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