concurrentMarkSweepGeneration.cpp revision 11945:6d3c44100184
1/* 2 * Copyright (c) 2001, 2016, 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/symbolTable.hpp" 29#include "classfile/systemDictionary.hpp" 30#include "code/codeCache.hpp" 31#include "gc/cms/cmsCollectorPolicy.hpp" 32#include "gc/cms/cmsOopClosures.inline.hpp" 33#include "gc/cms/compactibleFreeListSpace.hpp" 34#include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35#include "gc/cms/concurrentMarkSweepThread.hpp" 36#include "gc/cms/parNewGeneration.hpp" 37#include "gc/cms/vmCMSOperations.hpp" 38#include "gc/serial/genMarkSweep.hpp" 39#include "gc/serial/tenuredGeneration.hpp" 40#include "gc/shared/adaptiveSizePolicy.hpp" 41#include "gc/shared/cardGeneration.inline.hpp" 42#include "gc/shared/cardTableRS.hpp" 43#include "gc/shared/collectedHeap.inline.hpp" 44#include "gc/shared/collectorCounters.hpp" 45#include "gc/shared/collectorPolicy.hpp" 46#include "gc/shared/gcLocker.inline.hpp" 47#include "gc/shared/gcPolicyCounters.hpp" 48#include "gc/shared/gcTimer.hpp" 49#include "gc/shared/gcTrace.hpp" 50#include "gc/shared/gcTraceTime.inline.hpp" 51#include "gc/shared/genCollectedHeap.hpp" 52#include "gc/shared/genOopClosures.inline.hpp" 53#include "gc/shared/isGCActiveMark.hpp" 54#include "gc/shared/referencePolicy.hpp" 55#include "gc/shared/strongRootsScope.hpp" 56#include "gc/shared/taskqueue.inline.hpp" 57#include "logging/log.hpp" 58#include "memory/allocation.hpp" 59#include "memory/iterator.inline.hpp" 60#include "memory/padded.hpp" 61#include "memory/resourceArea.hpp" 62#include "oops/oop.inline.hpp" 63#include "prims/jvmtiExport.hpp" 64#include "runtime/atomic.hpp" 65#include "runtime/globals_extension.hpp" 66#include "runtime/handles.inline.hpp" 67#include "runtime/java.hpp" 68#include "runtime/orderAccess.inline.hpp" 69#include "runtime/timer.hpp" 70#include "runtime/vmThread.hpp" 71#include "services/memoryService.hpp" 72#include "services/runtimeService.hpp" 73#include "utilities/stack.inline.hpp" 74 75// statics 76CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77bool CMSCollector::_full_gc_requested = false; 78GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80////////////////////////////////////////////////////////////////// 81// In support of CMS/VM thread synchronization 82////////////////////////////////////////////////////////////////// 83// We split use of the CGC_lock into 2 "levels". 84// The low-level locking is of the usual CGC_lock monitor. We introduce 85// a higher level "token" (hereafter "CMS token") built on top of the 86// low level monitor (hereafter "CGC lock"). 87// The token-passing protocol gives priority to the VM thread. The 88// CMS-lock doesn't provide any fairness guarantees, but clients 89// should ensure that it is only held for very short, bounded 90// durations. 91// 92// When either of the CMS thread or the VM thread is involved in 93// collection operations during which it does not want the other 94// thread to interfere, it obtains the CMS token. 95// 96// If either thread tries to get the token while the other has 97// it, that thread waits. However, if the VM thread and CMS thread 98// both want the token, then the VM thread gets priority while the 99// CMS thread waits. This ensures, for instance, that the "concurrent" 100// phases of the CMS thread's work do not block out the VM thread 101// for long periods of time as the CMS thread continues to hog 102// the token. (See bug 4616232). 103// 104// The baton-passing functions are, however, controlled by the 105// flags _foregroundGCShouldWait and _foregroundGCIsActive, 106// and here the low-level CMS lock, not the high level token, 107// ensures mutual exclusion. 108// 109// Two important conditions that we have to satisfy: 110// 1. if a thread does a low-level wait on the CMS lock, then it 111// relinquishes the CMS token if it were holding that token 112// when it acquired the low-level CMS lock. 113// 2. any low-level notifications on the low-level lock 114// should only be sent when a thread has relinquished the token. 115// 116// In the absence of either property, we'd have potential deadlock. 117// 118// We protect each of the CMS (concurrent and sequential) phases 119// with the CMS _token_, not the CMS _lock_. 120// 121// The only code protected by CMS lock is the token acquisition code 122// itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123// baton-passing code. 124// 125// Unfortunately, i couldn't come up with a good abstraction to factor and 126// hide the naked CGC_lock manipulation in the baton-passing code 127// further below. That's something we should try to do. Also, the proof 128// of correctness of this 2-level locking scheme is far from obvious, 129// and potentially quite slippery. We have an uneasy suspicion, for instance, 130// that there may be a theoretical possibility of delay/starvation in the 131// low-level lock/wait/notify scheme used for the baton-passing because of 132// potential interference with the priority scheme embodied in the 133// CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134// invocation further below and marked with "XXX 20011219YSR". 135// Indeed, as we note elsewhere, this may become yet more slippery 136// in the presence of multiple CMS and/or multiple VM threads. XXX 137 138class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156}; 157 158// Convenience class that does a CMSTokenSync, and then acquires 159// upto three locks. 160class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173}; 174 175 176////////////////////////////////////////////////////////////////// 177// Concurrent Mark-Sweep Generation ///////////////////////////// 178////////////////////////////////////////////////////////////////// 179 180NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182// This struct contains per-thread things necessary to support parallel 183// young-gen collection. 184class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193}; 194 195ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200{ 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247} 248 249 250// The field "_initiating_occupancy" represents the occupancy percentage 251// at which we trigger a new collection cycle. Unless explicitly specified 252// via CMSInitiatingOccupancyFraction (argument "io" below), it 253// is calculated by: 254// 255// Let "f" be MinHeapFreeRatio in 256// 257// _initiating_occupancy = 100-f + 258// f * (CMSTriggerRatio/100) 259// where CMSTriggerRatio is the argument "tr" below. 260// 261// That is, if we assume the heap is at its desired maximum occupancy at the 262// end of a collection, we let CMSTriggerRatio of the (purported) free 263// space be allocated before initiating a new collection cycle. 264// 265void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274} 275 276void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279} 280 281void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296} 297 298AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301} 302 303void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314} 315 316CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318{ 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340} 341 342double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345} 346 347void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348} 349 350// If promotion failure handling is on use 351// the padded average size of the promotion for each 352// young generation collection. 353double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379} 380 381// Compare the duration of the cms collection to the 382// time remaining before the cms generation is empty. 383// Note that the time from the start of the cms collection 384// to the start of the cms sweep (less than the total 385// duration of the cms collection) can be used. This 386// has been tried and some applications experienced 387// promotion failures early in execution. This was 388// possibly because the averages were not accurate 389// enough at the beginning. 390double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408} 409 410#ifndef PRODUCT 411void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->cr(); 429} 430#endif // #ifndef PRODUCT 431 432CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434bool CMSCollector::_foregroundGCIsActive = false; 435bool CMSCollector::_foregroundGCShouldWait = false; 436 437CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474#ifndef PRODUCT 475 _num_par_pushes(0), 476#endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490{ 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 log_warning(gc)("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 log_warning(gc)("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 549 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 550 551 // Parallel task queues; these are shared for the 552 // concurrent and stop-world phases of CMS, but 553 // are not shared with parallel scavenge (ParNew). 554 { 555 uint i; 556 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 557 558 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 559 || ParallelRefProcEnabled) 560 && num_queues > 0) { 561 _task_queues = new OopTaskQueueSet(num_queues); 562 if (_task_queues == NULL) { 563 log_warning(gc)("task_queues allocation failure."); 564 return; 565 } 566 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 567 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 568 for (i = 0; i < num_queues; i++) { 569 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 570 if (q == NULL) { 571 log_warning(gc)("work_queue allocation failure."); 572 return; 573 } 574 _task_queues->register_queue(i, q); 575 } 576 for (i = 0; i < num_queues; i++) { 577 _task_queues->queue(i)->initialize(); 578 _hash_seed[i] = 17; // copied from ParNew 579 } 580 } 581 } 582 583 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 584 585 // Clip CMSBootstrapOccupancy between 0 and 100. 586 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 587 588 // Now tell CMS generations the identity of their collector 589 ConcurrentMarkSweepGeneration::set_collector(this); 590 591 // Create & start a CMS thread for this CMS collector 592 _cmsThread = ConcurrentMarkSweepThread::start(this); 593 assert(cmsThread() != NULL, "CMS Thread should have been created"); 594 assert(cmsThread()->collector() == this, 595 "CMS Thread should refer to this gen"); 596 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 597 598 // Support for parallelizing young gen rescan 599 GenCollectedHeap* gch = GenCollectedHeap::heap(); 600 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 601 _young_gen = (ParNewGeneration*)gch->young_gen(); 602 if (gch->supports_inline_contig_alloc()) { 603 _top_addr = gch->top_addr(); 604 _end_addr = gch->end_addr(); 605 assert(_young_gen != NULL, "no _young_gen"); 606 _eden_chunk_index = 0; 607 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 608 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 609 } 610 611 // Support for parallelizing survivor space rescan 612 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 613 const size_t max_plab_samples = 614 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 615 616 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 617 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 618 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 619 _survivor_chunk_capacity = max_plab_samples; 620 for (uint i = 0; i < ParallelGCThreads; i++) { 621 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 622 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 623 assert(cur->end() == 0, "Should be 0"); 624 assert(cur->array() == vec, "Should be vec"); 625 assert(cur->capacity() == max_plab_samples, "Error"); 626 } 627 } 628 629 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 630 _gc_counters = new CollectorCounters("CMS", 1); 631 _completed_initialization = true; 632 _inter_sweep_timer.start(); // start of time 633} 634 635const char* ConcurrentMarkSweepGeneration::name() const { 636 return "concurrent mark-sweep generation"; 637} 638void ConcurrentMarkSweepGeneration::update_counters() { 639 if (UsePerfData) { 640 _space_counters->update_all(); 641 _gen_counters->update_all(); 642 } 643} 644 645// this is an optimized version of update_counters(). it takes the 646// used value as a parameter rather than computing it. 647// 648void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 649 if (UsePerfData) { 650 _space_counters->update_used(used); 651 _space_counters->update_capacity(); 652 _gen_counters->update_all(); 653 } 654} 655 656void ConcurrentMarkSweepGeneration::print() const { 657 Generation::print(); 658 cmsSpace()->print(); 659} 660 661#ifndef PRODUCT 662void ConcurrentMarkSweepGeneration::print_statistics() { 663 cmsSpace()->printFLCensus(0); 664} 665#endif 666 667size_t 668ConcurrentMarkSweepGeneration::contiguous_available() const { 669 // dld proposes an improvement in precision here. If the committed 670 // part of the space ends in a free block we should add that to 671 // uncommitted size in the calculation below. Will make this 672 // change later, staying with the approximation below for the 673 // time being. -- ysr. 674 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 675} 676 677size_t 678ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 679 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 680} 681 682size_t ConcurrentMarkSweepGeneration::max_available() const { 683 return free() + _virtual_space.uncommitted_size(); 684} 685 686bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 687 size_t available = max_available(); 688 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 689 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 690 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 691 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 692 return res; 693} 694 695// At a promotion failure dump information on block layout in heap 696// (cms old generation). 697void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 698 Log(gc, promotion) log; 699 if (log.is_trace()) { 700 ResourceMark rm; 701 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 702 } 703} 704 705void ConcurrentMarkSweepGeneration::reset_after_compaction() { 706 // Clear the promotion information. These pointers can be adjusted 707 // along with all the other pointers into the heap but 708 // compaction is expected to be a rare event with 709 // a heap using cms so don't do it without seeing the need. 710 for (uint i = 0; i < ParallelGCThreads; i++) { 711 _par_gc_thread_states[i]->promo.reset(); 712 } 713} 714 715void ConcurrentMarkSweepGeneration::compute_new_size() { 716 assert_locked_or_safepoint(Heap_lock); 717 718 // If incremental collection failed, we just want to expand 719 // to the limit. 720 if (incremental_collection_failed()) { 721 clear_incremental_collection_failed(); 722 grow_to_reserved(); 723 return; 724 } 725 726 // The heap has been compacted but not reset yet. 727 // Any metric such as free() or used() will be incorrect. 728 729 CardGeneration::compute_new_size(); 730 731 // Reset again after a possible resizing 732 if (did_compact()) { 733 cmsSpace()->reset_after_compaction(); 734 } 735} 736 737void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 738 assert_locked_or_safepoint(Heap_lock); 739 740 // If incremental collection failed, we just want to expand 741 // to the limit. 742 if (incremental_collection_failed()) { 743 clear_incremental_collection_failed(); 744 grow_to_reserved(); 745 return; 746 } 747 748 double free_percentage = ((double) free()) / capacity(); 749 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 750 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 751 752 // compute expansion delta needed for reaching desired free percentage 753 if (free_percentage < desired_free_percentage) { 754 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 755 assert(desired_capacity >= capacity(), "invalid expansion size"); 756 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 757 Log(gc) log; 758 if (log.is_trace()) { 759 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 760 log.trace("From compute_new_size: "); 761 log.trace(" Free fraction %f", free_percentage); 762 log.trace(" Desired free fraction %f", desired_free_percentage); 763 log.trace(" Maximum free fraction %f", maximum_free_percentage); 764 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 765 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 766 GenCollectedHeap* gch = GenCollectedHeap::heap(); 767 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 768 size_t young_size = gch->young_gen()->capacity(); 769 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 770 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 771 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 772 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 773 } 774 // safe if expansion fails 775 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 776 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 777 } else { 778 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 779 assert(desired_capacity <= capacity(), "invalid expansion size"); 780 size_t shrink_bytes = capacity() - desired_capacity; 781 // Don't shrink unless the delta is greater than the minimum shrink we want 782 if (shrink_bytes >= MinHeapDeltaBytes) { 783 shrink_free_list_by(shrink_bytes); 784 } 785 } 786} 787 788Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 789 return cmsSpace()->freelistLock(); 790} 791 792HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 793 CMSSynchronousYieldRequest yr; 794 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 795 return have_lock_and_allocate(size, tlab); 796} 797 798HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 799 bool tlab /* ignored */) { 800 assert_lock_strong(freelistLock()); 801 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 802 HeapWord* res = cmsSpace()->allocate(adjustedSize); 803 // Allocate the object live (grey) if the background collector has 804 // started marking. This is necessary because the marker may 805 // have passed this address and consequently this object will 806 // not otherwise be greyed and would be incorrectly swept up. 807 // Note that if this object contains references, the writing 808 // of those references will dirty the card containing this object 809 // allowing the object to be blackened (and its references scanned) 810 // either during a preclean phase or at the final checkpoint. 811 if (res != NULL) { 812 // We may block here with an uninitialized object with 813 // its mark-bit or P-bits not yet set. Such objects need 814 // to be safely navigable by block_start(). 815 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 816 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 817 collector()->direct_allocated(res, adjustedSize); 818 _direct_allocated_words += adjustedSize; 819 // allocation counters 820 NOT_PRODUCT( 821 _numObjectsAllocated++; 822 _numWordsAllocated += (int)adjustedSize; 823 ) 824 } 825 return res; 826} 827 828// In the case of direct allocation by mutators in a generation that 829// is being concurrently collected, the object must be allocated 830// live (grey) if the background collector has started marking. 831// This is necessary because the marker may 832// have passed this address and consequently this object will 833// not otherwise be greyed and would be incorrectly swept up. 834// Note that if this object contains references, the writing 835// of those references will dirty the card containing this object 836// allowing the object to be blackened (and its references scanned) 837// either during a preclean phase or at the final checkpoint. 838void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 839 assert(_markBitMap.covers(start, size), "Out of bounds"); 840 if (_collectorState >= Marking) { 841 MutexLockerEx y(_markBitMap.lock(), 842 Mutex::_no_safepoint_check_flag); 843 // [see comments preceding SweepClosure::do_blk() below for details] 844 // 845 // Can the P-bits be deleted now? JJJ 846 // 847 // 1. need to mark the object as live so it isn't collected 848 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 849 // 3. need to mark the end of the object so marking, precleaning or sweeping 850 // can skip over uninitialized or unparsable objects. An allocated 851 // object is considered uninitialized for our purposes as long as 852 // its klass word is NULL. All old gen objects are parsable 853 // as soon as they are initialized.) 854 _markBitMap.mark(start); // object is live 855 _markBitMap.mark(start + 1); // object is potentially uninitialized? 856 _markBitMap.mark(start + size - 1); 857 // mark end of object 858 } 859 // check that oop looks uninitialized 860 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 861} 862 863void CMSCollector::promoted(bool par, HeapWord* start, 864 bool is_obj_array, size_t obj_size) { 865 assert(_markBitMap.covers(start), "Out of bounds"); 866 // See comment in direct_allocated() about when objects should 867 // be allocated live. 868 if (_collectorState >= Marking) { 869 // we already hold the marking bit map lock, taken in 870 // the prologue 871 if (par) { 872 _markBitMap.par_mark(start); 873 } else { 874 _markBitMap.mark(start); 875 } 876 // We don't need to mark the object as uninitialized (as 877 // in direct_allocated above) because this is being done with the 878 // world stopped and the object will be initialized by the 879 // time the marking, precleaning or sweeping get to look at it. 880 // But see the code for copying objects into the CMS generation, 881 // where we need to ensure that concurrent readers of the 882 // block offset table are able to safely navigate a block that 883 // is in flux from being free to being allocated (and in 884 // transition while being copied into) and subsequently 885 // becoming a bona-fide object when the copy/promotion is complete. 886 assert(SafepointSynchronize::is_at_safepoint(), 887 "expect promotion only at safepoints"); 888 889 if (_collectorState < Sweeping) { 890 // Mark the appropriate cards in the modUnionTable, so that 891 // this object gets scanned before the sweep. If this is 892 // not done, CMS generation references in the object might 893 // not get marked. 894 // For the case of arrays, which are otherwise precisely 895 // marked, we need to dirty the entire array, not just its head. 896 if (is_obj_array) { 897 // The [par_]mark_range() method expects mr.end() below to 898 // be aligned to the granularity of a bit's representation 899 // in the heap. In the case of the MUT below, that's a 900 // card size. 901 MemRegion mr(start, 902 (HeapWord*)round_to((intptr_t)(start + obj_size), 903 CardTableModRefBS::card_size /* bytes */)); 904 if (par) { 905 _modUnionTable.par_mark_range(mr); 906 } else { 907 _modUnionTable.mark_range(mr); 908 } 909 } else { // not an obj array; we can just mark the head 910 if (par) { 911 _modUnionTable.par_mark(start); 912 } else { 913 _modUnionTable.mark(start); 914 } 915 } 916 } 917 } 918} 919 920oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 921 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 922 // allocate, copy and if necessary update promoinfo -- 923 // delegate to underlying space. 924 assert_lock_strong(freelistLock()); 925 926#ifndef PRODUCT 927 if (GenCollectedHeap::heap()->promotion_should_fail()) { 928 return NULL; 929 } 930#endif // #ifndef PRODUCT 931 932 oop res = _cmsSpace->promote(obj, obj_size); 933 if (res == NULL) { 934 // expand and retry 935 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 936 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 937 // Since this is the old generation, we don't try to promote 938 // into a more senior generation. 939 res = _cmsSpace->promote(obj, obj_size); 940 } 941 if (res != NULL) { 942 // See comment in allocate() about when objects should 943 // be allocated live. 944 assert(obj->is_oop(), "Will dereference klass pointer below"); 945 collector()->promoted(false, // Not parallel 946 (HeapWord*)res, obj->is_objArray(), obj_size); 947 // promotion counters 948 NOT_PRODUCT( 949 _numObjectsPromoted++; 950 _numWordsPromoted += 951 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 952 ) 953 } 954 return res; 955} 956 957 958// IMPORTANT: Notes on object size recognition in CMS. 959// --------------------------------------------------- 960// A block of storage in the CMS generation is always in 961// one of three states. A free block (FREE), an allocated 962// object (OBJECT) whose size() method reports the correct size, 963// and an intermediate state (TRANSIENT) in which its size cannot 964// be accurately determined. 965// STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 966// ----------------------------------------------------- 967// FREE: klass_word & 1 == 1; mark_word holds block size 968// 969// OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 970// obj->size() computes correct size 971// 972// TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 973// 974// STATE IDENTIFICATION: (64 bit+COOPS) 975// ------------------------------------ 976// FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 977// 978// OBJECT: klass_word installed; klass_word != 0; 979// obj->size() computes correct size 980// 981// TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 982// 983// 984// STATE TRANSITION DIAGRAM 985// 986// mut / parnew mut / parnew 987// FREE --------------------> TRANSIENT ---------------------> OBJECT --| 988// ^ | 989// |------------------------ DEAD <------------------------------------| 990// sweep mut 991// 992// While a block is in TRANSIENT state its size cannot be determined 993// so readers will either need to come back later or stall until 994// the size can be determined. Note that for the case of direct 995// allocation, P-bits, when available, may be used to determine the 996// size of an object that may not yet have been initialized. 997 998// Things to support parallel young-gen collection. 999oop 1000ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1001 oop old, markOop m, 1002 size_t word_sz) { 1003#ifndef PRODUCT 1004 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1005 return NULL; 1006 } 1007#endif // #ifndef PRODUCT 1008 1009 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1010 PromotionInfo* promoInfo = &ps->promo; 1011 // if we are tracking promotions, then first ensure space for 1012 // promotion (including spooling space for saving header if necessary). 1013 // then allocate and copy, then track promoted info if needed. 1014 // When tracking (see PromotionInfo::track()), the mark word may 1015 // be displaced and in this case restoration of the mark word 1016 // occurs in the (oop_since_save_marks_)iterate phase. 1017 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1018 // Out of space for allocating spooling buffers; 1019 // try expanding and allocating spooling buffers. 1020 if (!expand_and_ensure_spooling_space(promoInfo)) { 1021 return NULL; 1022 } 1023 } 1024 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1025 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1026 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1027 if (obj_ptr == NULL) { 1028 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1029 if (obj_ptr == NULL) { 1030 return NULL; 1031 } 1032 } 1033 oop obj = oop(obj_ptr); 1034 OrderAccess::storestore(); 1035 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1036 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1037 // IMPORTANT: See note on object initialization for CMS above. 1038 // Otherwise, copy the object. Here we must be careful to insert the 1039 // klass pointer last, since this marks the block as an allocated object. 1040 // Except with compressed oops it's the mark word. 1041 HeapWord* old_ptr = (HeapWord*)old; 1042 // Restore the mark word copied above. 1043 obj->set_mark(m); 1044 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1045 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1046 OrderAccess::storestore(); 1047 1048 if (UseCompressedClassPointers) { 1049 // Copy gap missed by (aligned) header size calculation below 1050 obj->set_klass_gap(old->klass_gap()); 1051 } 1052 if (word_sz > (size_t)oopDesc::header_size()) { 1053 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1054 obj_ptr + oopDesc::header_size(), 1055 word_sz - oopDesc::header_size()); 1056 } 1057 1058 // Now we can track the promoted object, if necessary. We take care 1059 // to delay the transition from uninitialized to full object 1060 // (i.e., insertion of klass pointer) until after, so that it 1061 // atomically becomes a promoted object. 1062 if (promoInfo->tracking()) { 1063 promoInfo->track((PromotedObject*)obj, old->klass()); 1064 } 1065 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1066 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1067 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1068 1069 // Finally, install the klass pointer (this should be volatile). 1070 OrderAccess::storestore(); 1071 obj->set_klass(old->klass()); 1072 // We should now be able to calculate the right size for this object 1073 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1074 1075 collector()->promoted(true, // parallel 1076 obj_ptr, old->is_objArray(), word_sz); 1077 1078 NOT_PRODUCT( 1079 Atomic::inc_ptr(&_numObjectsPromoted); 1080 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1081 ) 1082 1083 return obj; 1084} 1085 1086void 1087ConcurrentMarkSweepGeneration:: 1088par_promote_alloc_done(int thread_num) { 1089 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1090 ps->lab.retire(thread_num); 1091} 1092 1093void 1094ConcurrentMarkSweepGeneration:: 1095par_oop_since_save_marks_iterate_done(int thread_num) { 1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1097 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1098 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1099 1100 // Because card-scanning has been completed, subsequent phases 1101 // (e.g., reference processing) will not need to recognize which 1102 // objects have been promoted during this GC. So, we can now disable 1103 // promotion tracking. 1104 ps->promo.stopTrackingPromotions(); 1105} 1106 1107bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1108 size_t size, 1109 bool tlab) 1110{ 1111 // We allow a STW collection only if a full 1112 // collection was requested. 1113 return full || should_allocate(size, tlab); // FIX ME !!! 1114 // This and promotion failure handling are connected at the 1115 // hip and should be fixed by untying them. 1116} 1117 1118bool CMSCollector::shouldConcurrentCollect() { 1119 LogTarget(Trace, gc) log; 1120 1121 if (_full_gc_requested) { 1122 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1123 return true; 1124 } 1125 1126 FreelistLocker x(this); 1127 // ------------------------------------------------------------------ 1128 // Print out lots of information which affects the initiation of 1129 // a collection. 1130 if (log.is_enabled() && stats().valid()) { 1131 log.print("CMSCollector shouldConcurrentCollect: "); 1132 1133 LogStream out(log); 1134 stats().print_on(&out); 1135 1136 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1137 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1138 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1139 log.print("promotion_rate=%g", stats().promotion_rate()); 1140 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1141 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1142 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1143 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1144 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1145 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1146 } 1147 // ------------------------------------------------------------------ 1148 1149 // If the estimated time to complete a cms collection (cms_duration()) 1150 // is less than the estimated time remaining until the cms generation 1151 // is full, start a collection. 1152 if (!UseCMSInitiatingOccupancyOnly) { 1153 if (stats().valid()) { 1154 if (stats().time_until_cms_start() == 0.0) { 1155 return true; 1156 } 1157 } else { 1158 // We want to conservatively collect somewhat early in order 1159 // to try and "bootstrap" our CMS/promotion statistics; 1160 // this branch will not fire after the first successful CMS 1161 // collection because the stats should then be valid. 1162 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1163 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1164 _cmsGen->occupancy(), _bootstrap_occupancy); 1165 return true; 1166 } 1167 } 1168 } 1169 1170 // Otherwise, we start a collection cycle if 1171 // old gen want a collection cycle started. Each may use 1172 // an appropriate criterion for making this decision. 1173 // XXX We need to make sure that the gen expansion 1174 // criterion dovetails well with this. XXX NEED TO FIX THIS 1175 if (_cmsGen->should_concurrent_collect()) { 1176 log.print("CMS old gen initiated"); 1177 return true; 1178 } 1179 1180 // We start a collection if we believe an incremental collection may fail; 1181 // this is not likely to be productive in practice because it's probably too 1182 // late anyway. 1183 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1184 assert(gch->collector_policy()->is_generation_policy(), 1185 "You may want to check the correctness of the following"); 1186 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1187 log.print("CMSCollector: collect because incremental collection will fail "); 1188 return true; 1189 } 1190 1191 if (MetaspaceGC::should_concurrent_collect()) { 1192 log.print("CMSCollector: collect for metadata allocation "); 1193 return true; 1194 } 1195 1196 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1197 if (CMSTriggerInterval >= 0) { 1198 if (CMSTriggerInterval == 0) { 1199 // Trigger always 1200 return true; 1201 } 1202 1203 // Check the CMS time since begin (we do not check the stats validity 1204 // as we want to be able to trigger the first CMS cycle as well) 1205 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1206 if (stats().valid()) { 1207 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1208 stats().cms_time_since_begin()); 1209 } else { 1210 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1211 } 1212 return true; 1213 } 1214 } 1215 1216 return false; 1217} 1218 1219void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1220 1221// Clear _expansion_cause fields of constituent generations 1222void CMSCollector::clear_expansion_cause() { 1223 _cmsGen->clear_expansion_cause(); 1224} 1225 1226// We should be conservative in starting a collection cycle. To 1227// start too eagerly runs the risk of collecting too often in the 1228// extreme. To collect too rarely falls back on full collections, 1229// which works, even if not optimum in terms of concurrent work. 1230// As a work around for too eagerly collecting, use the flag 1231// UseCMSInitiatingOccupancyOnly. This also has the advantage of 1232// giving the user an easily understandable way of controlling the 1233// collections. 1234// We want to start a new collection cycle if any of the following 1235// conditions hold: 1236// . our current occupancy exceeds the configured initiating occupancy 1237// for this generation, or 1238// . we recently needed to expand this space and have not, since that 1239// expansion, done a collection of this generation, or 1240// . the underlying space believes that it may be a good idea to initiate 1241// a concurrent collection (this may be based on criteria such as the 1242// following: the space uses linear allocation and linear allocation is 1243// going to fail, or there is believed to be excessive fragmentation in 1244// the generation, etc... or ... 1245// [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1246// the case of the old generation; see CR 6543076): 1247// we may be approaching a point at which allocation requests may fail because 1248// we will be out of sufficient free space given allocation rate estimates.] 1249bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1250 1251 assert_lock_strong(freelistLock()); 1252 if (occupancy() > initiating_occupancy()) { 1253 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1254 short_name(), occupancy(), initiating_occupancy()); 1255 return true; 1256 } 1257 if (UseCMSInitiatingOccupancyOnly) { 1258 return false; 1259 } 1260 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1261 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1262 return true; 1263 } 1264 return false; 1265} 1266 1267void ConcurrentMarkSweepGeneration::collect(bool full, 1268 bool clear_all_soft_refs, 1269 size_t size, 1270 bool tlab) 1271{ 1272 collector()->collect(full, clear_all_soft_refs, size, tlab); 1273} 1274 1275void CMSCollector::collect(bool full, 1276 bool clear_all_soft_refs, 1277 size_t size, 1278 bool tlab) 1279{ 1280 // The following "if" branch is present for defensive reasons. 1281 // In the current uses of this interface, it can be replaced with: 1282 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1283 // But I am not placing that assert here to allow future 1284 // generality in invoking this interface. 1285 if (GCLocker::is_active()) { 1286 // A consistency test for GCLocker 1287 assert(GCLocker::needs_gc(), "Should have been set already"); 1288 // Skip this foreground collection, instead 1289 // expanding the heap if necessary. 