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/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/referencePolicy.hpp" 54 #include "gc/shared/strongRootsScope.hpp" 55 #include "gc/shared/taskqueue.inline.hpp" 56 #include "memory/allocation.hpp" 57 #include "memory/iterator.inline.hpp" 58 #include "memory/padded.hpp" 59 #include "memory/resourceArea.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "prims/jvmtiExport.hpp" 62 #include "runtime/atomic.inline.hpp" 63 #include "runtime/globals_extension.hpp" 64 #include "runtime/handles.inline.hpp" 65 #include "runtime/java.hpp" 66 #include "runtime/orderAccess.inline.hpp" 67 #include "runtime/vmThread.hpp" 68 #include "services/memoryService.hpp" 69 #include "services/runtimeService.hpp" 70 #include "utilities/stack.inline.hpp" 71 72 // statics 73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 74 bool CMSCollector::_full_gc_requested = false; 75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 76 77 ////////////////////////////////////////////////////////////////// 78 // In support of CMS/VM thread synchronization 79 ////////////////////////////////////////////////////////////////// 80 // We split use of the CGC_lock into 2 "levels". 81 // The low-level locking is of the usual CGC_lock monitor. We introduce 82 // a higher level "token" (hereafter "CMS token") built on top of the 83 // low level monitor (hereafter "CGC lock"). 84 // The token-passing protocol gives priority to the VM thread. The 85 // CMS-lock doesn't provide any fairness guarantees, but clients 86 // should ensure that it is only held for very short, bounded 87 // durations. 88 // 89 // When either of the CMS thread or the VM thread is involved in 90 // collection operations during which it does not want the other 91 // thread to interfere, it obtains the CMS token. 92 // 93 // If either thread tries to get the token while the other has 94 // it, that thread waits. However, if the VM thread and CMS thread 95 // both want the token, then the VM thread gets priority while the 96 // CMS thread waits. This ensures, for instance, that the "concurrent" 97 // phases of the CMS thread's work do not block out the VM thread 98 // for long periods of time as the CMS thread continues to hog 99 // the token. (See bug 4616232). 100 // 101 // The baton-passing functions are, however, controlled by the 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 103 // and here the low-level CMS lock, not the high level token, 104 // ensures mutual exclusion. 105 // 106 // Two important conditions that we have to satisfy: 107 // 1. if a thread does a low-level wait on the CMS lock, then it 108 // relinquishes the CMS token if it were holding that token 109 // when it acquired the low-level CMS lock. 110 // 2. any low-level notifications on the low-level lock 111 // should only be sent when a thread has relinquished the token. 112 // 113 // In the absence of either property, we'd have potential deadlock. 114 // 115 // We protect each of the CMS (concurrent and sequential) phases 116 // with the CMS _token_, not the CMS _lock_. 117 // 118 // The only code protected by CMS lock is the token acquisition code 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 120 // baton-passing code. 121 // 122 // Unfortunately, i couldn't come up with a good abstraction to factor and 123 // hide the naked CGC_lock manipulation in the baton-passing code 124 // further below. That's something we should try to do. Also, the proof 125 // of correctness of this 2-level locking scheme is far from obvious, 126 // and potentially quite slippery. We have an uneasy suspicion, for instance, 127 // that there may be a theoretical possibility of delay/starvation in the 128 // low-level lock/wait/notify scheme used for the baton-passing because of 129 // potential interference with the priority scheme embodied in the 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 131 // invocation further below and marked with "XXX 20011219YSR". 132 // Indeed, as we note elsewhere, this may become yet more slippery 133 // in the presence of multiple CMS and/or multiple VM threads. XXX 134 135 class CMSTokenSync: public StackObj { 136 private: 137 bool _is_cms_thread; 138 public: 139 CMSTokenSync(bool is_cms_thread): 140 _is_cms_thread(is_cms_thread) { 141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 142 "Incorrect argument to constructor"); 143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 144 } 145 146 ~CMSTokenSync() { 147 assert(_is_cms_thread ? 148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 150 "Incorrect state"); 151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 152 } 153 }; 154 155 // Convenience class that does a CMSTokenSync, and then acquires 156 // upto three locks. 157 class CMSTokenSyncWithLocks: public CMSTokenSync { 158 private: 159 // Note: locks are acquired in textual declaration order 160 // and released in the opposite order 161 MutexLockerEx _locker1, _locker2, _locker3; 162 public: 163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 165 CMSTokenSync(is_cms_thread), 166 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 167 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 168 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 169 { } 170 }; 171 172 173 ////////////////////////////////////////////////////////////////// 174 // Concurrent Mark-Sweep Generation ///////////////////////////// 175 ////////////////////////////////////////////////////////////////// 176 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 178 179 // This struct contains per-thread things necessary to support parallel 180 // young-gen collection. 181 class CMSParGCThreadState: public CHeapObj<mtGC> { 182 public: 183 CFLS_LAB lab; 184 PromotionInfo promo; 185 186 // Constructor. 187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 188 promo.setSpace(cfls); 189 } 190 }; 191 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 193 ReservedSpace rs, size_t initial_byte_size, 194 CardTableRS* ct, bool use_adaptive_freelists, 195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 196 CardGeneration(rs, initial_byte_size, ct), 197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 198 _did_compact(false) 199 { 200 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 201 HeapWord* end = (HeapWord*) _virtual_space.high(); 202 203 _direct_allocated_words = 0; 204 NOT_PRODUCT( 205 _numObjectsPromoted = 0; 206 _numWordsPromoted = 0; 207 _numObjectsAllocated = 0; 208 _numWordsAllocated = 0; 209 ) 210 211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 212 use_adaptive_freelists, 213 dictionaryChoice); 214 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 215 _cmsSpace->_gen = this; 216 217 _gc_stats = new CMSGCStats(); 218 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 220 // offsets match. The ability to tell free chunks from objects 221 // depends on this property. 222 debug_only( 223 FreeChunk* junk = NULL; 224 assert(UseCompressedClassPointers || 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 226 "Offset of FreeChunk::_prev within FreeChunk must match" 227 " that of OopDesc::_klass within OopDesc"); 228 ) 229 230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 231 for (uint i = 0; i < ParallelGCThreads; i++) { 232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 233 } 234 235 _incremental_collection_failed = false; 236 // The "dilatation_factor" is the expansion that can occur on 237 // account of the fact that the minimum object size in the CMS 238 // generation may be larger than that in, say, a contiguous young 239 // generation. 240 // Ideally, in the calculation below, we'd compute the dilatation 241 // factor as: MinChunkSize/(promoting_gen's min object size) 242 // Since we do not have such a general query interface for the 243 // promoting generation, we'll instead just use the minimum 244 // object size (which today is a header's worth of space); 245 // note that all arithmetic is in units of HeapWords. 246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 247 assert(_dilatation_factor >= 1.0, "from previous assert"); 248 } 249 250 251 // The field "_initiating_occupancy" represents the occupancy percentage 252 // at which we trigger a new collection cycle. Unless explicitly specified 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it 254 // is calculated by: 255 // 256 // Let "f" be MinHeapFreeRatio in 257 // 258 // _initiating_occupancy = 100-f + 259 // f * (CMSTriggerRatio/100) 260 // where CMSTriggerRatio is the argument "tr" below. 261 // 262 // That is, if we assume the heap is at its desired maximum occupancy at the 263 // end of a collection, we let CMSTriggerRatio of the (purported) free 264 // space be allocated before initiating a new collection cycle. 265 // 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 267 assert(io <= 100 && tr <= 100, "Check the arguments"); 268 if (io >= 0) { 269 _initiating_occupancy = (double)io / 100.0; 270 } else { 271 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 272 (double)(tr * MinHeapFreeRatio) / 100.0) 273 / 100.0; 274 } 275 } 276 277 void ConcurrentMarkSweepGeneration::ref_processor_init() { 278 assert(collector() != NULL, "no collector"); 279 collector()->ref_processor_init(); 280 } 281 282 void CMSCollector::ref_processor_init() { 283 if (_ref_processor == NULL) { 284 // Allocate and initialize a reference processor 285 _ref_processor = 286 new ReferenceProcessor(_span, // span 287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 288 ParallelGCThreads, // mt processing degree 289 _cmsGen->refs_discovery_is_mt(), // mt discovery 290 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 292 &_is_alive_closure); // closure for liveness info 293 // Initialize the _ref_processor field of CMSGen 294 _cmsGen->set_ref_processor(_ref_processor); 295 296 } 297 } 298 299 AdaptiveSizePolicy* CMSCollector::size_policy() { 300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 301 return gch->gen_policy()->size_policy(); 302 } 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 308 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next minor collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 if (PrintGCDetails && Verbose) { 375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 377 cms_free, expected_promotion); 378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 379 cms_free_dbl, cms_consumption_rate() + 1.0); 380 } 381 // Add 1 in case the consumption rate goes to zero. 382 return cms_free_dbl / (cms_consumption_rate() + 1.0); 383 } 384 return 0.0; 385 } 386 387 // Compare the duration of the cms collection to the 388 // time remaining before the cms generation is empty. 389 // Note that the time from the start of the cms collection 390 // to the start of the cms sweep (less than the total 391 // duration of the cms collection) can be used. This 392 // has been tried and some applications experienced 393 // promotion failures early in execution. This was 394 // possibly because the averages were not accurate 395 // enough at the beginning. 396 double CMSStats::time_until_cms_start() const { 397 // We add "gc0_period" to the "work" calculation 398 // below because this query is done (mostly) at the 399 // end of a scavenge, so we need to conservatively 400 // account for that much possible delay 401 // in the query so as to avoid concurrent mode failures 402 // due to starting the collection just a wee bit too 403 // late. 404 double work = cms_duration() + gc0_period(); 405 double deadline = time_until_cms_gen_full(); 406 // If a concurrent mode failure occurred recently, we want to be 407 // more conservative and halve our expected time_until_cms_gen_full() 408 if (work > deadline) { 409 if (Verbose && PrintGCDetails) { 410 gclog_or_tty->print( 411 " CMSCollector: collect because of anticipated promotion " 412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 413 gc0_period(), time_until_cms_gen_full()); 414 } 415 return 0.0; 416 } 417 return work - deadline; 418 } 419 420 #ifndef PRODUCT 421 void CMSStats::print_on(outputStream *st) const { 422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 424 gc0_duration(), gc0_period(), gc0_promoted()); 425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 426 cms_duration(), cms_period(), cms_allocated()); 427 st->print(",cms_since_beg=%g,cms_since_end=%g", 428 cms_time_since_begin(), cms_time_since_end()); 429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 431 432 if (valid()) { 433 st->print(",promo_rate=%g,cms_alloc_rate=%g", 434 promotion_rate(), cms_allocation_rate()); 435 st->print(",cms_consumption_rate=%g,time_until_full=%g", 436 cms_consumption_rate(), time_until_cms_gen_full()); 437 } 438 st->print(" "); 439 } 440 #endif // #ifndef PRODUCT 441 442 CMSCollector::CollectorState CMSCollector::_collectorState = 443 CMSCollector::Idling; 444 bool CMSCollector::_foregroundGCIsActive = false; 445 bool CMSCollector::_foregroundGCShouldWait = false; 446 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 448 CardTableRS* ct, 449 ConcurrentMarkSweepPolicy* cp): 450 _cmsGen(cmsGen), 451 _ct(ct), 452 _ref_processor(NULL), // will be set later 453 _conc_workers(NULL), // may be set later 454 _abort_preclean(false), 455 _start_sampling(false), 456 _between_prologue_and_epilogue(false), 457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 458 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 459 -1 /* lock-free */, "No_lock" /* dummy */), 460 _modUnionClosurePar(&_modUnionTable), 461 // Adjust my span to cover old (cms) gen 462 _span(cmsGen->reserved()), 463 // Construct the is_alive_closure with _span & markBitMap 464 _is_alive_closure(_span, &_markBitMap), 465 _restart_addr(NULL), 466 _overflow_list(NULL), 467 _stats(cmsGen), 468 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 469 //verify that this lock should be acquired with safepoint check. 470 Monitor::_safepoint_check_sometimes)), 471 _eden_chunk_array(NULL), // may be set in ctor body 472 _eden_chunk_capacity(0), // -- ditto -- 473 _eden_chunk_index(0), // -- ditto -- 474 _survivor_plab_array(NULL), // -- ditto -- 475 _survivor_chunk_array(NULL), // -- ditto -- 476 _survivor_chunk_capacity(0), // -- ditto -- 477 _survivor_chunk_index(0), // -- ditto -- 478 _ser_pmc_preclean_ovflw(0), 479 _ser_kac_preclean_ovflw(0), 480 _ser_pmc_remark_ovflw(0), 481 _par_pmc_remark_ovflw(0), 482 _ser_kac_ovflw(0), 483 _par_kac_ovflw(0), 484 #ifndef PRODUCT 485 _num_par_pushes(0), 486 #endif 487 _collection_count_start(0), 488 _verifying(false), 489 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 490 _completed_initialization(false), 491 _collector_policy(cp), 492 _should_unload_classes(CMSClassUnloadingEnabled), 493 _concurrent_cycles_since_last_unload(0), 494 _roots_scanning_options(GenCollectedHeap::SO_None), 495 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 498 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 499 _cms_start_registered(false) 500 { 501 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 502 ExplicitGCInvokesConcurrent = true; 503 } 504 // Now expand the span and allocate the collection support structures 505 // (MUT, marking bit map etc.) to cover both generations subject to 506 // collection. 507 508 // For use by dirty card to oop closures. 509 _cmsGen->cmsSpace()->set_collector(this); 510 511 // Allocate MUT and marking bit map 512 { 513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 514 if (!_markBitMap.allocate(_span)) { 515 warning("Failed to allocate CMS Bit Map"); 516 return; 517 } 518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 519 } 520 { 521 _modUnionTable.allocate(_span); 522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 523 } 524 525 if (!_markStack.allocate(MarkStackSize)) { 526 warning("Failed to allocate CMS Marking Stack"); 527 return; 528 } 529 530 // Support for multi-threaded concurrent phases 531 if (CMSConcurrentMTEnabled) { 532 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 533 // just for now 534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 535 } 536 if (ConcGCThreads > 1) { 537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 538 ConcGCThreads, true); 539 if (_conc_workers == NULL) { 540 warning("GC/CMS: _conc_workers allocation failure: " 541 "forcing -CMSConcurrentMTEnabled"); 542 CMSConcurrentMTEnabled = false; 543 } else { 544 _conc_workers->initialize_workers(); 545 } 546 } else { 547 CMSConcurrentMTEnabled = false; 548 } 549 } 550 if (!CMSConcurrentMTEnabled) { 551 ConcGCThreads = 0; 552 } else { 553 // Turn off CMSCleanOnEnter optimization temporarily for 554 // the MT case where it's not fixed yet; see 6178663. 555 CMSCleanOnEnter = false; 556 } 557 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 558 "Inconsistency"); 559 560 // Parallel task queues; these are shared for the 561 // concurrent and stop-world phases of CMS, but 562 // are not shared with parallel scavenge (ParNew). 563 { 564 uint i; 565 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 566 567 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 568 || ParallelRefProcEnabled) 569 && num_queues > 0) { 570 _task_queues = new OopTaskQueueSet(num_queues); 571 if (_task_queues == NULL) { 572 warning("task_queues allocation failure."); 573 return; 574 } 575 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 576 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 577 for (i = 0; i < num_queues; i++) { 578 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 579 if (q == NULL) { 580 warning("work_queue allocation failure."); 581 return; 582 } 583 _task_queues->register_queue(i, q); 584 } 585 for (i = 0; i < num_queues; i++) { 586 _task_queues->queue(i)->initialize(); 587 _hash_seed[i] = 17; // copied from ParNew 588 } 589 } 590 } 591 592 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 593 594 // Clip CMSBootstrapOccupancy between 0 and 100. 595 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 596 597 // Now tell CMS generations the identity of their collector 598 ConcurrentMarkSweepGeneration::set_collector(this); 599 600 // Create & start a CMS thread for this CMS collector 601 _cmsThread = ConcurrentMarkSweepThread::start(this); 602 assert(cmsThread() != NULL, "CMS Thread should have been created"); 603 assert(cmsThread()->collector() == this, 604 "CMS Thread should refer to this gen"); 605 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 606 607 // Support for parallelizing young gen rescan 608 GenCollectedHeap* gch = GenCollectedHeap::heap(); 609 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 610 _young_gen = (ParNewGeneration*)gch->young_gen(); 611 if (gch->supports_inline_contig_alloc()) { 612 _top_addr = gch->top_addr(); 613 _end_addr = gch->end_addr(); 614 assert(_young_gen != NULL, "no _young_gen"); 615 _eden_chunk_index = 0; 616 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 617 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 618 } 619 620 // Support for parallelizing survivor space rescan 621 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 622 const size_t max_plab_samples = 623 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 624 625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 628 _survivor_chunk_capacity = max_plab_samples; 629 for (uint i = 0; i < ParallelGCThreads; i++) { 630 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 631 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 632 assert(cur->end() == 0, "Should be 0"); 633 assert(cur->array() == vec, "Should be vec"); 634 assert(cur->capacity() == max_plab_samples, "Error"); 635 } 636 } 637 638 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 639 _gc_counters = new CollectorCounters("CMS", 1); 640 _completed_initialization = true; 641 _inter_sweep_timer.start(); // start of time 642 } 643 644 const char* ConcurrentMarkSweepGeneration::name() const { 645 return "concurrent mark-sweep generation"; 646 } 647 void ConcurrentMarkSweepGeneration::update_counters() { 648 if (UsePerfData) { 649 _space_counters->update_all(); 650 _gen_counters->update_all(); 651 } 652 } 653 654 // this is an optimized version of update_counters(). it takes the 655 // used value as a parameter rather than computing it. 656 // 657 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 658 if (UsePerfData) { 659 _space_counters->update_used(used); 660 _space_counters->update_capacity(); 661 _gen_counters->update_all(); 662 } 663 } 664 665 void ConcurrentMarkSweepGeneration::print() const { 666 Generation::print(); 667 cmsSpace()->print(); 668 } 669 670 #ifndef PRODUCT 671 void ConcurrentMarkSweepGeneration::print_statistics() { 672 cmsSpace()->printFLCensus(0); 673 } 674 #endif 675 676 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 677 GenCollectedHeap* gch = GenCollectedHeap::heap(); 678 if (PrintGCDetails) { 679 // I didn't want to change the logging when removing the level concept, 680 // but I guess this logging could say "old" or something instead of "1". 681 assert(gch->is_old_gen(this), 682 "The CMS generation should be the old generation"); 683 uint level = 1; 684 if (Verbose) { 685 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]", 686 level, short_name(), s, used(), capacity()); 687 } else { 688 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]", 689 level, short_name(), s, used() / K, capacity() / K); 690 } 691 } 692 if (Verbose) { 693 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")", 694 gch->used(), gch->capacity()); 695 } else { 696 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)", 697 gch->used() / K, gch->capacity() / K); 698 } 699 } 700 701 size_t 702 ConcurrentMarkSweepGeneration::contiguous_available() const { 703 // dld proposes an improvement in precision here. If the committed 704 // part of the space ends in a free block we should add that to 705 // uncommitted size in the calculation below. Will make this 706 // change later, staying with the approximation below for the 707 // time being. -- ysr. 708 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 709 } 710 711 size_t 712 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 713 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 714 } 715 716 size_t ConcurrentMarkSweepGeneration::max_available() const { 717 return free() + _virtual_space.uncommitted_size(); 718 } 719 720 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 721 size_t available = max_available(); 722 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 723 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 724 if (Verbose && PrintGCDetails) { 725 gclog_or_tty->print_cr( 726 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT ")," 727 "max_promo(" SIZE_FORMAT ")", 728 res? "":" not", available, res? ">=":"<", 729 av_promo, max_promotion_in_bytes); 730 } 731 return res; 732 } 733 734 // At a promotion failure dump information on block layout in heap 735 // (cms old generation). 736 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 737 if (CMSDumpAtPromotionFailure) { 738 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 739 } 740 } 741 742 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 743 // Clear the promotion information. These pointers can be adjusted 744 // along with all the other pointers into the heap but 745 // compaction is expected to be a rare event with 746 // a heap using cms so don't do it without seeing the need. 747 for (uint i = 0; i < ParallelGCThreads; i++) { 748 _par_gc_thread_states[i]->promo.reset(); 749 } 750 } 751 752 void ConcurrentMarkSweepGeneration::compute_new_size() { 753 assert_locked_or_safepoint(Heap_lock); 754 755 // If incremental collection failed, we just want to expand 756 // to the limit. 757 if (incremental_collection_failed()) { 758 clear_incremental_collection_failed(); 759 grow_to_reserved(); 760 return; 761 } 762 763 // The heap has been compacted but not reset yet. 764 // Any metric such as free() or used() will be incorrect. 765 766 CardGeneration::compute_new_size(); 767 768 // Reset again after a possible resizing 769 if (did_compact()) { 770 cmsSpace()->reset_after_compaction(); 771 } 772 } 773 774 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 775 assert_locked_or_safepoint(Heap_lock); 776 777 // If incremental collection failed, we just want to expand 778 // to the limit. 779 if (incremental_collection_failed()) { 780 clear_incremental_collection_failed(); 781 grow_to_reserved(); 782 return; 783 } 784 785 double free_percentage = ((double) free()) / capacity(); 786 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 787 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 788 789 // compute expansion delta needed for reaching desired free percentage 790 if (free_percentage < desired_free_percentage) { 791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 792 assert(desired_capacity >= capacity(), "invalid expansion size"); 793 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 794 if (PrintGCDetails && Verbose) { 795 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 796 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 797 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 798 gclog_or_tty->print_cr(" Desired free fraction %f", 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 GenCollectedHeap* gch = GenCollectedHeap::heap(); 806 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 807 size_t young_size = gch->young_gen()->capacity(); 808 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000); 809 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, 810 unsafe_max_alloc_nogc()/1000); 811 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, 812 contiguous_available()/1000); 813 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", 814 expand_bytes); 815 } 816 // safe if expansion fails 817 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 818 if (PrintGCDetails && Verbose) { 819 gclog_or_tty->print_cr(" Expanded free fraction %f", 820 ((double) free()) / capacity()); 821 } 822 } else { 823 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 824 assert(desired_capacity <= capacity(), "invalid expansion size"); 825 size_t shrink_bytes = capacity() - desired_capacity; 826 // Don't shrink unless the delta is greater than the minimum shrink we want 827 if (shrink_bytes >= MinHeapDeltaBytes) { 828 shrink_free_list_by(shrink_bytes); 829 } 830 } 831 } 832 833 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 834 return cmsSpace()->freelistLock(); 835 } 836 837 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 838 bool tlab) { 839 CMSSynchronousYieldRequest yr; 840 MutexLockerEx x(freelistLock(), 841 Mutex::_no_safepoint_check_flag); 842 return have_lock_and_allocate(size, tlab); 843 } 844 845 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 846 bool tlab /* ignored */) { 847 assert_lock_strong(freelistLock()); 848 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 849 HeapWord* res = cmsSpace()->allocate(adjustedSize); 850 // Allocate the object live (grey) if the background collector has 851 // started marking. This is necessary because the marker may 852 // have passed this address and consequently this object will 853 // not otherwise be greyed and would be incorrectly swept up. 854 // Note that if this object contains references, the writing 855 // of those references will dirty the card containing this object 856 // allowing the object to be blackened (and its references scanned) 857 // either during a preclean phase or at the final checkpoint. 858 if (res != NULL) { 859 // We may block here with an uninitialized object with 860 // its mark-bit or P-bits not yet set. Such objects need 861 // to be safely navigable by block_start(). 862 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 863 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 864 collector()->direct_allocated(res, adjustedSize); 865 _direct_allocated_words += adjustedSize; 866 // allocation counters 867 NOT_PRODUCT( 868 _numObjectsAllocated++; 869 _numWordsAllocated += (int)adjustedSize; 870 ) 871 } 872 return res; 873 } 874 875 // In the case of direct allocation by mutators in a generation that 876 // is being concurrently collected, the object must be allocated 877 // live (grey) if the background collector has started marking. 878 // This is necessary because the marker may 879 // have passed this address and consequently this object will 880 // not otherwise be greyed and would be incorrectly swept up. 881 // Note that if this object contains references, the writing 882 // of those references will dirty the card containing this object 883 // allowing the object to be blackened (and its references scanned) 884 // either during a preclean phase or at the final checkpoint. 885 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 886 assert(_markBitMap.covers(start, size), "Out of bounds"); 887 if (_collectorState >= Marking) { 888 MutexLockerEx y(_markBitMap.lock(), 889 Mutex::_no_safepoint_check_flag); 890 // [see comments preceding SweepClosure::do_blk() below for details] 891 // 892 // Can the P-bits be deleted now? JJJ 893 // 894 // 1. need to mark the object as live so it isn't collected 895 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 896 // 3. need to mark the end of the object so marking, precleaning or sweeping 897 // can skip over uninitialized or unparsable objects. An allocated 898 // object is considered uninitialized for our purposes as long as 899 // its klass word is NULL. All old gen objects are parsable 900 // as soon as they are initialized.) 901 _markBitMap.mark(start); // object is live 902 _markBitMap.mark(start + 1); // object is potentially uninitialized? 903 _markBitMap.mark(start + size - 1); 904 // mark end of object 905 } 906 // check that oop looks uninitialized 907 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 908 } 909 910 void CMSCollector::promoted(bool par, HeapWord* start, 911 bool is_obj_array, size_t obj_size) { 912 assert(_markBitMap.covers(start), "Out of bounds"); 913 // See comment in direct_allocated() about when objects should 914 // be allocated live. 915 if (_collectorState >= Marking) { 916 // we already hold the marking bit map lock, taken in 917 // the prologue 918 if (par) { 919 _markBitMap.par_mark(start); 920 } else { 921 _markBitMap.mark(start); 922 } 923 // We don't need to mark the object as uninitialized (as 924 // in direct_allocated above) because this is being done with the 925 // world stopped and the object will be initialized by the 926 // time the marking, precleaning or sweeping get to look at it. 927 // But see the code for copying objects into the CMS generation, 928 // where we need to ensure that concurrent readers of the 929 // block offset table are able to safely navigate a block that 930 // is in flux from being free to being allocated (and in 931 // transition while being copied into) and subsequently 932 // becoming a bona-fide object when the copy/promotion is complete. 933 assert(SafepointSynchronize::is_at_safepoint(), 934 "expect promotion only at safepoints"); 935 936 if (_collectorState < Sweeping) { 937 // Mark the appropriate cards in the modUnionTable, so that 938 // this object gets scanned before the sweep. If this is 939 // not done, CMS generation references in the object might 940 // not get marked. 941 // For the case of arrays, which are otherwise precisely 942 // marked, we need to dirty the entire array, not just its head. 943 if (is_obj_array) { 944 // The [par_]mark_range() method expects mr.end() below to 945 // be aligned to the granularity of a bit's representation 946 // in the heap. In the case of the MUT below, that's a 947 // card size. 948 MemRegion mr(start, 949 (HeapWord*)round_to((intptr_t)(start + obj_size), 950 CardTableModRefBS::card_size /* bytes */)); 951 if (par) { 952 _modUnionTable.par_mark_range(mr); 953 } else { 954 _modUnionTable.mark_range(mr); 955 } 956 } else { // not an obj array; we can just mark the head 957 if (par) { 958 _modUnionTable.par_mark(start); 959 } else { 960 _modUnionTable.mark(start); 961 } 962 } 963 } 964 } 965 } 966 967 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 968 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 969 // allocate, copy and if necessary update promoinfo -- 970 // delegate to underlying space. 971 assert_lock_strong(freelistLock()); 972 973 #ifndef PRODUCT 974 if (GenCollectedHeap::heap()->promotion_should_fail()) { 975 return NULL; 976 } 977 #endif // #ifndef PRODUCT 978 979 oop res = _cmsSpace->promote(obj, obj_size); 980 if (res == NULL) { 981 // expand and retry 982 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 983 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 984 // Since this is the old generation, we don't try to promote 985 // into a more senior generation. 986 res = _cmsSpace->promote(obj, obj_size); 987 } 988 if (res != NULL) { 989 // See comment in allocate() about when objects should 990 // be allocated live. 991 assert(obj->is_oop(), "Will dereference klass pointer below"); 992 collector()->promoted(false, // Not parallel 993 (HeapWord*)res, obj->is_objArray(), obj_size); 994 // promotion counters 995 NOT_PRODUCT( 996 _numObjectsPromoted++; 997 _numWordsPromoted += 998 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 999 ) 1000 } 1001 return res; 1002 } 1003 1004 1005 // IMPORTANT: Notes on object size recognition in CMS. 1006 // --------------------------------------------------- 1007 // A block of storage in the CMS generation is always in 1008 // one of three states. A free block (FREE), an allocated 1009 // object (OBJECT) whose size() method reports the correct size, 1010 // and an intermediate state (TRANSIENT) in which its size cannot 1011 // be accurately determined. 1012 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1013 // ----------------------------------------------------- 1014 // FREE: klass_word & 1 == 1; mark_word holds block size 1015 // 1016 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1017 // obj->size() computes correct size 1018 // 1019 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1020 // 1021 // STATE IDENTIFICATION: (64 bit+COOPS) 1022 // ------------------------------------ 1023 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1024 // 1025 // OBJECT: klass_word installed; klass_word != 0; 1026 // obj->size() computes correct size 1027 // 1028 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1029 // 1030 // 1031 // STATE TRANSITION DIAGRAM 1032 // 1033 // mut / parnew mut / parnew 1034 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1035 // ^ | 1036 // |------------------------ DEAD <------------------------------------| 1037 // sweep mut 1038 // 1039 // While a block is in TRANSIENT state its size cannot be determined 1040 // so readers will either need to come back later or stall until 1041 // the size can be determined. Note that for the case of direct 1042 // allocation, P-bits, when available, may be used to determine the 1043 // size of an object that may not yet have been initialized. 1044 1045 // Things to support parallel young-gen collection. 1046 oop 1047 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1048 oop old, markOop m, 1049 size_t word_sz) { 1050 #ifndef PRODUCT 1051 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1052 return NULL; 1053 } 1054 #endif // #ifndef PRODUCT 1055 1056 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1057 PromotionInfo* promoInfo = &ps->promo; 1058 // if we are tracking promotions, then first ensure space for 1059 // promotion (including spooling space for saving header if necessary). 1060 // then allocate and copy, then track promoted info if needed. 1061 // When tracking (see PromotionInfo::track()), the mark word may 1062 // be displaced and in this case restoration of the mark word 1063 // occurs in the (oop_since_save_marks_)iterate phase. 1064 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1065 // Out of space for allocating spooling buffers; 1066 // try expanding and allocating spooling buffers. 1067 if (!expand_and_ensure_spooling_space(promoInfo)) { 1068 return NULL; 1069 } 1070 } 1071 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1072 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1073 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1074 if (obj_ptr == NULL) { 1075 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1076 if (obj_ptr == NULL) { 1077 return NULL; 1078 } 1079 } 1080 oop obj = oop(obj_ptr); 1081 OrderAccess::storestore(); 1082 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1083 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1084 // IMPORTANT: See note on object initialization for CMS above. 1085 // Otherwise, copy the object. Here we must be careful to insert the 1086 // klass pointer last, since this marks the block as an allocated object. 1087 // Except with compressed oops it's the mark word. 1088 HeapWord* old_ptr = (HeapWord*)old; 1089 // Restore the mark word copied above. 1090 obj->set_mark(m); 1091 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1092 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1093 OrderAccess::storestore(); 1094 1095 if (UseCompressedClassPointers) { 1096 // Copy gap missed by (aligned) header size calculation below 1097 obj->set_klass_gap(old->klass_gap()); 1098 } 1099 if (word_sz > (size_t)oopDesc::header_size()) { 1100 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1101 obj_ptr + oopDesc::header_size(), 1102 word_sz - oopDesc::header_size()); 1103 } 1104 1105 // Now we can track the promoted object, if necessary. We take care 1106 // to delay the transition from uninitialized to full object 1107 // (i.e., insertion of klass pointer) until after, so that it 1108 // atomically becomes a promoted object. 1109 if (promoInfo->tracking()) { 1110 promoInfo->track((PromotedObject*)obj, old->klass()); 1111 } 1112 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1113 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1114 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1115 1116 // Finally, install the klass pointer (this should be volatile). 1117 OrderAccess::storestore(); 1118 obj->set_klass(old->klass()); 1119 // We should now be able to calculate the right size for this object 1120 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1121 1122 collector()->promoted(true, // parallel 1123 obj_ptr, old->is_objArray(), word_sz); 1124 1125 NOT_PRODUCT( 1126 Atomic::inc_ptr(&_numObjectsPromoted); 1127 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1128 ) 1129 1130 return obj; 1131 } 1132 1133 void 1134 ConcurrentMarkSweepGeneration:: 1135 par_promote_alloc_done(int thread_num) { 1136 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1137 ps->lab.retire(thread_num); 1138 } 1139 1140 void 1141 ConcurrentMarkSweepGeneration:: 1142 par_oop_since_save_marks_iterate_done(int thread_num) { 1143 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1144 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1145 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1146 } 1147 1148 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1149 size_t size, 1150 bool tlab) 1151 { 1152 // We allow a STW collection only if a full 1153 // collection was requested. 1154 return full || should_allocate(size, tlab); // FIX ME !!! 1155 // This and promotion failure handling are connected at the 1156 // hip and should be fixed by untying them. 1157 } 1158 1159 bool CMSCollector::shouldConcurrentCollect() { 1160 if (_full_gc_requested) { 1161 if (Verbose && PrintGCDetails) { 1162 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1163 " gc request (or gc_locker)"); 1164 } 1165 return true; 1166 } 1167 1168 FreelistLocker x(this); 1169 // ------------------------------------------------------------------ 1170 // Print out lots of information which affects the initiation of 1171 // a collection. 