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, gc_tracer->gc_id()); 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 const GCId gc_id, 2829 bool print_cr = true); 2830 ~CMSPhaseAccounting(); 2831 2832 private: 2833 CMSCollector *_collector; 2834 const char *_phase; 2835 elapsedTimer _wallclock; 2836 bool _print_cr; 2837 const GCId _gc_id; 2838 2839 public: 2840 // Not MT-safe; so do not pass around these StackObj's 2841 // where they may be accessed by other threads. 2842 jlong wallclock_millis() { 2843 assert(_wallclock.is_active(), "Wall clock should not stop"); 2844 _wallclock.stop(); // to record time 2845 jlong ret = _wallclock.milliseconds(); 2846 _wallclock.start(); // restart 2847 return ret; 2848 } 2849 }; 2850 2851 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2852 const char *phase, 2853 const GCId gc_id, 2854 bool print_cr) : 2855 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2856 2857 if (PrintCMSStatistics != 0) { 2858 _collector->resetYields(); 2859 } 2860 if (PrintGCDetails) { 2861 gclog_or_tty->gclog_stamp(_gc_id); 2862 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2863 _collector->cmsGen()->short_name(), _phase); 2864 } 2865 _collector->resetTimer(); 2866 _wallclock.start(); 2867 _collector->startTimer(); 2868 } 2869 2870 CMSPhaseAccounting::~CMSPhaseAccounting() { 2871 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2872 _collector->stopTimer(); 2873 _wallclock.stop(); 2874 if (PrintGCDetails) { 2875 gclog_or_tty->gclog_stamp(_gc_id); 2876 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2877 _collector->cmsGen()->short_name(), 2878 _phase, _collector->timerValue(), _wallclock.seconds()); 2879 if (_print_cr) { 2880 gclog_or_tty->cr(); 2881 } 2882 if (PrintCMSStatistics != 0) { 2883 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2884 _collector->yields()); 2885 } 2886 } 2887 } 2888 2889 // CMS work 2890 2891 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2892 class CMSParMarkTask : public AbstractGangTask { 2893 protected: 2894 CMSCollector* _collector; 2895 uint _n_workers; 2896 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2897 AbstractGangTask(name), 2898 _collector(collector), 2899 _n_workers(n_workers) {} 2900 // Work method in support of parallel rescan ... of young gen spaces 2901 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2902 ContiguousSpace* space, 2903 HeapWord** chunk_array, size_t chunk_top); 2904 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2905 }; 2906 2907 // Parallel initial mark task 2908 class CMSParInitialMarkTask: public CMSParMarkTask { 2909 StrongRootsScope* _strong_roots_scope; 2910 public: 2911 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2912 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2913 _strong_roots_scope(strong_roots_scope) {} 2914 void work(uint worker_id); 2915 }; 2916 2917 // Checkpoint the roots into this generation from outside 2918 // this generation. [Note this initial checkpoint need only 2919 // be approximate -- we'll do a catch up phase subsequently.] 2920 void CMSCollector::checkpointRootsInitial() { 2921 assert(_collectorState == InitialMarking, "Wrong collector state"); 2922 check_correct_thread_executing(); 2923 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2924 2925 save_heap_summary(); 2926 report_heap_summary(GCWhen::BeforeGC); 2927 2928 ReferenceProcessor* rp = ref_processor(); 2929 assert(_restart_addr == NULL, "Control point invariant"); 2930 { 2931 // acquire locks for subsequent manipulations 2932 MutexLockerEx x(bitMapLock(), 2933 Mutex::_no_safepoint_check_flag); 2934 checkpointRootsInitialWork(); 2935 // enable ("weak") refs discovery 2936 rp->enable_discovery(); 2937 _collectorState = Marking; 2938 } 2939 } 2940 2941 void CMSCollector::checkpointRootsInitialWork() { 2942 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2943 assert(_collectorState == InitialMarking, "just checking"); 2944 2945 // Already have locks. 2946 assert_lock_strong(bitMapLock()); 2947 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2948 2949 // Setup the verification and class unloading state for this 2950 // CMS collection cycle. 2951 setup_cms_unloading_and_verification_state(); 2952 2953 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2954 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 2955 2956 // Reset all the PLAB chunk arrays if necessary. 2957 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2958 reset_survivor_plab_arrays(); 2959 } 2960 2961 ResourceMark rm; 2962 HandleMark hm; 2963 2964 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2965 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2966 2967 verify_work_stacks_empty(); 2968 verify_overflow_empty(); 2969 2970 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2971 // Update the saved marks which may affect the root scans. 2972 gch->save_marks(); 2973 2974 // weak reference processing has not started yet. 2975 ref_processor()->set_enqueuing_is_done(false); 2976 2977 // Need to remember all newly created CLDs, 2978 // so that we can guarantee that the remark finds them. 2979 ClassLoaderDataGraph::remember_new_clds(true); 2980 2981 // Whenever a CLD is found, it will be claimed before proceeding to mark 2982 // the klasses. The claimed marks need to be cleared before marking starts. 2983 ClassLoaderDataGraph::clear_claimed_marks(); 2984 2985 if (CMSPrintEdenSurvivorChunks) { 2986 print_eden_and_survivor_chunk_arrays(); 2987 } 2988 2989 { 2990 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2991 if (CMSParallelInitialMarkEnabled) { 2992 // The parallel version. 2993 WorkGang* workers = gch->workers(); 2994 assert(workers != NULL, "Need parallel worker threads."); 2995 uint n_workers = workers->active_workers(); 2996 2997 StrongRootsScope srs(n_workers); 2998 2999 CMSParInitialMarkTask tsk(this, &srs, n_workers); 3000 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3001 if (n_workers > 1) { 3002 workers->run_task(&tsk); 3003 } else { 3004 tsk.work(0); 3005 } 3006 } else { 3007 // The serial version. 3008 CLDToOopClosure cld_closure(¬Older, true); 3009 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3010 3011 StrongRootsScope srs(1); 3012 3013 gch->gen_process_roots(&srs, 3014 GenCollectedHeap::OldGen, 3015 true, // young gen as roots 3016 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3017 should_unload_classes(), 3018 ¬Older, 3019 NULL, 3020 &cld_closure); 3021 } 3022 } 3023 3024 // Clear mod-union table; it will be dirtied in the prologue of 3025 // CMS generation per each young generation collection. 3026 3027 assert(_modUnionTable.isAllClear(), 3028 "Was cleared in most recent final checkpoint phase" 3029 " or no bits are set in the gc_prologue before the start of the next " 3030 "subsequent marking phase."); 3031 3032 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3033 3034 // Save the end of the used_region of the constituent generations 3035 // to be used to limit the extent of sweep in each generation. 3036 save_sweep_limits(); 3037 verify_overflow_empty(); 3038 } 3039 3040 bool CMSCollector::markFromRoots() { 3041 // we might be tempted to assert that: 3042 // assert(!SafepointSynchronize::is_at_safepoint(), 3043 // "inconsistent argument?"); 3044 // However that wouldn't be right, because it's possible that 3045 // a safepoint is indeed in progress as a young generation 3046 // stop-the-world GC happens even as we mark in this generation. 3047 assert(_collectorState == Marking, "inconsistent state?"); 3048 check_correct_thread_executing(); 3049 verify_overflow_empty(); 3050 3051 // Weak ref discovery note: We may be discovering weak 3052 // refs in this generation concurrent (but interleaved) with 3053 // weak ref discovery by the young generation collector. 3054 3055 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3056 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3057 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3058 bool res = markFromRootsWork(); 3059 if (res) { 3060 _collectorState = Precleaning; 3061 } else { // We failed and a foreground collection wants to take over 3062 assert(_foregroundGCIsActive, "internal state inconsistency"); 3063 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3064 if (PrintGCDetails) { 3065 gclog_or_tty->print_cr("bailing out to foreground collection"); 3066 } 3067 } 3068 verify_overflow_empty(); 3069 return res; 3070 } 3071 3072 bool CMSCollector::markFromRootsWork() { 3073 // iterate over marked bits in bit map, doing a full scan and mark 3074 // from these roots using the following algorithm: 3075 // . if oop is to the right of the current scan pointer, 3076 // mark corresponding bit (we'll process it later) 3077 // . else (oop is to left of current scan pointer) 3078 // push oop on marking stack 3079 // . drain the marking stack 3080 3081 // Note that when we do a marking step we need to hold the 3082 // bit map lock -- recall that direct allocation (by mutators) 3083 // and promotion (by the young generation collector) is also 3084 // marking the bit map. [the so-called allocate live policy.] 3085 // Because the implementation of bit map marking is not 3086 // robust wrt simultaneous marking of bits in the same word, 3087 // we need to make sure that there is no such interference 3088 // between concurrent such updates. 3089 3090 // already have locks 3091 assert_lock_strong(bitMapLock()); 3092 3093 verify_work_stacks_empty(); 3094 verify_overflow_empty(); 3095 bool result = false; 3096 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3097 result = do_marking_mt(); 3098 } else { 3099 result = do_marking_st(); 3100 } 3101 return result; 3102 } 3103 3104 // Forward decl 3105 class CMSConcMarkingTask; 3106 3107 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3108 CMSCollector* _collector; 3109 CMSConcMarkingTask* _task; 3110 public: 3111 virtual void yield(); 3112 3113 // "n_threads" is the number of threads to be terminated. 3114 // "queue_set" is a set of work queues of other threads. 3115 // "collector" is the CMS collector associated with this task terminator. 3116 // "yield" indicates whether we need the gang as a whole to yield. 3117 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3118 ParallelTaskTerminator(n_threads, queue_set), 3119 _collector(collector) { } 3120 3121 void set_task(CMSConcMarkingTask* task) { 3122 _task = task; 3123 } 3124 }; 3125 3126 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3127 CMSConcMarkingTask* _task; 3128 public: 3129 bool should_exit_termination(); 3130 void set_task(CMSConcMarkingTask* task) { 3131 _task = task; 3132 } 3133 }; 3134 3135 // MT Concurrent Marking Task 3136 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3137 CMSCollector* _collector; 3138 uint _n_workers; // requested/desired # workers 3139 bool _result; 3140 CompactibleFreeListSpace* _cms_space; 3141 char _pad_front[64]; // padding to ... 3142 HeapWord* _global_finger; // ... avoid sharing cache line 3143 char _pad_back[64]; 3144 HeapWord* _restart_addr; 3145 3146 // Exposed here for yielding support 3147 Mutex* const _bit_map_lock; 3148 3149 // The per thread work queues, available here for stealing 3150 OopTaskQueueSet* _task_queues; 3151 3152 // Termination (and yielding) support 3153 CMSConcMarkingTerminator _term; 3154 CMSConcMarkingTerminatorTerminator _term_term; 3155 3156 public: 3157 CMSConcMarkingTask(CMSCollector* collector, 3158 CompactibleFreeListSpace* cms_space, 3159 YieldingFlexibleWorkGang* workers, 3160 OopTaskQueueSet* task_queues): 3161 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3162 _collector(collector), 3163 _cms_space(cms_space), 3164 _n_workers(0), _result(true), 3165 _task_queues(task_queues), 3166 _term(_n_workers, task_queues, _collector), 3167 _bit_map_lock(collector->bitMapLock()) 3168 { 3169 _requested_size = _n_workers; 3170 _term.set_task(this); 3171 _term_term.set_task(this); 3172 _restart_addr = _global_finger = _cms_space->bottom(); 3173 } 3174 3175 3176 OopTaskQueueSet* task_queues() { return _task_queues; } 3177 3178 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3179 3180 HeapWord** global_finger_addr() { return &_global_finger; } 3181 3182 CMSConcMarkingTerminator* terminator() { return &_term; } 3183 3184 virtual void set_for_termination(uint active_workers) { 3185 terminator()->reset_for_reuse(active_workers); 3186 } 3187 3188 void work(uint worker_id); 3189 bool should_yield() { 3190 return ConcurrentMarkSweepThread::should_yield() 3191 && !_collector->foregroundGCIsActive(); 3192 } 3193 3194 virtual void coordinator_yield(); // stuff done by coordinator 3195 bool result() { return _result; } 3196 3197 void reset(HeapWord* ra) { 3198 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3199 _restart_addr = _global_finger = ra; 3200 _term.reset_for_reuse(); 3201 } 3202 3203 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3204 OopTaskQueue* work_q); 3205 3206 private: 3207 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3208 void do_work_steal(int i); 3209 void bump_global_finger(HeapWord* f); 3210 }; 3211 3212 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3213 assert(_task != NULL, "Error"); 3214 return _task->yielding(); 3215 // Note that we do not need the disjunct || _task->should_yield() above 3216 // because we want terminating threads to yield only if the task 3217 // is already in the midst of yielding, which happens only after at least one 3218 // thread has yielded. 3219 } 3220 3221 void CMSConcMarkingTerminator::yield() { 3222 if (_task->should_yield()) { 3223 _task->yield(); 3224 } else { 3225 ParallelTaskTerminator::yield(); 3226 } 3227 } 3228 3229 //////////////////////////////////////////////////////////////// 3230 // Concurrent Marking Algorithm Sketch 3231 //////////////////////////////////////////////////////////////// 3232 // Until all tasks exhausted (both spaces): 3233 // -- claim next available chunk 3234 // -- bump global finger via CAS 3235 // -- find first object that starts in this chunk 3236 // and start scanning bitmap from that position 3237 // -- scan marked objects for oops 3238 // -- CAS-mark target, and if successful: 3239 // . if target oop is above global finger (volatile read) 3240 // nothing to do 3241 // . if target oop is in chunk and above local finger 3242 // then nothing to do 3243 // . else push on work-queue 3244 // -- Deal with possible overflow issues: 3245 // . local work-queue overflow causes stuff to be pushed on 3246 // global (common) overflow queue 3247 // . always first empty local work queue 3248 // . then get a batch of oops from global work queue if any 3249 // . then do work stealing 3250 // -- When all tasks claimed (both spaces) 3251 // and local work queue empty, 3252 // then in a loop do: 3253 // . check global overflow stack; steal a batch of oops and trace 3254 // . try to steal from other threads oif GOS is empty 3255 // . if neither is available, offer termination 3256 // -- Terminate and return result 3257 // 3258 void CMSConcMarkingTask::work(uint worker_id) { 3259 elapsedTimer _timer; 3260 ResourceMark rm; 3261 HandleMark hm; 3262 3263 DEBUG_ONLY(_collector->verify_overflow_empty();) 3264 3265 // Before we begin work, our work queue should be empty 3266 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3267 // Scan the bitmap covering _cms_space, tracing through grey objects. 3268 _timer.start(); 3269 do_scan_and_mark(worker_id, _cms_space); 3270 _timer.stop(); 3271 if (PrintCMSStatistics != 0) { 3272 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3273 worker_id, _timer.seconds()); 3274 // XXX: need xxx/xxx type of notation, two timers 3275 } 3276 3277 // ... do work stealing 3278 _timer.reset(); 3279 _timer.start(); 3280 do_work_steal(worker_id); 3281 _timer.stop(); 3282 if (PrintCMSStatistics != 0) { 3283 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3284 worker_id, _timer.seconds()); 3285 // XXX: need xxx/xxx type of notation, two timers 3286 } 3287 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3288 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3289 // Note that under the current task protocol, the 3290 // following assertion is true even of the spaces 3291 // expanded since the completion of the concurrent 3292 // marking. XXX This will likely change under a strict 3293 // ABORT semantics. 3294 // After perm removal the comparison was changed to 3295 // greater than or equal to from strictly greater than. 3296 // Before perm removal the highest address sweep would 3297 // have been at the end of perm gen but now is at the 3298 // end of the tenured gen. 3299 assert(_global_finger >= _cms_space->end(), 3300 "All tasks have been completed"); 3301 DEBUG_ONLY(_collector->verify_overflow_empty();) 3302 } 3303 3304 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3305 HeapWord* read = _global_finger; 3306 HeapWord* cur = read; 3307 while (f > read) { 3308 cur = read; 3309 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3310 if (cur == read) { 3311 // our cas succeeded 3312 assert(_global_finger >= f, "protocol consistency"); 3313 break; 3314 } 3315 } 3316 } 3317 3318 // This is really inefficient, and should be redone by 3319 // using (not yet available) block-read and -write interfaces to the 3320 // stack and the work_queue. XXX FIX ME !!! 3321 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3322 OopTaskQueue* work_q) { 3323 // Fast lock-free check 3324 if (ovflw_stk->length() == 0) { 3325 return false; 3326 } 3327 assert(work_q->size() == 0, "Shouldn't steal"); 3328 MutexLockerEx ml(ovflw_stk->par_lock(), 3329 Mutex::_no_safepoint_check_flag); 3330 // Grab up to 1/4 the size of the work queue 3331 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3332 (size_t)ParGCDesiredObjsFromOverflowList); 3333 num = MIN2(num, ovflw_stk->length()); 3334 for (int i = (int) num; i > 0; i--) { 3335 oop cur = ovflw_stk->pop(); 3336 assert(cur != NULL, "Counted wrong?"); 3337 work_q->push(cur); 3338 } 3339 return num > 0; 3340 } 3341 3342 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3343 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3344 int n_tasks = pst->n_tasks(); 3345 // We allow that there may be no tasks to do here because 3346 // we are restarting after a stack overflow. 3347 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3348 uint nth_task = 0; 3349 3350 HeapWord* aligned_start = sp->bottom(); 3351 if (sp->used_region().contains(_restart_addr)) { 3352 // Align down to a card boundary for the start of 0th task 3353 // for this space. 3354 aligned_start = 3355 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3356 CardTableModRefBS::card_size); 3357 } 3358 3359 size_t chunk_size = sp->marking_task_size(); 3360 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3361 // Having claimed the nth task in this space, 3362 // compute the chunk that it corresponds to: 3363 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3364 aligned_start + (nth_task+1)*chunk_size); 3365 // Try and bump the global finger via a CAS; 3366 // note that we need to do the global finger bump 3367 // _before_ taking the intersection below, because 3368 // the task corresponding to that region will be 3369 // deemed done even if the used_region() expands 3370 // because of allocation -- as it almost certainly will 3371 // during start-up while the threads yield in the 3372 // closure below. 3373 HeapWord* finger = span.end(); 3374 bump_global_finger(finger); // atomically 3375 // There are null tasks here corresponding to chunks 3376 // beyond the "top" address of the space. 3377 span = span.intersection(sp->used_region()); 3378 if (!span.is_empty()) { // Non-null task 3379 HeapWord* prev_obj; 3380 assert(!span.contains(_restart_addr) || nth_task == 0, 3381 "Inconsistency"); 3382 if (nth_task == 0) { 3383 // For the 0th task, we'll not need to compute a block_start. 3384 if (span.contains(_restart_addr)) { 3385 // In the case of a restart because of stack overflow, 3386 // we might additionally skip a chunk prefix. 3387 prev_obj = _restart_addr; 3388 } else { 3389 prev_obj = span.start(); 3390 } 3391 } else { 3392 // We want to skip the first object because 3393 // the protocol is to scan any object in its entirety 3394 // that _starts_ in this span; a fortiori, any 3395 // object starting in an earlier span is scanned 3396 // as part of an earlier claimed task. 3397 // Below we use the "careful" version of block_start 3398 // so we do not try to navigate uninitialized objects. 3399 prev_obj = sp->block_start_careful(span.start()); 3400 // Below we use a variant of block_size that uses the 3401 // Printezis bits to avoid waiting for allocated 3402 // objects to become initialized/parsable. 3403 while (prev_obj < span.start()) { 3404 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3405 if (sz > 0) { 3406 prev_obj += sz; 3407 } else { 3408 // In this case we may end up doing a bit of redundant 3409 // scanning, but that appears unavoidable, short of 3410 // locking the free list locks; see bug 6324141. 3411 break; 3412 } 3413 } 3414 } 3415 if (prev_obj < span.end()) { 3416 MemRegion my_span = MemRegion(prev_obj, span.end()); 3417 // Do the marking work within a non-empty span -- 3418 // the last argument to the constructor indicates whether the 3419 // iteration should be incremental with periodic yields. 3420 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3421 &_collector->_markBitMap, 3422 work_queue(i), 3423 &_collector->_markStack); 3424 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3425 } // else nothing to do for this task 3426 } // else nothing to do for this task 3427 } 3428 // We'd be tempted to assert here that since there are no 3429 // more tasks left to claim in this space, the global_finger 3430 // must exceed space->top() and a fortiori space->end(). However, 3431 // that would not quite be correct because the bumping of 3432 // global_finger occurs strictly after the claiming of a task, 3433 // so by the time we reach here the global finger may not yet 3434 // have been bumped up by the thread that claimed the last 3435 // task. 3436 pst->all_tasks_completed(); 3437 } 3438 3439 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3440 private: 3441 CMSCollector* _collector; 3442 CMSConcMarkingTask* _task; 3443 MemRegion _span; 3444 CMSBitMap* _bit_map; 3445 CMSMarkStack* _overflow_stack; 3446 OopTaskQueue* _work_queue; 3447 protected: 3448 DO_OOP_WORK_DEFN 3449 public: 3450 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3451 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3452 MetadataAwareOopClosure(collector->ref_processor()), 3453 _collector(collector), 3454 _task(task), 3455 _span(collector->_span), 3456 _work_queue(work_queue), 3457 _bit_map(bit_map), 3458 _overflow_stack(overflow_stack) 3459 { } 3460 virtual void do_oop(oop* p); 3461 virtual void do_oop(narrowOop* p); 3462 3463 void trim_queue(size_t max); 3464 void handle_stack_overflow(HeapWord* lost); 3465 void do_yield_check() { 3466 if (_task->should_yield()) { 3467 _task->yield(); 3468 } 3469 } 3470 }; 3471 3472 // Grey object scanning during work stealing phase -- 3473 // the salient assumption here is that any references 3474 // that are in these stolen objects being scanned must 3475 // already have been initialized (else they would not have 3476 // been published), so we do not need to check for 3477 // uninitialized objects before pushing here. 3478 void Par_ConcMarkingClosure::do_oop(oop obj) { 3479 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3480 HeapWord* addr = (HeapWord*)obj; 3481 // Check if oop points into the CMS generation 3482 // and is not marked 3483 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3484 // a white object ... 3485 // If we manage to "claim" the object, by being the 3486 // first thread to mark it, then we push it on our 3487 // marking stack 3488 if (_bit_map->par_mark(addr)) { // ... now grey 3489 // push on work queue (grey set) 3490 bool simulate_overflow = false; 3491 NOT_PRODUCT( 3492 if (CMSMarkStackOverflowALot && 3493 _collector->simulate_overflow()) { 3494 // simulate a stack overflow 3495 simulate_overflow = true; 3496 } 3497 ) 3498 if (simulate_overflow || 3499 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3500 // stack overflow 3501 if (PrintCMSStatistics != 0) { 3502 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3503 SIZE_FORMAT, _overflow_stack->capacity()); 3504 } 3505 // We cannot assert that the overflow stack is full because 3506 // it may have been emptied since. 3507 assert(simulate_overflow || 3508 _work_queue->size() == _work_queue->max_elems(), 3509 "Else push should have succeeded"); 3510 handle_stack_overflow(addr); 3511 } 3512 } // Else, some other thread got there first 3513 do_yield_check(); 3514 } 3515 } 3516 3517 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3518 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3519 3520 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3521 while (_work_queue->size() > max) { 3522 oop new_oop; 3523 if (_work_queue->pop_local(new_oop)) { 3524 assert(new_oop->is_oop(), "Should be an oop"); 3525 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3526 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3527 new_oop->oop_iterate(this); // do_oop() above 3528 do_yield_check(); 3529 } 3530 } 3531 } 3532 3533 // Upon stack overflow, we discard (part of) the stack, 3534 // remembering the least address amongst those discarded 3535 // in CMSCollector's _restart_address. 3536 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3537 // We need to do this under a mutex to prevent other 3538 // workers from interfering with the work done below. 3539 MutexLockerEx ml(_overflow_stack->par_lock(), 3540 Mutex::_no_safepoint_check_flag); 3541 // Remember the least grey address discarded 3542 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3543 _collector->lower_restart_addr(ra); 3544 _overflow_stack->reset(); // discard stack contents 3545 _overflow_stack->expand(); // expand the stack if possible 3546 } 3547 3548 3549 void CMSConcMarkingTask::do_work_steal(int i) { 3550 OopTaskQueue* work_q = work_queue(i); 3551 oop obj_to_scan; 3552 CMSBitMap* bm = &(_collector->_markBitMap); 3553 CMSMarkStack* ovflw = &(_collector->_markStack); 3554 int* seed = _collector->hash_seed(i); 3555 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3556 while (true) { 3557 cl.trim_queue(0); 3558 assert(work_q->size() == 0, "Should have been emptied above"); 3559 if (get_work_from_overflow_stack(ovflw, work_q)) { 3560 // Can't assert below because the work obtained from the 3561 // overflow stack may already have been stolen from us. 3562 // assert(work_q->size() > 0, "Work from overflow stack"); 3563 continue; 3564 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3565 assert(obj_to_scan->is_oop(), "Should be an oop"); 3566 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3567 obj_to_scan->oop_iterate(&cl); 3568 } else if (terminator()->offer_termination(&_term_term)) { 3569 assert(work_q->size() == 0, "Impossible!"); 3570 break; 3571 } else if (yielding() || should_yield()) { 3572 yield(); 3573 } 3574 } 3575 } 3576 3577 // This is run by the CMS (coordinator) thread. 