1290 // Need the free list locks for the call to free() in compute_new_size() 1291 compute_new_size(); 1292 return; 1293 } 1294 acquire_control_and_collect(full, clear_all_soft_refs); 1295} 1296 1297void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1298 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1299 unsigned int gc_count = gch->total_full_collections(); 1300 if (gc_count == full_gc_count) { 1301 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1302 _full_gc_requested = true; 1303 _full_gc_cause = cause; 1304 CGC_lock->notify(); // nudge CMS thread 1305 } else { 1306 assert(gc_count > full_gc_count, "Error: causal loop"); 1307 } 1308} 1309 1310bool CMSCollector::is_external_interruption() { 1311 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1312 return GCCause::is_user_requested_gc(cause) || 1313 GCCause::is_serviceability_requested_gc(cause); 1314} 1315 1316void CMSCollector::report_concurrent_mode_interruption() { 1317 if (is_external_interruption()) { 1318 log_debug(gc)("Concurrent mode interrupted"); 1319 } else { 1320 log_debug(gc)("Concurrent mode failure"); 1321 _gc_tracer_cm->report_concurrent_mode_failure(); 1322 } 1323} 1324 1325 1326// The foreground and background collectors need to coordinate in order 1327// to make sure that they do not mutually interfere with CMS collections. 1328// When a background collection is active, 1329// the foreground collector may need to take over (preempt) and 1330// synchronously complete an ongoing collection. Depending on the 1331// frequency of the background collections and the heap usage 1332// of the application, this preemption can be seldom or frequent. 1333// There are only certain 1334// points in the background collection that the "collection-baton" 1335// can be passed to the foreground collector. 1336// 1337// The foreground collector will wait for the baton before 1338// starting any part of the collection. The foreground collector 1339// will only wait at one location. 1340// 1341// The background collector will yield the baton before starting a new 1342// phase of the collection (e.g., before initial marking, marking from roots, 1343// precleaning, final re-mark, sweep etc.) This is normally done at the head 1344// of the loop which switches the phases. The background collector does some 1345// of the phases (initial mark, final re-mark) with the world stopped. 1346// Because of locking involved in stopping the world, 1347// the foreground collector should not block waiting for the background 1348// collector when it is doing a stop-the-world phase. The background 1349// collector will yield the baton at an additional point just before 1350// it enters a stop-the-world phase. Once the world is stopped, the 1351// background collector checks the phase of the collection. If the 1352// phase has not changed, it proceeds with the collection. If the 1353// phase has changed, it skips that phase of the collection. See 1354// the comments on the use of the Heap_lock in collect_in_background(). 1355// 1356// Variable used in baton passing. 1357// _foregroundGCIsActive - Set to true by the foreground collector when 1358// it wants the baton. The foreground clears it when it has finished 1359// the collection. 1360// _foregroundGCShouldWait - Set to true by the background collector 1361// when it is running. The foreground collector waits while 1362// _foregroundGCShouldWait is true. 1363// CGC_lock - monitor used to protect access to the above variables 1364// and to notify the foreground and background collectors. 1365// _collectorState - current state of the CMS collection. 1366// 1367// The foreground collector 1368// acquires the CGC_lock 1369// sets _foregroundGCIsActive 1370// waits on the CGC_lock for _foregroundGCShouldWait to be false 1371// various locks acquired in preparation for the collection 1372// are released so as not to block the background collector 1373// that is in the midst of a collection 1374// proceeds with the collection 1375// clears _foregroundGCIsActive 1376// returns 1377// 1378// The background collector in a loop iterating on the phases of the 1379// collection 1380// acquires the CGC_lock 1381// sets _foregroundGCShouldWait 1382// if _foregroundGCIsActive is set 1383// clears _foregroundGCShouldWait, notifies _CGC_lock 1384// waits on _CGC_lock for _foregroundGCIsActive to become false 1385// and exits the loop. 1386// otherwise 1387// proceed with that phase of the collection 1388// if the phase is a stop-the-world phase, 1389// yield the baton once more just before enqueueing 1390// the stop-world CMS operation (executed by the VM thread). 1391// returns after all phases of the collection are done 1392// 1393 1394void CMSCollector::acquire_control_and_collect(bool full, 1395 bool clear_all_soft_refs) { 1396 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1397 assert(!Thread::current()->is_ConcurrentGC_thread(), 1398 "shouldn't try to acquire control from self!"); 1399 1400 // Start the protocol for acquiring control of the 1401 // collection from the background collector (aka CMS thread). 1402 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1403 "VM thread should have CMS token"); 1404 // Remember the possibly interrupted state of an ongoing 1405 // concurrent collection 1406 CollectorState first_state = _collectorState; 1407 1408 // Signal to a possibly ongoing concurrent collection that 1409 // we want to do a foreground collection. 1410 _foregroundGCIsActive = true; 1411 1412 // release locks and wait for a notify from the background collector 1413 // releasing the locks in only necessary for phases which 1414 // do yields to improve the granularity of the collection. 1415 assert_lock_strong(bitMapLock()); 1416 // We need to lock the Free list lock for the space that we are 1417 // currently collecting. 1418 assert(haveFreelistLocks(), "Must be holding free list locks"); 1419 bitMapLock()->unlock(); 1420 releaseFreelistLocks(); 1421 { 1422 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1423 if (_foregroundGCShouldWait) { 1424 // We are going to be waiting for action for the CMS thread; 1425 // it had better not be gone (for instance at shutdown)! 1426 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1427 "CMS thread must be running"); 1428 // Wait here until the background collector gives us the go-ahead 1429 ConcurrentMarkSweepThread::clear_CMS_flag( 1430 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1431 // Get a possibly blocked CMS thread going: 1432 // Note that we set _foregroundGCIsActive true above, 1433 // without protection of the CGC_lock. 1434 CGC_lock->notify(); 1435 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1436 "Possible deadlock"); 1437 while (_foregroundGCShouldWait) { 1438 // wait for notification 1439 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1440 // Possibility of delay/starvation here, since CMS token does 1441 // not know to give priority to VM thread? Actually, i think 1442 // there wouldn't be any delay/starvation, but the proof of 1443 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1444 } 1445 ConcurrentMarkSweepThread::set_CMS_flag( 1446 ConcurrentMarkSweepThread::CMS_vm_has_token); 1447 } 1448 } 1449 // The CMS_token is already held. Get back the other locks. 1450 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1451 "VM thread should have CMS token"); 1452 getFreelistLocks(); 1453 bitMapLock()->lock_without_safepoint_check(); 1454 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1455 p2i(Thread::current()), first_state); 1456 log_debug(gc, state)(" gets control with state %d", _collectorState); 1457 1458 // Inform cms gen if this was due to partial collection failing. 1459 // The CMS gen may use this fact to determine its expansion policy. 1460 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1461 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1462 assert(!_cmsGen->incremental_collection_failed(), 1463 "Should have been noticed, reacted to and cleared"); 1464 _cmsGen->set_incremental_collection_failed(); 1465 } 1466 1467 if (first_state > Idling) { 1468 report_concurrent_mode_interruption(); 1469 } 1470 1471 set_did_compact(true); 1472 1473 // If the collection is being acquired from the background 1474 // collector, there may be references on the discovered 1475 // references lists. Abandon those references, since some 1476 // of them may have become unreachable after concurrent 1477 // discovery; the STW compacting collector will redo discovery 1478 // more precisely, without being subject to floating garbage. 1479 // Leaving otherwise unreachable references in the discovered 1480 // lists would require special handling. 1481 ref_processor()->disable_discovery(); 1482 ref_processor()->abandon_partial_discovery(); 1483 ref_processor()->verify_no_references_recorded(); 1484 1485 if (first_state > Idling) { 1486 save_heap_summary(); 1487 } 1488 1489 do_compaction_work(clear_all_soft_refs); 1490 1491 // Has the GC time limit been exceeded? 1492 size_t max_eden_size = _young_gen->max_eden_size(); 1493 GCCause::Cause gc_cause = gch->gc_cause(); 1494 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1495 _young_gen->eden()->used(), 1496 _cmsGen->max_capacity(), 1497 max_eden_size, 1498 full, 1499 gc_cause, 1500 gch->collector_policy()); 1501 1502 // Reset the expansion cause, now that we just completed 1503 // a collection cycle. 1504 clear_expansion_cause(); 1505 _foregroundGCIsActive = false; 1506 return; 1507} 1508 1509// Resize the tenured generation 1510// after obtaining the free list locks for the 1511// two generations. 1512void CMSCollector::compute_new_size() { 1513 assert_locked_or_safepoint(Heap_lock); 1514 FreelistLocker z(this); 1515 MetaspaceGC::compute_new_size(); 1516 _cmsGen->compute_new_size_free_list(); 1517} 1518 1519// A work method used by the foreground collector to do 1520// a mark-sweep-compact. 1521void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1522 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1523 1524 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1525 gc_timer->register_gc_start(); 1526 1527 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1528 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1529 1530 gch->pre_full_gc_dump(gc_timer); 1531 1532 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1533 1534 // Temporarily widen the span of the weak reference processing to 1535 // the entire heap. 1536 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1537 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1538 // Temporarily, clear the "is_alive_non_header" field of the 1539 // reference processor. 1540 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1541 // Temporarily make reference _processing_ single threaded (non-MT). 1542 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1543 // Temporarily make refs discovery atomic 1544 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1545 // Temporarily make reference _discovery_ single threaded (non-MT) 1546 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1547 1548 ref_processor()->set_enqueuing_is_done(false); 1549 ref_processor()->enable_discovery(); 1550 ref_processor()->setup_policy(clear_all_soft_refs); 1551 // If an asynchronous collection finishes, the _modUnionTable is 1552 // all clear. If we are assuming the collection from an asynchronous 1553 // collection, clear the _modUnionTable. 1554 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1555 "_modUnionTable should be clear if the baton was not passed"); 1556 _modUnionTable.clear_all(); 1557 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1558 "mod union for klasses should be clear if the baton was passed"); 1559 _ct->klass_rem_set()->clear_mod_union(); 1560 1561 // We must adjust the allocation statistics being maintained 1562 // in the free list space. We do so by reading and clearing 1563 // the sweep timer and updating the block flux rate estimates below. 1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1565 if (_inter_sweep_timer.is_active()) { 1566 _inter_sweep_timer.stop(); 1567 // Note that we do not use this sample to update the _inter_sweep_estimate. 1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1569 _inter_sweep_estimate.padded_average(), 1570 _intra_sweep_estimate.padded_average()); 1571 } 1572 1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1574 #ifdef ASSERT 1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1576 size_t free_size = cms_space->free(); 1577 assert(free_size == 1578 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1579 * HeapWordSize, 1580 "All the free space should be compacted into one chunk at top"); 1581 assert(cms_space->dictionary()->total_chunk_size( 1582 debug_only(cms_space->freelistLock())) == 0 || 1583 cms_space->totalSizeInIndexedFreeLists() == 0, 1584 "All the free space should be in a single chunk"); 1585 size_t num = cms_space->totalCount(); 1586 assert((free_size == 0 && num == 0) || 1587 (free_size > 0 && (num == 1 || num == 2)), 1588 "There should be at most 2 free chunks after compaction"); 1589 #endif // ASSERT 1590 _collectorState = Resetting; 1591 assert(_restart_addr == NULL, 1592 "Should have been NULL'd before baton was passed"); 1593 reset_stw(); 1594 _cmsGen->reset_after_compaction(); 1595 _concurrent_cycles_since_last_unload = 0; 1596 1597 // Clear any data recorded in the PLAB chunk arrays. 1598 if (_survivor_plab_array != NULL) { 1599 reset_survivor_plab_arrays(); 1600 } 1601 1602 // Adjust the per-size allocation stats for the next epoch. 1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1604 // Restart the "inter sweep timer" for the next epoch. 1605 _inter_sweep_timer.reset(); 1606 _inter_sweep_timer.start(); 1607 1608 // No longer a need to do a concurrent collection for Metaspace. 1609 MetaspaceGC::set_should_concurrent_collect(false); 1610 1611 gch->post_full_gc_dump(gc_timer); 1612 1613 gc_timer->register_gc_end(); 1614 1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1616 1617 // For a mark-sweep-compact, compute_new_size() will be called 1618 // in the heap's do_collection() method. 1619} 1620 1621void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1622 Log(gc, heap) log; 1623 if (!log.is_trace()) { 1624 return; 1625 } 1626 1627 ContiguousSpace* eden_space = _young_gen->eden(); 1628 ContiguousSpace* from_space = _young_gen->from(); 1629 ContiguousSpace* to_space = _young_gen->to(); 1630 // Eden 1631 if (_eden_chunk_array != NULL) { 1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1633 p2i(eden_space->bottom()), p2i(eden_space->top()), 1634 p2i(eden_space->end()), eden_space->capacity()); 1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1636 _eden_chunk_index, _eden_chunk_capacity); 1637 for (size_t i = 0; i < _eden_chunk_index; i++) { 1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1639 } 1640 } 1641 // Survivor 1642 if (_survivor_chunk_array != NULL) { 1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1644 p2i(from_space->bottom()), p2i(from_space->top()), 1645 p2i(from_space->end()), from_space->capacity()); 1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1647 _survivor_chunk_index, _survivor_chunk_capacity); 1648 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1650 } 1651 } 1652} 1653 1654void CMSCollector::getFreelistLocks() const { 1655 // Get locks for all free lists in all generations that this 1656 // collector is responsible for 1657 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1658} 1659 1660void CMSCollector::releaseFreelistLocks() const { 1661 // Release locks for all free lists in all generations that this 1662 // collector is responsible for 1663 _cmsGen->freelistLock()->unlock(); 1664} 1665 1666bool CMSCollector::haveFreelistLocks() const { 1667 // Check locks for all free lists in all generations that this 1668 // collector is responsible for 1669 assert_lock_strong(_cmsGen->freelistLock()); 1670 PRODUCT_ONLY(ShouldNotReachHere()); 1671 return true; 1672} 1673 1674// A utility class that is used by the CMS collector to 1675// temporarily "release" the foreground collector from its 1676// usual obligation to wait for the background collector to 1677// complete an ongoing phase before proceeding. 1678class ReleaseForegroundGC: public StackObj { 1679 private: 1680 CMSCollector* _c; 1681 public: 1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1685 // allow a potentially blocked foreground collector to proceed 1686 _c->_foregroundGCShouldWait = false; 1687 if (_c->_foregroundGCIsActive) { 1688 CGC_lock->notify(); 1689 } 1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1691 "Possible deadlock"); 1692 } 1693 1694 ~ReleaseForegroundGC() { 1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1697 _c->_foregroundGCShouldWait = true; 1698 } 1699}; 1700 1701void CMSCollector::collect_in_background(GCCause::Cause cause) { 1702 assert(Thread::current()->is_ConcurrentGC_thread(), 1703 "A CMS asynchronous collection is only allowed on a CMS thread."); 1704 1705 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1706 { 1707 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1708 MutexLockerEx hl(Heap_lock, safepoint_check); 1709 FreelistLocker fll(this); 1710 MutexLockerEx x(CGC_lock, safepoint_check); 1711 if (_foregroundGCIsActive) { 1712 // The foreground collector is. Skip this 1713 // background collection. 1714 assert(!_foregroundGCShouldWait, "Should be clear"); 1715 return; 1716 } else { 1717 assert(_collectorState == Idling, "Should be idling before start."); 1718 _collectorState = InitialMarking; 1719 register_gc_start(cause); 1720 // Reset the expansion cause, now that we are about to begin 1721 // a new cycle. 1722 clear_expansion_cause(); 1723 1724 // Clear the MetaspaceGC flag since a concurrent collection 1725 // is starting but also clear it after the collection. 1726 MetaspaceGC::set_should_concurrent_collect(false); 1727 } 1728 // Decide if we want to enable class unloading as part of the 1729 // ensuing concurrent GC cycle. 1730 update_should_unload_classes(); 1731 _full_gc_requested = false; // acks all outstanding full gc requests 1732 _full_gc_cause = GCCause::_no_gc; 1733 // Signal that we are about to start a collection 1734 gch->increment_total_full_collections(); // ... starting a collection cycle 1735 _collection_count_start = gch->total_full_collections(); 1736 } 1737 1738 size_t prev_used = _cmsGen->used(); 1739 1740 // The change of the collection state is normally done at this level; 1741 // the exceptions are phases that are executed while the world is 1742 // stopped. For those phases the change of state is done while the 1743 // world is stopped. For baton passing purposes this allows the 1744 // background collector to finish the phase and change state atomically. 1745 // The foreground collector cannot wait on a phase that is done 1746 // while the world is stopped because the foreground collector already 1747 // has the world stopped and would deadlock. 1748 while (_collectorState != Idling) { 1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1750 p2i(Thread::current()), _collectorState); 1751 // The foreground collector 1752 // holds the Heap_lock throughout its collection. 1753 // holds the CMS token (but not the lock) 1754 // except while it is waiting for the background collector to yield. 1755 // 1756 // The foreground collector should be blocked (not for long) 1757 // if the background collector is about to start a phase 1758 // executed with world stopped. If the background 1759 // collector has already started such a phase, the 1760 // foreground collector is blocked waiting for the 1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1762 // are executed in the VM thread. 1763 // 1764 // The locking order is 1765 // PendingListLock (PLL) -- if applicable (FinalMarking) 1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1767 // CMS token (claimed in 1768 // stop_world_and_do() --> 1769 // safepoint_synchronize() --> 1770 // CMSThread::synchronize()) 1771 1772 { 1773 // Check if the FG collector wants us to yield. 1774 CMSTokenSync x(true); // is cms thread 1775 if (waitForForegroundGC()) { 1776 // We yielded to a foreground GC, nothing more to be 1777 // done this round. 1778 assert(_foregroundGCShouldWait == false, "We set it to false in " 1779 "waitForForegroundGC()"); 1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1781 p2i(Thread::current()), _collectorState); 1782 return; 1783 } else { 1784 // The background collector can run but check to see if the 1785 // foreground collector has done a collection while the 1786 // background collector was waiting to get the CGC_lock 1787 // above. If yes, break so that _foregroundGCShouldWait 1788 // is cleared before returning. 1789 if (_collectorState == Idling) { 1790 break; 1791 } 1792 } 1793 } 1794 1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1796 "should be waiting"); 1797 1798 switch (_collectorState) { 1799 case InitialMarking: 1800 { 1801 ReleaseForegroundGC x(this); 1802 stats().record_cms_begin(); 1803 VM_CMS_Initial_Mark initial_mark_op(this); 1804 VMThread::execute(&initial_mark_op); 1805 } 1806 // The collector state may be any legal state at this point 1807 // since the background collector may have yielded to the 1808 // foreground collector. 1809 break; 1810 case Marking: 1811 // initial marking in checkpointRootsInitialWork has been completed 1812 if (markFromRoots()) { // we were successful 1813 assert(_collectorState == Precleaning, "Collector state should " 1814 "have changed"); 1815 } else { 1816 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1817 } 1818 break; 1819 case Precleaning: 1820 // marking from roots in markFromRoots has been completed 1821 preclean(); 1822 assert(_collectorState == AbortablePreclean || 1823 _collectorState == FinalMarking, 1824 "Collector state should have changed"); 1825 break; 1826 case AbortablePreclean: 1827 abortable_preclean(); 1828 assert(_collectorState == FinalMarking, "Collector state should " 1829 "have changed"); 1830 break; 1831 case FinalMarking: 1832 { 1833 ReleaseForegroundGC x(this); 1834 1835 VM_CMS_Final_Remark final_remark_op(this); 1836 VMThread::execute(&final_remark_op); 1837 } 1838 assert(_foregroundGCShouldWait, "block post-condition"); 1839 break; 1840 case Sweeping: 1841 // final marking in checkpointRootsFinal has been completed 1842 sweep(); 1843 assert(_collectorState == Resizing, "Collector state change " 1844 "to Resizing must be done under the free_list_lock"); 1845 1846 case Resizing: { 1847 // Sweeping has been completed... 1848 // At this point the background collection has completed. 1849 // Don't move the call to compute_new_size() down 1850 // into code that might be executed if the background 1851 // collection was preempted. 1852 { 1853 ReleaseForegroundGC x(this); // unblock FG collection 1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1855 CMSTokenSync z(true); // not strictly needed. 1856 if (_collectorState == Resizing) { 1857 compute_new_size(); 1858 save_heap_summary(); 1859 _collectorState = Resetting; 1860 } else { 1861 assert(_collectorState == Idling, "The state should only change" 1862 " because the foreground collector has finished the collection"); 1863 } 1864 } 1865 break; 1866 } 1867 case Resetting: 1868 // CMS heap resizing has been completed 1869 reset_concurrent(); 1870 assert(_collectorState == Idling, "Collector state should " 1871 "have changed"); 1872 1873 MetaspaceGC::set_should_concurrent_collect(false); 1874 1875 stats().record_cms_end(); 1876 // Don't move the concurrent_phases_end() and compute_new_size() 1877 // calls to here because a preempted background collection 1878 // has it's state set to "Resetting". 1879 break; 1880 case Idling: 1881 default: 1882 ShouldNotReachHere(); 1883 break; 1884 } 1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1886 p2i(Thread::current()), _collectorState); 1887 assert(_foregroundGCShouldWait, "block post-condition"); 1888 } 1889 1890 // Should this be in gc_epilogue? 1891 collector_policy()->counters()->update_counters(); 1892 1893 { 1894 // Clear _foregroundGCShouldWait and, in the event that the 1895 // foreground collector is waiting, notify it, before 1896 // returning. 1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1898 _foregroundGCShouldWait = false; 1899 if (_foregroundGCIsActive) { 1900 CGC_lock->notify(); 1901 } 1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1903 "Possible deadlock"); 1904 } 1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1906 p2i(Thread::current()), _collectorState); 1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1909} 1910 1911void CMSCollector::register_gc_start(GCCause::Cause cause) { 1912 _cms_start_registered = true; 1913 _gc_timer_cm->register_gc_start(); 1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1915} 1916 1917void CMSCollector::register_gc_end() { 1918 if (_cms_start_registered) { 1919 report_heap_summary(GCWhen::AfterGC); 1920 1921 _gc_timer_cm->register_gc_end(); 1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1923 _cms_start_registered = false; 1924 } 1925} 1926 1927void CMSCollector::save_heap_summary() { 1928 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1929 _last_heap_summary = gch->create_heap_summary(); 1930 _last_metaspace_summary = gch->create_metaspace_summary(); 1931} 1932 1933void CMSCollector::report_heap_summary(GCWhen::Type when) { 1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1936} 1937 1938bool CMSCollector::waitForForegroundGC() { 1939 bool res = false; 1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1941 "CMS thread should have CMS token"); 1942 // Block the foreground collector until the 1943 // background collectors decides whether to 1944 // yield. 1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1946 _foregroundGCShouldWait = true; 1947 if (_foregroundGCIsActive) { 1948 // The background collector yields to the 1949 // foreground collector and returns a value 1950 // indicating that it has yielded. The foreground 1951 // collector can proceed. 1952 res = true; 1953 _foregroundGCShouldWait = false; 1954 ConcurrentMarkSweepThread::clear_CMS_flag( 1955 ConcurrentMarkSweepThread::CMS_cms_has_token); 1956 ConcurrentMarkSweepThread::set_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1958 // Get a possibly blocked foreground thread going 1959 CGC_lock->notify(); 1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1961 p2i(Thread::current()), _collectorState); 1962 while (_foregroundGCIsActive) { 1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1964 } 1965 ConcurrentMarkSweepThread::set_CMS_flag( 1966 ConcurrentMarkSweepThread::CMS_cms_has_token); 1967 ConcurrentMarkSweepThread::clear_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1969 } 1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 return res; 1973} 1974 1975// Because of the need to lock the free lists and other structures in 1976// the collector, common to all the generations that the collector is 1977// collecting, we need the gc_prologues of individual CMS generations 1978// delegate to their collector. It may have been simpler had the 1979// current infrastructure allowed one to call a prologue on a 1980// collector. In the absence of that we have the generation's 1981// prologue delegate to the collector, which delegates back 1982// some "local" work to a worker method in the individual generations 1983// that it's responsible for collecting, while itself doing any 1984// work common to all generations it's responsible for. A similar 1985// comment applies to the gc_epilogue()'s. 1986// The role of the variable _between_prologue_and_epilogue is to 1987// enforce the invocation protocol. 1988void CMSCollector::gc_prologue(bool full) { 1989 // Call gc_prologue_work() for the CMSGen 1990 // we are responsible for. 1991 1992 // The following locking discipline assumes that we are only called 1993 // when the world is stopped. 1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1995 1996 // The CMSCollector prologue must call the gc_prologues for the 1997 // "generations" that it's responsible 1998 // for. 1999 2000 assert( Thread::current()->is_VM_thread() 2001 || ( CMSScavengeBeforeRemark 2002 && Thread::current()->is_ConcurrentGC_thread()), 2003 "Incorrect thread type for prologue execution"); 2004 2005 if (_between_prologue_and_epilogue) { 2006 // We have already been invoked; this is a gc_prologue delegation 2007 // from yet another CMS generation that we are responsible for, just 2008 // ignore it since all relevant work has already been done. 2009 return; 2010 } 2011 2012 // set a bit saying prologue has been called; cleared in epilogue 2013 _between_prologue_and_epilogue = true; 2014 // Claim locks for common data structures, then call gc_prologue_work() 2015 // for each CMSGen. 2016 2017 getFreelistLocks(); // gets free list locks on constituent spaces 2018 bitMapLock()->lock_without_safepoint_check(); 2019 2020 // Should call gc_prologue_work() for all cms gens we are responsible for 2021 bool duringMarking = _collectorState >= Marking 2022 && _collectorState < Sweeping; 2023 2024 // The young collections clear the modified oops state, which tells if 2025 // there are any modified oops in the class. The remark phase also needs 2026 // that information. Tell the young collection to save the union of all 2027 // modified klasses. 2028 if (duringMarking) { 2029 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2030 } 2031 2032 bool registerClosure = duringMarking; 2033 2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2035 2036 if (!full) { 2037 stats().record_gc0_begin(); 2038 } 2039} 2040 2041void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2042 2043 _capacity_at_prologue = capacity(); 2044 _used_at_prologue = used(); 2045 2046 // We enable promotion tracking so that card-scanning can recognize 2047 // which objects have been promoted during this GC and skip them. 2048 for (uint i = 0; i < ParallelGCThreads; i++) { 2049 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2050 } 2051 2052 // Delegate to CMScollector which knows how to coordinate between 2053 // this and any other CMS generations that it is responsible for 2054 // collecting. 2055 collector()->gc_prologue(full); 2056} 2057 2058// This is a "private" interface for use by this generation's CMSCollector. 2059// Not to be called directly by any other entity (for instance, 2060// GenCollectedHeap, which calls the "public" gc_prologue method above). 2061void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2062 bool registerClosure, ModUnionClosure* modUnionClosure) { 2063 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2065 "Should be NULL"); 2066 if (registerClosure) { 2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2068 } 2069 cmsSpace()->gc_prologue(); 2070 // Clear stat counters 2071 NOT_PRODUCT( 2072 assert(_numObjectsPromoted == 0, "check"); 2073 assert(_numWordsPromoted == 0, "check"); 2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2076 _numObjectsAllocated = 0; 2077 _numWordsAllocated = 0; 2078 ) 2079} 2080 2081void CMSCollector::gc_epilogue(bool full) { 2082 // The following locking discipline assumes that we are only called 2083 // when the world is stopped. 2084 assert(SafepointSynchronize::is_at_safepoint(), 2085 "world is stopped assumption"); 2086 2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2088 // if linear allocation blocks need to be appropriately marked to allow the 2089 // the blocks to be parsable. We also check here whether we need to nudge the 2090 // CMS collector thread to start a new cycle (if it's not already active). 2091 assert( Thread::current()->is_VM_thread() 2092 || ( CMSScavengeBeforeRemark 2093 && Thread::current()->is_ConcurrentGC_thread()), 2094 "Incorrect thread type for epilogue execution"); 2095 2096 if (!_between_prologue_and_epilogue) { 2097 // We have already been invoked; this is a gc_epilogue delegation 2098 // from yet another CMS generation that we are responsible for, just 2099 // ignore it since all relevant work has already been done. 2100 return; 2101 } 2102 assert(haveFreelistLocks(), "must have freelist locks"); 2103 assert_lock_strong(bitMapLock()); 2104 2105 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2106 2107 _cmsGen->gc_epilogue_work(full); 2108 2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2110 // in case sampling was not already enabled, enable it 2111 _start_sampling = true; 2112 } 2113 // reset _eden_chunk_array so sampling starts afresh 2114 _eden_chunk_index = 0; 2115 2116 size_t cms_used = _cmsGen->cmsSpace()->used(); 2117 2118 // update performance counters - this uses a special version of 2119 // update_counters() that allows the utilization to be passed as a 2120 // parameter, avoiding multiple calls to used(). 2121 // 2122 _cmsGen->update_counters(cms_used); 2123 2124 bitMapLock()->unlock(); 2125 releaseFreelistLocks(); 2126 2127 if (!CleanChunkPoolAsync) { 2128 Chunk::clean_chunk_pool(); 2129 } 2130 2131 set_did_compact(false); 2132 _between_prologue_and_epilogue = false; // ready for next cycle 2133} 2134 2135void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2136 collector()->gc_epilogue(full); 2137 2138 // When using ParNew, promotion tracking should have already been 2139 // disabled. However, the prologue (which enables promotion 2140 // tracking) and epilogue are called irrespective of the type of 2141 // GC. So they will also be called before and after Full GCs, during 2142 // which promotion tracking will not be explicitly disabled. So, 2143 // it's safer to also disable it here too (to be symmetric with 2144 // enabling it in the prologue). 2145 for (uint i = 0; i < ParallelGCThreads; i++) { 2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2147 } 2148} 2149 2150void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2151 assert(!incremental_collection_failed(), "Should have been cleared"); 2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2153 cmsSpace()->gc_epilogue(); 2154 // Print stat counters 2155 NOT_PRODUCT( 2156 assert(_numObjectsAllocated == 0, "check"); 2157 assert(_numWordsAllocated == 0, "check"); 2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2160 _numObjectsPromoted = 0; 2161 _numWordsPromoted = 0; 2162 ) 2163 2164 // Call down the chain in contiguous_available needs the freelistLock 2165 // so print this out before releasing the freeListLock. 2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2167} 2168 2169#ifndef PRODUCT 2170bool CMSCollector::have_cms_token() { 2171 Thread* thr = Thread::current(); 2172 if (thr->is_VM_thread()) { 2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2174 } else if (thr->is_ConcurrentGC_thread()) { 2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2176 } else if (thr->is_GC_task_thread()) { 2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2178 ParGCRareEvent_lock->owned_by_self(); 2179 } 2180 return false; 2181} 2182 2183// Check reachability of the given heap address in CMS generation, 2184// treating all other generations as roots. 2185bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2186 // We could "guarantee" below, rather than assert, but I'll 2187 // leave these as "asserts" so that an adventurous debugger 2188 // could try this in the product build provided some subset of 2189 // the conditions were met, provided they were interested in the 2190 // results and knew that the computation below wouldn't interfere 2191 // with other concurrent computations mutating the structures 2192 // being read or written. 2193 assert(SafepointSynchronize::is_at_safepoint(), 2194 "Else mutations in object graph will make answer suspect"); 2195 assert(have_cms_token(), "Should hold cms token"); 2196 assert(haveFreelistLocks(), "must hold free list locks"); 2197 assert_lock_strong(bitMapLock()); 2198 2199 // Clear the marking bit map array before starting, but, just 2200 // for kicks, first report if the given address is already marked 2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2202 _markBitMap.isMarked(addr) ? "" : " not"); 2203 2204 if (verify_after_remark()) { 2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2206 bool result = verification_mark_bm()->isMarked(addr); 2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2208 result ? "IS" : "is NOT"); 2209 return result; 2210 } else { 2211 tty->print_cr("Could not compute result"); 2212 return false; 2213 } 2214} 2215#endif 2216 2217void 2218CMSCollector::print_on_error(outputStream* st) { 2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2220 if (collector != NULL) { 2221 CMSBitMap* bitmap = &collector->_markBitMap; 2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2223 bitmap->print_on_error(st, " Bits: "); 2224 2225 st->cr(); 2226 2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2229 mut_bitmap->print_on_error(st, " Bits: "); 2230 } 2231} 2232 2233//////////////////////////////////////////////////////// 2234// CMS Verification Support 2235//////////////////////////////////////////////////////// 2236// Following the remark phase, the following invariant 2237// should hold -- each object in the CMS heap which is 2238// marked in markBitMap() should be marked in the verification_mark_bm(). 2239 2240class VerifyMarkedClosure: public BitMapClosure { 2241 CMSBitMap* _marks; 2242 bool _failed; 2243 2244 public: 2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2246 2247 bool do_bit(size_t offset) { 2248 HeapWord* addr = _marks->offsetToHeapWord(offset); 2249 if (!