1172 if (PrintCMSInitiationStatistics && stats().valid()) { 1173 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1174 gclog_or_tty->stamp(); 1175 gclog_or_tty->cr(); 1176 stats().print_on(gclog_or_tty); 1177 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1178 stats().time_until_cms_gen_full()); 1179 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free()); 1180 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT, 1181 _cmsGen->contiguous_available()); 1182 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1183 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1184 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1185 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1186 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1187 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1188 gclog_or_tty->print_cr("metadata initialized %d", 1189 MetaspaceGC::should_concurrent_collect()); 1190 } 1191 // ------------------------------------------------------------------ 1192 1193 // If the estimated time to complete a cms collection (cms_duration()) 1194 // is less than the estimated time remaining until the cms generation 1195 // is full, start a collection. 1196 if (!UseCMSInitiatingOccupancyOnly) { 1197 if (stats().valid()) { 1198 if (stats().time_until_cms_start() == 0.0) { 1199 return true; 1200 } 1201 } else { 1202 // We want to conservatively collect somewhat early in order 1203 // to try and "bootstrap" our CMS/promotion statistics; 1204 // this branch will not fire after the first successful CMS 1205 // collection because the stats should then be valid. 1206 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1207 if (Verbose && PrintGCDetails) { 1208 gclog_or_tty->print_cr( 1209 " CMSCollector: collect for bootstrapping statistics:" 1210 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1211 _bootstrap_occupancy); 1212 } 1213 return true; 1214 } 1215 } 1216 } 1217 1218 // Otherwise, we start a collection cycle if 1219 // old gen want a collection cycle started. Each may use 1220 // an appropriate criterion for making this decision. 1221 // XXX We need to make sure that the gen expansion 1222 // criterion dovetails well with this. XXX NEED TO FIX THIS 1223 if (_cmsGen->should_concurrent_collect()) { 1224 if (Verbose && PrintGCDetails) { 1225 gclog_or_tty->print_cr("CMS old gen initiated"); 1226 } 1227 return true; 1228 } 1229 1230 // We start a collection if we believe an incremental collection may fail; 1231 // this is not likely to be productive in practice because it's probably too 1232 // late anyway. 1233 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1234 assert(gch->collector_policy()->is_generation_policy(), 1235 "You may want to check the correctness of the following"); 1236 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1237 if (Verbose && PrintGCDetails) { 1238 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1239 } 1240 return true; 1241 } 1242 1243 if (MetaspaceGC::should_concurrent_collect()) { 1244 if (Verbose && PrintGCDetails) { 1245 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1246 } 1247 return true; 1248 } 1249 1250 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1251 if (CMSTriggerInterval >= 0) { 1252 if (CMSTriggerInterval == 0) { 1253 // Trigger always 1254 return true; 1255 } 1256 1257 // Check the CMS time since begin (we do not check the stats validity 1258 // as we want to be able to trigger the first CMS cycle as well) 1259 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1260 if (Verbose && PrintGCDetails) { 1261 if (stats().valid()) { 1262 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1263 stats().cms_time_since_begin()); 1264 } else { 1265 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1266 } 1267 } 1268 return true; 1269 } 1270 } 1271 1272 return false; 1273 } 1274 1275 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1276 1277 // Clear _expansion_cause fields of constituent generations 1278 void CMSCollector::clear_expansion_cause() { 1279 _cmsGen->clear_expansion_cause(); 1280 } 1281 1282 // We should be conservative in starting a collection cycle. To 1283 // start too eagerly runs the risk of collecting too often in the 1284 // extreme. To collect too rarely falls back on full collections, 1285 // which works, even if not optimum in terms of concurrent work. 1286 // As a work around for too eagerly collecting, use the flag 1287 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1288 // giving the user an easily understandable way of controlling the 1289 // collections. 1290 // We want to start a new collection cycle if any of the following 1291 // conditions hold: 1292 // . our current occupancy exceeds the configured initiating occupancy 1293 // for this generation, or 1294 // . we recently needed to expand this space and have not, since that 1295 // expansion, done a collection of this generation, or 1296 // . the underlying space believes that it may be a good idea to initiate 1297 // a concurrent collection (this may be based on criteria such as the 1298 // following: the space uses linear allocation and linear allocation is 1299 // going to fail, or there is believed to be excessive fragmentation in 1300 // the generation, etc... or ... 1301 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1302 // the case of the old generation; see CR 6543076): 1303 // we may be approaching a point at which allocation requests may fail because 1304 // we will be out of sufficient free space given allocation rate estimates.] 1305 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1306 1307 assert_lock_strong(freelistLock()); 1308 if (occupancy() > initiating_occupancy()) { 1309 if (PrintGCDetails && Verbose) { 1310 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1311 short_name(), occupancy(), initiating_occupancy()); 1312 } 1313 return true; 1314 } 1315 if (UseCMSInitiatingOccupancyOnly) { 1316 return false; 1317 } 1318 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1319 if (PrintGCDetails && Verbose) { 1320 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1321 short_name()); 1322 } 1323 return true; 1324 } 1325 if (_cmsSpace->should_concurrent_collect()) { 1326 if (PrintGCDetails && Verbose) { 1327 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1328 short_name()); 1329 } 1330 return true; 1331 } 1332 return false; 1333 } 1334 1335 void ConcurrentMarkSweepGeneration::collect(bool full, 1336 bool clear_all_soft_refs, 1337 size_t size, 1338 bool tlab) 1339 { 1340 collector()->collect(full, clear_all_soft_refs, size, tlab); 1341 } 1342 1343 void CMSCollector::collect(bool full, 1344 bool clear_all_soft_refs, 1345 size_t size, 1346 bool tlab) 1347 { 1348 // The following "if" branch is present for defensive reasons. 1349 // In the current uses of this interface, it can be replaced with: 1350 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1351 // But I am not placing that assert here to allow future 1352 // generality in invoking this interface. 1353 if (GC_locker::is_active()) { 1354 // A consistency test for GC_locker 1355 assert(GC_locker::needs_gc(), "Should have been set already"); 1356 // Skip this foreground collection, instead 1357 // expanding the heap if necessary. 1358 // Need the free list locks for the call to free() in compute_new_size() 1359 compute_new_size(); 1360 return; 1361 } 1362 acquire_control_and_collect(full, clear_all_soft_refs); 1363 } 1364 1365 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1366 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1367 unsigned int gc_count = gch->total_full_collections(); 1368 if (gc_count == full_gc_count) { 1369 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1370 _full_gc_requested = true; 1371 _full_gc_cause = cause; 1372 CGC_lock->notify(); // nudge CMS thread 1373 } else { 1374 assert(gc_count > full_gc_count, "Error: causal loop"); 1375 } 1376 } 1377 1378 bool CMSCollector::is_external_interruption() { 1379 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1380 return GCCause::is_user_requested_gc(cause) || 1381 GCCause::is_serviceability_requested_gc(cause); 1382 } 1383 1384 void CMSCollector::report_concurrent_mode_interruption() { 1385 if (is_external_interruption()) { 1386 if (PrintGCDetails) { 1387 gclog_or_tty->print(" (concurrent mode interrupted)"); 1388 } 1389 } else { 1390 if (PrintGCDetails) { 1391 gclog_or_tty->print(" (concurrent mode failure)"); 1392 } 1393 _gc_tracer_cm->report_concurrent_mode_failure(); 1394 } 1395 } 1396 1397 1398 // The foreground and background collectors need to coordinate in order 1399 // to make sure that they do not mutually interfere with CMS collections. 1400 // When a background collection is active, 1401 // the foreground collector may need to take over (preempt) and 1402 // synchronously complete an ongoing collection. Depending on the 1403 // frequency of the background collections and the heap usage 1404 // of the application, this preemption can be seldom or frequent. 1405 // There are only certain 1406 // points in the background collection that the "collection-baton" 1407 // can be passed to the foreground collector. 1408 // 1409 // The foreground collector will wait for the baton before 1410 // starting any part of the collection. The foreground collector 1411 // will only wait at one location. 1412 // 1413 // The background collector will yield the baton before starting a new 1414 // phase of the collection (e.g., before initial marking, marking from roots, 1415 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1416 // of the loop which switches the phases. The background collector does some 1417 // of the phases (initial mark, final re-mark) with the world stopped. 1418 // Because of locking involved in stopping the world, 1419 // the foreground collector should not block waiting for the background 1420 // collector when it is doing a stop-the-world phase. The background 1421 // collector will yield the baton at an additional point just before 1422 // it enters a stop-the-world phase. Once the world is stopped, the 1423 // background collector checks the phase of the collection. If the 1424 // phase has not changed, it proceeds with the collection. If the 1425 // phase has changed, it skips that phase of the collection. See 1426 // the comments on the use of the Heap_lock in collect_in_background(). 1427 // 1428 // Variable used in baton passing. 1429 // _foregroundGCIsActive - Set to true by the foreground collector when 1430 // it wants the baton. The foreground clears it when it has finished 1431 // the collection. 1432 // _foregroundGCShouldWait - Set to true by the background collector 1433 // when it is running. The foreground collector waits while 1434 // _foregroundGCShouldWait is true. 1435 // CGC_lock - monitor used to protect access to the above variables 1436 // and to notify the foreground and background collectors. 1437 // _collectorState - current state of the CMS collection. 1438 // 1439 // The foreground collector 1440 // acquires the CGC_lock 1441 // sets _foregroundGCIsActive 1442 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1443 // various locks acquired in preparation for the collection 1444 // are released so as not to block the background collector 1445 // that is in the midst of a collection 1446 // proceeds with the collection 1447 // clears _foregroundGCIsActive 1448 // returns 1449 // 1450 // The background collector in a loop iterating on the phases of the 1451 // collection 1452 // acquires the CGC_lock 1453 // sets _foregroundGCShouldWait 1454 // if _foregroundGCIsActive is set 1455 // clears _foregroundGCShouldWait, notifies _CGC_lock 1456 // waits on _CGC_lock for _foregroundGCIsActive to become false 1457 // and exits the loop. 1458 // otherwise 1459 // proceed with that phase of the collection 1460 // if the phase is a stop-the-world phase, 1461 // yield the baton once more just before enqueueing 1462 // the stop-world CMS operation (executed by the VM thread). 1463 // returns after all phases of the collection are done 1464 // 1465 1466 void CMSCollector::acquire_control_and_collect(bool full, 1467 bool clear_all_soft_refs) { 1468 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1469 assert(!Thread::current()->is_ConcurrentGC_thread(), 1470 "shouldn't try to acquire control from self!"); 1471 1472 // Start the protocol for acquiring control of the 1473 // collection from the background collector (aka CMS thread). 1474 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1475 "VM thread should have CMS token"); 1476 // Remember the possibly interrupted state of an ongoing 1477 // concurrent collection 1478 CollectorState first_state = _collectorState; 1479 1480 // Signal to a possibly ongoing concurrent collection that 1481 // we want to do a foreground collection. 1482 _foregroundGCIsActive = true; 1483 1484 // release locks and wait for a notify from the background collector 1485 // releasing the locks in only necessary for phases which 1486 // do yields to improve the granularity of the collection. 1487 assert_lock_strong(bitMapLock()); 1488 // We need to lock the Free list lock for the space that we are 1489 // currently collecting. 1490 assert(haveFreelistLocks(), "Must be holding free list locks"); 1491 bitMapLock()->unlock(); 1492 releaseFreelistLocks(); 1493 { 1494 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1495 if (_foregroundGCShouldWait) { 1496 // We are going to be waiting for action for the CMS thread; 1497 // it had better not be gone (for instance at shutdown)! 1498 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1499 "CMS thread must be running"); 1500 // Wait here until the background collector gives us the go-ahead 1501 ConcurrentMarkSweepThread::clear_CMS_flag( 1502 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1503 // Get a possibly blocked CMS thread going: 1504 // Note that we set _foregroundGCIsActive true above, 1505 // without protection of the CGC_lock. 1506 CGC_lock->notify(); 1507 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1508 "Possible deadlock"); 1509 while (_foregroundGCShouldWait) { 1510 // wait for notification 1511 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1512 // Possibility of delay/starvation here, since CMS token does 1513 // not know to give priority to VM thread? Actually, i think 1514 // there wouldn't be any delay/starvation, but the proof of 1515 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1516 } 1517 ConcurrentMarkSweepThread::set_CMS_flag( 1518 ConcurrentMarkSweepThread::CMS_vm_has_token); 1519 } 1520 } 1521 // The CMS_token is already held. Get back the other locks. 1522 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1523 "VM thread should have CMS token"); 1524 getFreelistLocks(); 1525 bitMapLock()->lock_without_safepoint_check(); 1526 if (TraceCMSState) { 1527 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1528 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1529 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1530 } 1531 1532 // Inform cms gen if this was due to partial collection failing. 1533 // The CMS gen may use this fact to determine its expansion policy. 1534 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1535 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1536 assert(!_cmsGen->incremental_collection_failed(), 1537 "Should have been noticed, reacted to and cleared"); 1538 _cmsGen->set_incremental_collection_failed(); 1539 } 1540 1541 if (first_state > Idling) { 1542 report_concurrent_mode_interruption(); 1543 } 1544 1545 set_did_compact(true); 1546 1547 // If the collection is being acquired from the background 1548 // collector, there may be references on the discovered 1549 // references lists. Abandon those references, since some 1550 // of them may have become unreachable after concurrent 1551 // discovery; the STW compacting collector will redo discovery 1552 // more precisely, without being subject to floating garbage. 1553 // Leaving otherwise unreachable references in the discovered 1554 // lists would require special handling. 1555 ref_processor()->disable_discovery(); 1556 ref_processor()->abandon_partial_discovery(); 1557 ref_processor()->verify_no_references_recorded(); 1558 1559 if (first_state > Idling) { 1560 save_heap_summary(); 1561 } 1562 1563 do_compaction_work(clear_all_soft_refs); 1564 1565 // Has the GC time limit been exceeded? 1566 size_t max_eden_size = _young_gen->max_capacity() - 1567 _young_gen->to()->capacity() - 1568 _young_gen->from()->capacity(); 1569 GCCause::Cause gc_cause = gch->gc_cause(); 1570 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1571 _young_gen->eden()->used(), 1572 _cmsGen->max_capacity(), 1573 max_eden_size, 1574 full, 1575 gc_cause, 1576 gch->collector_policy()); 1577 1578 // Reset the expansion cause, now that we just completed 1579 // a collection cycle. 1580 clear_expansion_cause(); 1581 _foregroundGCIsActive = false; 1582 return; 1583 } 1584 1585 // Resize the tenured generation 1586 // after obtaining the free list locks for the 1587 // two generations. 1588 void CMSCollector::compute_new_size() { 1589 assert_locked_or_safepoint(Heap_lock); 1590 FreelistLocker z(this); 1591 MetaspaceGC::compute_new_size(); 1592 _cmsGen->compute_new_size_free_list(); 1593 } 1594 1595 // A work method used by the foreground collector to do 1596 // a mark-sweep-compact. 1597 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1599 1600 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1601 gc_timer->register_gc_start(); 1602 1603 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1604 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1605 1606 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1607 1608 // Temporarily widen the span of the weak reference processing to 1609 // the entire heap. 1610 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1611 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1612 // Temporarily, clear the "is_alive_non_header" field of the 1613 // reference processor. 1614 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1615 // Temporarily make reference _processing_ single threaded (non-MT). 1616 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1617 // Temporarily make refs discovery atomic 1618 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1619 // Temporarily make reference _discovery_ single threaded (non-MT) 1620 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1621 1622 ref_processor()->set_enqueuing_is_done(false); 1623 ref_processor()->enable_discovery(); 1624 ref_processor()->setup_policy(clear_all_soft_refs); 1625 // If an asynchronous collection finishes, the _modUnionTable is 1626 // all clear. If we are assuming the collection from an asynchronous 1627 // collection, clear the _modUnionTable. 1628 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1629 "_modUnionTable should be clear if the baton was not passed"); 1630 _modUnionTable.clear_all(); 1631 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1632 "mod union for klasses should be clear if the baton was passed"); 1633 _ct->klass_rem_set()->clear_mod_union(); 1634 1635 // We must adjust the allocation statistics being maintained 1636 // in the free list space. We do so by reading and clearing 1637 // the sweep timer and updating the block flux rate estimates below. 1638 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1639 if (_inter_sweep_timer.is_active()) { 1640 _inter_sweep_timer.stop(); 1641 // Note that we do not use this sample to update the _inter_sweep_estimate. 1642 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1643 _inter_sweep_estimate.padded_average(), 1644 _intra_sweep_estimate.padded_average()); 1645 } 1646 1647 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1648 #ifdef ASSERT 1649 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1650 size_t free_size = cms_space->free(); 1651 assert(free_size == 1652 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1653 * HeapWordSize, 1654 "All the free space should be compacted into one chunk at top"); 1655 assert(cms_space->dictionary()->total_chunk_size( 1656 debug_only(cms_space->freelistLock())) == 0 || 1657 cms_space->totalSizeInIndexedFreeLists() == 0, 1658 "All the free space should be in a single chunk"); 1659 size_t num = cms_space->totalCount(); 1660 assert((free_size == 0 && num == 0) || 1661 (free_size > 0 && (num == 1 || num == 2)), 1662 "There should be at most 2 free chunks after compaction"); 1663 #endif // ASSERT 1664 _collectorState = Resetting; 1665 assert(_restart_addr == NULL, 1666 "Should have been NULL'd before baton was passed"); 1667 reset(false /* == !concurrent */); 1668 _cmsGen->reset_after_compaction(); 1669 _concurrent_cycles_since_last_unload = 0; 1670 1671 // Clear any data recorded in the PLAB chunk arrays. 1672 if (_survivor_plab_array != NULL) { 1673 reset_survivor_plab_arrays(); 1674 } 1675 1676 // Adjust the per-size allocation stats for the next epoch. 1677 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1678 // Restart the "inter sweep timer" for the next epoch. 1679 _inter_sweep_timer.reset(); 1680 _inter_sweep_timer.start(); 1681 1682 gc_timer->register_gc_end(); 1683 1684 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1685 1686 // For a mark-sweep-compact, compute_new_size() will be called 1687 // in the heap's do_collection() method. 1688 } 1689 1690 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1691 ContiguousSpace* eden_space = _young_gen->eden(); 1692 ContiguousSpace* from_space = _young_gen->from(); 1693 ContiguousSpace* to_space = _young_gen->to(); 1694 // Eden 1695 if (_eden_chunk_array != NULL) { 1696 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1697 p2i(eden_space->bottom()), p2i(eden_space->top()), 1698 p2i(eden_space->end()), eden_space->capacity()); 1699 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1700 "_eden_chunk_capacity=" SIZE_FORMAT, 1701 _eden_chunk_index, _eden_chunk_capacity); 1702 for (size_t i = 0; i < _eden_chunk_index; i++) { 1703 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1704 i, p2i(_eden_chunk_array[i])); 1705 } 1706 } 1707 // Survivor 1708 if (_survivor_chunk_array != NULL) { 1709 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1710 p2i(from_space->bottom()), p2i(from_space->top()), 1711 p2i(from_space->end()), from_space->capacity()); 1712 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1713 "_survivor_chunk_capacity=" SIZE_FORMAT, 1714 _survivor_chunk_index, _survivor_chunk_capacity); 1715 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1716 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1717 i, p2i(_survivor_chunk_array[i])); 1718 } 1719 } 1720 } 1721 1722 void CMSCollector::getFreelistLocks() const { 1723 // Get locks for all free lists in all generations that this 1724 // collector is responsible for 1725 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1726 } 1727 1728 void CMSCollector::releaseFreelistLocks() const { 1729 // Release locks for all free lists in all generations that this 1730 // collector is responsible for 1731 _cmsGen->freelistLock()->unlock(); 1732 } 1733 1734 bool CMSCollector::haveFreelistLocks() const { 1735 // Check locks for all free lists in all generations that this 1736 // collector is responsible for 1737 assert_lock_strong(_cmsGen->freelistLock()); 1738 PRODUCT_ONLY(ShouldNotReachHere()); 1739 return true; 1740 } 1741 1742 // A utility class that is used by the CMS collector to 1743 // temporarily "release" the foreground collector from its 1744 // usual obligation to wait for the background collector to 1745 // complete an ongoing phase before proceeding. 1746 class ReleaseForegroundGC: public StackObj { 1747 private: 1748 CMSCollector* _c; 1749 public: 1750 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1751 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1752 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1753 // allow a potentially blocked foreground collector to proceed 1754 _c->_foregroundGCShouldWait = false; 1755 if (_c->_foregroundGCIsActive) { 1756 CGC_lock->notify(); 1757 } 1758 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1759 "Possible deadlock"); 1760 } 1761 1762 ~ReleaseForegroundGC() { 1763 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1764 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1765 _c->_foregroundGCShouldWait = true; 1766 } 1767 }; 1768 1769 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1770 assert(Thread::current()->is_ConcurrentGC_thread(), 1771 "A CMS asynchronous collection is only allowed on a CMS thread."); 1772 1773 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1774 { 1775 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1776 MutexLockerEx hl(Heap_lock, safepoint_check); 1777 FreelistLocker fll(this); 1778 MutexLockerEx x(CGC_lock, safepoint_check); 1779 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1780 // The foreground collector is active or we're 1781 // not using asynchronous collections. Skip this 1782 // background collection. 1783 assert(!_foregroundGCShouldWait, "Should be clear"); 1784 return; 1785 } else { 1786 assert(_collectorState == Idling, "Should be idling before start."); 1787 _collectorState = InitialMarking; 1788 register_gc_start(cause); 1789 // Reset the expansion cause, now that we are about to begin 1790 // a new cycle. 1791 clear_expansion_cause(); 1792 1793 // Clear the MetaspaceGC flag since a concurrent collection 1794 // is starting but also clear it after the collection. 1795 MetaspaceGC::set_should_concurrent_collect(false); 1796 } 1797 // Decide if we want to enable class unloading as part of the 1798 // ensuing concurrent GC cycle. 1799 update_should_unload_classes(); 1800 _full_gc_requested = false; // acks all outstanding full gc requests 1801 _full_gc_cause = GCCause::_no_gc; 1802 // Signal that we are about to start a collection 1803 gch->increment_total_full_collections(); // ... starting a collection cycle 1804 _collection_count_start = gch->total_full_collections(); 1805 } 1806 1807 // Used for PrintGC 1808 size_t prev_used; 1809 if (PrintGC && Verbose) { 1810 prev_used = _cmsGen->used(); 1811 } 1812 1813 // The change of the collection state is normally done at this level; 1814 // the exceptions are phases that are executed while the world is 1815 // stopped. For those phases the change of state is done while the 1816 // world is stopped. For baton passing purposes this allows the 1817 // background collector to finish the phase and change state atomically. 1818 // The foreground collector cannot wait on a phase that is done 1819 // while the world is stopped because the foreground collector already 1820 // has the world stopped and would deadlock. 1821 while (_collectorState != Idling) { 1822 if (TraceCMSState) { 1823 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1824 p2i(Thread::current()), _collectorState); 1825 } 1826 // The foreground collector 1827 // holds the Heap_lock throughout its collection. 1828 // holds the CMS token (but not the lock) 1829 // except while it is waiting for the background collector to yield. 1830 // 1831 // The foreground collector should be blocked (not for long) 1832 // if the background collector is about to start a phase 1833 // executed with world stopped. If the background 1834 // collector has already started such a phase, the 1835 // foreground collector is blocked waiting for the 1836 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1837 // are executed in the VM thread. 1838 // 1839 // The locking order is 1840 // PendingListLock (PLL) -- if applicable (FinalMarking) 1841 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1842 // CMS token (claimed in 1843 // stop_world_and_do() --> 1844 // safepoint_synchronize() --> 1845 // CMSThread::synchronize()) 1846 1847 { 1848 // Check if the FG collector wants us to yield. 1849 CMSTokenSync x(true); // is cms thread 1850 if (waitForForegroundGC()) { 1851 // We yielded to a foreground GC, nothing more to be 1852 // done this round. 1853 assert(_foregroundGCShouldWait == false, "We set it to false in " 1854 "waitForForegroundGC()"); 1855 if (TraceCMSState) { 1856 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1857 " exiting collection CMS state %d", 1858 p2i(Thread::current()), _collectorState); 1859 } 1860 return; 1861 } else { 1862 // The background collector can run but check to see if the 1863 // foreground collector has done a collection while the 1864 // background collector was waiting to get the CGC_lock 1865 // above. If yes, break so that _foregroundGCShouldWait 1866 // is cleared before returning. 1867 if (_collectorState == Idling) { 1868 break; 1869 } 1870 } 1871 } 1872 1873 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1874 "should be waiting"); 1875 1876 switch (_collectorState) { 1877 case InitialMarking: 1878 { 1879 ReleaseForegroundGC x(this); 1880 stats().record_cms_begin(); 1881 VM_CMS_Initial_Mark initial_mark_op(this); 1882 VMThread::execute(&initial_mark_op); 1883 } 1884 // The collector state may be any legal state at this point 1885 // since the background collector may have yielded to the 1886 // foreground collector. 1887 break; 1888 case Marking: 1889 // initial marking in checkpointRootsInitialWork has been completed 1890 if (markFromRoots()) { // we were successful 1891 assert(_collectorState == Precleaning, "Collector state should " 1892 "have changed"); 1893 } else { 1894 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1895 } 1896 break; 1897 case Precleaning: 1898 // marking from roots in markFromRoots has been completed 1899 preclean(); 1900 assert(_collectorState == AbortablePreclean || 1901 _collectorState == FinalMarking, 1902 "Collector state should have changed"); 1903 break; 1904 case AbortablePreclean: 1905 abortable_preclean(); 1906 assert(_collectorState == FinalMarking, "Collector state should " 1907 "have changed"); 1908 break; 1909 case FinalMarking: 1910 { 1911 ReleaseForegroundGC x(this); 1912 1913 VM_CMS_Final_Remark final_remark_op(this); 1914 VMThread::execute(&final_remark_op); 1915 } 1916 assert(_foregroundGCShouldWait, "block post-condition"); 1917 break; 1918 case Sweeping: 1919 // final marking in checkpointRootsFinal has been completed 1920 sweep(); 1921 assert(_collectorState == Resizing, "Collector state change " 1922 "to Resizing must be done under the free_list_lock"); 1923 1924 case Resizing: { 1925 // Sweeping has been completed... 1926 // At this point the background collection has completed. 1927 // Don't move the call to compute_new_size() down 1928 // into code that might be executed if the background 1929 // collection was preempted. 1930 { 1931 ReleaseForegroundGC x(this); // unblock FG collection 1932 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1933 CMSTokenSync z(true); // not strictly needed. 1934 if (_collectorState == Resizing) { 1935 compute_new_size(); 1936 save_heap_summary(); 1937 _collectorState = Resetting; 1938 } else { 1939 assert(_collectorState == Idling, "The state should only change" 1940 " because the foreground collector has finished the collection"); 1941 } 1942 } 1943 break; 1944 } 1945 case Resetting: 1946 // CMS heap resizing has been completed 1947 reset(true); 1948 assert(_collectorState == Idling, "Collector state should " 1949 "have changed"); 1950 1951 MetaspaceGC::set_should_concurrent_collect(false); 1952 1953 stats().record_cms_end(); 1954 // Don't move the concurrent_phases_end() and compute_new_size() 1955 // calls to here because a preempted background collection 1956 // has it's state set to "Resetting". 1957 break; 1958 case Idling: 1959 default: 1960 ShouldNotReachHere(); 1961 break; 1962 } 1963 if (TraceCMSState) { 1964 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1965 p2i(Thread::current()), _collectorState); 1966 } 1967 assert(_foregroundGCShouldWait, "block post-condition"); 1968 } 1969 1970 // Should this be in gc_epilogue? 1971 collector_policy()->counters()->update_counters(); 1972 1973 { 1974 // Clear _foregroundGCShouldWait and, in the event that the 1975 // foreground collector is waiting, notify it, before 1976 // returning. 1977 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1978 _foregroundGCShouldWait = false; 1979 if (_foregroundGCIsActive) { 1980 CGC_lock->notify(); 1981 } 1982 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1983 "Possible deadlock"); 1984 } 1985 if (TraceCMSState) { 1986 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1987 " exiting collection CMS state %d", 1988 p2i(Thread::current()), _collectorState); 1989 } 1990 if (PrintGC && Verbose) { 1991 _cmsGen->print_heap_change(prev_used); 1992 } 1993 } 1994 1995 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1996 _cms_start_registered = true; 1997 _gc_timer_cm->register_gc_start(); 1998 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1999 } 2000 2001 void CMSCollector::register_gc_end() { 2002 if (_cms_start_registered) { 2003 report_heap_summary(GCWhen::AfterGC); 2004 2005 _gc_timer_cm->register_gc_end(); 2006 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2007 _cms_start_registered = false; 2008 } 2009 } 2010 2011 void CMSCollector::save_heap_summary() { 2012 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2013 _last_heap_summary = gch->create_heap_summary(); 2014 _last_metaspace_summary = gch->create_metaspace_summary(); 2015 } 2016 2017 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2018 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2019 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2020 } 2021 2022 bool CMSCollector::waitForForegroundGC() { 2023 bool res = false; 2024 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2025 "CMS thread should have CMS token"); 2026 // Block the foreground collector until the 2027 // background collectors decides whether to 2028 // yield. 2029 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2030 _foregroundGCShouldWait = true; 2031 if (_foregroundGCIsActive) { 2032 // The background collector yields to the 2033 // foreground collector and returns a value 2034 // indicating that it has yielded. The foreground 2035 // collector can proceed. 2036 res = true; 2037 _foregroundGCShouldWait = false; 2038 ConcurrentMarkSweepThread::clear_CMS_flag( 2039 ConcurrentMarkSweepThread::CMS_cms_has_token); 2040 ConcurrentMarkSweepThread::set_CMS_flag( 2041 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2042 // Get a possibly blocked foreground thread going 2043 CGC_lock->notify(); 2044 if (TraceCMSState) { 2045 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2046 p2i(Thread::current()), _collectorState); 2047 } 2048 while (_foregroundGCIsActive) { 2049 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2050 } 2051 ConcurrentMarkSweepThread::set_CMS_flag( 2052 ConcurrentMarkSweepThread::CMS_cms_has_token); 2053 ConcurrentMarkSweepThread::clear_CMS_flag( 2054 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2055 } 2056 if (TraceCMSState) { 2057 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2058 p2i(Thread::current()), _collectorState); 2059 } 2060 return res; 2061 } 2062 2063 // Because of the need to lock the free lists and other structures in 2064 // the collector, common to all the generations that the collector is 2065 // collecting, we need the gc_prologues of individual CMS generations 2066 // delegate to their collector. It may have been simpler had the 2067 // current infrastructure allowed one to call a prologue on a 2068 // collector. In the absence of that we have the generation's 2069 // prologue delegate to the collector, which delegates back 2070 // some "local" work to a worker method in the individual generations 2071 // that it's responsible for collecting, while itself doing any 2072 // work common to all generations it's responsible for. A similar 2073 // comment applies to the gc_epilogue()'s. 2074 // The role of the variable _between_prologue_and_epilogue is to 2075 // enforce the invocation protocol. 2076 void CMSCollector::gc_prologue(bool full) { 2077 // Call gc_prologue_work() for the CMSGen 2078 // we are responsible for. 2079 2080 // The following locking discipline assumes that we are only called 2081 // when the world is stopped. 2082 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2083 2084 // The CMSCollector prologue must call the gc_prologues for the 2085 // "generations" that it's responsible 2086 // for. 2087 2088 assert( Thread::current()->is_VM_thread() 2089 || ( CMSScavengeBeforeRemark 2090 && Thread::current()->is_ConcurrentGC_thread()), 2091 "Incorrect thread type for prologue execution"); 2092 2093 if (_between_prologue_and_epilogue) { 2094 // We have already been invoked; this is a gc_prologue delegation 2095 // from yet another CMS generation that we are responsible for, just 2096 // ignore it since all relevant work has already been done. 2097 return; 2098 } 2099 2100 // set a bit saying prologue has been called; cleared in epilogue 2101 _between_prologue_and_epilogue = true; 2102 // Claim locks for common data structures, then call gc_prologue_work() 2103 // for each CMSGen. 2104 2105 getFreelistLocks(); // gets free list locks on constituent spaces 2106 bitMapLock()->lock_without_safepoint_check(); 2107 2108 // Should call gc_prologue_work() for all cms gens we are responsible for 2109 bool duringMarking = _collectorState >= Marking 2110 && _collectorState < Sweeping; 2111 2112 // The young collections clear the modified oops state, which tells if 2113 // there are any modified oops in the class. The remark phase also needs 2114 // that information. Tell the young collection to save the union of all 2115 // modified klasses. 2116 if (duringMarking) { 2117 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2118 } 2119 2120 bool registerClosure = duringMarking; 2121 2122 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2123 2124 if (!full) { 2125 stats().record_gc0_begin(); 2126 } 2127 } 2128 2129 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2130 2131 _capacity_at_prologue = capacity(); 2132 _used_at_prologue = used(); 2133 2134 // Delegate to CMScollector which knows how to coordinate between 2135 // this and any other CMS generations that it is responsible for 2136 // collecting. 2137 collector()->gc_prologue(full); 2138 } 2139 2140 // This is a "private" interface for use by this generation's CMSCollector. 2141 // Not to be called directly by any other entity (for instance, 2142 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2143 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2144 bool registerClosure, ModUnionClosure* modUnionClosure) { 2145 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2146 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2147 "Should be NULL"); 2148 if (registerClosure) { 2149 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2150 } 2151 cmsSpace()->gc_prologue(); 2152 // Clear stat counters 2153 NOT_PRODUCT( 2154 assert(_numObjectsPromoted == 0, "check"); 2155 assert(_numWordsPromoted == 0, "check"); 2156 if (Verbose && PrintGC) { 2157 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, " 2158 SIZE_FORMAT " bytes concurrently", 2159 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2160 } 2161 _numObjectsAllocated = 0; 2162 _numWordsAllocated = 0; 2163 ) 2164 } 2165 2166 void CMSCollector::gc_epilogue(bool full) { 2167 // The following locking discipline assumes that we are only called 2168 // when the world is stopped. 2169 assert(SafepointSynchronize::is_at_safepoint(), 2170 "world is stopped assumption"); 2171 2172 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2173 // if linear allocation blocks need to be appropriately marked to allow the 2174 // the blocks to be parsable. We also check here whether we need to nudge the 2175 // CMS collector thread to start a new cycle (if it's not already active). 2176 assert( Thread::current()->is_VM_thread() 2177 || ( CMSScavengeBeforeRemark 2178 && Thread::current()->is_ConcurrentGC_thread()), 2179 "Incorrect thread type for epilogue execution"); 2180 2181 if (!