alignment problem"); 5854 assert(size >= 3, "Necessary for Printezis marks to work"); 5855 return size; 5856 } 5857 return 0; 5858 } 5859 5860 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5861 size_t sz = 0; 5862 oop p = (oop)addr; 5863 if (p->klass_or_null() != NULL) { 5864 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5865 } else { 5866 sz = block_size_using_printezis_bits(addr); 5867 } 5868 assert(sz > 0, "size must be nonzero"); 5869 HeapWord* next_block = addr + sz; 5870 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5871 CardTableModRefBS::card_size); 5872 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5873 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5874 "must be different cards"); 5875 return next_card; 5876 } 5877 5878 5879 // CMS Bit Map Wrapper ///////////////////////////////////////// 5880 5881 // Construct a CMS bit map infrastructure, but don't create the 5882 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5883 // further below. 5884 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5885 _bm(), 5886 _shifter(shifter), 5887 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5888 Monitor::_safepoint_check_sometimes) : NULL) 5889 { 5890 _bmStartWord = 0; 5891 _bmWordSize = 0; 5892 } 5893 5894 bool CMSBitMap::allocate(MemRegion mr) { 5895 _bmStartWord = mr.start(); 5896 _bmWordSize = mr.word_size(); 5897 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5898 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5899 if (!brs.is_reserved()) { 5900 warning("CMS bit map allocation failure"); 5901 return false; 5902 } 5903 // For now we'll just commit all of the bit map up front. 5904 // Later on we'll try to be more parsimonious with swap. 5905 if (!_virtual_space.initialize(brs, brs.size())) { 5906 warning("CMS bit map backing store failure"); 5907 return false; 5908 } 5909 assert(_virtual_space.committed_size() == brs.size(), 5910 "didn't reserve backing store for all of CMS bit map?"); 5911 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5912 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5913 _bmWordSize, "inconsistency in bit map sizing"); 5914 _bm.set_size(_bmWordSize >> _shifter); 5915 5916 // bm.clear(); // can we rely on getting zero'd memory? verify below 5917 assert(isAllClear(), 5918 "Expected zero'd memory from ReservedSpace constructor"); 5919 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5920 "consistency check"); 5921 return true; 5922 } 5923 5924 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5925 HeapWord *next_addr, *end_addr, *last_addr; 5926 assert_locked(); 5927 assert(covers(mr), "out-of-range error"); 5928 // XXX assert that start and end are appropriately aligned 5929 for (next_addr = mr.start(), end_addr = mr.end(); 5930 next_addr < end_addr; next_addr = last_addr) { 5931 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5932 last_addr = dirty_region.end(); 5933 if (!dirty_region.is_empty()) { 5934 cl->do_MemRegion(dirty_region); 5935 } else { 5936 assert(last_addr == end_addr, "program logic"); 5937 return; 5938 } 5939 } 5940 } 5941 5942 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5943 _bm.print_on_error(st, prefix); 5944 } 5945 5946 #ifndef PRODUCT 5947 void CMSBitMap::assert_locked() const { 5948 CMSLockVerifier::assert_locked(lock()); 5949 } 5950 5951 bool CMSBitMap::covers(MemRegion mr) const { 5952 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5953 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5954 "size inconsistency"); 5955 return (mr.start() >= _bmStartWord) && 5956 (mr.end() <= endWord()); 5957 } 5958 5959 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5960 return (start >= _bmStartWord && (start + size) <= endWord()); 5961 } 5962 5963 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5964 // verify that there are no 1 bits in the interval [left, right) 5965 FalseBitMapClosure falseBitMapClosure; 5966 iterate(&falseBitMapClosure, left, right); 5967 } 5968 5969 void CMSBitMap::region_invariant(MemRegion mr) 5970 { 5971 assert_locked(); 5972 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5973 assert(!mr.is_empty(), "unexpected empty region"); 5974 assert(covers(mr), "mr should be covered by bit map"); 5975 // convert address range into offset range 5976 size_t start_ofs = heapWordToOffset(mr.start()); 5977 // Make sure that end() is appropriately aligned 5978 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5979 (1 << (_shifter+LogHeapWordSize))), 5980 "Misaligned mr.end()"); 5981 size_t end_ofs = heapWordToOffset(mr.end()); 5982 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5983 } 5984 5985 #endif 5986 5987 bool CMSMarkStack::allocate(size_t size) { 5988 // allocate a stack of the requisite depth 5989 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5990 size * sizeof(oop))); 5991 if (!rs.is_reserved()) { 5992 warning("CMSMarkStack allocation failure"); 5993 return false; 5994 } 5995 if (!_virtual_space.initialize(rs, rs.size())) { 5996 warning("CMSMarkStack backing store failure"); 5997 return false; 5998 } 5999 assert(_virtual_space.committed_size() == rs.size(), 6000 "didn't reserve backing store for all of CMS stack?"); 6001 _base = (oop*)(_virtual_space.low()); 6002 _index = 0; 6003 _capacity = size; 6004 NOT_PRODUCT(_max_depth = 0); 6005 return true; 6006 } 6007 6008 // XXX FIX ME !!! In the MT case we come in here holding a 6009 // leaf lock. For printing we need to take a further lock 6010 // which has lower rank. We need to recalibrate the two 6011 // lock-ranks involved in order to be able to print the 6012 // messages below. (Or defer the printing to the caller. 6013 // For now we take the expedient path of just disabling the 6014 // messages for the problematic case.) 6015 void CMSMarkStack::expand() { 6016 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6017 if (_capacity == MarkStackSizeMax) { 6018 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6019 // We print a warning message only once per CMS cycle. 6020 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6021 } 6022 return; 6023 } 6024 // Double capacity if possible 6025 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6026 // Do not give up existing stack until we have managed to 6027 // get the double capacity that we desired. 6028 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6029 new_capacity * sizeof(oop))); 6030 if (rs.is_reserved()) { 6031 // Release the backing store associated with old stack 6032 _virtual_space.release(); 6033 // Reinitialize virtual space for new stack 6034 if (!_virtual_space.initialize(rs, rs.size())) { 6035 fatal("Not enough swap for expanded marking stack"); 6036 } 6037 _base = (oop*)(_virtual_space.low()); 6038 _index = 0; 6039 _capacity = new_capacity; 6040 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6041 // Failed to double capacity, continue; 6042 // we print a detail message only once per CMS cycle. 6043 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6044 SIZE_FORMAT"K", 6045 _capacity / K, new_capacity / K); 6046 } 6047 } 6048 6049 6050 // Closures 6051 // XXX: there seems to be a lot of code duplication here; 6052 // should refactor and consolidate common code. 6053 6054 // This closure is used to mark refs into the CMS generation in 6055 // the CMS bit map. Called at the first checkpoint. This closure 6056 // assumes that we do not need to re-mark dirty cards; if the CMS 6057 // generation on which this is used is not an oldest 6058 // generation then this will lose younger_gen cards! 6059 6060 MarkRefsIntoClosure::MarkRefsIntoClosure( 6061 MemRegion span, CMSBitMap* bitMap): 6062 _span(span), 6063 _bitMap(bitMap) 6064 { 6065 assert(_ref_processor == NULL, "deliberately left NULL"); 6066 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6067 } 6068 6069 void MarkRefsIntoClosure::do_oop(oop obj) { 6070 // if p points into _span, then mark corresponding bit in _markBitMap 6071 assert(obj->is_oop(), "expected an oop"); 6072 HeapWord* addr = (HeapWord*)obj; 6073 if (_span.contains(addr)) { 6074 // this should be made more efficient 6075 _bitMap->mark(addr); 6076 } 6077 } 6078 6079 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6080 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6081 6082 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6083 MemRegion span, CMSBitMap* bitMap): 6084 _span(span), 6085 _bitMap(bitMap) 6086 { 6087 assert(_ref_processor == NULL, "deliberately left NULL"); 6088 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6089 } 6090 6091 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6092 // if p points into _span, then mark corresponding bit in _markBitMap 6093 assert(obj->is_oop(), "expected an oop"); 6094 HeapWord* addr = (HeapWord*)obj; 6095 if (_span.contains(addr)) { 6096 // this should be made more efficient 6097 _bitMap->par_mark(addr); 6098 } 6099 } 6100 6101 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6102 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6103 6104 // A variant of the above, used for CMS marking verification. 6105 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6106 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6107 _span(span), 6108 _verification_bm(verification_bm), 6109 _cms_bm(cms_bm) 6110 { 6111 assert(_ref_processor == NULL, "deliberately left NULL"); 6112 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6113 } 6114 6115 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6116 // if p points into _span, then mark corresponding bit in _markBitMap 6117 assert(obj->is_oop(), "expected an oop"); 6118 HeapWord* addr = (HeapWord*)obj; 6119 if (_span.contains(addr)) { 6120 _verification_bm->mark(addr); 6121 if (!_cms_bm->isMarked(addr)) { 6122 oop(addr)->print(); 6123 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6124 fatal("... aborting"); 6125 } 6126 } 6127 } 6128 6129 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6130 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6131 6132 ////////////////////////////////////////////////// 6133 // MarkRefsIntoAndScanClosure 6134 ////////////////////////////////////////////////// 6135 6136 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6137 ReferenceProcessor* rp, 6138 CMSBitMap* bit_map, 6139 CMSBitMap* mod_union_table, 6140 CMSMarkStack* mark_stack, 6141 CMSCollector* collector, 6142 bool should_yield, 6143 bool concurrent_precleaning): 6144 _collector(collector), 6145 _span(span), 6146 _bit_map(bit_map), 6147 _mark_stack(mark_stack), 6148 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6149 mark_stack, concurrent_precleaning), 6150 _yield(should_yield), 6151 _concurrent_precleaning(concurrent_precleaning), 6152 _freelistLock(NULL) 6153 { 6154 _ref_processor = rp; 6155 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6156 } 6157 6158 // This closure is used to mark refs into the CMS generation at the 6159 // second (final) checkpoint, and to scan and transitively follow 6160 // the unmarked oops. It is also used during the concurrent precleaning 6161 // phase while scanning objects on dirty cards in the CMS generation. 6162 // The marks are made in the marking bit map and the marking stack is 6163 // used for keeping the (newly) grey objects during the scan. 6164 // The parallel version (Par_...) appears further below. 6165 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6166 if (obj != NULL) { 6167 assert(obj->is_oop(), "expected an oop"); 6168 HeapWord* addr = (HeapWord*)obj; 6169 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6170 assert(_collector->overflow_list_is_empty(), 6171 "overflow list should be empty"); 6172 if (_span.contains(addr) && 6173 !_bit_map->isMarked(addr)) { 6174 // mark bit map (object is now grey) 6175 _bit_map->mark(addr); 6176 // push on marking stack (stack should be empty), and drain the 6177 // stack by applying this closure to the oops in the oops popped 6178 // from the stack (i.e. blacken the grey objects) 6179 bool res = _mark_stack->push(obj); 6180 assert(res, "Should have space to push on empty stack"); 6181 do { 6182 oop new_oop = _mark_stack->pop(); 6183 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6184 assert(_bit_map->isMarked((HeapWord*)new_oop), 6185 "only grey objects on this stack"); 6186 // iterate over the oops in this oop, marking and pushing 6187 // the ones in CMS heap (i.e. in _span). 6188 new_oop->oop_iterate(&_pushAndMarkClosure); 6189 // check if it's time to yield 6190 do_yield_check(); 6191 } while (!_mark_stack->isEmpty() || 6192 (!_concurrent_precleaning && take_from_overflow_list())); 6193 // if marking stack is empty, and we are not doing this 6194 // during precleaning, then check the overflow list 6195 } 6196 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6197 assert(_collector->overflow_list_is_empty(), 6198 "overflow list was drained above"); 6199 // We could restore evacuated mark words, if any, used for 6200 // overflow list links here because the overflow list is 6201 // provably empty here. That would reduce the maximum 6202 // size requirements for preserved_{oop,mark}_stack. 