_marks->isMarked(addr)) { 2250 Log(gc, verify) log; 2251 ResourceMark rm; 2252 oop(addr)->print_on(log.error_stream()); 2253 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2254 _failed = true; 2255 } 2256 return true; 2257 } 2258 2259 bool failed() { return _failed; } 2260}; 2261 2262bool CMSCollector::verify_after_remark() { 2263 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2264 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2265 static bool init = false; 2266 2267 assert(SafepointSynchronize::is_at_safepoint(), 2268 "Else mutations in object graph will make answer suspect"); 2269 assert(have_cms_token(), 2270 "Else there may be mutual interference in use of " 2271 " verification data structures"); 2272 assert(_collectorState > Marking && _collectorState <= Sweeping, 2273 "Else marking info checked here may be obsolete"); 2274 assert(haveFreelistLocks(), "must hold free list locks"); 2275 assert_lock_strong(bitMapLock()); 2276 2277 2278 // Allocate marking bit map if not already allocated 2279 if (!init) { // first time 2280 if (!verification_mark_bm()->allocate(_span)) { 2281 return false; 2282 } 2283 init = true; 2284 } 2285 2286 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2287 2288 // Turn off refs discovery -- so we will be tracing through refs. 2289 // This is as intended, because by this time 2290 // GC must already have cleared any refs that need to be cleared, 2291 // and traced those that need to be marked; moreover, 2292 // the marking done here is not going to interfere in any 2293 // way with the marking information used by GC. 2294 NoRefDiscovery no_discovery(ref_processor()); 2295 2296#if defined(COMPILER2) || INCLUDE_JVMCI 2297 DerivedPointerTableDeactivate dpt_deact; 2298#endif 2299 2300 // Clear any marks from a previous round 2301 verification_mark_bm()->clear_all(); 2302 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2303 verify_work_stacks_empty(); 2304 2305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2306 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2307 // Update the saved marks which may affect the root scans. 2308 gch->save_marks(); 2309 2310 if (CMSRemarkVerifyVariant == 1) { 2311 // In this first variant of verification, we complete 2312 // all marking, then check if the new marks-vector is 2313 // a subset of the CMS marks-vector. 2314 verify_after_remark_work_1(); 2315 } else { 2316 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2317 // In this second variant of verification, we flag an error 2318 // (i.e. an object reachable in the new marks-vector not reachable 2319 // in the CMS marks-vector) immediately, also indicating the 2320 // identify of an object (A) that references the unmarked object (B) -- 2321 // presumably, a mutation to A failed to be picked up by preclean/remark? 2322 verify_after_remark_work_2(); 2323 } 2324 2325 return true; 2326} 2327 2328void CMSCollector::verify_after_remark_work_1() { 2329 ResourceMark rm; 2330 HandleMark hm; 2331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2332 2333 // Get a clear set of claim bits for the roots processing to work with. 2334 ClassLoaderDataGraph::clear_claimed_marks(); 2335 2336 // Mark from roots one level into CMS 2337 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2338 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2339 2340 { 2341 StrongRootsScope srs(1); 2342 2343 gch->gen_process_roots(&srs, 2344 GenCollectedHeap::OldGen, 2345 true, // young gen as roots 2346 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2347 should_unload_classes(), 2348 ¬Older, 2349 NULL, 2350 NULL); 2351 } 2352 2353 // Now mark from the roots 2354 MarkFromRootsClosure markFromRootsClosure(this, _span, 2355 verification_mark_bm(), verification_mark_stack(), 2356 false /* don't yield */, true /* verifying */); 2357 assert(_restart_addr == NULL, "Expected pre-condition"); 2358 verification_mark_bm()->iterate(&markFromRootsClosure); 2359 while (_restart_addr != NULL) { 2360 // Deal with stack overflow: by restarting at the indicated 2361 // address. 2362 HeapWord* ra = _restart_addr; 2363 markFromRootsClosure.reset(ra); 2364 _restart_addr = NULL; 2365 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2366 } 2367 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2368 verify_work_stacks_empty(); 2369 2370 // Marking completed -- now verify that each bit marked in 2371 // verification_mark_bm() is also marked in markBitMap(); flag all 2372 // errors by printing corresponding objects. 2373 VerifyMarkedClosure vcl(markBitMap()); 2374 verification_mark_bm()->iterate(&vcl); 2375 if (vcl.failed()) { 2376 Log(gc, verify) log; 2377 log.error("Failed marking verification after remark"); 2378 ResourceMark rm; 2379 gch->print_on(log.error_stream()); 2380 fatal("CMS: failed marking verification after remark"); 2381 } 2382} 2383 2384class VerifyKlassOopsKlassClosure : public KlassClosure { 2385 class VerifyKlassOopsClosure : public OopClosure { 2386 CMSBitMap* _bitmap; 2387 public: 2388 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2389 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2390 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2391 } _oop_closure; 2392 public: 2393 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2394 void do_klass(Klass* k) { 2395 k->oops_do(&_oop_closure); 2396 } 2397}; 2398 2399void CMSCollector::verify_after_remark_work_2() { 2400 ResourceMark rm; 2401 HandleMark hm; 2402 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2403 2404 // Get a clear set of claim bits for the roots processing to work with. 2405 ClassLoaderDataGraph::clear_claimed_marks(); 2406 2407 // Mark from roots one level into CMS 2408 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2409 markBitMap()); 2410 CLDToOopClosure cld_closure(¬Older, true); 2411 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 &cld_closure); 2425 } 2426 2427 // Now mark from the roots 2428 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2429 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2430 assert(_restart_addr == NULL, "Expected pre-condition"); 2431 verification_mark_bm()->iterate(&markFromRootsClosure); 2432 while (_restart_addr != NULL) { 2433 // Deal with stack overflow: by restarting at the indicated 2434 // address. 2435 HeapWord* ra = _restart_addr; 2436 markFromRootsClosure.reset(ra); 2437 _restart_addr = NULL; 2438 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2439 } 2440 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2441 verify_work_stacks_empty(); 2442 2443 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2444 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2445 2446 // Marking completed -- now verify that each bit marked in 2447 // verification_mark_bm() is also marked in markBitMap(); flag all 2448 // errors by printing corresponding objects. 2449 VerifyMarkedClosure vcl(markBitMap()); 2450 verification_mark_bm()->iterate(&vcl); 2451 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2452} 2453 2454void ConcurrentMarkSweepGeneration::save_marks() { 2455 // delegate to CMS space 2456 cmsSpace()->save_marks(); 2457} 2458 2459bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2460 return cmsSpace()->no_allocs_since_save_marks(); 2461} 2462 2463#define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2464 \ 2465void ConcurrentMarkSweepGeneration:: \ 2466oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2467 cl->set_generation(this); \ 2468 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2469 cl->reset_generation(); \ 2470 save_marks(); \ 2471} 2472 2473ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2474 2475void 2476ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2477 if (freelistLock()->owned_by_self()) { 2478 Generation::oop_iterate(cl); 2479 } else { 2480 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2481 Generation::oop_iterate(cl); 2482 } 2483} 2484 2485void 2486ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2487 if (freelistLock()->owned_by_self()) { 2488 Generation::object_iterate(cl); 2489 } else { 2490 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2491 Generation::object_iterate(cl); 2492 } 2493} 2494 2495void 2496ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2497 if (freelistLock()->owned_by_self()) { 2498 Generation::safe_object_iterate(cl); 2499 } else { 2500 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2501 Generation::safe_object_iterate(cl); 2502 } 2503} 2504 2505void 2506ConcurrentMarkSweepGeneration::post_compact() { 2507} 2508 2509void 2510ConcurrentMarkSweepGeneration::prepare_for_verify() { 2511 // Fix the linear allocation blocks to look like free blocks. 2512 2513 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2514 // are not called when the heap is verified during universe initialization and 2515 // at vm shutdown. 2516 if (freelistLock()->owned_by_self()) { 2517 cmsSpace()->prepare_for_verify(); 2518 } else { 2519 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2520 cmsSpace()->prepare_for_verify(); 2521 } 2522} 2523 2524void 2525ConcurrentMarkSweepGeneration::verify() { 2526 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2527 // are not called when the heap is verified during universe initialization and 2528 // at vm shutdown. 2529 if (freelistLock()->owned_by_self()) { 2530 cmsSpace()->verify(); 2531 } else { 2532 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2533 cmsSpace()->verify(); 2534 } 2535} 2536 2537void CMSCollector::verify() { 2538 _cmsGen->verify(); 2539} 2540 2541#ifndef PRODUCT 2542bool CMSCollector::overflow_list_is_empty() const { 2543 assert(_num_par_pushes >= 0, "Inconsistency"); 2544 if (_overflow_list == NULL) { 2545 assert(_num_par_pushes == 0, "Inconsistency"); 2546 } 2547 return _overflow_list == NULL; 2548} 2549 2550// The methods verify_work_stacks_empty() and verify_overflow_empty() 2551// merely consolidate assertion checks that appear to occur together frequently. 2552void CMSCollector::verify_work_stacks_empty() const { 2553 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2554 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2555} 2556 2557void CMSCollector::verify_overflow_empty() const { 2558 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2559 assert(no_preserved_marks(), "No preserved marks"); 2560} 2561#endif // PRODUCT 2562 2563// Decide if we want to enable class unloading as part of the 2564// ensuing concurrent GC cycle. We will collect and 2565// unload classes if it's the case that: 2566// (1) an explicit gc request has been made and the flag 2567// ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2568// (2) (a) class unloading is enabled at the command line, and 2569// (b) old gen is getting really full 2570// NOTE: Provided there is no change in the state of the heap between 2571// calls to this method, it should have idempotent results. Moreover, 2572// its results should be monotonically increasing (i.e. going from 0 to 1, 2573// but not 1 to 0) between successive calls between which the heap was 2574// not collected. For the implementation below, it must thus rely on 2575// the property that concurrent_cycles_since_last_unload() 2576// will not decrease unless a collection cycle happened and that 2577// _cmsGen->is_too_full() are 2578// themselves also monotonic in that sense. See check_monotonicity() 2579// below. 2580void CMSCollector::update_should_unload_classes() { 2581 _should_unload_classes = false; 2582 // Condition 1 above 2583 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2584 _should_unload_classes = true; 2585 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2586 // Disjuncts 2.b.(i,ii,iii) above 2587 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2588 CMSClassUnloadingMaxInterval) 2589 || _cmsGen->is_too_full(); 2590 } 2591} 2592 2593bool ConcurrentMarkSweepGeneration::is_too_full() const { 2594 bool res = should_concurrent_collect(); 2595 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2596 return res; 2597} 2598 2599void CMSCollector::setup_cms_unloading_and_verification_state() { 2600 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2601 || VerifyBeforeExit; 2602 const int rso = GenCollectedHeap::SO_AllCodeCache; 2603 2604 // We set the proper root for this CMS cycle here. 2605 if (should_unload_classes()) { // Should unload classes this cycle 2606 remove_root_scanning_option(rso); // Shrink the root set appropriately 2607 set_verifying(should_verify); // Set verification state for this cycle 2608 return; // Nothing else needs to be done at this time 2609 } 2610 2611 // Not unloading classes this cycle 2612 assert(!should_unload_classes(), "Inconsistency!"); 2613 2614 // If we are not unloading classes then add SO_AllCodeCache to root 2615 // scanning options. 2616 add_root_scanning_option(rso); 2617 2618 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2619 set_verifying(true); 2620 } else if (verifying() && !should_verify) { 2621 // We were verifying, but some verification flags got disabled. 2622 set_verifying(false); 2623 // Exclude symbols, strings and code cache elements from root scanning to 2624 // reduce IM and RM pauses. 2625 remove_root_scanning_option(rso); 2626 } 2627} 2628 2629 2630#ifndef PRODUCT 2631HeapWord* CMSCollector::block_start(const void* p) const { 2632 const HeapWord* addr = (HeapWord*)p; 2633 if (_span.contains(p)) { 2634 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2635 return _cmsGen->cmsSpace()->block_start(p); 2636 } 2637 } 2638 return NULL; 2639} 2640#endif 2641 2642HeapWord* 2643ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2644 bool tlab, 2645 bool parallel) { 2646 CMSSynchronousYieldRequest yr; 2647 assert(!tlab, "Can't deal with TLAB allocation"); 2648 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2649 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2650 if (GCExpandToAllocateDelayMillis > 0) { 2651 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2652 } 2653 return have_lock_and_allocate(word_size, tlab); 2654} 2655 2656void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2657 size_t bytes, 2658 size_t expand_bytes, 2659 CMSExpansionCause::Cause cause) 2660{ 2661 2662 bool success = expand(bytes, expand_bytes); 2663 2664 // remember why we expanded; this information is used 2665 // by shouldConcurrentCollect() when making decisions on whether to start 2666 // a new CMS cycle. 2667 if (success) { 2668 set_expansion_cause(cause); 2669 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2670 } 2671} 2672 2673HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2674 HeapWord* res = NULL; 2675 MutexLocker x(ParGCRareEvent_lock); 2676 while (true) { 2677 // Expansion by some other thread might make alloc OK now: 2678 res = ps->lab.alloc(word_sz); 2679 if (res != NULL) return res; 2680 // If there's not enough expansion space available, give up. 2681 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2682 return NULL; 2683 } 2684 // Otherwise, we try expansion. 2685 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2686 // Now go around the loop and try alloc again; 2687 // A competing par_promote might beat us to the expansion space, 2688 // so we may go around the loop again if promotion fails again. 2689 if (GCExpandToAllocateDelayMillis > 0) { 2690 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2691 } 2692 } 2693} 2694 2695 2696bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2697 PromotionInfo* promo) { 2698 MutexLocker x(ParGCRareEvent_lock); 2699 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2700 while (true) { 2701 // Expansion by some other thread might make alloc OK now: 2702 if (promo->ensure_spooling_space()) { 2703 assert(promo->has_spooling_space(), 2704 "Post-condition of successful ensure_spooling_space()"); 2705 return true; 2706 } 2707 // If there's not enough expansion space available, give up. 2708 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2709 return false; 2710 } 2711 // Otherwise, we try expansion. 2712 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2713 // Now go around the loop and try alloc again; 2714 // A competing allocation might beat us to the expansion space, 2715 // so we may go around the loop again if allocation fails again. 2716 if (GCExpandToAllocateDelayMillis > 0) { 2717 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2718 } 2719 } 2720} 2721 2722void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2723 // Only shrink if a compaction was done so that all the free space 2724 // in the generation is in a contiguous block at the end. 2725 if (did_compact()) { 2726 CardGeneration::shrink(bytes); 2727 } 2728} 2729 2730void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2731 assert_locked_or_safepoint(Heap_lock); 2732} 2733 2734void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2735 assert_locked_or_safepoint(Heap_lock); 2736 assert_lock_strong(freelistLock()); 2737 log_trace(gc)("Shrinking of CMS not yet implemented"); 2738 return; 2739} 2740 2741 2742// Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2743// phases. 2744class CMSPhaseAccounting: public StackObj { 2745 public: 2746 CMSPhaseAccounting(CMSCollector *collector, 2747 const char *title); 2748 ~CMSPhaseAccounting(); 2749 2750 private: 2751 CMSCollector *_collector; 2752 const char *_title; 2753 GCTraceConcTime(Info, gc) _trace_time; 2754 2755 public: 2756 // Not MT-safe; so do not pass around these StackObj's 2757 // where they may be accessed by other threads. 2758 double wallclock_millis() { 2759 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2760 } 2761}; 2762 2763CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2764 const char *title) : 2765 _collector(collector), _title(title), _trace_time(title) { 2766 2767 _collector->resetYields(); 2768 _collector->resetTimer(); 2769 _collector->startTimer(); 2770 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2771} 2772 2773CMSPhaseAccounting::~CMSPhaseAccounting() { 2774 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2775 _collector->stopTimer(); 2776 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2777 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2778} 2779 2780// CMS work 2781 2782// The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2783class CMSParMarkTask : public AbstractGangTask { 2784 protected: 2785 CMSCollector* _collector; 2786 uint _n_workers; 2787 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2788 AbstractGangTask(name), 2789 _collector(collector), 2790 _n_workers(n_workers) {} 2791 // Work method in support of parallel rescan ... of young gen spaces 2792 void do_young_space_rescan(OopsInGenClosure* cl, 2793 ContiguousSpace* space, 2794 HeapWord** chunk_array, size_t chunk_top); 2795 void work_on_young_gen_roots(OopsInGenClosure* cl); 2796}; 2797 2798// Parallel initial mark task 2799class CMSParInitialMarkTask: public CMSParMarkTask { 2800 StrongRootsScope* _strong_roots_scope; 2801 public: 2802 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2803 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2804 _strong_roots_scope(strong_roots_scope) {} 2805 void work(uint worker_id); 2806}; 2807 2808// Checkpoint the roots into this generation from outside 2809// this generation. [Note this initial checkpoint need only 2810// be approximate -- we'll do a catch up phase subsequently.] 2811void CMSCollector::checkpointRootsInitial() { 2812 assert(_collectorState == InitialMarking, "Wrong collector state"); 2813 check_correct_thread_executing(); 2814 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2815 2816 save_heap_summary(); 2817 report_heap_summary(GCWhen::BeforeGC); 2818 2819 ReferenceProcessor* rp = ref_processor(); 2820 assert(_restart_addr == NULL, "Control point invariant"); 2821 { 2822 // acquire locks for subsequent manipulations 2823 MutexLockerEx x(bitMapLock(), 2824 Mutex::_no_safepoint_check_flag); 2825 checkpointRootsInitialWork(); 2826 // enable ("weak") refs discovery 2827 rp->enable_discovery(); 2828 _collectorState = Marking; 2829 } 2830} 2831 2832void CMSCollector::checkpointRootsInitialWork() { 2833 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2834 assert(_collectorState == InitialMarking, "just checking"); 2835 2836 // Already have locks. 2837 assert_lock_strong(bitMapLock()); 2838 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2839 2840 // Setup the verification and class unloading state for this 2841 // CMS collection cycle. 2842 setup_cms_unloading_and_verification_state(); 2843 2844 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2845 2846 // Reset all the PLAB chunk arrays if necessary. 2847 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2848 reset_survivor_plab_arrays(); 2849 } 2850 2851 ResourceMark rm; 2852 HandleMark hm; 2853 2854 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2855 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2856 2857 verify_work_stacks_empty(); 2858 verify_overflow_empty(); 2859 2860 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2861 // Update the saved marks which may affect the root scans. 2862 gch->save_marks(); 2863 2864 // weak reference processing has not started yet. 2865 ref_processor()->set_enqueuing_is_done(false); 2866 2867 // Need to remember all newly created CLDs, 2868 // so that we can guarantee that the remark finds them. 2869 ClassLoaderDataGraph::remember_new_clds(true); 2870 2871 // Whenever a CLD is found, it will be claimed before proceeding to mark 2872 // the klasses. The claimed marks need to be cleared before marking starts. 2873 ClassLoaderDataGraph::clear_claimed_marks(); 2874 2875 print_eden_and_survivor_chunk_arrays(); 2876 2877 { 2878#if defined(COMPILER2) || INCLUDE_JVMCI 2879 DerivedPointerTableDeactivate dpt_deact; 2880#endif 2881 if (CMSParallelInitialMarkEnabled) { 2882 // The parallel version. 2883 WorkGang* workers = gch->workers(); 2884 assert(workers != NULL, "Need parallel worker threads."); 2885 uint n_workers = workers->active_workers(); 2886 2887 StrongRootsScope srs(n_workers); 2888 2889 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2890 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2891 // If the total workers is greater than 1, then multiple workers 2892 // may be used at some time and the initialization has been set 2893 // such that the single threaded path cannot be used. 2894 if (workers->total_workers() > 1) { 2895 workers->run_task(&tsk); 2896 } else { 2897 tsk.work(0); 2898 } 2899 } else { 2900 // The serial version. 2901 CLDToOopClosure cld_closure(¬Older, true); 2902 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2903 2904 StrongRootsScope srs(1); 2905 2906 gch->gen_process_roots(&srs, 2907 GenCollectedHeap::OldGen, 2908 true, // young gen as roots 2909 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2910 should_unload_classes(), 2911 ¬Older, 2912 NULL, 2913 &cld_closure); 2914 } 2915 } 2916 2917 // Clear mod-union table; it will be dirtied in the prologue of 2918 // CMS generation per each young generation collection. 2919 2920 assert(_modUnionTable.isAllClear(), 2921 "Was cleared in most recent final checkpoint phase" 2922 " or no bits are set in the gc_prologue before the start of the next " 2923 "subsequent marking phase."); 2924 2925 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2926 2927 // Save the end of the used_region of the constituent generations 2928 // to be used to limit the extent of sweep in each generation. 2929 save_sweep_limits(); 2930 verify_overflow_empty(); 2931} 2932 2933bool CMSCollector::markFromRoots() { 2934 // we might be tempted to assert that: 2935 // assert(!SafepointSynchronize::is_at_safepoint(), 2936 // "inconsistent argument?"); 2937 // However that wouldn't be right, because it's possible that 2938 // a safepoint is indeed in progress as a young generation 2939 // stop-the-world GC happens even as we mark in this generation. 2940 assert(_collectorState == Marking, "inconsistent state?"); 2941 check_correct_thread_executing(); 2942 verify_overflow_empty(); 2943 2944 // Weak ref discovery note: We may be discovering weak 2945 // refs in this generation concurrent (but interleaved) with 2946 // weak ref discovery by the young generation collector. 2947 2948 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2949 GCTraceCPUTime tcpu; 2950 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2951 bool res = markFromRootsWork(); 2952 if (res) { 2953 _collectorState = Precleaning; 2954 } else { // We failed and a foreground collection wants to take over 2955 assert(_foregroundGCIsActive, "internal state inconsistency"); 2956 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2957 log_debug(gc)("bailing out to foreground collection"); 2958 } 2959 verify_overflow_empty(); 2960 return res; 2961} 2962 2963bool CMSCollector::markFromRootsWork() { 2964 // iterate over marked bits in bit map, doing a full scan and mark 2965 // from these roots using the following algorithm: 2966 // . if oop is to the right of the current scan pointer, 2967 // mark corresponding bit (we'll process it later) 2968 // . else (oop is to left of current scan pointer) 2969 // push oop on marking stack 2970 // . drain the marking stack 2971 2972 // Note that when we do a marking step we need to hold the 2973 // bit map lock -- recall that direct allocation (by mutators) 2974 // and promotion (by the young generation collector) is also 2975 // marking the bit map. [the so-called allocate live policy.] 2976 // Because the implementation of bit map marking is not 2977 // robust wrt simultaneous marking of bits in the same word, 2978 // we need to make sure that there is no such interference 2979 // between concurrent such updates. 2980 2981 // already have locks 2982 assert_lock_strong(bitMapLock()); 2983 2984 verify_work_stacks_empty(); 2985 verify_overflow_empty(); 2986 bool result = false; 2987 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2988 result = do_marking_mt(); 2989 } else { 2990 result = do_marking_st(); 2991 } 2992 return result; 2993} 2994 2995// Forward decl 2996class CMSConcMarkingTask; 2997 2998class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2999 CMSCollector* _collector; 3000 CMSConcMarkingTask* _task; 3001 public: 3002 virtual void yield(); 3003 3004 // "n_threads" is the number of threads to be terminated. 3005 // "queue_set" is a set of work queues of other threads. 3006 // "collector" is the CMS collector associated with this task terminator. 3007 // "yield" indicates whether we need the gang as a whole to yield. 3008 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3009 ParallelTaskTerminator(n_threads, queue_set), 3010 _collector(collector) { } 3011 3012 void set_task(CMSConcMarkingTask* task) { 3013 _task = task; 3014 } 3015}; 3016 3017class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3018 CMSConcMarkingTask* _task; 3019 public: 3020 bool should_exit_termination(); 3021 void set_task(CMSConcMarkingTask* task) { 3022 _task = task; 3023 } 3024}; 3025 3026// MT Concurrent Marking Task 3027class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3028 CMSCollector* _collector; 3029 uint _n_workers; // requested/desired # workers 3030 bool _result; 3031 CompactibleFreeListSpace* _cms_space; 3032 char _pad_front[64]; // padding to ... 3033 HeapWord* _global_finger; // ... avoid sharing cache line 3034 char _pad_back[64]; 3035 HeapWord* _restart_addr; 3036 3037 // Exposed here for yielding support 3038 Mutex* const _bit_map_lock; 3039 3040 // The per thread work queues, available here for stealing 3041 OopTaskQueueSet* _task_queues; 3042 3043 // Termination (and yielding) support 3044 CMSConcMarkingTerminator _term; 3045 CMSConcMarkingTerminatorTerminator _term_term; 3046 3047 public: 3048 CMSConcMarkingTask(CMSCollector* collector, 3049 CompactibleFreeListSpace* cms_space, 3050 YieldingFlexibleWorkGang* workers, 3051 OopTaskQueueSet* task_queues): 3052 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3053 _collector(collector), 3054 _cms_space(cms_space), 3055 _n_workers(0), _result(true), 3056 _task_queues(task_queues), 3057 _term(_n_workers, task_queues, _collector), 3058 _bit_map_lock(collector->bitMapLock()) 3059 { 3060 _requested_size = _n_workers; 3061 _term.set_task(this); 3062 _term_term.set_task(this); 3063 _restart_addr = _global_finger = _cms_space->bottom(); 3064 } 3065 3066 3067 OopTaskQueueSet* task_queues() { return _task_queues; } 3068 3069 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3070 3071 HeapWord** global_finger_addr() { return &_global_finger; } 3072 3073 CMSConcMarkingTerminator* terminator() { return &_term; } 3074 3075 virtual void set_for_termination(uint active_workers) { 3076 terminator()->reset_for_reuse(active_workers); 3077 } 3078 3079 void work(uint worker_id); 3080 bool should_yield() { 3081 return ConcurrentMarkSweepThread::should_yield() 3082 && !_collector->foregroundGCIsActive(); 3083 } 3084 3085 virtual void coordinator_yield(); // stuff done by coordinator 3086 bool result() { return _result; } 3087 3088 void reset(HeapWord* ra) { 3089 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3090 _restart_addr = _global_finger = ra; 3091 _term.reset_for_reuse(); 3092 } 3093 3094 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3095 OopTaskQueue* work_q); 3096 3097 private: 3098 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3099 void do_work_steal(int i); 3100 void bump_global_finger(HeapWord* f); 3101}; 3102 3103bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3104 assert(_task != NULL, "Error"); 3105 return _task->yielding(); 3106 // Note that we do not need the disjunct || _task->should_yield() above 3107 // because we want terminating threads to yield only if the task 3108 // is already in the midst of yielding, which happens only after at least one 3109 // thread has yielded. 3110} 3111 3112void CMSConcMarkingTerminator::yield() { 3113 if (_task->should_yield()) { 3114 _task->yield(); 3115 } else { 3116 ParallelTaskTerminator::yield(); 3117 } 3118} 3119 3120//////////////////////////////////////////////////////////////// 3121// Concurrent Marking Algorithm Sketch 3122//////////////////////////////////////////////////////////////// 3123// Until all tasks exhausted (both spaces): 3124// -- claim next available chunk 3125// -- bump global finger via CAS 3126// -- find first object that starts in this chunk 3127// and start scanning bitmap from that position 3128// -- scan marked objects for oops 3129// -- CAS-mark target, and if successful: 3130// . if target oop is above global finger (volatile read) 3131// nothing to do 3132// . if target oop is in chunk and above local finger 3133// then nothing to do 3134// . else push on work-queue 3135// -- Deal with possible overflow issues: 3136// . local work-queue overflow causes stuff to be pushed on 3137// global (common) overflow queue 3138// . always first empty local work queue 3139// . then get a batch of oops from global work queue if any 3140// . then do work stealing 3141// -- When all tasks claimed (both spaces) 3142// and local work queue empty, 3143// then in a loop do: 3144// . check global overflow stack; steal a batch of oops and trace 3145// . try to steal from other threads oif GOS is empty 3146// . if neither is available, offer termination 3147// -- Terminate and return result 3148// 3149void CMSConcMarkingTask::work(uint worker_id) { 3150 elapsedTimer _timer; 3151 ResourceMark rm; 3152 HandleMark hm; 3153 3154 DEBUG_ONLY(_collector->verify_overflow_empty();) 3155 3156 // Before we begin work, our work queue should be empty 3157 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3158 // Scan the bitmap covering _cms_space, tracing through grey objects. 3159 _timer.start(); 3160 do_scan_and_mark(worker_id, _cms_space); 3161 _timer.stop(); 3162 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3163 3164 // ... do work stealing 3165 _timer.reset(); 3166 _timer.start(); 3167 do_work_steal(worker_id); 3168 _timer.stop(); 3169 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3170 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3171 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3172 // Note that under the current task protocol, the 3173 // following assertion is true even of the spaces 3174 // expanded since the completion of the concurrent 3175 // marking. XXX This will likely change under a strict 3176 // ABORT semantics. 3177 // After perm removal the comparison was changed to 3178 // greater than or equal to from strictly greater than. 3179 // Before perm removal the highest address sweep would 3180 // have been at the end of perm gen but now is at the 3181 // end of the tenured gen. 3182 assert(_global_finger >= _cms_space->end(), 3183 "All tasks have been completed"); 3184 DEBUG_ONLY(_collector->verify_overflow_empty();) 3185} 3186 3187void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3188 HeapWord* read = _global_finger; 3189 HeapWord* cur = read; 3190 while (f > read) { 3191 cur = read; 3192 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3193 if (cur == read) { 3194 // our cas succeeded 3195 assert(_global_finger >= f, "protocol consistency"); 3196 break; 3197 } 3198 } 3199} 3200 3201// This is really inefficient, and should be redone by 3202// using (not yet available) block-read and -write interfaces to the 3203// stack and the work_queue. XXX FIX ME !!! 3204bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3205 OopTaskQueue* work_q) { 3206 // Fast lock-free check 3207 if (ovflw_stk->length() == 0) { 3208 return false; 3209 } 3210 assert(work_q->size() == 0, "Shouldn't steal"); 3211 MutexLockerEx ml(ovflw_stk->par_lock(), 3212 Mutex::_no_safepoint_check_flag); 3213 // Grab up to 1/4 the size of the work queue 3214 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3215 (size_t)ParGCDesiredObjsFromOverflowList); 3216 num = MIN2(num, ovflw_stk->length()); 3217 for (int i = (int) num; i > 0; i--) { 3218 oop cur = ovflw_stk->pop(); 3219 assert(cur != NULL, "Counted wrong?"); 3220 work_q->push(cur); 3221 } 3222 return num > 0; 3223} 3224 3225void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3226 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3227 int n_tasks = pst->n_tasks(); 3228 // We allow that there may be no tasks to do here because 3229 // we are restarting after a stack overflow. 3230 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3231 uint nth_task = 0; 3232 3233 HeapWord* aligned_start = sp->bottom(); 3234 if (sp->used_region().contains(_restart_addr)) { 3235 // Align down to a card boundary for the start of 0th task 3236 // for this space. 3237 aligned_start = 3238 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3239 CardTableModRefBS::card_size); 3240 } 3241 3242 size_t chunk_size = sp->marking_task_size(); 3243 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3244 // Having claimed the nth task in this space, 3245 // compute the chunk that it corresponds to: 3246 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3247 aligned_start + (nth_task+1)*chunk_size); 3248 // Try and bump the global finger via a CAS; 3249 // note that we need to do the global finger bump 3250 // _before_ taking the intersection below, because 3251 // the task corresponding to that region will be 3252 // deemed done even if the used_region() expands 3253 // because of allocation -- as it almost certainly will 3254 // during start-up while the threads yield in the 3255 // closure below. 3256 HeapWord* finger = span.end(); 3257 bump_global_finger(finger); // atomically 3258 // There are null tasks here corresponding to chunks 3259 // beyond the "top" address of the space. 3260 span = span.intersection(sp->used_region()); 3261 if (!span.is_empty()) { // Non-null task 3262 HeapWord* prev_obj; 3263 assert(!span.contains(_restart_addr) || nth_task == 0, 3264 "Inconsistency"); 3265 if (nth_task == 0) { 3266 // For the 0th task, we'll not need to compute a block_start. 3267 if (span.contains(_restart_addr)) { 3268 // In the case of a restart because of stack overflow, 3269 // we might additionally skip a chunk prefix. 3270 prev_obj = _restart_addr; 3271 } else { 3272 prev_obj = span.start(); 3273 } 3274 } else { 3275 // We want to skip the first object because 3276 // the protocol is to scan any object in its entirety 3277 // that _starts_ in this span; a fortiori, any 3278 // object starting in an earlier span is scanned 3279 // as part of an earlier claimed task. 3280 // Below we use the "careful" version of block_start 3281 // so we do not try to navigate uninitialized objects. 3282 prev_obj = sp->block_start_careful(span.start()); 3283 // Below we use a variant of block_size that uses the 3284 // Printezis bits to avoid waiting for allocated 3285 // objects to become initialized/parsable. 3286 while (prev_obj < span.start()) { 3287 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3288 if (sz > 0) { 3289 prev_obj += sz; 3290 } else { 3291 // In this case we may end up doing a bit of redundant 3292 // scanning, but that appears unavoidable, short of 3293 // locking the free list locks; see bug 6324141. 3294 break; 3295 } 3296 } 3297 } 3298 if (prev_obj < span.end()) { 3299 MemRegion my_span = MemRegion(prev_obj, span.end()); 3300 // Do the marking work within a non-empty span -- 3301 // the last argument to the constructor indicates whether the 3302 // iteration should be incremental with periodic yields. 