_between_prologue_and_epilogue) { 2182 // We have already been invoked; this is a gc_epilogue delegation 2183 // from yet another CMS generation that we are responsible for, just 2184 // ignore it since all relevant work has already been done. 2185 return; 2186 } 2187 assert(haveFreelistLocks(), "must have freelist locks"); 2188 assert_lock_strong(bitMapLock()); 2189 2190 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2191 2192 _cmsGen->gc_epilogue_work(full); 2193 2194 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2195 // in case sampling was not already enabled, enable it 2196 _start_sampling = true; 2197 } 2198 // reset _eden_chunk_array so sampling starts afresh 2199 _eden_chunk_index = 0; 2200 2201 size_t cms_used = _cmsGen->cmsSpace()->used(); 2202 2203 // update performance counters - this uses a special version of 2204 // update_counters() that allows the utilization to be passed as a 2205 // parameter, avoiding multiple calls to used(). 2206 // 2207 _cmsGen->update_counters(cms_used); 2208 2209 bitMapLock()->unlock(); 2210 releaseFreelistLocks(); 2211 2212 if (!CleanChunkPoolAsync) { 2213 Chunk::clean_chunk_pool(); 2214 } 2215 2216 set_did_compact(false); 2217 _between_prologue_and_epilogue = false; // ready for next cycle 2218 } 2219 2220 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2221 collector()->gc_epilogue(full); 2222 2223 // Also reset promotion tracking in par gc thread states. 2224 for (uint i = 0; i < ParallelGCThreads; i++) { 2225 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2226 } 2227 } 2228 2229 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2230 assert(!incremental_collection_failed(), "Should have been cleared"); 2231 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2232 cmsSpace()->gc_epilogue(); 2233 // Print stat counters 2234 NOT_PRODUCT( 2235 assert(_numObjectsAllocated == 0, "check"); 2236 assert(_numWordsAllocated == 0, "check"); 2237 if (Verbose && PrintGC) { 2238 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, " 2239 SIZE_FORMAT " bytes", 2240 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2241 } 2242 _numObjectsPromoted = 0; 2243 _numWordsPromoted = 0; 2244 ) 2245 2246 if (PrintGC && Verbose) { 2247 // Call down the chain in contiguous_available needs the freelistLock 2248 // so print this out before releasing the freeListLock. 2249 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ", 2250 contiguous_available()); 2251 } 2252 } 2253 2254 #ifndef PRODUCT 2255 bool CMSCollector::have_cms_token() { 2256 Thread* thr = Thread::current(); 2257 if (thr->is_VM_thread()) { 2258 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2259 } else if (thr->is_ConcurrentGC_thread()) { 2260 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2261 } else if (thr->is_GC_task_thread()) { 2262 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2263 ParGCRareEvent_lock->owned_by_self(); 2264 } 2265 return false; 2266 } 2267 #endif 2268 2269 // Check reachability of the given heap address in CMS generation, 2270 // treating all other generations as roots. 2271 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2272 // We could "guarantee" below, rather than assert, but I'll 2273 // leave these as "asserts" so that an adventurous debugger 2274 // could try this in the product build provided some subset of 2275 // the conditions were met, provided they were interested in the 2276 // results and knew that the computation below wouldn't interfere 2277 // with other concurrent computations mutating the structures 2278 // being read or written. 2279 assert(SafepointSynchronize::is_at_safepoint(), 2280 "Else mutations in object graph will make answer suspect"); 2281 assert(have_cms_token(), "Should hold cms token"); 2282 assert(haveFreelistLocks(), "must hold free list locks"); 2283 assert_lock_strong(bitMapLock()); 2284 2285 // Clear the marking bit map array before starting, but, just 2286 // for kicks, first report if the given address is already marked 2287 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2288 _markBitMap.isMarked(addr) ? "" : " not"); 2289 2290 if (verify_after_remark()) { 2291 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2292 bool result = verification_mark_bm()->isMarked(addr); 2293 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2294 result ? "IS" : "is NOT"); 2295 return result; 2296 } else { 2297 gclog_or_tty->print_cr("Could not compute result"); 2298 return false; 2299 } 2300 } 2301 2302 2303 void 2304 CMSCollector::print_on_error(outputStream* st) { 2305 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2306 if (collector != NULL) { 2307 CMSBitMap* bitmap = &collector->_markBitMap; 2308 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2309 bitmap->print_on_error(st, " Bits: "); 2310 2311 st->cr(); 2312 2313 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2314 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2315 mut_bitmap->print_on_error(st, " Bits: "); 2316 } 2317 } 2318 2319 //////////////////////////////////////////////////////// 2320 // CMS Verification Support 2321 //////////////////////////////////////////////////////// 2322 // Following the remark phase, the following invariant 2323 // should hold -- each object in the CMS heap which is 2324 // marked in markBitMap() should be marked in the verification_mark_bm(). 2325 2326 class VerifyMarkedClosure: public BitMapClosure { 2327 CMSBitMap* _marks; 2328 bool _failed; 2329 2330 public: 2331 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2332 2333 bool do_bit(size_t offset) { 2334 HeapWord* addr = _marks->offsetToHeapWord(offset); 2335 if (!_marks->isMarked(addr)) { 2336 oop(addr)->print_on(gclog_or_tty); 2337 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2338 _failed = true; 2339 } 2340 return true; 2341 } 2342 2343 bool failed() { return _failed; } 2344 }; 2345 2346 bool CMSCollector::verify_after_remark(bool silent) { 2347 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2348 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2349 static bool init = false; 2350 2351 assert(SafepointSynchronize::is_at_safepoint(), 2352 "Else mutations in object graph will make answer suspect"); 2353 assert(have_cms_token(), 2354 "Else there may be mutual interference in use of " 2355 " verification data structures"); 2356 assert(_collectorState > Marking && _collectorState <= Sweeping, 2357 "Else marking info checked here may be obsolete"); 2358 assert(haveFreelistLocks(), "must hold free list locks"); 2359 assert_lock_strong(bitMapLock()); 2360 2361 2362 // Allocate marking bit map if not already allocated 2363 if (!init) { // first time 2364 if (!verification_mark_bm()->allocate(_span)) { 2365 return false; 2366 } 2367 init = true; 2368 } 2369 2370 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2371 2372 // Turn off refs discovery -- so we will be tracing through refs. 2373 // This is as intended, because by this time 2374 // GC must already have cleared any refs that need to be cleared, 2375 // and traced those that need to be marked; moreover, 2376 // the marking done here is not going to interfere in any 2377 // way with the marking information used by GC. 2378 NoRefDiscovery no_discovery(ref_processor()); 2379 2380 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2381 2382 // Clear any marks from a previous round 2383 verification_mark_bm()->clear_all(); 2384 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2385 verify_work_stacks_empty(); 2386 2387 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2388 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2389 // Update the saved marks which may affect the root scans. 2390 gch->save_marks(); 2391 2392 if (CMSRemarkVerifyVariant == 1) { 2393 // In this first variant of verification, we complete 2394 // all marking, then check if the new marks-vector is 2395 // a subset of the CMS marks-vector. 2396 verify_after_remark_work_1(); 2397 } else if (CMSRemarkVerifyVariant == 2) { 2398 // In this second variant of verification, we flag an error 2399 // (i.e. an object reachable in the new marks-vector not reachable 2400 // in the CMS marks-vector) immediately, also indicating the 2401 // identify of an object (A) that references the unmarked object (B) -- 2402 // presumably, a mutation to A failed to be picked up by preclean/remark? 2403 verify_after_remark_work_2(); 2404 } else { 2405 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2406 CMSRemarkVerifyVariant); 2407 } 2408 if (!silent) gclog_or_tty->print(" done] "); 2409 return true; 2410 } 2411 2412 void CMSCollector::verify_after_remark_work_1() { 2413 ResourceMark rm; 2414 HandleMark hm; 2415 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2416 2417 // Get a clear set of claim bits for the roots processing to work with. 2418 ClassLoaderDataGraph::clear_claimed_marks(); 2419 2420 // Mark from roots one level into CMS 2421 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2422 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2423 2424 { 2425 StrongRootsScope srs(1); 2426 2427 gch->gen_process_roots(&srs, 2428 GenCollectedHeap::OldGen, 2429 true, // younger gens are roots 2430 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2431 should_unload_classes(), 2432 ¬Older, 2433 NULL, 2434 NULL); 2435 } 2436 2437 // Now mark from the roots 2438 MarkFromRootsClosure markFromRootsClosure(this, _span, 2439 verification_mark_bm(), verification_mark_stack(), 2440 false /* don't yield */, true /* verifying */); 2441 assert(_restart_addr == NULL, "Expected pre-condition"); 2442 verification_mark_bm()->iterate(&markFromRootsClosure); 2443 while (_restart_addr != NULL) { 2444 // Deal with stack overflow: by restarting at the indicated 2445 // address. 2446 HeapWord* ra = _restart_addr; 2447 markFromRootsClosure.reset(ra); 2448 _restart_addr = NULL; 2449 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2450 } 2451 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2452 verify_work_stacks_empty(); 2453 2454 // Marking completed -- now verify that each bit marked in 2455 // verification_mark_bm() is also marked in markBitMap(); flag all 2456 // errors by printing corresponding objects. 2457 VerifyMarkedClosure vcl(markBitMap()); 2458 verification_mark_bm()->iterate(&vcl); 2459 if (vcl.failed()) { 2460 gclog_or_tty->print("Verification failed"); 2461 gch->print_on(gclog_or_tty); 2462 fatal("CMS: failed marking verification after remark"); 2463 } 2464 } 2465 2466 class VerifyKlassOopsKlassClosure : public KlassClosure { 2467 class VerifyKlassOopsClosure : public OopClosure { 2468 CMSBitMap* _bitmap; 2469 public: 2470 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2471 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2472 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2473 } _oop_closure; 2474 public: 2475 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2476 void do_klass(Klass* k) { 2477 k->oops_do(&_oop_closure); 2478 } 2479 }; 2480 2481 void CMSCollector::verify_after_remark_work_2() { 2482 ResourceMark rm; 2483 HandleMark hm; 2484 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2485 2486 // Get a clear set of claim bits for the roots processing to work with. 2487 ClassLoaderDataGraph::clear_claimed_marks(); 2488 2489 // Mark from roots one level into CMS 2490 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2491 markBitMap()); 2492 CLDToOopClosure cld_closure(¬Older, true); 2493 2494 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2495 2496 { 2497 StrongRootsScope srs(1); 2498 2499 gch->gen_process_roots(&srs, 2500 GenCollectedHeap::OldGen, 2501 true, // younger gens are roots 2502 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2503 should_unload_classes(), 2504 ¬Older, 2505 NULL, 2506 &cld_closure); 2507 } 2508 2509 // Now mark from the roots 2510 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2511 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2512 assert(_restart_addr == NULL, "Expected pre-condition"); 2513 verification_mark_bm()->iterate(&markFromRootsClosure); 2514 while (_restart_addr != NULL) { 2515 // Deal with stack overflow: by restarting at the indicated 2516 // address. 2517 HeapWord* ra = _restart_addr; 2518 markFromRootsClosure.reset(ra); 2519 _restart_addr = NULL; 2520 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2521 } 2522 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2523 verify_work_stacks_empty(); 2524 2525 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2526 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2527 2528 // Marking completed -- now verify that each bit marked in 2529 // verification_mark_bm() is also marked in markBitMap(); flag all 2530 // errors by printing corresponding objects. 2531 VerifyMarkedClosure vcl(markBitMap()); 2532 verification_mark_bm()->iterate(&vcl); 2533 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2534 } 2535 2536 void ConcurrentMarkSweepGeneration::save_marks() { 2537 // delegate to CMS space 2538 cmsSpace()->save_marks(); 2539 for (uint i = 0; i < ParallelGCThreads; i++) { 2540 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2541 } 2542 } 2543 2544 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2545 return cmsSpace()->no_allocs_since_save_marks(); 2546 } 2547 2548 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2549 \ 2550 void ConcurrentMarkSweepGeneration:: \ 2551 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2552 cl->set_generation(this); \ 2553 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2554 cl->reset_generation(); \ 2555 save_marks(); \ 2556 } 2557 2558 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2559 2560 void 2561 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2562 if (freelistLock()->owned_by_self()) { 2563 Generation::oop_iterate(cl); 2564 } else { 2565 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2566 Generation::oop_iterate(cl); 2567 } 2568 } 2569 2570 void 2571 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2572 if (freelistLock()->owned_by_self()) { 2573 Generation::object_iterate(cl); 2574 } else { 2575 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2576 Generation::object_iterate(cl); 2577 } 2578 } 2579 2580 void 2581 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2582 if (freelistLock()->owned_by_self()) { 2583 Generation::safe_object_iterate(cl); 2584 } else { 2585 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2586 Generation::safe_object_iterate(cl); 2587 } 2588 } 2589 2590 void 2591 ConcurrentMarkSweepGeneration::post_compact() { 2592 } 2593 2594 void 2595 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2596 // Fix the linear allocation blocks to look like free blocks. 2597 2598 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2599 // are not called when the heap is verified during universe initialization and 2600 // at vm shutdown. 2601 if (freelistLock()->owned_by_self()) { 2602 cmsSpace()->prepare_for_verify(); 2603 } else { 2604 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2605 cmsSpace()->prepare_for_verify(); 2606 } 2607 } 2608 2609 void 2610 ConcurrentMarkSweepGeneration::verify() { 2611 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2612 // are not called when the heap is verified during universe initialization and 2613 // at vm shutdown. 2614 if (freelistLock()->owned_by_self()) { 2615 cmsSpace()->verify(); 2616 } else { 2617 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2618 cmsSpace()->verify(); 2619 } 2620 } 2621 2622 void CMSCollector::verify() { 2623 _cmsGen->verify(); 2624 } 2625 2626 #ifndef PRODUCT 2627 bool CMSCollector::overflow_list_is_empty() const { 2628 assert(_num_par_pushes >= 0, "Inconsistency"); 2629 if (_overflow_list == NULL) { 2630 assert(_num_par_pushes == 0, "Inconsistency"); 2631 } 2632 return _overflow_list == NULL; 2633 } 2634 2635 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2636 // merely consolidate assertion checks that appear to occur together frequently. 2637 void CMSCollector::verify_work_stacks_empty() const { 2638 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2639 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2640 } 2641 2642 void CMSCollector::verify_overflow_empty() const { 2643 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2644 assert(no_preserved_marks(), "No preserved marks"); 2645 } 2646 #endif // PRODUCT 2647 2648 // Decide if we want to enable class unloading as part of the 2649 // ensuing concurrent GC cycle. We will collect and 2650 // unload classes if it's the case that: 2651 // (1) an explicit gc request has been made and the flag 2652 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2653 // (2) (a) class unloading is enabled at the command line, and 2654 // (b) old gen is getting really full 2655 // NOTE: Provided there is no change in the state of the heap between 2656 // calls to this method, it should have idempotent results. Moreover, 2657 // its results should be monotonically increasing (i.e. going from 0 to 1, 2658 // but not 1 to 0) between successive calls between which the heap was 2659 // not collected. For the implementation below, it must thus rely on 2660 // the property that concurrent_cycles_since_last_unload() 2661 // will not decrease unless a collection cycle happened and that 2662 // _cmsGen->is_too_full() are 2663 // themselves also monotonic in that sense. See check_monotonicity() 2664 // below. 2665 void CMSCollector::update_should_unload_classes() { 2666 _should_unload_classes = false; 2667 // Condition 1 above 2668 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2669 _should_unload_classes = true; 2670 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2671 // Disjuncts 2.b.(i,ii,iii) above 2672 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2673 CMSClassUnloadingMaxInterval) 2674 || _cmsGen->is_too_full(); 2675 } 2676 } 2677 2678 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2679 bool res = should_concurrent_collect(); 2680 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2681 return res; 2682 } 2683 2684 void CMSCollector::setup_cms_unloading_and_verification_state() { 2685 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2686 || VerifyBeforeExit; 2687 const int rso = GenCollectedHeap::SO_AllCodeCache; 2688 2689 // We set the proper root for this CMS cycle here. 2690 if (should_unload_classes()) { // Should unload classes this cycle 2691 remove_root_scanning_option(rso); // Shrink the root set appropriately 2692 set_verifying(should_verify); // Set verification state for this cycle 2693 return; // Nothing else needs to be done at this time 2694 } 2695 2696 // Not unloading classes this cycle 2697 assert(!should_unload_classes(), "Inconsistency!"); 2698 2699 // If we are not unloading classes then add SO_AllCodeCache to root 2700 // scanning options. 2701 add_root_scanning_option(rso); 2702 2703 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2704 set_verifying(true); 2705 } else if (verifying() && !should_verify) { 2706 // We were verifying, but some verification flags got disabled. 2707 set_verifying(false); 2708 // Exclude symbols, strings and code cache elements from root scanning to 2709 // reduce IM and RM pauses. 2710 remove_root_scanning_option(rso); 2711 } 2712 } 2713 2714 2715 #ifndef PRODUCT 2716 HeapWord* CMSCollector::block_start(const void* p) const { 2717 const HeapWord* addr = (HeapWord*)p; 2718 if (_span.contains(p)) { 2719 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2720 return _cmsGen->cmsSpace()->block_start(p); 2721 } 2722 } 2723 return NULL; 2724 } 2725 #endif 2726 2727 HeapWord* 2728 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2729 bool tlab, 2730 bool parallel) { 2731 CMSSynchronousYieldRequest yr; 2732 assert(!tlab, "Can't deal with TLAB allocation"); 2733 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2734 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2735 if (GCExpandToAllocateDelayMillis > 0) { 2736 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2737 } 2738 return have_lock_and_allocate(word_size, tlab); 2739 } 2740 2741 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2742 size_t bytes, 2743 size_t expand_bytes, 2744 CMSExpansionCause::Cause cause) 2745 { 2746 2747 bool success = expand(bytes, expand_bytes); 2748 2749 // remember why we expanded; this information is used 2750 // by shouldConcurrentCollect() when making decisions on whether to start 2751 // a new CMS cycle. 2752 if (success) { 2753 set_expansion_cause(cause); 2754 if (PrintGCDetails && Verbose) { 2755 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2756 CMSExpansionCause::to_string(cause)); 2757 } 2758 } 2759 } 2760 2761 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2762 HeapWord* res = NULL; 2763 MutexLocker x(ParGCRareEvent_lock); 2764 while (true) { 2765 // Expansion by some other thread might make alloc OK now: 2766 res = ps->lab.alloc(word_sz); 2767 if (res != NULL) return res; 2768 // If there's not enough expansion space available, give up. 2769 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2770 return NULL; 2771 } 2772 // Otherwise, we try expansion. 2773 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2774 // Now go around the loop and try alloc again; 2775 // A competing par_promote might beat us to the expansion space, 2776 // so we may go around the loop again if promotion fails again. 2777 if (GCExpandToAllocateDelayMillis > 0) { 2778 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2779 } 2780 } 2781 } 2782 2783 2784 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2785 PromotionInfo* promo) { 2786 MutexLocker x(ParGCRareEvent_lock); 2787 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2788 while (true) { 2789 // Expansion by some other thread might make alloc OK now: 2790 if (promo->ensure_spooling_space()) { 2791 assert(promo->has_spooling_space(), 2792 "Post-condition of successful ensure_spooling_space()"); 2793 return true; 2794 } 2795 // If there's not enough expansion space available, give up. 2796 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2797 return false; 2798 } 2799 // Otherwise, we try expansion. 2800 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2801 // Now go around the loop and try alloc again; 2802 // A competing allocation might beat us to the expansion space, 2803 // so we may go around the loop again if allocation fails again. 2804 if (GCExpandToAllocateDelayMillis > 0) { 2805 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2806 } 2807 } 2808 } 2809 2810 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2811 // Only shrink if a compaction was done so that all the free space 2812 // in the generation is in a contiguous block at the end. 2813 if (did_compact()) { 2814 CardGeneration::shrink(bytes); 2815 } 2816 } 2817 2818 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2819 assert_locked_or_safepoint(Heap_lock); 2820 } 2821 2822 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2823 assert_locked_or_safepoint(Heap_lock); 2824 assert_lock_strong(freelistLock()); 2825 if (PrintGCDetails && Verbose) { 2826 warning("Shrinking of CMS not yet implemented"); 2827 } 2828 return; 2829 } 2830 2831 2832 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2833 // phases. 2834 class CMSPhaseAccounting: public StackObj { 2835 public: 2836 CMSPhaseAccounting(CMSCollector *collector, 2837 const char *phase, 2838 const GCId gc_id, 2839 bool print_cr = true); 2840 ~CMSPhaseAccounting(); 2841 2842 private: 2843 CMSCollector *_collector; 2844 const char *_phase; 2845 elapsedTimer _wallclock; 2846 bool _print_cr; 2847 const GCId _gc_id; 2848 2849 public: 2850 // Not MT-safe; so do not pass around these StackObj's 2851 // where they may be accessed by other threads. 2852 jlong wallclock_millis() { 2853 assert(_wallclock.is_active(), "Wall clock should not stop"); 2854 _wallclock.stop(); // to record time 2855 jlong ret = _wallclock.milliseconds(); 2856 _wallclock.start(); // restart 2857 return ret; 2858 } 2859 }; 2860 2861 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2862 const char *phase, 2863 const GCId gc_id, 2864 bool print_cr) : 2865 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2866 2867 if (PrintCMSStatistics != 0) { 2868 _collector->resetYields(); 2869 } 2870 if (PrintGCDetails) { 2871 gclog_or_tty->gclog_stamp(_gc_id); 2872 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2873 _collector->cmsGen()->short_name(), _phase); 2874 } 2875 _collector->resetTimer(); 2876 _wallclock.start(); 2877 _collector->startTimer(); 2878 } 2879 2880 CMSPhaseAccounting::~CMSPhaseAccounting() { 2881 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2882 _collector->stopTimer(); 2883 _wallclock.stop(); 2884 if (PrintGCDetails) { 2885 gclog_or_tty->gclog_stamp(_gc_id); 2886 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2887 _collector->cmsGen()->short_name(), 2888 _phase, _collector->timerValue(), _wallclock.seconds()); 2889 if (_print_cr) { 2890 gclog_or_tty->cr(); 2891 } 2892 if (PrintCMSStatistics != 0) { 2893 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2894 _collector->yields()); 2895 } 2896 } 2897 } 2898 2899 // CMS work 2900 2901 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2902 class CMSParMarkTask : public AbstractGangTask { 2903 protected: 2904 CMSCollector* _collector; 2905 uint _n_workers; 2906 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2907 AbstractGangTask(name), 2908 _collector(collector), 2909 _n_workers(n_workers) {} 2910 // Work method in support of parallel rescan ... of young gen spaces 2911 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2912 ContiguousSpace* space, 2913 HeapWord** chunk_array, size_t chunk_top); 2914 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2915 }; 2916 2917 // Parallel initial mark task 2918 class CMSParInitialMarkTask: public CMSParMarkTask { 2919 StrongRootsScope* _strong_roots_scope; 2920 public: 2921 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2922 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2923 _strong_roots_scope(strong_roots_scope) {} 2924 void work(uint worker_id); 2925 }; 2926 2927 // Checkpoint the roots into this generation from outside 2928 // this generation. [Note this initial checkpoint need only 2929 // be approximate -- we'll do a catch up phase subsequently.] 2930 void CMSCollector::checkpointRootsInitial() { 2931 assert(_collectorState == InitialMarking, "Wrong collector state"); 2932 check_correct_thread_executing(); 2933 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2934 2935 save_heap_summary(); 2936 report_heap_summary(GCWhen::BeforeGC); 2937 2938 ReferenceProcessor* rp = ref_processor(); 2939 assert(_restart_addr == NULL, "Control point invariant"); 2940 { 2941 // acquire locks for subsequent manipulations 2942 MutexLockerEx x(bitMapLock(), 2943 Mutex::_no_safepoint_check_flag); 2944 checkpointRootsInitialWork(); 2945 // enable ("weak") refs discovery 2946 rp->enable_discovery(); 2947 _collectorState = Marking; 2948 } 2949 } 2950 2951 void CMSCollector::checkpointRootsInitialWork() { 2952 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2953 assert(_collectorState == InitialMarking, "just checking"); 2954 2955 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2956 // precede our marking with a collection of all 2957 // younger generations to keep floating garbage to a minimum. 2958 // XXX: we won't do this for now -- it's an optimization to be done later. 2959 2960 // already have locks 2961 assert_lock_strong(bitMapLock()); 2962 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2963 2964 // Setup the verification and class unloading state for this 2965 // CMS collection cycle. 2966 setup_cms_unloading_and_verification_state(); 2967 2968 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2969 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2970 2971 // Reset all the PLAB chunk arrays if necessary. 2972 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2973 reset_survivor_plab_arrays(); 2974 } 2975 2976 ResourceMark rm; 2977 HandleMark hm; 2978 2979 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2980 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2981 2982 verify_work_stacks_empty(); 2983 verify_overflow_empty(); 2984 2985 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2986 // Update the saved marks which may affect the root scans. 2987 gch->save_marks(); 2988 2989 // weak reference processing has not started yet. 2990 ref_processor()->set_enqueuing_is_done(false); 2991 2992 // Need to remember all newly created CLDs, 2993 // so that we can guarantee that the remark finds them. 2994 ClassLoaderDataGraph::remember_new_clds(true); 2995 2996 // Whenever a CLD is found, it will be claimed before proceeding to mark 2997 // the klasses. The claimed marks need to be cleared before marking starts. 2998 ClassLoaderDataGraph::clear_claimed_marks(); 2999 3000 if (CMSPrintEdenSurvivorChunks) { 3001 print_eden_and_survivor_chunk_arrays(); 3002 } 3003 3004 { 3005 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3006 if (CMSParallelInitialMarkEnabled) { 3007 // The parallel version. 3008 WorkGang* workers = gch->workers(); 3009 assert(workers != NULL, "Need parallel worker threads."); 3010 uint n_workers = workers->active_workers(); 3011 3012 StrongRootsScope srs(n_workers); 3013 3014 CMSParInitialMarkTask tsk(this, &srs, n_workers); 3015 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3016 if (n_workers > 1) { 3017 workers->run_task(&tsk); 3018 } else { 3019 tsk.work(0); 3020 } 3021 } else { 3022 // The serial version. 3023 CLDToOopClosure cld_closure(¬Older, true); 3024 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3025 3026 StrongRootsScope srs(1); 3027 3028 gch->gen_process_roots(&srs, 3029 GenCollectedHeap::OldGen, 3030 true, // younger gens are roots 3031 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3032 should_unload_classes(), 3033 ¬Older, 3034 NULL, 3035 &cld_closure); 3036 } 3037 } 3038 3039 // Clear mod-union table; it will be dirtied in the prologue of 3040 // CMS generation per each younger generation collection. 3041 3042 assert(_modUnionTable.isAllClear(), 3043 "Was cleared in most recent final checkpoint phase" 3044 " or no bits are set in the gc_prologue before the start of the next " 3045 "subsequent marking phase."); 3046 3047 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3048 3049 // Save the end of the used_region of the constituent generations 3050 // to be used to limit the extent of sweep in each generation. 3051 save_sweep_limits(); 3052 verify_overflow_empty(); 3053 } 3054 3055 bool CMSCollector::markFromRoots() { 3056 // we might be tempted to assert that: 3057 // assert(!SafepointSynchronize::is_at_safepoint(), 3058 // "inconsistent argument?"); 3059 // However that wouldn't be right, because it's possible that 3060 // a safepoint is indeed in progress as a younger generation 3061 // stop-the-world GC happens even as we mark in this generation. 3062 assert(_collectorState == Marking, "inconsistent state?"); 3063 check_correct_thread_executing(); 3064 verify_overflow_empty(); 3065 3066 // Weak ref discovery note: We may be discovering weak 3067 // refs in this generation concurrent (but interleaved) with 3068 // weak ref discovery by a younger generation collector. 3069 3070 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3071 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3072 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3073 bool res = markFromRootsWork(); 3074 if (res) { 3075 _collectorState = Precleaning; 3076 } else { // We failed and a foreground collection wants to take over 3077 assert(_foregroundGCIsActive, "internal state inconsistency"); 3078 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3079 if (PrintGCDetails) { 3080 gclog_or_tty->print_cr("bailing out to foreground collection"); 3081 } 3082 } 3083 verify_overflow_empty(); 3084 return res; 3085 } 3086 3087 bool CMSCollector::markFromRootsWork() { 3088 // iterate over marked bits in bit map, doing a full scan and mark 3089 // from these roots using the following algorithm: 3090 // . if oop is to the right of the current scan pointer, 3091 // mark corresponding bit (we'll process it later) 3092 // . else (oop is to left of current scan pointer) 3093 // push oop on marking stack 3094 // . drain the marking stack 3095 3096 // Note that when we do a marking step we need to hold the 3097 // bit map lock -- recall that direct allocation (by mutators) 3098 // and promotion (by younger generation collectors) is also 3099 // marking the bit map. [the so-called allocate live policy.] 3100 // Because the implementation of bit map marking is not 3101 // robust wrt simultaneous marking of bits in the same word, 3102 // we need to make sure that there is no such interference 3103 // between concurrent such updates. 3104 3105 // already have locks 3106 assert_lock_strong(bitMapLock()); 3107 3108 verify_work_stacks_empty(); 3109 verify_overflow_empty(); 3110 bool result = false; 3111 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3112 result = do_marking_mt(); 3113 } else { 3114 result = do_marking_st(); 3115 } 3116 return result; 3117 } 3118 3119 // Forward decl 3120 class CMSConcMarkingTask; 3121 3122 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3123 CMSCollector* _collector; 3124 CMSConcMarkingTask* _task; 3125 public: 3126 virtual void yield(); 3127 3128 // "n_threads" is the number of threads to be terminated. 3129 // "queue_set" is a set of work queues of other threads. 3130 // "collector" is the CMS collector associated with this task terminator. 3131 // "yield" indicates whether we need the gang as a whole to yield. 3132 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3133 ParallelTaskTerminator(n_threads, queue_set), 3134 _collector(collector) { } 3135 3136 void set_task(CMSConcMarkingTask* task) { 3137 _task = task; 3138 } 3139 }; 3140 3141 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3142 CMSConcMarkingTask* _task; 3143 public: 3144 bool should_exit_termination(); 3145 void set_task(CMSConcMarkingTask* task) { 3146 _task = task; 3147 } 3148 }; 3149 3150 // MT Concurrent Marking Task 3151 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3152 CMSCollector* _collector; 3153 uint _n_workers; // requested/desired # workers 3154 bool _result; 3155 CompactibleFreeListSpace* _cms_space; 3156 char _pad_front[64]; // padding to ... 3157 HeapWord* _global_finger; // ... avoid sharing cache line 3158 char _pad_back[64]; 3159 HeapWord* _restart_addr; 3160 3161 // Exposed here for yielding support 3162 Mutex* const _bit_map_lock; 3163 3164 // The per thread work queues, available here for stealing 3165 OopTaskQueueSet* _task_queues; 3166 3167 // Termination (and yielding) support 3168 CMSConcMarkingTerminator _term; 3169 CMSConcMarkingTerminatorTerminator _term_term; 3170 3171 public: 3172 CMSConcMarkingTask(CMSCollector* collector, 3173 CompactibleFreeListSpace* cms_space, 3174 YieldingFlexibleWorkGang* workers, 3175 OopTaskQueueSet* task_queues): 3176 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3177 _collector(collector), 3178 _cms_space(cms_space), 3179 _n_workers(0), _result(true), 3180 _task_queues(task_queues), 3181 _term(_n_workers, task_queues, _collector), 3182 _bit_map_lock(collector->bitMapLock()) 3183 { 3184 _requested_size = _n_workers; 3185 _term.set_task(this); 3186 _term_term.set_task(this); 3187 _restart_addr = _global_finger = _cms_space->bottom(); 3188 } 3189 3190 3191 OopTaskQueueSet* task_queues() { return _task_queues; } 3192 3193 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3194 3195 HeapWord** global_finger_addr() { return &_global_finger; } 3196 3197 CMSConcMarkingTerminator* terminator() { return &_term; } 3198 3199 virtual void set_for_termination(uint active_workers) { 3200 terminator()->reset_for_reuse(active_workers); 3201 } 3202 3203 void work(uint worker_id); 3204 bool should_yield() { 3205 return ConcurrentMarkSweepThread::should_yield() 3206 && !_collector->foregroundGCIsActive(); 3207 } 3208 3209 virtual void coordinator_yield(); // stuff done by coordinator 3210 bool result() { return _result; } 3211 3212 void reset(HeapWord* ra) { 3213 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3214 _restart_addr = _global_finger = ra; 3215 _term.reset_for_reuse(); 3216 } 3217 3218 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3219 OopTaskQueue* work_q); 3220 3221 private: 3222 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3223 void do_work_steal(int i); 3224 void bump_global_finger(HeapWord* f); 3225 }; 3226 3227 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3228 assert(_task != NULL, "Error"); 3229 return _task->yielding(); 3230 // Note that we do not need the disjunct || _task->should_yield() above 3231 // because we want terminating threads to yield only if the task 3232 // is already in the midst of yielding, which happens only after at least one 3233 // thread has yielded. 3234 } 3235 3236 void CMSConcMarkingTerminator::yield() { 3237 if (_task->should_yield()) { 3238 _task->yield(); 3239 } else { 3240 ParallelTaskTerminator::yield(); 3241 } 3242 } 3243 3244 //////////////////////////////////////////////////////////////// 3245 // Concurrent Marking Algorithm Sketch 3246 //////////////////////////////////////////////////////////////// 3247 // Until all tasks exhausted (both spaces): 3248 // -- claim next available chunk 3249 // -- bump global finger via CAS 3250 // -- find first object that starts in this chunk 3251 // and start scanning bitmap from that position 3252 // -- scan marked objects for oops 3253 // -- CAS-mark target, and if successful: 3254 // . if target oop is above global finger (volatile read) 3255 // nothing to do 3256 // . if target oop is in chunk and above local finger 3257 // then nothing to do 3258 // . else push on work-queue 3259 // -- Deal with possible overflow issues: 3260 // . local work-queue overflow causes stuff to be pushed on 3261 // global (common) overflow queue 3262 // . always first empty local work queue 3263 // . then get a batch of oops from global work queue if any 3264 // . then do work stealing 3265 // -- When all tasks claimed (both spaces) 3266 // and local work queue empty, 3267 // then in a loop do: 3268 // . check global overflow stack; steal a batch of oops and trace 3269 // . try to steal from other threads oif GOS is empty 3270 // . if neither is available, offer termination 3271 // -- Terminate and return result 3272 // 3273 void CMSConcMarkingTask::work(uint worker_id) { 3274 elapsedTimer _timer; 3275 ResourceMark rm; 3276 HandleMark hm; 3277 3278 DEBUG_ONLY(_collector->verify_overflow_empty();) 3279 3280 // Before we begin work, our work queue should be empty 3281 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3282 // Scan the bitmap covering _cms_space, tracing through grey objects. 3283 _timer.start(); 3284 do_scan_and_mark(worker_id, _cms_space); 3285 _timer.stop(); 3286 if (PrintCMSStatistics != 0) { 3287 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3288 worker_id, _timer.seconds()); 3289 // XXX: need xxx/xxx type of notation, two timers 3290 } 3291 3292 // ... do work stealing 3293 _timer.reset(); 3294 _timer.start(); 3295 do_work_steal(worker_id); 3296 _timer.stop(); 3297 if (PrintCMSStatistics != 0) { 3298 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3299 worker_id, _timer.seconds()); 3300 // XXX: need xxx/xxx type of notation, two timers 3301 } 3302 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3303 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3304 // Note that under the current task protocol, the 3305 // following assertion is true even of the spaces 3306 // expanded since the completion of the concurrent 3307 // marking. XXX This will likely change under a strict 3308 // ABORT semantics. 3309 // After perm removal the comparison was changed to 3310 // greater than or equal to from strictly greater than. 3311 // Before perm removal the highest address sweep would 3312 // have been at the end of perm gen but now is at the 3313 // end of the tenured gen. 3314 assert(_global_finger >= _cms_space->end(), 3315 "All tasks have been completed"); 3316 DEBUG_ONLY(_collector->verify_overflow_empty();) 3317 } 3318 3319 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3320 HeapWord* read = _global_finger; 3321 HeapWord* cur = read; 3322 while (f > read) { 3323 cur = read; 3324 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3325 if (cur == read) { 3326 // our cas succeeded 3327 assert(_global_finger >= f, "protocol consistency"); 3328 break; 3329 } 3330 } 3331 } 3332 3333 // This is really inefficient, and should be redone by 3334 // using (not yet available) block-read and -write interfaces to the 3335 // stack and the work_queue. XXX FIX ME !!! 