1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/cms/cmsCollectorPolicy.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/compactibleFreeListSpace.hpp" 33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 34 #include "gc/cms/concurrentMarkSweepThread.hpp" 35 #include "gc/cms/parNewGeneration.hpp" 36 #include "gc/cms/vmCMSOperations.hpp" 37 #include "gc/serial/genMarkSweep.hpp" 38 #include "gc/serial/tenuredGeneration.hpp" 39 #include "gc/shared/adaptiveSizePolicy.hpp" 40 #include "gc/shared/cardGeneration.inline.hpp" 41 #include "gc/shared/cardTableRS.hpp" 42 #include "gc/shared/collectedHeap.inline.hpp" 43 #include "gc/shared/collectorCounters.hpp" 44 #include "gc/shared/collectorPolicy.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/referencePolicy.hpp" 54 #include "gc/shared/strongRootsScope.hpp" 55 #include "gc/shared/taskqueue.inline.hpp" 56 #include "memory/allocation.hpp" 57 #include "memory/iterator.inline.hpp" 58 #include "memory/padded.hpp" 59 #include "memory/resourceArea.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "prims/jvmtiExport.hpp" 62 #include "runtime/atomic.inline.hpp" 63 #include "runtime/globals_extension.hpp" 64 #include "runtime/handles.inline.hpp" 65 #include "runtime/java.hpp" 66 #include "runtime/orderAccess.inline.hpp" 67 #include "runtime/vmThread.hpp" 68 #include "services/memoryService.hpp" 69 #include "services/runtimeService.hpp" 70 #include "utilities/stack.inline.hpp" 71 72 // statics 73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 74 bool CMSCollector::_full_gc_requested = false; 75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 76 77 ////////////////////////////////////////////////////////////////// 78 // In support of CMS/VM thread synchronization 79 ////////////////////////////////////////////////////////////////// 80 // We split use of the CGC_lock into 2 "levels". 81 // The low-level locking is of the usual CGC_lock monitor. We introduce 82 // a higher level "token" (hereafter "CMS token") built on top of the 83 // low level monitor (hereafter "CGC lock"). 84 // The token-passing protocol gives priority to the VM thread. The 85 // CMS-lock doesn't provide any fairness guarantees, but clients 86 // should ensure that it is only held for very short, bounded 87 // durations. 88 // 89 // When either of the CMS thread or the VM thread is involved in 90 // collection operations during which it does not want the other 91 // thread to interfere, it obtains the CMS token. 92 // 93 // If either thread tries to get the token while the other has 94 // it, that thread waits. However, if the VM thread and CMS thread 95 // both want the token, then the VM thread gets priority while the 96 // CMS thread waits. This ensures, for instance, that the "concurrent" 97 // phases of the CMS thread's work do not block out the VM thread 98 // for long periods of time as the CMS thread continues to hog 99 // the token. (See bug 4616232). 100 // 101 // The baton-passing functions are, however, controlled by the 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 103 // and here the low-level CMS lock, not the high level token, 104 // ensures mutual exclusion. 105 // 106 // Two important conditions that we have to satisfy: 107 // 1. if a thread does a low-level wait on the CMS lock, then it 108 // relinquishes the CMS token if it were holding that token 109 // when it acquired the low-level CMS lock. 110 // 2. any low-level notifications on the low-level lock 111 // should only be sent when a thread has relinquished the token. 112 // 113 // In the absence of either property, we'd have potential deadlock. 114 // 115 // We protect each of the CMS (concurrent and sequential) phases 116 // with the CMS _token_, not the CMS _lock_. 117 // 118 // The only code protected by CMS lock is the token acquisition code 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 120 // baton-passing code. 121 // 122 // Unfortunately, i couldn't come up with a good abstraction to factor and 123 // hide the naked CGC_lock manipulation in the baton-passing code 124 // further below. That's something we should try to do. Also, the proof 125 // of correctness of this 2-level locking scheme is far from obvious, 126 // and potentially quite slippery. We have an uneasy suspicion, for instance, 127 // that there may be a theoretical possibility of delay/starvation in the 128 // low-level lock/wait/notify scheme used for the baton-passing because of 129 // potential interference with the priority scheme embodied in the 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 131 // invocation further below and marked with "XXX 20011219YSR". 132 // Indeed, as we note elsewhere, this may become yet more slippery 133 // in the presence of multiple CMS and/or multiple VM threads. XXX 134 135 class CMSTokenSync: public StackObj { 136 private: 137 bool _is_cms_thread; 138 public: 139 CMSTokenSync(bool is_cms_thread): 140 _is_cms_thread(is_cms_thread) { 141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 142 "Incorrect argument to constructor"); 143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 144 } 145 146 ~CMSTokenSync() { 147 assert(_is_cms_thread ? 148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 150 "Incorrect state"); 151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 152 } 153 }; 154 155 // Convenience class that does a CMSTokenSync, and then acquires 156 // upto three locks. 157 class CMSTokenSyncWithLocks: public CMSTokenSync { 158 private: 159 // Note: locks are acquired in textual declaration order 160 // and released in the opposite order 161 MutexLockerEx _locker1, _locker2, _locker3; 162 public: 163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 165 CMSTokenSync(is_cms_thread), 166 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 167 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 168 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 169 { } 170 }; 171 172 173 ////////////////////////////////////////////////////////////////// 174 // Concurrent Mark-Sweep Generation ///////////////////////////// 175 ////////////////////////////////////////////////////////////////// 176 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 178 179 // This struct contains per-thread things necessary to support parallel 180 // young-gen collection. 181 class CMSParGCThreadState: public CHeapObj<mtGC> { 182 public: 183 CFLS_LAB lab; 184 PromotionInfo promo; 185 186 // Constructor. 187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 188 promo.setSpace(cfls); 189 } 190 }; 191 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 193 ReservedSpace rs, size_t initial_byte_size, 194 CardTableRS* ct, bool use_adaptive_freelists, 195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 196 CardGeneration(rs, initial_byte_size, ct), 197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 198 _did_compact(false) 199 { 200 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 201 HeapWord* end = (HeapWord*) _virtual_space.high(); 202 203 _direct_allocated_words = 0; 204 NOT_PRODUCT( 205 _numObjectsPromoted = 0; 206 _numWordsPromoted = 0; 207 _numObjectsAllocated = 0; 208 _numWordsAllocated = 0; 209 ) 210 211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 212 use_adaptive_freelists, 213 dictionaryChoice); 214 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 215 _cmsSpace->_gen = this; 216 217 _gc_stats = new CMSGCStats(); 218 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 220 // offsets match. The ability to tell free chunks from objects 221 // depends on this property. 222 debug_only( 223 FreeChunk* junk = NULL; 224 assert(UseCompressedClassPointers || 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 226 "Offset of FreeChunk::_prev within FreeChunk must match" 227 " that of OopDesc::_klass within OopDesc"); 228 ) 229 230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 231 for (uint i = 0; i < ParallelGCThreads; i++) { 232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 233 } 234 235 _incremental_collection_failed = false; 236 // The "dilatation_factor" is the expansion that can occur on 237 // account of the fact that the minimum object size in the CMS 238 // generation may be larger than that in, say, a contiguous young 239 // generation. 240 // Ideally, in the calculation below, we'd compute the dilatation 241 // factor as: MinChunkSize/(promoting_gen's min object size) 242 // Since we do not have such a general query interface for the 243 // promoting generation, we'll instead just use the minimum 244 // object size (which today is a header's worth of space); 245 // note that all arithmetic is in units of HeapWords. 246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 247 assert(_dilatation_factor >= 1.0, "from previous assert"); 248 } 249 250 251 // The field "_initiating_occupancy" represents the occupancy percentage 252 // at which we trigger a new collection cycle. Unless explicitly specified 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it 254 // is calculated by: 255 // 256 // Let "f" be MinHeapFreeRatio in 257 // 258 // _initiating_occupancy = 100-f + 259 // f * (CMSTriggerRatio/100) 260 // where CMSTriggerRatio is the argument "tr" below. 261 // 262 // That is, if we assume the heap is at its desired maximum occupancy at the 263 // end of a collection, we let CMSTriggerRatio of the (purported) free 264 // space be allocated before initiating a new collection cycle. 265 // 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 267 assert(io <= 100 && tr <= 100, "Check the arguments"); 268 if (io >= 0) { 269 _initiating_occupancy = (double)io / 100.0; 270 } else { 271 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 272 (double)(tr * MinHeapFreeRatio) / 100.0) 273 / 100.0; 274 } 275 } 276 277 void ConcurrentMarkSweepGeneration::ref_processor_init() { 278 assert(collector() != NULL, "no collector"); 279 collector()->ref_processor_init(); 280 } 281 282 void CMSCollector::ref_processor_init() { 283 if (_ref_processor == NULL) { 284 // Allocate and initialize a reference processor 285 _ref_processor = 286 new ReferenceProcessor(_span, // span 287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 288 ParallelGCThreads, // mt processing degree 289 _cmsGen->refs_discovery_is_mt(), // mt discovery 290 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 292 &_is_alive_closure); // closure for liveness info 293 // Initialize the _ref_processor field of CMSGen 294 _cmsGen->set_ref_processor(_ref_processor); 295 296 } 297 } 298 299 AdaptiveSizePolicy* CMSCollector::size_policy() { 300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 301 return gch->gen_policy()->size_policy(); 302 } 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 308 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next minor collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 if (PrintGCDetails && Verbose) { 375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 377 cms_free, expected_promotion); 378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 379 cms_free_dbl, cms_consumption_rate() + 1.0); 380 } 381 // Add 1 in case the consumption rate goes to zero. 382 return cms_free_dbl / (cms_consumption_rate() + 1.0); 383 } 384 return 0.0; 385 } 386 387 // Compare the duration of the cms collection to the 388 // time remaining before the cms generation is empty. 389 // Note that the time from the start of the cms collection 390 // to the start of the cms sweep (less than the total 391 // duration of the cms collection) can be used. This 392 // has been tried and some applications experienced 393 // promotion failures early in execution. This was 394 // possibly because the averages were not accurate 395 // enough at the beginning. 396 double CMSStats::time_until_cms_start() const { 397 // We add "gc0_period" to the "work" calculation 398 // below because this query is done (mostly) at the 399 // end of a scavenge, so we need to conservatively 400 // account for that much possible delay 401 // in the query so as to avoid concurrent mode failures 402 // due to starting the collection just a wee bit too 403 // late. 404 double work = cms_duration() + gc0_period(); 405 double deadline = time_until_cms_gen_full(); 406 // If a concurrent mode failure occurred recently, we want to be 407 // more conservative and halve our expected time_until_cms_gen_full() 408 if (work > deadline) { 409 if (Verbose && PrintGCDetails) { 410 gclog_or_tty->print( 411 " CMSCollector: collect because of anticipated promotion " 412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 413 gc0_period(), time_until_cms_gen_full()); 414 } 415 return 0.0; 416 } 417 return work - deadline; 418 } 419 420 #ifndef PRODUCT 421 void CMSStats::print_on(outputStream *st) const { 422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 424 gc0_duration(), gc0_period(), gc0_promoted()); 425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 426 cms_duration(), cms_period(), cms_allocated()); 427 st->print(",cms_since_beg=%g,cms_since_end=%g", 428 cms_time_since_begin(), cms_time_since_end()); 429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 431 432 if (valid()) { 433 st->print(",promo_rate=%g,cms_alloc_rate=%g", 434 promotion_rate(), cms_allocation_rate()); 435 st->print(",cms_consumption_rate=%g,time_until_full=%g", 436 cms_consumption_rate(), time_until_cms_gen_full()); 437 } 438 st->print(" "); 439 } 440 #endif // #ifndef PRODUCT 441 442 CMSCollector::CollectorState CMSCollector::_collectorState = 443 CMSCollector::Idling; 444 bool CMSCollector::_foregroundGCIsActive = false; 445 bool CMSCollector::_foregroundGCShouldWait = false; 446 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 448 CardTableRS* ct, 449 ConcurrentMarkSweepPolicy* cp): 450 _cmsGen(cmsGen), 451 _ct(ct), 452 _ref_processor(NULL), // will be set later 453 _conc_workers(NULL), // may be set later 454 _abort_preclean(false), 455 _start_sampling(false), 456 _between_prologue_and_epilogue(false), 457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock", 458 CMSParallelFullGC && ShareCMSMarkBitMapWithParallelFullGC), 459 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 460 -1 /* lock-free */, "No_lock" /* dummy */), 461 _modUnionClosurePar(&_modUnionTable), 462 // Adjust my span to cover old (cms) gen 463 _span(cmsGen->reserved()), 464 // Construct the is_alive_closure with _span & markBitMap 465 _is_alive_closure(_span, &_markBitMap), 466 _restart_addr(NULL), 467 _overflow_list(NULL), 468 _stats(cmsGen), 469 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 470 //verify that this lock should be acquired with safepoint check. 471 Monitor::_safepoint_check_sometimes)), 472 _eden_chunk_array(NULL), // may be set in ctor body 473 _eden_chunk_capacity(0), // -- ditto -- 474 _eden_chunk_index(0), // -- ditto -- 475 _survivor_plab_array(NULL), // -- ditto -- 476 _survivor_chunk_array(NULL), // -- ditto -- 477 _survivor_chunk_capacity(0), // -- ditto -- 478 _survivor_chunk_index(0), // -- ditto -- 479 _ser_pmc_preclean_ovflw(0), 480 _ser_kac_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_ovflw(0), 484 _par_kac_ovflw(0), 485 #ifndef PRODUCT 486 _num_par_pushes(0), 487 #endif 488 _collection_count_start(0), 489 _verifying(false), 490 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 491 _completed_initialization(false), 492 _collector_policy(cp), 493 _should_unload_classes(CMSClassUnloadingEnabled), 494 _concurrent_cycles_since_last_unload(0), 495 _roots_scanning_options(GenCollectedHeap::SO_None), 496 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 498 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 499 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 500 _cms_start_registered(false) 501 { 502 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 503 ExplicitGCInvokesConcurrent = true; 504 } 505 // Now expand the span and allocate the collection support structures 506 // (MUT, marking bit map etc.) to cover both generations subject to 507 // collection. 508 509 // For use by dirty card to oop closures. 510 _cmsGen->cmsSpace()->set_collector(this); 511 512 // Allocate MUT and marking bit map 513 { 514 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 515 if (!_markBitMap.allocate(_span)) { 516 warning("Failed to allocate CMS Bit Map"); 517 return; 518 } 519 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 520 } 521 { 522 _modUnionTable.allocate(_span); 523 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 524 } 525 526 if (!_markStack.allocate(MarkStackSize)) { 527 warning("Failed to allocate CMS Marking Stack"); 528 return; 529 } 530 531 // Support for multi-threaded concurrent phases 532 if (CMSConcurrentMTEnabled) { 533 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 534 // just for now 535 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 536 } 537 if (ConcGCThreads > 1) { 538 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 539 ConcGCThreads, true); 540 if (_conc_workers == NULL) { 541 warning("GC/CMS: _conc_workers allocation failure: " 542 "forcing -CMSConcurrentMTEnabled"); 543 CMSConcurrentMTEnabled = false; 544 } else { 545 _conc_workers->initialize_workers(); 546 } 547 } else { 548 CMSConcurrentMTEnabled = false; 549 } 550 } 551 if (!CMSConcurrentMTEnabled) { 552 ConcGCThreads = 0; 553 } else { 554 // Turn off CMSCleanOnEnter optimization temporarily for 555 // the MT case where it's not fixed yet; see 6178663. 556 CMSCleanOnEnter = false; 557 } 558 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 559 "Inconsistency"); 560 561 // Parallel task queues; these are shared for the 562 // concurrent and stop-world phases of CMS, but 563 // are not shared with parallel scavenge (ParNew). 564 { 565 uint i; 566 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 567 568 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 569 || ParallelRefProcEnabled) 570 && num_queues > 0) { 571 _task_queues = new OopTaskQueueSet(num_queues); 572 if (_task_queues == NULL) { 573 warning("task_queues allocation failure."); 574 return; 575 } 576 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 577 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 578 for (i = 0; i < num_queues; i++) { 579 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 580 if (q == NULL) { 581 warning("work_queue allocation failure."); 582 return; 583 } 584 _task_queues->register_queue(i, q); 585 } 586 for (i = 0; i < num_queues; i++) { 587 _task_queues->queue(i)->initialize(); 588 _hash_seed[i] = 17; // copied from ParNew 589 } 590 } 591 } 592 593 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 594 595 // Clip CMSBootstrapOccupancy between 0 and 100. 596 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 597 598 // Now tell CMS generations the identity of their collector 599 ConcurrentMarkSweepGeneration::set_collector(this); 600 601 // Create & start a CMS thread for this CMS collector 602 _cmsThread = ConcurrentMarkSweepThread::start(this); 603 assert(cmsThread() != NULL, "CMS Thread should have been created"); 604 assert(cmsThread()->collector() == this, 605 "CMS Thread should refer to this gen"); 606 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 607 608 // Support for parallelizing young gen rescan 609 GenCollectedHeap* gch = GenCollectedHeap::heap(); 610 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 611 _young_gen = (ParNewGeneration*)gch->young_gen(); 612 if (gch->supports_inline_contig_alloc()) { 613 _top_addr = gch->top_addr(); 614 _end_addr = gch->end_addr(); 615 assert(_young_gen != NULL, "no _young_gen"); 616 _eden_chunk_index = 0; 617 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 618 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 619 } 620 621 // Support for parallelizing survivor space rescan 622 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 623 const size_t max_plab_samples = 624 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 625 626 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 627 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 628 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 629 _survivor_chunk_capacity = 2*max_plab_samples; 630 for (uint i = 0; i < ParallelGCThreads; i++) { 631 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 632 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 633 assert(cur->end() == 0, "Should be 0"); 634 assert(cur->array() == vec, "Should be vec"); 635 assert(cur->capacity() == max_plab_samples, "Error"); 636 } 637 } 638 639 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 640 _gc_counters = new CollectorCounters("CMS", 1); 641 _completed_initialization = true; 642 _inter_sweep_timer.start(); // start of time 643 } 644 645 size_t CMSCollector::plab_sample_minimum_size() { 646 // The default value of MinTLABSize is 2k, but there is 647 // no way to get the default value if the flag has been overridden. 648 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 649 } 650 651 const char* ConcurrentMarkSweepGeneration::name() const { 652 return "concurrent mark-sweep generation"; 653 } 654 void ConcurrentMarkSweepGeneration::update_counters() { 655 if (UsePerfData) { 656 _space_counters->update_all(); 657 _gen_counters->update_all(); 658 } 659 } 660 661 // this is an optimized version of update_counters(). it takes the 662 // used value as a parameter rather than computing it. 663 // 664 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 665 if (UsePerfData) { 666 _space_counters->update_used(used); 667 _space_counters->update_capacity(); 668 _gen_counters->update_all(); 669 } 670 } 671 672 void ConcurrentMarkSweepGeneration::print() const { 673 Generation::print(); 674 cmsSpace()->print(); 675 } 676 677 #ifndef PRODUCT 678 void ConcurrentMarkSweepGeneration::print_statistics() { 679 cmsSpace()->printFLCensus(0); 680 } 681 #endif 682 683 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 684 GenCollectedHeap* gch = GenCollectedHeap::heap(); 685 if (PrintGCDetails) { 686 // I didn't want to change the logging when removing the level concept, 687 // but I guess this logging could say "old" or something instead of "1". 688 assert(gch->is_old_gen(this), 689 "The CMS generation should be the old generation"); 690 uint level = 1; 691 if (Verbose) { 692 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]", 693 level, short_name(), s, used(), capacity()); 694 } else { 695 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]", 696 level, short_name(), s, used() / K, capacity() / K); 697 } 698 } 699 if (Verbose) { 700 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")", 701 gch->used(), gch->capacity()); 702 } else { 703 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)", 704 gch->used() / K, gch->capacity() / K); 705 } 706 } 707 708 size_t 709 ConcurrentMarkSweepGeneration::contiguous_available() const { 710 // dld proposes an improvement in precision here. If the committed 711 // part of the space ends in a free block we should add that to 712 // uncommitted size in the calculation below. Will make this 713 // change later, staying with the approximation below for the 714 // time being. -- ysr. 715 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 716 } 717 718 size_t 719 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 720 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 721 } 722 723 size_t ConcurrentMarkSweepGeneration::max_available() const { 724 return free() + _virtual_space.uncommitted_size(); 725 } 726 727 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 728 size_t available = max_available(); 729 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 730 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 731 if (Verbose && PrintGCDetails) { 732 gclog_or_tty->print_cr( 733 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT ")," 734 "max_promo(" SIZE_FORMAT ")", 735 res? "":" not", available, res? ">=":"<", 736 av_promo, max_promotion_in_bytes); 737 } 738 return res; 739 } 740 741 // At a promotion failure dump information on block layout in heap 742 // (cms old generation). 743 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 744 if (CMSDumpAtPromotionFailure) { 745 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 746 } 747 } 748 749 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 750 // Clear the promotion information. These pointers can be adjusted 751 // along with all the other pointers into the heap but 752 // compaction is expected to be a rare event with 753 // a heap using cms so don't do it without seeing the need. 754 for (uint i = 0; i < ParallelGCThreads; i++) { 755 _par_gc_thread_states[i]->promo.reset(); 756 } 757 } 758 759 void ConcurrentMarkSweepGeneration::compute_new_size() { 760 assert_locked_or_safepoint(Heap_lock); 761 762 // If incremental collection failed, we just want to expand 763 // to the limit. 764 if (incremental_collection_failed()) { 765 clear_incremental_collection_failed(); 766 grow_to_reserved(); 767 return; 768 } 769 770 // The heap has been compacted but not reset yet. 771 // Any metric such as free() or used() will be incorrect. 772 773 CardGeneration::compute_new_size(); 774 775 // Reset again after a possible resizing 776 if (did_compact()) { 777 cmsSpace()->reset_after_compaction(); 778 } 779 } 780 781 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 782 assert_locked_or_safepoint(Heap_lock); 783 784 // If incremental collection failed, we just want to expand 785 // to the limit. 786 if (incremental_collection_failed()) { 787 clear_incremental_collection_failed(); 788 grow_to_reserved(); 789 return; 790 } 791 792 double free_percentage = ((double) free()) / capacity(); 793 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 794 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 795 796 // compute expansion delta needed for reaching desired free percentage 797 if (free_percentage < desired_free_percentage) { 798 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 799 assert(desired_capacity >= capacity(), "invalid expansion size"); 800 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 801 if (PrintGCDetails && Verbose) { 802 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 803 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 804 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 805 gclog_or_tty->print_cr(" Desired free fraction %f", 806 desired_free_percentage); 807 gclog_or_tty->print_cr(" Maximum free fraction %f", 808 maximum_free_percentage); 809 gclog_or_tty->print_cr(" Capacity " SIZE_FORMAT, capacity()/1000); 810 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT, 811 desired_capacity/1000); 812 GenCollectedHeap* gch = GenCollectedHeap::heap(); 813 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 814 size_t young_size = gch->young_gen()->capacity(); 815 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000); 816 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, 817 unsafe_max_alloc_nogc()/1000); 818 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, 819 contiguous_available()/1000); 820 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", 821 expand_bytes); 822 } 823 // safe if expansion fails 824 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 825 if (PrintGCDetails && Verbose) { 826 gclog_or_tty->print_cr(" Expanded free fraction %f", 827 ((double) free()) / capacity()); 828 } 829 } else { 830 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 831 assert(desired_capacity <= capacity(), "invalid expansion size"); 832 size_t shrink_bytes = capacity() - desired_capacity; 833 // Don't shrink unless the delta is greater than the minimum shrink we want 834 if (shrink_bytes >= MinHeapDeltaBytes) { 835 shrink_free_list_by(shrink_bytes); 836 } 837 } 838 } 839 840 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 841 return cmsSpace()->freelistLock(); 842 } 843 844 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 845 bool tlab) { 846 CMSSynchronousYieldRequest yr; 847 MutexLockerEx x(freelistLock(), 848 Mutex::_no_safepoint_check_flag); 849 return have_lock_and_allocate(size, tlab); 850 } 851 852 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 853 bool tlab /* ignored */) { 854 assert_lock_strong(freelistLock()); 855 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 856 HeapWord* res = cmsSpace()->allocate(adjustedSize); 857 // Allocate the object live (grey) if the background collector has 858 // started marking. This is necessary because the marker may 859 // have passed this address and consequently this object will 860 // not otherwise be greyed and would be incorrectly swept up. 861 // Note that if this object contains references, the writing 862 // of those references will dirty the card containing this object 863 // allowing the object to be blackened (and its references scanned) 864 // either during a preclean phase or at the final checkpoint. 865 if (res != NULL) { 866 // We may block here with an uninitialized object with 867 // its mark-bit or P-bits not yet set. Such objects need 868 // to be safely navigable by block_start(). 869 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 870 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 871 collector()->direct_allocated(res, adjustedSize); 872 _direct_allocated_words += adjustedSize; 873 // allocation counters 874 NOT_PRODUCT( 875 _numObjectsAllocated++; 876 _numWordsAllocated += (int)adjustedSize; 877 ) 878 } 879 return res; 880 } 881 882 // In the case of direct allocation by mutators in a generation that 883 // is being concurrently collected, the object must be allocated 884 // live (grey) if the background collector has started marking. 885 // This is necessary because the marker may 886 // have passed this address and consequently this object will 887 // not otherwise be greyed and would be incorrectly swept up. 888 // Note that if this object contains references, the writing 889 // of those references will dirty the card containing this object 890 // allowing the object to be blackened (and its references scanned) 891 // either during a preclean phase or at the final checkpoint. 892 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 893 assert(_markBitMap.covers(start, size), "Out of bounds"); 894 if (_collectorState >= Marking) { 895 MutexLockerEx y(_markBitMap.lock(), 896 Mutex::_no_safepoint_check_flag); 897 // [see comments preceding SweepClosure::do_blk() below for details] 898 // 899 // Can the P-bits be deleted now? JJJ 900 // 901 // 1. need to mark the object as live so it isn't collected 902 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 903 // 3. need to mark the end of the object so marking, precleaning or sweeping 904 // can skip over uninitialized or unparsable objects. An allocated 905 // object is considered uninitialized for our purposes as long as 906 // its klass word is NULL. All old gen objects are parsable 907 // as soon as they are initialized.) 908 _markBitMap.mark(start); // object is live 909 _markBitMap.mark(start + 1); // object is potentially uninitialized? 910 _markBitMap.mark(start + size - 1); 911 // mark end of object 912 } 913 // check that oop looks uninitialized 914 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 915 } 916 917 void CMSCollector::promoted(bool par, HeapWord* start, 918 bool is_obj_array, size_t obj_size) { 919 assert(_markBitMap.covers(start), "Out of bounds"); 920 // See comment in direct_allocated() about when objects should 921 // be allocated live. 922 if (_collectorState >= Marking) { 923 // we already hold the marking bit map lock, taken in 924 // the prologue 925 if (par) { 926 _markBitMap.par_mark(start); 927 } else { 928 _markBitMap.mark(start); 929 } 930 // We don't need to mark the object as uninitialized (as 931 // in direct_allocated above) because this is being done with the 932 // world stopped and the object will be initialized by the 933 // time the marking, precleaning or sweeping get to look at it. 934 // But see the code for copying objects into the CMS generation, 935 // where we need to ensure that concurrent readers of the 936 // block offset table are able to safely navigate a block that 937 // is in flux from being free to being allocated (and in 938 // transition while being copied into) and subsequently 939 // becoming a bona-fide object when the copy/promotion is complete. 940 assert(SafepointSynchronize::is_at_safepoint(), 941 "expect promotion only at safepoints"); 942 943 if (_collectorState < Sweeping) { 944 // Mark the appropriate cards in the modUnionTable, so that 945 // this object gets scanned before the sweep. If this is 946 // not done, CMS generation references in the object might 947 // not get marked. 948 // For the case of arrays, which are otherwise precisely 949 // marked, we need to dirty the entire array, not just its head. 950 if (is_obj_array) { 951 // The [par_]mark_range() method expects mr.end() below to 952 // be aligned to the granularity of a bit's representation 953 // in the heap. In the case of the MUT below, that's a 954 // card size. 955 MemRegion mr(start, 956 (HeapWord*)round_to((intptr_t)(start + obj_size), 957 CardTableModRefBS::card_size /* bytes */)); 958 if (par) { 959 _modUnionTable.par_mark_range(mr); 960 } else { 961 _modUnionTable.mark_range(mr); 962 } 963 } else { // not an obj array; we can just mark the head 964 if (par) { 965 _modUnionTable.par_mark(start); 966 } else { 967 _modUnionTable.mark(start); 968 } 969 } 970 } 971 } 972 } 973 974 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 975 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 976 // allocate, copy and if necessary update promoinfo -- 977 // delegate to underlying space. 978 assert_lock_strong(freelistLock()); 979 980 #ifndef PRODUCT 981 if (GenCollectedHeap::heap()->promotion_should_fail()) { 982 return NULL; 983 } 984 #endif // #ifndef PRODUCT 985 986 oop res = _cmsSpace->promote(obj, obj_size); 987 if (res == NULL) { 988 // expand and retry 989 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 990 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 991 // Since this is the old generation, we don't try to promote 992 // into a more senior generation. 993 res = _cmsSpace->promote(obj, obj_size); 994 } 995 if (res != NULL) { 996 // See comment in allocate() about when objects should 997 // be allocated live. 998 assert(obj->is_oop(), "Will dereference klass pointer below"); 999 collector()->promoted(false, // Not parallel 1000 (HeapWord*)res, obj->is_objArray(), obj_size); 1001 // promotion counters 1002 NOT_PRODUCT( 1003 _numObjectsPromoted++; 1004 _numWordsPromoted += 1005 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1006 ) 1007 } 1008 return res; 1009 } 1010 1011 1012 // IMPORTANT: Notes on object size recognition in CMS. 1013 // --------------------------------------------------- 1014 // A block of storage in the CMS generation is always in 1015 // one of three states. A free block (FREE), an allocated 1016 // object (OBJECT) whose size() method reports the correct size, 1017 // and an intermediate state (TRANSIENT) in which its size cannot 1018 // be accurately determined. 1019 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1020 // ----------------------------------------------------- 1021 // FREE: klass_word & 1 == 1; mark_word holds block size 1022 // 1023 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1024 // obj->size() computes correct size 1025 // 1026 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1027 // 1028 // STATE IDENTIFICATION: (64 bit+COOPS) 1029 // ------------------------------------ 1030 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1031 // 1032 // OBJECT: klass_word installed; klass_word != 0; 1033 // obj->size() computes correct size 1034 // 1035 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1036 // 1037 // 1038 // STATE TRANSITION DIAGRAM 1039 // 1040 // mut / parnew mut / parnew 1041 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1042 // ^ | 1043 // |------------------------ DEAD <------------------------------------| 1044 // sweep mut 1045 // 1046 // While a block is in TRANSIENT state its size cannot be determined 1047 // so readers will either need to come back later or stall until 1048 // the size can be determined. Note that for the case of direct 1049 // allocation, P-bits, when available, may be used to determine the 1050 // size of an object that may not yet have been initialized. 1051 1052 // Things to support parallel young-gen collection. 1053 oop 1054 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1055 oop old, markOop m, 1056 size_t word_sz) { 1057 #ifndef PRODUCT 1058 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1059 return NULL; 1060 } 1061 #endif // #ifndef PRODUCT 1062 1063 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1064 PromotionInfo* promoInfo = &ps->promo; 1065 // if we are tracking promotions, then first ensure space for 1066 // promotion (including spooling space for saving header if necessary). 1067 // then allocate and copy, then track promoted info if needed. 1068 // When tracking (see PromotionInfo::track()), the mark word may 1069 // be displaced and in this case restoration of the mark word 1070 // occurs in the (oop_since_save_marks_)iterate phase. 1071 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1072 // Out of space for allocating spooling buffers; 1073 // try expanding and allocating spooling buffers. 1074 if (!expand_and_ensure_spooling_space(promoInfo)) { 1075 return NULL; 1076 } 1077 } 1078 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1079 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1080 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1081 if (obj_ptr == NULL) { 1082 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1083 if (obj_ptr == NULL) { 1084 return NULL; 1085 } 1086 } 1087 oop obj = oop(obj_ptr); 1088 OrderAccess::storestore(); 1089 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1090 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1091 // IMPORTANT: See note on object initialization for CMS above. 1092 // Otherwise, copy the object. Here we must be careful to insert the 1093 // klass pointer last, since this marks the block as an allocated object. 1094 // Except with compressed oops it's the mark word. 1095 HeapWord* old_ptr = (HeapWord*)old; 1096 // Restore the mark word copied above. 1097 obj->set_mark(m); 1098 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1099 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1100 OrderAccess::storestore(); 1101 1102 if (UseCompressedClassPointers) { 1103 // Copy gap missed by (aligned) header size calculation below 1104 obj->set_klass_gap(old->klass_gap()); 1105 } 1106 if (word_sz > (size_t)oopDesc::header_size()) { 1107 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1108 obj_ptr + oopDesc::header_size(), 1109 word_sz - oopDesc::header_size()); 1110 } 1111 1112 // Now we can track the promoted object, if necessary. We take care 1113 // to delay the transition from uninitialized to full object 1114 // (i.e., insertion of klass pointer) until after, so that it 1115 // atomically becomes a promoted object. 1116 if (promoInfo->tracking()) { 1117 promoInfo->track((PromotedObject*)obj, old->klass()); 1118 } 1119 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1120 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1121 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1122 1123 // Finally, install the klass pointer (this should be volatile). 1124 OrderAccess::storestore(); 1125 obj->set_klass(old->klass()); 1126 // We should now be able to calculate the right size for this object 1127 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1128 1129 collector()->promoted(true, // parallel 1130 obj_ptr, old->is_objArray(), word_sz); 1131 1132 NOT_PRODUCT( 1133 Atomic::inc_ptr(&_numObjectsPromoted); 1134 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1135 ) 1136 1137 return obj; 1138 } 1139 1140 void 1141 ConcurrentMarkSweepGeneration:: 1142 par_promote_alloc_done(int thread_num) { 1143 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1144 ps->lab.retire(thread_num); 1145 } 1146 1147 void 1148 ConcurrentMarkSweepGeneration:: 1149 par_oop_since_save_marks_iterate_done(int thread_num) { 1150 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1151 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1152 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1153 } 1154 1155 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1156 size_t size, 1157 bool tlab) 1158 { 1159 // We allow a STW collection only if a full 1160 // collection was requested. 1161 return full || should_allocate(size, tlab); // FIX ME !!! 