6203 // But we'll just postpone it until we are all done 6204 // so we can just stream through. 6205 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6206 _collector->restore_preserved_marks_if_any(); 6207 assert(_collector->no_preserved_marks(), "No preserved marks"); 6208 } 6209 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6210 "All preserved marks should have been restored above"); 6211 } 6212 } 6213 6214 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6215 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6216 6217 void MarkRefsIntoAndScanClosure::do_yield_work() { 6218 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6219 "CMS thread should hold CMS token"); 6220 assert_lock_strong(_freelistLock); 6221 assert_lock_strong(_bit_map->lock()); 6222 // relinquish the free_list_lock and bitMaplock() 6223 _bit_map->lock()->unlock(); 6224 _freelistLock->unlock(); 6225 ConcurrentMarkSweepThread::desynchronize(true); 6226 _collector->stopTimer(); 6227 if (PrintCMSStatistics != 0) { 6228 _collector->incrementYields(); 6229 } 6230 6231 // See the comment in coordinator_yield() 6232 for (unsigned i = 0; 6233 i < CMSYieldSleepCount && 6234 ConcurrentMarkSweepThread::should_yield() && 6235 !CMSCollector::foregroundGCIsActive(); 6236 ++i) { 6237 os::sleep(Thread::current(), 1, false); 6238 } 6239 6240 ConcurrentMarkSweepThread::synchronize(true); 6241 _freelistLock->lock_without_safepoint_check(); 6242 _bit_map->lock()->lock_without_safepoint_check(); 6243 _collector->startTimer(); 6244 } 6245 6246 /////////////////////////////////////////////////////////// 6247 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6248 // MarkRefsIntoAndScanClosure 6249 /////////////////////////////////////////////////////////// 6250 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6251 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6252 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6253 _span(span), 6254 _bit_map(bit_map), 6255 _work_queue(work_queue), 6256 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6257 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6258 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6259 { 6260 _ref_processor = rp; 6261 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6262 } 6263 6264 // This closure is used to mark refs into the CMS generation at the 6265 // second (final) checkpoint, and to scan and transitively follow 6266 // the unmarked oops. The marks are made in the marking bit map and 6267 // the work_queue is used for keeping the (newly) grey objects during 6268 // the scan phase whence they are also available for stealing by parallel 6269 // threads. Since the marking bit map is shared, updates are 6270 // synchronized (via CAS). 6271 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6272 if (obj != NULL) { 6273 // Ignore mark word because this could be an already marked oop 6274 // that may be chained at the end of the overflow list. 6275 assert(obj->is_oop(true), "expected an oop"); 6276 HeapWord* addr = (HeapWord*)obj; 6277 if (_span.contains(addr) && 6278 !_bit_map->isMarked(addr)) { 6279 // mark bit map (object will become grey): 6280 // It is possible for several threads to be 6281 // trying to "claim" this object concurrently; 6282 // the unique thread that succeeds in marking the 6283 // object first will do the subsequent push on 6284 // to the work queue (or overflow list). 6285 if (_bit_map->par_mark(addr)) { 6286 // push on work_queue (which may not be empty), and trim the 6287 // queue to an appropriate length by applying this closure to 6288 // the oops in the oops popped from the stack (i.e. blacken the 6289 // grey objects) 6290 bool res = _work_queue->push(obj); 6291 assert(res, "Low water mark should be less than capacity?"); 6292 trim_queue(_low_water_mark); 6293 } // Else, another thread claimed the object 6294 } 6295 } 6296 } 6297 6298 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6299 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6300 6301 // This closure is used to rescan the marked objects on the dirty cards 6302 // in the mod union table and the card table proper. 6303 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6304 oop p, MemRegion mr) { 6305 6306 size_t size = 0; 6307 HeapWord* addr = (HeapWord*)p; 6308 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6309 assert(_span.contains(addr), "we are scanning the CMS generation"); 6310 // check if it's time to yield 6311 if (do_yield_check()) { 6312 // We yielded for some foreground stop-world work, 6313 // and we have been asked to abort this ongoing preclean cycle. 6314 return 0; 6315 } 6316 if (_bitMap->isMarked(addr)) { 6317 // it's marked; is it potentially uninitialized? 6318 if (p->klass_or_null() != NULL) { 6319 // an initialized object; ignore mark word in verification below 6320 // since we are running concurrent with mutators 6321 assert(p->is_oop(true), "should be an oop"); 6322 if (p->is_objArray()) { 6323 // objArrays are precisely marked; restrict scanning 6324 // to dirty cards only. 6325 size = CompactibleFreeListSpace::adjustObjectSize( 6326 p->oop_iterate(_scanningClosure, mr)); 6327 } else { 6328 // A non-array may have been imprecisely marked; we need 6329 // to scan object in its entirety. 6330 size = CompactibleFreeListSpace::adjustObjectSize( 6331 p->oop_iterate(_scanningClosure)); 6332 } 6333 #ifdef ASSERT 6334 size_t direct_size = 6335 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6336 assert(size == direct_size, "Inconsistency in size"); 6337 assert(size >= 3, "Necessary for Printezis marks to work"); 6338 if (!_bitMap->isMarked(addr+1)) { 6339 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6340 } else { 6341 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6342 assert(_bitMap->isMarked(addr+size-1), 6343 "inconsistent Printezis mark"); 6344 } 6345 #endif // ASSERT 6346 } else { 6347 // An uninitialized object. 6348 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6349 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6350 size = pointer_delta(nextOneAddr + 1, addr); 6351 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6352 "alignment problem"); 6353 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6354 // will dirty the card when the klass pointer is installed in the 6355 // object (signaling the completion of initialization). 6356 } 6357 } else { 6358 // Either a not yet marked object or an uninitialized object 6359 if (p->klass_or_null() == NULL) { 6360 // An uninitialized object, skip to the next card, since 6361 // we may not be able to read its P-bits yet. 6362 assert(size == 0, "Initial value"); 6363 } else { 6364 // An object not (yet) reached by marking: we merely need to 6365 // compute its size so as to go look at the next block. 6366 assert(p->is_oop(true), "should be an oop"); 6367 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6368 } 6369 } 6370 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6371 return size; 6372 } 6373 6374 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6375 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6376 "CMS thread should hold CMS token"); 6377 assert_lock_strong(_freelistLock); 6378 assert_lock_strong(_bitMap->lock()); 6379 // relinquish the free_list_lock and bitMaplock() 6380 _bitMap->lock()->unlock(); 6381 _freelistLock->unlock(); 6382 ConcurrentMarkSweepThread::desynchronize(true); 6383 _collector->stopTimer(); 6384 if (PrintCMSStatistics != 0) { 6385 _collector->incrementYields(); 6386 } 6387 6388 // See the comment in coordinator_yield() 6389 for (unsigned i = 0; i < CMSYieldSleepCount && 6390 ConcurrentMarkSweepThread::should_yield() && 6391 !CMSCollector::foregroundGCIsActive(); ++i) { 6392 os::sleep(Thread::current(), 1, false); 6393 } 6394 6395 ConcurrentMarkSweepThread::synchronize(true); 6396 _freelistLock->lock_without_safepoint_check(); 6397 _bitMap->lock()->lock_without_safepoint_check(); 6398 _collector->startTimer(); 6399 } 6400 6401 6402 ////////////////////////////////////////////////////////////////// 6403 // SurvivorSpacePrecleanClosure 6404 ////////////////////////////////////////////////////////////////// 6405 // This (single-threaded) closure is used to preclean the oops in 6406 // the survivor spaces. 6407 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6408 6409 HeapWord* addr = (HeapWord*)p; 6410 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6411 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6412 assert(p->klass_or_null() != NULL, "object should be initialized"); 6413 // an initialized object; ignore mark word in verification below 6414 // since we are running concurrent with mutators 6415 assert(p->is_oop(true), "should be an oop"); 6416 // Note that we do not yield while we iterate over 6417 // the interior oops of p, pushing the relevant ones 6418 // on our marking stack. 6419 size_t size = p->oop_iterate(_scanning_closure); 6420 do_yield_check(); 6421 // Observe that below, we do not abandon the preclean 6422 // phase as soon as we should; rather we empty the 6423 // marking stack before returning. This is to satisfy 6424 // some existing assertions. In general, it may be a 6425 // good idea to abort immediately and complete the marking 6426 // from the grey objects at a later time. 6427 while (!_mark_stack->isEmpty()) { 6428 oop new_oop = _mark_stack->pop(); 6429 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6430 assert(_bit_map->isMarked((HeapWord*)new_oop), 6431 "only grey objects on this stack"); 6432 // iterate over the oops in this oop, marking and pushing 6433 // the ones in CMS heap (i.e. in _span). 6434 new_oop->oop_iterate(_scanning_closure); 6435 // check if it's time to yield 6436 do_yield_check(); 6437 } 6438 unsigned int after_count = 6439 GenCollectedHeap::heap()->total_collections(); 6440 bool abort = (_before_count != after_count) || 6441 _collector->should_abort_preclean(); 6442 return abort ? 0 : size; 6443 } 6444 6445 void SurvivorSpacePrecleanClosure::do_yield_work() { 6446 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6447 "CMS thread should hold CMS token"); 6448 assert_lock_strong(_bit_map->lock()); 6449 // Relinquish the bit map lock 6450 _bit_map->lock()->unlock(); 6451 ConcurrentMarkSweepThread::desynchronize(true); 6452 _collector->stopTimer(); 6453 if (PrintCMSStatistics != 0) { 6454 _collector->incrementYields(); 6455 } 6456 6457 // See the comment in coordinator_yield() 6458 for (unsigned i = 0; i < CMSYieldSleepCount && 6459 ConcurrentMarkSweepThread::should_yield() && 6460 !CMSCollector::foregroundGCIsActive(); ++i) { 6461 os::sleep(Thread::current(), 1, false); 6462 } 6463 6464 ConcurrentMarkSweepThread::synchronize(true); 6465 _bit_map->lock()->lock_without_safepoint_check(); 6466 _collector->startTimer(); 6467 } 6468 6469 // This closure is used to rescan the marked objects on the dirty cards 6470 // in the mod union table and the card table proper. In the parallel 6471 // case, although the bitMap is shared, we do a single read so the 6472 // isMarked() query is "safe". 6473 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6474 // Ignore mark word because we are running concurrent with mutators 6475 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6476 HeapWord* addr = (HeapWord*)p; 6477 assert(_span.contains(addr), "we are scanning the CMS generation"); 6478 bool is_obj_array = false; 6479 #ifdef ASSERT 6480 if (!_parallel) { 6481 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6482 assert(_collector->overflow_list_is_empty(), 6483 "overflow list should be empty"); 6484 6485 } 6486 #endif // ASSERT 6487 if (_bit_map->isMarked(addr)) { 6488 // Obj arrays are precisely marked, non-arrays are not; 6489 // so we scan objArrays precisely and non-arrays in their 6490 // entirety. 6491 if (p->is_objArray()) { 6492 is_obj_array = true; 6493 if (_parallel) { 6494 p->oop_iterate(_par_scan_closure, mr); 6495 } else { 6496 p->oop_iterate(_scan_closure, mr); 6497 } 6498 } else { 6499 if (_parallel) { 6500 p->oop_iterate(_par_scan_closure); 6501 } else { 6502 p->oop_iterate(_scan_closure); 6503 } 6504 } 6505 } 6506 #ifdef ASSERT 6507 if (!_parallel) { 6508 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6509 assert(_collector->overflow_list_is_empty(), 6510 "overflow list should be empty"); 6511 6512 } 6513 #endif // ASSERT 6514 return is_obj_array; 6515 } 6516 6517 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6518 MemRegion span, 6519 CMSBitMap* bitMap, CMSMarkStack* markStack, 6520 bool should_yield, bool verifying): 6521 _collector(collector), 6522 _span(span), 6523 _bitMap(bitMap), 6524 _mut(&collector->_modUnionTable), 6525 _markStack(markStack), 6526 _yield(should_yield), 6527 _skipBits(0) 6528 { 6529 assert(_markStack->isEmpty(), "stack should be empty"); 6530 _finger = _bitMap->startWord(); 6531 _threshold = _finger; 6532 assert(_collector->_restart_addr == NULL, "Sanity check"); 6533 assert(_span.contains(_finger), "Out of bounds _finger?"); 6534 DEBUG_ONLY(_verifying = verifying;) 6535 } 6536 6537 void MarkFromRootsClosure::reset(HeapWord* addr) { 6538 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6539 assert(_span.contains(addr), "Out of bounds _finger?"); 6540 _finger = addr; 6541 _threshold = (HeapWord*)round_to( 6542 (intptr_t)_finger, CardTableModRefBS::card_size); 6543 } 6544 6545 // Should revisit to see if this should be restructured for 6546 // greater efficiency. 