3578 void CMSConcMarkingTask::coordinator_yield() { 3579 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3580 "CMS thread should hold CMS token"); 3581 // First give up the locks, then yield, then re-lock 3582 // We should probably use a constructor/destructor idiom to 3583 // do this unlock/lock or modify the MutexUnlocker class to 3584 // serve our purpose. XXX 3585 assert_lock_strong(_bit_map_lock); 3586 _bit_map_lock->unlock(); 3587 ConcurrentMarkSweepThread::desynchronize(true); 3588 _collector->stopTimer(); 3589 if (PrintCMSStatistics != 0) { 3590 _collector->incrementYields(); 3591 } 3592 3593 // It is possible for whichever thread initiated the yield request 3594 // not to get a chance to wake up and take the bitmap lock between 3595 // this thread releasing it and reacquiring it. So, while the 3596 // should_yield() flag is on, let's sleep for a bit to give the 3597 // other thread a chance to wake up. The limit imposed on the number 3598 // of iterations is defensive, to avoid any unforseen circumstances 3599 // putting us into an infinite loop. Since it's always been this 3600 // (coordinator_yield()) method that was observed to cause the 3601 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3602 // which is by default non-zero. For the other seven methods that 3603 // also perform the yield operation, as are using a different 3604 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3605 // can enable the sleeping for those methods too, if necessary. 3606 // See 6442774. 3607 // 3608 // We really need to reconsider the synchronization between the GC 3609 // thread and the yield-requesting threads in the future and we 3610 // should really use wait/notify, which is the recommended 3611 // way of doing this type of interaction. Additionally, we should 3612 // consolidate the eight methods that do the yield operation and they 3613 // are almost identical into one for better maintainability and 3614 // readability. See 6445193. 3615 // 3616 // Tony 2006.06.29 3617 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3618 ConcurrentMarkSweepThread::should_yield() && 3619 !CMSCollector::foregroundGCIsActive(); ++i) { 3620 os::sleep(Thread::current(), 1, false); 3621 } 3622 3623 ConcurrentMarkSweepThread::synchronize(true); 3624 _bit_map_lock->lock_without_safepoint_check(); 3625 _collector->startTimer(); 3626 } 3627 3628 bool CMSCollector::do_marking_mt() { 3629 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3630 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3631 conc_workers()->active_workers(), 3632 Threads::number_of_non_daemon_threads()); 3633 conc_workers()->set_active_workers(num_workers); 3634 3635 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3636 3637 CMSConcMarkingTask tsk(this, 3638 cms_space, 3639 conc_workers(), 3640 task_queues()); 3641 3642 // Since the actual number of workers we get may be different 3643 // from the number we requested above, do we need to do anything different 3644 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3645 // class?? XXX 3646 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3647 3648 // Refs discovery is already non-atomic. 3649 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3650 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3651 conc_workers()->start_task(&tsk); 3652 while (tsk.yielded()) { 3653 tsk.coordinator_yield(); 3654 conc_workers()->continue_task(&tsk); 3655 } 3656 // If the task was aborted, _restart_addr will be non-NULL 3657 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3658 while (_restart_addr != NULL) { 3659 // XXX For now we do not make use of ABORTED state and have not 3660 // yet implemented the right abort semantics (even in the original 3661 // single-threaded CMS case). That needs some more investigation 3662 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3663 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3664 // If _restart_addr is non-NULL, a marking stack overflow 3665 // occurred; we need to do a fresh marking iteration from the 3666 // indicated restart address. 3667 if (_foregroundGCIsActive) { 3668 // We may be running into repeated stack overflows, having 3669 // reached the limit of the stack size, while making very 3670 // slow forward progress. It may be best to bail out and 3671 // let the foreground collector do its job. 3672 // Clear _restart_addr, so that foreground GC 3673 // works from scratch. This avoids the headache of 3674 // a "rescan" which would otherwise be needed because 3675 // of the dirty mod union table & card table. 3676 _restart_addr = NULL; 3677 return false; 3678 } 3679 // Adjust the task to restart from _restart_addr 3680 tsk.reset(_restart_addr); 3681 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3682 _restart_addr); 3683 _restart_addr = NULL; 3684 // Get the workers going again 3685 conc_workers()->start_task(&tsk); 3686 while (tsk.yielded()) { 3687 tsk.coordinator_yield(); 3688 conc_workers()->continue_task(&tsk); 3689 } 3690 } 3691 assert(tsk.completed(), "Inconsistency"); 3692 assert(tsk.result() == true, "Inconsistency"); 3693 return true; 3694 } 3695 3696 bool CMSCollector::do_marking_st() { 3697 ResourceMark rm; 3698 HandleMark hm; 3699 3700 // Temporarily make refs discovery single threaded (non-MT) 3701 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3702 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3703 &_markStack, CMSYield); 3704 // the last argument to iterate indicates whether the iteration 3705 // should be incremental with periodic yields. 3706 _markBitMap.iterate(&markFromRootsClosure); 3707 // If _restart_addr is non-NULL, a marking stack overflow 3708 // occurred; we need to do a fresh iteration from the 3709 // indicated restart address. 3710 while (_restart_addr != NULL) { 3711 if (_foregroundGCIsActive) { 3712 // We may be running into repeated stack overflows, having 3713 // reached the limit of the stack size, while making very 3714 // slow forward progress. It may be best to bail out and 3715 // let the foreground collector do its job. 3716 // Clear _restart_addr, so that foreground GC 3717 // works from scratch. This avoids the headache of 3718 // a "rescan" which would otherwise be needed because 3719 // of the dirty mod union table & card table. 3720 _restart_addr = NULL; 3721 return false; // indicating failure to complete marking 3722 } 3723 // Deal with stack overflow: 3724 // we restart marking from _restart_addr 3725 HeapWord* ra = _restart_addr; 3726 markFromRootsClosure.reset(ra); 3727 _restart_addr = NULL; 3728 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3729 } 3730 return true; 3731 } 3732 3733 void CMSCollector::preclean() { 3734 check_correct_thread_executing(); 3735 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3736 verify_work_stacks_empty(); 3737 verify_overflow_empty(); 3738 _abort_preclean = false; 3739 if (CMSPrecleaningEnabled) { 3740 if (!CMSEdenChunksRecordAlways) { 3741 _eden_chunk_index = 0; 3742 } 3743 size_t used = get_eden_used(); 3744 size_t capacity = get_eden_capacity(); 3745 // Don't start sampling unless we will get sufficiently 3746 // many samples. 3747 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3748 * CMSScheduleRemarkEdenPenetration)) { 3749 _start_sampling = true; 3750 } else { 3751 _start_sampling = false; 3752 } 3753 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3754 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3755 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3756 } 3757 CMSTokenSync x(true); // is cms thread 3758 if (CMSPrecleaningEnabled) { 3759 sample_eden(); 3760 _collectorState = AbortablePreclean; 3761 } else { 3762 _collectorState = FinalMarking; 3763 } 3764 verify_work_stacks_empty(); 3765 verify_overflow_empty(); 3766 } 3767 3768 // Try and schedule the remark such that young gen 3769 // occupancy is CMSScheduleRemarkEdenPenetration %. 3770 void CMSCollector::abortable_preclean() { 3771 check_correct_thread_executing(); 3772 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3773 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3774 3775 // If Eden's current occupancy is below this threshold, 3776 // immediately schedule the remark; else preclean 3777 // past the next scavenge in an effort to 3778 // schedule the pause as described above. By choosing 3779 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3780 // we will never do an actual abortable preclean cycle. 3781 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3782 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3783 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3784 // We need more smarts in the abortable preclean 3785 // loop below to deal with cases where allocation 3786 // in young gen is very very slow, and our precleaning 3787 // is running a losing race against a horde of 3788 // mutators intent on flooding us with CMS updates 3789 // (dirty cards). 3790 // One, admittedly dumb, strategy is to give up 3791 // after a certain number of abortable precleaning loops 3792 // or after a certain maximum time. We want to make 3793 // this smarter in the next iteration. 3794 // XXX FIX ME!!! YSR 3795 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3796 while (!(should_abort_preclean() || 3797 ConcurrentMarkSweepThread::should_terminate())) { 3798 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3799 cumworkdone += workdone; 3800 loops++; 3801 // Voluntarily terminate abortable preclean phase if we have 3802 // been at it for too long. 3803 if ((CMSMaxAbortablePrecleanLoops != 0) && 3804 loops >= CMSMaxAbortablePrecleanLoops) { 3805 if (PrintGCDetails) { 3806 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3807 } 3808 break; 3809 } 3810 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3811 if (PrintGCDetails) { 3812 gclog_or_tty->print(" CMS: abort preclean due to time "); 3813 } 3814 break; 3815 } 3816 // If we are doing little work each iteration, we should 3817 // take a short break. 3818 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3819 // Sleep for some time, waiting for work to accumulate 3820 stopTimer(); 3821 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3822 startTimer(); 3823 waited++; 3824 } 3825 } 3826 if (PrintCMSStatistics > 0) { 3827 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3828 loops, waited, cumworkdone); 3829 } 3830 } 3831 CMSTokenSync x(true); // is cms thread 3832 if (_collectorState != Idling) { 3833 assert(_collectorState == AbortablePreclean, 3834 "Spontaneous state transition?"); 3835 _collectorState = FinalMarking; 3836 } // Else, a foreground collection completed this CMS cycle. 3837 return; 3838 } 3839 3840 // Respond to an Eden sampling opportunity 3841 void CMSCollector::sample_eden() { 3842 // Make sure a young gc cannot sneak in between our 3843 // reading and recording of a sample. 3844 assert(Thread::current()->is_ConcurrentGC_thread(), 3845 "Only the cms thread may collect Eden samples"); 3846 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3847 "Should collect samples while holding CMS token"); 3848 if (!_start_sampling) { 3849 return; 3850 } 3851 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3852 // is populated by the young generation. 3853 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3854 if (_eden_chunk_index < _eden_chunk_capacity) { 3855 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3856 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3857 "Unexpected state of Eden"); 3858 // We'd like to check that what we just sampled is an oop-start address; 3859 // however, we cannot do that here since the object may not yet have been 3860 // initialized. So we'll instead do the check when we _use_ this sample 3861 // later. 3862 if (_eden_chunk_index == 0 || 3863 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3864 _eden_chunk_array[_eden_chunk_index-1]) 3865 >= CMSSamplingGrain)) { 3866 _eden_chunk_index++; // commit sample 3867 } 3868 } 3869 } 3870 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3871 size_t used = get_eden_used(); 3872 size_t capacity = get_eden_capacity(); 3873 assert(used <= capacity, "Unexpected state of Eden"); 3874 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3875 _abort_preclean = true; 3876 } 3877 } 3878 } 3879 3880 3881 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3882 assert(_collectorState == Precleaning || 3883 _collectorState == AbortablePreclean, "incorrect state"); 3884 ResourceMark rm; 3885 HandleMark hm; 3886 3887 // Precleaning is currently not MT but the reference processor 3888 // may be set for MT. Disable it temporarily here. 3889 ReferenceProcessor* rp = ref_processor(); 3890 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3891 3892 // Do one pass of scrubbing the discovered reference lists 3893 // to remove any reference objects with strongly-reachable 3894 // referents. 3895 if (clean_refs) { 3896 CMSPrecleanRefsYieldClosure yield_cl(this); 3897 assert(rp->span().equals(_span), "Spans should be equal"); 3898 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3899 &_markStack, true /* preclean */); 3900 CMSDrainMarkingStackClosure complete_trace(this, 3901 _span, &_markBitMap, &_markStack, 3902 &keep_alive, true /* preclean */); 3903 3904 // We don't want this step to interfere with a young 3905 // collection because we don't want to take CPU 3906 // or memory bandwidth away from the young GC threads 3907 // (which may be as many as there are CPUs). 3908 // Note that we don't need to protect ourselves from 3909 // interference with mutators because they can't 3910 // manipulate the discovered reference lists nor affect 3911 // the computed reachability of the referents, the 3912 // only properties manipulated by the precleaning 3913 // of these reference lists. 3914 stopTimer(); 3915 CMSTokenSyncWithLocks x(true /* is cms thread */, 3916 bitMapLock()); 3917 startTimer(); 3918 sample_eden(); 3919 3920 // The following will yield to allow foreground 3921 // collection to proceed promptly. XXX YSR: 3922 // The code in this method may need further 3923 // tweaking for better performance and some restructuring 3924 // for cleaner interfaces. 3925 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3926 rp->preclean_discovered_references( 3927 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3928 gc_timer, _gc_tracer_cm->gc_id()); 3929 } 3930 3931 if (clean_survivor) { // preclean the active survivor space(s) 3932 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3933 &_markBitMap, &_modUnionTable, 3934 &_markStack, true /* precleaning phase */); 3935 stopTimer(); 3936 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3937 bitMapLock()); 3938 startTimer(); 3939 unsigned int before_count = 3940 GenCollectedHeap::heap()->total_collections(); 3941 SurvivorSpacePrecleanClosure 3942 sss_cl(this, _span, &_markBitMap, &_markStack, 3943 &pam_cl, before_count, CMSYield); 3944 _young_gen->from()->object_iterate_careful(&sss_cl); 3945 _young_gen->to()->object_iterate_careful(&sss_cl); 3946 } 3947 MarkRefsIntoAndScanClosure 3948 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3949 &_markStack, this, CMSYield, 3950 true /* precleaning phase */); 3951 // CAUTION: The following closure has persistent state that may need to 3952 // be reset upon a decrease in the sequence of addresses it 3953 // processes. 3954 ScanMarkedObjectsAgainCarefullyClosure 3955 smoac_cl(this, _span, 3956 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3957 3958 // Preclean dirty cards in ModUnionTable and CardTable using 3959 // appropriate convergence criterion; 3960 // repeat CMSPrecleanIter times unless we find that 3961 // we are losing. 3962 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3963 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3964 "Bad convergence multiplier"); 3965 assert(CMSPrecleanThreshold >= 100, 3966 "Unreasonably low CMSPrecleanThreshold"); 3967 3968 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3969 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3970 numIter < CMSPrecleanIter; 3971 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3972 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3973 if (Verbose && PrintGCDetails) { 3974 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3975 } 3976 // Either there are very few dirty cards, so re-mark 3977 // pause will be small anyway, or our pre-cleaning isn't 3978 // that much faster than the rate at which cards are being 3979 // dirtied, so we might as well stop and re-mark since 3980 // precleaning won't improve our re-mark time by much. 3981 if (curNumCards <= CMSPrecleanThreshold || 3982 (numIter > 0 && 3983 (curNumCards * CMSPrecleanDenominator > 3984 lastNumCards * CMSPrecleanNumerator))) { 3985 numIter++; 3986 cumNumCards += curNumCards; 3987 break; 3988 } 3989 } 3990 3991 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3992 3993 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3994 cumNumCards += curNumCards; 3995 if (PrintGCDetails && PrintCMSStatistics != 0) { 3996 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3997 curNumCards, cumNumCards, numIter); 3998 } 3999 return cumNumCards; // as a measure of useful work done 4000 } 4001 4002 // PRECLEANING NOTES: 4003 // Precleaning involves: 4004 // . reading the bits of the modUnionTable and clearing the set bits. 4005 // . For the cards corresponding to the set bits, we scan the 4006 // objects on those cards. This means we need the free_list_lock 4007 // so that we can safely iterate over the CMS space when scanning 4008 // for oops. 4009 // . When we scan the objects, we'll be both reading and setting 4010 // marks in the marking bit map, so we'll need the marking bit map. 4011 // . For protecting _collector_state transitions, we take the CGC_lock. 4012 // Note that any races in the reading of of card table entries by the 4013 // CMS thread on the one hand and the clearing of those entries by the 4014 // VM thread or the setting of those entries by the mutator threads on the 4015 // other are quite benign. However, for efficiency it makes sense to keep 4016 // the VM thread from racing with the CMS thread while the latter is 4017 // dirty card info to the modUnionTable. We therefore also use the 4018 // CGC_lock to protect the reading of the card table and the mod union 4019 // table by the CM thread. 4020 // . We run concurrently with mutator updates, so scanning 4021 // needs to be done carefully -- we should not try to scan 4022 // potentially uninitialized objects. 4023 // 4024 // Locking strategy: While holding the CGC_lock, we scan over and 4025 // reset a maximal dirty range of the mod union / card tables, then lock 4026 // the free_list_lock and bitmap lock to do a full marking, then 4027 // release these locks; and repeat the cycle. This allows for a 4028 // certain amount of fairness in the sharing of these locks between 4029 // the CMS collector on the one hand, and the VM thread and the 4030 // mutators on the other. 4031 4032 // NOTE: preclean_mod_union_table() and preclean_card_table() 4033 // further below are largely identical; if you need to modify 4034 // one of these methods, please check the other method too. 4035 4036 size_t CMSCollector::preclean_mod_union_table( 4037 ConcurrentMarkSweepGeneration* old_gen, 4038 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4039 verify_work_stacks_empty(); 4040 verify_overflow_empty(); 4041 4042 // strategy: starting with the first card, accumulate contiguous 4043 // ranges of dirty cards; clear these cards, then scan the region 4044 // covered by these cards. 4045 4046 // Since all of the MUT is committed ahead, we can just use 4047 // that, in case the generations expand while we are precleaning. 4048 // It might also be fine to just use the committed part of the 4049 // generation, but we might potentially miss cards when the 4050 // generation is rapidly expanding while we are in the midst 4051 // of precleaning. 4052 HeapWord* startAddr = old_gen->reserved().start(); 4053 HeapWord* endAddr = old_gen->reserved().end(); 4054 4055 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4056 4057 size_t numDirtyCards, cumNumDirtyCards; 4058 HeapWord *nextAddr, *lastAddr; 4059 for (cumNumDirtyCards = numDirtyCards = 0, 4060 nextAddr = lastAddr = startAddr; 4061 nextAddr < endAddr; 4062 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4063 4064 ResourceMark rm; 4065 HandleMark hm; 4066 4067 MemRegion dirtyRegion; 4068 { 4069 stopTimer(); 4070 // Potential yield point 4071 CMSTokenSync ts(true); 4072 startTimer(); 4073 sample_eden(); 4074 // Get dirty region starting at nextOffset (inclusive), 4075 // simultaneously clearing it. 4076 dirtyRegion = 4077 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4078 assert(dirtyRegion.start() >= nextAddr, 4079 "returned region inconsistent?"); 4080 } 4081 // Remember where the next search should begin. 4082 // The returned region (if non-empty) is a right open interval, 4083 // so lastOffset is obtained from the right end of that 4084 // interval. 4085 lastAddr = dirtyRegion.end(); 4086 // Should do something more transparent and less hacky XXX 4087 numDirtyCards = 4088 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4089 4090 // We'll scan the cards in the dirty region (with periodic 4091 // yields for foreground GC as needed). 4092 if (!dirtyRegion.is_empty()) { 4093 assert(numDirtyCards > 0, "consistency check"); 4094 HeapWord* stop_point = NULL; 4095 stopTimer(); 4096 // Potential yield point 4097 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4098 bitMapLock()); 4099 startTimer(); 4100 { 4101 verify_work_stacks_empty(); 4102 verify_overflow_empty(); 4103 sample_eden(); 4104 stop_point = 4105 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4106 } 4107 if (stop_point != NULL) { 4108 // The careful iteration stopped early either because it found an 4109 // uninitialized object, or because we were in the midst of an 4110 // "abortable preclean", which should now be aborted. Redirty 4111 // the bits corresponding to the partially-scanned or unscanned 4112 // cards. We'll either restart at the next block boundary or 4113 // abort the preclean. 4114 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4115 "Should only be AbortablePreclean."); 4116 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4117 if (should_abort_preclean()) { 4118 break; // out of preclean loop 4119 } else { 4120 // Compute the next address at which preclean should pick up; 4121 // might need bitMapLock in order to read P-bits. 4122 lastAddr = next_card_start_after_block(stop_point); 4123 } 4124 } 4125 } else { 4126 assert(lastAddr == endAddr, "consistency check"); 4127 assert(numDirtyCards == 0, "consistency check"); 4128 break; 4129 } 4130 } 4131 verify_work_stacks_empty(); 4132 verify_overflow_empty(); 4133 return cumNumDirtyCards; 4134 } 4135 4136 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4137 // below are largely identical; if you need to modify 4138 // one of these methods, please check the other method too. 4139 4140 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4141 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4142 // strategy: it's similar to precleamModUnionTable above, in that 4143 // we accumulate contiguous ranges of dirty cards, mark these cards 4144 // precleaned, then scan the region covered by these cards. 4145 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4146 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4147 4148 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4149 4150 size_t numDirtyCards, cumNumDirtyCards; 4151 HeapWord *lastAddr, *nextAddr; 4152 4153 for (cumNumDirtyCards = numDirtyCards = 0, 4154 nextAddr = lastAddr = startAddr; 4155 nextAddr < endAddr; 4156 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4157 4158 ResourceMark rm; 4159 HandleMark hm; 4160 4161 MemRegion dirtyRegion; 4162 { 4163 // See comments in "Precleaning notes" above on why we 4164 // do this locking. XXX Could the locking overheads be 4165 // too high when dirty cards are sparse? [I don't think so.] 4166 stopTimer(); 4167 CMSTokenSync x(true); // is cms thread 4168 startTimer(); 4169 sample_eden(); 4170 // Get and clear dirty region from card table 4171 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4172 MemRegion(nextAddr, endAddr), 4173 true, 4174 CardTableModRefBS::precleaned_card_val()); 4175 4176 assert(dirtyRegion.start() >= nextAddr, 4177 "returned region inconsistent?"); 4178 } 4179 lastAddr = dirtyRegion.end(); 4180 numDirtyCards = 4181 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4182 4183 if (!dirtyRegion.is_empty()) { 4184 stopTimer(); 4185 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4186 startTimer(); 4187 sample_eden(); 4188 verify_work_stacks_empty(); 4189 verify_overflow_empty(); 4190 HeapWord* stop_point = 4191 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4192 if (stop_point != NULL) { 4193 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4194 "Should only be AbortablePreclean."); 4195 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4196 if (should_abort_preclean()) { 4197 break; // out of preclean loop 4198 } else { 4199 // Compute the next address at which preclean should pick up. 4200 lastAddr = next_card_start_after_block(stop_point); 4201 } 4202 } 4203 } else { 4204 break; 4205 } 4206 } 4207 verify_work_stacks_empty(); 4208 verify_overflow_empty(); 4209 return cumNumDirtyCards; 4210 } 4211 4212 class PrecleanKlassClosure : public KlassClosure { 4213 KlassToOopClosure _cm_klass_closure; 4214 public: 4215 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4216 void do_klass(Klass* k) { 4217 if (k->has_accumulated_modified_oops()) { 4218 k->clear_accumulated_modified_oops(); 4219 4220 _cm_klass_closure.do_klass(k); 4221 } 4222 } 4223 }; 4224 4225 // The freelist lock is needed to prevent asserts, is it really needed? 4226 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4227 4228 cl->set_freelistLock(freelistLock); 4229 4230 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4231 4232 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4233 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4234 PrecleanKlassClosure preclean_klass_closure(cl); 4235 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4236 4237 verify_work_stacks_empty(); 4238 verify_overflow_empty(); 4239 } 4240 4241 void CMSCollector::checkpointRootsFinal() { 4242 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4243 check_correct_thread_executing(); 4244 // world is stopped at this checkpoint 4245 assert(SafepointSynchronize::is_at_safepoint(), 4246 "world should be stopped"); 4247 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4248 4249 verify_work_stacks_empty(); 4250 verify_overflow_empty(); 4251 4252 if (PrintGCDetails) { 4253 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]", 4254 _young_gen->used() / K, 4255 _young_gen->capacity() / K); 4256 } 4257 { 4258 if (CMSScavengeBeforeRemark) { 4259 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4260 // Temporarily set flag to false, GCH->do_collection will 4261 // expect it to be false and set to true 4262 FlagSetting fl(gch->_is_gc_active, false); 4263 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4264 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4265 gch->do_collection(true, // full (i.e. force, see below) 4266 false, // !clear_all_soft_refs 4267 0, // size 4268 false, // is_tlab 4269 GenCollectedHeap::YoungGen // type 4270 ); 4271 } 4272 FreelistLocker x(this); 4273 MutexLockerEx y(bitMapLock(), 4274 Mutex::_no_safepoint_check_flag); 4275 checkpointRootsFinalWork(); 4276 } 4277 verify_work_stacks_empty(); 4278 verify_overflow_empty(); 4279 } 4280 4281 void CMSCollector::checkpointRootsFinalWork() { 4282 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4283 4284 assert(haveFreelistLocks(), "must have free list locks"); 4285 assert_lock_strong(bitMapLock()); 4286 4287 ResourceMark rm; 4288 HandleMark hm; 4289 4290 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4291 4292 if (should_unload_classes()) { 4293 CodeCache::gc_prologue(); 4294 } 4295 assert(haveFreelistLocks(), "must have free list locks"); 4296 assert_lock_strong(bitMapLock()); 4297 4298 // We might assume that we need not fill TLAB's when 4299 // CMSScavengeBeforeRemark is set, because we may have just done 4300 // a scavenge which would have filled all TLAB's -- and besides 4301 // Eden would be empty. This however may not always be the case -- 4302 // for instance although we asked for a scavenge, it may not have 4303 // happened because of a JNI critical section. We probably need 4304 // a policy for deciding whether we can in that case wait until 4305 // the critical section releases and then do the remark following 4306 // the scavenge, and skip it here. In the absence of that policy, 4307 // or of an indication of whether the scavenge did indeed occur, 4308 // we cannot rely on TLAB's having been filled and must do 4309 // so here just in case a scavenge did not happen. 