1162 // This and promotion failure handling are connected at the 1163 // hip and should be fixed by untying them. 1164 } 1165 1166 bool CMSCollector::shouldConcurrentCollect() { 1167 if (_full_gc_requested) { 1168 if (Verbose && PrintGCDetails) { 1169 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1170 " gc request (or gc_locker)"); 1171 } 1172 return true; 1173 } 1174 1175 FreelistLocker x(this); 1176 // ------------------------------------------------------------------ 1177 // Print out lots of information which affects the initiation of 1178 // a collection. 1179 if (PrintCMSInitiationStatistics && stats().valid()) { 1180 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1181 gclog_or_tty->stamp(); 1182 gclog_or_tty->cr(); 1183 stats().print_on(gclog_or_tty); 1184 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1185 stats().time_until_cms_gen_full()); 1186 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free()); 1187 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT, 1188 _cmsGen->contiguous_available()); 1189 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1190 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1191 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1192 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1193 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1194 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1195 gclog_or_tty->print_cr("metadata initialized %d", 1196 MetaspaceGC::should_concurrent_collect()); 1197 } 1198 // ------------------------------------------------------------------ 1199 1200 // If the estimated time to complete a cms collection (cms_duration()) 1201 // is less than the estimated time remaining until the cms generation 1202 // is full, start a collection. 1203 if (!UseCMSInitiatingOccupancyOnly) { 1204 if (stats().valid()) { 1205 if (stats().time_until_cms_start() == 0.0) { 1206 return true; 1207 } 1208 } else { 1209 // We want to conservatively collect somewhat early in order 1210 // to try and "bootstrap" our CMS/promotion statistics; 1211 // this branch will not fire after the first successful CMS 1212 // collection because the stats should then be valid. 1213 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1214 if (Verbose && PrintGCDetails) { 1215 gclog_or_tty->print_cr( 1216 " CMSCollector: collect for bootstrapping statistics:" 1217 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1218 _bootstrap_occupancy); 1219 } 1220 return true; 1221 } 1222 } 1223 } 1224 1225 // Otherwise, we start a collection cycle if 1226 // old gen want a collection cycle started. Each may use 1227 // an appropriate criterion for making this decision. 1228 // XXX We need to make sure that the gen expansion 1229 // criterion dovetails well with this. XXX NEED TO FIX THIS 1230 if (_cmsGen->should_concurrent_collect()) { 1231 if (Verbose && PrintGCDetails) { 1232 gclog_or_tty->print_cr("CMS old gen initiated"); 1233 } 1234 return true; 1235 } 1236 1237 // We start a collection if we believe an incremental collection may fail; 1238 // this is not likely to be productive in practice because it's probably too 1239 // late anyway. 1240 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1241 assert(gch->collector_policy()->is_generation_policy(), 1242 "You may want to check the correctness of the following"); 1243 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1244 if (Verbose && PrintGCDetails) { 1245 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1246 } 1247 return true; 1248 } 1249 1250 if (MetaspaceGC::should_concurrent_collect()) { 1251 if (Verbose && PrintGCDetails) { 1252 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1253 } 1254 return true; 1255 } 1256 1257 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1258 if (CMSTriggerInterval >= 0) { 1259 if (CMSTriggerInterval == 0) { 1260 // Trigger always 1261 return true; 1262 } 1263 1264 // Check the CMS time since begin (we do not check the stats validity 1265 // as we want to be able to trigger the first CMS cycle as well) 1266 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1267 if (Verbose && PrintGCDetails) { 1268 if (stats().valid()) { 1269 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1270 stats().cms_time_since_begin()); 1271 } else { 1272 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1273 } 1274 } 1275 return true; 1276 } 1277 } 1278 1279 return false; 1280 } 1281 1282 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1283 1284 // Clear _expansion_cause fields of constituent generations 1285 void CMSCollector::clear_expansion_cause() { 1286 _cmsGen->clear_expansion_cause(); 1287 } 1288 1289 // We should be conservative in starting a collection cycle. To 1290 // start too eagerly runs the risk of collecting too often in the 1291 // extreme. To collect too rarely falls back on full collections, 1292 // which works, even if not optimum in terms of concurrent work. 1293 // As a work around for too eagerly collecting, use the flag 1294 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1295 // giving the user an easily understandable way of controlling the 1296 // collections. 1297 // We want to start a new collection cycle if any of the following 1298 // conditions hold: 1299 // . our current occupancy exceeds the configured initiating occupancy 1300 // for this generation, or 1301 // . we recently needed to expand this space and have not, since that 1302 // expansion, done a collection of this generation, or 1303 // . the underlying space believes that it may be a good idea to initiate 1304 // a concurrent collection (this may be based on criteria such as the 1305 // following: the space uses linear allocation and linear allocation is 1306 // going to fail, or there is believed to be excessive fragmentation in 1307 // the generation, etc... or ... 1308 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1309 // the case of the old generation; see CR 6543076): 1310 // we may be approaching a point at which allocation requests may fail because 1311 // we will be out of sufficient free space given allocation rate estimates.] 1312 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1313 1314 assert_lock_strong(freelistLock()); 1315 if (occupancy() > initiating_occupancy()) { 1316 if (PrintGCDetails && Verbose) { 1317 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1318 short_name(), occupancy(), initiating_occupancy()); 1319 } 1320 return true; 1321 } 1322 if (UseCMSInitiatingOccupancyOnly) { 1323 return false; 1324 } 1325 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1326 if (PrintGCDetails && Verbose) { 1327 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1328 short_name()); 1329 } 1330 return true; 1331 } 1332 if (_cmsSpace->should_concurrent_collect()) { 1333 if (PrintGCDetails && Verbose) { 1334 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1335 short_name()); 1336 } 1337 return true; 1338 } 1339 return false; 1340 } 1341 1342 void ConcurrentMarkSweepGeneration::collect(bool full, 1343 bool clear_all_soft_refs, 1344 size_t size, 1345 bool tlab) 1346 { 1347 collector()->collect(full, clear_all_soft_refs, size, tlab); 1348 } 1349 1350 void CMSCollector::collect(bool full, 1351 bool clear_all_soft_refs, 1352 size_t size, 1353 bool tlab) 1354 { 1355 // The following "if" branch is present for defensive reasons. 1356 // In the current uses of this interface, it can be replaced with: 1357 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1358 // But I am not placing that assert here to allow future 1359 // generality in invoking this interface. 1360 if (GC_locker::is_active()) { 1361 // A consistency test for GC_locker 1362 assert(GC_locker::needs_gc(), "Should have been set already"); 1363 // Skip this foreground collection, instead 1364 // expanding the heap if necessary. 1365 // Need the free list locks for the call to free() in compute_new_size() 1366 compute_new_size(); 1367 return; 1368 } 1369 acquire_control_and_collect(full, clear_all_soft_refs); 1370 } 1371 1372 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1373 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1374 unsigned int gc_count = gch->total_full_collections(); 1375 if (gc_count == full_gc_count) { 1376 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1377 _full_gc_requested = true; 1378 _full_gc_cause = cause; 1379 CGC_lock->notify(); // nudge CMS thread 1380 } else { 1381 assert(gc_count > full_gc_count, "Error: causal loop"); 1382 } 1383 } 1384 1385 bool CMSCollector::is_external_interruption() { 1386 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1387 return GCCause::is_user_requested_gc(cause) || 1388 GCCause::is_serviceability_requested_gc(cause); 1389 } 1390 1391 void CMSCollector::report_concurrent_mode_interruption() { 1392 if (is_external_interruption()) { 1393 if (PrintGCDetails) { 1394 gclog_or_tty->print(" (concurrent mode interrupted)"); 1395 } 1396 } else { 1397 if (PrintGCDetails) { 1398 gclog_or_tty->print(" (concurrent mode failure)"); 1399 } 1400 _gc_tracer_cm->report_concurrent_mode_failure(); 1401 } 1402 } 1403 1404 1405 // The foreground and background collectors need to coordinate in order 1406 // to make sure that they do not mutually interfere with CMS collections. 1407 // When a background collection is active, 1408 // the foreground collector may need to take over (preempt) and 1409 // synchronously complete an ongoing collection. Depending on the 1410 // frequency of the background collections and the heap usage 1411 // of the application, this preemption can be seldom or frequent. 1412 // There are only certain 1413 // points in the background collection that the "collection-baton" 1414 // can be passed to the foreground collector. 1415 // 1416 // The foreground collector will wait for the baton before 1417 // starting any part of the collection. The foreground collector 1418 // will only wait at one location. 1419 // 1420 // The background collector will yield the baton before starting a new 1421 // phase of the collection (e.g., before initial marking, marking from roots, 1422 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1423 // of the loop which switches the phases. The background collector does some 1424 // of the phases (initial mark, final re-mark) with the world stopped. 1425 // Because of locking involved in stopping the world, 1426 // the foreground collector should not block waiting for the background 1427 // collector when it is doing a stop-the-world phase. The background 1428 // collector will yield the baton at an additional point just before 1429 // it enters a stop-the-world phase. Once the world is stopped, the 1430 // background collector checks the phase of the collection. If the 1431 // phase has not changed, it proceeds with the collection. If the 1432 // phase has changed, it skips that phase of the collection. See 1433 // the comments on the use of the Heap_lock in collect_in_background(). 1434 // 1435 // Variable used in baton passing. 1436 // _foregroundGCIsActive - Set to true by the foreground collector when 1437 // it wants the baton. The foreground clears it when it has finished 1438 // the collection. 1439 // _foregroundGCShouldWait - Set to true by the background collector 1440 // when it is running. The foreground collector waits while 1441 // _foregroundGCShouldWait is true. 1442 // CGC_lock - monitor used to protect access to the above variables 1443 // and to notify the foreground and background collectors. 1444 // _collectorState - current state of the CMS collection. 1445 // 1446 // The foreground collector 1447 // acquires the CGC_lock 1448 // sets _foregroundGCIsActive 1449 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1450 // various locks acquired in preparation for the collection 1451 // are released so as not to block the background collector 1452 // that is in the midst of a collection 1453 // proceeds with the collection 1454 // clears _foregroundGCIsActive 1455 // returns 1456 // 1457 // The background collector in a loop iterating on the phases of the 1458 // collection 1459 // acquires the CGC_lock 1460 // sets _foregroundGCShouldWait 1461 // if _foregroundGCIsActive is set 1462 // clears _foregroundGCShouldWait, notifies _CGC_lock 1463 // waits on _CGC_lock for _foregroundGCIsActive to become false 1464 // and exits the loop. 1465 // otherwise 1466 // proceed with that phase of the collection 1467 // if the phase is a stop-the-world phase, 1468 // yield the baton once more just before enqueueing 1469 // the stop-world CMS operation (executed by the VM thread). 1470 // returns after all phases of the collection are done 1471 // 1472 1473 void CMSCollector::acquire_control_and_collect(bool full, 1474 bool clear_all_soft_refs) { 1475 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1476 assert(!Thread::current()->is_ConcurrentGC_thread(), 1477 "shouldn't try to acquire control from self!"); 1478 1479 // Start the protocol for acquiring control of the 1480 // collection from the background collector (aka CMS thread). 1481 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1482 "VM thread should have CMS token"); 1483 // Remember the possibly interrupted state of an ongoing 1484 // concurrent collection 1485 CollectorState first_state = _collectorState; 1486 1487 // Signal to a possibly ongoing concurrent collection that 1488 // we want to do a foreground collection. 1489 _foregroundGCIsActive = true; 1490 1491 // release locks and wait for a notify from the background collector 1492 // releasing the locks in only necessary for phases which 1493 // do yields to improve the granularity of the collection. 1494 assert_lock_strong(bitMapLock()); 1495 // We need to lock the Free list lock for the space that we are 1496 // currently collecting. 1497 assert(haveFreelistLocks(), "Must be holding free list locks"); 1498 bitMapLock()->unlock(); 1499 releaseFreelistLocks(); 1500 { 1501 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1502 if (_foregroundGCShouldWait) { 1503 // We are going to be waiting for action for the CMS thread; 1504 // it had better not be gone (for instance at shutdown)! 1505 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1506 "CMS thread must be running"); 1507 // Wait here until the background collector gives us the go-ahead 1508 ConcurrentMarkSweepThread::clear_CMS_flag( 1509 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1510 // Get a possibly blocked CMS thread going: 1511 // Note that we set _foregroundGCIsActive true above, 1512 // without protection of the CGC_lock. 1513 CGC_lock->notify(); 1514 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1515 "Possible deadlock"); 1516 while (_foregroundGCShouldWait) { 1517 // wait for notification 1518 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1519 // Possibility of delay/starvation here, since CMS token does 1520 // not know to give priority to VM thread? Actually, i think 1521 // there wouldn't be any delay/starvation, but the proof of 1522 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1523 } 1524 ConcurrentMarkSweepThread::set_CMS_flag( 1525 ConcurrentMarkSweepThread::CMS_vm_has_token); 1526 } 1527 } 1528 // The CMS_token is already held. Get back the other locks. 1529 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1530 "VM thread should have CMS token"); 1531 getFreelistLocks(); 1532 bitMapLock()->lock_without_safepoint_check(); 1533 if (TraceCMSState) { 1534 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1535 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1536 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1537 } 1538 1539 // Inform cms gen if this was due to partial collection failing. 1540 // The CMS gen may use this fact to determine its expansion policy. 1541 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1542 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1543 assert(!_cmsGen->incremental_collection_failed(), 1544 "Should have been noticed, reacted to and cleared"); 1545 _cmsGen->set_incremental_collection_failed(); 1546 } 1547 1548 if (first_state > Idling) { 1549 report_concurrent_mode_interruption(); 1550 } 1551 1552 set_did_compact(true); 1553 1554 // If the collection is being acquired from the background 1555 // collector, there may be references on the discovered 1556 // references lists. Abandon those references, since some 1557 // of them may have become unreachable after concurrent 1558 // discovery; the STW compacting collector will redo discovery 1559 // more precisely, without being subject to floating garbage. 1560 // Leaving otherwise unreachable references in the discovered 1561 // lists would require special handling. 1562 ref_processor()->disable_discovery(); 1563 ref_processor()->abandon_partial_discovery(); 1564 ref_processor()->verify_no_references_recorded(); 1565 1566 if (first_state > Idling) { 1567 save_heap_summary(); 1568 } 1569 1570 do_compaction_work(clear_all_soft_refs); 1571 1572 // Has the GC time limit been exceeded? 1573 size_t max_eden_size = _young_gen->max_capacity() - 1574 _young_gen->to()->capacity() - 1575 _young_gen->from()->capacity(); 1576 GCCause::Cause gc_cause = gch->gc_cause(); 1577 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1578 _young_gen->eden()->used(), 1579 _cmsGen->max_capacity(), 1580 max_eden_size, 1581 full, 1582 gc_cause, 1583 gch->collector_policy()); 1584 1585 // Reset the expansion cause, now that we just completed 1586 // a collection cycle. 1587 clear_expansion_cause(); 1588 _foregroundGCIsActive = false; 1589 return; 1590 } 1591 1592 // Resize the tenured generation 1593 // after obtaining the free list locks for the 1594 // two generations. 1595 void CMSCollector::compute_new_size() { 1596 assert_locked_or_safepoint(Heap_lock); 1597 FreelistLocker z(this); 1598 MetaspaceGC::compute_new_size(); 1599 _cmsGen->compute_new_size_free_list(); 1600 } 1601 1602 // A work method used by the foreground collector to do 1603 // a mark-sweep-compact. 1604 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1605 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1606 1607 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1608 gc_timer->register_gc_start(); 1609 1610 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1611 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1612 1613 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1614 1615 // Temporarily widen the span of the weak reference processing to 1616 // the entire heap. 1617 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1618 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1619 // Temporarily, clear the "is_alive_non_header" field of the 1620 // reference processor. 1621 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1622 // Temporarily make reference _processing_ single threaded or MT as 1623 // per CMSParallelFullGC 1624 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), CMSParallelFullGC); 1625 // Temporarily make refs discovery atomic 1626 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1627 // Temporarily make reference _discovery_ single threaded or MT as 1628 // per CMSParallelFullGC 1629 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), CMSParallelFullGC); 1630 1631 ref_processor()->set_enqueuing_is_done(false); 1632 ref_processor()->enable_discovery(); 1633 ref_processor()->setup_policy(clear_all_soft_refs); 1634 // If an asynchronous collection finishes, the _modUnionTable is 1635 // all clear. If we are assuming the collection from an asynchronous 1636 // collection, clear the _modUnionTable. 1637 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1638 "_modUnionTable should be clear if the baton was not passed"); 1639 _modUnionTable.clear_all(); 1640 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1641 "mod union for klasses should be clear if the baton was passed"); 1642 _ct->klass_rem_set()->clear_mod_union(); 1643 1644 // We must adjust the allocation statistics being maintained 1645 // in the free list space. We do so by reading and clearing 1646 // the sweep timer and updating the block flux rate estimates below. 1647 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1648 if (_inter_sweep_timer.is_active()) { 1649 _inter_sweep_timer.stop(); 1650 // Note that we do not use this sample to update the _inter_sweep_estimate. 1651 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1652 _inter_sweep_estimate.padded_average(), 1653 _intra_sweep_estimate.padded_average()); 1654 } 1655 1656 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1657 #ifdef ASSERT 1658 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1659 size_t free_size = cms_space->free(); 1660 assert(free_size == 1661 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1662 * HeapWordSize, 1663 "All the free space should be compacted into one chunk at top"); 1664 assert(cms_space->dictionary()->total_chunk_size( 1665 debug_only(cms_space->freelistLock())) == 0 || 1666 cms_space->totalSizeInIndexedFreeLists() == 0, 1667 "All the free space should be in a single chunk"); 1668 size_t num = cms_space->totalCount(); 1669 assert((free_size == 0 && num == 0) || 1670 (free_size > 0 && (num == 1 || num == 2)), 1671 "There should be at most 2 free chunks after compaction"); 1672 #endif // ASSERT 1673 _collectorState = Resetting; 1674 assert(_restart_addr == NULL, 1675 "Should have been NULL'd before baton was passed"); 1676 reset(false /* == !concurrent */); 1677 _cmsGen->reset_after_compaction(); 1678 _concurrent_cycles_since_last_unload = 0; 1679 1680 // Clear any data recorded in the PLAB chunk arrays. 1681 if (_survivor_plab_array != NULL) { 1682 reset_survivor_plab_arrays(); 1683 } 1684 1685 // Adjust the per-size allocation stats for the next epoch. 1686 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1687 // Restart the "inter sweep timer" for the next epoch. 1688 _inter_sweep_timer.reset(); 1689 _inter_sweep_timer.start(); 1690 1691 gc_timer->register_gc_end(); 1692 1693 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1694 1695 // For a mark-sweep-compact, compute_new_size() will be called 1696 // in the heap's do_collection() method. 1697 } 1698 1699 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1700 ContiguousSpace* eden_space = _young_gen->eden(); 1701 ContiguousSpace* from_space = _young_gen->from(); 1702 ContiguousSpace* to_space = _young_gen->to(); 1703 // Eden 1704 if (_eden_chunk_array != NULL) { 1705 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1706 p2i(eden_space->bottom()), p2i(eden_space->top()), 1707 p2i(eden_space->end()), eden_space->capacity()); 1708 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1709 "_eden_chunk_capacity=" SIZE_FORMAT, 1710 _eden_chunk_index, _eden_chunk_capacity); 1711 for (size_t i = 0; i < _eden_chunk_index; i++) { 1712 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1713 i, p2i(_eden_chunk_array[i])); 1714 } 1715 } 1716 // Survivor 1717 if (_survivor_chunk_array != NULL) { 1718 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1719 p2i(from_space->bottom()), p2i(from_space->top()), 1720 p2i(from_space->end()), from_space->capacity()); 1721 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1722 "_survivor_chunk_capacity=" SIZE_FORMAT, 1723 _survivor_chunk_index, _survivor_chunk_capacity); 1724 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1725 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1726 i, p2i(_survivor_chunk_array[i])); 1727 } 1728 } 1729 } 1730 1731 void CMSCollector::getFreelistLocks() const { 1732 // Get locks for all free lists in all generations that this 1733 // collector is responsible for 1734 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1735 } 1736 1737 void CMSCollector::releaseFreelistLocks() const { 1738 // Release locks for all free lists in all generations that this 1739 // collector is responsible for 1740 _cmsGen->freelistLock()->unlock(); 1741 } 1742 1743 bool CMSCollector::haveFreelistLocks() const { 1744 // Check locks for all free lists in all generations that this 1745 // collector is responsible for 1746 assert_lock_strong(_cmsGen->freelistLock()); 1747 PRODUCT_ONLY(ShouldNotReachHere()); 1748 return true; 1749 } 1750 1751 // A utility class that is used by the CMS collector to 1752 // temporarily "release" the foreground collector from its 1753 // usual obligation to wait for the background collector to 1754 // complete an ongoing phase before proceeding. 1755 class ReleaseForegroundGC: public StackObj { 1756 private: 1757 CMSCollector* _c; 1758 public: 1759 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1760 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1761 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1762 // allow a potentially blocked foreground collector to proceed 1763 _c->_foregroundGCShouldWait = false; 1764 if (_c->_foregroundGCIsActive) { 1765 CGC_lock->notify(); 1766 } 1767 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1768 "Possible deadlock"); 1769 } 1770 1771 ~ReleaseForegroundGC() { 1772 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1773 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1774 _c->_foregroundGCShouldWait = true; 1775 } 1776 }; 1777 1778 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1779 assert(Thread::current()->is_ConcurrentGC_thread(), 1780 "A CMS asynchronous collection is only allowed on a CMS thread."); 1781 1782 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1783 { 1784 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1785 MutexLockerEx hl(Heap_lock, safepoint_check); 1786 FreelistLocker fll(this); 1787 MutexLockerEx x(CGC_lock, safepoint_check); 1788 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1789 // The foreground collector is active or we're 1790 // not using asynchronous collections. Skip this 1791 // background collection. 1792 assert(!_foregroundGCShouldWait, "Should be clear"); 1793 return; 1794 } else { 1795 assert(_collectorState == Idling, "Should be idling before start."); 1796 _collectorState = InitialMarking; 1797 register_gc_start(cause); 1798 // Reset the expansion cause, now that we are about to begin 1799 // a new cycle. 1800 clear_expansion_cause(); 1801 1802 // Clear the MetaspaceGC flag since a concurrent collection 1803 // is starting but also clear it after the collection. 1804 MetaspaceGC::set_should_concurrent_collect(false); 1805 } 1806 // Decide if we want to enable class unloading as part of the 1807 // ensuing concurrent GC cycle. 1808 update_should_unload_classes(); 1809 _full_gc_requested = false; // acks all outstanding full gc requests 1810 _full_gc_cause = GCCause::_no_gc; 1811 // Signal that we are about to start a collection 1812 gch->increment_total_full_collections(); // ... starting a collection cycle 1813 _collection_count_start = gch->total_full_collections(); 1814 } 1815 1816 // Used for PrintGC 1817 size_t prev_used; 1818 if (PrintGC && Verbose) { 1819 prev_used = _cmsGen->used(); 1820 } 1821 1822 // The change of the collection state is normally done at this level; 1823 // the exceptions are phases that are executed while the world is 1824 // stopped. For those phases the change of state is done while the 1825 // world is stopped. For baton passing purposes this allows the 1826 // background collector to finish the phase and change state atomically. 1827 // The foreground collector cannot wait on a phase that is done 1828 // while the world is stopped because the foreground collector already 1829 // has the world stopped and would deadlock. 1830 while (_collectorState != Idling) { 1831 if (TraceCMSState) { 1832 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1833 p2i(Thread::current()), _collectorState); 1834 } 1835 // The foreground collector 1836 // holds the Heap_lock throughout its collection. 1837 // holds the CMS token (but not the lock) 1838 // except while it is waiting for the background collector to yield. 1839 // 1840 // The foreground collector should be blocked (not for long) 1841 // if the background collector is about to start a phase 1842 // executed with world stopped. If the background 1843 // collector has already started such a phase, the 1844 // foreground collector is blocked waiting for the 1845 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1846 // are executed in the VM thread. 1847 // 1848 // The locking order is 1849 // PendingListLock (PLL) -- if applicable (FinalMarking) 1850 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1851 // CMS token (claimed in 1852 // stop_world_and_do() --> 1853 // safepoint_synchronize() --> 1854 // CMSThread::synchronize()) 1855 1856 { 1857 // Check if the FG collector wants us to yield. 1858 CMSTokenSync x(true); // is cms thread 1859 if (waitForForegroundGC()) { 1860 // We yielded to a foreground GC, nothing more to be 1861 // done this round. 1862 assert(_foregroundGCShouldWait == false, "We set it to false in " 1863 "waitForForegroundGC()"); 1864 if (TraceCMSState) { 1865 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1866 " exiting collection CMS state %d", 1867 p2i(Thread::current()), _collectorState); 1868 } 1869 return; 1870 } else { 1871 // The background collector can run but check to see if the 1872 // foreground collector has done a collection while the 1873 // background collector was waiting to get the CGC_lock 1874 // above. If yes, break so that _foregroundGCShouldWait 1875 // is cleared before returning. 1876 if (_collectorState == Idling) { 1877 break; 1878 } 1879 } 1880 } 1881 1882 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1883 "should be waiting"); 1884 1885 switch (_collectorState) { 1886 case InitialMarking: 1887 { 1888 ReleaseForegroundGC x(this); 1889 stats().record_cms_begin(); 1890 VM_CMS_Initial_Mark initial_mark_op(this); 1891 VMThread::execute(&initial_mark_op); 1892 } 1893 // The collector state may be any legal state at this point 1894 // since the background collector may have yielded to the 1895 // foreground collector. 1896 break; 1897 case Marking: 1898 // initial marking in checkpointRootsInitialWork has been completed 1899 if (markFromRoots()) { // we were successful 1900 assert(_collectorState == Precleaning, "Collector state should " 1901 "have changed"); 1902 } else { 1903 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1904 } 1905 break; 1906 case Precleaning: 1907 // marking from roots in markFromRoots has been completed 1908 preclean(); 1909 assert(_collectorState == AbortablePreclean || 1910 _collectorState == FinalMarking, 1911 "Collector state should have changed"); 1912 break; 1913 case AbortablePreclean: 1914 abortable_preclean(); 1915 assert(_collectorState == FinalMarking, "Collector state should " 1916 "have changed"); 1917 break; 1918 case FinalMarking: 1919 { 1920 ReleaseForegroundGC x(this); 1921 1922 VM_CMS_Final_Remark final_remark_op(this); 1923 VMThread::execute(&final_remark_op); 1924 } 1925 assert(_foregroundGCShouldWait, "block post-condition"); 1926 break; 1927 case Sweeping: 1928 // final marking in checkpointRootsFinal has been completed 1929 sweep(); 1930 assert(_collectorState == Resizing, "Collector state change " 1931 "to Resizing must be done under the free_list_lock"); 1932 1933 case Resizing: { 1934 // Sweeping has been completed... 1935 // At this point the background collection has completed. 1936 // Don't move the call to compute_new_size() down 1937 // into code that might be executed if the background 1938 // collection was preempted. 1939 { 1940 ReleaseForegroundGC x(this); // unblock FG collection 1941 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1942 CMSTokenSync z(true); // not strictly needed. 1943 if (_collectorState == Resizing) { 1944 compute_new_size(); 1945 save_heap_summary(); 1946 _collectorState = Resetting; 1947 } else { 1948 assert(_collectorState == Idling, "The state should only change" 1949 " because the foreground collector has finished the collection"); 1950 } 1951 } 1952 break; 1953 } 1954 case Resetting: 1955 // CMS heap resizing has been completed 1956 reset(true); 1957 assert(_collectorState == Idling, "Collector state should " 1958 "have changed"); 1959 1960 MetaspaceGC::set_should_concurrent_collect(false); 1961 1962 stats().record_cms_end(); 1963 // Don't move the concurrent_phases_end() and compute_new_size() 1964 // calls to here because a preempted background collection 1965 // has it's state set to "Resetting". 1966 break; 1967 case Idling: 1968 default: 1969 ShouldNotReachHere(); 1970 break; 1971 } 1972 if (TraceCMSState) { 1973 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1974 p2i(Thread::current()), _collectorState); 1975 } 1976 assert(_foregroundGCShouldWait, "block post-condition"); 1977 } 1978 1979 // Should this be in gc_epilogue? 1980 collector_policy()->counters()->update_counters(); 1981 1982 { 1983 // Clear _foregroundGCShouldWait and, in the event that the 1984 // foreground collector is waiting, notify it, before 1985 // returning. 1986 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1987 _foregroundGCShouldWait = false; 1988 if (_foregroundGCIsActive) { 1989 CGC_lock->notify(); 1990 } 1991 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1992 "Possible deadlock"); 1993 } 1994 if (TraceCMSState) { 1995 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1996 " exiting collection CMS state %d", 1997 p2i(Thread::current()), _collectorState); 1998 } 1999 if (PrintGC && Verbose) { 2000 _cmsGen->print_heap_change(prev_used); 2001 } 2002 } 2003 2004 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2005 _cms_start_registered = true; 2006 _gc_timer_cm->register_gc_start(); 2007 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2008 } 2009 2010 void CMSCollector::register_gc_end() { 2011 if (_cms_start_registered) { 2012 report_heap_summary(GCWhen::AfterGC); 2013 2014 _gc_timer_cm->register_gc_end(); 2015 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2016 _cms_start_registered = false; 2017 } 2018 } 2019 2020 void CMSCollector::save_heap_summary() { 2021 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2022 _last_heap_summary = gch->create_heap_summary(); 2023 _last_metaspace_summary = gch->create_metaspace_summary(); 2024 } 2025 2026 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2027 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2028 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2029 } 2030 2031 bool CMSCollector::waitForForegroundGC() { 2032 bool res = false; 2033 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2034 "CMS thread should have CMS token"); 2035 // Block the foreground collector until the 2036 // background collectors decides whether to 2037 // yield. 2038 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2039 _foregroundGCShouldWait = true; 2040 if (_foregroundGCIsActive) { 2041 // The background collector yields to the 2042 // foreground collector and returns a value 2043 // indicating that it has yielded. The foreground 2044 // collector can proceed. 2045 res = true; 2046 _foregroundGCShouldWait = false; 2047 ConcurrentMarkSweepThread::clear_CMS_flag( 2048 ConcurrentMarkSweepThread::CMS_cms_has_token); 2049 ConcurrentMarkSweepThread::set_CMS_flag( 2050 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2051 // Get a possibly blocked foreground thread going 2052 CGC_lock->notify(); 2053 if (TraceCMSState) { 2054 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2055 p2i(Thread::current()), _collectorState); 2056 } 2057 while (_foregroundGCIsActive) { 2058 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2059 } 2060 ConcurrentMarkSweepThread::set_CMS_flag( 2061 ConcurrentMarkSweepThread::CMS_cms_has_token); 2062 ConcurrentMarkSweepThread::clear_CMS_flag( 2063 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2064 } 2065 if (TraceCMSState) { 2066 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2067 p2i(Thread::current()), _collectorState); 2068 } 2069 return res; 2070 } 2071 2072 // Because of the need to lock the free lists and other structures in 2073 // the collector, common to all the generations that the collector is 2074 // collecting, we need the gc_prologues of individual CMS generations 2075 // delegate to their collector. It may have been simpler had the 2076 // current infrastructure allowed one to call a prologue on a 2077 // collector. In the absence of that we have the generation's 2078 // prologue delegate to the collector, which delegates back 2079 // some "local" work to a worker method in the individual generations 2080 // that it's responsible for collecting, while itself doing any 2081 // work common to all generations it's responsible for. A similar 2082 // comment applies to the gc_epilogue()'s. 2083 // The role of the variable _between_prologue_and_epilogue is to 2084 // enforce the invocation protocol. 2085 void CMSCollector::gc_prologue(bool full) { 2086 // Call gc_prologue_work() for the CMSGen 2087 // we are responsible for. 2088 2089 // The following locking discipline assumes that we are only called 2090 // when the world is stopped. 2091 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2092 2093 // The CMSCollector prologue must call the gc_prologues for the 2094 // "generations" that it's responsible 2095 // for. 2096 2097 assert( Thread::current()->is_VM_thread() 2098 || ( CMSScavengeBeforeRemark 2099 && Thread::current()->is_ConcurrentGC_thread()), 2100 "Incorrect thread type for prologue execution"); 2101 2102 if (_between_prologue_and_epilogue) { 2103 // We have already been invoked; this is a gc_prologue delegation 2104 // from yet another CMS generation that we are responsible for, just 2105 // ignore it since all relevant work has already been done. 2106 return; 2107 } 2108 2109 // set a bit saying prologue has been called; cleared in epilogue 2110 _between_prologue_and_epilogue = true; 2111 // Claim locks for common data structures, then call gc_prologue_work() 2112 // for each CMSGen. 2113 2114 getFreelistLocks(); // gets free list locks on constituent spaces 2115 bitMapLock()->lock_without_safepoint_check(); 2116 2117 // Should call gc_prologue_work() for all cms gens we are responsible for 2118 bool duringMarking = _collectorState >= Marking 2119 && _collectorState < Sweeping; 2120 2121 // The young collections clear the modified oops state, which tells if 2122 // there are any modified oops in the class. The remark phase also needs 2123 // that information. Tell the young collection to save the union of all 2124 // modified klasses. 2125 if (duringMarking) { 2126 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2127 } 2128 2129 bool registerClosure = duringMarking; 2130 2131 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2132 2133 if (!full) { 2134 stats().record_gc0_begin(); 2135 } 2136 } 2137 2138 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2139 2140 _capacity_at_prologue = capacity(); 2141 _used_at_prologue = used(); 2142 2143 // Delegate to CMScollector which knows how to coordinate between 2144 // this and any other CMS generations that it is responsible for 2145 // collecting. 2146 collector()->gc_prologue(full); 2147 } 2148 2149 // This is a "private" interface for use by this generation's CMSCollector. 2150 // Not to be called directly by any other entity (for instance, 2151 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2152 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2153 bool registerClosure, ModUnionClosure* modUnionClosure) { 2154 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2155 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2156 "Should be NULL"); 2157 if (registerClosure) { 2158 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2159 } 2160 cmsSpace()->gc_prologue(); 2161 // Clear stat counters 2162 NOT_PRODUCT( 2163 assert(_numObjectsPromoted == 0, "check"); 2164 assert(_numWordsPromoted == 0, "check"); 2165 if (Verbose && PrintGC) { 2166 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, " 2167 SIZE_FORMAT " bytes concurrently", 2168 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2169 } 2170 _numObjectsAllocated = 0; 2171 _numWordsAllocated = 0; 2172 ) 2173 } 2174 2175 void CMSCollector::gc_epilogue(bool full) { 2176 // The following locking discipline assumes that we are only called 2177 // when the world is stopped. 2178 assert(SafepointSynchronize::is_at_safepoint(), 2179 "world is stopped assumption"); 2180 2181 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2182 // if linear allocation blocks need to be appropriately marked to allow the 2183 // the blocks to be parsable. We also check here whether we need to nudge the 2184 // CMS collector thread to start a new cycle (if it's not already active). 2185 assert( Thread::current()->is_VM_thread() 2186 || ( CMSScavengeBeforeRemark 2187 && Thread::current()->is_ConcurrentGC_thread()), 2188 "Incorrect thread type for epilogue execution"); 2189 2190 if (!