6547 bool MarkFromRootsClosure::do_bit(size_t offset) { 6548 if (_skipBits > 0) { 6549 _skipBits--; 6550 return true; 6551 } 6552 // convert offset into a HeapWord* 6553 HeapWord* addr = _bitMap->startWord() + offset; 6554 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6555 "address out of range"); 6556 assert(_bitMap->isMarked(addr), "tautology"); 6557 if (_bitMap->isMarked(addr+1)) { 6558 // this is an allocated but not yet initialized object 6559 assert(_skipBits == 0, "tautology"); 6560 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6561 oop p = oop(addr); 6562 if (p->klass_or_null() == NULL) { 6563 DEBUG_ONLY(if (!_verifying) {) 6564 // We re-dirty the cards on which this object lies and increase 6565 // the _threshold so that we'll come back to scan this object 6566 // during the preclean or remark phase. (CMSCleanOnEnter) 6567 if (CMSCleanOnEnter) { 6568 size_t sz = _collector->block_size_using_printezis_bits(addr); 6569 HeapWord* end_card_addr = (HeapWord*)round_to( 6570 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6571 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6572 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6573 // Bump _threshold to end_card_addr; note that 6574 // _threshold cannot possibly exceed end_card_addr, anyhow. 6575 // This prevents future clearing of the card as the scan proceeds 6576 // to the right. 6577 assert(_threshold <= end_card_addr, 6578 "Because we are just scanning into this object"); 6579 if (_threshold < end_card_addr) { 6580 _threshold = end_card_addr; 6581 } 6582 if (p->klass_or_null() != NULL) { 6583 // Redirty the range of cards... 6584 _mut->mark_range(redirty_range); 6585 } // ...else the setting of klass will dirty the card anyway. 6586 } 6587 DEBUG_ONLY(}) 6588 return true; 6589 } 6590 } 6591 scanOopsInOop(addr); 6592 return true; 6593 } 6594 6595 // We take a break if we've been at this for a while, 6596 // so as to avoid monopolizing the locks involved. 6597 void MarkFromRootsClosure::do_yield_work() { 6598 // First give up the locks, then yield, then re-lock 6599 // We should probably use a constructor/destructor idiom to 6600 // do this unlock/lock or modify the MutexUnlocker class to 6601 // serve our purpose. XXX 6602 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6603 "CMS thread should hold CMS token"); 6604 assert_lock_strong(_bitMap->lock()); 6605 _bitMap->lock()->unlock(); 6606 ConcurrentMarkSweepThread::desynchronize(true); 6607 _collector->stopTimer(); 6608 if (PrintCMSStatistics != 0) { 6609 _collector->incrementYields(); 6610 } 6611 6612 // See the comment in coordinator_yield() 6613 for (unsigned i = 0; i < CMSYieldSleepCount && 6614 ConcurrentMarkSweepThread::should_yield() && 6615 !CMSCollector::foregroundGCIsActive(); ++i) { 6616 os::sleep(Thread::current(), 1, false); 6617 } 6618 6619 ConcurrentMarkSweepThread::synchronize(true); 6620 _bitMap->lock()->lock_without_safepoint_check(); 6621 _collector->startTimer(); 6622 } 6623 6624 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6625 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6626 assert(_markStack->isEmpty(), 6627 "should drain stack to limit stack usage"); 6628 // convert ptr to an oop preparatory to scanning 6629 oop obj = oop(ptr); 6630 // Ignore mark word in verification below, since we 6631 // may be running concurrent with mutators. 6632 assert(obj->is_oop(true), "should be an oop"); 6633 assert(_finger <= ptr, "_finger runneth ahead"); 6634 // advance the finger to right end of this object 6635 _finger = ptr + obj->size(); 6636 assert(_finger > ptr, "we just incremented it above"); 6637 // On large heaps, it may take us some time to get through 6638 // the marking phase. During 6639 // this time it's possible that a lot of mutations have 6640 // accumulated in the card table and the mod union table -- 6641 // these mutation records are redundant until we have 6642 // actually traced into the corresponding card. 6643 // Here, we check whether advancing the finger would make 6644 // us cross into a new card, and if so clear corresponding 6645 // cards in the MUT (preclean them in the card-table in the 6646 // future). 6647 6648 DEBUG_ONLY(if (!_verifying) {) 6649 // The clean-on-enter optimization is disabled by default, 6650 // until we fix 6178663. 6651 if (CMSCleanOnEnter && (_finger > _threshold)) { 6652 // [_threshold, _finger) represents the interval 6653 // of cards to be cleared in MUT (or precleaned in card table). 6654 // The set of cards to be cleared is all those that overlap 6655 // with the interval [_threshold, _finger); note that 6656 // _threshold is always kept card-aligned but _finger isn't 6657 // always card-aligned. 6658 HeapWord* old_threshold = _threshold; 6659 assert(old_threshold == (HeapWord*)round_to( 6660 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6661 "_threshold should always be card-aligned"); 6662 _threshold = (HeapWord*)round_to( 6663 (intptr_t)_finger, CardTableModRefBS::card_size); 6664 MemRegion mr(old_threshold, _threshold); 6665 assert(!mr.is_empty(), "Control point invariant"); 6666 assert(_span.contains(mr), "Should clear within span"); 6667 _mut->clear_range(mr); 6668 } 6669 DEBUG_ONLY(}) 6670 // Note: the finger doesn't advance while we drain 6671 // the stack below. 6672 PushOrMarkClosure pushOrMarkClosure(_collector, 6673 _span, _bitMap, _markStack, 6674 _finger, this); 6675 bool res = _markStack->push(obj); 6676 assert(res, "Empty non-zero size stack should have space for single push"); 6677 while (!_markStack->isEmpty()) { 6678 oop new_oop = _markStack->pop(); 6679 // Skip verifying header mark word below because we are 6680 // running concurrent with mutators. 6681 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6682 // now scan this oop's oops 6683 new_oop->oop_iterate(&pushOrMarkClosure); 6684 do_yield_check(); 6685 } 6686 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6687 } 6688 6689 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6690 CMSCollector* collector, MemRegion span, 6691 CMSBitMap* bit_map, 6692 OopTaskQueue* work_queue, 6693 CMSMarkStack* overflow_stack): 6694 _collector(collector), 6695 _whole_span(collector->_span), 6696 _span(span), 6697 _bit_map(bit_map), 6698 _mut(&collector->_modUnionTable), 6699 _work_queue(work_queue), 6700 _overflow_stack(overflow_stack), 6701 _skip_bits(0), 6702 _task(task) 6703 { 6704 assert(_work_queue->size() == 0, "work_queue should be empty"); 6705 _finger = span.start(); 6706 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6707 assert(_span.contains(_finger), "Out of bounds _finger?"); 6708 } 6709 6710 // Should revisit to see if this should be restructured for 6711 // greater efficiency. 6712 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6713 if (_skip_bits > 0) { 6714 _skip_bits--; 6715 return true; 6716 } 6717 // convert offset into a HeapWord* 6718 HeapWord* addr = _bit_map->startWord() + offset; 6719 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6720 "address out of range"); 6721 assert(_bit_map->isMarked(addr), "tautology"); 6722 if (_bit_map->isMarked(addr+1)) { 6723 // this is an allocated object that might not yet be initialized 6724 assert(_skip_bits == 0, "tautology"); 6725 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6726 oop p = oop(addr); 6727 if (p->klass_or_null() == NULL) { 6728 // in the case of Clean-on-Enter optimization, redirty card 6729 // and avoid clearing card by increasing the threshold. 6730 return true; 6731 } 6732 } 6733 scan_oops_in_oop(addr); 6734 return true; 6735 } 6736 6737 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6738 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6739 // Should we assert that our work queue is empty or 6740 // below some drain limit? 6741 assert(_work_queue->size() == 0, 6742 "should drain stack to limit stack usage"); 6743 // convert ptr to an oop preparatory to scanning 6744 oop obj = oop(ptr); 6745 // Ignore mark word in verification below, since we 6746 // may be running concurrent with mutators. 6747 assert(obj->is_oop(true), "should be an oop"); 6748 assert(_finger <= ptr, "_finger runneth ahead"); 6749 // advance the finger to right end of this object 6750 _finger = ptr + obj->size(); 6751 assert(_finger > ptr, "we just incremented it above"); 6752 // On large heaps, it may take us some time to get through 6753 // the marking phase. During 6754 // this time it's possible that a lot of mutations have 6755 // accumulated in the card table and the mod union table -- 6756 // these mutation records are redundant until we have 6757 // actually traced into the corresponding card. 6758 // Here, we check whether advancing the finger would make 6759 // us cross into a new card, and if so clear corresponding 6760 // cards in the MUT (preclean them in the card-table in the 6761 // future). 6762 6763 // The clean-on-enter optimization is disabled by default, 6764 // until we fix 6178663. 6765 if (CMSCleanOnEnter && (_finger > _threshold)) { 6766 // [_threshold, _finger) represents the interval 6767 // of cards to be cleared in MUT (or precleaned in card table). 6768 // The set of cards to be cleared is all those that overlap 6769 // with the interval [_threshold, _finger); note that 6770 // _threshold is always kept card-aligned but _finger isn't 6771 // always card-aligned. 6772 HeapWord* old_threshold = _threshold; 6773 assert(old_threshold == (HeapWord*)round_to( 6774 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6775 "_threshold should always be card-aligned"); 6776 _threshold = (HeapWord*)round_to( 6777 (intptr_t)_finger, CardTableModRefBS::card_size); 6778 MemRegion mr(old_threshold, _threshold); 6779 assert(!mr.is_empty(), "Control point invariant"); 6780 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6781 _mut->clear_range(mr); 6782 } 6783 6784 // Note: the local finger doesn't advance while we drain 6785 // the stack below, but the global finger sure can and will. 6786 HeapWord** gfa = _task->global_finger_addr(); 6787 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6788 _span, _bit_map, 6789 _work_queue, 6790 _overflow_stack, 6791 _finger, 6792 gfa, this); 6793 bool res = _work_queue->push(obj); // overflow could occur here 6794 assert(res, "Will hold once we use workqueues"); 6795 while (true) { 6796 oop new_oop; 6797 if (!_work_queue->pop_local(new_oop)) { 6798 // We emptied our work_queue; check if there's stuff that can 6799 // be gotten from the overflow stack. 6800 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6801 _overflow_stack, _work_queue)) { 6802 do_yield_check(); 6803 continue; 6804 } else { // done 6805 break; 6806 } 6807 } 6808 // Skip verifying header mark word below because we are 6809 // running concurrent with mutators. 6810 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6811 // now scan this oop's oops 6812 new_oop->oop_iterate(&pushOrMarkClosure); 6813 do_yield_check(); 6814 } 6815 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6816 } 6817 6818 // Yield in response to a request from VM Thread or 6819 // from mutators. 6820 void Par_MarkFromRootsClosure::do_yield_work() { 6821 assert(_task != NULL, "sanity"); 6822 _task->yield(); 6823 } 6824 6825 // A variant of the above used for verifying CMS marking work. 6826 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6827 MemRegion span, 6828 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6829 CMSMarkStack* mark_stack): 6830 _collector(collector), 6831 _span(span), 6832 _verification_bm(verification_bm), 6833 _cms_bm(cms_bm), 6834 _mark_stack(mark_stack), 6835 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6836 mark_stack) 6837 { 6838 assert(_mark_stack->isEmpty(), "stack should be empty"); 6839 _finger = _verification_bm->startWord(); 6840 assert(_collector->_restart_addr == NULL, "Sanity check"); 6841 assert(_span.contains(_finger), "Out of bounds _finger?"); 6842 } 6843 6844 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6845 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6846 assert(_span.contains(addr), "Out of bounds _finger?"); 6847 _finger = addr; 6848 } 6849 6850 // Should revisit to see if this should be restructured for 6851 // greater efficiency. 6852 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6853 // convert offset into a HeapWord* 6854 HeapWord* addr = _verification_bm->startWord() + offset; 6855 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6856 "address out of range"); 6857 assert(_verification_bm->isMarked(addr), "tautology"); 6858 assert(_cms_bm->isMarked(addr), "tautology"); 6859 6860 assert(_mark_stack->isEmpty(), 6861 "should drain stack to limit stack usage"); 6862 // convert addr to an oop preparatory to scanning 6863 oop obj = oop(addr); 6864 assert(obj->is_oop(), "should be an oop"); 6865 assert(_finger <= addr, "_finger runneth ahead"); 6866 // advance the finger to right end of this object 6867 _finger = addr + obj->size(); 6868 assert(_finger > addr, "we just incremented it above"); 6869 // Note: the finger doesn't advance while we drain 6870 // the stack below. 