4310 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4311 // Update the saved marks which may affect the root scans. 4312 gch->save_marks(); 4313 4314 if (CMSPrintEdenSurvivorChunks) { 4315 print_eden_and_survivor_chunk_arrays(); 4316 } 4317 4318 { 4319 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4320 4321 // Note on the role of the mod union table: 4322 // Since the marker in "markFromRoots" marks concurrently with 4323 // mutators, it is possible for some reachable objects not to have been 4324 // scanned. For instance, an only reference to an object A was 4325 // placed in object B after the marker scanned B. Unless B is rescanned, 4326 // A would be collected. Such updates to references in marked objects 4327 // are detected via the mod union table which is the set of all cards 4328 // dirtied since the first checkpoint in this GC cycle and prior to 4329 // the most recent young generation GC, minus those cleaned up by the 4330 // concurrent precleaning. 4331 if (CMSParallelRemarkEnabled) { 4332 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4333 do_remark_parallel(); 4334 } else { 4335 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4336 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4337 do_remark_non_parallel(); 4338 } 4339 } 4340 verify_work_stacks_empty(); 4341 verify_overflow_empty(); 4342 4343 { 4344 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4345 refProcessingWork(); 4346 } 4347 verify_work_stacks_empty(); 4348 verify_overflow_empty(); 4349 4350 if (should_unload_classes()) { 4351 CodeCache::gc_epilogue(); 4352 } 4353 JvmtiExport::gc_epilogue(); 4354 4355 // If we encountered any (marking stack / work queue) overflow 4356 // events during the current CMS cycle, take appropriate 4357 // remedial measures, where possible, so as to try and avoid 4358 // recurrence of that condition. 4359 assert(_markStack.isEmpty(), "No grey objects"); 4360 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4361 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4362 if (ser_ovflw > 0) { 4363 if (PrintCMSStatistics != 0) { 4364 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4365 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT 4366 ", kac_preclean=" SIZE_FORMAT ")", 4367 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4368 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4369 } 4370 _markStack.expand(); 4371 _ser_pmc_remark_ovflw = 0; 4372 _ser_pmc_preclean_ovflw = 0; 4373 _ser_kac_preclean_ovflw = 0; 4374 _ser_kac_ovflw = 0; 4375 } 4376 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4377 if (PrintCMSStatistics != 0) { 4378 gclog_or_tty->print_cr("Work queue overflow (benign) " 4379 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4380 _par_pmc_remark_ovflw, _par_kac_ovflw); 4381 } 4382 _par_pmc_remark_ovflw = 0; 4383 _par_kac_ovflw = 0; 4384 } 4385 if (PrintCMSStatistics != 0) { 4386 if (_markStack._hit_limit > 0) { 4387 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4388 _markStack._hit_limit); 4389 } 4390 if (_markStack._failed_double > 0) { 4391 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT ")," 4392 " current capacity " SIZE_FORMAT, 4393 _markStack._failed_double, 4394 _markStack.capacity()); 4395 } 4396 } 4397 _markStack._hit_limit = 0; 4398 _markStack._failed_double = 0; 4399 4400 if ((VerifyAfterGC || VerifyDuringGC) && 4401 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4402 verify_after_remark(); 4403 } 4404 4405 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4406 4407 // Change under the freelistLocks. 4408 _collectorState = Sweeping; 4409 // Call isAllClear() under bitMapLock 4410 assert(_modUnionTable.isAllClear(), 4411 "Should be clear by end of the final marking"); 4412 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4413 "Should be clear by end of the final marking"); 4414 } 4415 4416 void CMSParInitialMarkTask::work(uint worker_id) { 4417 elapsedTimer _timer; 4418 ResourceMark rm; 4419 HandleMark hm; 4420 4421 // ---------- scan from roots -------------- 4422 _timer.start(); 4423 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4424 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4425 4426 // ---------- young gen roots -------------- 4427 { 4428 work_on_young_gen_roots(worker_id, &par_mri_cl); 4429 _timer.stop(); 4430 if (PrintCMSStatistics != 0) { 4431 gclog_or_tty->print_cr( 4432 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4433 worker_id, _timer.seconds()); 4434 } 4435 } 4436 4437 // ---------- remaining roots -------------- 4438 _timer.reset(); 4439 _timer.start(); 4440 4441 CLDToOopClosure cld_closure(&par_mri_cl, true); 4442 4443 gch->gen_process_roots(_strong_roots_scope, 4444 GenCollectedHeap::OldGen, 4445 false, // yg was scanned above 4446 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4447 _collector->should_unload_classes(), 4448 &par_mri_cl, 4449 NULL, 4450 &cld_closure); 4451 assert(_collector->should_unload_classes() 4452 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4453 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4454 _timer.stop(); 4455 if (PrintCMSStatistics != 0) { 4456 gclog_or_tty->print_cr( 4457 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4458 worker_id, _timer.seconds()); 4459 } 4460 } 4461 4462 // Parallel remark task 4463 class CMSParRemarkTask: public CMSParMarkTask { 4464 CompactibleFreeListSpace* _cms_space; 4465 4466 // The per-thread work queues, available here for stealing. 4467 OopTaskQueueSet* _task_queues; 4468 ParallelTaskTerminator _term; 4469 StrongRootsScope* _strong_roots_scope; 4470 4471 public: 4472 // A value of 0 passed to n_workers will cause the number of 4473 // workers to be taken from the active workers in the work gang. 4474 CMSParRemarkTask(CMSCollector* collector, 4475 CompactibleFreeListSpace* cms_space, 4476 uint n_workers, WorkGang* workers, 4477 OopTaskQueueSet* task_queues, 4478 StrongRootsScope* strong_roots_scope): 4479 CMSParMarkTask("Rescan roots and grey objects in parallel", 4480 collector, n_workers), 4481 _cms_space(cms_space), 4482 _task_queues(task_queues), 4483 _term(n_workers, task_queues), 4484 _strong_roots_scope(strong_roots_scope) { } 4485 4486 OopTaskQueueSet* task_queues() { return _task_queues; } 4487 4488 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4489 4490 ParallelTaskTerminator* terminator() { return &_term; } 4491 uint n_workers() { return _n_workers; } 4492 4493 void work(uint worker_id); 4494 4495 private: 4496 // ... of dirty cards in old space 4497 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4498 Par_MarkRefsIntoAndScanClosure* cl); 4499 4500 // ... work stealing for the above 4501 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4502 }; 4503 4504 class RemarkKlassClosure : public KlassClosure { 4505 KlassToOopClosure _cm_klass_closure; 4506 public: 4507 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4508 void do_klass(Klass* k) { 4509 // Check if we have modified any oops in the Klass during the concurrent marking. 4510 if (k->has_accumulated_modified_oops()) { 4511 k->clear_accumulated_modified_oops(); 4512 4513 // We could have transfered the current modified marks to the accumulated marks, 4514 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4515 } else if (k->has_modified_oops()) { 4516 // Don't clear anything, this info is needed by the next young collection. 4517 } else { 4518 // No modified oops in the Klass. 4519 return; 4520 } 4521 4522 // The klass has modified fields, need to scan the klass. 4523 _cm_klass_closure.do_klass(k); 4524 } 4525 }; 4526 4527 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4528 ParNewGeneration* young_gen = _collector->_young_gen; 4529 ContiguousSpace* eden_space = young_gen->eden(); 4530 ContiguousSpace* from_space = young_gen->from(); 4531 ContiguousSpace* to_space = young_gen->to(); 4532 4533 HeapWord** eca = _collector->_eden_chunk_array; 4534 size_t ect = _collector->_eden_chunk_index; 4535 HeapWord** sca = _collector->_survivor_chunk_array; 4536 size_t sct = _collector->_survivor_chunk_index; 4537 4538 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4539 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4540 4541 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4542 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4543 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4544 } 4545 4546 // work_queue(i) is passed to the closure 4547 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4548 // also is passed to do_dirty_card_rescan_tasks() and to 4549 // do_work_steal() to select the i-th task_queue. 4550 4551 void CMSParRemarkTask::work(uint worker_id) { 4552 elapsedTimer _timer; 4553 ResourceMark rm; 4554 HandleMark hm; 4555 4556 // ---------- rescan from roots -------------- 4557 _timer.start(); 4558 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4559 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4560 _collector->_span, _collector->ref_processor(), 4561 &(_collector->_markBitMap), 4562 work_queue(worker_id)); 4563 4564 // Rescan young gen roots first since these are likely 4565 // coarsely partitioned and may, on that account, constitute 4566 // the critical path; thus, it's best to start off that 4567 // work first. 4568 // ---------- young gen roots -------------- 4569 { 4570 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4571 _timer.stop(); 4572 if (PrintCMSStatistics != 0) { 4573 gclog_or_tty->print_cr( 4574 "Finished young gen rescan work in %dth thread: %3.3f sec", 4575 worker_id, _timer.seconds()); 4576 } 4577 } 4578 4579 // ---------- remaining roots -------------- 4580 _timer.reset(); 4581 _timer.start(); 4582 gch->gen_process_roots(_strong_roots_scope, 4583 GenCollectedHeap::OldGen, 4584 false, // yg was scanned above 4585 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4586 _collector->should_unload_classes(), 4587 &par_mrias_cl, 4588 NULL, 4589 NULL); // The dirty klasses will be handled below 4590 4591 assert(_collector->should_unload_classes() 4592 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4593 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4594 _timer.stop(); 4595 if (PrintCMSStatistics != 0) { 4596 gclog_or_tty->print_cr( 4597 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4598 worker_id, _timer.seconds()); 4599 } 4600 4601 // ---------- unhandled CLD scanning ---------- 4602 if (worker_id == 0) { // Single threaded at the moment. 4603 _timer.reset(); 4604 _timer.start(); 4605 4606 // Scan all new class loader data objects and new dependencies that were 4607 // introduced during concurrent marking. 4608 ResourceMark rm; 4609 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4610 for (int i = 0; i < array->length(); i++) { 4611 par_mrias_cl.do_cld_nv(array->at(i)); 4612 } 4613 4614 // We don't need to keep track of new CLDs anymore. 4615 ClassLoaderDataGraph::remember_new_clds(false); 4616 4617 _timer.stop(); 4618 if (PrintCMSStatistics != 0) { 4619 gclog_or_tty->print_cr( 4620 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4621 worker_id, _timer.seconds()); 4622 } 4623 } 4624 4625 // ---------- dirty klass scanning ---------- 4626 if (worker_id == 0) { // Single threaded at the moment. 4627 _timer.reset(); 4628 _timer.start(); 4629 4630 // Scan all classes that was dirtied during the concurrent marking phase. 4631 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4632 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4633 4634 _timer.stop(); 4635 if (PrintCMSStatistics != 0) { 4636 gclog_or_tty->print_cr( 4637 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4638 worker_id, _timer.seconds()); 4639 } 4640 } 4641 4642 // We might have added oops to ClassLoaderData::_handles during the 4643 // concurrent marking phase. These oops point to newly allocated objects 4644 // that are guaranteed to be kept alive. Either by the direct allocation 4645 // code, or when the young collector processes the roots. Hence, 4646 // we don't have to revisit the _handles block during the remark phase. 4647 4648 // ---------- rescan dirty cards ------------ 4649 _timer.reset(); 4650 _timer.start(); 4651 4652 // Do the rescan tasks for each of the two spaces 4653 // (cms_space) in turn. 4654 // "worker_id" is passed to select the task_queue for "worker_id" 4655 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4656 _timer.stop(); 4657 if (PrintCMSStatistics != 0) { 4658 gclog_or_tty->print_cr( 4659 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4660 worker_id, _timer.seconds()); 4661 } 4662 4663 // ---------- steal work from other threads ... 4664 // ---------- ... and drain overflow list. 4665 _timer.reset(); 4666 _timer.start(); 4667 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4668 _timer.stop(); 4669 if (PrintCMSStatistics != 0) { 4670 gclog_or_tty->print_cr( 4671 "Finished work stealing in %dth thread: %3.3f sec", 4672 worker_id, _timer.seconds()); 4673 } 4674 } 4675 4676 // Note that parameter "i" is not used. 4677 void 4678 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4679 OopsInGenClosure* cl, ContiguousSpace* space, 4680 HeapWord** chunk_array, size_t chunk_top) { 4681 // Until all tasks completed: 4682 // . claim an unclaimed task 4683 // . compute region boundaries corresponding to task claimed 4684 // using chunk_array 4685 // . par_oop_iterate(cl) over that region 4686 4687 ResourceMark rm; 4688 HandleMark hm; 4689 4690 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4691 4692 uint nth_task = 0; 4693 uint n_tasks = pst->n_tasks(); 4694 4695 if (n_tasks > 0) { 4696 assert(pst->valid(), "Uninitialized use?"); 4697 HeapWord *start, *end; 4698 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4699 // We claimed task # nth_task; compute its boundaries. 4700 if (chunk_top == 0) { // no samples were taken 4701 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4702 start = space->bottom(); 4703 end = space->top(); 4704 } else if (nth_task == 0) { 4705 start = space->bottom(); 4706 end = chunk_array[nth_task]; 4707 } else if (nth_task < (uint)chunk_top) { 4708 assert(nth_task >= 1, "Control point invariant"); 4709 start = chunk_array[nth_task - 1]; 4710 end = chunk_array[nth_task]; 4711 } else { 4712 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4713 start = chunk_array[chunk_top - 1]; 4714 end = space->top(); 4715 } 4716 MemRegion mr(start, end); 4717 // Verify that mr is in space 4718 assert(mr.is_empty() || space->used_region().contains(mr), 4719 "Should be in space"); 4720 // Verify that "start" is an object boundary 4721 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4722 "Should be an oop"); 4723 space->par_oop_iterate(mr, cl); 4724 } 4725 pst->all_tasks_completed(); 4726 } 4727 } 4728 4729 void 4730 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4731 CompactibleFreeListSpace* sp, int i, 4732 Par_MarkRefsIntoAndScanClosure* cl) { 4733 // Until all tasks completed: 4734 // . claim an unclaimed task 4735 // . compute region boundaries corresponding to task claimed 4736 // . transfer dirty bits ct->mut for that region 4737 // . apply rescanclosure to dirty mut bits for that region 4738 4739 ResourceMark rm; 4740 HandleMark hm; 4741 4742 OopTaskQueue* work_q = work_queue(i); 4743 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4744 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4745 // CAUTION: This closure has state that persists across calls to 4746 // the work method dirty_range_iterate_clear() in that it has 4747 // embedded in it a (subtype of) UpwardsObjectClosure. The 4748 // use of that state in the embedded UpwardsObjectClosure instance 4749 // assumes that the cards are always iterated (even if in parallel 4750 // by several threads) in monotonically increasing order per each 4751 // thread. This is true of the implementation below which picks 4752 // card ranges (chunks) in monotonically increasing order globally 4753 // and, a-fortiori, in monotonically increasing order per thread 4754 // (the latter order being a subsequence of the former). 4755 // If the work code below is ever reorganized into a more chaotic 4756 // work-partitioning form than the current "sequential tasks" 4757 // paradigm, the use of that persistent state will have to be 4758 // revisited and modified appropriately. See also related 4759 // bug 4756801 work on which should examine this code to make 4760 // sure that the changes there do not run counter to the 4761 // assumptions made here and necessary for correctness and 4762 // efficiency. Note also that this code might yield inefficient 4763 // behavior in the case of very large objects that span one or 4764 // more work chunks. Such objects would potentially be scanned 4765 // several times redundantly. Work on 4756801 should try and 4766 // address that performance anomaly if at all possible. XXX 4767 MemRegion full_span = _collector->_span; 4768 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4769 MarkFromDirtyCardsClosure 4770 greyRescanClosure(_collector, full_span, // entire span of interest 4771 sp, bm, work_q, cl); 4772 4773 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4774 assert(pst->valid(), "Uninitialized use?"); 4775 uint nth_task = 0; 4776 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4777 MemRegion span = sp->used_region(); 4778 HeapWord* start_addr = span.start(); 4779 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4780 alignment); 4781 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4782 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4783 start_addr, "Check alignment"); 4784 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4785 chunk_size, "Check alignment"); 4786 4787 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4788 // Having claimed the nth_task, compute corresponding mem-region, 4789 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4790 // The alignment restriction ensures that we do not need any 4791 // synchronization with other gang-workers while setting or 4792 // clearing bits in thus chunk of the MUT. 4793 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4794 start_addr + (nth_task+1)*chunk_size); 4795 // The last chunk's end might be way beyond end of the 4796 // used region. In that case pull back appropriately. 4797 if (this_span.end() > end_addr) { 4798 this_span.set_end(end_addr); 4799 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4800 } 4801 // Iterate over the dirty cards covering this chunk, marking them 4802 // precleaned, and setting the corresponding bits in the mod union 4803 // table. Since we have been careful to partition at Card and MUT-word 4804 // boundaries no synchronization is needed between parallel threads. 4805 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4806 &modUnionClosure); 4807 4808 // Having transferred these marks into the modUnionTable, 4809 // rescan the marked objects on the dirty cards in the modUnionTable. 4810 // Even if this is at a synchronous collection, the initial marking 4811 // may have been done during an asynchronous collection so there 4812 // may be dirty bits in the mod-union table. 4813 _collector->_modUnionTable.dirty_range_iterate_clear( 4814 this_span, &greyRescanClosure); 4815 _collector->_modUnionTable.verifyNoOneBitsInRange( 4816 this_span.start(), 4817 this_span.end()); 4818 } 4819 pst->all_tasks_completed(); // declare that i am done 4820 } 4821 4822 // . see if we can share work_queues with ParNew? XXX 4823 void 4824 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4825 int* seed) { 4826 OopTaskQueue* work_q = work_queue(i); 4827 NOT_PRODUCT(int num_steals = 0;) 4828 oop obj_to_scan; 4829 CMSBitMap* bm = &(_collector->_markBitMap); 4830 4831 while (true) { 4832 // Completely finish any left over work from (an) earlier round(s) 4833 cl->trim_queue(0); 4834 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4835 (size_t)ParGCDesiredObjsFromOverflowList); 4836 // Now check if there's any work in the overflow list 4837 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4838 // only affects the number of attempts made to get work from the 4839 // overflow list and does not affect the number of workers. Just 4840 // pass ParallelGCThreads so this behavior is unchanged. 4841 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4842 work_q, 4843 ParallelGCThreads)) { 4844 // found something in global overflow list; 4845 // not yet ready to go stealing work from others. 4846 // We'd like to assert(work_q->size() != 0, ...) 4847 // because we just took work from the overflow list, 4848 // but of course we can't since all of that could have 4849 // been already stolen from us. 4850 // "He giveth and He taketh away." 4851 continue; 4852 } 4853 // Verify that we have no work before we resort to stealing 4854 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4855 // Try to steal from other queues that have work 4856 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4857 NOT_PRODUCT(num_steals++;) 4858 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4859 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4860 // Do scanning work 4861 obj_to_scan->oop_iterate(cl); 4862 // Loop around, finish this work, and try to steal some more 4863 } else if (terminator()->offer_termination()) { 4864 break; // nirvana from the infinite cycle 4865 } 4866 } 4867 NOT_PRODUCT( 4868 if (PrintCMSStatistics != 0) { 4869 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4870 } 4871 ) 4872 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4873 "Else our work is not yet done"); 4874 } 4875 4876 // Record object boundaries in _eden_chunk_array by sampling the eden 4877 // top in the slow-path eden object allocation code path and record 4878 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4879 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4880 // sampling in sample_eden() that activates during the part of the 4881 // preclean phase. 4882 void CMSCollector::sample_eden_chunk() { 4883 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4884 if (_eden_chunk_lock->try_lock()) { 4885 // Record a sample. This is the critical section. The contents 4886 // of the _eden_chunk_array have to be non-decreasing in the 4887 // address order. 4888 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4889 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4890 "Unexpected state of Eden"); 4891 if (_eden_chunk_index == 0 || 4892 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4893 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4894 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4895 _eden_chunk_index++; // commit sample 4896 } 4897 _eden_chunk_lock->unlock(); 4898 } 4899 } 4900 } 4901 4902 // Return a thread-local PLAB recording array, as appropriate. 4903 void* CMSCollector::get_data_recorder(int thr_num) { 4904 if (_survivor_plab_array != NULL && 4905 (CMSPLABRecordAlways || 4906 (_collectorState > Marking && _collectorState < FinalMarking))) { 4907 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4908 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4909 ca->reset(); // clear it so that fresh data is recorded 4910 return (void*) ca; 4911 } else { 4912 return NULL; 4913 } 4914 } 4915 4916 // Reset all the thread-local PLAB recording arrays 4917 void CMSCollector::reset_survivor_plab_arrays() { 4918 for (uint i = 0; i < ParallelGCThreads; i++) { 4919 _survivor_plab_array[i].reset(); 4920 } 4921 } 4922 4923 // Merge the per-thread plab arrays into the global survivor chunk 4924 // array which will provide the partitioning of the survivor space 4925 // for CMS initial scan and rescan. 4926 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4927 int no_of_gc_threads) { 4928 assert(_survivor_plab_array != NULL, "Error"); 4929 assert(_survivor_chunk_array != NULL, "Error"); 4930 assert(_collectorState == FinalMarking || 4931 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4932 for (int j = 0; j < no_of_gc_threads; j++) { 4933 _cursor[j] = 0; 4934 } 4935 HeapWord* top = surv->top(); 4936 size_t i; 4937 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4938 HeapWord* min_val = top; // Higher than any PLAB address 4939 uint min_tid = 0; // position of min_val this round 4940 for (int j = 0; j < no_of_gc_threads; j++) { 4941 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4942 if (_cursor[j] == cur_sca->end()) { 4943 continue; 4944 } 4945 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4946 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4947 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4948 if (cur_val < min_val) { 4949 min_tid = j; 4950 min_val = cur_val; 4951 } else { 4952 assert(cur_val < top, "All recorded addresses should be less"); 4953 } 4954 } 4955 // At this point min_val and min_tid are respectively 4956 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4957 // and the thread (j) that witnesses that address. 4958 // We record this address in the _survivor_chunk_array[i] 4959 // and increment _cursor[min_tid] prior to the next round i. 4960 if (min_val == top) { 4961 break; 4962 } 4963 _survivor_chunk_array[i] = min_val; 4964 _cursor[min_tid]++; 4965 } 4966 // We are all done; record the size of the _survivor_chunk_array 4967 _survivor_chunk_index = i; // exclusive: [0, i) 4968 if (PrintCMSStatistics > 0) { 4969 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4970 } 4971 // Verify that we used up all the recorded entries 4972 #ifdef ASSERT 4973 size_t total = 0; 4974 for (int j = 0; j < no_of_gc_threads; j++) { 4975 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4976 total += _cursor[j]; 4977 } 4978 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4979 // Check that the merged array is in sorted order 4980 if (total > 0) { 4981 for (size_t i = 0; i < total - 1; i++) { 4982 if (PrintCMSStatistics > 0) { 4983 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4984 i, p2i(_survivor_chunk_array[i])); 4985 } 4986 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4987 "Not sorted"); 4988 } 4989 } 4990 #endif // ASSERT 4991 } 4992 4993 // Set up the space's par_seq_tasks structure for work claiming 4994 // for parallel initial scan and rescan of young gen. 4995 // See ParRescanTask where this is currently used. 4996 void 4997 CMSCollector:: 4998 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4999 assert(n_threads > 0, "Unexpected n_threads argument"); 5000 5001 // Eden space 5002 if (!