_between_prologue_and_epilogue) { 2191 // We have already been invoked; this is a gc_epilogue delegation 2192 // from yet another CMS generation that we are responsible for, just 2193 // ignore it since all relevant work has already been done. 2194 return; 2195 } 2196 assert(haveFreelistLocks(), "must have freelist locks"); 2197 assert_lock_strong(bitMapLock()); 2198 2199 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2200 2201 _cmsGen->gc_epilogue_work(full); 2202 2203 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2204 // in case sampling was not already enabled, enable it 2205 _start_sampling = true; 2206 } 2207 // reset _eden_chunk_array so sampling starts afresh 2208 _eden_chunk_index = 0; 2209 2210 size_t cms_used = _cmsGen->cmsSpace()->used(); 2211 2212 // update performance counters - this uses a special version of 2213 // update_counters() that allows the utilization to be passed as a 2214 // parameter, avoiding multiple calls to used(). 2215 // 2216 _cmsGen->update_counters(cms_used); 2217 2218 bitMapLock()->unlock(); 2219 releaseFreelistLocks(); 2220 2221 if (!CleanChunkPoolAsync) { 2222 Chunk::clean_chunk_pool(); 2223 } 2224 2225 set_did_compact(false); 2226 _between_prologue_and_epilogue = false; // ready for next cycle 2227 } 2228 2229 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2230 collector()->gc_epilogue(full); 2231 2232 // Also reset promotion tracking in par gc thread states. 2233 for (uint i = 0; i < ParallelGCThreads; i++) { 2234 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2235 } 2236 } 2237 2238 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2239 assert(!incremental_collection_failed(), "Should have been cleared"); 2240 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2241 cmsSpace()->gc_epilogue(); 2242 // Print stat counters 2243 NOT_PRODUCT( 2244 assert(_numObjectsAllocated == 0, "check"); 2245 assert(_numWordsAllocated == 0, "check"); 2246 if (Verbose && PrintGC) { 2247 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, " 2248 SIZE_FORMAT " bytes", 2249 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2250 } 2251 _numObjectsPromoted = 0; 2252 _numWordsPromoted = 0; 2253 ) 2254 2255 if (PrintGC && Verbose) { 2256 // Call down the chain in contiguous_available needs the freelistLock 2257 // so print this out before releasing the freeListLock. 2258 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ", 2259 contiguous_available()); 2260 } 2261 } 2262 2263 #ifndef PRODUCT 2264 bool CMSCollector::have_cms_token() { 2265 Thread* thr = Thread::current(); 2266 if (thr->is_VM_thread()) { 2267 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2268 } else if (thr->is_ConcurrentGC_thread()) { 2269 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2270 } else if (thr->is_GC_task_thread()) { 2271 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2272 ParGCRareEvent_lock->owned_by_self(); 2273 } 2274 return false; 2275 } 2276 #endif 2277 2278 // Check reachability of the given heap address in CMS generation, 2279 // treating all other generations as roots. 2280 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2281 // We could "guarantee" below, rather than assert, but I'll 2282 // leave these as "asserts" so that an adventurous debugger 2283 // could try this in the product build provided some subset of 2284 // the conditions were met, provided they were interested in the 2285 // results and knew that the computation below wouldn't interfere 2286 // with other concurrent computations mutating the structures 2287 // being read or written. 2288 assert(SafepointSynchronize::is_at_safepoint(), 2289 "Else mutations in object graph will make answer suspect"); 2290 assert(have_cms_token(), "Should hold cms token"); 2291 assert(haveFreelistLocks(), "must hold free list locks"); 2292 assert_lock_strong(bitMapLock()); 2293 2294 // Clear the marking bit map array before starting, but, just 2295 // for kicks, first report if the given address is already marked 2296 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2297 _markBitMap.isMarked(addr) ? "" : " not"); 2298 2299 if (verify_after_remark()) { 2300 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2301 bool result = verification_mark_bm()->isMarked(addr); 2302 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2303 result ? "IS" : "is NOT"); 2304 return result; 2305 } else { 2306 gclog_or_tty->print_cr("Could not compute result"); 2307 return false; 2308 } 2309 } 2310 2311 2312 void 2313 CMSCollector::print_on_error(outputStream* st) { 2314 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2315 if (collector != NULL) { 2316 CMSBitMap* bitmap = &collector->_markBitMap; 2317 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2318 bitmap->print_on_error(st, " Bits: "); 2319 2320 st->cr(); 2321 2322 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2323 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2324 mut_bitmap->print_on_error(st, " Bits: "); 2325 } 2326 } 2327 2328 //////////////////////////////////////////////////////// 2329 // CMS Verification Support 2330 //////////////////////////////////////////////////////// 2331 // Following the remark phase, the following invariant 2332 // should hold -- each object in the CMS heap which is 2333 // marked in markBitMap() should be marked in the verification_mark_bm(). 2334 2335 class VerifyMarkedClosure: public BitMapClosure { 2336 CMSBitMap* _marks; 2337 bool _failed; 2338 2339 public: 2340 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2341 2342 bool do_bit(size_t offset) { 2343 HeapWord* addr = _marks->offsetToHeapWord(offset); 2344 if (!_marks->isMarked(addr)) { 2345 oop(addr)->print_on(gclog_or_tty); 2346 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2347 _failed = true; 2348 } 2349 return true; 2350 } 2351 2352 bool failed() { return _failed; } 2353 }; 2354 2355 bool CMSCollector::verify_after_remark(bool silent) { 2356 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2357 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2358 static bool init = false; 2359 2360 assert(SafepointSynchronize::is_at_safepoint(), 2361 "Else mutations in object graph will make answer suspect"); 2362 assert(have_cms_token(), 2363 "Else there may be mutual interference in use of " 2364 " verification data structures"); 2365 assert(_collectorState > Marking && _collectorState <= Sweeping, 2366 "Else marking info checked here may be obsolete"); 2367 assert(haveFreelistLocks(), "must hold free list locks"); 2368 assert_lock_strong(bitMapLock()); 2369 2370 2371 // Allocate marking bit map if not already allocated 2372 if (!init) { // first time 2373 if (!verification_mark_bm()->allocate(_span)) { 2374 return false; 2375 } 2376 init = true; 2377 } 2378 2379 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2380 2381 // Turn off refs discovery -- so we will be tracing through refs. 2382 // This is as intended, because by this time 2383 // GC must already have cleared any refs that need to be cleared, 2384 // and traced those that need to be marked; moreover, 2385 // the marking done here is not going to interfere in any 2386 // way with the marking information used by GC. 2387 NoRefDiscovery no_discovery(ref_processor()); 2388 2389 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2390 2391 // Clear any marks from a previous round 2392 verification_mark_bm()->clear_all(); 2393 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2394 verify_work_stacks_empty(); 2395 2396 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2397 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2398 // Update the saved marks which may affect the root scans. 2399 gch->save_marks(); 2400 2401 if (CMSRemarkVerifyVariant == 1) { 2402 // In this first variant of verification, we complete 2403 // all marking, then check if the new marks-vector is 2404 // a subset of the CMS marks-vector. 2405 verify_after_remark_work_1(); 2406 } else if (CMSRemarkVerifyVariant == 2) { 2407 // In this second variant of verification, we flag an error 2408 // (i.e. an object reachable in the new marks-vector not reachable 2409 // in the CMS marks-vector) immediately, also indicating the 2410 // identify of an object (A) that references the unmarked object (B) -- 2411 // presumably, a mutation to A failed to be picked up by preclean/remark? 2412 verify_after_remark_work_2(); 2413 } else { 2414 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2415 CMSRemarkVerifyVariant); 2416 } 2417 if (!silent) gclog_or_tty->print(" done] "); 2418 return true; 2419 } 2420 2421 void CMSCollector::verify_after_remark_work_1() { 2422 ResourceMark rm; 2423 HandleMark hm; 2424 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2425 2426 // Get a clear set of claim bits for the roots processing to work with. 2427 ClassLoaderDataGraph::clear_claimed_marks(); 2428 2429 // Mark from roots one level into CMS 2430 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2431 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2432 2433 { 2434 StrongRootsScope srs(1); 2435 2436 gch->gen_process_roots(&srs, 2437 GenCollectedHeap::OldGen, 2438 true, // younger gens are roots 2439 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2440 should_unload_classes(), 2441 ¬Older, 2442 NULL, 2443 NULL); 2444 } 2445 2446 // Now mark from the roots 2447 MarkFromRootsClosure markFromRootsClosure(this, _span, 2448 verification_mark_bm(), verification_mark_stack(), 2449 false /* don't yield */, true /* verifying */); 2450 assert(_restart_addr == NULL, "Expected pre-condition"); 2451 verification_mark_bm()->iterate(&markFromRootsClosure); 2452 while (_restart_addr != NULL) { 2453 // Deal with stack overflow: by restarting at the indicated 2454 // address. 2455 HeapWord* ra = _restart_addr; 2456 markFromRootsClosure.reset(ra); 2457 _restart_addr = NULL; 2458 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2459 } 2460 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2461 verify_work_stacks_empty(); 2462 2463 // Marking completed -- now verify that each bit marked in 2464 // verification_mark_bm() is also marked in markBitMap(); flag all 2465 // errors by printing corresponding objects. 2466 VerifyMarkedClosure vcl(markBitMap()); 2467 verification_mark_bm()->iterate(&vcl); 2468 if (vcl.failed()) { 2469 gclog_or_tty->print("Verification failed"); 2470 gch->print_on(gclog_or_tty); 2471 fatal("CMS: failed marking verification after remark"); 2472 } 2473 } 2474 2475 class VerifyKlassOopsKlassClosure : public KlassClosure { 2476 class VerifyKlassOopsClosure : public OopClosure { 2477 CMSBitMap* _bitmap; 2478 public: 2479 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2480 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2481 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2482 } _oop_closure; 2483 public: 2484 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2485 void do_klass(Klass* k) { 2486 k->oops_do(&_oop_closure); 2487 } 2488 }; 2489 2490 void CMSCollector::verify_after_remark_work_2() { 2491 ResourceMark rm; 2492 HandleMark hm; 2493 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2494 2495 // Get a clear set of claim bits for the roots processing to work with. 2496 ClassLoaderDataGraph::clear_claimed_marks(); 2497 2498 // Mark from roots one level into CMS 2499 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2500 markBitMap()); 2501 CLDToOopClosure cld_closure(¬Older, true); 2502 2503 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2504 2505 { 2506 StrongRootsScope srs(1); 2507 2508 gch->gen_process_roots(&srs, 2509 GenCollectedHeap::OldGen, 2510 true, // younger gens are roots 2511 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2512 should_unload_classes(), 2513 ¬Older, 2514 NULL, 2515 &cld_closure); 2516 } 2517 2518 // Now mark from the roots 2519 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2520 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2521 assert(_restart_addr == NULL, "Expected pre-condition"); 2522 verification_mark_bm()->iterate(&markFromRootsClosure); 2523 while (_restart_addr != NULL) { 2524 // Deal with stack overflow: by restarting at the indicated 2525 // address. 2526 HeapWord* ra = _restart_addr; 2527 markFromRootsClosure.reset(ra); 2528 _restart_addr = NULL; 2529 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2530 } 2531 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2532 verify_work_stacks_empty(); 2533 2534 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2535 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2536 2537 // Marking completed -- now verify that each bit marked in 2538 // verification_mark_bm() is also marked in markBitMap(); flag all 2539 // errors by printing corresponding objects. 2540 VerifyMarkedClosure vcl(markBitMap()); 2541 verification_mark_bm()->iterate(&vcl); 2542 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2543 } 2544 2545 void ConcurrentMarkSweepGeneration::save_marks() { 2546 // delegate to CMS space 2547 cmsSpace()->save_marks(); 2548 for (uint i = 0; i < ParallelGCThreads; i++) { 2549 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2550 } 2551 } 2552 2553 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2554 return cmsSpace()->no_allocs_since_save_marks(); 2555 } 2556 2557 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2558 \ 2559 void ConcurrentMarkSweepGeneration:: \ 2560 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2561 cl->set_generation(this); \ 2562 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2563 cl->reset_generation(); \ 2564 save_marks(); \ 2565 } 2566 2567 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2568 2569 void 2570 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2571 if (freelistLock()->owned_by_self()) { 2572 Generation::oop_iterate(cl); 2573 } else { 2574 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2575 Generation::oop_iterate(cl); 2576 } 2577 } 2578 2579 void 2580 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2581 if (freelistLock()->owned_by_self()) { 2582 Generation::object_iterate(cl); 2583 } else { 2584 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2585 Generation::object_iterate(cl); 2586 } 2587 } 2588 2589 void 2590 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2591 if (freelistLock()->owned_by_self()) { 2592 Generation::safe_object_iterate(cl); 2593 } else { 2594 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2595 Generation::safe_object_iterate(cl); 2596 } 2597 } 2598 2599 void 2600 ConcurrentMarkSweepGeneration::post_compact() { 2601 } 2602 2603 void 2604 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2605 // Fix the linear allocation blocks to look like free blocks. 2606 2607 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2608 // are not called when the heap is verified during universe initialization and 2609 // at vm shutdown. 2610 if (freelistLock()->owned_by_self()) { 2611 cmsSpace()->prepare_for_verify(); 2612 } else { 2613 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2614 cmsSpace()->prepare_for_verify(); 2615 } 2616 } 2617 2618 void 2619 ConcurrentMarkSweepGeneration::verify() { 2620 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2621 // are not called when the heap is verified during universe initialization and 2622 // at vm shutdown. 2623 if (freelistLock()->owned_by_self()) { 2624 cmsSpace()->verify(); 2625 } else { 2626 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2627 cmsSpace()->verify(); 2628 } 2629 } 2630 2631 void CMSCollector::verify() { 2632 _cmsGen->verify(); 2633 } 2634 2635 #ifndef PRODUCT 2636 bool CMSCollector::overflow_list_is_empty() const { 2637 assert(_num_par_pushes >= 0, "Inconsistency"); 2638 if (_overflow_list == NULL) { 2639 assert(_num_par_pushes == 0, "Inconsistency"); 2640 } 2641 return _overflow_list == NULL; 2642 } 2643 2644 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2645 // merely consolidate assertion checks that appear to occur together frequently. 2646 void CMSCollector::verify_work_stacks_empty() const { 2647 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2648 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2649 } 2650 2651 void CMSCollector::verify_overflow_empty() const { 2652 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2653 assert(no_preserved_marks(), "No preserved marks"); 2654 } 2655 #endif // PRODUCT 2656 2657 // Decide if we want to enable class unloading as part of the 2658 // ensuing concurrent GC cycle. We will collect and 2659 // unload classes if it's the case that: 2660 // (1) an explicit gc request has been made and the flag 2661 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2662 // (2) (a) class unloading is enabled at the command line, and 2663 // (b) old gen is getting really full 2664 // NOTE: Provided there is no change in the state of the heap between 2665 // calls to this method, it should have idempotent results. Moreover, 2666 // its results should be monotonically increasing (i.e. going from 0 to 1, 2667 // but not 1 to 0) between successive calls between which the heap was 2668 // not collected. For the implementation below, it must thus rely on 2669 // the property that concurrent_cycles_since_last_unload() 2670 // will not decrease unless a collection cycle happened and that 2671 // _cmsGen->is_too_full() are 2672 // themselves also monotonic in that sense. See check_monotonicity() 2673 // below. 2674 void CMSCollector::update_should_unload_classes() { 2675 _should_unload_classes = false; 2676 // Condition 1 above 2677 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2678 _should_unload_classes = true; 2679 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2680 // Disjuncts 2.b.(i,ii,iii) above 2681 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2682 CMSClassUnloadingMaxInterval) 2683 || _cmsGen->is_too_full(); 2684 } 2685 } 2686 2687 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2688 bool res = should_concurrent_collect(); 2689 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2690 return res; 2691 } 2692 2693 void CMSCollector::setup_cms_unloading_and_verification_state() { 2694 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2695 || VerifyBeforeExit; 2696 const int rso = GenCollectedHeap::SO_AllCodeCache; 2697 2698 // We set the proper root for this CMS cycle here. 2699 if (should_unload_classes()) { // Should unload classes this cycle 2700 remove_root_scanning_option(rso); // Shrink the root set appropriately 2701 set_verifying(should_verify); // Set verification state for this cycle 2702 return; // Nothing else needs to be done at this time 2703 } 2704 2705 // Not unloading classes this cycle 2706 assert(!should_unload_classes(), "Inconsistency!"); 2707 2708 // If we are not unloading classes then add SO_AllCodeCache to root 2709 // scanning options. 2710 add_root_scanning_option(rso); 2711 2712 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2713 set_verifying(true); 2714 } else if (verifying() && !should_verify) { 2715 // We were verifying, but some verification flags got disabled. 2716 set_verifying(false); 2717 // Exclude symbols, strings and code cache elements from root scanning to 2718 // reduce IM and RM pauses. 2719 remove_root_scanning_option(rso); 2720 } 2721 } 2722 2723 2724 #ifndef PRODUCT 2725 HeapWord* CMSCollector::block_start(const void* p) const { 2726 const HeapWord* addr = (HeapWord*)p; 2727 if (_span.contains(p)) { 2728 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2729 return _cmsGen->cmsSpace()->block_start(p); 2730 } 2731 } 2732 return NULL; 2733 } 2734 #endif 2735 2736 HeapWord* 2737 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2738 bool tlab, 2739 bool parallel) { 2740 CMSSynchronousYieldRequest yr; 2741 assert(!tlab, "Can't deal with TLAB allocation"); 2742 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2743 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2744 if (GCExpandToAllocateDelayMillis > 0) { 2745 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2746 } 2747 return have_lock_and_allocate(word_size, tlab); 2748 } 2749 2750 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2751 size_t bytes, 2752 size_t expand_bytes, 2753 CMSExpansionCause::Cause cause) 2754 { 2755 2756 bool success = expand(bytes, expand_bytes); 2757 2758 // remember why we expanded; this information is used 2759 // by shouldConcurrentCollect() when making decisions on whether to start 2760 // a new CMS cycle. 2761 if (success) { 2762 set_expansion_cause(cause); 2763 if (PrintGCDetails && Verbose) { 2764 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2765 CMSExpansionCause::to_string(cause)); 2766 } 2767 } 2768 } 2769 2770 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2771 HeapWord* res = NULL; 2772 MutexLocker x(ParGCRareEvent_lock); 2773 while (true) { 2774 // Expansion by some other thread might make alloc OK now: 2775 res = ps->lab.alloc(word_sz); 2776 if (res != NULL) return res; 2777 // If there's not enough expansion space available, give up. 2778 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2779 return NULL; 2780 } 2781 // Otherwise, we try expansion. 2782 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2783 // Now go around the loop and try alloc again; 2784 // A competing par_promote might beat us to the expansion space, 2785 // so we may go around the loop again if promotion fails again. 2786 if (GCExpandToAllocateDelayMillis > 0) { 2787 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2788 } 2789 } 2790 } 2791 2792 2793 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2794 PromotionInfo* promo) { 2795 MutexLocker x(ParGCRareEvent_lock); 2796 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2797 while (true) { 2798 // Expansion by some other thread might make alloc OK now: 2799 if (promo->ensure_spooling_space()) { 2800 assert(promo->has_spooling_space(), 2801 "Post-condition of successful ensure_spooling_space()"); 2802 return true; 2803 } 2804 // If there's not enough expansion space available, give up. 2805 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2806 return false; 2807 } 2808 // Otherwise, we try expansion. 2809 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2810 // Now go around the loop and try alloc again; 2811 // A competing allocation might beat us to the expansion space, 2812 // so we may go around the loop again if allocation fails again. 2813 if (GCExpandToAllocateDelayMillis > 0) { 2814 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2815 } 2816 } 2817 } 2818 2819 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2820 // Only shrink if a compaction was done so that all the free space 2821 // in the generation is in a contiguous block at the end. 2822 if (did_compact()) { 2823 CardGeneration::shrink(bytes); 2824 } 2825 } 2826 2827 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2828 assert_locked_or_safepoint(Heap_lock); 2829 } 2830 2831 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2832 assert_locked_or_safepoint(Heap_lock); 2833 assert_lock_strong(freelistLock()); 2834 if (PrintGCDetails && Verbose) { 2835 warning("Shrinking of CMS not yet implemented"); 2836 } 2837 return; 2838 } 2839 2840 2841 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2842 // phases. 2843 class CMSPhaseAccounting: public StackObj { 2844 public: 2845 CMSPhaseAccounting(CMSCollector *collector, 2846 const char *phase, 2847 const GCId gc_id, 2848 bool print_cr = true); 2849 ~CMSPhaseAccounting(); 2850 2851 private: 2852 CMSCollector *_collector; 2853 const char *_phase; 2854 elapsedTimer _wallclock; 2855 bool _print_cr; 2856 const GCId _gc_id; 2857 2858 public: 2859 // Not MT-safe; so do not pass around these StackObj's 2860 // where they may be accessed by other threads. 2861 jlong wallclock_millis() { 2862 assert(_wallclock.is_active(), "Wall clock should not stop"); 2863 _wallclock.stop(); // to record time 2864 jlong ret = _wallclock.milliseconds(); 2865 _wallclock.start(); // restart 2866 return ret; 2867 } 2868 }; 2869 2870 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2871 const char *phase, 2872 const GCId gc_id, 2873 bool print_cr) : 2874 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2875 2876 if (PrintCMSStatistics != 0) { 2877 _collector->resetYields(); 2878 } 2879 if (PrintGCDetails) { 2880 gclog_or_tty->gclog_stamp(_gc_id); 2881 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2882 _collector->cmsGen()->short_name(), _phase); 2883 } 2884 _collector->resetTimer(); 2885 _wallclock.start(); 2886 _collector->startTimer(); 2887 } 2888 2889 CMSPhaseAccounting::~CMSPhaseAccounting() { 2890 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2891 _collector->stopTimer(); 2892 _wallclock.stop(); 2893 if (PrintGCDetails) { 2894 gclog_or_tty->gclog_stamp(_gc_id); 2895 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2896 _collector->cmsGen()->short_name(), 2897 _phase, _collector->timerValue(), _wallclock.seconds()); 2898 if (_print_cr) { 2899 gclog_or_tty->cr(); 2900 } 2901 if (PrintCMSStatistics != 0) { 2902 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2903 _collector->yields()); 2904 } 2905 } 2906 } 2907 2908 // CMS work 2909 2910 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2911 class CMSParMarkTask : public AbstractGangTask { 2912 protected: 2913 CMSCollector* _collector; 2914 uint _n_workers; 2915 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2916 AbstractGangTask(name), 2917 _collector(collector), 2918 _n_workers(n_workers) {} 2919 // Work method in support of parallel rescan ... of young gen spaces 2920 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2921 ContiguousSpace* space, 2922 HeapWord** chunk_array, size_t chunk_top); 2923 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2924 }; 2925 2926 // Parallel initial mark task 2927 class CMSParInitialMarkTask: public CMSParMarkTask { 2928 StrongRootsScope* _strong_roots_scope; 2929 public: 2930 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2931 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2932 _strong_roots_scope(strong_roots_scope) {} 2933 void work(uint worker_id); 2934 }; 2935 2936 // Checkpoint the roots into this generation from outside 2937 // this generation. [Note this initial checkpoint need only 2938 // be approximate -- we'll do a catch up phase subsequently.] 2939 void CMSCollector::checkpointRootsInitial() { 2940 assert(_collectorState == InitialMarking, "Wrong collector state"); 2941 check_correct_thread_executing(); 2942 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2943 2944 save_heap_summary(); 2945 report_heap_summary(GCWhen::BeforeGC); 2946 2947 ReferenceProcessor* rp = ref_processor(); 2948 assert(_restart_addr == NULL, "Control point invariant"); 2949 { 2950 // acquire locks for subsequent manipulations 2951 MutexLockerEx x(bitMapLock(), 2952 Mutex::_no_safepoint_check_flag); 2953 checkpointRootsInitialWork(); 2954 // enable ("weak") refs discovery 2955 rp->enable_discovery(); 2956 _collectorState = Marking; 2957 } 2958 } 2959 2960 void CMSCollector::checkpointRootsInitialWork() { 2961 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2962 assert(_collectorState == InitialMarking, "just checking"); 2963 2964 // If there has not been a GC[n-1] since last GC[n] cycle completed, 2965 // precede our marking with a collection of all 2966 // younger generations to keep floating garbage to a minimum. 2967 // XXX: we won't do this for now -- it's an optimization to be done later. 2968 2969 // already have locks 2970 assert_lock_strong(bitMapLock()); 2971 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2972 2973 // Setup the verification and class unloading state for this 2974 // CMS collection cycle. 2975 setup_cms_unloading_and_verification_state(); 2976 2977 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2978 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2979 2980 // Reset all the PLAB chunk arrays if necessary. 2981 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2982 reset_survivor_plab_arrays(); 2983 } 2984 2985 ResourceMark rm; 2986 HandleMark hm; 2987 2988 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2989 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2990 2991 verify_work_stacks_empty(); 2992 verify_overflow_empty(); 2993 2994 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2995 // Update the saved marks which may affect the root scans. 2996 gch->save_marks(); 2997 2998 // weak reference processing has not started yet. 2999 ref_processor()->set_enqueuing_is_done(false); 3000 3001 // Need to remember all newly created CLDs, 3002 // so that we can guarantee that the remark finds them. 3003 ClassLoaderDataGraph::remember_new_clds(true); 3004 3005 // Whenever a CLD is found, it will be claimed before proceeding to mark 3006 // the klasses. The claimed marks need to be cleared before marking starts. 3007 ClassLoaderDataGraph::clear_claimed_marks(); 3008 3009 if (CMSPrintEdenSurvivorChunks) { 3010 print_eden_and_survivor_chunk_arrays(); 3011 } 3012 3013 { 3014 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3015 if (CMSParallelInitialMarkEnabled) { 3016 // The parallel version. 3017 FlexibleWorkGang* workers = gch->workers(); 3018 assert(workers != NULL, "Need parallel worker threads."); 3019 uint n_workers = workers->active_workers(); 3020 3021 StrongRootsScope srs(n_workers); 3022 3023 CMSParInitialMarkTask tsk(this, &srs, n_workers); 3024 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3025 if (n_workers > 1) { 3026 workers->run_task(&tsk); 3027 } else { 3028 tsk.work(0); 3029 } 3030 } else { 3031 // The serial version. 3032 CLDToOopClosure cld_closure(¬Older, true); 3033 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3034 3035 StrongRootsScope srs(1); 3036 3037 gch->gen_process_roots(&srs, 3038 GenCollectedHeap::OldGen, 3039 true, // younger gens are roots 3040 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3041 should_unload_classes(), 3042 ¬Older, 3043 NULL, 3044 &cld_closure); 3045 } 3046 } 3047 3048 // Clear mod-union table; it will be dirtied in the prologue of 3049 // CMS generation per each younger generation collection. 3050 3051 assert(_modUnionTable.isAllClear(), 3052 "Was cleared in most recent final checkpoint phase" 3053 " or no bits are set in the gc_prologue before the start of the next " 3054 "subsequent marking phase."); 3055 3056 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3057 3058 // Save the end of the used_region of the constituent generations 3059 // to be used to limit the extent of sweep in each generation. 3060 save_sweep_limits(); 3061 verify_overflow_empty(); 3062 } 3063 3064 bool CMSCollector::markFromRoots() { 3065 // we might be tempted to assert that: 3066 // assert(!SafepointSynchronize::is_at_safepoint(), 3067 // "inconsistent argument?"); 3068 // However that wouldn't be right, because it's possible that 3069 // a safepoint is indeed in progress as a younger generation 3070 // stop-the-world GC happens even as we mark in this generation. 3071 assert(_collectorState == Marking, "inconsistent state?"); 3072 check_correct_thread_executing(); 3073 verify_overflow_empty(); 3074 3075 // Weak ref discovery note: We may be discovering weak 3076 // refs in this generation concurrent (but interleaved) with 3077 // weak ref discovery by a younger generation collector. 3078 3079 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3080 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3081 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3082 bool res = markFromRootsWork(); 3083 if (res) { 3084 _collectorState = Precleaning; 3085 } else { // We failed and a foreground collection wants to take over 3086 assert(_foregroundGCIsActive, "internal state inconsistency"); 3087 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3088 if (PrintGCDetails) { 3089 gclog_or_tty->print_cr("bailing out to foreground collection"); 3090 } 3091 } 3092 verify_overflow_empty(); 3093 return res; 3094 } 3095 3096 bool CMSCollector::markFromRootsWork() { 3097 // iterate over marked bits in bit map, doing a full scan and mark 3098 // from these roots using the following algorithm: 3099 // . if oop is to the right of the current scan pointer, 3100 // mark corresponding bit (we'll process it later) 3101 // . else (oop is to left of current scan pointer) 3102 // push oop on marking stack 3103 // . drain the marking stack 3104 3105 // Note that when we do a marking step we need to hold the 3106 // bit map lock -- recall that direct allocation (by mutators) 3107 // and promotion (by younger generation collectors) is also 3108 // marking the bit map. [the so-called allocate live policy.] 3109 // Because the implementation of bit map marking is not 3110 // robust wrt simultaneous marking of bits in the same word, 3111 // we need to make sure that there is no such interference 3112 // between concurrent such updates. 3113 3114 // already have locks 3115 assert_lock_strong(bitMapLock()); 3116 3117 verify_work_stacks_empty(); 3118 verify_overflow_empty(); 3119 bool result = false; 3120 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3121 result = do_marking_mt(); 3122 } else { 3123 result = do_marking_st(); 3124 } 3125 return result; 3126 } 3127 3128 // Forward decl 3129 class CMSConcMarkingTask; 3130 3131 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3132 CMSCollector* _collector; 3133 CMSConcMarkingTask* _task; 3134 public: 3135 virtual void yield(); 3136 3137 // "n_threads" is the number of threads to be terminated. 3138 // "queue_set" is a set of work queues of other threads. 3139 // "collector" is the CMS collector associated with this task terminator. 3140 // "yield" indicates whether we need the gang as a whole to yield. 3141 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3142 ParallelTaskTerminator(n_threads, queue_set), 3143 _collector(collector) { } 3144 3145 void set_task(CMSConcMarkingTask* task) { 3146 _task = task; 3147 } 3148 }; 3149 3150 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3151 CMSConcMarkingTask* _task; 3152 public: 3153 bool should_exit_termination(); 3154 void set_task(CMSConcMarkingTask* task) { 3155 _task = task; 3156 } 3157 }; 3158 3159 // MT Concurrent Marking Task 3160 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3161 CMSCollector* _collector; 3162 uint _n_workers; // requested/desired # workers 3163 bool _result; 3164 CompactibleFreeListSpace* _cms_space; 3165 char _pad_front[64]; // padding to ... 3166 HeapWord* _global_finger; // ... avoid sharing cache line 3167 char _pad_back[64]; 3168 HeapWord* _restart_addr; 3169 3170 // Exposed here for yielding support 3171 Mutex* const _bit_map_lock; 3172 3173 // The per thread work queues, available here for stealing 3174 OopTaskQueueSet* _task_queues; 3175 3176 // Termination (and yielding) support 3177 CMSConcMarkingTerminator _term; 3178 CMSConcMarkingTerminatorTerminator _term_term; 3179 3180 public: 3181 CMSConcMarkingTask(CMSCollector* collector, 3182 CompactibleFreeListSpace* cms_space, 3183 YieldingFlexibleWorkGang* workers, 3184 OopTaskQueueSet* task_queues): 3185 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3186 _collector(collector), 3187 _cms_space(cms_space), 3188 _n_workers(0), _result(true), 3189 _task_queues(task_queues), 3190 _term(_n_workers, task_queues, _collector), 3191 _bit_map_lock(collector->bitMapLock()) 3192 { 3193 _requested_size = _n_workers; 3194 _term.set_task(this); 3195 _term_term.set_task(this); 3196 _restart_addr = _global_finger = _cms_space->bottom(); 3197 } 3198 3199 3200 OopTaskQueueSet* task_queues() { return _task_queues; } 3201 3202 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3203 3204 HeapWord** global_finger_addr() { return &_global_finger; } 3205 3206 CMSConcMarkingTerminator* terminator() { return &_term; } 3207 3208 virtual void set_for_termination(uint active_workers) { 3209 terminator()->reset_for_reuse(active_workers); 3210 } 3211 3212 void work(uint worker_id); 3213 bool should_yield() { 3214 return ConcurrentMarkSweepThread::should_yield() 3215 && !_collector->foregroundGCIsActive(); 3216 } 3217 3218 virtual void coordinator_yield(); // stuff done by coordinator 3219 bool result() { return _result; } 3220 3221 void reset(HeapWord* ra) { 3222 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3223 _restart_addr = _global_finger = ra; 3224 _term.reset_for_reuse(); 3225 } 3226 3227 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3228 OopTaskQueue* work_q); 3229 3230 private: 3231 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3232 void do_work_steal(int i); 3233 void bump_global_finger(HeapWord* f); 3234 }; 3235 3236 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3237 assert(_task != NULL, "Error"); 3238 return _task->yielding(); 3239 // Note that we do not need the disjunct || _task->should_yield() above 3240 // because we want terminating threads to yield only if the task 3241 // is already in the midst of yielding, which happens only after at least one 3242 // thread has yielded. 3243 } 3244 3245 void CMSConcMarkingTerminator::yield() { 3246 if (_task->should_yield()) { 3247 _task->yield(); 3248 } else { 3249 ParallelTaskTerminator::yield(); 3250 } 3251 } 3252 3253 //////////////////////////////////////////////////////////////// 3254 // Concurrent Marking Algorithm Sketch 3255 //////////////////////////////////////////////////////////////// 3256 // Until all tasks exhausted (both spaces): 3257 // -- claim next available chunk 3258 // -- bump global finger via CAS 3259 // -- find first object that starts in this chunk 3260 // and start scanning bitmap from that position 3261 // -- scan marked objects for oops 3262 // -- CAS-mark target, and if successful: 3263 // . if target oop is above global finger (volatile read) 3264 // nothing to do 3265 // . if target oop is in chunk and above local finger 3266 // then nothing to do 3267 // . else push on work-queue 3268 // -- Deal with possible overflow issues: 3269 // . local work-queue overflow causes stuff to be pushed on 3270 // global (common) overflow queue 3271 // . always first empty local work queue 3272 // . then get a batch of oops from global work queue if any 3273 // . then do work stealing 3274 // -- When all tasks claimed (both spaces) 3275 // and local work queue empty, 3276 // then in a loop do: 3277 // . check global overflow stack; steal a batch of oops and trace 3278 // . try to steal from other threads oif GOS is empty 3279 // . if neither is available, offer termination 3280 // -- Terminate and return result 3281 // 3282 void CMSConcMarkingTask::work(uint worker_id) { 3283 elapsedTimer _timer; 3284 ResourceMark rm; 3285 HandleMark hm; 3286 3287 DEBUG_ONLY(_collector->verify_overflow_empty();) 3288 3289 // Before we begin work, our work queue should be empty 3290 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3291 // Scan the bitmap covering _cms_space, tracing through grey objects. 3292 _timer.start(); 3293 do_scan_and_mark(worker_id, _cms_space); 3294 _timer.stop(); 3295 if (PrintCMSStatistics != 0) { 3296 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3297 worker_id, _timer.seconds()); 3298 // XXX: need xxx/xxx type of notation, two timers 3299 } 3300 3301 // ... do work stealing 3302 _timer.reset(); 3303 _timer.start(); 3304 do_work_steal(worker_id); 3305 _timer.stop(); 3306 if (PrintCMSStatistics != 0) { 3307 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3308 worker_id, _timer.seconds()); 3309 // XXX: need xxx/xxx type of notation, two timers 3310 } 3311 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3312 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3313 // Note that under the current task protocol, the 3314 // following assertion is true even of the spaces 3315 // expanded since the completion of the concurrent 3316 // marking. XXX This will likely change under a strict 3317 // ABORT semantics. 3318 // After perm removal the comparison was changed to 3319 // greater than or equal to from strictly greater than. 3320 // Before perm removal the highest address sweep would 3321 // have been at the end of perm gen but now is at the 3322 // end of the tenured gen. 3323 assert(_global_finger >= _cms_space->end(), 3324 "All tasks have been completed"); 3325 DEBUG_ONLY(_collector->verify_overflow_empty();) 3326 } 3327 3328 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3329 HeapWord* read = _global_finger; 3330 HeapWord* cur = read; 3331 while (f > read) { 3332 cur = read; 3333 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3334 if (cur == read) { 3335 // our cas succeeded 3336 assert(_global_finger >= f, "protocol consistency"); 3337 break; 3338 } 3339 } 3340 } 3341 3342 // This is really inefficient, and should be redone by 3343 // using (not yet available) block-read and -write interfaces to the 3344 // stack and the work_queue. XXX FIX ME !!! 