1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/cms/cmsCollectorPolicy.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/compactibleFreeListSpace.hpp" 33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 34 #include "gc/cms/concurrentMarkSweepThread.hpp" 35 #include "gc/cms/parNewGeneration.hpp" 36 #include "gc/cms/vmCMSOperations.hpp" 37 #include "gc/serial/genMarkSweep.hpp" 38 #include "gc/serial/tenuredGeneration.hpp" 39 #include "gc/shared/adaptiveSizePolicy.hpp" 40 #include "gc/shared/cardGeneration.inline.hpp" 41 #include "gc/shared/cardTableRS.hpp" 42 #include "gc/shared/collectedHeap.inline.hpp" 43 #include "gc/shared/collectorCounters.hpp" 44 #include "gc/shared/collectorPolicy.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcTimer.hpp" 47 #include "gc/shared/gcTrace.hpp" 48 #include "gc/shared/gcTraceTime.hpp" 49 #include "gc/shared/genCollectedHeap.hpp" 50 #include "gc/shared/genOopClosures.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/referencePolicy.hpp" 53 #include "gc/shared/strongRootsScope.hpp" 54 #include "gc/shared/taskqueue.inline.hpp" 55 #include "memory/allocation.hpp" 56 #include "memory/iterator.inline.hpp" 57 #include "memory/padded.hpp" 58 #include "memory/resourceArea.hpp" 59 #include "oops/oop.inline.hpp" 60 #include "prims/jvmtiExport.hpp" 61 #include "runtime/atomic.inline.hpp" 62 #include "runtime/globals_extension.hpp" 63 #include "runtime/handles.inline.hpp" 64 #include "runtime/java.hpp" 65 #include "runtime/orderAccess.inline.hpp" 66 #include "runtime/vmThread.hpp" 67 #include "services/memoryService.hpp" 68 #include "services/runtimeService.hpp" 69 #include "utilities/stack.inline.hpp" 70 71 // statics 72 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 73 bool CMSCollector::_full_gc_requested = false; 74 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 75 76 ////////////////////////////////////////////////////////////////// 77 // In support of CMS/VM thread synchronization 78 ////////////////////////////////////////////////////////////////// 79 // We split use of the CGC_lock into 2 "levels". 80 // The low-level locking is of the usual CGC_lock monitor. We introduce 81 // a higher level "token" (hereafter "CMS token") built on top of the 82 // low level monitor (hereafter "CGC lock"). 83 // The token-passing protocol gives priority to the VM thread. The 84 // CMS-lock doesn't provide any fairness guarantees, but clients 85 // should ensure that it is only held for very short, bounded 86 // durations. 87 // 88 // When either of the CMS thread or the VM thread is involved in 89 // collection operations during which it does not want the other 90 // thread to interfere, it obtains the CMS token. 91 // 92 // If either thread tries to get the token while the other has 93 // it, that thread waits. However, if the VM thread and CMS thread 94 // both want the token, then the VM thread gets priority while the 95 // CMS thread waits. This ensures, for instance, that the "concurrent" 96 // phases of the CMS thread's work do not block out the VM thread 97 // for long periods of time as the CMS thread continues to hog 98 // the token. (See bug 4616232). 99 // 100 // The baton-passing functions are, however, controlled by the 101 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 102 // and here the low-level CMS lock, not the high level token, 103 // ensures mutual exclusion. 104 // 105 // Two important conditions that we have to satisfy: 106 // 1. if a thread does a low-level wait on the CMS lock, then it 107 // relinquishes the CMS token if it were holding that token 108 // when it acquired the low-level CMS lock. 109 // 2. any low-level notifications on the low-level lock 110 // should only be sent when a thread has relinquished the token. 111 // 112 // In the absence of either property, we'd have potential deadlock. 113 // 114 // We protect each of the CMS (concurrent and sequential) phases 115 // with the CMS _token_, not the CMS _lock_. 116 // 117 // The only code protected by CMS lock is the token acquisition code 118 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 119 // baton-passing code. 120 // 121 // Unfortunately, i couldn't come up with a good abstraction to factor and 122 // hide the naked CGC_lock manipulation in the baton-passing code 123 // further below. That's something we should try to do. Also, the proof 124 // of correctness of this 2-level locking scheme is far from obvious, 125 // and potentially quite slippery. We have an uneasy suspicion, for instance, 126 // that there may be a theoretical possibility of delay/starvation in the 127 // low-level lock/wait/notify scheme used for the baton-passing because of 128 // potential interference with the priority scheme embodied in the 129 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 130 // invocation further below and marked with "XXX 20011219YSR". 131 // Indeed, as we note elsewhere, this may become yet more slippery 132 // in the presence of multiple CMS and/or multiple VM threads. XXX 133 134 class CMSTokenSync: public StackObj { 135 private: 136 bool _is_cms_thread; 137 public: 138 CMSTokenSync(bool is_cms_thread): 139 _is_cms_thread(is_cms_thread) { 140 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 141 "Incorrect argument to constructor"); 142 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 143 } 144 145 ~CMSTokenSync() { 146 assert(_is_cms_thread ? 147 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 148 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 149 "Incorrect state"); 150 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 151 } 152 }; 153 154 // Convenience class that does a CMSTokenSync, and then acquires 155 // upto three locks. 156 class CMSTokenSyncWithLocks: public CMSTokenSync { 157 private: 158 // Note: locks are acquired in textual declaration order 159 // and released in the opposite order 160 MutexLockerEx _locker1, _locker2, _locker3; 161 public: 162 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 163 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 164 CMSTokenSync(is_cms_thread), 165 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 166 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 167 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 168 { } 169 }; 170 171 172 ////////////////////////////////////////////////////////////////// 173 // Concurrent Mark-Sweep Generation ///////////////////////////// 174 ////////////////////////////////////////////////////////////////// 175 176 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 177 178 // This struct contains per-thread things necessary to support parallel 179 // young-gen collection. 180 class CMSParGCThreadState: public CHeapObj<mtGC> { 181 public: 182 CFLS_LAB lab; 183 PromotionInfo promo; 184 185 // Constructor. 186 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 187 promo.setSpace(cfls); 188 } 189 }; 190 191 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 192 ReservedSpace rs, size_t initial_byte_size, int level, 193 CardTableRS* ct, bool use_adaptive_freelists, 194 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 195 CardGeneration(rs, initial_byte_size, level, ct), 196 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 (int) ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 307 308 // Generation Counters - generation 1, 1 subspace 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, 310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 311 312 _space_counters = new GSpaceCounters(gen_name, 0, 313 _virtual_space.reserved_size(), 314 this, _gen_counters); 315 } 316 317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 318 _cms_gen(cms_gen) 319 { 320 assert(alpha <= 100, "bad value"); 321 _saved_alpha = alpha; 322 323 // Initialize the alphas to the bootstrap value of 100. 324 _gc0_alpha = _cms_alpha = 100; 325 326 _cms_begin_time.update(); 327 _cms_end_time.update(); 328 329 _gc0_duration = 0.0; 330 _gc0_period = 0.0; 331 _gc0_promoted = 0; 332 333 _cms_duration = 0.0; 334 _cms_period = 0.0; 335 _cms_allocated = 0; 336 337 _cms_used_at_gc0_begin = 0; 338 _cms_used_at_gc0_end = 0; 339 _allow_duty_cycle_reduction = false; 340 _valid_bits = 0; 341 } 342 343 double CMSStats::cms_free_adjustment_factor(size_t free) const { 344 // TBD: CR 6909490 345 return 1.0; 346 } 347 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 349 } 350 351 // If promotion failure handling is on use 352 // the padded average size of the promotion for each 353 // young generation collection. 354 double CMSStats::time_until_cms_gen_full() const { 355 size_t cms_free = _cms_gen->cmsSpace()->free(); 356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 359 if (cms_free > expected_promotion) { 360 // Start a cms collection if there isn't enough space to promote 361 // for the next minor collection. Use the padded average as 362 // a safety factor. 363 cms_free -= expected_promotion; 364 365 // Adjust by the safety factor. 366 double cms_free_dbl = (double)cms_free; 367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 368 // Apply a further correction factor which tries to adjust 369 // for recent occurance of concurrent mode failures. 370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 371 cms_free_dbl = cms_free_dbl * cms_adjustment; 372 373 if (PrintGCDetails && Verbose) { 374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 376 cms_free, expected_promotion); 377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 378 cms_free_dbl, cms_consumption_rate() + 1.0); 379 } 380 // Add 1 in case the consumption rate goes to zero. 381 return cms_free_dbl / (cms_consumption_rate() + 1.0); 382 } 383 return 0.0; 384 } 385 386 // Compare the duration of the cms collection to the 387 // time remaining before the cms generation is empty. 388 // Note that the time from the start of the cms collection 389 // to the start of the cms sweep (less than the total 390 // duration of the cms collection) can be used. This 391 // has been tried and some applications experienced 392 // promotion failures early in execution. This was 393 // possibly because the averages were not accurate 394 // enough at the beginning. 395 double CMSStats::time_until_cms_start() const { 396 // We add "gc0_period" to the "work" calculation 397 // below because this query is done (mostly) at the 398 // end of a scavenge, so we need to conservatively 399 // account for that much possible delay 400 // in the query so as to avoid concurrent mode failures 401 // due to starting the collection just a wee bit too 402 // late. 403 double work = cms_duration() + gc0_period(); 404 double deadline = time_until_cms_gen_full(); 405 // If a concurrent mode failure occurred recently, we want to be 406 // more conservative and halve our expected time_until_cms_gen_full() 407 if (work > deadline) { 408 if (Verbose && PrintGCDetails) { 409 gclog_or_tty->print( 410 " CMSCollector: collect because of anticipated promotion " 411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 412 gc0_period(), time_until_cms_gen_full()); 413 } 414 return 0.0; 415 } 416 return work - deadline; 417 } 418 419 #ifndef PRODUCT 420 void CMSStats::print_on(outputStream *st) const { 421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 423 gc0_duration(), gc0_period(), gc0_promoted()); 424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 425 cms_duration(), cms_period(), cms_allocated()); 426 st->print(",cms_since_beg=%g,cms_since_end=%g", 427 cms_time_since_begin(), cms_time_since_end()); 428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 430 431 if (valid()) { 432 st->print(",promo_rate=%g,cms_alloc_rate=%g", 433 promotion_rate(), cms_allocation_rate()); 434 st->print(",cms_consumption_rate=%g,time_until_full=%g", 435 cms_consumption_rate(), time_until_cms_gen_full()); 436 } 437 st->print(" "); 438 } 439 #endif // #ifndef PRODUCT 440 441 CMSCollector::CollectorState CMSCollector::_collectorState = 442 CMSCollector::Idling; 443 bool CMSCollector::_foregroundGCIsActive = false; 444 bool CMSCollector::_foregroundGCShouldWait = false; 445 446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 447 CardTableRS* ct, 448 ConcurrentMarkSweepPolicy* cp): 449 _cmsGen(cmsGen), 450 _ct(ct), 451 _ref_processor(NULL), // will be set later 452 _conc_workers(NULL), // may be set later 453 _abort_preclean(false), 454 _start_sampling(false), 455 _between_prologue_and_epilogue(false), 456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 458 -1 /* lock-free */, "No_lock" /* dummy */), 459 _modUnionClosurePar(&_modUnionTable), 460 // Adjust my span to cover old (cms) gen 461 _span(cmsGen->reserved()), 462 // Construct the is_alive_closure with _span & markBitMap 463 _is_alive_closure(_span, &_markBitMap), 464 _restart_addr(NULL), 465 _overflow_list(NULL), 466 _stats(cmsGen), 467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 468 //verify that this lock should be acquired with safepoint check. 469 Monitor::_safepoint_check_sometimes)), 470 _eden_chunk_array(NULL), // may be set in ctor body 471 _eden_chunk_capacity(0), // -- ditto -- 472 _eden_chunk_index(0), // -- ditto -- 473 _survivor_plab_array(NULL), // -- ditto -- 474 _survivor_chunk_array(NULL), // -- ditto -- 475 _survivor_chunk_capacity(0), // -- ditto -- 476 _survivor_chunk_index(0), // -- ditto -- 477 _ser_pmc_preclean_ovflw(0), 478 _ser_kac_preclean_ovflw(0), 479 _ser_pmc_remark_ovflw(0), 480 _par_pmc_remark_ovflw(0), 481 _ser_kac_ovflw(0), 482 _par_kac_ovflw(0), 483 #ifndef PRODUCT 484 _num_par_pushes(0), 485 #endif 486 _collection_count_start(0), 487 _verifying(false), 488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 489 _completed_initialization(false), 490 _collector_policy(cp), 491 _should_unload_classes(CMSClassUnloadingEnabled), 492 _concurrent_cycles_since_last_unload(0), 493 _roots_scanning_options(GenCollectedHeap::SO_None), 494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 498 _cms_start_registered(false) 499 { 500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 501 ExplicitGCInvokesConcurrent = true; 502 } 503 // Now expand the span and allocate the collection support structures 504 // (MUT, marking bit map etc.) to cover both generations subject to 505 // collection. 506 507 // For use by dirty card to oop closures. 508 _cmsGen->cmsSpace()->set_collector(this); 509 510 // Allocate MUT and marking bit map 511 { 512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 513 if (!_markBitMap.allocate(_span)) { 514 warning("Failed to allocate CMS Bit Map"); 515 return; 516 } 517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 518 } 519 { 520 _modUnionTable.allocate(_span); 521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 522 } 523 524 if (!_markStack.allocate(MarkStackSize)) { 525 warning("Failed to allocate CMS Marking Stack"); 526 return; 527 } 528 529 // Support for multi-threaded concurrent phases 530 if (CMSConcurrentMTEnabled) { 531 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 532 // just for now 533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 534 } 535 if (ConcGCThreads > 1) { 536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 537 ConcGCThreads, true); 538 if (_conc_workers == NULL) { 539 warning("GC/CMS: _conc_workers allocation failure: " 540 "forcing -CMSConcurrentMTEnabled"); 541 CMSConcurrentMTEnabled = false; 542 } else { 543 _conc_workers->initialize_workers(); 544 } 545 } else { 546 CMSConcurrentMTEnabled = false; 547 } 548 } 549 if (!CMSConcurrentMTEnabled) { 550 ConcGCThreads = 0; 551 } else { 552 // Turn off CMSCleanOnEnter optimization temporarily for 553 // the MT case where it's not fixed yet; see 6178663. 554 CMSCleanOnEnter = false; 555 } 556 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 557 "Inconsistency"); 558 559 // Parallel task queues; these are shared for the 560 // concurrent and stop-world phases of CMS, but 561 // are not shared with parallel scavenge (ParNew). 562 { 563 uint i; 564 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 565 566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 567 || ParallelRefProcEnabled) 568 && num_queues > 0) { 569 _task_queues = new OopTaskQueueSet(num_queues); 570 if (_task_queues == NULL) { 571 warning("task_queues allocation failure."); 572 return; 573 } 574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 576 for (i = 0; i < num_queues; i++) { 577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 578 if (q == NULL) { 579 warning("work_queue allocation failure."); 580 return; 581 } 582 _task_queues->register_queue(i, q); 583 } 584 for (i = 0; i < num_queues; i++) { 585 _task_queues->queue(i)->initialize(); 586 _hash_seed[i] = 17; // copied from ParNew 587 } 588 } 589 } 590 591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 592 593 // Clip CMSBootstrapOccupancy between 0 and 100. 594 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 595 596 // Now tell CMS generations the identity of their collector 597 ConcurrentMarkSweepGeneration::set_collector(this); 598 599 // Create & start a CMS thread for this CMS collector 600 _cmsThread = ConcurrentMarkSweepThread::start(this); 601 assert(cmsThread() != NULL, "CMS Thread should have been created"); 602 assert(cmsThread()->collector() == this, 603 "CMS Thread should refer to this gen"); 604 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 605 606 // Support for parallelizing young gen rescan 607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 609 _young_gen = (ParNewGeneration*)gch->young_gen(); 610 if (gch->supports_inline_contig_alloc()) { 611 _top_addr = gch->top_addr(); 612 _end_addr = gch->end_addr(); 613 assert(_young_gen != NULL, "no _young_gen"); 614 _eden_chunk_index = 0; 615 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 617 } 618 619 // Support for parallelizing survivor space rescan 620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 621 const size_t max_plab_samples = 622 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 623 624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 627 _survivor_chunk_capacity = 2*max_plab_samples; 628 for (uint i = 0; i < ParallelGCThreads; i++) { 629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 631 assert(cur->end() == 0, "Should be 0"); 632 assert(cur->array() == vec, "Should be vec"); 633 assert(cur->capacity() == max_plab_samples, "Error"); 634 } 635 } 636 637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 638 _gc_counters = new CollectorCounters("CMS", 1); 639 _completed_initialization = true; 640 _inter_sweep_timer.start(); // start of time 641 } 642 643 size_t CMSCollector::plab_sample_minimum_size() { 644 // The default value of MinTLABSize is 2k, but there is 645 // no way to get the default value if the flag has been overridden. 646 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 647 } 648 649 const char* ConcurrentMarkSweepGeneration::name() const { 650 return "concurrent mark-sweep generation"; 651 } 652 void ConcurrentMarkSweepGeneration::update_counters() { 653 if (UsePerfData) { 654 _space_counters->update_all(); 655 _gen_counters->update_all(); 656 } 657 } 658 659 // this is an optimized version of update_counters(). it takes the 660 // used value as a parameter rather than computing it. 661 // 662 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 663 if (UsePerfData) { 664 _space_counters->update_used(used); 665 _space_counters->update_capacity(); 666 _gen_counters->update_all(); 667 } 668 } 669 670 void ConcurrentMarkSweepGeneration::print() const { 671 Generation::print(); 672 cmsSpace()->print(); 673 } 674 675 #ifndef PRODUCT 676 void ConcurrentMarkSweepGeneration::print_statistics() { 677 cmsSpace()->printFLCensus(0); 678 } 679 #endif 680 681 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 682 GenCollectedHeap* gch = GenCollectedHeap::heap(); 683 if (PrintGCDetails) { 684 if (Verbose) { 685 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 686 level(), short_name(), s, used(), capacity()); 687 } else { 688 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 689 level(), short_name(), s, used() / K, capacity() / K); 690 } 691 } 692 if (Verbose) { 693 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 694 gch->used(), gch->capacity()); 695 } else { 696 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 697 gch->used() / K, gch->capacity() / K); 698 } 699 } 700 701 size_t 702 ConcurrentMarkSweepGeneration::contiguous_available() const { 703 // dld proposes an improvement in precision here. If the committed 704 // part of the space ends in a free block we should add that to 705 // uncommitted size in the calculation below. Will make this 706 // change later, staying with the approximation below for the 707 // time being. -- ysr. 708 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 709 } 710 711 size_t 712 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 713 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 714 } 715 716 size_t ConcurrentMarkSweepGeneration::max_available() const { 717 return free() + _virtual_space.uncommitted_size(); 718 } 719 720 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 721 size_t available = max_available(); 722 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 723 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 724 if (Verbose && PrintGCDetails) { 725 gclog_or_tty->print_cr( 726 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 727 "max_promo("SIZE_FORMAT")", 728 res? "":" not", available, res? ">=":"<", 729 av_promo, max_promotion_in_bytes); 730 } 731 return res; 732 } 733 734 // At a promotion failure dump information on block layout in heap 735 // (cms old generation). 736 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 737 if (CMSDumpAtPromotionFailure) { 738 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 739 } 740 } 741 742 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 743 // Clear the promotion information. These pointers can be adjusted 744 // along with all the other pointers into the heap but 745 // compaction is expected to be a rare event with 746 // a heap using cms so don't do it without seeing the need. 747 for (uint i = 0; i < ParallelGCThreads; i++) { 748 _par_gc_thread_states[i]->promo.reset(); 749 } 750 } 751 752 void ConcurrentMarkSweepGeneration::compute_new_size() { 753 assert_locked_or_safepoint(Heap_lock); 754 755 // If incremental collection failed, we just want to expand 756 // to the limit. 757 if (incremental_collection_failed()) { 758 clear_incremental_collection_failed(); 759 grow_to_reserved(); 760 return; 761 } 762 763 // The heap has been compacted but not reset yet. 764 // Any metric such as free() or used() will be incorrect. 765 766 CardGeneration::compute_new_size(); 767 768 // Reset again after a possible resizing 769 if (did_compact()) { 770 cmsSpace()->reset_after_compaction(); 771 } 772 } 773 774 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 775 assert_locked_or_safepoint(Heap_lock); 776 777 // If incremental collection failed, we just want to expand 778 // to the limit. 779 if (incremental_collection_failed()) { 780 clear_incremental_collection_failed(); 781 grow_to_reserved(); 782 return; 783 } 784 785 double free_percentage = ((double) free()) / capacity(); 786 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 787 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 788 789 // compute expansion delta needed for reaching desired free percentage 790 if (free_percentage < desired_free_percentage) { 791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 792 assert(desired_capacity >= capacity(), "invalid expansion size"); 793 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 794 if (PrintGCDetails && Verbose) { 795 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 796 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 797 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 798 gclog_or_tty->print_cr(" Desired free fraction %f", 799 desired_free_percentage); 800 gclog_or_tty->print_cr(" Maximum free fraction %f", 801 maximum_free_percentage); 802 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 803 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 804 desired_capacity/1000); 805 int prev_level = level() - 1; 806 if (prev_level >= 0) { 807 size_t prev_size = 0; 808 GenCollectedHeap* gch = GenCollectedHeap::heap(); 809 Generation* prev_gen = gch->young_gen(); 810 prev_size = prev_gen->capacity(); 811 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 812 prev_size/1000); 813 } 814 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 815 unsafe_max_alloc_nogc()/1000); 816 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 817 contiguous_available()/1000); 818 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 819 expand_bytes); 820 } 821 // safe if expansion fails 822 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 823 if (PrintGCDetails && Verbose) { 824 gclog_or_tty->print_cr(" Expanded free fraction %f", 825 ((double) free()) / capacity()); 826 } 827 } else { 828 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 829 assert(desired_capacity <= capacity(), "invalid expansion size"); 830 size_t shrink_bytes = capacity() - desired_capacity; 831 // Don't shrink unless the delta is greater than the minimum shrink we want 832 if (shrink_bytes >= MinHeapDeltaBytes) { 833 shrink_free_list_by(shrink_bytes); 834 } 835 } 836 } 837 838 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 839 return cmsSpace()->freelistLock(); 840 } 841 842 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 843 bool tlab) { 844 CMSSynchronousYieldRequest yr; 845 MutexLockerEx x(freelistLock(), 846 Mutex::_no_safepoint_check_flag); 847 return have_lock_and_allocate(size, tlab); 848 } 849 850 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 851 bool tlab /* ignored */) { 852 assert_lock_strong(freelistLock()); 853 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 854 HeapWord* res = cmsSpace()->allocate(adjustedSize); 855 // Allocate the object live (grey) if the background collector has 856 // started marking. This is necessary because the marker may 857 // have passed this address and consequently this object will 858 // not otherwise be greyed and would be incorrectly swept up. 859 // Note that if this object contains references, the writing 860 // of those references will dirty the card containing this object 861 // allowing the object to be blackened (and its references scanned) 862 // either during a preclean phase or at the final checkpoint. 863 if (res != NULL) { 864 // We may block here with an uninitialized object with 865 // its mark-bit or P-bits not yet set. Such objects need 866 // to be safely navigable by block_start(). 867 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 868 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 869 collector()->direct_allocated(res, adjustedSize); 870 _direct_allocated_words += adjustedSize; 871 // allocation counters 872 NOT_PRODUCT( 873 _numObjectsAllocated++; 874 _numWordsAllocated += (int)adjustedSize; 875 ) 876 } 877 return res; 878 } 879 880 // In the case of direct allocation by mutators in a generation that 881 // is being concurrently collected, the object must be allocated 882 // live (grey) if the background collector has started marking. 883 // This is necessary because the marker may 884 // have passed this address and consequently this object will 885 // not otherwise be greyed and would be incorrectly swept up. 886 // Note that if this object contains references, the writing 887 // of those references will dirty the card containing this object 888 // allowing the object to be blackened (and its references scanned) 889 // either during a preclean phase or at the final checkpoint. 890 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 891 assert(_markBitMap.covers(start, size), "Out of bounds"); 892 if (_collectorState >= Marking) { 893 MutexLockerEx y(_markBitMap.lock(), 894 Mutex::_no_safepoint_check_flag); 895 // [see comments preceding SweepClosure::do_blk() below for details] 896 // 897 // Can the P-bits be deleted now? JJJ 898 // 899 // 1. need to mark the object as live so it isn't collected 900 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 901 // 3. need to mark the end of the object so marking, precleaning or sweeping 902 // can skip over uninitialized or unparsable objects. An allocated 903 // object is considered uninitialized for our purposes as long as 904 // its klass word is NULL. All old gen objects are parsable 905 // as soon as they are initialized.) 906 _markBitMap.mark(start); // object is live 907 _markBitMap.mark(start + 1); // object is potentially uninitialized? 908 _markBitMap.mark(start + size - 1); 909 // mark end of object 910 } 911 // check that oop looks uninitialized 912 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 913 } 914 915 void CMSCollector::promoted(bool par, HeapWord* start, 916 bool is_obj_array, size_t obj_size) { 917 assert(_markBitMap.covers(start), "Out of bounds"); 918 // See comment in direct_allocated() about when objects should 919 // be allocated live. 920 if (_collectorState >= Marking) { 921 // we already hold the marking bit map lock, taken in 922 // the prologue 923 if (par) { 924 _markBitMap.par_mark(start); 925 } else { 926 _markBitMap.mark(start); 927 } 928 // We don't need to mark the object as uninitialized (as 929 // in direct_allocated above) because this is being done with the 930 // world stopped and the object will be initialized by the 931 // time the marking, precleaning or sweeping get to look at it. 932 // But see the code for copying objects into the CMS generation, 933 // where we need to ensure that concurrent readers of the 934 // block offset table are able to safely navigate a block that 935 // is in flux from being free to being allocated (and in 936 // transition while being copied into) and subsequently 937 // becoming a bona-fide object when the copy/promotion is complete. 938 assert(SafepointSynchronize::is_at_safepoint(), 939 "expect promotion only at safepoints"); 940 941 if (_collectorState < Sweeping) { 942 // Mark the appropriate cards in the modUnionTable, so that 943 // this object gets scanned before the sweep. If this is 944 // not done, CMS generation references in the object might 945 // not get marked. 946 // For the case of arrays, which are otherwise precisely 947 // marked, we need to dirty the entire array, not just its head. 948 if (is_obj_array) { 949 // The [par_]mark_range() method expects mr.end() below to 950 // be aligned to the granularity of a bit's representation 951 // in the heap. In the case of the MUT below, that's a 952 // card size. 953 MemRegion mr(start, 954 (HeapWord*)round_to((intptr_t)(start + obj_size), 955 CardTableModRefBS::card_size /* bytes */)); 956 if (par) { 957 _modUnionTable.par_mark_range(mr); 958 } else { 959 _modUnionTable.mark_range(mr); 960 } 961 } else { // not an obj array; we can just mark the head 962 if (par) { 963 _modUnionTable.par_mark(start); 964 } else { 965 _modUnionTable.mark(start); 966 } 967 } 968 } 969 } 970 } 971 972 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 973 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 974 // allocate, copy and if necessary update promoinfo -- 975 // delegate to underlying space. 976 assert_lock_strong(freelistLock()); 977 978 #ifndef PRODUCT 979 if (GenCollectedHeap::heap()->promotion_should_fail()) { 980 return NULL; 981 } 982 #endif // #ifndef PRODUCT 983 984 oop res = _cmsSpace->promote(obj, obj_size); 985 if (res == NULL) { 986 // expand and retry 987 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 988 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 989 // Since this is the old generation, we don't try to promote 990 // into a more senior generation. 991 res = _cmsSpace->promote(obj, obj_size); 992 } 993 if (res != NULL) { 994 // See comment in allocate() about when objects should 995 // be allocated live. 996 assert(obj->is_oop(), "Will dereference klass pointer below"); 997 collector()->promoted(false, // Not parallel 998 (HeapWord*)res, obj->is_objArray(), obj_size); 999 // promotion counters 1000 NOT_PRODUCT( 1001 _numObjectsPromoted++; 1002 _numWordsPromoted += 1003 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1004 ) 1005 } 1006 return res; 1007 } 1008 1009 1010 // IMPORTANT: Notes on object size recognition in CMS. 1011 // --------------------------------------------------- 1012 // A block of storage in the CMS generation is always in 1013 // one of three states. A free block (FREE), an allocated 1014 // object (OBJECT) whose size() method reports the correct size, 1015 // and an intermediate state (TRANSIENT) in which its size cannot 1016 // be accurately determined. 1017 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1018 // ----------------------------------------------------- 1019 // FREE: klass_word & 1 == 1; mark_word holds block size 1020 // 1021 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1022 // obj->size() computes correct size 1023 // 1024 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1025 // 1026 // STATE IDENTIFICATION: (64 bit+COOPS) 1027 // ------------------------------------ 1028 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1029 // 1030 // OBJECT: klass_word installed; klass_word != 0; 1031 // obj->size() computes correct size 1032 // 1033 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1034 // 1035 // 1036 // STATE TRANSITION DIAGRAM 1037 // 1038 // mut / parnew mut / parnew 1039 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1040 // ^ | 1041 // |------------------------ DEAD <------------------------------------| 1042 // sweep mut 1043 // 1044 // While a block is in TRANSIENT state its size cannot be determined 1045 // so readers will either need to come back later or stall until 1046 // the size can be determined. Note that for the case of direct 1047 // allocation, P-bits, when available, may be used to determine the 1048 // size of an object that may not yet have been initialized. 1049 1050 // Things to support parallel young-gen collection. 1051 oop 1052 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1053 oop old, markOop m, 1054 size_t word_sz) { 1055 #ifndef PRODUCT 1056 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1057 return NULL; 1058 } 1059 #endif // #ifndef PRODUCT 1060 1061 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1062 PromotionInfo* promoInfo = &ps->promo; 1063 // if we are tracking promotions, then first ensure space for 1064 // promotion (including spooling space for saving header if necessary). 1065 // then allocate and copy, then track promoted info if needed. 1066 // When tracking (see PromotionInfo::track()), the mark word may 1067 // be displaced and in this case restoration of the mark word 1068 // occurs in the (oop_since_save_marks_)iterate phase. 1069 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1070 // Out of space for allocating spooling buffers; 1071 // try expanding and allocating spooling buffers. 1072 if (!expand_and_ensure_spooling_space(promoInfo)) { 1073 return NULL; 1074 } 1075 } 1076 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1077 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1078 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1079 if (obj_ptr == NULL) { 1080 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1081 if (obj_ptr == NULL) { 1082 return NULL; 1083 } 1084 } 1085 oop obj = oop(obj_ptr); 1086 OrderAccess::storestore(); 1087 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1088 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1089 // IMPORTANT: See note on object initialization for CMS above. 1090 // Otherwise, copy the object. Here we must be careful to insert the 1091 // klass pointer last, since this marks the block as an allocated object. 1092 // Except with compressed oops it's the mark word. 1093 HeapWord* old_ptr = (HeapWord*)old; 1094 // Restore the mark word copied above. 1095 obj->set_mark(m); 1096 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1097 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1098 OrderAccess::storestore(); 1099 1100 if (UseCompressedClassPointers) { 1101 // Copy gap missed by (aligned) header size calculation below 1102 obj->set_klass_gap(old->klass_gap()); 1103 } 1104 if (word_sz > (size_t)oopDesc::header_size()) { 1105 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1106 obj_ptr + oopDesc::header_size(), 1107 word_sz - oopDesc::header_size()); 1108 } 1109 1110 // Now we can track the promoted object, if necessary. We take care 1111 // to delay the transition from uninitialized to full object 1112 // (i.e., insertion of klass pointer) until after, so that it 1113 // atomically becomes a promoted object. 1114 if (promoInfo->tracking()) { 1115 promoInfo->track((PromotedObject*)obj, old->klass()); 1116 } 1117 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1118 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1119 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1120 1121 // Finally, install the klass pointer (this should be volatile). 1122 OrderAccess::storestore(); 1123 obj->set_klass(old->klass()); 1124 // We should now be able to calculate the right size for this object 1125 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1126 1127 collector()->promoted(true, // parallel 1128 obj_ptr, old->is_objArray(), word_sz); 1129 1130 NOT_PRODUCT( 1131 Atomic::inc_ptr(&_numObjectsPromoted); 1132 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1133 ) 1134 1135 return obj; 1136 } 1137 1138 void 1139 ConcurrentMarkSweepGeneration:: 1140 par_promote_alloc_done(int thread_num) { 1141 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1142 ps->lab.retire(thread_num); 1143 } 1144 1145 void 1146 ConcurrentMarkSweepGeneration:: 1147 par_oop_since_save_marks_iterate_done(int thread_num) { 1148 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1149 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1150 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1151 } 1152 1153 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1154 size_t size, 1155 bool tlab) 1156 { 1157 // We allow a STW collection only if a full 1158 // collection was requested. 1159 return full || should_allocate(size, tlab); // FIX ME !!! 1160 // This and promotion failure handling are connected at the 1161 // hip and should be fixed by untying them. 1162 } 1163 1164 bool CMSCollector::shouldConcurrentCollect() { 1165 if (_full_gc_requested) { 1166 if (Verbose && PrintGCDetails) { 1167 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1168 " gc request (or gc_locker)"); 1169 } 1170 return true; 1171 } 1172 1173 FreelistLocker x(this); 1174 // ------------------------------------------------------------------ 1175 // Print out lots of information which affects the initiation of 1176 // a collection. 1177 if (PrintCMSInitiationStatistics && stats().valid()) { 1178 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1179 gclog_or_tty->stamp(); 1180 gclog_or_tty->cr(); 1181 stats().print_on(gclog_or_tty); 1182 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1183 stats().time_until_cms_gen_full()); 1184 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1185 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1186 _cmsGen->contiguous_available()); 1187 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1188 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1189 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1190 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1191 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1192 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1193 gclog_or_tty->print_cr("metadata initialized %d", 1194 MetaspaceGC::should_concurrent_collect()); 1195 } 1196 // ------------------------------------------------------------------ 1197 1198 // If the estimated time to complete a cms collection (cms_duration()) 1199 // is less than the estimated time remaining until the cms generation 1200 // is full, start a collection. 1201 if (!UseCMSInitiatingOccupancyOnly) { 1202 if (stats().valid()) { 1203 if (stats().time_until_cms_start() == 0.0) { 1204 return true; 1205 } 1206 } else { 1207 // We want to conservatively collect somewhat early in order 1208 // to try and "bootstrap" our CMS/promotion statistics; 1209 // this branch will not fire after the first successful CMS 1210 // collection because the stats should then be valid. 