6871 bool res = _mark_stack->push(obj); 6872 assert(res, "Empty non-zero size stack should have space for single push"); 6873 while (!_mark_stack->isEmpty()) { 6874 oop new_oop = _mark_stack->pop(); 6875 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6876 // now scan this oop's oops 6877 new_oop->oop_iterate(&_pam_verify_closure); 6878 } 6879 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6880 return true; 6881 } 6882 6883 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6884 CMSCollector* collector, MemRegion span, 6885 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6886 CMSMarkStack* mark_stack): 6887 MetadataAwareOopClosure(collector->ref_processor()), 6888 _collector(collector), 6889 _span(span), 6890 _verification_bm(verification_bm), 6891 _cms_bm(cms_bm), 6892 _mark_stack(mark_stack) 6893 { } 6894 6895 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6896 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6897 6898 // Upon stack overflow, we discard (part of) the stack, 6899 // remembering the least address amongst those discarded 6900 // in CMSCollector's _restart_address. 6901 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6902 // Remember the least grey address discarded 6903 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6904 _collector->lower_restart_addr(ra); 6905 _mark_stack->reset(); // discard stack contents 6906 _mark_stack->expand(); // expand the stack if possible 6907 } 6908 6909 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6910 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6911 HeapWord* addr = (HeapWord*)obj; 6912 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6913 // Oop lies in _span and isn't yet grey or black 6914 _verification_bm->mark(addr); // now grey 6915 if (!_cms_bm->isMarked(addr)) { 6916 oop(addr)->print(); 6917 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6918 p2i(addr)); 6919 fatal("... aborting"); 6920 } 6921 6922 if (!_mark_stack->push(obj)) { // stack overflow 6923 if (PrintCMSStatistics != 0) { 6924 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6925 SIZE_FORMAT, _mark_stack->capacity()); 6926 } 6927 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6928 handle_stack_overflow(addr); 6929 } 6930 // anything including and to the right of _finger 6931 // will be scanned as we iterate over the remainder of the 6932 // bit map 6933 } 6934 } 6935 6936 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6937 MemRegion span, 6938 CMSBitMap* bitMap, CMSMarkStack* markStack, 6939 HeapWord* finger, MarkFromRootsClosure* parent) : 6940 MetadataAwareOopClosure(collector->ref_processor()), 6941 _collector(collector), 6942 _span(span), 6943 _bitMap(bitMap), 6944 _markStack(markStack), 6945 _finger(finger), 6946 _parent(parent) 6947 { } 6948 6949 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6950 MemRegion span, 6951 CMSBitMap* bit_map, 6952 OopTaskQueue* work_queue, 6953 CMSMarkStack* overflow_stack, 6954 HeapWord* finger, 6955 HeapWord** global_finger_addr, 6956 Par_MarkFromRootsClosure* parent) : 6957 MetadataAwareOopClosure(collector->ref_processor()), 6958 _collector(collector), 6959 _whole_span(collector->_span), 6960 _span(span), 6961 _bit_map(bit_map), 6962 _work_queue(work_queue), 6963 _overflow_stack(overflow_stack), 6964 _finger(finger), 6965 _global_finger_addr(global_finger_addr), 6966 _parent(parent) 6967 { } 6968 6969 // Assumes thread-safe access by callers, who are 6970 // responsible for mutual exclusion. 6971 void CMSCollector::lower_restart_addr(HeapWord* low) { 6972 assert(_span.contains(low), "Out of bounds addr"); 6973 if (_restart_addr == NULL) { 6974 _restart_addr = low; 6975 } else { 6976 _restart_addr = MIN2(_restart_addr, low); 6977 } 6978 } 6979 6980 // Upon stack overflow, we discard (part of) the stack, 6981 // remembering the least address amongst those discarded 6982 // in CMSCollector's _restart_address. 6983 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6984 // Remember the least grey address discarded 6985 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6986 _collector->lower_restart_addr(ra); 6987 _markStack->reset(); // discard stack contents 6988 _markStack->expand(); // expand the stack if possible 6989 } 6990 6991 // Upon stack overflow, we discard (part of) the stack, 6992 // remembering the least address amongst those discarded 6993 // in CMSCollector's _restart_address. 6994 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6995 // We need to do this under a mutex to prevent other 6996 // workers from interfering with the work done below. 6997 MutexLockerEx ml(_overflow_stack->par_lock(), 6998 Mutex::_no_safepoint_check_flag); 6999 // Remember the least grey address discarded 7000 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7001 _collector->lower_restart_addr(ra); 7002 _overflow_stack->reset(); // discard stack contents 7003 _overflow_stack->expand(); // expand the stack if possible 7004 } 7005 7006 void PushOrMarkClosure::do_oop(oop obj) { 7007 // Ignore mark word because we are running concurrent with mutators. 7008 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7009 HeapWord* addr = (HeapWord*)obj; 7010 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7011 // Oop lies in _span and isn't yet grey or black 7012 _bitMap->mark(addr); // now grey 7013 if (addr < _finger) { 7014 // the bit map iteration has already either passed, or 7015 // sampled, this bit in the bit map; we'll need to 7016 // use the marking stack to scan this oop's oops. 7017 bool simulate_overflow = false; 7018 NOT_PRODUCT( 7019 if (CMSMarkStackOverflowALot && 7020 _collector->simulate_overflow()) { 7021 // simulate a stack overflow 7022 simulate_overflow = true; 7023 } 7024 ) 7025 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7026 if (PrintCMSStatistics != 0) { 7027 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7028 SIZE_FORMAT, _markStack->capacity()); 7029 } 7030 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7031 handle_stack_overflow(addr); 7032 } 7033 } 7034 // anything including and to the right of _finger 7035 // will be scanned as we iterate over the remainder of the 7036 // bit map 7037 do_yield_check(); 7038 } 7039 } 7040 7041 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7042 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7043 7044 void Par_PushOrMarkClosure::do_oop(oop obj) { 7045 // Ignore mark word because we are running concurrent with mutators. 7046 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7047 HeapWord* addr = (HeapWord*)obj; 7048 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7049 // Oop lies in _span and isn't yet grey or black 7050 // We read the global_finger (volatile read) strictly after marking oop 7051 bool res = _bit_map->par_mark(addr); // now grey 7052 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7053 // Should we push this marked oop on our stack? 7054 // -- if someone else marked it, nothing to do 7055 // -- if target oop is above global finger nothing to do 7056 // -- if target oop is in chunk and above local finger 7057 // then nothing to do 7058 // -- else push on work queue 7059 if ( !res // someone else marked it, they will deal with it 7060 || (addr >= *gfa) // will be scanned in a later task 7061 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7062 return; 7063 } 7064 // the bit map iteration has already either passed, or 7065 // sampled, this bit in the bit map; we'll need to 7066 // use the marking stack to scan this oop's oops. 7067 bool simulate_overflow = false; 7068 NOT_PRODUCT( 7069 if (CMSMarkStackOverflowALot && 7070 _collector->simulate_overflow()) { 7071 // simulate a stack overflow 7072 simulate_overflow = true; 7073 } 7074 ) 7075 if (simulate_overflow || 7076 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7077 // stack overflow 7078 if (PrintCMSStatistics != 0) { 7079 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7080 SIZE_FORMAT, _overflow_stack->capacity()); 7081 } 7082 // We cannot assert that the overflow stack is full because 7083 // it may have been emptied since. 7084 assert(simulate_overflow || 7085 _work_queue->size() == _work_queue->max_elems(), 7086 "Else push should have succeeded"); 7087 handle_stack_overflow(addr); 7088 } 7089 do_yield_check(); 7090 } 7091 } 7092 7093 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7094 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7095 7096 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7097 MemRegion span, 7098 ReferenceProcessor* rp, 7099 CMSBitMap* bit_map, 7100 CMSBitMap* mod_union_table, 7101 CMSMarkStack* mark_stack, 7102 bool concurrent_precleaning): 7103 MetadataAwareOopClosure(rp), 7104 _collector(collector), 7105 _span(span), 7106 _bit_map(bit_map), 7107 _mod_union_table(mod_union_table), 7108 _mark_stack(mark_stack), 7109 _concurrent_precleaning(concurrent_precleaning) 7110 { 7111 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7112 } 7113 7114 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7115 // the non-parallel version (the parallel version appears further below.) 7116 void PushAndMarkClosure::do_oop(oop obj) { 7117 // Ignore mark word verification. If during concurrent precleaning, 7118 // the object monitor may be locked. If during the checkpoint 7119 // phases, the object may already have been reached by a different 7120 // path and may be at the end of the global overflow list (so 7121 // the mark word may be NULL). 7122 assert(obj->is_oop_or_null(true /* ignore mark word */), 7123 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7124 HeapWord* addr = (HeapWord*)obj; 7125 // Check if oop points into the CMS generation 7126 // and is not marked 7127 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7128 // a white object ... 7129 _bit_map->mark(addr); // ... now grey 7130 // push on the marking stack (grey set) 7131 bool simulate_overflow = false; 7132 NOT_PRODUCT( 7133 if (CMSMarkStackOverflowALot && 7134 _collector->simulate_overflow()) { 7135 // simulate a stack overflow 7136 simulate_overflow = true; 7137 } 7138 ) 7139 if (simulate_overflow || !_mark_stack->push(obj)) { 7140 if (_concurrent_precleaning) { 7141 // During precleaning we can just dirty the appropriate card(s) 7142 // in the mod union table, thus ensuring that the object remains 7143 // in the grey set and continue. In the case of object arrays 7144 // we need to dirty all of the cards that the object spans, 7145 // since the rescan of object arrays will be limited to the 7146 // dirty cards. 7147 // Note that no one can be interfering with us in this action 7148 // of dirtying the mod union table, so no locking or atomics 7149 // are required. 7150 if (obj->is_objArray()) { 7151 size_t sz = obj->size(); 7152 HeapWord* end_card_addr = (HeapWord*)round_to( 7153 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7154 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7155 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7156 _mod_union_table->mark_range(redirty_range); 7157 } else { 7158 _mod_union_table->mark(addr); 7159 } 7160 _collector->_ser_pmc_preclean_ovflw++; 7161 } else { 7162 // During the remark phase, we need to remember this oop 7163 // in the overflow list. 7164 _collector->push_on_overflow_list(obj); 7165 _collector->_ser_pmc_remark_ovflw++; 7166 } 7167 } 7168 } 7169 } 7170 7171 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7172 MemRegion span, 7173 ReferenceProcessor* rp, 7174 CMSBitMap* bit_map, 7175 OopTaskQueue* work_queue): 7176 MetadataAwareOopClosure(rp), 7177 _collector(collector), 7178 _span(span), 7179 _bit_map(bit_map), 7180 _work_queue(work_queue) 7181 { 7182 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7183 } 7184 7185 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7186 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7187 7188 // Grey object rescan during second checkpoint phase -- 7189 // the parallel version. 7190 void Par_PushAndMarkClosure::do_oop(oop obj) { 7191 // In the assert below, we ignore the mark word because 7192 // this oop may point to an already visited object that is 7193 // on the overflow stack (in which case the mark word has 7194 // been hijacked for chaining into the overflow stack -- 7195 // if this is the last object in the overflow stack then 7196 // its mark word will be NULL). Because this object may 7197 // have been subsequently popped off the global overflow 7198 // stack, and the mark word possibly restored to the prototypical 7199 // value, by the time we get to examined this failing assert in 7200 // the debugger, is_oop_or_null(false) may subsequently start 7201 // to hold. 7202 assert(obj->is_oop_or_null(true), 7203 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7204 HeapWord* addr = (HeapWord*)obj; 7205 // Check if oop points into the CMS generation 7206 // and is not marked 7207 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7208 // a white object ... 