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