1211 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1212 if (Verbose && PrintGCDetails) { 1213 gclog_or_tty->print_cr( 1214 " CMSCollector: collect for bootstrapping statistics:" 1215 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1216 _bootstrap_occupancy); 1217 } 1218 return true; 1219 } 1220 } 1221 } 1222 1223 // Otherwise, we start a collection cycle if 1224 // old gen want a collection cycle started. Each may use 1225 // an appropriate criterion for making this decision. 1226 // XXX We need to make sure that the gen expansion 1227 // criterion dovetails well with this. XXX NEED TO FIX THIS 1228 if (_cmsGen->should_concurrent_collect()) { 1229 if (Verbose && PrintGCDetails) { 1230 gclog_or_tty->print_cr("CMS old gen initiated"); 1231 } 1232 return true; 1233 } 1234 1235 // We start a collection if we believe an incremental collection may fail; 1236 // this is not likely to be productive in practice because it's probably too 1237 // late anyway. 1238 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1239 assert(gch->collector_policy()->is_generation_policy(), 1240 "You may want to check the correctness of the following"); 1241 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1242 if (Verbose && PrintGCDetails) { 1243 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1244 } 1245 return true; 1246 } 1247 1248 if (MetaspaceGC::should_concurrent_collect()) { 1249 if (Verbose && PrintGCDetails) { 1250 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1251 } 1252 return true; 1253 } 1254 1255 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1256 if (CMSTriggerInterval >= 0) { 1257 if (CMSTriggerInterval == 0) { 1258 // Trigger always 1259 return true; 1260 } 1261 1262 // Check the CMS time since begin (we do not check the stats validity 1263 // as we want to be able to trigger the first CMS cycle as well) 1264 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1265 if (Verbose && PrintGCDetails) { 1266 if (stats().valid()) { 1267 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1268 stats().cms_time_since_begin()); 1269 } else { 1270 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1271 } 1272 } 1273 return true; 1274 } 1275 } 1276 1277 return false; 1278 } 1279 1280 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1281 1282 // Clear _expansion_cause fields of constituent generations 1283 void CMSCollector::clear_expansion_cause() { 1284 _cmsGen->clear_expansion_cause(); 1285 } 1286 1287 // We should be conservative in starting a collection cycle. To 1288 // start too eagerly runs the risk of collecting too often in the 1289 // extreme. To collect too rarely falls back on full collections, 1290 // which works, even if not optimum in terms of concurrent work. 1291 // As a work around for too eagerly collecting, use the flag 1292 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1293 // giving the user an easily understandable way of controlling the 1294 // collections. 1295 // We want to start a new collection cycle if any of the following 1296 // conditions hold: 1297 // . our current occupancy exceeds the configured initiating occupancy 1298 // for this generation, or 1299 // . we recently needed to expand this space and have not, since that 1300 // expansion, done a collection of this generation, or 1301 // . the underlying space believes that it may be a good idea to initiate 1302 // a concurrent collection (this may be based on criteria such as the 1303 // following: the space uses linear allocation and linear allocation is 1304 // going to fail, or there is believed to be excessive fragmentation in 1305 // the generation, etc... or ... 1306 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1307 // the case of the old generation; see CR 6543076): 1308 // we may be approaching a point at which allocation requests may fail because 1309 // we will be out of sufficient free space given allocation rate estimates.] 1310 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1311 1312 assert_lock_strong(freelistLock()); 1313 if (occupancy() > initiating_occupancy()) { 1314 if (PrintGCDetails && Verbose) { 1315 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1316 short_name(), occupancy(), initiating_occupancy()); 1317 } 1318 return true; 1319 } 1320 if (UseCMSInitiatingOccupancyOnly) { 1321 return false; 1322 } 1323 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1324 if (PrintGCDetails && Verbose) { 1325 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1326 short_name()); 1327 } 1328 return true; 1329 } 1330 if (_cmsSpace->should_concurrent_collect()) { 1331 if (PrintGCDetails && Verbose) { 1332 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1333 short_name()); 1334 } 1335 return true; 1336 } 1337 return false; 1338 } 1339 1340 void ConcurrentMarkSweepGeneration::collect(bool full, 1341 bool clear_all_soft_refs, 1342 size_t size, 1343 bool tlab) 1344 { 1345 collector()->collect(full, clear_all_soft_refs, size, tlab); 1346 } 1347 1348 void CMSCollector::collect(bool full, 1349 bool clear_all_soft_refs, 1350 size_t size, 1351 bool tlab) 1352 { 1353 // The following "if" branch is present for defensive reasons. 1354 // In the current uses of this interface, it can be replaced with: 1355 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1356 // But I am not placing that assert here to allow future 1357 // generality in invoking this interface. 1358 if (GC_locker::is_active()) { 1359 // A consistency test for GC_locker 1360 assert(GC_locker::needs_gc(), "Should have been set already"); 1361 // Skip this foreground collection, instead 1362 // expanding the heap if necessary. 1363 // Need the free list locks for the call to free() in compute_new_size() 1364 compute_new_size(); 1365 return; 1366 } 1367 acquire_control_and_collect(full, clear_all_soft_refs); 1368 } 1369 1370 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1371 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1372 unsigned int gc_count = gch->total_full_collections(); 1373 if (gc_count == full_gc_count) { 1374 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1375 _full_gc_requested = true; 1376 _full_gc_cause = cause; 1377 CGC_lock->notify(); // nudge CMS thread 1378 } else { 1379 assert(gc_count > full_gc_count, "Error: causal loop"); 1380 } 1381 } 1382 1383 bool CMSCollector::is_external_interruption() { 1384 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1385 return GCCause::is_user_requested_gc(cause) || 1386 GCCause::is_serviceability_requested_gc(cause); 1387 } 1388 1389 void CMSCollector::report_concurrent_mode_interruption() { 1390 if (is_external_interruption()) { 1391 if (PrintGCDetails) { 1392 gclog_or_tty->print(" (concurrent mode interrupted)"); 1393 } 1394 } else { 1395 if (PrintGCDetails) { 1396 gclog_or_tty->print(" (concurrent mode failure)"); 1397 } 1398 _gc_tracer_cm->report_concurrent_mode_failure(); 1399 } 1400 } 1401 1402 1403 // The foreground and background collectors need to coordinate in order 1404 // to make sure that they do not mutually interfere with CMS collections. 1405 // When a background collection is active, 1406 // the foreground collector may need to take over (preempt) and 1407 // synchronously complete an ongoing collection. Depending on the 1408 // frequency of the background collections and the heap usage 1409 // of the application, this preemption can be seldom or frequent. 1410 // There are only certain 1411 // points in the background collection that the "collection-baton" 1412 // can be passed to the foreground collector. 1413 // 1414 // The foreground collector will wait for the baton before 1415 // starting any part of the collection. The foreground collector 1416 // will only wait at one location. 1417 // 1418 // The background collector will yield the baton before starting a new 1419 // phase of the collection (e.g., before initial marking, marking from roots, 1420 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1421 // of the loop which switches the phases. The background collector does some 1422 // of the phases (initial mark, final re-mark) with the world stopped. 1423 // Because of locking involved in stopping the world, 1424 // the foreground collector should not block waiting for the background 1425 // collector when it is doing a stop-the-world phase. The background 1426 // collector will yield the baton at an additional point just before 1427 // it enters a stop-the-world phase. Once the world is stopped, the 1428 // background collector checks the phase of the collection. If the 1429 // phase has not changed, it proceeds with the collection. If the 1430 // phase has changed, it skips that phase of the collection. See 1431 // the comments on the use of the Heap_lock in collect_in_background(). 1432 // 1433 // Variable used in baton passing. 1434 // _foregroundGCIsActive - Set to true by the foreground collector when 1435 // it wants the baton. The foreground clears it when it has finished 1436 // the collection. 1437 // _foregroundGCShouldWait - Set to true by the background collector 1438 // when it is running. The foreground collector waits while 1439 // _foregroundGCShouldWait is true. 1440 // CGC_lock - monitor used to protect access to the above variables 1441 // and to notify the foreground and background collectors. 1442 // _collectorState - current state of the CMS collection. 1443 // 1444 // The foreground collector 1445 // acquires the CGC_lock 1446 // sets _foregroundGCIsActive 1447 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1448 // various locks acquired in preparation for the collection 1449 // are released so as not to block the background collector 1450 // that is in the midst of a collection 1451 // proceeds with the collection 1452 // clears _foregroundGCIsActive 1453 // returns 1454 // 1455 // The background collector in a loop iterating on the phases of the 1456 // collection 1457 // acquires the CGC_lock 1458 // sets _foregroundGCShouldWait 1459 // if _foregroundGCIsActive is set 1460 // clears _foregroundGCShouldWait, notifies _CGC_lock 1461 // waits on _CGC_lock for _foregroundGCIsActive to become false 1462 // and exits the loop. 1463 // otherwise 1464 // proceed with that phase of the collection 1465 // if the phase is a stop-the-world phase, 1466 // yield the baton once more just before enqueueing 1467 // the stop-world CMS operation (executed by the VM thread). 1468 // returns after all phases of the collection are done 1469 // 1470 1471 void CMSCollector::acquire_control_and_collect(bool full, 1472 bool clear_all_soft_refs) { 1473 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1474 assert(!Thread::current()->is_ConcurrentGC_thread(), 1475 "shouldn't try to acquire control from self!"); 1476 1477 // Start the protocol for acquiring control of the 1478 // collection from the background collector (aka CMS thread). 1479 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1480 "VM thread should have CMS token"); 1481 // Remember the possibly interrupted state of an ongoing 1482 // concurrent collection 1483 CollectorState first_state = _collectorState; 1484 1485 // Signal to a possibly ongoing concurrent collection that 1486 // we want to do a foreground collection. 1487 _foregroundGCIsActive = true; 1488 1489 // release locks and wait for a notify from the background collector 1490 // releasing the locks in only necessary for phases which 1491 // do yields to improve the granularity of the collection. 1492 assert_lock_strong(bitMapLock()); 1493 // We need to lock the Free list lock for the space that we are 1494 // currently collecting. 1495 assert(haveFreelistLocks(), "Must be holding free list locks"); 1496 bitMapLock()->unlock(); 1497 releaseFreelistLocks(); 1498 { 1499 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1500 if (_foregroundGCShouldWait) { 1501 // We are going to be waiting for action for the CMS thread; 1502 // it had better not be gone (for instance at shutdown)! 1503 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1504 "CMS thread must be running"); 1505 // Wait here until the background collector gives us the go-ahead 1506 ConcurrentMarkSweepThread::clear_CMS_flag( 1507 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1508 // Get a possibly blocked CMS thread going: 1509 // Note that we set _foregroundGCIsActive true above, 1510 // without protection of the CGC_lock. 1511 CGC_lock->notify(); 1512 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1513 "Possible deadlock"); 1514 while (_foregroundGCShouldWait) { 1515 // wait for notification 1516 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1517 // Possibility of delay/starvation here, since CMS token does 1518 // not know to give priority to VM thread? Actually, i think 1519 // there wouldn't be any delay/starvation, but the proof of 1520 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1521 } 1522 ConcurrentMarkSweepThread::set_CMS_flag( 1523 ConcurrentMarkSweepThread::CMS_vm_has_token); 1524 } 1525 } 1526 // The CMS_token is already held. Get back the other locks. 1527 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1528 "VM thread should have CMS token"); 1529 getFreelistLocks(); 1530 bitMapLock()->lock_without_safepoint_check(); 1531 if (TraceCMSState) { 1532 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1533 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1534 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1535 } 1536 1537 // Inform cms gen if this was due to partial collection failing. 1538 // The CMS gen may use this fact to determine its expansion policy. 1539 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1540 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1541 assert(!_cmsGen->incremental_collection_failed(), 1542 "Should have been noticed, reacted to and cleared"); 1543 _cmsGen->set_incremental_collection_failed(); 1544 } 1545 1546 if (first_state > Idling) { 1547 report_concurrent_mode_interruption(); 1548 } 1549 1550 set_did_compact(true); 1551 1552 // If the collection is being acquired from the background 1553 // collector, there may be references on the discovered 1554 // references lists. Abandon those references, since some 1555 // of them may have become unreachable after concurrent 1556 // discovery; the STW compacting collector will redo discovery 1557 // more precisely, without being subject to floating garbage. 1558 // Leaving otherwise unreachable references in the discovered 1559 // lists would require special handling. 1560 ref_processor()->disable_discovery(); 1561 ref_processor()->abandon_partial_discovery(); 1562 ref_processor()->verify_no_references_recorded(); 1563 1564 if (first_state > Idling) { 1565 save_heap_summary(); 1566 } 1567 1568 do_compaction_work(clear_all_soft_refs); 1569 1570 // Has the GC time limit been exceeded? 1571 size_t max_eden_size = _young_gen->max_capacity() - 1572 _young_gen->to()->capacity() - 1573 _young_gen->from()->capacity(); 1574 GCCause::Cause gc_cause = gch->gc_cause(); 1575 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1576 _young_gen->eden()->used(), 1577 _cmsGen->max_capacity(), 1578 max_eden_size, 1579 full, 1580 gc_cause, 1581 gch->collector_policy()); 1582 1583 // Reset the expansion cause, now that we just completed 1584 // a collection cycle. 1585 clear_expansion_cause(); 1586 _foregroundGCIsActive = false; 1587 return; 1588 } 1589 1590 // Resize the tenured generation 1591 // after obtaining the free list locks for the 1592 // two generations. 1593 void CMSCollector::compute_new_size() { 1594 assert_locked_or_safepoint(Heap_lock); 1595 FreelistLocker z(this); 1596 MetaspaceGC::compute_new_size(); 1597 _cmsGen->compute_new_size_free_list(); 1598 } 1599 1600 // A work method used by the foreground collector to do 1601 // a mark-sweep-compact. 1602 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1603 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1604 1605 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1606 gc_timer->register_gc_start(); 1607 1608 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1609 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1610 1611 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1612 1613 // Temporarily widen the span of the weak reference processing to 1614 // the entire heap. 1615 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1616 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1617 // Temporarily, clear the "is_alive_non_header" field of the 1618 // reference processor. 1619 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1620 // Temporarily make reference _processing_ single threaded (non-MT). 1621 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1622 // Temporarily make refs discovery atomic 1623 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1624 // Temporarily make reference _discovery_ single threaded (non-MT) 1625 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1626 1627 ref_processor()->set_enqueuing_is_done(false); 1628 ref_processor()->enable_discovery(); 1629 ref_processor()->setup_policy(clear_all_soft_refs); 1630 // If an asynchronous collection finishes, the _modUnionTable is 1631 // all clear. If we are assuming the collection from an asynchronous 1632 // collection, clear the _modUnionTable. 1633 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1634 "_modUnionTable should be clear if the baton was not passed"); 1635 _modUnionTable.clear_all(); 1636 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1637 "mod union for klasses should be clear if the baton was passed"); 1638 _ct->klass_rem_set()->clear_mod_union(); 1639 1640 // We must adjust the allocation statistics being maintained 1641 // in the free list space. We do so by reading and clearing 1642 // the sweep timer and updating the block flux rate estimates below. 1643 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1644 if (_inter_sweep_timer.is_active()) { 1645 _inter_sweep_timer.stop(); 1646 // Note that we do not use this sample to update the _inter_sweep_estimate. 1647 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1648 _inter_sweep_estimate.padded_average(), 1649 _intra_sweep_estimate.padded_average()); 1650 } 1651 1652 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1653 ref_processor(), clear_all_soft_refs); 1654 #ifdef ASSERT 1655 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1656 size_t free_size = cms_space->free(); 1657 assert(free_size == 1658 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1659 * HeapWordSize, 1660 "All the free space should be compacted into one chunk at top"); 1661 assert(cms_space->dictionary()->total_chunk_size( 1662 debug_only(cms_space->freelistLock())) == 0 || 1663 cms_space->totalSizeInIndexedFreeLists() == 0, 1664 "All the free space should be in a single chunk"); 1665 size_t num = cms_space->totalCount(); 1666 assert((free_size == 0 && num == 0) || 1667 (free_size > 0 && (num == 1 || num == 2)), 1668 "There should be at most 2 free chunks after compaction"); 1669 #endif // ASSERT 1670 _collectorState = Resetting; 1671 assert(_restart_addr == NULL, 1672 "Should have been NULL'd before baton was passed"); 1673 reset(false /* == !concurrent */); 1674 _cmsGen->reset_after_compaction(); 1675 _concurrent_cycles_since_last_unload = 0; 1676 1677 // Clear any data recorded in the PLAB chunk arrays. 1678 if (_survivor_plab_array != NULL) { 1679 reset_survivor_plab_arrays(); 1680 } 1681 1682 // Adjust the per-size allocation stats for the next epoch. 1683 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1684 // Restart the "inter sweep timer" for the next epoch. 1685 _inter_sweep_timer.reset(); 1686 _inter_sweep_timer.start(); 1687 1688 gc_timer->register_gc_end(); 1689 1690 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1691 1692 // For a mark-sweep-compact, compute_new_size() will be called 1693 // in the heap's do_collection() method. 1694 } 1695 1696 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1697 ContiguousSpace* eden_space = _young_gen->eden(); 1698 ContiguousSpace* from_space = _young_gen->from(); 1699 ContiguousSpace* to_space = _young_gen->to(); 1700 // Eden 1701 if (_eden_chunk_array != NULL) { 1702 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1703 p2i(eden_space->bottom()), p2i(eden_space->top()), 1704 p2i(eden_space->end()), eden_space->capacity()); 1705 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1706 "_eden_chunk_capacity=" SIZE_FORMAT, 1707 _eden_chunk_index, _eden_chunk_capacity); 1708 for (size_t i = 0; i < _eden_chunk_index; i++) { 1709 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1710 i, p2i(_eden_chunk_array[i])); 1711 } 1712 } 1713 // Survivor 1714 if (_survivor_chunk_array != NULL) { 1715 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1716 p2i(from_space->bottom()), p2i(from_space->top()), 1717 p2i(from_space->end()), from_space->capacity()); 1718 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1719 "_survivor_chunk_capacity=" SIZE_FORMAT, 1720 _survivor_chunk_index, _survivor_chunk_capacity); 1721 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1722 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1723 i, p2i(_survivor_chunk_array[i])); 1724 } 1725 } 1726 } 1727 1728 void CMSCollector::getFreelistLocks() const { 1729 // Get locks for all free lists in all generations that this 1730 // collector is responsible for 1731 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1732 } 1733 1734 void CMSCollector::releaseFreelistLocks() const { 1735 // Release locks for all free lists in all generations that this 1736 // collector is responsible for 1737 _cmsGen->freelistLock()->unlock(); 1738 } 1739 1740 bool CMSCollector::haveFreelistLocks() const { 1741 // Check locks for all free lists in all generations that this 1742 // collector is responsible for 1743 assert_lock_strong(_cmsGen->freelistLock()); 1744 PRODUCT_ONLY(ShouldNotReachHere()); 1745 return true; 1746 } 1747 1748 // A utility class that is used by the CMS collector to 1749 // temporarily "release" the foreground collector from its 1750 // usual obligation to wait for the background collector to 1751 // complete an ongoing phase before proceeding. 1752 class ReleaseForegroundGC: public StackObj { 1753 private: 1754 CMSCollector* _c; 1755 public: 1756 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1757 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1758 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1759 // allow a potentially blocked foreground collector to proceed 1760 _c->_foregroundGCShouldWait = false; 1761 if (_c->_foregroundGCIsActive) { 1762 CGC_lock->notify(); 1763 } 1764 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1765 "Possible deadlock"); 1766 } 1767 1768 ~ReleaseForegroundGC() { 1769 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1771 _c->_foregroundGCShouldWait = true; 1772 } 1773 }; 1774 1775 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1776 assert(Thread::current()->is_ConcurrentGC_thread(), 1777 "A CMS asynchronous collection is only allowed on a CMS thread."); 1778 1779 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1780 { 1781 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1782 MutexLockerEx hl(Heap_lock, safepoint_check); 1783 FreelistLocker fll(this); 1784 MutexLockerEx x(CGC_lock, safepoint_check); 1785 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1786 // The foreground collector is active or we're 1787 // not using asynchronous collections. Skip this 1788 // background collection. 1789 assert(!_foregroundGCShouldWait, "Should be clear"); 1790 return; 1791 } else { 1792 assert(_collectorState == Idling, "Should be idling before start."); 1793 _collectorState = InitialMarking; 1794 register_gc_start(cause); 1795 // Reset the expansion cause, now that we are about to begin 1796 // a new cycle. 1797 clear_expansion_cause(); 1798 1799 // Clear the MetaspaceGC flag since a concurrent collection 1800 // is starting but also clear it after the collection. 1801 MetaspaceGC::set_should_concurrent_collect(false); 1802 } 1803 // Decide if we want to enable class unloading as part of the 1804 // ensuing concurrent GC cycle. 1805 update_should_unload_classes(); 1806 _full_gc_requested = false; // acks all outstanding full gc requests 1807 _full_gc_cause = GCCause::_no_gc; 1808 // Signal that we are about to start a collection 1809 gch->increment_total_full_collections(); // ... starting a collection cycle 1810 _collection_count_start = gch->total_full_collections(); 1811 } 1812 1813 // Used for PrintGC 1814 size_t prev_used; 1815 if (PrintGC && Verbose) { 1816 prev_used = _cmsGen->used(); 1817 } 1818 1819 // The change of the collection state is normally done at this level; 1820 // the exceptions are phases that are executed while the world is 1821 // stopped. For those phases the change of state is done while the 1822 // world is stopped. For baton passing purposes this allows the 1823 // background collector to finish the phase and change state atomically. 1824 // The foreground collector cannot wait on a phase that is done 1825 // while the world is stopped because the foreground collector already 1826 // has the world stopped and would deadlock. 1827 while (_collectorState != Idling) { 1828 if (TraceCMSState) { 1829 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1830 p2i(Thread::current()), _collectorState); 1831 } 1832 // The foreground collector 1833 // holds the Heap_lock throughout its collection. 1834 // holds the CMS token (but not the lock) 1835 // except while it is waiting for the background collector to yield. 1836 // 1837 // The foreground collector should be blocked (not for long) 1838 // if the background collector is about to start a phase 1839 // executed with world stopped. If the background 1840 // collector has already started such a phase, the 1841 // foreground collector is blocked waiting for the 1842 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1843 // are executed in the VM thread. 1844 // 1845 // The locking order is 1846 // PendingListLock (PLL) -- if applicable (FinalMarking) 1847 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1848 // CMS token (claimed in 1849 // stop_world_and_do() --> 1850 // safepoint_synchronize() --> 1851 // CMSThread::synchronize()) 1852 1853 { 1854 // Check if the FG collector wants us to yield. 1855 CMSTokenSync x(true); // is cms thread 1856 if (waitForForegroundGC()) { 1857 // We yielded to a foreground GC, nothing more to be 1858 // done this round. 1859 assert(_foregroundGCShouldWait == false, "We set it to false in " 1860 "waitForForegroundGC()"); 1861 if (TraceCMSState) { 1862 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1863 " exiting collection CMS state %d", 1864 p2i(Thread::current()), _collectorState); 1865 } 1866 return; 1867 } else { 1868 // The background collector can run but check to see if the 1869 // foreground collector has done a collection while the 1870 // background collector was waiting to get the CGC_lock 1871 // above. If yes, break so that _foregroundGCShouldWait 1872 // is cleared before returning. 1873 if (_collectorState == Idling) { 1874 break; 1875 } 1876 } 1877 } 1878 1879 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1880 "should be waiting"); 1881 1882 switch (_collectorState) { 1883 case InitialMarking: 1884 { 1885 ReleaseForegroundGC x(this); 1886 stats().record_cms_begin(); 1887 VM_CMS_Initial_Mark initial_mark_op(this); 1888 VMThread::execute(&initial_mark_op); 1889 } 1890 // The collector state may be any legal state at this point 1891 // since the background collector may have yielded to the 1892 // foreground collector. 1893 break; 1894 case Marking: 1895 // initial marking in checkpointRootsInitialWork has been completed 1896 if (markFromRoots()) { // we were successful 1897 assert(_collectorState == Precleaning, "Collector state should " 1898 "have changed"); 1899 } else { 1900 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1901 } 1902 break; 1903 case Precleaning: 1904 // marking from roots in markFromRoots has been completed 1905 preclean(); 1906 assert(_collectorState == AbortablePreclean || 1907 _collectorState == FinalMarking, 1908 "Collector state should have changed"); 1909 break; 1910 case AbortablePreclean: 1911 abortable_preclean(); 1912 assert(_collectorState == FinalMarking, "Collector state should " 1913 "have changed"); 1914 break; 1915 case FinalMarking: 1916 { 1917 ReleaseForegroundGC x(this); 1918 1919 VM_CMS_Final_Remark final_remark_op(this); 1920 VMThread::execute(&final_remark_op); 1921 } 1922 assert(_foregroundGCShouldWait, "block post-condition"); 1923 break; 1924 case Sweeping: 1925 // final marking in checkpointRootsFinal has been completed 1926 sweep(); 1927 assert(_collectorState == Resizing, "Collector state change " 1928 "to Resizing must be done under the free_list_lock"); 1929 1930 case Resizing: { 1931 // Sweeping has been completed... 1932 // At this point the background collection has completed. 1933 // Don't move the call to compute_new_size() down 1934 // into code that might be executed if the background 1935 // collection was preempted. 1936 { 1937 ReleaseForegroundGC x(this); // unblock FG collection 1938 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1939 CMSTokenSync z(true); // not strictly needed. 1940 if (_collectorState == Resizing) { 1941 compute_new_size(); 1942 save_heap_summary(); 1943 _collectorState = Resetting; 1944 } else { 1945 assert(_collectorState == Idling, "The state should only change" 1946 " because the foreground collector has finished the collection"); 1947 } 1948 } 1949 break; 1950 } 1951 case Resetting: 1952 // CMS heap resizing has been completed 1953 reset(true); 1954 assert(_collectorState == Idling, "Collector state should " 1955 "have changed"); 1956 1957 MetaspaceGC::set_should_concurrent_collect(false); 1958 1959 stats().record_cms_end(); 1960 // Don't move the concurrent_phases_end() and compute_new_size() 1961 // calls to here because a preempted background collection 1962 // has it's state set to "Resetting". 1963 break; 1964 case Idling: 1965 default: 1966 ShouldNotReachHere(); 1967 break; 1968 } 1969 if (TraceCMSState) { 1970 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 } 1973 assert(_foregroundGCShouldWait, "block post-condition"); 1974 } 1975 1976 // Should this be in gc_epilogue? 1977 collector_policy()->counters()->update_counters(); 1978 1979 { 1980 // Clear _foregroundGCShouldWait and, in the event that the 1981 // foreground collector is waiting, notify it, before 1982 // returning. 1983 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1984 _foregroundGCShouldWait = false; 1985 if (_foregroundGCIsActive) { 1986 CGC_lock->notify(); 1987 } 1988 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1989 "Possible deadlock"); 1990 } 1991 if (TraceCMSState) { 1992 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1993 " exiting collection CMS state %d", 1994 p2i(Thread::current()), _collectorState); 1995 } 1996 if (PrintGC && Verbose) { 1997 _cmsGen->print_heap_change(prev_used); 1998 } 1999 } 2000 2001 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2002 _cms_start_registered = true; 2003 _gc_timer_cm->register_gc_start(); 2004 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2005 } 2006 2007 void CMSCollector::register_gc_end() { 2008 if (_cms_start_registered) { 2009 report_heap_summary(GCWhen::AfterGC); 2010 2011 _gc_timer_cm->register_gc_end(); 2012 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2013 _cms_start_registered = false; 2014 } 2015 } 2016 2017 void CMSCollector::save_heap_summary() { 2018 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2019 _last_heap_summary = gch->create_heap_summary(); 2020 _last_metaspace_summary = gch->create_metaspace_summary(); 2021 } 2022 2023 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2024 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2025 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2026 } 2027 2028 bool CMSCollector::waitForForegroundGC() { 2029 bool res = false; 2030 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2031 "CMS thread should have CMS token"); 2032 // Block the foreground collector until the 2033 // background collectors decides whether to 2034 // yield. 2035 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2036 _foregroundGCShouldWait = true; 2037 if (_foregroundGCIsActive) { 2038 // The background collector yields to the 2039 // foreground collector and returns a value 2040 // indicating that it has yielded. The foreground 2041 // collector can proceed. 2042 res = true; 2043 _foregroundGCShouldWait = false; 2044 ConcurrentMarkSweepThread::clear_CMS_flag( 2045 ConcurrentMarkSweepThread::CMS_cms_has_token); 2046 ConcurrentMarkSweepThread::set_CMS_flag( 2047 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2048 // Get a possibly blocked foreground thread going 2049 CGC_lock->notify(); 2050 if (TraceCMSState) { 2051 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2052 p2i(Thread::current()), _collectorState); 2053 } 2054 while (_foregroundGCIsActive) { 2055 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2056 } 2057 ConcurrentMarkSweepThread::set_CMS_flag( 2058 ConcurrentMarkSweepThread::CMS_cms_has_token); 2059 ConcurrentMarkSweepThread::clear_CMS_flag( 2060 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2061 } 2062 if (TraceCMSState) { 2063 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2064 p2i(Thread::current()), _collectorState); 2065 } 2066 return res; 2067 } 2068 2069 // Because of the need to lock the free lists and other structures in 2070 // the collector, common to all the generations that the collector is 2071 // collecting, we need the gc_prologues of individual CMS generations 2072 // delegate to their collector. It may have been simpler had the 2073 // current infrastructure allowed one to call a prologue on a 2074 // collector. In the absence of that we have the generation's 2075 // prologue delegate to the collector, which delegates back 2076 // some "local" work to a worker method in the individual generations 2077 // that it's responsible for collecting, while itself doing any 2078 // work common to all generations it's responsible for. A similar 2079 // comment applies to the gc_epilogue()'s. 2080 // The role of the variable _between_prologue_and_epilogue is to 2081 // enforce the invocation protocol. 2082 void CMSCollector::gc_prologue(bool full) { 2083 // Call gc_prologue_work() for the CMSGen 2084 // we are responsible for. 2085 2086 // The following locking discipline assumes that we are only called 2087 // when the world is stopped. 2088 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2089 2090 // The CMSCollector prologue must call the gc_prologues for the 2091 // "generations" that it's responsible 2092 // for. 2093 2094 assert( Thread::current()->is_VM_thread() 2095 || ( CMSScavengeBeforeRemark 2096 && Thread::current()->is_ConcurrentGC_thread()), 2097 "Incorrect thread type for prologue execution"); 2098 2099 if (_between_prologue_and_epilogue) { 2100 // We have already been invoked; this is a gc_prologue delegation 2101 // from yet another CMS generation that we are responsible for, just 2102 // ignore it since all relevant work has already been done. 2103 return; 2104 } 2105 2106 // set a bit saying prologue has been called; cleared in epilogue 2107 _between_prologue_and_epilogue = true; 2108 // Claim locks for common data structures, then call gc_prologue_work() 2109 // for each CMSGen. 2110 2111 getFreelistLocks(); // gets free list locks on constituent spaces 2112 bitMapLock()->lock_without_safepoint_check(); 2113 2114 // Should call gc_prologue_work() for all cms gens we are responsible for 2115 bool duringMarking = _collectorState >= Marking 2116 && _collectorState < Sweeping; 2117 2118 // The young collections clear the modified oops state, which tells if 2119 // there are any modified oops in the class. The remark phase also needs 2120 // that information. Tell the young collection to save the union of all 2121 // modified klasses. 2122 if (duringMarking) { 2123 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2124 } 2125 2126 bool registerClosure = duringMarking; 2127 2128 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2129 2130 if (!full) { 2131 stats().record_gc0_begin(); 2132 } 2133 } 2134 2135 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2136 2137 _capacity_at_prologue = capacity(); 2138 _used_at_prologue = used(); 2139 2140 // Delegate to CMScollector which knows how to coordinate between 2141 // this and any other CMS generations that it is responsible for 2142 // collecting. 2143 collector()->gc_prologue(full); 2144 } 2145 2146 // This is a "private" interface for use by this generation's CMSCollector. 2147 // Not to be called directly by any other entity (for instance, 2148 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2149 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2150 bool registerClosure, ModUnionClosure* modUnionClosure) { 2151 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2152 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2153 "Should be NULL"); 2154 if (registerClosure) { 2155 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2156 } 2157 cmsSpace()->gc_prologue(); 2158 // Clear stat counters 2159 NOT_PRODUCT( 2160 assert(_numObjectsPromoted == 0, "check"); 2161 assert(_numWordsPromoted == 0, "check"); 2162 if (Verbose && PrintGC) { 2163 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2164 SIZE_FORMAT" bytes concurrently", 2165 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2166 } 2167 _numObjectsAllocated = 0; 2168 _numWordsAllocated = 0; 2169 ) 2170 } 2171 2172 void CMSCollector::gc_epilogue(bool full) { 2173 // The following locking discipline assumes that we are only called 2174 // when the world is stopped. 2175 assert(SafepointSynchronize::is_at_safepoint(), 2176 "world is stopped assumption"); 2177 2178 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2179 // if linear allocation blocks need to be appropriately marked to allow the 2180 // the blocks to be parsable. We also check here whether we need to nudge the 2181 // CMS collector thread to start a new cycle (if it's not already active). 2182 assert( Thread::current()->is_VM_thread() 2183 || ( CMSScavengeBeforeRemark 2184 && Thread::current()->is_ConcurrentGC_thread()), 2185 "Incorrect thread type for epilogue execution"); 2186 2187 if (!_between_prologue_and_epilogue) { 2188 // We have already been invoked; this is a gc_epilogue delegation 2189 // from yet another CMS generation that we are responsible for, just 2190 // ignore it since all relevant work has already been done. 2191 return; 2192 } 2193 assert(haveFreelistLocks(), "must have freelist locks"); 2194 assert_lock_strong(bitMapLock()); 2195 2196 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2197 2198 _cmsGen->gc_epilogue_work(full); 2199 2200 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2201 // in case sampling was not already enabled, enable it 2202 _start_sampling = true; 2203 } 2204 // reset _eden_chunk_array so sampling starts afresh 2205 _eden_chunk_index = 0; 2206 2207 size_t cms_used = _cmsGen->cmsSpace()->used(); 2208 2209 // update performance counters - this uses a special version of 2210 // update_counters() that allows the utilization to be passed as a 2211 // parameter, avoiding multiple calls to used(). 2212 // 2213 _cmsGen->update_counters(cms_used); 2214 2215 bitMapLock()->unlock(); 2216 releaseFreelistLocks(); 2217 2218 if (!CleanChunkPoolAsync) { 2219 Chunk::clean_chunk_pool(); 2220 } 2221 2222 set_did_compact(false); 2223 _between_prologue_and_epilogue = false; // ready for next cycle 2224 } 2225 2226 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2227 collector()->gc_epilogue(full); 2228 2229 // Also reset promotion tracking in par gc thread states. 