3345 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3346 OopTaskQueue* work_q) { 3347 // Fast lock-free check 3348 if (ovflw_stk->length() == 0) { 3349 return false; 3350 } 3351 assert(work_q->size() == 0, "Shouldn't steal"); 3352 MutexLockerEx ml(ovflw_stk->par_lock(), 3353 Mutex::_no_safepoint_check_flag); 3354 // Grab up to 1/4 the size of the work queue 3355 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3356 (size_t)ParGCDesiredObjsFromOverflowList); 3357 num = MIN2(num, ovflw_stk->length()); 3358 for (int i = (int) num; i > 0; i--) { 3359 oop cur = ovflw_stk->pop(); 3360 assert(cur != NULL, "Counted wrong?"); 3361 work_q->push(cur); 3362 } 3363 return num > 0; 3364 } 3365 3366 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3367 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3368 int n_tasks = pst->n_tasks(); 3369 // We allow that there may be no tasks to do here because 3370 // we are restarting after a stack overflow. 3371 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3372 uint nth_task = 0; 3373 3374 HeapWord* aligned_start = sp->bottom(); 3375 if (sp->used_region().contains(_restart_addr)) { 3376 // Align down to a card boundary for the start of 0th task 3377 // for this space. 3378 aligned_start = 3379 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3380 CardTableModRefBS::card_size); 3381 } 3382 3383 size_t chunk_size = sp->marking_task_size(); 3384 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3385 // Having claimed the nth task in this space, 3386 // compute the chunk that it corresponds to: 3387 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3388 aligned_start + (nth_task+1)*chunk_size); 3389 // Try and bump the global finger via a CAS; 3390 // note that we need to do the global finger bump 3391 // _before_ taking the intersection below, because 3392 // the task corresponding to that region will be 3393 // deemed done even if the used_region() expands 3394 // because of allocation -- as it almost certainly will 3395 // during start-up while the threads yield in the 3396 // closure below. 3397 HeapWord* finger = span.end(); 3398 bump_global_finger(finger); // atomically 3399 // There are null tasks here corresponding to chunks 3400 // beyond the "top" address of the space. 3401 span = span.intersection(sp->used_region()); 3402 if (!span.is_empty()) { // Non-null task 3403 HeapWord* prev_obj; 3404 assert(!span.contains(_restart_addr) || nth_task == 0, 3405 "Inconsistency"); 3406 if (nth_task == 0) { 3407 // For the 0th task, we'll not need to compute a block_start. 3408 if (span.contains(_restart_addr)) { 3409 // In the case of a restart because of stack overflow, 3410 // we might additionally skip a chunk prefix. 3411 prev_obj = _restart_addr; 3412 } else { 3413 prev_obj = span.start(); 3414 } 3415 } else { 3416 // We want to skip the first object because 3417 // the protocol is to scan any object in its entirety 3418 // that _starts_ in this span; a fortiori, any 3419 // object starting in an earlier span is scanned 3420 // as part of an earlier claimed task. 3421 // Below we use the "careful" version of block_start 3422 // so we do not try to navigate uninitialized objects. 3423 prev_obj = sp->block_start_careful(span.start()); 3424 // Below we use a variant of block_size that uses the 3425 // Printezis bits to avoid waiting for allocated 3426 // objects to become initialized/parsable. 3427 while (prev_obj < span.start()) { 3428 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3429 if (sz > 0) { 3430 prev_obj += sz; 3431 } else { 3432 // In this case we may end up doing a bit of redundant 3433 // scanning, but that appears unavoidable, short of 3434 // locking the free list locks; see bug 6324141. 3435 break; 3436 } 3437 } 3438 } 3439 if (prev_obj < span.end()) { 3440 MemRegion my_span = MemRegion(prev_obj, span.end()); 3441 // Do the marking work within a non-empty span -- 3442 // the last argument to the constructor indicates whether the 3443 // iteration should be incremental with periodic yields. 3444 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3445 &_collector->_markBitMap, 3446 work_queue(i), 3447 &_collector->_markStack); 3448 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3449 } // else nothing to do for this task 3450 } // else nothing to do for this task 3451 } 3452 // We'd be tempted to assert here that since there are no 3453 // more tasks left to claim in this space, the global_finger 3454 // must exceed space->top() and a fortiori space->end(). However, 3455 // that would not quite be correct because the bumping of 3456 // global_finger occurs strictly after the claiming of a task, 3457 // so by the time we reach here the global finger may not yet 3458 // have been bumped up by the thread that claimed the last 3459 // task. 3460 pst->all_tasks_completed(); 3461 } 3462 3463 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3464 private: 3465 CMSCollector* _collector; 3466 CMSConcMarkingTask* _task; 3467 MemRegion _span; 3468 CMSBitMap* _bit_map; 3469 CMSMarkStack* _overflow_stack; 3470 OopTaskQueue* _work_queue; 3471 protected: 3472 DO_OOP_WORK_DEFN 3473 public: 3474 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3475 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3476 MetadataAwareOopClosure(collector->ref_processor()), 3477 _collector(collector), 3478 _task(task), 3479 _span(collector->_span), 3480 _work_queue(work_queue), 3481 _bit_map(bit_map), 3482 _overflow_stack(overflow_stack) 3483 { } 3484 virtual void do_oop(oop* p); 3485 virtual void do_oop(narrowOop* p); 3486 3487 void trim_queue(size_t max); 3488 void handle_stack_overflow(HeapWord* lost); 3489 void do_yield_check() { 3490 if (_task->should_yield()) { 3491 _task->yield(); 3492 } 3493 } 3494 }; 3495 3496 // Grey object scanning during work stealing phase -- 3497 // the salient assumption here is that any references 3498 // that are in these stolen objects being scanned must 3499 // already have been initialized (else they would not have 3500 // been published), so we do not need to check for 3501 // uninitialized objects before pushing here. 3502 void Par_ConcMarkingClosure::do_oop(oop obj) { 3503 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3504 HeapWord* addr = (HeapWord*)obj; 3505 // Check if oop points into the CMS generation 3506 // and is not marked 3507 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3508 // a white object ... 3509 // If we manage to "claim" the object, by being the 3510 // first thread to mark it, then we push it on our 3511 // marking stack 3512 if (_bit_map->par_mark(addr)) { // ... now grey 3513 // push on work queue (grey set) 3514 bool simulate_overflow = false; 3515 NOT_PRODUCT( 3516 if (CMSMarkStackOverflowALot && 3517 _collector->simulate_overflow()) { 3518 // simulate a stack overflow 3519 simulate_overflow = true; 3520 } 3521 ) 3522 if (simulate_overflow || 3523 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3524 // stack overflow 3525 if (PrintCMSStatistics != 0) { 3526 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3527 SIZE_FORMAT, _overflow_stack->capacity()); 3528 } 3529 // We cannot assert that the overflow stack is full because 3530 // it may have been emptied since. 3531 assert(simulate_overflow || 3532 _work_queue->size() == _work_queue->max_elems(), 3533 "Else push should have succeeded"); 3534 handle_stack_overflow(addr); 3535 } 3536 } // Else, some other thread got there first 3537 do_yield_check(); 3538 } 3539 } 3540 3541 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3542 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3543 3544 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3545 while (_work_queue->size() > max) { 3546 oop new_oop; 3547 if (_work_queue->pop_local(new_oop)) { 3548 assert(new_oop->is_oop(), "Should be an oop"); 3549 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3550 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3551 new_oop->oop_iterate(this); // do_oop() above 3552 do_yield_check(); 3553 } 3554 } 3555 } 3556 3557 // Upon stack overflow, we discard (part of) the stack, 3558 // remembering the least address amongst those discarded 3559 // in CMSCollector's _restart_address. 3560 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3561 // We need to do this under a mutex to prevent other 3562 // workers from interfering with the work done below. 3563 MutexLockerEx ml(_overflow_stack->par_lock(), 3564 Mutex::_no_safepoint_check_flag); 3565 // Remember the least grey address discarded 3566 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3567 _collector->lower_restart_addr(ra); 3568 _overflow_stack->reset(); // discard stack contents 3569 _overflow_stack->expand(); // expand the stack if possible 3570 } 3571 3572 3573 void CMSConcMarkingTask::do_work_steal(int i) { 3574 OopTaskQueue* work_q = work_queue(i); 3575 oop obj_to_scan; 3576 CMSBitMap* bm = &(_collector->_markBitMap); 3577 CMSMarkStack* ovflw = &(_collector->_markStack); 3578 int* seed = _collector->hash_seed(i); 3579 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3580 while (true) { 3581 cl.trim_queue(0); 3582 assert(work_q->size() == 0, "Should have been emptied above"); 3583 if (get_work_from_overflow_stack(ovflw, work_q)) { 3584 // Can't assert below because the work obtained from the 3585 // overflow stack may already have been stolen from us. 3586 // assert(work_q->size() > 0, "Work from overflow stack"); 3587 continue; 3588 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3589 assert(obj_to_scan->is_oop(), "Should be an oop"); 3590 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3591 obj_to_scan->oop_iterate(&cl); 3592 } else if (terminator()->offer_termination(&_term_term)) { 3593 assert(work_q->size() == 0, "Impossible!"); 3594 break; 3595 } else if (yielding() || should_yield()) { 3596 yield(); 3597 } 3598 } 3599 } 3600 3601 // This is run by the CMS (coordinator) thread. 3602 void CMSConcMarkingTask::coordinator_yield() { 3603 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3604 "CMS thread should hold CMS token"); 3605 // First give up the locks, then yield, then re-lock 3606 // We should probably use a constructor/destructor idiom to 3607 // do this unlock/lock or modify the MutexUnlocker class to 3608 // serve our purpose. XXX 3609 assert_lock_strong(_bit_map_lock); 3610 _bit_map_lock->unlock(); 3611 ConcurrentMarkSweepThread::desynchronize(true); 3612 _collector->stopTimer(); 3613 if (PrintCMSStatistics != 0) { 3614 _collector->incrementYields(); 3615 } 3616 3617 // It is possible for whichever thread initiated the yield request 3618 // not to get a chance to wake up and take the bitmap lock between 3619 // this thread releasing it and reacquiring it. So, while the 3620 // should_yield() flag is on, let's sleep for a bit to give the 3621 // other thread a chance to wake up. The limit imposed on the number 3622 // of iterations is defensive, to avoid any unforseen circumstances 3623 // putting us into an infinite loop. Since it's always been this 3624 // (coordinator_yield()) method that was observed to cause the 3625 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3626 // which is by default non-zero. For the other seven methods that 3627 // also perform the yield operation, as are using a different 3628 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3629 // can enable the sleeping for those methods too, if necessary. 3630 // See 6442774. 3631 // 3632 // We really need to reconsider the synchronization between the GC 3633 // thread and the yield-requesting threads in the future and we 3634 // should really use wait/notify, which is the recommended 3635 // way of doing this type of interaction. Additionally, we should 3636 // consolidate the eight methods that do the yield operation and they 3637 // are almost identical into one for better maintainability and 3638 // readability. See 6445193. 3639 // 3640 // Tony 2006.06.29 3641 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3642 ConcurrentMarkSweepThread::should_yield() && 3643 !CMSCollector::foregroundGCIsActive(); ++i) { 3644 os::sleep(Thread::current(), 1, false); 3645 } 3646 3647 ConcurrentMarkSweepThread::synchronize(true); 3648 _bit_map_lock->lock_without_safepoint_check(); 3649 _collector->startTimer(); 3650 } 3651 3652 bool CMSCollector::do_marking_mt() { 3653 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3654 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3655 conc_workers()->active_workers(), 3656 Threads::number_of_non_daemon_threads()); 3657 conc_workers()->set_active_workers(num_workers); 3658 3659 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3660 3661 CMSConcMarkingTask tsk(this, 3662 cms_space, 3663 conc_workers(), 3664 task_queues()); 3665 3666 // Since the actual number of workers we get may be different 3667 // from the number we requested above, do we need to do anything different 3668 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3669 // class?? XXX 3670 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3671 3672 // Refs discovery is already non-atomic. 3673 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3674 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3675 conc_workers()->start_task(&tsk); 3676 while (tsk.yielded()) { 3677 tsk.coordinator_yield(); 3678 conc_workers()->continue_task(&tsk); 3679 } 3680 // If the task was aborted, _restart_addr will be non-NULL 3681 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3682 while (_restart_addr != NULL) { 3683 // XXX For now we do not make use of ABORTED state and have not 3684 // yet implemented the right abort semantics (even in the original 3685 // single-threaded CMS case). That needs some more investigation 3686 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3687 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3688 // If _restart_addr is non-NULL, a marking stack overflow 3689 // occurred; we need to do a fresh marking iteration from the 3690 // indicated restart address. 3691 if (_foregroundGCIsActive) { 3692 // We may be running into repeated stack overflows, having 3693 // reached the limit of the stack size, while making very 3694 // slow forward progress. It may be best to bail out and 3695 // let the foreground collector do its job. 3696 // Clear _restart_addr, so that foreground GC 3697 // works from scratch. This avoids the headache of 3698 // a "rescan" which would otherwise be needed because 3699 // of the dirty mod union table & card table. 3700 _restart_addr = NULL; 3701 return false; 3702 } 3703 // Adjust the task to restart from _restart_addr 3704 tsk.reset(_restart_addr); 3705 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3706 _restart_addr); 3707 _restart_addr = NULL; 3708 // Get the workers going again 3709 conc_workers()->start_task(&tsk); 3710 while (tsk.yielded()) { 3711 tsk.coordinator_yield(); 3712 conc_workers()->continue_task(&tsk); 3713 } 3714 } 3715 assert(tsk.completed(), "Inconsistency"); 3716 assert(tsk.result() == true, "Inconsistency"); 3717 return true; 3718 } 3719 3720 bool CMSCollector::do_marking_st() { 3721 ResourceMark rm; 3722 HandleMark hm; 3723 3724 // Temporarily make refs discovery single threaded (non-MT) 3725 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3726 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3727 &_markStack, CMSYield); 3728 // the last argument to iterate indicates whether the iteration 3729 // should be incremental with periodic yields. 3730 _markBitMap.iterate(&markFromRootsClosure); 3731 // If _restart_addr is non-NULL, a marking stack overflow 3732 // occurred; we need to do a fresh iteration from the 3733 // indicated restart address. 3734 while (_restart_addr != NULL) { 3735 if (_foregroundGCIsActive) { 3736 // We may be running into repeated stack overflows, having 3737 // reached the limit of the stack size, while making very 3738 // slow forward progress. It may be best to bail out and 3739 // let the foreground collector do its job. 3740 // Clear _restart_addr, so that foreground GC 3741 // works from scratch. This avoids the headache of 3742 // a "rescan" which would otherwise be needed because 3743 // of the dirty mod union table & card table. 3744 _restart_addr = NULL; 3745 return false; // indicating failure to complete marking 3746 } 3747 // Deal with stack overflow: 3748 // we restart marking from _restart_addr 3749 HeapWord* ra = _restart_addr; 3750 markFromRootsClosure.reset(ra); 3751 _restart_addr = NULL; 3752 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3753 } 3754 return true; 3755 } 3756 3757 void CMSCollector::preclean() { 3758 check_correct_thread_executing(); 3759 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3760 verify_work_stacks_empty(); 3761 verify_overflow_empty(); 3762 _abort_preclean = false; 3763 if (CMSPrecleaningEnabled) { 3764 if (!CMSEdenChunksRecordAlways) { 3765 _eden_chunk_index = 0; 3766 } 3767 size_t used = get_eden_used(); 3768 size_t capacity = get_eden_capacity(); 3769 // Don't start sampling unless we will get sufficiently 3770 // many samples. 3771 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3772 * CMSScheduleRemarkEdenPenetration)) { 3773 _start_sampling = true; 3774 } else { 3775 _start_sampling = false; 3776 } 3777 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3778 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3779 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3780 } 3781 CMSTokenSync x(true); // is cms thread 3782 if (CMSPrecleaningEnabled) { 3783 sample_eden(); 3784 _collectorState = AbortablePreclean; 3785 } else { 3786 _collectorState = FinalMarking; 3787 } 3788 verify_work_stacks_empty(); 3789 verify_overflow_empty(); 3790 } 3791 3792 // Try and schedule the remark such that young gen 3793 // occupancy is CMSScheduleRemarkEdenPenetration %. 3794 void CMSCollector::abortable_preclean() { 3795 check_correct_thread_executing(); 3796 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3797 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3798 3799 // If Eden's current occupancy is below this threshold, 3800 // immediately schedule the remark; else preclean 3801 // past the next scavenge in an effort to 3802 // schedule the pause as described above. By choosing 3803 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3804 // we will never do an actual abortable preclean cycle. 3805 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3806 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3807 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3808 // We need more smarts in the abortable preclean 3809 // loop below to deal with cases where allocation 3810 // in young gen is very very slow, and our precleaning 3811 // is running a losing race against a horde of 3812 // mutators intent on flooding us with CMS updates 3813 // (dirty cards). 3814 // One, admittedly dumb, strategy is to give up 3815 // after a certain number of abortable precleaning loops 3816 // or after a certain maximum time. We want to make 3817 // this smarter in the next iteration. 3818 // XXX FIX ME!!! YSR 3819 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3820 while (!(should_abort_preclean() || 3821 ConcurrentMarkSweepThread::should_terminate())) { 3822 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3823 cumworkdone += workdone; 3824 loops++; 3825 // Voluntarily terminate abortable preclean phase if we have 3826 // been at it for too long. 3827 if ((CMSMaxAbortablePrecleanLoops != 0) && 3828 loops >= CMSMaxAbortablePrecleanLoops) { 3829 if (PrintGCDetails) { 3830 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3831 } 3832 break; 3833 } 3834 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3835 if (PrintGCDetails) { 3836 gclog_or_tty->print(" CMS: abort preclean due to time "); 3837 } 3838 break; 3839 } 3840 // If we are doing little work each iteration, we should 3841 // take a short break. 3842 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3843 // Sleep for some time, waiting for work to accumulate 3844 stopTimer(); 3845 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3846 startTimer(); 3847 waited++; 3848 } 3849 } 3850 if (PrintCMSStatistics > 0) { 3851 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3852 loops, waited, cumworkdone); 3853 } 3854 } 3855 CMSTokenSync x(true); // is cms thread 3856 if (_collectorState != Idling) { 3857 assert(_collectorState == AbortablePreclean, 3858 "Spontaneous state transition?"); 3859 _collectorState = FinalMarking; 3860 } // Else, a foreground collection completed this CMS cycle. 3861 return; 3862 } 3863 3864 // Respond to an Eden sampling opportunity 3865 void CMSCollector::sample_eden() { 3866 // Make sure a young gc cannot sneak in between our 3867 // reading and recording of a sample. 3868 assert(Thread::current()->is_ConcurrentGC_thread(), 3869 "Only the cms thread may collect Eden samples"); 3870 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3871 "Should collect samples while holding CMS token"); 3872 if (!_start_sampling) { 3873 return; 3874 } 3875 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3876 // is populated by the young generation. 3877 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3878 if (_eden_chunk_index < _eden_chunk_capacity) { 3879 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3880 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3881 "Unexpected state of Eden"); 3882 // We'd like to check that what we just sampled is an oop-start address; 3883 // however, we cannot do that here since the object may not yet have been 3884 // initialized. So we'll instead do the check when we _use_ this sample 3885 // later. 3886 if (_eden_chunk_index == 0 || 3887 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3888 _eden_chunk_array[_eden_chunk_index-1]) 3889 >= CMSSamplingGrain)) { 3890 _eden_chunk_index++; // commit sample 3891 } 3892 } 3893 } 3894 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3895 size_t used = get_eden_used(); 3896 size_t capacity = get_eden_capacity(); 3897 assert(used <= capacity, "Unexpected state of Eden"); 3898 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3899 _abort_preclean = true; 3900 } 3901 } 3902 } 3903 3904 3905 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3906 assert(_collectorState == Precleaning || 3907 _collectorState == AbortablePreclean, "incorrect state"); 3908 ResourceMark rm; 3909 HandleMark hm; 3910 3911 // Precleaning is currently not MT but the reference processor 3912 // may be set for MT. Disable it temporarily here. 3913 ReferenceProcessor* rp = ref_processor(); 3914 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3915 3916 // Do one pass of scrubbing the discovered reference lists 3917 // to remove any reference objects with strongly-reachable 3918 // referents. 3919 if (clean_refs) { 3920 CMSPrecleanRefsYieldClosure yield_cl(this); 3921 assert(rp->span().equals(_span), "Spans should be equal"); 3922 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3923 &_markStack, true /* preclean */); 3924 CMSDrainMarkingStackClosure complete_trace(this, 3925 _span, &_markBitMap, &_markStack, 3926 &keep_alive, true /* preclean */); 3927 3928 // We don't want this step to interfere with a young 3929 // collection because we don't want to take CPU 3930 // or memory bandwidth away from the young GC threads 3931 // (which may be as many as there are CPUs). 3932 // Note that we don't need to protect ourselves from 3933 // interference with mutators because they can't 3934 // manipulate the discovered reference lists nor affect 3935 // the computed reachability of the referents, the 3936 // only properties manipulated by the precleaning 3937 // of these reference lists. 3938 stopTimer(); 3939 CMSTokenSyncWithLocks x(true /* is cms thread */, 3940 bitMapLock()); 3941 startTimer(); 3942 sample_eden(); 3943 3944 // The following will yield to allow foreground 3945 // collection to proceed promptly. XXX YSR: 3946 // The code in this method may need further 3947 // tweaking for better performance and some restructuring 3948 // for cleaner interfaces. 3949 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3950 rp->preclean_discovered_references( 3951 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3952 gc_timer, _gc_tracer_cm->gc_id()); 3953 } 3954 3955 if (clean_survivor) { // preclean the active survivor space(s) 3956 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3957 &_markBitMap, &_modUnionTable, 3958 &_markStack, true /* precleaning phase */); 3959 stopTimer(); 3960 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3961 bitMapLock()); 3962 startTimer(); 3963 unsigned int before_count = 3964 GenCollectedHeap::heap()->total_collections(); 3965 SurvivorSpacePrecleanClosure 3966 sss_cl(this, _span, &_markBitMap, &_markStack, 3967 &pam_cl, before_count, CMSYield); 3968 _young_gen->from()->object_iterate_careful(&sss_cl); 3969 _young_gen->to()->object_iterate_careful(&sss_cl); 3970 } 3971 MarkRefsIntoAndScanClosure 3972 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3973 &_markStack, this, CMSYield, 3974 true /* precleaning phase */); 3975 // CAUTION: The following closure has persistent state that may need to 3976 // be reset upon a decrease in the sequence of addresses it 3977 // processes. 3978 ScanMarkedObjectsAgainCarefullyClosure 3979 smoac_cl(this, _span, 3980 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3981 3982 // Preclean dirty cards in ModUnionTable and CardTable using 3983 // appropriate convergence criterion; 3984 // repeat CMSPrecleanIter times unless we find that 3985 // we are losing. 3986 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3987 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3988 "Bad convergence multiplier"); 3989 assert(CMSPrecleanThreshold >= 100, 3990 "Unreasonably low CMSPrecleanThreshold"); 3991 3992 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3993 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3994 numIter < CMSPrecleanIter; 3995 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3996 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3997 if (Verbose && PrintGCDetails) { 3998 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3999 } 4000 // Either there are very few dirty cards, so re-mark 4001 // pause will be small anyway, or our pre-cleaning isn't 4002 // that much faster than the rate at which cards are being 4003 // dirtied, so we might as well stop and re-mark since 4004 // precleaning won't improve our re-mark time by much. 4005 if (curNumCards <= CMSPrecleanThreshold || 4006 (numIter > 0 && 4007 (curNumCards * CMSPrecleanDenominator > 4008 lastNumCards * CMSPrecleanNumerator))) { 4009 numIter++; 4010 cumNumCards += curNumCards; 4011 break; 4012 } 4013 } 4014 4015 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4016 4017 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4018 cumNumCards += curNumCards; 4019 if (PrintGCDetails && PrintCMSStatistics != 0) { 4020 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4021 curNumCards, cumNumCards, numIter); 4022 } 4023 return cumNumCards; // as a measure of useful work done 4024 } 4025 4026 // PRECLEANING NOTES: 4027 // Precleaning involves: 4028 // . reading the bits of the modUnionTable and clearing the set bits. 4029 // . For the cards corresponding to the set bits, we scan the 4030 // objects on those cards. This means we need the free_list_lock 4031 // so that we can safely iterate over the CMS space when scanning 4032 // for oops. 4033 // . When we scan the objects, we'll be both reading and setting 4034 // marks in the marking bit map, so we'll need the marking bit map. 4035 // . For protecting _collector_state transitions, we take the CGC_lock. 4036 // Note that any races in the reading of of card table entries by the 4037 // CMS thread on the one hand and the clearing of those entries by the 4038 // VM thread or the setting of those entries by the mutator threads on the 4039 // other are quite benign. However, for efficiency it makes sense to keep 4040 // the VM thread from racing with the CMS thread while the latter is 4041 // dirty card info to the modUnionTable. We therefore also use the 4042 // CGC_lock to protect the reading of the card table and the mod union 4043 // table by the CM thread. 4044 // . We run concurrently with mutator updates, so scanning 4045 // needs to be done carefully -- we should not try to scan 4046 // potentially uninitialized objects. 4047 // 4048 // Locking strategy: While holding the CGC_lock, we scan over and 4049 // reset a maximal dirty range of the mod union / card tables, then lock 4050 // the free_list_lock and bitmap lock to do a full marking, then 4051 // release these locks; and repeat the cycle. This allows for a 4052 // certain amount of fairness in the sharing of these locks between 4053 // the CMS collector on the one hand, and the VM thread and the 4054 // mutators on the other. 4055 4056 // NOTE: preclean_mod_union_table() and preclean_card_table() 4057 // further below are largely identical; if you need to modify 4058 // one of these methods, please check the other method too. 4059 4060 size_t CMSCollector::preclean_mod_union_table( 4061 ConcurrentMarkSweepGeneration* gen, 4062 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4063 verify_work_stacks_empty(); 4064 verify_overflow_empty(); 4065 4066 // strategy: starting with the first card, accumulate contiguous 4067 // ranges of dirty cards; clear these cards, then scan the region 4068 // covered by these cards. 4069 4070 // Since all of the MUT is committed ahead, we can just use 4071 // that, in case the generations expand while we are precleaning. 4072 // It might also be fine to just use the committed part of the 4073 // generation, but we might potentially miss cards when the 4074 // generation is rapidly expanding while we are in the midst 4075 // of precleaning. 4076 HeapWord* startAddr = gen->reserved().start(); 4077 HeapWord* endAddr = gen->reserved().end(); 4078 4079 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4080 4081 size_t numDirtyCards, cumNumDirtyCards; 4082 HeapWord *nextAddr, *lastAddr; 4083 for (cumNumDirtyCards = numDirtyCards = 0, 4084 nextAddr = lastAddr = startAddr; 4085 nextAddr < endAddr; 4086 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4087 4088 ResourceMark rm; 4089 HandleMark hm; 4090 4091 MemRegion dirtyRegion; 4092 { 4093 stopTimer(); 4094 // Potential yield point 4095 CMSTokenSync ts(true); 4096 startTimer(); 4097 sample_eden(); 4098 // Get dirty region starting at nextOffset (inclusive), 4099 // simultaneously clearing it. 4100 dirtyRegion = 4101 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4102 assert(dirtyRegion.start() >= nextAddr, 4103 "returned region inconsistent?"); 4104 } 4105 // Remember where the next search should begin. 4106 // The returned region (if non-empty) is a right open interval, 4107 // so lastOffset is obtained from the right end of that 4108 // interval. 4109 lastAddr = dirtyRegion.end(); 4110 // Should do something more transparent and less hacky XXX 4111 numDirtyCards = 4112 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4113 4114 // We'll scan the cards in the dirty region (with periodic 4115 // yields for foreground GC as needed). 4116 if (!dirtyRegion.is_empty()) { 4117 assert(numDirtyCards > 0, "consistency check"); 4118 HeapWord* stop_point = NULL; 4119 stopTimer(); 4120 // Potential yield point 4121 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4122 bitMapLock()); 4123 startTimer(); 4124 { 4125 verify_work_stacks_empty(); 4126 verify_overflow_empty(); 4127 sample_eden(); 4128 stop_point = 4129 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4130 } 4131 if (stop_point != NULL) { 4132 // The careful iteration stopped early either because it found an 4133 // uninitialized object, or because we were in the midst of an 4134 // "abortable preclean", which should now be aborted. Redirty 4135 // the bits corresponding to the partially-scanned or unscanned 4136 // cards. We'll either restart at the next block boundary or 4137 // abort the preclean. 4138 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4139 "Should only be AbortablePreclean."); 4140 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4141 if (should_abort_preclean()) { 4142 break; // out of preclean loop 4143 } else { 4144 // Compute the next address at which preclean should pick up; 4145 // might need bitMapLock in order to read P-bits. 4146 lastAddr = next_card_start_after_block(stop_point); 4147 } 4148 } 4149 } else { 4150 assert(lastAddr == endAddr, "consistency check"); 4151 assert(numDirtyCards == 0, "consistency check"); 4152 break; 4153 } 4154 } 4155 verify_work_stacks_empty(); 4156 verify_overflow_empty(); 4157 return cumNumDirtyCards; 4158 } 4159 4160 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4161 // below are largely identical; if you need to modify 4162 // one of these methods, please check the other method too. 4163 4164 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4165 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4166 // strategy: it's similar to precleamModUnionTable above, in that 4167 // we accumulate contiguous ranges of dirty cards, mark these cards 4168 // precleaned, then scan the region covered by these cards. 4169 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4170 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4171 4172 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4173 4174 size_t numDirtyCards, cumNumDirtyCards; 4175 HeapWord *lastAddr, *nextAddr; 4176 4177 for (cumNumDirtyCards = numDirtyCards = 0, 4178 nextAddr = lastAddr = startAddr; 4179 nextAddr < endAddr; 4180 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4181 4182 ResourceMark rm; 4183 HandleMark hm; 4184 4185 MemRegion dirtyRegion; 4186 { 4187 // See comments in "Precleaning notes" above on why we 4188 // do this locking. XXX Could the locking overheads be 4189 // too high when dirty cards are sparse? [I don't think so.] 4190 stopTimer(); 4191 CMSTokenSync x(true); // is cms thread 4192 startTimer(); 4193 sample_eden(); 4194 // Get and clear dirty region from card table 4195 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4196 MemRegion(nextAddr, endAddr), 4197 true, 4198 CardTableModRefBS::precleaned_card_val()); 4199 4200 assert(dirtyRegion.start() >= nextAddr, 4201 "returned region inconsistent?"); 4202 } 4203 lastAddr = dirtyRegion.end(); 4204 numDirtyCards = 4205 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4206 4207 if (!dirtyRegion.is_empty()) { 4208 stopTimer(); 4209 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4210 startTimer(); 4211 sample_eden(); 4212 verify_work_stacks_empty(); 4213 verify_overflow_empty(); 4214 HeapWord* stop_point = 4215 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4216 if (stop_point != NULL) { 4217 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4218 "Should only be AbortablePreclean."); 4219 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4220 if (should_abort_preclean()) { 4221 break; // out of preclean loop 4222 } else { 4223 // Compute the next address at which preclean should pick up. 4224 lastAddr = next_card_start_after_block(stop_point); 4225 } 4226 } 4227 } else { 4228 break; 4229 } 4230 } 4231 verify_work_stacks_empty(); 4232 verify_overflow_empty(); 4233 return cumNumDirtyCards; 4234 } 4235 4236 class PrecleanKlassClosure : public KlassClosure { 4237 KlassToOopClosure _cm_klass_closure; 4238 public: 4239 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4240 void do_klass(Klass* k) { 4241 if (k->has_accumulated_modified_oops()) { 4242 k->clear_accumulated_modified_oops(); 4243 4244 _cm_klass_closure.do_klass(k); 4245 } 4246 } 4247 }; 4248 4249 // The freelist lock is needed to prevent asserts, is it really needed? 4250 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4251 4252 cl->set_freelistLock(freelistLock); 4253 4254 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4255 4256 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4257 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4258 PrecleanKlassClosure preclean_klass_closure(cl); 4259 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4260 4261 verify_work_stacks_empty(); 4262 verify_overflow_empty(); 4263 } 4264 4265 void CMSCollector::checkpointRootsFinal() { 4266 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4267 check_correct_thread_executing(); 4268 // world is stopped at this checkpoint 4269 assert(SafepointSynchronize::is_at_safepoint(), 4270 "world should be stopped"); 4271 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4272 4273 verify_work_stacks_empty(); 4274 verify_overflow_empty(); 4275 4276 if (PrintGCDetails) { 4277 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]", 4278 _young_gen->used() / K, 4279 _young_gen->capacity() / K); 4280 } 4281 { 4282 if (CMSScavengeBeforeRemark) { 4283 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4284 // Temporarily set flag to false, GCH->do_collection will 4285 // expect it to be false and set to true 4286 FlagSetting fl(gch->_is_gc_active, false); 4287 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4288 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4289 gch->do_collection(true, // full (i.e. force, see below) 4290 false, // !clear_all_soft_refs 4291 0, // size 4292 false, // is_tlab 4293 GenCollectedHeap::YoungGen // type 4294 ); 4295 } 4296 FreelistLocker x(this); 4297 MutexLockerEx y(bitMapLock(), 4298 Mutex::_no_safepoint_check_flag); 4299 checkpointRootsFinalWork(); 4300 } 4301 verify_work_stacks_empty(); 4302 verify_overflow_empty(); 4303 } 4304 4305 void CMSCollector::checkpointRootsFinalWork() { 4306 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4307 4308 assert(haveFreelistLocks(), "must have free list locks"); 4309 assert_lock_strong(bitMapLock()); 4310 4311 ResourceMark rm; 4312 HandleMark hm; 4313 4314 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4315 4316 if (should_unload_classes()) { 4317 CodeCache::gc_prologue(); 4318 } 4319 assert(haveFreelistLocks(), "must have free list locks"); 4320 assert_lock_strong(bitMapLock()); 4321 4322 // We might assume that we need not fill TLAB's when 4323 // CMSScavengeBeforeRemark is set, because we may have just done 4324 // a scavenge which would have filled all TLAB's -- and besides 4325 // Eden would be empty. This however may not always be the case -- 4326 // for instance although we asked for a scavenge, it may not have 4327 // happened because of a JNI critical section. We probably need 4328 // a policy for deciding whether we can in that case wait until 4329 // the critical section releases and then do the remark following 4330 // the scavenge, and skip it here. In the absence of that policy, 4331 // or of an indication of whether the scavenge did indeed occur, 4332 // we cannot rely on TLAB's having been filled and must do 4333 // so here just in case a scavenge did not happen. 4334 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4335 // Update the saved marks which may affect the root scans. 4336 gch->save_marks(); 4337 4338 if (CMSPrintEdenSurvivorChunks) { 4339 print_eden_and_survivor_chunk_arrays(); 4340 } 4341 4342 { 4343 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4344 4345 // Note on the role of the mod union table: 4346 // Since the marker in "markFromRoots" marks concurrently with 4347 // mutators, it is possible for some reachable objects not to have been 4348 // scanned. For instance, an only reference to an object A was 4349 // placed in object B after the marker scanned B. Unless B is rescanned, 4350 // A would be collected. Such updates to references in marked objects 4351 // are detected via the mod union table which is the set of all cards 4352 // dirtied since the first checkpoint in this GC cycle and prior to 4353 // the most recent young generation GC, minus those cleaned up by the 4354 // concurrent precleaning. 4355 if (CMSParallelRemarkEnabled) { 4356 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4357 do_remark_parallel(); 4358 } else { 4359 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4360 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4361 do_remark_non_parallel(); 4362 } 4363 } 4364 verify_work_stacks_empty(); 4365 verify_overflow_empty(); 4366 4367 { 4368 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4369 refProcessingWork(); 4370 } 4371 verify_work_stacks_empty(); 4372 verify_overflow_empty(); 4373 4374 if (should_unload_classes()) { 4375 CodeCache::gc_epilogue(); 4376 } 4377 JvmtiExport::gc_epilogue(); 4378 4379 // If we encountered any (marking stack / work queue) overflow 4380 // events during the current CMS cycle, take appropriate 4381 // remedial measures, where possible, so as to try and avoid 4382 // recurrence of that condition. 