_young_gen->eden()->is_empty()) { 5003 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5004 assert(!pst->valid(), "Clobbering existing data?"); 5005 // Each valid entry in [0, _eden_chunk_index) represents a task. 5006 size_t n_tasks = _eden_chunk_index + 1; 5007 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5008 // Sets the condition for completion of the subtask (how many threads 5009 // need to finish in order to be done). 5010 pst->set_n_threads(n_threads); 5011 pst->set_n_tasks((int)n_tasks); 5012 } 5013 5014 // Merge the survivor plab arrays into _survivor_chunk_array 5015 if (_survivor_plab_array != NULL) { 5016 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5017 } else { 5018 assert(_survivor_chunk_index == 0, "Error"); 5019 } 5020 5021 // To space 5022 { 5023 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5024 assert(!pst->valid(), "Clobbering existing data?"); 5025 // Sets the condition for completion of the subtask (how many threads 5026 // need to finish in order to be done). 5027 pst->set_n_threads(n_threads); 5028 pst->set_n_tasks(1); 5029 assert(pst->valid(), "Error"); 5030 } 5031 5032 // From space 5033 { 5034 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5035 assert(!pst->valid(), "Clobbering existing data?"); 5036 size_t n_tasks = _survivor_chunk_index + 1; 5037 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5038 // Sets the condition for completion of the subtask (how many threads 5039 // need to finish in order to be done). 5040 pst->set_n_threads(n_threads); 5041 pst->set_n_tasks((int)n_tasks); 5042 assert(pst->valid(), "Error"); 5043 } 5044 } 5045 5046 // Parallel version of remark 5047 void CMSCollector::do_remark_parallel() { 5048 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5049 WorkGang* workers = gch->workers(); 5050 assert(workers != NULL, "Need parallel worker threads."); 5051 // Choose to use the number of GC workers most recently set 5052 // into "active_workers". 5053 uint n_workers = workers->active_workers(); 5054 5055 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5056 5057 StrongRootsScope srs(n_workers); 5058 5059 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5060 5061 // We won't be iterating over the cards in the card table updating 5062 // the younger_gen cards, so we shouldn't call the following else 5063 // the verification code as well as subsequent younger_refs_iterate 5064 // code would get confused. XXX 5065 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5066 5067 // The young gen rescan work will not be done as part of 5068 // process_roots (which currently doesn't know how to 5069 // parallelize such a scan), but rather will be broken up into 5070 // a set of parallel tasks (via the sampling that the [abortable] 5071 // preclean phase did of eden, plus the [two] tasks of 5072 // scanning the [two] survivor spaces. Further fine-grain 5073 // parallelization of the scanning of the survivor spaces 5074 // themselves, and of precleaning of the young gen itself 5075 // is deferred to the future. 5076 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5077 5078 // The dirty card rescan work is broken up into a "sequence" 5079 // of parallel tasks (per constituent space) that are dynamically 5080 // claimed by the parallel threads. 5081 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5082 5083 // It turns out that even when we're using 1 thread, doing the work in a 5084 // separate thread causes wide variance in run times. We can't help this 5085 // in the multi-threaded case, but we special-case n=1 here to get 5086 // repeatable measurements of the 1-thread overhead of the parallel code. 5087 if (n_workers > 1) { 5088 // Make refs discovery MT-safe, if it isn't already: it may not 5089 // necessarily be so, since it's possible that we are doing 5090 // ST marking. 5091 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5092 workers->run_task(&tsk); 5093 } else { 5094 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5095 tsk.work(0); 5096 } 5097 5098 // restore, single-threaded for now, any preserved marks 5099 // as a result of work_q overflow 5100 restore_preserved_marks_if_any(); 5101 } 5102 5103 // Non-parallel version of remark 5104 void CMSCollector::do_remark_non_parallel() { 5105 ResourceMark rm; 5106 HandleMark hm; 5107 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5108 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5109 5110 MarkRefsIntoAndScanClosure 5111 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5112 &_markStack, this, 5113 false /* should_yield */, false /* not precleaning */); 5114 MarkFromDirtyCardsClosure 5115 markFromDirtyCardsClosure(this, _span, 5116 NULL, // space is set further below 5117 &_markBitMap, &_markStack, &mrias_cl); 5118 { 5119 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5120 // Iterate over the dirty cards, setting the corresponding bits in the 5121 // mod union table. 5122 { 5123 ModUnionClosure modUnionClosure(&_modUnionTable); 5124 _ct->ct_bs()->dirty_card_iterate( 5125 _cmsGen->used_region(), 5126 &modUnionClosure); 5127 } 5128 // Having transferred these marks into the modUnionTable, we just need 5129 // to rescan the marked objects on the dirty cards in the modUnionTable. 5130 // The initial marking may have been done during an asynchronous 5131 // collection so there may be dirty bits in the mod-union table. 5132 const int alignment = 5133 CardTableModRefBS::card_size * BitsPerWord; 5134 { 5135 // ... First handle dirty cards in CMS gen 5136 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5137 MemRegion ur = _cmsGen->used_region(); 5138 HeapWord* lb = ur.start(); 5139 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5140 MemRegion cms_span(lb, ub); 5141 _modUnionTable.dirty_range_iterate_clear(cms_span, 5142 &markFromDirtyCardsClosure); 5143 verify_work_stacks_empty(); 5144 if (PrintCMSStatistics != 0) { 5145 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", 5146 markFromDirtyCardsClosure.num_dirty_cards()); 5147 } 5148 } 5149 } 5150 if (VerifyDuringGC && 5151 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5152 HandleMark hm; // Discard invalid handles created during verification 5153 Universe::verify(); 5154 } 5155 { 5156 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5157 5158 verify_work_stacks_empty(); 5159 5160 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5161 StrongRootsScope srs(1); 5162 5163 gch->gen_process_roots(&srs, 5164 GenCollectedHeap::OldGen, 5165 true, // young gen as roots 5166 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5167 should_unload_classes(), 5168 &mrias_cl, 5169 NULL, 5170 NULL); // The dirty klasses will be handled below 5171 5172 assert(should_unload_classes() 5173 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5174 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5175 } 5176 5177 { 5178 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5179 5180 verify_work_stacks_empty(); 5181 5182 // Scan all class loader data objects that might have been introduced 5183 // during concurrent marking. 5184 ResourceMark rm; 5185 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5186 for (int i = 0; i < array->length(); i++) { 5187 mrias_cl.do_cld_nv(array->at(i)); 5188 } 5189 5190 // We don't need to keep track of new CLDs anymore. 5191 ClassLoaderDataGraph::remember_new_clds(false); 5192 5193 verify_work_stacks_empty(); 5194 } 5195 5196 { 5197 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5198 5199 verify_work_stacks_empty(); 5200 5201 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5202 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5203 5204 verify_work_stacks_empty(); 5205 } 5206 5207 // We might have added oops to ClassLoaderData::_handles during the 5208 // concurrent marking phase. These oops point to newly allocated objects 5209 // that are guaranteed to be kept alive. Either by the direct allocation 5210 // code, or when the young collector processes the roots. Hence, 5211 // we don't have to revisit the _handles block during the remark phase. 5212 5213 verify_work_stacks_empty(); 5214 // Restore evacuated mark words, if any, used for overflow list links 5215 if (!CMSOverflowEarlyRestoration) { 5216 restore_preserved_marks_if_any(); 5217 } 5218 verify_overflow_empty(); 5219 } 5220 5221 //////////////////////////////////////////////////////// 5222 // Parallel Reference Processing Task Proxy Class 5223 //////////////////////////////////////////////////////// 5224 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5225 OopTaskQueueSet* _queues; 5226 ParallelTaskTerminator _terminator; 5227 public: 5228 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5229 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5230 ParallelTaskTerminator* terminator() { return &_terminator; } 5231 OopTaskQueueSet* queues() { return _queues; } 5232 }; 5233 5234 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5235 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5236 CMSCollector* _collector; 5237 CMSBitMap* _mark_bit_map; 5238 const MemRegion _span; 5239 ProcessTask& _task; 5240 5241 public: 5242 CMSRefProcTaskProxy(ProcessTask& task, 5243 CMSCollector* collector, 5244 const MemRegion& span, 5245 CMSBitMap* mark_bit_map, 5246 AbstractWorkGang* workers, 5247 OopTaskQueueSet* task_queues): 5248 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5249 task_queues, 5250 workers->active_workers()), 5251 _task(task), 5252 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5253 { 5254 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5255 "Inconsistency in _span"); 5256 } 5257 5258 OopTaskQueueSet* task_queues() { return queues(); } 5259 5260 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5261 5262 void do_work_steal(int i, 5263 CMSParDrainMarkingStackClosure* drain, 5264 CMSParKeepAliveClosure* keep_alive, 5265 int* seed); 5266 5267 virtual void work(uint worker_id); 5268 }; 5269 5270 void CMSRefProcTaskProxy::work(uint worker_id) { 5271 ResourceMark rm; 5272 HandleMark hm; 5273 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5274 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5275 _mark_bit_map, 5276 work_queue(worker_id)); 5277 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5278 _mark_bit_map, 5279 work_queue(worker_id)); 5280 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5281 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5282 if (_task.marks_oops_alive()) { 5283 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5284 _collector->hash_seed(worker_id)); 5285 } 5286 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5287 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5288 } 5289 5290 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5291 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5292 EnqueueTask& _task; 5293 5294 public: 5295 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5296 : AbstractGangTask("Enqueue reference objects in parallel"), 5297 _task(task) 5298 { } 5299 5300 virtual void work(uint worker_id) 5301 { 5302 _task.work(worker_id); 5303 } 5304 }; 5305 5306 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5307 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5308 _span(span), 5309 _bit_map(bit_map), 5310 _work_queue(work_queue), 5311 _mark_and_push(collector, span, bit_map, work_queue), 5312 _low_water_mark(MIN2((work_queue->max_elems()/4), 5313 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5314 { } 5315 5316 // . see if we can share work_queues with ParNew? XXX 5317 void CMSRefProcTaskProxy::do_work_steal(int i, 5318 CMSParDrainMarkingStackClosure* drain, 5319 CMSParKeepAliveClosure* keep_alive, 5320 int* seed) { 5321 OopTaskQueue* work_q = work_queue(i); 5322 NOT_PRODUCT(int num_steals = 0;) 5323 oop obj_to_scan; 5324 5325 while (true) { 5326 // Completely finish any left over work from (an) earlier round(s) 5327 drain->trim_queue(0); 5328 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5329 (size_t)ParGCDesiredObjsFromOverflowList); 5330 // Now check if there's any work in the overflow list 5331 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5332 // only affects the number of attempts made to get work from the 5333 // overflow list and does not affect the number of workers. Just 5334 // pass ParallelGCThreads so this behavior is unchanged. 5335 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5336 work_q, 5337 ParallelGCThreads)) { 5338 // Found something in global overflow list; 5339 // not yet ready to go stealing work from others. 5340 // We'd like to assert(work_q->size() != 0, ...) 5341 // because we just took work from the overflow list, 5342 // but of course we can't, since all of that might have 5343 // been already stolen from us. 5344 continue; 5345 } 5346 // Verify that we have no work before we resort to stealing 5347 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5348 // Try to steal from other queues that have work 5349 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5350 NOT_PRODUCT(num_steals++;) 5351 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5352 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5353 // Do scanning work 5354 obj_to_scan->oop_iterate(keep_alive); 5355 // Loop around, finish this work, and try to steal some more 5356 } else if (terminator()->offer_termination()) { 5357 break; // nirvana from the infinite cycle 5358 } 5359 } 5360 NOT_PRODUCT( 5361 if (PrintCMSStatistics != 0) { 5362 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5363 } 5364 ) 5365 } 5366 5367 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5368 { 5369 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5370 WorkGang* workers = gch->workers(); 5371 assert(workers != NULL, "Need parallel worker threads."); 5372 CMSRefProcTaskProxy rp_task(task, &_collector, 5373 _collector.ref_processor()->span(), 5374 _collector.markBitMap(), 5375 workers, _collector.task_queues()); 5376 workers->run_task(&rp_task); 5377 } 5378 5379 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5380 { 5381 5382 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5383 WorkGang* workers = gch->workers(); 5384 assert(workers != NULL, "Need parallel worker threads."); 5385 CMSRefEnqueueTaskProxy enq_task(task); 5386 workers->run_task(&enq_task); 5387 } 5388 5389 void CMSCollector::refProcessingWork() { 5390 ResourceMark rm; 5391 HandleMark hm; 5392 5393 ReferenceProcessor* rp = ref_processor(); 5394 assert(rp->span().equals(_span), "Spans should be equal"); 5395 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5396 // Process weak references. 5397 rp->setup_policy(false); 5398 verify_work_stacks_empty(); 5399 5400 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5401 &_markStack, false /* !preclean */); 5402 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5403 _span, &_markBitMap, &_markStack, 5404 &cmsKeepAliveClosure, false /* !preclean */); 5405 { 5406 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5407 5408 ReferenceProcessorStats stats; 5409 if (rp->processing_is_mt()) { 5410 // Set the degree of MT here. If the discovery is done MT, there 5411 // may have been a different number of threads doing the discovery 5412 // and a different number of discovered lists may have Ref objects. 5413 // That is OK as long as the Reference lists are balanced (see 5414 // balance_all_queues() and balance_queues()). 5415 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5416 uint active_workers = ParallelGCThreads; 5417 WorkGang* workers = gch->workers(); 5418 if (workers != NULL) { 5419 active_workers = workers->active_workers(); 5420 // The expectation is that active_workers will have already 5421 // been set to a reasonable value. If it has not been set, 5422 // investigate. 5423 assert(active_workers > 0, "Should have been set during scavenge"); 5424 } 5425 rp->set_active_mt_degree(active_workers); 5426 CMSRefProcTaskExecutor task_executor(*this); 5427 stats = rp->process_discovered_references(&_is_alive_closure, 5428 &cmsKeepAliveClosure, 5429 &cmsDrainMarkingStackClosure, 5430 &task_executor, 5431 _gc_timer_cm, 5432 _gc_tracer_cm->gc_id()); 5433 } else { 5434 stats = rp->process_discovered_references(&_is_alive_closure, 5435 &cmsKeepAliveClosure, 5436 &cmsDrainMarkingStackClosure, 5437 NULL, 5438 _gc_timer_cm, 5439 _gc_tracer_cm->gc_id()); 5440 } 5441 _gc_tracer_cm->report_gc_reference_stats(stats); 5442 5443 } 5444 5445 // This is the point where the entire marking should have completed. 5446 verify_work_stacks_empty(); 5447 5448 if (should_unload_classes()) { 5449 { 5450 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5451 5452 // Unload classes and purge the SystemDictionary. 5453 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5454 5455 // Unload nmethods. 5456 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5457 5458 // Prune dead klasses from subklass/sibling/implementor lists. 5459 Klass::clean_weak_klass_links(&_is_alive_closure); 5460 } 5461 5462 { 5463 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5464 // Clean up unreferenced symbols in symbol table. 5465 SymbolTable::unlink(); 5466 } 5467 5468 { 5469 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5470 // Delete entries for dead interned strings. 5471 StringTable::unlink(&_is_alive_closure); 5472 } 5473 } 5474 5475 5476 // Restore any preserved marks as a result of mark stack or 5477 // work queue overflow 5478 restore_preserved_marks_if_any(); // done single-threaded for now 5479 5480 rp->set_enqueuing_is_done(true); 5481 if (rp->processing_is_mt()) { 5482 rp->balance_all_queues(); 5483 CMSRefProcTaskExecutor task_executor(*this); 5484 rp->enqueue_discovered_references(&task_executor); 5485 } else { 5486 rp->enqueue_discovered_references(NULL); 5487 } 5488 rp->verify_no_references_recorded(); 5489 assert(!rp->discovery_enabled(), "should have been disabled"); 5490 } 5491 5492 #ifndef PRODUCT 5493 void CMSCollector::check_correct_thread_executing() { 5494 Thread* t = Thread::current(); 5495 // Only the VM thread or the CMS thread should be here. 5496 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5497 "Unexpected thread type"); 5498 // If this is the vm thread, the foreground process 5499 // should not be waiting. Note that _foregroundGCIsActive is 5500 // true while the foreground collector is waiting. 5501 if (_foregroundGCShouldWait) { 5502 // We cannot be the VM thread 5503 assert(t->is_ConcurrentGC_thread(), 5504 "Should be CMS thread"); 5505 } else { 5506 // We can be the CMS thread only if we are in a stop-world 5507 // phase of CMS collection. 5508 if (t->is_ConcurrentGC_thread()) { 5509 assert(_collectorState == InitialMarking || 5510 _collectorState == FinalMarking, 5511 "Should be a stop-world phase"); 5512 // The CMS thread should be holding the CMS_token. 5513 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5514 "Potential interference with concurrently " 5515 "executing VM thread"); 5516 } 5517 } 5518 } 5519 #endif 5520 5521 void CMSCollector::sweep() { 5522 assert(_collectorState == Sweeping, "just checking"); 5523 check_correct_thread_executing(); 5524 verify_work_stacks_empty(); 5525 verify_overflow_empty(); 5526 increment_sweep_count(); 5527 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5528 5529 _inter_sweep_timer.stop(); 5530 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5531 5532 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5533 _intra_sweep_timer.reset(); 5534 _intra_sweep_timer.start(); 5535 { 5536 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5537 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5538 // First sweep the old gen 5539 { 5540 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5541 bitMapLock()); 5542 sweepWork(_cmsGen); 5543 } 5544 5545 // Update Universe::_heap_*_at_gc figures. 5546 // We need all the free list locks to make the abstract state 5547 // transition from Sweeping to Resetting. See detailed note 5548 // further below. 5549 { 5550 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5551 // Update heap occupancy information which is used as 5552 // input to soft ref clearing policy at the next gc. 5553 Universe::update_heap_info_at_gc(); 5554 _collectorState = Resizing; 5555 } 5556 } 5557 verify_work_stacks_empty(); 5558 verify_overflow_empty(); 5559 5560 if (should_unload_classes()) { 5561 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5562 // requires that the virtual spaces are stable and not deleted. 5563 ClassLoaderDataGraph::set_should_purge(true); 5564 } 5565 5566 _intra_sweep_timer.stop(); 5567 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5568 5569 _inter_sweep_timer.reset(); 5570 _inter_sweep_timer.start(); 5571 5572 // We need to use a monotonically non-decreasing time in ms 5573 // or we will see time-warp warnings and os::javaTimeMillis() 5574 // does not guarantee monotonicity. 5575 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5576 update_time_of_last_gc(now); 5577 5578 // NOTE on abstract state transitions: 5579 // Mutators allocate-live and/or mark the mod-union table dirty 5580 // based on the state of the collection. The former is done in 5581 // the interval [Marking, Sweeping] and the latter in the interval 5582 // [Marking, Sweeping). Thus the transitions into the Marking state 5583 // and out of the Sweeping state must be synchronously visible 5584 // globally to the mutators. 5585 // The transition into the Marking state happens with the world 5586 // stopped so the mutators will globally see it. Sweeping is 5587 // done asynchronously by the background collector so the transition 5588 // from the Sweeping state to the Resizing state must be done 5589 // under the freelistLock (as is the check for whether to 5590 // allocate-live and whether to dirty the mod-union table). 5591 assert(_collectorState == Resizing, "Change of collector state to" 5592 " Resizing must be done under the freelistLocks (plural)"); 5593 5594 // Now that sweeping has been completed, we clear 5595 // the incremental_collection_failed flag, 5596 // thus inviting a younger gen collection to promote into 5597 // this generation. If such a promotion may still fail, 5598 // the flag will be set again when a young collection is 5599 // attempted. 5600 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5601 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5602 gch->update_full_collections_completed(_collection_count_start); 5603 } 5604 5605 // FIX ME!!! Looks like this belongs in CFLSpace, with 5606 // CMSGen merely delegating to it. 5607 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5608 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5609 HeapWord* minAddr = _cmsSpace->bottom(); 5610 HeapWord* largestAddr = 5611 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5612 if (largestAddr == NULL) { 5613 // The dictionary appears to be empty. In this case 5614 // try to coalesce at the end of the heap. 5615 largestAddr = _cmsSpace->end(); 5616 } 5617 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5618 size_t nearLargestOffset = 5619 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5620 if (PrintFLSStatistics != 0) { 5621 gclog_or_tty->print_cr( 5622 "CMS: Large Block: " PTR_FORMAT ";" 5623 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5624 p2i(largestAddr), 5625 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5626 } 5627 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5628 } 5629 5630 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5631 return addr >= _cmsSpace->nearLargestChunk(); 5632 } 5633 5634 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5635 return _cmsSpace->find_chunk_at_end(); 5636 } 5637 5638 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5639 bool full) { 5640 // If the young generation has been collected, gather any statistics 5641 // that are of interest at this point. 5642 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5643 if (!full && current_is_young) { 5644 // Gather statistics on the young generation collection. 5645 collector()->stats().record_gc0_end(used()); 5646 } 5647 } 5648 5649 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5650 // We iterate over the space(s) underlying this generation, 5651 // checking the mark bit map to see if the bits corresponding 5652 // to specific blocks are marked or not. Blocks that are 5653 // marked are live and are not swept up. All remaining blocks 5654 // are swept up, with coalescing on-the-fly as we sweep up 5655 // contiguous free and/or garbage blocks: 5656 // We need to ensure that the sweeper synchronizes with allocators 5657 // and stop-the-world collectors. In particular, the following 5658 // locks are used: 5659 // . CMS token: if this is held, a stop the world collection cannot occur 5660 // . freelistLock: if this is held no allocation can occur from this 5661 // generation by another thread 5662 // . bitMapLock: if this is held, no other thread can access or update 5663 // 5664 5665 // Note that we need to hold the freelistLock if we use 5666 // block iterate below; else the iterator might go awry if 5667 // a mutator (or promotion) causes block contents to change 5668 // (for instance if the allocator divvies up a block). 5669 // If we hold the free list lock, for all practical purposes 5670 // young generation GC's can't occur (they'll usually need to 5671 // promote), so we might as well prevent all young generation 5672 // GC's while we do a sweeping step. For the same reason, we might 5673 // as well take the bit map lock for the entire duration 5674 5675 // check that we hold the requisite locks 5676 assert(have_cms_token(), "Should hold cms token"); 5677 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5678 assert_lock_strong(old_gen->freelistLock()); 5679 assert_lock_strong(bitMapLock()); 5680 5681 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5682 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5683 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5684 _inter_sweep_estimate.padded_average(), 5685 _intra_sweep_estimate.padded_average()); 5686 old_gen->setNearLargestChunk(); 5687 5688 { 5689 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5690 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5691 // We need to free-up/coalesce garbage/blocks from a 5692 // co-terminal free run. This is done in the SweepClosure 5693 // destructor; so, do not remove this scope, else the 5694 // end-of-sweep-census below will be off by a little bit. 5695 } 5696 old_gen->cmsSpace()->sweep_completed(); 5697 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5698 if (should_unload_classes()) { // unloaded classes this cycle, 5699 _concurrent_cycles_since_last_unload = 0; // ... reset count 5700 } else { // did not unload classes, 5701 _concurrent_cycles_since_last_unload++; // ... increment count 5702 } 5703 } 5704 5705 // Reset CMS data structures (for now just the marking bit map) 5706 // preparatory for the next cycle. 