2230 for (uint i = 0; i < ParallelGCThreads; i++) { 2231 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2232 } 2233 } 2234 2235 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2236 assert(!incremental_collection_failed(), "Should have been cleared"); 2237 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2238 cmsSpace()->gc_epilogue(); 2239 // Print stat counters 2240 NOT_PRODUCT( 2241 assert(_numObjectsAllocated == 0, "check"); 2242 assert(_numWordsAllocated == 0, "check"); 2243 if (Verbose && PrintGC) { 2244 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2245 SIZE_FORMAT" bytes", 2246 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2247 } 2248 _numObjectsPromoted = 0; 2249 _numWordsPromoted = 0; 2250 ) 2251 2252 if (PrintGC && Verbose) { 2253 // Call down the chain in contiguous_available needs the freelistLock 2254 // so print this out before releasing the freeListLock. 2255 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2256 contiguous_available()); 2257 } 2258 } 2259 2260 #ifndef PRODUCT 2261 bool CMSCollector::have_cms_token() { 2262 Thread* thr = Thread::current(); 2263 if (thr->is_VM_thread()) { 2264 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2265 } else if (thr->is_ConcurrentGC_thread()) { 2266 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2267 } else if (thr->is_GC_task_thread()) { 2268 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2269 ParGCRareEvent_lock->owned_by_self(); 2270 } 2271 return false; 2272 } 2273 #endif 2274 2275 // Check reachability of the given heap address in CMS generation, 2276 // treating all other generations as roots. 2277 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2278 // We could "guarantee" below, rather than assert, but I'll 2279 // leave these as "asserts" so that an adventurous debugger 2280 // could try this in the product build provided some subset of 2281 // the conditions were met, provided they were interested in the 2282 // results and knew that the computation below wouldn't interfere 2283 // with other concurrent computations mutating the structures 2284 // being read or written. 2285 assert(SafepointSynchronize::is_at_safepoint(), 2286 "Else mutations in object graph will make answer suspect"); 2287 assert(have_cms_token(), "Should hold cms token"); 2288 assert(haveFreelistLocks(), "must hold free list locks"); 2289 assert_lock_strong(bitMapLock()); 2290 2291 // Clear the marking bit map array before starting, but, just 2292 // for kicks, first report if the given address is already marked 2293 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2294 _markBitMap.isMarked(addr) ? "" : " not"); 2295 2296 if (verify_after_remark()) { 2297 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2298 bool result = verification_mark_bm()->isMarked(addr); 2299 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2300 result ? "IS" : "is NOT"); 2301 return result; 2302 } else { 2303 gclog_or_tty->print_cr("Could not compute result"); 2304 return false; 2305 } 2306 } 2307 2308 2309 void 2310 CMSCollector::print_on_error(outputStream* st) { 2311 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2312 if (collector != NULL) { 2313 CMSBitMap* bitmap = &collector->_markBitMap; 2314 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2315 bitmap->print_on_error(st, " Bits: "); 2316 2317 st->cr(); 2318 2319 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2320 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2321 mut_bitmap->print_on_error(st, " Bits: "); 2322 } 2323 } 2324 2325 //////////////////////////////////////////////////////// 2326 // CMS Verification Support 2327 //////////////////////////////////////////////////////// 2328 // Following the remark phase, the following invariant 2329 // should hold -- each object in the CMS heap which is 2330 // marked in markBitMap() should be marked in the verification_mark_bm(). 2331 2332 class VerifyMarkedClosure: public BitMapClosure { 2333 CMSBitMap* _marks; 2334 bool _failed; 2335 2336 public: 2337 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2338 2339 bool do_bit(size_t offset) { 2340 HeapWord* addr = _marks->offsetToHeapWord(offset); 2341 if (!_marks->isMarked(addr)) { 2342 oop(addr)->print_on(gclog_or_tty); 2343 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2344 _failed = true; 2345 } 2346 return true; 2347 } 2348 2349 bool failed() { return _failed; } 2350 }; 2351 2352 bool CMSCollector::verify_after_remark(bool silent) { 2353 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2354 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2355 static bool init = false; 2356 2357 assert(SafepointSynchronize::is_at_safepoint(), 2358 "Else mutations in object graph will make answer suspect"); 2359 assert(have_cms_token(), 2360 "Else there may be mutual interference in use of " 2361 " verification data structures"); 2362 assert(_collectorState > Marking && _collectorState <= Sweeping, 2363 "Else marking info checked here may be obsolete"); 2364 assert(haveFreelistLocks(), "must hold free list locks"); 2365 assert_lock_strong(bitMapLock()); 2366 2367 2368 // Allocate marking bit map if not already allocated 2369 if (!init) { // first time 2370 if (!verification_mark_bm()->allocate(_span)) { 2371 return false; 2372 } 2373 init = true; 2374 } 2375 2376 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2377 2378 // Turn off refs discovery -- so we will be tracing through refs. 2379 // This is as intended, because by this time 2380 // GC must already have cleared any refs that need to be cleared, 2381 // and traced those that need to be marked; moreover, 2382 // the marking done here is not going to interfere in any 2383 // way with the marking information used by GC. 2384 NoRefDiscovery no_discovery(ref_processor()); 2385 2386 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2387 2388 // Clear any marks from a previous round 2389 verification_mark_bm()->clear_all(); 2390 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2391 verify_work_stacks_empty(); 2392 2393 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2394 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2395 // Update the saved marks which may affect the root scans. 2396 gch->save_marks(); 2397 2398 if (CMSRemarkVerifyVariant == 1) { 2399 // In this first variant of verification, we complete 2400 // all marking, then check if the new marks-vector is 2401 // a subset of the CMS marks-vector. 2402 verify_after_remark_work_1(); 2403 } else if (CMSRemarkVerifyVariant == 2) { 2404 // In this second variant of verification, we flag an error 2405 // (i.e. an object reachable in the new marks-vector not reachable 2406 // in the CMS marks-vector) immediately, also indicating the 2407 // identify of an object (A) that references the unmarked object (B) -- 2408 // presumably, a mutation to A failed to be picked up by preclean/remark? 2409 verify_after_remark_work_2(); 2410 } else { 2411 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2412 CMSRemarkVerifyVariant); 2413 } 2414 if (!silent) gclog_or_tty->print(" done] "); 2415 return true; 2416 } 2417 2418 void CMSCollector::verify_after_remark_work_1() { 2419 ResourceMark rm; 2420 HandleMark hm; 2421 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2422 2423 // Get a clear set of claim bits for the roots processing to work with. 2424 ClassLoaderDataGraph::clear_claimed_marks(); 2425 2426 // Mark from roots one level into CMS 2427 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2428 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2429 2430 gch->gen_process_roots(_cmsGen->level(), 2431 true, // younger gens are roots 2432 true, // activate StrongRootsScope 2433 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2434 should_unload_classes(), 2435 ¬Older, 2436 NULL, 2437 NULL); // SSS: Provide correct closure 2438 2439 // Now mark from the roots 2440 MarkFromRootsClosure markFromRootsClosure(this, _span, 2441 verification_mark_bm(), verification_mark_stack(), 2442 false /* don't yield */, true /* verifying */); 2443 assert(_restart_addr == NULL, "Expected pre-condition"); 2444 verification_mark_bm()->iterate(&markFromRootsClosure); 2445 while (_restart_addr != NULL) { 2446 // Deal with stack overflow: by restarting at the indicated 2447 // address. 2448 HeapWord* ra = _restart_addr; 2449 markFromRootsClosure.reset(ra); 2450 _restart_addr = NULL; 2451 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2452 } 2453 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2454 verify_work_stacks_empty(); 2455 2456 // Marking completed -- now verify that each bit marked in 2457 // verification_mark_bm() is also marked in markBitMap(); flag all 2458 // errors by printing corresponding objects. 2459 VerifyMarkedClosure vcl(markBitMap()); 2460 verification_mark_bm()->iterate(&vcl); 2461 if (vcl.failed()) { 2462 gclog_or_tty->print("Verification failed"); 2463 gch->print_on(gclog_or_tty); 2464 fatal("CMS: failed marking verification after remark"); 2465 } 2466 } 2467 2468 class VerifyKlassOopsKlassClosure : public KlassClosure { 2469 class VerifyKlassOopsClosure : public OopClosure { 2470 CMSBitMap* _bitmap; 2471 public: 2472 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2473 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2474 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2475 } _oop_closure; 2476 public: 2477 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2478 void do_klass(Klass* k) { 2479 k->oops_do(&_oop_closure); 2480 } 2481 }; 2482 2483 void CMSCollector::verify_after_remark_work_2() { 2484 ResourceMark rm; 2485 HandleMark hm; 2486 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2487 2488 // Get a clear set of claim bits for the roots processing to work with. 2489 ClassLoaderDataGraph::clear_claimed_marks(); 2490 2491 // Mark from roots one level into CMS 2492 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2493 markBitMap()); 2494 CLDToOopClosure cld_closure(¬Older, true); 2495 2496 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2497 2498 gch->gen_process_roots(_cmsGen->level(), 2499 true, // younger gens are roots 2500 true, // activate StrongRootsScope 2501 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2502 should_unload_classes(), 2503 ¬Older, 2504 NULL, 2505 &cld_closure); 2506 2507 // Now mark from the roots 2508 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2509 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2510 assert(_restart_addr == NULL, "Expected pre-condition"); 2511 verification_mark_bm()->iterate(&markFromRootsClosure); 2512 while (_restart_addr != NULL) { 2513 // Deal with stack overflow: by restarting at the indicated 2514 // address. 2515 HeapWord* ra = _restart_addr; 2516 markFromRootsClosure.reset(ra); 2517 _restart_addr = NULL; 2518 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2519 } 2520 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2521 verify_work_stacks_empty(); 2522 2523 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2524 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2525 2526 // Marking completed -- now verify that each bit marked in 2527 // verification_mark_bm() is also marked in markBitMap(); flag all 2528 // errors by printing corresponding objects. 2529 VerifyMarkedClosure vcl(markBitMap()); 2530 verification_mark_bm()->iterate(&vcl); 2531 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2532 } 2533 2534 void ConcurrentMarkSweepGeneration::save_marks() { 2535 // delegate to CMS space 2536 cmsSpace()->save_marks(); 2537 for (uint i = 0; i < ParallelGCThreads; i++) { 2538 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2539 } 2540 } 2541 2542 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2543 return cmsSpace()->no_allocs_since_save_marks(); 2544 } 2545 2546 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2547 \ 2548 void ConcurrentMarkSweepGeneration:: \ 2549 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2550 cl->set_generation(this); \ 2551 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2552 cl->reset_generation(); \ 2553 save_marks(); \ 2554 } 2555 2556 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2557 2558 void 2559 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2560 if (freelistLock()->owned_by_self()) { 2561 Generation::oop_iterate(cl); 2562 } else { 2563 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2564 Generation::oop_iterate(cl); 2565 } 2566 } 2567 2568 void 2569 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2570 if (freelistLock()->owned_by_self()) { 2571 Generation::object_iterate(cl); 2572 } else { 2573 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2574 Generation::object_iterate(cl); 2575 } 2576 } 2577 2578 void 2579 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2580 if (freelistLock()->owned_by_self()) { 2581 Generation::safe_object_iterate(cl); 2582 } else { 2583 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2584 Generation::safe_object_iterate(cl); 2585 } 2586 } 2587 2588 void 2589 ConcurrentMarkSweepGeneration::post_compact() { 2590 } 2591 2592 void 2593 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2594 // Fix the linear allocation blocks to look like free blocks. 2595 2596 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2597 // are not called when the heap is verified during universe initialization and 2598 // at vm shutdown. 2599 if (freelistLock()->owned_by_self()) { 2600 cmsSpace()->prepare_for_verify(); 2601 } else { 2602 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2603 cmsSpace()->prepare_for_verify(); 2604 } 2605 } 2606 2607 void 2608 ConcurrentMarkSweepGeneration::verify() { 2609 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2610 // are not called when the heap is verified during universe initialization and 2611 // at vm shutdown. 2612 if (freelistLock()->owned_by_self()) { 2613 cmsSpace()->verify(); 2614 } else { 2615 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2616 cmsSpace()->verify(); 2617 } 2618 } 2619 2620 void CMSCollector::verify() { 2621 _cmsGen->verify(); 2622 } 2623 2624 #ifndef PRODUCT 2625 bool CMSCollector::overflow_list_is_empty() const { 2626 assert(_num_par_pushes >= 0, "Inconsistency"); 2627 if (_overflow_list == NULL) { 2628 assert(_num_par_pushes == 0, "Inconsistency"); 2629 } 2630 return _overflow_list == NULL; 2631 } 2632 2633 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2634 // merely consolidate assertion checks that appear to occur together frequently. 2635 void CMSCollector::verify_work_stacks_empty() const { 2636 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2637 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2638 } 2639 2640 void CMSCollector::verify_overflow_empty() const { 2641 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2642 assert(no_preserved_marks(), "No preserved marks"); 2643 } 2644 #endif // PRODUCT 2645 2646 // Decide if we want to enable class unloading as part of the 2647 // ensuing concurrent GC cycle. We will collect and 2648 // unload classes if it's the case that: 2649 // (1) an explicit gc request has been made and the flag 2650 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2651 // (2) (a) class unloading is enabled at the command line, and 2652 // (b) old gen is getting really full 2653 // NOTE: Provided there is no change in the state of the heap between 2654 // calls to this method, it should have idempotent results. Moreover, 2655 // its results should be monotonically increasing (i.e. going from 0 to 1, 2656 // but not 1 to 0) between successive calls between which the heap was 2657 // not collected. For the implementation below, it must thus rely on 2658 // the property that concurrent_cycles_since_last_unload() 2659 // will not decrease unless a collection cycle happened and that 2660 // _cmsGen->is_too_full() are 2661 // themselves also monotonic in that sense. See check_monotonicity() 2662 // below. 2663 void CMSCollector::update_should_unload_classes() { 2664 _should_unload_classes = false; 2665 // Condition 1 above 2666 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2667 _should_unload_classes = true; 2668 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2669 // Disjuncts 2.b.(i,ii,iii) above 2670 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2671 CMSClassUnloadingMaxInterval) 2672 || _cmsGen->is_too_full(); 2673 } 2674 } 2675 2676 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2677 bool res = should_concurrent_collect(); 2678 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2679 return res; 2680 } 2681 2682 void CMSCollector::setup_cms_unloading_and_verification_state() { 2683 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2684 || VerifyBeforeExit; 2685 const int rso = GenCollectedHeap::SO_AllCodeCache; 2686 2687 // We set the proper root for this CMS cycle here. 2688 if (should_unload_classes()) { // Should unload classes this cycle 2689 remove_root_scanning_option(rso); // Shrink the root set appropriately 2690 set_verifying(should_verify); // Set verification state for this cycle 2691 return; // Nothing else needs to be done at this time 2692 } 2693 2694 // Not unloading classes this cycle 2695 assert(!should_unload_classes(), "Inconsistency!"); 2696 2697 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2698 // Include symbols, strings and code cache elements to prevent their resurrection. 2699 add_root_scanning_option(rso); 2700 set_verifying(true); 2701 } else if (verifying() && !should_verify) { 2702 // We were verifying, but some verification flags got disabled. 2703 set_verifying(false); 2704 // Exclude symbols, strings and code cache elements from root scanning to 2705 // reduce IM and RM pauses. 2706 remove_root_scanning_option(rso); 2707 } 2708 } 2709 2710 2711 #ifndef PRODUCT 2712 HeapWord* CMSCollector::block_start(const void* p) const { 2713 const HeapWord* addr = (HeapWord*)p; 2714 if (_span.contains(p)) { 2715 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2716 return _cmsGen->cmsSpace()->block_start(p); 2717 } 2718 } 2719 return NULL; 2720 } 2721 #endif 2722 2723 HeapWord* 2724 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2725 bool tlab, 2726 bool parallel) { 2727 CMSSynchronousYieldRequest yr; 2728 assert(!tlab, "Can't deal with TLAB allocation"); 2729 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2730 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2731 if (GCExpandToAllocateDelayMillis > 0) { 2732 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2733 } 2734 return have_lock_and_allocate(word_size, tlab); 2735 } 2736 2737 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2738 size_t bytes, 2739 size_t expand_bytes, 2740 CMSExpansionCause::Cause cause) 2741 { 2742 2743 bool success = expand(bytes, expand_bytes); 2744 2745 // remember why we expanded; this information is used 2746 // by shouldConcurrentCollect() when making decisions on whether to start 2747 // a new CMS cycle. 2748 if (success) { 2749 set_expansion_cause(cause); 2750 if (PrintGCDetails && Verbose) { 2751 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2752 CMSExpansionCause::to_string(cause)); 2753 } 2754 } 2755 } 2756 2757 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2758 HeapWord* res = NULL; 2759 MutexLocker x(ParGCRareEvent_lock); 2760 while (true) { 2761 // Expansion by some other thread might make alloc OK now: 2762 res = ps->lab.alloc(word_sz); 2763 if (res != NULL) return res; 2764 // If there's not enough expansion space available, give up. 2765 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2766 return NULL; 2767 } 2768 // Otherwise, we try expansion. 2769 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2770 // Now go around the loop and try alloc again; 2771 // A competing par_promote might beat us to the expansion space, 2772 // so we may go around the loop again if promotion fails again. 2773 if (GCExpandToAllocateDelayMillis > 0) { 2774 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2775 } 2776 } 2777 } 2778 2779 2780 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2781 PromotionInfo* promo) { 2782 MutexLocker x(ParGCRareEvent_lock); 2783 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2784 while (true) { 2785 // Expansion by some other thread might make alloc OK now: 2786 if (promo->ensure_spooling_space()) { 2787 assert(promo->has_spooling_space(), 2788 "Post-condition of successful ensure_spooling_space()"); 2789 return true; 2790 } 2791 // If there's not enough expansion space available, give up. 2792 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2793 return false; 2794 } 2795 // Otherwise, we try expansion. 2796 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2797 // Now go around the loop and try alloc again; 2798 // A competing allocation might beat us to the expansion space, 2799 // so we may go around the loop again if allocation fails again. 2800 if (GCExpandToAllocateDelayMillis > 0) { 2801 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2802 } 2803 } 2804 } 2805 2806 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2807 // Only shrink if a compaction was done so that all the free space 2808 // in the generation is in a contiguous block at the end. 2809 if (did_compact()) { 2810 CardGeneration::shrink(bytes); 2811 } 2812 } 2813 2814 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2815 assert_locked_or_safepoint(Heap_lock); 2816 } 2817 2818 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2819 assert_locked_or_safepoint(Heap_lock); 2820 assert_lock_strong(freelistLock()); 2821 if (PrintGCDetails && Verbose) { 2822 warning("Shrinking of CMS not yet implemented"); 2823 } 2824 return; 2825 } 2826 2827 2828 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2829 // phases. 2830 class CMSPhaseAccounting: public StackObj { 2831 public: 2832 CMSPhaseAccounting(CMSCollector *collector, 2833 const char *phase, 2834 const GCId gc_id, 2835 bool print_cr = true); 2836 ~CMSPhaseAccounting(); 2837 2838 private: 2839 CMSCollector *_collector; 2840 const char *_phase; 2841 elapsedTimer _wallclock; 2842 bool _print_cr; 2843 const GCId _gc_id; 2844 2845 public: 2846 // Not MT-safe; so do not pass around these StackObj's 2847 // where they may be accessed by other threads. 2848 jlong wallclock_millis() { 2849 assert(_wallclock.is_active(), "Wall clock should not stop"); 2850 _wallclock.stop(); // to record time 2851 jlong ret = _wallclock.milliseconds(); 2852 _wallclock.start(); // restart 2853 return ret; 2854 } 2855 }; 2856 2857 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2858 const char *phase, 2859 const GCId gc_id, 2860 bool print_cr) : 2861 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2862 2863 if (PrintCMSStatistics != 0) { 2864 _collector->resetYields(); 2865 } 2866 if (PrintGCDetails) { 2867 gclog_or_tty->gclog_stamp(_gc_id); 2868 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2869 _collector->cmsGen()->short_name(), _phase); 2870 } 2871 _collector->resetTimer(); 2872 _wallclock.start(); 2873 _collector->startTimer(); 2874 } 2875 2876 CMSPhaseAccounting::~CMSPhaseAccounting() { 2877 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2878 _collector->stopTimer(); 2879 _wallclock.stop(); 2880 if (PrintGCDetails) { 2881 gclog_or_tty->gclog_stamp(_gc_id); 2882 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2883 _collector->cmsGen()->short_name(), 2884 _phase, _collector->timerValue(), _wallclock.seconds()); 2885 if (_print_cr) { 2886 gclog_or_tty->cr(); 2887 } 2888 if (PrintCMSStatistics != 0) { 2889 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2890 _collector->yields()); 2891 } 2892 } 2893 } 2894 2895 // CMS work 2896 2897 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2898 class CMSParMarkTask : public AbstractGangTask { 2899 protected: 2900 CMSCollector* _collector; 2901 uint _n_workers; 2902 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2903 AbstractGangTask(name), 2904 _collector(collector), 2905 _n_workers(n_workers) {} 2906 // Work method in support of parallel rescan ... of young gen spaces 2907 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2908 ContiguousSpace* space, 2909 HeapWord** chunk_array, size_t chunk_top); 2910 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2911 }; 2912 2913 // Parallel initial mark task 2914 class CMSParInitialMarkTask: public CMSParMarkTask { 2915 public: 2916 CMSParInitialMarkTask(CMSCollector* collector, uint n_workers) : 2917 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2918 collector, n_workers) {} 2919 void work(uint worker_id); 2920 }; 2921 2922 // Checkpoint the roots into this generation from outside 2923 // this generation. [Note this initial checkpoint need only 2924 // be approximate -- we'll do a catch up phase subsequently.] 2925 void CMSCollector::checkpointRootsInitial() { 2926 assert(_collectorState == InitialMarking, "Wrong collector state"); 2927 check_correct_thread_executing(); 2928 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2929 2930 save_heap_summary(); 2931 report_heap_summary(GCWhen::BeforeGC); 2932 2933 ReferenceProcessor* rp = ref_processor(); 2934 assert(_restart_addr == NULL, "Control point invariant"); 2935 { 2936 // acquire locks for subsequent manipulations 2937 MutexLockerEx x(bitMapLock(), 2938 Mutex::_no_safepoint_check_flag); 2939 checkpointRootsInitialWork(); 2940 // enable ("weak") refs discovery 2941 rp->enable_discovery(); 2942 _collectorState = Marking; 2943 } 2944 } 2945 2946 void CMSCollector::checkpointRootsInitialWork() { 2947 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2948 assert(_collectorState == InitialMarking, "just checking"); 2949 2950 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2951 // precede our marking with a collection of all 2952 // younger generations to keep floating garbage to a minimum. 2953 // XXX: we won't do this for now -- it's an optimization to be done later. 2954 2955 // already have locks 2956 assert_lock_strong(bitMapLock()); 2957 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2958 2959 // Setup the verification and class unloading state for this 2960 // CMS collection cycle. 2961 setup_cms_unloading_and_verification_state(); 2962 2963 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2964 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2965 2966 // Reset all the PLAB chunk arrays if necessary. 2967 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2968 reset_survivor_plab_arrays(); 2969 } 2970 2971 ResourceMark rm; 2972 HandleMark hm; 2973 2974 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2975 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2976 2977 verify_work_stacks_empty(); 2978 verify_overflow_empty(); 2979 2980 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2981 // Update the saved marks which may affect the root scans. 2982 gch->save_marks(); 2983 2984 // weak reference processing has not started yet. 2985 ref_processor()->set_enqueuing_is_done(false); 2986 2987 // Need to remember all newly created CLDs, 2988 // so that we can guarantee that the remark finds them. 2989 ClassLoaderDataGraph::remember_new_clds(true); 2990 2991 // Whenever a CLD is found, it will be claimed before proceeding to mark 2992 // the klasses. The claimed marks need to be cleared before marking starts. 2993 ClassLoaderDataGraph::clear_claimed_marks(); 2994 2995 if (CMSPrintEdenSurvivorChunks) { 2996 print_eden_and_survivor_chunk_arrays(); 2997 } 2998 2999 { 3000 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3001 if (CMSParallelInitialMarkEnabled) { 3002 // The parallel version. 3003 FlexibleWorkGang* workers = gch->workers(); 3004 assert(workers != NULL, "Need parallel worker threads."); 3005 uint n_workers = workers->active_workers(); 3006 CMSParInitialMarkTask tsk(this, n_workers); 3007 gch->set_par_threads(n_workers); 3008 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3009 if (n_workers > 1) { 3010 StrongRootsScope srs; 3011 workers->run_task(&tsk); 3012 } else { 3013 StrongRootsScope srs; 3014 tsk.work(0); 3015 } 3016 gch->set_par_threads(0); 3017 } else { 3018 // The serial version. 3019 CLDToOopClosure cld_closure(¬Older, true); 3020 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3021 gch->gen_process_roots(_cmsGen->level(), 3022 true, // younger gens are roots 3023 true, // activate StrongRootsScope 3024 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3025 should_unload_classes(), 3026 ¬Older, 3027 NULL, 3028 &cld_closure); 3029 } 3030 } 3031 3032 // Clear mod-union table; it will be dirtied in the prologue of 3033 // CMS generation per each younger generation collection. 3034 3035 assert(_modUnionTable.isAllClear(), 3036 "Was cleared in most recent final checkpoint phase" 3037 " or no bits are set in the gc_prologue before the start of the next " 3038 "subsequent marking phase."); 3039 3040 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3041 3042 // Save the end of the used_region of the constituent generations 3043 // to be used to limit the extent of sweep in each generation. 3044 save_sweep_limits(); 3045 verify_overflow_empty(); 3046 } 3047 3048 bool CMSCollector::markFromRoots() { 3049 // we might be tempted to assert that: 3050 // assert(!SafepointSynchronize::is_at_safepoint(), 3051 // "inconsistent argument?"); 3052 // However that wouldn't be right, because it's possible that 3053 // a safepoint is indeed in progress as a younger generation 3054 // stop-the-world GC happens even as we mark in this generation. 3055 assert(_collectorState == Marking, "inconsistent state?"); 3056 check_correct_thread_executing(); 3057 verify_overflow_empty(); 3058 3059 // Weak ref discovery note: We may be discovering weak 3060 // refs in this generation concurrent (but interleaved) with 3061 // weak ref discovery by a younger generation collector. 3062 3063 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3064 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3065 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3066 bool res = markFromRootsWork(); 3067 if (res) { 3068 _collectorState = Precleaning; 3069 } else { // We failed and a foreground collection wants to take over 3070 assert(_foregroundGCIsActive, "internal state inconsistency"); 3071 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3072 if (PrintGCDetails) { 3073 gclog_or_tty->print_cr("bailing out to foreground collection"); 3074 } 3075 } 3076 verify_overflow_empty(); 3077 return res; 3078 } 3079 3080 bool CMSCollector::markFromRootsWork() { 3081 // iterate over marked bits in bit map, doing a full scan and mark 3082 // from these roots using the following algorithm: 3083 // . if oop is to the right of the current scan pointer, 3084 // mark corresponding bit (we'll process it later) 3085 // . else (oop is to left of current scan pointer) 3086 // push oop on marking stack 3087 // . drain the marking stack 3088 3089 // Note that when we do a marking step we need to hold the 3090 // bit map lock -- recall that direct allocation (by mutators) 3091 // and promotion (by younger generation collectors) is also 3092 // marking the bit map. [the so-called allocate live policy.] 3093 // Because the implementation of bit map marking is not 3094 // robust wrt simultaneous marking of bits in the same word, 3095 // we need to make sure that there is no such interference 3096 // between concurrent such updates. 3097 3098 // already have locks 3099 assert_lock_strong(bitMapLock()); 3100 3101 verify_work_stacks_empty(); 3102 verify_overflow_empty(); 3103 bool result = false; 3104 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3105 result = do_marking_mt(); 3106 } else { 3107 result = do_marking_st(); 3108 } 3109 return result; 3110 } 3111 3112 // Forward decl 3113 class CMSConcMarkingTask; 3114 3115 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3116 CMSCollector* _collector; 3117 CMSConcMarkingTask* _task; 3118 public: 3119 virtual void yield(); 3120 3121 // "n_threads" is the number of threads to be terminated. 3122 // "queue_set" is a set of work queues of other threads. 3123 // "collector" is the CMS collector associated with this task terminator. 3124 // "yield" indicates whether we need the gang as a whole to yield. 3125 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3126 ParallelTaskTerminator(n_threads, queue_set), 3127 _collector(collector) { } 3128 3129 void set_task(CMSConcMarkingTask* task) { 3130 _task = task; 3131 } 3132 }; 3133 3134 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3135 CMSConcMarkingTask* _task; 3136 public: 3137 bool should_exit_termination(); 3138 void set_task(CMSConcMarkingTask* task) { 3139 _task = task; 3140 } 3141 }; 3142 3143 // MT Concurrent Marking Task 3144 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3145 CMSCollector* _collector; 3146 uint _n_workers; // requested/desired # workers 3147 bool _result; 3148 CompactibleFreeListSpace* _cms_space; 3149 char _pad_front[64]; // padding to ... 3150 HeapWord* _global_finger; // ... avoid sharing cache line 3151 char _pad_back[64]; 3152 HeapWord* _restart_addr; 3153 3154 // Exposed here for yielding support 3155 Mutex* const _bit_map_lock; 3156 3157 // The per thread work queues, available here for stealing 3158 OopTaskQueueSet* _task_queues; 3159 3160 // Termination (and yielding) support 3161 CMSConcMarkingTerminator _term; 3162 CMSConcMarkingTerminatorTerminator _term_term; 3163 3164 public: 3165 CMSConcMarkingTask(CMSCollector* collector, 3166 CompactibleFreeListSpace* cms_space, 3167 YieldingFlexibleWorkGang* workers, 3168 OopTaskQueueSet* task_queues): 3169 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3170 _collector(collector), 3171 _cms_space(cms_space), 3172 _n_workers(0), _result(true), 3173 _task_queues(task_queues), 3174 _term(_n_workers, task_queues, _collector), 3175 _bit_map_lock(collector->bitMapLock()) 3176 { 3177 _requested_size = _n_workers; 3178 _term.set_task(this); 3179 _term_term.set_task(this); 3180 _restart_addr = _global_finger = _cms_space->bottom(); 3181 } 3182 3183 3184 OopTaskQueueSet* task_queues() { return _task_queues; } 3185 3186 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3187 3188 HeapWord** global_finger_addr() { return &_global_finger; } 3189 3190 CMSConcMarkingTerminator* terminator() { return &_term; } 3191 3192 virtual void set_for_termination(uint active_workers) { 3193 terminator()->reset_for_reuse(active_workers); 3194 } 3195 3196 void work(uint worker_id); 3197 bool should_yield() { 3198 return ConcurrentMarkSweepThread::should_yield() 3199 && !_collector->foregroundGCIsActive(); 3200 } 3201 3202 virtual void coordinator_yield(); // stuff done by coordinator 3203 bool result() { return _result; } 3204 3205 void reset(HeapWord* ra) { 3206 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3207 _restart_addr = _global_finger = ra; 3208 _term.reset_for_reuse(); 3209 } 3210 3211 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3212 OopTaskQueue* work_q); 3213 3214 private: 3215 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3216 void do_work_steal(int i); 3217 void bump_global_finger(HeapWord* f); 3218 }; 3219 3220 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3221 assert(_task != NULL, "Error"); 3222 return _task->yielding(); 3223 // Note that we do not need the disjunct || _task->should_yield() above 3224 // because we want terminating threads to yield only if the task 3225 // is already in the midst of yielding, which happens only after at least one 3226 // thread has yielded. 3227 } 3228 3229 void CMSConcMarkingTerminator::yield() { 3230 if (_task->should_yield()) { 3231 _task->yield(); 3232 } else { 3233 ParallelTaskTerminator::yield(); 3234 } 3235 } 3236 3237 //////////////////////////////////////////////////////////////// 3238 // Concurrent Marking Algorithm Sketch 3239 //////////////////////////////////////////////////////////////// 3240 // Until all tasks exhausted (both spaces): 3241 // -- claim next available chunk 3242 // -- bump global finger via CAS 3243 // -- find first object that starts in this chunk 3244 // and start scanning bitmap from that position 3245 // -- scan marked objects for oops 3246 // -- CAS-mark target, and if successful: 3247 // . if target oop is above global finger (volatile read) 3248 // nothing to do 3249 // . if target oop is in chunk and above local finger 3250 // then nothing to do 3251 // . else push on work-queue 3252 // -- Deal with possible overflow issues: 3253 // . local work-queue overflow causes stuff to be pushed on 3254 // global (common) overflow queue 3255 // . always first empty local work queue 3256 // . then get a batch of oops from global work queue if any 3257 // . then do work stealing 3258 // -- When all tasks claimed (both spaces) 3259 // and local work queue empty, 3260 // then in a loop do: 3261 // . check global overflow stack; steal a batch of oops and trace 3262 // . try to steal from other threads oif GOS is empty 3263 // . if neither is available, offer termination 3264 // -- Terminate and return result 3265 // 3266 void CMSConcMarkingTask::work(uint worker_id) { 3267 elapsedTimer _timer; 3268 ResourceMark rm; 3269 HandleMark hm; 3270 3271 DEBUG_ONLY(_collector->verify_overflow_empty();) 3272 3273 // Before we begin work, our work queue should be empty 3274 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3275 // Scan the bitmap covering _cms_space, tracing through grey objects. 3276 _timer.start(); 3277 do_scan_and_mark(worker_id, _cms_space); 3278 _timer.stop(); 3279 if (PrintCMSStatistics != 0) { 3280 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3281 worker_id, _timer.seconds()); 3282 // XXX: need xxx/xxx type of notation, two timers 3283 } 3284 3285 // ... do work stealing 3286 _timer.reset(); 3287 _timer.start(); 3288 do_work_steal(worker_id); 3289 _timer.stop(); 3290 if (PrintCMSStatistics != 0) { 3291 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3292 worker_id, _timer.seconds()); 3293 // XXX: need xxx/xxx type of notation, two timers 3294 } 3295 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3296 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3297 // Note that under the current task protocol, the 3298 // following assertion is true even of the spaces 3299 // expanded since the completion of the concurrent 3300 // marking. XXX This will likely change under a strict 3301 // ABORT semantics. 3302 // After perm removal the comparison was changed to 3303 // greater than or equal to from strictly greater than. 3304 // Before perm removal the highest address sweep would 3305 // have been at the end of perm gen but now is at the 3306 // end of the tenured gen. 3307 assert(_global_finger >= _cms_space->end(), 3308 "All tasks have been completed"); 3309 DEBUG_ONLY(_collector->verify_overflow_empty();) 3310 } 3311 3312 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3313 HeapWord* read = _global_finger; 3314 HeapWord* cur = read; 3315 while (f > read) { 3316 cur = read; 3317 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3318 if (cur == read) { 3319 // our cas succeeded 3320 assert(_global_finger >= f, "protocol consistency"); 3321 break; 3322 } 3323 } 3324 } 3325 3326 // This is really inefficient, and should be redone by 3327 // using (not yet available) block-read and -write interfaces to the 3328 // stack and the work_queue. XXX FIX ME !!! 3329 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3330 OopTaskQueue* work_q) { 3331 // Fast lock-free check 3332 if (ovflw_stk->length() == 0) { 3333 return false; 3334 } 3335 assert(work_q->size() == 0, "Shouldn't steal"); 3336 MutexLockerEx ml(ovflw_stk->par_lock(), 3337 Mutex::_no_safepoint_check_flag); 3338 // Grab up to 1/4 the size of the work queue 3339 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3340 (size_t)ParGCDesiredObjsFromOverflowList); 3341 num = MIN2(num, ovflw_stk->length()); 3342 for (int i = (int) num; i > 0; i--) { 3343 oop cur = ovflw_stk->pop(); 3344 assert(cur != NULL, "Counted wrong?"); 3345 work_q->push(cur); 3346 } 3347 return num > 0; 3348 } 3349 3350 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3351 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3352 int n_tasks = pst->n_tasks(); 3353 // We allow that there may be no tasks to do here because 3354 // we are restarting after a stack overflow. 