7209 // If we manage to "claim" the object, by being the 7210 // first thread to mark it, then we push it on our 7211 // marking stack 7212 if (_bit_map->par_mark(addr)) { // ... now grey 7213 // push on work queue (grey set) 7214 bool simulate_overflow = false; 7215 NOT_PRODUCT( 7216 if (CMSMarkStackOverflowALot && 7217 _collector->par_simulate_overflow()) { 7218 // simulate a stack overflow 7219 simulate_overflow = true; 7220 } 7221 ) 7222 if (simulate_overflow || !_work_queue->push(obj)) { 7223 _collector->par_push_on_overflow_list(obj); 7224 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7225 } 7226 } // Else, some other thread got there first 7227 } 7228 } 7229 7230 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7231 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7232 7233 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7234 Mutex* bml = _collector->bitMapLock(); 7235 assert_lock_strong(bml); 7236 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7237 "CMS thread should hold CMS token"); 7238 7239 bml->unlock(); 7240 ConcurrentMarkSweepThread::desynchronize(true); 7241 7242 _collector->stopTimer(); 7243 if (PrintCMSStatistics != 0) { 7244 _collector->incrementYields(); 7245 } 7246 7247 // See the comment in coordinator_yield() 7248 for (unsigned i = 0; i < CMSYieldSleepCount && 7249 ConcurrentMarkSweepThread::should_yield() && 7250 !CMSCollector::foregroundGCIsActive(); ++i) { 7251 os::sleep(Thread::current(), 1, false); 7252 } 7253 7254 ConcurrentMarkSweepThread::synchronize(true); 7255 bml->lock(); 7256 7257 _collector->startTimer(); 7258 } 7259 7260 bool CMSPrecleanRefsYieldClosure::should_return() { 7261 if (ConcurrentMarkSweepThread::should_yield()) { 7262 do_yield_work(); 7263 } 7264 return _collector->foregroundGCIsActive(); 7265 } 7266 7267 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7268 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7269 "mr should be aligned to start at a card boundary"); 7270 // We'd like to assert: 7271 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7272 // "mr should be a range of cards"); 7273 // However, that would be too strong in one case -- the last 7274 // partition ends at _unallocated_block which, in general, can be 7275 // an arbitrary boundary, not necessarily card aligned. 7276 if (PrintCMSStatistics != 0) { 7277 _num_dirty_cards += 7278 mr.word_size()/CardTableModRefBS::card_size_in_words; 7279 } 7280 _space->object_iterate_mem(mr, &_scan_cl); 7281 } 7282 7283 SweepClosure::SweepClosure(CMSCollector* collector, 7284 ConcurrentMarkSweepGeneration* g, 7285 CMSBitMap* bitMap, bool should_yield) : 7286 _collector(collector), 7287 _g(g), 7288 _sp(g->cmsSpace()), 7289 _limit(_sp->sweep_limit()), 7290 _freelistLock(_sp->freelistLock()), 7291 _bitMap(bitMap), 7292 _yield(should_yield), 7293 _inFreeRange(false), // No free range at beginning of sweep 7294 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7295 _lastFreeRangeCoalesced(false), 7296 _freeFinger(g->used_region().start()) 7297 { 7298 NOT_PRODUCT( 7299 _numObjectsFreed = 0; 7300 _numWordsFreed = 0; 7301 _numObjectsLive = 0; 7302 _numWordsLive = 0; 7303 _numObjectsAlreadyFree = 0; 7304 _numWordsAlreadyFree = 0; 7305 _last_fc = NULL; 7306 7307 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7308 _sp->dictionary()->initialize_dict_returned_bytes(); 7309 ) 7310 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7311 "sweep _limit out of bounds"); 7312 if (CMSTraceSweeper) { 7313 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7314 p2i(_limit)); 7315 } 7316 } 7317 7318 void SweepClosure::print_on(outputStream* st) const { 7319 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7320 p2i(_sp->bottom()), p2i(_sp->end())); 7321 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7322 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7323 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7324 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7325 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7326 } 7327 7328 #ifndef PRODUCT 7329 // Assertion checking only: no useful work in product mode -- 7330 // however, if any of the flags below become product flags, 7331 // you may need to review this code to see if it needs to be 7332 // enabled in product mode. 7333 SweepClosure::~SweepClosure() { 7334 assert_lock_strong(_freelistLock); 7335 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7336 "sweep _limit out of bounds"); 7337 if (inFreeRange()) { 7338 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7339 print(); 7340 ShouldNotReachHere(); 7341 } 7342 if (Verbose && PrintGC) { 7343 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7344 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7345 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7346 SIZE_FORMAT" bytes " 7347 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7348 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7349 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7350 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7351 * sizeof(HeapWord); 7352 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7353 7354 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7355 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7356 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7357 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7358 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7359 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7360 indexListReturnedBytes); 7361 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7362 dict_returned_bytes); 7363 } 7364 } 7365 if (CMSTraceSweeper) { 7366 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7367 p2i(_limit)); 7368 } 7369 } 7370 #endif // PRODUCT 7371 7372 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7373 bool freeRangeInFreeLists) { 7374 if (CMSTraceSweeper) { 7375 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7376 p2i(freeFinger), freeRangeInFreeLists); 7377 } 7378 assert(!inFreeRange(), "Trampling existing free range"); 7379 set_inFreeRange(true); 7380 set_lastFreeRangeCoalesced(false); 7381 7382 set_freeFinger(freeFinger); 7383 set_freeRangeInFreeLists(freeRangeInFreeLists); 7384 if (CMSTestInFreeList) { 7385 if (freeRangeInFreeLists) { 7386 FreeChunk* fc = (FreeChunk*) freeFinger; 7387 assert(fc->is_free(), "A chunk on the free list should be free."); 7388 assert(fc->size() > 0, "Free range should have a size"); 7389 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7390 } 7391 } 7392 } 7393 7394 // Note that the sweeper runs concurrently with mutators. Thus, 7395 // it is possible for direct allocation in this generation to happen 7396 // in the middle of the sweep. Note that the sweeper also coalesces 7397 // contiguous free blocks. Thus, unless the sweeper and the allocator 7398 // synchronize appropriately freshly allocated blocks may get swept up. 7399 // This is accomplished by the sweeper locking the free lists while 7400 // it is sweeping. Thus blocks that are determined to be free are 7401 // indeed free. There is however one additional complication: 7402 // blocks that have been allocated since the final checkpoint and 7403 // mark, will not have been marked and so would be treated as 7404 // unreachable and swept up. To prevent this, the allocator marks 7405 // the bit map when allocating during the sweep phase. This leads, 7406 // however, to a further complication -- objects may have been allocated 7407 // but not yet initialized -- in the sense that the header isn't yet 7408 // installed. The sweeper can not then determine the size of the block 7409 // in order to skip over it. To deal with this case, we use a technique 7410 // (due to Printezis) to encode such uninitialized block sizes in the 7411 // bit map. Since the bit map uses a bit per every HeapWord, but the 7412 // CMS generation has a minimum object size of 3 HeapWords, it follows 7413 // that "normal marks" won't be adjacent in the bit map (there will 7414 // always be at least two 0 bits between successive 1 bits). We make use 7415 // of these "unused" bits to represent uninitialized blocks -- the bit 7416 // corresponding to the start of the uninitialized object and the next 7417 // bit are both set. Finally, a 1 bit marks the end of the object that 7418 // started with the two consecutive 1 bits to indicate its potentially 7419 // uninitialized state. 7420 7421 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7422 FreeChunk* fc = (FreeChunk*)addr; 7423 size_t res; 7424 7425 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7426 // than "addr == _limit" because although _limit was a block boundary when 7427 // we started the sweep, it may no longer be one because heap expansion 7428 // may have caused us to coalesce the block ending at the address _limit 7429 // with a newly expanded chunk (this happens when _limit was set to the 7430 // previous _end of the space), so we may have stepped past _limit: 7431 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7432 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7433 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7434 "sweep _limit out of bounds"); 7435 assert(addr < _sp->end(), "addr out of bounds"); 7436 // Flush any free range we might be holding as a single 7437 // coalesced chunk to the appropriate free list. 7438 if (inFreeRange()) { 7439 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7440 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7441 flush_cur_free_chunk(freeFinger(), 7442 pointer_delta(addr, freeFinger())); 7443 if (CMSTraceSweeper) { 7444 gclog_or_tty->print("Sweep: last chunk: "); 7445 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7446 "[coalesced:%d]\n", 7447 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7448 lastFreeRangeCoalesced() ? 1 : 0); 7449 } 7450 } 7451 7452 // help the iterator loop finish 7453 return pointer_delta(_sp->end(), addr); 7454 } 7455 7456 assert(addr < _limit, "sweep invariant"); 7457 // check if we should yield 7458 do_yield_check(addr); 7459 if (fc->is_free()) { 7460 // Chunk that is already free 7461 res = fc->size(); 7462 do_already_free_chunk(fc); 7463 debug_only(_sp->verifyFreeLists()); 7464 // If we flush the chunk at hand in lookahead_and_flush() 7465 // and it's coalesced with a preceding chunk, then the 7466 // process of "mangling" the payload of the coalesced block 7467 // will cause erasure of the size information from the 7468 // (erstwhile) header of all the coalesced blocks but the 7469 // first, so the first disjunct in the assert will not hold 7470 // in that specific case (in which case the second disjunct 7471 // will hold). 7472 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7473 "Otherwise the size info doesn't change at this step"); 7474 NOT_PRODUCT( 7475 _numObjectsAlreadyFree++; 7476 _numWordsAlreadyFree += res; 7477 ) 7478 NOT_PRODUCT(_last_fc = fc;) 7479 } else if (!_bitMap->isMarked(addr)) { 7480 // Chunk is fresh garbage 7481 res = do_garbage_chunk(fc); 7482 debug_only(_sp->verifyFreeLists()); 7483 NOT_PRODUCT( 7484 _numObjectsFreed++; 7485 _numWordsFreed += res; 7486 ) 7487 } else { 7488 // Chunk that is alive. 7489 res = do_live_chunk(fc); 7490 debug_only(_sp->verifyFreeLists()); 7491 NOT_PRODUCT( 7492 _numObjectsLive++; 7493 _numWordsLive += res; 7494 ) 7495 } 7496 return res; 7497 } 7498 7499 // For the smart allocation, record following 7500 // split deaths - a free chunk is removed from its free list because 7501 // it is being split into two or more chunks. 7502 // split birth - a free chunk is being added to its free list because 7503 // a larger free chunk has been split and resulted in this free chunk. 7504 // coal death - a free chunk is being removed from its free list because 7505 // it is being coalesced into a large free chunk. 7506 // coal birth - a free chunk is being added to its free list because 7507 // it was created when two or more free chunks where coalesced into 7508 // this free chunk. 7509 // 7510 // These statistics are used to determine the desired number of free 7511 // chunks of a given size. The desired number is chosen to be relative 7512 // to the end of a CMS sweep. The desired number at the end of a sweep 7513 // is the 7514 // count-at-end-of-previous-sweep (an amount that was enough) 7515 // - count-at-beginning-of-current-sweep (the excess) 7516 // + split-births (gains in this size during interval) 7517 // - split-deaths (demands on this size during interval) 7518 // where the interval is from the end of one sweep to the end of the 7519 // next. 7520 // 7521 // When sweeping the sweeper maintains an accumulated chunk which is 7522 // the chunk that is made up of chunks that have been coalesced. That 7523 // will be termed the left-hand chunk. A new chunk of garbage that 7524 // is being considered for coalescing will be referred to as the 7525 // right-hand chunk. 