5707 void CMSCollector::reset(bool concurrent) { 5708 if (concurrent) { 5709 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5710 5711 // If the state is not "Resetting", the foreground thread 5712 // has done a collection and the resetting. 5713 if (_collectorState != Resetting) { 5714 assert(_collectorState == Idling, "The state should only change" 5715 " because the foreground collector has finished the collection"); 5716 return; 5717 } 5718 5719 // Clear the mark bitmap (no grey objects to start with) 5720 // for the next cycle. 5721 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5722 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5723 5724 HeapWord* curAddr = _markBitMap.startWord(); 5725 while (curAddr < _markBitMap.endWord()) { 5726 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5727 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5728 _markBitMap.clear_large_range(chunk); 5729 if (ConcurrentMarkSweepThread::should_yield() && 5730 !foregroundGCIsActive() && 5731 CMSYield) { 5732 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5733 "CMS thread should hold CMS token"); 5734 assert_lock_strong(bitMapLock()); 5735 bitMapLock()->unlock(); 5736 ConcurrentMarkSweepThread::desynchronize(true); 5737 stopTimer(); 5738 if (PrintCMSStatistics != 0) { 5739 incrementYields(); 5740 } 5741 5742 // See the comment in coordinator_yield() 5743 for (unsigned i = 0; i < CMSYieldSleepCount && 5744 ConcurrentMarkSweepThread::should_yield() && 5745 !CMSCollector::foregroundGCIsActive(); ++i) { 5746 os::sleep(Thread::current(), 1, false); 5747 } 5748 5749 ConcurrentMarkSweepThread::synchronize(true); 5750 bitMapLock()->lock_without_safepoint_check(); 5751 startTimer(); 5752 } 5753 curAddr = chunk.end(); 5754 } 5755 // A successful mostly concurrent collection has been done. 5756 // Because only the full (i.e., concurrent mode failure) collections 5757 // are being measured for gc overhead limits, clean the "near" flag 5758 // and count. 5759 size_policy()->reset_gc_overhead_limit_count(); 5760 _collectorState = Idling; 5761 } else { 5762 // already have the lock 5763 assert(_collectorState == Resetting, "just checking"); 5764 assert_lock_strong(bitMapLock()); 5765 _markBitMap.clear_all(); 5766 _collectorState = Idling; 5767 } 5768 5769 register_gc_end(); 5770 } 5771 5772 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5773 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5774 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5775 TraceCollectorStats tcs(counters()); 5776 5777 switch (op) { 5778 case CMS_op_checkpointRootsInitial: { 5779 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5780 checkpointRootsInitial(); 5781 if (PrintGC) { 5782 _cmsGen->printOccupancy("initial-mark"); 5783 } 5784 break; 5785 } 5786 case CMS_op_checkpointRootsFinal: { 5787 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5788 checkpointRootsFinal(); 5789 if (PrintGC) { 5790 _cmsGen->printOccupancy("remark"); 5791 } 5792 break; 5793 } 5794 default: 5795 fatal("No such CMS_op"); 5796 } 5797 } 5798 5799 #ifndef PRODUCT 5800 size_t const CMSCollector::skip_header_HeapWords() { 5801 return FreeChunk::header_size(); 5802 } 5803 5804 // Try and collect here conditions that should hold when 5805 // CMS thread is exiting. The idea is that the foreground GC 5806 // thread should not be blocked if it wants to terminate 5807 // the CMS thread and yet continue to run the VM for a while 5808 // after that. 5809 void CMSCollector::verify_ok_to_terminate() const { 5810 assert(Thread::current()->is_ConcurrentGC_thread(), 5811 "should be called by CMS thread"); 5812 assert(!_foregroundGCShouldWait, "should be false"); 5813 // We could check here that all the various low-level locks 5814 // are not held by the CMS thread, but that is overkill; see 5815 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5816 // is checked. 5817 } 5818 #endif 5819 5820 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5821 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5822 "missing Printezis mark?"); 5823 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5824 size_t size = pointer_delta(nextOneAddr + 1, addr); 5825 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5826 "alignment problem"); 5827 assert(size >= 3, "Necessary for Printezis marks to work"); 5828 return size; 5829 } 5830 5831 // A variant of the above (block_size_using_printezis_bits()) except 5832 // that we return 0 if the P-bits are not yet set. 5833 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5834 if (_markBitMap.isMarked(addr + 1)) { 5835 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5836 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5837 size_t size = pointer_delta(nextOneAddr + 1, addr); 5838 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5839 "alignment problem"); 5840 assert(size >= 3, "Necessary for Printezis marks to work"); 5841 return size; 5842 } 5843 return 0; 5844 } 5845 5846 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5847 size_t sz = 0; 5848 oop p = (oop)addr; 5849 if (p->klass_or_null() != NULL) { 5850 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5851 } else { 5852 sz = block_size_using_printezis_bits(addr); 5853 } 5854 assert(sz > 0, "size must be nonzero"); 5855 HeapWord* next_block = addr + sz; 5856 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5857 CardTableModRefBS::card_size); 5858 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5859 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5860 "must be different cards"); 5861 return next_card; 5862 } 5863 5864 5865 // CMS Bit Map Wrapper ///////////////////////////////////////// 5866 5867 // Construct a CMS bit map infrastructure, but don't create the 5868 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5869 // further below. 5870 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5871 _bm(), 5872 _shifter(shifter), 5873 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5874 Monitor::_safepoint_check_sometimes) : NULL) 5875 { 5876 _bmStartWord = 0; 5877 _bmWordSize = 0; 5878 } 5879 5880 bool CMSBitMap::allocate(MemRegion mr) { 5881 _bmStartWord = mr.start(); 5882 _bmWordSize = mr.word_size(); 5883 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5884 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5885 if (!brs.is_reserved()) { 5886 warning("CMS bit map allocation failure"); 5887 return false; 5888 } 5889 // For now we'll just commit all of the bit map up front. 5890 // Later on we'll try to be more parsimonious with swap. 5891 if (!_virtual_space.initialize(brs, brs.size())) { 5892 warning("CMS bit map backing store failure"); 5893 return false; 5894 } 5895 assert(_virtual_space.committed_size() == brs.size(), 5896 "didn't reserve backing store for all of CMS bit map?"); 5897 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5898 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5899 _bmWordSize, "inconsistency in bit map sizing"); 5900 _bm.set_size(_bmWordSize >> _shifter); 5901 5902 // bm.clear(); // can we rely on getting zero'd memory? verify below 5903 assert(isAllClear(), 5904 "Expected zero'd memory from ReservedSpace constructor"); 5905 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5906 "consistency check"); 5907 return true; 5908 } 5909 5910 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5911 HeapWord *next_addr, *end_addr, *last_addr; 5912 assert_locked(); 5913 assert(covers(mr), "out-of-range error"); 5914 // XXX assert that start and end are appropriately aligned 5915 for (next_addr = mr.start(), end_addr = mr.end(); 5916 next_addr < end_addr; next_addr = last_addr) { 5917 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5918 last_addr = dirty_region.end(); 5919 if (!dirty_region.is_empty()) { 5920 cl->do_MemRegion(dirty_region); 5921 } else { 5922 assert(last_addr == end_addr, "program logic"); 5923 return; 5924 } 5925 } 5926 } 5927 5928 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5929 _bm.print_on_error(st, prefix); 5930 } 5931 5932 #ifndef PRODUCT 5933 void CMSBitMap::assert_locked() const { 5934 CMSLockVerifier::assert_locked(lock()); 5935 } 5936 5937 bool CMSBitMap::covers(MemRegion mr) const { 5938 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5939 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5940 "size inconsistency"); 5941 return (mr.start() >= _bmStartWord) && 5942 (mr.end() <= endWord()); 5943 } 5944 5945 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5946 return (start >= _bmStartWord && (start + size) <= endWord()); 5947 } 5948 5949 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5950 // verify that there are no 1 bits in the interval [left, right) 5951 FalseBitMapClosure falseBitMapClosure; 5952 iterate(&falseBitMapClosure, left, right); 5953 } 5954 5955 void CMSBitMap::region_invariant(MemRegion mr) 5956 { 5957 assert_locked(); 5958 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5959 assert(!mr.is_empty(), "unexpected empty region"); 5960 assert(covers(mr), "mr should be covered by bit map"); 5961 // convert address range into offset range 5962 size_t start_ofs = heapWordToOffset(mr.start()); 5963 // Make sure that end() is appropriately aligned 5964 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5965 (1 << (_shifter+LogHeapWordSize))), 5966 "Misaligned mr.end()"); 5967 size_t end_ofs = heapWordToOffset(mr.end()); 5968 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5969 } 5970 5971 #endif 5972 5973 bool CMSMarkStack::allocate(size_t size) { 5974 // allocate a stack of the requisite depth 5975 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5976 size * sizeof(oop))); 5977 if (!rs.is_reserved()) { 5978 warning("CMSMarkStack allocation failure"); 5979 return false; 5980 } 5981 if (!_virtual_space.initialize(rs, rs.size())) { 5982 warning("CMSMarkStack backing store failure"); 5983 return false; 5984 } 5985 assert(_virtual_space.committed_size() == rs.size(), 5986 "didn't reserve backing store for all of CMS stack?"); 5987 _base = (oop*)(_virtual_space.low()); 5988 _index = 0; 5989 _capacity = size; 5990 NOT_PRODUCT(_max_depth = 0); 5991 return true; 5992 } 5993 5994 // XXX FIX ME !!! In the MT case we come in here holding a 5995 // leaf lock. For printing we need to take a further lock 5996 // which has lower rank. We need to recalibrate the two 5997 // lock-ranks involved in order to be able to print the 5998 // messages below. (Or defer the printing to the caller. 5999 // For now we take the expedient path of just disabling the 6000 // messages for the problematic case.) 6001 void CMSMarkStack::expand() { 6002 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6003 if (_capacity == MarkStackSizeMax) { 6004 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6005 // We print a warning message only once per CMS cycle. 6006 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6007 } 6008 return; 6009 } 6010 // Double capacity if possible 6011 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6012 // Do not give up existing stack until we have managed to 6013 // get the double capacity that we desired. 6014 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6015 new_capacity * sizeof(oop))); 6016 if (rs.is_reserved()) { 6017 // Release the backing store associated with old stack 6018 _virtual_space.release(); 6019 // Reinitialize virtual space for new stack 6020 if (!_virtual_space.initialize(rs, rs.size())) { 6021 fatal("Not enough swap for expanded marking stack"); 6022 } 6023 _base = (oop*)(_virtual_space.low()); 6024 _index = 0; 6025 _capacity = new_capacity; 6026 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6027 // Failed to double capacity, continue; 6028 // we print a detail message only once per CMS cycle. 6029 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " 6030 SIZE_FORMAT "K", 6031 _capacity / K, new_capacity / K); 6032 } 6033 } 6034 6035 6036 // Closures 6037 // XXX: there seems to be a lot of code duplication here; 6038 // should refactor and consolidate common code. 6039 6040 // This closure is used to mark refs into the CMS generation in 6041 // the CMS bit map. Called at the first checkpoint. This closure 6042 // assumes that we do not need to re-mark dirty cards; if the CMS 6043 // generation on which this is used is not an oldest 6044 // generation then this will lose younger_gen cards! 6045 6046 MarkRefsIntoClosure::MarkRefsIntoClosure( 6047 MemRegion span, CMSBitMap* bitMap): 6048 _span(span), 6049 _bitMap(bitMap) 6050 { 6051 assert(_ref_processor == NULL, "deliberately left NULL"); 6052 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6053 } 6054 6055 void MarkRefsIntoClosure::do_oop(oop obj) { 6056 // if p points into _span, then mark corresponding bit in _markBitMap 6057 assert(obj->is_oop(), "expected an oop"); 6058 HeapWord* addr = (HeapWord*)obj; 6059 if (_span.contains(addr)) { 6060 // this should be made more efficient 6061 _bitMap->mark(addr); 6062 } 6063 } 6064 6065 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6066 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6067 6068 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6069 MemRegion span, CMSBitMap* bitMap): 6070 _span(span), 6071 _bitMap(bitMap) 6072 { 6073 assert(_ref_processor == NULL, "deliberately left NULL"); 6074 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6075 } 6076 6077 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6078 // if p points into _span, then mark corresponding bit in _markBitMap 6079 assert(obj->is_oop(), "expected an oop"); 6080 HeapWord* addr = (HeapWord*)obj; 6081 if (_span.contains(addr)) { 6082 // this should be made more efficient 6083 _bitMap->par_mark(addr); 6084 } 6085 } 6086 6087 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6088 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6089 6090 // A variant of the above, used for CMS marking verification. 6091 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6092 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6093 _span(span), 6094 _verification_bm(verification_bm), 6095 _cms_bm(cms_bm) 6096 { 6097 assert(_ref_processor == NULL, "deliberately left NULL"); 6098 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6099 } 6100 6101 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6102 // if p points into _span, then mark corresponding bit in _markBitMap 6103 assert(obj->is_oop(), "expected an oop"); 6104 HeapWord* addr = (HeapWord*)obj; 6105 if (_span.contains(addr)) { 6106 _verification_bm->mark(addr); 6107 if (!_cms_bm->isMarked(addr)) { 6108 oop(addr)->print(); 6109 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6110 fatal("... aborting"); 6111 } 6112 } 6113 } 6114 6115 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6116 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6117 6118 ////////////////////////////////////////////////// 6119 // MarkRefsIntoAndScanClosure 6120 ////////////////////////////////////////////////// 6121 6122 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6123 ReferenceProcessor* rp, 6124 CMSBitMap* bit_map, 6125 CMSBitMap* mod_union_table, 6126 CMSMarkStack* mark_stack, 6127 CMSCollector* collector, 6128 bool should_yield, 6129 bool concurrent_precleaning): 6130 _collector(collector), 6131 _span(span), 6132 _bit_map(bit_map), 6133 _mark_stack(mark_stack), 6134 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6135 mark_stack, concurrent_precleaning), 6136 _yield(should_yield), 6137 _concurrent_precleaning(concurrent_precleaning), 6138 _freelistLock(NULL) 6139 { 6140 _ref_processor = rp; 6141 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6142 } 6143 6144 // This closure is used to mark refs into the CMS generation at the 6145 // second (final) checkpoint, and to scan and transitively follow 6146 // the unmarked oops. It is also used during the concurrent precleaning 6147 // phase while scanning objects on dirty cards in the CMS generation. 6148 // The marks are made in the marking bit map and the marking stack is 6149 // used for keeping the (newly) grey objects during the scan. 6150 // The parallel version (Par_...) appears further below. 6151 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6152 if (obj != NULL) { 6153 assert(obj->is_oop(), "expected an oop"); 6154 HeapWord* addr = (HeapWord*)obj; 6155 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6156 assert(_collector->overflow_list_is_empty(), 6157 "overflow list should be empty"); 6158 if (_span.contains(addr) && 6159 !_bit_map->isMarked(addr)) { 6160 // mark bit map (object is now grey) 6161 _bit_map->mark(addr); 6162 // push on marking stack (stack should be empty), and drain the 6163 // stack by applying this closure to the oops in the oops popped 6164 // from the stack (i.e. blacken the grey objects) 6165 bool res = _mark_stack->push(obj); 6166 assert(res, "Should have space to push on empty stack"); 6167 do { 6168 oop new_oop = _mark_stack->pop(); 6169 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6170 assert(_bit_map->isMarked((HeapWord*)new_oop), 6171 "only grey objects on this stack"); 6172 // iterate over the oops in this oop, marking and pushing 6173 // the ones in CMS heap (i.e. in _span). 6174 new_oop->oop_iterate(&_pushAndMarkClosure); 6175 // check if it's time to yield 6176 do_yield_check(); 6177 } while (!_mark_stack->isEmpty() || 6178 (!_concurrent_precleaning && take_from_overflow_list())); 6179 // if marking stack is empty, and we are not doing this 6180 // during precleaning, then check the overflow list 6181 } 6182 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6183 assert(_collector->overflow_list_is_empty(), 6184 "overflow list was drained above"); 6185 // We could restore evacuated mark words, if any, used for 6186 // overflow list links here because the overflow list is 6187 // provably empty here. That would reduce the maximum 6188 // size requirements for preserved_{oop,mark}_stack. 6189 // But we'll just postpone it until we are all done 6190 // so we can just stream through. 6191 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6192 _collector->restore_preserved_marks_if_any(); 6193 assert(_collector->no_preserved_marks(), "No preserved marks"); 6194 } 6195 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6196 "All preserved marks should have been restored above"); 6197 } 6198 } 6199 6200 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6201 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6202 6203 void MarkRefsIntoAndScanClosure::do_yield_work() { 6204 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6205 "CMS thread should hold CMS token"); 6206 assert_lock_strong(_freelistLock); 6207 assert_lock_strong(_bit_map->lock()); 6208 // relinquish the free_list_lock and bitMaplock() 6209 _bit_map->lock()->unlock(); 6210 _freelistLock->unlock(); 6211 ConcurrentMarkSweepThread::desynchronize(true); 6212 _collector->stopTimer(); 6213 if (PrintCMSStatistics != 0) { 6214 _collector->incrementYields(); 6215 } 6216 6217 // See the comment in coordinator_yield() 6218 for (unsigned i = 0; 6219 i < CMSYieldSleepCount && 6220 ConcurrentMarkSweepThread::should_yield() && 6221 !CMSCollector::foregroundGCIsActive(); 6222 ++i) { 6223 os::sleep(Thread::current(), 1, false); 6224 } 6225 6226 ConcurrentMarkSweepThread::synchronize(true); 6227 _freelistLock->lock_without_safepoint_check(); 6228 _bit_map->lock()->lock_without_safepoint_check(); 6229 _collector->startTimer(); 6230 } 6231 6232 /////////////////////////////////////////////////////////// 6233 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6234 // MarkRefsIntoAndScanClosure 6235 /////////////////////////////////////////////////////////// 6236 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6237 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6238 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6239 _span(span), 6240 _bit_map(bit_map), 6241 _work_queue(work_queue), 6242 _low_water_mark(MIN2((work_queue->max_elems()/4), 6243 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6244 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6245 { 6246 _ref_processor = rp; 6247 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6248 } 6249 6250 // This closure is used to mark refs into the CMS generation at the 6251 // second (final) checkpoint, and to scan and transitively follow 6252 // the unmarked oops. The marks are made in the marking bit map and 6253 // the work_queue is used for keeping the (newly) grey objects during 6254 // the scan phase whence they are also available for stealing by parallel 6255 // threads. Since the marking bit map is shared, updates are 6256 // synchronized (via CAS). 6257 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6258 if (obj != NULL) { 6259 // Ignore mark word because this could be an already marked oop 6260 // that may be chained at the end of the overflow list. 6261 assert(obj->is_oop(true), "expected an oop"); 6262 HeapWord* addr = (HeapWord*)obj; 6263 if (_span.contains(addr) && 6264 !_bit_map->isMarked(addr)) { 6265 // mark bit map (object will become grey): 6266 // It is possible for several threads to be 6267 // trying to "claim" this object concurrently; 6268 // the unique thread that succeeds in marking the 6269 // object first will do the subsequent push on 6270 // to the work queue (or overflow list). 6271 if (_bit_map->par_mark(addr)) { 6272 // push on work_queue (which may not be empty), and trim the 6273 // queue to an appropriate length by applying this closure to 6274 // the oops in the oops popped from the stack (i.e. blacken the 6275 // grey objects) 6276 bool res = _work_queue->push(obj); 6277 assert(res, "Low water mark should be less than capacity?"); 6278 trim_queue(_low_water_mark); 6279 } // Else, another thread claimed the object 6280 } 6281 } 6282 } 6283 6284 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6285 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6286 6287 // This closure is used to rescan the marked objects on the dirty cards 6288 // in the mod union table and the card table proper. 6289 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6290 oop p, MemRegion mr) { 6291 6292 size_t size = 0; 6293 HeapWord* addr = (HeapWord*)p; 6294 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6295 assert(_span.contains(addr), "we are scanning the CMS generation"); 6296 // check if it's time to yield 6297 if (do_yield_check()) { 6298 // We yielded for some foreground stop-world work, 6299 // and we have been asked to abort this ongoing preclean cycle. 6300 return 0; 6301 } 6302 if (_bitMap->isMarked(addr)) { 6303 // it's marked; is it potentially uninitialized? 6304 if (p->klass_or_null() != NULL) { 6305 // an initialized object; ignore mark word in verification below 6306 // since we are running concurrent with mutators 6307 assert(p->is_oop(true), "should be an oop"); 6308 if (p->is_objArray()) { 6309 // objArrays are precisely marked; restrict scanning 6310 // to dirty cards only. 6311 size = CompactibleFreeListSpace::adjustObjectSize( 6312 p->oop_iterate_size(_scanningClosure, mr)); 6313 } else { 6314 // A non-array may have been imprecisely marked; we need 6315 // to scan object in its entirety. 6316 size = CompactibleFreeListSpace::adjustObjectSize( 6317 p->oop_iterate_size(_scanningClosure)); 6318 } 6319 #ifdef ASSERT 6320 size_t direct_size = 6321 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6322 assert(size == direct_size, "Inconsistency in size"); 6323 assert(size >= 3, "Necessary for Printezis marks to work"); 6324 if (!_bitMap->isMarked(addr+1)) { 6325 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6326 } else { 6327 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6328 assert(_bitMap->isMarked(addr+size-1), 6329 "inconsistent Printezis mark"); 6330 } 6331 #endif // ASSERT 6332 } else { 6333 // An uninitialized object. 6334 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6335 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6336 size = pointer_delta(nextOneAddr + 1, addr); 6337 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6338 "alignment problem"); 6339 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6340 // will dirty the card when the klass pointer is installed in the 6341 // object (signaling the completion of initialization). 6342 } 6343 } else { 6344 // Either a not yet marked object or an uninitialized object 6345 if (p->klass_or_null() == NULL) { 6346 // An uninitialized object, skip to the next card, since 6347 // we may not be able to read its P-bits yet. 6348 assert(size == 0, "Initial value"); 6349 } else { 6350 // An object not (yet) reached by marking: we merely need to 6351 // compute its size so as to go look at the next block. 6352 assert(p->is_oop(true), "should be an oop"); 6353 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6354 } 6355 } 6356 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6357 return size; 6358 } 6359 6360 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6361 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6362 "CMS thread should hold CMS token"); 6363 assert_lock_strong(_freelistLock); 6364 assert_lock_strong(_bitMap->lock()); 6365 // relinquish the free_list_lock and bitMaplock() 6366 _bitMap->lock()->unlock(); 6367 _freelistLock->unlock(); 6368 ConcurrentMarkSweepThread::desynchronize(true); 6369 _collector->stopTimer(); 6370 if (PrintCMSStatistics != 0) { 6371 _collector->incrementYields(); 6372 } 6373 6374 // See the comment in coordinator_yield() 6375 for (unsigned i = 0; i < CMSYieldSleepCount && 6376 ConcurrentMarkSweepThread::should_yield() && 6377 !CMSCollector::foregroundGCIsActive(); ++i) { 6378 os::sleep(Thread::current(), 1, false); 6379 } 6380 6381 ConcurrentMarkSweepThread::synchronize(true); 6382 _freelistLock->lock_without_safepoint_check(); 6383 _bitMap->lock()->lock_without_safepoint_check(); 6384 _collector->startTimer(); 6385 } 6386 6387 6388 ////////////////////////////////////////////////////////////////// 6389 // SurvivorSpacePrecleanClosure 6390 ////////////////////////////////////////////////////////////////// 6391 // This (single-threaded) closure is used to preclean the oops in 6392 // the survivor spaces. 6393 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6394 6395 HeapWord* addr = (HeapWord*)p; 6396 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6397 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6398 assert(p->klass_or_null() != NULL, "object should be initialized"); 6399 // an initialized object; ignore mark word in verification below 6400 // since we are running concurrent with mutators 6401 assert(p->is_oop(true), "should be an oop"); 6402 // Note that we do not yield while we iterate over 6403 // the interior oops of p, pushing the relevant ones 6404 // on our marking stack. 6405 size_t size = p->oop_iterate_size(_scanning_closure); 6406 do_yield_check(); 6407 // Observe that below, we do not abandon the preclean 6408 // phase as soon as we should; rather we empty the 6409 // marking stack before returning. This is to satisfy 6410 // some existing assertions. In general, it may be a 6411 // good idea to abort immediately and complete the marking 6412 // from the grey objects at a later time. 6413 while (!_mark_stack->isEmpty()) { 6414 oop new_oop = _mark_stack->pop(); 6415 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6416 assert(_bit_map->isMarked((HeapWord*)new_oop), 6417 "only grey objects on this stack"); 6418 // iterate over the oops in this oop, marking and pushing 6419 // the ones in CMS heap (i.e. in _span). 