3355 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3356 uint nth_task = 0; 3357 3358 HeapWord* aligned_start = sp->bottom(); 3359 if (sp->used_region().contains(_restart_addr)) { 3360 // Align down to a card boundary for the start of 0th task 3361 // for this space. 3362 aligned_start = 3363 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3364 CardTableModRefBS::card_size); 3365 } 3366 3367 size_t chunk_size = sp->marking_task_size(); 3368 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3369 // Having claimed the nth task in this space, 3370 // compute the chunk that it corresponds to: 3371 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3372 aligned_start + (nth_task+1)*chunk_size); 3373 // Try and bump the global finger via a CAS; 3374 // note that we need to do the global finger bump 3375 // _before_ taking the intersection below, because 3376 // the task corresponding to that region will be 3377 // deemed done even if the used_region() expands 3378 // because of allocation -- as it almost certainly will 3379 // during start-up while the threads yield in the 3380 // closure below. 3381 HeapWord* finger = span.end(); 3382 bump_global_finger(finger); // atomically 3383 // There are null tasks here corresponding to chunks 3384 // beyond the "top" address of the space. 3385 span = span.intersection(sp->used_region()); 3386 if (!span.is_empty()) { // Non-null task 3387 HeapWord* prev_obj; 3388 assert(!span.contains(_restart_addr) || nth_task == 0, 3389 "Inconsistency"); 3390 if (nth_task == 0) { 3391 // For the 0th task, we'll not need to compute a block_start. 3392 if (span.contains(_restart_addr)) { 3393 // In the case of a restart because of stack overflow, 3394 // we might additionally skip a chunk prefix. 3395 prev_obj = _restart_addr; 3396 } else { 3397 prev_obj = span.start(); 3398 } 3399 } else { 3400 // We want to skip the first object because 3401 // the protocol is to scan any object in its entirety 3402 // that _starts_ in this span; a fortiori, any 3403 // object starting in an earlier span is scanned 3404 // as part of an earlier claimed task. 3405 // Below we use the "careful" version of block_start 3406 // so we do not try to navigate uninitialized objects. 3407 prev_obj = sp->block_start_careful(span.start()); 3408 // Below we use a variant of block_size that uses the 3409 // Printezis bits to avoid waiting for allocated 3410 // objects to become initialized/parsable. 3411 while (prev_obj < span.start()) { 3412 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3413 if (sz > 0) { 3414 prev_obj += sz; 3415 } else { 3416 // In this case we may end up doing a bit of redundant 3417 // scanning, but that appears unavoidable, short of 3418 // locking the free list locks; see bug 6324141. 3419 break; 3420 } 3421 } 3422 } 3423 if (prev_obj < span.end()) { 3424 MemRegion my_span = MemRegion(prev_obj, span.end()); 3425 // Do the marking work within a non-empty span -- 3426 // the last argument to the constructor indicates whether the 3427 // iteration should be incremental with periodic yields. 3428 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3429 &_collector->_markBitMap, 3430 work_queue(i), 3431 &_collector->_markStack); 3432 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3433 } // else nothing to do for this task 3434 } // else nothing to do for this task 3435 } 3436 // We'd be tempted to assert here that since there are no 3437 // more tasks left to claim in this space, the global_finger 3438 // must exceed space->top() and a fortiori space->end(). However, 3439 // that would not quite be correct because the bumping of 3440 // global_finger occurs strictly after the claiming of a task, 3441 // so by the time we reach here the global finger may not yet 3442 // have been bumped up by the thread that claimed the last 3443 // task. 3444 pst->all_tasks_completed(); 3445 } 3446 3447 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3448 private: 3449 CMSCollector* _collector; 3450 CMSConcMarkingTask* _task; 3451 MemRegion _span; 3452 CMSBitMap* _bit_map; 3453 CMSMarkStack* _overflow_stack; 3454 OopTaskQueue* _work_queue; 3455 protected: 3456 DO_OOP_WORK_DEFN 3457 public: 3458 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3459 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3460 MetadataAwareOopClosure(collector->ref_processor()), 3461 _collector(collector), 3462 _task(task), 3463 _span(collector->_span), 3464 _work_queue(work_queue), 3465 _bit_map(bit_map), 3466 _overflow_stack(overflow_stack) 3467 { } 3468 virtual void do_oop(oop* p); 3469 virtual void do_oop(narrowOop* p); 3470 3471 void trim_queue(size_t max); 3472 void handle_stack_overflow(HeapWord* lost); 3473 void do_yield_check() { 3474 if (_task->should_yield()) { 3475 _task->yield(); 3476 } 3477 } 3478 }; 3479 3480 // Grey object scanning during work stealing phase -- 3481 // the salient assumption here is that any references 3482 // that are in these stolen objects being scanned must 3483 // already have been initialized (else they would not have 3484 // been published), so we do not need to check for 3485 // uninitialized objects before pushing here. 3486 void Par_ConcMarkingClosure::do_oop(oop obj) { 3487 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3488 HeapWord* addr = (HeapWord*)obj; 3489 // Check if oop points into the CMS generation 3490 // and is not marked 3491 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3492 // a white object ... 3493 // If we manage to "claim" the object, by being the 3494 // first thread to mark it, then we push it on our 3495 // marking stack 3496 if (_bit_map->par_mark(addr)) { // ... now grey 3497 // push on work queue (grey set) 3498 bool simulate_overflow = false; 3499 NOT_PRODUCT( 3500 if (CMSMarkStackOverflowALot && 3501 _collector->simulate_overflow()) { 3502 // simulate a stack overflow 3503 simulate_overflow = true; 3504 } 3505 ) 3506 if (simulate_overflow || 3507 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3508 // stack overflow 3509 if (PrintCMSStatistics != 0) { 3510 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3511 SIZE_FORMAT, _overflow_stack->capacity()); 3512 } 3513 // We cannot assert that the overflow stack is full because 3514 // it may have been emptied since. 3515 assert(simulate_overflow || 3516 _work_queue->size() == _work_queue->max_elems(), 3517 "Else push should have succeeded"); 3518 handle_stack_overflow(addr); 3519 } 3520 } // Else, some other thread got there first 3521 do_yield_check(); 3522 } 3523 } 3524 3525 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3526 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3527 3528 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3529 while (_work_queue->size() > max) { 3530 oop new_oop; 3531 if (_work_queue->pop_local(new_oop)) { 3532 assert(new_oop->is_oop(), "Should be an oop"); 3533 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3534 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3535 new_oop->oop_iterate(this); // do_oop() above 3536 do_yield_check(); 3537 } 3538 } 3539 } 3540 3541 // Upon stack overflow, we discard (part of) the stack, 3542 // remembering the least address amongst those discarded 3543 // in CMSCollector's _restart_address. 3544 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3545 // We need to do this under a mutex to prevent other 3546 // workers from interfering with the work done below. 3547 MutexLockerEx ml(_overflow_stack->par_lock(), 3548 Mutex::_no_safepoint_check_flag); 3549 // Remember the least grey address discarded 3550 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3551 _collector->lower_restart_addr(ra); 3552 _overflow_stack->reset(); // discard stack contents 3553 _overflow_stack->expand(); // expand the stack if possible 3554 } 3555 3556 3557 void CMSConcMarkingTask::do_work_steal(int i) { 3558 OopTaskQueue* work_q = work_queue(i); 3559 oop obj_to_scan; 3560 CMSBitMap* bm = &(_collector->_markBitMap); 3561 CMSMarkStack* ovflw = &(_collector->_markStack); 3562 int* seed = _collector->hash_seed(i); 3563 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3564 while (true) { 3565 cl.trim_queue(0); 3566 assert(work_q->size() == 0, "Should have been emptied above"); 3567 if (get_work_from_overflow_stack(ovflw, work_q)) { 3568 // Can't assert below because the work obtained from the 3569 // overflow stack may already have been stolen from us. 3570 // assert(work_q->size() > 0, "Work from overflow stack"); 3571 continue; 3572 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3573 assert(obj_to_scan->is_oop(), "Should be an oop"); 3574 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3575 obj_to_scan->oop_iterate(&cl); 3576 } else if (terminator()->offer_termination(&_term_term)) { 3577 assert(work_q->size() == 0, "Impossible!"); 3578 break; 3579 } else if (yielding() || should_yield()) { 3580 yield(); 3581 } 3582 } 3583 } 3584 3585 // This is run by the CMS (coordinator) thread. 3586 void CMSConcMarkingTask::coordinator_yield() { 3587 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3588 "CMS thread should hold CMS token"); 3589 // First give up the locks, then yield, then re-lock 3590 // We should probably use a constructor/destructor idiom to 3591 // do this unlock/lock or modify the MutexUnlocker class to 3592 // serve our purpose. XXX 3593 assert_lock_strong(_bit_map_lock); 3594 _bit_map_lock->unlock(); 3595 ConcurrentMarkSweepThread::desynchronize(true); 3596 _collector->stopTimer(); 3597 if (PrintCMSStatistics != 0) { 3598 _collector->incrementYields(); 3599 } 3600 3601 // It is possible for whichever thread initiated the yield request 3602 // not to get a chance to wake up and take the bitmap lock between 3603 // this thread releasing it and reacquiring it. So, while the 3604 // should_yield() flag is on, let's sleep for a bit to give the 3605 // other thread a chance to wake up. The limit imposed on the number 3606 // of iterations is defensive, to avoid any unforseen circumstances 3607 // putting us into an infinite loop. Since it's always been this 3608 // (coordinator_yield()) method that was observed to cause the 3609 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3610 // which is by default non-zero. For the other seven methods that 3611 // also perform the yield operation, as are using a different 3612 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3613 // can enable the sleeping for those methods too, if necessary. 3614 // See 6442774. 3615 // 3616 // We really need to reconsider the synchronization between the GC 3617 // thread and the yield-requesting threads in the future and we 3618 // should really use wait/notify, which is the recommended 3619 // way of doing this type of interaction. Additionally, we should 3620 // consolidate the eight methods that do the yield operation and they 3621 // are almost identical into one for better maintainability and 3622 // readability. See 6445193. 3623 // 3624 // Tony 2006.06.29 3625 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3626 ConcurrentMarkSweepThread::should_yield() && 3627 !CMSCollector::foregroundGCIsActive(); ++i) { 3628 os::sleep(Thread::current(), 1, false); 3629 } 3630 3631 ConcurrentMarkSweepThread::synchronize(true); 3632 _bit_map_lock->lock_without_safepoint_check(); 3633 _collector->startTimer(); 3634 } 3635 3636 bool CMSCollector::do_marking_mt() { 3637 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3638 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3639 conc_workers()->active_workers(), 3640 Threads::number_of_non_daemon_threads()); 3641 conc_workers()->set_active_workers(num_workers); 3642 3643 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3644 3645 CMSConcMarkingTask tsk(this, 3646 cms_space, 3647 conc_workers(), 3648 task_queues()); 3649 3650 // Since the actual number of workers we get may be different 3651 // from the number we requested above, do we need to do anything different 3652 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3653 // class?? XXX 3654 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3655 3656 // Refs discovery is already non-atomic. 3657 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3658 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3659 conc_workers()->start_task(&tsk); 3660 while (tsk.yielded()) { 3661 tsk.coordinator_yield(); 3662 conc_workers()->continue_task(&tsk); 3663 } 3664 // If the task was aborted, _restart_addr will be non-NULL 3665 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3666 while (_restart_addr != NULL) { 3667 // XXX For now we do not make use of ABORTED state and have not 3668 // yet implemented the right abort semantics (even in the original 3669 // single-threaded CMS case). That needs some more investigation 3670 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3671 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3672 // If _restart_addr is non-NULL, a marking stack overflow 3673 // occurred; we need to do a fresh marking iteration from the 3674 // indicated restart address. 3675 if (_foregroundGCIsActive) { 3676 // We may be running into repeated stack overflows, having 3677 // reached the limit of the stack size, while making very 3678 // slow forward progress. It may be best to bail out and 3679 // let the foreground collector do its job. 3680 // Clear _restart_addr, so that foreground GC 3681 // works from scratch. This avoids the headache of 3682 // a "rescan" which would otherwise be needed because 3683 // of the dirty mod union table & card table. 3684 _restart_addr = NULL; 3685 return false; 3686 } 3687 // Adjust the task to restart from _restart_addr 3688 tsk.reset(_restart_addr); 3689 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3690 _restart_addr); 3691 _restart_addr = NULL; 3692 // Get the workers going again 3693 conc_workers()->start_task(&tsk); 3694 while (tsk.yielded()) { 3695 tsk.coordinator_yield(); 3696 conc_workers()->continue_task(&tsk); 3697 } 3698 } 3699 assert(tsk.completed(), "Inconsistency"); 3700 assert(tsk.result() == true, "Inconsistency"); 3701 return true; 3702 } 3703 3704 bool CMSCollector::do_marking_st() { 3705 ResourceMark rm; 3706 HandleMark hm; 3707 3708 // Temporarily make refs discovery single threaded (non-MT) 3709 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3710 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3711 &_markStack, CMSYield); 3712 // the last argument to iterate indicates whether the iteration 3713 // should be incremental with periodic yields. 3714 _markBitMap.iterate(&markFromRootsClosure); 3715 // If _restart_addr is non-NULL, a marking stack overflow 3716 // occurred; we need to do a fresh iteration from the 3717 // indicated restart address. 3718 while (_restart_addr != NULL) { 3719 if (_foregroundGCIsActive) { 3720 // We may be running into repeated stack overflows, having 3721 // reached the limit of the stack size, while making very 3722 // slow forward progress. It may be best to bail out and 3723 // let the foreground collector do its job. 3724 // Clear _restart_addr, so that foreground GC 3725 // works from scratch. This avoids the headache of 3726 // a "rescan" which would otherwise be needed because 3727 // of the dirty mod union table & card table. 3728 _restart_addr = NULL; 3729 return false; // indicating failure to complete marking 3730 } 3731 // Deal with stack overflow: 3732 // we restart marking from _restart_addr 3733 HeapWord* ra = _restart_addr; 3734 markFromRootsClosure.reset(ra); 3735 _restart_addr = NULL; 3736 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3737 } 3738 return true; 3739 } 3740 3741 void CMSCollector::preclean() { 3742 check_correct_thread_executing(); 3743 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3744 verify_work_stacks_empty(); 3745 verify_overflow_empty(); 3746 _abort_preclean = false; 3747 if (CMSPrecleaningEnabled) { 3748 if (!CMSEdenChunksRecordAlways) { 3749 _eden_chunk_index = 0; 3750 } 3751 size_t used = get_eden_used(); 3752 size_t capacity = get_eden_capacity(); 3753 // Don't start sampling unless we will get sufficiently 3754 // many samples. 3755 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3756 * CMSScheduleRemarkEdenPenetration)) { 3757 _start_sampling = true; 3758 } else { 3759 _start_sampling = false; 3760 } 3761 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3762 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3763 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3764 } 3765 CMSTokenSync x(true); // is cms thread 3766 if (CMSPrecleaningEnabled) { 3767 sample_eden(); 3768 _collectorState = AbortablePreclean; 3769 } else { 3770 _collectorState = FinalMarking; 3771 } 3772 verify_work_stacks_empty(); 3773 verify_overflow_empty(); 3774 } 3775 3776 // Try and schedule the remark such that young gen 3777 // occupancy is CMSScheduleRemarkEdenPenetration %. 3778 void CMSCollector::abortable_preclean() { 3779 check_correct_thread_executing(); 3780 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3781 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3782 3783 // If Eden's current occupancy is below this threshold, 3784 // immediately schedule the remark; else preclean 3785 // past the next scavenge in an effort to 3786 // schedule the pause as described above. By choosing 3787 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3788 // we will never do an actual abortable preclean cycle. 3789 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3790 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3791 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3792 // We need more smarts in the abortable preclean 3793 // loop below to deal with cases where allocation 3794 // in young gen is very very slow, and our precleaning 3795 // is running a losing race against a horde of 3796 // mutators intent on flooding us with CMS updates 3797 // (dirty cards). 3798 // One, admittedly dumb, strategy is to give up 3799 // after a certain number of abortable precleaning loops 3800 // or after a certain maximum time. We want to make 3801 // this smarter in the next iteration. 3802 // XXX FIX ME!!! YSR 3803 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3804 while (!(should_abort_preclean() || 3805 ConcurrentMarkSweepThread::should_terminate())) { 3806 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3807 cumworkdone += workdone; 3808 loops++; 3809 // Voluntarily terminate abortable preclean phase if we have 3810 // been at it for too long. 3811 if ((CMSMaxAbortablePrecleanLoops != 0) && 3812 loops >= CMSMaxAbortablePrecleanLoops) { 3813 if (PrintGCDetails) { 3814 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3815 } 3816 break; 3817 } 3818 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3819 if (PrintGCDetails) { 3820 gclog_or_tty->print(" CMS: abort preclean due to time "); 3821 } 3822 break; 3823 } 3824 // If we are doing little work each iteration, we should 3825 // take a short break. 3826 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3827 // Sleep for some time, waiting for work to accumulate 3828 stopTimer(); 3829 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3830 startTimer(); 3831 waited++; 3832 } 3833 } 3834 if (PrintCMSStatistics > 0) { 3835 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3836 loops, waited, cumworkdone); 3837 } 3838 } 3839 CMSTokenSync x(true); // is cms thread 3840 if (_collectorState != Idling) { 3841 assert(_collectorState == AbortablePreclean, 3842 "Spontaneous state transition?"); 3843 _collectorState = FinalMarking; 3844 } // Else, a foreground collection completed this CMS cycle. 3845 return; 3846 } 3847 3848 // Respond to an Eden sampling opportunity 3849 void CMSCollector::sample_eden() { 3850 // Make sure a young gc cannot sneak in between our 3851 // reading and recording of a sample. 3852 assert(Thread::current()->is_ConcurrentGC_thread(), 3853 "Only the cms thread may collect Eden samples"); 3854 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3855 "Should collect samples while holding CMS token"); 3856 if (!_start_sampling) { 3857 return; 3858 } 3859 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3860 // is populated by the young generation. 3861 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3862 if (_eden_chunk_index < _eden_chunk_capacity) { 3863 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3864 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3865 "Unexpected state of Eden"); 3866 // We'd like to check that what we just sampled is an oop-start address; 3867 // however, we cannot do that here since the object may not yet have been 3868 // initialized. So we'll instead do the check when we _use_ this sample 3869 // later. 3870 if (_eden_chunk_index == 0 || 3871 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3872 _eden_chunk_array[_eden_chunk_index-1]) 3873 >= CMSSamplingGrain)) { 3874 _eden_chunk_index++; // commit sample 3875 } 3876 } 3877 } 3878 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3879 size_t used = get_eden_used(); 3880 size_t capacity = get_eden_capacity(); 3881 assert(used <= capacity, "Unexpected state of Eden"); 3882 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3883 _abort_preclean = true; 3884 } 3885 } 3886 } 3887 3888 3889 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3890 assert(_collectorState == Precleaning || 3891 _collectorState == AbortablePreclean, "incorrect state"); 3892 ResourceMark rm; 3893 HandleMark hm; 3894 3895 // Precleaning is currently not MT but the reference processor 3896 // may be set for MT. Disable it temporarily here. 3897 ReferenceProcessor* rp = ref_processor(); 3898 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3899 3900 // Do one pass of scrubbing the discovered reference lists 3901 // to remove any reference objects with strongly-reachable 3902 // referents. 3903 if (clean_refs) { 3904 CMSPrecleanRefsYieldClosure yield_cl(this); 3905 assert(rp->span().equals(_span), "Spans should be equal"); 3906 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3907 &_markStack, true /* preclean */); 3908 CMSDrainMarkingStackClosure complete_trace(this, 3909 _span, &_markBitMap, &_markStack, 3910 &keep_alive, true /* preclean */); 3911 3912 // We don't want this step to interfere with a young 3913 // collection because we don't want to take CPU 3914 // or memory bandwidth away from the young GC threads 3915 // (which may be as many as there are CPUs). 3916 // Note that we don't need to protect ourselves from 3917 // interference with mutators because they can't 3918 // manipulate the discovered reference lists nor affect 3919 // the computed reachability of the referents, the 3920 // only properties manipulated by the precleaning 3921 // of these reference lists. 3922 stopTimer(); 3923 CMSTokenSyncWithLocks x(true /* is cms thread */, 3924 bitMapLock()); 3925 startTimer(); 3926 sample_eden(); 3927 3928 // The following will yield to allow foreground 3929 // collection to proceed promptly. XXX YSR: 3930 // The code in this method may need further 3931 // tweaking for better performance and some restructuring 3932 // for cleaner interfaces. 3933 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3934 rp->preclean_discovered_references( 3935 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3936 gc_timer, _gc_tracer_cm->gc_id()); 3937 } 3938 3939 if (clean_survivor) { // preclean the active survivor space(s) 3940 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3941 &_markBitMap, &_modUnionTable, 3942 &_markStack, true /* precleaning phase */); 3943 stopTimer(); 3944 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3945 bitMapLock()); 3946 startTimer(); 3947 unsigned int before_count = 3948 GenCollectedHeap::heap()->total_collections(); 3949 SurvivorSpacePrecleanClosure 3950 sss_cl(this, _span, &_markBitMap, &_markStack, 3951 &pam_cl, before_count, CMSYield); 3952 _young_gen->from()->object_iterate_careful(&sss_cl); 3953 _young_gen->to()->object_iterate_careful(&sss_cl); 3954 } 3955 MarkRefsIntoAndScanClosure 3956 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3957 &_markStack, this, CMSYield, 3958 true /* precleaning phase */); 3959 // CAUTION: The following closure has persistent state that may need to 3960 // be reset upon a decrease in the sequence of addresses it 3961 // processes. 3962 ScanMarkedObjectsAgainCarefullyClosure 3963 smoac_cl(this, _span, 3964 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3965 3966 // Preclean dirty cards in ModUnionTable and CardTable using 3967 // appropriate convergence criterion; 3968 // repeat CMSPrecleanIter times unless we find that 3969 // we are losing. 3970 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3971 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3972 "Bad convergence multiplier"); 3973 assert(CMSPrecleanThreshold >= 100, 3974 "Unreasonably low CMSPrecleanThreshold"); 3975 3976 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3977 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3978 numIter < CMSPrecleanIter; 3979 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3980 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3981 if (Verbose && PrintGCDetails) { 3982 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3983 } 3984 // Either there are very few dirty cards, so re-mark 3985 // pause will be small anyway, or our pre-cleaning isn't 3986 // that much faster than the rate at which cards are being 3987 // dirtied, so we might as well stop and re-mark since 3988 // precleaning won't improve our re-mark time by much. 3989 if (curNumCards <= CMSPrecleanThreshold || 3990 (numIter > 0 && 3991 (curNumCards * CMSPrecleanDenominator > 3992 lastNumCards * CMSPrecleanNumerator))) { 3993 numIter++; 3994 cumNumCards += curNumCards; 3995 break; 3996 } 3997 } 3998 3999 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4000 4001 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4002 cumNumCards += curNumCards; 4003 if (PrintGCDetails && PrintCMSStatistics != 0) { 4004 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4005 curNumCards, cumNumCards, numIter); 4006 } 4007 return cumNumCards; // as a measure of useful work done 4008 } 4009 4010 // PRECLEANING NOTES: 4011 // Precleaning involves: 4012 // . reading the bits of the modUnionTable and clearing the set bits. 4013 // . For the cards corresponding to the set bits, we scan the 4014 // objects on those cards. This means we need the free_list_lock 4015 // so that we can safely iterate over the CMS space when scanning 4016 // for oops. 4017 // . When we scan the objects, we'll be both reading and setting 4018 // marks in the marking bit map, so we'll need the marking bit map. 4019 // . For protecting _collector_state transitions, we take the CGC_lock. 4020 // Note that any races in the reading of of card table entries by the 4021 // CMS thread on the one hand and the clearing of those entries by the 4022 // VM thread or the setting of those entries by the mutator threads on the 4023 // other are quite benign. However, for efficiency it makes sense to keep 4024 // the VM thread from racing with the CMS thread while the latter is 4025 // dirty card info to the modUnionTable. We therefore also use the 4026 // CGC_lock to protect the reading of the card table and the mod union 4027 // table by the CM thread. 4028 // . We run concurrently with mutator updates, so scanning 4029 // needs to be done carefully -- we should not try to scan 4030 // potentially uninitialized objects. 4031 // 4032 // Locking strategy: While holding the CGC_lock, we scan over and 4033 // reset a maximal dirty range of the mod union / card tables, then lock 4034 // the free_list_lock and bitmap lock to do a full marking, then 4035 // release these locks; and repeat the cycle. This allows for a 4036 // certain amount of fairness in the sharing of these locks between 4037 // the CMS collector on the one hand, and the VM thread and the 4038 // mutators on the other. 4039 4040 // NOTE: preclean_mod_union_table() and preclean_card_table() 4041 // further below are largely identical; if you need to modify 4042 // one of these methods, please check the other method too. 4043 4044 size_t CMSCollector::preclean_mod_union_table( 4045 ConcurrentMarkSweepGeneration* gen, 4046 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4047 verify_work_stacks_empty(); 4048 verify_overflow_empty(); 4049 4050 // strategy: starting with the first card, accumulate contiguous 4051 // ranges of dirty cards; clear these cards, then scan the region 4052 // covered by these cards. 4053 4054 // Since all of the MUT is committed ahead, we can just use 4055 // that, in case the generations expand while we are precleaning. 4056 // It might also be fine to just use the committed part of the 4057 // generation, but we might potentially miss cards when the 4058 // generation is rapidly expanding while we are in the midst 4059 // of precleaning. 4060 HeapWord* startAddr = gen->reserved().start(); 4061 HeapWord* endAddr = gen->reserved().end(); 4062 4063 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4064 4065 size_t numDirtyCards, cumNumDirtyCards; 4066 HeapWord *nextAddr, *lastAddr; 4067 for (cumNumDirtyCards = numDirtyCards = 0, 4068 nextAddr = lastAddr = startAddr; 4069 nextAddr < endAddr; 4070 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4071 4072 ResourceMark rm; 4073 HandleMark hm; 4074 4075 MemRegion dirtyRegion; 4076 { 4077 stopTimer(); 4078 // Potential yield point 4079 CMSTokenSync ts(true); 4080 startTimer(); 4081 sample_eden(); 4082 // Get dirty region starting at nextOffset (inclusive), 4083 // simultaneously clearing it. 4084 dirtyRegion = 4085 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4086 assert(dirtyRegion.start() >= nextAddr, 4087 "returned region inconsistent?"); 4088 } 4089 // Remember where the next search should begin. 4090 // The returned region (if non-empty) is a right open interval, 4091 // so lastOffset is obtained from the right end of that 4092 // interval. 4093 lastAddr = dirtyRegion.end(); 4094 // Should do something more transparent and less hacky XXX 4095 numDirtyCards = 4096 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4097 4098 // We'll scan the cards in the dirty region (with periodic 4099 // yields for foreground GC as needed). 4100 if (!dirtyRegion.is_empty()) { 4101 assert(numDirtyCards > 0, "consistency check"); 4102 HeapWord* stop_point = NULL; 4103 stopTimer(); 4104 // Potential yield point 4105 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4106 bitMapLock()); 4107 startTimer(); 4108 { 4109 verify_work_stacks_empty(); 4110 verify_overflow_empty(); 4111 sample_eden(); 4112 stop_point = 4113 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4114 } 4115 if (stop_point != NULL) { 4116 // The careful iteration stopped early either because it found an 4117 // uninitialized object, or because we were in the midst of an 4118 // "abortable preclean", which should now be aborted. Redirty 4119 // the bits corresponding to the partially-scanned or unscanned 4120 // cards. We'll either restart at the next block boundary or 4121 // abort the preclean. 4122 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4123 "Should only be AbortablePreclean."); 4124 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4125 if (should_abort_preclean()) { 4126 break; // out of preclean loop 4127 } else { 4128 // Compute the next address at which preclean should pick up; 4129 // might need bitMapLock in order to read P-bits. 4130 lastAddr = next_card_start_after_block(stop_point); 4131 } 4132 } 4133 } else { 4134 assert(lastAddr == endAddr, "consistency check"); 4135 assert(numDirtyCards == 0, "consistency check"); 4136 break; 4137 } 4138 } 4139 verify_work_stacks_empty(); 4140 verify_overflow_empty(); 4141 return cumNumDirtyCards; 4142 } 4143 4144 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4145 // below are largely identical; if you need to modify 4146 // one of these methods, please check the other method too. 4147 4148 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4149 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4150 // strategy: it's similar to precleamModUnionTable above, in that 4151 // we accumulate contiguous ranges of dirty cards, mark these cards 4152 // precleaned, then scan the region covered by these cards. 4153 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4154 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4155 4156 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4157 4158 size_t numDirtyCards, cumNumDirtyCards; 4159 HeapWord *lastAddr, *nextAddr; 4160 4161 for (cumNumDirtyCards = numDirtyCards = 0, 4162 nextAddr = lastAddr = startAddr; 4163 nextAddr < endAddr; 4164 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4165 4166 ResourceMark rm; 4167 HandleMark hm; 4168 4169 MemRegion dirtyRegion; 4170 { 4171 // See comments in "Precleaning notes" above on why we 4172 // do this locking. XXX Could the locking overheads be 4173 // too high when dirty cards are sparse? [I don't think so.] 4174 stopTimer(); 4175 CMSTokenSync x(true); // is cms thread 4176 startTimer(); 4177 sample_eden(); 4178 // Get and clear dirty region from card table 4179 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4180 MemRegion(nextAddr, endAddr), 4181 true, 4182 CardTableModRefBS::precleaned_card_val()); 4183 4184 assert(dirtyRegion.start() >= nextAddr, 4185 "returned region inconsistent?"); 4186 } 4187 lastAddr = dirtyRegion.end(); 4188 numDirtyCards = 4189 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4190 4191 if (!dirtyRegion.is_empty()) { 4192 stopTimer(); 4193 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4194 startTimer(); 4195 sample_eden(); 4196 verify_work_stacks_empty(); 4197 verify_overflow_empty(); 4198 HeapWord* stop_point = 4199 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4200 if (stop_point != NULL) { 4201 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4202 "Should only be AbortablePreclean."); 4203 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4204 if (should_abort_preclean()) { 4205 break; // out of preclean loop 4206 } else { 4207 // Compute the next address at which preclean should pick up. 4208 lastAddr = next_card_start_after_block(stop_point); 4209 } 4210 } 4211 } else { 4212 break; 4213 } 4214 } 4215 verify_work_stacks_empty(); 4216 verify_overflow_empty(); 4217 return cumNumDirtyCards; 4218 } 4219 4220 class PrecleanKlassClosure : public KlassClosure { 4221 KlassToOopClosure _cm_klass_closure; 4222 public: 4223 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4224 void do_klass(Klass* k) { 4225 if (k->has_accumulated_modified_oops()) { 4226 k->clear_accumulated_modified_oops(); 4227 4228 _cm_klass_closure.do_klass(k); 4229 } 4230 } 4231 }; 4232 4233 // The freelist lock is needed to prevent asserts, is it really needed? 4234 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4235 4236 cl->set_freelistLock(freelistLock); 4237 4238 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4239 4240 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4241 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4242 PrecleanKlassClosure preclean_klass_closure(cl); 4243 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4244 4245 verify_work_stacks_empty(); 4246 verify_overflow_empty(); 4247 } 4248 4249 void CMSCollector::checkpointRootsFinal() { 4250 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4251 check_correct_thread_executing(); 4252 // world is stopped at this checkpoint 4253 assert(SafepointSynchronize::is_at_safepoint(), 4254 "world should be stopped"); 4255 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4256 4257 verify_work_stacks_empty(); 4258 verify_overflow_empty(); 4259 4260 if (PrintGCDetails) { 4261 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4262 _young_gen->used() / K, 4263 _young_gen->capacity() / K); 4264 } 4265 { 4266 if (CMSScavengeBeforeRemark) { 4267 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4268 // Temporarily set flag to false, GCH->do_collection will 4269 // expect it to be false and set to true 4270 FlagSetting fl(gch->_is_gc_active, false); 4271 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4272 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4273 int level = _cmsGen->level() - 1; 4274 if (level >= 0) { 4275 gch->do_collection(true, // full (i.e. force, see below) 4276 false, // !clear_all_soft_refs 4277 0, // size 4278 false, // is_tlab 4279 level // max_level 4280 ); 4281 } 4282 } 4283 FreelistLocker x(this); 4284 MutexLockerEx y(bitMapLock(), 4285 Mutex::_no_safepoint_check_flag); 4286 checkpointRootsFinalWork(); 4287 } 4288 verify_work_stacks_empty(); 4289 verify_overflow_empty(); 4290 } 4291 4292 void CMSCollector::checkpointRootsFinalWork() { 4293 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4294 4295 assert(haveFreelistLocks(), "must have free list locks"); 4296 assert_lock_strong(bitMapLock()); 4297 4298 ResourceMark rm; 4299 HandleMark hm; 4300 4301 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4302 4303 if (should_unload_classes()) { 4304 CodeCache::gc_prologue(); 4305 } 4306 assert(haveFreelistLocks(), "must have free list locks"); 4307 assert_lock_strong(bitMapLock()); 4308 4309 // We might assume that we need not fill TLAB's when 4310 // CMSScavengeBeforeRemark is set, because we may have just done 4311 // a scavenge which would have filled all TLAB's -- and besides 4312 // Eden would be empty. This however may not always be the case -- 4313 // for instance although we asked for a scavenge, it may not have 4314 // happened because of a JNI critical section. We probably need 4315 // a policy for deciding whether we can in that case wait until 4316 // the critical section releases and then do the remark following 4317 // the scavenge, and skip it here. In the absence of that policy, 4318 // or of an indication of whether the scavenge did indeed occur, 4319 // we cannot rely on TLAB's having been filled and must do 4320 // so here just in case a scavenge did not happen. 4321 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4322 // Update the saved marks which may affect the root scans. 4323 gch->save_marks(); 4324 4325 if (CMSPrintEdenSurvivorChunks) { 4326 print_eden_and_survivor_chunk_arrays(); 4327 } 4328 4329 { 4330 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4331 4332 // Note on the role of the mod union table: 4333 // Since the marker in "markFromRoots" marks concurrently with 4334 // mutators, it is possible for some reachable objects not to have been 4335 // scanned. For instance, an only reference to an object A was 4336 // placed in object B after the marker scanned B. Unless B is rescanned, 4337 // A would be collected. Such updates to references in marked objects 4338 // are detected via the mod union table which is the set of all cards 4339 // dirtied since the first checkpoint in this GC cycle and prior to 4340 // the most recent young generation GC, minus those cleaned up by the 4341 // concurrent precleaning. 4342 if (CMSParallelRemarkEnabled) { 4343 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4344 do_remark_parallel(); 4345 } else { 4346 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4347 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4348 do_remark_non_parallel(); 4349 } 4350 } 4351 verify_work_stacks_empty(); 4352 verify_overflow_empty(); 4353 4354 { 4355 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4356 refProcessingWork(); 4357 } 4358 verify_work_stacks_empty(); 4359 verify_overflow_empty(); 4360 4361 if (should_unload_classes()) { 4362 CodeCache::gc_epilogue(); 4363 } 4364 JvmtiExport::gc_epilogue(); 4365 4366 // If we encountered any (marking stack / work queue) overflow 4367 // events during the current CMS cycle, take appropriate 4368 // remedial measures, where possible, so as to try and avoid 4369 // recurrence of that condition. 