3336 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3337 OopTaskQueue* work_q) { 3338 // Fast lock-free check 3339 if (ovflw_stk->length() == 0) { 3340 return false; 3341 } 3342 assert(work_q->size() == 0, "Shouldn't steal"); 3343 MutexLockerEx ml(ovflw_stk->par_lock(), 3344 Mutex::_no_safepoint_check_flag); 3345 // Grab up to 1/4 the size of the work queue 3346 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3347 (size_t)ParGCDesiredObjsFromOverflowList); 3348 num = MIN2(num, ovflw_stk->length()); 3349 for (int i = (int) num; i > 0; i--) { 3350 oop cur = ovflw_stk->pop(); 3351 assert(cur != NULL, "Counted wrong?"); 3352 work_q->push(cur); 3353 } 3354 return num > 0; 3355 } 3356 3357 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3358 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3359 int n_tasks = pst->n_tasks(); 3360 // We allow that there may be no tasks to do here because 3361 // we are restarting after a stack overflow. 3362 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3363 uint nth_task = 0; 3364 3365 HeapWord* aligned_start = sp->bottom(); 3366 if (sp->used_region().contains(_restart_addr)) { 3367 // Align down to a card boundary for the start of 0th task 3368 // for this space. 3369 aligned_start = 3370 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3371 CardTableModRefBS::card_size); 3372 } 3373 3374 size_t chunk_size = sp->marking_task_size(); 3375 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3376 // Having claimed the nth task in this space, 3377 // compute the chunk that it corresponds to: 3378 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3379 aligned_start + (nth_task+1)*chunk_size); 3380 // Try and bump the global finger via a CAS; 3381 // note that we need to do the global finger bump 3382 // _before_ taking the intersection below, because 3383 // the task corresponding to that region will be 3384 // deemed done even if the used_region() expands 3385 // because of allocation -- as it almost certainly will 3386 // during start-up while the threads yield in the 3387 // closure below. 3388 HeapWord* finger = span.end(); 3389 bump_global_finger(finger); // atomically 3390 // There are null tasks here corresponding to chunks 3391 // beyond the "top" address of the space. 3392 span = span.intersection(sp->used_region()); 3393 if (!span.is_empty()) { // Non-null task 3394 HeapWord* prev_obj; 3395 assert(!span.contains(_restart_addr) || nth_task == 0, 3396 "Inconsistency"); 3397 if (nth_task == 0) { 3398 // For the 0th task, we'll not need to compute a block_start. 3399 if (span.contains(_restart_addr)) { 3400 // In the case of a restart because of stack overflow, 3401 // we might additionally skip a chunk prefix. 3402 prev_obj = _restart_addr; 3403 } else { 3404 prev_obj = span.start(); 3405 } 3406 } else { 3407 // We want to skip the first object because 3408 // the protocol is to scan any object in its entirety 3409 // that _starts_ in this span; a fortiori, any 3410 // object starting in an earlier span is scanned 3411 // as part of an earlier claimed task. 3412 // Below we use the "careful" version of block_start 3413 // so we do not try to navigate uninitialized objects. 3414 prev_obj = sp->block_start_careful(span.start()); 3415 // Below we use a variant of block_size that uses the 3416 // Printezis bits to avoid waiting for allocated 3417 // objects to become initialized/parsable. 3418 while (prev_obj < span.start()) { 3419 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3420 if (sz > 0) { 3421 prev_obj += sz; 3422 } else { 3423 // In this case we may end up doing a bit of redundant 3424 // scanning, but that appears unavoidable, short of 3425 // locking the free list locks; see bug 6324141. 3426 break; 3427 } 3428 } 3429 } 3430 if (prev_obj < span.end()) { 3431 MemRegion my_span = MemRegion(prev_obj, span.end()); 3432 // Do the marking work within a non-empty span -- 3433 // the last argument to the constructor indicates whether the 3434 // iteration should be incremental with periodic yields. 3435 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3436 &_collector->_markBitMap, 3437 work_queue(i), 3438 &_collector->_markStack); 3439 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3440 } // else nothing to do for this task 3441 } // else nothing to do for this task 3442 } 3443 // We'd be tempted to assert here that since there are no 3444 // more tasks left to claim in this space, the global_finger 3445 // must exceed space->top() and a fortiori space->end(). However, 3446 // that would not quite be correct because the bumping of 3447 // global_finger occurs strictly after the claiming of a task, 3448 // so by the time we reach here the global finger may not yet 3449 // have been bumped up by the thread that claimed the last 3450 // task. 3451 pst->all_tasks_completed(); 3452 } 3453 3454 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3455 private: 3456 CMSCollector* _collector; 3457 CMSConcMarkingTask* _task; 3458 MemRegion _span; 3459 CMSBitMap* _bit_map; 3460 CMSMarkStack* _overflow_stack; 3461 OopTaskQueue* _work_queue; 3462 protected: 3463 DO_OOP_WORK_DEFN 3464 public: 3465 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3466 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3467 MetadataAwareOopClosure(collector->ref_processor()), 3468 _collector(collector), 3469 _task(task), 3470 _span(collector->_span), 3471 _work_queue(work_queue), 3472 _bit_map(bit_map), 3473 _overflow_stack(overflow_stack) 3474 { } 3475 virtual void do_oop(oop* p); 3476 virtual void do_oop(narrowOop* p); 3477 3478 void trim_queue(size_t max); 3479 void handle_stack_overflow(HeapWord* lost); 3480 void do_yield_check() { 3481 if (_task->should_yield()) { 3482 _task->yield(); 3483 } 3484 } 3485 }; 3486 3487 // Grey object scanning during work stealing phase -- 3488 // the salient assumption here is that any references 3489 // that are in these stolen objects being scanned must 3490 // already have been initialized (else they would not have 3491 // been published), so we do not need to check for 3492 // uninitialized objects before pushing here. 3493 void Par_ConcMarkingClosure::do_oop(oop obj) { 3494 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3495 HeapWord* addr = (HeapWord*)obj; 3496 // Check if oop points into the CMS generation 3497 // and is not marked 3498 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3499 // a white object ... 3500 // If we manage to "claim" the object, by being the 3501 // first thread to mark it, then we push it on our 3502 // marking stack 3503 if (_bit_map->par_mark(addr)) { // ... now grey 3504 // push on work queue (grey set) 3505 bool simulate_overflow = false; 3506 NOT_PRODUCT( 3507 if (CMSMarkStackOverflowALot && 3508 _collector->simulate_overflow()) { 3509 // simulate a stack overflow 3510 simulate_overflow = true; 3511 } 3512 ) 3513 if (simulate_overflow || 3514 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3515 // stack overflow 3516 if (PrintCMSStatistics != 0) { 3517 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3518 SIZE_FORMAT, _overflow_stack->capacity()); 3519 } 3520 // We cannot assert that the overflow stack is full because 3521 // it may have been emptied since. 3522 assert(simulate_overflow || 3523 _work_queue->size() == _work_queue->max_elems(), 3524 "Else push should have succeeded"); 3525 handle_stack_overflow(addr); 3526 } 3527 } // Else, some other thread got there first 3528 do_yield_check(); 3529 } 3530 } 3531 3532 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3533 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3534 3535 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3536 while (_work_queue->size() > max) { 3537 oop new_oop; 3538 if (_work_queue->pop_local(new_oop)) { 3539 assert(new_oop->is_oop(), "Should be an oop"); 3540 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3541 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3542 new_oop->oop_iterate(this); // do_oop() above 3543 do_yield_check(); 3544 } 3545 } 3546 } 3547 3548 // Upon stack overflow, we discard (part of) the stack, 3549 // remembering the least address amongst those discarded 3550 // in CMSCollector's _restart_address. 3551 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3552 // We need to do this under a mutex to prevent other 3553 // workers from interfering with the work done below. 3554 MutexLockerEx ml(_overflow_stack->par_lock(), 3555 Mutex::_no_safepoint_check_flag); 3556 // Remember the least grey address discarded 3557 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3558 _collector->lower_restart_addr(ra); 3559 _overflow_stack->reset(); // discard stack contents 3560 _overflow_stack->expand(); // expand the stack if possible 3561 } 3562 3563 3564 void CMSConcMarkingTask::do_work_steal(int i) { 3565 OopTaskQueue* work_q = work_queue(i); 3566 oop obj_to_scan; 3567 CMSBitMap* bm = &(_collector->_markBitMap); 3568 CMSMarkStack* ovflw = &(_collector->_markStack); 3569 int* seed = _collector->hash_seed(i); 3570 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3571 while (true) { 3572 cl.trim_queue(0); 3573 assert(work_q->size() == 0, "Should have been emptied above"); 3574 if (get_work_from_overflow_stack(ovflw, work_q)) { 3575 // Can't assert below because the work obtained from the 3576 // overflow stack may already have been stolen from us. 3577 // assert(work_q->size() > 0, "Work from overflow stack"); 3578 continue; 3579 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3580 assert(obj_to_scan->is_oop(), "Should be an oop"); 3581 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3582 obj_to_scan->oop_iterate(&cl); 3583 } else if (terminator()->offer_termination(&_term_term)) { 3584 assert(work_q->size() == 0, "Impossible!"); 3585 break; 3586 } else if (yielding() || should_yield()) { 3587 yield(); 3588 } 3589 } 3590 } 3591 3592 // This is run by the CMS (coordinator) thread. 3593 void CMSConcMarkingTask::coordinator_yield() { 3594 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3595 "CMS thread should hold CMS token"); 3596 // First give up the locks, then yield, then re-lock 3597 // We should probably use a constructor/destructor idiom to 3598 // do this unlock/lock or modify the MutexUnlocker class to 3599 // serve our purpose. XXX 3600 assert_lock_strong(_bit_map_lock); 3601 _bit_map_lock->unlock(); 3602 ConcurrentMarkSweepThread::desynchronize(true); 3603 _collector->stopTimer(); 3604 if (PrintCMSStatistics != 0) { 3605 _collector->incrementYields(); 3606 } 3607 3608 // It is possible for whichever thread initiated the yield request 3609 // not to get a chance to wake up and take the bitmap lock between 3610 // this thread releasing it and reacquiring it. So, while the 3611 // should_yield() flag is on, let's sleep for a bit to give the 3612 // other thread a chance to wake up. The limit imposed on the number 3613 // of iterations is defensive, to avoid any unforseen circumstances 3614 // putting us into an infinite loop. Since it's always been this 3615 // (coordinator_yield()) method that was observed to cause the 3616 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3617 // which is by default non-zero. For the other seven methods that 3618 // also perform the yield operation, as are using a different 3619 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3620 // can enable the sleeping for those methods too, if necessary. 3621 // See 6442774. 3622 // 3623 // We really need to reconsider the synchronization between the GC 3624 // thread and the yield-requesting threads in the future and we 3625 // should really use wait/notify, which is the recommended 3626 // way of doing this type of interaction. Additionally, we should 3627 // consolidate the eight methods that do the yield operation and they 3628 // are almost identical into one for better maintainability and 3629 // readability. See 6445193. 3630 // 3631 // Tony 2006.06.29 3632 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3633 ConcurrentMarkSweepThread::should_yield() && 3634 !CMSCollector::foregroundGCIsActive(); ++i) { 3635 os::sleep(Thread::current(), 1, false); 3636 } 3637 3638 ConcurrentMarkSweepThread::synchronize(true); 3639 _bit_map_lock->lock_without_safepoint_check(); 3640 _collector->startTimer(); 3641 } 3642 3643 bool CMSCollector::do_marking_mt() { 3644 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3645 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3646 conc_workers()->active_workers(), 3647 Threads::number_of_non_daemon_threads()); 3648 conc_workers()->set_active_workers(num_workers); 3649 3650 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3651 3652 CMSConcMarkingTask tsk(this, 3653 cms_space, 3654 conc_workers(), 3655 task_queues()); 3656 3657 // Since the actual number of workers we get may be different 3658 // from the number we requested above, do we need to do anything different 3659 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3660 // class?? XXX 3661 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3662 3663 // Refs discovery is already non-atomic. 3664 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3665 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3666 conc_workers()->start_task(&tsk); 3667 while (tsk.yielded()) { 3668 tsk.coordinator_yield(); 3669 conc_workers()->continue_task(&tsk); 3670 } 3671 // If the task was aborted, _restart_addr will be non-NULL 3672 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3673 while (_restart_addr != NULL) { 3674 // XXX For now we do not make use of ABORTED state and have not 3675 // yet implemented the right abort semantics (even in the original 3676 // single-threaded CMS case). That needs some more investigation 3677 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3678 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3679 // If _restart_addr is non-NULL, a marking stack overflow 3680 // occurred; we need to do a fresh marking iteration from the 3681 // indicated restart address. 3682 if (_foregroundGCIsActive) { 3683 // We may be running into repeated stack overflows, having 3684 // reached the limit of the stack size, while making very 3685 // slow forward progress. It may be best to bail out and 3686 // let the foreground collector do its job. 3687 // Clear _restart_addr, so that foreground GC 3688 // works from scratch. This avoids the headache of 3689 // a "rescan" which would otherwise be needed because 3690 // of the dirty mod union table & card table. 3691 _restart_addr = NULL; 3692 return false; 3693 } 3694 // Adjust the task to restart from _restart_addr 3695 tsk.reset(_restart_addr); 3696 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3697 _restart_addr); 3698 _restart_addr = NULL; 3699 // Get the workers going again 3700 conc_workers()->start_task(&tsk); 3701 while (tsk.yielded()) { 3702 tsk.coordinator_yield(); 3703 conc_workers()->continue_task(&tsk); 3704 } 3705 } 3706 assert(tsk.completed(), "Inconsistency"); 3707 assert(tsk.result() == true, "Inconsistency"); 3708 return true; 3709 } 3710 3711 bool CMSCollector::do_marking_st() { 3712 ResourceMark rm; 3713 HandleMark hm; 3714 3715 // Temporarily make refs discovery single threaded (non-MT) 3716 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3717 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3718 &_markStack, CMSYield); 3719 // the last argument to iterate indicates whether the iteration 3720 // should be incremental with periodic yields. 3721 _markBitMap.iterate(&markFromRootsClosure); 3722 // If _restart_addr is non-NULL, a marking stack overflow 3723 // occurred; we need to do a fresh iteration from the 3724 // indicated restart address. 3725 while (_restart_addr != NULL) { 3726 if (_foregroundGCIsActive) { 3727 // We may be running into repeated stack overflows, having 3728 // reached the limit of the stack size, while making very 3729 // slow forward progress. It may be best to bail out and 3730 // let the foreground collector do its job. 3731 // Clear _restart_addr, so that foreground GC 3732 // works from scratch. This avoids the headache of 3733 // a "rescan" which would otherwise be needed because 3734 // of the dirty mod union table & card table. 3735 _restart_addr = NULL; 3736 return false; // indicating failure to complete marking 3737 } 3738 // Deal with stack overflow: 3739 // we restart marking from _restart_addr 3740 HeapWord* ra = _restart_addr; 3741 markFromRootsClosure.reset(ra); 3742 _restart_addr = NULL; 3743 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3744 } 3745 return true; 3746 } 3747 3748 void CMSCollector::preclean() { 3749 check_correct_thread_executing(); 3750 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3751 verify_work_stacks_empty(); 3752 verify_overflow_empty(); 3753 _abort_preclean = false; 3754 if (CMSPrecleaningEnabled) { 3755 if (!CMSEdenChunksRecordAlways) { 3756 _eden_chunk_index = 0; 3757 } 3758 size_t used = get_eden_used(); 3759 size_t capacity = get_eden_capacity(); 3760 // Don't start sampling unless we will get sufficiently 3761 // many samples. 3762 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3763 * CMSScheduleRemarkEdenPenetration)) { 3764 _start_sampling = true; 3765 } else { 3766 _start_sampling = false; 3767 } 3768 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3769 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3770 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3771 } 3772 CMSTokenSync x(true); // is cms thread 3773 if (CMSPrecleaningEnabled) { 3774 sample_eden(); 3775 _collectorState = AbortablePreclean; 3776 } else { 3777 _collectorState = FinalMarking; 3778 } 3779 verify_work_stacks_empty(); 3780 verify_overflow_empty(); 3781 } 3782 3783 // Try and schedule the remark such that young gen 3784 // occupancy is CMSScheduleRemarkEdenPenetration %. 3785 void CMSCollector::abortable_preclean() { 3786 check_correct_thread_executing(); 3787 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3788 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3789 3790 // If Eden's current occupancy is below this threshold, 3791 // immediately schedule the remark; else preclean 3792 // past the next scavenge in an effort to 3793 // schedule the pause as described above. By choosing 3794 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3795 // we will never do an actual abortable preclean cycle. 3796 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3797 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3798 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3799 // We need more smarts in the abortable preclean 3800 // loop below to deal with cases where allocation 3801 // in young gen is very very slow, and our precleaning 3802 // is running a losing race against a horde of 3803 // mutators intent on flooding us with CMS updates 3804 // (dirty cards). 3805 // One, admittedly dumb, strategy is to give up 3806 // after a certain number of abortable precleaning loops 3807 // or after a certain maximum time. We want to make 3808 // this smarter in the next iteration. 3809 // XXX FIX ME!!! YSR 3810 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3811 while (!(should_abort_preclean() || 3812 ConcurrentMarkSweepThread::should_terminate())) { 3813 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3814 cumworkdone += workdone; 3815 loops++; 3816 // Voluntarily terminate abortable preclean phase if we have 3817 // been at it for too long. 3818 if ((CMSMaxAbortablePrecleanLoops != 0) && 3819 loops >= CMSMaxAbortablePrecleanLoops) { 3820 if (PrintGCDetails) { 3821 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3822 } 3823 break; 3824 } 3825 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3826 if (PrintGCDetails) { 3827 gclog_or_tty->print(" CMS: abort preclean due to time "); 3828 } 3829 break; 3830 } 3831 // If we are doing little work each iteration, we should 3832 // take a short break. 3833 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3834 // Sleep for some time, waiting for work to accumulate 3835 stopTimer(); 3836 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3837 startTimer(); 3838 waited++; 3839 } 3840 } 3841 if (PrintCMSStatistics > 0) { 3842 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3843 loops, waited, cumworkdone); 3844 } 3845 } 3846 CMSTokenSync x(true); // is cms thread 3847 if (_collectorState != Idling) { 3848 assert(_collectorState == AbortablePreclean, 3849 "Spontaneous state transition?"); 3850 _collectorState = FinalMarking; 3851 } // Else, a foreground collection completed this CMS cycle. 3852 return; 3853 } 3854 3855 // Respond to an Eden sampling opportunity 3856 void CMSCollector::sample_eden() { 3857 // Make sure a young gc cannot sneak in between our 3858 // reading and recording of a sample. 3859 assert(Thread::current()->is_ConcurrentGC_thread(), 3860 "Only the cms thread may collect Eden samples"); 3861 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3862 "Should collect samples while holding CMS token"); 3863 if (!_start_sampling) { 3864 return; 3865 } 3866 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3867 // is populated by the young generation. 3868 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3869 if (_eden_chunk_index < _eden_chunk_capacity) { 3870 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3871 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3872 "Unexpected state of Eden"); 3873 // We'd like to check that what we just sampled is an oop-start address; 3874 // however, we cannot do that here since the object may not yet have been 3875 // initialized. So we'll instead do the check when we _use_ this sample 3876 // later. 3877 if (_eden_chunk_index == 0 || 3878 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3879 _eden_chunk_array[_eden_chunk_index-1]) 3880 >= CMSSamplingGrain)) { 3881 _eden_chunk_index++; // commit sample 3882 } 3883 } 3884 } 3885 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3886 size_t used = get_eden_used(); 3887 size_t capacity = get_eden_capacity(); 3888 assert(used <= capacity, "Unexpected state of Eden"); 3889 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3890 _abort_preclean = true; 3891 } 3892 } 3893 } 3894 3895 3896 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3897 assert(_collectorState == Precleaning || 3898 _collectorState == AbortablePreclean, "incorrect state"); 3899 ResourceMark rm; 3900 HandleMark hm; 3901 3902 // Precleaning is currently not MT but the reference processor 3903 // may be set for MT. Disable it temporarily here. 3904 ReferenceProcessor* rp = ref_processor(); 3905 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3906 3907 // Do one pass of scrubbing the discovered reference lists 3908 // to remove any reference objects with strongly-reachable 3909 // referents. 3910 if (clean_refs) { 3911 CMSPrecleanRefsYieldClosure yield_cl(this); 3912 assert(rp->span().equals(_span), "Spans should be equal"); 3913 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3914 &_markStack, true /* preclean */); 3915 CMSDrainMarkingStackClosure complete_trace(this, 3916 _span, &_markBitMap, &_markStack, 3917 &keep_alive, true /* preclean */); 3918 3919 // We don't want this step to interfere with a young 3920 // collection because we don't want to take CPU 3921 // or memory bandwidth away from the young GC threads 3922 // (which may be as many as there are CPUs). 3923 // Note that we don't need to protect ourselves from 3924 // interference with mutators because they can't 3925 // manipulate the discovered reference lists nor affect 3926 // the computed reachability of the referents, the 3927 // only properties manipulated by the precleaning 3928 // of these reference lists. 3929 stopTimer(); 3930 CMSTokenSyncWithLocks x(true /* is cms thread */, 3931 bitMapLock()); 3932 startTimer(); 3933 sample_eden(); 3934 3935 // The following will yield to allow foreground 3936 // collection to proceed promptly. XXX YSR: 3937 // The code in this method may need further 3938 // tweaking for better performance and some restructuring 3939 // for cleaner interfaces. 3940 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3941 rp->preclean_discovered_references( 3942 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3943 gc_timer, _gc_tracer_cm->gc_id()); 3944 } 3945 3946 if (clean_survivor) { // preclean the active survivor space(s) 3947 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3948 &_markBitMap, &_modUnionTable, 3949 &_markStack, true /* precleaning phase */); 3950 stopTimer(); 3951 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3952 bitMapLock()); 3953 startTimer(); 3954 unsigned int before_count = 3955 GenCollectedHeap::heap()->total_collections(); 3956 SurvivorSpacePrecleanClosure 3957 sss_cl(this, _span, &_markBitMap, &_markStack, 3958 &pam_cl, before_count, CMSYield); 3959 _young_gen->from()->object_iterate_careful(&sss_cl); 3960 _young_gen->to()->object_iterate_careful(&sss_cl); 3961 } 3962 MarkRefsIntoAndScanClosure 3963 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3964 &_markStack, this, CMSYield, 3965 true /* precleaning phase */); 3966 // CAUTION: The following closure has persistent state that may need to 3967 // be reset upon a decrease in the sequence of addresses it 3968 // processes. 3969 ScanMarkedObjectsAgainCarefullyClosure 3970 smoac_cl(this, _span, 3971 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3972 3973 // Preclean dirty cards in ModUnionTable and CardTable using 3974 // appropriate convergence criterion; 3975 // repeat CMSPrecleanIter times unless we find that 3976 // we are losing. 3977 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3978 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3979 "Bad convergence multiplier"); 3980 assert(CMSPrecleanThreshold >= 100, 3981 "Unreasonably low CMSPrecleanThreshold"); 3982 3983 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3984 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3985 numIter < CMSPrecleanIter; 3986 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3987 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3988 if (Verbose && PrintGCDetails) { 3989 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3990 } 3991 // Either there are very few dirty cards, so re-mark 3992 // pause will be small anyway, or our pre-cleaning isn't 3993 // that much faster than the rate at which cards are being 3994 // dirtied, so we might as well stop and re-mark since 3995 // precleaning won't improve our re-mark time by much. 3996 if (curNumCards <= CMSPrecleanThreshold || 3997 (numIter > 0 && 3998 (curNumCards * CMSPrecleanDenominator > 3999 lastNumCards * CMSPrecleanNumerator))) { 4000 numIter++; 4001 cumNumCards += curNumCards; 4002 break; 4003 } 4004 } 4005 4006 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4007 4008 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4009 cumNumCards += curNumCards; 4010 if (PrintGCDetails && PrintCMSStatistics != 0) { 4011 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4012 curNumCards, cumNumCards, numIter); 4013 } 4014 return cumNumCards; // as a measure of useful work done 4015 } 4016 4017 // PRECLEANING NOTES: 4018 // Precleaning involves: 4019 // . reading the bits of the modUnionTable and clearing the set bits. 4020 // . For the cards corresponding to the set bits, we scan the 4021 // objects on those cards. This means we need the free_list_lock 4022 // so that we can safely iterate over the CMS space when scanning 4023 // for oops. 4024 // . When we scan the objects, we'll be both reading and setting 4025 // marks in the marking bit map, so we'll need the marking bit map. 4026 // . For protecting _collector_state transitions, we take the CGC_lock. 4027 // Note that any races in the reading of of card table entries by the 4028 // CMS thread on the one hand and the clearing of those entries by the 4029 // VM thread or the setting of those entries by the mutator threads on the 4030 // other are quite benign. However, for efficiency it makes sense to keep 4031 // the VM thread from racing with the CMS thread while the latter is 4032 // dirty card info to the modUnionTable. We therefore also use the 4033 // CGC_lock to protect the reading of the card table and the mod union 4034 // table by the CM thread. 4035 // . We run concurrently with mutator updates, so scanning 4036 // needs to be done carefully -- we should not try to scan 4037 // potentially uninitialized objects. 4038 // 4039 // Locking strategy: While holding the CGC_lock, we scan over and 4040 // reset a maximal dirty range of the mod union / card tables, then lock 4041 // the free_list_lock and bitmap lock to do a full marking, then 4042 // release these locks; and repeat the cycle. This allows for a 4043 // certain amount of fairness in the sharing of these locks between 4044 // the CMS collector on the one hand, and the VM thread and the 4045 // mutators on the other. 4046 4047 // NOTE: preclean_mod_union_table() and preclean_card_table() 4048 // further below are largely identical; if you need to modify 4049 // one of these methods, please check the other method too. 4050 4051 size_t CMSCollector::preclean_mod_union_table( 4052 ConcurrentMarkSweepGeneration* gen, 4053 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4054 verify_work_stacks_empty(); 4055 verify_overflow_empty(); 4056 4057 // strategy: starting with the first card, accumulate contiguous 4058 // ranges of dirty cards; clear these cards, then scan the region 4059 // covered by these cards. 4060 4061 // Since all of the MUT is committed ahead, we can just use 4062 // that, in case the generations expand while we are precleaning. 4063 // It might also be fine to just use the committed part of the 4064 // generation, but we might potentially miss cards when the 4065 // generation is rapidly expanding while we are in the midst 4066 // of precleaning. 4067 HeapWord* startAddr = gen->reserved().start(); 4068 HeapWord* endAddr = gen->reserved().end(); 4069 4070 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4071 4072 size_t numDirtyCards, cumNumDirtyCards; 4073 HeapWord *nextAddr, *lastAddr; 4074 for (cumNumDirtyCards = numDirtyCards = 0, 4075 nextAddr = lastAddr = startAddr; 4076 nextAddr < endAddr; 4077 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4078 4079 ResourceMark rm; 4080 HandleMark hm; 4081 4082 MemRegion dirtyRegion; 4083 { 4084 stopTimer(); 4085 // Potential yield point 4086 CMSTokenSync ts(true); 4087 startTimer(); 4088 sample_eden(); 4089 // Get dirty region starting at nextOffset (inclusive), 4090 // simultaneously clearing it. 4091 dirtyRegion = 4092 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4093 assert(dirtyRegion.start() >= nextAddr, 4094 "returned region inconsistent?"); 4095 } 4096 // Remember where the next search should begin. 4097 // The returned region (if non-empty) is a right open interval, 4098 // so lastOffset is obtained from the right end of that 4099 // interval. 4100 lastAddr = dirtyRegion.end(); 4101 // Should do something more transparent and less hacky XXX 4102 numDirtyCards = 4103 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4104 4105 // We'll scan the cards in the dirty region (with periodic 4106 // yields for foreground GC as needed). 4107 if (!dirtyRegion.is_empty()) { 4108 assert(numDirtyCards > 0, "consistency check"); 4109 HeapWord* stop_point = NULL; 4110 stopTimer(); 4111 // Potential yield point 4112 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4113 bitMapLock()); 4114 startTimer(); 4115 { 4116 verify_work_stacks_empty(); 4117 verify_overflow_empty(); 4118 sample_eden(); 4119 stop_point = 4120 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4121 } 4122 if (stop_point != NULL) { 4123 // The careful iteration stopped early either because it found an 4124 // uninitialized object, or because we were in the midst of an 4125 // "abortable preclean", which should now be aborted. Redirty 4126 // the bits corresponding to the partially-scanned or unscanned 4127 // cards. We'll either restart at the next block boundary or 4128 // abort the preclean. 4129 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4130 "Should only be AbortablePreclean."); 4131 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4132 if (should_abort_preclean()) { 4133 break; // out of preclean loop 4134 } else { 4135 // Compute the next address at which preclean should pick up; 4136 // might need bitMapLock in order to read P-bits. 4137 lastAddr = next_card_start_after_block(stop_point); 4138 } 4139 } 4140 } else { 4141 assert(lastAddr == endAddr, "consistency check"); 4142 assert(numDirtyCards == 0, "consistency check"); 4143 break; 4144 } 4145 } 4146 verify_work_stacks_empty(); 4147 verify_overflow_empty(); 4148 return cumNumDirtyCards; 4149 } 4150 4151 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4152 // below are largely identical; if you need to modify 4153 // one of these methods, please check the other method too. 4154 4155 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4156 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4157 // strategy: it's similar to precleamModUnionTable above, in that 4158 // we accumulate contiguous ranges of dirty cards, mark these cards 4159 // precleaned, then scan the region covered by these cards. 4160 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4161 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4162 4163 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4164 4165 size_t numDirtyCards, cumNumDirtyCards; 4166 HeapWord *lastAddr, *nextAddr; 4167 4168 for (cumNumDirtyCards = numDirtyCards = 0, 4169 nextAddr = lastAddr = startAddr; 4170 nextAddr < endAddr; 4171 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4172 4173 ResourceMark rm; 4174 HandleMark hm; 4175 4176 MemRegion dirtyRegion; 4177 { 4178 // See comments in "Precleaning notes" above on why we 4179 // do this locking. XXX Could the locking overheads be 4180 // too high when dirty cards are sparse? [I don't think so.] 4181 stopTimer(); 4182 CMSTokenSync x(true); // is cms thread 4183 startTimer(); 4184 sample_eden(); 4185 // Get and clear dirty region from card table 4186 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4187 MemRegion(nextAddr, endAddr), 4188 true, 4189 CardTableModRefBS::precleaned_card_val()); 4190 4191 assert(dirtyRegion.start() >= nextAddr, 4192 "returned region inconsistent?"); 4193 } 4194 lastAddr = dirtyRegion.end(); 4195 numDirtyCards = 4196 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4197 4198 if (!dirtyRegion.is_empty()) { 4199 stopTimer(); 4200 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4201 startTimer(); 4202 sample_eden(); 4203 verify_work_stacks_empty(); 4204 verify_overflow_empty(); 4205 HeapWord* stop_point = 4206 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4207 if (stop_point != NULL) { 4208 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4209 "Should only be AbortablePreclean."); 4210 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4211 if (should_abort_preclean()) { 4212 break; // out of preclean loop 4213 } else { 4214 // Compute the next address at which preclean should pick up. 4215 lastAddr = next_card_start_after_block(stop_point); 4216 } 4217 } 4218 } else { 4219 break; 4220 } 4221 } 4222 verify_work_stacks_empty(); 4223 verify_overflow_empty(); 4224 return cumNumDirtyCards; 4225 } 4226 4227 class PrecleanKlassClosure : public KlassClosure { 4228 KlassToOopClosure _cm_klass_closure; 4229 public: 4230 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4231 void do_klass(Klass* k) { 4232 if (k->has_accumulated_modified_oops()) { 4233 k->clear_accumulated_modified_oops(); 4234 4235 _cm_klass_closure.do_klass(k); 4236 } 4237 } 4238 }; 4239 4240 // The freelist lock is needed to prevent asserts, is it really needed? 4241 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4242 4243 cl->set_freelistLock(freelistLock); 4244 4245 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4246 4247 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4248 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4249 PrecleanKlassClosure preclean_klass_closure(cl); 4250 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4251 4252 verify_work_stacks_empty(); 4253 verify_overflow_empty(); 4254 } 4255 4256 void CMSCollector::checkpointRootsFinal() { 4257 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4258 check_correct_thread_executing(); 4259 // world is stopped at this checkpoint 4260 assert(SafepointSynchronize::is_at_safepoint(), 4261 "world should be stopped"); 4262 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4263 4264 verify_work_stacks_empty(); 4265 verify_overflow_empty(); 4266 4267 if (PrintGCDetails) { 4268 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]", 4269 _young_gen->used() / K, 4270 _young_gen->capacity() / K); 4271 } 4272 { 4273 if (CMSScavengeBeforeRemark) { 4274 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4275 // Temporarily set flag to false, GCH->do_collection will 4276 // expect it to be false and set to true 4277 FlagSetting fl(gch->_is_gc_active, false); 4278 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4279 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4280 gch->do_collection(true, // full (i.e. force, see below) 4281 false, // !clear_all_soft_refs 4282 0, // size 4283 false, // is_tlab 4284 GenCollectedHeap::YoungGen // type 4285 ); 4286 } 4287 FreelistLocker x(this); 4288 MutexLockerEx y(bitMapLock(), 4289 Mutex::_no_safepoint_check_flag); 4290 checkpointRootsFinalWork(); 4291 } 4292 verify_work_stacks_empty(); 4293 verify_overflow_empty(); 4294 } 4295 4296 void CMSCollector::checkpointRootsFinalWork() { 4297 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4298 4299 assert(haveFreelistLocks(), "must have free list locks"); 4300 assert_lock_strong(bitMapLock()); 4301 4302 ResourceMark rm; 4303 HandleMark hm; 4304 4305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4306 4307 if (should_unload_classes()) { 4308 CodeCache::gc_prologue(); 4309 } 4310 assert(haveFreelistLocks(), "must have free list locks"); 4311 assert_lock_strong(bitMapLock()); 4312 4313 // We might assume that we need not fill TLAB's when 4314 // CMSScavengeBeforeRemark is set, because we may have just done 4315 // a scavenge which would have filled all TLAB's -- and besides 4316 // Eden would be empty. This however may not always be the case -- 4317 // for instance although we asked for a scavenge, it may not have 4318 // happened because of a JNI critical section. We probably need 4319 // a policy for deciding whether we can in that case wait until 4320 // the critical section releases and then do the remark following 4321 // the scavenge, and skip it here. In the absence of that policy, 4322 // or of an indication of whether the scavenge did indeed occur, 4323 // we cannot rely on TLAB's having been filled and must do 4324 // so here just in case a scavenge did not happen. 4325 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4326 // Update the saved marks which may affect the root scans. 4327 gch->save_marks(); 4328 4329 if (CMSPrintEdenSurvivorChunks) { 4330 print_eden_and_survivor_chunk_arrays(); 4331 } 4332 4333 { 4334 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4335 4336 // Note on the role of the mod union table: 4337 // Since the marker in "markFromRoots" marks concurrently with 4338 // mutators, it is possible for some reachable objects not to have been 4339 // scanned. For instance, an only reference to an object A was 4340 // placed in object B after the marker scanned B. Unless B is rescanned, 4341 // A would be collected. Such updates to references in marked objects 4342 // are detected via the mod union table which is the set of all cards 4343 // dirtied since the first checkpoint in this GC cycle and prior to 4344 // the most recent young generation GC, minus those cleaned up by the 4345 // concurrent precleaning. 