4383 assert(_markStack.isEmpty(), "No grey objects"); 4384 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4385 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4386 if (ser_ovflw > 0) { 4387 if (PrintCMSStatistics != 0) { 4388 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4389 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT 4390 ", kac_preclean=" SIZE_FORMAT ")", 4391 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4392 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4393 } 4394 _markStack.expand(); 4395 _ser_pmc_remark_ovflw = 0; 4396 _ser_pmc_preclean_ovflw = 0; 4397 _ser_kac_preclean_ovflw = 0; 4398 _ser_kac_ovflw = 0; 4399 } 4400 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4401 if (PrintCMSStatistics != 0) { 4402 gclog_or_tty->print_cr("Work queue overflow (benign) " 4403 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4404 _par_pmc_remark_ovflw, _par_kac_ovflw); 4405 } 4406 _par_pmc_remark_ovflw = 0; 4407 _par_kac_ovflw = 0; 4408 } 4409 if (PrintCMSStatistics != 0) { 4410 if (_markStack._hit_limit > 0) { 4411 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4412 _markStack._hit_limit); 4413 } 4414 if (_markStack._failed_double > 0) { 4415 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT ")," 4416 " current capacity " SIZE_FORMAT, 4417 _markStack._failed_double, 4418 _markStack.capacity()); 4419 } 4420 } 4421 _markStack._hit_limit = 0; 4422 _markStack._failed_double = 0; 4423 4424 if ((VerifyAfterGC || VerifyDuringGC) && 4425 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4426 verify_after_remark(); 4427 } 4428 4429 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4430 4431 // Change under the freelistLocks. 4432 _collectorState = Sweeping; 4433 // Call isAllClear() under bitMapLock 4434 assert(_modUnionTable.isAllClear(), 4435 "Should be clear by end of the final marking"); 4436 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4437 "Should be clear by end of the final marking"); 4438 } 4439 4440 void CMSParInitialMarkTask::work(uint worker_id) { 4441 elapsedTimer _timer; 4442 ResourceMark rm; 4443 HandleMark hm; 4444 4445 // ---------- scan from roots -------------- 4446 _timer.start(); 4447 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4448 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4449 4450 // ---------- young gen roots -------------- 4451 { 4452 work_on_young_gen_roots(worker_id, &par_mri_cl); 4453 _timer.stop(); 4454 if (PrintCMSStatistics != 0) { 4455 gclog_or_tty->print_cr( 4456 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4457 worker_id, _timer.seconds()); 4458 } 4459 } 4460 4461 // ---------- remaining roots -------------- 4462 _timer.reset(); 4463 _timer.start(); 4464 4465 CLDToOopClosure cld_closure(&par_mri_cl, true); 4466 4467 gch->gen_process_roots(_strong_roots_scope, 4468 GenCollectedHeap::OldGen, 4469 false, // yg was scanned above 4470 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4471 _collector->should_unload_classes(), 4472 &par_mri_cl, 4473 NULL, 4474 &cld_closure); 4475 assert(_collector->should_unload_classes() 4476 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4477 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4478 _timer.stop(); 4479 if (PrintCMSStatistics != 0) { 4480 gclog_or_tty->print_cr( 4481 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4482 worker_id, _timer.seconds()); 4483 } 4484 } 4485 4486 // Parallel remark task 4487 class CMSParRemarkTask: public CMSParMarkTask { 4488 CompactibleFreeListSpace* _cms_space; 4489 4490 // The per-thread work queues, available here for stealing. 4491 OopTaskQueueSet* _task_queues; 4492 ParallelTaskTerminator _term; 4493 StrongRootsScope* _strong_roots_scope; 4494 4495 public: 4496 // A value of 0 passed to n_workers will cause the number of 4497 // workers to be taken from the active workers in the work gang. 4498 CMSParRemarkTask(CMSCollector* collector, 4499 CompactibleFreeListSpace* cms_space, 4500 uint n_workers, FlexibleWorkGang* workers, 4501 OopTaskQueueSet* task_queues, 4502 StrongRootsScope* strong_roots_scope): 4503 CMSParMarkTask("Rescan roots and grey objects in parallel", 4504 collector, n_workers), 4505 _cms_space(cms_space), 4506 _task_queues(task_queues), 4507 _term(n_workers, task_queues), 4508 _strong_roots_scope(strong_roots_scope) { } 4509 4510 OopTaskQueueSet* task_queues() { return _task_queues; } 4511 4512 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4513 4514 ParallelTaskTerminator* terminator() { return &_term; } 4515 uint n_workers() { return _n_workers; } 4516 4517 void work(uint worker_id); 4518 4519 private: 4520 // ... of dirty cards in old space 4521 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4522 Par_MarkRefsIntoAndScanClosure* cl); 4523 4524 // ... work stealing for the above 4525 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4526 }; 4527 4528 class RemarkKlassClosure : public KlassClosure { 4529 KlassToOopClosure _cm_klass_closure; 4530 public: 4531 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4532 void do_klass(Klass* k) { 4533 // Check if we have modified any oops in the Klass during the concurrent marking. 4534 if (k->has_accumulated_modified_oops()) { 4535 k->clear_accumulated_modified_oops(); 4536 4537 // We could have transfered the current modified marks to the accumulated marks, 4538 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4539 } else if (k->has_modified_oops()) { 4540 // Don't clear anything, this info is needed by the next young collection. 4541 } else { 4542 // No modified oops in the Klass. 4543 return; 4544 } 4545 4546 // The klass has modified fields, need to scan the klass. 4547 _cm_klass_closure.do_klass(k); 4548 } 4549 }; 4550 4551 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4552 ParNewGeneration* young_gen = _collector->_young_gen; 4553 ContiguousSpace* eden_space = young_gen->eden(); 4554 ContiguousSpace* from_space = young_gen->from(); 4555 ContiguousSpace* to_space = young_gen->to(); 4556 4557 HeapWord** eca = _collector->_eden_chunk_array; 4558 size_t ect = _collector->_eden_chunk_index; 4559 HeapWord** sca = _collector->_survivor_chunk_array; 4560 size_t sct = _collector->_survivor_chunk_index; 4561 4562 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4563 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4564 4565 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4566 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4567 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4568 } 4569 4570 // work_queue(i) is passed to the closure 4571 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4572 // also is passed to do_dirty_card_rescan_tasks() and to 4573 // do_work_steal() to select the i-th task_queue. 4574 4575 void CMSParRemarkTask::work(uint worker_id) { 4576 elapsedTimer _timer; 4577 ResourceMark rm; 4578 HandleMark hm; 4579 4580 // ---------- rescan from roots -------------- 4581 _timer.start(); 4582 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4583 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4584 _collector->_span, _collector->ref_processor(), 4585 &(_collector->_markBitMap), 4586 work_queue(worker_id)); 4587 4588 // Rescan young gen roots first since these are likely 4589 // coarsely partitioned and may, on that account, constitute 4590 // the critical path; thus, it's best to start off that 4591 // work first. 4592 // ---------- young gen roots -------------- 4593 { 4594 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4595 _timer.stop(); 4596 if (PrintCMSStatistics != 0) { 4597 gclog_or_tty->print_cr( 4598 "Finished young gen rescan work in %dth thread: %3.3f sec", 4599 worker_id, _timer.seconds()); 4600 } 4601 } 4602 4603 // ---------- remaining roots -------------- 4604 _timer.reset(); 4605 _timer.start(); 4606 gch->gen_process_roots(_strong_roots_scope, 4607 GenCollectedHeap::OldGen, 4608 false, // yg was scanned above 4609 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4610 _collector->should_unload_classes(), 4611 &par_mrias_cl, 4612 NULL, 4613 NULL); // The dirty klasses will be handled below 4614 4615 assert(_collector->should_unload_classes() 4616 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4617 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4618 _timer.stop(); 4619 if (PrintCMSStatistics != 0) { 4620 gclog_or_tty->print_cr( 4621 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4622 worker_id, _timer.seconds()); 4623 } 4624 4625 // ---------- unhandled CLD scanning ---------- 4626 if (worker_id == 0) { // Single threaded at the moment. 4627 _timer.reset(); 4628 _timer.start(); 4629 4630 // Scan all new class loader data objects and new dependencies that were 4631 // introduced during concurrent marking. 4632 ResourceMark rm; 4633 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4634 for (int i = 0; i < array->length(); i++) { 4635 par_mrias_cl.do_class_loader_data(array->at(i)); 4636 } 4637 4638 // We don't need to keep track of new CLDs anymore. 4639 ClassLoaderDataGraph::remember_new_clds(false); 4640 4641 _timer.stop(); 4642 if (PrintCMSStatistics != 0) { 4643 gclog_or_tty->print_cr( 4644 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4645 worker_id, _timer.seconds()); 4646 } 4647 } 4648 4649 // ---------- dirty klass scanning ---------- 4650 if (worker_id == 0) { // Single threaded at the moment. 4651 _timer.reset(); 4652 _timer.start(); 4653 4654 // Scan all classes that was dirtied during the concurrent marking phase. 4655 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4656 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4657 4658 _timer.stop(); 4659 if (PrintCMSStatistics != 0) { 4660 gclog_or_tty->print_cr( 4661 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4662 worker_id, _timer.seconds()); 4663 } 4664 } 4665 4666 // We might have added oops to ClassLoaderData::_handles during the 4667 // concurrent marking phase. These oops point to newly allocated objects 4668 // that are guaranteed to be kept alive. Either by the direct allocation 4669 // code, or when the young collector processes the roots. Hence, 4670 // we don't have to revisit the _handles block during the remark phase. 4671 4672 // ---------- rescan dirty cards ------------ 4673 _timer.reset(); 4674 _timer.start(); 4675 4676 // Do the rescan tasks for each of the two spaces 4677 // (cms_space) in turn. 4678 // "worker_id" is passed to select the task_queue for "worker_id" 4679 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4680 _timer.stop(); 4681 if (PrintCMSStatistics != 0) { 4682 gclog_or_tty->print_cr( 4683 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4684 worker_id, _timer.seconds()); 4685 } 4686 4687 // ---------- steal work from other threads ... 4688 // ---------- ... and drain overflow list. 4689 _timer.reset(); 4690 _timer.start(); 4691 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4692 _timer.stop(); 4693 if (PrintCMSStatistics != 0) { 4694 gclog_or_tty->print_cr( 4695 "Finished work stealing in %dth thread: %3.3f sec", 4696 worker_id, _timer.seconds()); 4697 } 4698 } 4699 4700 // Note that parameter "i" is not used. 4701 void 4702 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4703 OopsInGenClosure* cl, ContiguousSpace* space, 4704 HeapWord** chunk_array, size_t chunk_top) { 4705 // Until all tasks completed: 4706 // . claim an unclaimed task 4707 // . compute region boundaries corresponding to task claimed 4708 // using chunk_array 4709 // . par_oop_iterate(cl) over that region 4710 4711 ResourceMark rm; 4712 HandleMark hm; 4713 4714 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4715 4716 uint nth_task = 0; 4717 uint n_tasks = pst->n_tasks(); 4718 4719 if (n_tasks > 0) { 4720 assert(pst->valid(), "Uninitialized use?"); 4721 HeapWord *start, *end; 4722 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4723 // We claimed task # nth_task; compute its boundaries. 4724 if (chunk_top == 0) { // no samples were taken 4725 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4726 start = space->bottom(); 4727 end = space->top(); 4728 } else if (nth_task == 0) { 4729 start = space->bottom(); 4730 end = chunk_array[nth_task]; 4731 } else if (nth_task < (uint)chunk_top) { 4732 assert(nth_task >= 1, "Control point invariant"); 4733 start = chunk_array[nth_task - 1]; 4734 end = chunk_array[nth_task]; 4735 } else { 4736 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4737 start = chunk_array[chunk_top - 1]; 4738 end = space->top(); 4739 } 4740 MemRegion mr(start, end); 4741 // Verify that mr is in space 4742 assert(mr.is_empty() || space->used_region().contains(mr), 4743 "Should be in space"); 4744 // Verify that "start" is an object boundary 4745 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4746 "Should be an oop"); 4747 space->par_oop_iterate(mr, cl); 4748 } 4749 pst->all_tasks_completed(); 4750 } 4751 } 4752 4753 void 4754 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4755 CompactibleFreeListSpace* sp, int i, 4756 Par_MarkRefsIntoAndScanClosure* cl) { 4757 // Until all tasks completed: 4758 // . claim an unclaimed task 4759 // . compute region boundaries corresponding to task claimed 4760 // . transfer dirty bits ct->mut for that region 4761 // . apply rescanclosure to dirty mut bits for that region 4762 4763 ResourceMark rm; 4764 HandleMark hm; 4765 4766 OopTaskQueue* work_q = work_queue(i); 4767 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4768 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4769 // CAUTION: This closure has state that persists across calls to 4770 // the work method dirty_range_iterate_clear() in that it has 4771 // embedded in it a (subtype of) UpwardsObjectClosure. The 4772 // use of that state in the embedded UpwardsObjectClosure instance 4773 // assumes that the cards are always iterated (even if in parallel 4774 // by several threads) in monotonically increasing order per each 4775 // thread. This is true of the implementation below which picks 4776 // card ranges (chunks) in monotonically increasing order globally 4777 // and, a-fortiori, in monotonically increasing order per thread 4778 // (the latter order being a subsequence of the former). 4779 // If the work code below is ever reorganized into a more chaotic 4780 // work-partitioning form than the current "sequential tasks" 4781 // paradigm, the use of that persistent state will have to be 4782 // revisited and modified appropriately. See also related 4783 // bug 4756801 work on which should examine this code to make 4784 // sure that the changes there do not run counter to the 4785 // assumptions made here and necessary for correctness and 4786 // efficiency. Note also that this code might yield inefficient 4787 // behavior in the case of very large objects that span one or 4788 // more work chunks. Such objects would potentially be scanned 4789 // several times redundantly. Work on 4756801 should try and 4790 // address that performance anomaly if at all possible. XXX 4791 MemRegion full_span = _collector->_span; 4792 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4793 MarkFromDirtyCardsClosure 4794 greyRescanClosure(_collector, full_span, // entire span of interest 4795 sp, bm, work_q, cl); 4796 4797 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4798 assert(pst->valid(), "Uninitialized use?"); 4799 uint nth_task = 0; 4800 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4801 MemRegion span = sp->used_region(); 4802 HeapWord* start_addr = span.start(); 4803 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4804 alignment); 4805 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4806 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4807 start_addr, "Check alignment"); 4808 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4809 chunk_size, "Check alignment"); 4810 4811 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4812 // Having claimed the nth_task, compute corresponding mem-region, 4813 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4814 // The alignment restriction ensures that we do not need any 4815 // synchronization with other gang-workers while setting or 4816 // clearing bits in thus chunk of the MUT. 4817 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4818 start_addr + (nth_task+1)*chunk_size); 4819 // The last chunk's end might be way beyond end of the 4820 // used region. In that case pull back appropriately. 4821 if (this_span.end() > end_addr) { 4822 this_span.set_end(end_addr); 4823 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4824 } 4825 // Iterate over the dirty cards covering this chunk, marking them 4826 // precleaned, and setting the corresponding bits in the mod union 4827 // table. Since we have been careful to partition at Card and MUT-word 4828 // boundaries no synchronization is needed between parallel threads. 4829 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4830 &modUnionClosure); 4831 4832 // Having transferred these marks into the modUnionTable, 4833 // rescan the marked objects on the dirty cards in the modUnionTable. 4834 // Even if this is at a synchronous collection, the initial marking 4835 // may have been done during an asynchronous collection so there 4836 // may be dirty bits in the mod-union table. 4837 _collector->_modUnionTable.dirty_range_iterate_clear( 4838 this_span, &greyRescanClosure); 4839 _collector->_modUnionTable.verifyNoOneBitsInRange( 4840 this_span.start(), 4841 this_span.end()); 4842 } 4843 pst->all_tasks_completed(); // declare that i am done 4844 } 4845 4846 // . see if we can share work_queues with ParNew? XXX 4847 void 4848 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4849 int* seed) { 4850 OopTaskQueue* work_q = work_queue(i); 4851 NOT_PRODUCT(int num_steals = 0;) 4852 oop obj_to_scan; 4853 CMSBitMap* bm = &(_collector->_markBitMap); 4854 4855 while (true) { 4856 // Completely finish any left over work from (an) earlier round(s) 4857 cl->trim_queue(0); 4858 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4859 (size_t)ParGCDesiredObjsFromOverflowList); 4860 // Now check if there's any work in the overflow list 4861 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4862 // only affects the number of attempts made to get work from the 4863 // overflow list and does not affect the number of workers. Just 4864 // pass ParallelGCThreads so this behavior is unchanged. 4865 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4866 work_q, 4867 ParallelGCThreads)) { 4868 // found something in global overflow list; 4869 // not yet ready to go stealing work from others. 4870 // We'd like to assert(work_q->size() != 0, ...) 4871 // because we just took work from the overflow list, 4872 // but of course we can't since all of that could have 4873 // been already stolen from us. 4874 // "He giveth and He taketh away." 4875 continue; 4876 } 4877 // Verify that we have no work before we resort to stealing 4878 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4879 // Try to steal from other queues that have work 4880 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4881 NOT_PRODUCT(num_steals++;) 4882 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4883 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4884 // Do scanning work 4885 obj_to_scan->oop_iterate(cl); 4886 // Loop around, finish this work, and try to steal some more 4887 } else if (terminator()->offer_termination()) { 4888 break; // nirvana from the infinite cycle 4889 } 4890 } 4891 NOT_PRODUCT( 4892 if (PrintCMSStatistics != 0) { 4893 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4894 } 4895 ) 4896 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4897 "Else our work is not yet done"); 4898 } 4899 4900 // Record object boundaries in _eden_chunk_array by sampling the eden 4901 // top in the slow-path eden object allocation code path and record 4902 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4903 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4904 // sampling in sample_eden() that activates during the part of the 4905 // preclean phase. 4906 void CMSCollector::sample_eden_chunk() { 4907 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4908 if (_eden_chunk_lock->try_lock()) { 4909 // Record a sample. This is the critical section. The contents 4910 // of the _eden_chunk_array have to be non-decreasing in the 4911 // address order. 4912 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4913 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4914 "Unexpected state of Eden"); 4915 if (_eden_chunk_index == 0 || 4916 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4917 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4918 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4919 _eden_chunk_index++; // commit sample 4920 } 4921 _eden_chunk_lock->unlock(); 4922 } 4923 } 4924 } 4925 4926 // Return a thread-local PLAB recording array, as appropriate. 4927 void* CMSCollector::get_data_recorder(int thr_num) { 4928 if (_survivor_plab_array != NULL && 4929 (CMSPLABRecordAlways || 4930 (_collectorState > Marking && _collectorState < FinalMarking))) { 4931 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4932 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4933 ca->reset(); // clear it so that fresh data is recorded 4934 return (void*) ca; 4935 } else { 4936 return NULL; 4937 } 4938 } 4939 4940 // Reset all the thread-local PLAB recording arrays 4941 void CMSCollector::reset_survivor_plab_arrays() { 4942 for (uint i = 0; i < ParallelGCThreads; i++) { 4943 _survivor_plab_array[i].reset(); 4944 } 4945 } 4946 4947 // Merge the per-thread plab arrays into the global survivor chunk 4948 // array which will provide the partitioning of the survivor space 4949 // for CMS initial scan and rescan. 4950 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4951 int no_of_gc_threads) { 4952 assert(_survivor_plab_array != NULL, "Error"); 4953 assert(_survivor_chunk_array != NULL, "Error"); 4954 assert(_collectorState == FinalMarking || 4955 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4956 for (int j = 0; j < no_of_gc_threads; j++) { 4957 _cursor[j] = 0; 4958 } 4959 HeapWord* top = surv->top(); 4960 size_t i; 4961 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4962 HeapWord* min_val = top; // Higher than any PLAB address 4963 uint min_tid = 0; // position of min_val this round 4964 for (int j = 0; j < no_of_gc_threads; j++) { 4965 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4966 if (_cursor[j] == cur_sca->end()) { 4967 continue; 4968 } 4969 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4970 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4971 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4972 if (cur_val < min_val) { 4973 min_tid = j; 4974 min_val = cur_val; 4975 } else { 4976 assert(cur_val < top, "All recorded addresses should be less"); 4977 } 4978 } 4979 // At this point min_val and min_tid are respectively 4980 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4981 // and the thread (j) that witnesses that address. 4982 // We record this address in the _survivor_chunk_array[i] 4983 // and increment _cursor[min_tid] prior to the next round i. 4984 if (min_val == top) { 4985 break; 4986 } 4987 _survivor_chunk_array[i] = min_val; 4988 _cursor[min_tid]++; 4989 } 4990 // We are all done; record the size of the _survivor_chunk_array 4991 _survivor_chunk_index = i; // exclusive: [0, i) 4992 if (PrintCMSStatistics > 0) { 4993 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4994 } 4995 // Verify that we used up all the recorded entries 4996 #ifdef ASSERT 4997 size_t total = 0; 4998 for (int j = 0; j < no_of_gc_threads; j++) { 4999 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5000 total += _cursor[j]; 5001 } 5002 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5003 // Check that the merged array is in sorted order 5004 if (total > 0) { 5005 for (size_t i = 0; i < total - 1; i++) { 5006 if (PrintCMSStatistics > 0) { 5007 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5008 i, p2i(_survivor_chunk_array[i])); 5009 } 5010 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5011 "Not sorted"); 5012 } 5013 } 5014 #endif // ASSERT 5015 } 5016 5017 // Set up the space's par_seq_tasks structure for work claiming 5018 // for parallel initial scan and rescan of young gen. 5019 // See ParRescanTask where this is currently used. 5020 void 5021 CMSCollector:: 5022 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5023 assert(n_threads > 0, "Unexpected n_threads argument"); 5024 5025 // Eden space 5026 if (!_young_gen->eden()->is_empty()) { 5027 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5028 assert(!pst->valid(), "Clobbering existing data?"); 5029 // Each valid entry in [0, _eden_chunk_index) represents a task. 5030 size_t n_tasks = _eden_chunk_index + 1; 5031 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5032 // Sets the condition for completion of the subtask (how many threads 5033 // need to finish in order to be done). 5034 pst->set_n_threads(n_threads); 5035 pst->set_n_tasks((int)n_tasks); 5036 } 5037 5038 // Merge the survivor plab arrays into _survivor_chunk_array 5039 if (_survivor_plab_array != NULL) { 5040 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5041 } else { 5042 assert(_survivor_chunk_index == 0, "Error"); 5043 } 5044 5045 // To space 5046 { 5047 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5048 assert(!pst->valid(), "Clobbering existing data?"); 5049 // Sets the condition for completion of the subtask (how many threads 5050 // need to finish in order to be done). 5051 pst->set_n_threads(n_threads); 5052 pst->set_n_tasks(1); 5053 assert(pst->valid(), "Error"); 5054 } 5055 5056 // From space 5057 { 5058 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5059 assert(!pst->valid(), "Clobbering existing data?"); 5060 size_t n_tasks = _survivor_chunk_index + 1; 5061 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5062 // Sets the condition for completion of the subtask (how many threads 5063 // need to finish in order to be done). 5064 pst->set_n_threads(n_threads); 5065 pst->set_n_tasks((int)n_tasks); 5066 assert(pst->valid(), "Error"); 5067 } 5068 } 5069 5070 // Parallel version of remark 5071 void CMSCollector::do_remark_parallel() { 5072 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5073 FlexibleWorkGang* workers = gch->workers(); 5074 assert(workers != NULL, "Need parallel worker threads."); 5075 // Choose to use the number of GC workers most recently set 5076 // into "active_workers". 5077 uint n_workers = workers->active_workers(); 5078 5079 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5080 5081 StrongRootsScope srs(n_workers); 5082 5083 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5084 5085 // We won't be iterating over the cards in the card table updating 5086 // the younger_gen cards, so we shouldn't call the following else 5087 // the verification code as well as subsequent younger_refs_iterate 5088 // code would get confused. XXX 5089 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5090 5091 // The young gen rescan work will not be done as part of 5092 // process_roots (which currently doesn't know how to 5093 // parallelize such a scan), but rather will be broken up into 5094 // a set of parallel tasks (via the sampling that the [abortable] 5095 // preclean phase did of eden, plus the [two] tasks of 5096 // scanning the [two] survivor spaces. Further fine-grain 5097 // parallelization of the scanning of the survivor spaces 5098 // themselves, and of precleaning of the younger gen itself 5099 // is deferred to the future. 5100 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5101 5102 // The dirty card rescan work is broken up into a "sequence" 5103 // of parallel tasks (per constituent space) that are dynamically 5104 // claimed by the parallel threads. 5105 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5106 5107 // It turns out that even when we're using 1 thread, doing the work in a 5108 // separate thread causes wide variance in run times. We can't help this 5109 // in the multi-threaded case, but we special-case n=1 here to get 5110 // repeatable measurements of the 1-thread overhead of the parallel code. 5111 if (n_workers > 1) { 5112 // Make refs discovery MT-safe, if it isn't already: it may not 5113 // necessarily be so, since it's possible that we are doing 5114 // ST marking. 5115 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5116 workers->run_task(&tsk); 5117 } else { 5118 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5119 tsk.work(0); 5120 } 5121 5122 // restore, single-threaded for now, any preserved marks 5123 // as a result of work_q overflow 5124 restore_preserved_marks_if_any(); 5125 } 5126 5127 // Non-parallel version of remark 5128 void CMSCollector::do_remark_non_parallel() { 5129 ResourceMark rm; 5130 HandleMark hm; 5131 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5132 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5133 5134 MarkRefsIntoAndScanClosure 5135 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5136 &_markStack, this, 5137 false /* should_yield */, false /* not precleaning */); 5138 MarkFromDirtyCardsClosure 5139 markFromDirtyCardsClosure(this, _span, 5140 NULL, // space is set further below 5141 &_markBitMap, &_markStack, &mrias_cl); 5142 { 5143 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5144 // Iterate over the dirty cards, setting the corresponding bits in the 5145 // mod union table. 5146 { 5147 ModUnionClosure modUnionClosure(&_modUnionTable); 5148 _ct->ct_bs()->dirty_card_iterate( 5149 _cmsGen->used_region(), 5150 &modUnionClosure); 5151 } 5152 // Having transferred these marks into the modUnionTable, we just need 5153 // to rescan the marked objects on the dirty cards in the modUnionTable. 5154 // The initial marking may have been done during an asynchronous 5155 // collection so there may be dirty bits in the mod-union table. 5156 const int alignment = 5157 CardTableModRefBS::card_size * BitsPerWord; 5158 { 5159 // ... First handle dirty cards in CMS gen 5160 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5161 MemRegion ur = _cmsGen->used_region(); 5162 HeapWord* lb = ur.start(); 5163 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5164 MemRegion cms_span(lb, ub); 5165 _modUnionTable.dirty_range_iterate_clear(cms_span, 5166 &markFromDirtyCardsClosure); 5167 verify_work_stacks_empty(); 5168 if (PrintCMSStatistics != 0) { 5169 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", 5170 markFromDirtyCardsClosure.num_dirty_cards()); 5171 } 5172 } 5173 } 5174 if (VerifyDuringGC && 5175 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5176 HandleMark hm; // Discard invalid handles created during verification 5177 Universe::verify(); 5178 } 5179 { 5180 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5181 5182 verify_work_stacks_empty(); 5183 5184 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5185 StrongRootsScope srs(1); 5186 5187 gch->gen_process_roots(&srs, 5188 GenCollectedHeap::OldGen, 5189 true, // younger gens as roots 5190 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5191 should_unload_classes(), 5192 &mrias_cl, 5193 NULL, 5194 NULL); // The dirty klasses will be handled below 5195 5196 assert(should_unload_classes() 5197 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5198 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5199 } 5200 5201 { 5202 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5203 5204 verify_work_stacks_empty(); 5205 5206 // Scan all class loader data objects that might have been introduced 5207 // during concurrent marking. 5208 ResourceMark rm; 5209 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5210 for (int i = 0; i < array->length(); i++) { 5211 mrias_cl.do_class_loader_data(array->at(i)); 5212 } 5213 5214 // We don't need to keep track of new CLDs anymore. 5215 ClassLoaderDataGraph::remember_new_clds(false); 5216 5217 verify_work_stacks_empty(); 5218 } 5219 5220 { 5221 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5222 5223 verify_work_stacks_empty(); 5224 5225 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5226 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5227 5228 verify_work_stacks_empty(); 5229 } 5230 5231 // We might have added oops to ClassLoaderData::_handles during the 5232 // concurrent marking phase. These oops point to newly allocated objects 5233 // that are guaranteed to be kept alive. Either by the direct allocation 5234 // code, or when the young collector processes the roots. Hence, 5235 // we don't have to revisit the _handles block during the remark phase. 5236 5237 verify_work_stacks_empty(); 5238 // Restore evacuated mark words, if any, used for overflow list links 5239 if (!CMSOverflowEarlyRestoration) { 5240 restore_preserved_marks_if_any(); 5241 } 5242 verify_overflow_empty(); 5243 } 5244 5245 //////////////////////////////////////////////////////// 5246 // Parallel Reference Processing Task Proxy Class 5247 //////////////////////////////////////////////////////// 5248 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5249 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5250 CMSCollector* _collector; 5251 CMSBitMap* _mark_bit_map; 5252 const MemRegion _span; 5253 ProcessTask& _task; 5254 5255 public: 5256 CMSRefProcTaskProxy(ProcessTask& task, 5257 CMSCollector* collector, 5258 const MemRegion& span, 5259 CMSBitMap* mark_bit_map, 5260 AbstractWorkGang* workers, 5261 OopTaskQueueSet* task_queues): 5262 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5263 task_queues, 5264 workers->active_workers()), 5265 _task(task), 5266 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5267 { 5268 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5269 "Inconsistency in _span"); 5270 } 5271 5272 OopTaskQueueSet* task_queues() { return queues(); } 5273 5274 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5275 5276 void do_work_steal(int i, 5277 CMSParDrainMarkingStackClosure* drain, 5278 CMSParKeepAliveClosure* keep_alive, 5279 int* seed); 5280 5281 virtual void work(uint worker_id); 5282 }; 5283 5284 void CMSRefProcTaskProxy::work(uint worker_id) { 5285 ResourceMark rm; 5286 HandleMark hm; 5287 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5288 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5289 _mark_bit_map, 5290 work_queue(worker_id)); 5291 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5292 _mark_bit_map, 5293 work_queue(worker_id)); 5294 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5295 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5296 if (_task.marks_oops_alive()) { 5297 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5298 _collector->hash_seed(worker_id)); 5299 } 5300 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5301 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5302 } 5303 5304 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5305 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5306 EnqueueTask& _task; 5307 5308 public: 5309 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5310 : AbstractGangTask("Enqueue reference objects in parallel"), 5311 _task(task) 5312 { } 5313 5314 virtual void work(uint worker_id) 5315 { 5316 _task.work(worker_id); 5317 } 5318 }; 5319 5320 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5321 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5322 _span(span), 5323 _bit_map(bit_map), 5324 _work_queue(work_queue), 5325 _mark_and_push(collector, span, bit_map, work_queue), 5326 _low_water_mark(MIN2((work_queue->max_elems()/4), 5327 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5328 { } 5329 5330 // . see if we can share work_queues with ParNew? XXX 5331 void CMSRefProcTaskProxy::do_work_steal(int i, 5332 CMSParDrainMarkingStackClosure* drain, 5333 CMSParKeepAliveClosure* keep_alive, 5334 int* seed) { 5335 OopTaskQueue* work_q = work_queue(i); 5336 NOT_PRODUCT(int num_steals = 0;) 5337 oop obj_to_scan; 5338 5339 while (true) { 5340 // Completely finish any left over work from (an) earlier round(s) 5341 drain->trim_queue(0); 5342 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5343 (size_t)ParGCDesiredObjsFromOverflowList); 5344 // Now check if there's any work in the overflow list 5345 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5346 // only affects the number of attempts made to get work from the 5347 // overflow list and does not affect the number of workers. Just 5348 // pass ParallelGCThreads so this behavior is unchanged. 5349 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5350 work_q, 5351 ParallelGCThreads)) { 5352 // Found something in global overflow list; 5353 // not yet ready to go stealing work from others. 5354 // We'd like to assert(work_q->size() != 0, ...) 5355 // because we just took work from the overflow list, 5356 // but of course we can't, since all of that might have 5357 // been already stolen from us. 5358 continue; 5359 } 5360 // Verify that we have no work before we resort to stealing 5361 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5362 // Try to steal from other queues that have work 5363 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5364 NOT_PRODUCT(num_steals++;) 5365 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5366 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5367 // Do scanning work 5368 obj_to_scan->oop_iterate(keep_alive); 5369 // Loop around, finish this work, and try to steal some more 5370 } else if (terminator()->offer_termination()) { 5371 break; // nirvana from the infinite cycle 5372 } 5373 } 5374 NOT_PRODUCT( 5375 if (PrintCMSStatistics != 0) { 5376 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5377 } 5378 ) 5379 } 5380 5381 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5382 { 5383 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5384 FlexibleWorkGang* workers = gch->workers(); 5385 assert(workers != NULL, "Need parallel worker threads."); 5386 CMSRefProcTaskProxy rp_task(task, &_collector, 5387 _collector.ref_processor()->span(), 5388 _collector.markBitMap(), 5389 workers, _collector.task_queues()); 5390 workers->run_task(&rp_task); 5391 } 5392 5393 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5394 { 5395 5396 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5397 FlexibleWorkGang* workers = gch->workers(); 5398 assert(workers != NULL, "Need parallel worker threads."); 5399 CMSRefEnqueueTaskProxy enq_task(task); 5400 workers->run_task(&enq_task); 5401 } 5402 5403 void CMSCollector::refProcessingWork() { 5404 ResourceMark rm; 5405 HandleMark hm; 5406 5407 ReferenceProcessor* rp = ref_processor(); 5408 assert(rp->span().equals(_span), "Spans should be equal"); 5409 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5410 // Process weak references. 5411 rp->setup_policy(false); 5412 verify_work_stacks_empty(); 5413 5414 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5415 &_markStack, false /* !preclean */); 5416 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5417 _span, &_markBitMap, &_markStack, 5418 &cmsKeepAliveClosure, false /* !preclean */); 5419 { 5420 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5421 5422 ReferenceProcessorStats stats; 5423 if (rp->processing_is_mt()) { 5424 // Set the degree of MT here. If the discovery is done MT, there 5425 // may have been a different number of threads doing the discovery 5426 // and a different number of discovered lists may have Ref objects. 5427 // That is OK as long as the Reference lists are balanced (see 5428 // balance_all_queues() and balance_queues()). 5429 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5430 uint active_workers = ParallelGCThreads; 5431 FlexibleWorkGang* workers = gch->workers(); 5432 if (workers != NULL) { 5433 active_workers = workers->active_workers(); 5434 // The expectation is that active_workers will have already 5435 // been set to a reasonable value. If it has not been set, 5436 // investigate. 5437 assert(active_workers > 0, "Should have been set during scavenge"); 5438 } 5439 rp->set_active_mt_degree(active_workers); 5440 CMSRefProcTaskExecutor task_executor(*this); 5441 stats = rp->process_discovered_references(&_is_alive_closure, 5442 &cmsKeepAliveClosure, 5443 &cmsDrainMarkingStackClosure, 5444 &task_executor, 5445 _gc_timer_cm, 5446 _gc_tracer_cm->gc_id()); 5447 } else { 5448 stats = rp->process_discovered_references(&_is_alive_closure, 5449 &cmsKeepAliveClosure, 5450 &cmsDrainMarkingStackClosure, 5451 NULL, 5452 _gc_timer_cm, 5453 _gc_tracer_cm->gc_id()); 5454 } 5455 _gc_tracer_cm->report_gc_reference_stats(stats); 5456 5457 } 5458 5459 // This is the point where the entire marking should have completed. 5460 verify_work_stacks_empty(); 5461 5462 if (should_unload_classes()) { 5463 { 5464 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5465 5466 // Unload classes and purge the SystemDictionary. 5467 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5468 5469 // Unload nmethods. 5470 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5471 5472 // Prune dead klasses from subklass/sibling/implementor lists. 