4370 assert(_markStack.isEmpty(), "No grey objects"); 4371 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4372 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4373 if (ser_ovflw > 0) { 4374 if (PrintCMSStatistics != 0) { 4375 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4376 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4377 ", kac_preclean="SIZE_FORMAT")", 4378 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4379 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4380 } 4381 _markStack.expand(); 4382 _ser_pmc_remark_ovflw = 0; 4383 _ser_pmc_preclean_ovflw = 0; 4384 _ser_kac_preclean_ovflw = 0; 4385 _ser_kac_ovflw = 0; 4386 } 4387 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4388 if (PrintCMSStatistics != 0) { 4389 gclog_or_tty->print_cr("Work queue overflow (benign) " 4390 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4391 _par_pmc_remark_ovflw, _par_kac_ovflw); 4392 } 4393 _par_pmc_remark_ovflw = 0; 4394 _par_kac_ovflw = 0; 4395 } 4396 if (PrintCMSStatistics != 0) { 4397 if (_markStack._hit_limit > 0) { 4398 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4399 _markStack._hit_limit); 4400 } 4401 if (_markStack._failed_double > 0) { 4402 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4403 " current capacity "SIZE_FORMAT, 4404 _markStack._failed_double, 4405 _markStack.capacity()); 4406 } 4407 } 4408 _markStack._hit_limit = 0; 4409 _markStack._failed_double = 0; 4410 4411 if ((VerifyAfterGC || VerifyDuringGC) && 4412 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4413 verify_after_remark(); 4414 } 4415 4416 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4417 4418 // Change under the freelistLocks. 4419 _collectorState = Sweeping; 4420 // Call isAllClear() under bitMapLock 4421 assert(_modUnionTable.isAllClear(), 4422 "Should be clear by end of the final marking"); 4423 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4424 "Should be clear by end of the final marking"); 4425 } 4426 4427 void CMSParInitialMarkTask::work(uint worker_id) { 4428 elapsedTimer _timer; 4429 ResourceMark rm; 4430 HandleMark hm; 4431 4432 // ---------- scan from roots -------------- 4433 _timer.start(); 4434 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4435 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4436 4437 // ---------- young gen roots -------------- 4438 { 4439 work_on_young_gen_roots(worker_id, &par_mri_cl); 4440 _timer.stop(); 4441 if (PrintCMSStatistics != 0) { 4442 gclog_or_tty->print_cr( 4443 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4444 worker_id, _timer.seconds()); 4445 } 4446 } 4447 4448 // ---------- remaining roots -------------- 4449 _timer.reset(); 4450 _timer.start(); 4451 4452 CLDToOopClosure cld_closure(&par_mri_cl, true); 4453 4454 gch->gen_process_roots(_collector->_cmsGen->level(), 4455 false, // yg was scanned above 4456 false, // this is parallel code 4457 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4458 _collector->should_unload_classes(), 4459 &par_mri_cl, 4460 NULL, 4461 &cld_closure); 4462 assert(_collector->should_unload_classes() 4463 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4464 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4465 _timer.stop(); 4466 if (PrintCMSStatistics != 0) { 4467 gclog_or_tty->print_cr( 4468 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4469 worker_id, _timer.seconds()); 4470 } 4471 } 4472 4473 // Parallel remark task 4474 class CMSParRemarkTask: public CMSParMarkTask { 4475 CompactibleFreeListSpace* _cms_space; 4476 4477 // The per-thread work queues, available here for stealing. 4478 OopTaskQueueSet* _task_queues; 4479 ParallelTaskTerminator _term; 4480 4481 public: 4482 // A value of 0 passed to n_workers will cause the number of 4483 // workers to be taken from the active workers in the work gang. 4484 CMSParRemarkTask(CMSCollector* collector, 4485 CompactibleFreeListSpace* cms_space, 4486 uint n_workers, FlexibleWorkGang* workers, 4487 OopTaskQueueSet* task_queues): 4488 CMSParMarkTask("Rescan roots and grey objects in parallel", 4489 collector, n_workers), 4490 _cms_space(cms_space), 4491 _task_queues(task_queues), 4492 _term(n_workers, task_queues) { } 4493 4494 OopTaskQueueSet* task_queues() { return _task_queues; } 4495 4496 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4497 4498 ParallelTaskTerminator* terminator() { return &_term; } 4499 uint n_workers() { return _n_workers; } 4500 4501 void work(uint worker_id); 4502 4503 private: 4504 // ... of dirty cards in old space 4505 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4506 Par_MarkRefsIntoAndScanClosure* cl); 4507 4508 // ... work stealing for the above 4509 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4510 }; 4511 4512 class RemarkKlassClosure : public KlassClosure { 4513 KlassToOopClosure _cm_klass_closure; 4514 public: 4515 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4516 void do_klass(Klass* k) { 4517 // Check if we have modified any oops in the Klass during the concurrent marking. 4518 if (k->has_accumulated_modified_oops()) { 4519 k->clear_accumulated_modified_oops(); 4520 4521 // We could have transfered the current modified marks to the accumulated marks, 4522 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4523 } else if (k->has_modified_oops()) { 4524 // Don't clear anything, this info is needed by the next young collection. 4525 } else { 4526 // No modified oops in the Klass. 4527 return; 4528 } 4529 4530 // The klass has modified fields, need to scan the klass. 4531 _cm_klass_closure.do_klass(k); 4532 } 4533 }; 4534 4535 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4536 ParNewGeneration* young_gen = _collector->_young_gen; 4537 ContiguousSpace* eden_space = young_gen->eden(); 4538 ContiguousSpace* from_space = young_gen->from(); 4539 ContiguousSpace* to_space = young_gen->to(); 4540 4541 HeapWord** eca = _collector->_eden_chunk_array; 4542 size_t ect = _collector->_eden_chunk_index; 4543 HeapWord** sca = _collector->_survivor_chunk_array; 4544 size_t sct = _collector->_survivor_chunk_index; 4545 4546 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4547 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4548 4549 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4550 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4551 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4552 } 4553 4554 // work_queue(i) is passed to the closure 4555 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4556 // also is passed to do_dirty_card_rescan_tasks() and to 4557 // do_work_steal() to select the i-th task_queue. 4558 4559 void CMSParRemarkTask::work(uint worker_id) { 4560 elapsedTimer _timer; 4561 ResourceMark rm; 4562 HandleMark hm; 4563 4564 // ---------- rescan from roots -------------- 4565 _timer.start(); 4566 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4567 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4568 _collector->_span, _collector->ref_processor(), 4569 &(_collector->_markBitMap), 4570 work_queue(worker_id)); 4571 4572 // Rescan young gen roots first since these are likely 4573 // coarsely partitioned and may, on that account, constitute 4574 // the critical path; thus, it's best to start off that 4575 // work first. 4576 // ---------- young gen roots -------------- 4577 { 4578 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4579 _timer.stop(); 4580 if (PrintCMSStatistics != 0) { 4581 gclog_or_tty->print_cr( 4582 "Finished young gen rescan work in %dth thread: %3.3f sec", 4583 worker_id, _timer.seconds()); 4584 } 4585 } 4586 4587 // ---------- remaining roots -------------- 4588 _timer.reset(); 4589 _timer.start(); 4590 gch->gen_process_roots(_collector->_cmsGen->level(), 4591 false, // yg was scanned above 4592 false, // this is parallel code 4593 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4594 _collector->should_unload_classes(), 4595 &par_mrias_cl, 4596 NULL, 4597 NULL); // The dirty klasses will be handled below 4598 4599 assert(_collector->should_unload_classes() 4600 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4601 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4602 _timer.stop(); 4603 if (PrintCMSStatistics != 0) { 4604 gclog_or_tty->print_cr( 4605 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4606 worker_id, _timer.seconds()); 4607 } 4608 4609 // ---------- unhandled CLD scanning ---------- 4610 if (worker_id == 0) { // Single threaded at the moment. 4611 _timer.reset(); 4612 _timer.start(); 4613 4614 // Scan all new class loader data objects and new dependencies that were 4615 // introduced during concurrent marking. 4616 ResourceMark rm; 4617 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4618 for (int i = 0; i < array->length(); i++) { 4619 par_mrias_cl.do_class_loader_data(array->at(i)); 4620 } 4621 4622 // We don't need to keep track of new CLDs anymore. 4623 ClassLoaderDataGraph::remember_new_clds(false); 4624 4625 _timer.stop(); 4626 if (PrintCMSStatistics != 0) { 4627 gclog_or_tty->print_cr( 4628 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4629 worker_id, _timer.seconds()); 4630 } 4631 } 4632 4633 // ---------- dirty klass scanning ---------- 4634 if (worker_id == 0) { // Single threaded at the moment. 4635 _timer.reset(); 4636 _timer.start(); 4637 4638 // Scan all classes that was dirtied during the concurrent marking phase. 4639 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4640 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4641 4642 _timer.stop(); 4643 if (PrintCMSStatistics != 0) { 4644 gclog_or_tty->print_cr( 4645 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4646 worker_id, _timer.seconds()); 4647 } 4648 } 4649 4650 // We might have added oops to ClassLoaderData::_handles during the 4651 // concurrent marking phase. These oops point to newly allocated objects 4652 // that are guaranteed to be kept alive. Either by the direct allocation 4653 // code, or when the young collector processes the roots. Hence, 4654 // we don't have to revisit the _handles block during the remark phase. 4655 4656 // ---------- rescan dirty cards ------------ 4657 _timer.reset(); 4658 _timer.start(); 4659 4660 // Do the rescan tasks for each of the two spaces 4661 // (cms_space) in turn. 4662 // "worker_id" is passed to select the task_queue for "worker_id" 4663 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4664 _timer.stop(); 4665 if (PrintCMSStatistics != 0) { 4666 gclog_or_tty->print_cr( 4667 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4668 worker_id, _timer.seconds()); 4669 } 4670 4671 // ---------- steal work from other threads ... 4672 // ---------- ... and drain overflow list. 4673 _timer.reset(); 4674 _timer.start(); 4675 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4676 _timer.stop(); 4677 if (PrintCMSStatistics != 0) { 4678 gclog_or_tty->print_cr( 4679 "Finished work stealing in %dth thread: %3.3f sec", 4680 worker_id, _timer.seconds()); 4681 } 4682 } 4683 4684 // Note that parameter "i" is not used. 4685 void 4686 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4687 OopsInGenClosure* cl, ContiguousSpace* space, 4688 HeapWord** chunk_array, size_t chunk_top) { 4689 // Until all tasks completed: 4690 // . claim an unclaimed task 4691 // . compute region boundaries corresponding to task claimed 4692 // using chunk_array 4693 // . par_oop_iterate(cl) over that region 4694 4695 ResourceMark rm; 4696 HandleMark hm; 4697 4698 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4699 4700 uint nth_task = 0; 4701 uint n_tasks = pst->n_tasks(); 4702 4703 if (n_tasks > 0) { 4704 assert(pst->valid(), "Uninitialized use?"); 4705 HeapWord *start, *end; 4706 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4707 // We claimed task # nth_task; compute its boundaries. 4708 if (chunk_top == 0) { // no samples were taken 4709 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4710 start = space->bottom(); 4711 end = space->top(); 4712 } else if (nth_task == 0) { 4713 start = space->bottom(); 4714 end = chunk_array[nth_task]; 4715 } else if (nth_task < (uint)chunk_top) { 4716 assert(nth_task >= 1, "Control point invariant"); 4717 start = chunk_array[nth_task - 1]; 4718 end = chunk_array[nth_task]; 4719 } else { 4720 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4721 start = chunk_array[chunk_top - 1]; 4722 end = space->top(); 4723 } 4724 MemRegion mr(start, end); 4725 // Verify that mr is in space 4726 assert(mr.is_empty() || space->used_region().contains(mr), 4727 "Should be in space"); 4728 // Verify that "start" is an object boundary 4729 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4730 "Should be an oop"); 4731 space->par_oop_iterate(mr, cl); 4732 } 4733 pst->all_tasks_completed(); 4734 } 4735 } 4736 4737 void 4738 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4739 CompactibleFreeListSpace* sp, int i, 4740 Par_MarkRefsIntoAndScanClosure* cl) { 4741 // Until all tasks completed: 4742 // . claim an unclaimed task 4743 // . compute region boundaries corresponding to task claimed 4744 // . transfer dirty bits ct->mut for that region 4745 // . apply rescanclosure to dirty mut bits for that region 4746 4747 ResourceMark rm; 4748 HandleMark hm; 4749 4750 OopTaskQueue* work_q = work_queue(i); 4751 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4752 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4753 // CAUTION: This closure has state that persists across calls to 4754 // the work method dirty_range_iterate_clear() in that it has 4755 // embedded in it a (subtype of) UpwardsObjectClosure. The 4756 // use of that state in the embedded UpwardsObjectClosure instance 4757 // assumes that the cards are always iterated (even if in parallel 4758 // by several threads) in monotonically increasing order per each 4759 // thread. This is true of the implementation below which picks 4760 // card ranges (chunks) in monotonically increasing order globally 4761 // and, a-fortiori, in monotonically increasing order per thread 4762 // (the latter order being a subsequence of the former). 4763 // If the work code below is ever reorganized into a more chaotic 4764 // work-partitioning form than the current "sequential tasks" 4765 // paradigm, the use of that persistent state will have to be 4766 // revisited and modified appropriately. See also related 4767 // bug 4756801 work on which should examine this code to make 4768 // sure that the changes there do not run counter to the 4769 // assumptions made here and necessary for correctness and 4770 // efficiency. Note also that this code might yield inefficient 4771 // behavior in the case of very large objects that span one or 4772 // more work chunks. Such objects would potentially be scanned 4773 // several times redundantly. Work on 4756801 should try and 4774 // address that performance anomaly if at all possible. XXX 4775 MemRegion full_span = _collector->_span; 4776 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4777 MarkFromDirtyCardsClosure 4778 greyRescanClosure(_collector, full_span, // entire span of interest 4779 sp, bm, work_q, cl); 4780 4781 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4782 assert(pst->valid(), "Uninitialized use?"); 4783 uint nth_task = 0; 4784 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4785 MemRegion span = sp->used_region(); 4786 HeapWord* start_addr = span.start(); 4787 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4788 alignment); 4789 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4790 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4791 start_addr, "Check alignment"); 4792 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4793 chunk_size, "Check alignment"); 4794 4795 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4796 // Having claimed the nth_task, compute corresponding mem-region, 4797 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4798 // The alignment restriction ensures that we do not need any 4799 // synchronization with other gang-workers while setting or 4800 // clearing bits in thus chunk of the MUT. 4801 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4802 start_addr + (nth_task+1)*chunk_size); 4803 // The last chunk's end might be way beyond end of the 4804 // used region. In that case pull back appropriately. 4805 if (this_span.end() > end_addr) { 4806 this_span.set_end(end_addr); 4807 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4808 } 4809 // Iterate over the dirty cards covering this chunk, marking them 4810 // precleaned, and setting the corresponding bits in the mod union 4811 // table. Since we have been careful to partition at Card and MUT-word 4812 // boundaries no synchronization is needed between parallel threads. 4813 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4814 &modUnionClosure); 4815 4816 // Having transferred these marks into the modUnionTable, 4817 // rescan the marked objects on the dirty cards in the modUnionTable. 4818 // Even if this is at a synchronous collection, the initial marking 4819 // may have been done during an asynchronous collection so there 4820 // may be dirty bits in the mod-union table. 4821 _collector->_modUnionTable.dirty_range_iterate_clear( 4822 this_span, &greyRescanClosure); 4823 _collector->_modUnionTable.verifyNoOneBitsInRange( 4824 this_span.start(), 4825 this_span.end()); 4826 } 4827 pst->all_tasks_completed(); // declare that i am done 4828 } 4829 4830 // . see if we can share work_queues with ParNew? XXX 4831 void 4832 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4833 int* seed) { 4834 OopTaskQueue* work_q = work_queue(i); 4835 NOT_PRODUCT(int num_steals = 0;) 4836 oop obj_to_scan; 4837 CMSBitMap* bm = &(_collector->_markBitMap); 4838 4839 while (true) { 4840 // Completely finish any left over work from (an) earlier round(s) 4841 cl->trim_queue(0); 4842 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4843 (size_t)ParGCDesiredObjsFromOverflowList); 4844 // Now check if there's any work in the overflow list 4845 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4846 // only affects the number of attempts made to get work from the 4847 // overflow list and does not affect the number of workers. Just 4848 // pass ParallelGCThreads so this behavior is unchanged. 4849 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4850 work_q, 4851 ParallelGCThreads)) { 4852 // found something in global overflow list; 4853 // not yet ready to go stealing work from others. 4854 // We'd like to assert(work_q->size() != 0, ...) 4855 // because we just took work from the overflow list, 4856 // but of course we can't since all of that could have 4857 // been already stolen from us. 4858 // "He giveth and He taketh away." 4859 continue; 4860 } 4861 // Verify that we have no work before we resort to stealing 4862 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4863 // Try to steal from other queues that have work 4864 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4865 NOT_PRODUCT(num_steals++;) 4866 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4867 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4868 // Do scanning work 4869 obj_to_scan->oop_iterate(cl); 4870 // Loop around, finish this work, and try to steal some more 4871 } else if (terminator()->offer_termination()) { 4872 break; // nirvana from the infinite cycle 4873 } 4874 } 4875 NOT_PRODUCT( 4876 if (PrintCMSStatistics != 0) { 4877 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4878 } 4879 ) 4880 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4881 "Else our work is not yet done"); 4882 } 4883 4884 // Record object boundaries in _eden_chunk_array by sampling the eden 4885 // top in the slow-path eden object allocation code path and record 4886 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4887 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4888 // sampling in sample_eden() that activates during the part of the 4889 // preclean phase. 4890 void CMSCollector::sample_eden_chunk() { 4891 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4892 if (_eden_chunk_lock->try_lock()) { 4893 // Record a sample. This is the critical section. The contents 4894 // of the _eden_chunk_array have to be non-decreasing in the 4895 // address order. 4896 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4897 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4898 "Unexpected state of Eden"); 4899 if (_eden_chunk_index == 0 || 4900 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4901 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4902 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4903 _eden_chunk_index++; // commit sample 4904 } 4905 _eden_chunk_lock->unlock(); 4906 } 4907 } 4908 } 4909 4910 // Return a thread-local PLAB recording array, as appropriate. 4911 void* CMSCollector::get_data_recorder(int thr_num) { 4912 if (_survivor_plab_array != NULL && 4913 (CMSPLABRecordAlways || 4914 (_collectorState > Marking && _collectorState < FinalMarking))) { 4915 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4916 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4917 ca->reset(); // clear it so that fresh data is recorded 4918 return (void*) ca; 4919 } else { 4920 return NULL; 4921 } 4922 } 4923 4924 // Reset all the thread-local PLAB recording arrays 4925 void CMSCollector::reset_survivor_plab_arrays() { 4926 for (uint i = 0; i < ParallelGCThreads; i++) { 4927 _survivor_plab_array[i].reset(); 4928 } 4929 } 4930 4931 // Merge the per-thread plab arrays into the global survivor chunk 4932 // array which will provide the partitioning of the survivor space 4933 // for CMS initial scan and rescan. 4934 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4935 int no_of_gc_threads) { 4936 assert(_survivor_plab_array != NULL, "Error"); 4937 assert(_survivor_chunk_array != NULL, "Error"); 4938 assert(_collectorState == FinalMarking || 4939 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4940 for (int j = 0; j < no_of_gc_threads; j++) { 4941 _cursor[j] = 0; 4942 } 4943 HeapWord* top = surv->top(); 4944 size_t i; 4945 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4946 HeapWord* min_val = top; // Higher than any PLAB address 4947 uint min_tid = 0; // position of min_val this round 4948 for (int j = 0; j < no_of_gc_threads; j++) { 4949 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4950 if (_cursor[j] == cur_sca->end()) { 4951 continue; 4952 } 4953 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4954 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4955 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4956 if (cur_val < min_val) { 4957 min_tid = j; 4958 min_val = cur_val; 4959 } else { 4960 assert(cur_val < top, "All recorded addresses should be less"); 4961 } 4962 } 4963 // At this point min_val and min_tid are respectively 4964 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4965 // and the thread (j) that witnesses that address. 4966 // We record this address in the _survivor_chunk_array[i] 4967 // and increment _cursor[min_tid] prior to the next round i. 4968 if (min_val == top) { 4969 break; 4970 } 4971 _survivor_chunk_array[i] = min_val; 4972 _cursor[min_tid]++; 4973 } 4974 // We are all done; record the size of the _survivor_chunk_array 4975 _survivor_chunk_index = i; // exclusive: [0, i) 4976 if (PrintCMSStatistics > 0) { 4977 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4978 } 4979 // Verify that we used up all the recorded entries 4980 #ifdef ASSERT 4981 size_t total = 0; 4982 for (int j = 0; j < no_of_gc_threads; j++) { 4983 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4984 total += _cursor[j]; 4985 } 4986 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4987 // Check that the merged array is in sorted order 4988 if (total > 0) { 4989 for (size_t i = 0; i < total - 1; i++) { 4990 if (PrintCMSStatistics > 0) { 4991 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4992 i, p2i(_survivor_chunk_array[i])); 4993 } 4994 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4995 "Not sorted"); 4996 } 4997 } 4998 #endif // ASSERT 4999 } 5000 5001 // Set up the space's par_seq_tasks structure for work claiming 5002 // for parallel initial scan and rescan of young gen. 5003 // See ParRescanTask where this is currently used. 5004 void 5005 CMSCollector:: 5006 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5007 assert(n_threads > 0, "Unexpected n_threads argument"); 5008 5009 // Eden space 5010 if (!_young_gen->eden()->is_empty()) { 5011 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5012 assert(!pst->valid(), "Clobbering existing data?"); 5013 // Each valid entry in [0, _eden_chunk_index) represents a task. 5014 size_t n_tasks = _eden_chunk_index + 1; 5015 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5016 // Sets the condition for completion of the subtask (how many threads 5017 // need to finish in order to be done). 5018 pst->set_n_threads(n_threads); 5019 pst->set_n_tasks((int)n_tasks); 5020 } 5021 5022 // Merge the survivor plab arrays into _survivor_chunk_array 5023 if (_survivor_plab_array != NULL) { 5024 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5025 } else { 5026 assert(_survivor_chunk_index == 0, "Error"); 5027 } 5028 5029 // To space 5030 { 5031 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5032 assert(!pst->valid(), "Clobbering existing data?"); 5033 // Sets the condition for completion of the subtask (how many threads 5034 // need to finish in order to be done). 5035 pst->set_n_threads(n_threads); 5036 pst->set_n_tasks(1); 5037 assert(pst->valid(), "Error"); 5038 } 5039 5040 // From space 5041 { 5042 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5043 assert(!pst->valid(), "Clobbering existing data?"); 5044 size_t n_tasks = _survivor_chunk_index + 1; 5045 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5046 // Sets the condition for completion of the subtask (how many threads 5047 // need to finish in order to be done). 5048 pst->set_n_threads(n_threads); 5049 pst->set_n_tasks((int)n_tasks); 5050 assert(pst->valid(), "Error"); 5051 } 5052 } 5053 5054 // Parallel version of remark 5055 void CMSCollector::do_remark_parallel() { 5056 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5057 FlexibleWorkGang* workers = gch->workers(); 5058 assert(workers != NULL, "Need parallel worker threads."); 5059 // Choose to use the number of GC workers most recently set 5060 // into "active_workers". If active_workers is not set, set it 5061 // to ParallelGCThreads. 5062 uint n_workers = workers->active_workers(); 5063 if (n_workers == 0) { 5064 assert(n_workers > 0, "Should have been set during scavenge"); 5065 n_workers = ParallelGCThreads; 5066 workers->set_active_workers(n_workers); 5067 } 5068 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5069 5070 CMSParRemarkTask tsk(this, 5071 cms_space, 5072 n_workers, workers, task_queues()); 5073 5074 // Set up for parallel process_roots work. 5075 gch->set_par_threads(n_workers); 5076 // We won't be iterating over the cards in the card table updating 5077 // the younger_gen cards, so we shouldn't call the following else 5078 // the verification code as well as subsequent younger_refs_iterate 5079 // code would get confused. XXX 5080 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5081 5082 // The young gen rescan work will not be done as part of 5083 // process_roots (which currently doesn't know how to 5084 // parallelize such a scan), but rather will be broken up into 5085 // a set of parallel tasks (via the sampling that the [abortable] 5086 // preclean phase did of eden, plus the [two] tasks of 5087 // scanning the [two] survivor spaces. Further fine-grain 5088 // parallelization of the scanning of the survivor spaces 5089 // themselves, and of precleaning of the younger gen itself 5090 // is deferred to the future. 5091 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5092 5093 // The dirty card rescan work is broken up into a "sequence" 5094 // of parallel tasks (per constituent space) that are dynamically 5095 // claimed by the parallel threads. 5096 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5097 5098 // It turns out that even when we're using 1 thread, doing the work in a 5099 // separate thread causes wide variance in run times. We can't help this 5100 // in the multi-threaded case, but we special-case n=1 here to get 5101 // repeatable measurements of the 1-thread overhead of the parallel code. 5102 if (n_workers > 1) { 5103 // Make refs discovery MT-safe, if it isn't already: it may not 5104 // necessarily be so, since it's possible that we are doing 5105 // ST marking. 5106 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5107 StrongRootsScope srs; 5108 workers->run_task(&tsk); 5109 } else { 5110 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5111 StrongRootsScope srs; 5112 tsk.work(0); 5113 } 5114 5115 gch->set_par_threads(0); // 0 ==> non-parallel. 5116 // restore, single-threaded for now, any preserved marks 5117 // as a result of work_q overflow 5118 restore_preserved_marks_if_any(); 5119 } 5120 5121 // Non-parallel version of remark 5122 void CMSCollector::do_remark_non_parallel() { 5123 ResourceMark rm; 5124 HandleMark hm; 5125 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5126 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5127 5128 MarkRefsIntoAndScanClosure 5129 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5130 &_markStack, this, 5131 false /* should_yield */, false /* not precleaning */); 5132 MarkFromDirtyCardsClosure 5133 markFromDirtyCardsClosure(this, _span, 5134 NULL, // space is set further below 5135 &_markBitMap, &_markStack, &mrias_cl); 5136 { 5137 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5138 // Iterate over the dirty cards, setting the corresponding bits in the 5139 // mod union table. 5140 { 5141 ModUnionClosure modUnionClosure(&_modUnionTable); 5142 _ct->ct_bs()->dirty_card_iterate( 5143 _cmsGen->used_region(), 5144 &modUnionClosure); 5145 } 5146 // Having transferred these marks into the modUnionTable, we just need 5147 // to rescan the marked objects on the dirty cards in the modUnionTable. 5148 // The initial marking may have been done during an asynchronous 5149 // collection so there may be dirty bits in the mod-union table. 5150 const int alignment = 5151 CardTableModRefBS::card_size * BitsPerWord; 5152 { 5153 // ... First handle dirty cards in CMS gen 5154 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5155 MemRegion ur = _cmsGen->used_region(); 5156 HeapWord* lb = ur.start(); 5157 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5158 MemRegion cms_span(lb, ub); 5159 _modUnionTable.dirty_range_iterate_clear(cms_span, 5160 &markFromDirtyCardsClosure); 5161 verify_work_stacks_empty(); 5162 if (PrintCMSStatistics != 0) { 5163 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5164 markFromDirtyCardsClosure.num_dirty_cards()); 5165 } 5166 } 5167 } 5168 if (VerifyDuringGC && 5169 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5170 HandleMark hm; // Discard invalid handles created during verification 5171 Universe::verify(); 5172 } 5173 { 5174 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5175 5176 verify_work_stacks_empty(); 5177 5178 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5179 StrongRootsScope srs; 5180 5181 gch->gen_process_roots(_cmsGen->level(), 5182 true, // younger gens as roots 5183 false, // use the local StrongRootsScope 5184 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5185 should_unload_classes(), 5186 &mrias_cl, 5187 NULL, 5188 NULL); // The dirty klasses will be handled below 5189 5190 assert(should_unload_classes() 5191 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5192 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5193 } 5194 5195 { 5196 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5197 5198 verify_work_stacks_empty(); 5199 5200 // Scan all class loader data objects that might have been introduced 5201 // during concurrent marking. 5202 ResourceMark rm; 5203 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5204 for (int i = 0; i < array->length(); i++) { 5205 mrias_cl.do_class_loader_data(array->at(i)); 5206 } 5207 5208 // We don't need to keep track of new CLDs anymore. 5209 ClassLoaderDataGraph::remember_new_clds(false); 5210 5211 verify_work_stacks_empty(); 5212 } 5213 5214 { 5215 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5216 5217 verify_work_stacks_empty(); 5218 5219 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5220 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5221 5222 verify_work_stacks_empty(); 5223 } 5224 5225 // We might have added oops to ClassLoaderData::_handles during the 5226 // concurrent marking phase. These oops point to newly allocated objects 5227 // that are guaranteed to be kept alive. Either by the direct allocation 5228 // code, or when the young collector processes the roots. Hence, 5229 // we don't have to revisit the _handles block during the remark phase. 5230 5231 verify_work_stacks_empty(); 5232 // Restore evacuated mark words, if any, used for overflow list links 5233 if (!CMSOverflowEarlyRestoration) { 5234 restore_preserved_marks_if_any(); 5235 } 5236 verify_overflow_empty(); 5237 } 5238 5239 //////////////////////////////////////////////////////// 5240 // Parallel Reference Processing Task Proxy Class 5241 //////////////////////////////////////////////////////// 5242 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5243 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5244 CMSCollector* _collector; 5245 CMSBitMap* _mark_bit_map; 5246 const MemRegion _span; 5247 ProcessTask& _task; 5248 5249 public: 5250 CMSRefProcTaskProxy(ProcessTask& task, 5251 CMSCollector* collector, 5252 const MemRegion& span, 5253 CMSBitMap* mark_bit_map, 5254 AbstractWorkGang* workers, 5255 OopTaskQueueSet* task_queues): 5256 // XXX Should superclass AGTWOQ also know about AWG since it knows 5257 // about the task_queues used by the AWG? Then it could initialize 5258 // the terminator() object. See 6984287. The set_for_termination() 5259 // below is a temporary band-aid for the regression in 6984287. 5260 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5261 task_queues), 5262 _task(task), 5263 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5264 { 5265 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5266 "Inconsistency in _span"); 5267 set_for_termination(workers->active_workers()); 5268 } 5269 5270 OopTaskQueueSet* task_queues() { return queues(); } 5271 5272 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5273 5274 void do_work_steal(int i, 5275 CMSParDrainMarkingStackClosure* drain, 5276 CMSParKeepAliveClosure* keep_alive, 5277 int* seed); 5278 5279 virtual void work(uint worker_id); 5280 }; 5281 5282 void CMSRefProcTaskProxy::work(uint worker_id) { 5283 ResourceMark rm; 5284 HandleMark hm; 5285 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5286 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5287 _mark_bit_map, 5288 work_queue(worker_id)); 5289 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5290 _mark_bit_map, 5291 work_queue(worker_id)); 5292 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5293 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5294 if (_task.marks_oops_alive()) { 5295 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5296 _collector->hash_seed(worker_id)); 5297 } 5298 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5299 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5300 } 5301 5302 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5303 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5304 EnqueueTask& _task; 5305 5306 public: 5307 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5308 : AbstractGangTask("Enqueue reference objects in parallel"), 5309 _task(task) 5310 { } 5311 5312 virtual void work(uint worker_id) 5313 { 5314 _task.work(worker_id); 5315 } 5316 }; 5317 5318 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5319 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5320 _span(span), 5321 _bit_map(bit_map), 5322 _work_queue(work_queue), 5323 _mark_and_push(collector, span, bit_map, work_queue), 5324 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5325 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5326 { } 5327 5328 // . see if we can share work_queues with ParNew? XXX 5329 void CMSRefProcTaskProxy::do_work_steal(int i, 5330 CMSParDrainMarkingStackClosure* drain, 5331 CMSParKeepAliveClosure* keep_alive, 5332 int* seed) { 5333 OopTaskQueue* work_q = work_queue(i); 5334 NOT_PRODUCT(int num_steals = 0;) 5335 oop obj_to_scan; 5336 5337 while (true) { 5338 // Completely finish any left over work from (an) earlier round(s) 5339 drain->trim_queue(0); 5340 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5341 (size_t)ParGCDesiredObjsFromOverflowList); 5342 // Now check if there's any work in the overflow list 5343 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5344 // only affects the number of attempts made to get work from the 5345 // overflow list and does not affect the number of workers. Just 5346 // pass ParallelGCThreads so this behavior is unchanged. 5347 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5348 work_q, 5349 ParallelGCThreads)) { 5350 // Found something in global overflow list; 5351 // not yet ready to go stealing work from others. 5352 // We'd like to assert(work_q->size() != 0, ...) 5353 // because we just took work from the overflow list, 5354 // but of course we can't, since all of that might have 5355 // been already stolen from us. 5356 continue; 5357 } 5358 // Verify that we have no work before we resort to stealing 5359 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5360 // Try to steal from other queues that have work 5361 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5362 NOT_PRODUCT(num_steals++;) 5363 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5364 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5365 // Do scanning work 5366 obj_to_scan->oop_iterate(keep_alive); 5367 // Loop around, finish this work, and try to steal some more 5368 } else if (terminator()->offer_termination()) { 5369 break; // nirvana from the infinite cycle 5370 } 5371 } 5372 NOT_PRODUCT( 5373 if (PrintCMSStatistics != 0) { 5374 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5375 } 5376 ) 5377 } 5378 5379 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5380 { 5381 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5382 FlexibleWorkGang* workers = gch->workers(); 5383 assert(workers != NULL, "Need parallel worker threads."); 5384 CMSRefProcTaskProxy rp_task(task, &_collector, 5385 _collector.ref_processor()->span(), 5386 _collector.markBitMap(), 5387 workers, _collector.task_queues()); 5388 workers->run_task(&rp_task); 5389 } 5390 5391 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5392 { 5393 5394 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5395 FlexibleWorkGang* workers = gch->workers(); 5396 assert(workers != NULL, "Need parallel worker threads."); 5397 CMSRefEnqueueTaskProxy enq_task(task); 5398 workers->run_task(&enq_task); 5399 } 5400 5401 void CMSCollector::refProcessingWork() { 5402 ResourceMark rm; 5403 HandleMark hm; 5404 5405 ReferenceProcessor* rp = ref_processor(); 5406 assert(rp->span().equals(_span), "Spans should be equal"); 5407 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5408 // Process weak references. 5409 rp->setup_policy(false); 5410 verify_work_stacks_empty(); 5411 5412 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5413 &_markStack, false /* !preclean */); 5414 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5415 _span, &_markBitMap, &_markStack, 5416 &cmsKeepAliveClosure, false /* !preclean */); 5417 { 5418 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5419 5420 ReferenceProcessorStats stats; 5421 if (rp->processing_is_mt()) { 5422 // Set the degree of MT here. If the discovery is done MT, there 5423 // may have been a different number of threads doing the discovery 5424 // and a different number of discovered lists may have Ref objects. 5425 // That is OK as long as the Reference lists are balanced (see 5426 // balance_all_queues() and balance_queues()). 5427 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5428 uint active_workers = ParallelGCThreads; 5429 FlexibleWorkGang* workers = gch->workers(); 5430 if (workers != NULL) { 5431 active_workers = workers->active_workers(); 5432 // The expectation is that active_workers will have already 5433 // been set to a reasonable value. If it has not been set, 5434 // investigate. 