3303 ParMarkFromRootsClosure cl(this, _collector, my_span, 3304 &_collector->_markBitMap, 3305 work_queue(i), 3306 &_collector->_markStack); 3307 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3308 } // else nothing to do for this task 3309 } // else nothing to do for this task 3310 } 3311 // We'd be tempted to assert here that since there are no 3312 // more tasks left to claim in this space, the global_finger 3313 // must exceed space->top() and a fortiori space->end(). However, 3314 // that would not quite be correct because the bumping of 3315 // global_finger occurs strictly after the claiming of a task, 3316 // so by the time we reach here the global finger may not yet 3317 // have been bumped up by the thread that claimed the last 3318 // task. 3319 pst->all_tasks_completed(); 3320} 3321 3322class ParConcMarkingClosure: public MetadataAwareOopClosure { 3323 private: 3324 CMSCollector* _collector; 3325 CMSConcMarkingTask* _task; 3326 MemRegion _span; 3327 CMSBitMap* _bit_map; 3328 CMSMarkStack* _overflow_stack; 3329 OopTaskQueue* _work_queue; 3330 protected: 3331 DO_OOP_WORK_DEFN 3332 public: 3333 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3334 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3335 MetadataAwareOopClosure(collector->ref_processor()), 3336 _collector(collector), 3337 _task(task), 3338 _span(collector->_span), 3339 _work_queue(work_queue), 3340 _bit_map(bit_map), 3341 _overflow_stack(overflow_stack) 3342 { } 3343 virtual void do_oop(oop* p); 3344 virtual void do_oop(narrowOop* p); 3345 3346 void trim_queue(size_t max); 3347 void handle_stack_overflow(HeapWord* lost); 3348 void do_yield_check() { 3349 if (_task->should_yield()) { 3350 _task->yield(); 3351 } 3352 } 3353}; 3354 3355DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3356 3357// Grey object scanning during work stealing phase -- 3358// the salient assumption here is that any references 3359// that are in these stolen objects being scanned must 3360// already have been initialized (else they would not have 3361// been published), so we do not need to check for 3362// uninitialized objects before pushing here. 3363void ParConcMarkingClosure::do_oop(oop obj) { 3364 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3365 HeapWord* addr = (HeapWord*)obj; 3366 // Check if oop points into the CMS generation 3367 // and is not marked 3368 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3369 // a white object ... 3370 // If we manage to "claim" the object, by being the 3371 // first thread to mark it, then we push it on our 3372 // marking stack 3373 if (_bit_map->par_mark(addr)) { // ... now grey 3374 // push on work queue (grey set) 3375 bool simulate_overflow = false; 3376 NOT_PRODUCT( 3377 if (CMSMarkStackOverflowALot && 3378 _collector->simulate_overflow()) { 3379 // simulate a stack overflow 3380 simulate_overflow = true; 3381 } 3382 ) 3383 if (simulate_overflow || 3384 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3385 // stack overflow 3386 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3387 // We cannot assert that the overflow stack is full because 3388 // it may have been emptied since. 3389 assert(simulate_overflow || 3390 _work_queue->size() == _work_queue->max_elems(), 3391 "Else push should have succeeded"); 3392 handle_stack_overflow(addr); 3393 } 3394 } // Else, some other thread got there first 3395 do_yield_check(); 3396 } 3397} 3398 3399void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3400void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3401 3402void ParConcMarkingClosure::trim_queue(size_t max) { 3403 while (_work_queue->size() > max) { 3404 oop new_oop; 3405 if (_work_queue->pop_local(new_oop)) { 3406 assert(new_oop->is_oop(), "Should be an oop"); 3407 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3408 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3409 new_oop->oop_iterate(this); // do_oop() above 3410 do_yield_check(); 3411 } 3412 } 3413} 3414 3415// Upon stack overflow, we discard (part of) the stack, 3416// remembering the least address amongst those discarded 3417// in CMSCollector's _restart_address. 3418void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3419 // We need to do this under a mutex to prevent other 3420 // workers from interfering with the work done below. 3421 MutexLockerEx ml(_overflow_stack->par_lock(), 3422 Mutex::_no_safepoint_check_flag); 3423 // Remember the least grey address discarded 3424 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3425 _collector->lower_restart_addr(ra); 3426 _overflow_stack->reset(); // discard stack contents 3427 _overflow_stack->expand(); // expand the stack if possible 3428} 3429 3430 3431void CMSConcMarkingTask::do_work_steal(int i) { 3432 OopTaskQueue* work_q = work_queue(i); 3433 oop obj_to_scan; 3434 CMSBitMap* bm = &(_collector->_markBitMap); 3435 CMSMarkStack* ovflw = &(_collector->_markStack); 3436 int* seed = _collector->hash_seed(i); 3437 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3438 while (true) { 3439 cl.trim_queue(0); 3440 assert(work_q->size() == 0, "Should have been emptied above"); 3441 if (get_work_from_overflow_stack(ovflw, work_q)) { 3442 // Can't assert below because the work obtained from the 3443 // overflow stack may already have been stolen from us. 3444 // assert(work_q->size() > 0, "Work from overflow stack"); 3445 continue; 3446 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3447 assert(obj_to_scan->is_oop(), "Should be an oop"); 3448 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3449 obj_to_scan->oop_iterate(&cl); 3450 } else if (terminator()->offer_termination(&_term_term)) { 3451 assert(work_q->size() == 0, "Impossible!"); 3452 break; 3453 } else if (yielding() || should_yield()) { 3454 yield(); 3455 } 3456 } 3457} 3458 3459// This is run by the CMS (coordinator) thread. 3460void CMSConcMarkingTask::coordinator_yield() { 3461 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3462 "CMS thread should hold CMS token"); 3463 // First give up the locks, then yield, then re-lock 3464 // We should probably use a constructor/destructor idiom to 3465 // do this unlock/lock or modify the MutexUnlocker class to 3466 // serve our purpose. XXX 3467 assert_lock_strong(_bit_map_lock); 3468 _bit_map_lock->unlock(); 3469 ConcurrentMarkSweepThread::desynchronize(true); 3470 _collector->stopTimer(); 3471 _collector->incrementYields(); 3472 3473 // It is possible for whichever thread initiated the yield request 3474 // not to get a chance to wake up and take the bitmap lock between 3475 // this thread releasing it and reacquiring it. So, while the 3476 // should_yield() flag is on, let's sleep for a bit to give the 3477 // other thread a chance to wake up. The limit imposed on the number 3478 // of iterations is defensive, to avoid any unforseen circumstances 3479 // putting us into an infinite loop. Since it's always been this 3480 // (coordinator_yield()) method that was observed to cause the 3481 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3482 // which is by default non-zero. For the other seven methods that 3483 // also perform the yield operation, as are using a different 3484 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3485 // can enable the sleeping for those methods too, if necessary. 3486 // See 6442774. 3487 // 3488 // We really need to reconsider the synchronization between the GC 3489 // thread and the yield-requesting threads in the future and we 3490 // should really use wait/notify, which is the recommended 3491 // way of doing this type of interaction. Additionally, we should 3492 // consolidate the eight methods that do the yield operation and they 3493 // are almost identical into one for better maintainability and 3494 // readability. See 6445193. 3495 // 3496 // Tony 2006.06.29 3497 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3498 ConcurrentMarkSweepThread::should_yield() && 3499 !CMSCollector::foregroundGCIsActive(); ++i) { 3500 os::sleep(Thread::current(), 1, false); 3501 } 3502 3503 ConcurrentMarkSweepThread::synchronize(true); 3504 _bit_map_lock->lock_without_safepoint_check(); 3505 _collector->startTimer(); 3506} 3507 3508bool CMSCollector::do_marking_mt() { 3509 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3510 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3511 conc_workers()->active_workers(), 3512 Threads::number_of_non_daemon_threads()); 3513 num_workers = conc_workers()->update_active_workers(num_workers); 3514 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3515 3516 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3517 3518 CMSConcMarkingTask tsk(this, 3519 cms_space, 3520 conc_workers(), 3521 task_queues()); 3522 3523 // Since the actual number of workers we get may be different 3524 // from the number we requested above, do we need to do anything different 3525 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3526 // class?? XXX 3527 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3528 3529 // Refs discovery is already non-atomic. 3530 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3531 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3532 conc_workers()->start_task(&tsk); 3533 while (tsk.yielded()) { 3534 tsk.coordinator_yield(); 3535 conc_workers()->continue_task(&tsk); 3536 } 3537 // If the task was aborted, _restart_addr will be non-NULL 3538 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3539 while (_restart_addr != NULL) { 3540 // XXX For now we do not make use of ABORTED state and have not 3541 // yet implemented the right abort semantics (even in the original 3542 // single-threaded CMS case). That needs some more investigation 3543 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3544 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3545 // If _restart_addr is non-NULL, a marking stack overflow 3546 // occurred; we need to do a fresh marking iteration from the 3547 // indicated restart address. 3548 if (_foregroundGCIsActive) { 3549 // We may be running into repeated stack overflows, having 3550 // reached the limit of the stack size, while making very 3551 // slow forward progress. It may be best to bail out and 3552 // let the foreground collector do its job. 3553 // Clear _restart_addr, so that foreground GC 3554 // works from scratch. This avoids the headache of 3555 // a "rescan" which would otherwise be needed because 3556 // of the dirty mod union table & card table. 3557 _restart_addr = NULL; 3558 return false; 3559 } 3560 // Adjust the task to restart from _restart_addr 3561 tsk.reset(_restart_addr); 3562 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3563 _restart_addr); 3564 _restart_addr = NULL; 3565 // Get the workers going again 3566 conc_workers()->start_task(&tsk); 3567 while (tsk.yielded()) { 3568 tsk.coordinator_yield(); 3569 conc_workers()->continue_task(&tsk); 3570 } 3571 } 3572 assert(tsk.completed(), "Inconsistency"); 3573 assert(tsk.result() == true, "Inconsistency"); 3574 return true; 3575} 3576 3577bool CMSCollector::do_marking_st() { 3578 ResourceMark rm; 3579 HandleMark hm; 3580 3581 // Temporarily make refs discovery single threaded (non-MT) 3582 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3583 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3584 &_markStack, CMSYield); 3585 // the last argument to iterate indicates whether the iteration 3586 // should be incremental with periodic yields. 3587 _markBitMap.iterate(&markFromRootsClosure); 3588 // If _restart_addr is non-NULL, a marking stack overflow 3589 // occurred; we need to do a fresh iteration from the 3590 // indicated restart address. 3591 while (_restart_addr != NULL) { 3592 if (_foregroundGCIsActive) { 3593 // We may be running into repeated stack overflows, having 3594 // reached the limit of the stack size, while making very 3595 // slow forward progress. It may be best to bail out and 3596 // let the foreground collector do its job. 3597 // Clear _restart_addr, so that foreground GC 3598 // works from scratch. This avoids the headache of 3599 // a "rescan" which would otherwise be needed because 3600 // of the dirty mod union table & card table. 3601 _restart_addr = NULL; 3602 return false; // indicating failure to complete marking 3603 } 3604 // Deal with stack overflow: 3605 // we restart marking from _restart_addr 3606 HeapWord* ra = _restart_addr; 3607 markFromRootsClosure.reset(ra); 3608 _restart_addr = NULL; 3609 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3610 } 3611 return true; 3612} 3613 3614void CMSCollector::preclean() { 3615 check_correct_thread_executing(); 3616 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3617 verify_work_stacks_empty(); 3618 verify_overflow_empty(); 3619 _abort_preclean = false; 3620 if (CMSPrecleaningEnabled) { 3621 if (!CMSEdenChunksRecordAlways) { 3622 _eden_chunk_index = 0; 3623 } 3624 size_t used = get_eden_used(); 3625 size_t capacity = get_eden_capacity(); 3626 // Don't start sampling unless we will get sufficiently 3627 // many samples. 3628 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3629 * CMSScheduleRemarkEdenPenetration)) { 3630 _start_sampling = true; 3631 } else { 3632 _start_sampling = false; 3633 } 3634 GCTraceCPUTime tcpu; 3635 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3636 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3637 } 3638 CMSTokenSync x(true); // is cms thread 3639 if (CMSPrecleaningEnabled) { 3640 sample_eden(); 3641 _collectorState = AbortablePreclean; 3642 } else { 3643 _collectorState = FinalMarking; 3644 } 3645 verify_work_stacks_empty(); 3646 verify_overflow_empty(); 3647} 3648 3649// Try and schedule the remark such that young gen 3650// occupancy is CMSScheduleRemarkEdenPenetration %. 3651void CMSCollector::abortable_preclean() { 3652 check_correct_thread_executing(); 3653 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3654 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3655 3656 // If Eden's current occupancy is below this threshold, 3657 // immediately schedule the remark; else preclean 3658 // past the next scavenge in an effort to 3659 // schedule the pause as described above. By choosing 3660 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3661 // we will never do an actual abortable preclean cycle. 3662 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3663 GCTraceCPUTime tcpu; 3664 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3665 // We need more smarts in the abortable preclean 3666 // loop below to deal with cases where allocation 3667 // in young gen is very very slow, and our precleaning 3668 // is running a losing race against a horde of 3669 // mutators intent on flooding us with CMS updates 3670 // (dirty cards). 3671 // One, admittedly dumb, strategy is to give up 3672 // after a certain number of abortable precleaning loops 3673 // or after a certain maximum time. We want to make 3674 // this smarter in the next iteration. 3675 // XXX FIX ME!!! YSR 3676 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3677 while (!(should_abort_preclean() || 3678 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3679 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3680 cumworkdone += workdone; 3681 loops++; 3682 // Voluntarily terminate abortable preclean phase if we have 3683 // been at it for too long. 3684 if ((CMSMaxAbortablePrecleanLoops != 0) && 3685 loops >= CMSMaxAbortablePrecleanLoops) { 3686 log_debug(gc)(" CMS: abort preclean due to loops "); 3687 break; 3688 } 3689 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3690 log_debug(gc)(" CMS: abort preclean due to time "); 3691 break; 3692 } 3693 // If we are doing little work each iteration, we should 3694 // take a short break. 3695 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3696 // Sleep for some time, waiting for work to accumulate 3697 stopTimer(); 3698 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3699 startTimer(); 3700 waited++; 3701 } 3702 } 3703 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3704 loops, waited, cumworkdone); 3705 } 3706 CMSTokenSync x(true); // is cms thread 3707 if (_collectorState != Idling) { 3708 assert(_collectorState == AbortablePreclean, 3709 "Spontaneous state transition?"); 3710 _collectorState = FinalMarking; 3711 } // Else, a foreground collection completed this CMS cycle. 3712 return; 3713} 3714 3715// Respond to an Eden sampling opportunity 3716void CMSCollector::sample_eden() { 3717 // Make sure a young gc cannot sneak in between our 3718 // reading and recording of a sample. 3719 assert(Thread::current()->is_ConcurrentGC_thread(), 3720 "Only the cms thread may collect Eden samples"); 3721 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3722 "Should collect samples while holding CMS token"); 3723 if (!_start_sampling) { 3724 return; 3725 } 3726 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3727 // is populated by the young generation. 3728 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3729 if (_eden_chunk_index < _eden_chunk_capacity) { 3730 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3731 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3732 "Unexpected state of Eden"); 3733 // We'd like to check that what we just sampled is an oop-start address; 3734 // however, we cannot do that here since the object may not yet have been 3735 // initialized. So we'll instead do the check when we _use_ this sample 3736 // later. 3737 if (_eden_chunk_index == 0 || 3738 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3739 _eden_chunk_array[_eden_chunk_index-1]) 3740 >= CMSSamplingGrain)) { 3741 _eden_chunk_index++; // commit sample 3742 } 3743 } 3744 } 3745 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3746 size_t used = get_eden_used(); 3747 size_t capacity = get_eden_capacity(); 3748 assert(used <= capacity, "Unexpected state of Eden"); 3749 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3750 _abort_preclean = true; 3751 } 3752 } 3753} 3754 3755 3756size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3757 assert(_collectorState == Precleaning || 3758 _collectorState == AbortablePreclean, "incorrect state"); 3759 ResourceMark rm; 3760 HandleMark hm; 3761 3762 // Precleaning is currently not MT but the reference processor 3763 // may be set for MT. Disable it temporarily here. 3764 ReferenceProcessor* rp = ref_processor(); 3765 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3766 3767 // Do one pass of scrubbing the discovered reference lists 3768 // to remove any reference objects with strongly-reachable 3769 // referents. 3770 if (clean_refs) { 3771 CMSPrecleanRefsYieldClosure yield_cl(this); 3772 assert(rp->span().equals(_span), "Spans should be equal"); 3773 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3774 &_markStack, true /* preclean */); 3775 CMSDrainMarkingStackClosure complete_trace(this, 3776 _span, &_markBitMap, &_markStack, 3777 &keep_alive, true /* preclean */); 3778 3779 // We don't want this step to interfere with a young 3780 // collection because we don't want to take CPU 3781 // or memory bandwidth away from the young GC threads 3782 // (which may be as many as there are CPUs). 3783 // Note that we don't need to protect ourselves from 3784 // interference with mutators because they can't 3785 // manipulate the discovered reference lists nor affect 3786 // the computed reachability of the referents, the 3787 // only properties manipulated by the precleaning 3788 // of these reference lists. 3789 stopTimer(); 3790 CMSTokenSyncWithLocks x(true /* is cms thread */, 3791 bitMapLock()); 3792 startTimer(); 3793 sample_eden(); 3794 3795 // The following will yield to allow foreground 3796 // collection to proceed promptly. XXX YSR: 3797 // The code in this method may need further 3798 // tweaking for better performance and some restructuring 3799 // for cleaner interfaces. 3800 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3801 rp->preclean_discovered_references( 3802 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3803 gc_timer); 3804 } 3805 3806 if (clean_survivor) { // preclean the active survivor space(s) 3807 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3808 &_markBitMap, &_modUnionTable, 3809 &_markStack, true /* precleaning phase */); 3810 stopTimer(); 3811 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3812 bitMapLock()); 3813 startTimer(); 3814 unsigned int before_count = 3815 GenCollectedHeap::heap()->total_collections(); 3816 SurvivorSpacePrecleanClosure 3817 sss_cl(this, _span, &_markBitMap, &_markStack, 3818 &pam_cl, before_count, CMSYield); 3819 _young_gen->from()->object_iterate_careful(&sss_cl); 3820 _young_gen->to()->object_iterate_careful(&sss_cl); 3821 } 3822 MarkRefsIntoAndScanClosure 3823 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3824 &_markStack, this, CMSYield, 3825 true /* precleaning phase */); 3826 // CAUTION: The following closure has persistent state that may need to 3827 // be reset upon a decrease in the sequence of addresses it 3828 // processes. 3829 ScanMarkedObjectsAgainCarefullyClosure 3830 smoac_cl(this, _span, 3831 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3832 3833 // Preclean dirty cards in ModUnionTable and CardTable using 3834 // appropriate convergence criterion; 3835 // repeat CMSPrecleanIter times unless we find that 3836 // we are losing. 3837 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3838 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3839 "Bad convergence multiplier"); 3840 assert(CMSPrecleanThreshold >= 100, 3841 "Unreasonably low CMSPrecleanThreshold"); 3842 3843 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3844 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3845 numIter < CMSPrecleanIter; 3846 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3847 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3848 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3849 // Either there are very few dirty cards, so re-mark 3850 // pause will be small anyway, or our pre-cleaning isn't 3851 // that much faster than the rate at which cards are being 3852 // dirtied, so we might as well stop and re-mark since 3853 // precleaning won't improve our re-mark time by much. 3854 if (curNumCards <= CMSPrecleanThreshold || 3855 (numIter > 0 && 3856 (curNumCards * CMSPrecleanDenominator > 3857 lastNumCards * CMSPrecleanNumerator))) { 3858 numIter++; 3859 cumNumCards += curNumCards; 3860 break; 3861 } 3862 } 3863 3864 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3865 3866 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3867 cumNumCards += curNumCards; 3868 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3869 curNumCards, cumNumCards, numIter); 3870 return cumNumCards; // as a measure of useful work done 3871} 3872 3873// PRECLEANING NOTES: 3874// Precleaning involves: 3875// . reading the bits of the modUnionTable and clearing the set bits. 3876// . For the cards corresponding to the set bits, we scan the 3877// objects on those cards. This means we need the free_list_lock 3878// so that we can safely iterate over the CMS space when scanning 3879// for oops. 3880// . When we scan the objects, we'll be both reading and setting 3881// marks in the marking bit map, so we'll need the marking bit map. 3882// . For protecting _collector_state transitions, we take the CGC_lock. 3883// Note that any races in the reading of of card table entries by the 3884// CMS thread on the one hand and the clearing of those entries by the 3885// VM thread or the setting of those entries by the mutator threads on the 3886// other are quite benign. However, for efficiency it makes sense to keep 3887// the VM thread from racing with the CMS thread while the latter is 3888// dirty card info to the modUnionTable. We therefore also use the 3889// CGC_lock to protect the reading of the card table and the mod union 3890// table by the CM thread. 3891// . We run concurrently with mutator updates, so scanning 3892// needs to be done carefully -- we should not try to scan 3893// potentially uninitialized objects. 3894// 3895// Locking strategy: While holding the CGC_lock, we scan over and 3896// reset a maximal dirty range of the mod union / card tables, then lock 3897// the free_list_lock and bitmap lock to do a full marking, then 3898// release these locks; and repeat the cycle. This allows for a 3899// certain amount of fairness in the sharing of these locks between 3900// the CMS collector on the one hand, and the VM thread and the 3901// mutators on the other. 3902 3903// NOTE: preclean_mod_union_table() and preclean_card_table() 3904// further below are largely identical; if you need to modify 3905// one of these methods, please check the other method too. 3906 3907size_t CMSCollector::preclean_mod_union_table( 3908 ConcurrentMarkSweepGeneration* old_gen, 3909 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3910 verify_work_stacks_empty(); 3911 verify_overflow_empty(); 3912 3913 // strategy: starting with the first card, accumulate contiguous 3914 // ranges of dirty cards; clear these cards, then scan the region 3915 // covered by these cards. 3916 3917 // Since all of the MUT is committed ahead, we can just use 3918 // that, in case the generations expand while we are precleaning. 3919 // It might also be fine to just use the committed part of the 3920 // generation, but we might potentially miss cards when the 3921 // generation is rapidly expanding while we are in the midst 3922 // of precleaning. 3923 HeapWord* startAddr = old_gen->reserved().start(); 3924 HeapWord* endAddr = old_gen->reserved().end(); 3925 3926 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3927 3928 size_t numDirtyCards, cumNumDirtyCards; 3929 HeapWord *nextAddr, *lastAddr; 3930 for (cumNumDirtyCards = numDirtyCards = 0, 3931 nextAddr = lastAddr = startAddr; 3932 nextAddr < endAddr; 3933 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3934 3935 ResourceMark rm; 3936 HandleMark hm; 3937 3938 MemRegion dirtyRegion; 3939 { 3940 stopTimer(); 3941 // Potential yield point 3942 CMSTokenSync ts(true); 3943 startTimer(); 3944 sample_eden(); 3945 // Get dirty region starting at nextOffset (inclusive), 3946 // simultaneously clearing it. 3947 dirtyRegion = 3948 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3949 assert(dirtyRegion.start() >= nextAddr, 3950 "returned region inconsistent?"); 3951 } 3952 // Remember where the next search should begin. 3953 // The returned region (if non-empty) is a right open interval, 3954 // so lastOffset is obtained from the right end of that 3955 // interval. 3956 lastAddr = dirtyRegion.end(); 3957 // Should do something more transparent and less hacky XXX 3958 numDirtyCards = 3959 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3960 3961 // We'll scan the cards in the dirty region (with periodic 3962 // yields for foreground GC as needed). 3963 if (!dirtyRegion.is_empty()) { 3964 assert(numDirtyCards > 0, "consistency check"); 3965 HeapWord* stop_point = NULL; 3966 stopTimer(); 3967 // Potential yield point 3968 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3969 bitMapLock()); 3970 startTimer(); 3971 { 3972 verify_work_stacks_empty(); 3973 verify_overflow_empty(); 3974 sample_eden(); 3975 stop_point = 3976 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3977 } 3978 if (stop_point != NULL) { 3979 // The careful iteration stopped early either because it found an 3980 // uninitialized object, or because we were in the midst of an 3981 // "abortable preclean", which should now be aborted. Redirty 3982 // the bits corresponding to the partially-scanned or unscanned 3983 // cards. We'll either restart at the next block boundary or 3984 // abort the preclean. 3985 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3986 "Should only be AbortablePreclean."); 3987 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3988 if (should_abort_preclean()) { 3989 break; // out of preclean loop 3990 } else { 3991 // Compute the next address at which preclean should pick up; 3992 // might need bitMapLock in order to read P-bits. 3993 lastAddr = next_card_start_after_block(stop_point); 3994 } 3995 } 3996 } else { 3997 assert(lastAddr == endAddr, "consistency check"); 3998 assert(numDirtyCards == 0, "consistency check"); 3999 break; 4000 } 4001 } 4002 verify_work_stacks_empty(); 4003 verify_overflow_empty(); 4004 return cumNumDirtyCards; 4005} 4006 4007// NOTE: preclean_mod_union_table() above and preclean_card_table() 4008// below are largely identical; if you need to modify 4009// one of these methods, please check the other method too. 4010 4011size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4012 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4013 // strategy: it's similar to precleamModUnionTable above, in that 4014 // we accumulate contiguous ranges of dirty cards, mark these cards 4015 // precleaned, then scan the region covered by these cards. 4016 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4017 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4018 4019 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4020 4021 size_t numDirtyCards, cumNumDirtyCards; 4022 HeapWord *lastAddr, *nextAddr; 4023 4024 for (cumNumDirtyCards = numDirtyCards = 0, 4025 nextAddr = lastAddr = startAddr; 4026 nextAddr < endAddr; 4027 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4028 4029 ResourceMark rm; 4030 HandleMark hm; 4031 4032 MemRegion dirtyRegion; 4033 { 4034 // See comments in "Precleaning notes" above on why we 4035 // do this locking. XXX Could the locking overheads be 4036 // too high when dirty cards are sparse? [I don't think so.] 4037 stopTimer(); 4038 CMSTokenSync x(true); // is cms thread 4039 startTimer(); 4040 sample_eden(); 4041 // Get and clear dirty region from card table 4042 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4043 MemRegion(nextAddr, endAddr), 4044 true, 4045 CardTableModRefBS::precleaned_card_val()); 4046 4047 assert(dirtyRegion.start() >= nextAddr, 4048 "returned region inconsistent?"); 4049 } 4050 lastAddr = dirtyRegion.end(); 4051 numDirtyCards = 4052 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4053 4054 if (!dirtyRegion.is_empty()) { 4055 stopTimer(); 4056 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4057 startTimer(); 4058 sample_eden(); 4059 verify_work_stacks_empty(); 4060 verify_overflow_empty(); 4061 HeapWord* stop_point = 4062 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4063 if (stop_point != NULL) { 4064 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4065 "Should only be AbortablePreclean."); 4066 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4067 if (should_abort_preclean()) { 4068 break; // out of preclean loop 4069 } else { 4070 // Compute the next address at which preclean should pick up. 4071 lastAddr = next_card_start_after_block(stop_point); 4072 } 4073 } 4074 } else { 4075 break; 4076 } 4077 } 4078 verify_work_stacks_empty(); 4079 verify_overflow_empty(); 4080 return cumNumDirtyCards; 4081} 4082 4083class PrecleanKlassClosure : public KlassClosure { 4084 KlassToOopClosure _cm_klass_closure; 4085 public: 4086 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4087 void do_klass(Klass* k) { 4088 if (k->has_accumulated_modified_oops()) { 4089 k->clear_accumulated_modified_oops(); 4090 4091 _cm_klass_closure.do_klass(k); 4092 } 4093 } 4094}; 4095 4096// The freelist lock is needed to prevent asserts, is it really needed? 4097void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4098 4099 cl->set_freelistLock(freelistLock); 4100 4101 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4102 4103 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4104 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4105 PrecleanKlassClosure preclean_klass_closure(cl); 4106 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4107 4108 verify_work_stacks_empty(); 4109 verify_overflow_empty(); 4110} 4111 4112void CMSCollector::checkpointRootsFinal() { 4113 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4114 check_correct_thread_executing(); 4115 // world is stopped at this checkpoint 4116 assert(SafepointSynchronize::is_at_safepoint(), 4117 "world should be stopped"); 4118 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4119 4120 verify_work_stacks_empty(); 4121 verify_overflow_empty(); 4122 4123 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4124 _young_gen->used() / K, _young_gen->capacity() / K); 4125 { 4126 if (CMSScavengeBeforeRemark) { 4127 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4128 // Temporarily set flag to false, GCH->do_collection will 4129 // expect it to be false and set to true 4130 FlagSetting fl(gch->_is_gc_active, false); 4131 4132 gch->do_collection(true, // full (i.e. force, see below) 4133 false, // !clear_all_soft_refs 4134 0, // size 4135 false, // is_tlab 4136 GenCollectedHeap::YoungGen // type 4137 ); 4138 } 4139 FreelistLocker x(this); 4140 MutexLockerEx y(bitMapLock(), 4141 Mutex::_no_safepoint_check_flag); 4142 checkpointRootsFinalWork(); 4143 } 4144 verify_work_stacks_empty(); 4145 verify_overflow_empty(); 4146} 4147 4148void CMSCollector::checkpointRootsFinalWork() { 4149 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4150 4151 assert(haveFreelistLocks(), "must have free list locks"); 4152 assert_lock_strong(bitMapLock()); 4153 4154 ResourceMark rm; 4155 HandleMark hm; 4156 4157 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4158 4159 if (should_unload_classes()) { 4160 CodeCache::gc_prologue(); 4161 } 4162 assert(haveFreelistLocks(), "must have free list locks"); 4163 assert_lock_strong(bitMapLock()); 4164 4165 // We might assume that we need not fill TLAB's when 4166 // CMSScavengeBeforeRemark is set, because we may have just done 4167 // a scavenge which would have filled all TLAB's -- and besides 4168 // Eden would be empty. This however may not always be the case -- 4169 // for instance although we asked for a scavenge, it may not have 4170 // happened because of a JNI critical section. We probably need 4171 // a policy for deciding whether we can in that case wait until 4172 // the critical section releases and then do the remark following 4173 // the scavenge, and skip it here. In the absence of that policy, 4174 // or of an indication of whether the scavenge did indeed occur, 4175 // we cannot rely on TLAB's having been filled and must do 4176 // so here just in case a scavenge did not happen. 4177 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4178 // Update the saved marks which may affect the root scans. 4179 gch->save_marks(); 4180 4181 print_eden_and_survivor_chunk_arrays(); 4182 4183 { 4184#if defined(COMPILER2) || INCLUDE_JVMCI 4185 DerivedPointerTableDeactivate dpt_deact; 4186#endif 4187 4188 // Note on the role of the mod union table: 4189 // Since the marker in "markFromRoots" marks concurrently with 4190 // mutators, it is possible for some reachable objects not to have been 4191 // scanned. For instance, an only reference to an object A was 4192 // placed in object B after the marker scanned B. Unless B is rescanned, 4193 // A would be collected. Such updates to references in marked objects 4194 // are detected via the mod union table which is the set of all cards 4195 // dirtied since the first checkpoint in this GC cycle and prior to 4196 // the most recent young generation GC, minus those cleaned up by the 4197 // concurrent precleaning. 4198 if (CMSParallelRemarkEnabled) { 4199 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4200 do_remark_parallel(); 4201 } else { 4202 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4203 do_remark_non_parallel(); 4204 } 4205 } 4206 verify_work_stacks_empty(); 4207 verify_overflow_empty(); 4208 4209 { 4210 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4211 refProcessingWork(); 4212 } 4213 verify_work_stacks_empty(); 4214 verify_overflow_empty(); 4215 4216 if (should_unload_classes()) { 4217 CodeCache::gc_epilogue(); 4218 } 4219 JvmtiExport::gc_epilogue(); 4220 4221 // If we encountered any (marking stack / work queue) overflow 4222 // events during the current CMS cycle, take appropriate 4223 // remedial measures, where possible, so as to try and avoid 4224 // recurrence of that condition. 4225 assert(_markStack.isEmpty(), "No grey objects"); 4226 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4227 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4228 if (ser_ovflw > 0) { 4229 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4230 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4231 _markStack.expand(); 4232 _ser_pmc_remark_ovflw = 0; 4233 _ser_pmc_preclean_ovflw = 0; 4234 _ser_kac_preclean_ovflw = 0; 4235 _ser_kac_ovflw = 0; 4236 } 4237 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4238 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4239 _par_pmc_remark_ovflw, _par_kac_ovflw); 4240 _par_pmc_remark_ovflw = 0; 4241 _par_kac_ovflw = 0; 4242 } 4243 if (_markStack._hit_limit > 0) { 4244 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4245 _markStack._hit_limit); 4246 } 4247 if (_markStack._failed_double > 0) { 4248 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4249 _markStack._failed_double, _markStack.capacity()); 4250 } 4251 _markStack._hit_limit = 0; 4252 _markStack._failed_double = 0; 4253 4254 if ((VerifyAfterGC || VerifyDuringGC) && 4255 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4256 verify_after_remark(); 4257 } 4258 4259 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4260 4261 // Change under the freelistLocks. 