7526 // 7527 // When making a decision on whether to coalesce a right-hand chunk with 7528 // the current left-hand chunk, the current count vs. the desired count 7529 // of the left-hand chunk is considered. Also if the right-hand chunk 7530 // is near the large chunk at the end of the heap (see 7531 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7532 // left-hand chunk is coalesced. 7533 // 7534 // When making a decision about whether to split a chunk, the desired count 7535 // vs. the current count of the candidate to be split is also considered. 7536 // If the candidate is underpopulated (currently fewer chunks than desired) 7537 // a chunk of an overpopulated (currently more chunks than desired) size may 7538 // be chosen. The "hint" associated with a free list, if non-null, points 7539 // to a free list which may be overpopulated. 7540 // 7541 7542 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7543 const size_t size = fc->size(); 7544 // Chunks that cannot be coalesced are not in the 7545 // free lists. 7546 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7547 assert(_sp->verify_chunk_in_free_list(fc), 7548 "free chunk should be in free lists"); 7549 } 7550 // a chunk that is already free, should not have been 7551 // marked in the bit map 7552 HeapWord* const addr = (HeapWord*) fc; 7553 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7554 // Verify that the bit map has no bits marked between 7555 // addr and purported end of this block. 7556 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7557 7558 // Some chunks cannot be coalesced under any circumstances. 7559 // See the definition of cantCoalesce(). 7560 if (!fc->cantCoalesce()) { 7561 // This chunk can potentially be coalesced. 7562 if (_sp->adaptive_freelists()) { 7563 // All the work is done in 7564 do_post_free_or_garbage_chunk(fc, size); 7565 } else { // Not adaptive free lists 7566 // this is a free chunk that can potentially be coalesced by the sweeper; 7567 if (!inFreeRange()) { 7568 // if the next chunk is a free block that can't be coalesced 7569 // it doesn't make sense to remove this chunk from the free lists 7570 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7571 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7572 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7573 nextChunk->is_free() && // ... which is free... 7574 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7575 // nothing to do 7576 } else { 7577 // Potentially the start of a new free range: 7578 // Don't eagerly remove it from the free lists. 7579 // No need to remove it if it will just be put 7580 // back again. (Also from a pragmatic point of view 7581 // if it is a free block in a region that is beyond 7582 // any allocated blocks, an assertion will fail) 7583 // Remember the start of a free run. 7584 initialize_free_range(addr, true); 7585 // end - can coalesce with next chunk 7586 } 7587 } else { 7588 // the midst of a free range, we are coalescing 7589 print_free_block_coalesced(fc); 7590 if (CMSTraceSweeper) { 7591 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7592 } 7593 // remove it from the free lists 7594 _sp->removeFreeChunkFromFreeLists(fc); 7595 set_lastFreeRangeCoalesced(true); 7596 // If the chunk is being coalesced and the current free range is 7597 // in the free lists, remove the current free range so that it 7598 // will be returned to the free lists in its entirety - all 7599 // the coalesced pieces included. 7600 if (freeRangeInFreeLists()) { 7601 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7602 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7603 "Size of free range is inconsistent with chunk size."); 7604 if (CMSTestInFreeList) { 7605 assert(_sp->verify_chunk_in_free_list(ffc), 7606 "free range is not in free lists"); 7607 } 7608 _sp->removeFreeChunkFromFreeLists(ffc); 7609 set_freeRangeInFreeLists(false); 7610 } 7611 } 7612 } 7613 // Note that if the chunk is not coalescable (the else arm 7614 // below), we unconditionally flush, without needing to do 7615 // a "lookahead," as we do below. 7616 if (inFreeRange()) lookahead_and_flush(fc, size); 7617 } else { 7618 // Code path common to both original and adaptive free lists. 7619 7620 // cant coalesce with previous block; this should be treated 7621 // as the end of a free run if any 7622 if (inFreeRange()) { 7623 // we kicked some butt; time to pick up the garbage 7624 assert(freeFinger() < addr, "freeFinger points too high"); 7625 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7626 } 7627 // else, nothing to do, just continue 7628 } 7629 } 7630 7631 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7632 // This is a chunk of garbage. It is not in any free list. 7633 // Add it to a free list or let it possibly be coalesced into 7634 // a larger chunk. 7635 HeapWord* const addr = (HeapWord*) fc; 7636 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7637 7638 if (_sp->adaptive_freelists()) { 7639 // Verify that the bit map has no bits marked between 7640 // addr and purported end of just dead object. 7641 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7642 7643 do_post_free_or_garbage_chunk(fc, size); 7644 } else { 7645 if (!inFreeRange()) { 7646 // start of a new free range 7647 assert(size > 0, "A free range should have a size"); 7648 initialize_free_range(addr, false); 7649 } else { 7650 // this will be swept up when we hit the end of the 7651 // free range 7652 if (CMSTraceSweeper) { 7653 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7654 } 7655 // If the chunk is being coalesced and the current free range is 7656 // in the free lists, remove the current free range so that it 7657 // will be returned to the free lists in its entirety - all 7658 // the coalesced pieces included. 7659 if (freeRangeInFreeLists()) { 7660 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7661 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7662 "Size of free range is inconsistent with chunk size."); 7663 if (CMSTestInFreeList) { 7664 assert(_sp->verify_chunk_in_free_list(ffc), 7665 "free range is not in free lists"); 7666 } 7667 _sp->removeFreeChunkFromFreeLists(ffc); 7668 set_freeRangeInFreeLists(false); 7669 } 7670 set_lastFreeRangeCoalesced(true); 7671 } 7672 // this will be swept up when we hit the end of the free range 7673 7674 // Verify that the bit map has no bits marked between 7675 // addr and purported end of just dead object. 7676 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7677 } 7678 assert(_limit >= addr + size, 7679 "A freshly garbage chunk can't possibly straddle over _limit"); 7680 if (inFreeRange()) lookahead_and_flush(fc, size); 7681 return size; 7682 } 7683 7684 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7685 HeapWord* addr = (HeapWord*) fc; 7686 // The sweeper has just found a live object. Return any accumulated 7687 // left hand chunk to the free lists. 7688 if (inFreeRange()) { 7689 assert(freeFinger() < addr, "freeFinger points too high"); 7690 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7691 } 7692 7693 // This object is live: we'd normally expect this to be 7694 // an oop, and like to assert the following: 7695 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7696 // However, as we commented above, this may be an object whose 7697 // header hasn't yet been initialized. 7698 size_t size; 7699 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7700 if (_bitMap->isMarked(addr + 1)) { 7701 // Determine the size from the bit map, rather than trying to 7702 // compute it from the object header. 7703 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7704 size = pointer_delta(nextOneAddr + 1, addr); 7705 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7706 "alignment problem"); 7707 7708 #ifdef ASSERT 7709 if (oop(addr)->klass_or_null() != NULL) { 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 assert(size == 7713 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7714 "P-mark and computed size do not agree"); 7715 } 7716 #endif 7717 7718 } else { 7719 // This should be an initialized object that's alive. 7720 assert(oop(addr)->klass_or_null() != NULL, 7721 "Should be an initialized object"); 7722 // Ignore mark word because we are running concurrent with mutators 7723 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7724 // Verify that the bit map has no bits marked between 7725 // addr and purported end of this block. 7726 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7727 assert(size >= 3, "Necessary for Printezis marks to work"); 7728 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7729 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7730 } 7731 return size; 7732 } 7733 7734 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7735 size_t chunkSize) { 7736 // do_post_free_or_garbage_chunk() should only be called in the case 7737 // of the adaptive free list allocator. 7738 const bool fcInFreeLists = fc->is_free(); 7739 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7740 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7741 if (CMSTestInFreeList && fcInFreeLists) { 7742 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7743 } 7744 7745 if (CMSTraceSweeper) { 7746 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7747 } 7748 7749 HeapWord* const fc_addr = (HeapWord*) fc; 7750 7751 bool coalesce; 7752 const size_t left = pointer_delta(fc_addr, freeFinger()); 7753 const size_t right = chunkSize; 7754 switch (FLSCoalescePolicy) { 7755 // numeric value forms a coalition aggressiveness metric 7756 case 0: { // never coalesce 7757 coalesce = false; 7758 break; 7759 } 7760 case 1: { // coalesce if left & right chunks on overpopulated lists 7761 coalesce = _sp->coalOverPopulated(left) && 7762 _sp->coalOverPopulated(right); 7763 break; 7764 } 7765 case 2: { // coalesce if left chunk on overpopulated list (default) 7766 coalesce = _sp->coalOverPopulated(left); 7767 break; 7768 } 7769 case 3: { // coalesce if left OR right chunk on overpopulated list 7770 coalesce = _sp->coalOverPopulated(left) || 7771 _sp->coalOverPopulated(right); 7772 break; 7773 } 7774 case 4: { // always coalesce 7775 coalesce = true; 7776 break; 7777 } 7778 default: 7779 ShouldNotReachHere(); 7780 } 7781 7782 // Should the current free range be coalesced? 7783 // If the chunk is in a free range and either we decided to coalesce above 7784 // or the chunk is near the large block at the end of the heap 7785 // (isNearLargestChunk() returns true), then coalesce this chunk. 7786 const bool doCoalesce = inFreeRange() 7787 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7788 if (doCoalesce) { 7789 // Coalesce the current free range on the left with the new 7790 // chunk on the right. If either is on a free list, 7791 // it must be removed from the list and stashed in the closure. 7792 if (freeRangeInFreeLists()) { 7793 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7794 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7795 "Size of free range is inconsistent with chunk size."); 7796 if (CMSTestInFreeList) { 7797 assert(_sp->verify_chunk_in_free_list(ffc), 7798 "Chunk is not in free lists"); 7799 } 7800 _sp->coalDeath(ffc->size()); 7801 _sp->removeFreeChunkFromFreeLists(ffc); 7802 set_freeRangeInFreeLists(false); 7803 } 7804 if (fcInFreeLists) { 7805 _sp->coalDeath(chunkSize); 7806 assert(fc->size() == chunkSize, 7807 "The chunk has the wrong size or is not in the free lists"); 7808 _sp->removeFreeChunkFromFreeLists(fc); 7809 } 7810 set_lastFreeRangeCoalesced(true); 7811 print_free_block_coalesced(fc); 7812 } else { // not in a free range and/or should not coalesce 7813 // Return the current free range and start a new one. 7814 if (inFreeRange()) { 7815 // In a free range but cannot coalesce with the right hand chunk. 7816 // Put the current free range into the free lists. 7817 flush_cur_free_chunk(freeFinger(), 7818 pointer_delta(fc_addr, freeFinger())); 7819 } 7820 // Set up for new free range. Pass along whether the right hand 7821 // chunk is in the free lists. 7822 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7823 } 7824 } 7825 7826 // Lookahead flush: 7827 // If we are tracking a free range, and this is the last chunk that 7828 // we'll look at because its end crosses past _limit, we'll preemptively 7829 // flush it along with any free range we may be holding on to. Note that 7830 // this can be the case only for an already free or freshly garbage 7831 // chunk. If this block is an object, it can never straddle 7832 // over _limit. The "straddling" occurs when _limit is set at 7833 // the previous end of the space when this cycle started, and 7834 // a subsequent heap expansion caused the previously co-terminal 7835 // free block to be coalesced with the newly expanded portion, 7836 // thus rendering _limit a non-block-boundary making it dangerous 7837 // for the sweeper to step over and examine. 