5435 assert(active_workers > 0, "Should have been set during scavenge"); 5436 } 5437 rp->set_active_mt_degree(active_workers); 5438 CMSRefProcTaskExecutor task_executor(*this); 5439 stats = rp->process_discovered_references(&_is_alive_closure, 5440 &cmsKeepAliveClosure, 5441 &cmsDrainMarkingStackClosure, 5442 &task_executor, 5443 _gc_timer_cm, 5444 _gc_tracer_cm->gc_id()); 5445 } else { 5446 stats = rp->process_discovered_references(&_is_alive_closure, 5447 &cmsKeepAliveClosure, 5448 &cmsDrainMarkingStackClosure, 5449 NULL, 5450 _gc_timer_cm, 5451 _gc_tracer_cm->gc_id()); 5452 } 5453 _gc_tracer_cm->report_gc_reference_stats(stats); 5454 5455 } 5456 5457 // This is the point where the entire marking should have completed. 5458 verify_work_stacks_empty(); 5459 5460 if (should_unload_classes()) { 5461 { 5462 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5463 5464 // Unload classes and purge the SystemDictionary. 5465 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5466 5467 // Unload nmethods. 5468 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5469 5470 // Prune dead klasses from subklass/sibling/implementor lists. 5471 Klass::clean_weak_klass_links(&_is_alive_closure); 5472 } 5473 5474 { 5475 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5476 // Clean up unreferenced symbols in symbol table. 5477 SymbolTable::unlink(); 5478 } 5479 5480 { 5481 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5482 // Delete entries for dead interned strings. 5483 StringTable::unlink(&_is_alive_closure); 5484 } 5485 } 5486 5487 5488 // Restore any preserved marks as a result of mark stack or 5489 // work queue overflow 5490 restore_preserved_marks_if_any(); // done single-threaded for now 5491 5492 rp->set_enqueuing_is_done(true); 5493 if (rp->processing_is_mt()) { 5494 rp->balance_all_queues(); 5495 CMSRefProcTaskExecutor task_executor(*this); 5496 rp->enqueue_discovered_references(&task_executor); 5497 } else { 5498 rp->enqueue_discovered_references(NULL); 5499 } 5500 rp->verify_no_references_recorded(); 5501 assert(!rp->discovery_enabled(), "should have been disabled"); 5502 } 5503 5504 #ifndef PRODUCT 5505 void CMSCollector::check_correct_thread_executing() { 5506 Thread* t = Thread::current(); 5507 // Only the VM thread or the CMS thread should be here. 5508 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5509 "Unexpected thread type"); 5510 // If this is the vm thread, the foreground process 5511 // should not be waiting. Note that _foregroundGCIsActive is 5512 // true while the foreground collector is waiting. 5513 if (_foregroundGCShouldWait) { 5514 // We cannot be the VM thread 5515 assert(t->is_ConcurrentGC_thread(), 5516 "Should be CMS thread"); 5517 } else { 5518 // We can be the CMS thread only if we are in a stop-world 5519 // phase of CMS collection. 5520 if (t->is_ConcurrentGC_thread()) { 5521 assert(_collectorState == InitialMarking || 5522 _collectorState == FinalMarking, 5523 "Should be a stop-world phase"); 5524 // The CMS thread should be holding the CMS_token. 5525 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5526 "Potential interference with concurrently " 5527 "executing VM thread"); 5528 } 5529 } 5530 } 5531 #endif 5532 5533 void CMSCollector::sweep() { 5534 assert(_collectorState == Sweeping, "just checking"); 5535 check_correct_thread_executing(); 5536 verify_work_stacks_empty(); 5537 verify_overflow_empty(); 5538 increment_sweep_count(); 5539 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5540 5541 _inter_sweep_timer.stop(); 5542 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5543 5544 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5545 _intra_sweep_timer.reset(); 5546 _intra_sweep_timer.start(); 5547 { 5548 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5549 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5550 // First sweep the old gen 5551 { 5552 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5553 bitMapLock()); 5554 sweepWork(_cmsGen); 5555 } 5556 5557 // Update Universe::_heap_*_at_gc figures. 5558 // We need all the free list locks to make the abstract state 5559 // transition from Sweeping to Resetting. See detailed note 5560 // further below. 5561 { 5562 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5563 // Update heap occupancy information which is used as 5564 // input to soft ref clearing policy at the next gc. 5565 Universe::update_heap_info_at_gc(); 5566 _collectorState = Resizing; 5567 } 5568 } 5569 verify_work_stacks_empty(); 5570 verify_overflow_empty(); 5571 5572 if (should_unload_classes()) { 5573 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5574 // requires that the virtual spaces are stable and not deleted. 5575 ClassLoaderDataGraph::set_should_purge(true); 5576 } 5577 5578 _intra_sweep_timer.stop(); 5579 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5580 5581 _inter_sweep_timer.reset(); 5582 _inter_sweep_timer.start(); 5583 5584 // We need to use a monotonically non-decreasing time in ms 5585 // or we will see time-warp warnings and os::javaTimeMillis() 5586 // does not guarantee monotonicity. 5587 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5588 update_time_of_last_gc(now); 5589 5590 // NOTE on abstract state transitions: 5591 // Mutators allocate-live and/or mark the mod-union table dirty 5592 // based on the state of the collection. The former is done in 5593 // the interval [Marking, Sweeping] and the latter in the interval 5594 // [Marking, Sweeping). Thus the transitions into the Marking state 5595 // and out of the Sweeping state must be synchronously visible 5596 // globally to the mutators. 5597 // The transition into the Marking state happens with the world 5598 // stopped so the mutators will globally see it. Sweeping is 5599 // done asynchronously by the background collector so the transition 5600 // from the Sweeping state to the Resizing state must be done 5601 // under the freelistLock (as is the check for whether to 5602 // allocate-live and whether to dirty the mod-union table). 5603 assert(_collectorState == Resizing, "Change of collector state to" 5604 " Resizing must be done under the freelistLocks (plural)"); 5605 5606 // Now that sweeping has been completed, we clear 5607 // the incremental_collection_failed flag, 5608 // thus inviting a younger gen collection to promote into 5609 // this generation. If such a promotion may still fail, 5610 // the flag will be set again when a young collection is 5611 // attempted. 5612 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5613 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5614 gch->update_full_collections_completed(_collection_count_start); 5615 } 5616 5617 // FIX ME!!! Looks like this belongs in CFLSpace, with 5618 // CMSGen merely delegating to it. 5619 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5620 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5621 HeapWord* minAddr = _cmsSpace->bottom(); 5622 HeapWord* largestAddr = 5623 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5624 if (largestAddr == NULL) { 5625 // The dictionary appears to be empty. In this case 5626 // try to coalesce at the end of the heap. 5627 largestAddr = _cmsSpace->end(); 5628 } 5629 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5630 size_t nearLargestOffset = 5631 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5632 if (PrintFLSStatistics != 0) { 5633 gclog_or_tty->print_cr( 5634 "CMS: Large Block: " PTR_FORMAT ";" 5635 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5636 p2i(largestAddr), 5637 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5638 } 5639 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5640 } 5641 5642 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5643 return addr >= _cmsSpace->nearLargestChunk(); 5644 } 5645 5646 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5647 return _cmsSpace->find_chunk_at_end(); 5648 } 5649 5650 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5651 bool full) { 5652 // The next lower level has been collected. Gather any statistics 5653 // that are of interest at this point. 5654 if (!full && (current_level + 1) == level()) { 5655 // Gather statistics on the young generation collection. 5656 collector()->stats().record_gc0_end(used()); 5657 } 5658 } 5659 5660 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5661 // We iterate over the space(s) underlying this generation, 5662 // checking the mark bit map to see if the bits corresponding 5663 // to specific blocks are marked or not. Blocks that are 5664 // marked are live and are not swept up. All remaining blocks 5665 // are swept up, with coalescing on-the-fly as we sweep up 5666 // contiguous free and/or garbage blocks: 5667 // We need to ensure that the sweeper synchronizes with allocators 5668 // and stop-the-world collectors. In particular, the following 5669 // locks are used: 5670 // . CMS token: if this is held, a stop the world collection cannot occur 5671 // . freelistLock: if this is held no allocation can occur from this 5672 // generation by another thread 5673 // . bitMapLock: if this is held, no other thread can access or update 5674 // 5675 5676 // Note that we need to hold the freelistLock if we use 5677 // block iterate below; else the iterator might go awry if 5678 // a mutator (or promotion) causes block contents to change 5679 // (for instance if the allocator divvies up a block). 5680 // If we hold the free list lock, for all practical purposes 5681 // young generation GC's can't occur (they'll usually need to 5682 // promote), so we might as well prevent all young generation 5683 // GC's while we do a sweeping step. For the same reason, we might 5684 // as well take the bit map lock for the entire duration 5685 5686 // check that we hold the requisite locks 5687 assert(have_cms_token(), "Should hold cms token"); 5688 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5689 assert_lock_strong(gen->freelistLock()); 5690 assert_lock_strong(bitMapLock()); 5691 5692 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5693 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5694 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5695 _inter_sweep_estimate.padded_average(), 5696 _intra_sweep_estimate.padded_average()); 5697 gen->setNearLargestChunk(); 5698 5699 { 5700 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5701 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5702 // We need to free-up/coalesce garbage/blocks from a 5703 // co-terminal free run. This is done in the SweepClosure 5704 // destructor; so, do not remove this scope, else the 5705 // end-of-sweep-census below will be off by a little bit. 5706 } 5707 gen->cmsSpace()->sweep_completed(); 5708 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5709 if (should_unload_classes()) { // unloaded classes this cycle, 5710 _concurrent_cycles_since_last_unload = 0; // ... reset count 5711 } else { // did not unload classes, 5712 _concurrent_cycles_since_last_unload++; // ... increment count 5713 } 5714 } 5715 5716 // Reset CMS data structures (for now just the marking bit map) 5717 // preparatory for the next cycle. 5718 void CMSCollector::reset(bool concurrent) { 5719 if (concurrent) { 5720 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5721 5722 // If the state is not "Resetting", the foreground thread 5723 // has done a collection and the resetting. 5724 if (_collectorState != Resetting) { 5725 assert(_collectorState == Idling, "The state should only change" 5726 " because the foreground collector has finished the collection"); 5727 return; 5728 } 5729 5730 // Clear the mark bitmap (no grey objects to start with) 5731 // for the next cycle. 5732 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5733 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5734 5735 HeapWord* curAddr = _markBitMap.startWord(); 5736 while (curAddr < _markBitMap.endWord()) { 5737 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5738 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5739 _markBitMap.clear_large_range(chunk); 5740 if (ConcurrentMarkSweepThread::should_yield() && 5741 !foregroundGCIsActive() && 5742 CMSYield) { 5743 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5744 "CMS thread should hold CMS token"); 5745 assert_lock_strong(bitMapLock()); 5746 bitMapLock()->unlock(); 5747 ConcurrentMarkSweepThread::desynchronize(true); 5748 stopTimer(); 5749 if (PrintCMSStatistics != 0) { 5750 incrementYields(); 5751 } 5752 5753 // See the comment in coordinator_yield() 5754 for (unsigned i = 0; i < CMSYieldSleepCount && 5755 ConcurrentMarkSweepThread::should_yield() && 5756 !CMSCollector::foregroundGCIsActive(); ++i) { 5757 os::sleep(Thread::current(), 1, false); 5758 } 5759 5760 ConcurrentMarkSweepThread::synchronize(true); 5761 bitMapLock()->lock_without_safepoint_check(); 5762 startTimer(); 5763 } 5764 curAddr = chunk.end(); 5765 } 5766 // A successful mostly concurrent collection has been done. 5767 // Because only the full (i.e., concurrent mode failure) collections 5768 // are being measured for gc overhead limits, clean the "near" flag 5769 // and count. 5770 size_policy()->reset_gc_overhead_limit_count(); 5771 _collectorState = Idling; 5772 } else { 5773 // already have the lock 5774 assert(_collectorState == Resetting, "just checking"); 5775 assert_lock_strong(bitMapLock()); 5776 _markBitMap.clear_all(); 5777 _collectorState = Idling; 5778 } 5779 5780 register_gc_end(); 5781 } 5782 5783 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5784 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5785 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5786 TraceCollectorStats tcs(counters()); 5787 5788 switch (op) { 5789 case CMS_op_checkpointRootsInitial: { 5790 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5791 checkpointRootsInitial(); 5792 if (PrintGC) { 5793 _cmsGen->printOccupancy("initial-mark"); 5794 } 5795 break; 5796 } 5797 case CMS_op_checkpointRootsFinal: { 5798 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5799 checkpointRootsFinal(); 5800 if (PrintGC) { 5801 _cmsGen->printOccupancy("remark"); 5802 } 5803 break; 5804 } 5805 default: 5806 fatal("No such CMS_op"); 5807 } 5808 } 5809 5810 #ifndef PRODUCT 5811 size_t const CMSCollector::skip_header_HeapWords() { 5812 return FreeChunk::header_size(); 5813 } 5814 5815 // Try and collect here conditions that should hold when 5816 // CMS thread is exiting. The idea is that the foreground GC 5817 // thread should not be blocked if it wants to terminate 5818 // the CMS thread and yet continue to run the VM for a while 5819 // after that. 5820 void CMSCollector::verify_ok_to_terminate() const { 5821 assert(Thread::current()->is_ConcurrentGC_thread(), 5822 "should be called by CMS thread"); 5823 assert(!_foregroundGCShouldWait, "should be false"); 5824 // We could check here that all the various low-level locks 5825 // are not held by the CMS thread, but that is overkill; see 5826 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5827 // is checked. 5828 } 5829 #endif 5830 5831 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5832 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5833 "missing Printezis mark?"); 5834 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5835 size_t size = pointer_delta(nextOneAddr + 1, addr); 5836 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5837 "alignment problem"); 5838 assert(size >= 3, "Necessary for Printezis marks to work"); 5839 return size; 5840 } 5841 5842 // A variant of the above (block_size_using_printezis_bits()) except 5843 // that we return 0 if the P-bits are not yet set. 5844 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5845 if (_markBitMap.isMarked(addr + 1)) { 5846 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5847 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5848 size_t size = pointer_delta(nextOneAddr + 1, addr); 5849 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5850 "alignment problem"); 5851 assert(size >= 3, "Necessary for Printezis marks to work"); 5852 return size; 5853 } 5854 return 0; 5855 } 5856 5857 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5858 size_t sz = 0; 5859 oop p = (oop)addr; 5860 if (p->klass_or_null() != NULL) { 5861 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5862 } else { 5863 sz = block_size_using_printezis_bits(addr); 5864 } 5865 assert(sz > 0, "size must be nonzero"); 5866 HeapWord* next_block = addr + sz; 5867 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5868 CardTableModRefBS::card_size); 5869 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5870 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5871 "must be different cards"); 5872 return next_card; 5873 } 5874 5875 5876 // CMS Bit Map Wrapper ///////////////////////////////////////// 5877 5878 // Construct a CMS bit map infrastructure, but don't create the 5879 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5880 // further below. 5881 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5882 _bm(), 5883 _shifter(shifter), 5884 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5885 Monitor::_safepoint_check_sometimes) : NULL) 5886 { 5887 _bmStartWord = 0; 5888 _bmWordSize = 0; 5889 } 5890 5891 bool CMSBitMap::allocate(MemRegion mr) { 5892 _bmStartWord = mr.start(); 5893 _bmWordSize = mr.word_size(); 5894 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5895 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5896 if (!brs.is_reserved()) { 5897 warning("CMS bit map allocation failure"); 5898 return false; 5899 } 5900 // For now we'll just commit all of the bit map up front. 5901 // Later on we'll try to be more parsimonious with swap. 5902 if (!_virtual_space.initialize(brs, brs.size())) { 5903 warning("CMS bit map backing store failure"); 5904 return false; 5905 } 5906 assert(_virtual_space.committed_size() == brs.size(), 5907 "didn't reserve backing store for all of CMS bit map?"); 5908 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5909 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5910 _bmWordSize, "inconsistency in bit map sizing"); 5911 _bm.set_size(_bmWordSize >> _shifter); 5912 5913 // bm.clear(); // can we rely on getting zero'd memory? verify below 5914 assert(isAllClear(), 5915 "Expected zero'd memory from ReservedSpace constructor"); 5916 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5917 "consistency check"); 5918 return true; 5919 } 5920 5921 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5922 HeapWord *next_addr, *end_addr, *last_addr; 5923 assert_locked(); 5924 assert(covers(mr), "out-of-range error"); 5925 // XXX assert that start and end are appropriately aligned 5926 for (next_addr = mr.start(), end_addr = mr.end(); 5927 next_addr < end_addr; next_addr = last_addr) { 5928 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5929 last_addr = dirty_region.end(); 5930 if (!dirty_region.is_empty()) { 5931 cl->do_MemRegion(dirty_region); 5932 } else { 5933 assert(last_addr == end_addr, "program logic"); 5934 return; 5935 } 5936 } 5937 } 5938 5939 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5940 _bm.print_on_error(st, prefix); 5941 } 5942 5943 #ifndef PRODUCT 5944 void CMSBitMap::assert_locked() const { 5945 CMSLockVerifier::assert_locked(lock()); 5946 } 5947 5948 bool CMSBitMap::covers(MemRegion mr) const { 5949 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5950 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5951 "size inconsistency"); 5952 return (mr.start() >= _bmStartWord) && 5953 (mr.end() <= endWord()); 5954 } 5955 5956 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5957 return (start >= _bmStartWord && (start + size) <= endWord()); 5958 } 5959 5960 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5961 // verify that there are no 1 bits in the interval [left, right) 5962 FalseBitMapClosure falseBitMapClosure; 5963 iterate(&falseBitMapClosure, left, right); 5964 } 5965 5966 void CMSBitMap::region_invariant(MemRegion mr) 5967 { 5968 assert_locked(); 5969 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5970 assert(!mr.is_empty(), "unexpected empty region"); 5971 assert(covers(mr), "mr should be covered by bit map"); 5972 // convert address range into offset range 5973 size_t start_ofs = heapWordToOffset(mr.start()); 5974 // Make sure that end() is appropriately aligned 5975 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5976 (1 << (_shifter+LogHeapWordSize))), 5977 "Misaligned mr.end()"); 5978 size_t end_ofs = heapWordToOffset(mr.end()); 5979 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5980 } 5981 5982 #endif 5983 5984 bool CMSMarkStack::allocate(size_t size) { 5985 // allocate a stack of the requisite depth 5986 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5987 size * sizeof(oop))); 5988 if (!rs.is_reserved()) { 5989 warning("CMSMarkStack allocation failure"); 5990 return false; 5991 } 5992 if (!_virtual_space.initialize(rs, rs.size())) { 5993 warning("CMSMarkStack backing store failure"); 5994 return false; 5995 } 5996 assert(_virtual_space.committed_size() == rs.size(), 5997 "didn't reserve backing store for all of CMS stack?"); 5998 _base = (oop*)(_virtual_space.low()); 5999 _index = 0; 6000 _capacity = size; 6001 NOT_PRODUCT(_max_depth = 0); 6002 return true; 6003 } 6004 6005 // XXX FIX ME !!! In the MT case we come in here holding a 6006 // leaf lock. For printing we need to take a further lock 6007 // which has lower rank. We need to recalibrate the two 6008 // lock-ranks involved in order to be able to print the 6009 // messages below. (Or defer the printing to the caller. 6010 // For now we take the expedient path of just disabling the 6011 // messages for the problematic case.) 6012 void CMSMarkStack::expand() { 6013 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6014 if (_capacity == MarkStackSizeMax) { 6015 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6016 // We print a warning message only once per CMS cycle. 6017 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6018 } 6019 return; 6020 } 6021 // Double capacity if possible 6022 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6023 // Do not give up existing stack until we have managed to 6024 // get the double capacity that we desired. 6025 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6026 new_capacity * sizeof(oop))); 6027 if (rs.is_reserved()) { 6028 // Release the backing store associated with old stack 6029 _virtual_space.release(); 6030 // Reinitialize virtual space for new stack 6031 if (!_virtual_space.initialize(rs, rs.size())) { 6032 fatal("Not enough swap for expanded marking stack"); 6033 } 6034 _base = (oop*)(_virtual_space.low()); 6035 _index = 0; 6036 _capacity = new_capacity; 6037 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6038 // Failed to double capacity, continue; 6039 // we print a detail message only once per CMS cycle. 6040 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6041 SIZE_FORMAT"K", 6042 _capacity / K, new_capacity / K); 6043 } 6044 } 6045 6046 6047 // Closures 6048 // XXX: there seems to be a lot of code duplication here; 6049 // should refactor and consolidate common code. 6050 6051 // This closure is used to mark refs into the CMS generation in 6052 // the CMS bit map. Called at the first checkpoint. This closure 6053 // assumes that we do not need to re-mark dirty cards; if the CMS 6054 // generation on which this is used is not an oldest 6055 // generation then this will lose younger_gen cards! 6056 6057 MarkRefsIntoClosure::MarkRefsIntoClosure( 6058 MemRegion span, CMSBitMap* bitMap): 6059 _span(span), 6060 _bitMap(bitMap) 6061 { 6062 assert(_ref_processor == NULL, "deliberately left NULL"); 6063 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6064 } 6065 6066 void MarkRefsIntoClosure::do_oop(oop obj) { 6067 // if p points into _span, then mark corresponding bit in _markBitMap 6068 assert(obj->is_oop(), "expected an oop"); 6069 HeapWord* addr = (HeapWord*)obj; 6070 if (_span.contains(addr)) { 6071 // this should be made more efficient 6072 _bitMap->mark(addr); 6073 } 6074 } 6075 6076 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6077 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6078 6079 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6080 MemRegion span, CMSBitMap* bitMap): 6081 _span(span), 6082 _bitMap(bitMap) 6083 { 6084 assert(_ref_processor == NULL, "deliberately left NULL"); 6085 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6086 } 6087 6088 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6089 // if p points into _span, then mark corresponding bit in _markBitMap 6090 assert(obj->is_oop(), "expected an oop"); 6091 HeapWord* addr = (HeapWord*)obj; 6092 if (_span.contains(addr)) { 6093 // this should be made more efficient 6094 _bitMap->par_mark(addr); 6095 } 6096 } 6097 6098 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6099 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6100 6101 // A variant of the above, used for CMS marking verification. 6102 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6103 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6104 _span(span), 6105 _verification_bm(verification_bm), 6106 _cms_bm(cms_bm) 6107 { 6108 assert(_ref_processor == NULL, "deliberately left NULL"); 6109 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6110 } 6111 6112 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6113 // if p points into _span, then mark corresponding bit in _markBitMap 6114 assert(obj->is_oop(), "expected an oop"); 6115 HeapWord* addr = (HeapWord*)obj; 6116 if (_span.contains(addr)) { 6117 _verification_bm->mark(addr); 6118 if (!_cms_bm->isMarked(addr)) { 6119 oop(addr)->print(); 6120 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6121 fatal("... aborting"); 6122 } 6123 } 6124 } 6125 6126 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6127 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6128 6129 ////////////////////////////////////////////////// 6130 // MarkRefsIntoAndScanClosure 6131 ////////////////////////////////////////////////// 6132 6133 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6134 ReferenceProcessor* rp, 6135 CMSBitMap* bit_map, 6136 CMSBitMap* mod_union_table, 6137 CMSMarkStack* mark_stack, 6138 CMSCollector* collector, 6139 bool should_yield, 6140 bool concurrent_precleaning): 6141 _collector(collector), 6142 _span(span), 6143 _bit_map(bit_map), 6144 _mark_stack(mark_stack), 6145 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6146 mark_stack, concurrent_precleaning), 6147 _yield(should_yield), 6148 _concurrent_precleaning(concurrent_precleaning), 6149 _freelistLock(NULL) 6150 { 6151 _ref_processor = rp; 6152 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6153 } 6154 6155 // This closure is used to mark refs into the CMS generation at the 6156 // second (final) checkpoint, and to scan and transitively follow 6157 // the unmarked oops. It is also used during the concurrent precleaning 6158 // phase while scanning objects on dirty cards in the CMS generation. 6159 // The marks are made in the marking bit map and the marking stack is 6160 // used for keeping the (newly) grey objects during the scan. 6161 // The parallel version (Par_...) appears further below. 6162 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6163 if (obj != NULL) { 6164 assert(obj->is_oop(), "expected an oop"); 6165 HeapWord* addr = (HeapWord*)obj; 6166 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6167 assert(_collector->overflow_list_is_empty(), 6168 "overflow list should be empty"); 6169 if (_span.contains(addr) && 6170 !_bit_map->isMarked(addr)) { 6171 // mark bit map (object is now grey) 6172 _bit_map->mark(addr); 6173 // push on marking stack (stack should be empty), and drain the 6174 // stack by applying this closure to the oops in the oops popped 6175 // from the stack (i.e. blacken the grey objects) 6176 bool res = _mark_stack->push(obj); 6177 assert(res, "Should have space to push on empty stack"); 6178 do { 6179 oop new_oop = _mark_stack->pop(); 6180 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6181 assert(_bit_map->isMarked((HeapWord*)new_oop), 6182 "only grey objects on this stack"); 6183 // iterate over the oops in this oop, marking and pushing 6184 // the ones in CMS heap (i.e. in _span). 6185 new_oop->oop_iterate(&_pushAndMarkClosure); 6186 // check if it's time to yield 6187 do_yield_check(); 6188 } while (!_mark_stack->isEmpty() || 6189 (!_concurrent_precleaning && take_from_overflow_list())); 6190 // if marking stack is empty, and we are not doing this 6191 // during precleaning, then check the overflow list 6192 } 6193 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6194 assert(_collector->overflow_list_is_empty(), 6195 "overflow list was drained above"); 6196 // We could restore evacuated mark words, if any, used for 6197 // overflow list links here because the overflow list is 6198 // provably empty here. That would reduce the maximum 6199 // size requirements for preserved_{oop,mark}_stack. 6200 // But we'll just postpone it until we are all done 6201 // so we can just stream through. 6202 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6203 _collector->restore_preserved_marks_if_any(); 6204 assert(_collector->no_preserved_marks(), "No preserved marks"); 6205 } 6206 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6207 "All preserved marks should have been restored above"); 6208 } 6209 } 6210 6211 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6212 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6213 6214 void MarkRefsIntoAndScanClosure::do_yield_work() { 6215 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6216 "CMS thread should hold CMS token"); 6217 assert_lock_strong(_freelistLock); 6218 assert_lock_strong(_bit_map->lock()); 6219 // relinquish the free_list_lock and bitMaplock() 6220 _bit_map->lock()->unlock(); 6221 _freelistLock->unlock(); 6222 ConcurrentMarkSweepThread::desynchronize(true); 6223 _collector->stopTimer(); 6224 if (PrintCMSStatistics != 0) { 6225 _collector->incrementYields(); 6226 } 6227 6228 // See the comment in coordinator_yield() 6229 for (unsigned i = 0; 6230 i < CMSYieldSleepCount && 6231 ConcurrentMarkSweepThread::should_yield() && 6232 !CMSCollector::foregroundGCIsActive(); 6233 ++i) { 6234 os::sleep(Thread::current(), 1, false); 6235 } 6236 6237 ConcurrentMarkSweepThread::synchronize(true); 6238 _freelistLock->lock_without_safepoint_check(); 6239 _bit_map->lock()->lock_without_safepoint_check(); 6240 _collector->startTimer(); 6241 } 6242 6243 /////////////////////////////////////////////////////////// 6244 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6245 // MarkRefsIntoAndScanClosure 6246 /////////////////////////////////////////////////////////// 6247 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6248 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6249 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6250 _span(span), 6251 _bit_map(bit_map), 6252 _work_queue(work_queue), 6253 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6254 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6255 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6256 { 6257 _ref_processor = rp; 6258 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6259 } 6260 6261 // This closure is used to mark refs into the CMS generation at the 6262 // second (final) checkpoint, and to scan and transitively follow 6263 // the unmarked oops. The marks are made in the marking bit map and 6264 // the work_queue is used for keeping the (newly) grey objects during 6265 // the scan phase whence they are also available for stealing by parallel 6266 // threads. Since the marking bit map is shared, updates are 6267 // synchronized (via CAS). 6268 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6269 if (obj != NULL) { 6270 // Ignore mark word because this could be an already marked oop 6271 // that may be chained at the end of the overflow list. 6272 assert(obj->is_oop(true), "expected an oop"); 6273 HeapWord* addr = (HeapWord*)obj; 6274 if (_span.contains(addr) && 6275 !_bit_map->isMarked(addr)) { 6276 // mark bit map (object will become grey): 6277 // It is possible for several threads to be 6278 // trying to "claim" this object concurrently; 6279 // the unique thread that succeeds in marking the 6280 // object first will do the subsequent push on 6281 // to the work queue (or overflow list). 6282 if (_bit_map->par_mark(addr)) { 6283 // push on work_queue (which may not be empty), and trim the 6284 // queue to an appropriate length by applying this closure to 6285 // the oops in the oops popped from the stack (i.e. blacken the 6286 // grey objects) 6287 bool res = _work_queue->push(obj); 6288 assert(res, "Low water mark should be less than capacity?"); 6289 trim_queue(_low_water_mark); 6290 } // Else, another thread claimed the object 6291 } 6292 } 6293 } 6294 6295 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6296 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6297 6298 // This closure is used to rescan the marked objects on the dirty cards 6299 // in the mod union table and the card table proper. 6300 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6301 oop p, MemRegion mr) { 6302 6303 size_t size = 0; 6304 HeapWord* addr = (HeapWord*)p; 6305 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6306 assert(_span.contains(addr), "we are scanning the CMS generation"); 6307 // check if it's time to yield 6308 if (do_yield_check()) { 6309 // We yielded for some foreground stop-world work, 6310 // and we have been asked to abort this ongoing preclean cycle. 6311 return 0; 6312 } 6313 if (_bitMap->isMarked(addr)) { 6314 // it's marked; is it potentially uninitialized? 6315 if (p->klass_or_null() != NULL) { 6316 // an initialized object; ignore mark word in verification below 6317 // since we are running concurrent with mutators 6318 assert(p->is_oop(true), "should be an oop"); 6319 if (p->is_objArray()) { 6320 // objArrays are precisely marked; restrict scanning 6321 // to dirty cards only. 6322 size = CompactibleFreeListSpace::adjustObjectSize( 6323 p->oop_iterate(_scanningClosure, mr)); 6324 } else { 6325 // A non-array may have been imprecisely marked; we need 6326 // to scan object in its entirety. 6327 size = CompactibleFreeListSpace::adjustObjectSize( 6328 p->oop_iterate(_scanningClosure)); 6329 } 6330 #ifdef ASSERT 6331 size_t direct_size = 6332 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6333 assert(size == direct_size, "Inconsistency in size"); 6334 assert(size >= 3, "Necessary for Printezis marks to work"); 6335 if (!_bitMap->isMarked(addr+1)) { 6336 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6337 } else { 6338 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6339 assert(_bitMap->isMarked(addr+size-1), 6340 "inconsistent Printezis mark"); 6341 } 6342 #endif // ASSERT 6343 } else { 6344 // An uninitialized object. 6345 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6346 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6347 size = pointer_delta(nextOneAddr + 1, addr); 6348 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6349 "alignment problem"); 6350 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6351 // will dirty the card when the klass pointer is installed in the 6352 // object (signaling the completion of initialization). 6353 } 6354 } else { 6355 // Either a not yet marked object or an uninitialized object 6356 if (p->klass_or_null() == NULL) { 6357 // An uninitialized object, skip to the next card, since 6358 // we may not be able to read its P-bits yet. 6359 assert(size == 0, "Initial value"); 6360 } else { 6361 // An object not (yet) reached by marking: we merely need to 6362 // compute its size so as to go look at the next block. 6363 assert(p->is_oop(true), "should be an oop"); 6364 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6365 } 6366 } 6367 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6368 return size; 6369 } 6370 6371 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6372 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6373 "CMS thread should hold CMS token"); 6374 assert_lock_strong(_freelistLock); 6375 assert_lock_strong(_bitMap->lock()); 6376 // relinquish the free_list_lock and bitMaplock() 6377 _bitMap->lock()->unlock(); 6378 _freelistLock->unlock(); 6379 ConcurrentMarkSweepThread::desynchronize(true); 6380 _collector->stopTimer(); 6381 if (PrintCMSStatistics != 0) { 6382 _collector->incrementYields(); 6383 } 6384 6385 // See the comment in coordinator_yield() 6386 for (unsigned i = 0; i < CMSYieldSleepCount && 6387 ConcurrentMarkSweepThread::should_yield() && 6388 !CMSCollector::foregroundGCIsActive(); ++i) { 6389 os::sleep(Thread::current(), 1, false); 6390 } 6391 6392 ConcurrentMarkSweepThread::synchronize(true); 6393 _freelistLock->lock_without_safepoint_check(); 6394 _bitMap->lock()->lock_without_safepoint_check(); 6395 _collector->startTimer(); 6396 } 6397 6398 6399 ////////////////////////////////////////////////////////////////// 6400 // SurvivorSpacePrecleanClosure 6401 ////////////////////////////////////////////////////////////////// 6402 // This (single-threaded) closure is used to preclean the oops in 6403 // the survivor spaces. 6404 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6405 6406 HeapWord* addr = (HeapWord*)p; 6407 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6408 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6409 assert(p->klass_or_null() != NULL, "object should be initialized"); 6410 // an initialized object; ignore mark word in verification below 6411 // since we are running concurrent with mutators 6412 assert(p->is_oop(true), "should be an oop"); 6413 // Note that we do not yield while we iterate over 6414 // the interior oops of p, pushing the relevant ones 6415 // on our marking stack. 6416 size_t size = p->oop_iterate(_scanning_closure); 6417 do_yield_check(); 6418 // Observe that below, we do not abandon the preclean 6419 // phase as soon as we should; rather we empty the 6420 // marking stack before returning. This is to satisfy 6421 // some existing assertions. In general, it may be a 6422 // good idea to abort immediately and complete the marking 6423 // from the grey objects at a later time. 6424 while (!_mark_stack->isEmpty()) { 6425 oop new_oop = _mark_stack->pop(); 6426 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6427 assert(_bit_map->isMarked((HeapWord*)new_oop), 6428 "only grey objects on this stack"); 6429 // iterate over the oops in this oop, marking and pushing 6430 // the ones in CMS heap (i.e. in _span). 6431 new_oop->oop_iterate(_scanning_closure); 6432 // check if it's time to yield 6433 do_yield_check(); 6434 } 6435 unsigned int after_count = 6436 GenCollectedHeap::heap()->total_collections(); 6437 bool abort = (_before_count != after_count) || 6438 _collector->should_abort_preclean(); 6439 return abort ? 0 : size; 6440 } 6441 6442 void SurvivorSpacePrecleanClosure::do_yield_work() { 6443 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6444 "CMS thread should hold CMS token"); 6445 assert_lock_strong(_bit_map->lock()); 6446 // Relinquish the bit map lock 6447 _bit_map->lock()->unlock(); 6448 ConcurrentMarkSweepThread::desynchronize(true); 6449 _collector->stopTimer(); 6450 if (PrintCMSStatistics != 0) { 6451 _collector->incrementYields(); 6452 } 6453 6454 // See the comment in coordinator_yield() 6455 for (unsigned i = 0; i < CMSYieldSleepCount && 6456 ConcurrentMarkSweepThread::should_yield() && 6457 !CMSCollector::foregroundGCIsActive(); ++i) { 6458 os::sleep(Thread::current(), 1, false); 6459 } 6460 6461 ConcurrentMarkSweepThread::synchronize(true); 6462 _bit_map->lock()->lock_without_safepoint_check(); 6463 _collector->startTimer(); 6464 } 6465 6466 // This closure is used to rescan the marked objects on the dirty cards 6467 // in the mod union table and the card table proper. In the parallel 6468 // case, although the bitMap is shared, we do a single read so the 6469 // isMarked() query is "safe". 6470 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6471 // Ignore mark word because we are running concurrent with mutators 6472 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6473 HeapWord* addr = (HeapWord*)p; 6474 assert(_span.contains(addr), "we are scanning the CMS generation"); 6475 bool is_obj_array = false; 6476 #ifdef ASSERT 6477 if (!