4262 _collectorState = Sweeping; 4263 // Call isAllClear() under bitMapLock 4264 assert(_modUnionTable.isAllClear(), 4265 "Should be clear by end of the final marking"); 4266 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4267 "Should be clear by end of the final marking"); 4268} 4269 4270void CMSParInitialMarkTask::work(uint worker_id) { 4271 elapsedTimer _timer; 4272 ResourceMark rm; 4273 HandleMark hm; 4274 4275 // ---------- scan from roots -------------- 4276 _timer.start(); 4277 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4278 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4279 4280 // ---------- young gen roots -------------- 4281 { 4282 work_on_young_gen_roots(&par_mri_cl); 4283 _timer.stop(); 4284 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4285 } 4286 4287 // ---------- remaining roots -------------- 4288 _timer.reset(); 4289 _timer.start(); 4290 4291 CLDToOopClosure cld_closure(&par_mri_cl, true); 4292 4293 gch->gen_process_roots(_strong_roots_scope, 4294 GenCollectedHeap::OldGen, 4295 false, // yg was scanned above 4296 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4297 _collector->should_unload_classes(), 4298 &par_mri_cl, 4299 NULL, 4300 &cld_closure); 4301 assert(_collector->should_unload_classes() 4302 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4303 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4304 _timer.stop(); 4305 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4306} 4307 4308// Parallel remark task 4309class CMSParRemarkTask: public CMSParMarkTask { 4310 CompactibleFreeListSpace* _cms_space; 4311 4312 // The per-thread work queues, available here for stealing. 4313 OopTaskQueueSet* _task_queues; 4314 ParallelTaskTerminator _term; 4315 StrongRootsScope* _strong_roots_scope; 4316 4317 public: 4318 // A value of 0 passed to n_workers will cause the number of 4319 // workers to be taken from the active workers in the work gang. 4320 CMSParRemarkTask(CMSCollector* collector, 4321 CompactibleFreeListSpace* cms_space, 4322 uint n_workers, WorkGang* workers, 4323 OopTaskQueueSet* task_queues, 4324 StrongRootsScope* strong_roots_scope): 4325 CMSParMarkTask("Rescan roots and grey objects in parallel", 4326 collector, n_workers), 4327 _cms_space(cms_space), 4328 _task_queues(task_queues), 4329 _term(n_workers, task_queues), 4330 _strong_roots_scope(strong_roots_scope) { } 4331 4332 OopTaskQueueSet* task_queues() { return _task_queues; } 4333 4334 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4335 4336 ParallelTaskTerminator* terminator() { return &_term; } 4337 uint n_workers() { return _n_workers; } 4338 4339 void work(uint worker_id); 4340 4341 private: 4342 // ... of dirty cards in old space 4343 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4344 ParMarkRefsIntoAndScanClosure* cl); 4345 4346 // ... work stealing for the above 4347 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4348}; 4349 4350class RemarkKlassClosure : public KlassClosure { 4351 KlassToOopClosure _cm_klass_closure; 4352 public: 4353 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4354 void do_klass(Klass* k) { 4355 // Check if we have modified any oops in the Klass during the concurrent marking. 4356 if (k->has_accumulated_modified_oops()) { 4357 k->clear_accumulated_modified_oops(); 4358 4359 // We could have transfered the current modified marks to the accumulated marks, 4360 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4361 } else if (k->has_modified_oops()) { 4362 // Don't clear anything, this info is needed by the next young collection. 4363 } else { 4364 // No modified oops in the Klass. 4365 return; 4366 } 4367 4368 // The klass has modified fields, need to scan the klass. 4369 _cm_klass_closure.do_klass(k); 4370 } 4371}; 4372 4373void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4374 ParNewGeneration* young_gen = _collector->_young_gen; 4375 ContiguousSpace* eden_space = young_gen->eden(); 4376 ContiguousSpace* from_space = young_gen->from(); 4377 ContiguousSpace* to_space = young_gen->to(); 4378 4379 HeapWord** eca = _collector->_eden_chunk_array; 4380 size_t ect = _collector->_eden_chunk_index; 4381 HeapWord** sca = _collector->_survivor_chunk_array; 4382 size_t sct = _collector->_survivor_chunk_index; 4383 4384 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4385 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4386 4387 do_young_space_rescan(cl, to_space, NULL, 0); 4388 do_young_space_rescan(cl, from_space, sca, sct); 4389 do_young_space_rescan(cl, eden_space, eca, ect); 4390} 4391 4392// work_queue(i) is passed to the closure 4393// ParMarkRefsIntoAndScanClosure. The "i" parameter 4394// also is passed to do_dirty_card_rescan_tasks() and to 4395// do_work_steal() to select the i-th task_queue. 4396 4397void CMSParRemarkTask::work(uint worker_id) { 4398 elapsedTimer _timer; 4399 ResourceMark rm; 4400 HandleMark hm; 4401 4402 // ---------- rescan from roots -------------- 4403 _timer.start(); 4404 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4405 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4406 _collector->_span, _collector->ref_processor(), 4407 &(_collector->_markBitMap), 4408 work_queue(worker_id)); 4409 4410 // Rescan young gen roots first since these are likely 4411 // coarsely partitioned and may, on that account, constitute 4412 // the critical path; thus, it's best to start off that 4413 // work first. 4414 // ---------- young gen roots -------------- 4415 { 4416 work_on_young_gen_roots(&par_mrias_cl); 4417 _timer.stop(); 4418 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4419 } 4420 4421 // ---------- remaining roots -------------- 4422 _timer.reset(); 4423 _timer.start(); 4424 gch->gen_process_roots(_strong_roots_scope, 4425 GenCollectedHeap::OldGen, 4426 false, // yg was scanned above 4427 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4428 _collector->should_unload_classes(), 4429 &par_mrias_cl, 4430 NULL, 4431 NULL); // The dirty klasses will be handled below 4432 4433 assert(_collector->should_unload_classes() 4434 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4435 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4436 _timer.stop(); 4437 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4438 4439 // ---------- unhandled CLD scanning ---------- 4440 if (worker_id == 0) { // Single threaded at the moment. 4441 _timer.reset(); 4442 _timer.start(); 4443 4444 // Scan all new class loader data objects and new dependencies that were 4445 // introduced during concurrent marking. 4446 ResourceMark rm; 4447 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4448 for (int i = 0; i < array->length(); i++) { 4449 par_mrias_cl.do_cld_nv(array->at(i)); 4450 } 4451 4452 // We don't need to keep track of new CLDs anymore. 4453 ClassLoaderDataGraph::remember_new_clds(false); 4454 4455 _timer.stop(); 4456 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4457 } 4458 4459 // ---------- dirty klass scanning ---------- 4460 if (worker_id == 0) { // Single threaded at the moment. 4461 _timer.reset(); 4462 _timer.start(); 4463 4464 // Scan all classes that was dirtied during the concurrent marking phase. 4465 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4466 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4467 4468 _timer.stop(); 4469 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4470 } 4471 4472 // We might have added oops to ClassLoaderData::_handles during the 4473 // concurrent marking phase. These oops point to newly allocated objects 4474 // that are guaranteed to be kept alive. Either by the direct allocation 4475 // code, or when the young collector processes the roots. Hence, 4476 // we don't have to revisit the _handles block during the remark phase. 4477 4478 // ---------- rescan dirty cards ------------ 4479 _timer.reset(); 4480 _timer.start(); 4481 4482 // Do the rescan tasks for each of the two spaces 4483 // (cms_space) in turn. 4484 // "worker_id" is passed to select the task_queue for "worker_id" 4485 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4486 _timer.stop(); 4487 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4488 4489 // ---------- steal work from other threads ... 4490 // ---------- ... and drain overflow list. 4491 _timer.reset(); 4492 _timer.start(); 4493 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4494 _timer.stop(); 4495 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4496} 4497 4498void 4499CMSParMarkTask::do_young_space_rescan( 4500 OopsInGenClosure* cl, ContiguousSpace* space, 4501 HeapWord** chunk_array, size_t chunk_top) { 4502 // Until all tasks completed: 4503 // . claim an unclaimed task 4504 // . compute region boundaries corresponding to task claimed 4505 // using chunk_array 4506 // . par_oop_iterate(cl) over that region 4507 4508 ResourceMark rm; 4509 HandleMark hm; 4510 4511 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4512 4513 uint nth_task = 0; 4514 uint n_tasks = pst->n_tasks(); 4515 4516 if (n_tasks > 0) { 4517 assert(pst->valid(), "Uninitialized use?"); 4518 HeapWord *start, *end; 4519 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4520 // We claimed task # nth_task; compute its boundaries. 4521 if (chunk_top == 0) { // no samples were taken 4522 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4523 start = space->bottom(); 4524 end = space->top(); 4525 } else if (nth_task == 0) { 4526 start = space->bottom(); 4527 end = chunk_array[nth_task]; 4528 } else if (nth_task < (uint)chunk_top) { 4529 assert(nth_task >= 1, "Control point invariant"); 4530 start = chunk_array[nth_task - 1]; 4531 end = chunk_array[nth_task]; 4532 } else { 4533 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4534 start = chunk_array[chunk_top - 1]; 4535 end = space->top(); 4536 } 4537 MemRegion mr(start, end); 4538 // Verify that mr is in space 4539 assert(mr.is_empty() || space->used_region().contains(mr), 4540 "Should be in space"); 4541 // Verify that "start" is an object boundary 4542 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4543 "Should be an oop"); 4544 space->par_oop_iterate(mr, cl); 4545 } 4546 pst->all_tasks_completed(); 4547 } 4548} 4549 4550void 4551CMSParRemarkTask::do_dirty_card_rescan_tasks( 4552 CompactibleFreeListSpace* sp, int i, 4553 ParMarkRefsIntoAndScanClosure* cl) { 4554 // Until all tasks completed: 4555 // . claim an unclaimed task 4556 // . compute region boundaries corresponding to task claimed 4557 // . transfer dirty bits ct->mut for that region 4558 // . apply rescanclosure to dirty mut bits for that region 4559 4560 ResourceMark rm; 4561 HandleMark hm; 4562 4563 OopTaskQueue* work_q = work_queue(i); 4564 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4565 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4566 // CAUTION: This closure has state that persists across calls to 4567 // the work method dirty_range_iterate_clear() in that it has 4568 // embedded in it a (subtype of) UpwardsObjectClosure. The 4569 // use of that state in the embedded UpwardsObjectClosure instance 4570 // assumes that the cards are always iterated (even if in parallel 4571 // by several threads) in monotonically increasing order per each 4572 // thread. This is true of the implementation below which picks 4573 // card ranges (chunks) in monotonically increasing order globally 4574 // and, a-fortiori, in monotonically increasing order per thread 4575 // (the latter order being a subsequence of the former). 4576 // If the work code below is ever reorganized into a more chaotic 4577 // work-partitioning form than the current "sequential tasks" 4578 // paradigm, the use of that persistent state will have to be 4579 // revisited and modified appropriately. See also related 4580 // bug 4756801 work on which should examine this code to make 4581 // sure that the changes there do not run counter to the 4582 // assumptions made here and necessary for correctness and 4583 // efficiency. Note also that this code might yield inefficient 4584 // behavior in the case of very large objects that span one or 4585 // more work chunks. Such objects would potentially be scanned 4586 // several times redundantly. Work on 4756801 should try and 4587 // address that performance anomaly if at all possible. XXX 4588 MemRegion full_span = _collector->_span; 4589 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4590 MarkFromDirtyCardsClosure 4591 greyRescanClosure(_collector, full_span, // entire span of interest 4592 sp, bm, work_q, cl); 4593 4594 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4595 assert(pst->valid(), "Uninitialized use?"); 4596 uint nth_task = 0; 4597 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4598 MemRegion span = sp->used_region(); 4599 HeapWord* start_addr = span.start(); 4600 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4601 alignment); 4602 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4603 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4604 start_addr, "Check alignment"); 4605 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4606 chunk_size, "Check alignment"); 4607 4608 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4609 // Having claimed the nth_task, compute corresponding mem-region, 4610 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4611 // The alignment restriction ensures that we do not need any 4612 // synchronization with other gang-workers while setting or 4613 // clearing bits in thus chunk of the MUT. 4614 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4615 start_addr + (nth_task+1)*chunk_size); 4616 // The last chunk's end might be way beyond end of the 4617 // used region. In that case pull back appropriately. 4618 if (this_span.end() > end_addr) { 4619 this_span.set_end(end_addr); 4620 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4621 } 4622 // Iterate over the dirty cards covering this chunk, marking them 4623 // precleaned, and setting the corresponding bits in the mod union 4624 // table. Since we have been careful to partition at Card and MUT-word 4625 // boundaries no synchronization is needed between parallel threads. 4626 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4627 &modUnionClosure); 4628 4629 // Having transferred these marks into the modUnionTable, 4630 // rescan the marked objects on the dirty cards in the modUnionTable. 4631 // Even if this is at a synchronous collection, the initial marking 4632 // may have been done during an asynchronous collection so there 4633 // may be dirty bits in the mod-union table. 4634 _collector->_modUnionTable.dirty_range_iterate_clear( 4635 this_span, &greyRescanClosure); 4636 _collector->_modUnionTable.verifyNoOneBitsInRange( 4637 this_span.start(), 4638 this_span.end()); 4639 } 4640 pst->all_tasks_completed(); // declare that i am done 4641} 4642 4643// . see if we can share work_queues with ParNew? XXX 4644void 4645CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4646 int* seed) { 4647 OopTaskQueue* work_q = work_queue(i); 4648 NOT_PRODUCT(int num_steals = 0;) 4649 oop obj_to_scan; 4650 CMSBitMap* bm = &(_collector->_markBitMap); 4651 4652 while (true) { 4653 // Completely finish any left over work from (an) earlier round(s) 4654 cl->trim_queue(0); 4655 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4656 (size_t)ParGCDesiredObjsFromOverflowList); 4657 // Now check if there's any work in the overflow list 4658 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4659 // only affects the number of attempts made to get work from the 4660 // overflow list and does not affect the number of workers. Just 4661 // pass ParallelGCThreads so this behavior is unchanged. 4662 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4663 work_q, 4664 ParallelGCThreads)) { 4665 // found something in global overflow list; 4666 // not yet ready to go stealing work from others. 4667 // We'd like to assert(work_q->size() != 0, ...) 4668 // because we just took work from the overflow list, 4669 // but of course we can't since all of that could have 4670 // been already stolen from us. 4671 // "He giveth and He taketh away." 4672 continue; 4673 } 4674 // Verify that we have no work before we resort to stealing 4675 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4676 // Try to steal from other queues that have work 4677 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4678 NOT_PRODUCT(num_steals++;) 4679 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4680 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4681 // Do scanning work 4682 obj_to_scan->oop_iterate(cl); 4683 // Loop around, finish this work, and try to steal some more 4684 } else if (terminator()->offer_termination()) { 4685 break; // nirvana from the infinite cycle 4686 } 4687 } 4688 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4689 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4690 "Else our work is not yet done"); 4691} 4692 4693// Record object boundaries in _eden_chunk_array by sampling the eden 4694// top in the slow-path eden object allocation code path and record 4695// the boundaries, if CMSEdenChunksRecordAlways is true. If 4696// CMSEdenChunksRecordAlways is false, we use the other asynchronous 4697// sampling in sample_eden() that activates during the part of the 4698// preclean phase. 4699void CMSCollector::sample_eden_chunk() { 4700 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4701 if (_eden_chunk_lock->try_lock()) { 4702 // Record a sample. This is the critical section. The contents 4703 // of the _eden_chunk_array have to be non-decreasing in the 4704 // address order. 4705 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4706 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4707 "Unexpected state of Eden"); 4708 if (_eden_chunk_index == 0 || 4709 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4710 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4711 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4712 _eden_chunk_index++; // commit sample 4713 } 4714 _eden_chunk_lock->unlock(); 4715 } 4716 } 4717} 4718 4719// Return a thread-local PLAB recording array, as appropriate. 4720void* CMSCollector::get_data_recorder(int thr_num) { 4721 if (_survivor_plab_array != NULL && 4722 (CMSPLABRecordAlways || 4723 (_collectorState > Marking && _collectorState < FinalMarking))) { 4724 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4725 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4726 ca->reset(); // clear it so that fresh data is recorded 4727 return (void*) ca; 4728 } else { 4729 return NULL; 4730 } 4731} 4732 4733// Reset all the thread-local PLAB recording arrays 4734void CMSCollector::reset_survivor_plab_arrays() { 4735 for (uint i = 0; i < ParallelGCThreads; i++) { 4736 _survivor_plab_array[i].reset(); 4737 } 4738} 4739 4740// Merge the per-thread plab arrays into the global survivor chunk 4741// array which will provide the partitioning of the survivor space 4742// for CMS initial scan and rescan. 4743void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4744 int no_of_gc_threads) { 4745 assert(_survivor_plab_array != NULL, "Error"); 4746 assert(_survivor_chunk_array != NULL, "Error"); 4747 assert(_collectorState == FinalMarking || 4748 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4749 for (int j = 0; j < no_of_gc_threads; j++) { 4750 _cursor[j] = 0; 4751 } 4752 HeapWord* top = surv->top(); 4753 size_t i; 4754 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4755 HeapWord* min_val = top; // Higher than any PLAB address 4756 uint min_tid = 0; // position of min_val this round 4757 for (int j = 0; j < no_of_gc_threads; j++) { 4758 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4759 if (_cursor[j] == cur_sca->end()) { 4760 continue; 4761 } 4762 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4763 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4764 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4765 if (cur_val < min_val) { 4766 min_tid = j; 4767 min_val = cur_val; 4768 } else { 4769 assert(cur_val < top, "All recorded addresses should be less"); 4770 } 4771 } 4772 // At this point min_val and min_tid are respectively 4773 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4774 // and the thread (j) that witnesses that address. 4775 // We record this address in the _survivor_chunk_array[i] 4776 // and increment _cursor[min_tid] prior to the next round i. 4777 if (min_val == top) { 4778 break; 4779 } 4780 _survivor_chunk_array[i] = min_val; 4781 _cursor[min_tid]++; 4782 } 4783 // We are all done; record the size of the _survivor_chunk_array 4784 _survivor_chunk_index = i; // exclusive: [0, i) 4785 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4786 // Verify that we used up all the recorded entries 4787 #ifdef ASSERT 4788 size_t total = 0; 4789 for (int j = 0; j < no_of_gc_threads; j++) { 4790 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4791 total += _cursor[j]; 4792 } 4793 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4794 // Check that the merged array is in sorted order 4795 if (total > 0) { 4796 for (size_t i = 0; i < total - 1; i++) { 4797 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4798 i, p2i(_survivor_chunk_array[i])); 4799 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4800 "Not sorted"); 4801 } 4802 } 4803 #endif // ASSERT 4804} 4805 4806// Set up the space's par_seq_tasks structure for work claiming 4807// for parallel initial scan and rescan of young gen. 4808// See ParRescanTask where this is currently used. 4809void 4810CMSCollector:: 4811initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4812 assert(n_threads > 0, "Unexpected n_threads argument"); 4813 4814 // Eden space 4815 if (!_young_gen->eden()->is_empty()) { 4816 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 // Each valid entry in [0, _eden_chunk_index) represents a task. 4819 size_t n_tasks = _eden_chunk_index + 1; 4820 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4821 // Sets the condition for completion of the subtask (how many threads 4822 // need to finish in order to be done). 4823 pst->set_n_threads(n_threads); 4824 pst->set_n_tasks((int)n_tasks); 4825 } 4826 4827 // Merge the survivor plab arrays into _survivor_chunk_array 4828 if (_survivor_plab_array != NULL) { 4829 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4830 } else { 4831 assert(_survivor_chunk_index == 0, "Error"); 4832 } 4833 4834 // To space 4835 { 4836 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4837 assert(!pst->valid(), "Clobbering existing data?"); 4838 // Sets the condition for completion of the subtask (how many threads 4839 // need to finish in order to be done). 4840 pst->set_n_threads(n_threads); 4841 pst->set_n_tasks(1); 4842 assert(pst->valid(), "Error"); 4843 } 4844 4845 // From space 4846 { 4847 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4848 assert(!pst->valid(), "Clobbering existing data?"); 4849 size_t n_tasks = _survivor_chunk_index + 1; 4850 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4851 // Sets the condition for completion of the subtask (how many threads 4852 // need to finish in order to be done). 4853 pst->set_n_threads(n_threads); 4854 pst->set_n_tasks((int)n_tasks); 4855 assert(pst->valid(), "Error"); 4856 } 4857} 4858 4859// Parallel version of remark 4860void CMSCollector::do_remark_parallel() { 4861 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4862 WorkGang* workers = gch->workers(); 4863 assert(workers != NULL, "Need parallel worker threads."); 4864 // Choose to use the number of GC workers most recently set 4865 // into "active_workers". 4866 uint n_workers = workers->active_workers(); 4867 4868 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4869 4870 StrongRootsScope srs(n_workers); 4871 4872 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4873 4874 // We won't be iterating over the cards in the card table updating 4875 // the younger_gen cards, so we shouldn't call the following else 4876 // the verification code as well as subsequent younger_refs_iterate 4877 // code would get confused. XXX 4878 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4879 4880 // The young gen rescan work will not be done as part of 4881 // process_roots (which currently doesn't know how to 4882 // parallelize such a scan), but rather will be broken up into 4883 // a set of parallel tasks (via the sampling that the [abortable] 4884 // preclean phase did of eden, plus the [two] tasks of 4885 // scanning the [two] survivor spaces. Further fine-grain 4886 // parallelization of the scanning of the survivor spaces 4887 // themselves, and of precleaning of the young gen itself 4888 // is deferred to the future. 4889 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4890 4891 // The dirty card rescan work is broken up into a "sequence" 4892 // of parallel tasks (per constituent space) that are dynamically 4893 // claimed by the parallel threads. 4894 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4895 4896 // It turns out that even when we're using 1 thread, doing the work in a 4897 // separate thread causes wide variance in run times. We can't help this 4898 // in the multi-threaded case, but we special-case n=1 here to get 4899 // repeatable measurements of the 1-thread overhead of the parallel code. 4900 if (n_workers > 1) { 4901 // Make refs discovery MT-safe, if it isn't already: it may not 4902 // necessarily be so, since it's possible that we are doing 4903 // ST marking. 4904 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4905 workers->run_task(&tsk); 4906 } else { 4907 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4908 tsk.work(0); 4909 } 4910 4911 // restore, single-threaded for now, any preserved marks 4912 // as a result of work_q overflow 4913 restore_preserved_marks_if_any(); 4914} 4915 4916// Non-parallel version of remark 4917void CMSCollector::do_remark_non_parallel() { 4918 ResourceMark rm; 4919 HandleMark hm; 4920 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4921 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4922 4923 MarkRefsIntoAndScanClosure 4924 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4925 &_markStack, this, 4926 false /* should_yield */, false /* not precleaning */); 4927 MarkFromDirtyCardsClosure 4928 markFromDirtyCardsClosure(this, _span, 4929 NULL, // space is set further below 4930 &_markBitMap, &_markStack, &mrias_cl); 4931 { 4932 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4933 // Iterate over the dirty cards, setting the corresponding bits in the 4934 // mod union table. 4935 { 4936 ModUnionClosure modUnionClosure(&_modUnionTable); 4937 _ct->ct_bs()->dirty_card_iterate( 4938 _cmsGen->used_region(), 4939 &modUnionClosure); 4940 } 4941 // Having transferred these marks into the modUnionTable, we just need 4942 // to rescan the marked objects on the dirty cards in the modUnionTable. 4943 // The initial marking may have been done during an asynchronous 4944 // collection so there may be dirty bits in the mod-union table. 4945 const int alignment = 4946 CardTableModRefBS::card_size * BitsPerWord; 4947 { 4948 // ... First handle dirty cards in CMS gen 4949 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4950 MemRegion ur = _cmsGen->used_region(); 4951 HeapWord* lb = ur.start(); 4952 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4953 MemRegion cms_span(lb, ub); 4954 _modUnionTable.dirty_range_iterate_clear(cms_span, 4955 &markFromDirtyCardsClosure); 4956 verify_work_stacks_empty(); 4957 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4958 } 4959 } 4960 if (VerifyDuringGC && 4961 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4962 HandleMark hm; // Discard invalid handles created during verification 4963 Universe::verify(); 4964 } 4965 { 4966 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4967 4968 verify_work_stacks_empty(); 4969 4970 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4971 StrongRootsScope srs(1); 4972 4973 gch->gen_process_roots(&srs, 4974 GenCollectedHeap::OldGen, 4975 true, // young gen as roots 4976 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4977 should_unload_classes(), 4978 &mrias_cl, 4979 NULL, 4980 NULL); // The dirty klasses will be handled below 4981 4982 assert(should_unload_classes() 4983 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4984 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4985 } 4986 4987 { 4988 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4989 4990 verify_work_stacks_empty(); 4991 4992 // Scan all class loader data objects that might have been introduced 4993 // during concurrent marking. 4994 ResourceMark rm; 4995 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4996 for (int i = 0; i < array->length(); i++) { 4997 mrias_cl.do_cld_nv(array->at(i)); 4998 } 4999 5000 // We don't need to keep track of new CLDs anymore. 5001 ClassLoaderDataGraph::remember_new_clds(false); 5002 5003 verify_work_stacks_empty(); 5004 } 5005 5006 { 5007 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm); 5008 5009 verify_work_stacks_empty(); 5010 5011 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5012 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5013 5014 verify_work_stacks_empty(); 5015 } 5016 5017 // We might have added oops to ClassLoaderData::_handles during the 5018 // concurrent marking phase. These oops point to newly allocated objects 5019 // that are guaranteed to be kept alive. Either by the direct allocation 5020 // code, or when the young collector processes the roots. Hence, 5021 // we don't have to revisit the _handles block during the remark phase. 5022 5023 verify_work_stacks_empty(); 5024 // Restore evacuated mark words, if any, used for overflow list links 5025 restore_preserved_marks_if_any(); 5026 5027 verify_overflow_empty(); 5028} 5029 5030//////////////////////////////////////////////////////// 5031// Parallel Reference Processing Task Proxy Class 5032//////////////////////////////////////////////////////// 5033class AbstractGangTaskWOopQueues : public AbstractGangTask { 5034 OopTaskQueueSet* _queues; 5035 ParallelTaskTerminator _terminator; 5036 public: 5037 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5038 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5039 ParallelTaskTerminator* terminator() { return &_terminator; } 5040 OopTaskQueueSet* queues() { return _queues; } 5041}; 5042 5043class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5044 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5045 CMSCollector* _collector; 5046 CMSBitMap* _mark_bit_map; 5047 const MemRegion _span; 5048 ProcessTask& _task; 5049 5050public: 5051 CMSRefProcTaskProxy(ProcessTask& task, 5052 CMSCollector* collector, 5053 const MemRegion& span, 5054 CMSBitMap* mark_bit_map, 5055 AbstractWorkGang* workers, 5056 OopTaskQueueSet* task_queues): 5057 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5058 task_queues, 5059 workers->active_workers()), 5060 _task(task), 5061 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5062 { 5063 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5064 "Inconsistency in _span"); 5065 } 5066 5067 OopTaskQueueSet* task_queues() { return queues(); } 5068 5069 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5070 5071 void do_work_steal(int i, 5072 CMSParDrainMarkingStackClosure* drain, 5073 CMSParKeepAliveClosure* keep_alive, 5074 int* seed); 5075 5076 virtual void work(uint worker_id); 5077}; 5078 5079void CMSRefProcTaskProxy::work(uint worker_id) { 5080 ResourceMark rm; 5081 HandleMark hm; 5082 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5083 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5084 _mark_bit_map, 5085 work_queue(worker_id)); 5086 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5087 _mark_bit_map, 5088 work_queue(worker_id)); 5089 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5090 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5091 if (_task.marks_oops_alive()) { 5092 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5093 _collector->hash_seed(worker_id)); 5094 } 5095 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5096 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5097} 5098 5099class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5100 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5101 EnqueueTask& _task; 5102 5103public: 5104 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5105 : AbstractGangTask("Enqueue reference objects in parallel"), 5106 _task(task) 5107 { } 5108 5109 virtual void work(uint worker_id) 5110 { 5111 _task.work(worker_id); 5112 } 5113}; 5114 5115CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5116 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5117 _span(span), 5118 _bit_map(bit_map), 5119 _work_queue(work_queue), 5120 _mark_and_push(collector, span, bit_map, work_queue), 5121 _low_water_mark(MIN2((work_queue->max_elems()/4), 5122 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5123{ } 5124 5125// . see if we can share work_queues with ParNew? XXX 5126void CMSRefProcTaskProxy::do_work_steal(int i, 5127 CMSParDrainMarkingStackClosure* drain, 5128 CMSParKeepAliveClosure* keep_alive, 5129 int* seed) { 5130 OopTaskQueue* work_q = work_queue(i); 5131 NOT_PRODUCT(int num_steals = 0;) 5132 oop obj_to_scan; 5133 5134 while (true) { 5135 // Completely finish any left over work from (an) earlier round(s) 5136 drain->trim_queue(0); 5137 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5138 (size_t)ParGCDesiredObjsFromOverflowList); 5139 // Now check if there's any work in the overflow list 5140 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5141 // only affects the number of attempts made to get work from the 5142 // overflow list and does not affect the number of workers. Just 5143 // pass ParallelGCThreads so this behavior is unchanged. 5144 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5145 work_q, 5146 ParallelGCThreads)) { 5147 // Found something in global overflow list; 5148 // not yet ready to go stealing work from others. 5149 // We'd like to assert(work_q->size() != 0, ...) 5150 // because we just took work from the overflow list, 5151 // but of course we can't, since all of that might have 5152 // been already stolen from us. 5153 continue; 5154 } 5155 // Verify that we have no work before we resort to stealing 5156 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5157 // Try to steal from other queues that have work 5158 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5159 NOT_PRODUCT(num_steals++;) 5160 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5161 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5162 // Do scanning work 5163 obj_to_scan->oop_iterate(keep_alive); 5164 // Loop around, finish this work, and try to steal some more 5165 } else if (terminator()->offer_termination()) { 5166 break; // nirvana from the infinite cycle 5167 } 5168 } 5169 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5170} 5171 5172void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5173{ 5174 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5175 WorkGang* workers = gch->workers(); 5176 assert(workers != NULL, "Need parallel worker threads."); 5177 CMSRefProcTaskProxy rp_task(task, &_collector, 5178 _collector.ref_processor()->span(), 5179 _collector.markBitMap(), 5180 workers, _collector.task_queues()); 5181 workers->run_task(&rp_task); 5182} 5183 5184void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5185{ 5186 5187 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5188 WorkGang* workers = gch->workers(); 5189 assert(workers != NULL, "Need parallel worker threads."); 5190 CMSRefEnqueueTaskProxy enq_task(task); 5191 workers->run_task(&enq_task); 5192} 5193 5194void CMSCollector::refProcessingWork() { 5195 ResourceMark rm; 5196 HandleMark hm; 5197 5198 ReferenceProcessor* rp = ref_processor(); 5199 assert(rp->span().equals(_span), "Spans should be equal"); 5200 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5201 // Process weak references. 5202 rp->setup_policy(false); 5203 verify_work_stacks_empty(); 5204 5205 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5206 &_markStack, false /* !preclean */); 5207 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5208 _span, &_markBitMap, &_markStack, 5209 &cmsKeepAliveClosure, false /* !preclean */); 5210 { 5211 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5212 5213 ReferenceProcessorStats stats; 5214 if (rp->processing_is_mt()) { 5215 // Set the degree of MT here. If the discovery is done MT, there 5216 // may have been a different number of threads doing the discovery 5217 // and a different number of discovered lists may have Ref objects. 5218 // That is OK as long as the Reference lists are balanced (see 5219 // balance_all_queues() and balance_queues()). 5220 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5221 uint active_workers = ParallelGCThreads; 5222 WorkGang* workers = gch->workers(); 5223 if (workers != NULL) { 5224 active_workers = workers->active_workers(); 5225 // The expectation is that active_workers will have already 5226 // been set to a reasonable value. If it has not been set, 5227 // investigate. 5228 assert(active_workers > 0, "Should have been set during scavenge"); 5229 } 5230 rp->set_active_mt_degree(active_workers); 5231 CMSRefProcTaskExecutor task_executor(*this); 5232 stats = rp->process_discovered_references(&_is_alive_closure, 5233 &cmsKeepAliveClosure, 5234 &cmsDrainMarkingStackClosure, 5235 &task_executor, 5236 _gc_timer_cm); 5237 } else { 5238 stats = rp->process_discovered_references(&_is_alive_closure, 5239 &cmsKeepAliveClosure, 5240 &cmsDrainMarkingStackClosure, 5241 NULL, 5242 _gc_timer_cm); 5243 } 5244 _gc_tracer_cm->report_gc_reference_stats(stats); 5245 5246 } 5247 5248 // This is the point where the entire marking should have completed. 