7838 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7839 assert(inFreeRange(), "Should only be called if currently in a free range."); 7840 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7841 assert(_sp->used_region().contains(eob - 1), 7842 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7843 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7844 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7845 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7846 if (eob >= _limit) { 7847 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7848 if (CMSTraceSweeper) { 7849 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7850 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7851 "[" PTR_FORMAT "," PTR_FORMAT ")", 7852 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7853 } 7854 // Return the storage we are tracking back into the free lists. 7855 if (CMSTraceSweeper) { 7856 gclog_or_tty->print_cr("Flushing ... "); 7857 } 7858 assert(freeFinger() < eob, "Error"); 7859 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7860 } 7861 } 7862 7863 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7864 assert(inFreeRange(), "Should only be called if currently in a free range."); 7865 assert(size > 0, 7866 "A zero sized chunk cannot be added to the free lists."); 7867 if (!freeRangeInFreeLists()) { 7868 if (CMSTestInFreeList) { 7869 FreeChunk* fc = (FreeChunk*) chunk; 7870 fc->set_size(size); 7871 assert(!_sp->verify_chunk_in_free_list(fc), 7872 "chunk should not be in free lists yet"); 7873 } 7874 if (CMSTraceSweeper) { 7875 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7876 p2i(chunk), size); 7877 } 7878 // A new free range is going to be starting. The current 7879 // free range has not been added to the free lists yet or 7880 // was removed so add it back. 7881 // If the current free range was coalesced, then the death 7882 // of the free range was recorded. Record a birth now. 7883 if (lastFreeRangeCoalesced()) { 7884 _sp->coalBirth(size); 7885 } 7886 _sp->addChunkAndRepairOffsetTable(chunk, size, 7887 lastFreeRangeCoalesced()); 7888 } else if (CMSTraceSweeper) { 7889 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7890 } 7891 set_inFreeRange(false); 7892 set_freeRangeInFreeLists(false); 7893 } 7894 7895 // We take a break if we've been at this for a while, 7896 // so as to avoid monopolizing the locks involved. 7897 void SweepClosure::do_yield_work(HeapWord* addr) { 7898 // Return current free chunk being used for coalescing (if any) 7899 // to the appropriate freelist. After yielding, the next 7900 // free block encountered will start a coalescing range of 7901 // free blocks. If the next free block is adjacent to the 7902 // chunk just flushed, they will need to wait for the next 7903 // sweep to be coalesced. 7904 if (inFreeRange()) { 7905 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7906 } 7907 7908 // First give up the locks, then yield, then re-lock. 7909 // We should probably use a constructor/destructor idiom to 7910 // do this unlock/lock or modify the MutexUnlocker class to 7911 // serve our purpose. XXX 7912 assert_lock_strong(_bitMap->lock()); 7913 assert_lock_strong(_freelistLock); 7914 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7915 "CMS thread should hold CMS token"); 7916 _bitMap->lock()->unlock(); 7917 _freelistLock->unlock(); 7918 ConcurrentMarkSweepThread::desynchronize(true); 7919 _collector->stopTimer(); 7920 if (PrintCMSStatistics != 0) { 7921 _collector->incrementYields(); 7922 } 7923 7924 // See the comment in coordinator_yield() 7925 for (unsigned i = 0; i < CMSYieldSleepCount && 7926 ConcurrentMarkSweepThread::should_yield() && 7927 !CMSCollector::foregroundGCIsActive(); ++i) { 7928 os::sleep(Thread::current(), 1, false); 7929 } 7930 7931 ConcurrentMarkSweepThread::synchronize(true); 7932 _freelistLock->lock(); 7933 _bitMap->lock()->lock_without_safepoint_check(); 7934 _collector->startTimer(); 7935 } 7936 7937 #ifndef PRODUCT 7938 // This is actually very useful in a product build if it can 7939 // be called from the debugger. Compile it into the product 7940 // as needed. 7941 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7942 return debug_cms_space->verify_chunk_in_free_list(fc); 7943 } 7944 #endif 7945 7946 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7947 if (CMSTraceSweeper) { 7948 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7949 p2i(fc), fc->size()); 7950 } 7951 } 7952 7953 // CMSIsAliveClosure 7954 bool CMSIsAliveClosure::do_object_b(oop obj) { 7955 HeapWord* addr = (HeapWord*)obj; 7956 return addr != NULL && 7957 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7958 } 7959 7960 7961 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7962 MemRegion span, 7963 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7964 bool cpc): 7965 _collector(collector), 7966 _span(span), 7967 _bit_map(bit_map), 7968 _mark_stack(mark_stack), 7969 _concurrent_precleaning(cpc) { 7970 assert(!_span.is_empty(), "Empty span could spell trouble"); 7971 } 7972 7973 7974 // CMSKeepAliveClosure: the serial version 7975 void CMSKeepAliveClosure::do_oop(oop obj) { 7976 HeapWord* addr = (HeapWord*)obj; 7977 if (_span.contains(addr) && 7978 !_bit_map->isMarked(addr)) { 7979 _bit_map->mark(addr); 7980 bool simulate_overflow = false; 7981 NOT_PRODUCT( 7982 if (CMSMarkStackOverflowALot && 7983 _collector->simulate_overflow()) { 7984 // simulate a stack overflow 7985 simulate_overflow = true; 7986 } 7987 ) 7988 if (simulate_overflow || !_mark_stack->push(obj)) { 7989 if (_concurrent_precleaning) { 7990 // We dirty the overflown object and let the remark 7991 // phase deal with it. 7992 assert(_collector->overflow_list_is_empty(), "Error"); 7993 // In the case of object arrays, we need to dirty all of 7994 // the cards that the object spans. No locking or atomics 7995 // are needed since no one else can be mutating the mod union 7996 // table. 7997 if (obj->is_objArray()) { 7998 size_t sz = obj->size(); 7999 HeapWord* end_card_addr = 8000 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8001 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8002 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8003 _collector->_modUnionTable.mark_range(redirty_range); 8004 } else { 8005 _collector->_modUnionTable.mark(addr); 8006 } 8007 _collector->_ser_kac_preclean_ovflw++; 8008 } else { 8009 _collector->push_on_overflow_list(obj); 8010 _collector->_ser_kac_ovflw++; 8011 } 8012 } 8013 } 8014 } 8015 8016 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8017 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8018 8019 // CMSParKeepAliveClosure: a parallel version of the above. 8020 // The work queues are private to each closure (thread), 8021 // but (may be) available for stealing by other threads. 8022 void CMSParKeepAliveClosure::do_oop(oop obj) { 8023 HeapWord* addr = (HeapWord*)obj; 8024 if (_span.contains(addr) && 8025 !_bit_map->isMarked(addr)) { 8026 // In general, during recursive tracing, several threads 8027 // may be concurrently getting here; the first one to 8028 // "tag" it, claims it. 8029 if (_bit_map->par_mark(addr)) { 8030 bool res = _work_queue->push(obj); 8031 assert(res, "Low water mark should be much less than capacity"); 8032 // Do a recursive trim in the hope that this will keep 8033 // stack usage lower, but leave some oops for potential stealers 8034 trim_queue(_low_water_mark); 8035 } // Else, another thread got there first 8036 } 8037 } 8038 8039 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8040 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8041 8042 void CMSParKeepAliveClosure::trim_queue(uint max) { 8043 while (_work_queue->size() > max) { 8044 oop new_oop; 8045 if (_work_queue->pop_local(new_oop)) { 8046 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8047 assert(_bit_map->isMarked((HeapWord*)new_oop), 8048 "no white objects on this stack!"); 8049 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8050 // iterate over the oops in this oop, marking and pushing 8051 // the ones in CMS heap (i.e. in _span). 8052 new_oop->oop_iterate(&_mark_and_push); 8053 } 8054 } 8055 } 8056 8057 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8058 CMSCollector* collector, 8059 MemRegion span, CMSBitMap* bit_map, 8060 OopTaskQueue* work_queue): 8061 _collector(collector), 8062 _span(span), 8063 _bit_map(bit_map), 8064 _work_queue(work_queue) { } 8065 8066 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8067 HeapWord* addr = (HeapWord*)obj; 8068 if (_span.contains(addr) && 8069 !_bit_map->isMarked(addr)) { 8070 if (_bit_map->par_mark(addr)) { 8071 bool simulate_overflow = false; 8072 NOT_PRODUCT( 8073 if (CMSMarkStackOverflowALot && 8074 _collector->par_simulate_overflow()) { 8075 // simulate a stack overflow 8076 simulate_overflow = true; 8077 } 8078 ) 8079 if (simulate_overflow || !_work_queue->push(obj)) { 8080 _collector->par_push_on_overflow_list(obj); 8081 _collector->_par_kac_ovflw++; 8082 } 8083 } // Else another thread got there already 8084 } 8085 } 8086 8087 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8088 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8089 8090 ////////////////////////////////////////////////////////////////// 8091 // CMSExpansionCause ///////////////////////////// 8092 ////////////////////////////////////////////////////////////////// 8093 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8094 switch (cause) { 8095 case _no_expansion: 8096 return "No expansion"; 8097 case _satisfy_free_ratio: 8098 return "Free ratio"; 8099 case _satisfy_promotion: 8100 return "Satisfy promotion"; 8101 case _satisfy_allocation: 8102 return "allocation"; 8103 case _allocate_par_lab: 8104 return "Par LAB"; 8105 case _allocate_par_spooling_space: 8106 return "Par Spooling Space"; 8107 case _adaptive_size_policy: 8108 return "Ergonomics"; 8109 default: 8110 return "unknown"; 8111 } 8112 } 8113 8114 void CMSDrainMarkingStackClosure::do_void() { 8115 // the max number to take from overflow list at a time 8116 const size_t num = _mark_stack->capacity()/4; 8117 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8118 "Overflow list should be NULL during concurrent phases"); 8119 while (!_mark_stack->isEmpty() || 8120 // if stack is empty, check the overflow list 8121 _collector->take_from_overflow_list(num, _mark_stack)) { 8122 oop obj = _mark_stack->pop(); 8123 HeapWord* addr = (HeapWord*)obj; 8124 assert(_span.contains(addr), "Should be within span"); 8125 assert(_bit_map->isMarked(addr), "Should be marked"); 8126 assert(obj->is_oop(), "Should be an oop"); 8127 obj->oop_iterate(_keep_alive); 8128 } 8129 } 8130 8131 void CMSParDrainMarkingStackClosure::do_void() { 8132 // drain queue 8133 trim_queue(0); 8134 } 8135 8136 // Trim our work_queue so its length is below max at return 8137 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8138 while (_work_queue->size() > max) { 8139 oop new_oop; 8140 if (_work_queue->pop_local(new_oop)) { 8141 assert(new_oop->is_oop(), "Expected an oop"); 8142 assert(_bit_map->isMarked((HeapWord*)new_oop), 8143 "no white objects on this stack!"); 8144 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8145 // iterate over the oops in this oop, marking and pushing 8146 // the ones in CMS heap (i.e. in _span). 8147 new_oop->oop_iterate(&_mark_and_push); 8148 } 8149 } 8150 } 8151 8152 //////////////////////////////////////////////////////////////////// 8153 // Support for Marking Stack Overflow list handling and related code 8154 //////////////////////////////////////////////////////////////////// 8155 // Much of the following code is similar in shape and spirit to the 8156 // code used in ParNewGC. We should try and share that code 8157 // as much as possible in the future. 8158 8159 #ifndef PRODUCT 8160 // Debugging support for CMSStackOverflowALot 8161 8162 // It's OK to call this multi-threaded; the worst thing 8163 // that can happen is that we'll get a bunch of closely 8164 // spaced simulated overflows, but that's OK, in fact 8165 // probably good as it would exercise the overflow code 8166 // under contention. 8167 bool CMSCollector::simulate_overflow() { 8168 if (_overflow_counter-- <= 0) { // just being defensive 8169 _overflow_counter = CMSMarkStackOverflowInterval; 8170 return true; 8171 } else { 8172 return false; 8173 } 8174 } 8175 8176 bool CMSCollector::par_simulate_overflow() { 8177 return simulate_overflow(); 8178 } 8179 #endif 8180 8181 // Single-threaded 8182 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8183 assert(stack->isEmpty(), "Expected precondition"); 8184 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8185 size_t i = num; 8186 oop cur = _overflow_list; 8187 const markOop proto = markOopDesc::prototype(); 8188 NOT_PRODUCT(ssize_t n = 0;) 8189 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8190 next = oop(cur->mark()); 8191 cur->set_mark(proto); // until proven otherwise 8192 assert(cur->is_oop(), "Should be an oop"); 8193 bool res = stack->push(cur); 8194 assert(res, "Bit off more than can chew?"); 8195 NOT_PRODUCT(n++;) 8196 } 8197 _overflow_list = cur; 8198 #ifndef PRODUCT 8199 assert(_num_par_pushes >= n, "Too many pops?"); 8200 _num_par_pushes -=n; 8201 #endif 8202 return !stack->isEmpty(); 8203 } 8204 8205 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8206 // (MT-safe) Get a prefix of at most "num" from the list. 8207 // The overflow list is chained through the mark word of 8208 // each object in the list. We fetch the entire list, 8209 // break off a prefix of the right size and return the 8210 // remainder. If other threads try to take objects from 8211 // the overflow list at that time, they will wait for 8212 // some time to see if data becomes available. If (and 8213 // only if) another thread places one or more object(s) 8214 // on the global list before we have returned the suffix 8215 // to the global list, we will walk down our local list 8216 // to find its end and append the global list to 8217 // our suffix before returning it. This suffix walk can 8218 // prove to be expensive (quadratic in the amount of traffic) 8219 // when there are many objects in the overflow list and 8220 // there is much producer-consumer contention on the list. 8221 // *NOTE*: The overflow list manipulation code here and 8222 // in ParNewGeneration:: are very similar in shape, 8223 // except that in the ParNew case we use the old (from/eden) 8224 // copy of the object to thread the list via its klass word. 8225 // Because of the common code, if you make any changes in 8226 // the code below, please check the ParNew version to see if 8227 // similar changes might be needed. 