_parallel) { 6478 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6479 assert(_collector->overflow_list_is_empty(), 6480 "overflow list should be empty"); 6481 6482 } 6483 #endif // ASSERT 6484 if (_bit_map->isMarked(addr)) { 6485 // Obj arrays are precisely marked, non-arrays are not; 6486 // so we scan objArrays precisely and non-arrays in their 6487 // entirety. 6488 if (p->is_objArray()) { 6489 is_obj_array = true; 6490 if (_parallel) { 6491 p->oop_iterate(_par_scan_closure, mr); 6492 } else { 6493 p->oop_iterate(_scan_closure, mr); 6494 } 6495 } else { 6496 if (_parallel) { 6497 p->oop_iterate(_par_scan_closure); 6498 } else { 6499 p->oop_iterate(_scan_closure); 6500 } 6501 } 6502 } 6503 #ifdef ASSERT 6504 if (!_parallel) { 6505 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6506 assert(_collector->overflow_list_is_empty(), 6507 "overflow list should be empty"); 6508 6509 } 6510 #endif // ASSERT 6511 return is_obj_array; 6512 } 6513 6514 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6515 MemRegion span, 6516 CMSBitMap* bitMap, CMSMarkStack* markStack, 6517 bool should_yield, bool verifying): 6518 _collector(collector), 6519 _span(span), 6520 _bitMap(bitMap), 6521 _mut(&collector->_modUnionTable), 6522 _markStack(markStack), 6523 _yield(should_yield), 6524 _skipBits(0) 6525 { 6526 assert(_markStack->isEmpty(), "stack should be empty"); 6527 _finger = _bitMap->startWord(); 6528 _threshold = _finger; 6529 assert(_collector->_restart_addr == NULL, "Sanity check"); 6530 assert(_span.contains(_finger), "Out of bounds _finger?"); 6531 DEBUG_ONLY(_verifying = verifying;) 6532 } 6533 6534 void MarkFromRootsClosure::reset(HeapWord* addr) { 6535 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6536 assert(_span.contains(addr), "Out of bounds _finger?"); 6537 _finger = addr; 6538 _threshold = (HeapWord*)round_to( 6539 (intptr_t)_finger, CardTableModRefBS::card_size); 6540 } 6541 6542 // Should revisit to see if this should be restructured for 6543 // greater efficiency. 6544 bool MarkFromRootsClosure::do_bit(size_t offset) { 6545 if (_skipBits > 0) { 6546 _skipBits--; 6547 return true; 6548 } 6549 // convert offset into a HeapWord* 6550 HeapWord* addr = _bitMap->startWord() + offset; 6551 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6552 "address out of range"); 6553 assert(_bitMap->isMarked(addr), "tautology"); 6554 if (_bitMap->isMarked(addr+1)) { 6555 // this is an allocated but not yet initialized object 6556 assert(_skipBits == 0, "tautology"); 6557 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6558 oop p = oop(addr); 6559 if (p->klass_or_null() == NULL) { 6560 DEBUG_ONLY(if (!_verifying) {) 6561 // We re-dirty the cards on which this object lies and increase 6562 // the _threshold so that we'll come back to scan this object 6563 // during the preclean or remark phase. (CMSCleanOnEnter) 6564 if (CMSCleanOnEnter) { 6565 size_t sz = _collector->block_size_using_printezis_bits(addr); 6566 HeapWord* end_card_addr = (HeapWord*)round_to( 6567 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6568 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6569 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6570 // Bump _threshold to end_card_addr; note that 6571 // _threshold cannot possibly exceed end_card_addr, anyhow. 6572 // This prevents future clearing of the card as the scan proceeds 6573 // to the right. 6574 assert(_threshold <= end_card_addr, 6575 "Because we are just scanning into this object"); 6576 if (_threshold < end_card_addr) { 6577 _threshold = end_card_addr; 6578 } 6579 if (p->klass_or_null() != NULL) { 6580 // Redirty the range of cards... 6581 _mut->mark_range(redirty_range); 6582 } // ...else the setting of klass will dirty the card anyway. 6583 } 6584 DEBUG_ONLY(}) 6585 return true; 6586 } 6587 } 6588 scanOopsInOop(addr); 6589 return true; 6590 } 6591 6592 // We take a break if we've been at this for a while, 6593 // so as to avoid monopolizing the locks involved. 6594 void MarkFromRootsClosure::do_yield_work() { 6595 // First give up the locks, then yield, then re-lock 6596 // We should probably use a constructor/destructor idiom to 6597 // do this unlock/lock or modify the MutexUnlocker class to 6598 // serve our purpose. XXX 6599 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6600 "CMS thread should hold CMS token"); 6601 assert_lock_strong(_bitMap->lock()); 6602 _bitMap->lock()->unlock(); 6603 ConcurrentMarkSweepThread::desynchronize(true); 6604 _collector->stopTimer(); 6605 if (PrintCMSStatistics != 0) { 6606 _collector->incrementYields(); 6607 } 6608 6609 // See the comment in coordinator_yield() 6610 for (unsigned i = 0; i < CMSYieldSleepCount && 6611 ConcurrentMarkSweepThread::should_yield() && 6612 !CMSCollector::foregroundGCIsActive(); ++i) { 6613 os::sleep(Thread::current(), 1, false); 6614 } 6615 6616 ConcurrentMarkSweepThread::synchronize(true); 6617 _bitMap->lock()->lock_without_safepoint_check(); 6618 _collector->startTimer(); 6619 } 6620 6621 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6622 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6623 assert(_markStack->isEmpty(), 6624 "should drain stack to limit stack usage"); 6625 // convert ptr to an oop preparatory to scanning 6626 oop obj = oop(ptr); 6627 // Ignore mark word in verification below, since we 6628 // may be running concurrent with mutators. 6629 assert(obj->is_oop(true), "should be an oop"); 6630 assert(_finger <= ptr, "_finger runneth ahead"); 6631 // advance the finger to right end of this object 6632 _finger = ptr + obj->size(); 6633 assert(_finger > ptr, "we just incremented it above"); 6634 // On large heaps, it may take us some time to get through 6635 // the marking phase. During 6636 // this time it's possible that a lot of mutations have 6637 // accumulated in the card table and the mod union table -- 6638 // these mutation records are redundant until we have 6639 // actually traced into the corresponding card. 6640 // Here, we check whether advancing the finger would make 6641 // us cross into a new card, and if so clear corresponding 6642 // cards in the MUT (preclean them in the card-table in the 6643 // future). 6644 6645 DEBUG_ONLY(if (!_verifying) {) 6646 // The clean-on-enter optimization is disabled by default, 6647 // until we fix 6178663. 6648 if (CMSCleanOnEnter && (_finger > _threshold)) { 6649 // [_threshold, _finger) represents the interval 6650 // of cards to be cleared in MUT (or precleaned in card table). 6651 // The set of cards to be cleared is all those that overlap 6652 // with the interval [_threshold, _finger); note that 6653 // _threshold is always kept card-aligned but _finger isn't 6654 // always card-aligned. 6655 HeapWord* old_threshold = _threshold; 6656 assert(old_threshold == (HeapWord*)round_to( 6657 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6658 "_threshold should always be card-aligned"); 6659 _threshold = (HeapWord*)round_to( 6660 (intptr_t)_finger, CardTableModRefBS::card_size); 6661 MemRegion mr(old_threshold, _threshold); 6662 assert(!mr.is_empty(), "Control point invariant"); 6663 assert(_span.contains(mr), "Should clear within span"); 6664 _mut->clear_range(mr); 6665 } 6666 DEBUG_ONLY(}) 6667 // Note: the finger doesn't advance while we drain 6668 // the stack below. 6669 PushOrMarkClosure pushOrMarkClosure(_collector, 6670 _span, _bitMap, _markStack, 6671 _finger, this); 6672 bool res = _markStack->push(obj); 6673 assert(res, "Empty non-zero size stack should have space for single push"); 6674 while (!_markStack->isEmpty()) { 6675 oop new_oop = _markStack->pop(); 6676 // Skip verifying header mark word below because we are 6677 // running concurrent with mutators. 6678 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6679 // now scan this oop's oops 6680 new_oop->oop_iterate(&pushOrMarkClosure); 6681 do_yield_check(); 6682 } 6683 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6684 } 6685 6686 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6687 CMSCollector* collector, MemRegion span, 6688 CMSBitMap* bit_map, 6689 OopTaskQueue* work_queue, 6690 CMSMarkStack* overflow_stack): 6691 _collector(collector), 6692 _whole_span(collector->_span), 6693 _span(span), 6694 _bit_map(bit_map), 6695 _mut(&collector->_modUnionTable), 6696 _work_queue(work_queue), 6697 _overflow_stack(overflow_stack), 6698 _skip_bits(0), 6699 _task(task) 6700 { 6701 assert(_work_queue->size() == 0, "work_queue should be empty"); 6702 _finger = span.start(); 6703 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6704 assert(_span.contains(_finger), "Out of bounds _finger?"); 6705 } 6706 6707 // Should revisit to see if this should be restructured for 6708 // greater efficiency. 6709 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6710 if (_skip_bits > 0) { 6711 _skip_bits--; 6712 return true; 6713 } 6714 // convert offset into a HeapWord* 6715 HeapWord* addr = _bit_map->startWord() + offset; 6716 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6717 "address out of range"); 6718 assert(_bit_map->isMarked(addr), "tautology"); 6719 if (_bit_map->isMarked(addr+1)) { 6720 // this is an allocated object that might not yet be initialized 6721 assert(_skip_bits == 0, "tautology"); 6722 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6723 oop p = oop(addr); 6724 if (p->klass_or_null() == NULL) { 6725 // in the case of Clean-on-Enter optimization, redirty card 6726 // and avoid clearing card by increasing the threshold. 6727 return true; 6728 } 6729 } 6730 scan_oops_in_oop(addr); 6731 return true; 6732 } 6733 6734 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6735 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6736 // Should we assert that our work queue is empty or 6737 // below some drain limit? 6738 assert(_work_queue->size() == 0, 6739 "should drain stack to limit stack usage"); 6740 // convert ptr to an oop preparatory to scanning 6741 oop obj = oop(ptr); 6742 // Ignore mark word in verification below, since we 6743 // may be running concurrent with mutators. 6744 assert(obj->is_oop(true), "should be an oop"); 6745 assert(_finger <= ptr, "_finger runneth ahead"); 6746 // advance the finger to right end of this object 6747 _finger = ptr + obj->size(); 6748 assert(_finger > ptr, "we just incremented it above"); 6749 // On large heaps, it may take us some time to get through 6750 // the marking phase. During 6751 // this time it's possible that a lot of mutations have 6752 // accumulated in the card table and the mod union table -- 6753 // these mutation records are redundant until we have 6754 // actually traced into the corresponding card. 6755 // Here, we check whether advancing the finger would make 6756 // us cross into a new card, and if so clear corresponding 6757 // cards in the MUT (preclean them in the card-table in the 6758 // future). 6759 6760 // The clean-on-enter optimization is disabled by default, 6761 // until we fix 6178663. 6762 if (CMSCleanOnEnter && (_finger > _threshold)) { 6763 // [_threshold, _finger) represents the interval 6764 // of cards to be cleared in MUT (or precleaned in card table). 6765 // The set of cards to be cleared is all those that overlap 6766 // with the interval [_threshold, _finger); note that 6767 // _threshold is always kept card-aligned but _finger isn't 6768 // always card-aligned. 6769 HeapWord* old_threshold = _threshold; 6770 assert(old_threshold == (HeapWord*)round_to( 6771 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6772 "_threshold should always be card-aligned"); 6773 _threshold = (HeapWord*)round_to( 6774 (intptr_t)_finger, CardTableModRefBS::card_size); 6775 MemRegion mr(old_threshold, _threshold); 6776 assert(!mr.is_empty(), "Control point invariant"); 6777 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6778 _mut->clear_range(mr); 6779 } 6780 6781 // Note: the local finger doesn't advance while we drain 6782 // the stack below, but the global finger sure can and will. 6783 HeapWord** gfa = _task->global_finger_addr(); 6784 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6785 _span, _bit_map, 6786 _work_queue, 6787 _overflow_stack, 6788 _finger, 6789 gfa, this); 6790 bool res = _work_queue->push(obj); // overflow could occur here 6791 assert(res, "Will hold once we use workqueues"); 6792 while (true) { 6793 oop new_oop; 6794 if (!_work_queue->pop_local(new_oop)) { 6795 // We emptied our work_queue; check if there's stuff that can 6796 // be gotten from the overflow stack. 6797 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6798 _overflow_stack, _work_queue)) { 6799 do_yield_check(); 6800 continue; 6801 } else { // done 6802 break; 6803 } 6804 } 6805 // Skip verifying header mark word below because we are 6806 // running concurrent with mutators. 6807 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6808 // now scan this oop's oops 6809 new_oop->oop_iterate(&pushOrMarkClosure); 6810 do_yield_check(); 6811 } 6812 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6813 } 6814 6815 // Yield in response to a request from VM Thread or 6816 // from mutators. 6817 void Par_MarkFromRootsClosure::do_yield_work() { 6818 assert(_task != NULL, "sanity"); 6819 _task->yield(); 6820 } 6821 6822 // A variant of the above used for verifying CMS marking work. 6823 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6824 MemRegion span, 6825 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6826 CMSMarkStack* mark_stack): 6827 _collector(collector), 6828 _span(span), 6829 _verification_bm(verification_bm), 6830 _cms_bm(cms_bm), 6831 _mark_stack(mark_stack), 6832 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6833 mark_stack) 6834 { 6835 assert(_mark_stack->isEmpty(), "stack should be empty"); 6836 _finger = _verification_bm->startWord(); 6837 assert(_collector->_restart_addr == NULL, "Sanity check"); 6838 assert(_span.contains(_finger), "Out of bounds _finger?"); 6839 } 6840 6841 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6842 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6843 assert(_span.contains(addr), "Out of bounds _finger?"); 6844 _finger = addr; 6845 } 6846 6847 // Should revisit to see if this should be restructured for 6848 // greater efficiency. 6849 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6850 // convert offset into a HeapWord* 6851 HeapWord* addr = _verification_bm->startWord() + offset; 6852 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6853 "address out of range"); 6854 assert(_verification_bm->isMarked(addr), "tautology"); 6855 assert(_cms_bm->isMarked(addr), "tautology"); 6856 6857 assert(_mark_stack->isEmpty(), 6858 "should drain stack to limit stack usage"); 6859 // convert addr to an oop preparatory to scanning 6860 oop obj = oop(addr); 6861 assert(obj->is_oop(), "should be an oop"); 6862 assert(_finger <= addr, "_finger runneth ahead"); 6863 // advance the finger to right end of this object 6864 _finger = addr + obj->size(); 6865 assert(_finger > addr, "we just incremented it above"); 6866 // Note: the finger doesn't advance while we drain 6867 // the stack below. 6868 bool res = _mark_stack->push(obj); 6869 assert(res, "Empty non-zero size stack should have space for single push"); 6870 while (!_mark_stack->isEmpty()) { 6871 oop new_oop = _mark_stack->pop(); 6872 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6873 // now scan this oop's oops 6874 new_oop->oop_iterate(&_pam_verify_closure); 6875 } 6876 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6877 return true; 6878 } 6879 6880 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6881 CMSCollector* collector, MemRegion span, 6882 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6883 CMSMarkStack* mark_stack): 6884 MetadataAwareOopClosure(collector->ref_processor()), 6885 _collector(collector), 6886 _span(span), 6887 _verification_bm(verification_bm), 6888 _cms_bm(cms_bm), 6889 _mark_stack(mark_stack) 6890 { } 6891 6892 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6893 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6894 6895 // Upon stack overflow, we discard (part of) the stack, 6896 // remembering the least address amongst those discarded 6897 // in CMSCollector's _restart_address. 6898 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6899 // Remember the least grey address discarded 6900 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6901 _collector->lower_restart_addr(ra); 6902 _mark_stack->reset(); // discard stack contents 6903 _mark_stack->expand(); // expand the stack if possible 6904 } 6905 6906 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6907 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6908 HeapWord* addr = (HeapWord*)obj; 6909 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6910 // Oop lies in _span and isn't yet grey or black 6911 _verification_bm->mark(addr); // now grey 6912 if (!_cms_bm->isMarked(addr)) { 6913 oop(addr)->print(); 6914 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6915 p2i(addr)); 6916 fatal("... aborting"); 6917 } 6918 6919 if (!_mark_stack->push(obj)) { // stack overflow 6920 if (PrintCMSStatistics != 0) { 6921 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6922 SIZE_FORMAT, _mark_stack->capacity()); 6923 } 6924 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6925 handle_stack_overflow(addr); 6926 } 6927 // anything including and to the right of _finger 6928 // will be scanned as we iterate over the remainder of the 6929 // bit map 6930 } 6931 } 6932 6933 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6934 MemRegion span, 6935 CMSBitMap* bitMap, CMSMarkStack* markStack, 6936 HeapWord* finger, MarkFromRootsClosure* parent) : 6937 MetadataAwareOopClosure(collector->ref_processor()), 6938 _collector(collector), 6939 _span(span), 6940 _bitMap(bitMap), 6941 _markStack(markStack), 6942 _finger(finger), 6943 _parent(parent) 6944 { } 6945 6946 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6947 MemRegion span, 6948 CMSBitMap* bit_map, 6949 OopTaskQueue* work_queue, 6950 CMSMarkStack* overflow_stack, 6951 HeapWord* finger, 6952 HeapWord** global_finger_addr, 6953 Par_MarkFromRootsClosure* parent) : 6954 MetadataAwareOopClosure(collector->ref_processor()), 6955 _collector(collector), 6956 _whole_span(collector->_span), 6957 _span(span), 6958 _bit_map(bit_map), 6959 _work_queue(work_queue), 6960 _overflow_stack(overflow_stack), 6961 _finger(finger), 6962 _global_finger_addr(global_finger_addr), 6963 _parent(parent) 6964 { } 6965 6966 // Assumes thread-safe access by callers, who are 6967 // responsible for mutual exclusion. 6968 void CMSCollector::lower_restart_addr(HeapWord* low) { 6969 assert(_span.contains(low), "Out of bounds addr"); 6970 if (_restart_addr == NULL) { 6971 _restart_addr = low; 6972 } else { 6973 _restart_addr = MIN2(_restart_addr, low); 6974 } 6975 } 6976 6977 // Upon stack overflow, we discard (part of) the stack, 6978 // remembering the least address amongst those discarded 6979 // in CMSCollector's _restart_address. 6980 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6981 // Remember the least grey address discarded 6982 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6983 _collector->lower_restart_addr(ra); 6984 _markStack->reset(); // discard stack contents 6985 _markStack->expand(); // expand the stack if possible 6986 } 6987 6988 // Upon stack overflow, we discard (part of) the stack, 6989 // remembering the least address amongst those discarded 6990 // in CMSCollector's _restart_address. 6991 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6992 // We need to do this under a mutex to prevent other 6993 // workers from interfering with the work done below. 6994 MutexLockerEx ml(_overflow_stack->par_lock(), 6995 Mutex::_no_safepoint_check_flag); 6996 // Remember the least grey address discarded 6997 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6998 _collector->lower_restart_addr(ra); 6999 _overflow_stack->reset(); // discard stack contents 7000 _overflow_stack->expand(); // expand the stack if possible 7001 } 7002 7003 void PushOrMarkClosure::do_oop(oop obj) { 7004 // Ignore mark word because we are running concurrent with mutators. 7005 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7006 HeapWord* addr = (HeapWord*)obj; 7007 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7008 // Oop lies in _span and isn't yet grey or black 7009 _bitMap->mark(addr); // now grey 7010 if (addr < _finger) { 7011 // the bit map iteration has already either passed, or 7012 // sampled, this bit in the bit map; we'll need to 7013 // use the marking stack to scan this oop's oops. 7014 bool simulate_overflow = false; 7015 NOT_PRODUCT( 7016 if (CMSMarkStackOverflowALot && 7017 _collector->simulate_overflow()) { 7018 // simulate a stack overflow 7019 simulate_overflow = true; 7020 } 7021 ) 7022 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7023 if (PrintCMSStatistics != 0) { 7024 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7025 SIZE_FORMAT, _markStack->capacity()); 7026 } 7027 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7028 handle_stack_overflow(addr); 7029 } 7030 } 7031 // anything including and to the right of _finger 7032 // will be scanned as we iterate over the remainder of the 7033 // bit map 7034 do_yield_check(); 7035 } 7036 } 7037 7038 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7039 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7040 7041 void Par_PushOrMarkClosure::do_oop(oop obj) { 7042 // Ignore mark word because we are running concurrent with mutators. 7043 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7044 HeapWord* addr = (HeapWord*)obj; 7045 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7046 // Oop lies in _span and isn't yet grey or black 7047 // We read the global_finger (volatile read) strictly after marking oop 7048 bool res = _bit_map->par_mark(addr); // now grey 7049 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7050 // Should we push this marked oop on our stack? 7051 // -- if someone else marked it, nothing to do 7052 // -- if target oop is above global finger nothing to do 7053 // -- if target oop is in chunk and above local finger 7054 // then nothing to do 7055 // -- else push on work queue 7056 if ( !res // someone else marked it, they will deal with it 7057 || (addr >= *gfa) // will be scanned in a later task 7058 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7059 return; 7060 } 7061 // the bit map iteration has already either passed, or 7062 // sampled, this bit in the bit map; we'll need to 7063 // use the marking stack to scan this oop's oops. 7064 bool simulate_overflow = false; 7065 NOT_PRODUCT( 7066 if (CMSMarkStackOverflowALot && 7067 _collector->simulate_overflow()) { 7068 // simulate a stack overflow 7069 simulate_overflow = true; 7070 } 7071 ) 7072 if (simulate_overflow || 7073 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7074 // stack overflow 7075 if (PrintCMSStatistics != 0) { 7076 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7077 SIZE_FORMAT, _overflow_stack->capacity()); 7078 } 7079 // We cannot assert that the overflow stack is full because 7080 // it may have been emptied since. 7081 assert(simulate_overflow || 7082 _work_queue->size() == _work_queue->max_elems(), 7083 "Else push should have succeeded"); 7084 handle_stack_overflow(addr); 7085 } 7086 do_yield_check(); 7087 } 7088 } 7089 7090 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7091 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7092 7093 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7094 MemRegion span, 7095 ReferenceProcessor* rp, 7096 CMSBitMap* bit_map, 7097 CMSBitMap* mod_union_table, 7098 CMSMarkStack* mark_stack, 7099 bool concurrent_precleaning): 7100 MetadataAwareOopClosure(rp), 7101 _collector(collector), 7102 _span(span), 7103 _bit_map(bit_map), 7104 _mod_union_table(mod_union_table), 7105 _mark_stack(mark_stack), 7106 _concurrent_precleaning(concurrent_precleaning) 7107 { 7108 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7109 } 7110 7111 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7112 // the non-parallel version (the parallel version appears further below.) 7113 void PushAndMarkClosure::do_oop(oop obj) { 7114 // Ignore mark word verification. If during concurrent precleaning, 7115 // the object monitor may be locked. If during the checkpoint 7116 // phases, the object may already have been reached by a different 7117 // path and may be at the end of the global overflow list (so 7118 // the mark word may be NULL). 7119 assert(obj->is_oop_or_null(true /* ignore mark word */), 7120 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7121 HeapWord* addr = (HeapWord*)obj; 7122 // Check if oop points into the CMS generation 7123 // and is not marked 7124 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7125 // a white object ... 7126 _bit_map->mark(addr); // ... now grey 7127 // push on the marking stack (grey set) 7128 bool simulate_overflow = false; 7129 NOT_PRODUCT( 7130 if (CMSMarkStackOverflowALot && 7131 _collector->simulate_overflow()) { 7132 // simulate a stack overflow 7133 simulate_overflow = true; 7134 } 7135 ) 7136 if (simulate_overflow || !_mark_stack->push(obj)) { 7137 if (_concurrent_precleaning) { 7138 // During precleaning we can just dirty the appropriate card(s) 7139 // in the mod union table, thus ensuring that the object remains 7140 // in the grey set and continue. In the case of object arrays 7141 // we need to dirty all of the cards that the object spans, 7142 // since the rescan of object arrays will be limited to the 7143 // dirty cards. 7144 // Note that no one can be interfering with us in this action 7145 // of dirtying the mod union table, so no locking or atomics 7146 // are required. 7147 if (obj->is_objArray()) { 7148 size_t sz = obj->size(); 7149 HeapWord* end_card_addr = (HeapWord*)round_to( 7150 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7151 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7152 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7153 _mod_union_table->mark_range(redirty_range); 7154 } else { 7155 _mod_union_table->mark(addr); 7156 } 7157 _collector->_ser_pmc_preclean_ovflw++; 7158 } else { 7159 // During the remark phase, we need to remember this oop 7160 // in the overflow list. 7161 _collector->push_on_overflow_list(obj); 7162 _collector->_ser_pmc_remark_ovflw++; 7163 } 7164 } 7165 } 7166 } 7167 7168 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7169 MemRegion span, 7170 ReferenceProcessor* rp, 7171 CMSBitMap* bit_map, 7172 OopTaskQueue* work_queue): 7173 MetadataAwareOopClosure(rp), 7174 _collector(collector), 7175 _span(span), 7176 _bit_map(bit_map), 7177 _work_queue(work_queue) 7178 { 7179 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7180 } 7181 7182 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7183 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7184 7185 // Grey object rescan during second checkpoint phase -- 7186 // the parallel version. 7187 void Par_PushAndMarkClosure::do_oop(oop obj) { 7188 // In the assert below, we ignore the mark word because 7189 // this oop may point to an already visited object that is 7190 // on the overflow stack (in which case the mark word has 7191 // been hijacked for chaining into the overflow stack -- 7192 // if this is the last object in the overflow stack then 7193 // its mark word will be NULL). Because this object may 7194 // have been subsequently popped off the global overflow 7195 // stack, and the mark word possibly restored to the prototypical 7196 // value, by the time we get to examined this failing assert in 7197 // the debugger, is_oop_or_null(false) may subsequently start 7198 // to hold. 7199 assert(obj->is_oop_or_null(true), 7200 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7201 HeapWord* addr = (HeapWord*)obj; 7202 // Check if oop points into the CMS generation 7203 // and is not marked 7204 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7205 // a white object ... 7206 // If we manage to "claim" the object, by being the 7207 // first thread to mark it, then we push it on our 7208 // marking stack 7209 if (_bit_map->par_mark(addr)) { // ... now grey 7210 // push on work queue (grey set) 7211 bool simulate_overflow = false; 7212 NOT_PRODUCT( 7213 if (CMSMarkStackOverflowALot && 7214 _collector->par_simulate_overflow()) { 7215 // simulate a stack overflow 7216 simulate_overflow = true; 7217 } 7218 ) 7219 if (simulate_overflow || !_work_queue->push(obj)) { 7220 _collector->par_push_on_overflow_list(obj); 7221 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7222 } 7223 } // Else, some other thread got there first 7224 } 7225 } 7226 7227 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7228 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7229 7230 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7231 Mutex* bml = _collector->bitMapLock(); 7232 assert_lock_strong(bml); 7233 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7234 "CMS thread should hold CMS token"); 7235 7236 bml->unlock(); 7237 ConcurrentMarkSweepThread::desynchronize(true); 7238 7239 _collector->stopTimer(); 7240 if (PrintCMSStatistics != 0) { 7241 _collector->incrementYields(); 7242 } 7243 7244 // See the comment in coordinator_yield() 7245 for (unsigned i = 0; i < CMSYieldSleepCount && 7246 ConcurrentMarkSweepThread::should_yield() && 7247 !CMSCollector::foregroundGCIsActive(); ++i) { 7248 os::sleep(Thread::current(), 1, false); 7249 } 7250 7251 ConcurrentMarkSweepThread::synchronize(true); 7252 bml->lock(); 7253 7254 _collector->startTimer(); 7255 } 7256 7257 bool CMSPrecleanRefsYieldClosure::should_return() { 7258 if (ConcurrentMarkSweepThread::should_yield()) { 7259 do_yield_work(); 7260 } 7261 return _collector->foregroundGCIsActive(); 7262 } 7263 7264 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7265 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7266 "mr should be aligned to start at a card boundary"); 7267 // We'd like to assert: 7268 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7269 // "mr should be a range of cards"); 7270 // However, that would be too strong in one case -- the last 7271 // partition ends at _unallocated_block which, in general, can be 7272 // an arbitrary boundary, not necessarily card aligned. 7273 if (PrintCMSStatistics != 0) { 7274 _num_dirty_cards += 7275 mr.word_size()/CardTableModRefBS::card_size_in_words; 7276 } 7277 _space->object_iterate_mem(mr, &_scan_cl); 7278 } 7279 7280 SweepClosure::SweepClosure(CMSCollector* collector, 7281 ConcurrentMarkSweepGeneration* g, 7282 CMSBitMap* bitMap, bool should_yield) : 7283 _collector(collector), 7284 _g(g), 7285 _sp(g->cmsSpace()), 7286 _limit(_sp->sweep_limit()), 7287 _freelistLock(_sp->freelistLock()), 7288 _bitMap(bitMap), 7289 _yield(should_yield), 7290 _inFreeRange(false), // No free range at beginning of sweep 7291 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7292 _lastFreeRangeCoalesced(false), 7293 _freeFinger(g->used_region().start()) 7294 { 7295 NOT_PRODUCT( 7296 _numObjectsFreed = 0; 7297 _numWordsFreed = 0; 7298 _numObjectsLive = 0; 7299 _numWordsLive = 0; 7300 _numObjectsAlreadyFree = 0; 7301 _numWordsAlreadyFree = 0; 7302 _last_fc = NULL; 7303 7304 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7305 _sp->dictionary()->initialize_dict_returned_bytes(); 7306 ) 7307 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7308 "sweep _limit out of bounds"); 7309 if (CMSTraceSweeper) { 7310 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7311 p2i(_limit)); 7312 } 7313 } 7314 7315 void SweepClosure::print_on(outputStream* st) const { 7316 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7317 p2i(_sp->bottom()), p2i(_sp->end())); 7318 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7319 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7320 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7321 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7322 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7323 } 7324 7325 #ifndef PRODUCT 7326 // Assertion checking only: no useful work in product mode -- 7327 // however, if any of the flags below become product flags, 7328 // you may need to review this code to see if it needs to be 7329 // enabled in product mode. 7330 SweepClosure::~SweepClosure() { 7331 assert_lock_strong(_freelistLock); 7332 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7333 "sweep _limit out of bounds"); 7334 if (inFreeRange()) { 7335 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7336 print(); 7337 ShouldNotReachHere(); 7338 } 7339 if (Verbose && PrintGC) { 7340 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7341 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7342 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7343 SIZE_FORMAT" bytes " 7344 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7345 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7346 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7347 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7348 * sizeof(HeapWord); 7349 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7350 7351 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7352 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7353 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7354 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7355 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7356 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7357 indexListReturnedBytes); 7358 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7359 dict_returned_bytes); 7360 } 7361 } 7362 if (CMSTraceSweeper) { 7363 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7364 p2i(_limit)); 7365 } 7366 } 7367 #endif // PRODUCT 7368 7369 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7370 bool freeRangeInFreeLists) { 7371 if (CMSTraceSweeper) { 7372 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7373 p2i(freeFinger), freeRangeInFreeLists); 7374 } 7375 assert(!inFreeRange(), "Trampling existing free range"); 7376 set_inFreeRange(true); 7377 set_lastFreeRangeCoalesced(false); 7378 7379 set_freeFinger(freeFinger); 7380 set_freeRangeInFreeLists(freeRangeInFreeLists); 7381 if (CMSTestInFreeList) { 7382 if (freeRangeInFreeLists) { 7383 FreeChunk* fc = (FreeChunk*) freeFinger; 7384 assert(fc->is_free(), "A chunk on the free list should be free."); 7385 assert(fc->size() > 0, "Free range should have a size"); 7386 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7387 } 7388 } 7389 } 7390 7391 // Note that the sweeper runs concurrently with mutators. Thus, 7392 // it is possible for direct allocation in this generation to happen 7393 // in the middle of the sweep. Note that the sweeper also coalesces 7394 // contiguous free blocks. Thus, unless the sweeper and the allocator 7395 // synchronize appropriately freshly allocated blocks may get swept up. 7396 // This is accomplished by the sweeper locking the free lists while 7397 // it is sweeping. Thus blocks that are determined to be free are 7398 // indeed free. There is however one additional complication: 7399 // blocks that have been allocated since the final checkpoint and 7400 // mark, will not have been marked and so would be treated as 7401 // unreachable and swept up. To prevent this, the allocator marks 7402 // the bit map when allocating during the sweep phase. This leads, 7403 // however, to a further complication -- objects may have been allocated 7404 // but not yet initialized -- in the sense that the header isn't yet 7405 // installed. The sweeper can not then determine the size of the block 7406 // in order to skip over it. To deal with this case, we use a technique 7407 // (due to Printezis) to encode such uninitialized block sizes in the 7408 // bit map. Since the bit map uses a bit per every HeapWord, but the 7409 // CMS generation has a minimum object size of 3 HeapWords, it follows 7410 // that "normal marks" won't be adjacent in the bit map (there will 7411 // always be at least two 0 bits between successive 1 bits). We make use 7412 // of these "unused" bits to represent uninitialized blocks -- the bit 7413 // corresponding to the start of the uninitialized object and the next 7414 // bit are both set. Finally, a 1 bit marks the end of the object that 7415 // started with the two consecutive 1 bits to indicate its potentially 7416 // uninitialized state. 7417 7418 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7419 FreeChunk* fc = (FreeChunk*)addr; 7420 size_t res; 7421 7422 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7423 // than "addr == _limit" because although _limit was a block boundary when 7424 // we started the sweep, it may no longer be one because heap expansion 7425 // may have caused us to coalesce the block ending at the address _limit 7426 // with a newly expanded chunk (this happens when _limit was set to the 7427 // previous _end of the space), so we may have stepped past _limit: 7428 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7429 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7430 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7431 "sweep _limit out of bounds"); 7432 assert(addr < _sp->end(), "addr out of bounds"); 7433 // Flush any free range we might be holding as a single 7434 // coalesced chunk to the appropriate free list. 7435 if (inFreeRange()) { 7436 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7437 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7438 flush_cur_free_chunk(freeFinger(), 7439 pointer_delta(addr, freeFinger())); 7440 if (CMSTraceSweeper) { 7441 gclog_or_tty->print("Sweep: last chunk: "); 7442 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7443 "[coalesced:%d]\n", 7444 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7445 lastFreeRangeCoalesced() ? 1 : 0); 7446 } 7447 } 7448 7449 // help the iterator loop finish 7450 return pointer_delta(_sp->end(), addr); 7451 } 7452 7453 assert(addr < _limit, "sweep invariant"); 7454 // check if we should yield 7455 do_yield_check(addr); 7456 if (fc->is_free()) { 7457 // Chunk that is already free 7458 res = fc->size(); 7459 do_already_free_chunk(fc); 7460 debug_only(_sp->verifyFreeLists()); 7461 // If we flush the chunk at hand in lookahead_and_flush() 7462 // and it's coalesced with a preceding chunk, then the 7463 // process of "mangling" the payload of the coalesced block 7464 // will cause erasure of the size information from the 7465 // (erstwhile) header of all the coalesced blocks but the 7466 // first, so the first disjunct in the assert will not hold 7467 // in that specific case (in which case the second disjunct 7468 // will hold). 7469 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7470 "Otherwise the size info doesn't change at this step"); 7471 NOT_PRODUCT( 7472 _numObjectsAlreadyFree++; 7473 _numWordsAlreadyFree += res; 7474 ) 7475 NOT_PRODUCT(_last_fc = fc;) 7476 } else if (!_bitMap->isMarked(addr)) { 7477 // Chunk is fresh garbage 7478 res = do_garbage_chunk(fc); 7479 debug_only(_sp->verifyFreeLists()); 7480 NOT_PRODUCT( 7481 _numObjectsFreed++; 7482 _numWordsFreed += res; 7483 ) 7484 } else { 7485 // Chunk that is alive. 7486 res = do_live_chunk(fc); 7487 debug_only(_sp->verifyFreeLists()); 7488 NOT_PRODUCT( 7489 _numObjectsLive++; 7490 _numWordsLive += res; 7491 ) 7492 } 7493 return res; 7494 } 7495 7496 // For the smart allocation, record following 7497 // split deaths - a free chunk is removed from its free list because 7498 // it is being split into two or more chunks. 7499 // split birth - a free chunk is being added to its free list because 7500 // a larger free chunk has been split and resulted in this free chunk. 7501 // coal death - a free chunk is being removed from its free list because 7502 // it is being coalesced into a large free chunk. 7503 // coal birth - a free chunk is being added to its free list because 7504 // it was created when two or more free chunks where coalesced into 7505 // this free chunk. 