4346 if (CMSParallelRemarkEnabled) { 4347 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4348 do_remark_parallel(); 4349 } else { 4350 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4351 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4352 do_remark_non_parallel(); 4353 } 4354 } 4355 verify_work_stacks_empty(); 4356 verify_overflow_empty(); 4357 4358 { 4359 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4360 refProcessingWork(); 4361 } 4362 verify_work_stacks_empty(); 4363 verify_overflow_empty(); 4364 4365 if (should_unload_classes()) { 4366 CodeCache::gc_epilogue(); 4367 } 4368 JvmtiExport::gc_epilogue(); 4369 4370 // If we encountered any (marking stack / work queue) overflow 4371 // events during the current CMS cycle, take appropriate 4372 // remedial measures, where possible, so as to try and avoid 4373 // recurrence of that condition. 4374 assert(_markStack.isEmpty(), "No grey objects"); 4375 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4376 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4377 if (ser_ovflw > 0) { 4378 if (PrintCMSStatistics != 0) { 4379 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4380 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT 4381 ", kac_preclean=" SIZE_FORMAT ")", 4382 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4383 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4384 } 4385 _markStack.expand(); 4386 _ser_pmc_remark_ovflw = 0; 4387 _ser_pmc_preclean_ovflw = 0; 4388 _ser_kac_preclean_ovflw = 0; 4389 _ser_kac_ovflw = 0; 4390 } 4391 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4392 if (PrintCMSStatistics != 0) { 4393 gclog_or_tty->print_cr("Work queue overflow (benign) " 4394 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4395 _par_pmc_remark_ovflw, _par_kac_ovflw); 4396 } 4397 _par_pmc_remark_ovflw = 0; 4398 _par_kac_ovflw = 0; 4399 } 4400 if (PrintCMSStatistics != 0) { 4401 if (_markStack._hit_limit > 0) { 4402 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4403 _markStack._hit_limit); 4404 } 4405 if (_markStack._failed_double > 0) { 4406 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT ")," 4407 " current capacity " SIZE_FORMAT, 4408 _markStack._failed_double, 4409 _markStack.capacity()); 4410 } 4411 } 4412 _markStack._hit_limit = 0; 4413 _markStack._failed_double = 0; 4414 4415 if ((VerifyAfterGC || VerifyDuringGC) && 4416 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4417 verify_after_remark(); 4418 } 4419 4420 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4421 4422 // Change under the freelistLocks. 4423 _collectorState = Sweeping; 4424 // Call isAllClear() under bitMapLock 4425 assert(_modUnionTable.isAllClear(), 4426 "Should be clear by end of the final marking"); 4427 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4428 "Should be clear by end of the final marking"); 4429 } 4430 4431 void CMSParInitialMarkTask::work(uint worker_id) { 4432 elapsedTimer _timer; 4433 ResourceMark rm; 4434 HandleMark hm; 4435 4436 // ---------- scan from roots -------------- 4437 _timer.start(); 4438 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4439 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4440 4441 // ---------- young gen roots -------------- 4442 { 4443 work_on_young_gen_roots(worker_id, &par_mri_cl); 4444 _timer.stop(); 4445 if (PrintCMSStatistics != 0) { 4446 gclog_or_tty->print_cr( 4447 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4448 worker_id, _timer.seconds()); 4449 } 4450 } 4451 4452 // ---------- remaining roots -------------- 4453 _timer.reset(); 4454 _timer.start(); 4455 4456 CLDToOopClosure cld_closure(&par_mri_cl, true); 4457 4458 gch->gen_process_roots(_strong_roots_scope, 4459 GenCollectedHeap::OldGen, 4460 false, // yg was scanned above 4461 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4462 _collector->should_unload_classes(), 4463 &par_mri_cl, 4464 NULL, 4465 &cld_closure); 4466 assert(_collector->should_unload_classes() 4467 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4468 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4469 _timer.stop(); 4470 if (PrintCMSStatistics != 0) { 4471 gclog_or_tty->print_cr( 4472 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4473 worker_id, _timer.seconds()); 4474 } 4475 } 4476 4477 // Parallel remark task 4478 class CMSParRemarkTask: public CMSParMarkTask { 4479 CompactibleFreeListSpace* _cms_space; 4480 4481 // The per-thread work queues, available here for stealing. 4482 OopTaskQueueSet* _task_queues; 4483 ParallelTaskTerminator _term; 4484 StrongRootsScope* _strong_roots_scope; 4485 4486 public: 4487 // A value of 0 passed to n_workers will cause the number of 4488 // workers to be taken from the active workers in the work gang. 4489 CMSParRemarkTask(CMSCollector* collector, 4490 CompactibleFreeListSpace* cms_space, 4491 uint n_workers, WorkGang* workers, 4492 OopTaskQueueSet* task_queues, 4493 StrongRootsScope* strong_roots_scope): 4494 CMSParMarkTask("Rescan roots and grey objects in parallel", 4495 collector, n_workers), 4496 _cms_space(cms_space), 4497 _task_queues(task_queues), 4498 _term(n_workers, task_queues), 4499 _strong_roots_scope(strong_roots_scope) { } 4500 4501 OopTaskQueueSet* task_queues() { return _task_queues; } 4502 4503 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4504 4505 ParallelTaskTerminator* terminator() { return &_term; } 4506 uint n_workers() { return _n_workers; } 4507 4508 void work(uint worker_id); 4509 4510 private: 4511 // ... of dirty cards in old space 4512 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4513 Par_MarkRefsIntoAndScanClosure* cl); 4514 4515 // ... work stealing for the above 4516 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4517 }; 4518 4519 class RemarkKlassClosure : public KlassClosure { 4520 KlassToOopClosure _cm_klass_closure; 4521 public: 4522 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4523 void do_klass(Klass* k) { 4524 // Check if we have modified any oops in the Klass during the concurrent marking. 4525 if (k->has_accumulated_modified_oops()) { 4526 k->clear_accumulated_modified_oops(); 4527 4528 // We could have transfered the current modified marks to the accumulated marks, 4529 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4530 } else if (k->has_modified_oops()) { 4531 // Don't clear anything, this info is needed by the next young collection. 4532 } else { 4533 // No modified oops in the Klass. 4534 return; 4535 } 4536 4537 // The klass has modified fields, need to scan the klass. 4538 _cm_klass_closure.do_klass(k); 4539 } 4540 }; 4541 4542 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4543 ParNewGeneration* young_gen = _collector->_young_gen; 4544 ContiguousSpace* eden_space = young_gen->eden(); 4545 ContiguousSpace* from_space = young_gen->from(); 4546 ContiguousSpace* to_space = young_gen->to(); 4547 4548 HeapWord** eca = _collector->_eden_chunk_array; 4549 size_t ect = _collector->_eden_chunk_index; 4550 HeapWord** sca = _collector->_survivor_chunk_array; 4551 size_t sct = _collector->_survivor_chunk_index; 4552 4553 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4554 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4555 4556 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4557 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4558 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4559 } 4560 4561 // work_queue(i) is passed to the closure 4562 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4563 // also is passed to do_dirty_card_rescan_tasks() and to 4564 // do_work_steal() to select the i-th task_queue. 4565 4566 void CMSParRemarkTask::work(uint worker_id) { 4567 elapsedTimer _timer; 4568 ResourceMark rm; 4569 HandleMark hm; 4570 4571 // ---------- rescan from roots -------------- 4572 _timer.start(); 4573 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4574 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4575 _collector->_span, _collector->ref_processor(), 4576 &(_collector->_markBitMap), 4577 work_queue(worker_id)); 4578 4579 // Rescan young gen roots first since these are likely 4580 // coarsely partitioned and may, on that account, constitute 4581 // the critical path; thus, it's best to start off that 4582 // work first. 4583 // ---------- young gen roots -------------- 4584 { 4585 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4586 _timer.stop(); 4587 if (PrintCMSStatistics != 0) { 4588 gclog_or_tty->print_cr( 4589 "Finished young gen rescan work in %dth thread: %3.3f sec", 4590 worker_id, _timer.seconds()); 4591 } 4592 } 4593 4594 // ---------- remaining roots -------------- 4595 _timer.reset(); 4596 _timer.start(); 4597 gch->gen_process_roots(_strong_roots_scope, 4598 GenCollectedHeap::OldGen, 4599 false, // yg was scanned above 4600 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4601 _collector->should_unload_classes(), 4602 &par_mrias_cl, 4603 NULL, 4604 NULL); // The dirty klasses will be handled below 4605 4606 assert(_collector->should_unload_classes() 4607 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4608 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4609 _timer.stop(); 4610 if (PrintCMSStatistics != 0) { 4611 gclog_or_tty->print_cr( 4612 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4613 worker_id, _timer.seconds()); 4614 } 4615 4616 // ---------- unhandled CLD scanning ---------- 4617 if (worker_id == 0) { // Single threaded at the moment. 4618 _timer.reset(); 4619 _timer.start(); 4620 4621 // Scan all new class loader data objects and new dependencies that were 4622 // introduced during concurrent marking. 4623 ResourceMark rm; 4624 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4625 for (int i = 0; i < array->length(); i++) { 4626 par_mrias_cl.do_class_loader_data(array->at(i)); 4627 } 4628 4629 // We don't need to keep track of new CLDs anymore. 4630 ClassLoaderDataGraph::remember_new_clds(false); 4631 4632 _timer.stop(); 4633 if (PrintCMSStatistics != 0) { 4634 gclog_or_tty->print_cr( 4635 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4636 worker_id, _timer.seconds()); 4637 } 4638 } 4639 4640 // ---------- dirty klass scanning ---------- 4641 if (worker_id == 0) { // Single threaded at the moment. 4642 _timer.reset(); 4643 _timer.start(); 4644 4645 // Scan all classes that was dirtied during the concurrent marking phase. 4646 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4647 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4648 4649 _timer.stop(); 4650 if (PrintCMSStatistics != 0) { 4651 gclog_or_tty->print_cr( 4652 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4653 worker_id, _timer.seconds()); 4654 } 4655 } 4656 4657 // We might have added oops to ClassLoaderData::_handles during the 4658 // concurrent marking phase. These oops point to newly allocated objects 4659 // that are guaranteed to be kept alive. Either by the direct allocation 4660 // code, or when the young collector processes the roots. Hence, 4661 // we don't have to revisit the _handles block during the remark phase. 4662 4663 // ---------- rescan dirty cards ------------ 4664 _timer.reset(); 4665 _timer.start(); 4666 4667 // Do the rescan tasks for each of the two spaces 4668 // (cms_space) in turn. 4669 // "worker_id" is passed to select the task_queue for "worker_id" 4670 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4671 _timer.stop(); 4672 if (PrintCMSStatistics != 0) { 4673 gclog_or_tty->print_cr( 4674 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4675 worker_id, _timer.seconds()); 4676 } 4677 4678 // ---------- steal work from other threads ... 4679 // ---------- ... and drain overflow list. 4680 _timer.reset(); 4681 _timer.start(); 4682 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4683 _timer.stop(); 4684 if (PrintCMSStatistics != 0) { 4685 gclog_or_tty->print_cr( 4686 "Finished work stealing in %dth thread: %3.3f sec", 4687 worker_id, _timer.seconds()); 4688 } 4689 } 4690 4691 // Note that parameter "i" is not used. 4692 void 4693 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4694 OopsInGenClosure* cl, ContiguousSpace* space, 4695 HeapWord** chunk_array, size_t chunk_top) { 4696 // Until all tasks completed: 4697 // . claim an unclaimed task 4698 // . compute region boundaries corresponding to task claimed 4699 // using chunk_array 4700 // . par_oop_iterate(cl) over that region 4701 4702 ResourceMark rm; 4703 HandleMark hm; 4704 4705 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4706 4707 uint nth_task = 0; 4708 uint n_tasks = pst->n_tasks(); 4709 4710 if (n_tasks > 0) { 4711 assert(pst->valid(), "Uninitialized use?"); 4712 HeapWord *start, *end; 4713 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4714 // We claimed task # nth_task; compute its boundaries. 4715 if (chunk_top == 0) { // no samples were taken 4716 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4717 start = space->bottom(); 4718 end = space->top(); 4719 } else if (nth_task == 0) { 4720 start = space->bottom(); 4721 end = chunk_array[nth_task]; 4722 } else if (nth_task < (uint)chunk_top) { 4723 assert(nth_task >= 1, "Control point invariant"); 4724 start = chunk_array[nth_task - 1]; 4725 end = chunk_array[nth_task]; 4726 } else { 4727 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4728 start = chunk_array[chunk_top - 1]; 4729 end = space->top(); 4730 } 4731 MemRegion mr(start, end); 4732 // Verify that mr is in space 4733 assert(mr.is_empty() || space->used_region().contains(mr), 4734 "Should be in space"); 4735 // Verify that "start" is an object boundary 4736 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4737 "Should be an oop"); 4738 space->par_oop_iterate(mr, cl); 4739 } 4740 pst->all_tasks_completed(); 4741 } 4742 } 4743 4744 void 4745 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4746 CompactibleFreeListSpace* sp, int i, 4747 Par_MarkRefsIntoAndScanClosure* cl) { 4748 // Until all tasks completed: 4749 // . claim an unclaimed task 4750 // . compute region boundaries corresponding to task claimed 4751 // . transfer dirty bits ct->mut for that region 4752 // . apply rescanclosure to dirty mut bits for that region 4753 4754 ResourceMark rm; 4755 HandleMark hm; 4756 4757 OopTaskQueue* work_q = work_queue(i); 4758 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4759 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4760 // CAUTION: This closure has state that persists across calls to 4761 // the work method dirty_range_iterate_clear() in that it has 4762 // embedded in it a (subtype of) UpwardsObjectClosure. The 4763 // use of that state in the embedded UpwardsObjectClosure instance 4764 // assumes that the cards are always iterated (even if in parallel 4765 // by several threads) in monotonically increasing order per each 4766 // thread. This is true of the implementation below which picks 4767 // card ranges (chunks) in monotonically increasing order globally 4768 // and, a-fortiori, in monotonically increasing order per thread 4769 // (the latter order being a subsequence of the former). 4770 // If the work code below is ever reorganized into a more chaotic 4771 // work-partitioning form than the current "sequential tasks" 4772 // paradigm, the use of that persistent state will have to be 4773 // revisited and modified appropriately. See also related 4774 // bug 4756801 work on which should examine this code to make 4775 // sure that the changes there do not run counter to the 4776 // assumptions made here and necessary for correctness and 4777 // efficiency. Note also that this code might yield inefficient 4778 // behavior in the case of very large objects that span one or 4779 // more work chunks. Such objects would potentially be scanned 4780 // several times redundantly. Work on 4756801 should try and 4781 // address that performance anomaly if at all possible. XXX 4782 MemRegion full_span = _collector->_span; 4783 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4784 MarkFromDirtyCardsClosure 4785 greyRescanClosure(_collector, full_span, // entire span of interest 4786 sp, bm, work_q, cl); 4787 4788 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4789 assert(pst->valid(), "Uninitialized use?"); 4790 uint nth_task = 0; 4791 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4792 MemRegion span = sp->used_region(); 4793 HeapWord* start_addr = span.start(); 4794 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4795 alignment); 4796 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4797 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4798 start_addr, "Check alignment"); 4799 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4800 chunk_size, "Check alignment"); 4801 4802 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4803 // Having claimed the nth_task, compute corresponding mem-region, 4804 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4805 // The alignment restriction ensures that we do not need any 4806 // synchronization with other gang-workers while setting or 4807 // clearing bits in thus chunk of the MUT. 4808 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4809 start_addr + (nth_task+1)*chunk_size); 4810 // The last chunk's end might be way beyond end of the 4811 // used region. In that case pull back appropriately. 4812 if (this_span.end() > end_addr) { 4813 this_span.set_end(end_addr); 4814 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4815 } 4816 // Iterate over the dirty cards covering this chunk, marking them 4817 // precleaned, and setting the corresponding bits in the mod union 4818 // table. Since we have been careful to partition at Card and MUT-word 4819 // boundaries no synchronization is needed between parallel threads. 4820 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4821 &modUnionClosure); 4822 4823 // Having transferred these marks into the modUnionTable, 4824 // rescan the marked objects on the dirty cards in the modUnionTable. 4825 // Even if this is at a synchronous collection, the initial marking 4826 // may have been done during an asynchronous collection so there 4827 // may be dirty bits in the mod-union table. 4828 _collector->_modUnionTable.dirty_range_iterate_clear( 4829 this_span, &greyRescanClosure); 4830 _collector->_modUnionTable.verifyNoOneBitsInRange( 4831 this_span.start(), 4832 this_span.end()); 4833 } 4834 pst->all_tasks_completed(); // declare that i am done 4835 } 4836 4837 // . see if we can share work_queues with ParNew? XXX 4838 void 4839 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4840 int* seed) { 4841 OopTaskQueue* work_q = work_queue(i); 4842 NOT_PRODUCT(int num_steals = 0;) 4843 oop obj_to_scan; 4844 CMSBitMap* bm = &(_collector->_markBitMap); 4845 4846 while (true) { 4847 // Completely finish any left over work from (an) earlier round(s) 4848 cl->trim_queue(0); 4849 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4850 (size_t)ParGCDesiredObjsFromOverflowList); 4851 // Now check if there's any work in the overflow list 4852 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4853 // only affects the number of attempts made to get work from the 4854 // overflow list and does not affect the number of workers. Just 4855 // pass ParallelGCThreads so this behavior is unchanged. 4856 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4857 work_q, 4858 ParallelGCThreads)) { 4859 // found something in global overflow list; 4860 // not yet ready to go stealing work from others. 4861 // We'd like to assert(work_q->size() != 0, ...) 4862 // because we just took work from the overflow list, 4863 // but of course we can't since all of that could have 4864 // been already stolen from us. 4865 // "He giveth and He taketh away." 4866 continue; 4867 } 4868 // Verify that we have no work before we resort to stealing 4869 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4870 // Try to steal from other queues that have work 4871 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4872 NOT_PRODUCT(num_steals++;) 4873 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4874 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4875 // Do scanning work 4876 obj_to_scan->oop_iterate(cl); 4877 // Loop around, finish this work, and try to steal some more 4878 } else if (terminator()->offer_termination()) { 4879 break; // nirvana from the infinite cycle 4880 } 4881 } 4882 NOT_PRODUCT( 4883 if (PrintCMSStatistics != 0) { 4884 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4885 } 4886 ) 4887 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4888 "Else our work is not yet done"); 4889 } 4890 4891 // Record object boundaries in _eden_chunk_array by sampling the eden 4892 // top in the slow-path eden object allocation code path and record 4893 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4894 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4895 // sampling in sample_eden() that activates during the part of the 4896 // preclean phase. 4897 void CMSCollector::sample_eden_chunk() { 4898 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4899 if (_eden_chunk_lock->try_lock()) { 4900 // Record a sample. This is the critical section. The contents 4901 // of the _eden_chunk_array have to be non-decreasing in the 4902 // address order. 4903 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4904 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4905 "Unexpected state of Eden"); 4906 if (_eden_chunk_index == 0 || 4907 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4908 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4909 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4910 _eden_chunk_index++; // commit sample 4911 } 4912 _eden_chunk_lock->unlock(); 4913 } 4914 } 4915 } 4916 4917 // Return a thread-local PLAB recording array, as appropriate. 4918 void* CMSCollector::get_data_recorder(int thr_num) { 4919 if (_survivor_plab_array != NULL && 4920 (CMSPLABRecordAlways || 4921 (_collectorState > Marking && _collectorState < FinalMarking))) { 4922 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4923 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4924 ca->reset(); // clear it so that fresh data is recorded 4925 return (void*) ca; 4926 } else { 4927 return NULL; 4928 } 4929 } 4930 4931 // Reset all the thread-local PLAB recording arrays 4932 void CMSCollector::reset_survivor_plab_arrays() { 4933 for (uint i = 0; i < ParallelGCThreads; i++) { 4934 _survivor_plab_array[i].reset(); 4935 } 4936 } 4937 4938 // Merge the per-thread plab arrays into the global survivor chunk 4939 // array which will provide the partitioning of the survivor space 4940 // for CMS initial scan and rescan. 4941 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4942 int no_of_gc_threads) { 4943 assert(_survivor_plab_array != NULL, "Error"); 4944 assert(_survivor_chunk_array != NULL, "Error"); 4945 assert(_collectorState == FinalMarking || 4946 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4947 for (int j = 0; j < no_of_gc_threads; j++) { 4948 _cursor[j] = 0; 4949 } 4950 HeapWord* top = surv->top(); 4951 size_t i; 4952 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4953 HeapWord* min_val = top; // Higher than any PLAB address 4954 uint min_tid = 0; // position of min_val this round 4955 for (int j = 0; j < no_of_gc_threads; j++) { 4956 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4957 if (_cursor[j] == cur_sca->end()) { 4958 continue; 4959 } 4960 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4961 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4962 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4963 if (cur_val < min_val) { 4964 min_tid = j; 4965 min_val = cur_val; 4966 } else { 4967 assert(cur_val < top, "All recorded addresses should be less"); 4968 } 4969 } 4970 // At this point min_val and min_tid are respectively 4971 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4972 // and the thread (j) that witnesses that address. 4973 // We record this address in the _survivor_chunk_array[i] 4974 // and increment _cursor[min_tid] prior to the next round i. 4975 if (min_val == top) { 4976 break; 4977 } 4978 _survivor_chunk_array[i] = min_val; 4979 _cursor[min_tid]++; 4980 } 4981 // We are all done; record the size of the _survivor_chunk_array 4982 _survivor_chunk_index = i; // exclusive: [0, i) 4983 if (PrintCMSStatistics > 0) { 4984 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4985 } 4986 // Verify that we used up all the recorded entries 4987 #ifdef ASSERT 4988 size_t total = 0; 4989 for (int j = 0; j < no_of_gc_threads; j++) { 4990 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4991 total += _cursor[j]; 4992 } 4993 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4994 // Check that the merged array is in sorted order 4995 if (total > 0) { 4996 for (size_t i = 0; i < total - 1; i++) { 4997 if (PrintCMSStatistics > 0) { 4998 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4999 i, p2i(_survivor_chunk_array[i])); 5000 } 5001 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5002 "Not sorted"); 5003 } 5004 } 5005 #endif // ASSERT 5006 } 5007 5008 // Set up the space's par_seq_tasks structure for work claiming 5009 // for parallel initial scan and rescan of young gen. 5010 // See ParRescanTask where this is currently used. 5011 void 5012 CMSCollector:: 5013 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5014 assert(n_threads > 0, "Unexpected n_threads argument"); 5015 5016 // Eden space 5017 if (!_young_gen->eden()->is_empty()) { 5018 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5019 assert(!pst->valid(), "Clobbering existing data?"); 5020 // Each valid entry in [0, _eden_chunk_index) represents a task. 5021 size_t n_tasks = _eden_chunk_index + 1; 5022 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5023 // Sets the condition for completion of the subtask (how many threads 5024 // need to finish in order to be done). 5025 pst->set_n_threads(n_threads); 5026 pst->set_n_tasks((int)n_tasks); 5027 } 5028 5029 // Merge the survivor plab arrays into _survivor_chunk_array 5030 if (_survivor_plab_array != NULL) { 5031 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5032 } else { 5033 assert(_survivor_chunk_index == 0, "Error"); 5034 } 5035 5036 // To space 5037 { 5038 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5039 assert(!pst->valid(), "Clobbering existing data?"); 5040 // Sets the condition for completion of the subtask (how many threads 5041 // need to finish in order to be done). 5042 pst->set_n_threads(n_threads); 5043 pst->set_n_tasks(1); 5044 assert(pst->valid(), "Error"); 5045 } 5046 5047 // From space 5048 { 5049 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5050 assert(!pst->valid(), "Clobbering existing data?"); 5051 size_t n_tasks = _survivor_chunk_index + 1; 5052 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5053 // Sets the condition for completion of the subtask (how many threads 5054 // need to finish in order to be done). 5055 pst->set_n_threads(n_threads); 5056 pst->set_n_tasks((int)n_tasks); 5057 assert(pst->valid(), "Error"); 5058 } 5059 } 5060 5061 // Parallel version of remark 5062 void CMSCollector::do_remark_parallel() { 5063 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5064 WorkGang* workers = gch->workers(); 5065 assert(workers != NULL, "Need parallel worker threads."); 5066 // Choose to use the number of GC workers most recently set 5067 // into "active_workers". 5068 uint n_workers = workers->active_workers(); 5069 5070 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5071 5072 StrongRootsScope srs(n_workers); 5073 5074 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5075 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 workers->run_task(&tsk); 5108 } else { 5109 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5110 tsk.work(0); 5111 } 5112 5113 // restore, single-threaded for now, any preserved marks 5114 // as a result of work_q overflow 5115 restore_preserved_marks_if_any(); 5116 } 5117 5118 // Non-parallel version of remark 5119 void CMSCollector::do_remark_non_parallel() { 5120 ResourceMark rm; 5121 HandleMark hm; 5122 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5123 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5124 5125 MarkRefsIntoAndScanClosure 5126 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5127 &_markStack, this, 5128 false /* should_yield */, false /* not precleaning */); 5129 MarkFromDirtyCardsClosure 5130 markFromDirtyCardsClosure(this, _span, 5131 NULL, // space is set further below 5132 &_markBitMap, &_markStack, &mrias_cl); 5133 { 5134 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5135 // Iterate over the dirty cards, setting the corresponding bits in the 5136 // mod union table. 5137 { 5138 ModUnionClosure modUnionClosure(&_modUnionTable); 5139 _ct->ct_bs()->dirty_card_iterate( 5140 _cmsGen->used_region(), 5141 &modUnionClosure); 5142 } 5143 // Having transferred these marks into the modUnionTable, we just need 5144 // to rescan the marked objects on the dirty cards in the modUnionTable. 5145 // The initial marking may have been done during an asynchronous 5146 // collection so there may be dirty bits in the mod-union table. 5147 const int alignment = 5148 CardTableModRefBS::card_size * BitsPerWord; 5149 { 5150 // ... First handle dirty cards in CMS gen 5151 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5152 MemRegion ur = _cmsGen->used_region(); 5153 HeapWord* lb = ur.start(); 5154 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5155 MemRegion cms_span(lb, ub); 5156 _modUnionTable.dirty_range_iterate_clear(cms_span, 5157 &markFromDirtyCardsClosure); 5158 verify_work_stacks_empty(); 5159 if (PrintCMSStatistics != 0) { 5160 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", 5161 markFromDirtyCardsClosure.num_dirty_cards()); 5162 } 5163 } 5164 } 5165 if (VerifyDuringGC && 5166 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5167 HandleMark hm; // Discard invalid handles created during verification 5168 Universe::verify(); 5169 } 5170 { 5171 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5172 5173 verify_work_stacks_empty(); 5174 5175 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5176 StrongRootsScope srs(1); 5177 5178 gch->gen_process_roots(&srs, 5179 GenCollectedHeap::OldGen, 5180 true, // younger gens as roots 5181 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5182 should_unload_classes(), 5183 &mrias_cl, 5184 NULL, 5185 NULL); // The dirty klasses will be handled below 5186 5187 assert(should_unload_classes() 5188 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5189 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5190 } 5191 5192 { 5193 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5194 5195 verify_work_stacks_empty(); 5196 5197 // Scan all class loader data objects that might have been introduced 5198 // during concurrent marking. 5199 ResourceMark rm; 5200 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5201 for (int i = 0; i < array->length(); i++) { 5202 mrias_cl.do_class_loader_data(array->at(i)); 5203 } 5204 5205 // We don't need to keep track of new CLDs anymore. 5206 ClassLoaderDataGraph::remember_new_clds(false); 5207 5208 verify_work_stacks_empty(); 5209 } 5210 5211 { 5212 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5213 5214 verify_work_stacks_empty(); 5215 5216 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5217 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5218 5219 verify_work_stacks_empty(); 5220 } 5221 5222 // We might have added oops to ClassLoaderData::_handles during the 5223 // concurrent marking phase. These oops point to newly allocated objects 5224 // that are guaranteed to be kept alive. Either by the direct allocation 5225 // code, or when the young collector processes the roots. Hence, 5226 // we don't have to revisit the _handles block during the remark phase. 5227 5228 verify_work_stacks_empty(); 5229 // Restore evacuated mark words, if any, used for overflow list links 5230 if (!CMSOverflowEarlyRestoration) { 5231 restore_preserved_marks_if_any(); 5232 } 5233 verify_overflow_empty(); 5234 } 5235 5236 //////////////////////////////////////////////////////// 5237 // Parallel Reference Processing Task Proxy Class 5238 //////////////////////////////////////////////////////// 5239 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5240 OopTaskQueueSet* _queues; 5241 ParallelTaskTerminator _terminator; 5242 public: 5243 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5244 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5245 ParallelTaskTerminator* terminator() { return &_terminator; } 5246 OopTaskQueueSet* queues() { return _queues; } 5247 }; 5248 5249 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5250 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5251 CMSCollector* _collector; 5252 CMSBitMap* _mark_bit_map; 5253 const MemRegion _span; 5254 ProcessTask& _task; 5255 5256 public: 5257 CMSRefProcTaskProxy(ProcessTask& task, 5258 CMSCollector* collector, 5259 const MemRegion& span, 5260 CMSBitMap* mark_bit_map, 5261 AbstractWorkGang* workers, 5262 OopTaskQueueSet* task_queues): 5263 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5264 task_queues, 5265 workers->active_workers()), 5266 _task(task), 5267 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5268 { 5269 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5270 "Inconsistency in _span"); 5271 } 5272 5273 OopTaskQueueSet* task_queues() { return queues(); } 5274 5275 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5276 5277 void do_work_steal(int i, 5278 CMSParDrainMarkingStackClosure* drain, 5279 CMSParKeepAliveClosure* keep_alive, 5280 int* seed); 5281 5282 virtual void work(uint worker_id); 5283 }; 5284 5285 void CMSRefProcTaskProxy::work(uint worker_id) { 5286 ResourceMark rm; 5287 HandleMark hm; 5288 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5289 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5290 _mark_bit_map, 5291 work_queue(worker_id)); 5292 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5293 _mark_bit_map, 5294 work_queue(worker_id)); 5295 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5296 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5297 if (_task.marks_oops_alive()) { 5298 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5299 _collector->hash_seed(worker_id)); 5300 } 5301 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5302 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5303 } 5304 5305 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5306 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5307 EnqueueTask& _task; 5308 5309 public: 5310 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5311 : AbstractGangTask("Enqueue reference objects in parallel"), 5312 _task(task) 5313 { } 5314 5315 virtual void work(uint worker_id) 5316 { 5317 _task.work(worker_id); 5318 } 5319 }; 5320 5321 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5322 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5323 _span(span), 5324 _bit_map(bit_map), 5325 _work_queue(work_queue), 5326 _mark_and_push(collector, span, bit_map, work_queue), 5327 _low_water_mark(MIN2((work_queue->max_elems()/4), 5328 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5329 { } 5330 5331 // . see if we can share work_queues with ParNew? XXX 5332 void CMSRefProcTaskProxy::do_work_steal(int i, 5333 CMSParDrainMarkingStackClosure* drain, 5334 CMSParKeepAliveClosure* keep_alive, 5335 int* seed) { 5336 OopTaskQueue* work_q = work_queue(i); 5337 NOT_PRODUCT(int num_steals = 0;) 5338 oop obj_to_scan; 5339 5340 while (true) { 5341 // Completely finish any left over work from (an) earlier round(s) 5342 drain->trim_queue(0); 5343 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5344 (size_t)ParGCDesiredObjsFromOverflowList); 5345 // Now check if there's any work in the overflow list 5346 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5347 // only affects the number of attempts made to get work from the 5348 // overflow list and does not affect the number of workers. Just 5349 // pass ParallelGCThreads so this behavior is unchanged. 5350 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5351 work_q, 5352 ParallelGCThreads)) { 5353 // Found something in global overflow list; 5354 // not yet ready to go stealing work from others. 5355 // We'd like to assert(work_q->size() != 0, ...) 5356 // because we just took work from the overflow list, 5357 // but of course we can't, since all of that might have 5358 // been already stolen from us. 5359 continue; 5360 } 5361 // Verify that we have no work before we resort to stealing 5362 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5363 // Try to steal from other queues that have work 5364 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5365 NOT_PRODUCT(num_steals++;) 5366 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5367 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5368 // Do scanning work 5369 obj_to_scan->oop_iterate(keep_alive); 5370 // Loop around, finish this work, and try to steal some more 5371 } else if (terminator()->offer_termination()) { 5372 break; // nirvana from the infinite cycle 5373 } 5374 } 5375 NOT_PRODUCT( 5376 if (PrintCMSStatistics != 0) { 5377 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5378 } 5379 ) 5380 } 5381 5382 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5383 { 5384 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5385 WorkGang* workers = gch->workers(); 5386 assert(workers != NULL, "Need parallel worker threads."); 5387 CMSRefProcTaskProxy rp_task(task, &_collector, 5388 _collector.ref_processor()->span(), 5389 _collector.markBitMap(), 5390 workers, _collector.task_queues()); 5391 workers->run_task(&rp_task); 5392 } 5393 5394 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5395 { 5396 5397 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5398 WorkGang* workers = gch->workers(); 5399 assert(workers != NULL, "Need parallel worker threads."); 5400 CMSRefEnqueueTaskProxy enq_task(task); 5401 workers->run_task(&enq_task); 5402 } 5403 5404 void CMSCollector::refProcessingWork() { 5405 ResourceMark rm; 5406 HandleMark hm; 5407 5408 ReferenceProcessor* rp = ref_processor(); 5409 assert(rp->span().equals(_span), "Spans should be equal"); 5410 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5411 // Process weak references. 5412 rp->setup_policy(false); 5413 verify_work_stacks_empty(); 5414 5415 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5416 &_markStack, false /* !preclean */); 5417 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5418 _span, &_markBitMap, &_markStack, 5419 &cmsKeepAliveClosure, false /* !preclean */); 5420 { 5421 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5422 5423 ReferenceProcessorStats stats; 5424 if (rp->processing_is_mt()) { 5425 // Set the degree of MT here. If the discovery is done MT, there 5426 // may have been a different number of threads doing the discovery 5427 // and a different number of discovered lists may have Ref objects. 5428 // That is OK as long as the Reference lists are balanced (see 5429 // balance_all_queues() and balance_queues()). 5430 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5431 uint active_workers = ParallelGCThreads; 5432 WorkGang* workers = gch->workers(); 5433 if (workers != NULL) { 5434 active_workers = workers->active_workers(); 5435 // The expectation is that active_workers will have already 5436 // been set to a reasonable value. If it has not been set, 5437 // investigate. 5438 assert(active_workers > 0, "Should have been set during scavenge"); 5439 } 5440 rp->set_active_mt_degree(active_workers); 5441 CMSRefProcTaskExecutor task_executor(*this); 5442 stats = rp->process_discovered_references(&_is_alive_closure, 5443 &cmsKeepAliveClosure, 5444 &cmsDrainMarkingStackClosure, 5445 &task_executor, 5446 _gc_timer_cm, 5447 _gc_tracer_cm->gc_id()); 5448 } else { 5449 stats = rp->process_discovered_references(&_is_alive_closure, 5450 &cmsKeepAliveClosure, 5451 &cmsDrainMarkingStackClosure, 5452 NULL, 5453 _gc_timer_cm, 5454 _gc_tracer_cm->gc_id()); 5455 } 5456 _gc_tracer_cm->report_gc_reference_stats(stats); 5457 5458 } 5459 5460 // This is the point where the entire marking should have completed. 