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