8228 // CR 6797058 has been filed to consolidate the common code. 8229 bool CMSCollector::par_take_from_overflow_list(size_t num, 8230 OopTaskQueue* work_q, 8231 int no_of_gc_threads) { 8232 assert(work_q->size() == 0, "First empty local work queue"); 8233 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8234 if (_overflow_list == NULL) { 8235 return false; 8236 } 8237 // Grab the entire list; we'll put back a suffix 8238 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8239 Thread* tid = Thread::current(); 8240 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8241 // set to ParallelGCThreads. 8242 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8243 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8244 // If the list is busy, we spin for a short while, 8245 // sleeping between attempts to get the list. 8246 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8247 os::sleep(tid, sleep_time_millis, false); 8248 if (_overflow_list == NULL) { 8249 // Nothing left to take 8250 return false; 8251 } else if (_overflow_list != BUSY) { 8252 // Try and grab the prefix 8253 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8254 } 8255 } 8256 // If the list was found to be empty, or we spun long 8257 // enough, we give up and return empty-handed. If we leave 8258 // the list in the BUSY state below, it must be the case that 8259 // some other thread holds the overflow list and will set it 8260 // to a non-BUSY state in the future. 8261 if (prefix == NULL || prefix == BUSY) { 8262 // Nothing to take or waited long enough 8263 if (prefix == NULL) { 8264 // Write back the NULL in case we overwrote it with BUSY above 8265 // and it is still the same value. 8266 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8267 } 8268 return false; 8269 } 8270 assert(prefix != NULL && prefix != BUSY, "Error"); 8271 size_t i = num; 8272 oop cur = prefix; 8273 // Walk down the first "num" objects, unless we reach the end. 8274 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8275 if (cur->mark() == NULL) { 8276 // We have "num" or fewer elements in the list, so there 8277 // is nothing to return to the global list. 8278 // Write back the NULL in lieu of the BUSY we wrote 8279 // above, if it is still the same value. 8280 if (_overflow_list == BUSY) { 8281 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8282 } 8283 } else { 8284 // Chop off the suffix and return it to the global list. 8285 assert(cur->mark() != BUSY, "Error"); 8286 oop suffix_head = cur->mark(); // suffix will be put back on global list 8287 cur->set_mark(NULL); // break off suffix 8288 // It's possible that the list is still in the empty(busy) state 8289 // we left it in a short while ago; in that case we may be 8290 // able to place back the suffix without incurring the cost 8291 // of a walk down the list. 8292 oop observed_overflow_list = _overflow_list; 8293 oop cur_overflow_list = observed_overflow_list; 8294 bool attached = false; 8295 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8296 observed_overflow_list = 8297 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8298 if (cur_overflow_list == observed_overflow_list) { 8299 attached = true; 8300 break; 8301 } else cur_overflow_list = observed_overflow_list; 8302 } 8303 if (!attached) { 8304 // Too bad, someone else sneaked in (at least) an element; we'll need 8305 // to do a splice. Find tail of suffix so we can prepend suffix to global 8306 // list. 8307 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8308 oop suffix_tail = cur; 8309 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8310 "Tautology"); 8311 observed_overflow_list = _overflow_list; 8312 do { 8313 cur_overflow_list = observed_overflow_list; 8314 if (cur_overflow_list != BUSY) { 8315 // Do the splice ... 8316 suffix_tail->set_mark(markOop(cur_overflow_list)); 8317 } else { // cur_overflow_list == BUSY 8318 suffix_tail->set_mark(NULL); 8319 } 8320 // ... and try to place spliced list back on overflow_list ... 8321 observed_overflow_list = 8322 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8323 } while (cur_overflow_list != observed_overflow_list); 8324 // ... until we have succeeded in doing so. 8325 } 8326 } 8327 8328 // Push the prefix elements on work_q 8329 assert(prefix != NULL, "control point invariant"); 8330 const markOop proto = markOopDesc::prototype(); 8331 oop next; 8332 NOT_PRODUCT(ssize_t n = 0;) 8333 for (cur = prefix; cur != NULL; cur = next) { 8334 next = oop(cur->mark()); 8335 cur->set_mark(proto); // until proven otherwise 8336 assert(cur->is_oop(), "Should be an oop"); 8337 bool res = work_q->push(cur); 8338 assert(res, "Bit off more than we can chew?"); 8339 NOT_PRODUCT(n++;) 8340 } 8341 #ifndef PRODUCT 8342 assert(_num_par_pushes >= n, "Too many pops?"); 8343 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8344 #endif 8345 return true; 8346 } 8347 8348 // Single-threaded 8349 void CMSCollector::push_on_overflow_list(oop p) { 8350 NOT_PRODUCT(_num_par_pushes++;) 8351 assert(p->is_oop(), "Not an oop"); 8352 preserve_mark_if_necessary(p); 8353 p->set_mark((markOop)_overflow_list); 8354 _overflow_list = p; 8355 } 8356 8357 // Multi-threaded; use CAS to prepend to overflow list 8358 void CMSCollector::par_push_on_overflow_list(oop p) { 8359 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8360 assert(p->is_oop(), "Not an oop"); 8361 par_preserve_mark_if_necessary(p); 8362 oop observed_overflow_list = _overflow_list; 8363 oop cur_overflow_list; 8364 do { 8365 cur_overflow_list = observed_overflow_list; 8366 if (cur_overflow_list != BUSY) { 8367 p->set_mark(markOop(cur_overflow_list)); 8368 } else { 8369 p->set_mark(NULL); 8370 } 8371 observed_overflow_list = 8372 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8373 } while (cur_overflow_list != observed_overflow_list); 8374 } 8375 #undef BUSY 8376 8377 // Single threaded 8378 // General Note on GrowableArray: pushes may silently fail 8379 // because we are (temporarily) out of C-heap for expanding 8380 // the stack. The problem is quite ubiquitous and affects 8381 // a lot of code in the JVM. The prudent thing for GrowableArray 8382 // to do (for now) is to exit with an error. However, that may 8383 // be too draconian in some cases because the caller may be 8384 // able to recover without much harm. For such cases, we 8385 // should probably introduce a "soft_push" method which returns 8386 // an indication of success or failure with the assumption that 8387 // the caller may be able to recover from a failure; code in 8388 // the VM can then be changed, incrementally, to deal with such 8389 // failures where possible, thus, incrementally hardening the VM 8390 // in such low resource situations. 8391 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8392 _preserved_oop_stack.push(p); 8393 _preserved_mark_stack.push(m); 8394 assert(m == p->mark(), "Mark word changed"); 8395 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8396 "bijection"); 8397 } 8398 8399 // Single threaded 8400 void CMSCollector::preserve_mark_if_necessary(oop p) { 8401 markOop m = p->mark(); 8402 if (m->must_be_preserved(p)) { 8403 preserve_mark_work(p, m); 8404 } 8405 } 8406 8407 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8408 markOop m = p->mark(); 8409 if (m->must_be_preserved(p)) { 8410 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8411 // Even though we read the mark word without holding 8412 // the lock, we are assured that it will not change 8413 // because we "own" this oop, so no other thread can 8414 // be trying to push it on the overflow list; see 8415 // the assertion in preserve_mark_work() that checks 8416 // that m == p->mark(). 8417 preserve_mark_work(p, m); 8418 } 8419 } 8420 8421 // We should be able to do this multi-threaded, 8422 // a chunk of stack being a task (this is 8423 // correct because each oop only ever appears 8424 // once in the overflow list. However, it's 8425 // not very easy to completely overlap this with 8426 // other operations, so will generally not be done 8427 // until all work's been completed. Because we 8428 // expect the preserved oop stack (set) to be small, 8429 // it's probably fine to do this single-threaded. 8430 // We can explore cleverer concurrent/overlapped/parallel 8431 // processing of preserved marks if we feel the 8432 // need for this in the future. Stack overflow should 8433 // be so rare in practice and, when it happens, its 8434 // effect on performance so great that this will 8435 // likely just be in the noise anyway. 8436 void CMSCollector::restore_preserved_marks_if_any() { 8437 assert(SafepointSynchronize::is_at_safepoint(), 8438 "world should be stopped"); 8439 assert(Thread::current()->is_ConcurrentGC_thread() || 8440 Thread::current()->is_VM_thread(), 8441 "should be single-threaded"); 8442 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8443 "bijection"); 8444 8445 while (!_preserved_oop_stack.is_empty()) { 8446 oop p = _preserved_oop_stack.pop(); 8447 assert(p->is_oop(), "Should be an oop"); 8448 assert(_span.contains(p), "oop should be in _span"); 8449 assert(p->mark() == markOopDesc::prototype(), 8450 "Set when taken from overflow list"); 8451 markOop m = _preserved_mark_stack.pop(); 8452 p->set_mark(m); 8453 } 8454 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8455 "stacks were cleared above"); 8456 } 8457 8458 #ifndef PRODUCT 8459 bool CMSCollector::no_preserved_marks() const { 8460 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8461 } 8462 #endif 8463 8464 // Transfer some number of overflown objects to usual marking 8465 // stack. Return true if some objects were transferred. 8466 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8467 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8468 (size_t)ParGCDesiredObjsFromOverflowList); 8469 8470 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8471 assert(_collector->overflow_list_is_empty() || res, 8472 "If list is not empty, we should have taken something"); 8473 assert(!res || !_mark_stack->isEmpty(), 8474 "If we took something, it should now be on our stack"); 8475 return res; 8476 } 8477 8478 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8479 size_t res = _sp->block_size_no_stall(addr, _collector); 8480 if (_sp->block_is_obj(addr)) { 8481 if (_live_bit_map->isMarked(addr)) { 8482 // It can't have been dead in a previous cycle 8483 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8484 } else { 8485 _dead_bit_map->mark(addr); // mark the dead object 8486 } 8487 } 8488 // Could be 0, if the block size could not be computed without stalling. 8489 return res; 8490 } 8491 8492 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8493 8494 switch (phase) { 8495 case CMSCollector::InitialMarking: 8496 initialize(true /* fullGC */ , 8497 cause /* cause of the GC */, 8498 true /* recordGCBeginTime */, 8499 true /* recordPreGCUsage */, 8500 false /* recordPeakUsage */, 8501 false /* recordPostGCusage */, 8502 true /* recordAccumulatedGCTime */, 8503 false /* recordGCEndTime */, 8504 false /* countCollection */ ); 8505 break; 8506 8507 case CMSCollector::FinalMarking: 8508 initialize(true /* fullGC */ , 8509 cause /* cause of the GC */, 8510 false /* recordGCBeginTime */, 8511 false /* recordPreGCUsage */, 8512 false /* recordPeakUsage */, 8513 false /* recordPostGCusage */, 8514 true /* recordAccumulatedGCTime */, 8515 false /* recordGCEndTime */, 8516 false /* countCollection */ ); 8517 break; 8518 8519 case CMSCollector::Sweeping: 8520 initialize(true /* fullGC */ , 8521 cause /* cause of the GC */, 8522 false /* recordGCBeginTime */, 8523 false /* recordPreGCUsage */, 8524 true /* recordPeakUsage */, 8525 true /* recordPostGCusage */, 8526 false /* recordAccumulatedGCTime */, 8527 true /* recordGCEndTime */, 8528 true /* countCollection */ ); 8529 break; 8530 8531 default: 8532 ShouldNotReachHere(); 8533 } 8534 }