6420 new_oop->oop_iterate(_scanning_closure); 6421 // check if it's time to yield 6422 do_yield_check(); 6423 } 6424 unsigned int after_count = 6425 GenCollectedHeap::heap()->total_collections(); 6426 bool abort = (_before_count != after_count) || 6427 _collector->should_abort_preclean(); 6428 return abort ? 0 : size; 6429 } 6430 6431 void SurvivorSpacePrecleanClosure::do_yield_work() { 6432 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6433 "CMS thread should hold CMS token"); 6434 assert_lock_strong(_bit_map->lock()); 6435 // Relinquish the bit map lock 6436 _bit_map->lock()->unlock(); 6437 ConcurrentMarkSweepThread::desynchronize(true); 6438 _collector->stopTimer(); 6439 if (PrintCMSStatistics != 0) { 6440 _collector->incrementYields(); 6441 } 6442 6443 // See the comment in coordinator_yield() 6444 for (unsigned i = 0; i < CMSYieldSleepCount && 6445 ConcurrentMarkSweepThread::should_yield() && 6446 !CMSCollector::foregroundGCIsActive(); ++i) { 6447 os::sleep(Thread::current(), 1, false); 6448 } 6449 6450 ConcurrentMarkSweepThread::synchronize(true); 6451 _bit_map->lock()->lock_without_safepoint_check(); 6452 _collector->startTimer(); 6453 } 6454 6455 // This closure is used to rescan the marked objects on the dirty cards 6456 // in the mod union table and the card table proper. In the parallel 6457 // case, although the bitMap is shared, we do a single read so the 6458 // isMarked() query is "safe". 6459 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6460 // Ignore mark word because we are running concurrent with mutators 6461 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6462 HeapWord* addr = (HeapWord*)p; 6463 assert(_span.contains(addr), "we are scanning the CMS generation"); 6464 bool is_obj_array = false; 6465 #ifdef ASSERT 6466 if (!_parallel) { 6467 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6468 assert(_collector->overflow_list_is_empty(), 6469 "overflow list should be empty"); 6470 6471 } 6472 #endif // ASSERT 6473 if (_bit_map->isMarked(addr)) { 6474 // Obj arrays are precisely marked, non-arrays are not; 6475 // so we scan objArrays precisely and non-arrays in their 6476 // entirety. 6477 if (p->is_objArray()) { 6478 is_obj_array = true; 6479 if (_parallel) { 6480 p->oop_iterate(_par_scan_closure, mr); 6481 } else { 6482 p->oop_iterate(_scan_closure, mr); 6483 } 6484 } else { 6485 if (_parallel) { 6486 p->oop_iterate(_par_scan_closure); 6487 } else { 6488 p->oop_iterate(_scan_closure); 6489 } 6490 } 6491 } 6492 #ifdef ASSERT 6493 if (!_parallel) { 6494 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6495 assert(_collector->overflow_list_is_empty(), 6496 "overflow list should be empty"); 6497 6498 } 6499 #endif // ASSERT 6500 return is_obj_array; 6501 } 6502 6503 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6504 MemRegion span, 6505 CMSBitMap* bitMap, CMSMarkStack* markStack, 6506 bool should_yield, bool verifying): 6507 _collector(collector), 6508 _span(span), 6509 _bitMap(bitMap), 6510 _mut(&collector->_modUnionTable), 6511 _markStack(markStack), 6512 _yield(should_yield), 6513 _skipBits(0) 6514 { 6515 assert(_markStack->isEmpty(), "stack should be empty"); 6516 _finger = _bitMap->startWord(); 6517 _threshold = _finger; 6518 assert(_collector->_restart_addr == NULL, "Sanity check"); 6519 assert(_span.contains(_finger), "Out of bounds _finger?"); 6520 DEBUG_ONLY(_verifying = verifying;) 6521 } 6522 6523 void MarkFromRootsClosure::reset(HeapWord* addr) { 6524 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6525 assert(_span.contains(addr), "Out of bounds _finger?"); 6526 _finger = addr; 6527 _threshold = (HeapWord*)round_to( 6528 (intptr_t)_finger, CardTableModRefBS::card_size); 6529 } 6530 6531 // Should revisit to see if this should be restructured for 6532 // greater efficiency. 6533 bool MarkFromRootsClosure::do_bit(size_t offset) { 6534 if (_skipBits > 0) { 6535 _skipBits--; 6536 return true; 6537 } 6538 // convert offset into a HeapWord* 6539 HeapWord* addr = _bitMap->startWord() + offset; 6540 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6541 "address out of range"); 6542 assert(_bitMap->isMarked(addr), "tautology"); 6543 if (_bitMap->isMarked(addr+1)) { 6544 // this is an allocated but not yet initialized object 6545 assert(_skipBits == 0, "tautology"); 6546 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6547 oop p = oop(addr); 6548 if (p->klass_or_null() == NULL) { 6549 DEBUG_ONLY(if (!_verifying) {) 6550 // We re-dirty the cards on which this object lies and increase 6551 // the _threshold so that we'll come back to scan this object 6552 // during the preclean or remark phase. (CMSCleanOnEnter) 6553 if (CMSCleanOnEnter) { 6554 size_t sz = _collector->block_size_using_printezis_bits(addr); 6555 HeapWord* end_card_addr = (HeapWord*)round_to( 6556 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6557 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6558 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6559 // Bump _threshold to end_card_addr; note that 6560 // _threshold cannot possibly exceed end_card_addr, anyhow. 6561 // This prevents future clearing of the card as the scan proceeds 6562 // to the right. 6563 assert(_threshold <= end_card_addr, 6564 "Because we are just scanning into this object"); 6565 if (_threshold < end_card_addr) { 6566 _threshold = end_card_addr; 6567 } 6568 if (p->klass_or_null() != NULL) { 6569 // Redirty the range of cards... 6570 _mut->mark_range(redirty_range); 6571 } // ...else the setting of klass will dirty the card anyway. 6572 } 6573 DEBUG_ONLY(}) 6574 return true; 6575 } 6576 } 6577 scanOopsInOop(addr); 6578 return true; 6579 } 6580 6581 // We take a break if we've been at this for a while, 6582 // so as to avoid monopolizing the locks involved. 6583 void MarkFromRootsClosure::do_yield_work() { 6584 // First give up the locks, then yield, then re-lock 6585 // We should probably use a constructor/destructor idiom to 6586 // do this unlock/lock or modify the MutexUnlocker class to 6587 // serve our purpose. XXX 6588 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6589 "CMS thread should hold CMS token"); 6590 assert_lock_strong(_bitMap->lock()); 6591 _bitMap->lock()->unlock(); 6592 ConcurrentMarkSweepThread::desynchronize(true); 6593 _collector->stopTimer(); 6594 if (PrintCMSStatistics != 0) { 6595 _collector->incrementYields(); 6596 } 6597 6598 // See the comment in coordinator_yield() 6599 for (unsigned i = 0; i < CMSYieldSleepCount && 6600 ConcurrentMarkSweepThread::should_yield() && 6601 !CMSCollector::foregroundGCIsActive(); ++i) { 6602 os::sleep(Thread::current(), 1, false); 6603 } 6604 6605 ConcurrentMarkSweepThread::synchronize(true); 6606 _bitMap->lock()->lock_without_safepoint_check(); 6607 _collector->startTimer(); 6608 } 6609 6610 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6611 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6612 assert(_markStack->isEmpty(), 6613 "should drain stack to limit stack usage"); 6614 // convert ptr to an oop preparatory to scanning 6615 oop obj = oop(ptr); 6616 // Ignore mark word in verification below, since we 6617 // may be running concurrent with mutators. 6618 assert(obj->is_oop(true), "should be an oop"); 6619 assert(_finger <= ptr, "_finger runneth ahead"); 6620 // advance the finger to right end of this object 6621 _finger = ptr + obj->size(); 6622 assert(_finger > ptr, "we just incremented it above"); 6623 // On large heaps, it may take us some time to get through 6624 // the marking phase. During 6625 // this time it's possible that a lot of mutations have 6626 // accumulated in the card table and the mod union table -- 6627 // these mutation records are redundant until we have 6628 // actually traced into the corresponding card. 6629 // Here, we check whether advancing the finger would make 6630 // us cross into a new card, and if so clear corresponding 6631 // cards in the MUT (preclean them in the card-table in the 6632 // future). 6633 6634 DEBUG_ONLY(if (!_verifying) {) 6635 // The clean-on-enter optimization is disabled by default, 6636 // until we fix 6178663. 6637 if (CMSCleanOnEnter && (_finger > _threshold)) { 6638 // [_threshold, _finger) represents the interval 6639 // of cards to be cleared in MUT (or precleaned in card table). 6640 // The set of cards to be cleared is all those that overlap 6641 // with the interval [_threshold, _finger); note that 6642 // _threshold is always kept card-aligned but _finger isn't 6643 // always card-aligned. 6644 HeapWord* old_threshold = _threshold; 6645 assert(old_threshold == (HeapWord*)round_to( 6646 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6647 "_threshold should always be card-aligned"); 6648 _threshold = (HeapWord*)round_to( 6649 (intptr_t)_finger, CardTableModRefBS::card_size); 6650 MemRegion mr(old_threshold, _threshold); 6651 assert(!mr.is_empty(), "Control point invariant"); 6652 assert(_span.contains(mr), "Should clear within span"); 6653 _mut->clear_range(mr); 6654 } 6655 DEBUG_ONLY(}) 6656 // Note: the finger doesn't advance while we drain 6657 // the stack below. 6658 PushOrMarkClosure pushOrMarkClosure(_collector, 6659 _span, _bitMap, _markStack, 6660 _finger, this); 6661 bool res = _markStack->push(obj); 6662 assert(res, "Empty non-zero size stack should have space for single push"); 6663 while (!_markStack->isEmpty()) { 6664 oop new_oop = _markStack->pop(); 6665 // Skip verifying header mark word below because we are 6666 // running concurrent with mutators. 6667 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6668 // now scan this oop's oops 6669 new_oop->oop_iterate(&pushOrMarkClosure); 6670 do_yield_check(); 6671 } 6672 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6673 } 6674 6675 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6676 CMSCollector* collector, MemRegion span, 6677 CMSBitMap* bit_map, 6678 OopTaskQueue* work_queue, 6679 CMSMarkStack* overflow_stack): 6680 _collector(collector), 6681 _whole_span(collector->_span), 6682 _span(span), 6683 _bit_map(bit_map), 6684 _mut(&collector->_modUnionTable), 6685 _work_queue(work_queue), 6686 _overflow_stack(overflow_stack), 6687 _skip_bits(0), 6688 _task(task) 6689 { 6690 assert(_work_queue->size() == 0, "work_queue should be empty"); 6691 _finger = span.start(); 6692 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6693 assert(_span.contains(_finger), "Out of bounds _finger?"); 6694 } 6695 6696 // Should revisit to see if this should be restructured for 6697 // greater efficiency. 6698 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6699 if (_skip_bits > 0) { 6700 _skip_bits--; 6701 return true; 6702 } 6703 // convert offset into a HeapWord* 6704 HeapWord* addr = _bit_map->startWord() + offset; 6705 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6706 "address out of range"); 6707 assert(_bit_map->isMarked(addr), "tautology"); 6708 if (_bit_map->isMarked(addr+1)) { 6709 // this is an allocated object that might not yet be initialized 6710 assert(_skip_bits == 0, "tautology"); 6711 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6712 oop p = oop(addr); 6713 if (p->klass_or_null() == NULL) { 6714 // in the case of Clean-on-Enter optimization, redirty card 6715 // and avoid clearing card by increasing the threshold. 6716 return true; 6717 } 6718 } 6719 scan_oops_in_oop(addr); 6720 return true; 6721 } 6722 6723 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6724 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6725 // Should we assert that our work queue is empty or 6726 // below some drain limit? 6727 assert(_work_queue->size() == 0, 6728 "should drain stack to limit stack usage"); 6729 // convert ptr to an oop preparatory to scanning 6730 oop obj = oop(ptr); 6731 // Ignore mark word in verification below, since we 6732 // may be running concurrent with mutators. 6733 assert(obj->is_oop(true), "should be an oop"); 6734 assert(_finger <= ptr, "_finger runneth ahead"); 6735 // advance the finger to right end of this object 6736 _finger = ptr + obj->size(); 6737 assert(_finger > ptr, "we just incremented it above"); 6738 // On large heaps, it may take us some time to get through 6739 // the marking phase. During 6740 // this time it's possible that a lot of mutations have 6741 // accumulated in the card table and the mod union table -- 6742 // these mutation records are redundant until we have 6743 // actually traced into the corresponding card. 6744 // Here, we check whether advancing the finger would make 6745 // us cross into a new card, and if so clear corresponding 6746 // cards in the MUT (preclean them in the card-table in the 6747 // future). 6748 6749 // The clean-on-enter optimization is disabled by default, 6750 // until we fix 6178663. 6751 if (CMSCleanOnEnter && (_finger > _threshold)) { 6752 // [_threshold, _finger) represents the interval 6753 // of cards to be cleared in MUT (or precleaned in card table). 6754 // The set of cards to be cleared is all those that overlap 6755 // with the interval [_threshold, _finger); note that 6756 // _threshold is always kept card-aligned but _finger isn't 6757 // always card-aligned. 6758 HeapWord* old_threshold = _threshold; 6759 assert(old_threshold == (HeapWord*)round_to( 6760 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6761 "_threshold should always be card-aligned"); 6762 _threshold = (HeapWord*)round_to( 6763 (intptr_t)_finger, CardTableModRefBS::card_size); 6764 MemRegion mr(old_threshold, _threshold); 6765 assert(!mr.is_empty(), "Control point invariant"); 6766 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6767 _mut->clear_range(mr); 6768 } 6769 6770 // Note: the local finger doesn't advance while we drain 6771 // the stack below, but the global finger sure can and will. 6772 HeapWord** gfa = _task->global_finger_addr(); 6773 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6774 _span, _bit_map, 6775 _work_queue, 6776 _overflow_stack, 6777 _finger, 6778 gfa, this); 6779 bool res = _work_queue->push(obj); // overflow could occur here 6780 assert(res, "Will hold once we use workqueues"); 6781 while (true) { 6782 oop new_oop; 6783 if (!_work_queue->pop_local(new_oop)) { 6784 // We emptied our work_queue; check if there's stuff that can 6785 // be gotten from the overflow stack. 6786 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6787 _overflow_stack, _work_queue)) { 6788 do_yield_check(); 6789 continue; 6790 } else { // done 6791 break; 6792 } 6793 } 6794 // Skip verifying header mark word below because we are 6795 // running concurrent with mutators. 6796 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6797 // now scan this oop's oops 6798 new_oop->oop_iterate(&pushOrMarkClosure); 6799 do_yield_check(); 6800 } 6801 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6802 } 6803 6804 // Yield in response to a request from VM Thread or 6805 // from mutators. 6806 void Par_MarkFromRootsClosure::do_yield_work() { 6807 assert(_task != NULL, "sanity"); 6808 _task->yield(); 6809 } 6810 6811 // A variant of the above used for verifying CMS marking work. 6812 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6813 MemRegion span, 6814 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6815 CMSMarkStack* mark_stack): 6816 _collector(collector), 6817 _span(span), 6818 _verification_bm(verification_bm), 6819 _cms_bm(cms_bm), 6820 _mark_stack(mark_stack), 6821 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6822 mark_stack) 6823 { 6824 assert(_mark_stack->isEmpty(), "stack should be empty"); 6825 _finger = _verification_bm->startWord(); 6826 assert(_collector->_restart_addr == NULL, "Sanity check"); 6827 assert(_span.contains(_finger), "Out of bounds _finger?"); 6828 } 6829 6830 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6831 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6832 assert(_span.contains(addr), "Out of bounds _finger?"); 6833 _finger = addr; 6834 } 6835 6836 // Should revisit to see if this should be restructured for 6837 // greater efficiency. 6838 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6839 // convert offset into a HeapWord* 6840 HeapWord* addr = _verification_bm->startWord() + offset; 6841 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6842 "address out of range"); 6843 assert(_verification_bm->isMarked(addr), "tautology"); 6844 assert(_cms_bm->isMarked(addr), "tautology"); 6845 6846 assert(_mark_stack->isEmpty(), 6847 "should drain stack to limit stack usage"); 6848 // convert addr to an oop preparatory to scanning 6849 oop obj = oop(addr); 6850 assert(obj->is_oop(), "should be an oop"); 6851 assert(_finger <= addr, "_finger runneth ahead"); 6852 // advance the finger to right end of this object 6853 _finger = addr + obj->size(); 6854 assert(_finger > addr, "we just incremented it above"); 6855 // Note: the finger doesn't advance while we drain 6856 // the stack below. 6857 bool res = _mark_stack->push(obj); 6858 assert(res, "Empty non-zero size stack should have space for single push"); 6859 while (!_mark_stack->isEmpty()) { 6860 oop new_oop = _mark_stack->pop(); 6861 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6862 // now scan this oop's oops 6863 new_oop->oop_iterate(&_pam_verify_closure); 6864 } 6865 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6866 return true; 6867 } 6868 6869 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6870 CMSCollector* collector, MemRegion span, 6871 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6872 CMSMarkStack* mark_stack): 6873 MetadataAwareOopClosure(collector->ref_processor()), 6874 _collector(collector), 6875 _span(span), 6876 _verification_bm(verification_bm), 6877 _cms_bm(cms_bm), 6878 _mark_stack(mark_stack) 6879 { } 6880 6881 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6882 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6883 6884 // Upon stack overflow, we discard (part of) the stack, 6885 // remembering the least address amongst those discarded 6886 // in CMSCollector's _restart_address. 6887 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6888 // Remember the least grey address discarded 6889 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6890 _collector->lower_restart_addr(ra); 6891 _mark_stack->reset(); // discard stack contents 6892 _mark_stack->expand(); // expand the stack if possible 6893 } 6894 6895 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6896 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6897 HeapWord* addr = (HeapWord*)obj; 6898 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6899 // Oop lies in _span and isn't yet grey or black 6900 _verification_bm->mark(addr); // now grey 6901 if (!_cms_bm->isMarked(addr)) { 6902 oop(addr)->print(); 6903 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6904 p2i(addr)); 6905 fatal("... aborting"); 6906 } 6907 6908 if (!_mark_stack->push(obj)) { // stack overflow 6909 if (PrintCMSStatistics != 0) { 6910 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6911 SIZE_FORMAT, _mark_stack->capacity()); 6912 } 6913 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6914 handle_stack_overflow(addr); 6915 } 6916 // anything including and to the right of _finger 6917 // will be scanned as we iterate over the remainder of the 6918 // bit map 6919 } 6920 } 6921 6922 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6923 MemRegion span, 6924 CMSBitMap* bitMap, CMSMarkStack* markStack, 6925 HeapWord* finger, MarkFromRootsClosure* parent) : 6926 MetadataAwareOopClosure(collector->ref_processor()), 6927 _collector(collector), 6928 _span(span), 6929 _bitMap(bitMap), 6930 _markStack(markStack), 6931 _finger(finger), 6932 _parent(parent) 6933 { } 6934 6935 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6936 MemRegion span, 6937 CMSBitMap* bit_map, 6938 OopTaskQueue* work_queue, 6939 CMSMarkStack* overflow_stack, 6940 HeapWord* finger, 6941 HeapWord** global_finger_addr, 6942 Par_MarkFromRootsClosure* parent) : 6943 MetadataAwareOopClosure(collector->ref_processor()), 6944 _collector(collector), 6945 _whole_span(collector->_span), 6946 _span(span), 6947 _bit_map(bit_map), 6948 _work_queue(work_queue), 6949 _overflow_stack(overflow_stack), 6950 _finger(finger), 6951 _global_finger_addr(global_finger_addr), 6952 _parent(parent) 6953 { } 6954 6955 // Assumes thread-safe access by callers, who are 6956 // responsible for mutual exclusion. 6957 void CMSCollector::lower_restart_addr(HeapWord* low) { 6958 assert(_span.contains(low), "Out of bounds addr"); 6959 if (_restart_addr == NULL) { 6960 _restart_addr = low; 6961 } else { 6962 _restart_addr = MIN2(_restart_addr, low); 6963 } 6964 } 6965 6966 // Upon stack overflow, we discard (part of) the stack, 6967 // remembering the least address amongst those discarded 6968 // in CMSCollector's _restart_address. 6969 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6970 // Remember the least grey address discarded 6971 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6972 _collector->lower_restart_addr(ra); 6973 _markStack->reset(); // discard stack contents 6974 _markStack->expand(); // expand the stack if possible 6975 } 6976 6977 // Upon stack overflow, we discard (part of) the stack, 6978 // remembering the least address amongst those discarded 6979 // in CMSCollector's _restart_address. 6980 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6981 // We need to do this under a mutex to prevent other 6982 // workers from interfering with the work done below. 6983 MutexLockerEx ml(_overflow_stack->par_lock(), 6984 Mutex::_no_safepoint_check_flag); 6985 // Remember the least grey address discarded 6986 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6987 _collector->lower_restart_addr(ra); 6988 _overflow_stack->reset(); // discard stack contents 6989 _overflow_stack->expand(); // expand the stack if possible 6990 } 6991 6992 void PushOrMarkClosure::do_oop(oop obj) { 6993 // Ignore mark word because we are running concurrent with mutators. 6994 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6995 HeapWord* addr = (HeapWord*)obj; 6996 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6997 // Oop lies in _span and isn't yet grey or black 6998 _bitMap->mark(addr); // now grey 6999 if (addr < _finger) { 7000 // the bit map iteration has already either passed, or 7001 // sampled, this bit in the bit map; we'll need to 7002 // use the marking stack to scan this oop's oops. 7003 bool simulate_overflow = false; 7004 NOT_PRODUCT( 7005 if (CMSMarkStackOverflowALot && 7006 _collector->simulate_overflow()) { 7007 // simulate a stack overflow 7008 simulate_overflow = true; 7009 } 7010 ) 7011 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7012 if (PrintCMSStatistics != 0) { 7013 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7014 SIZE_FORMAT, _markStack->capacity()); 7015 } 7016 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7017 handle_stack_overflow(addr); 7018 } 7019 } 7020 // anything including and to the right of _finger 7021 // will be scanned as we iterate over the remainder of the 7022 // bit map 7023 do_yield_check(); 7024 } 7025 } 7026 7027 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7028 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7029 7030 void Par_PushOrMarkClosure::do_oop(oop obj) { 7031 // Ignore mark word because we are running concurrent with mutators. 7032 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7033 HeapWord* addr = (HeapWord*)obj; 7034 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7035 // Oop lies in _span and isn't yet grey or black 7036 // We read the global_finger (volatile read) strictly after marking oop 7037 bool res = _bit_map->par_mark(addr); // now grey 7038 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7039 // Should we push this marked oop on our stack? 7040 // -- if someone else marked it, nothing to do 7041 // -- if target oop is above global finger nothing to do 7042 // -- if target oop is in chunk and above local finger 7043 // then nothing to do 7044 // -- else push on work queue 7045 if ( !res // someone else marked it, they will deal with it 7046 || (addr >= *gfa) // will be scanned in a later task 7047 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7048 return; 7049 } 7050 // the bit map iteration has already either passed, or 7051 // sampled, this bit in the bit map; we'll need to 7052 // use the marking stack to scan this oop's oops. 7053 bool simulate_overflow = false; 7054 NOT_PRODUCT( 7055 if (CMSMarkStackOverflowALot && 7056 _collector->simulate_overflow()) { 7057 // simulate a stack overflow 7058 simulate_overflow = true; 7059 } 7060 ) 7061 if (simulate_overflow || 7062 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7063 // stack overflow 7064 if (PrintCMSStatistics != 0) { 7065 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7066 SIZE_FORMAT, _overflow_stack->capacity()); 7067 } 7068 // We cannot assert that the overflow stack is full because 7069 // it may have been emptied since. 7070 assert(simulate_overflow || 7071 _work_queue->size() == _work_queue->max_elems(), 7072 "Else push should have succeeded"); 7073 handle_stack_overflow(addr); 7074 } 7075 do_yield_check(); 7076 } 7077 } 7078 7079 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7080 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7081 7082 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7083 MemRegion span, 7084 ReferenceProcessor* rp, 7085 CMSBitMap* bit_map, 7086 CMSBitMap* mod_union_table, 7087 CMSMarkStack* mark_stack, 7088 bool concurrent_precleaning): 7089 MetadataAwareOopClosure(rp), 7090 _collector(collector), 7091 _span(span), 7092 _bit_map(bit_map), 7093 _mod_union_table(mod_union_table), 7094 _mark_stack(mark_stack), 7095 _concurrent_precleaning(concurrent_precleaning) 7096 { 7097 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7098 } 7099 7100 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7101 // the non-parallel version (the parallel version appears further below.) 7102 void PushAndMarkClosure::do_oop(oop obj) { 7103 // Ignore mark word verification. If during concurrent precleaning, 7104 // the object monitor may be locked. If during the checkpoint 7105 // phases, the object may already have been reached by a different 7106 // path and may be at the end of the global overflow list (so 7107 // the mark word may be NULL). 7108 assert(obj->is_oop_or_null(true /* ignore mark word */), 7109 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7110 HeapWord* addr = (HeapWord*)obj; 7111 // Check if oop points into the CMS generation 7112 // and is not marked 7113 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7114 // a white object ... 7115 _bit_map->mark(addr); // ... now grey 7116 // push on the marking stack (grey set) 7117 bool simulate_overflow = false; 7118 NOT_PRODUCT( 7119 if (CMSMarkStackOverflowALot && 7120 _collector->simulate_overflow()) { 7121 // simulate a stack overflow 7122 simulate_overflow = true; 7123 } 7124 ) 7125 if (simulate_overflow || !_mark_stack->push(obj)) { 7126 if (_concurrent_precleaning) { 7127 // During precleaning we can just dirty the appropriate card(s) 7128 // in the mod union table, thus ensuring that the object remains 7129 // in the grey set and continue. In the case of object arrays 7130 // we need to dirty all of the cards that the object spans, 7131 // since the rescan of object arrays will be limited to the 7132 // dirty cards. 