5461 verify_work_stacks_empty(); 5462 5463 if (should_unload_classes()) { 5464 { 5465 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5466 5467 // Unload classes and purge the SystemDictionary. 5468 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5469 5470 // Unload nmethods. 5471 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5472 5473 // Prune dead klasses from subklass/sibling/implementor lists. 5474 Klass::clean_weak_klass_links(&_is_alive_closure); 5475 } 5476 5477 { 5478 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5479 // Clean up unreferenced symbols in symbol table. 5480 SymbolTable::unlink(); 5481 } 5482 5483 { 5484 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5485 // Delete entries for dead interned strings. 5486 StringTable::unlink(&_is_alive_closure); 5487 } 5488 } 5489 5490 5491 // Restore any preserved marks as a result of mark stack or 5492 // work queue overflow 5493 restore_preserved_marks_if_any(); // done single-threaded for now 5494 5495 rp->set_enqueuing_is_done(true); 5496 if (rp->processing_is_mt()) { 5497 rp->balance_all_queues(); 5498 CMSRefProcTaskExecutor task_executor(*this); 5499 rp->enqueue_discovered_references(&task_executor); 5500 } else { 5501 rp->enqueue_discovered_references(NULL); 5502 } 5503 rp->verify_no_references_recorded(); 5504 assert(!rp->discovery_enabled(), "should have been disabled"); 5505 } 5506 5507 #ifndef PRODUCT 5508 void CMSCollector::check_correct_thread_executing() { 5509 Thread* t = Thread::current(); 5510 // Only the VM thread or the CMS thread should be here. 5511 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5512 "Unexpected thread type"); 5513 // If this is the vm thread, the foreground process 5514 // should not be waiting. Note that _foregroundGCIsActive is 5515 // true while the foreground collector is waiting. 5516 if (_foregroundGCShouldWait) { 5517 // We cannot be the VM thread 5518 assert(t->is_ConcurrentGC_thread(), 5519 "Should be CMS thread"); 5520 } else { 5521 // We can be the CMS thread only if we are in a stop-world 5522 // phase of CMS collection. 5523 if (t->is_ConcurrentGC_thread()) { 5524 assert(_collectorState == InitialMarking || 5525 _collectorState == FinalMarking, 5526 "Should be a stop-world phase"); 5527 // The CMS thread should be holding the CMS_token. 5528 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5529 "Potential interference with concurrently " 5530 "executing VM thread"); 5531 } 5532 } 5533 } 5534 #endif 5535 5536 void CMSCollector::sweep() { 5537 assert(_collectorState == Sweeping, "just checking"); 5538 check_correct_thread_executing(); 5539 verify_work_stacks_empty(); 5540 verify_overflow_empty(); 5541 increment_sweep_count(); 5542 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5543 5544 _inter_sweep_timer.stop(); 5545 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5546 5547 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5548 _intra_sweep_timer.reset(); 5549 _intra_sweep_timer.start(); 5550 { 5551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5552 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5553 // First sweep the old gen 5554 { 5555 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5556 bitMapLock()); 5557 sweepWork(_cmsGen); 5558 } 5559 5560 // Update Universe::_heap_*_at_gc figures. 5561 // We need all the free list locks to make the abstract state 5562 // transition from Sweeping to Resetting. See detailed note 5563 // further below. 5564 { 5565 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5566 // Update heap occupancy information which is used as 5567 // input to soft ref clearing policy at the next gc. 5568 Universe::update_heap_info_at_gc(); 5569 _collectorState = Resizing; 5570 } 5571 } 5572 verify_work_stacks_empty(); 5573 verify_overflow_empty(); 5574 5575 if (should_unload_classes()) { 5576 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5577 // requires that the virtual spaces are stable and not deleted. 5578 ClassLoaderDataGraph::set_should_purge(true); 5579 } 5580 5581 _intra_sweep_timer.stop(); 5582 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5583 5584 _inter_sweep_timer.reset(); 5585 _inter_sweep_timer.start(); 5586 5587 // We need to use a monotonically non-decreasing time in ms 5588 // or we will see time-warp warnings and os::javaTimeMillis() 5589 // does not guarantee monotonicity. 5590 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5591 update_time_of_last_gc(now); 5592 5593 // NOTE on abstract state transitions: 5594 // Mutators allocate-live and/or mark the mod-union table dirty 5595 // based on the state of the collection. The former is done in 5596 // the interval [Marking, Sweeping] and the latter in the interval 5597 // [Marking, Sweeping). Thus the transitions into the Marking state 5598 // and out of the Sweeping state must be synchronously visible 5599 // globally to the mutators. 5600 // The transition into the Marking state happens with the world 5601 // stopped so the mutators will globally see it. Sweeping is 5602 // done asynchronously by the background collector so the transition 5603 // from the Sweeping state to the Resizing state must be done 5604 // under the freelistLock (as is the check for whether to 5605 // allocate-live and whether to dirty the mod-union table). 5606 assert(_collectorState == Resizing, "Change of collector state to" 5607 " Resizing must be done under the freelistLocks (plural)"); 5608 5609 // Now that sweeping has been completed, we clear 5610 // the incremental_collection_failed flag, 5611 // thus inviting a younger gen collection to promote into 5612 // this generation. If such a promotion may still fail, 5613 // the flag will be set again when a young collection is 5614 // attempted. 5615 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5616 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5617 gch->update_full_collections_completed(_collection_count_start); 5618 } 5619 5620 // FIX ME!!! Looks like this belongs in CFLSpace, with 5621 // CMSGen merely delegating to it. 5622 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5623 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5624 HeapWord* minAddr = _cmsSpace->bottom(); 5625 HeapWord* largestAddr = 5626 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5627 if (largestAddr == NULL) { 5628 // The dictionary appears to be empty. In this case 5629 // try to coalesce at the end of the heap. 5630 largestAddr = _cmsSpace->end(); 5631 } 5632 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5633 size_t nearLargestOffset = 5634 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5635 if (PrintFLSStatistics != 0) { 5636 gclog_or_tty->print_cr( 5637 "CMS: Large Block: " PTR_FORMAT ";" 5638 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5639 p2i(largestAddr), 5640 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5641 } 5642 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5643 } 5644 5645 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5646 return addr >= _cmsSpace->nearLargestChunk(); 5647 } 5648 5649 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5650 return _cmsSpace->find_chunk_at_end(); 5651 } 5652 5653 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5654 bool full) { 5655 // If the young generation has been collected, gather any statistics 5656 // that are of interest at this point. 5657 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5658 if (!full && current_is_young) { 5659 // Gather statistics on the young generation collection. 5660 collector()->stats().record_gc0_end(used()); 5661 } 5662 } 5663 5664 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5665 // We iterate over the space(s) underlying this generation, 5666 // checking the mark bit map to see if the bits corresponding 5667 // to specific blocks are marked or not. Blocks that are 5668 // marked are live and are not swept up. All remaining blocks 5669 // are swept up, with coalescing on-the-fly as we sweep up 5670 // contiguous free and/or garbage blocks: 5671 // We need to ensure that the sweeper synchronizes with allocators 5672 // and stop-the-world collectors. In particular, the following 5673 // locks are used: 5674 // . CMS token: if this is held, a stop the world collection cannot occur 5675 // . freelistLock: if this is held no allocation can occur from this 5676 // generation by another thread 5677 // . bitMapLock: if this is held, no other thread can access or update 5678 // 5679 5680 // Note that we need to hold the freelistLock if we use 5681 // block iterate below; else the iterator might go awry if 5682 // a mutator (or promotion) causes block contents to change 5683 // (for instance if the allocator divvies up a block). 5684 // If we hold the free list lock, for all practical purposes 5685 // young generation GC's can't occur (they'll usually need to 5686 // promote), so we might as well prevent all young generation 5687 // GC's while we do a sweeping step. For the same reason, we might 5688 // as well take the bit map lock for the entire duration 5689 5690 // check that we hold the requisite locks 5691 assert(have_cms_token(), "Should hold cms token"); 5692 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5693 assert_lock_strong(gen->freelistLock()); 5694 assert_lock_strong(bitMapLock()); 5695 5696 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5697 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5698 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5699 _inter_sweep_estimate.padded_average(), 5700 _intra_sweep_estimate.padded_average()); 5701 gen->setNearLargestChunk(); 5702 5703 { 5704 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5705 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5706 // We need to free-up/coalesce garbage/blocks from a 5707 // co-terminal free run. This is done in the SweepClosure 5708 // destructor; so, do not remove this scope, else the 5709 // end-of-sweep-census below will be off by a little bit. 5710 } 5711 gen->cmsSpace()->sweep_completed(); 5712 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5713 if (should_unload_classes()) { // unloaded classes this cycle, 5714 _concurrent_cycles_since_last_unload = 0; // ... reset count 5715 } else { // did not unload classes, 5716 _concurrent_cycles_since_last_unload++; // ... increment count 5717 } 5718 } 5719 5720 // Reset CMS data structures (for now just the marking bit map) 5721 // preparatory for the next cycle. 5722 void CMSCollector::reset(bool concurrent) { 5723 if (concurrent) { 5724 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5725 5726 // If the state is not "Resetting", the foreground thread 5727 // has done a collection and the resetting. 5728 if (_collectorState != Resetting) { 5729 assert(_collectorState == Idling, "The state should only change" 5730 " because the foreground collector has finished the collection"); 5731 return; 5732 } 5733 5734 // Clear the mark bitmap (no grey objects to start with) 5735 // for the next cycle. 5736 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5737 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5738 5739 HeapWord* curAddr = _markBitMap.startWord(); 5740 while (curAddr < _markBitMap.endWord()) { 5741 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5742 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5743 _markBitMap.clear_large_range(chunk); 5744 if (ConcurrentMarkSweepThread::should_yield() && 5745 !foregroundGCIsActive() && 5746 CMSYield) { 5747 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5748 "CMS thread should hold CMS token"); 5749 assert_lock_strong(bitMapLock()); 5750 bitMapLock()->unlock(); 5751 ConcurrentMarkSweepThread::desynchronize(true); 5752 stopTimer(); 5753 if (PrintCMSStatistics != 0) { 5754 incrementYields(); 5755 } 5756 5757 // See the comment in coordinator_yield() 5758 for (unsigned i = 0; i < CMSYieldSleepCount && 5759 ConcurrentMarkSweepThread::should_yield() && 5760 !CMSCollector::foregroundGCIsActive(); ++i) { 5761 os::sleep(Thread::current(), 1, false); 5762 } 5763 5764 ConcurrentMarkSweepThread::synchronize(true); 5765 bitMapLock()->lock_without_safepoint_check(); 5766 startTimer(); 5767 } 5768 curAddr = chunk.end(); 5769 } 5770 // A successful mostly concurrent collection has been done. 5771 // Because only the full (i.e., concurrent mode failure) collections 5772 // are being measured for gc overhead limits, clean the "near" flag 5773 // and count. 5774 size_policy()->reset_gc_overhead_limit_count(); 5775 _collectorState = Idling; 5776 } else { 5777 // already have the lock 5778 assert(_collectorState == Resetting, "just checking"); 5779 assert_lock_strong(bitMapLock()); 5780 _markBitMap.clear_all(); 5781 _collectorState = Idling; 5782 } 5783 5784 register_gc_end(); 5785 } 5786 5787 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5788 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5789 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5790 TraceCollectorStats tcs(counters()); 5791 5792 switch (op) { 5793 case CMS_op_checkpointRootsInitial: { 5794 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5795 checkpointRootsInitial(); 5796 if (PrintGC) { 5797 _cmsGen->printOccupancy("initial-mark"); 5798 } 5799 break; 5800 } 5801 case CMS_op_checkpointRootsFinal: { 5802 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5803 checkpointRootsFinal(); 5804 if (PrintGC) { 5805 _cmsGen->printOccupancy("remark"); 5806 } 5807 break; 5808 } 5809 default: 5810 fatal("No such CMS_op"); 5811 } 5812 } 5813 5814 #ifndef PRODUCT 5815 size_t const CMSCollector::skip_header_HeapWords() { 5816 return FreeChunk::header_size(); 5817 } 5818 5819 // Try and collect here conditions that should hold when 5820 // CMS thread is exiting. The idea is that the foreground GC 5821 // thread should not be blocked if it wants to terminate 5822 // the CMS thread and yet continue to run the VM for a while 5823 // after that. 5824 void CMSCollector::verify_ok_to_terminate() const { 5825 assert(Thread::current()->is_ConcurrentGC_thread(), 5826 "should be called by CMS thread"); 5827 assert(!_foregroundGCShouldWait, "should be false"); 5828 // We could check here that all the various low-level locks 5829 // are not held by the CMS thread, but that is overkill; see 5830 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5831 // is checked. 5832 } 5833 #endif 5834 5835 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5836 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5837 "missing Printezis mark?"); 5838 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5839 size_t size = pointer_delta(nextOneAddr + 1, addr); 5840 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5841 "alignment problem"); 5842 assert(size >= 3, "Necessary for Printezis marks to work"); 5843 return size; 5844 } 5845 5846 // A variant of the above (block_size_using_printezis_bits()) except 5847 // that we return 0 if the P-bits are not yet set. 5848 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5849 if (_markBitMap.isMarked(addr + 1)) { 5850 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5851 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5852 size_t size = pointer_delta(nextOneAddr + 1, addr); 5853 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5854 "alignment problem"); 5855 assert(size >= 3, "Necessary for Printezis marks to work"); 5856 return size; 5857 } 5858 return 0; 5859 } 5860 5861 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5862 size_t sz = 0; 5863 oop p = (oop)addr; 5864 if (p->klass_or_null() != NULL) { 5865 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5866 } else { 5867 sz = block_size_using_printezis_bits(addr); 5868 } 5869 assert(sz > 0, "size must be nonzero"); 5870 HeapWord* next_block = addr + sz; 5871 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5872 CardTableModRefBS::card_size); 5873 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5874 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5875 "must be different cards"); 5876 return next_card; 5877 } 5878 5879 5880 // CMS Bit Map Wrapper ///////////////////////////////////////// 5881 5882 // Construct a CMS bit map infrastructure, but don't create the 5883 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5884 // further below. 5885 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5886 _bm(), 5887 _shifter(shifter), 5888 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5889 Monitor::_safepoint_check_sometimes) : NULL) 5890 { 5891 _bmStartWord = 0; 5892 _bmWordSize = 0; 5893 } 5894 5895 bool CMSBitMap::allocate(MemRegion mr) { 5896 _bmStartWord = mr.start(); 5897 _bmWordSize = mr.word_size(); 5898 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5899 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5900 if (!brs.is_reserved()) { 5901 warning("CMS bit map allocation failure"); 5902 return false; 5903 } 5904 // For now we'll just commit all of the bit map up front. 5905 // Later on we'll try to be more parsimonious with swap. 5906 if (!_virtual_space.initialize(brs, brs.size())) { 5907 warning("CMS bit map backing store failure"); 5908 return false; 5909 } 5910 assert(_virtual_space.committed_size() == brs.size(), 5911 "didn't reserve backing store for all of CMS bit map?"); 5912 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5913 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5914 _bmWordSize, "inconsistency in bit map sizing"); 5915 _bm.set_size(_bmWordSize >> _shifter); 5916 5917 // bm.clear(); // can we rely on getting zero'd memory? verify below 5918 assert(isAllClear(), 5919 "Expected zero'd memory from ReservedSpace constructor"); 5920 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5921 "consistency check"); 5922 return true; 5923 } 5924 5925 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5926 HeapWord *next_addr, *end_addr, *last_addr; 5927 assert_locked(); 5928 assert(covers(mr), "out-of-range error"); 5929 // XXX assert that start and end are appropriately aligned 5930 for (next_addr = mr.start(), end_addr = mr.end(); 5931 next_addr < end_addr; next_addr = last_addr) { 5932 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5933 last_addr = dirty_region.end(); 5934 if (!dirty_region.is_empty()) { 5935 cl->do_MemRegion(dirty_region); 5936 } else { 5937 assert(last_addr == end_addr, "program logic"); 5938 return; 5939 } 5940 } 5941 } 5942 5943 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5944 _bm.print_on_error(st, prefix); 5945 } 5946 5947 #ifndef PRODUCT 5948 void CMSBitMap::assert_locked() const { 5949 CMSLockVerifier::assert_locked(lock()); 5950 } 5951 5952 bool CMSBitMap::covers(MemRegion mr) const { 5953 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5954 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5955 "size inconsistency"); 5956 return (mr.start() >= _bmStartWord) && 5957 (mr.end() <= endWord()); 5958 } 5959 5960 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5961 return (start >= _bmStartWord && (start + size) <= endWord()); 5962 } 5963 5964 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5965 // verify that there are no 1 bits in the interval [left, right) 5966 FalseBitMapClosure falseBitMapClosure; 5967 iterate(&falseBitMapClosure, left, right); 5968 } 5969 5970 void CMSBitMap::region_invariant(MemRegion mr) 5971 { 5972 assert_locked(); 5973 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5974 assert(!mr.is_empty(), "unexpected empty region"); 5975 assert(covers(mr), "mr should be covered by bit map"); 5976 // convert address range into offset range 5977 size_t start_ofs = heapWordToOffset(mr.start()); 5978 // Make sure that end() is appropriately aligned 5979 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5980 (1 << (_shifter+LogHeapWordSize))), 5981 "Misaligned mr.end()"); 5982 size_t end_ofs = heapWordToOffset(mr.end()); 5983 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5984 } 5985 5986 #endif 5987 5988 bool CMSMarkStack::allocate(size_t size) { 5989 // allocate a stack of the requisite depth 5990 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5991 size * sizeof(oop))); 5992 if (!rs.is_reserved()) { 5993 warning("CMSMarkStack allocation failure"); 5994 return false; 5995 } 5996 if (!_virtual_space.initialize(rs, rs.size())) { 5997 warning("CMSMarkStack backing store failure"); 5998 return false; 5999 } 6000 assert(_virtual_space.committed_size() == rs.size(), 6001 "didn't reserve backing store for all of CMS stack?"); 6002 _base = (oop*)(_virtual_space.low()); 6003 _index = 0; 6004 _capacity = size; 6005 NOT_PRODUCT(_max_depth = 0); 6006 return true; 6007 } 6008 6009 // XXX FIX ME !!! In the MT case we come in here holding a 6010 // leaf lock. For printing we need to take a further lock 6011 // which has lower rank. We need to recalibrate the two 6012 // lock-ranks involved in order to be able to print the 6013 // messages below. (Or defer the printing to the caller. 6014 // For now we take the expedient path of just disabling the 6015 // messages for the problematic case.) 6016 void CMSMarkStack::expand() { 6017 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6018 if (_capacity == MarkStackSizeMax) { 6019 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6020 // We print a warning message only once per CMS cycle. 6021 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6022 } 6023 return; 6024 } 6025 // Double capacity if possible 6026 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6027 // Do not give up existing stack until we have managed to 6028 // get the double capacity that we desired. 6029 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6030 new_capacity * sizeof(oop))); 6031 if (rs.is_reserved()) { 6032 // Release the backing store associated with old stack 6033 _virtual_space.release(); 6034 // Reinitialize virtual space for new stack 6035 if (!_virtual_space.initialize(rs, rs.size())) { 6036 fatal("Not enough swap for expanded marking stack"); 6037 } 6038 _base = (oop*)(_virtual_space.low()); 6039 _index = 0; 6040 _capacity = new_capacity; 6041 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6042 // Failed to double capacity, continue; 6043 // we print a detail message only once per CMS cycle. 6044 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " 6045 SIZE_FORMAT "K", 6046 _capacity / K, new_capacity / K); 6047 } 6048 } 6049 6050 6051 // Closures 6052 // XXX: there seems to be a lot of code duplication here; 6053 // should refactor and consolidate common code. 6054 6055 // This closure is used to mark refs into the CMS generation in 6056 // the CMS bit map. Called at the first checkpoint. This closure 6057 // assumes that we do not need to re-mark dirty cards; if the CMS 6058 // generation on which this is used is not an oldest 6059 // generation then this will lose younger_gen cards! 6060 6061 MarkRefsIntoClosure::MarkRefsIntoClosure( 6062 MemRegion span, CMSBitMap* bitMap): 6063 _span(span), 6064 _bitMap(bitMap) 6065 { 6066 assert(_ref_processor == NULL, "deliberately left NULL"); 6067 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6068 } 6069 6070 void MarkRefsIntoClosure::do_oop(oop obj) { 6071 // if p points into _span, then mark corresponding bit in _markBitMap 6072 assert(obj->is_oop(), "expected an oop"); 6073 HeapWord* addr = (HeapWord*)obj; 6074 if (_span.contains(addr)) { 6075 // this should be made more efficient 6076 _bitMap->mark(addr); 6077 } 6078 } 6079 6080 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6081 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6082 6083 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6084 MemRegion span, CMSBitMap* bitMap): 6085 _span(span), 6086 _bitMap(bitMap) 6087 { 6088 assert(_ref_processor == NULL, "deliberately left NULL"); 6089 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6090 } 6091 6092 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6093 // if p points into _span, then mark corresponding bit in _markBitMap 6094 assert(obj->is_oop(), "expected an oop"); 6095 HeapWord* addr = (HeapWord*)obj; 6096 if (_span.contains(addr)) { 6097 // this should be made more efficient 6098 _bitMap->par_mark(addr); 6099 } 6100 } 6101 6102 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6103 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6104 6105 // A variant of the above, used for CMS marking verification. 6106 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6107 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6108 _span(span), 6109 _verification_bm(verification_bm), 6110 _cms_bm(cms_bm) 6111 { 6112 assert(_ref_processor == NULL, "deliberately left NULL"); 6113 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6114 } 6115 6116 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6117 // if p points into _span, then mark corresponding bit in _markBitMap 6118 assert(obj->is_oop(), "expected an oop"); 6119 HeapWord* addr = (HeapWord*)obj; 6120 if (_span.contains(addr)) { 6121 _verification_bm->mark(addr); 6122 if (!_cms_bm->isMarked(addr)) { 6123 oop(addr)->print(); 6124 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6125 fatal("... aborting"); 6126 } 6127 } 6128 } 6129 6130 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6131 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6132 6133 ////////////////////////////////////////////////// 6134 // MarkRefsIntoAndScanClosure 6135 ////////////////////////////////////////////////// 6136 6137 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6138 ReferenceProcessor* rp, 6139 CMSBitMap* bit_map, 6140 CMSBitMap* mod_union_table, 6141 CMSMarkStack* mark_stack, 6142 CMSCollector* collector, 6143 bool should_yield, 6144 bool concurrent_precleaning): 6145 _collector(collector), 6146 _span(span), 6147 _bit_map(bit_map), 6148 _mark_stack(mark_stack), 6149 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6150 mark_stack, concurrent_precleaning), 6151 _yield(should_yield), 6152 _concurrent_precleaning(concurrent_precleaning), 6153 _freelistLock(NULL) 6154 { 6155 _ref_processor = rp; 6156 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6157 } 6158 6159 // This closure is used to mark refs into the CMS generation at the 6160 // second (final) checkpoint, and to scan and transitively follow 6161 // the unmarked oops. It is also used during the concurrent precleaning 6162 // phase while scanning objects on dirty cards in the CMS generation. 6163 // The marks are made in the marking bit map and the marking stack is 6164 // used for keeping the (newly) grey objects during the scan. 6165 // The parallel version (Par_...) appears further below. 6166 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6167 if (obj != NULL) { 6168 assert(obj->is_oop(), "expected an oop"); 6169 HeapWord* addr = (HeapWord*)obj; 6170 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6171 assert(_collector->overflow_list_is_empty(), 6172 "overflow list should be empty"); 6173 if (_span.contains(addr) && 6174 !_bit_map->isMarked(addr)) { 6175 // mark bit map (object is now grey) 6176 _bit_map->mark(addr); 6177 // push on marking stack (stack should be empty), and drain the 6178 // stack by applying this closure to the oops in the oops popped 6179 // from the stack (i.e. blacken the grey objects) 6180 bool res = _mark_stack->push(obj); 6181 assert(res, "Should have space to push on empty stack"); 6182 do { 6183 oop new_oop = _mark_stack->pop(); 6184 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6185 assert(_bit_map->isMarked((HeapWord*)new_oop), 6186 "only grey objects on this stack"); 6187 // iterate over the oops in this oop, marking and pushing 6188 // the ones in CMS heap (i.e. in _span). 6189 new_oop->oop_iterate(&_pushAndMarkClosure); 6190 // check if it's time to yield 6191 do_yield_check(); 6192 } while (!_mark_stack->isEmpty() || 6193 (!_concurrent_precleaning && take_from_overflow_list())); 6194 // if marking stack is empty, and we are not doing this 6195 // during precleaning, then check the overflow list 6196 } 6197 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6198 assert(_collector->overflow_list_is_empty(), 6199 "overflow list was drained above"); 6200 // We could restore evacuated mark words, if any, used for 6201 // overflow list links here because the overflow list is 6202 // provably empty here. That would reduce the maximum 6203 // size requirements for preserved_{oop,mark}_stack. 6204 // But we'll just postpone it until we are all done 6205 // so we can just stream through. 6206 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6207 _collector->restore_preserved_marks_if_any(); 6208 assert(_collector->no_preserved_marks(), "No preserved marks"); 6209 } 6210 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6211 "All preserved marks should have been restored above"); 6212 } 6213 } 6214 6215 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6216 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6217 6218 void MarkRefsIntoAndScanClosure::do_yield_work() { 6219 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6220 "CMS thread should hold CMS token"); 6221 assert_lock_strong(_freelistLock); 6222 assert_lock_strong(_bit_map->lock()); 6223 // relinquish the free_list_lock and bitMaplock() 6224 _bit_map->lock()->unlock(); 6225 _freelistLock->unlock(); 6226 ConcurrentMarkSweepThread::desynchronize(true); 6227 _collector->stopTimer(); 6228 if (PrintCMSStatistics != 0) { 6229 _collector->incrementYields(); 6230 } 6231 6232 // See the comment in coordinator_yield() 6233 for (unsigned i = 0; 6234 i < CMSYieldSleepCount && 6235 ConcurrentMarkSweepThread::should_yield() && 6236 !CMSCollector::foregroundGCIsActive(); 6237 ++i) { 6238 os::sleep(Thread::current(), 1, false); 6239 } 6240 6241 ConcurrentMarkSweepThread::synchronize(true); 6242 _freelistLock->lock_without_safepoint_check(); 6243 _bit_map->lock()->lock_without_safepoint_check(); 6244 _collector->startTimer(); 6245 } 6246 6247 /////////////////////////////////////////////////////////// 6248 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6249 // MarkRefsIntoAndScanClosure 6250 /////////////////////////////////////////////////////////// 6251 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6252 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6253 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6254 _span(span), 6255 _bit_map(bit_map), 6256 _work_queue(work_queue), 6257 _low_water_mark(MIN2((work_queue->max_elems()/4), 6258 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6259 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6260 { 6261 _ref_processor = rp; 6262 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6263 } 6264 6265 // This closure is used to mark refs into the CMS generation at the 6266 // second (final) checkpoint, and to scan and transitively follow 6267 // the unmarked oops. The marks are made in the marking bit map and 6268 // the work_queue is used for keeping the (newly) grey objects during 6269 // the scan phase whence they are also available for stealing by parallel 6270 // threads. Since the marking bit map is shared, updates are 6271 // synchronized (via CAS). 6272 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6273 if (obj != NULL) { 6274 // Ignore mark word because this could be an already marked oop 6275 // that may be chained at the end of the overflow list. 6276 assert(obj->is_oop(true), "expected an oop"); 6277 HeapWord* addr = (HeapWord*)obj; 6278 if (_span.contains(addr) && 6279 !_bit_map->isMarked(addr)) { 6280 // mark bit map (object will become grey): 6281 // It is possible for several threads to be 6282 // trying to "claim" this object concurrently; 6283 // the unique thread that succeeds in marking the 6284 // object first will do the subsequent push on 6285 // to the work queue (or overflow list). 6286 if (_bit_map->par_mark(addr)) { 6287 // push on work_queue (which may not be empty), and trim the 6288 // queue to an appropriate length by applying this closure to 6289 // the oops in the oops popped from the stack (i.e. blacken the 6290 // grey objects) 6291 bool res = _work_queue->push(obj); 6292 assert(res, "Low water mark should be less than capacity?"); 6293 trim_queue(_low_water_mark); 6294 } // Else, another thread claimed the object 6295 } 6296 } 6297 } 6298 6299 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6300 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6301 6302 // This closure is used to rescan the marked objects on the dirty cards 6303 // in the mod union table and the card table proper. 6304 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6305 oop p, MemRegion mr) { 6306 6307 size_t size = 0; 6308 HeapWord* addr = (HeapWord*)p; 6309 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6310 assert(_span.contains(addr), "we are scanning the CMS generation"); 6311 // check if it's time to yield 6312 if (do_yield_check()) { 6313 // We yielded for some foreground stop-world work, 6314 // and we have been asked to abort this ongoing preclean cycle. 6315 return 0; 6316 } 6317 if (_bitMap->isMarked(addr)) { 6318 // it's marked; is it potentially uninitialized? 6319 if (p->klass_or_null() != NULL) { 6320 // an initialized object; ignore mark word in verification below 6321 // since we are running concurrent with mutators 6322 assert(p->is_oop(true), "should be an oop"); 6323 if (p->is_objArray()) { 6324 // objArrays are precisely marked; restrict scanning 6325 // to dirty cards only. 6326 size = CompactibleFreeListSpace::adjustObjectSize( 6327 p->oop_iterate(_scanningClosure, mr)); 6328 } else { 6329 // A non-array may have been imprecisely marked; we need 6330 // to scan object in its entirety. 6331 size = CompactibleFreeListSpace::adjustObjectSize( 6332 p->oop_iterate(_scanningClosure)); 6333 } 6334 #ifdef ASSERT 6335 size_t direct_size = 6336 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6337 assert(size == direct_size, "Inconsistency in size"); 6338 assert(size >= 3, "Necessary for Printezis marks to work"); 6339 if (!_bitMap->isMarked(addr+1)) { 6340 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6341 } else { 6342 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6343 assert(_bitMap->isMarked(addr+size-1), 6344 "inconsistent Printezis mark"); 6345 } 6346 #endif // ASSERT 6347 } else { 6348 // An uninitialized object. 6349 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6350 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6351 size = pointer_delta(nextOneAddr + 1, addr); 6352 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6353 "alignment problem"); 6354 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6355 // will dirty the card when the klass pointer is installed in the 6356 // object (signaling the completion of initialization). 6357 } 6358 } else { 6359 // Either a not yet marked object or an uninitialized object 6360 if (p->klass_or_null() == NULL) { 6361 // An uninitialized object, skip to the next card, since 6362 // we may not be able to read its P-bits yet. 6363 assert(size == 0, "Initial value"); 6364 } else { 6365 // An object not (yet) reached by marking: we merely need to 6366 // compute its size so as to go look at the next block. 6367 assert(p->is_oop(true), "should be an oop"); 6368 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6369 } 6370 } 6371 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6372 return size; 6373 } 6374 6375 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6376 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6377 "CMS thread should hold CMS token"); 6378 assert_lock_strong(_freelistLock); 6379 assert_lock_strong(_bitMap->lock()); 6380 // relinquish the free_list_lock and bitMaplock() 6381 _bitMap->lock()->unlock(); 6382 _freelistLock->unlock(); 6383 ConcurrentMarkSweepThread::desynchronize(true); 6384 _collector->stopTimer(); 6385 if (PrintCMSStatistics != 0) { 6386 _collector->incrementYields(); 6387 } 6388 6389 // See the comment in coordinator_yield() 6390 for (unsigned i = 0; i < CMSYieldSleepCount && 6391 ConcurrentMarkSweepThread::should_yield() && 6392 !CMSCollector::foregroundGCIsActive(); ++i) { 6393 os::sleep(Thread::current(), 1, false); 6394 } 6395 6396 ConcurrentMarkSweepThread::synchronize(true); 6397 _freelistLock->lock_without_safepoint_check(); 6398 _bitMap->lock()->lock_without_safepoint_check(); 6399 _collector->startTimer(); 6400 } 6401 6402 6403 ////////////////////////////////////////////////////////////////// 6404 // SurvivorSpacePrecleanClosure 6405 ////////////////////////////////////////////////////////////////// 6406 // This (single-threaded) closure is used to preclean the oops in 6407 // the survivor spaces. 6408 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6409 6410 HeapWord* addr = (HeapWord*)p; 6411 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6412 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6413 assert(p->klass_or_null() != NULL, "object should be initialized"); 6414 // an initialized object; ignore mark word in verification below 6415 // since we are running concurrent with mutators 6416 assert(p->is_oop(true), "should be an oop"); 6417 // Note that we do not yield while we iterate over 6418 // the interior oops of p, pushing the relevant ones 6419 // on our marking stack. 6420 size_t size = p->oop_iterate(_scanning_closure); 6421 do_yield_check(); 6422 // Observe that below, we do not abandon the preclean 6423 // phase as soon as we should; rather we empty the 6424 // marking stack before returning. This is to satisfy 6425 // some existing assertions. In general, it may be a 6426 // good idea to abort immediately and complete the marking 6427 // from the grey objects at a later time. 6428 while (!_mark_stack->isEmpty()) { 6429 oop new_oop = _mark_stack->pop(); 6430 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6431 assert(_bit_map->isMarked((HeapWord*)new_oop), 6432 "only grey objects on this stack"); 6433 // iterate over the oops in this oop, marking and pushing 6434 // the ones in CMS heap (i.e. in _span). 6435 new_oop->oop_iterate(_scanning_closure); 6436 // check if it's time to yield 6437 do_yield_check(); 6438 } 6439 unsigned int after_count = 6440 GenCollectedHeap::heap()->total_collections(); 6441 bool abort = (_before_count != after_count) || 6442 _collector->should_abort_preclean(); 6443 return abort ? 0 : size; 6444 } 6445 6446 void SurvivorSpacePrecleanClosure::do_yield_work() { 6447 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6448 "CMS thread should hold CMS token"); 6449 assert_lock_strong(_bit_map->lock()); 6450 // Relinquish the bit map lock 6451 _bit_map->lock()->unlock(); 6452 ConcurrentMarkSweepThread::desynchronize(true); 6453 _collector->stopTimer(); 6454 if (PrintCMSStatistics != 0) { 6455 _collector->incrementYields(); 6456 } 6457 6458 // See the comment in coordinator_yield() 6459 for (unsigned i = 0; i < CMSYieldSleepCount && 6460 ConcurrentMarkSweepThread::should_yield() && 6461 !CMSCollector::foregroundGCIsActive(); ++i) { 6462 os::sleep(Thread::current(), 1, false); 6463 } 6464 6465 ConcurrentMarkSweepThread::synchronize(true); 6466 _bit_map->lock()->lock_without_safepoint_check(); 6467 _collector->startTimer(); 6468 } 6469 6470 // This closure is used to rescan the marked objects on the dirty cards 6471 // in the mod union table and the card table proper. In the parallel 6472 // case, although the bitMap is shared, we do a single read so the 6473 // isMarked() query is "safe". 6474 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6475 // Ignore mark word because we are running concurrent with mutators 6476 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6477 HeapWord* addr = (HeapWord*)p; 6478 assert(_span.contains(addr), "we are scanning the CMS generation"); 6479 bool is_obj_array = false; 6480 #ifdef ASSERT 6481 if (!