5249 verify_work_stacks_empty(); 5250 5251 if (should_unload_classes()) { 5252 { 5253 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5254 5255 // Unload classes and purge the SystemDictionary. 5256 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5257 5258 // Unload nmethods. 5259 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5260 5261 // Prune dead klasses from subklass/sibling/implementor lists. 5262 Klass::clean_weak_klass_links(&_is_alive_closure); 5263 } 5264 5265 { 5266 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5267 // Clean up unreferenced symbols in symbol table. 5268 SymbolTable::unlink(); 5269 } 5270 5271 { 5272 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5273 // Delete entries for dead interned strings. 5274 StringTable::unlink(&_is_alive_closure); 5275 } 5276 } 5277 5278 5279 // Restore any preserved marks as a result of mark stack or 5280 // work queue overflow 5281 restore_preserved_marks_if_any(); // done single-threaded for now 5282 5283 rp->set_enqueuing_is_done(true); 5284 if (rp->processing_is_mt()) { 5285 rp->balance_all_queues(); 5286 CMSRefProcTaskExecutor task_executor(*this); 5287 rp->enqueue_discovered_references(&task_executor); 5288 } else { 5289 rp->enqueue_discovered_references(NULL); 5290 } 5291 rp->verify_no_references_recorded(); 5292 assert(!rp->discovery_enabled(), "should have been disabled"); 5293} 5294 5295#ifndef PRODUCT 5296void CMSCollector::check_correct_thread_executing() { 5297 Thread* t = Thread::current(); 5298 // Only the VM thread or the CMS thread should be here. 5299 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5300 "Unexpected thread type"); 5301 // If this is the vm thread, the foreground process 5302 // should not be waiting. Note that _foregroundGCIsActive is 5303 // true while the foreground collector is waiting. 5304 if (_foregroundGCShouldWait) { 5305 // We cannot be the VM thread 5306 assert(t->is_ConcurrentGC_thread(), 5307 "Should be CMS thread"); 5308 } else { 5309 // We can be the CMS thread only if we are in a stop-world 5310 // phase of CMS collection. 5311 if (t->is_ConcurrentGC_thread()) { 5312 assert(_collectorState == InitialMarking || 5313 _collectorState == FinalMarking, 5314 "Should be a stop-world phase"); 5315 // The CMS thread should be holding the CMS_token. 5316 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5317 "Potential interference with concurrently " 5318 "executing VM thread"); 5319 } 5320 } 5321} 5322#endif 5323 5324void CMSCollector::sweep() { 5325 assert(_collectorState == Sweeping, "just checking"); 5326 check_correct_thread_executing(); 5327 verify_work_stacks_empty(); 5328 verify_overflow_empty(); 5329 increment_sweep_count(); 5330 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5331 5332 _inter_sweep_timer.stop(); 5333 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5334 5335 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5336 _intra_sweep_timer.reset(); 5337 _intra_sweep_timer.start(); 5338 { 5339 GCTraceCPUTime tcpu; 5340 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5341 // First sweep the old gen 5342 { 5343 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5344 bitMapLock()); 5345 sweepWork(_cmsGen); 5346 } 5347 5348 // Update Universe::_heap_*_at_gc figures. 5349 // We need all the free list locks to make the abstract state 5350 // transition from Sweeping to Resetting. See detailed note 5351 // further below. 5352 { 5353 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5354 // Update heap occupancy information which is used as 5355 // input to soft ref clearing policy at the next gc. 5356 Universe::update_heap_info_at_gc(); 5357 _collectorState = Resizing; 5358 } 5359 } 5360 verify_work_stacks_empty(); 5361 verify_overflow_empty(); 5362 5363 if (should_unload_classes()) { 5364 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5365 // requires that the virtual spaces are stable and not deleted. 5366 ClassLoaderDataGraph::set_should_purge(true); 5367 } 5368 5369 _intra_sweep_timer.stop(); 5370 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5371 5372 _inter_sweep_timer.reset(); 5373 _inter_sweep_timer.start(); 5374 5375 // We need to use a monotonically non-decreasing time in ms 5376 // or we will see time-warp warnings and os::javaTimeMillis() 5377 // does not guarantee monotonicity. 5378 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5379 update_time_of_last_gc(now); 5380 5381 // NOTE on abstract state transitions: 5382 // Mutators allocate-live and/or mark the mod-union table dirty 5383 // based on the state of the collection. The former is done in 5384 // the interval [Marking, Sweeping] and the latter in the interval 5385 // [Marking, Sweeping). Thus the transitions into the Marking state 5386 // and out of the Sweeping state must be synchronously visible 5387 // globally to the mutators. 5388 // The transition into the Marking state happens with the world 5389 // stopped so the mutators will globally see it. Sweeping is 5390 // done asynchronously by the background collector so the transition 5391 // from the Sweeping state to the Resizing state must be done 5392 // under the freelistLock (as is the check for whether to 5393 // allocate-live and whether to dirty the mod-union table). 5394 assert(_collectorState == Resizing, "Change of collector state to" 5395 " Resizing must be done under the freelistLocks (plural)"); 5396 5397 // Now that sweeping has been completed, we clear 5398 // the incremental_collection_failed flag, 5399 // thus inviting a younger gen collection to promote into 5400 // this generation. If such a promotion may still fail, 5401 // the flag will be set again when a young collection is 5402 // attempted. 5403 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5404 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5405 gch->update_full_collections_completed(_collection_count_start); 5406} 5407 5408// FIX ME!!! Looks like this belongs in CFLSpace, with 5409// CMSGen merely delegating to it. 5410void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5411 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5412 HeapWord* minAddr = _cmsSpace->bottom(); 5413 HeapWord* largestAddr = 5414 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5415 if (largestAddr == NULL) { 5416 // The dictionary appears to be empty. In this case 5417 // try to coalesce at the end of the heap. 5418 largestAddr = _cmsSpace->end(); 5419 } 5420 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5421 size_t nearLargestOffset = 5422 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5423 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5424 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5425 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5426} 5427 5428bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5429 return addr >= _cmsSpace->nearLargestChunk(); 5430} 5431 5432FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5433 return _cmsSpace->find_chunk_at_end(); 5434} 5435 5436void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5437 bool full) { 5438 // If the young generation has been collected, gather any statistics 5439 // that are of interest at this point. 5440 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5441 if (!full && current_is_young) { 5442 // Gather statistics on the young generation collection. 5443 collector()->stats().record_gc0_end(used()); 5444 } 5445} 5446 5447void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5448 // We iterate over the space(s) underlying this generation, 5449 // checking the mark bit map to see if the bits corresponding 5450 // to specific blocks are marked or not. Blocks that are 5451 // marked are live and are not swept up. All remaining blocks 5452 // are swept up, with coalescing on-the-fly as we sweep up 5453 // contiguous free and/or garbage blocks: 5454 // We need to ensure that the sweeper synchronizes with allocators 5455 // and stop-the-world collectors. In particular, the following 5456 // locks are used: 5457 // . CMS token: if this is held, a stop the world collection cannot occur 5458 // . freelistLock: if this is held no allocation can occur from this 5459 // generation by another thread 5460 // . bitMapLock: if this is held, no other thread can access or update 5461 // 5462 5463 // Note that we need to hold the freelistLock if we use 5464 // block iterate below; else the iterator might go awry if 5465 // a mutator (or promotion) causes block contents to change 5466 // (for instance if the allocator divvies up a block). 5467 // If we hold the free list lock, for all practical purposes 5468 // young generation GC's can't occur (they'll usually need to 5469 // promote), so we might as well prevent all young generation 5470 // GC's while we do a sweeping step. For the same reason, we might 5471 // as well take the bit map lock for the entire duration 5472 5473 // check that we hold the requisite locks 5474 assert(have_cms_token(), "Should hold cms token"); 5475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5476 assert_lock_strong(old_gen->freelistLock()); 5477 assert_lock_strong(bitMapLock()); 5478 5479 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5480 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5481 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5482 _inter_sweep_estimate.padded_average(), 5483 _intra_sweep_estimate.padded_average()); 5484 old_gen->setNearLargestChunk(); 5485 5486 { 5487 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5488 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5489 // We need to free-up/coalesce garbage/blocks from a 5490 // co-terminal free run. This is done in the SweepClosure 5491 // destructor; so, do not remove this scope, else the 5492 // end-of-sweep-census below will be off by a little bit. 5493 } 5494 old_gen->cmsSpace()->sweep_completed(); 5495 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5496 if (should_unload_classes()) { // unloaded classes this cycle, 5497 _concurrent_cycles_since_last_unload = 0; // ... reset count 5498 } else { // did not unload classes, 5499 _concurrent_cycles_since_last_unload++; // ... increment count 5500 } 5501} 5502 5503// Reset CMS data structures (for now just the marking bit map) 5504// preparatory for the next cycle. 5505void CMSCollector::reset_concurrent() { 5506 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5507 5508 // If the state is not "Resetting", the foreground thread 5509 // has done a collection and the resetting. 5510 if (_collectorState != Resetting) { 5511 assert(_collectorState == Idling, "The state should only change" 5512 " because the foreground collector has finished the collection"); 5513 return; 5514 } 5515 5516 { 5517 // Clear the mark bitmap (no grey objects to start with) 5518 // for the next cycle. 5519 GCTraceCPUTime tcpu; 5520 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5521 5522 HeapWord* curAddr = _markBitMap.startWord(); 5523 while (curAddr < _markBitMap.endWord()) { 5524 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5525 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5526 _markBitMap.clear_large_range(chunk); 5527 if (ConcurrentMarkSweepThread::should_yield() && 5528 !foregroundGCIsActive() && 5529 CMSYield) { 5530 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5531 "CMS thread should hold CMS token"); 5532 assert_lock_strong(bitMapLock()); 5533 bitMapLock()->unlock(); 5534 ConcurrentMarkSweepThread::desynchronize(true); 5535 stopTimer(); 5536 incrementYields(); 5537 5538 // See the comment in coordinator_yield() 5539 for (unsigned i = 0; i < CMSYieldSleepCount && 5540 ConcurrentMarkSweepThread::should_yield() && 5541 !CMSCollector::foregroundGCIsActive(); ++i) { 5542 os::sleep(Thread::current(), 1, false); 5543 } 5544 5545 ConcurrentMarkSweepThread::synchronize(true); 5546 bitMapLock()->lock_without_safepoint_check(); 5547 startTimer(); 5548 } 5549 curAddr = chunk.end(); 5550 } 5551 // A successful mostly concurrent collection has been done. 5552 // Because only the full (i.e., concurrent mode failure) collections 5553 // are being measured for gc overhead limits, clean the "near" flag 5554 // and count. 5555 size_policy()->reset_gc_overhead_limit_count(); 5556 _collectorState = Idling; 5557 } 5558 5559 register_gc_end(); 5560} 5561 5562// Same as above but for STW paths 5563void CMSCollector::reset_stw() { 5564 // already have the lock 5565 assert(_collectorState == Resetting, "just checking"); 5566 assert_lock_strong(bitMapLock()); 5567 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5568 _markBitMap.clear_all(); 5569 _collectorState = Idling; 5570 register_gc_end(); 5571} 5572 5573void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5574 GCTraceCPUTime tcpu; 5575 TraceCollectorStats tcs(counters()); 5576 5577 switch (op) { 5578 case CMS_op_checkpointRootsInitial: { 5579 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5580 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5581 checkpointRootsInitial(); 5582 break; 5583 } 5584 case CMS_op_checkpointRootsFinal: { 5585 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5586 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5587 checkpointRootsFinal(); 5588 break; 5589 } 5590 default: 5591 fatal("No such CMS_op"); 5592 } 5593} 5594 5595#ifndef PRODUCT 5596size_t const CMSCollector::skip_header_HeapWords() { 5597 return FreeChunk::header_size(); 5598} 5599 5600// Try and collect here conditions that should hold when 5601// CMS thread is exiting. The idea is that the foreground GC 5602// thread should not be blocked if it wants to terminate 5603// the CMS thread and yet continue to run the VM for a while 5604// after that. 5605void CMSCollector::verify_ok_to_terminate() const { 5606 assert(Thread::current()->is_ConcurrentGC_thread(), 5607 "should be called by CMS thread"); 5608 assert(!_foregroundGCShouldWait, "should be false"); 5609 // We could check here that all the various low-level locks 5610 // are not held by the CMS thread, but that is overkill; see 5611 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5612 // is checked. 5613} 5614#endif 5615 5616size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5617 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5618 "missing Printezis mark?"); 5619 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5620 size_t size = pointer_delta(nextOneAddr + 1, addr); 5621 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5622 "alignment problem"); 5623 assert(size >= 3, "Necessary for Printezis marks to work"); 5624 return size; 5625} 5626 5627// A variant of the above (block_size_using_printezis_bits()) except 5628// that we return 0 if the P-bits are not yet set. 5629size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5630 if (_markBitMap.isMarked(addr + 1)) { 5631 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5632 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5633 size_t size = pointer_delta(nextOneAddr + 1, addr); 5634 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5635 "alignment problem"); 5636 assert(size >= 3, "Necessary for Printezis marks to work"); 5637 return size; 5638 } 5639 return 0; 5640} 5641 5642HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5643 size_t sz = 0; 5644 oop p = (oop)addr; 5645 if (p->klass_or_null() != NULL) { 5646 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5647 } else { 5648 sz = block_size_using_printezis_bits(addr); 5649 } 5650 assert(sz > 0, "size must be nonzero"); 5651 HeapWord* next_block = addr + sz; 5652 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5653 CardTableModRefBS::card_size); 5654 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5655 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5656 "must be different cards"); 5657 return next_card; 5658} 5659 5660 5661// CMS Bit Map Wrapper ///////////////////////////////////////// 5662 5663// Construct a CMS bit map infrastructure, but don't create the 5664// bit vector itself. That is done by a separate call CMSBitMap::allocate() 5665// further below. 5666CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5667 _bm(), 5668 _shifter(shifter), 5669 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5670 Monitor::_safepoint_check_sometimes) : NULL) 5671{ 5672 _bmStartWord = 0; 5673 _bmWordSize = 0; 5674} 5675 5676bool CMSBitMap::allocate(MemRegion mr) { 5677 _bmStartWord = mr.start(); 5678 _bmWordSize = mr.word_size(); 5679 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5680 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5681 if (!brs.is_reserved()) { 5682 log_warning(gc)("CMS bit map allocation failure"); 5683 return false; 5684 } 5685 // For now we'll just commit all of the bit map up front. 5686 // Later on we'll try to be more parsimonious with swap. 5687 if (!_virtual_space.initialize(brs, brs.size())) { 5688 log_warning(gc)("CMS bit map backing store failure"); 5689 return false; 5690 } 5691 assert(_virtual_space.committed_size() == brs.size(), 5692 "didn't reserve backing store for all of CMS bit map?"); 5693 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5694 _bmWordSize, "inconsistency in bit map sizing"); 5695 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5696 5697 // bm.clear(); // can we rely on getting zero'd memory? verify below 5698 assert(isAllClear(), 5699 "Expected zero'd memory from ReservedSpace constructor"); 5700 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5701 "consistency check"); 5702 return true; 5703} 5704 5705void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5706 HeapWord *next_addr, *end_addr, *last_addr; 5707 assert_locked(); 5708 assert(covers(mr), "out-of-range error"); 5709 // XXX assert that start and end are appropriately aligned 5710 for (next_addr = mr.start(), end_addr = mr.end(); 5711 next_addr < end_addr; next_addr = last_addr) { 5712 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5713 last_addr = dirty_region.end(); 5714 if (!dirty_region.is_empty()) { 5715 cl->do_MemRegion(dirty_region); 5716 } else { 5717 assert(last_addr == end_addr, "program logic"); 5718 return; 5719 } 5720 } 5721} 5722 5723void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5724 _bm.print_on_error(st, prefix); 5725} 5726 5727#ifndef PRODUCT 5728void CMSBitMap::assert_locked() const { 5729 CMSLockVerifier::assert_locked(lock()); 5730} 5731 5732bool CMSBitMap::covers(MemRegion mr) const { 5733 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5734 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5735 "size inconsistency"); 5736 return (mr.start() >= _bmStartWord) && 5737 (mr.end() <= endWord()); 5738} 5739 5740bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5741 return (start >= _bmStartWord && (start + size) <= endWord()); 5742} 5743 5744void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5745 // verify that there are no 1 bits in the interval [left, right) 5746 FalseBitMapClosure falseBitMapClosure; 5747 iterate(&falseBitMapClosure, left, right); 5748} 5749 5750void CMSBitMap::region_invariant(MemRegion mr) 5751{ 5752 assert_locked(); 5753 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5754 assert(!mr.is_empty(), "unexpected empty region"); 5755 assert(covers(mr), "mr should be covered by bit map"); 5756 // convert address range into offset range 5757 size_t start_ofs = heapWordToOffset(mr.start()); 5758 // Make sure that end() is appropriately aligned 5759 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5760 (1 << (_shifter+LogHeapWordSize))), 5761 "Misaligned mr.end()"); 5762 size_t end_ofs = heapWordToOffset(mr.end()); 5763 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5764} 5765 5766#endif 5767 5768bool CMSMarkStack::allocate(size_t size) { 5769 // allocate a stack of the requisite depth 5770 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5771 size * sizeof(oop))); 5772 if (!rs.is_reserved()) { 5773 log_warning(gc)("CMSMarkStack allocation failure"); 5774 return false; 5775 } 5776 if (!_virtual_space.initialize(rs, rs.size())) { 5777 log_warning(gc)("CMSMarkStack backing store failure"); 5778 return false; 5779 } 5780 assert(_virtual_space.committed_size() == rs.size(), 5781 "didn't reserve backing store for all of CMS stack?"); 5782 _base = (oop*)(_virtual_space.low()); 5783 _index = 0; 5784 _capacity = size; 5785 NOT_PRODUCT(_max_depth = 0); 5786 return true; 5787} 5788 5789// XXX FIX ME !!! In the MT case we come in here holding a 5790// leaf lock. For printing we need to take a further lock 5791// which has lower rank. We need to recalibrate the two 5792// lock-ranks involved in order to be able to print the 5793// messages below. (Or defer the printing to the caller. 5794// For now we take the expedient path of just disabling the 5795// messages for the problematic case.) 5796void CMSMarkStack::expand() { 5797 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5798 if (_capacity == MarkStackSizeMax) { 5799 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5800 // We print a warning message only once per CMS cycle. 5801 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5802 } 5803 return; 5804 } 5805 // Double capacity if possible 5806 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5807 // Do not give up existing stack until we have managed to 5808 // get the double capacity that we desired. 5809 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5810 new_capacity * sizeof(oop))); 5811 if (rs.is_reserved()) { 5812 // Release the backing store associated with old stack 5813 _virtual_space.release(); 5814 // Reinitialize virtual space for new stack 5815 if (!_virtual_space.initialize(rs, rs.size())) { 5816 fatal("Not enough swap for expanded marking stack"); 5817 } 5818 _base = (oop*)(_virtual_space.low()); 5819 _index = 0; 5820 _capacity = new_capacity; 5821 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5822 // Failed to double capacity, continue; 5823 // we print a detail message only once per CMS cycle. 5824 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5825 _capacity / K, new_capacity / K); 5826 } 5827} 5828 5829 5830// Closures 5831// XXX: there seems to be a lot of code duplication here; 5832// should refactor and consolidate common code. 5833 5834// This closure is used to mark refs into the CMS generation in 5835// the CMS bit map. Called at the first checkpoint. This closure 5836// assumes that we do not need to re-mark dirty cards; if the CMS 5837// generation on which this is used is not an oldest 5838// generation then this will lose younger_gen cards! 5839 5840MarkRefsIntoClosure::MarkRefsIntoClosure( 5841 MemRegion span, CMSBitMap* bitMap): 5842 _span(span), 5843 _bitMap(bitMap) 5844{ 5845 assert(ref_processor() == NULL, "deliberately left NULL"); 5846 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5847} 5848 5849void MarkRefsIntoClosure::do_oop(oop obj) { 5850 // if p points into _span, then mark corresponding bit in _markBitMap 5851 assert(obj->is_oop(), "expected an oop"); 5852 HeapWord* addr = (HeapWord*)obj; 5853 if (_span.contains(addr)) { 5854 // this should be made more efficient 5855 _bitMap->mark(addr); 5856 } 5857} 5858 5859void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5860void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5861 5862ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5863 MemRegion span, CMSBitMap* bitMap): 5864 _span(span), 5865 _bitMap(bitMap) 5866{ 5867 assert(ref_processor() == NULL, "deliberately left NULL"); 5868 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5869} 5870 5871void ParMarkRefsIntoClosure::do_oop(oop obj) { 5872 // if p points into _span, then mark corresponding bit in _markBitMap 5873 assert(obj->is_oop(), "expected an oop"); 5874 HeapWord* addr = (HeapWord*)obj; 5875 if (_span.contains(addr)) { 5876 // this should be made more efficient 5877 _bitMap->par_mark(addr); 5878 } 5879} 5880 5881void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5882void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5883 5884// A variant of the above, used for CMS marking verification. 5885MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5886 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5887 _span(span), 5888 _verification_bm(verification_bm), 5889 _cms_bm(cms_bm) 5890{ 5891 assert(ref_processor() == NULL, "deliberately left NULL"); 5892 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5893} 5894 5895void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5896 // if p points into _span, then mark corresponding bit in _markBitMap 5897 assert(obj->is_oop(), "expected an oop"); 5898 HeapWord* addr = (HeapWord*)obj; 5899 if (_span.contains(addr)) { 5900 _verification_bm->mark(addr); 5901 if (!_cms_bm->isMarked(addr)) { 5902 Log(gc, verify) log; 5903 ResourceMark rm; 5904 oop(addr)->print_on(log.error_stream()); 5905 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5906 fatal("... aborting"); 5907 } 5908 } 5909} 5910 5911void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5912void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5913 5914////////////////////////////////////////////////// 5915// MarkRefsIntoAndScanClosure 5916////////////////////////////////////////////////// 5917 5918MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5919 ReferenceProcessor* rp, 5920 CMSBitMap* bit_map, 5921 CMSBitMap* mod_union_table, 5922 CMSMarkStack* mark_stack, 5923 CMSCollector* collector, 5924 bool should_yield, 5925 bool concurrent_precleaning): 5926 _collector(collector), 5927 _span(span), 5928 _bit_map(bit_map), 5929 _mark_stack(mark_stack), 5930 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5931 mark_stack, concurrent_precleaning), 5932 _yield(should_yield), 5933 _concurrent_precleaning(concurrent_precleaning), 5934 _freelistLock(NULL) 5935{ 5936 // FIXME: Should initialize in base class constructor. 5937 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5938 set_ref_processor_internal(rp); 5939} 5940 5941// This closure is used to mark refs into the CMS generation at the 5942// second (final) checkpoint, and to scan and transitively follow 5943// the unmarked oops. It is also used during the concurrent precleaning 5944// phase while scanning objects on dirty cards in the CMS generation. 5945// The marks are made in the marking bit map and the marking stack is 5946// used for keeping the (newly) grey objects during the scan. 5947// The parallel version (Par_...) appears further below. 5948void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5949 if (obj != NULL) { 5950 assert(obj->is_oop(), "expected an oop"); 5951 HeapWord* addr = (HeapWord*)obj; 5952 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5953 assert(_collector->overflow_list_is_empty(), 5954 "overflow list should be empty"); 5955 if (_span.contains(addr) && 5956 !_bit_map->isMarked(addr)) { 5957 // mark bit map (object is now grey) 5958 _bit_map->mark(addr); 5959 // push on marking stack (stack should be empty), and drain the 5960 // stack by applying this closure to the oops in the oops popped 5961 // from the stack (i.e. blacken the grey objects) 5962 bool res = _mark_stack->push(obj); 5963 assert(res, "Should have space to push on empty stack"); 5964 do { 5965 oop new_oop = _mark_stack->pop(); 5966 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5967 assert(_bit_map->isMarked((HeapWord*)new_oop), 5968 "only grey objects on this stack"); 5969 // iterate over the oops in this oop, marking and pushing 5970 // the ones in CMS heap (i.e. in _span). 5971 new_oop->oop_iterate(&_pushAndMarkClosure); 5972 // check if it's time to yield 5973 do_yield_check(); 5974 } while (!_mark_stack->isEmpty() || 5975 (!_concurrent_precleaning && take_from_overflow_list())); 5976 // if marking stack is empty, and we are not doing this 5977 // during precleaning, then check the overflow list 5978 } 5979 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5980 assert(_collector->overflow_list_is_empty(), 5981 "overflow list was drained above"); 5982 5983 assert(_collector->no_preserved_marks(), 5984 "All preserved marks should have been restored above"); 5985 } 5986} 5987 5988void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5989void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5990 5991void MarkRefsIntoAndScanClosure::do_yield_work() { 5992 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5993 "CMS thread should hold CMS token"); 5994 assert_lock_strong(_freelistLock); 5995 assert_lock_strong(_bit_map->lock()); 5996 // relinquish the free_list_lock and bitMaplock() 5997 _bit_map->lock()->unlock(); 5998 _freelistLock->unlock(); 5999 ConcurrentMarkSweepThread::desynchronize(true); 6000 _collector->stopTimer(); 6001 _collector->incrementYields(); 6002 6003 // See the comment in coordinator_yield() 6004 for (unsigned i = 0; 6005 i < CMSYieldSleepCount && 6006 ConcurrentMarkSweepThread::should_yield() && 6007 !CMSCollector::foregroundGCIsActive(); 6008 ++i) { 6009 os::sleep(Thread::current(), 1, false); 6010 } 6011 6012 ConcurrentMarkSweepThread::synchronize(true); 6013 _freelistLock->lock_without_safepoint_check(); 6014 _bit_map->lock()->lock_without_safepoint_check(); 6015 _collector->startTimer(); 6016} 6017 6018/////////////////////////////////////////////////////////// 6019// ParMarkRefsIntoAndScanClosure: a parallel version of 6020// MarkRefsIntoAndScanClosure 6021/////////////////////////////////////////////////////////// 6022ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6023 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6024 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6025 _span(span), 6026 _bit_map(bit_map), 6027 _work_queue(work_queue), 6028 _low_water_mark(MIN2((work_queue->max_elems()/4), 6029 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6030 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6031{ 6032 // FIXME: Should initialize in base class constructor. 6033 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6034 set_ref_processor_internal(rp); 6035} 6036 6037// This closure is used to mark refs into the CMS generation at the 6038// second (final) checkpoint, and to scan and transitively follow 6039// the unmarked oops. The marks are made in the marking bit map and 6040// the work_queue is used for keeping the (newly) grey objects during 6041// the scan phase whence they are also available for stealing by parallel 6042// threads. Since the marking bit map is shared, updates are 6043// synchronized (via CAS). 6044void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6045 if (obj != NULL) { 6046 // Ignore mark word because this could be an already marked oop 6047 // that may be chained at the end of the overflow list. 6048 assert(obj->is_oop(true), "expected an oop"); 6049 HeapWord* addr = (HeapWord*)obj; 6050 if (_span.contains(addr) && 6051 !_bit_map->isMarked(addr)) { 6052 // mark bit map (object will become grey): 6053 // It is possible for several threads to be 6054 // trying to "claim" this object concurrently; 6055 // the unique thread that succeeds in marking the 6056 // object first will do the subsequent push on 6057 // to the work queue (or overflow list). 6058 if (_bit_map->par_mark(addr)) { 6059 // push on work_queue (which may not be empty), and trim the 6060 // queue to an appropriate length by applying this closure to 6061 // the oops in the oops popped from the stack (i.e. blacken the 6062 // grey objects) 6063 bool res = _work_queue->push(obj); 6064 assert(res, "Low water mark should be less than capacity?"); 6065 trim_queue(_low_water_mark); 6066 } // Else, another thread claimed the object 6067 } 6068 } 6069} 6070 6071void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6072void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6073 6074// This closure is used to rescan the marked objects on the dirty cards 6075// in the mod union table and the card table proper. 6076size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6077 oop p, MemRegion mr) { 6078 6079 size_t size = 0; 6080 HeapWord* addr = (HeapWord*)p; 6081 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6082 assert(_span.contains(addr), "we are scanning the CMS generation"); 6083 // check if it's time to yield 6084 if (do_yield_check()) { 6085 // We yielded for some foreground stop-world work, 6086 // and we have been asked to abort this ongoing preclean cycle. 6087 return 0; 6088 } 6089 if (_bitMap->isMarked(addr)) { 6090 // it's marked; is it potentially uninitialized? 6091 if (p->klass_or_null() != NULL) { 6092 // an initialized object; ignore mark word in verification below 6093 // since we are running concurrent with mutators 6094 assert(p->is_oop(true), "should be an oop"); 6095 if (p->is_objArray()) { 6096 // objArrays are precisely marked; restrict scanning 6097 // to dirty cards only. 6098 size = CompactibleFreeListSpace::adjustObjectSize( 6099 p->oop_iterate_size(_scanningClosure, mr)); 6100 } else { 6101 // A non-array may have been imprecisely marked; we need 6102 // to scan object in its entirety. 6103 size = CompactibleFreeListSpace::adjustObjectSize( 6104 p->oop_iterate_size(_scanningClosure)); 6105 } 6106 #ifdef ASSERT 6107 size_t direct_size = 6108 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6109 assert(size == direct_size, "Inconsistency in size"); 6110 assert(size >= 3, "Necessary for Printezis marks to work"); 6111 HeapWord* start_pbit = addr + 1; 6112 HeapWord* end_pbit = addr + size - 1; 6113 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6114 "inconsistent Printezis mark"); 6115 // Verify inner mark bits (between Printezis bits) are clear, 6116 // but don't repeat if there are multiple dirty regions for 6117 // the same object, to avoid potential O(N^2) performance. 6118 if (addr != _last_scanned_object) { 6119 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6120 _last_scanned_object = addr; 6121 } 6122 #endif // ASSERT 6123 } else { 6124 // An uninitialized object. 6125 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6126 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6127 size = pointer_delta(nextOneAddr + 1, addr); 6128 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6129 "alignment problem"); 6130 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6131 // will dirty the card when the klass pointer is installed in the 6132 // object (signaling the completion of initialization). 6133 } 6134 } else { 6135 // Either a not yet marked object or an uninitialized object 6136 if (p->klass_or_null() == NULL) { 6137 // An uninitialized object, skip to the next card, since 6138 // we may not be able to read its P-bits yet. 6139 assert(size == 0, "Initial value"); 6140 } else { 6141 // An object not (yet) reached by marking: we merely need to 6142 // compute its size so as to go look at the next block. 6143 assert(p->is_oop(true), "should be an oop"); 6144 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6145 } 6146 } 6147 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6148 return size; 6149} 6150 6151void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6152 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6153 "CMS thread should hold CMS token"); 6154 assert_lock_strong(_freelistLock); 6155 assert_lock_strong(_bitMap->lock()); 6156 // relinquish the free_list_lock and bitMaplock() 6157 _bitMap->lock()->unlock(); 6158 _freelistLock->unlock(); 6159 ConcurrentMarkSweepThread::desynchronize(true); 6160 _collector->stopTimer(); 6161 _collector->incrementYields(); 6162 6163 // See the comment in coordinator_yield() 6164 for (unsigned i = 0; i < CMSYieldSleepCount && 6165 ConcurrentMarkSweepThread::should_yield() && 6166 !CMSCollector::foregroundGCIsActive(); ++i) { 6167 os::sleep(Thread::current(), 1, false); 6168 } 6169 6170 ConcurrentMarkSweepThread::synchronize(true); 6171 _freelistLock->lock_without_safepoint_check(); 6172 _bitMap->lock()->lock_without_safepoint_check(); 6173 _collector->startTimer(); 6174} 6175 6176 6177////////////////////////////////////////////////////////////////// 6178// SurvivorSpacePrecleanClosure 6179////////////////////////////////////////////////////////////////// 6180// This (single-threaded) closure is used to preclean the oops in 6181// the survivor spaces. 6182size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6183 6184 HeapWord* addr = (HeapWord*)p; 6185 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6186 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6187 assert(p->klass_or_null() != NULL, "object should be initialized"); 6188 // an initialized object; ignore mark word in verification below 6189 // since we are running concurrent with mutators 6190 assert(p->is_oop(true), "should be an oop"); 6191 // Note that we do not yield while we iterate over 6192 // the interior oops of p, pushing the relevant ones 6193 // on our marking stack. 6194 size_t size = p->oop_iterate_size(_scanning_closure); 6195 do_yield_check(); 6196 // Observe that below, we do not abandon the preclean 6197 // phase as soon as we should; rather we empty the 6198 // marking stack before returning. This is to satisfy 6199 // some existing assertions. In general, it may be a 6200 // good idea to abort immediately and complete the marking 6201 // from the grey objects at a later time. 6202 while (!_mark_stack->isEmpty()) { 6203 oop new_oop = _mark_stack->pop(); 6204 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6205 assert(_bit_map->isMarked((HeapWord*)new_oop), 6206 "only grey objects on this stack"); 6207 // iterate over the oops in this oop, marking and pushing 6208 // the ones in CMS heap (i.e. in _span). 