7133 // Note that no one can be interfering with us in this action 7134 // of dirtying the mod union table, so no locking or atomics 7135 // are required. 7136 if (obj->is_objArray()) { 7137 size_t sz = obj->size(); 7138 HeapWord* end_card_addr = (HeapWord*)round_to( 7139 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7140 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7141 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7142 _mod_union_table->mark_range(redirty_range); 7143 } else { 7144 _mod_union_table->mark(addr); 7145 } 7146 _collector->_ser_pmc_preclean_ovflw++; 7147 } else { 7148 // During the remark phase, we need to remember this oop 7149 // in the overflow list. 7150 _collector->push_on_overflow_list(obj); 7151 _collector->_ser_pmc_remark_ovflw++; 7152 } 7153 } 7154 } 7155 } 7156 7157 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7158 MemRegion span, 7159 ReferenceProcessor* rp, 7160 CMSBitMap* bit_map, 7161 OopTaskQueue* work_queue): 7162 MetadataAwareOopClosure(rp), 7163 _collector(collector), 7164 _span(span), 7165 _bit_map(bit_map), 7166 _work_queue(work_queue) 7167 { 7168 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7169 } 7170 7171 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7172 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7173 7174 // Grey object rescan during second checkpoint phase -- 7175 // the parallel version. 7176 void Par_PushAndMarkClosure::do_oop(oop obj) { 7177 // In the assert below, we ignore the mark word because 7178 // this oop may point to an already visited object that is 7179 // on the overflow stack (in which case the mark word has 7180 // been hijacked for chaining into the overflow stack -- 7181 // if this is the last object in the overflow stack then 7182 // its mark word will be NULL). Because this object may 7183 // have been subsequently popped off the global overflow 7184 // stack, and the mark word possibly restored to the prototypical 7185 // value, by the time we get to examined this failing assert in 7186 // the debugger, is_oop_or_null(false) may subsequently start 7187 // to hold. 7188 assert(obj->is_oop_or_null(true), 7189 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7190 HeapWord* addr = (HeapWord*)obj; 7191 // Check if oop points into the CMS generation 7192 // and is not marked 7193 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7194 // a white object ... 7195 // If we manage to "claim" the object, by being the 7196 // first thread to mark it, then we push it on our 7197 // marking stack 7198 if (_bit_map->par_mark(addr)) { // ... now grey 7199 // push on work queue (grey set) 7200 bool simulate_overflow = false; 7201 NOT_PRODUCT( 7202 if (CMSMarkStackOverflowALot && 7203 _collector->par_simulate_overflow()) { 7204 // simulate a stack overflow 7205 simulate_overflow = true; 7206 } 7207 ) 7208 if (simulate_overflow || !_work_queue->push(obj)) { 7209 _collector->par_push_on_overflow_list(obj); 7210 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7211 } 7212 } // Else, some other thread got there first 7213 } 7214 } 7215 7216 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7217 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7218 7219 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7220 Mutex* bml = _collector->bitMapLock(); 7221 assert_lock_strong(bml); 7222 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7223 "CMS thread should hold CMS token"); 7224 7225 bml->unlock(); 7226 ConcurrentMarkSweepThread::desynchronize(true); 7227 7228 _collector->stopTimer(); 7229 if (PrintCMSStatistics != 0) { 7230 _collector->incrementYields(); 7231 } 7232 7233 // See the comment in coordinator_yield() 7234 for (unsigned i = 0; i < CMSYieldSleepCount && 7235 ConcurrentMarkSweepThread::should_yield() && 7236 !CMSCollector::foregroundGCIsActive(); ++i) { 7237 os::sleep(Thread::current(), 1, false); 7238 } 7239 7240 ConcurrentMarkSweepThread::synchronize(true); 7241 bml->lock(); 7242 7243 _collector->startTimer(); 7244 } 7245 7246 bool CMSPrecleanRefsYieldClosure::should_return() { 7247 if (ConcurrentMarkSweepThread::should_yield()) { 7248 do_yield_work(); 7249 } 7250 return _collector->foregroundGCIsActive(); 7251 } 7252 7253 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7254 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7255 "mr should be aligned to start at a card boundary"); 7256 // We'd like to assert: 7257 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7258 // "mr should be a range of cards"); 7259 // However, that would be too strong in one case -- the last 7260 // partition ends at _unallocated_block which, in general, can be 7261 // an arbitrary boundary, not necessarily card aligned. 7262 if (PrintCMSStatistics != 0) { 7263 _num_dirty_cards += 7264 mr.word_size()/CardTableModRefBS::card_size_in_words; 7265 } 7266 _space->object_iterate_mem(mr, &_scan_cl); 7267 } 7268 7269 SweepClosure::SweepClosure(CMSCollector* collector, 7270 ConcurrentMarkSweepGeneration* g, 7271 CMSBitMap* bitMap, bool should_yield) : 7272 _collector(collector), 7273 _g(g), 7274 _sp(g->cmsSpace()), 7275 _limit(_sp->sweep_limit()), 7276 _freelistLock(_sp->freelistLock()), 7277 _bitMap(bitMap), 7278 _yield(should_yield), 7279 _inFreeRange(false), // No free range at beginning of sweep 7280 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7281 _lastFreeRangeCoalesced(false), 7282 _freeFinger(g->used_region().start()) 7283 { 7284 NOT_PRODUCT( 7285 _numObjectsFreed = 0; 7286 _numWordsFreed = 0; 7287 _numObjectsLive = 0; 7288 _numWordsLive = 0; 7289 _numObjectsAlreadyFree = 0; 7290 _numWordsAlreadyFree = 0; 7291 _last_fc = NULL; 7292 7293 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7294 _sp->dictionary()->initialize_dict_returned_bytes(); 7295 ) 7296 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7297 "sweep _limit out of bounds"); 7298 if (CMSTraceSweeper) { 7299 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7300 p2i(_limit)); 7301 } 7302 } 7303 7304 void SweepClosure::print_on(outputStream* st) const { 7305 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7306 p2i(_sp->bottom()), p2i(_sp->end())); 7307 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7308 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7309 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7310 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7311 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7312 } 7313 7314 #ifndef PRODUCT 7315 // Assertion checking only: no useful work in product mode -- 7316 // however, if any of the flags below become product flags, 7317 // you may need to review this code to see if it needs to be 7318 // enabled in product mode. 7319 SweepClosure::~SweepClosure() { 7320 assert_lock_strong(_freelistLock); 7321 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7322 "sweep _limit out of bounds"); 7323 if (inFreeRange()) { 7324 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7325 print(); 7326 ShouldNotReachHere(); 7327 } 7328 if (Verbose && PrintGC) { 7329 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7330 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7331 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, " 7332 SIZE_FORMAT " bytes " 7333 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7334 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7335 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7336 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7337 * sizeof(HeapWord); 7338 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7339 7340 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7341 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7342 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7343 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7344 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes); 7345 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes", 7346 indexListReturnedBytes); 7347 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes", 7348 dict_returned_bytes); 7349 } 7350 } 7351 if (CMSTraceSweeper) { 7352 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7353 p2i(_limit)); 7354 } 7355 } 7356 #endif // PRODUCT 7357 7358 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7359 bool freeRangeInFreeLists) { 7360 if (CMSTraceSweeper) { 7361 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7362 p2i(freeFinger), freeRangeInFreeLists); 7363 } 7364 assert(!inFreeRange(), "Trampling existing free range"); 7365 set_inFreeRange(true); 7366 set_lastFreeRangeCoalesced(false); 7367 7368 set_freeFinger(freeFinger); 7369 set_freeRangeInFreeLists(freeRangeInFreeLists); 7370 if (CMSTestInFreeList) { 7371 if (freeRangeInFreeLists) { 7372 FreeChunk* fc = (FreeChunk*) freeFinger; 7373 assert(fc->is_free(), "A chunk on the free list should be free."); 7374 assert(fc->size() > 0, "Free range should have a size"); 7375 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7376 } 7377 } 7378 } 7379 7380 // Note that the sweeper runs concurrently with mutators. Thus, 7381 // it is possible for direct allocation in this generation to happen 7382 // in the middle of the sweep. Note that the sweeper also coalesces 7383 // contiguous free blocks. Thus, unless the sweeper and the allocator 7384 // synchronize appropriately freshly allocated blocks may get swept up. 7385 // This is accomplished by the sweeper locking the free lists while 7386 // it is sweeping. Thus blocks that are determined to be free are 7387 // indeed free. There is however one additional complication: 7388 // blocks that have been allocated since the final checkpoint and 7389 // mark, will not have been marked and so would be treated as 7390 // unreachable and swept up. To prevent this, the allocator marks 7391 // the bit map when allocating during the sweep phase. This leads, 7392 // however, to a further complication -- objects may have been allocated 7393 // but not yet initialized -- in the sense that the header isn't yet 7394 // installed. The sweeper can not then determine the size of the block 7395 // in order to skip over it. To deal with this case, we use a technique 7396 // (due to Printezis) to encode such uninitialized block sizes in the 7397 // bit map. Since the bit map uses a bit per every HeapWord, but the 7398 // CMS generation has a minimum object size of 3 HeapWords, it follows 7399 // that "normal marks" won't be adjacent in the bit map (there will 7400 // always be at least two 0 bits between successive 1 bits). We make use 7401 // of these "unused" bits to represent uninitialized blocks -- the bit 7402 // corresponding to the start of the uninitialized object and the next 7403 // bit are both set. Finally, a 1 bit marks the end of the object that 7404 // started with the two consecutive 1 bits to indicate its potentially 7405 // uninitialized state. 7406 7407 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7408 FreeChunk* fc = (FreeChunk*)addr; 7409 size_t res; 7410 7411 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7412 // than "addr == _limit" because although _limit was a block boundary when 7413 // we started the sweep, it may no longer be one because heap expansion 7414 // may have caused us to coalesce the block ending at the address _limit 7415 // with a newly expanded chunk (this happens when _limit was set to the 7416 // previous _end of the space), so we may have stepped past _limit: 7417 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7418 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7419 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7420 "sweep _limit out of bounds"); 7421 assert(addr < _sp->end(), "addr out of bounds"); 7422 // Flush any free range we might be holding as a single 7423 // coalesced chunk to the appropriate free list. 7424 if (inFreeRange()) { 7425 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7426 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()))); 7427 flush_cur_free_chunk(freeFinger(), 7428 pointer_delta(addr, freeFinger())); 7429 if (CMSTraceSweeper) { 7430 gclog_or_tty->print("Sweep: last chunk: "); 7431 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") " 7432 "[coalesced:%d]\n", 7433 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7434 lastFreeRangeCoalesced() ? 1 : 0); 7435 } 7436 } 7437 7438 // help the iterator loop finish 7439 return pointer_delta(_sp->end(), addr); 7440 } 7441 7442 assert(addr < _limit, "sweep invariant"); 7443 // check if we should yield 7444 do_yield_check(addr); 7445 if (fc->is_free()) { 7446 // Chunk that is already free 7447 res = fc->size(); 7448 do_already_free_chunk(fc); 7449 debug_only(_sp->verifyFreeLists()); 7450 // If we flush the chunk at hand in lookahead_and_flush() 7451 // and it's coalesced with a preceding chunk, then the 7452 // process of "mangling" the payload of the coalesced block 7453 // will cause erasure of the size information from the 7454 // (erstwhile) header of all the coalesced blocks but the 7455 // first, so the first disjunct in the assert will not hold 7456 // in that specific case (in which case the second disjunct 7457 // will hold). 7458 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7459 "Otherwise the size info doesn't change at this step"); 7460 NOT_PRODUCT( 7461 _numObjectsAlreadyFree++; 7462 _numWordsAlreadyFree += res; 7463 ) 7464 NOT_PRODUCT(_last_fc = fc;) 7465 } else if (!_bitMap->isMarked(addr)) { 7466 // Chunk is fresh garbage 7467 res = do_garbage_chunk(fc); 7468 debug_only(_sp->verifyFreeLists()); 7469 NOT_PRODUCT( 7470 _numObjectsFreed++; 7471 _numWordsFreed += res; 7472 ) 7473 } else { 7474 // Chunk that is alive. 7475 res = do_live_chunk(fc); 7476 debug_only(_sp->verifyFreeLists()); 7477 NOT_PRODUCT( 7478 _numObjectsLive++; 7479 _numWordsLive += res; 7480 ) 7481 } 7482 return res; 7483 } 7484 7485 // For the smart allocation, record following 7486 // split deaths - a free chunk is removed from its free list because 7487 // it is being split into two or more chunks. 7488 // split birth - a free chunk is being added to its free list because 7489 // a larger free chunk has been split and resulted in this free chunk. 7490 // coal death - a free chunk is being removed from its free list because 7491 // it is being coalesced into a large free chunk. 7492 // coal birth - a free chunk is being added to its free list because 7493 // it was created when two or more free chunks where coalesced into 7494 // this free chunk. 7495 // 7496 // These statistics are used to determine the desired number of free 7497 // chunks of a given size. The desired number is chosen to be relative 7498 // to the end of a CMS sweep. The desired number at the end of a sweep 7499 // is the 7500 // count-at-end-of-previous-sweep (an amount that was enough) 7501 // - count-at-beginning-of-current-sweep (the excess) 7502 // + split-births (gains in this size during interval) 7503 // - split-deaths (demands on this size during interval) 7504 // where the interval is from the end of one sweep to the end of the 7505 // next. 7506 // 7507 // When sweeping the sweeper maintains an accumulated chunk which is 7508 // the chunk that is made up of chunks that have been coalesced. That 7509 // will be termed the left-hand chunk. A new chunk of garbage that 7510 // is being considered for coalescing will be referred to as the 7511 // right-hand chunk. 7512 // 7513 // When making a decision on whether to coalesce a right-hand chunk with 7514 // the current left-hand chunk, the current count vs. the desired count 7515 // of the left-hand chunk is considered. Also if the right-hand chunk 7516 // is near the large chunk at the end of the heap (see 7517 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7518 // left-hand chunk is coalesced. 7519 // 7520 // When making a decision about whether to split a chunk, the desired count 7521 // vs. the current count of the candidate to be split is also considered. 7522 // If the candidate is underpopulated (currently fewer chunks than desired) 7523 // a chunk of an overpopulated (currently more chunks than desired) size may 7524 // be chosen. The "hint" associated with a free list, if non-null, points 7525 // to a free list which may be overpopulated. 7526 // 7527 7528 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7529 const size_t size = fc->size(); 7530 // Chunks that cannot be coalesced are not in the 7531 // free lists. 7532 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7533 assert(_sp->verify_chunk_in_free_list(fc), 7534 "free chunk should be in free lists"); 7535 } 7536 // a chunk that is already free, should not have been 7537 // marked in the bit map 7538 HeapWord* const addr = (HeapWord*) fc; 7539 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7540 // Verify that the bit map has no bits marked between 7541 // addr and purported end of this block. 7542 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7543 7544 // Some chunks cannot be coalesced under any circumstances. 7545 // See the definition of cantCoalesce(). 7546 if (!fc->cantCoalesce()) { 7547 // This chunk can potentially be coalesced. 7548 if (_sp->adaptive_freelists()) { 7549 // All the work is done in 7550 do_post_free_or_garbage_chunk(fc, size); 7551 } else { // Not adaptive free lists 7552 // this is a free chunk that can potentially be coalesced by the sweeper; 7553 if (!inFreeRange()) { 7554 // if the next chunk is a free block that can't be coalesced 7555 // it doesn't make sense to remove this chunk from the free lists 7556 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7557 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7558 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7559 nextChunk->is_free() && // ... which is free... 7560 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7561 // nothing to do 7562 } else { 7563 // Potentially the start of a new free range: 7564 // Don't eagerly remove it from the free lists. 7565 // No need to remove it if it will just be put 7566 // back again. (Also from a pragmatic point of view 7567 // if it is a free block in a region that is beyond 7568 // any allocated blocks, an assertion will fail) 7569 // Remember the start of a free run. 7570 initialize_free_range(addr, true); 7571 // end - can coalesce with next chunk 7572 } 7573 } else { 7574 // the midst of a free range, we are coalescing 7575 print_free_block_coalesced(fc); 7576 if (CMSTraceSweeper) { 7577 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7578 } 7579 // remove it from the free lists 7580 _sp->removeFreeChunkFromFreeLists(fc); 7581 set_lastFreeRangeCoalesced(true); 7582 // If the chunk is being coalesced and the current free range is 7583 // in the free lists, remove the current free range so that it 7584 // will be returned to the free lists in its entirety - all 7585 // the coalesced pieces included. 7586 if (freeRangeInFreeLists()) { 7587 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7588 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7589 "Size of free range is inconsistent with chunk size."); 7590 if (CMSTestInFreeList) { 7591 assert(_sp->verify_chunk_in_free_list(ffc), 7592 "free range is not in free lists"); 7593 } 7594 _sp->removeFreeChunkFromFreeLists(ffc); 7595 set_freeRangeInFreeLists(false); 7596 } 7597 } 7598 } 7599 // Note that if the chunk is not coalescable (the else arm 7600 // below), we unconditionally flush, without needing to do 7601 // a "lookahead," as we do below. 7602 if (inFreeRange()) lookahead_and_flush(fc, size); 7603 } else { 7604 // Code path common to both original and adaptive free lists. 7605 7606 // cant coalesce with previous block; this should be treated 7607 // as the end of a free run if any 7608 if (inFreeRange()) { 7609 // we kicked some butt; time to pick up the garbage 7610 assert(freeFinger() < addr, "freeFinger points too high"); 7611 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7612 } 7613 // else, nothing to do, just continue 7614 } 7615 } 7616 7617 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7618 // This is a chunk of garbage. It is not in any free list. 7619 // Add it to a free list or let it possibly be coalesced into 7620 // a larger chunk. 7621 HeapWord* const addr = (HeapWord*) fc; 7622 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7623 7624 if (_sp->adaptive_freelists()) { 7625 // Verify that the bit map has no bits marked between 7626 // addr and purported end of just dead object. 7627 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7628 7629 do_post_free_or_garbage_chunk(fc, size); 7630 } else { 7631 if (!inFreeRange()) { 7632 // start of a new free range 7633 assert(size > 0, "A free range should have a size"); 7634 initialize_free_range(addr, false); 7635 } else { 7636 // this will be swept up when we hit the end of the 7637 // free range 7638 if (CMSTraceSweeper) { 7639 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7640 } 7641 // If the chunk is being coalesced and the current free range is 7642 // in the free lists, remove the current free range so that it 7643 // will be returned to the free lists in its entirety - all 7644 // the coalesced pieces included. 7645 if (freeRangeInFreeLists()) { 7646 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7647 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7648 "Size of free range is inconsistent with chunk size."); 7649 if (CMSTestInFreeList) { 7650 assert(_sp->verify_chunk_in_free_list(ffc), 7651 "free range is not in free lists"); 7652 } 7653 _sp->removeFreeChunkFromFreeLists(ffc); 7654 set_freeRangeInFreeLists(false); 7655 } 7656 set_lastFreeRangeCoalesced(true); 7657 } 7658 // this will be swept up when we hit the end of the free range 7659 7660 // Verify that the bit map has no bits marked between 7661 // addr and purported end of just dead object. 7662 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7663 } 7664 assert(_limit >= addr + size, 7665 "A freshly garbage chunk can't possibly straddle over _limit"); 7666 if (inFreeRange()) lookahead_and_flush(fc, size); 7667 return size; 7668 } 7669 7670 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7671 HeapWord* addr = (HeapWord*) fc; 7672 // The sweeper has just found a live object. Return any accumulated 7673 // left hand chunk to the free lists. 7674 if (inFreeRange()) { 7675 assert(freeFinger() < addr, "freeFinger points too high"); 7676 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7677 } 7678 7679 // This object is live: we'd normally expect this to be 7680 // an oop, and like to assert the following: 7681 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7682 // However, as we commented above, this may be an object whose 7683 // header hasn't yet been initialized. 7684 size_t size; 7685 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7686 if (_bitMap->isMarked(addr + 1)) { 7687 // Determine the size from the bit map, rather than trying to 7688 // compute it from the object header. 7689 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7690 size = pointer_delta(nextOneAddr + 1, addr); 7691 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7692 "alignment problem"); 7693 7694 #ifdef ASSERT 7695 if (oop(addr)->klass_or_null() != NULL) { 7696 // Ignore mark word because we are running concurrent with mutators 7697 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7698 assert(size == 7699 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7700 "P-mark and computed size do not agree"); 7701 } 7702 #endif 7703 7704 } else { 7705 // This should be an initialized object that's alive. 7706 assert(oop(addr)->klass_or_null() != NULL, 7707 "Should be an initialized object"); 7708 // Ignore mark word because we are running concurrent with mutators 7709 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7710 // Verify that the bit map has no bits marked between 7711 // addr and purported end of this block. 7712 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7713 assert(size >= 3, "Necessary for Printezis marks to work"); 7714 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7715 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7716 } 7717 return size; 7718 } 7719 7720 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7721 size_t chunkSize) { 7722 // do_post_free_or_garbage_chunk() should only be called in the case 7723 // of the adaptive free list allocator. 7724 const bool fcInFreeLists = fc->is_free(); 7725 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7726 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7727 if (CMSTestInFreeList && fcInFreeLists) { 7728 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7729 } 7730 7731 if (CMSTraceSweeper) { 7732 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7733 } 7734 7735 HeapWord* const fc_addr = (HeapWord*) fc; 7736 7737 bool coalesce; 7738 const size_t left = pointer_delta(fc_addr, freeFinger()); 7739 const size_t right = chunkSize; 7740 switch (FLSCoalescePolicy) { 7741 // numeric value forms a coalition aggressiveness metric 7742 case 0: { // never coalesce 7743 coalesce = false; 7744 break; 7745 } 7746 case 1: { // coalesce if left & right chunks on overpopulated lists 7747 coalesce = _sp->coalOverPopulated(left) && 7748 _sp->coalOverPopulated(right); 7749 break; 7750 } 7751 case 2: { // coalesce if left chunk on overpopulated list (default) 7752 coalesce = _sp->coalOverPopulated(left); 7753 break; 7754 } 7755 case 3: { // coalesce if left OR right chunk on overpopulated list 7756 coalesce = _sp->coalOverPopulated(left) || 7757 _sp->coalOverPopulated(right); 7758 break; 7759 } 7760 case 4: { // always coalesce 7761 coalesce = true; 7762 break; 7763 } 7764 default: 7765 ShouldNotReachHere(); 7766 } 7767 7768 // Should the current free range be coalesced? 7769 // If the chunk is in a free range and either we decided to coalesce above 7770 // or the chunk is near the large block at the end of the heap 7771 // (isNearLargestChunk() returns true), then coalesce this chunk. 7772 const bool doCoalesce = inFreeRange() 7773 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7774 if (doCoalesce) { 7775 // Coalesce the current free range on the left with the new 7776 // chunk on the right. If either is on a free list, 7777 // it must be removed from the list and stashed in the closure. 7778 if (freeRangeInFreeLists()) { 7779 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7780 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7781 "Size of free range is inconsistent with chunk size."); 7782 if (CMSTestInFreeList) { 7783 assert(_sp->verify_chunk_in_free_list(ffc), 7784 "Chunk is not in free lists"); 7785 } 7786 _sp->coalDeath(ffc->size()); 7787 _sp->removeFreeChunkFromFreeLists(ffc); 7788 set_freeRangeInFreeLists(false); 7789 } 7790 if (fcInFreeLists) { 7791 _sp->coalDeath(chunkSize); 7792 assert(fc->size() == chunkSize, 7793 "The chunk has the wrong size or is not in the free lists"); 7794 _sp->removeFreeChunkFromFreeLists(fc); 7795 } 7796 set_lastFreeRangeCoalesced(true); 7797 print_free_block_coalesced(fc); 7798 } else { // not in a free range and/or should not coalesce 7799 // Return the current free range and start a new one. 7800 if (inFreeRange()) { 7801 // In a free range but cannot coalesce with the right hand chunk. 7802 // Put the current free range into the free lists. 7803 flush_cur_free_chunk(freeFinger(), 7804 pointer_delta(fc_addr, freeFinger())); 7805 } 7806 // Set up for new free range. Pass along whether the right hand 7807 // chunk is in the free lists. 