_parallel) { 6482 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6483 assert(_collector->overflow_list_is_empty(), 6484 "overflow list should be empty"); 6485 6486 } 6487 #endif // ASSERT 6488 if (_bit_map->isMarked(addr)) { 6489 // Obj arrays are precisely marked, non-arrays are not; 6490 // so we scan objArrays precisely and non-arrays in their 6491 // entirety. 6492 if (p->is_objArray()) { 6493 is_obj_array = true; 6494 if (_parallel) { 6495 p->oop_iterate(_par_scan_closure, mr); 6496 } else { 6497 p->oop_iterate(_scan_closure, mr); 6498 } 6499 } else { 6500 if (_parallel) { 6501 p->oop_iterate(_par_scan_closure); 6502 } else { 6503 p->oop_iterate(_scan_closure); 6504 } 6505 } 6506 } 6507 #ifdef ASSERT 6508 if (!_parallel) { 6509 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6510 assert(_collector->overflow_list_is_empty(), 6511 "overflow list should be empty"); 6512 6513 } 6514 #endif // ASSERT 6515 return is_obj_array; 6516 } 6517 6518 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6519 MemRegion span, 6520 CMSBitMap* bitMap, CMSMarkStack* markStack, 6521 bool should_yield, bool verifying): 6522 _collector(collector), 6523 _span(span), 6524 _bitMap(bitMap), 6525 _mut(&collector->_modUnionTable), 6526 _markStack(markStack), 6527 _yield(should_yield), 6528 _skipBits(0) 6529 { 6530 assert(_markStack->isEmpty(), "stack should be empty"); 6531 _finger = _bitMap->startWord(); 6532 _threshold = _finger; 6533 assert(_collector->_restart_addr == NULL, "Sanity check"); 6534 assert(_span.contains(_finger), "Out of bounds _finger?"); 6535 DEBUG_ONLY(_verifying = verifying;) 6536 } 6537 6538 void MarkFromRootsClosure::reset(HeapWord* addr) { 6539 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6540 assert(_span.contains(addr), "Out of bounds _finger?"); 6541 _finger = addr; 6542 _threshold = (HeapWord*)round_to( 6543 (intptr_t)_finger, CardTableModRefBS::card_size); 6544 } 6545 6546 // Should revisit to see if this should be restructured for 6547 // greater efficiency. 6548 bool MarkFromRootsClosure::do_bit(size_t offset) { 6549 if (_skipBits > 0) { 6550 _skipBits--; 6551 return true; 6552 } 6553 // convert offset into a HeapWord* 6554 HeapWord* addr = _bitMap->startWord() + offset; 6555 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6556 "address out of range"); 6557 assert(_bitMap->isMarked(addr), "tautology"); 6558 if (_bitMap->isMarked(addr+1)) { 6559 // this is an allocated but not yet initialized object 6560 assert(_skipBits == 0, "tautology"); 6561 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6562 oop p = oop(addr); 6563 if (p->klass_or_null() == NULL) { 6564 DEBUG_ONLY(if (!_verifying) {) 6565 // We re-dirty the cards on which this object lies and increase 6566 // the _threshold so that we'll come back to scan this object 6567 // during the preclean or remark phase. (CMSCleanOnEnter) 6568 if (CMSCleanOnEnter) { 6569 size_t sz = _collector->block_size_using_printezis_bits(addr); 6570 HeapWord* end_card_addr = (HeapWord*)round_to( 6571 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6572 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6573 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6574 // Bump _threshold to end_card_addr; note that 6575 // _threshold cannot possibly exceed end_card_addr, anyhow. 6576 // This prevents future clearing of the card as the scan proceeds 6577 // to the right. 6578 assert(_threshold <= end_card_addr, 6579 "Because we are just scanning into this object"); 6580 if (_threshold < end_card_addr) { 6581 _threshold = end_card_addr; 6582 } 6583 if (p->klass_or_null() != NULL) { 6584 // Redirty the range of cards... 6585 _mut->mark_range(redirty_range); 6586 } // ...else the setting of klass will dirty the card anyway. 6587 } 6588 DEBUG_ONLY(}) 6589 return true; 6590 } 6591 } 6592 scanOopsInOop(addr); 6593 return true; 6594 } 6595 6596 // We take a break if we've been at this for a while, 6597 // so as to avoid monopolizing the locks involved. 6598 void MarkFromRootsClosure::do_yield_work() { 6599 // First give up the locks, then yield, then re-lock 6600 // We should probably use a constructor/destructor idiom to 6601 // do this unlock/lock or modify the MutexUnlocker class to 6602 // serve our purpose. XXX 6603 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6604 "CMS thread should hold CMS token"); 6605 assert_lock_strong(_bitMap->lock()); 6606 _bitMap->lock()->unlock(); 6607 ConcurrentMarkSweepThread::desynchronize(true); 6608 _collector->stopTimer(); 6609 if (PrintCMSStatistics != 0) { 6610 _collector->incrementYields(); 6611 } 6612 6613 // See the comment in coordinator_yield() 6614 for (unsigned i = 0; i < CMSYieldSleepCount && 6615 ConcurrentMarkSweepThread::should_yield() && 6616 !CMSCollector::foregroundGCIsActive(); ++i) { 6617 os::sleep(Thread::current(), 1, false); 6618 } 6619 6620 ConcurrentMarkSweepThread::synchronize(true); 6621 _bitMap->lock()->lock_without_safepoint_check(); 6622 _collector->startTimer(); 6623 } 6624 6625 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6626 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6627 assert(_markStack->isEmpty(), 6628 "should drain stack to limit stack usage"); 6629 // convert ptr to an oop preparatory to scanning 6630 oop obj = oop(ptr); 6631 // Ignore mark word in verification below, since we 6632 // may be running concurrent with mutators. 6633 assert(obj->is_oop(true), "should be an oop"); 6634 assert(_finger <= ptr, "_finger runneth ahead"); 6635 // advance the finger to right end of this object 6636 _finger = ptr + obj->size(); 6637 assert(_finger > ptr, "we just incremented it above"); 6638 // On large heaps, it may take us some time to get through 6639 // the marking phase. During 6640 // this time it's possible that a lot of mutations have 6641 // accumulated in the card table and the mod union table -- 6642 // these mutation records are redundant until we have 6643 // actually traced into the corresponding card. 6644 // Here, we check whether advancing the finger would make 6645 // us cross into a new card, and if so clear corresponding 6646 // cards in the MUT (preclean them in the card-table in the 6647 // future). 6648 6649 DEBUG_ONLY(if (!_verifying) {) 6650 // The clean-on-enter optimization is disabled by default, 6651 // until we fix 6178663. 6652 if (CMSCleanOnEnter && (_finger > _threshold)) { 6653 // [_threshold, _finger) represents the interval 6654 // of cards to be cleared in MUT (or precleaned in card table). 6655 // The set of cards to be cleared is all those that overlap 6656 // with the interval [_threshold, _finger); note that 6657 // _threshold is always kept card-aligned but _finger isn't 6658 // always card-aligned. 6659 HeapWord* old_threshold = _threshold; 6660 assert(old_threshold == (HeapWord*)round_to( 6661 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6662 "_threshold should always be card-aligned"); 6663 _threshold = (HeapWord*)round_to( 6664 (intptr_t)_finger, CardTableModRefBS::card_size); 6665 MemRegion mr(old_threshold, _threshold); 6666 assert(!mr.is_empty(), "Control point invariant"); 6667 assert(_span.contains(mr), "Should clear within span"); 6668 _mut->clear_range(mr); 6669 } 6670 DEBUG_ONLY(}) 6671 // Note: the finger doesn't advance while we drain 6672 // the stack below. 6673 PushOrMarkClosure pushOrMarkClosure(_collector, 6674 _span, _bitMap, _markStack, 6675 _finger, this); 6676 bool res = _markStack->push(obj); 6677 assert(res, "Empty non-zero size stack should have space for single push"); 6678 while (!_markStack->isEmpty()) { 6679 oop new_oop = _markStack->pop(); 6680 // Skip verifying header mark word below because we are 6681 // running concurrent with mutators. 6682 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6683 // now scan this oop's oops 6684 new_oop->oop_iterate(&pushOrMarkClosure); 6685 do_yield_check(); 6686 } 6687 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6688 } 6689 6690 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6691 CMSCollector* collector, MemRegion span, 6692 CMSBitMap* bit_map, 6693 OopTaskQueue* work_queue, 6694 CMSMarkStack* overflow_stack): 6695 _collector(collector), 6696 _whole_span(collector->_span), 6697 _span(span), 6698 _bit_map(bit_map), 6699 _mut(&collector->_modUnionTable), 6700 _work_queue(work_queue), 6701 _overflow_stack(overflow_stack), 6702 _skip_bits(0), 6703 _task(task) 6704 { 6705 assert(_work_queue->size() == 0, "work_queue should be empty"); 6706 _finger = span.start(); 6707 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6708 assert(_span.contains(_finger), "Out of bounds _finger?"); 6709 } 6710 6711 // Should revisit to see if this should be restructured for 6712 // greater efficiency. 6713 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6714 if (_skip_bits > 0) { 6715 _skip_bits--; 6716 return true; 6717 } 6718 // convert offset into a HeapWord* 6719 HeapWord* addr = _bit_map->startWord() + offset; 6720 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6721 "address out of range"); 6722 assert(_bit_map->isMarked(addr), "tautology"); 6723 if (_bit_map->isMarked(addr+1)) { 6724 // this is an allocated object that might not yet be initialized 6725 assert(_skip_bits == 0, "tautology"); 6726 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6727 oop p = oop(addr); 6728 if (p->klass_or_null() == NULL) { 6729 // in the case of Clean-on-Enter optimization, redirty card 6730 // and avoid clearing card by increasing the threshold. 6731 return true; 6732 } 6733 } 6734 scan_oops_in_oop(addr); 6735 return true; 6736 } 6737 6738 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6739 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6740 // Should we assert that our work queue is empty or 6741 // below some drain limit? 6742 assert(_work_queue->size() == 0, 6743 "should drain stack to limit stack usage"); 6744 // convert ptr to an oop preparatory to scanning 6745 oop obj = oop(ptr); 6746 // Ignore mark word in verification below, since we 6747 // may be running concurrent with mutators. 6748 assert(obj->is_oop(true), "should be an oop"); 6749 assert(_finger <= ptr, "_finger runneth ahead"); 6750 // advance the finger to right end of this object 6751 _finger = ptr + obj->size(); 6752 assert(_finger > ptr, "we just incremented it above"); 6753 // On large heaps, it may take us some time to get through 6754 // the marking phase. During 6755 // this time it's possible that a lot of mutations have 6756 // accumulated in the card table and the mod union table -- 6757 // these mutation records are redundant until we have 6758 // actually traced into the corresponding card. 6759 // Here, we check whether advancing the finger would make 6760 // us cross into a new card, and if so clear corresponding 6761 // cards in the MUT (preclean them in the card-table in the 6762 // future). 6763 6764 // The clean-on-enter optimization is disabled by default, 6765 // until we fix 6178663. 6766 if (CMSCleanOnEnter && (_finger > _threshold)) { 6767 // [_threshold, _finger) represents the interval 6768 // of cards to be cleared in MUT (or precleaned in card table). 6769 // The set of cards to be cleared is all those that overlap 6770 // with the interval [_threshold, _finger); note that 6771 // _threshold is always kept card-aligned but _finger isn't 6772 // always card-aligned. 6773 HeapWord* old_threshold = _threshold; 6774 assert(old_threshold == (HeapWord*)round_to( 6775 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6776 "_threshold should always be card-aligned"); 6777 _threshold = (HeapWord*)round_to( 6778 (intptr_t)_finger, CardTableModRefBS::card_size); 6779 MemRegion mr(old_threshold, _threshold); 6780 assert(!mr.is_empty(), "Control point invariant"); 6781 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6782 _mut->clear_range(mr); 6783 } 6784 6785 // Note: the local finger doesn't advance while we drain 6786 // the stack below, but the global finger sure can and will. 6787 HeapWord** gfa = _task->global_finger_addr(); 6788 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6789 _span, _bit_map, 6790 _work_queue, 6791 _overflow_stack, 6792 _finger, 6793 gfa, this); 6794 bool res = _work_queue->push(obj); // overflow could occur here 6795 assert(res, "Will hold once we use workqueues"); 6796 while (true) { 6797 oop new_oop; 6798 if (!_work_queue->pop_local(new_oop)) { 6799 // We emptied our work_queue; check if there's stuff that can 6800 // be gotten from the overflow stack. 6801 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6802 _overflow_stack, _work_queue)) { 6803 do_yield_check(); 6804 continue; 6805 } else { // done 6806 break; 6807 } 6808 } 6809 // Skip verifying header mark word below because we are 6810 // running concurrent with mutators. 6811 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6812 // now scan this oop's oops 6813 new_oop->oop_iterate(&pushOrMarkClosure); 6814 do_yield_check(); 6815 } 6816 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6817 } 6818 6819 // Yield in response to a request from VM Thread or 6820 // from mutators. 6821 void Par_MarkFromRootsClosure::do_yield_work() { 6822 assert(_task != NULL, "sanity"); 6823 _task->yield(); 6824 } 6825 6826 // A variant of the above used for verifying CMS marking work. 6827 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6828 MemRegion span, 6829 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6830 CMSMarkStack* mark_stack): 6831 _collector(collector), 6832 _span(span), 6833 _verification_bm(verification_bm), 6834 _cms_bm(cms_bm), 6835 _mark_stack(mark_stack), 6836 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6837 mark_stack) 6838 { 6839 assert(_mark_stack->isEmpty(), "stack should be empty"); 6840 _finger = _verification_bm->startWord(); 6841 assert(_collector->_restart_addr == NULL, "Sanity check"); 6842 assert(_span.contains(_finger), "Out of bounds _finger?"); 6843 } 6844 6845 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6846 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6847 assert(_span.contains(addr), "Out of bounds _finger?"); 6848 _finger = addr; 6849 } 6850 6851 // Should revisit to see if this should be restructured for 6852 // greater efficiency. 6853 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6854 // convert offset into a HeapWord* 6855 HeapWord* addr = _verification_bm->startWord() + offset; 6856 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6857 "address out of range"); 6858 assert(_verification_bm->isMarked(addr), "tautology"); 6859 assert(_cms_bm->isMarked(addr), "tautology"); 6860 6861 assert(_mark_stack->isEmpty(), 6862 "should drain stack to limit stack usage"); 6863 // convert addr to an oop preparatory to scanning 6864 oop obj = oop(addr); 6865 assert(obj->is_oop(), "should be an oop"); 6866 assert(_finger <= addr, "_finger runneth ahead"); 6867 // advance the finger to right end of this object 6868 _finger = addr + obj->size(); 6869 assert(_finger > addr, "we just incremented it above"); 6870 // Note: the finger doesn't advance while we drain 6871 // the stack below. 6872 bool res = _mark_stack->push(obj); 6873 assert(res, "Empty non-zero size stack should have space for single push"); 6874 while (!_mark_stack->isEmpty()) { 6875 oop new_oop = _mark_stack->pop(); 6876 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6877 // now scan this oop's oops 6878 new_oop->oop_iterate(&_pam_verify_closure); 6879 } 6880 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6881 return true; 6882 } 6883 6884 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6885 CMSCollector* collector, MemRegion span, 6886 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6887 CMSMarkStack* mark_stack): 6888 MetadataAwareOopClosure(collector->ref_processor()), 6889 _collector(collector), 6890 _span(span), 6891 _verification_bm(verification_bm), 6892 _cms_bm(cms_bm), 6893 _mark_stack(mark_stack) 6894 { } 6895 6896 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6897 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6898 6899 // Upon stack overflow, we discard (part of) the stack, 6900 // remembering the least address amongst those discarded 6901 // in CMSCollector's _restart_address. 6902 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6903 // Remember the least grey address discarded 6904 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6905 _collector->lower_restart_addr(ra); 6906 _mark_stack->reset(); // discard stack contents 6907 _mark_stack->expand(); // expand the stack if possible 6908 } 6909 6910 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6911 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6912 HeapWord* addr = (HeapWord*)obj; 6913 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6914 // Oop lies in _span and isn't yet grey or black 6915 _verification_bm->mark(addr); // now grey 6916 if (!_cms_bm->isMarked(addr)) { 6917 oop(addr)->print(); 6918 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6919 p2i(addr)); 6920 fatal("... aborting"); 6921 } 6922 6923 if (!_mark_stack->push(obj)) { // stack overflow 6924 if (PrintCMSStatistics != 0) { 6925 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6926 SIZE_FORMAT, _mark_stack->capacity()); 6927 } 6928 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6929 handle_stack_overflow(addr); 6930 } 6931 // anything including and to the right of _finger 6932 // will be scanned as we iterate over the remainder of the 6933 // bit map 6934 } 6935 } 6936 6937 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6938 MemRegion span, 6939 CMSBitMap* bitMap, CMSMarkStack* markStack, 6940 HeapWord* finger, MarkFromRootsClosure* parent) : 6941 MetadataAwareOopClosure(collector->ref_processor()), 6942 _collector(collector), 6943 _span(span), 6944 _bitMap(bitMap), 6945 _markStack(markStack), 6946 _finger(finger), 6947 _parent(parent) 6948 { } 6949 6950 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6951 MemRegion span, 6952 CMSBitMap* bit_map, 6953 OopTaskQueue* work_queue, 6954 CMSMarkStack* overflow_stack, 6955 HeapWord* finger, 6956 HeapWord** global_finger_addr, 6957 Par_MarkFromRootsClosure* parent) : 6958 MetadataAwareOopClosure(collector->ref_processor()), 6959 _collector(collector), 6960 _whole_span(collector->_span), 6961 _span(span), 6962 _bit_map(bit_map), 6963 _work_queue(work_queue), 6964 _overflow_stack(overflow_stack), 6965 _finger(finger), 6966 _global_finger_addr(global_finger_addr), 6967 _parent(parent) 6968 { } 6969 6970 // Assumes thread-safe access by callers, who are 6971 // responsible for mutual exclusion. 6972 void CMSCollector::lower_restart_addr(HeapWord* low) { 6973 assert(_span.contains(low), "Out of bounds addr"); 6974 if (_restart_addr == NULL) { 6975 _restart_addr = low; 6976 } else { 6977 _restart_addr = MIN2(_restart_addr, low); 6978 } 6979 } 6980 6981 // Upon stack overflow, we discard (part of) the stack, 6982 // remembering the least address amongst those discarded 6983 // in CMSCollector's _restart_address. 6984 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6985 // Remember the least grey address discarded 6986 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6987 _collector->lower_restart_addr(ra); 6988 _markStack->reset(); // discard stack contents 6989 _markStack->expand(); // expand the stack if possible 6990 } 6991 6992 // Upon stack overflow, we discard (part of) the stack, 6993 // remembering the least address amongst those discarded 6994 // in CMSCollector's _restart_address. 6995 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6996 // We need to do this under a mutex to prevent other 6997 // workers from interfering with the work done below. 6998 MutexLockerEx ml(_overflow_stack->par_lock(), 6999 Mutex::_no_safepoint_check_flag); 7000 // Remember the least grey address discarded 7001 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7002 _collector->lower_restart_addr(ra); 7003 _overflow_stack->reset(); // discard stack contents 7004 _overflow_stack->expand(); // expand the stack if possible 7005 } 7006 7007 void PushOrMarkClosure::do_oop(oop obj) { 7008 // Ignore mark word because we are running concurrent with mutators. 7009 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7010 HeapWord* addr = (HeapWord*)obj; 7011 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7012 // Oop lies in _span and isn't yet grey or black 7013 _bitMap->mark(addr); // now grey 7014 if (addr < _finger) { 7015 // the bit map iteration has already either passed, or 7016 // sampled, this bit in the bit map; we'll need to 7017 // use the marking stack to scan this oop's oops. 7018 bool simulate_overflow = false; 7019 NOT_PRODUCT( 7020 if (CMSMarkStackOverflowALot && 7021 _collector->simulate_overflow()) { 7022 // simulate a stack overflow 7023 simulate_overflow = true; 7024 } 7025 ) 7026 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7027 if (PrintCMSStatistics != 0) { 7028 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7029 SIZE_FORMAT, _markStack->capacity()); 7030 } 7031 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7032 handle_stack_overflow(addr); 7033 } 7034 } 7035 // anything including and to the right of _finger 7036 // will be scanned as we iterate over the remainder of the 7037 // bit map 7038 do_yield_check(); 7039 } 7040 } 7041 7042 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7043 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7044 7045 void Par_PushOrMarkClosure::do_oop(oop obj) { 7046 // Ignore mark word because we are running concurrent with mutators. 7047 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7048 HeapWord* addr = (HeapWord*)obj; 7049 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7050 // Oop lies in _span and isn't yet grey or black 7051 // We read the global_finger (volatile read) strictly after marking oop 7052 bool res = _bit_map->par_mark(addr); // now grey 7053 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7054 // Should we push this marked oop on our stack? 7055 // -- if someone else marked it, nothing to do 7056 // -- if target oop is above global finger nothing to do 7057 // -- if target oop is in chunk and above local finger 7058 // then nothing to do 7059 // -- else push on work queue 7060 if ( !res // someone else marked it, they will deal with it 7061 || (addr >= *gfa) // will be scanned in a later task 7062 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7063 return; 7064 } 7065 // the bit map iteration has already either passed, or 7066 // sampled, this bit in the bit map; we'll need to 7067 // use the marking stack to scan this oop's oops. 7068 bool simulate_overflow = false; 7069 NOT_PRODUCT( 7070 if (CMSMarkStackOverflowALot && 7071 _collector->simulate_overflow()) { 7072 // simulate a stack overflow 7073 simulate_overflow = true; 7074 } 7075 ) 7076 if (simulate_overflow || 7077 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7078 // stack overflow 7079 if (PrintCMSStatistics != 0) { 7080 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7081 SIZE_FORMAT, _overflow_stack->capacity()); 7082 } 7083 // We cannot assert that the overflow stack is full because 7084 // it may have been emptied since. 7085 assert(simulate_overflow || 7086 _work_queue->size() == _work_queue->max_elems(), 7087 "Else push should have succeeded"); 7088 handle_stack_overflow(addr); 7089 } 7090 do_yield_check(); 7091 } 7092 } 7093 7094 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7095 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7096 7097 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7098 MemRegion span, 7099 ReferenceProcessor* rp, 7100 CMSBitMap* bit_map, 7101 CMSBitMap* mod_union_table, 7102 CMSMarkStack* mark_stack, 7103 bool concurrent_precleaning): 7104 MetadataAwareOopClosure(rp), 7105 _collector(collector), 7106 _span(span), 7107 _bit_map(bit_map), 7108 _mod_union_table(mod_union_table), 7109 _mark_stack(mark_stack), 7110 _concurrent_precleaning(concurrent_precleaning) 7111 { 7112 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7113 } 7114 7115 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7116 // the non-parallel version (the parallel version appears further below.) 7117 void PushAndMarkClosure::do_oop(oop obj) { 7118 // Ignore mark word verification. If during concurrent precleaning, 7119 // the object monitor may be locked. If during the checkpoint 7120 // phases, the object may already have been reached by a different 7121 // path and may be at the end of the global overflow list (so 7122 // the mark word may be NULL). 7123 assert(obj->is_oop_or_null(true /* ignore mark word */), 7124 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7125 HeapWord* addr = (HeapWord*)obj; 7126 // Check if oop points into the CMS generation 7127 // and is not marked 7128 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7129 // a white object ... 7130 _bit_map->mark(addr); // ... now grey 7131 // push on the marking stack (grey set) 7132 bool simulate_overflow = false; 7133 NOT_PRODUCT( 7134 if (CMSMarkStackOverflowALot && 7135 _collector->simulate_overflow()) { 7136 // simulate a stack overflow 7137 simulate_overflow = true; 7138 } 7139 ) 7140 if (simulate_overflow || !_mark_stack->push(obj)) { 7141 if (_concurrent_precleaning) { 7142 // During precleaning we can just dirty the appropriate card(s) 7143 // in the mod union table, thus ensuring that the object remains 7144 // in the grey set and continue. In the case of object arrays 7145 // we need to dirty all of the cards that the object spans, 7146 // since the rescan of object arrays will be limited to the 7147 // dirty cards. 7148 // Note that no one can be interfering with us in this action 7149 // of dirtying the mod union table, so no locking or atomics 7150 // are required. 7151 if (obj->is_objArray()) { 7152 size_t sz = obj->size(); 7153 HeapWord* end_card_addr = (HeapWord*)round_to( 7154 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7155 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7156 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7157 _mod_union_table->mark_range(redirty_range); 7158 } else { 7159 _mod_union_table->mark(addr); 7160 } 7161 _collector->_ser_pmc_preclean_ovflw++; 7162 } else { 7163 // During the remark phase, we need to remember this oop 7164 // in the overflow list. 7165 _collector->push_on_overflow_list(obj); 7166 _collector->_ser_pmc_remark_ovflw++; 7167 } 7168 } 7169 } 7170 } 7171 7172 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7173 MemRegion span, 7174 ReferenceProcessor* rp, 7175 CMSBitMap* bit_map, 7176 OopTaskQueue* work_queue): 7177 MetadataAwareOopClosure(rp), 7178 _collector(collector), 7179 _span(span), 7180 _bit_map(bit_map), 7181 _work_queue(work_queue) 7182 { 7183 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7184 } 7185 7186 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7187 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7188 7189 // Grey object rescan during second checkpoint phase -- 7190 // the parallel version. 7191 void Par_PushAndMarkClosure::do_oop(oop obj) { 7192 // In the assert below, we ignore the mark word because 7193 // this oop may point to an already visited object that is 7194 // on the overflow stack (in which case the mark word has 7195 // been hijacked for chaining into the overflow stack -- 7196 // if this is the last object in the overflow stack then 7197 // its mark word will be NULL). Because this object may 7198 // have been subsequently popped off the global overflow 7199 // stack, and the mark word possibly restored to the prototypical 7200 // value, by the time we get to examined this failing assert in 7201 // the debugger, is_oop_or_null(false) may subsequently start 7202 // to hold. 7203 assert(obj->is_oop_or_null(true), 7204 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7205 HeapWord* addr = (HeapWord*)obj; 7206 // Check if oop points into the CMS generation 7207 // and is not marked 7208 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7209 // a white object ... 7210 // If we manage to "claim" the object, by being the 7211 // first thread to mark it, then we push it on our 7212 // marking stack 7213 if (_bit_map->par_mark(addr)) { // ... now grey 7214 // push on work queue (grey set) 7215 bool simulate_overflow = false; 7216 NOT_PRODUCT( 7217 if (CMSMarkStackOverflowALot && 7218 _collector->par_simulate_overflow()) { 7219 // simulate a stack overflow 7220 simulate_overflow = true; 7221 } 7222 ) 7223 if (simulate_overflow || !_work_queue->push(obj)) { 7224 _collector->par_push_on_overflow_list(obj); 7225 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7226 } 7227 } // Else, some other thread got there first 7228 } 7229 } 7230 7231 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7232 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7233 7234 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7235 Mutex* bml = _collector->bitMapLock(); 7236 assert_lock_strong(bml); 7237 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7238 "CMS thread should hold CMS token"); 7239 7240 bml->unlock(); 7241 ConcurrentMarkSweepThread::desynchronize(true); 7242 7243 _collector->stopTimer(); 7244 if (PrintCMSStatistics != 0) { 7245 _collector->incrementYields(); 7246 } 7247 7248 // See the comment in coordinator_yield() 7249 for (unsigned i = 0; i < CMSYieldSleepCount && 7250 ConcurrentMarkSweepThread::should_yield() && 7251 !CMSCollector::foregroundGCIsActive(); ++i) { 7252 os::sleep(Thread::current(), 1, false); 7253 } 7254 7255 ConcurrentMarkSweepThread::synchronize(true); 7256 bml->lock(); 7257 7258 _collector->startTimer(); 7259 } 7260 7261 bool CMSPrecleanRefsYieldClosure::should_return() { 7262 if (ConcurrentMarkSweepThread::should_yield()) { 7263 do_yield_work(); 7264 } 7265 return _collector->foregroundGCIsActive(); 7266 } 7267 7268 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7269 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7270 "mr should be aligned to start at a card boundary"); 7271 // We'd like to assert: 7272 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7273 // "mr should be a range of cards"); 7274 // However, that would be too strong in one case -- the last 7275 // partition ends at _unallocated_block which, in general, can be 7276 // an arbitrary boundary, not necessarily card aligned. 7277 if (PrintCMSStatistics != 0) { 7278 _num_dirty_cards += 7279 mr.word_size()/CardTableModRefBS::card_size_in_words; 7280 } 7281 _space->object_iterate_mem(mr, &_scan_cl); 7282 } 7283 7284 SweepClosure::SweepClosure(CMSCollector* collector, 7285 ConcurrentMarkSweepGeneration* g, 7286 CMSBitMap* bitMap, bool should_yield) : 7287 _collector(collector), 7288 _g(g), 7289 _sp(g->cmsSpace()), 7290 _limit(_sp->sweep_limit()), 7291 _freelistLock(_sp->freelistLock()), 7292 _bitMap(bitMap), 7293 _yield(should_yield), 7294 _inFreeRange(false), // No free range at beginning of sweep 7295 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7296 _lastFreeRangeCoalesced(false), 7297 _freeFinger(g->used_region().start()) 7298 { 7299 NOT_PRODUCT( 7300 _numObjectsFreed = 0; 7301 _numWordsFreed = 0; 7302 _numObjectsLive = 0; 7303 _numWordsLive = 0; 7304 _numObjectsAlreadyFree = 0; 7305 _numWordsAlreadyFree = 0; 7306 _last_fc = NULL; 7307 7308 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7309 _sp->dictionary()->initialize_dict_returned_bytes(); 7310 ) 7311 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7312 "sweep _limit out of bounds"); 7313 if (CMSTraceSweeper) { 7314 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7315 p2i(_limit)); 7316 } 7317 } 7318 7319 void SweepClosure::print_on(outputStream* st) const { 7320 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7321 p2i(_sp->bottom()), p2i(_sp->end())); 7322 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7323 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7324 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7325 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7326 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7327 } 7328 7329 #ifndef PRODUCT 7330 // Assertion checking only: no useful work in product mode -- 7331 // however, if any of the flags below become product flags, 7332 // you may need to review this code to see if it needs to be 7333 // enabled in product mode. 7334 SweepClosure::~SweepClosure() { 7335 assert_lock_strong(_freelistLock); 7336 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7337 "sweep _limit out of bounds"); 7338 if (inFreeRange()) { 7339 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7340 print(); 7341 ShouldNotReachHere(); 7342 } 7343 if (Verbose && PrintGC) { 7344 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7345 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7346 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, " 7347 SIZE_FORMAT " bytes " 7348 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7349 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7350 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7351 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7352 * sizeof(HeapWord); 7353 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7354 7355 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7356 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7357 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7358 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7359 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes); 7360 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes", 7361 indexListReturnedBytes); 7362 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes", 7363 dict_returned_bytes); 7364 } 7365 } 7366 if (CMSTraceSweeper) { 7367 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7368 p2i(_limit)); 7369 } 7370 } 7371 #endif // PRODUCT 7372 7373 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7374 bool freeRangeInFreeLists) { 7375 if (CMSTraceSweeper) { 7376 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7377 p2i(freeFinger), freeRangeInFreeLists); 7378 } 7379 assert(!inFreeRange(), "Trampling existing free range"); 7380 set_inFreeRange(true); 7381 set_lastFreeRangeCoalesced(false); 7382 7383 set_freeFinger(freeFinger); 7384 set_freeRangeInFreeLists(freeRangeInFreeLists); 7385 if (CMSTestInFreeList) { 7386 if (freeRangeInFreeLists) { 7387 FreeChunk* fc = (FreeChunk*) freeFinger; 7388 assert(fc->is_free(), "A chunk on the free list should be free."); 7389 assert(fc->size() > 0, "Free range should have a size"); 7390 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7391 } 7392 } 7393 } 7394 7395 // Note that the sweeper runs concurrently with mutators. Thus, 7396 // it is possible for direct allocation in this generation to happen 7397 // in the middle of the sweep. Note that the sweeper also coalesces 7398 // contiguous free blocks. Thus, unless the sweeper and the allocator 7399 // synchronize appropriately freshly allocated blocks may get swept up. 7400 // This is accomplished by the sweeper locking the free lists while 7401 // it is sweeping. Thus blocks that are determined to be free are 7402 // indeed free. There is however one additional complication: 7403 // blocks that have been allocated since the final checkpoint and 7404 // mark, will not have been marked and so would be treated as 7405 // unreachable and swept up. To prevent this, the allocator marks 7406 // the bit map when allocating during the sweep phase. This leads, 7407 // however, to a further complication -- objects may have been allocated 7408 // but not yet initialized -- in the sense that the header isn't yet 7409 // installed. The sweeper can not then determine the size of the block 7410 // in order to skip over it. To deal with this case, we use a technique 7411 // (due to Printezis) to encode such uninitialized block sizes in the 7412 // bit map. Since the bit map uses a bit per every HeapWord, but the 7413 // CMS generation has a minimum object size of 3 HeapWords, it follows 7414 // that "normal marks" won't be adjacent in the bit map (there will 7415 // always be at least two 0 bits between successive 1 bits). We make use 7416 // of these "unused" bits to represent uninitialized blocks -- the bit 7417 // corresponding to the start of the uninitialized object and the next 7418 // bit are both set. Finally, a 1 bit marks the end of the object that 7419 // started with the two consecutive 1 bits to indicate its potentially 7420 // uninitialized state. 7421 7422 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7423 FreeChunk* fc = (FreeChunk*)addr; 7424 size_t res; 7425 7426 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7427 // than "addr == _limit" because although _limit was a block boundary when 7428 // we started the sweep, it may no longer be one because heap expansion 7429 // may have caused us to coalesce the block ending at the address _limit 7430 // with a newly expanded chunk (this happens when _limit was set to the 7431 // previous _end of the space), so we may have stepped past _limit: 7432 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7433 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7434 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7435 "sweep _limit out of bounds"); 7436 assert(addr < _sp->end(), "addr out of bounds"); 7437 // Flush any free range we might be holding as a single 7438 // coalesced chunk to the appropriate free list. 7439 if (inFreeRange()) { 7440 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7441 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()))); 7442 flush_cur_free_chunk(freeFinger(), 7443 pointer_delta(addr, freeFinger())); 7444 if (CMSTraceSweeper) { 7445 gclog_or_tty->print("Sweep: last chunk: "); 7446 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") " 7447 "[coalesced:%d]\n", 7448 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7449 lastFreeRangeCoalesced() ? 1 : 0); 7450 } 7451 } 7452 7453 // help the iterator loop finish 7454 return pointer_delta(_sp->end(), addr); 7455 } 7456 7457 assert(addr < _limit, "sweep invariant"); 7458 // check if we should yield 7459 do_yield_check(addr); 7460 if (fc->is_free()) { 7461 // Chunk that is already free 7462 res = fc->size(); 7463 do_already_free_chunk(fc); 7464 debug_only(_sp->verifyFreeLists()); 7465 // If we flush the chunk at hand in lookahead_and_flush() 7466 // and it's coalesced with a preceding chunk, then the 7467 // process of "mangling" the payload of the coalesced block 7468 // will cause erasure of the size information from the 7469 // (erstwhile) header of all the coalesced blocks but the 7470 // first, so the first disjunct in the assert will not hold 7471 // in that specific case (in which case the second disjunct 7472 // will hold). 7473 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7474 "Otherwise the size info doesn't change at this step"); 7475 NOT_PRODUCT( 7476 _numObjectsAlreadyFree++; 7477 _numWordsAlreadyFree += res; 7478 ) 7479 NOT_PRODUCT(_last_fc = fc;) 7480 } else if (!_bitMap->isMarked(addr)) { 7481 // Chunk is fresh garbage 7482 res = do_garbage_chunk(fc); 7483 debug_only(_sp->verifyFreeLists()); 7484 NOT_PRODUCT( 7485 _numObjectsFreed++; 7486 _numWordsFreed += res; 7487 ) 7488 } else { 7489 // Chunk that is alive. 7490 res = do_live_chunk(fc); 7491 debug_only(_sp->verifyFreeLists()); 7492 NOT_PRODUCT( 7493 _numObjectsLive++; 7494 _numWordsLive += res; 7495 ) 7496 } 7497 return res; 7498 } 7499 7500 // For the smart allocation, record following 7501 // split deaths - a free chunk is removed from its free list because 7502 // it is being split into two or more chunks. 7503 // split birth - a free chunk is being added to its free list because 7504 // a larger free chunk has been split and resulted in this free chunk. 7505 // coal death - a free chunk is being removed from its free list because 7506 // it is being coalesced into a large free chunk. 7507 // coal birth - a free chunk is being added to its free list because 7508 // it was created when two or more free chunks where coalesced into 7509 // this free chunk. 