6209 new_oop->oop_iterate(_scanning_closure); 6210 // check if it's time to yield 6211 do_yield_check(); 6212 } 6213 unsigned int after_count = 6214 GenCollectedHeap::heap()->total_collections(); 6215 bool abort = (_before_count != after_count) || 6216 _collector->should_abort_preclean(); 6217 return abort ? 0 : size; 6218} 6219 6220void SurvivorSpacePrecleanClosure::do_yield_work() { 6221 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6222 "CMS thread should hold CMS token"); 6223 assert_lock_strong(_bit_map->lock()); 6224 // Relinquish the bit map lock 6225 _bit_map->lock()->unlock(); 6226 ConcurrentMarkSweepThread::desynchronize(true); 6227 _collector->stopTimer(); 6228 _collector->incrementYields(); 6229 6230 // See the comment in coordinator_yield() 6231 for (unsigned i = 0; i < CMSYieldSleepCount && 6232 ConcurrentMarkSweepThread::should_yield() && 6233 !CMSCollector::foregroundGCIsActive(); ++i) { 6234 os::sleep(Thread::current(), 1, false); 6235 } 6236 6237 ConcurrentMarkSweepThread::synchronize(true); 6238 _bit_map->lock()->lock_without_safepoint_check(); 6239 _collector->startTimer(); 6240} 6241 6242// This closure is used to rescan the marked objects on the dirty cards 6243// in the mod union table and the card table proper. In the parallel 6244// case, although the bitMap is shared, we do a single read so the 6245// isMarked() query is "safe". 6246bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6247 // Ignore mark word because we are running concurrent with mutators 6248 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6249 HeapWord* addr = (HeapWord*)p; 6250 assert(_span.contains(addr), "we are scanning the CMS generation"); 6251 bool is_obj_array = false; 6252 #ifdef ASSERT 6253 if (!_parallel) { 6254 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6255 assert(_collector->overflow_list_is_empty(), 6256 "overflow list should be empty"); 6257 6258 } 6259 #endif // ASSERT 6260 if (_bit_map->isMarked(addr)) { 6261 // Obj arrays are precisely marked, non-arrays are not; 6262 // so we scan objArrays precisely and non-arrays in their 6263 // entirety. 6264 if (p->is_objArray()) { 6265 is_obj_array = true; 6266 if (_parallel) { 6267 p->oop_iterate(_par_scan_closure, mr); 6268 } else { 6269 p->oop_iterate(_scan_closure, mr); 6270 } 6271 } else { 6272 if (_parallel) { 6273 p->oop_iterate(_par_scan_closure); 6274 } else { 6275 p->oop_iterate(_scan_closure); 6276 } 6277 } 6278 } 6279 #ifdef ASSERT 6280 if (!_parallel) { 6281 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6282 assert(_collector->overflow_list_is_empty(), 6283 "overflow list should be empty"); 6284 6285 } 6286 #endif // ASSERT 6287 return is_obj_array; 6288} 6289 6290MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6291 MemRegion span, 6292 CMSBitMap* bitMap, CMSMarkStack* markStack, 6293 bool should_yield, bool verifying): 6294 _collector(collector), 6295 _span(span), 6296 _bitMap(bitMap), 6297 _mut(&collector->_modUnionTable), 6298 _markStack(markStack), 6299 _yield(should_yield), 6300 _skipBits(0) 6301{ 6302 assert(_markStack->isEmpty(), "stack should be empty"); 6303 _finger = _bitMap->startWord(); 6304 _threshold = _finger; 6305 assert(_collector->_restart_addr == NULL, "Sanity check"); 6306 assert(_span.contains(_finger), "Out of bounds _finger?"); 6307 DEBUG_ONLY(_verifying = verifying;) 6308} 6309 6310void MarkFromRootsClosure::reset(HeapWord* addr) { 6311 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6312 assert(_span.contains(addr), "Out of bounds _finger?"); 6313 _finger = addr; 6314 _threshold = (HeapWord*)round_to( 6315 (intptr_t)_finger, CardTableModRefBS::card_size); 6316} 6317 6318// Should revisit to see if this should be restructured for 6319// greater efficiency. 6320bool MarkFromRootsClosure::do_bit(size_t offset) { 6321 if (_skipBits > 0) { 6322 _skipBits--; 6323 return true; 6324 } 6325 // convert offset into a HeapWord* 6326 HeapWord* addr = _bitMap->startWord() + offset; 6327 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6328 "address out of range"); 6329 assert(_bitMap->isMarked(addr), "tautology"); 6330 if (_bitMap->isMarked(addr+1)) { 6331 // this is an allocated but not yet initialized object 6332 assert(_skipBits == 0, "tautology"); 6333 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6334 oop p = oop(addr); 6335 if (p->klass_or_null() == NULL) { 6336 DEBUG_ONLY(if (!_verifying) {) 6337 // We re-dirty the cards on which this object lies and increase 6338 // the _threshold so that we'll come back to scan this object 6339 // during the preclean or remark phase. (CMSCleanOnEnter) 6340 if (CMSCleanOnEnter) { 6341 size_t sz = _collector->block_size_using_printezis_bits(addr); 6342 HeapWord* end_card_addr = (HeapWord*)round_to( 6343 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6344 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6345 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6346 // Bump _threshold to end_card_addr; note that 6347 // _threshold cannot possibly exceed end_card_addr, anyhow. 6348 // This prevents future clearing of the card as the scan proceeds 6349 // to the right. 6350 assert(_threshold <= end_card_addr, 6351 "Because we are just scanning into this object"); 6352 if (_threshold < end_card_addr) { 6353 _threshold = end_card_addr; 6354 } 6355 if (p->klass_or_null() != NULL) { 6356 // Redirty the range of cards... 6357 _mut->mark_range(redirty_range); 6358 } // ...else the setting of klass will dirty the card anyway. 6359 } 6360 DEBUG_ONLY(}) 6361 return true; 6362 } 6363 } 6364 scanOopsInOop(addr); 6365 return true; 6366} 6367 6368// We take a break if we've been at this for a while, 6369// so as to avoid monopolizing the locks involved. 6370void MarkFromRootsClosure::do_yield_work() { 6371 // First give up the locks, then yield, then re-lock 6372 // We should probably use a constructor/destructor idiom to 6373 // do this unlock/lock or modify the MutexUnlocker class to 6374 // serve our purpose. XXX 6375 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6376 "CMS thread should hold CMS token"); 6377 assert_lock_strong(_bitMap->lock()); 6378 _bitMap->lock()->unlock(); 6379 ConcurrentMarkSweepThread::desynchronize(true); 6380 _collector->stopTimer(); 6381 _collector->incrementYields(); 6382 6383 // See the comment in coordinator_yield() 6384 for (unsigned i = 0; i < CMSYieldSleepCount && 6385 ConcurrentMarkSweepThread::should_yield() && 6386 !CMSCollector::foregroundGCIsActive(); ++i) { 6387 os::sleep(Thread::current(), 1, false); 6388 } 6389 6390 ConcurrentMarkSweepThread::synchronize(true); 6391 _bitMap->lock()->lock_without_safepoint_check(); 6392 _collector->startTimer(); 6393} 6394 6395void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6396 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6397 assert(_markStack->isEmpty(), 6398 "should drain stack to limit stack usage"); 6399 // convert ptr to an oop preparatory to scanning 6400 oop obj = oop(ptr); 6401 // Ignore mark word in verification below, since we 6402 // may be running concurrent with mutators. 6403 assert(obj->is_oop(true), "should be an oop"); 6404 assert(_finger <= ptr, "_finger runneth ahead"); 6405 // advance the finger to right end of this object 6406 _finger = ptr + obj->size(); 6407 assert(_finger > ptr, "we just incremented it above"); 6408 // On large heaps, it may take us some time to get through 6409 // the marking phase. During 6410 // this time it's possible that a lot of mutations have 6411 // accumulated in the card table and the mod union table -- 6412 // these mutation records are redundant until we have 6413 // actually traced into the corresponding card. 6414 // Here, we check whether advancing the finger would make 6415 // us cross into a new card, and if so clear corresponding 6416 // cards in the MUT (preclean them in the card-table in the 6417 // future). 6418 6419 DEBUG_ONLY(if (!_verifying) {) 6420 // The clean-on-enter optimization is disabled by default, 6421 // until we fix 6178663. 6422 if (CMSCleanOnEnter && (_finger > _threshold)) { 6423 // [_threshold, _finger) represents the interval 6424 // of cards to be cleared in MUT (or precleaned in card table). 6425 // The set of cards to be cleared is all those that overlap 6426 // with the interval [_threshold, _finger); note that 6427 // _threshold is always kept card-aligned but _finger isn't 6428 // always card-aligned. 6429 HeapWord* old_threshold = _threshold; 6430 assert(old_threshold == (HeapWord*)round_to( 6431 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6432 "_threshold should always be card-aligned"); 6433 _threshold = (HeapWord*)round_to( 6434 (intptr_t)_finger, CardTableModRefBS::card_size); 6435 MemRegion mr(old_threshold, _threshold); 6436 assert(!mr.is_empty(), "Control point invariant"); 6437 assert(_span.contains(mr), "Should clear within span"); 6438 _mut->clear_range(mr); 6439 } 6440 DEBUG_ONLY(}) 6441 // Note: the finger doesn't advance while we drain 6442 // the stack below. 6443 PushOrMarkClosure pushOrMarkClosure(_collector, 6444 _span, _bitMap, _markStack, 6445 _finger, this); 6446 bool res = _markStack->push(obj); 6447 assert(res, "Empty non-zero size stack should have space for single push"); 6448 while (!_markStack->isEmpty()) { 6449 oop new_oop = _markStack->pop(); 6450 // Skip verifying header mark word below because we are 6451 // running concurrent with mutators. 6452 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6453 // now scan this oop's oops 6454 new_oop->oop_iterate(&pushOrMarkClosure); 6455 do_yield_check(); 6456 } 6457 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6458} 6459 6460ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6461 CMSCollector* collector, MemRegion span, 6462 CMSBitMap* bit_map, 6463 OopTaskQueue* work_queue, 6464 CMSMarkStack* overflow_stack): 6465 _collector(collector), 6466 _whole_span(collector->_span), 6467 _span(span), 6468 _bit_map(bit_map), 6469 _mut(&collector->_modUnionTable), 6470 _work_queue(work_queue), 6471 _overflow_stack(overflow_stack), 6472 _skip_bits(0), 6473 _task(task) 6474{ 6475 assert(_work_queue->size() == 0, "work_queue should be empty"); 6476 _finger = span.start(); 6477 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6478 assert(_span.contains(_finger), "Out of bounds _finger?"); 6479} 6480 6481// Should revisit to see if this should be restructured for 6482// greater efficiency. 6483bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6484 if (_skip_bits > 0) { 6485 _skip_bits--; 6486 return true; 6487 } 6488 // convert offset into a HeapWord* 6489 HeapWord* addr = _bit_map->startWord() + offset; 6490 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6491 "address out of range"); 6492 assert(_bit_map->isMarked(addr), "tautology"); 6493 if (_bit_map->isMarked(addr+1)) { 6494 // this is an allocated object that might not yet be initialized 6495 assert(_skip_bits == 0, "tautology"); 6496 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6497 oop p = oop(addr); 6498 if (p->klass_or_null() == NULL) { 6499 // in the case of Clean-on-Enter optimization, redirty card 6500 // and avoid clearing card by increasing the threshold. 6501 return true; 6502 } 6503 } 6504 scan_oops_in_oop(addr); 6505 return true; 6506} 6507 6508void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6509 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6510 // Should we assert that our work queue is empty or 6511 // below some drain limit? 6512 assert(_work_queue->size() == 0, 6513 "should drain stack to limit stack usage"); 6514 // convert ptr to an oop preparatory to scanning 6515 oop obj = oop(ptr); 6516 // Ignore mark word in verification below, since we 6517 // may be running concurrent with mutators. 6518 assert(obj->is_oop(true), "should be an oop"); 6519 assert(_finger <= ptr, "_finger runneth ahead"); 6520 // advance the finger to right end of this object 6521 _finger = ptr + obj->size(); 6522 assert(_finger > ptr, "we just incremented it above"); 6523 // On large heaps, it may take us some time to get through 6524 // the marking phase. During 6525 // this time it's possible that a lot of mutations have 6526 // accumulated in the card table and the mod union table -- 6527 // these mutation records are redundant until we have 6528 // actually traced into the corresponding card. 6529 // Here, we check whether advancing the finger would make 6530 // us cross into a new card, and if so clear corresponding 6531 // cards in the MUT (preclean them in the card-table in the 6532 // future). 6533 6534 // The clean-on-enter optimization is disabled by default, 6535 // until we fix 6178663. 6536 if (CMSCleanOnEnter && (_finger > _threshold)) { 6537 // [_threshold, _finger) represents the interval 6538 // of cards to be cleared in MUT (or precleaned in card table). 6539 // The set of cards to be cleared is all those that overlap 6540 // with the interval [_threshold, _finger); note that 6541 // _threshold is always kept card-aligned but _finger isn't 6542 // always card-aligned. 6543 HeapWord* old_threshold = _threshold; 6544 assert(old_threshold == (HeapWord*)round_to( 6545 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6546 "_threshold should always be card-aligned"); 6547 _threshold = (HeapWord*)round_to( 6548 (intptr_t)_finger, CardTableModRefBS::card_size); 6549 MemRegion mr(old_threshold, _threshold); 6550 assert(!mr.is_empty(), "Control point invariant"); 6551 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6552 _mut->clear_range(mr); 6553 } 6554 6555 // Note: the local finger doesn't advance while we drain 6556 // the stack below, but the global finger sure can and will. 6557 HeapWord** gfa = _task->global_finger_addr(); 6558 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6559 _span, _bit_map, 6560 _work_queue, 6561 _overflow_stack, 6562 _finger, 6563 gfa, this); 6564 bool res = _work_queue->push(obj); // overflow could occur here 6565 assert(res, "Will hold once we use workqueues"); 6566 while (true) { 6567 oop new_oop; 6568 if (!_work_queue->pop_local(new_oop)) { 6569 // We emptied our work_queue; check if there's stuff that can 6570 // be gotten from the overflow stack. 6571 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6572 _overflow_stack, _work_queue)) { 6573 do_yield_check(); 6574 continue; 6575 } else { // done 6576 break; 6577 } 6578 } 6579 // Skip verifying header mark word below because we are 6580 // running concurrent with mutators. 6581 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6582 // now scan this oop's oops 6583 new_oop->oop_iterate(&pushOrMarkClosure); 6584 do_yield_check(); 6585 } 6586 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6587} 6588 6589// Yield in response to a request from VM Thread or 6590// from mutators. 6591void ParMarkFromRootsClosure::do_yield_work() { 6592 assert(_task != NULL, "sanity"); 6593 _task->yield(); 6594} 6595 6596// A variant of the above used for verifying CMS marking work. 6597MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6598 MemRegion span, 6599 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6600 CMSMarkStack* mark_stack): 6601 _collector(collector), 6602 _span(span), 6603 _verification_bm(verification_bm), 6604 _cms_bm(cms_bm), 6605 _mark_stack(mark_stack), 6606 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6607 mark_stack) 6608{ 6609 assert(_mark_stack->isEmpty(), "stack should be empty"); 6610 _finger = _verification_bm->startWord(); 6611 assert(_collector->_restart_addr == NULL, "Sanity check"); 6612 assert(_span.contains(_finger), "Out of bounds _finger?"); 6613} 6614 6615void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6616 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6617 assert(_span.contains(addr), "Out of bounds _finger?"); 6618 _finger = addr; 6619} 6620 6621// Should revisit to see if this should be restructured for 6622// greater efficiency. 6623bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6624 // convert offset into a HeapWord* 6625 HeapWord* addr = _verification_bm->startWord() + offset; 6626 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6627 "address out of range"); 6628 assert(_verification_bm->isMarked(addr), "tautology"); 6629 assert(_cms_bm->isMarked(addr), "tautology"); 6630 6631 assert(_mark_stack->isEmpty(), 6632 "should drain stack to limit stack usage"); 6633 // convert addr to an oop preparatory to scanning 6634 oop obj = oop(addr); 6635 assert(obj->is_oop(), "should be an oop"); 6636 assert(_finger <= addr, "_finger runneth ahead"); 6637 // advance the finger to right end of this object 6638 _finger = addr + obj->size(); 6639 assert(_finger > addr, "we just incremented it above"); 6640 // Note: the finger doesn't advance while we drain 6641 // the stack below. 6642 bool res = _mark_stack->push(obj); 6643 assert(res, "Empty non-zero size stack should have space for single push"); 6644 while (!_mark_stack->isEmpty()) { 6645 oop new_oop = _mark_stack->pop(); 6646 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6647 // now scan this oop's oops 6648 new_oop->oop_iterate(&_pam_verify_closure); 6649 } 6650 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6651 return true; 6652} 6653 6654PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6655 CMSCollector* collector, MemRegion span, 6656 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6657 CMSMarkStack* mark_stack): 6658 MetadataAwareOopClosure(collector->ref_processor()), 6659 _collector(collector), 6660 _span(span), 6661 _verification_bm(verification_bm), 6662 _cms_bm(cms_bm), 6663 _mark_stack(mark_stack) 6664{ } 6665 6666void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6667void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6668 6669// Upon stack overflow, we discard (part of) the stack, 6670// remembering the least address amongst those discarded 6671// in CMSCollector's _restart_address. 6672void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6673 // Remember the least grey address discarded 6674 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6675 _collector->lower_restart_addr(ra); 6676 _mark_stack->reset(); // discard stack contents 6677 _mark_stack->expand(); // expand the stack if possible 6678} 6679 6680void PushAndMarkVerifyClosure::do_oop(oop obj) { 6681 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6682 HeapWord* addr = (HeapWord*)obj; 6683 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6684 // Oop lies in _span and isn't yet grey or black 6685 _verification_bm->mark(addr); // now grey 6686 if (!_cms_bm->isMarked(addr)) { 6687 Log(gc, verify) log; 6688 ResourceMark rm; 6689 oop(addr)->print_on(log.error_stream()); 6690 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6691 fatal("... aborting"); 6692 } 6693 6694 if (!_mark_stack->push(obj)) { // stack overflow 6695 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6696 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6697 handle_stack_overflow(addr); 6698 } 6699 // anything including and to the right of _finger 6700 // will be scanned as we iterate over the remainder of the 6701 // bit map 6702 } 6703} 6704 6705PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6706 MemRegion span, 6707 CMSBitMap* bitMap, CMSMarkStack* markStack, 6708 HeapWord* finger, MarkFromRootsClosure* parent) : 6709 MetadataAwareOopClosure(collector->ref_processor()), 6710 _collector(collector), 6711 _span(span), 6712 _bitMap(bitMap), 6713 _markStack(markStack), 6714 _finger(finger), 6715 _parent(parent) 6716{ } 6717 6718ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6719 MemRegion span, 6720 CMSBitMap* bit_map, 6721 OopTaskQueue* work_queue, 6722 CMSMarkStack* overflow_stack, 6723 HeapWord* finger, 6724 HeapWord** global_finger_addr, 6725 ParMarkFromRootsClosure* parent) : 6726 MetadataAwareOopClosure(collector->ref_processor()), 6727 _collector(collector), 6728 _whole_span(collector->_span), 6729 _span(span), 6730 _bit_map(bit_map), 6731 _work_queue(work_queue), 6732 _overflow_stack(overflow_stack), 6733 _finger(finger), 6734 _global_finger_addr(global_finger_addr), 6735 _parent(parent) 6736{ } 6737 6738// Assumes thread-safe access by callers, who are 6739// responsible for mutual exclusion. 6740void CMSCollector::lower_restart_addr(HeapWord* low) { 6741 assert(_span.contains(low), "Out of bounds addr"); 6742 if (_restart_addr == NULL) { 6743 _restart_addr = low; 6744 } else { 6745 _restart_addr = MIN2(_restart_addr, low); 6746 } 6747} 6748 6749// Upon stack overflow, we discard (part of) the stack, 6750// remembering the least address amongst those discarded 6751// in CMSCollector's _restart_address. 6752void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6753 // Remember the least grey address discarded 6754 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6755 _collector->lower_restart_addr(ra); 6756 _markStack->reset(); // discard stack contents 6757 _markStack->expand(); // expand the stack if possible 6758} 6759 6760// Upon stack overflow, we discard (part of) the stack, 6761// remembering the least address amongst those discarded 6762// in CMSCollector's _restart_address. 6763void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6764 // We need to do this under a mutex to prevent other 6765 // workers from interfering with the work done below. 6766 MutexLockerEx ml(_overflow_stack->par_lock(), 6767 Mutex::_no_safepoint_check_flag); 6768 // Remember the least grey address discarded 6769 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6770 _collector->lower_restart_addr(ra); 6771 _overflow_stack->reset(); // discard stack contents 6772 _overflow_stack->expand(); // expand the stack if possible 6773} 6774 6775void PushOrMarkClosure::do_oop(oop obj) { 6776 // Ignore mark word because we are running concurrent with mutators. 6777 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6778 HeapWord* addr = (HeapWord*)obj; 6779 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6780 // Oop lies in _span and isn't yet grey or black 6781 _bitMap->mark(addr); // now grey 6782 if (addr < _finger) { 6783 // the bit map iteration has already either passed, or 6784 // sampled, this bit in the bit map; we'll need to 6785 // use the marking stack to scan this oop's oops. 6786 bool simulate_overflow = false; 6787 NOT_PRODUCT( 6788 if (CMSMarkStackOverflowALot && 6789 _collector->simulate_overflow()) { 6790 // simulate a stack overflow 6791 simulate_overflow = true; 6792 } 6793 ) 6794 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6795 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6796 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6797 handle_stack_overflow(addr); 6798 } 6799 } 6800 // anything including and to the right of _finger 6801 // will be scanned as we iterate over the remainder of the 6802 // bit map 6803 do_yield_check(); 6804 } 6805} 6806 6807void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6808void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6809 6810void ParPushOrMarkClosure::do_oop(oop obj) { 6811 // Ignore mark word because we are running concurrent with mutators. 6812 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6813 HeapWord* addr = (HeapWord*)obj; 6814 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6815 // Oop lies in _span and isn't yet grey or black 6816 // We read the global_finger (volatile read) strictly after marking oop 6817 bool res = _bit_map->par_mark(addr); // now grey 6818 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6819 // Should we push this marked oop on our stack? 6820 // -- if someone else marked it, nothing to do 6821 // -- if target oop is above global finger nothing to do 6822 // -- if target oop is in chunk and above local finger 6823 // then nothing to do 6824 // -- else push on work queue 6825 if ( !res // someone else marked it, they will deal with it 6826 || (addr >= *gfa) // will be scanned in a later task 6827 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6828 return; 6829 } 6830 // the bit map iteration has already either passed, or 6831 // sampled, this bit in the bit map; we'll need to 6832 // use the marking stack to scan this oop's oops. 6833 bool simulate_overflow = false; 6834 NOT_PRODUCT( 6835 if (CMSMarkStackOverflowALot && 6836 _collector->simulate_overflow()) { 6837 // simulate a stack overflow 6838 simulate_overflow = true; 6839 } 6840 ) 6841 if (simulate_overflow || 6842 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6843 // stack overflow 6844 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6845 // We cannot assert that the overflow stack is full because 6846 // it may have been emptied since. 6847 assert(simulate_overflow || 6848 _work_queue->size() == _work_queue->max_elems(), 6849 "Else push should have succeeded"); 6850 handle_stack_overflow(addr); 6851 } 6852 do_yield_check(); 6853 } 6854} 6855 6856void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6857void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6858 6859PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6860 MemRegion span, 6861 ReferenceProcessor* rp, 6862 CMSBitMap* bit_map, 6863 CMSBitMap* mod_union_table, 6864 CMSMarkStack* mark_stack, 6865 bool concurrent_precleaning): 6866 MetadataAwareOopClosure(rp), 6867 _collector(collector), 6868 _span(span), 6869 _bit_map(bit_map), 6870 _mod_union_table(mod_union_table), 6871 _mark_stack(mark_stack), 6872 _concurrent_precleaning(concurrent_precleaning) 6873{ 6874 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6875} 6876 6877// Grey object rescan during pre-cleaning and second checkpoint phases -- 6878// the non-parallel version (the parallel version appears further below.) 6879void PushAndMarkClosure::do_oop(oop obj) { 6880 // Ignore mark word verification. If during concurrent precleaning, 6881 // the object monitor may be locked. If during the checkpoint 6882 // phases, the object may already have been reached by a different 6883 // path and may be at the end of the global overflow list (so 6884 // the mark word may be NULL). 6885 assert(obj->is_oop_or_null(true /* ignore mark word */), 6886 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6887 HeapWord* addr = (HeapWord*)obj; 6888 // Check if oop points into the CMS generation 6889 // and is not marked 6890 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6891 // a white object ... 6892 _bit_map->mark(addr); // ... now grey 6893 // push on the marking stack (grey set) 6894 bool simulate_overflow = false; 6895 NOT_PRODUCT( 6896 if (CMSMarkStackOverflowALot && 6897 _collector->simulate_overflow()) { 6898 // simulate a stack overflow 6899 simulate_overflow = true; 6900 } 6901 ) 6902 if (simulate_overflow || !_mark_stack->push(obj)) { 6903 if (_concurrent_precleaning) { 6904 // During precleaning we can just dirty the appropriate card(s) 6905 // in the mod union table, thus ensuring that the object remains 6906 // in the grey set and continue. In the case of object arrays 6907 // we need to dirty all of the cards that the object spans, 6908 // since the rescan of object arrays will be limited to the 6909 // dirty cards. 6910 // Note that no one can be interfering with us in this action 6911 // of dirtying the mod union table, so no locking or atomics 6912 // are required. 6913 if (obj->is_objArray()) { 6914 size_t sz = obj->size(); 6915 HeapWord* end_card_addr = (HeapWord*)round_to( 6916 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6917 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6918 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6919 _mod_union_table->mark_range(redirty_range); 6920 } else { 6921 _mod_union_table->mark(addr); 6922 } 6923 _collector->_ser_pmc_preclean_ovflw++; 6924 } else { 6925 // During the remark phase, we need to remember this oop 6926 // in the overflow list. 6927 _collector->push_on_overflow_list(obj); 6928 _collector->_ser_pmc_remark_ovflw++; 6929 } 6930 } 6931 } 6932} 6933 6934ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6935 MemRegion span, 6936 ReferenceProcessor* rp, 6937 CMSBitMap* bit_map, 6938 OopTaskQueue* work_queue): 6939 MetadataAwareOopClosure(rp), 6940 _collector(collector), 6941 _span(span), 6942 _bit_map(bit_map), 6943 _work_queue(work_queue) 6944{ 6945 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6946} 6947 6948void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6949void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6950 6951// Grey object rescan during second checkpoint phase -- 6952// the parallel version. 6953void ParPushAndMarkClosure::do_oop(oop obj) { 6954 // In the assert below, we ignore the mark word because 6955 // this oop may point to an already visited object that is 6956 // on the overflow stack (in which case the mark word has 6957 // been hijacked for chaining into the overflow stack -- 6958 // if this is the last object in the overflow stack then 6959 // its mark word will be NULL). Because this object may 6960 // have been subsequently popped off the global overflow 6961 // stack, and the mark word possibly restored to the prototypical 6962 // value, by the time we get to examined this failing assert in 6963 // the debugger, is_oop_or_null(false) may subsequently start 6964 // to hold. 6965 assert(obj->is_oop_or_null(true), 6966 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6967 HeapWord* addr = (HeapWord*)obj; 6968 // Check if oop points into the CMS generation 6969 // and is not marked 6970 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6971 // a white object ... 6972 // If we manage to "claim" the object, by being the 6973 // first thread to mark it, then we push it on our 6974 // marking stack 6975 if (_bit_map->par_mark(addr)) { // ... now grey 6976 // push on work queue (grey set) 6977 bool simulate_overflow = false; 6978 NOT_PRODUCT( 6979 if (CMSMarkStackOverflowALot && 6980 _collector->par_simulate_overflow()) { 6981 // simulate a stack overflow 6982 simulate_overflow = true; 6983 } 6984 ) 6985 if (simulate_overflow || !_work_queue->push(obj)) { 6986 _collector->par_push_on_overflow_list(obj); 6987 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6988 } 6989 } // Else, some other thread got there first 6990 } 6991} 6992 6993void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6994void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6995 6996void CMSPrecleanRefsYieldClosure::do_yield_work() { 6997 Mutex* bml = _collector->bitMapLock(); 6998 assert_lock_strong(bml); 6999 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7000 "CMS thread should hold CMS token"); 7001 7002 bml->unlock(); 7003 ConcurrentMarkSweepThread::desynchronize(true); 7004 7005 _collector->stopTimer(); 7006 _collector->incrementYields(); 7007 7008 // See the comment in coordinator_yield() 7009 for (unsigned i = 0; i < CMSYieldSleepCount && 7010 ConcurrentMarkSweepThread::should_yield() && 7011 !CMSCollector::foregroundGCIsActive(); ++i) { 7012 os::sleep(Thread::current(), 1, false); 7013 } 7014 7015 ConcurrentMarkSweepThread::synchronize(true); 7016 bml->lock(); 7017 7018 _collector->startTimer(); 7019} 7020 7021bool CMSPrecleanRefsYieldClosure::should_return() { 7022 if (ConcurrentMarkSweepThread::should_yield()) { 7023 do_yield_work(); 7024 } 7025 return _collector->foregroundGCIsActive(); 7026} 7027 7028void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7029 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7030 "mr should be aligned to start at a card boundary"); 7031 // We'd like to assert: 7032 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7033 // "mr should be a range of cards"); 7034 // However, that would be too strong in one case -- the last 7035 // partition ends at _unallocated_block which, in general, can be 7036 // an arbitrary boundary, not necessarily card aligned. 7037 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7038 _space->object_iterate_mem(mr, &_scan_cl); 7039} 7040 7041SweepClosure::SweepClosure(CMSCollector* collector, 7042 ConcurrentMarkSweepGeneration* g, 7043 CMSBitMap* bitMap, bool should_yield) : 7044 _collector(collector), 7045 _g(g), 7046 _sp(g->cmsSpace()), 7047 _limit(_sp->sweep_limit()), 7048 _freelistLock(_sp->freelistLock()), 7049 _bitMap(bitMap), 7050 _yield(should_yield), 7051 _inFreeRange(false), // No free range at beginning of sweep 7052 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7053 _lastFreeRangeCoalesced(false), 7054 _freeFinger(g->used_region().start()) 7055{ 7056 NOT_PRODUCT( 7057 _numObjectsFreed = 0; 7058 _numWordsFreed = 0; 7059 _numObjectsLive = 0; 7060 _numWordsLive = 0; 7061 _numObjectsAlreadyFree = 0; 7062 _numWordsAlreadyFree = 0; 7063 _last_fc = NULL; 7064 7065 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7066 _sp->dictionary()->initialize_dict_returned_bytes(); 7067 ) 7068 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7069 "sweep _limit out of bounds"); 7070 log_develop_trace(gc, sweep)("===================="); 7071 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7072} 7073 7074void SweepClosure::print_on(outputStream* st) const { 7075 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7076 p2i(_sp->bottom()), p2i(_sp->end())); 7077 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7078 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7079 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7080 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7081 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7082} 7083 7084#ifndef PRODUCT 7085// Assertion checking only: no useful work in product mode -- 7086// however, if any of the flags below become product flags, 7087// you may need to review this code to see if it needs to be 7088// enabled in product mode. 7089SweepClosure::~SweepClosure() { 7090 assert_lock_strong(_freelistLock); 7091 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7092 "sweep _limit out of bounds"); 7093 if (inFreeRange()) { 7094 Log(gc, sweep) log; 7095 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7096 ResourceMark rm; 7097 print_on(log.error_stream()); 7098 ShouldNotReachHere(); 7099 } 7100 7101 if (log_is_enabled(Debug, gc, sweep)) { 7102 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7103 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7104 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7105 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7106 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7107 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7108 } 7109 7110 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7111 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7112 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7113 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7114 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7115 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7116 } 7117 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7118 log_develop_trace(gc, sweep)("================"); 7119} 7120#endif // PRODUCT 7121 7122void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7123 bool freeRangeInFreeLists) { 7124 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7125 p2i(freeFinger), freeRangeInFreeLists); 7126 assert(!inFreeRange(), "Trampling existing free range"); 7127 set_inFreeRange(true); 7128 set_lastFreeRangeCoalesced(false); 7129 7130 set_freeFinger(freeFinger); 7131 set_freeRangeInFreeLists(freeRangeInFreeLists); 7132 if (CMSTestInFreeList) { 7133 if (freeRangeInFreeLists) { 7134 FreeChunk* fc = (FreeChunk*) freeFinger; 7135 assert(fc->is_free(), "A chunk on the free list should be free."); 7136 assert(fc->size() > 0, "Free range should have a size"); 7137 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7138 } 7139 } 7140} 7141 7142// Note that the sweeper runs concurrently with mutators. Thus, 7143// it is possible for direct allocation in this generation to happen 7144// in the middle of the sweep. Note that the sweeper also coalesces 7145// contiguous free blocks. Thus, unless the sweeper and the allocator 7146// synchronize appropriately freshly allocated blocks may get swept up. 7147// This is accomplished by the sweeper locking the free lists while 7148// it is sweeping. Thus blocks that are determined to be free are 7149// indeed free. There is however one additional complication: 7150// blocks that have been allocated since the final checkpoint and 7151// mark, will not have been marked and so would be treated as 7152// unreachable and swept up. To prevent this, the allocator marks 7153// the bit map when allocating during the sweep phase. This leads, 7154// however, to a further complication -- objects may have been allocated 7155// but not yet initialized -- in the sense that the header isn't yet 7156// installed. The sweeper can not then determine the size of the block 7157// in order to skip over it. To deal with this case, we use a technique 7158// (due to Printezis) to encode such uninitialized block sizes in the 7159// bit map. Since the bit map uses a bit per every HeapWord, but the 7160// CMS generation has a minimum object size of 3 HeapWords, it follows 7161// that "normal marks" won't be adjacent in the bit map (there will 7162// always be at least two 0 bits between successive 1 bits). We make use 7163// of these "unused" bits to represent uninitialized blocks -- the bit 7164// corresponding to the start of the uninitialized object and the next 7165// bit are both set. Finally, a 1 bit marks the end of the object that 7166// started with the two consecutive 1 bits to indicate its potentially 7167// uninitialized state. 