7808 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7809 } 7810 } 7811 7812 // Lookahead flush: 7813 // If we are tracking a free range, and this is the last chunk that 7814 // we'll look at because its end crosses past _limit, we'll preemptively 7815 // flush it along with any free range we may be holding on to. Note that 7816 // this can be the case only for an already free or freshly garbage 7817 // chunk. If this block is an object, it can never straddle 7818 // over _limit. The "straddling" occurs when _limit is set at 7819 // the previous end of the space when this cycle started, and 7820 // a subsequent heap expansion caused the previously co-terminal 7821 // free block to be coalesced with the newly expanded portion, 7822 // thus rendering _limit a non-block-boundary making it dangerous 7823 // for the sweeper to step over and examine. 7824 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7825 assert(inFreeRange(), "Should only be called if currently in a free range."); 7826 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7827 assert(_sp->used_region().contains(eob - 1), 7828 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7829 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7830 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7831 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7832 if (eob >= _limit) { 7833 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7834 if (CMSTraceSweeper) { 7835 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7836 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7837 "[" PTR_FORMAT "," PTR_FORMAT ")", 7838 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7839 } 7840 // Return the storage we are tracking back into the free lists. 7841 if (CMSTraceSweeper) { 7842 gclog_or_tty->print_cr("Flushing ... "); 7843 } 7844 assert(freeFinger() < eob, "Error"); 7845 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7846 } 7847 } 7848 7849 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7850 assert(inFreeRange(), "Should only be called if currently in a free range."); 7851 assert(size > 0, 7852 "A zero sized chunk cannot be added to the free lists."); 7853 if (!freeRangeInFreeLists()) { 7854 if (CMSTestInFreeList) { 7855 FreeChunk* fc = (FreeChunk*) chunk; 7856 fc->set_size(size); 7857 assert(!_sp->verify_chunk_in_free_list(fc), 7858 "chunk should not be in free lists yet"); 7859 } 7860 if (CMSTraceSweeper) { 7861 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7862 p2i(chunk), size); 7863 } 7864 // A new free range is going to be starting. The current 7865 // free range has not been added to the free lists yet or 7866 // was removed so add it back. 7867 // If the current free range was coalesced, then the death 7868 // of the free range was recorded. Record a birth now. 7869 if (lastFreeRangeCoalesced()) { 7870 _sp->coalBirth(size); 7871 } 7872 _sp->addChunkAndRepairOffsetTable(chunk, size, 7873 lastFreeRangeCoalesced()); 7874 } else if (CMSTraceSweeper) { 7875 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7876 } 7877 set_inFreeRange(false); 7878 set_freeRangeInFreeLists(false); 7879 } 7880 7881 // We take a break if we've been at this for a while, 7882 // so as to avoid monopolizing the locks involved. 7883 void SweepClosure::do_yield_work(HeapWord* addr) { 7884 // Return current free chunk being used for coalescing (if any) 7885 // to the appropriate freelist. After yielding, the next 7886 // free block encountered will start a coalescing range of 7887 // free blocks. If the next free block is adjacent to the 7888 // chunk just flushed, they will need to wait for the next 7889 // sweep to be coalesced. 7890 if (inFreeRange()) { 7891 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7892 } 7893 7894 // First give up the locks, then yield, then re-lock. 7895 // We should probably use a constructor/destructor idiom to 7896 // do this unlock/lock or modify the MutexUnlocker class to 7897 // serve our purpose. XXX 7898 assert_lock_strong(_bitMap->lock()); 7899 assert_lock_strong(_freelistLock); 7900 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7901 "CMS thread should hold CMS token"); 7902 _bitMap->lock()->unlock(); 7903 _freelistLock->unlock(); 7904 ConcurrentMarkSweepThread::desynchronize(true); 7905 _collector->stopTimer(); 7906 if (PrintCMSStatistics != 0) { 7907 _collector->incrementYields(); 7908 } 7909 7910 // See the comment in coordinator_yield() 7911 for (unsigned i = 0; i < CMSYieldSleepCount && 7912 ConcurrentMarkSweepThread::should_yield() && 7913 !CMSCollector::foregroundGCIsActive(); ++i) { 7914 os::sleep(Thread::current(), 1, false); 7915 } 7916 7917 ConcurrentMarkSweepThread::synchronize(true); 7918 _freelistLock->lock(); 7919 _bitMap->lock()->lock_without_safepoint_check(); 7920 _collector->startTimer(); 7921 } 7922 7923 #ifndef PRODUCT 7924 // This is actually very useful in a product build if it can 7925 // be called from the debugger. Compile it into the product 7926 // as needed. 7927 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7928 return debug_cms_space->verify_chunk_in_free_list(fc); 7929 } 7930 #endif 7931 7932 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7933 if (CMSTraceSweeper) { 7934 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7935 p2i(fc), fc->size()); 7936 } 7937 } 7938 7939 // CMSIsAliveClosure 7940 bool CMSIsAliveClosure::do_object_b(oop obj) { 7941 HeapWord* addr = (HeapWord*)obj; 7942 return addr != NULL && 7943 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7944 } 7945 7946 7947 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7948 MemRegion span, 7949 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7950 bool cpc): 7951 _collector(collector), 7952 _span(span), 7953 _bit_map(bit_map), 7954 _mark_stack(mark_stack), 7955 _concurrent_precleaning(cpc) { 7956 assert(!_span.is_empty(), "Empty span could spell trouble"); 7957 } 7958 7959 7960 // CMSKeepAliveClosure: the serial version 7961 void CMSKeepAliveClosure::do_oop(oop obj) { 7962 HeapWord* addr = (HeapWord*)obj; 7963 if (_span.contains(addr) && 7964 !_bit_map->isMarked(addr)) { 7965 _bit_map->mark(addr); 7966 bool simulate_overflow = false; 7967 NOT_PRODUCT( 7968 if (CMSMarkStackOverflowALot && 7969 _collector->simulate_overflow()) { 7970 // simulate a stack overflow 7971 simulate_overflow = true; 7972 } 7973 ) 7974 if (simulate_overflow || !_mark_stack->push(obj)) { 7975 if (_concurrent_precleaning) { 7976 // We dirty the overflown object and let the remark 7977 // phase deal with it. 7978 assert(_collector->overflow_list_is_empty(), "Error"); 7979 // In the case of object arrays, we need to dirty all of 7980 // the cards that the object spans. No locking or atomics 7981 // are needed since no one else can be mutating the mod union 7982 // table. 7983 if (obj->is_objArray()) { 7984 size_t sz = obj->size(); 7985 HeapWord* end_card_addr = 7986 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7987 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7988 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7989 _collector->_modUnionTable.mark_range(redirty_range); 7990 } else { 7991 _collector->_modUnionTable.mark(addr); 7992 } 7993 _collector->_ser_kac_preclean_ovflw++; 7994 } else { 7995 _collector->push_on_overflow_list(obj); 7996 _collector->_ser_kac_ovflw++; 7997 } 7998 } 7999 } 8000 } 8001 8002 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8003 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8004 8005 // CMSParKeepAliveClosure: a parallel version of the above. 8006 // The work queues are private to each closure (thread), 8007 // but (may be) available for stealing by other threads. 8008 void CMSParKeepAliveClosure::do_oop(oop obj) { 8009 HeapWord* addr = (HeapWord*)obj; 8010 if (_span.contains(addr) && 8011 !_bit_map->isMarked(addr)) { 8012 // In general, during recursive tracing, several threads 8013 // may be concurrently getting here; the first one to 8014 // "tag" it, claims it. 8015 if (_bit_map->par_mark(addr)) { 8016 bool res = _work_queue->push(obj); 8017 assert(res, "Low water mark should be much less than capacity"); 8018 // Do a recursive trim in the hope that this will keep 8019 // stack usage lower, but leave some oops for potential stealers 8020 trim_queue(_low_water_mark); 8021 } // Else, another thread got there first 8022 } 8023 } 8024 8025 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8026 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8027 8028 void CMSParKeepAliveClosure::trim_queue(uint max) { 8029 while (_work_queue->size() > max) { 8030 oop new_oop; 8031 if (_work_queue->pop_local(new_oop)) { 8032 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8033 assert(_bit_map->isMarked((HeapWord*)new_oop), 8034 "no white objects on this stack!"); 8035 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8036 // iterate over the oops in this oop, marking and pushing 8037 // the ones in CMS heap (i.e. in _span). 8038 new_oop->oop_iterate(&_mark_and_push); 8039 } 8040 } 8041 } 8042 8043 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8044 CMSCollector* collector, 8045 MemRegion span, CMSBitMap* bit_map, 8046 OopTaskQueue* work_queue): 8047 _collector(collector), 8048 _span(span), 8049 _bit_map(bit_map), 8050 _work_queue(work_queue) { } 8051 8052 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8053 HeapWord* addr = (HeapWord*)obj; 8054 if (_span.contains(addr) && 8055 !_bit_map->isMarked(addr)) { 8056 if (_bit_map->par_mark(addr)) { 8057 bool simulate_overflow = false; 8058 NOT_PRODUCT( 8059 if (CMSMarkStackOverflowALot && 8060 _collector->par_simulate_overflow()) { 8061 // simulate a stack overflow 8062 simulate_overflow = true; 8063 } 8064 ) 8065 if (simulate_overflow || !_work_queue->push(obj)) { 8066 _collector->par_push_on_overflow_list(obj); 8067 _collector->_par_kac_ovflw++; 8068 } 8069 } // Else another thread got there already 8070 } 8071 } 8072 8073 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8074 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8075 8076 ////////////////////////////////////////////////////////////////// 8077 // CMSExpansionCause ///////////////////////////// 8078 ////////////////////////////////////////////////////////////////// 8079 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8080 switch (cause) { 8081 case _no_expansion: 8082 return "No expansion"; 8083 case _satisfy_free_ratio: 8084 return "Free ratio"; 8085 case _satisfy_promotion: 8086 return "Satisfy promotion"; 8087 case _satisfy_allocation: 8088 return "allocation"; 8089 case _allocate_par_lab: 8090 return "Par LAB"; 8091 case _allocate_par_spooling_space: 8092 return "Par Spooling Space"; 8093 case _adaptive_size_policy: 8094 return "Ergonomics"; 8095 default: 8096 return "unknown"; 8097 } 8098 } 8099 8100 void CMSDrainMarkingStackClosure::do_void() { 8101 // the max number to take from overflow list at a time 8102 const size_t num = _mark_stack->capacity()/4; 8103 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8104 "Overflow list should be NULL during concurrent phases"); 8105 while (!_mark_stack->isEmpty() || 8106 // if stack is empty, check the overflow list 8107 _collector->take_from_overflow_list(num, _mark_stack)) { 8108 oop obj = _mark_stack->pop(); 8109 HeapWord* addr = (HeapWord*)obj; 8110 assert(_span.contains(addr), "Should be within span"); 8111 assert(_bit_map->isMarked(addr), "Should be marked"); 8112 assert(obj->is_oop(), "Should be an oop"); 8113 obj->oop_iterate(_keep_alive); 8114 } 8115 } 8116 8117 void CMSParDrainMarkingStackClosure::do_void() { 8118 // drain queue 8119 trim_queue(0); 8120 } 8121 8122 // Trim our work_queue so its length is below max at return 8123 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8124 while (_work_queue->size() > max) { 8125 oop new_oop; 8126 if (_work_queue->pop_local(new_oop)) { 8127 assert(new_oop->is_oop(), "Expected an oop"); 8128 assert(_bit_map->isMarked((HeapWord*)new_oop), 8129 "no white objects on this stack!"); 8130 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8131 // iterate over the oops in this oop, marking and pushing 8132 // the ones in CMS heap (i.e. in _span). 8133 new_oop->oop_iterate(&_mark_and_push); 8134 } 8135 } 8136 } 8137 8138 //////////////////////////////////////////////////////////////////// 8139 // Support for Marking Stack Overflow list handling and related code 8140 //////////////////////////////////////////////////////////////////// 8141 // Much of the following code is similar in shape and spirit to the 8142 // code used in ParNewGC. We should try and share that code 8143 // as much as possible in the future. 8144 8145 #ifndef PRODUCT 8146 // Debugging support for CMSStackOverflowALot 8147 8148 // It's OK to call this multi-threaded; the worst thing 8149 // that can happen is that we'll get a bunch of closely 8150 // spaced simulated overflows, but that's OK, in fact 8151 // probably good as it would exercise the overflow code 8152 // under contention. 8153 bool CMSCollector::simulate_overflow() { 8154 if (_overflow_counter-- <= 0) { // just being defensive 8155 _overflow_counter = CMSMarkStackOverflowInterval; 8156 return true; 8157 } else { 8158 return false; 8159 } 8160 } 8161 8162 bool CMSCollector::par_simulate_overflow() { 8163 return simulate_overflow(); 8164 } 8165 #endif 8166 8167 // Single-threaded 8168 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8169 assert(stack->isEmpty(), "Expected precondition"); 8170 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8171 size_t i = num; 8172 oop cur = _overflow_list; 8173 const markOop proto = markOopDesc::prototype(); 8174 NOT_PRODUCT(ssize_t n = 0;) 8175 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8176 next = oop(cur->mark()); 8177 cur->set_mark(proto); // until proven otherwise 8178 assert(cur->is_oop(), "Should be an oop"); 8179 bool res = stack->push(cur); 8180 assert(res, "Bit off more than can chew?"); 8181 NOT_PRODUCT(n++;) 8182 } 8183 _overflow_list = cur; 8184 #ifndef PRODUCT 8185 assert(_num_par_pushes >= n, "Too many pops?"); 8186 _num_par_pushes -=n; 8187 #endif 8188 return !stack->isEmpty(); 8189 } 8190 8191 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8192 // (MT-safe) Get a prefix of at most "num" from the list. 8193 // The overflow list is chained through the mark word of 8194 // each object in the list. We fetch the entire list, 8195 // break off a prefix of the right size and return the 8196 // remainder. If other threads try to take objects from 8197 // the overflow list at that time, they will wait for 8198 // some time to see if data becomes available. If (and 8199 // only if) another thread places one or more object(s) 8200 // on the global list before we have returned the suffix 8201 // to the global list, we will walk down our local list 8202 // to find its end and append the global list to 8203 // our suffix before returning it. This suffix walk can 8204 // prove to be expensive (quadratic in the amount of traffic) 8205 // when there are many objects in the overflow list and 8206 // there is much producer-consumer contention on the list. 8207 // *NOTE*: The overflow list manipulation code here and 8208 // in ParNewGeneration:: are very similar in shape, 8209 // except that in the ParNew case we use the old (from/eden) 8210 // copy of the object to thread the list via its klass word. 8211 // Because of the common code, if you make any changes in 8212 // the code below, please check the ParNew version to see if 8213 // similar changes might be needed. 8214 // CR 6797058 has been filed to consolidate the common code. 8215 bool CMSCollector::par_take_from_overflow_list(size_t num, 8216 OopTaskQueue* work_q, 8217 int no_of_gc_threads) { 8218 assert(work_q->size() == 0, "First empty local work queue"); 8219 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8220 if (_overflow_list == NULL) { 8221 return false; 8222 } 8223 // Grab the entire list; we'll put back a suffix 8224 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8225 Thread* tid = Thread::current(); 8226 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8227 // set to ParallelGCThreads. 8228 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8229 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8230 // If the list is busy, we spin for a short while, 8231 // sleeping between attempts to get the list. 8232 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8233 os::sleep(tid, sleep_time_millis, false); 8234 if (_overflow_list == NULL) { 8235 // Nothing left to take 8236 return false; 8237 } else if (_overflow_list != BUSY) { 8238 // Try and grab the prefix 8239 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8240 } 8241 } 8242 // If the list was found to be empty, or we spun long 8243 // enough, we give up and return empty-handed. If we leave 8244 // the list in the BUSY state below, it must be the case that 8245 // some other thread holds the overflow list and will set it 8246 // to a non-BUSY state in the future. 8247 if (prefix == NULL || prefix == BUSY) { 8248 // Nothing to take or waited long enough 8249 if (prefix == NULL) { 8250 // Write back the NULL in case we overwrote it with BUSY above 8251 // and it is still the same value. 8252 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8253 } 8254 return false; 8255 } 8256 assert(prefix != NULL && prefix != BUSY, "Error"); 8257 size_t i = num; 8258 oop cur = prefix; 8259 // Walk down the first "num" objects, unless we reach the end. 8260 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8261 if (cur->mark() == NULL) { 8262 // We have "num" or fewer elements in the list, so there 8263 // is nothing to return to the global list. 8264 // Write back the NULL in lieu of the BUSY we wrote 8265 // above, if it is still the same value. 8266 if (_overflow_list == BUSY) { 8267 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8268 } 8269 } else { 8270 // Chop off the suffix and return it to the global list. 8271 assert(cur->mark() != BUSY, "Error"); 8272 oop suffix_head = cur->mark(); // suffix will be put back on global list 8273 cur->set_mark(NULL); // break off suffix 8274 // It's possible that the list is still in the empty(busy) state 8275 // we left it in a short while ago; in that case we may be 8276 // able to place back the suffix without incurring the cost 8277 // of a walk down the list. 8278 oop observed_overflow_list = _overflow_list; 8279 oop cur_overflow_list = observed_overflow_list; 8280 bool attached = false; 8281 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8282 observed_overflow_list = 8283 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8284 if (cur_overflow_list == observed_overflow_list) { 8285 attached = true; 8286 break; 8287 } else cur_overflow_list = observed_overflow_list; 8288 } 8289 if (!attached) { 8290 // Too bad, someone else sneaked in (at least) an element; we'll need 8291 // to do a splice. Find tail of suffix so we can prepend suffix to global 8292 // list. 8293 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8294 oop suffix_tail = cur; 8295 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8296 "Tautology"); 8297 observed_overflow_list = _overflow_list; 8298 do { 8299 cur_overflow_list = observed_overflow_list; 8300 if (cur_overflow_list != BUSY) { 8301 // Do the splice ... 8302 suffix_tail->set_mark(markOop(cur_overflow_list)); 8303 } else { // cur_overflow_list == BUSY 8304 suffix_tail->set_mark(NULL); 8305 } 8306 // ... and try to place spliced list back on overflow_list ... 8307 observed_overflow_list = 8308 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8309 } while (cur_overflow_list != observed_overflow_list); 8310 // ... until we have succeeded in doing so. 8311 } 8312 } 8313 8314 // Push the prefix elements on work_q 8315 assert(prefix != NULL, "control point invariant"); 8316 const markOop proto = markOopDesc::prototype(); 8317 oop next; 8318 NOT_PRODUCT(ssize_t n = 0;) 8319 for (cur = prefix; cur != NULL; cur = next) { 8320 next = oop(cur->mark()); 8321 cur->set_mark(proto); // until proven otherwise 8322 assert(cur->is_oop(), "Should be an oop"); 8323 bool res = work_q->push(cur); 8324 assert(res, "Bit off more than we can chew?"); 8325 NOT_PRODUCT(n++;) 8326 } 8327 #ifndef PRODUCT 8328 assert(_num_par_pushes >= n, "Too many pops?"); 8329 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8330 #endif 8331 return true; 8332 } 8333 8334 // Single-threaded 8335 void CMSCollector::push_on_overflow_list(oop p) { 8336 NOT_PRODUCT(_num_par_pushes++;) 8337 assert(p->is_oop(), "Not an oop"); 8338 preserve_mark_if_necessary(p); 8339 p->set_mark((markOop)_overflow_list); 8340 _overflow_list = p; 8341 } 8342 8343 // Multi-threaded; use CAS to prepend to overflow list 8344 void CMSCollector::par_push_on_overflow_list(oop p) { 8345 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8346 assert(p->is_oop(), "Not an oop"); 8347 par_preserve_mark_if_necessary(p); 8348 oop observed_overflow_list = _overflow_list; 8349 oop cur_overflow_list; 8350 do { 8351 cur_overflow_list = observed_overflow_list; 8352 if (cur_overflow_list != BUSY) { 8353 p->set_mark(markOop(cur_overflow_list)); 8354 } else { 8355 p->set_mark(NULL); 8356 } 8357 observed_overflow_list = 8358 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8359 } while (cur_overflow_list != observed_overflow_list); 8360 } 8361 #undef BUSY 8362 8363 // Single threaded 8364 // General Note on GrowableArray: pushes may silently fail 8365 // because we are (temporarily) out of C-heap for expanding 8366 // the stack. The problem is quite ubiquitous and affects 8367 // a lot of code in the JVM. The prudent thing for GrowableArray 8368 // to do (for now) is to exit with an error. However, that may 8369 // be too draconian in some cases because the caller may be 8370 // able to recover without much harm. For such cases, we 8371 // should probably introduce a "soft_push" method which returns 8372 // an indication of success or failure with the assumption that 8373 // the caller may be able to recover from a failure; code in 8374 // the VM can then be changed, incrementally, to deal with such 8375 // failures where possible, thus, incrementally hardening the VM 8376 // in such low resource situations. 8377 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8378 _preserved_oop_stack.push(p); 8379 _preserved_mark_stack.push(m); 8380 assert(m == p->mark(), "Mark word changed"); 8381 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8382 "bijection"); 8383 } 8384 8385 // Single threaded 8386 void CMSCollector::preserve_mark_if_necessary(oop p) { 8387 markOop m = p->mark(); 8388 if (m->must_be_preserved(p)) { 8389 preserve_mark_work(p, m); 8390 } 8391 } 8392 8393 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8394 markOop m = p->mark(); 8395 if (m->must_be_preserved(p)) { 8396 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8397 // Even though we read the mark word without holding 8398 // the lock, we are assured that it will not change 8399 // because we "own" this oop, so no other thread can 8400 // be trying to push it on the overflow list; see 8401 // the assertion in preserve_mark_work() that checks 8402 // that m == p->mark(). 8403 preserve_mark_work(p, m); 8404 } 8405 } 8406 8407 // We should be able to do this multi-threaded, 8408 // a chunk of stack being a task (this is 8409 // correct because each oop only ever appears 8410 // once in the overflow list. However, it's 8411 // not very easy to completely overlap this with 8412 // other operations, so will generally not be done 8413 // until all work's been completed. Because we 8414 // expect the preserved oop stack (set) to be small, 8415 // it's probably fine to do this single-threaded. 8416 // We can explore cleverer concurrent/overlapped/parallel 8417 // processing of preserved marks if we feel the 8418 // need for this in the future. Stack overflow should 8419 // be so rare in practice and, when it happens, its 8420 // effect on performance so great that this will 8421 // likely just be in the noise anyway. 8422 void CMSCollector::restore_preserved_marks_if_any() { 8423 assert(SafepointSynchronize::is_at_safepoint(), 8424 "world should be stopped"); 8425 assert(Thread::current()->is_ConcurrentGC_thread() || 8426 Thread::current()->is_VM_thread(), 8427 "should be single-threaded"); 8428 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8429 "bijection"); 8430 8431 while (!_preserved_oop_stack.is_empty()) { 8432 oop p = _preserved_oop_stack.pop(); 8433 assert(p->is_oop(), "Should be an oop"); 8434 assert(_span.contains(p), "oop should be in _span"); 8435 assert(p->mark() == markOopDesc::prototype(), 8436 "Set when taken from overflow list"); 8437 markOop m = _preserved_mark_stack.pop(); 8438 p->set_mark(m); 8439 } 8440 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8441 "stacks were cleared above"); 8442 } 8443 8444 #ifndef PRODUCT 8445 bool CMSCollector::no_preserved_marks() const { 8446 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8447 } 8448 #endif 8449 8450 // Transfer some number of overflown objects to usual marking 8451 // stack. Return true if some objects were transferred. 8452 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8453 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8454 (size_t)ParGCDesiredObjsFromOverflowList); 8455 8456 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8457 assert(_collector->overflow_list_is_empty() || res, 8458 "If list is not empty, we should have taken something"); 8459 assert(!res || !_mark_stack->isEmpty(), 8460 "If we took something, it should now be on our stack"); 8461 return res; 8462 } 8463 8464 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8465 size_t res = _sp->block_size_no_stall(addr, _collector); 8466 if (_sp->block_is_obj(addr)) { 8467 if (_live_bit_map->isMarked(addr)) { 8468 // It can't have been dead in a previous cycle 8469 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8470 } else { 8471 _dead_bit_map->mark(addr); // mark the dead object 8472 } 8473 } 8474 // Could be 0, if the block size could not be computed without stalling. 8475 return res; 8476 } 8477 8478 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8479 8480 switch (phase) { 8481 case CMSCollector::InitialMarking: 8482 initialize(true /* fullGC */ , 8483 cause /* cause of the GC */, 8484 true /* recordGCBeginTime */, 8485 true /* recordPreGCUsage */, 8486 false /* recordPeakUsage */, 8487 false /* recordPostGCusage */, 8488 true /* recordAccumulatedGCTime */, 8489 false /* recordGCEndTime */, 8490 false /* countCollection */ ); 8491 break; 8492 8493 case CMSCollector::FinalMarking: 8494 initialize(true /* fullGC */ , 8495 cause /* cause of the GC */, 8496 false /* recordGCBeginTime */, 8497 false /* recordPreGCUsage */, 8498 false /* recordPeakUsage */, 8499 false /* recordPostGCusage */, 8500 true /* recordAccumulatedGCTime */, 8501 false /* recordGCEndTime */, 8502 false /* countCollection */ ); 8503 break; 8504 8505 case CMSCollector::Sweeping: 8506 initialize(true /* fullGC */ , 8507 cause /* cause of the GC */, 8508 false /* recordGCBeginTime */, 8509 false /* recordPreGCUsage */, 8510 true /* recordPeakUsage */, 8511 true /* recordPostGCusage */, 8512 false /* recordAccumulatedGCTime */, 8513 true /* recordGCEndTime */, 8514 true /* countCollection */ ); 8515 break; 8516 8517 default: 8518 ShouldNotReachHere(); 8519 } 8520 }