7510 // 7511 // These statistics are used to determine the desired number of free 7512 // chunks of a given size. The desired number is chosen to be relative 7513 // to the end of a CMS sweep. The desired number at the end of a sweep 7514 // is the 7515 // count-at-end-of-previous-sweep (an amount that was enough) 7516 // - count-at-beginning-of-current-sweep (the excess) 7517 // + split-births (gains in this size during interval) 7518 // - split-deaths (demands on this size during interval) 7519 // where the interval is from the end of one sweep to the end of the 7520 // next. 7521 // 7522 // When sweeping the sweeper maintains an accumulated chunk which is 7523 // the chunk that is made up of chunks that have been coalesced. That 7524 // will be termed the left-hand chunk. A new chunk of garbage that 7525 // is being considered for coalescing will be referred to as the 7526 // right-hand chunk. 7527 // 7528 // When making a decision on whether to coalesce a right-hand chunk with 7529 // the current left-hand chunk, the current count vs. the desired count 7530 // of the left-hand chunk is considered. Also if the right-hand chunk 7531 // is near the large chunk at the end of the heap (see 7532 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7533 // left-hand chunk is coalesced. 7534 // 7535 // When making a decision about whether to split a chunk, the desired count 7536 // vs. the current count of the candidate to be split is also considered. 7537 // If the candidate is underpopulated (currently fewer chunks than desired) 7538 // a chunk of an overpopulated (currently more chunks than desired) size may 7539 // be chosen. The "hint" associated with a free list, if non-null, points 7540 // to a free list which may be overpopulated. 7541 // 7542 7543 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7544 const size_t size = fc->size(); 7545 // Chunks that cannot be coalesced are not in the 7546 // free lists. 7547 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7548 assert(_sp->verify_chunk_in_free_list(fc), 7549 "free chunk should be in free lists"); 7550 } 7551 // a chunk that is already free, should not have been 7552 // marked in the bit map 7553 HeapWord* const addr = (HeapWord*) fc; 7554 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7555 // Verify that the bit map has no bits marked between 7556 // addr and purported end of this block. 7557 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7558 7559 // Some chunks cannot be coalesced under any circumstances. 7560 // See the definition of cantCoalesce(). 7561 if (!fc->cantCoalesce()) { 7562 // This chunk can potentially be coalesced. 7563 if (_sp->adaptive_freelists()) { 7564 // All the work is done in 7565 do_post_free_or_garbage_chunk(fc, size); 7566 } else { // Not adaptive free lists 7567 // this is a free chunk that can potentially be coalesced by the sweeper; 7568 if (!inFreeRange()) { 7569 // if the next chunk is a free block that can't be coalesced 7570 // it doesn't make sense to remove this chunk from the free lists 7571 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7572 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7573 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7574 nextChunk->is_free() && // ... which is free... 7575 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7576 // nothing to do 7577 } else { 7578 // Potentially the start of a new free range: 7579 // Don't eagerly remove it from the free lists. 7580 // No need to remove it if it will just be put 7581 // back again. (Also from a pragmatic point of view 7582 // if it is a free block in a region that is beyond 7583 // any allocated blocks, an assertion will fail) 7584 // Remember the start of a free run. 7585 initialize_free_range(addr, true); 7586 // end - can coalesce with next chunk 7587 } 7588 } else { 7589 // the midst of a free range, we are coalescing 7590 print_free_block_coalesced(fc); 7591 if (CMSTraceSweeper) { 7592 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7593 } 7594 // remove it from the free lists 7595 _sp->removeFreeChunkFromFreeLists(fc); 7596 set_lastFreeRangeCoalesced(true); 7597 // If the chunk is being coalesced and the current free range is 7598 // in the free lists, remove the current free range so that it 7599 // will be returned to the free lists in its entirety - all 7600 // the coalesced pieces included. 7601 if (freeRangeInFreeLists()) { 7602 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7603 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7604 "Size of free range is inconsistent with chunk size."); 7605 if (CMSTestInFreeList) { 7606 assert(_sp->verify_chunk_in_free_list(ffc), 7607 "free range is not in free lists"); 7608 } 7609 _sp->removeFreeChunkFromFreeLists(ffc); 7610 set_freeRangeInFreeLists(false); 7611 } 7612 } 7613 } 7614 // Note that if the chunk is not coalescable (the else arm 7615 // below), we unconditionally flush, without needing to do 7616 // a "lookahead," as we do below. 7617 if (inFreeRange()) lookahead_and_flush(fc, size); 7618 } else { 7619 // Code path common to both original and adaptive free lists. 7620 7621 // cant coalesce with previous block; this should be treated 7622 // as the end of a free run if any 7623 if (inFreeRange()) { 7624 // we kicked some butt; time to pick up the garbage 7625 assert(freeFinger() < addr, "freeFinger points too high"); 7626 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7627 } 7628 // else, nothing to do, just continue 7629 } 7630 } 7631 7632 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7633 // This is a chunk of garbage. It is not in any free list. 7634 // Add it to a free list or let it possibly be coalesced into 7635 // a larger chunk. 7636 HeapWord* const addr = (HeapWord*) fc; 7637 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7638 7639 if (_sp->adaptive_freelists()) { 7640 // Verify that the bit map has no bits marked between 7641 // addr and purported end of just dead object. 7642 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7643 7644 do_post_free_or_garbage_chunk(fc, size); 7645 } else { 7646 if (!inFreeRange()) { 7647 // start of a new free range 7648 assert(size > 0, "A free range should have a size"); 7649 initialize_free_range(addr, false); 7650 } else { 7651 // this will be swept up when we hit the end of the 7652 // free range 7653 if (CMSTraceSweeper) { 7654 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7655 } 7656 // If the chunk is being coalesced and the current free range is 7657 // in the free lists, remove the current free range so that it 7658 // will be returned to the free lists in its entirety - all 7659 // the coalesced pieces included. 7660 if (freeRangeInFreeLists()) { 7661 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7662 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7663 "Size of free range is inconsistent with chunk size."); 7664 if (CMSTestInFreeList) { 7665 assert(_sp->verify_chunk_in_free_list(ffc), 7666 "free range is not in free lists"); 7667 } 7668 _sp->removeFreeChunkFromFreeLists(ffc); 7669 set_freeRangeInFreeLists(false); 7670 } 7671 set_lastFreeRangeCoalesced(true); 7672 } 7673 // this will be swept up when we hit the end of the free range 7674 7675 // Verify that the bit map has no bits marked between 7676 // addr and purported end of just dead object. 7677 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7678 } 7679 assert(_limit >= addr + size, 7680 "A freshly garbage chunk can't possibly straddle over _limit"); 7681 if (inFreeRange()) lookahead_and_flush(fc, size); 7682 return size; 7683 } 7684 7685 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7686 HeapWord* addr = (HeapWord*) fc; 7687 // The sweeper has just found a live object. Return any accumulated 7688 // left hand chunk to the free lists. 7689 if (inFreeRange()) { 7690 assert(freeFinger() < addr, "freeFinger points too high"); 7691 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7692 } 7693 7694 // This object is live: we'd normally expect this to be 7695 // an oop, and like to assert the following: 7696 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7697 // However, as we commented above, this may be an object whose 7698 // header hasn't yet been initialized. 7699 size_t size; 7700 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7701 if (_bitMap->isMarked(addr + 1)) { 7702 // Determine the size from the bit map, rather than trying to 7703 // compute it from the object header. 7704 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7705 size = pointer_delta(nextOneAddr + 1, addr); 7706 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7707 "alignment problem"); 7708 7709 #ifdef ASSERT 7710 if (oop(addr)->klass_or_null() != NULL) { 7711 // Ignore mark word because we are running concurrent with mutators 7712 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7713 assert(size == 7714 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7715 "P-mark and computed size do not agree"); 7716 } 7717 #endif 7718 7719 } else { 7720 // This should be an initialized object that's alive. 7721 assert(oop(addr)->klass_or_null() != NULL, 7722 "Should be an initialized object"); 7723 // Ignore mark word because we are running concurrent with mutators 7724 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7725 // Verify that the bit map has no bits marked between 7726 // addr and purported end of this block. 7727 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7728 assert(size >= 3, "Necessary for Printezis marks to work"); 7729 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7730 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7731 } 7732 return size; 7733 } 7734 7735 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7736 size_t chunkSize) { 7737 // do_post_free_or_garbage_chunk() should only be called in the case 7738 // of the adaptive free list allocator. 7739 const bool fcInFreeLists = fc->is_free(); 7740 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7741 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7742 if (CMSTestInFreeList && fcInFreeLists) { 7743 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7744 } 7745 7746 if (CMSTraceSweeper) { 7747 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7748 } 7749 7750 HeapWord* const fc_addr = (HeapWord*) fc; 7751 7752 bool coalesce; 7753 const size_t left = pointer_delta(fc_addr, freeFinger()); 7754 const size_t right = chunkSize; 7755 switch (FLSCoalescePolicy) { 7756 // numeric value forms a coalition aggressiveness metric 7757 case 0: { // never coalesce 7758 coalesce = false; 7759 break; 7760 } 7761 case 1: { // coalesce if left & right chunks on overpopulated lists 7762 coalesce = _sp->coalOverPopulated(left) && 7763 _sp->coalOverPopulated(right); 7764 break; 7765 } 7766 case 2: { // coalesce if left chunk on overpopulated list (default) 7767 coalesce = _sp->coalOverPopulated(left); 7768 break; 7769 } 7770 case 3: { // coalesce if left OR right chunk on overpopulated list 7771 coalesce = _sp->coalOverPopulated(left) || 7772 _sp->coalOverPopulated(right); 7773 break; 7774 } 7775 case 4: { // always coalesce 7776 coalesce = true; 7777 break; 7778 } 7779 default: 7780 ShouldNotReachHere(); 7781 } 7782 7783 // Should the current free range be coalesced? 7784 // If the chunk is in a free range and either we decided to coalesce above 7785 // or the chunk is near the large block at the end of the heap 7786 // (isNearLargestChunk() returns true), then coalesce this chunk. 7787 const bool doCoalesce = inFreeRange() 7788 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7789 if (doCoalesce) { 7790 // Coalesce the current free range on the left with the new 7791 // chunk on the right. If either is on a free list, 7792 // it must be removed from the list and stashed in the closure. 7793 if (freeRangeInFreeLists()) { 7794 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7795 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7796 "Size of free range is inconsistent with chunk size."); 7797 if (CMSTestInFreeList) { 7798 assert(_sp->verify_chunk_in_free_list(ffc), 7799 "Chunk is not in free lists"); 7800 } 7801 _sp->coalDeath(ffc->size()); 7802 _sp->removeFreeChunkFromFreeLists(ffc); 7803 set_freeRangeInFreeLists(false); 7804 } 7805 if (fcInFreeLists) { 7806 _sp->coalDeath(chunkSize); 7807 assert(fc->size() == chunkSize, 7808 "The chunk has the wrong size or is not in the free lists"); 7809 _sp->removeFreeChunkFromFreeLists(fc); 7810 } 7811 set_lastFreeRangeCoalesced(true); 7812 print_free_block_coalesced(fc); 7813 } else { // not in a free range and/or should not coalesce 7814 // Return the current free range and start a new one. 7815 if (inFreeRange()) { 7816 // In a free range but cannot coalesce with the right hand chunk. 7817 // Put the current free range into the free lists. 7818 flush_cur_free_chunk(freeFinger(), 7819 pointer_delta(fc_addr, freeFinger())); 7820 } 7821 // Set up for new free range. Pass along whether the right hand 7822 // chunk is in the free lists. 7823 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7824 } 7825 } 7826 7827 // Lookahead flush: 7828 // If we are tracking a free range, and this is the last chunk that 7829 // we'll look at because its end crosses past _limit, we'll preemptively 7830 // flush it along with any free range we may be holding on to. Note that 7831 // this can be the case only for an already free or freshly garbage 7832 // chunk. If this block is an object, it can never straddle 7833 // over _limit. The "straddling" occurs when _limit is set at 7834 // the previous end of the space when this cycle started, and 7835 // a subsequent heap expansion caused the previously co-terminal 7836 // free block to be coalesced with the newly expanded portion, 7837 // thus rendering _limit a non-block-boundary making it dangerous 7838 // for the sweeper to step over and examine. 7839 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7840 assert(inFreeRange(), "Should only be called if currently in a free range."); 7841 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7842 assert(_sp->used_region().contains(eob - 1), 7843 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7844 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7845 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7846 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7847 if (eob >= _limit) { 7848 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7849 if (CMSTraceSweeper) { 7850 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7851 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7852 "[" PTR_FORMAT "," PTR_FORMAT ")", 7853 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7854 } 7855 // Return the storage we are tracking back into the free lists. 7856 if (CMSTraceSweeper) { 7857 gclog_or_tty->print_cr("Flushing ... "); 7858 } 7859 assert(freeFinger() < eob, "Error"); 7860 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7861 } 7862 } 7863 7864 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7865 assert(inFreeRange(), "Should only be called if currently in a free range."); 7866 assert(size > 0, 7867 "A zero sized chunk cannot be added to the free lists."); 7868 if (!freeRangeInFreeLists()) { 7869 if (CMSTestInFreeList) { 7870 FreeChunk* fc = (FreeChunk*) chunk; 7871 fc->set_size(size); 7872 assert(!_sp->verify_chunk_in_free_list(fc), 7873 "chunk should not be in free lists yet"); 7874 } 7875 if (CMSTraceSweeper) { 7876 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7877 p2i(chunk), size); 7878 } 7879 // A new free range is going to be starting. The current 7880 // free range has not been added to the free lists yet or 7881 // was removed so add it back. 7882 // If the current free range was coalesced, then the death 7883 // of the free range was recorded. Record a birth now. 7884 if (lastFreeRangeCoalesced()) { 7885 _sp->coalBirth(size); 7886 } 7887 _sp->addChunkAndRepairOffsetTable(chunk, size, 7888 lastFreeRangeCoalesced()); 7889 } else if (CMSTraceSweeper) { 7890 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7891 } 7892 set_inFreeRange(false); 7893 set_freeRangeInFreeLists(false); 7894 } 7895 7896 // We take a break if we've been at this for a while, 7897 // so as to avoid monopolizing the locks involved. 7898 void SweepClosure::do_yield_work(HeapWord* addr) { 7899 // Return current free chunk being used for coalescing (if any) 7900 // to the appropriate freelist. After yielding, the next 7901 // free block encountered will start a coalescing range of 7902 // free blocks. If the next free block is adjacent to the 7903 // chunk just flushed, they will need to wait for the next 7904 // sweep to be coalesced. 7905 if (inFreeRange()) { 7906 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7907 } 7908 7909 // First give up the locks, then yield, then re-lock. 7910 // We should probably use a constructor/destructor idiom to 7911 // do this unlock/lock or modify the MutexUnlocker class to 7912 // serve our purpose. XXX 7913 assert_lock_strong(_bitMap->lock()); 7914 assert_lock_strong(_freelistLock); 7915 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7916 "CMS thread should hold CMS token"); 7917 _bitMap->lock()->unlock(); 7918 _freelistLock->unlock(); 7919 ConcurrentMarkSweepThread::desynchronize(true); 7920 _collector->stopTimer(); 7921 if (PrintCMSStatistics != 0) { 7922 _collector->incrementYields(); 7923 } 7924 7925 // See the comment in coordinator_yield() 7926 for (unsigned i = 0; i < CMSYieldSleepCount && 7927 ConcurrentMarkSweepThread::should_yield() && 7928 !CMSCollector::foregroundGCIsActive(); ++i) { 7929 os::sleep(Thread::current(), 1, false); 7930 } 7931 7932 ConcurrentMarkSweepThread::synchronize(true); 7933 _freelistLock->lock(); 7934 _bitMap->lock()->lock_without_safepoint_check(); 7935 _collector->startTimer(); 7936 } 7937 7938 #ifndef PRODUCT 7939 // This is actually very useful in a product build if it can 7940 // be called from the debugger. Compile it into the product 7941 // as needed. 7942 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7943 return debug_cms_space->verify_chunk_in_free_list(fc); 7944 } 7945 #endif 7946 7947 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7948 if (CMSTraceSweeper) { 7949 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7950 p2i(fc), fc->size()); 7951 } 7952 } 7953 7954 // CMSIsAliveClosure 7955 bool CMSIsAliveClosure::do_object_b(oop obj) { 7956 HeapWord* addr = (HeapWord*)obj; 7957 return addr != NULL && 7958 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7959 } 7960 7961 7962 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7963 MemRegion span, 7964 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7965 bool cpc): 7966 _collector(collector), 7967 _span(span), 7968 _bit_map(bit_map), 7969 _mark_stack(mark_stack), 7970 _concurrent_precleaning(cpc) { 7971 assert(!_span.is_empty(), "Empty span could spell trouble"); 7972 } 7973 7974 7975 // CMSKeepAliveClosure: the serial version 7976 void CMSKeepAliveClosure::do_oop(oop obj) { 7977 HeapWord* addr = (HeapWord*)obj; 7978 if (_span.contains(addr) && 7979 !_bit_map->isMarked(addr)) { 7980 _bit_map->mark(addr); 7981 bool simulate_overflow = false; 7982 NOT_PRODUCT( 7983 if (CMSMarkStackOverflowALot && 7984 _collector->simulate_overflow()) { 7985 // simulate a stack overflow 7986 simulate_overflow = true; 7987 } 7988 ) 7989 if (simulate_overflow || !_mark_stack->push(obj)) { 7990 if (_concurrent_precleaning) { 7991 // We dirty the overflown object and let the remark 7992 // phase deal with it. 7993 assert(_collector->overflow_list_is_empty(), "Error"); 7994 // In the case of object arrays, we need to dirty all of 7995 // the cards that the object spans. No locking or atomics 7996 // are needed since no one else can be mutating the mod union 7997 // table. 7998 if (obj->is_objArray()) { 7999 size_t sz = obj->size(); 8000 HeapWord* end_card_addr = 8001 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8002 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8003 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8004 _collector->_modUnionTable.mark_range(redirty_range); 8005 } else { 8006 _collector->_modUnionTable.mark(addr); 8007 } 8008 _collector->_ser_kac_preclean_ovflw++; 8009 } else { 8010 _collector->push_on_overflow_list(obj); 8011 _collector->_ser_kac_ovflw++; 8012 } 8013 } 8014 } 8015 } 8016 8017 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8018 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8019 8020 // CMSParKeepAliveClosure: a parallel version of the above. 8021 // The work queues are private to each closure (thread), 8022 // but (may be) available for stealing by other threads. 8023 void CMSParKeepAliveClosure::do_oop(oop obj) { 8024 HeapWord* addr = (HeapWord*)obj; 8025 if (_span.contains(addr) && 8026 !_bit_map->isMarked(addr)) { 8027 // In general, during recursive tracing, several threads 8028 // may be concurrently getting here; the first one to 8029 // "tag" it, claims it. 8030 if (_bit_map->par_mark(addr)) { 8031 bool res = _work_queue->push(obj); 8032 assert(res, "Low water mark should be much less than capacity"); 8033 // Do a recursive trim in the hope that this will keep 8034 // stack usage lower, but leave some oops for potential stealers 8035 trim_queue(_low_water_mark); 8036 } // Else, another thread got there first 8037 } 8038 } 8039 8040 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8041 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8042 8043 void CMSParKeepAliveClosure::trim_queue(uint max) { 8044 while (_work_queue->size() > max) { 8045 oop new_oop; 8046 if (_work_queue->pop_local(new_oop)) { 8047 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8048 assert(_bit_map->isMarked((HeapWord*)new_oop), 8049 "no white objects on this stack!"); 8050 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8051 // iterate over the oops in this oop, marking and pushing 8052 // the ones in CMS heap (i.e. in _span). 8053 new_oop->oop_iterate(&_mark_and_push); 8054 } 8055 } 8056 } 8057 8058 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8059 CMSCollector* collector, 8060 MemRegion span, CMSBitMap* bit_map, 8061 OopTaskQueue* work_queue): 8062 _collector(collector), 8063 _span(span), 8064 _bit_map(bit_map), 8065 _work_queue(work_queue) { } 8066 8067 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8068 HeapWord* addr = (HeapWord*)obj; 8069 if (_span.contains(addr) && 8070 !_bit_map->isMarked(addr)) { 8071 if (_bit_map->par_mark(addr)) { 8072 bool simulate_overflow = false; 8073 NOT_PRODUCT( 8074 if (CMSMarkStackOverflowALot && 8075 _collector->par_simulate_overflow()) { 8076 // simulate a stack overflow 8077 simulate_overflow = true; 8078 } 8079 ) 8080 if (simulate_overflow || !_work_queue->push(obj)) { 8081 _collector->par_push_on_overflow_list(obj); 8082 _collector->_par_kac_ovflw++; 8083 } 8084 } // Else another thread got there already 8085 } 8086 } 8087 8088 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8089 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8090 8091 ////////////////////////////////////////////////////////////////// 8092 // CMSExpansionCause ///////////////////////////// 8093 ////////////////////////////////////////////////////////////////// 8094 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8095 switch (cause) { 8096 case _no_expansion: 8097 return "No expansion"; 8098 case _satisfy_free_ratio: 8099 return "Free ratio"; 8100 case _satisfy_promotion: 8101 return "Satisfy promotion"; 8102 case _satisfy_allocation: 8103 return "allocation"; 8104 case _allocate_par_lab: 8105 return "Par LAB"; 8106 case _allocate_par_spooling_space: 8107 return "Par Spooling Space"; 8108 case _adaptive_size_policy: 8109 return "Ergonomics"; 8110 default: 8111 return "unknown"; 8112 } 8113 } 8114 8115 void CMSDrainMarkingStackClosure::do_void() { 8116 // the max number to take from overflow list at a time 8117 const size_t num = _mark_stack->capacity()/4; 8118 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8119 "Overflow list should be NULL during concurrent phases"); 8120 while (!_mark_stack->isEmpty() || 8121 // if stack is empty, check the overflow list 8122 _collector->take_from_overflow_list(num, _mark_stack)) { 8123 oop obj = _mark_stack->pop(); 8124 HeapWord* addr = (HeapWord*)obj; 8125 assert(_span.contains(addr), "Should be within span"); 8126 assert(_bit_map->isMarked(addr), "Should be marked"); 8127 assert(obj->is_oop(), "Should be an oop"); 8128 obj->oop_iterate(_keep_alive); 8129 } 8130 } 8131 8132 void CMSParDrainMarkingStackClosure::do_void() { 8133 // drain queue 8134 trim_queue(0); 8135 } 8136 8137 // Trim our work_queue so its length is below max at return 8138 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8139 while (_work_queue->size() > max) { 8140 oop new_oop; 8141 if (_work_queue->pop_local(new_oop)) { 8142 assert(new_oop->is_oop(), "Expected an oop"); 8143 assert(_bit_map->isMarked((HeapWord*)new_oop), 8144 "no white objects on this stack!"); 8145 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8146 // iterate over the oops in this oop, marking and pushing 8147 // the ones in CMS heap (i.e. in _span). 8148 new_oop->oop_iterate(&_mark_and_push); 8149 } 8150 } 8151 } 8152 8153 //////////////////////////////////////////////////////////////////// 8154 // Support for Marking Stack Overflow list handling and related code 8155 //////////////////////////////////////////////////////////////////// 8156 // Much of the following code is similar in shape and spirit to the 8157 // code used in ParNewGC. We should try and share that code 8158 // as much as possible in the future. 8159 8160 #ifndef PRODUCT 8161 // Debugging support for CMSStackOverflowALot 8162 8163 // It's OK to call this multi-threaded; the worst thing 8164 // that can happen is that we'll get a bunch of closely 8165 // spaced simulated overflows, but that's OK, in fact 8166 // probably good as it would exercise the overflow code 8167 // under contention. 8168 bool CMSCollector::simulate_overflow() { 8169 if (_overflow_counter-- <= 0) { // just being defensive 8170 _overflow_counter = CMSMarkStackOverflowInterval; 8171 return true; 8172 } else { 8173 return false; 8174 } 8175 } 8176 8177 bool CMSCollector::par_simulate_overflow() { 8178 return simulate_overflow(); 8179 } 8180 #endif 8181 8182 // Single-threaded 8183 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8184 assert(stack->isEmpty(), "Expected precondition"); 8185 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8186 size_t i = num; 8187 oop cur = _overflow_list; 8188 const markOop proto = markOopDesc::prototype(); 8189 NOT_PRODUCT(ssize_t n = 0;) 8190 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8191 next = oop(cur->mark()); 8192 cur->set_mark(proto); // until proven otherwise 8193 assert(cur->is_oop(), "Should be an oop"); 8194 bool res = stack->push(cur); 8195 assert(res, "Bit off more than can chew?"); 8196 NOT_PRODUCT(n++;) 8197 } 8198 _overflow_list = cur; 8199 #ifndef PRODUCT 8200 assert(_num_par_pushes >= n, "Too many pops?"); 8201 _num_par_pushes -=n; 8202 #endif 8203 return !stack->isEmpty(); 8204 } 8205 8206 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8207 // (MT-safe) Get a prefix of at most "num" from the list. 8208 // The overflow list is chained through the mark word of 8209 // each object in the list. We fetch the entire list, 8210 // break off a prefix of the right size and return the 8211 // remainder. If other threads try to take objects from 8212 // the overflow list at that time, they will wait for 8213 // some time to see if data becomes available. If (and 8214 // only if) another thread places one or more object(s) 8215 // on the global list before we have returned the suffix 8216 // to the global list, we will walk down our local list 8217 // to find its end and append the global list to 8218 // our suffix before returning it. This suffix walk can 8219 // prove to be expensive (quadratic in the amount of traffic) 8220 // when there are many objects in the overflow list and 8221 // there is much producer-consumer contention on the list. 8222 // *NOTE*: The overflow list manipulation code here and 8223 // in ParNewGeneration:: are very similar in shape, 8224 // except that in the ParNew case we use the old (from/eden) 8225 // copy of the object to thread the list via its klass word. 8226 // Because of the common code, if you make any changes in 8227 // the code below, please check the ParNew version to see if 8228 // similar changes might be needed. 8229 // CR 6797058 has been filed to consolidate the common code. 8230 bool CMSCollector::par_take_from_overflow_list(size_t num, 8231 OopTaskQueue* work_q, 8232 int no_of_gc_threads) { 8233 assert(work_q->size() == 0, "First empty local work queue"); 8234 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8235 if (_overflow_list == NULL) { 8236 return false; 8237 } 8238 // Grab the entire list; we'll put back a suffix 8239 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8240 Thread* tid = Thread::current(); 8241 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8242 // set to ParallelGCThreads. 8243 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8244 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8245 // If the list is busy, we spin for a short while, 8246 // sleeping between attempts to get the list. 8247 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8248 os::sleep(tid, sleep_time_millis, false); 8249 if (_overflow_list == NULL) { 8250 // Nothing left to take 8251 return false; 8252 } else if (_overflow_list != BUSY) { 8253 // Try and grab the prefix 8254 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8255 } 8256 } 8257 // If the list was found to be empty, or we spun long 8258 // enough, we give up and return empty-handed. If we leave 8259 // the list in the BUSY state below, it must be the case that 8260 // some other thread holds the overflow list and will set it 8261 // to a non-BUSY state in the future. 8262 if (prefix == NULL || prefix == BUSY) { 8263 // Nothing to take or waited long enough 8264 if (prefix == NULL) { 8265 // Write back the NULL in case we overwrote it with BUSY above 8266 // and it is still the same value. 8267 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8268 } 8269 return false; 8270 } 8271 assert(prefix != NULL && prefix != BUSY, "Error"); 8272 size_t i = num; 8273 oop cur = prefix; 8274 // Walk down the first "num" objects, unless we reach the end. 8275 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8276 if (cur->mark() == NULL) { 8277 // We have "num" or fewer elements in the list, so there 8278 // is nothing to return to the global list. 8279 // Write back the NULL in lieu of the BUSY we wrote 8280 // above, if it is still the same value. 8281 if (_overflow_list == BUSY) { 8282 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8283 } 8284 } else { 8285 // Chop off the suffix and return it to the global list. 8286 assert(cur->mark() != BUSY, "Error"); 8287 oop suffix_head = cur->mark(); // suffix will be put back on global list 8288 cur->set_mark(NULL); // break off suffix 8289 // It's possible that the list is still in the empty(busy) state 8290 // we left it in a short while ago; in that case we may be 8291 // able to place back the suffix without incurring the cost 8292 // of a walk down the list. 8293 oop observed_overflow_list = _overflow_list; 8294 oop cur_overflow_list = observed_overflow_list; 8295 bool attached = false; 8296 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8297 observed_overflow_list = 8298 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8299 if (cur_overflow_list == observed_overflow_list) { 8300 attached = true; 8301 break; 8302 } else cur_overflow_list = observed_overflow_list; 8303 } 8304 if (!attached) { 8305 // Too bad, someone else sneaked in (at least) an element; we'll need 8306 // to do a splice. Find tail of suffix so we can prepend suffix to global 8307 // list. 8308 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8309 oop suffix_tail = cur; 8310 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8311 "Tautology"); 8312 observed_overflow_list = _overflow_list; 8313 do { 8314 cur_overflow_list = observed_overflow_list; 8315 if (cur_overflow_list != BUSY) { 8316 // Do the splice ... 8317 suffix_tail->set_mark(markOop(cur_overflow_list)); 8318 } else { // cur_overflow_list == BUSY 8319 suffix_tail->set_mark(NULL); 8320 } 8321 // ... and try to place spliced list back on overflow_list ... 8322 observed_overflow_list = 8323 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8324 } while (cur_overflow_list != observed_overflow_list); 8325 // ... until we have succeeded in doing so. 8326 } 8327 } 8328 8329 // Push the prefix elements on work_q 8330 assert(prefix != NULL, "control point invariant"); 8331 const markOop proto = markOopDesc::prototype(); 8332 oop next; 8333 NOT_PRODUCT(ssize_t n = 0;) 8334 for (cur = prefix; cur != NULL; cur = next) { 8335 next = oop(cur->mark()); 8336 cur->set_mark(proto); // until proven otherwise 8337 assert(cur->is_oop(), "Should be an oop"); 8338 bool res = work_q->push(cur); 8339 assert(res, "Bit off more than we can chew?"); 8340 NOT_PRODUCT(n++;) 8341 } 8342 #ifndef PRODUCT 8343 assert(_num_par_pushes >= n, "Too many pops?"); 8344 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8345 #endif 8346 return true; 8347 } 8348 8349 // Single-threaded 8350 void CMSCollector::push_on_overflow_list(oop p) { 8351 NOT_PRODUCT(_num_par_pushes++;) 8352 assert(p->is_oop(), "Not an oop"); 8353 preserve_mark_if_necessary(p); 8354 p->set_mark((markOop)_overflow_list); 8355 _overflow_list = p; 8356 } 8357 8358 // Multi-threaded; use CAS to prepend to overflow list 8359 void CMSCollector::par_push_on_overflow_list(oop p) { 8360 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8361 assert(p->is_oop(), "Not an oop"); 8362 par_preserve_mark_if_necessary(p); 8363 oop observed_overflow_list = _overflow_list; 8364 oop cur_overflow_list; 8365 do { 8366 cur_overflow_list = observed_overflow_list; 8367 if (cur_overflow_list != BUSY) { 8368 p->set_mark(markOop(cur_overflow_list)); 8369 } else { 8370 p->set_mark(NULL); 8371 } 8372 observed_overflow_list = 8373 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8374 } while (cur_overflow_list != observed_overflow_list); 8375 } 8376 #undef BUSY 8377 8378 // Single threaded 8379 // General Note on GrowableArray: pushes may silently fail 8380 // because we are (temporarily) out of C-heap for expanding 8381 // the stack. The problem is quite ubiquitous and affects 8382 // a lot of code in the JVM. The prudent thing for GrowableArray 8383 // to do (for now) is to exit with an error. However, that may 8384 // be too draconian in some cases because the caller may be 8385 // able to recover without much harm. For such cases, we 8386 // should probably introduce a "soft_push" method which returns 8387 // an indication of success or failure with the assumption that 8388 // the caller may be able to recover from a failure; code in 8389 // the VM can then be changed, incrementally, to deal with such 8390 // failures where possible, thus, incrementally hardening the VM 8391 // in such low resource situations. 8392 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8393 _preserved_oop_stack.push(p); 8394 _preserved_mark_stack.push(m); 8395 assert(m == p->mark(), "Mark word changed"); 8396 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8397 "bijection"); 8398 } 8399 8400 // Single threaded 8401 void CMSCollector::preserve_mark_if_necessary(oop p) { 8402 markOop m = p->mark(); 8403 if (m->must_be_preserved(p)) { 8404 preserve_mark_work(p, m); 8405 } 8406 } 8407 8408 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8409 markOop m = p->mark(); 8410 if (m->must_be_preserved(p)) { 8411 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8412 // Even though we read the mark word without holding 8413 // the lock, we are assured that it will not change 8414 // because we "own" this oop, so no other thread can 8415 // be trying to push it on the overflow list; see 8416 // the assertion in preserve_mark_work() that checks 8417 // that m == p->mark(). 8418 preserve_mark_work(p, m); 8419 } 8420 } 8421 8422 // We should be able to do this multi-threaded, 8423 // a chunk of stack being a task (this is 8424 // correct because each oop only ever appears 8425 // once in the overflow list. However, it's 8426 // not very easy to completely overlap this with 8427 // other operations, so will generally not be done 8428 // until all work's been completed. Because we 8429 // expect the preserved oop stack (set) to be small, 8430 // it's probably fine to do this single-threaded. 8431 // We can explore cleverer concurrent/overlapped/parallel 8432 // processing of preserved marks if we feel the 8433 // need for this in the future. Stack overflow should 8434 // be so rare in practice and, when it happens, its 8435 // effect on performance so great that this will 8436 // likely just be in the noise anyway. 8437 void CMSCollector::restore_preserved_marks_if_any() { 8438 assert(SafepointSynchronize::is_at_safepoint(), 8439 "world should be stopped"); 8440 assert(Thread::current()->is_ConcurrentGC_thread() || 8441 Thread::current()->is_VM_thread(), 8442 "should be single-threaded"); 8443 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8444 "bijection"); 8445 8446 while (!_preserved_oop_stack.is_empty()) { 8447 oop p = _preserved_oop_stack.pop(); 8448 assert(p->is_oop(), "Should be an oop"); 8449 assert(_span.contains(p), "oop should be in _span"); 8450 assert(p->mark() == markOopDesc::prototype(), 8451 "Set when taken from overflow list"); 8452 markOop m = _preserved_mark_stack.pop(); 8453 p->set_mark(m); 8454 } 8455 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8456 "stacks were cleared above"); 8457 } 8458 8459 #ifndef PRODUCT 8460 bool CMSCollector::no_preserved_marks() const { 8461 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8462 } 8463 #endif 8464 8465 // Transfer some number of overflown objects to usual marking 8466 // stack. Return true if some objects were transferred. 8467 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8468 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8469 (size_t)ParGCDesiredObjsFromOverflowList); 8470 8471 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8472 assert(_collector->overflow_list_is_empty() || res, 8473 "If list is not empty, we should have taken something"); 8474 assert(!res || !_mark_stack->isEmpty(), 8475 "If we took something, it should now be on our stack"); 8476 return res; 8477 } 8478 8479 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8480 size_t res = _sp->block_size_no_stall(addr, _collector); 8481 if (_sp->block_is_obj(addr)) { 8482 if (_live_bit_map->isMarked(addr)) { 8483 // It can't have been dead in a previous cycle 8484 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8485 } else { 8486 _dead_bit_map->mark(addr); // mark the dead object 8487 } 8488 } 8489 // Could be 0, if the block size could not be computed without stalling. 8490 return res; 8491 } 8492 8493 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8494 8495 switch (phase) { 8496 case CMSCollector::InitialMarking: 8497 initialize(true /* fullGC */ , 8498 cause /* cause of the GC */, 8499 true /* recordGCBeginTime */, 8500 true /* recordPreGCUsage */, 8501 false /* recordPeakUsage */, 8502 false /* recordPostGCusage */, 8503 true /* recordAccumulatedGCTime */, 8504 false /* recordGCEndTime */, 8505 false /* countCollection */ ); 8506 break; 8507 8508 case CMSCollector::FinalMarking: 8509 initialize(true /* fullGC */ , 8510 cause /* cause of the GC */, 8511 false /* recordGCBeginTime */, 8512 false /* recordPreGCUsage */, 8513 false /* recordPeakUsage */, 8514 false /* recordPostGCusage */, 8515 true /* recordAccumulatedGCTime */, 8516 false /* recordGCEndTime */, 8517 false /* countCollection */ ); 8518 break; 8519 8520 case CMSCollector::Sweeping: 8521 initialize(true /* fullGC */ , 8522 cause /* cause of the GC */, 8523 false /* recordGCBeginTime */, 8524 false /* recordPreGCUsage */, 8525 true /* recordPeakUsage */, 8526 true /* recordPostGCusage */, 8527 false /* recordAccumulatedGCTime */, 8528 true /* recordGCEndTime */, 8529 true /* countCollection */ ); 8530 break; 8531 8532 default: 8533 ShouldNotReachHere(); 8534 } 8535 }