5473 Klass::clean_weak_klass_links(&_is_alive_closure); 5474 } 5475 5476 { 5477 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5478 // Clean up unreferenced symbols in symbol table. 5479 SymbolTable::unlink(); 5480 } 5481 5482 { 5483 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5484 // Delete entries for dead interned strings. 5485 StringTable::unlink(&_is_alive_closure); 5486 } 5487 } 5488 5489 5490 // Restore any preserved marks as a result of mark stack or 5491 // work queue overflow 5492 restore_preserved_marks_if_any(); // done single-threaded for now 5493 5494 rp->set_enqueuing_is_done(true); 5495 if (rp->processing_is_mt()) { 5496 rp->balance_all_queues(); 5497 CMSRefProcTaskExecutor task_executor(*this); 5498 rp->enqueue_discovered_references(&task_executor); 5499 } else { 5500 rp->enqueue_discovered_references(NULL); 5501 } 5502 rp->verify_no_references_recorded(); 5503 assert(!rp->discovery_enabled(), "should have been disabled"); 5504 } 5505 5506 #ifndef PRODUCT 5507 void CMSCollector::check_correct_thread_executing() { 5508 Thread* t = Thread::current(); 5509 // Only the VM thread or the CMS thread should be here. 5510 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5511 "Unexpected thread type"); 5512 // If this is the vm thread, the foreground process 5513 // should not be waiting. Note that _foregroundGCIsActive is 5514 // true while the foreground collector is waiting. 5515 if (_foregroundGCShouldWait) { 5516 // We cannot be the VM thread 5517 assert(t->is_ConcurrentGC_thread(), 5518 "Should be CMS thread"); 5519 } else { 5520 // We can be the CMS thread only if we are in a stop-world 5521 // phase of CMS collection. 5522 if (t->is_ConcurrentGC_thread()) { 5523 assert(_collectorState == InitialMarking || 5524 _collectorState == FinalMarking, 5525 "Should be a stop-world phase"); 5526 // The CMS thread should be holding the CMS_token. 5527 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5528 "Potential interference with concurrently " 5529 "executing VM thread"); 5530 } 5531 } 5532 } 5533 #endif 5534 5535 void CMSCollector::sweep() { 5536 assert(_collectorState == Sweeping, "just checking"); 5537 check_correct_thread_executing(); 5538 verify_work_stacks_empty(); 5539 verify_overflow_empty(); 5540 increment_sweep_count(); 5541 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5542 5543 _inter_sweep_timer.stop(); 5544 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5545 5546 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5547 _intra_sweep_timer.reset(); 5548 _intra_sweep_timer.start(); 5549 { 5550 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5551 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5552 // First sweep the old gen 5553 { 5554 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5555 bitMapLock()); 5556 sweepWork(_cmsGen); 5557 } 5558 5559 // Update Universe::_heap_*_at_gc figures. 5560 // We need all the free list locks to make the abstract state 5561 // transition from Sweeping to Resetting. See detailed note 5562 // further below. 5563 { 5564 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5565 // Update heap occupancy information which is used as 5566 // input to soft ref clearing policy at the next gc. 5567 Universe::update_heap_info_at_gc(); 5568 _collectorState = Resizing; 5569 } 5570 } 5571 verify_work_stacks_empty(); 5572 verify_overflow_empty(); 5573 5574 if (should_unload_classes()) { 5575 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5576 // requires that the virtual spaces are stable and not deleted. 5577 ClassLoaderDataGraph::set_should_purge(true); 5578 } 5579 5580 _intra_sweep_timer.stop(); 5581 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5582 5583 _inter_sweep_timer.reset(); 5584 _inter_sweep_timer.start(); 5585 5586 // We need to use a monotonically non-decreasing time in ms 5587 // or we will see time-warp warnings and os::javaTimeMillis() 5588 // does not guarantee monotonicity. 5589 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5590 update_time_of_last_gc(now); 5591 5592 // NOTE on abstract state transitions: 5593 // Mutators allocate-live and/or mark the mod-union table dirty 5594 // based on the state of the collection. The former is done in 5595 // the interval [Marking, Sweeping] and the latter in the interval 5596 // [Marking, Sweeping). Thus the transitions into the Marking state 5597 // and out of the Sweeping state must be synchronously visible 5598 // globally to the mutators. 5599 // The transition into the Marking state happens with the world 5600 // stopped so the mutators will globally see it. Sweeping is 5601 // done asynchronously by the background collector so the transition 5602 // from the Sweeping state to the Resizing state must be done 5603 // under the freelistLock (as is the check for whether to 5604 // allocate-live and whether to dirty the mod-union table). 5605 assert(_collectorState == Resizing, "Change of collector state to" 5606 " Resizing must be done under the freelistLocks (plural)"); 5607 5608 // Now that sweeping has been completed, we clear 5609 // the incremental_collection_failed flag, 5610 // thus inviting a younger gen collection to promote into 5611 // this generation. If such a promotion may still fail, 5612 // the flag will be set again when a young collection is 5613 // attempted. 5614 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5615 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5616 gch->update_full_collections_completed(_collection_count_start); 5617 } 5618 5619 // FIX ME!!! Looks like this belongs in CFLSpace, with 5620 // CMSGen merely delegating to it. 5621 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5622 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5623 HeapWord* minAddr = _cmsSpace->bottom(); 5624 HeapWord* largestAddr = 5625 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5626 if (largestAddr == NULL) { 5627 // The dictionary appears to be empty. In this case 5628 // try to coalesce at the end of the heap. 5629 largestAddr = _cmsSpace->end(); 5630 } 5631 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5632 size_t nearLargestOffset = 5633 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5634 if (PrintFLSStatistics != 0) { 5635 gclog_or_tty->print_cr( 5636 "CMS: Large Block: " PTR_FORMAT ";" 5637 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5638 p2i(largestAddr), 5639 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5640 } 5641 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5642 } 5643 5644 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5645 return addr >= _cmsSpace->nearLargestChunk(); 5646 } 5647 5648 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5649 return _cmsSpace->find_chunk_at_end(); 5650 } 5651 5652 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5653 bool full) { 5654 // If the young generation has been collected, gather any statistics 5655 // that are of interest at this point. 5656 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5657 if (!full && current_is_young) { 5658 // Gather statistics on the young generation collection. 5659 collector()->stats().record_gc0_end(used()); 5660 } 5661 } 5662 5663 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5664 // We iterate over the space(s) underlying this generation, 5665 // checking the mark bit map to see if the bits corresponding 5666 // to specific blocks are marked or not. Blocks that are 5667 // marked are live and are not swept up. All remaining blocks 5668 // are swept up, with coalescing on-the-fly as we sweep up 5669 // contiguous free and/or garbage blocks: 5670 // We need to ensure that the sweeper synchronizes with allocators 5671 // and stop-the-world collectors. In particular, the following 5672 // locks are used: 5673 // . CMS token: if this is held, a stop the world collection cannot occur 5674 // . freelistLock: if this is held no allocation can occur from this 5675 // generation by another thread 5676 // . bitMapLock: if this is held, no other thread can access or update 5677 // 5678 5679 // Note that we need to hold the freelistLock if we use 5680 // block iterate below; else the iterator might go awry if 5681 // a mutator (or promotion) causes block contents to change 5682 // (for instance if the allocator divvies up a block). 5683 // If we hold the free list lock, for all practical purposes 5684 // young generation GC's can't occur (they'll usually need to 5685 // promote), so we might as well prevent all young generation 5686 // GC's while we do a sweeping step. For the same reason, we might 5687 // as well take the bit map lock for the entire duration 5688 5689 // check that we hold the requisite locks 5690 assert(have_cms_token(), "Should hold cms token"); 5691 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5692 assert_lock_strong(gen->freelistLock()); 5693 assert_lock_strong(bitMapLock()); 5694 5695 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5696 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5697 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5698 _inter_sweep_estimate.padded_average(), 5699 _intra_sweep_estimate.padded_average()); 5700 gen->setNearLargestChunk(); 5701 5702 { 5703 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5704 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5705 // We need to free-up/coalesce garbage/blocks from a 5706 // co-terminal free run. This is done in the SweepClosure 5707 // destructor; so, do not remove this scope, else the 5708 // end-of-sweep-census below will be off by a little bit. 5709 } 5710 gen->cmsSpace()->sweep_completed(); 5711 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5712 if (should_unload_classes()) { // unloaded classes this cycle, 5713 _concurrent_cycles_since_last_unload = 0; // ... reset count 5714 } else { // did not unload classes, 5715 _concurrent_cycles_since_last_unload++; // ... increment count 5716 } 5717 } 5718 5719 // Reset CMS data structures (for now just the marking bit map) 5720 // preparatory for the next cycle. 5721 void CMSCollector::reset(bool concurrent) { 5722 if (concurrent) { 5723 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5724 5725 // If the state is not "Resetting", the foreground thread 5726 // has done a collection and the resetting. 5727 if (_collectorState != Resetting) { 5728 assert(_collectorState == Idling, "The state should only change" 5729 " because the foreground collector has finished the collection"); 5730 return; 5731 } 5732 5733 // Clear the mark bitmap (no grey objects to start with) 5734 // for the next cycle. 5735 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5736 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5737 5738 HeapWord* curAddr = _markBitMap.startWord(); 5739 while (curAddr < _markBitMap.endWord()) { 5740 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5741 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5742 _markBitMap.clear_large_range(chunk); 5743 if (ConcurrentMarkSweepThread::should_yield() && 5744 !foregroundGCIsActive() && 5745 CMSYield) { 5746 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5747 "CMS thread should hold CMS token"); 5748 assert_lock_strong(bitMapLock()); 5749 bitMapLock()->unlock(); 5750 ConcurrentMarkSweepThread::desynchronize(true); 5751 stopTimer(); 5752 if (PrintCMSStatistics != 0) { 5753 incrementYields(); 5754 } 5755 5756 // See the comment in coordinator_yield() 5757 for (unsigned i = 0; i < CMSYieldSleepCount && 5758 ConcurrentMarkSweepThread::should_yield() && 5759 !CMSCollector::foregroundGCIsActive(); ++i) { 5760 os::sleep(Thread::current(), 1, false); 5761 } 5762 5763 ConcurrentMarkSweepThread::synchronize(true); 5764 bitMapLock()->lock_without_safepoint_check(); 5765 startTimer(); 5766 } 5767 curAddr = chunk.end(); 5768 } 5769 // A successful mostly concurrent collection has been done. 5770 // Because only the full (i.e., concurrent mode failure) collections 5771 // are being measured for gc overhead limits, clean the "near" flag 5772 // and count. 5773 size_policy()->reset_gc_overhead_limit_count(); 5774 _collectorState = Idling; 5775 } else { 5776 // already have the lock 5777 assert(_collectorState == Resetting, "just checking"); 5778 assert_lock_strong(bitMapLock()); 5779 _markBitMap.clear_all(); 5780 _collectorState = Idling; 5781 } 5782 5783 register_gc_end(); 5784 } 5785 5786 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5787 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5788 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5789 TraceCollectorStats tcs(counters()); 5790 5791 switch (op) { 5792 case CMS_op_checkpointRootsInitial: { 5793 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5794 checkpointRootsInitial(); 5795 if (PrintGC) { 5796 _cmsGen->printOccupancy("initial-mark"); 5797 } 5798 break; 5799 } 5800 case CMS_op_checkpointRootsFinal: { 5801 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5802 checkpointRootsFinal(); 5803 if (PrintGC) { 5804 _cmsGen->printOccupancy("remark"); 5805 } 5806 break; 5807 } 5808 default: 5809 fatal("No such CMS_op"); 5810 } 5811 } 5812 5813 #ifndef PRODUCT 5814 size_t const CMSCollector::skip_header_HeapWords() { 5815 return FreeChunk::header_size(); 5816 } 5817 5818 // Try and collect here conditions that should hold when 5819 // CMS thread is exiting. The idea is that the foreground GC 5820 // thread should not be blocked if it wants to terminate 5821 // the CMS thread and yet continue to run the VM for a while 5822 // after that. 5823 void CMSCollector::verify_ok_to_terminate() const { 5824 assert(Thread::current()->is_ConcurrentGC_thread(), 5825 "should be called by CMS thread"); 5826 assert(!_foregroundGCShouldWait, "should be false"); 5827 // We could check here that all the various low-level locks 5828 // are not held by the CMS thread, but that is overkill; see 5829 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5830 // is checked. 5831 } 5832 #endif 5833 5834 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5835 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5836 "missing Printezis mark?"); 5837 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5838 size_t size = pointer_delta(nextOneAddr + 1, addr); 5839 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5840 "alignment problem"); 5841 assert(size >= 3, "Necessary for Printezis marks to work"); 5842 return size; 5843 } 5844 5845 // A variant of the above (block_size_using_printezis_bits()) except 5846 // that we return 0 if the P-bits are not yet set. 5847 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5848 if (_markBitMap.isMarked(addr + 1)) { 5849 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5850 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5851 size_t size = pointer_delta(nextOneAddr + 1, addr); 5852 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5853 "alignment problem"); 5854 assert(size >= 3, "Necessary for Printezis marks to work"); 5855 return size; 5856 } 5857 return 0; 5858 } 5859 5860 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5861 size_t sz = 0; 5862 oop p = (oop)addr; 5863 if (p->klass_or_null() != NULL) { 5864 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5865 } else { 5866 sz = block_size_using_printezis_bits(addr); 5867 } 5868 assert(sz > 0, "size must be nonzero"); 5869 HeapWord* next_block = addr + sz; 5870 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5871 CardTableModRefBS::card_size); 5872 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5873 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5874 "must be different cards"); 5875 return next_card; 5876 } 5877 5878 5879 // CMS Bit Map Wrapper ///////////////////////////////////////// 5880 5881 // Construct a CMS bit map infrastructure, but don't create the 5882 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5883 // further below. 5884 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name, 5885 bool allocate_for_entire_heap): 5886 _bm(), 5887 _shifter(shifter), 5888 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5889 Monitor::_safepoint_check_sometimes) : NULL), 5890 _allocate_for_entire_heap(allocate_for_entire_heap) 5891 { 5892 _bmStartWord = 0; 5893 _bmWordSize = 0; 5894 } 5895 5896 bool CMSBitMap::allocate(MemRegion mr) { 5897 _bmStartWord = mr.start(); 5898 _bmWordSize = mr.word_size(); 5899 size_t reserved_size; 5900 MemRegion heap_mr = GenCollectedHeap::heap()->reserved_region(); 5901 size_t heapWordSize = heap_mr.word_size(); 5902 if (!_allocate_for_entire_heap) { 5903 reserved_size = ReservedSpace::allocation_align_size_up( 5904 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1); 5905 } else { 5906 // Allocate a bit map memory large enough to cover the entire heap 5907 // (as opposed to the OG and PG only.) 5908 assert(CMSParallelFullGC && ShareCMSMarkBitMapWithParallelFullGC, 5909 "Otherwise this code must not reachable."); 5910 assert(_shifter == 0, "Valid only for the mark bit map, not the mod union table."); 5911 assert(heap_mr.contains(mr), 5912 "The give MR (OG and PG) must be part of the heap"); 5913 reserved_size = ReservedSpace::allocation_align_size_up( 5914 (heapWordSize >> (_shifter + LogBitsPerByte)) + 1); 5915 } 5916 ReservedSpace brs(reserved_size); 5917 if (!brs.is_reserved()) { 5918 warning("CMS bit map allocation failure"); 5919 return false; 5920 } 5921 // For now we'll just commit all of the bit map up front. 5922 // Later on we'll try to be more parsimonious with swap. 5923 if (!_virtual_space.initialize(brs, brs.size())) { 5924 warning("CMS bit map backing store failure"); 5925 return false; 5926 } 5927 assert(_virtual_space.committed_size() == brs.size(), 5928 "didn't reserve backing store for all of CMS bit map?"); 5929 if (!_allocate_for_entire_heap) { 5930 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5931 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5932 _bmWordSize, "inconsistency in bit map sizing"); 5933 } else { 5934 // First, verify that there is no alignment issue between the 5935 // heap, the generation boundaries, and the bit map because each 5936 // word in the bit map covers BitsPerWord heap words. 5937 assert(CMSParallelFullGC && ShareCMSMarkBitMapWithParallelFullGC, 5938 "Otherwise this code must not reachable."); 5939 assert(_shifter == 0, "Valid only for the mark bit map, not the mod union table."); 5940 assert((((intptr_t) heap_mr.start()) % BitsPerWord) == 0, 5941 "Assume the heap base is aligned by BitsPerWord"); 5942 assert((((intptr_t) _bmStartWord) % BitsPerWord) == 0, 5943 "Assume the yg-og boundary is aligned by BitsPerWord"); 5944 assert(heapWordSize % BitsPerByte == 0, 5945 "heap_mr.word_size() should be aligned by BitsPerByte"); 5946 assert(_bmWordSize % BitsPerByte == 0, 5947 "_bmWordSize should be aligned by BitsPerByte"); 5948 // Second, let _bm point to the OG and PG portion of the allocated 5949 // memory. 5950 size_t payload_size = (_bmWordSize >> (_shifter + LogBitsPerByte)); 5951 size_t total_size = (heapWordSize >> (_shifter + LogBitsPerByte)); 5952 size_t non_payload_size = total_size - payload_size; 5953 _bm.set_map((BitMap::bm_word_t*)(_virtual_space.low() + non_payload_size)); 5954 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5955 heapWordSize, "inconsistency in bit map sizing"); 5956 } 5957 _bm.set_size(_bmWordSize >> _shifter); 5958 5959 // bm.clear(); // can we rely on getting zero'd memory? verify below 5960 assert(isAllClear(), 5961 "Expected zero'd memory from ReservedSpace constructor"); 5962 #ifdef ASSERT 5963 if (CMSParallelFullGC && ShareCMSMarkBitMapWithParallelFullGC) { 5964 // The above isAllClear() checks only for the portion visible to 5965 // _bm (OG and PG). Check that the entire backing memory including 5966 // the YG portion is all zero. 5967 char* brs_base = brs.base(); 5968 size_t brs_size = brs.size(); 5969 for (size_t p = 0; p < brs_size; p++) { 5970 char* a = brs_base + p; 5971 assert(*a == 0, "The backing memory of the CMSBitMap must be all zero at init."); 5972 } 5973 } 5974 #endif 5975 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5976 "consistency check"); 5977 return true; 5978 } 5979 5980 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5981 HeapWord *next_addr, *end_addr, *last_addr; 5982 assert_locked(); 5983 assert(covers(mr), "out-of-range error"); 5984 // XXX assert that start and end are appropriately aligned 5985 for (next_addr = mr.start(), end_addr = mr.end(); 5986 next_addr < end_addr; next_addr = last_addr) { 5987 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5988 last_addr = dirty_region.end(); 5989 if (!dirty_region.is_empty()) { 5990 cl->do_MemRegion(dirty_region); 5991 } else { 5992 assert(last_addr == end_addr, "program logic"); 5993 return; 5994 } 5995 } 5996 } 5997 5998 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5999 _bm.print_on_error(st, prefix); 6000 } 6001 6002 #ifndef PRODUCT 6003 void CMSBitMap::assert_locked() const { 6004 CMSLockVerifier::assert_locked(lock()); 6005 } 6006 6007 bool CMSBitMap::covers(MemRegion mr) const { 6008 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6009 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6010 "size inconsistency"); 6011 return (mr.start() >= _bmStartWord) && 6012 (mr.end() <= endWord()); 6013 } 6014 6015 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6016 return (start >= _bmStartWord && (start + size) <= endWord()); 6017 } 6018 6019 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6020 // verify that there are no 1 bits in the interval [left, right) 6021 FalseBitMapClosure falseBitMapClosure; 6022 iterate(&falseBitMapClosure, left, right); 6023 } 6024 6025 void CMSBitMap::region_invariant(MemRegion mr) 6026 { 6027 assert_locked(); 6028 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6029 assert(!mr.is_empty(), "unexpected empty region"); 6030 assert(covers(mr), "mr should be covered by bit map"); 6031 // convert address range into offset range 6032 size_t start_ofs = heapWordToOffset(mr.start()); 6033 // Make sure that end() is appropriately aligned 6034 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6035 (1 << (_shifter+LogHeapWordSize))), 6036 "Misaligned mr.end()"); 6037 size_t end_ofs = heapWordToOffset(mr.end()); 6038 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6039 } 6040 6041 #endif 6042 6043 bool CMSMarkStack::allocate(size_t size) { 6044 // allocate a stack of the requisite depth 6045 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6046 size * sizeof(oop))); 6047 if (!rs.is_reserved()) { 6048 warning("CMSMarkStack allocation failure"); 6049 return false; 6050 } 6051 if (!_virtual_space.initialize(rs, rs.size())) { 6052 warning("CMSMarkStack backing store failure"); 6053 return false; 6054 } 6055 assert(_virtual_space.committed_size() == rs.size(), 6056 "didn't reserve backing store for all of CMS stack?"); 6057 _base = (oop*)(_virtual_space.low()); 6058 _index = 0; 6059 _capacity = size; 6060 NOT_PRODUCT(_max_depth = 0); 6061 return true; 6062 } 6063 6064 // XXX FIX ME !!! In the MT case we come in here holding a 6065 // leaf lock. For printing we need to take a further lock 6066 // which has lower rank. We need to recalibrate the two 6067 // lock-ranks involved in order to be able to print the 6068 // messages below. (Or defer the printing to the caller. 6069 // For now we take the expedient path of just disabling the 6070 // messages for the problematic case.) 6071 void CMSMarkStack::expand() { 6072 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6073 if (_capacity == MarkStackSizeMax) { 6074 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6075 // We print a warning message only once per CMS cycle. 6076 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6077 } 6078 return; 6079 } 6080 // Double capacity if possible 6081 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6082 // Do not give up existing stack until we have managed to 6083 // get the double capacity that we desired. 6084 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6085 new_capacity * sizeof(oop))); 6086 if (rs.is_reserved()) { 6087 // Release the backing store associated with old stack 6088 _virtual_space.release(); 6089 // Reinitialize virtual space for new stack 6090 if (!_virtual_space.initialize(rs, rs.size())) { 6091 fatal("Not enough swap for expanded marking stack"); 6092 } 6093 _base = (oop*)(_virtual_space.low()); 6094 _index = 0; 6095 _capacity = new_capacity; 6096 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6097 // Failed to double capacity, continue; 6098 // we print a detail message only once per CMS cycle. 6099 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " 6100 SIZE_FORMAT "K", 6101 _capacity / K, new_capacity / K); 6102 } 6103 } 6104 6105 6106 // Closures 6107 // XXX: there seems to be a lot of code duplication here; 6108 // should refactor and consolidate common code. 6109 6110 // This closure is used to mark refs into the CMS generation in 6111 // the CMS bit map. Called at the first checkpoint. This closure 6112 // assumes that we do not need to re-mark dirty cards; if the CMS 6113 // generation on which this is used is not an oldest 6114 // generation then this will lose younger_gen cards! 6115 6116 MarkRefsIntoClosure::MarkRefsIntoClosure( 6117 MemRegion span, CMSBitMap* bitMap): 6118 _span(span), 6119 _bitMap(bitMap) 6120 { 6121 assert(_ref_processor == NULL, "deliberately left NULL"); 6122 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6123 } 6124 6125 void MarkRefsIntoClosure::do_oop(oop obj) { 6126 // if p points into _span, then mark corresponding bit in _markBitMap 6127 assert(obj->is_oop(), "expected an oop"); 6128 HeapWord* addr = (HeapWord*)obj; 6129 if (_span.contains(addr)) { 6130 // this should be made more efficient 6131 _bitMap->mark(addr); 6132 } 6133 } 6134 6135 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6136 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6137 6138 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6139 MemRegion span, CMSBitMap* bitMap): 6140 _span(span), 6141 _bitMap(bitMap) 6142 { 6143 assert(_ref_processor == NULL, "deliberately left NULL"); 6144 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6145 } 6146 6147 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6148 // if p points into _span, then mark corresponding bit in _markBitMap 6149 assert(obj->is_oop(), "expected an oop"); 6150 HeapWord* addr = (HeapWord*)obj; 6151 if (_span.contains(addr)) { 6152 // this should be made more efficient 6153 _bitMap->par_mark(addr); 6154 } 6155 } 6156 6157 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6158 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6159 6160 // A variant of the above, used for CMS marking verification. 6161 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6162 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6163 _span(span), 6164 _verification_bm(verification_bm), 6165 _cms_bm(cms_bm) 6166 { 6167 assert(_ref_processor == NULL, "deliberately left NULL"); 6168 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6169 } 6170 6171 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6172 // if p points into _span, then mark corresponding bit in _markBitMap 6173 assert(obj->is_oop(), "expected an oop"); 6174 HeapWord* addr = (HeapWord*)obj; 6175 if (_span.contains(addr)) { 6176 _verification_bm->mark(addr); 6177 if (!_cms_bm->isMarked(addr)) { 6178 oop(addr)->print(); 6179 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6180 fatal("... aborting"); 6181 } 6182 } 6183 } 6184 6185 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6186 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6187 6188 ////////////////////////////////////////////////// 6189 // MarkRefsIntoAndScanClosure 6190 ////////////////////////////////////////////////// 6191 6192 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6193 ReferenceProcessor* rp, 6194 CMSBitMap* bit_map, 6195 CMSBitMap* mod_union_table, 6196 CMSMarkStack* mark_stack, 6197 CMSCollector* collector, 6198 bool should_yield, 6199 bool concurrent_precleaning): 6200 _collector(collector), 6201 _span(span), 6202 _bit_map(bit_map), 6203 _mark_stack(mark_stack), 6204 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6205 mark_stack, concurrent_precleaning), 6206 _yield(should_yield), 6207 _concurrent_precleaning(concurrent_precleaning), 6208 _freelistLock(NULL) 6209 { 6210 _ref_processor = rp; 6211 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6212 } 6213 6214 // This closure is used to mark refs into the CMS generation at the 6215 // second (final) checkpoint, and to scan and transitively follow 6216 // the unmarked oops. It is also used during the concurrent precleaning 6217 // phase while scanning objects on dirty cards in the CMS generation. 6218 // The marks are made in the marking bit map and the marking stack is 6219 // used for keeping the (newly) grey objects during the scan. 6220 // The parallel version (Par_...) appears further below. 6221 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6222 if (obj != NULL) { 6223 assert(obj->is_oop(), "expected an oop"); 6224 HeapWord* addr = (HeapWord*)obj; 6225 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6226 assert(_collector->overflow_list_is_empty(), 6227 "overflow list should be empty"); 6228 if (_span.contains(addr) && 6229 !_bit_map->isMarked(addr)) { 6230 // mark bit map (object is now grey) 6231 _bit_map->mark(addr); 6232 // push on marking stack (stack should be empty), and drain the 6233 // stack by applying this closure to the oops in the oops popped 6234 // from the stack (i.e. blacken the grey objects) 6235 bool res = _mark_stack->push(obj); 6236 assert(res, "Should have space to push on empty stack"); 6237 do { 6238 oop new_oop = _mark_stack->pop(); 6239 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6240 assert(_bit_map->isMarked((HeapWord*)new_oop), 6241 "only grey objects on this stack"); 6242 // iterate over the oops in this oop, marking and pushing 6243 // the ones in CMS heap (i.e. in _span). 6244 new_oop->oop_iterate(&_pushAndMarkClosure); 6245 // check if it's time to yield 6246 do_yield_check(); 6247 } while (!_mark_stack->isEmpty() || 6248 (!_concurrent_precleaning && take_from_overflow_list())); 6249 // if marking stack is empty, and we are not doing this 6250 // during precleaning, then check the overflow list 6251 } 6252 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6253 assert(_collector->overflow_list_is_empty(), 6254 "overflow list was drained above"); 6255 // We could restore evacuated mark words, if any, used for 6256 // overflow list links here because the overflow list is 6257 // provably empty here. That would reduce the maximum 6258 // size requirements for preserved_{oop,mark}_stack. 6259 // But we'll just postpone it until we are all done 6260 // so we can just stream through. 6261 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6262 _collector->restore_preserved_marks_if_any(); 6263 assert(_collector->no_preserved_marks(), "No preserved marks"); 6264 } 6265 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6266 "All preserved marks should have been restored above"); 6267 } 6268 } 6269 6270 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6271 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6272 6273 void MarkRefsIntoAndScanClosure::do_yield_work() { 6274 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6275 "CMS thread should hold CMS token"); 6276 assert_lock_strong(_freelistLock); 6277 assert_lock_strong(_bit_map->lock()); 6278 // relinquish the free_list_lock and bitMaplock() 6279 _bit_map->lock()->unlock(); 6280 _freelistLock->unlock(); 6281 ConcurrentMarkSweepThread::desynchronize(true); 6282 _collector->stopTimer(); 6283 if (PrintCMSStatistics != 0) { 6284 _collector->incrementYields(); 6285 } 6286 6287 // See the comment in coordinator_yield() 6288 for (unsigned i = 0; 6289 i < CMSYieldSleepCount && 6290 ConcurrentMarkSweepThread::should_yield() && 6291 !CMSCollector::foregroundGCIsActive(); 6292 ++i) { 6293 os::sleep(Thread::current(), 1, false); 6294 } 6295 6296 ConcurrentMarkSweepThread::synchronize(true); 6297 _freelistLock->lock_without_safepoint_check(); 6298 _bit_map->lock()->lock_without_safepoint_check(); 6299 _collector->startTimer(); 6300 } 6301 6302 /////////////////////////////////////////////////////////// 6303 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6304 // MarkRefsIntoAndScanClosure 6305 /////////////////////////////////////////////////////////// 6306 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6307 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6308 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6309 _span(span), 6310 _bit_map(bit_map), 6311 _work_queue(work_queue), 6312 _low_water_mark(MIN2((work_queue->max_elems()/4), 6313 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6314 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6315 { 6316 _ref_processor = rp; 6317 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6318 } 6319 6320 // This closure is used to mark refs into the CMS generation at the 6321 // second (final) checkpoint, and to scan and transitively follow 6322 // the unmarked oops. The marks are made in the marking bit map and 6323 // the work_queue is used for keeping the (newly) grey objects during 6324 // the scan phase whence they are also available for stealing by parallel 6325 // threads. Since the marking bit map is shared, updates are 6326 // synchronized (via CAS). 6327 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6328 if (obj != NULL) { 6329 // Ignore mark word because this could be an already marked oop 6330 // that may be chained at the end of the overflow list. 6331 assert(obj->is_oop(true), "expected an oop"); 6332 HeapWord* addr = (HeapWord*)obj; 6333 if (_span.contains(addr) && 6334 !_bit_map->isMarked(addr)) { 6335 // mark bit map (object will become grey): 6336 // It is possible for several threads to be 6337 // trying to "claim" this object concurrently; 6338 // the unique thread that succeeds in marking the 6339 // object first will do the subsequent push on 6340 // to the work queue (or overflow list). 6341 if (_bit_map->par_mark(addr)) { 6342 // push on work_queue (which may not be empty), and trim the 6343 // queue to an appropriate length by applying this closure to 6344 // the oops in the oops popped from the stack (i.e. blacken the 6345 // grey objects) 6346 bool res = _work_queue->push(obj); 6347 assert(res, "Low water mark should be less than capacity?"); 6348 trim_queue(_low_water_mark); 6349 } // Else, another thread claimed the object 6350 } 6351 } 6352 } 6353 6354 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6355 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6356 6357 // This closure is used to rescan the marked objects on the dirty cards 6358 // in the mod union table and the card table proper. 6359 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6360 oop p, MemRegion mr) { 6361 6362 size_t size = 0; 6363 HeapWord* addr = (HeapWord*)p; 6364 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6365 assert(_span.contains(addr), "we are scanning the CMS generation"); 6366 // check if it's time to yield 6367 if (do_yield_check()) { 6368 // We yielded for some foreground stop-world work, 6369 // and we have been asked to abort this ongoing preclean cycle. 6370 return 0; 6371 } 6372 if (_bitMap->isMarked(addr)) { 6373 // it's marked; is it potentially uninitialized? 6374 if (p->klass_or_null() != NULL) { 6375 // an initialized object; ignore mark word in verification below 6376 // since we are running concurrent with mutators 6377 assert(p->is_oop(true), "should be an oop"); 6378 if (p->is_objArray()) { 6379 // objArrays are precisely marked; restrict scanning 6380 // to dirty cards only. 6381 size = CompactibleFreeListSpace::adjustObjectSize( 6382 p->oop_iterate(_scanningClosure, mr)); 6383 } else { 6384 // A non-array may have been imprecisely marked; we need 6385 // to scan object in its entirety. 6386 size = CompactibleFreeListSpace::adjustObjectSize( 6387 p->oop_iterate(_scanningClosure)); 6388 } 6389 #ifdef ASSERT 6390 size_t direct_size = 6391 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6392 assert(size == direct_size, "Inconsistency in size"); 6393 assert(size >= 3, "Necessary for Printezis marks to work"); 6394 if (!_bitMap->isMarked(addr+1)) { 6395 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6396 } else { 6397 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6398 assert(_bitMap->isMarked(addr+size-1), 6399 "inconsistent Printezis mark"); 6400 } 6401 #endif // ASSERT 6402 } else { 6403 // An uninitialized object. 6404 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6405 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6406 size = pointer_delta(nextOneAddr + 1, addr); 6407 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6408 "alignment problem"); 6409 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6410 // will dirty the card when the klass pointer is installed in the 6411 // object (signaling the completion of initialization). 6412 } 6413 } else { 6414 // Either a not yet marked object or an uninitialized object 6415 if (p->klass_or_null() == NULL) { 6416 // An uninitialized object, skip to the next card, since 6417 // we may not be able to read its P-bits yet. 6418 assert(size == 0, "Initial value"); 6419 } else { 6420 // An object not (yet) reached by marking: we merely need to 6421 // compute its size so as to go look at the next block. 6422 assert(p->is_oop(true), "should be an oop"); 6423 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6424 } 6425 } 6426 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6427 return size; 6428 } 6429 6430 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6431 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6432 "CMS thread should hold CMS token"); 6433 assert_lock_strong(_freelistLock); 6434 assert_lock_strong(_bitMap->lock()); 6435 // relinquish the free_list_lock and bitMaplock() 6436 _bitMap->lock()->unlock(); 6437 _freelistLock->unlock(); 6438 ConcurrentMarkSweepThread::desynchronize(true); 6439 _collector->stopTimer(); 6440 if (PrintCMSStatistics != 0) { 6441 _collector->incrementYields(); 6442 } 6443 6444 // See the comment in coordinator_yield() 6445 for (unsigned i = 0; i < CMSYieldSleepCount && 6446 ConcurrentMarkSweepThread::should_yield() && 6447 !CMSCollector::foregroundGCIsActive(); ++i) { 6448 os::sleep(Thread::current(), 1, false); 6449 } 6450 6451 ConcurrentMarkSweepThread::synchronize(true); 6452 _freelistLock->lock_without_safepoint_check(); 6453 _bitMap->lock()->lock_without_safepoint_check(); 6454 _collector->startTimer(); 6455 } 6456 6457 6458 ////////////////////////////////////////////////////////////////// 6459 // SurvivorSpacePrecleanClosure 6460 ////////////////////////////////////////////////////////////////// 6461 // This (single-threaded) closure is used to preclean the oops in 6462 // the survivor spaces. 6463 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6464 6465 HeapWord* addr = (HeapWord*)p; 6466 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6467 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6468 assert(p->klass_or_null() != NULL, "object should be initialized"); 6469 // an initialized object; ignore mark word in verification below 6470 // since we are running concurrent with mutators 6471 assert(p->is_oop(true), "should be an oop"); 6472 // Note that we do not yield while we iterate over 6473 // the interior oops of p, pushing the relevant ones 6474 // on our marking stack. 6475 size_t size = p->oop_iterate(_scanning_closure); 6476 do_yield_check(); 6477 // Observe that below, we do not abandon the preclean 6478 // phase as soon as we should; rather we empty the 6479 // marking stack before returning. This is to satisfy 6480 // some existing assertions. In general, it may be a 6481 // good idea to abort immediately and complete the marking 6482 // from the grey objects at a later time. 6483 while (!_mark_stack->isEmpty()) { 6484 oop new_oop = _mark_stack->pop(); 6485 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6486 assert(_bit_map->isMarked((HeapWord*)new_oop), 6487 "only grey objects on this stack"); 6488 // iterate over the oops in this oop, marking and pushing 6489 // the ones in CMS heap (i.e. in _span). 6490 new_oop->oop_iterate(_scanning_closure); 6491 // check if it's time to yield 6492 do_yield_check(); 6493 } 6494 unsigned int after_count = 6495 GenCollectedHeap::heap()->total_collections(); 6496 bool abort = (_before_count != after_count) || 6497 _collector->should_abort_preclean(); 6498 return abort ? 