7506 // 7507 // These statistics are used to determine the desired number of free 7508 // chunks of a given size. The desired number is chosen to be relative 7509 // to the end of a CMS sweep. The desired number at the end of a sweep 7510 // is the 7511 // count-at-end-of-previous-sweep (an amount that was enough) 7512 // - count-at-beginning-of-current-sweep (the excess) 7513 // + split-births (gains in this size during interval) 7514 // - split-deaths (demands on this size during interval) 7515 // where the interval is from the end of one sweep to the end of the 7516 // next. 7517 // 7518 // When sweeping the sweeper maintains an accumulated chunk which is 7519 // the chunk that is made up of chunks that have been coalesced. That 7520 // will be termed the left-hand chunk. A new chunk of garbage that 7521 // is being considered for coalescing will be referred to as the 7522 // right-hand chunk. 7523 // 7524 // When making a decision on whether to coalesce a right-hand chunk with 7525 // the current left-hand chunk, the current count vs. the desired count 7526 // of the left-hand chunk is considered. Also if the right-hand chunk 7527 // is near the large chunk at the end of the heap (see 7528 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7529 // left-hand chunk is coalesced. 7530 // 7531 // When making a decision about whether to split a chunk, the desired count 7532 // vs. the current count of the candidate to be split is also considered. 7533 // If the candidate is underpopulated (currently fewer chunks than desired) 7534 // a chunk of an overpopulated (currently more chunks than desired) size may 7535 // be chosen. The "hint" associated with a free list, if non-null, points 7536 // to a free list which may be overpopulated. 7537 // 7538 7539 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7540 const size_t size = fc->size(); 7541 // Chunks that cannot be coalesced are not in the 7542 // free lists. 7543 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7544 assert(_sp->verify_chunk_in_free_list(fc), 7545 "free chunk should be in free lists"); 7546 } 7547 // a chunk that is already free, should not have been 7548 // marked in the bit map 7549 HeapWord* const addr = (HeapWord*) fc; 7550 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7551 // Verify that the bit map has no bits marked between 7552 // addr and purported end of this block. 7553 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7554 7555 // Some chunks cannot be coalesced under any circumstances. 7556 // See the definition of cantCoalesce(). 7557 if (!fc->cantCoalesce()) { 7558 // This chunk can potentially be coalesced. 7559 if (_sp->adaptive_freelists()) { 7560 // All the work is done in 7561 do_post_free_or_garbage_chunk(fc, size); 7562 } else { // Not adaptive free lists 7563 // this is a free chunk that can potentially be coalesced by the sweeper; 7564 if (!inFreeRange()) { 7565 // if the next chunk is a free block that can't be coalesced 7566 // it doesn't make sense to remove this chunk from the free lists 7567 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7568 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7569 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7570 nextChunk->is_free() && // ... which is free... 7571 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7572 // nothing to do 7573 } else { 7574 // Potentially the start of a new free range: 7575 // Don't eagerly remove it from the free lists. 7576 // No need to remove it if it will just be put 7577 // back again. (Also from a pragmatic point of view 7578 // if it is a free block in a region that is beyond 7579 // any allocated blocks, an assertion will fail) 7580 // Remember the start of a free run. 7581 initialize_free_range(addr, true); 7582 // end - can coalesce with next chunk 7583 } 7584 } else { 7585 // the midst of a free range, we are coalescing 7586 print_free_block_coalesced(fc); 7587 if (CMSTraceSweeper) { 7588 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7589 } 7590 // remove it from the free lists 7591 _sp->removeFreeChunkFromFreeLists(fc); 7592 set_lastFreeRangeCoalesced(true); 7593 // If the chunk is being coalesced and the current free range is 7594 // in the free lists, remove the current free range so that it 7595 // will be returned to the free lists in its entirety - all 7596 // the coalesced pieces included. 7597 if (freeRangeInFreeLists()) { 7598 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7599 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7600 "Size of free range is inconsistent with chunk size."); 7601 if (CMSTestInFreeList) { 7602 assert(_sp->verify_chunk_in_free_list(ffc), 7603 "free range is not in free lists"); 7604 } 7605 _sp->removeFreeChunkFromFreeLists(ffc); 7606 set_freeRangeInFreeLists(false); 7607 } 7608 } 7609 } 7610 // Note that if the chunk is not coalescable (the else arm 7611 // below), we unconditionally flush, without needing to do 7612 // a "lookahead," as we do below. 7613 if (inFreeRange()) lookahead_and_flush(fc, size); 7614 } else { 7615 // Code path common to both original and adaptive free lists. 7616 7617 // cant coalesce with previous block; this should be treated 7618 // as the end of a free run if any 7619 if (inFreeRange()) { 7620 // we kicked some butt; time to pick up the garbage 7621 assert(freeFinger() < addr, "freeFinger points too high"); 7622 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7623 } 7624 // else, nothing to do, just continue 7625 } 7626 } 7627 7628 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7629 // This is a chunk of garbage. It is not in any free list. 7630 // Add it to a free list or let it possibly be coalesced into 7631 // a larger chunk. 7632 HeapWord* const addr = (HeapWord*) fc; 7633 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7634 7635 if (_sp->adaptive_freelists()) { 7636 // Verify that the bit map has no bits marked between 7637 // addr and purported end of just dead object. 7638 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7639 7640 do_post_free_or_garbage_chunk(fc, size); 7641 } else { 7642 if (!inFreeRange()) { 7643 // start of a new free range 7644 assert(size > 0, "A free range should have a size"); 7645 initialize_free_range(addr, false); 7646 } else { 7647 // this will be swept up when we hit the end of the 7648 // free range 7649 if (CMSTraceSweeper) { 7650 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7651 } 7652 // If the chunk is being coalesced and the current free range is 7653 // in the free lists, remove the current free range so that it 7654 // will be returned to the free lists in its entirety - all 7655 // the coalesced pieces included. 7656 if (freeRangeInFreeLists()) { 7657 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7658 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7659 "Size of free range is inconsistent with chunk size."); 7660 if (CMSTestInFreeList) { 7661 assert(_sp->verify_chunk_in_free_list(ffc), 7662 "free range is not in free lists"); 7663 } 7664 _sp->removeFreeChunkFromFreeLists(ffc); 7665 set_freeRangeInFreeLists(false); 7666 } 7667 set_lastFreeRangeCoalesced(true); 7668 } 7669 // this will be swept up when we hit the end of the free range 7670 7671 // Verify that the bit map has no bits marked between 7672 // addr and purported end of just dead object. 7673 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7674 } 7675 assert(_limit >= addr + size, 7676 "A freshly garbage chunk can't possibly straddle over _limit"); 7677 if (inFreeRange()) lookahead_and_flush(fc, size); 7678 return size; 7679 } 7680 7681 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7682 HeapWord* addr = (HeapWord*) fc; 7683 // The sweeper has just found a live object. Return any accumulated 7684 // left hand chunk to the free lists. 7685 if (inFreeRange()) { 7686 assert(freeFinger() < addr, "freeFinger points too high"); 7687 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7688 } 7689 7690 // This object is live: we'd normally expect this to be 7691 // an oop, and like to assert the following: 7692 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7693 // However, as we commented above, this may be an object whose 7694 // header hasn't yet been initialized. 7695 size_t size; 7696 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7697 if (_bitMap->isMarked(addr + 1)) { 7698 // Determine the size from the bit map, rather than trying to 7699 // compute it from the object header. 7700 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7701 size = pointer_delta(nextOneAddr + 1, addr); 7702 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7703 "alignment problem"); 7704 7705 #ifdef ASSERT 7706 if (oop(addr)->klass_or_null() != NULL) { 7707 // Ignore mark word because we are running concurrent with mutators 7708 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7709 assert(size == 7710 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7711 "P-mark and computed size do not agree"); 7712 } 7713 #endif 7714 7715 } else { 7716 // This should be an initialized object that's alive. 7717 assert(oop(addr)->klass_or_null() != NULL, 7718 "Should be an initialized object"); 7719 // Ignore mark word because we are running concurrent with mutators 7720 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7721 // Verify that the bit map has no bits marked between 7722 // addr and purported end of this block. 7723 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7724 assert(size >= 3, "Necessary for Printezis marks to work"); 7725 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7726 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7727 } 7728 return size; 7729 } 7730 7731 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7732 size_t chunkSize) { 7733 // do_post_free_or_garbage_chunk() should only be called in the case 7734 // of the adaptive free list allocator. 7735 const bool fcInFreeLists = fc->is_free(); 7736 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7737 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7738 if (CMSTestInFreeList && fcInFreeLists) { 7739 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7740 } 7741 7742 if (CMSTraceSweeper) { 7743 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7744 } 7745 7746 HeapWord* const fc_addr = (HeapWord*) fc; 7747 7748 bool coalesce; 7749 const size_t left = pointer_delta(fc_addr, freeFinger()); 7750 const size_t right = chunkSize; 7751 switch (FLSCoalescePolicy) { 7752 // numeric value forms a coalition aggressiveness metric 7753 case 0: { // never coalesce 7754 coalesce = false; 7755 break; 7756 } 7757 case 1: { // coalesce if left & right chunks on overpopulated lists 7758 coalesce = _sp->coalOverPopulated(left) && 7759 _sp->coalOverPopulated(right); 7760 break; 7761 } 7762 case 2: { // coalesce if left chunk on overpopulated list (default) 7763 coalesce = _sp->coalOverPopulated(left); 7764 break; 7765 } 7766 case 3: { // coalesce if left OR right chunk on overpopulated list 7767 coalesce = _sp->coalOverPopulated(left) || 7768 _sp->coalOverPopulated(right); 7769 break; 7770 } 7771 case 4: { // always coalesce 7772 coalesce = true; 7773 break; 7774 } 7775 default: 7776 ShouldNotReachHere(); 7777 } 7778 7779 // Should the current free range be coalesced? 7780 // If the chunk is in a free range and either we decided to coalesce above 7781 // or the chunk is near the large block at the end of the heap 7782 // (isNearLargestChunk() returns true), then coalesce this chunk. 7783 const bool doCoalesce = inFreeRange() 7784 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7785 if (doCoalesce) { 7786 // Coalesce the current free range on the left with the new 7787 // chunk on the right. If either is on a free list, 7788 // it must be removed from the list and stashed in the closure. 7789 if (freeRangeInFreeLists()) { 7790 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7791 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7792 "Size of free range is inconsistent with chunk size."); 7793 if (CMSTestInFreeList) { 7794 assert(_sp->verify_chunk_in_free_list(ffc), 7795 "Chunk is not in free lists"); 7796 } 7797 _sp->coalDeath(ffc->size()); 7798 _sp->removeFreeChunkFromFreeLists(ffc); 7799 set_freeRangeInFreeLists(false); 7800 } 7801 if (fcInFreeLists) { 7802 _sp->coalDeath(chunkSize); 7803 assert(fc->size() == chunkSize, 7804 "The chunk has the wrong size or is not in the free lists"); 7805 _sp->removeFreeChunkFromFreeLists(fc); 7806 } 7807 set_lastFreeRangeCoalesced(true); 7808 print_free_block_coalesced(fc); 7809 } else { // not in a free range and/or should not coalesce 7810 // Return the current free range and start a new one. 7811 if (inFreeRange()) { 7812 // In a free range but cannot coalesce with the right hand chunk. 7813 // Put the current free range into the free lists. 7814 flush_cur_free_chunk(freeFinger(), 7815 pointer_delta(fc_addr, freeFinger())); 7816 } 7817 // Set up for new free range. Pass along whether the right hand 7818 // chunk is in the free lists. 7819 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7820 } 7821 } 7822 7823 // Lookahead flush: 7824 // If we are tracking a free range, and this is the last chunk that 7825 // we'll look at because its end crosses past _limit, we'll preemptively 7826 // flush it along with any free range we may be holding on to. Note that 7827 // this can be the case only for an already free or freshly garbage 7828 // chunk. If this block is an object, it can never straddle 7829 // over _limit. The "straddling" occurs when _limit is set at 7830 // the previous end of the space when this cycle started, and 7831 // a subsequent heap expansion caused the previously co-terminal 7832 // free block to be coalesced with the newly expanded portion, 7833 // thus rendering _limit a non-block-boundary making it dangerous 7834 // for the sweeper to step over and examine. 7835 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7836 assert(inFreeRange(), "Should only be called if currently in a free range."); 7837 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7838 assert(_sp->used_region().contains(eob - 1), 7839 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7840 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7841 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7842 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7843 if (eob >= _limit) { 7844 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7845 if (CMSTraceSweeper) { 7846 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7847 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7848 "[" PTR_FORMAT "," PTR_FORMAT ")", 7849 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7850 } 7851 // Return the storage we are tracking back into the free lists. 7852 if (CMSTraceSweeper) { 7853 gclog_or_tty->print_cr("Flushing ... "); 7854 } 7855 assert(freeFinger() < eob, "Error"); 7856 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7857 } 7858 } 7859 7860 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7861 assert(inFreeRange(), "Should only be called if currently in a free range."); 7862 assert(size > 0, 7863 "A zero sized chunk cannot be added to the free lists."); 7864 if (!freeRangeInFreeLists()) { 7865 if (CMSTestInFreeList) { 7866 FreeChunk* fc = (FreeChunk*) chunk; 7867 fc->set_size(size); 7868 assert(!_sp->verify_chunk_in_free_list(fc), 7869 "chunk should not be in free lists yet"); 7870 } 7871 if (CMSTraceSweeper) { 7872 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7873 p2i(chunk), size); 7874 } 7875 // A new free range is going to be starting. The current 7876 // free range has not been added to the free lists yet or 7877 // was removed so add it back. 7878 // If the current free range was coalesced, then the death 7879 // of the free range was recorded. Record a birth now. 7880 if (lastFreeRangeCoalesced()) { 7881 _sp->coalBirth(size); 7882 } 7883 _sp->addChunkAndRepairOffsetTable(chunk, size, 7884 lastFreeRangeCoalesced()); 7885 } else if (CMSTraceSweeper) { 7886 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7887 } 7888 set_inFreeRange(false); 7889 set_freeRangeInFreeLists(false); 7890 } 7891 7892 // We take a break if we've been at this for a while, 7893 // so as to avoid monopolizing the locks involved. 7894 void SweepClosure::do_yield_work(HeapWord* addr) { 7895 // Return current free chunk being used for coalescing (if any) 7896 // to the appropriate freelist. After yielding, the next 7897 // free block encountered will start a coalescing range of 7898 // free blocks. If the next free block is adjacent to the 7899 // chunk just flushed, they will need to wait for the next 7900 // sweep to be coalesced. 7901 if (inFreeRange()) { 7902 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7903 } 7904 7905 // First give up the locks, then yield, then re-lock. 7906 // We should probably use a constructor/destructor idiom to 7907 // do this unlock/lock or modify the MutexUnlocker class to 7908 // serve our purpose. XXX 7909 assert_lock_strong(_bitMap->lock()); 7910 assert_lock_strong(_freelistLock); 7911 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7912 "CMS thread should hold CMS token"); 7913 _bitMap->lock()->unlock(); 7914 _freelistLock->unlock(); 7915 ConcurrentMarkSweepThread::desynchronize(true); 7916 _collector->stopTimer(); 7917 if (PrintCMSStatistics != 0) { 7918 _collector->incrementYields(); 7919 } 7920 7921 // See the comment in coordinator_yield() 7922 for (unsigned i = 0; i < CMSYieldSleepCount && 7923 ConcurrentMarkSweepThread::should_yield() && 7924 !CMSCollector::foregroundGCIsActive(); ++i) { 7925 os::sleep(Thread::current(), 1, false); 7926 } 7927 7928 ConcurrentMarkSweepThread::synchronize(true); 7929 _freelistLock->lock(); 7930 _bitMap->lock()->lock_without_safepoint_check(); 7931 _collector->startTimer(); 7932 } 7933 7934 #ifndef PRODUCT 7935 // This is actually very useful in a product build if it can 7936 // be called from the debugger. Compile it into the product 7937 // as needed. 7938 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7939 return debug_cms_space->verify_chunk_in_free_list(fc); 7940 } 7941 #endif 7942 7943 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7944 if (CMSTraceSweeper) { 7945 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7946 p2i(fc), fc->size()); 7947 } 7948 } 7949 7950 // CMSIsAliveClosure 7951 bool CMSIsAliveClosure::do_object_b(oop obj) { 7952 HeapWord* addr = (HeapWord*)obj; 7953 return addr != NULL && 7954 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7955 } 7956 7957 7958 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7959 MemRegion span, 7960 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7961 bool cpc): 7962 _collector(collector), 7963 _span(span), 7964 _bit_map(bit_map), 7965 _mark_stack(mark_stack), 7966 _concurrent_precleaning(cpc) { 7967 assert(!_span.is_empty(), "Empty span could spell trouble"); 7968 } 7969 7970 7971 // CMSKeepAliveClosure: the serial version 7972 void CMSKeepAliveClosure::do_oop(oop obj) { 7973 HeapWord* addr = (HeapWord*)obj; 7974 if (_span.contains(addr) && 7975 !_bit_map->isMarked(addr)) { 7976 _bit_map->mark(addr); 7977 bool simulate_overflow = false; 7978 NOT_PRODUCT( 7979 if (CMSMarkStackOverflowALot && 7980 _collector->simulate_overflow()) { 7981 // simulate a stack overflow 7982 simulate_overflow = true; 7983 } 7984 ) 7985 if (simulate_overflow || !_mark_stack->push(obj)) { 7986 if (_concurrent_precleaning) { 7987 // We dirty the overflown object and let the remark 7988 // phase deal with it. 7989 assert(_collector->overflow_list_is_empty(), "Error"); 7990 // In the case of object arrays, we need to dirty all of 7991 // the cards that the object spans. No locking or atomics 7992 // are needed since no one else can be mutating the mod union 7993 // table. 7994 if (obj->is_objArray()) { 7995 size_t sz = obj->size(); 7996 HeapWord* end_card_addr = 7997 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7998 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7999 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8000 _collector->_modUnionTable.mark_range(redirty_range); 8001 } else { 8002 _collector->_modUnionTable.mark(addr); 8003 } 8004 _collector->_ser_kac_preclean_ovflw++; 8005 } else { 8006 _collector->push_on_overflow_list(obj); 8007 _collector->_ser_kac_ovflw++; 8008 } 8009 } 8010 } 8011 } 8012 8013 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8014 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8015 8016 // CMSParKeepAliveClosure: a parallel version of the above. 8017 // The work queues are private to each closure (thread), 8018 // but (may be) available for stealing by other threads. 8019 void CMSParKeepAliveClosure::do_oop(oop obj) { 8020 HeapWord* addr = (HeapWord*)obj; 8021 if (_span.contains(addr) && 8022 !_bit_map->isMarked(addr)) { 8023 // In general, during recursive tracing, several threads 8024 // may be concurrently getting here; the first one to 8025 // "tag" it, claims it. 8026 if (_bit_map->par_mark(addr)) { 8027 bool res = _work_queue->push(obj); 8028 assert(res, "Low water mark should be much less than capacity"); 8029 // Do a recursive trim in the hope that this will keep 8030 // stack usage lower, but leave some oops for potential stealers 8031 trim_queue(_low_water_mark); 8032 } // Else, another thread got there first 8033 } 8034 } 8035 8036 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8037 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8038 8039 void CMSParKeepAliveClosure::trim_queue(uint max) { 8040 while (_work_queue->size() > max) { 8041 oop new_oop; 8042 if (_work_queue->pop_local(new_oop)) { 8043 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8044 assert(_bit_map->isMarked((HeapWord*)new_oop), 8045 "no white objects on this stack!"); 8046 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8047 // iterate over the oops in this oop, marking and pushing 8048 // the ones in CMS heap (i.e. in _span). 8049 new_oop->oop_iterate(&_mark_and_push); 8050 } 8051 } 8052 } 8053 8054 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8055 CMSCollector* collector, 8056 MemRegion span, CMSBitMap* bit_map, 8057 OopTaskQueue* work_queue): 8058 _collector(collector), 8059 _span(span), 8060 _bit_map(bit_map), 8061 _work_queue(work_queue) { } 8062 8063 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8064 HeapWord* addr = (HeapWord*)obj; 8065 if (_span.contains(addr) && 8066 !_bit_map->isMarked(addr)) { 8067 if (_bit_map->par_mark(addr)) { 8068 bool simulate_overflow = false; 8069 NOT_PRODUCT( 8070 if (CMSMarkStackOverflowALot && 8071 _collector->par_simulate_overflow()) { 8072 // simulate a stack overflow 8073 simulate_overflow = true; 8074 } 8075 ) 8076 if (simulate_overflow || !_work_queue->push(obj)) { 8077 _collector->par_push_on_overflow_list(obj); 8078 _collector->_par_kac_ovflw++; 8079 } 8080 } // Else another thread got there already 8081 } 8082 } 8083 8084 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8085 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8086 8087 ////////////////////////////////////////////////////////////////// 8088 // CMSExpansionCause ///////////////////////////// 8089 ////////////////////////////////////////////////////////////////// 8090 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8091 switch (cause) { 8092 case _no_expansion: 8093 return "No expansion"; 8094 case _satisfy_free_ratio: 8095 return "Free ratio"; 8096 case _satisfy_promotion: 8097 return "Satisfy promotion"; 8098 case _satisfy_allocation: 8099 return "allocation"; 8100 case _allocate_par_lab: 8101 return "Par LAB"; 8102 case _allocate_par_spooling_space: 8103 return "Par Spooling Space"; 8104 case _adaptive_size_policy: 8105 return "Ergonomics"; 8106 default: 8107 return "unknown"; 8108 } 8109 } 8110 8111 void CMSDrainMarkingStackClosure::do_void() { 8112 // the max number to take from overflow list at a time 8113 const size_t num = _mark_stack->capacity()/4; 8114 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8115 "Overflow list should be NULL during concurrent phases"); 8116 while (!_mark_stack->isEmpty() || 8117 // if stack is empty, check the overflow list 8118 _collector->take_from_overflow_list(num, _mark_stack)) { 8119 oop obj = _mark_stack->pop(); 8120 HeapWord* addr = (HeapWord*)obj; 8121 assert(_span.contains(addr), "Should be within span"); 8122 assert(_bit_map->isMarked(addr), "Should be marked"); 8123 assert(obj->is_oop(), "Should be an oop"); 8124 obj->oop_iterate(_keep_alive); 8125 } 8126 } 8127 8128 void CMSParDrainMarkingStackClosure::do_void() { 8129 // drain queue 8130 trim_queue(0); 8131 } 8132 8133 // Trim our work_queue so its length is below max at return 8134 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8135 while (_work_queue->size() > max) { 8136 oop new_oop; 8137 if (_work_queue->pop_local(new_oop)) { 8138 assert(new_oop->is_oop(), "Expected an oop"); 8139 assert(_bit_map->isMarked((HeapWord*)new_oop), 8140 "no white objects on this stack!"); 8141 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8142 // iterate over the oops in this oop, marking and pushing 8143 // the ones in CMS heap (i.e. in _span). 8144 new_oop->oop_iterate(&_mark_and_push); 8145 } 8146 } 8147 } 8148 8149 //////////////////////////////////////////////////////////////////// 8150 // Support for Marking Stack Overflow list handling and related code 8151 //////////////////////////////////////////////////////////////////// 8152 // Much of the following code is similar in shape and spirit to the 8153 // code used in ParNewGC. We should try and share that code 8154 // as much as possible in the future. 8155 8156 #ifndef PRODUCT 8157 // Debugging support for CMSStackOverflowALot 8158 8159 // It's OK to call this multi-threaded; the worst thing 8160 // that can happen is that we'll get a bunch of closely 8161 // spaced simulated overflows, but that's OK, in fact 8162 // probably good as it would exercise the overflow code 8163 // under contention. 8164 bool CMSCollector::simulate_overflow() { 8165 if (_overflow_counter-- <= 0) { // just being defensive 8166 _overflow_counter = CMSMarkStackOverflowInterval; 8167 return true; 8168 } else { 8169 return false; 8170 } 8171 } 8172 8173 bool CMSCollector::par_simulate_overflow() { 8174 return simulate_overflow(); 8175 } 8176 #endif 8177 8178 // Single-threaded 8179 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8180 assert(stack->isEmpty(), "Expected precondition"); 8181 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8182 size_t i = num; 8183 oop cur = _overflow_list; 8184 const markOop proto = markOopDesc::prototype(); 8185 NOT_PRODUCT(ssize_t n = 0;) 8186 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8187 next = oop(cur->mark()); 8188 cur->set_mark(proto); // until proven otherwise 8189 assert(cur->is_oop(), "Should be an oop"); 8190 bool res = stack->push(cur); 8191 assert(res, "Bit off more than can chew?"); 8192 NOT_PRODUCT(n++;) 8193 } 8194 _overflow_list = cur; 8195 #ifndef PRODUCT 8196 assert(_num_par_pushes >= n, "Too many pops?"); 8197 _num_par_pushes -=n; 8198 #endif 8199 return !stack->isEmpty(); 8200 } 8201 8202 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8203 // (MT-safe) Get a prefix of at most "num" from the list. 8204 // The overflow list is chained through the mark word of 8205 // each object in the list. We fetch the entire list, 8206 // break off a prefix of the right size and return the 8207 // remainder. If other threads try to take objects from 8208 // the overflow list at that time, they will wait for 8209 // some time to see if data becomes available. If (and 8210 // only if) another thread places one or more object(s) 8211 // on the global list before we have returned the suffix 8212 // to the global list, we will walk down our local list 8213 // to find its end and append the global list to 8214 // our suffix before returning it. This suffix walk can 8215 // prove to be expensive (quadratic in the amount of traffic) 8216 // when there are many objects in the overflow list and 8217 // there is much producer-consumer contention on the list. 8218 // *NOTE*: The overflow list manipulation code here and 8219 // in ParNewGeneration:: are very similar in shape, 8220 // except that in the ParNew case we use the old (from/eden) 8221 // copy of the object to thread the list via its klass word. 8222 // Because of the common code, if you make any changes in 8223 // the code below, please check the ParNew version to see if 8224 // similar changes might be needed. 8225 // CR 6797058 has been filed to consolidate the common code. 8226 bool CMSCollector::par_take_from_overflow_list(size_t num, 8227 OopTaskQueue* work_q, 8228 int no_of_gc_threads) { 8229 assert(work_q->size() == 0, "First empty local work queue"); 8230 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8231 if (_overflow_list == NULL) { 8232 return false; 8233 } 8234 // Grab the entire list; we'll put back a suffix 8235 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8236 Thread* tid = Thread::current(); 8237 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8238 // set to ParallelGCThreads. 8239 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8240 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8241 // If the list is busy, we spin for a short while, 8242 // sleeping between attempts to get the list. 8243 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8244 os::sleep(tid, sleep_time_millis, false); 8245 if (_overflow_list == NULL) { 8246 // Nothing left to take 8247 return false; 8248 } else if (_overflow_list != BUSY) { 8249 // Try and grab the prefix 8250 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8251 } 8252 } 8253 // If the list was found to be empty, or we spun long 8254 // enough, we give up and return empty-handed. If we leave 8255 // the list in the BUSY state below, it must be the case that 8256 // some other thread holds the overflow list and will set it 8257 // to a non-BUSY state in the future. 8258 if (prefix == NULL || prefix == BUSY) { 8259 // Nothing to take or waited long enough 8260 if (prefix == NULL) { 8261 // Write back the NULL in case we overwrote it with BUSY above 8262 // and it is still the same value. 8263 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8264 } 8265 return false; 8266 } 8267 assert(prefix != NULL && prefix != BUSY, "Error"); 8268 size_t i = num; 8269 oop cur = prefix; 8270 // Walk down the first "num" objects, unless we reach the end. 8271 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8272 if (cur->mark() == NULL) { 8273 // We have "num" or fewer elements in the list, so there 8274 // is nothing to return to the global list. 8275 // Write back the NULL in lieu of the BUSY we wrote 8276 // above, if it is still the same value. 8277 if (_overflow_list == BUSY) { 8278 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8279 } 8280 } else { 8281 // Chop off the suffix and return it to the global list. 8282 assert(cur->mark() != BUSY, "Error"); 8283 oop suffix_head = cur->mark(); // suffix will be put back on global list 8284 cur->set_mark(NULL); // break off suffix 8285 // It's possible that the list is still in the empty(busy) state 8286 // we left it in a short while ago; in that case we may be 8287 // able to place back the suffix without incurring the cost 8288 // of a walk down the list. 8289 oop observed_overflow_list = _overflow_list; 8290 oop cur_overflow_list = observed_overflow_list; 8291 bool attached = false; 8292 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8293 observed_overflow_list = 8294 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8295 if (cur_overflow_list == observed_overflow_list) { 8296 attached = true; 8297 break; 8298 } else cur_overflow_list = observed_overflow_list; 8299 } 8300 if (!attached) { 8301 // Too bad, someone else sneaked in (at least) an element; we'll need 8302 // to do a splice. Find tail of suffix so we can prepend suffix to global 8303 // list. 8304 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8305 oop suffix_tail = cur; 8306 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8307 "Tautology"); 8308 observed_overflow_list = _overflow_list; 8309 do { 8310 cur_overflow_list = observed_overflow_list; 8311 if (cur_overflow_list != BUSY) { 8312 // Do the splice ... 8313 suffix_tail->set_mark(markOop(cur_overflow_list)); 8314 } else { // cur_overflow_list == BUSY 8315 suffix_tail->set_mark(NULL); 8316 } 8317 // ... and try to place spliced list back on overflow_list ... 8318 observed_overflow_list = 8319 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8320 } while (cur_overflow_list != observed_overflow_list); 8321 // ... until we have succeeded in doing so. 8322 } 8323 } 8324 8325 // Push the prefix elements on work_q 8326 assert(prefix != NULL, "control point invariant"); 8327 const markOop proto = markOopDesc::prototype(); 8328 oop next; 8329 NOT_PRODUCT(ssize_t n = 0;) 8330 for (cur = prefix; cur != NULL; cur = next) { 8331 next = oop(cur->mark()); 8332 cur->set_mark(proto); // until proven otherwise 8333 assert(cur->is_oop(), "Should be an oop"); 8334 bool res = work_q->push(cur); 8335 assert(res, "Bit off more than we can chew?"); 8336 NOT_PRODUCT(n++;) 8337 } 8338 #ifndef PRODUCT 8339 assert(_num_par_pushes >= n, "Too many pops?"); 8340 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8341 #endif 8342 return true; 8343 } 8344 8345 // Single-threaded 8346 void CMSCollector::push_on_overflow_list(oop p) { 8347 NOT_PRODUCT(_num_par_pushes++;) 8348 assert(p->is_oop(), "Not an oop"); 8349 preserve_mark_if_necessary(p); 8350 p->set_mark((markOop)_overflow_list); 8351 _overflow_list = p; 8352 } 8353 8354 // Multi-threaded; use CAS to prepend to overflow list 8355 void CMSCollector::par_push_on_overflow_list(oop p) { 8356 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8357 assert(p->is_oop(), "Not an oop"); 8358 par_preserve_mark_if_necessary(p); 8359 oop observed_overflow_list = _overflow_list; 8360 oop cur_overflow_list; 8361 do { 8362 cur_overflow_list = observed_overflow_list; 8363 if (cur_overflow_list != BUSY) { 8364 p->set_mark(markOop(cur_overflow_list)); 8365 } else { 8366 p->set_mark(NULL); 8367 } 8368 observed_overflow_list = 8369 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8370 } while (cur_overflow_list != observed_overflow_list); 8371 } 8372 #undef BUSY 8373 8374 // Single threaded 8375 // General Note on GrowableArray: pushes may silently fail 8376 // because we are (temporarily) out of C-heap for expanding 8377 // the stack. The problem is quite ubiquitous and affects 8378 // a lot of code in the JVM. The prudent thing for GrowableArray 8379 // to do (for now) is to exit with an error. However, that may 8380 // be too draconian in some cases because the caller may be 8381 // able to recover without much harm. For such cases, we 8382 // should probably introduce a "soft_push" method which returns 8383 // an indication of success or failure with the assumption that 8384 // the caller may be able to recover from a failure; code in 8385 // the VM can then be changed, incrementally, to deal with such 8386 // failures where possible, thus, incrementally hardening the VM 8387 // in such low resource situations. 8388 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8389 _preserved_oop_stack.push(p); 8390 _preserved_mark_stack.push(m); 8391 assert(m == p->mark(), "Mark word changed"); 8392 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8393 "bijection"); 8394 } 8395 8396 // Single threaded 8397 void CMSCollector::preserve_mark_if_necessary(oop p) { 8398 markOop m = p->mark(); 8399 if (m->must_be_preserved(p)) { 8400 preserve_mark_work(p, m); 8401 } 8402 } 8403 8404 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8405 markOop m = p->mark(); 8406 if (m->must_be_preserved(p)) { 8407 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8408 // Even though we read the mark word without holding 8409 // the lock, we are assured that it will not change 8410 // because we "own" this oop, so no other thread can 8411 // be trying to push it on the overflow list; see 8412 // the assertion in preserve_mark_work() that checks 8413 // that m == p->mark(). 8414 preserve_mark_work(p, m); 8415 } 8416 } 8417 8418 // We should be able to do this multi-threaded, 8419 // a chunk of stack being a task (this is 8420 // correct because each oop only ever appears 8421 // once in the overflow list. However, it's 8422 // not very easy to completely overlap this with 8423 // other operations, so will generally not be done 8424 // until all work's been completed. Because we 8425 // expect the preserved oop stack (set) to be small, 8426 // it's probably fine to do this single-threaded. 8427 // We can explore cleverer concurrent/overlapped/parallel 8428 // processing of preserved marks if we feel the 8429 // need for this in the future. Stack overflow should 8430 // be so rare in practice and, when it happens, its 8431 // effect on performance so great that this will 8432 // likely just be in the noise anyway. 8433 void CMSCollector::restore_preserved_marks_if_any() { 8434 assert(SafepointSynchronize::is_at_safepoint(), 8435 "world should be stopped"); 8436 assert(Thread::current()->is_ConcurrentGC_thread() || 8437 Thread::current()->is_VM_thread(), 8438 "should be single-threaded"); 8439 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8440 "bijection"); 8441 8442 while (!_preserved_oop_stack.is_empty()) { 8443 oop p = _preserved_oop_stack.pop(); 8444 assert(p->is_oop(), "Should be an oop"); 8445 assert(_span.contains(p), "oop should be in _span"); 8446 assert(p->mark() == markOopDesc::prototype(), 8447 "Set when taken from overflow list"); 8448 markOop m = _preserved_mark_stack.pop(); 8449 p->set_mark(m); 8450 } 8451 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8452 "stacks were cleared above"); 8453 } 8454 8455 #ifndef PRODUCT 8456 bool CMSCollector::no_preserved_marks() const { 8457 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8458 } 8459 #endif 8460 8461 // Transfer some number of overflown objects to usual marking 8462 // stack. Return true if some objects were transferred. 8463 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8464 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8465 (size_t)ParGCDesiredObjsFromOverflowList); 8466 8467 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8468 assert(_collector->overflow_list_is_empty() || res, 8469 "If list is not empty, we should have taken something"); 8470 assert(!res || !_mark_stack->isEmpty(), 8471 "If we took something, it should now be on our stack"); 8472 return res; 8473 } 8474 8475 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8476 size_t res = _sp->block_size_no_stall(addr, _collector); 8477 if (_sp->block_is_obj(addr)) { 8478 if (_live_bit_map->isMarked(addr)) { 8479 // It can't have been dead in a previous cycle 8480 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8481 } else { 8482 _dead_bit_map->mark(addr); // mark the dead object 8483 } 8484 } 8485 // Could be 0, if the block size could not be computed without stalling. 8486 return res; 8487 } 8488 8489 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8490 8491 switch (phase) { 8492 case CMSCollector::InitialMarking: 8493 initialize(true /* fullGC */ , 8494 cause /* cause of the GC */, 8495 true /* recordGCBeginTime */, 8496 true /* recordPreGCUsage */, 8497 false /* recordPeakUsage */, 8498 false /* recordPostGCusage */, 8499 true /* recordAccumulatedGCTime */, 8500 false /* recordGCEndTime */, 8501 false /* countCollection */ ); 8502 break; 8503 8504 case CMSCollector::FinalMarking: 8505 initialize(true /* fullGC */ , 8506 cause /* cause of the GC */, 8507 false /* recordGCBeginTime */, 8508 false /* recordPreGCUsage */, 8509 false /* recordPeakUsage */, 8510 false /* recordPostGCusage */, 8511 true /* recordAccumulatedGCTime */, 8512 false /* recordGCEndTime */, 8513 false /* countCollection */ ); 8514 break; 8515 8516 case CMSCollector::Sweeping: 8517 initialize(true /* fullGC */ , 8518 cause /* cause of the GC */, 8519 false /* recordGCBeginTime */, 8520 false /* recordPreGCUsage */, 8521 true /* recordPeakUsage */, 8522 true /* recordPostGCusage */, 8523 false /* recordAccumulatedGCTime */, 8524 true /* recordGCEndTime */, 8525 true /* countCollection */ ); 8526 break; 8527 8528 default: 8529 ShouldNotReachHere(); 8530 } 8531 }