0 : size; 6499 } 6500 6501 void SurvivorSpacePrecleanClosure::do_yield_work() { 6502 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6503 "CMS thread should hold CMS token"); 6504 assert_lock_strong(_bit_map->lock()); 6505 // Relinquish the bit map lock 6506 _bit_map->lock()->unlock(); 6507 ConcurrentMarkSweepThread::desynchronize(true); 6508 _collector->stopTimer(); 6509 if (PrintCMSStatistics != 0) { 6510 _collector->incrementYields(); 6511 } 6512 6513 // See the comment in coordinator_yield() 6514 for (unsigned i = 0; i < CMSYieldSleepCount && 6515 ConcurrentMarkSweepThread::should_yield() && 6516 !CMSCollector::foregroundGCIsActive(); ++i) { 6517 os::sleep(Thread::current(), 1, false); 6518 } 6519 6520 ConcurrentMarkSweepThread::synchronize(true); 6521 _bit_map->lock()->lock_without_safepoint_check(); 6522 _collector->startTimer(); 6523 } 6524 6525 // This closure is used to rescan the marked objects on the dirty cards 6526 // in the mod union table and the card table proper. In the parallel 6527 // case, although the bitMap is shared, we do a single read so the 6528 // isMarked() query is "safe". 6529 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6530 // Ignore mark word because we are running concurrent with mutators 6531 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6532 HeapWord* addr = (HeapWord*)p; 6533 assert(_span.contains(addr), "we are scanning the CMS generation"); 6534 bool is_obj_array = false; 6535 #ifdef ASSERT 6536 if (!_parallel) { 6537 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6538 assert(_collector->overflow_list_is_empty(), 6539 "overflow list should be empty"); 6540 6541 } 6542 #endif // ASSERT 6543 if (_bit_map->isMarked(addr)) { 6544 // Obj arrays are precisely marked, non-arrays are not; 6545 // so we scan objArrays precisely and non-arrays in their 6546 // entirety. 6547 if (p->is_objArray()) { 6548 is_obj_array = true; 6549 if (_parallel) { 6550 p->oop_iterate(_par_scan_closure, mr); 6551 } else { 6552 p->oop_iterate(_scan_closure, mr); 6553 } 6554 } else { 6555 if (_parallel) { 6556 p->oop_iterate(_par_scan_closure); 6557 } else { 6558 p->oop_iterate(_scan_closure); 6559 } 6560 } 6561 } 6562 #ifdef ASSERT 6563 if (!_parallel) { 6564 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6565 assert(_collector->overflow_list_is_empty(), 6566 "overflow list should be empty"); 6567 6568 } 6569 #endif // ASSERT 6570 return is_obj_array; 6571 } 6572 6573 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6574 MemRegion span, 6575 CMSBitMap* bitMap, CMSMarkStack* markStack, 6576 bool should_yield, bool verifying): 6577 _collector(collector), 6578 _span(span), 6579 _bitMap(bitMap), 6580 _mut(&collector->_modUnionTable), 6581 _markStack(markStack), 6582 _yield(should_yield), 6583 _skipBits(0) 6584 { 6585 assert(_markStack->isEmpty(), "stack should be empty"); 6586 _finger = _bitMap->startWord(); 6587 _threshold = _finger; 6588 assert(_collector->_restart_addr == NULL, "Sanity check"); 6589 assert(_span.contains(_finger), "Out of bounds _finger?"); 6590 DEBUG_ONLY(_verifying = verifying;) 6591 } 6592 6593 void MarkFromRootsClosure::reset(HeapWord* addr) { 6594 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6595 assert(_span.contains(addr), "Out of bounds _finger?"); 6596 _finger = addr; 6597 _threshold = (HeapWord*)round_to( 6598 (intptr_t)_finger, CardTableModRefBS::card_size); 6599 } 6600 6601 // Should revisit to see if this should be restructured for 6602 // greater efficiency. 6603 bool MarkFromRootsClosure::do_bit(size_t offset) { 6604 if (_skipBits > 0) { 6605 _skipBits--; 6606 return true; 6607 } 6608 // convert offset into a HeapWord* 6609 HeapWord* addr = _bitMap->startWord() + offset; 6610 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6611 "address out of range"); 6612 assert(_bitMap->isMarked(addr), "tautology"); 6613 if (_bitMap->isMarked(addr+1)) { 6614 // this is an allocated but not yet initialized object 6615 assert(_skipBits == 0, "tautology"); 6616 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6617 oop p = oop(addr); 6618 if (p->klass_or_null() == NULL) { 6619 DEBUG_ONLY(if (!_verifying) {) 6620 // We re-dirty the cards on which this object lies and increase 6621 // the _threshold so that we'll come back to scan this object 6622 // during the preclean or remark phase. (CMSCleanOnEnter) 6623 if (CMSCleanOnEnter) { 6624 size_t sz = _collector->block_size_using_printezis_bits(addr); 6625 HeapWord* end_card_addr = (HeapWord*)round_to( 6626 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6627 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6628 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6629 // Bump _threshold to end_card_addr; note that 6630 // _threshold cannot possibly exceed end_card_addr, anyhow. 6631 // This prevents future clearing of the card as the scan proceeds 6632 // to the right. 6633 assert(_threshold <= end_card_addr, 6634 "Because we are just scanning into this object"); 6635 if (_threshold < end_card_addr) { 6636 _threshold = end_card_addr; 6637 } 6638 if (p->klass_or_null() != NULL) { 6639 // Redirty the range of cards... 6640 _mut->mark_range(redirty_range); 6641 } // ...else the setting of klass will dirty the card anyway. 6642 } 6643 DEBUG_ONLY(}) 6644 return true; 6645 } 6646 } 6647 scanOopsInOop(addr); 6648 return true; 6649 } 6650 6651 // We take a break if we've been at this for a while, 6652 // so as to avoid monopolizing the locks involved. 6653 void MarkFromRootsClosure::do_yield_work() { 6654 // First give up the locks, then yield, then re-lock 6655 // We should probably use a constructor/destructor idiom to 6656 // do this unlock/lock or modify the MutexUnlocker class to 6657 // serve our purpose. XXX 6658 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6659 "CMS thread should hold CMS token"); 6660 assert_lock_strong(_bitMap->lock()); 6661 _bitMap->lock()->unlock(); 6662 ConcurrentMarkSweepThread::desynchronize(true); 6663 _collector->stopTimer(); 6664 if (PrintCMSStatistics != 0) { 6665 _collector->incrementYields(); 6666 } 6667 6668 // See the comment in coordinator_yield() 6669 for (unsigned i = 0; i < CMSYieldSleepCount && 6670 ConcurrentMarkSweepThread::should_yield() && 6671 !CMSCollector::foregroundGCIsActive(); ++i) { 6672 os::sleep(Thread::current(), 1, false); 6673 } 6674 6675 ConcurrentMarkSweepThread::synchronize(true); 6676 _bitMap->lock()->lock_without_safepoint_check(); 6677 _collector->startTimer(); 6678 } 6679 6680 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6681 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6682 assert(_markStack->isEmpty(), 6683 "should drain stack to limit stack usage"); 6684 // convert ptr to an oop preparatory to scanning 6685 oop obj = oop(ptr); 6686 // Ignore mark word in verification below, since we 6687 // may be running concurrent with mutators. 6688 assert(obj->is_oop(true), "should be an oop"); 6689 assert(_finger <= ptr, "_finger runneth ahead"); 6690 // advance the finger to right end of this object 6691 _finger = ptr + obj->size(); 6692 assert(_finger > ptr, "we just incremented it above"); 6693 // On large heaps, it may take us some time to get through 6694 // the marking phase. During 6695 // this time it's possible that a lot of mutations have 6696 // accumulated in the card table and the mod union table -- 6697 // these mutation records are redundant until we have 6698 // actually traced into the corresponding card. 6699 // Here, we check whether advancing the finger would make 6700 // us cross into a new card, and if so clear corresponding 6701 // cards in the MUT (preclean them in the card-table in the 6702 // future). 6703 6704 DEBUG_ONLY(if (!_verifying) {) 6705 // The clean-on-enter optimization is disabled by default, 6706 // until we fix 6178663. 6707 if (CMSCleanOnEnter && (_finger > _threshold)) { 6708 // [_threshold, _finger) represents the interval 6709 // of cards to be cleared in MUT (or precleaned in card table). 6710 // The set of cards to be cleared is all those that overlap 6711 // with the interval [_threshold, _finger); note that 6712 // _threshold is always kept card-aligned but _finger isn't 6713 // always card-aligned. 6714 HeapWord* old_threshold = _threshold; 6715 assert(old_threshold == (HeapWord*)round_to( 6716 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6717 "_threshold should always be card-aligned"); 6718 _threshold = (HeapWord*)round_to( 6719 (intptr_t)_finger, CardTableModRefBS::card_size); 6720 MemRegion mr(old_threshold, _threshold); 6721 assert(!mr.is_empty(), "Control point invariant"); 6722 assert(_span.contains(mr), "Should clear within span"); 6723 _mut->clear_range(mr); 6724 } 6725 DEBUG_ONLY(}) 6726 // Note: the finger doesn't advance while we drain 6727 // the stack below. 6728 PushOrMarkClosure pushOrMarkClosure(_collector, 6729 _span, _bitMap, _markStack, 6730 _finger, this); 6731 bool res = _markStack->push(obj); 6732 assert(res, "Empty non-zero size stack should have space for single push"); 6733 while (!_markStack->isEmpty()) { 6734 oop new_oop = _markStack->pop(); 6735 // Skip verifying header mark word below because we are 6736 // running concurrent with mutators. 6737 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6738 // now scan this oop's oops 6739 new_oop->oop_iterate(&pushOrMarkClosure); 6740 do_yield_check(); 6741 } 6742 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6743 } 6744 6745 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6746 CMSCollector* collector, MemRegion span, 6747 CMSBitMap* bit_map, 6748 OopTaskQueue* work_queue, 6749 CMSMarkStack* overflow_stack): 6750 _collector(collector), 6751 _whole_span(collector->_span), 6752 _span(span), 6753 _bit_map(bit_map), 6754 _mut(&collector->_modUnionTable), 6755 _work_queue(work_queue), 6756 _overflow_stack(overflow_stack), 6757 _skip_bits(0), 6758 _task(task) 6759 { 6760 assert(_work_queue->size() == 0, "work_queue should be empty"); 6761 _finger = span.start(); 6762 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6763 assert(_span.contains(_finger), "Out of bounds _finger?"); 6764 } 6765 6766 // Should revisit to see if this should be restructured for 6767 // greater efficiency. 6768 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6769 if (_skip_bits > 0) { 6770 _skip_bits--; 6771 return true; 6772 } 6773 // convert offset into a HeapWord* 6774 HeapWord* addr = _bit_map->startWord() + offset; 6775 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6776 "address out of range"); 6777 assert(_bit_map->isMarked(addr), "tautology"); 6778 if (_bit_map->isMarked(addr+1)) { 6779 // this is an allocated object that might not yet be initialized 6780 assert(_skip_bits == 0, "tautology"); 6781 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6782 oop p = oop(addr); 6783 if (p->klass_or_null() == NULL) { 6784 // in the case of Clean-on-Enter optimization, redirty card 6785 // and avoid clearing card by increasing the threshold. 6786 return true; 6787 } 6788 } 6789 scan_oops_in_oop(addr); 6790 return true; 6791 } 6792 6793 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6794 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6795 // Should we assert that our work queue is empty or 6796 // below some drain limit? 6797 assert(_work_queue->size() == 0, 6798 "should drain stack to limit stack usage"); 6799 // convert ptr to an oop preparatory to scanning 6800 oop obj = oop(ptr); 6801 // Ignore mark word in verification below, since we 6802 // may be running concurrent with mutators. 6803 assert(obj->is_oop(true), "should be an oop"); 6804 assert(_finger <= ptr, "_finger runneth ahead"); 6805 // advance the finger to right end of this object 6806 _finger = ptr + obj->size(); 6807 assert(_finger > ptr, "we just incremented it above"); 6808 // On large heaps, it may take us some time to get through 6809 // the marking phase. During 6810 // this time it's possible that a lot of mutations have 6811 // accumulated in the card table and the mod union table -- 6812 // these mutation records are redundant until we have 6813 // actually traced into the corresponding card. 6814 // Here, we check whether advancing the finger would make 6815 // us cross into a new card, and if so clear corresponding 6816 // cards in the MUT (preclean them in the card-table in the 6817 // future). 6818 6819 // The clean-on-enter optimization is disabled by default, 6820 // until we fix 6178663. 6821 if (CMSCleanOnEnter && (_finger > _threshold)) { 6822 // [_threshold, _finger) represents the interval 6823 // of cards to be cleared in MUT (or precleaned in card table). 6824 // The set of cards to be cleared is all those that overlap 6825 // with the interval [_threshold, _finger); note that 6826 // _threshold is always kept card-aligned but _finger isn't 6827 // always card-aligned. 6828 HeapWord* old_threshold = _threshold; 6829 assert(old_threshold == (HeapWord*)round_to( 6830 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6831 "_threshold should always be card-aligned"); 6832 _threshold = (HeapWord*)round_to( 6833 (intptr_t)_finger, CardTableModRefBS::card_size); 6834 MemRegion mr(old_threshold, _threshold); 6835 assert(!mr.is_empty(), "Control point invariant"); 6836 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6837 _mut->clear_range(mr); 6838 } 6839 6840 // Note: the local finger doesn't advance while we drain 6841 // the stack below, but the global finger sure can and will. 6842 HeapWord** gfa = _task->global_finger_addr(); 6843 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6844 _span, _bit_map, 6845 _work_queue, 6846 _overflow_stack, 6847 _finger, 6848 gfa, this); 6849 bool res = _work_queue->push(obj); // overflow could occur here 6850 assert(res, "Will hold once we use workqueues"); 6851 while (true) { 6852 oop new_oop; 6853 if (!_work_queue->pop_local(new_oop)) { 6854 // We emptied our work_queue; check if there's stuff that can 6855 // be gotten from the overflow stack. 6856 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6857 _overflow_stack, _work_queue)) { 6858 do_yield_check(); 6859 continue; 6860 } else { // done 6861 break; 6862 } 6863 } 6864 // Skip verifying header mark word below because we are 6865 // running concurrent with mutators. 6866 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6867 // now scan this oop's oops 6868 new_oop->oop_iterate(&pushOrMarkClosure); 6869 do_yield_check(); 6870 } 6871 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6872 } 6873 6874 // Yield in response to a request from VM Thread or 6875 // from mutators. 6876 void Par_MarkFromRootsClosure::do_yield_work() { 6877 assert(_task != NULL, "sanity"); 6878 _task->yield(); 6879 } 6880 6881 // A variant of the above used for verifying CMS marking work. 6882 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6883 MemRegion span, 6884 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6885 CMSMarkStack* mark_stack): 6886 _collector(collector), 6887 _span(span), 6888 _verification_bm(verification_bm), 6889 _cms_bm(cms_bm), 6890 _mark_stack(mark_stack), 6891 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6892 mark_stack) 6893 { 6894 assert(_mark_stack->isEmpty(), "stack should be empty"); 6895 _finger = _verification_bm->startWord(); 6896 assert(_collector->_restart_addr == NULL, "Sanity check"); 6897 assert(_span.contains(_finger), "Out of bounds _finger?"); 6898 } 6899 6900 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6901 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6902 assert(_span.contains(addr), "Out of bounds _finger?"); 6903 _finger = addr; 6904 } 6905 6906 // Should revisit to see if this should be restructured for 6907 // greater efficiency. 6908 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6909 // convert offset into a HeapWord* 6910 HeapWord* addr = _verification_bm->startWord() + offset; 6911 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6912 "address out of range"); 6913 assert(_verification_bm->isMarked(addr), "tautology"); 6914 assert(_cms_bm->isMarked(addr), "tautology"); 6915 6916 assert(_mark_stack->isEmpty(), 6917 "should drain stack to limit stack usage"); 6918 // convert addr to an oop preparatory to scanning 6919 oop obj = oop(addr); 6920 assert(obj->is_oop(), "should be an oop"); 6921 assert(_finger <= addr, "_finger runneth ahead"); 6922 // advance the finger to right end of this object 6923 _finger = addr + obj->size(); 6924 assert(_finger > addr, "we just incremented it above"); 6925 // Note: the finger doesn't advance while we drain 6926 // the stack below. 6927 bool res = _mark_stack->push(obj); 6928 assert(res, "Empty non-zero size stack should have space for single push"); 6929 while (!_mark_stack->isEmpty()) { 6930 oop new_oop = _mark_stack->pop(); 6931 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6932 // now scan this oop's oops 6933 new_oop->oop_iterate(&_pam_verify_closure); 6934 } 6935 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6936 return true; 6937 } 6938 6939 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6940 CMSCollector* collector, MemRegion span, 6941 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6942 CMSMarkStack* mark_stack): 6943 MetadataAwareOopClosure(collector->ref_processor()), 6944 _collector(collector), 6945 _span(span), 6946 _verification_bm(verification_bm), 6947 _cms_bm(cms_bm), 6948 _mark_stack(mark_stack) 6949 { } 6950 6951 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6952 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6953 6954 // Upon stack overflow, we discard (part of) the stack, 6955 // remembering the least address amongst those discarded 6956 // in CMSCollector's _restart_address. 6957 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6958 // Remember the least grey address discarded 6959 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6960 _collector->lower_restart_addr(ra); 6961 _mark_stack->reset(); // discard stack contents 6962 _mark_stack->expand(); // expand the stack if possible 6963 } 6964 6965 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6966 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6967 HeapWord* addr = (HeapWord*)obj; 6968 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6969 // Oop lies in _span and isn't yet grey or black 6970 _verification_bm->mark(addr); // now grey 6971 if (!_cms_bm->isMarked(addr)) { 6972 oop(addr)->print(); 6973 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6974 p2i(addr)); 6975 fatal("... aborting"); 6976 } 6977 6978 if (!_mark_stack->push(obj)) { // stack overflow 6979 if (PrintCMSStatistics != 0) { 6980 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6981 SIZE_FORMAT, _mark_stack->capacity()); 6982 } 6983 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6984 handle_stack_overflow(addr); 6985 } 6986 // anything including and to the right of _finger 6987 // will be scanned as we iterate over the remainder of the 6988 // bit map 6989 } 6990 } 6991 6992 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6993 MemRegion span, 6994 CMSBitMap* bitMap, CMSMarkStack* markStack, 6995 HeapWord* finger, MarkFromRootsClosure* parent) : 6996 MetadataAwareOopClosure(collector->ref_processor()), 6997 _collector(collector), 6998 _span(span), 6999 _bitMap(bitMap), 7000 _markStack(markStack), 7001 _finger(finger), 7002 _parent(parent) 7003 { } 7004 7005 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7006 MemRegion span, 7007 CMSBitMap* bit_map, 7008 OopTaskQueue* work_queue, 7009 CMSMarkStack* overflow_stack, 7010 HeapWord* finger, 7011 HeapWord** global_finger_addr, 7012 Par_MarkFromRootsClosure* parent) : 7013 MetadataAwareOopClosure(collector->ref_processor()), 7014 _collector(collector), 7015 _whole_span(collector->_span), 7016 _span(span), 7017 _bit_map(bit_map), 7018 _work_queue(work_queue), 7019 _overflow_stack(overflow_stack), 7020 _finger(finger), 7021 _global_finger_addr(global_finger_addr), 7022 _parent(parent) 7023 { } 7024 7025 // Assumes thread-safe access by callers, who are 7026 // responsible for mutual exclusion. 7027 void CMSCollector::lower_restart_addr(HeapWord* low) { 7028 assert(_span.contains(low), "Out of bounds addr"); 7029 if (_restart_addr == NULL) { 7030 _restart_addr = low; 7031 } else { 7032 _restart_addr = MIN2(_restart_addr, low); 7033 } 7034 } 7035 7036 // Upon stack overflow, we discard (part of) the stack, 7037 // remembering the least address amongst those discarded 7038 // in CMSCollector's _restart_address. 7039 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7040 // Remember the least grey address discarded 7041 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7042 _collector->lower_restart_addr(ra); 7043 _markStack->reset(); // discard stack contents 7044 _markStack->expand(); // expand the stack if possible 7045 } 7046 7047 // Upon stack overflow, we discard (part of) the stack, 7048 // remembering the least address amongst those discarded 7049 // in CMSCollector's _restart_address. 7050 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7051 // We need to do this under a mutex to prevent other 7052 // workers from interfering with the work done below. 7053 MutexLockerEx ml(_overflow_stack->par_lock(), 7054 Mutex::_no_safepoint_check_flag); 7055 // Remember the least grey address discarded 7056 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7057 _collector->lower_restart_addr(ra); 7058 _overflow_stack->reset(); // discard stack contents 7059 _overflow_stack->expand(); // expand the stack if possible 7060 } 7061 7062 void PushOrMarkClosure::do_oop(oop obj) { 7063 // Ignore mark word because we are running concurrent with mutators. 7064 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7065 HeapWord* addr = (HeapWord*)obj; 7066 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7067 // Oop lies in _span and isn't yet grey or black 7068 _bitMap->mark(addr); // now grey 7069 if (addr < _finger) { 7070 // the bit map iteration has already either passed, or 7071 // sampled, this bit in the bit map; we'll need to 7072 // use the marking stack to scan this oop's oops. 7073 bool simulate_overflow = false; 7074 NOT_PRODUCT( 7075 if (CMSMarkStackOverflowALot && 7076 _collector->simulate_overflow()) { 7077 // simulate a stack overflow 7078 simulate_overflow = true; 7079 } 7080 ) 7081 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7082 if (PrintCMSStatistics != 0) { 7083 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7084 SIZE_FORMAT, _markStack->capacity()); 7085 } 7086 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7087 handle_stack_overflow(addr); 7088 } 7089 } 7090 // anything including and to the right of _finger 7091 // will be scanned as we iterate over the remainder of the 7092 // bit map 7093 do_yield_check(); 7094 } 7095 } 7096 7097 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7098 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7099 7100 void Par_PushOrMarkClosure::do_oop(oop obj) { 7101 // Ignore mark word because we are running concurrent with mutators. 7102 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7103 HeapWord* addr = (HeapWord*)obj; 7104 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7105 // Oop lies in _span and isn't yet grey or black 7106 // We read the global_finger (volatile read) strictly after marking oop 7107 bool res = _bit_map->par_mark(addr); // now grey 7108 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7109 // Should we push this marked oop on our stack? 7110 // -- if someone else marked it, nothing to do 7111 // -- if target oop is above global finger nothing to do 7112 // -- if target oop is in chunk and above local finger 7113 // then nothing to do 7114 // -- else push on work queue 7115 if ( !res // someone else marked it, they will deal with it 7116 || (addr >= *gfa) // will be scanned in a later task 7117 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7118 return; 7119 } 7120 // the bit map iteration has already either passed, or 7121 // sampled, this bit in the bit map; we'll need to 7122 // use the marking stack to scan this oop's oops. 7123 bool simulate_overflow = false; 7124 NOT_PRODUCT( 7125 if (CMSMarkStackOverflowALot && 7126 _collector->simulate_overflow()) { 7127 // simulate a stack overflow 7128 simulate_overflow = true; 7129 } 7130 ) 7131 if (simulate_overflow || 7132 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7133 // stack overflow 7134 if (PrintCMSStatistics != 0) { 7135 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7136 SIZE_FORMAT, _overflow_stack->capacity()); 7137 } 7138 // We cannot assert that the overflow stack is full because 7139 // it may have been emptied since. 7140 assert(simulate_overflow || 7141 _work_queue->size() == _work_queue->max_elems(), 7142 "Else push should have succeeded"); 7143 handle_stack_overflow(addr); 7144 } 7145 do_yield_check(); 7146 } 7147 } 7148 7149 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7150 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7151 7152 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7153 MemRegion span, 7154 ReferenceProcessor* rp, 7155 CMSBitMap* bit_map, 7156 CMSBitMap* mod_union_table, 7157 CMSMarkStack* mark_stack, 7158 bool concurrent_precleaning): 7159 MetadataAwareOopClosure(rp), 7160 _collector(collector), 7161 _span(span), 7162 _bit_map(bit_map), 7163 _mod_union_table(mod_union_table), 7164 _mark_stack(mark_stack), 7165 _concurrent_precleaning(concurrent_precleaning) 7166 { 7167 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7168 } 7169 7170 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7171 // the non-parallel version (the parallel version appears further below.) 7172 void PushAndMarkClosure::do_oop(oop obj) { 7173 // Ignore mark word verification. If during concurrent precleaning, 7174 // the object monitor may be locked. If during the checkpoint 7175 // phases, the object may already have been reached by a different 7176 // path and may be at the end of the global overflow list (so 7177 // the mark word may be NULL). 7178 assert(obj->is_oop_or_null(true /* ignore mark word */), 7179 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7180 HeapWord* addr = (HeapWord*)obj; 7181 // Check if oop points into the CMS generation 7182 // and is not marked 7183 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7184 // a white object ... 7185 _bit_map->mark(addr); // ... now grey 7186 // push on the marking stack (grey set) 7187 bool simulate_overflow = false; 7188 NOT_PRODUCT( 7189 if (CMSMarkStackOverflowALot && 7190 _collector->simulate_overflow()) { 7191 // simulate a stack overflow 7192 simulate_overflow = true; 7193 } 7194 ) 7195 if (simulate_overflow || !_mark_stack->push(obj)) { 7196 if (_concurrent_precleaning) { 7197 // During precleaning we can just dirty the appropriate card(s) 7198 // in the mod union table, thus ensuring that the object remains 7199 // in the grey set and continue. In the case of object arrays 7200 // we need to dirty all of the cards that the object spans, 7201 // since the rescan of object arrays will be limited to the 7202 // dirty cards. 7203 // Note that no one can be interfering with us in this action 7204 // of dirtying the mod union table, so no locking or atomics 7205 // are required. 7206 if (obj->is_objArray()) { 7207 size_t sz = obj->size(); 7208 HeapWord* end_card_addr = (HeapWord*)round_to( 7209 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7210 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7211 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7212 _mod_union_table->mark_range(redirty_range); 7213 } else { 7214 _mod_union_table->mark(addr); 7215 } 7216 _collector->_ser_pmc_preclean_ovflw++; 7217 } else { 7218 // During the remark phase, we need to remember this oop 7219 // in the overflow list. 7220 _collector->push_on_overflow_list(obj); 7221 _collector->_ser_pmc_remark_ovflw++; 7222 } 7223 } 7224 } 7225 } 7226 7227 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7228 MemRegion span, 7229 ReferenceProcessor* rp, 7230 CMSBitMap* bit_map, 7231 OopTaskQueue* work_queue): 7232 MetadataAwareOopClosure(rp), 7233 _collector(collector), 7234 _span(span), 7235 _bit_map(bit_map), 7236 _work_queue(work_queue) 7237 { 7238 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7239 } 7240 7241 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7242 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7243 7244 // Grey object rescan during second checkpoint phase -- 7245 // the parallel version. 7246 void Par_PushAndMarkClosure::do_oop(oop obj) { 7247 // In the assert below, we ignore the mark word because 7248 // this oop may point to an already visited object that is 7249 // on the overflow stack (in which case the mark word has 7250 // been hijacked for chaining into the overflow stack -- 7251 // if this is the last object in the overflow stack then 7252 // its mark word will be NULL). Because this object may 7253 // have been subsequently popped off the global overflow 7254 // stack, and the mark word possibly restored to the prototypical 7255 // value, by the time we get to examined this failing assert in 7256 // the debugger, is_oop_or_null(false) may subsequently start 7257 // to hold. 7258 assert(obj->is_oop_or_null(true), 7259 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7260 HeapWord* addr = (HeapWord*)obj; 7261 // Check if oop points into the CMS generation 7262 // and is not marked 7263 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7264 // a white object ... 7265 // If we manage to "claim" the object, by being the 7266 // first thread to mark it, then we push it on our 7267 // marking stack 7268 if (_bit_map->par_mark(addr)) { // ... now grey 7269 // push on work queue (grey set) 7270 bool simulate_overflow = false; 7271 NOT_PRODUCT( 7272 if (CMSMarkStackOverflowALot && 7273 _collector->par_simulate_overflow()) { 7274 // simulate a stack overflow 7275 simulate_overflow = true; 7276 } 7277 ) 7278 if (simulate_overflow || !_work_queue->push(obj)) { 7279 _collector->par_push_on_overflow_list(obj); 7280 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7281 } 7282 } // Else, some other thread got there first 7283 } 7284 } 7285 7286 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7287 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7288 7289 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7290 Mutex* bml = _collector->bitMapLock(); 7291 assert_lock_strong(bml); 7292 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7293 "CMS thread should hold CMS token"); 7294 7295 bml->unlock(); 7296 ConcurrentMarkSweepThread::desynchronize(true); 7297 7298 _collector->stopTimer(); 7299 if (PrintCMSStatistics != 0) { 7300 _collector->incrementYields(); 7301 } 7302 7303 // See the comment in coordinator_yield() 7304 for (unsigned i = 0; i < CMSYieldSleepCount && 7305 ConcurrentMarkSweepThread::should_yield() && 7306 !CMSCollector::foregroundGCIsActive(); ++i) { 7307 os::sleep(Thread::current(), 1, false); 7308 } 7309 7310 ConcurrentMarkSweepThread::synchronize(true); 7311 bml->lock(); 7312 7313 _collector->startTimer(); 7314 } 7315 7316 bool CMSPrecleanRefsYieldClosure::should_return() { 7317 if (ConcurrentMarkSweepThread::should_yield()) { 7318 do_yield_work(); 7319 } 7320 return _collector->foregroundGCIsActive(); 7321 } 7322 7323 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7324 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7325 "mr should be aligned to start at a card boundary"); 7326 // We'd like to assert: 7327 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7328 // "mr should be a range of cards"); 7329 // However, that would be too strong in one case -- the last 7330 // partition ends at _unallocated_block which, in general, can be 7331 // an arbitrary boundary, not necessarily card aligned. 7332 if (PrintCMSStatistics != 0) { 7333 _num_dirty_cards += 7334 mr.word_size()/CardTableModRefBS::card_size_in_words; 7335 } 7336 _space->object_iterate_mem(mr, &_scan_cl); 7337 } 7338 7339 SweepClosure::SweepClosure(CMSCollector* collector, 7340 ConcurrentMarkSweepGeneration* g, 7341 CMSBitMap* bitMap, bool should_yield) : 7342 _collector(collector), 7343 _g(g), 7344 _sp(g->cmsSpace()), 7345 _limit(_sp->sweep_limit()), 7346 _freelistLock(_sp->freelistLock()), 7347 _bitMap(bitMap), 7348 _yield(should_yield), 7349 _inFreeRange(false), // No free range at beginning of sweep 7350 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7351 _lastFreeRangeCoalesced(false), 7352 _freeFinger(g->used_region().start()) 7353 { 7354 NOT_PRODUCT( 7355 _numObjectsFreed = 0; 7356 _numWordsFreed = 0; 7357 _numObjectsLive = 0; 7358 _numWordsLive = 0; 7359 _numObjectsAlreadyFree = 0; 7360 _numWordsAlreadyFree = 0; 7361 _last_fc = NULL; 7362 7363 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7364 _sp->dictionary()->initialize_dict_returned_bytes(); 7365 ) 7366 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7367 "sweep _limit out of bounds"); 7368 if (CMSTraceSweeper) { 7369 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7370 p2i(_limit)); 7371 } 7372 } 7373 7374 void SweepClosure::print_on(outputStream* st) const { 7375 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7376 p2i(_sp->bottom()), p2i(_sp->end())); 7377 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7378 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7379 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7380 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7381 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7382 } 7383 7384 #ifndef PRODUCT 7385 // Assertion checking only: no useful work in product mode -- 7386 // however, if any of the flags below become product flags, 7387 // you may need to review this code to see if it needs to be 7388 // enabled in product mode. 7389 SweepClosure::~SweepClosure() { 7390 assert_lock_strong(_freelistLock); 7391 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7392 "sweep _limit out of bounds"); 7393 if (inFreeRange()) { 7394 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7395 print(); 7396 ShouldNotReachHere(); 7397 } 7398 if (Verbose && PrintGC) { 7399 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7400 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7401 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, " 7402 SIZE_FORMAT " bytes " 7403 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7404 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7405 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7406 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7407 * sizeof(HeapWord); 7408 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7409 7410 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7411 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7412 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7413 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7414 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes); 7415 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes", 7416 indexListReturnedBytes); 7417 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes", 7418 dict_returned_bytes); 7419 } 7420 } 7421 if (CMSTraceSweeper) { 7422 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7423 p2i(_limit)); 7424 } 7425 } 7426 #endif // PRODUCT 7427 7428 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7429 bool freeRangeInFreeLists) { 7430 if (CMSTraceSweeper) { 7431 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7432 p2i(freeFinger), freeRangeInFreeLists); 7433 } 7434 assert(!inFreeRange(), "Trampling existing free range"); 7435 set_inFreeRange(true); 7436 set_lastFreeRangeCoalesced(false); 7437 7438 set_freeFinger(freeFinger); 7439 set_freeRangeInFreeLists(freeRangeInFreeLists); 7440 if (CMSTestInFreeList) { 7441 if (freeRangeInFreeLists) { 7442 FreeChunk* fc = (FreeChunk*) freeFinger; 7443 assert(fc->is_free(), "A chunk on the free list should be free."); 7444 assert(fc->size() > 0, "Free range should have a size"); 7445 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7446 } 7447 } 7448 } 7449 7450 // Note that the sweeper runs concurrently with mutators. Thus, 7451 // it is possible for direct allocation in this generation to happen 7452 // in the middle of the sweep. Note that the sweeper also coalesces 7453 // contiguous free blocks. Thus, unless the sweeper and the allocator 7454 // synchronize appropriately freshly allocated blocks may get swept up. 7455 // This is accomplished by the sweeper locking the free lists while 7456 // it is sweeping. Thus blocks that are determined to be free are 7457 // indeed free. There is however one additional complication: 7458 // blocks that have been allocated since the final checkpoint and 7459 // mark, will not have been marked and so would be treated as 7460 // unreachable and swept up. To prevent this, the allocator marks 7461 // the bit map when allocating during the sweep phase. This leads, 7462 // however, to a further complication -- objects may have been allocated 7463 // but not yet initialized -- in the sense that the header isn't yet 7464 // installed. The sweeper can not then determine the size of the block 7465 // in order to skip over it. To deal with this case, we use a technique 7466 // (due to Printezis) to encode such uninitialized block sizes in the 7467 // bit map. Since the bit map uses a bit per every HeapWord, but the 7468 // CMS generation has a minimum object size of 3 HeapWords, it follows 7469 // that "normal marks" won't be adjacent in the bit map (there will 7470 // always be at least two 0 bits between successive 1 bits). We make use 7471 // of these "unused" bits to represent uninitialized blocks -- the bit 7472 // corresponding to the start of the uninitialized object and the next 7473 // bit are both set. Finally, a 1 bit marks the end of the object that 7474 // started with the two consecutive 1 bits to indicate its potentially 7475 // uninitialized state. 7476 7477 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7478 FreeChunk* fc = (FreeChunk*)addr; 7479 size_t res; 7480 7481 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7482 // than "addr == _limit" because although _limit was a block boundary when 7483 // we started the sweep, it may no longer be one because heap expansion 7484 // may have caused us to coalesce the block ending at the address _limit 7485 // with a newly expanded chunk (this happens when _limit was set to the 7486 // previous _end of the space), so we may have stepped past _limit: 7487 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7488 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7489 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7490 "sweep _limit out of bounds"); 7491 assert(addr < _sp->end(), "addr out of bounds"); 7492 // Flush any free range we might be holding as a single 7493 // coalesced chunk to the appropriate free list. 7494 if (inFreeRange()) { 7495 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7496 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()))); 7497 flush_cur_free_chunk(freeFinger(), 7498 pointer_delta(addr, freeFinger())); 7499 if (CMSTraceSweeper) { 7500 gclog_or_tty->print("Sweep: last chunk: "); 7501 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") " 7502 "[coalesced:%d]\n", 7503 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7504 lastFreeRangeCoalesced() ? 1 : 0); 7505 } 7506 } 7507 7508 // help the iterator loop finish 7509 return pointer_delta(_sp->end(), addr); 7510 } 7511 7512 assert(addr < _limit, "sweep invariant"); 7513 // check if we should yield 7514 do_yield_check(addr); 7515 if (fc->is_free()) { 7516 // Chunk that is already free 7517 res = fc->size(); 7518 do_already_free_chunk(fc); 7519 debug_only(_sp->verifyFreeLists()); 7520 // If we flush the chunk at hand in lookahead_and_flush() 7521 // and it's coalesced with a preceding chunk, then the 7522 // process of "mangling" the payload of the coalesced block 7523 // will cause erasure of the size information from the 7524 // (erstwhile) header of all the coalesced blocks but the 7525 // first, so the first disjunct in the assert will not hold 7526 // in that specific case (in which case the second disjunct 7527 // will hold). 7528 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7529 "Otherwise the size info doesn't change at this step"); 7530 NOT_PRODUCT( 7531 _numObjectsAlreadyFree++; 7532 _numWordsAlreadyFree += res; 7533 ) 7534 NOT_PRODUCT(_last_fc = fc;) 7535 } else if (!_bitMap->isMarked(addr)) { 7536 // Chunk is fresh garbage 7537 res = do_garbage_chunk(fc); 7538 debug_only(_sp->verifyFreeLists()); 7539 NOT_PRODUCT( 7540 _numObjectsFreed++; 7541 _numWordsFreed += res; 7542 ) 7543 } else { 7544 // Chunk that is alive. 7545 res = do_live_chunk(fc); 7546 debug_only(_sp->verifyFreeLists()); 7547 NOT_PRODUCT( 7548 _numObjectsLive++; 7549 _numWordsLive += res; 7550 ) 7551 } 7552 return res; 7553 } 7554 7555 // For the smart allocation, record following 7556 // split deaths - a free chunk is removed from its free list because 7557 // it is being split into two or more chunks. 7558 // split birth - a free chunk is being added to its free list because 7559 // a larger free chunk has been split and resulted in this free chunk. 7560 // coal death - a free chunk is being removed from its free list because 7561 // it is being coalesced into a large free chunk. 7562 // coal birth - a free chunk is being added to its free list because 7563 // it was created when two or more free chunks where coalesced into 7564 // this free chunk. 7565 // 7566 // These statistics are used to determine the desired number of free 7567 // chunks of a given size. The desired number is chosen to be relative 7568 // to the end of a CMS sweep. The desired number at the end of a sweep 7569 // is the 7570 // count-at-end-of-previous-sweep (an amount that was enough) 7571 // - count-at-beginning-of-current-sweep (the excess) 7572 // + split-births (gains in this size during interval) 7573 // - split-deaths (demands on this size during interval) 7574 // where the interval is from the end of one sweep to the end of the 7575 // next. 7576 // 7577 // When sweeping the sweeper maintains an accumulated chunk which is 7578 // the chunk that is made up of chunks that have been coalesced. That 7579 // will be termed the left-hand chunk. A new chunk of garbage that 7580 // is being considered for coalescing will be referred to as the 7581 // right-hand chunk. 7582 // 7583 // When making a decision on whether to coalesce a right-hand chunk with 7584 // the current left-hand chunk, the current count vs. the desired count 7585 // of the left-hand chunk is considered. Also if the right-hand chunk 7586 // is near the large chunk at the end of the heap (see 7587 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7588 // left-hand chunk is coalesced. 7589 // 7590 // When making a decision about whether to split a chunk, the desired count 7591 // vs. the current count of the candidate to be split is also considered. 7592 // If the candidate is underpopulated (currently fewer chunks than desired) 7593 // a chunk of an overpopulated (currently more chunks than desired) size may 7594 // be chosen. The "hint" associated with a free list, if non-null, points 7595 // to a free list which may be overpopulated. 7596 // 7597 7598 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7599 const size_t size = fc->size(); 7600 // Chunks that cannot be coalesced are not in the 7601 // free lists. 7602 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7603 assert(_sp->verify_chunk_in_free_list(fc), 7604 "free chunk should be in free lists"); 7605 } 7606 // a chunk that is already free, should not have been 7607 // marked in the bit map 7608 HeapWord* const addr = (HeapWord*) fc; 7609 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7610 // Verify that the bit map has no bits marked between 7611 // addr and purported end of this block. 7612 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7613 7614 // Some chunks cannot be coalesced under any circumstances. 7615 // See the definition of cantCoalesce(). 7616 if (!fc->cantCoalesce()) { 7617 // This chunk can potentially be coalesced. 7618 if (_sp->adaptive_freelists()) { 7619 // All the work is done in 7620 do_post_free_or_garbage_chunk(fc, size); 7621 } else { // Not adaptive free lists 7622 // this is a free chunk that can potentially be coalesced by the sweeper; 7623 if (!inFreeRange()) { 7624 // if the next chunk is a free block that can't be coalesced 7625 // it doesn't make sense to remove this chunk from the free lists 7626 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7627 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7628 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7629 nextChunk->is_free() && // ... which is free... 7630 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7631 // nothing to do 7632 } else { 7633 // Potentially the start of a new free range: 7634 // Don't eagerly remove it from the free lists. 7635 // No need to remove it if it will just be put 7636 // back again. (Also from a pragmatic point of view 7637 // if it is a free block in a region that is beyond 7638 // any allocated blocks, an assertion will fail) 7639 // Remember the start of a free run. 7640 initialize_free_range(addr, true); 7641 // end - can coalesce with next chunk 7642 } 7643 } else { 7644 // the midst of a free range, we are coalescing 7645 print_free_block_coalesced(fc); 7646 if (CMSTraceSweeper) { 7647 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7648 } 7649 // remove it from the free lists 7650 _sp->removeFreeChunkFromFreeLists(fc); 7651 set_lastFreeRangeCoalesced(true); 7652 // If the chunk is being coalesced and the current free range is 7653 // in the free lists, remove the current free range so that it 7654 // will be returned to the free lists in its entirety - all 7655 // the coalesced pieces included. 7656 if (freeRangeInFreeLists()) { 7657 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7658 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7659 "Size of free range is inconsistent with chunk size."); 7660 if (CMSTestInFreeList) { 7661 assert(_sp->verify_chunk_in_free_list(ffc), 7662 "free range is not in free lists"); 7663 } 7664 _sp->removeFreeChunkFromFreeLists(ffc); 7665 set_freeRangeInFreeLists(false); 7666 } 7667 } 7668 } 7669 // Note that if the chunk is not coalescable (the else arm 7670 // below), we unconditionally flush, without needing to do 7671 // a "lookahead," as we do below. 7672 if (inFreeRange()) lookahead_and_flush(fc, size); 7673 } else { 7674 // Code path common to both original and adaptive free lists. 7675 7676 // cant coalesce with previous block; this should be treated 7677 // as the end of a free run if any 7678 if (inFreeRange()) { 7679 // we kicked some butt; time to pick up the garbage 7680 assert(freeFinger() < addr, "freeFinger points too high"); 7681 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7682 } 7683 // else, nothing to do, just continue 7684 } 7685 } 7686 7687 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7688 // This is a chunk of garbage. It is not in any free list. 7689 // Add it to a free list or let it possibly be coalesced into 7690 // a larger chunk. 7691 HeapWord* const addr = (HeapWord*) fc; 7692 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7693 7694 if (_sp->adaptive_freelists()) { 7695 // Verify that the bit map has no bits marked between 7696 // addr and purported end of just dead object. 7697 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7698 7699 do_post_free_or_garbage_chunk(fc, size); 7700 } else { 7701 if (!inFreeRange()) { 7702 // start of a new free range 7703 assert(size > 0, "A free range should have a size"); 7704 initialize_free_range(addr, false); 7705 } else { 7706 // this will be swept up when we hit the end of the 7707 // free range 7708 if (CMSTraceSweeper) { 7709 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7710 } 7711 // If the chunk is being coalesced and the current free range is 7712 // in the free lists, remove the current free range so that it 7713 // will be returned to the free lists in its entirety - all 7714 // the coalesced pieces included. 7715 if (freeRangeInFreeLists()) { 7716 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7717 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7718 "Size of free range is inconsistent with chunk size."); 7719 if (CMSTestInFreeList) { 7720 assert(_sp->verify_chunk_in_free_list(ffc), 7721 "free range is not in free lists"); 7722 } 7723 _sp->removeFreeChunkFromFreeLists(ffc); 7724 set_freeRangeInFreeLists(false); 7725 } 7726 set_lastFreeRangeCoalesced(true); 7727 } 7728 // this will be swept up when we hit the end of the free range 7729 7730 // Verify that the bit map has no bits marked between 7731 // addr and purported end of just dead object. 7732 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7733 } 7734 assert(_limit >= addr + size, 7735 "A freshly garbage chunk can't possibly straddle over _limit"); 7736 if (inFreeRange()) lookahead_and_flush(fc, size); 7737 return size; 7738 } 7739 7740 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7741 HeapWord* addr = (HeapWord*) fc; 7742 // The sweeper has just found a live object. Return any accumulated 7743 // left hand chunk to the free lists. 7744 if (inFreeRange()) { 7745 assert(freeFinger() < addr, "freeFinger points too high"); 7746 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7747 } 7748 7749 // This object is live: we'd normally expect this to be 7750 // an oop, and like to assert the following: 7751 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7752 // However, as we commented above, this may be an object whose 7753 // header hasn't yet been initialized. 7754 size_t size; 7755 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7756 if (_bitMap->isMarked(addr + 1)) { 7757 // Determine the size from the bit map, rather than trying to 7758 // compute it from the object header. 7759 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7760 size = pointer_delta(nextOneAddr + 1, addr); 7761 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7762 "alignment problem"); 7763 7764 #ifdef ASSERT 7765 if (oop(addr)->klass_or_null() != NULL) { 7766 // Ignore mark word because we are running concurrent with mutators 7767 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7768 assert(size == 7769 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7770 "P-mark and computed size do not agree"); 7771 } 7772 #endif 7773 7774 } else { 7775 // This should be an initialized object that's alive. 7776 assert(oop(addr)->klass_or_null() != NULL, 7777 "Should be an initialized object"); 7778 // Ignore mark word because we are running concurrent with mutators 7779 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7780 // Verify that the bit map has no bits marked between 7781 // addr and purported end of this block. 7782 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7783 assert(size >= 3, "Necessary for Printezis marks to work"); 7784 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7785 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7786 } 7787 return size; 7788 } 7789 7790 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7791 size_t chunkSize) { 7792 // do_post_free_or_garbage_chunk() should only be called in the case 7793 // of the adaptive free list allocator. 7794 const bool fcInFreeLists = fc->is_free(); 7795 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7796 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7797 if (CMSTestInFreeList && fcInFreeLists) { 7798 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7799 } 7800 7801 if (CMSTraceSweeper) { 7802 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7803 } 7804 7805 HeapWord* const fc_addr = (HeapWord*) fc; 7806 7807 bool coalesce; 7808 const size_t left = pointer_delta(fc_addr, freeFinger()); 7809 const size_t right = chunkSize; 7810 switch (FLSCoalescePolicy) { 7811 // numeric value forms a coalition aggressiveness metric 7812 case 0: { // never coalesce 7813 coalesce = false; 7814 break; 7815 } 7816 case 1: { // coalesce if left & right chunks on overpopulated lists 7817 coalesce = _sp->coalOverPopulated(left) && 7818 _sp->coalOverPopulated(right); 7819 break; 7820 } 7821 case 2: { // coalesce if left chunk on overpopulated list (default) 7822 coalesce = _sp->coalOverPopulated(left); 7823 break; 7824 } 7825 case 3: { // coalesce if left OR right chunk on overpopulated list 7826 coalesce = _sp->coalOverPopulated(left) || 7827 _sp->coalOverPopulated(right); 7828 break; 7829 } 7830 case 4: { // always coalesce 7831 coalesce = true; 7832 break; 7833 } 7834 default: 7835 ShouldNotReachHere(); 7836 } 7837 7838 // Should the current free range be coalesced? 7839 // If the chunk is in a free range and either we decided to coalesce above 7840 // or the chunk is near the large block at the end of the heap 7841 // (isNearLargestChunk() returns true), then coalesce this chunk. 7842 const bool doCoalesce = inFreeRange() 7843 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7844 if (doCoalesce) { 7845 // Coalesce the current free range on the left with the new 7846 // chunk on the right. If either is on a free list, 7847 // it must be removed from the list and stashed in the closure. 7848 if (freeRangeInFreeLists()) { 7849 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7850 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7851 "Size of free range is inconsistent with chunk size."); 7852 if (CMSTestInFreeList) { 7853 assert(_sp->verify_chunk_in_free_list(ffc), 7854 "Chunk is not in free lists"); 7855 } 7856 _sp->coalDeath(ffc->size()); 7857 _sp->removeFreeChunkFromFreeLists(ffc); 7858 set_freeRangeInFreeLists(false); 7859 } 7860 if (fcInFreeLists) { 7861 _sp->coalDeath(chunkSize); 7862 assert(fc->size() == chunkSize, 7863 "The chunk has the wrong size or is not in the free lists"); 7864 _sp->removeFreeChunkFromFreeLists(fc); 7865 } 7866 set_lastFreeRangeCoalesced(true); 7867 print_free_block_coalesced(fc); 7868 } else { // not in a free range and/or should not coalesce 7869 // Return the current free range and start a new one. 7870 if (inFreeRange()) { 7871 // In a free range but cannot coalesce with the right hand chunk. 7872 // Put the current free range into the free lists. 7873 flush_cur_free_chunk(freeFinger(), 7874 pointer_delta(fc_addr, freeFinger())); 7875 } 7876 // Set up for new free range. Pass along whether the right hand 7877 // chunk is in the free lists. 7878 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7879 } 7880 } 7881 7882 // Lookahead flush: 7883 // If we are tracking a free range, and this is the last chunk that 7884 // we'll look at because its end crosses past _limit, we'll preemptively 7885 // flush it along with any free range we may be holding on to. Note that 7886 // this can be the case only for an already free or freshly garbage 7887 // chunk. If this block is an object, it can never straddle 7888 // over _limit. The "straddling" occurs when _limit is set at 7889 // the previous end of the space when this cycle started, and 7890 // a subsequent heap expansion caused the previously co-terminal 7891 // free block to be coalesced with the newly expanded portion, 7892 // thus rendering _limit a non-block-boundary making it dangerous 7893 // for the sweeper to step over and examine. 7894 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7895 assert(inFreeRange(), "Should only be called if currently in a free range."); 7896 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7897 assert(_sp->used_region().contains(eob - 1), 7898 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7899 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7900 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7901 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7902 if (eob >= _limit) { 7903 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7904 if (CMSTraceSweeper) { 7905 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7906 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7907 "[" PTR_FORMAT "," PTR_FORMAT ")", 7908 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7909 } 7910 // Return the storage we are tracking back into the free lists. 7911 if (CMSTraceSweeper) { 7912 gclog_or_tty->print_cr("Flushing ... "); 7913 } 7914 assert(freeFinger() < eob, "Error"); 7915 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7916 } 7917 } 7918 7919 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7920 assert(inFreeRange(), "Should only be called if currently in a free range."); 7921 assert(size > 0, 7922 "A zero sized chunk cannot be added to the free lists."); 7923 if (!freeRangeInFreeLists()) { 7924 if (CMSTestInFreeList) { 7925 FreeChunk* fc = (FreeChunk*) chunk; 7926 fc->set_size(size); 7927 assert(!_sp->verify_chunk_in_free_list(fc), 7928 "chunk should not be in free lists yet"); 7929 } 7930 if (CMSTraceSweeper) { 7931 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7932 p2i(chunk), size); 7933 } 7934 // A new free range is going to be starting. The current 7935 // free range has not been added to the free lists yet or 7936 // was removed so add it back. 7937 // If the current free range was coalesced, then the death 7938 // of the free range was recorded. Record a birth now. 7939 if (lastFreeRangeCoalesced()) { 7940 _sp->coalBirth(size); 7941 } 7942 _sp->addChunkAndRepairOffsetTable(chunk, size, 7943 lastFreeRangeCoalesced()); 7944 } else if (CMSTraceSweeper) { 7945 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7946 } 7947 set_inFreeRange(false); 7948 set_freeRangeInFreeLists(false); 7949 } 7950 7951 // We take a break if we've been at this for a while, 7952 // so as to avoid monopolizing the locks involved. 7953 void SweepClosure::do_yield_work(HeapWord* addr) { 7954 // Return current free chunk being used for coalescing (if any) 7955 // to the appropriate freelist. After yielding, the next 7956 // free block encountered will start a coalescing range of 7957 // free blocks. If the next free block is adjacent to the 7958 // chunk just flushed, they will need to wait for the next 7959 // sweep to be coalesced. 7960 if (inFreeRange()) { 7961 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7962 } 7963 7964 // First give up the locks, then yield, then re-lock. 7965 // We should probably use a constructor/destructor idiom to 7966 // do this unlock/lock or modify the MutexUnlocker class to 7967 // serve our purpose. XXX 7968 assert_lock_strong(_bitMap->lock()); 7969 assert_lock_strong(_freelistLock); 7970 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7971 "CMS thread should hold CMS token"); 7972 _bitMap->lock()->unlock(); 7973 _freelistLock->unlock(); 7974 ConcurrentMarkSweepThread::desynchronize(true); 7975 _collector->stopTimer(); 7976 if (PrintCMSStatistics != 0) { 7977 _collector->incrementYields(); 7978 } 7979 7980 // See the comment in coordinator_yield() 7981 for (unsigned i = 0; i < CMSYieldSleepCount && 7982 ConcurrentMarkSweepThread::should_yield() && 7983 !CMSCollector::foregroundGCIsActive(); ++i) { 7984 os::sleep(Thread::current(), 1, false); 7985 } 7986 7987 ConcurrentMarkSweepThread::synchronize(true); 7988 _freelistLock->lock(); 7989 _bitMap->lock()->lock_without_safepoint_check(); 7990 _collector->startTimer(); 7991 } 7992 7993 #ifndef PRODUCT 7994 // This is actually very useful in a product build if it can 7995 // be called from the debugger. Compile it into the product 7996 // as needed. 7997 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7998 return debug_cms_space->verify_chunk_in_free_list(fc); 7999 } 8000 #endif 8001 8002 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8003 if (CMSTraceSweeper) { 8004 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8005 p2i(fc), fc->size()); 8006 } 8007 } 8008 8009 // CMSIsAliveClosure 8010 bool CMSIsAliveClosure::do_object_b(oop obj) { 8011 HeapWord* addr = (HeapWord*)obj; 8012 return addr != NULL && 8013 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8014 } 8015 8016 8017 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8018 MemRegion span, 8019 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8020 bool cpc): 8021 _collector(collector), 8022 _span(span), 8023 _bit_map(bit_map), 8024 _mark_stack(mark_stack), 8025 _concurrent_precleaning(cpc) { 8026 assert(!_span.is_empty(), "Empty span could spell trouble"); 8027 } 8028 8029 8030 // CMSKeepAliveClosure: the serial version 8031 void CMSKeepAliveClosure::do_oop(oop obj) { 8032 HeapWord* addr = (HeapWord*)obj; 8033 if (_span.contains(addr) && 8034 !_bit_map->isMarked(addr)) { 8035 _bit_map->mark(addr); 8036 bool simulate_overflow = false; 8037 NOT_PRODUCT( 8038 if (CMSMarkStackOverflowALot && 8039 _collector->simulate_overflow()) { 8040 // simulate a stack overflow 8041 simulate_overflow = true; 8042 } 8043 ) 8044 if (simulate_overflow || !_mark_stack->push(obj)) { 8045 if (_concurrent_precleaning) { 8046 // We dirty the overflown object and let the remark 8047 // phase deal with it. 8048 assert(_collector->overflow_list_is_empty(), "Error"); 8049 // In the case of object arrays, we need to dirty all of 8050 // the cards that the object spans. No locking or atomics 8051 // are needed since no one else can be mutating the mod union 8052 // table. 8053 if (obj->is_objArray()) { 8054 size_t sz = obj->size(); 8055 HeapWord* end_card_addr = 8056 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8057 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8058 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8059 _collector->_modUnionTable.mark_range(redirty_range); 8060 } else { 8061 _collector->_modUnionTable.mark(addr); 8062 } 8063 _collector->_ser_kac_preclean_ovflw++; 8064 } else { 8065 _collector->push_on_overflow_list(obj); 8066 _collector->_ser_kac_ovflw++; 8067 } 8068 } 8069 } 8070 } 8071 8072 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8073 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8074 8075 // CMSParKeepAliveClosure: a parallel version of the above. 8076 // The work queues are private to each closure (thread), 8077 // but (may be) available for stealing by other threads. 8078 void CMSParKeepAliveClosure::do_oop(oop obj) { 8079 HeapWord* addr = (HeapWord*)obj; 8080 if (_span.contains(addr) && 8081 !_bit_map->isMarked(addr)) { 8082 // In general, during recursive tracing, several threads 8083 // may be concurrently getting here; the first one to 8084 // "tag" it, claims it. 8085 if (_bit_map->par_mark(addr)) { 8086 bool res = _work_queue->push(obj); 8087 assert(res, "Low water mark should be much less than capacity"); 8088 // Do a recursive trim in the hope that this will keep 8089 // stack usage lower, but leave some oops for potential stealers 8090 trim_queue(_low_water_mark); 8091 } // Else, another thread got there first 8092 } 8093 } 8094 8095 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8096 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8097 8098 void CMSParKeepAliveClosure::trim_queue(uint max) { 8099 while (_work_queue->size() > max) { 8100 oop new_oop; 8101 if (_work_queue->pop_local(new_oop)) { 8102 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8103 assert(_bit_map->isMarked((HeapWord*)new_oop), 8104 "no white objects on this stack!"); 8105 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8106 // iterate over the oops in this oop, marking and pushing 8107 // the ones in CMS heap (i.e. in _span). 8108 new_oop->oop_iterate(&_mark_and_push); 8109 } 8110 } 8111 } 8112 8113 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8114 CMSCollector* collector, 8115 MemRegion span, CMSBitMap* bit_map, 8116 OopTaskQueue* work_queue): 8117 _collector(collector), 8118 _span(span), 8119 _bit_map(bit_map), 8120 _work_queue(work_queue) { } 8121 8122 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8123 HeapWord* addr = (HeapWord*)obj; 8124 if (_span.contains(addr) && 8125 !_bit_map->isMarked(addr)) { 8126 if (_bit_map->par_mark(addr)) { 8127 bool simulate_overflow = false; 8128 NOT_PRODUCT( 8129 if (CMSMarkStackOverflowALot && 8130 _collector->par_simulate_overflow()) { 8131 // simulate a stack overflow 8132 simulate_overflow = true; 8133 } 8134 ) 8135 if (simulate_overflow || !_work_queue->push(obj)) { 8136 _collector->par_push_on_overflow_list(obj); 8137 _collector->_par_kac_ovflw++; 8138 } 8139 } // Else another thread got there already 8140 } 8141 } 8142 8143 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8144 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8145 8146 ////////////////////////////////////////////////////////////////// 8147 // CMSExpansionCause ///////////////////////////// 8148 ////////////////////////////////////////////////////////////////// 8149 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8150 switch (cause) { 8151 case _no_expansion: 8152 return "No expansion"; 8153 case _satisfy_free_ratio: 8154 return "Free ratio"; 8155 case _satisfy_promotion: 8156 return "Satisfy promotion"; 8157 case _satisfy_allocation: 8158 return "allocation"; 8159 case _allocate_par_lab: 8160 return "Par LAB"; 8161 case _allocate_par_spooling_space: 8162 return "Par Spooling Space"; 8163 case _adaptive_size_policy: 8164 return "Ergonomics"; 8165 default: 8166 return "unknown"; 8167 } 8168 } 8169 8170 void CMSDrainMarkingStackClosure::do_void() { 8171 // the max number to take from overflow list at a time 8172 const size_t num = _mark_stack->capacity()/4; 8173 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8174 "Overflow list should be NULL during concurrent phases"); 8175 while (!_mark_stack->isEmpty() || 8176 // if stack is empty, check the overflow list 8177 _collector->take_from_overflow_list(num, _mark_stack)) { 8178 oop obj = _mark_stack->pop(); 8179 HeapWord* addr = (HeapWord*)obj; 8180 assert(_span.contains(addr), "Should be within span"); 8181 assert(_bit_map->isMarked(addr), "Should be marked"); 8182 assert(obj->is_oop(), "Should be an oop"); 8183 obj->oop_iterate(_keep_alive); 8184 } 8185 } 8186 8187 void CMSParDrainMarkingStackClosure::do_void() { 8188 // drain queue 8189 trim_queue(0); 8190 } 8191 8192 // Trim our work_queue so its length is below max at return 8193 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8194 while (_work_queue->size() > max) { 8195 oop new_oop; 8196 if (_work_queue->pop_local(new_oop)) { 8197 assert(new_oop->is_oop(), "Expected an oop"); 8198 assert(_bit_map->isMarked((HeapWord*)new_oop), 8199 "no white objects on this stack!"); 8200 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8201 // iterate over the oops in this oop, marking and pushing 8202 // the ones in CMS heap (i.e. in _span). 8203 new_oop->oop_iterate(&_mark_and_push); 8204 } 8205 } 8206 } 8207 8208 //////////////////////////////////////////////////////////////////// 8209 // Support for Marking Stack Overflow list handling and related code 8210 //////////////////////////////////////////////////////////////////// 8211 // Much of the following code is similar in shape and spirit to the 8212 // code used in ParNewGC. We should try and share that code 8213 // as much as possible in the future. 8214 8215 #ifndef PRODUCT 8216 // Debugging support for CMSStackOverflowALot 8217 8218 // It's OK to call this multi-threaded; the worst thing 8219 // that can happen is that we'll get a bunch of closely 8220 // spaced simulated overflows, but that's OK, in fact 8221 // probably good as it would exercise the overflow code 8222 // under contention. 8223 bool CMSCollector::simulate_overflow() { 8224 if (_overflow_counter-- <= 0) { // just being defensive 8225 _overflow_counter = CMSMarkStackOverflowInterval; 8226 return true; 8227 } else { 8228 return false; 8229 } 8230 } 8231 8232 bool CMSCollector::par_simulate_overflow() { 8233 return simulate_overflow(); 8234 } 8235 #endif 8236 8237 // Single-threaded 8238 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8239 assert(stack->isEmpty(), "Expected precondition"); 8240 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8241 size_t i = num; 8242 oop cur = _overflow_list; 8243 const markOop proto = markOopDesc::prototype(); 8244 NOT_PRODUCT(ssize_t n = 0;) 8245 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8246 next = oop(cur->mark()); 8247 cur->set_mark(proto); // until proven otherwise 8248 assert(cur->is_oop(), "Should be an oop"); 8249 bool res = stack->push(cur); 8250 assert(res, "Bit off more than can chew?"); 8251 NOT_PRODUCT(n++;) 8252 } 8253 _overflow_list = cur; 8254 #ifndef PRODUCT 8255 assert(_num_par_pushes >= n, "Too many pops?"); 8256 _num_par_pushes -=n; 8257 #endif 8258 return !stack->isEmpty(); 8259 } 8260 8261 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8262 // (MT-safe) Get a prefix of at most "num" from the list. 8263 // The overflow list is chained through the mark word of 8264 // each object in the list. We fetch the entire list, 8265 // break off a prefix of the right size and return the 8266 // remainder. If other threads try to take objects from 8267 // the overflow list at that time, they will wait for 8268 // some time to see if data becomes available. If (and 8269 // only if) another thread places one or more object(s) 8270 // on the global list before we have returned the suffix 8271 // to the global list, we will walk down our local list 8272 // to find its end and append the global list to 8273 // our suffix before returning it. This suffix walk can 8274 // prove to be expensive (quadratic in the amount of traffic) 8275 // when there are many objects in the overflow list and 8276 // there is much producer-consumer contention on the list. 8277 // *NOTE*: The overflow list manipulation code here and 8278 // in ParNewGeneration:: are very similar in shape, 8279 // except that in the ParNew case we use the old (from/eden) 8280 // copy of the object to thread the list via its klass word. 8281 // Because of the common code, if you make any changes in 8282 // the code below, please check the ParNew version to see if 8283 // similar changes might be needed. 8284 // CR 6797058 has been filed to consolidate the common code. 8285 bool CMSCollector::par_take_from_overflow_list(size_t num, 8286 OopTaskQueue* work_q, 8287 int no_of_gc_threads) { 8288 assert(work_q->size() == 0, "First empty local work queue"); 8289 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8290 if (_overflow_list == NULL) { 8291 return false; 8292 } 8293 // Grab the entire list; we'll put back a suffix 8294 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8295 Thread* tid = Thread::current(); 8296 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8297 // set to ParallelGCThreads. 8298 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8299 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8300 // If the list is busy, we spin for a short while, 8301 // sleeping between attempts to get the list. 8302 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8303 os::sleep(tid, sleep_time_millis, false); 8304 if (_overflow_list == NULL) { 8305 // Nothing left to take 8306 return false; 8307 } else if (_overflow_list != BUSY) { 8308 // Try and grab the prefix 8309 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8310 } 8311 } 8312 // If the list was found to be empty, or we spun long 8313 // enough, we give up and return empty-handed. If we leave 8314 // the list in the BUSY state below, it must be the case that 8315 // some other thread holds the overflow list and will set it 8316 // to a non-BUSY state in the future. 8317 if (prefix == NULL || prefix == BUSY) { 8318 // Nothing to take or waited long enough 8319 if (prefix == NULL) { 8320 // Write back the NULL in case we overwrote it with BUSY above 8321 // and it is still the same value. 8322 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8323 } 8324 return false; 8325 } 8326 assert(prefix != NULL && prefix != BUSY, "Error"); 8327 size_t i = num; 8328 oop cur = prefix; 8329 // Walk down the first "num" objects, unless we reach the end. 8330 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8331 if (cur->mark() == NULL) { 8332 // We have "num" or fewer elements in the list, so there 8333 // is nothing to return to the global list. 8334 // Write back the NULL in lieu of the BUSY we wrote 8335 // above, if it is still the same value. 8336 if (_overflow_list == BUSY) { 8337 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8338 } 8339 } else { 8340 // Chop off the suffix and return it to the global list. 8341 assert(cur->mark() != BUSY, "Error"); 8342 oop suffix_head = cur->mark(); // suffix will be put back on global list 8343 cur->set_mark(NULL); // break off suffix 8344 // It's possible that the list is still in the empty(busy) state 8345 // we left it in a short while ago; in that case we may be 8346 // able to place back the suffix without incurring the cost 8347 // of a walk down the list. 8348 oop observed_overflow_list = _overflow_list; 8349 oop cur_overflow_list = observed_overflow_list; 8350 bool attached = false; 8351 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8352 observed_overflow_list = 8353 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8354 if (cur_overflow_list == observed_overflow_list) { 8355 attached = true; 8356 break; 8357 } else cur_overflow_list = observed_overflow_list; 8358 } 8359 if (!attached) { 8360 // Too bad, someone else sneaked in (at least) an element; we'll need 8361 // to do a splice. Find tail of suffix so we can prepend suffix to global 8362 // list. 8363 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8364 oop suffix_tail = cur; 8365 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8366 "Tautology"); 8367 observed_overflow_list = _overflow_list; 8368 do { 8369 cur_overflow_list = observed_overflow_list; 8370 if (cur_overflow_list != BUSY) { 8371 // Do the splice ... 8372 suffix_tail->set_mark(markOop(cur_overflow_list)); 8373 } else { // cur_overflow_list == BUSY 8374 suffix_tail->set_mark(NULL); 8375 } 8376 // ... and try to place spliced list back on overflow_list ... 8377 observed_overflow_list = 8378 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8379 } while (cur_overflow_list != observed_overflow_list); 8380 // ... until we have succeeded in doing so. 8381 } 8382 } 8383 8384 // Push the prefix elements on work_q 8385 assert(prefix != NULL, "control point invariant"); 8386 const markOop proto = markOopDesc::prototype(); 8387 oop next; 8388 NOT_PRODUCT(ssize_t n = 0;) 8389 for (cur = prefix; cur != NULL; cur = next) { 8390 next = oop(cur->mark()); 8391 cur->set_mark(proto); // until proven otherwise 8392 assert(cur->is_oop(), "Should be an oop"); 8393 bool res = work_q->push(cur); 8394 assert(res, "Bit off more than we can chew?"); 8395 NOT_PRODUCT(n++;) 8396 } 8397 #ifndef PRODUCT 8398 assert(_num_par_pushes >= n, "Too many pops?"); 8399 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8400 #endif 8401 return true; 8402 } 8403 8404 // Single-threaded 8405 void CMSCollector::push_on_overflow_list(oop p) { 8406 NOT_PRODUCT(_num_par_pushes++;) 8407 assert(p->is_oop(), "Not an oop"); 8408 preserve_mark_if_necessary(p); 8409 p->set_mark((markOop)_overflow_list); 8410 _overflow_list = p; 8411 } 8412 8413 // Multi-threaded; use CAS to prepend to overflow list 8414 void CMSCollector::par_push_on_overflow_list(oop p) { 8415 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8416 assert(p->is_oop(), "Not an oop"); 8417 par_preserve_mark_if_necessary(p); 8418 oop observed_overflow_list = _overflow_list; 8419 oop cur_overflow_list; 8420 do { 8421 cur_overflow_list = observed_overflow_list; 8422 if (cur_overflow_list != BUSY) { 8423 p->set_mark(markOop(cur_overflow_list)); 8424 } else { 8425 p->set_mark(NULL); 8426 } 8427 observed_overflow_list = 8428 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8429 } while (cur_overflow_list != observed_overflow_list); 8430 } 8431 #undef BUSY 8432 8433 // Single threaded 8434 // General Note on GrowableArray: pushes may silently fail 8435 // because we are (temporarily) out of C-heap for expanding 8436 // the stack. The problem is quite ubiquitous and affects 8437 // a lot of code in the JVM. The prudent thing for GrowableArray 8438 // to do (for now) is to exit with an error. However, that may 8439 // be too draconian in some cases because the caller may be 8440 // able to recover without much harm. For such cases, we 8441 // should probably introduce a "soft_push" method which returns 8442 // an indication of success or failure with the assumption that 8443 // the caller may be able to recover from a failure; code in 8444 // the VM can then be changed, incrementally, to deal with such 8445 // failures where possible, thus, incrementally hardening the VM 8446 // in such low resource situations. 8447 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8448 _preserved_oop_stack.push(p); 8449 _preserved_mark_stack.push(m); 8450 assert(m == p->mark(), "Mark word changed"); 8451 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8452 "bijection"); 8453 } 8454 8455 // Single threaded 8456 void CMSCollector::preserve_mark_if_necessary(oop p) { 8457 markOop m = p->mark(); 8458 if (m->must_be_preserved(p)) { 8459 preserve_mark_work(p, m); 8460 } 8461 } 8462 8463 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8464 markOop m = p->mark(); 8465 if (m->must_be_preserved(p)) { 8466 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8467 // Even though we read the mark word without holding 8468 // the lock, we are assured that it will not change 8469 // because we "own" this oop, so no other thread can 8470 // be trying to push it on the overflow list; see 8471 // the assertion in preserve_mark_work() that checks 8472 // that m == p->mark(). 8473 preserve_mark_work(p, m); 8474 } 8475 } 8476 8477 // We should be able to do this multi-threaded, 8478 // a chunk of stack being a task (this is 8479 // correct because each oop only ever appears 8480 // once in the overflow list. However, it's 8481 // not very easy to completely overlap this with 8482 // other operations, so will generally not be done 8483 // until all work's been completed. Because we 8484 // expect the preserved oop stack (set) to be small, 8485 // it's probably fine to do this single-threaded. 8486 // We can explore cleverer concurrent/overlapped/parallel 8487 // processing of preserved marks if we feel the 8488 // need for this in the future. Stack overflow should 8489 // be so rare in practice and, when it happens, its 8490 // effect on performance so great that this will 8491 // likely just be in the noise anyway. 8492 void CMSCollector::restore_preserved_marks_if_any() { 8493 assert(SafepointSynchronize::is_at_safepoint(), 8494 "world should be stopped"); 8495 assert(Thread::current()->is_ConcurrentGC_thread() || 8496 Thread::current()->is_VM_thread(), 8497 "should be single-threaded"); 8498 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8499 "bijection"); 8500 8501 while (!_preserved_oop_stack.is_empty()) { 8502 oop p = _preserved_oop_stack.pop(); 8503 assert(p->is_oop(), "Should be an oop"); 8504 assert(_span.contains(p), "oop should be in _span"); 8505 assert(p->mark() == markOopDesc::prototype(), 8506 "Set when taken from overflow list"); 8507 markOop m = _preserved_mark_stack.pop(); 8508 p->set_mark(m); 8509 } 8510 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8511 "stacks were cleared above"); 8512 } 8513 8514 #ifndef PRODUCT 8515 bool CMSCollector::no_preserved_marks() const { 8516 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8517 } 8518 #endif 8519 8520 // Transfer some number of overflown objects to usual marking 8521 // stack. Return true if some objects were transferred. 8522 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8523 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8524 (size_t)ParGCDesiredObjsFromOverflowList); 8525 8526 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8527 assert(_collector->overflow_list_is_empty() || res, 8528 "If list is not empty, we should have taken something"); 8529 assert(!res || !_mark_stack->isEmpty(), 8530 "If we took something, it should now be on our stack"); 8531 return res; 8532 } 8533 8534 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8535 size_t res = _sp->block_size_no_stall(addr, _collector); 8536 if (_sp->block_is_obj(addr)) { 8537 if (_live_bit_map->isMarked(addr)) { 8538 // It can't have been dead in a previous cycle 8539 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8540 } else { 8541 _dead_bit_map->mark(addr); // mark the dead object 8542 } 8543 } 8544 // Could be 0, if the block size could not be computed without stalling. 8545 return res; 8546 } 8547 8548 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8549 8550 switch (phase) { 8551 case CMSCollector::InitialMarking: 8552 initialize(true /* fullGC */ , 8553 cause /* cause of the GC */, 8554 true /* recordGCBeginTime */, 8555 true /* recordPreGCUsage */, 8556 false /* recordPeakUsage */, 8557 false /* recordPostGCusage */, 8558 true /* recordAccumulatedGCTime */, 8559 false /* recordGCEndTime */, 8560 false /* countCollection */ ); 8561 break; 8562 8563 case CMSCollector::FinalMarking: 8564 initialize(true /* fullGC */ , 8565 cause /* cause of the GC */, 8566 false /* recordGCBeginTime */, 8567 false /* recordPreGCUsage */, 8568 false /* recordPeakUsage */, 8569 false /* recordPostGCusage */, 8570 true /* recordAccumulatedGCTime */, 8571 false /* recordGCEndTime */, 8572 false /* countCollection */ ); 8573 break; 8574 8575 case CMSCollector::Sweeping: 8576 initialize(true /* fullGC */ , 8577 cause /* cause of the GC */, 8578 false /* recordGCBeginTime */, 8579 false /* recordPreGCUsage */, 8580 true /* recordPeakUsage */, 8581 true /* recordPostGCusage */, 8582 false /* recordAccumulatedGCTime */, 8583 true /* recordGCEndTime */, 8584 true /* countCollection */ ); 8585 break; 8586 8587 default: 8588 ShouldNotReachHere(); 8589 } 8590 }