7168 7169size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7170 FreeChunk* fc = (FreeChunk*)addr; 7171 size_t res; 7172 7173 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7174 // than "addr == _limit" because although _limit was a block boundary when 7175 // we started the sweep, it may no longer be one because heap expansion 7176 // may have caused us to coalesce the block ending at the address _limit 7177 // with a newly expanded chunk (this happens when _limit was set to the 7178 // previous _end of the space), so we may have stepped past _limit: 7179 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7180 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7181 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7182 "sweep _limit out of bounds"); 7183 assert(addr < _sp->end(), "addr out of bounds"); 7184 // Flush any free range we might be holding as a single 7185 // coalesced chunk to the appropriate free list. 7186 if (inFreeRange()) { 7187 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7188 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7189 flush_cur_free_chunk(freeFinger(), 7190 pointer_delta(addr, freeFinger())); 7191 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7192 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7193 lastFreeRangeCoalesced() ? 1 : 0); 7194 } 7195 7196 // help the iterator loop finish 7197 return pointer_delta(_sp->end(), addr); 7198 } 7199 7200 assert(addr < _limit, "sweep invariant"); 7201 // check if we should yield 7202 do_yield_check(addr); 7203 if (fc->is_free()) { 7204 // Chunk that is already free 7205 res = fc->size(); 7206 do_already_free_chunk(fc); 7207 debug_only(_sp->verifyFreeLists()); 7208 // If we flush the chunk at hand in lookahead_and_flush() 7209 // and it's coalesced with a preceding chunk, then the 7210 // process of "mangling" the payload of the coalesced block 7211 // will cause erasure of the size information from the 7212 // (erstwhile) header of all the coalesced blocks but the 7213 // first, so the first disjunct in the assert will not hold 7214 // in that specific case (in which case the second disjunct 7215 // will hold). 7216 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7217 "Otherwise the size info doesn't change at this step"); 7218 NOT_PRODUCT( 7219 _numObjectsAlreadyFree++; 7220 _numWordsAlreadyFree += res; 7221 ) 7222 NOT_PRODUCT(_last_fc = fc;) 7223 } else if (!_bitMap->isMarked(addr)) { 7224 // Chunk is fresh garbage 7225 res = do_garbage_chunk(fc); 7226 debug_only(_sp->verifyFreeLists()); 7227 NOT_PRODUCT( 7228 _numObjectsFreed++; 7229 _numWordsFreed += res; 7230 ) 7231 } else { 7232 // Chunk that is alive. 7233 res = do_live_chunk(fc); 7234 debug_only(_sp->verifyFreeLists()); 7235 NOT_PRODUCT( 7236 _numObjectsLive++; 7237 _numWordsLive += res; 7238 ) 7239 } 7240 return res; 7241} 7242 7243// For the smart allocation, record following 7244// split deaths - a free chunk is removed from its free list because 7245// it is being split into two or more chunks. 7246// split birth - a free chunk is being added to its free list because 7247// a larger free chunk has been split and resulted in this free chunk. 7248// coal death - a free chunk is being removed from its free list because 7249// it is being coalesced into a large free chunk. 7250// coal birth - a free chunk is being added to its free list because 7251// it was created when two or more free chunks where coalesced into 7252// this free chunk. 7253// 7254// These statistics are used to determine the desired number of free 7255// chunks of a given size. The desired number is chosen to be relative 7256// to the end of a CMS sweep. The desired number at the end of a sweep 7257// is the 7258// count-at-end-of-previous-sweep (an amount that was enough) 7259// - count-at-beginning-of-current-sweep (the excess) 7260// + split-births (gains in this size during interval) 7261// - split-deaths (demands on this size during interval) 7262// where the interval is from the end of one sweep to the end of the 7263// next. 7264// 7265// When sweeping the sweeper maintains an accumulated chunk which is 7266// the chunk that is made up of chunks that have been coalesced. That 7267// will be termed the left-hand chunk. A new chunk of garbage that 7268// is being considered for coalescing will be referred to as the 7269// right-hand chunk. 7270// 7271// When making a decision on whether to coalesce a right-hand chunk with 7272// the current left-hand chunk, the current count vs. the desired count 7273// of the left-hand chunk is considered. Also if the right-hand chunk 7274// is near the large chunk at the end of the heap (see 7275// ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7276// left-hand chunk is coalesced. 7277// 7278// When making a decision about whether to split a chunk, the desired count 7279// vs. the current count of the candidate to be split is also considered. 7280// If the candidate is underpopulated (currently fewer chunks than desired) 7281// a chunk of an overpopulated (currently more chunks than desired) size may 7282// be chosen. The "hint" associated with a free list, if non-null, points 7283// to a free list which may be overpopulated. 7284// 7285 7286void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7287 const size_t size = fc->size(); 7288 // Chunks that cannot be coalesced are not in the 7289 // free lists. 7290 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7291 assert(_sp->verify_chunk_in_free_list(fc), 7292 "free chunk should be in free lists"); 7293 } 7294 // a chunk that is already free, should not have been 7295 // marked in the bit map 7296 HeapWord* const addr = (HeapWord*) fc; 7297 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7298 // Verify that the bit map has no bits marked between 7299 // addr and purported end of this block. 7300 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7301 7302 // Some chunks cannot be coalesced under any circumstances. 7303 // See the definition of cantCoalesce(). 7304 if (!fc->cantCoalesce()) { 7305 // This chunk can potentially be coalesced. 7306 // All the work is done in 7307 do_post_free_or_garbage_chunk(fc, size); 7308 // Note that if the chunk is not coalescable (the else arm 7309 // below), we unconditionally flush, without needing to do 7310 // a "lookahead," as we do below. 7311 if (inFreeRange()) lookahead_and_flush(fc, size); 7312 } else { 7313 // Code path common to both original and adaptive free lists. 7314 7315 // cant coalesce with previous block; this should be treated 7316 // as the end of a free run if any 7317 if (inFreeRange()) { 7318 // we kicked some butt; time to pick up the garbage 7319 assert(freeFinger() < addr, "freeFinger points too high"); 7320 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7321 } 7322 // else, nothing to do, just continue 7323 } 7324} 7325 7326size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7327 // This is a chunk of garbage. It is not in any free list. 7328 // Add it to a free list or let it possibly be coalesced into 7329 // a larger chunk. 7330 HeapWord* const addr = (HeapWord*) fc; 7331 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7332 7333 // Verify that the bit map has no bits marked between 7334 // addr and purported end of just dead object. 7335 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7336 do_post_free_or_garbage_chunk(fc, size); 7337 7338 assert(_limit >= addr + size, 7339 "A freshly garbage chunk can't possibly straddle over _limit"); 7340 if (inFreeRange()) lookahead_and_flush(fc, size); 7341 return size; 7342} 7343 7344size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7345 HeapWord* addr = (HeapWord*) fc; 7346 // The sweeper has just found a live object. Return any accumulated 7347 // left hand chunk to the free lists. 7348 if (inFreeRange()) { 7349 assert(freeFinger() < addr, "freeFinger points too high"); 7350 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7351 } 7352 7353 // This object is live: we'd normally expect this to be 7354 // an oop, and like to assert the following: 7355 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7356 // However, as we commented above, this may be an object whose 7357 // header hasn't yet been initialized. 7358 size_t size; 7359 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7360 if (_bitMap->isMarked(addr + 1)) { 7361 // Determine the size from the bit map, rather than trying to 7362 // compute it from the object header. 7363 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7364 size = pointer_delta(nextOneAddr + 1, addr); 7365 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7366 "alignment problem"); 7367 7368#ifdef ASSERT 7369 if (oop(addr)->klass_or_null() != NULL) { 7370 // Ignore mark word because we are running concurrent with mutators 7371 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7372 assert(size == 7373 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7374 "P-mark and computed size do not agree"); 7375 } 7376#endif 7377 7378 } else { 7379 // This should be an initialized object that's alive. 7380 assert(oop(addr)->klass_or_null() != NULL, 7381 "Should be an initialized object"); 7382 // Ignore mark word because we are running concurrent with mutators 7383 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7384 // Verify that the bit map has no bits marked between 7385 // addr and purported end of this block. 7386 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7387 assert(size >= 3, "Necessary for Printezis marks to work"); 7388 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7389 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7390 } 7391 return size; 7392} 7393 7394void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7395 size_t chunkSize) { 7396 // do_post_free_or_garbage_chunk() should only be called in the case 7397 // of the adaptive free list allocator. 7398 const bool fcInFreeLists = fc->is_free(); 7399 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7400 if (CMSTestInFreeList && fcInFreeLists) { 7401 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7402 } 7403 7404 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7405 7406 HeapWord* const fc_addr = (HeapWord*) fc; 7407 7408 bool coalesce = false; 7409 const size_t left = pointer_delta(fc_addr, freeFinger()); 7410 const size_t right = chunkSize; 7411 switch (FLSCoalescePolicy) { 7412 // numeric value forms a coalition aggressiveness metric 7413 case 0: { // never coalesce 7414 coalesce = false; 7415 break; 7416 } 7417 case 1: { // coalesce if left & right chunks on overpopulated lists 7418 coalesce = _sp->coalOverPopulated(left) && 7419 _sp->coalOverPopulated(right); 7420 break; 7421 } 7422 case 2: { // coalesce if left chunk on overpopulated list (default) 7423 coalesce = _sp->coalOverPopulated(left); 7424 break; 7425 } 7426 case 3: { // coalesce if left OR right chunk on overpopulated list 7427 coalesce = _sp->coalOverPopulated(left) || 7428 _sp->coalOverPopulated(right); 7429 break; 7430 } 7431 case 4: { // always coalesce 7432 coalesce = true; 7433 break; 7434 } 7435 default: 7436 ShouldNotReachHere(); 7437 } 7438 7439 // Should the current free range be coalesced? 7440 // If the chunk is in a free range and either we decided to coalesce above 7441 // or the chunk is near the large block at the end of the heap 7442 // (isNearLargestChunk() returns true), then coalesce this chunk. 7443 const bool doCoalesce = inFreeRange() 7444 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7445 if (doCoalesce) { 7446 // Coalesce the current free range on the left with the new 7447 // chunk on the right. If either is on a free list, 7448 // it must be removed from the list and stashed in the closure. 7449 if (freeRangeInFreeLists()) { 7450 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7451 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7452 "Size of free range is inconsistent with chunk size."); 7453 if (CMSTestInFreeList) { 7454 assert(_sp->verify_chunk_in_free_list(ffc), 7455 "Chunk is not in free lists"); 7456 } 7457 _sp->coalDeath(ffc->size()); 7458 _sp->removeFreeChunkFromFreeLists(ffc); 7459 set_freeRangeInFreeLists(false); 7460 } 7461 if (fcInFreeLists) { 7462 _sp->coalDeath(chunkSize); 7463 assert(fc->size() == chunkSize, 7464 "The chunk has the wrong size or is not in the free lists"); 7465 _sp->removeFreeChunkFromFreeLists(fc); 7466 } 7467 set_lastFreeRangeCoalesced(true); 7468 print_free_block_coalesced(fc); 7469 } else { // not in a free range and/or should not coalesce 7470 // Return the current free range and start a new one. 7471 if (inFreeRange()) { 7472 // In a free range but cannot coalesce with the right hand chunk. 7473 // Put the current free range into the free lists. 7474 flush_cur_free_chunk(freeFinger(), 7475 pointer_delta(fc_addr, freeFinger())); 7476 } 7477 // Set up for new free range. Pass along whether the right hand 7478 // chunk is in the free lists. 7479 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7480 } 7481} 7482 7483// Lookahead flush: 7484// If we are tracking a free range, and this is the last chunk that 7485// we'll look at because its end crosses past _limit, we'll preemptively 7486// flush it along with any free range we may be holding on to. Note that 7487// this can be the case only for an already free or freshly garbage 7488// chunk. If this block is an object, it can never straddle 7489// over _limit. The "straddling" occurs when _limit is set at 7490// the previous end of the space when this cycle started, and 7491// a subsequent heap expansion caused the previously co-terminal 7492// free block to be coalesced with the newly expanded portion, 7493// thus rendering _limit a non-block-boundary making it dangerous 7494// for the sweeper to step over and examine. 7495void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7496 assert(inFreeRange(), "Should only be called if currently in a free range."); 7497 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7498 assert(_sp->used_region().contains(eob - 1), 7499 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7500 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7501 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7502 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7503 if (eob >= _limit) { 7504 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7505 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7506 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7507 "[" PTR_FORMAT "," PTR_FORMAT ")", 7508 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7509 // Return the storage we are tracking back into the free lists. 7510 log_develop_trace(gc, sweep)("Flushing ... "); 7511 assert(freeFinger() < eob, "Error"); 7512 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7513 } 7514} 7515 7516void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7517 assert(inFreeRange(), "Should only be called if currently in a free range."); 7518 assert(size > 0, 7519 "A zero sized chunk cannot be added to the free lists."); 7520 if (!freeRangeInFreeLists()) { 7521 if (CMSTestInFreeList) { 7522 FreeChunk* fc = (FreeChunk*) chunk; 7523 fc->set_size(size); 7524 assert(!_sp->verify_chunk_in_free_list(fc), 7525 "chunk should not be in free lists yet"); 7526 } 7527 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7528 // A new free range is going to be starting. The current 7529 // free range has not been added to the free lists yet or 7530 // was removed so add it back. 7531 // If the current free range was coalesced, then the death 7532 // of the free range was recorded. Record a birth now. 7533 if (lastFreeRangeCoalesced()) { 7534 _sp->coalBirth(size); 7535 } 7536 _sp->addChunkAndRepairOffsetTable(chunk, size, 7537 lastFreeRangeCoalesced()); 7538 } else { 7539 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7540 } 7541 set_inFreeRange(false); 7542 set_freeRangeInFreeLists(false); 7543} 7544 7545// We take a break if we've been at this for a while, 7546// so as to avoid monopolizing the locks involved. 7547void SweepClosure::do_yield_work(HeapWord* addr) { 7548 // Return current free chunk being used for coalescing (if any) 7549 // to the appropriate freelist. After yielding, the next 7550 // free block encountered will start a coalescing range of 7551 // free blocks. If the next free block is adjacent to the 7552 // chunk just flushed, they will need to wait for the next 7553 // sweep to be coalesced. 7554 if (inFreeRange()) { 7555 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7556 } 7557 7558 // First give up the locks, then yield, then re-lock. 7559 // We should probably use a constructor/destructor idiom to 7560 // do this unlock/lock or modify the MutexUnlocker class to 7561 // serve our purpose. XXX 7562 assert_lock_strong(_bitMap->lock()); 7563 assert_lock_strong(_freelistLock); 7564 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7565 "CMS thread should hold CMS token"); 7566 _bitMap->lock()->unlock(); 7567 _freelistLock->unlock(); 7568 ConcurrentMarkSweepThread::desynchronize(true); 7569 _collector->stopTimer(); 7570 _collector->incrementYields(); 7571 7572 // See the comment in coordinator_yield() 7573 for (unsigned i = 0; i < CMSYieldSleepCount && 7574 ConcurrentMarkSweepThread::should_yield() && 7575 !CMSCollector::foregroundGCIsActive(); ++i) { 7576 os::sleep(Thread::current(), 1, false); 7577 } 7578 7579 ConcurrentMarkSweepThread::synchronize(true); 7580 _freelistLock->lock(); 7581 _bitMap->lock()->lock_without_safepoint_check(); 7582 _collector->startTimer(); 7583} 7584 7585#ifndef PRODUCT 7586// This is actually very useful in a product build if it can 7587// be called from the debugger. Compile it into the product 7588// as needed. 7589bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7590 return debug_cms_space->verify_chunk_in_free_list(fc); 7591} 7592#endif 7593 7594void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7595 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7596 p2i(fc), fc->size()); 7597} 7598 7599// CMSIsAliveClosure 7600bool CMSIsAliveClosure::do_object_b(oop obj) { 7601 HeapWord* addr = (HeapWord*)obj; 7602 return addr != NULL && 7603 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7604} 7605 7606 7607CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7608 MemRegion span, 7609 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7610 bool cpc): 7611 _collector(collector), 7612 _span(span), 7613 _bit_map(bit_map), 7614 _mark_stack(mark_stack), 7615 _concurrent_precleaning(cpc) { 7616 assert(!_span.is_empty(), "Empty span could spell trouble"); 7617} 7618 7619 7620// CMSKeepAliveClosure: the serial version 7621void CMSKeepAliveClosure::do_oop(oop obj) { 7622 HeapWord* addr = (HeapWord*)obj; 7623 if (_span.contains(addr) && 7624 !_bit_map->isMarked(addr)) { 7625 _bit_map->mark(addr); 7626 bool simulate_overflow = false; 7627 NOT_PRODUCT( 7628 if (CMSMarkStackOverflowALot && 7629 _collector->simulate_overflow()) { 7630 // simulate a stack overflow 7631 simulate_overflow = true; 7632 } 7633 ) 7634 if (simulate_overflow || !_mark_stack->push(obj)) { 7635 if (_concurrent_precleaning) { 7636 // We dirty the overflown object and let the remark 7637 // phase deal with it. 7638 assert(_collector->overflow_list_is_empty(), "Error"); 7639 // In the case of object arrays, we need to dirty all of 7640 // the cards that the object spans. No locking or atomics 7641 // are needed since no one else can be mutating the mod union 7642 // table. 7643 if (obj->is_objArray()) { 7644 size_t sz = obj->size(); 7645 HeapWord* end_card_addr = 7646 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7647 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7648 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7649 _collector->_modUnionTable.mark_range(redirty_range); 7650 } else { 7651 _collector->_modUnionTable.mark(addr); 7652 } 7653 _collector->_ser_kac_preclean_ovflw++; 7654 } else { 7655 _collector->push_on_overflow_list(obj); 7656 _collector->_ser_kac_ovflw++; 7657 } 7658 } 7659 } 7660} 7661 7662void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7663void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7664 7665// CMSParKeepAliveClosure: a parallel version of the above. 7666// The work queues are private to each closure (thread), 7667// but (may be) available for stealing by other threads. 7668void CMSParKeepAliveClosure::do_oop(oop obj) { 7669 HeapWord* addr = (HeapWord*)obj; 7670 if (_span.contains(addr) && 7671 !_bit_map->isMarked(addr)) { 7672 // In general, during recursive tracing, several threads 7673 // may be concurrently getting here; the first one to 7674 // "tag" it, claims it. 7675 if (_bit_map->par_mark(addr)) { 7676 bool res = _work_queue->push(obj); 7677 assert(res, "Low water mark should be much less than capacity"); 7678 // Do a recursive trim in the hope that this will keep 7679 // stack usage lower, but leave some oops for potential stealers 7680 trim_queue(_low_water_mark); 7681 } // Else, another thread got there first 7682 } 7683} 7684 7685void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7686void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7687 7688void CMSParKeepAliveClosure::trim_queue(uint max) { 7689 while (_work_queue->size() > max) { 7690 oop new_oop; 7691 if (_work_queue->pop_local(new_oop)) { 7692 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7693 assert(_bit_map->isMarked((HeapWord*)new_oop), 7694 "no white objects on this stack!"); 7695 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7696 // iterate over the oops in this oop, marking and pushing 7697 // the ones in CMS heap (i.e. in _span). 7698 new_oop->oop_iterate(&_mark_and_push); 7699 } 7700 } 7701} 7702 7703CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7704 CMSCollector* collector, 7705 MemRegion span, CMSBitMap* bit_map, 7706 OopTaskQueue* work_queue): 7707 _collector(collector), 7708 _span(span), 7709 _bit_map(bit_map), 7710 _work_queue(work_queue) { } 7711 7712void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7713 HeapWord* addr = (HeapWord*)obj; 7714 if (_span.contains(addr) && 7715 !_bit_map->isMarked(addr)) { 7716 if (_bit_map->par_mark(addr)) { 7717 bool simulate_overflow = false; 7718 NOT_PRODUCT( 7719 if (CMSMarkStackOverflowALot && 7720 _collector->par_simulate_overflow()) { 7721 // simulate a stack overflow 7722 simulate_overflow = true; 7723 } 7724 ) 7725 if (simulate_overflow || !_work_queue->push(obj)) { 7726 _collector->par_push_on_overflow_list(obj); 7727 _collector->_par_kac_ovflw++; 7728 } 7729 } // Else another thread got there already 7730 } 7731} 7732 7733void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7734void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7735 7736////////////////////////////////////////////////////////////////// 7737// CMSExpansionCause ///////////////////////////// 7738////////////////////////////////////////////////////////////////// 7739const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7740 switch (cause) { 7741 case _no_expansion: 7742 return "No expansion"; 7743 case _satisfy_free_ratio: 7744 return "Free ratio"; 7745 case _satisfy_promotion: 7746 return "Satisfy promotion"; 7747 case _satisfy_allocation: 7748 return "allocation"; 7749 case _allocate_par_lab: 7750 return "Par LAB"; 7751 case _allocate_par_spooling_space: 7752 return "Par Spooling Space"; 7753 case _adaptive_size_policy: 7754 return "Ergonomics"; 7755 default: 7756 return "unknown"; 7757 } 7758} 7759 7760void CMSDrainMarkingStackClosure::do_void() { 7761 // the max number to take from overflow list at a time 7762 const size_t num = _mark_stack->capacity()/4; 7763 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7764 "Overflow list should be NULL during concurrent phases"); 7765 while (!_mark_stack->isEmpty() || 7766 // if stack is empty, check the overflow list 7767 _collector->take_from_overflow_list(num, _mark_stack)) { 7768 oop obj = _mark_stack->pop(); 7769 HeapWord* addr = (HeapWord*)obj; 7770 assert(_span.contains(addr), "Should be within span"); 7771 assert(_bit_map->isMarked(addr), "Should be marked"); 7772 assert(obj->is_oop(), "Should be an oop"); 7773 obj->oop_iterate(_keep_alive); 7774 } 7775} 7776 7777void CMSParDrainMarkingStackClosure::do_void() { 7778 // drain queue 7779 trim_queue(0); 7780} 7781 7782// Trim our work_queue so its length is below max at return 7783void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7784 while (_work_queue->size() > max) { 7785 oop new_oop; 7786 if (_work_queue->pop_local(new_oop)) { 7787 assert(new_oop->is_oop(), "Expected an oop"); 7788 assert(_bit_map->isMarked((HeapWord*)new_oop), 7789 "no white objects on this stack!"); 7790 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7791 // iterate over the oops in this oop, marking and pushing 7792 // the ones in CMS heap (i.e. in _span). 7793 new_oop->oop_iterate(&_mark_and_push); 7794 } 7795 } 7796} 7797 7798//////////////////////////////////////////////////////////////////// 7799// Support for Marking Stack Overflow list handling and related code 7800//////////////////////////////////////////////////////////////////// 7801// Much of the following code is similar in shape and spirit to the 7802// code used in ParNewGC. We should try and share that code 7803// as much as possible in the future. 7804 7805#ifndef PRODUCT 7806// Debugging support for CMSStackOverflowALot 7807 7808// It's OK to call this multi-threaded; the worst thing 7809// that can happen is that we'll get a bunch of closely 7810// spaced simulated overflows, but that's OK, in fact 7811// probably good as it would exercise the overflow code 7812// under contention. 7813bool CMSCollector::simulate_overflow() { 7814 if (_overflow_counter-- <= 0) { // just being defensive 7815 _overflow_counter = CMSMarkStackOverflowInterval; 7816 return true; 7817 } else { 7818 return false; 7819 } 7820} 7821 7822bool CMSCollector::par_simulate_overflow() { 7823 return simulate_overflow(); 7824} 7825#endif 7826 7827// Single-threaded 7828bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7829 assert(stack->isEmpty(), "Expected precondition"); 7830 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7831 size_t i = num; 7832 oop cur = _overflow_list; 7833 const markOop proto = markOopDesc::prototype(); 7834 NOT_PRODUCT(ssize_t n = 0;) 7835 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7836 next = oop(cur->mark()); 7837 cur->set_mark(proto); // until proven otherwise 7838 assert(cur->is_oop(), "Should be an oop"); 7839 bool res = stack->push(cur); 7840 assert(res, "Bit off more than can chew?"); 7841 NOT_PRODUCT(n++;) 7842 } 7843 _overflow_list = cur; 7844#ifndef PRODUCT 7845 assert(_num_par_pushes >= n, "Too many pops?"); 7846 _num_par_pushes -=n; 7847#endif 7848 return !stack->isEmpty(); 7849} 7850 7851#define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7852// (MT-safe) Get a prefix of at most "num" from the list. 7853// The overflow list is chained through the mark word of 7854// each object in the list. We fetch the entire list, 7855// break off a prefix of the right size and return the 7856// remainder. If other threads try to take objects from 7857// the overflow list at that time, they will wait for 7858// some time to see if data becomes available. If (and 7859// only if) another thread places one or more object(s) 7860// on the global list before we have returned the suffix 7861// to the global list, we will walk down our local list 7862// to find its end and append the global list to 7863// our suffix before returning it. This suffix walk can 7864// prove to be expensive (quadratic in the amount of traffic) 7865// when there are many objects in the overflow list and 7866// there is much producer-consumer contention on the list. 7867// *NOTE*: The overflow list manipulation code here and 7868// in ParNewGeneration:: are very similar in shape, 7869// except that in the ParNew case we use the old (from/eden) 7870// copy of the object to thread the list via its klass word. 7871// Because of the common code, if you make any changes in 7872// the code below, please check the ParNew version to see if 7873// similar changes might be needed. 7874// CR 6797058 has been filed to consolidate the common code. 7875bool CMSCollector::par_take_from_overflow_list(size_t num, 7876 OopTaskQueue* work_q, 7877 int no_of_gc_threads) { 7878 assert(work_q->size() == 0, "First empty local work queue"); 7879 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7880 if (_overflow_list == NULL) { 7881 return false; 7882 } 7883 // Grab the entire list; we'll put back a suffix 7884 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7885 Thread* tid = Thread::current(); 7886 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7887 // set to ParallelGCThreads. 7888 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7889 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7890 // If the list is busy, we spin for a short while, 7891 // sleeping between attempts to get the list. 7892 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7893 os::sleep(tid, sleep_time_millis, false); 7894 if (_overflow_list == NULL) { 7895 // Nothing left to take 7896 return false; 7897 } else if (_overflow_list != BUSY) { 7898 // Try and grab the prefix 7899 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7900 } 7901 } 7902 // If the list was found to be empty, or we spun long 7903 // enough, we give up and return empty-handed. If we leave 7904 // the list in the BUSY state below, it must be the case that 7905 // some other thread holds the overflow list and will set it 7906 // to a non-BUSY state in the future. 7907 if (prefix == NULL || prefix == BUSY) { 7908 // Nothing to take or waited long enough 7909 if (prefix == NULL) { 7910 // Write back the NULL in case we overwrote it with BUSY above 7911 // and it is still the same value. 7912 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7913 } 7914 return false; 7915 } 7916 assert(prefix != NULL && prefix != BUSY, "Error"); 7917 size_t i = num; 7918 oop cur = prefix; 7919 // Walk down the first "num" objects, unless we reach the end. 7920 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7921 if (cur->mark() == NULL) { 7922 // We have "num" or fewer elements in the list, so there 7923 // is nothing to return to the global list. 7924 // Write back the NULL in lieu of the BUSY we wrote 7925 // above, if it is still the same value. 7926 if (_overflow_list == BUSY) { 7927 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7928 } 7929 } else { 7930 // Chop off the suffix and return it to the global list. 7931 assert(cur->mark() != BUSY, "Error"); 7932 oop suffix_head = cur->mark(); // suffix will be put back on global list 7933 cur->set_mark(NULL); // break off suffix 7934 // It's possible that the list is still in the empty(busy) state 7935 // we left it in a short while ago; in that case we may be 7936 // able to place back the suffix without incurring the cost 7937 // of a walk down the list. 7938 oop observed_overflow_list = _overflow_list; 7939 oop cur_overflow_list = observed_overflow_list; 7940 bool attached = false; 7941 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7942 observed_overflow_list = 7943 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7944 if (cur_overflow_list == observed_overflow_list) { 7945 attached = true; 7946 break; 7947 } else cur_overflow_list = observed_overflow_list; 7948 } 7949 if (!attached) { 7950 // Too bad, someone else sneaked in (at least) an element; we'll need 7951 // to do a splice. Find tail of suffix so we can prepend suffix to global 7952 // list. 7953 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7954 oop suffix_tail = cur; 7955 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7956 "Tautology"); 7957 observed_overflow_list = _overflow_list; 7958 do { 7959 cur_overflow_list = observed_overflow_list; 7960 if (cur_overflow_list != BUSY) { 7961 // Do the splice ... 7962 suffix_tail->set_mark(markOop(cur_overflow_list)); 7963 } else { // cur_overflow_list == BUSY 7964 suffix_tail->set_mark(NULL); 7965 } 7966 // ... and try to place spliced list back on overflow_list ... 7967 observed_overflow_list = 7968 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7969 } while (cur_overflow_list != observed_overflow_list); 7970 // ... until we have succeeded in doing so. 7971 } 7972 } 7973 7974 // Push the prefix elements on work_q 7975 assert(prefix != NULL, "control point invariant"); 7976 const markOop proto = markOopDesc::prototype(); 7977 oop next; 7978 NOT_PRODUCT(ssize_t n = 0;) 7979 for (cur = prefix; cur != NULL; cur = next) { 7980 next = oop(cur->mark()); 7981 cur->set_mark(proto); // until proven otherwise 7982 assert(cur->is_oop(), "Should be an oop"); 7983 bool res = work_q->push(cur); 7984 assert(res, "Bit off more than we can chew?"); 7985 NOT_PRODUCT(n++;) 7986 } 7987#ifndef PRODUCT 7988 assert(_num_par_pushes >= n, "Too many pops?"); 7989 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7990#endif 7991 return true; 7992} 7993 7994// Single-threaded 7995void CMSCollector::push_on_overflow_list(oop p) { 7996 NOT_PRODUCT(_num_par_pushes++;) 7997 assert(p->is_oop(), "Not an oop"); 7998 preserve_mark_if_necessary(p); 7999 p->set_mark((markOop)_overflow_list); 8000 _overflow_list = p; 8001} 8002 8003// Multi-threaded; use CAS to prepend to overflow list 8004void CMSCollector::par_push_on_overflow_list(oop p) { 8005 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8006 assert(p->is_oop(), "Not an oop"); 8007 par_preserve_mark_if_necessary(p); 8008 oop observed_overflow_list = _overflow_list; 8009 oop cur_overflow_list; 8010 do { 8011 cur_overflow_list = observed_overflow_list; 8012 if (cur_overflow_list != BUSY) { 8013 p->set_mark(markOop(cur_overflow_list)); 8014 } else { 8015 p->set_mark(NULL); 8016 } 8017 observed_overflow_list = 8018 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8019 } while (cur_overflow_list != observed_overflow_list); 8020} 8021#undef BUSY 8022 8023// Single threaded 8024// General Note on GrowableArray: pushes may silently fail 8025// because we are (temporarily) out of C-heap for expanding 8026// the stack. The problem is quite ubiquitous and affects 8027// a lot of code in the JVM. The prudent thing for GrowableArray 8028// to do (for now) is to exit with an error. However, that may 8029// be too draconian in some cases because the caller may be 8030// able to recover without much harm. For such cases, we 8031// should probably introduce a "soft_push" method which returns 8032// an indication of success or failure with the assumption that 8033// the caller may be able to recover from a failure; code in 8034// the VM can then be changed, incrementally, to deal with such 8035// failures where possible, thus, incrementally hardening the VM 8036// in such low resource situations. 8037void CMSCollector::preserve_mark_work(oop p, markOop m) { 8038 _preserved_oop_stack.push(p); 8039 _preserved_mark_stack.push(m); 8040 assert(m == p->mark(), "Mark word changed"); 8041 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8042 "bijection"); 8043} 8044 8045// Single threaded 8046void CMSCollector::preserve_mark_if_necessary(oop p) { 8047 markOop m = p->mark(); 8048 if (m->must_be_preserved(p)) { 8049 preserve_mark_work(p, m); 8050 } 8051} 8052 8053void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8054 markOop m = p->mark(); 8055 if (m->must_be_preserved(p)) { 8056 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8057 // Even though we read the mark word without holding 8058 // the lock, we are assured that it will not change 8059 // because we "own" this oop, so no other thread can 8060 // be trying to push it on the overflow list; see 8061 // the assertion in preserve_mark_work() that checks 8062 // that m == p->mark(). 8063 preserve_mark_work(p, m); 8064 } 8065} 8066 8067// We should be able to do this multi-threaded, 8068// a chunk of stack being a task (this is 8069// correct because each oop only ever appears 8070// once in the overflow list. However, it's 8071// not very easy to completely overlap this with 8072// other operations, so will generally not be done 8073// until all work's been completed. Because we 8074// expect the preserved oop stack (set) to be small, 8075// it's probably fine to do this single-threaded. 8076// We can explore cleverer concurrent/overlapped/parallel 8077// processing of preserved marks if we feel the 8078// need for this in the future. Stack overflow should 8079// be so rare in practice and, when it happens, its 8080// effect on performance so great that this will 8081// likely just be in the noise anyway. 8082void CMSCollector::restore_preserved_marks_if_any() { 8083 assert(SafepointSynchronize::is_at_safepoint(), 8084 "world should be stopped"); 8085 assert(Thread::current()->is_ConcurrentGC_thread() || 8086 Thread::current()->is_VM_thread(), 8087 "should be single-threaded"); 8088 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8089 "bijection"); 8090 8091 while (!_preserved_oop_stack.is_empty()) { 8092 oop p = _preserved_oop_stack.pop(); 8093 assert(p->is_oop(), "Should be an oop"); 8094 assert(_span.contains(p), "oop should be in _span"); 8095 assert(p->mark() == markOopDesc::prototype(), 8096 "Set when taken from overflow list"); 8097 markOop m = _preserved_mark_stack.pop(); 8098 p->set_mark(m); 8099 } 8100 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8101 "stacks were cleared above"); 8102} 8103 8104#ifndef PRODUCT 8105bool CMSCollector::no_preserved_marks() const { 8106 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8107} 8108#endif 8109 8110// Transfer some number of overflown objects to usual marking 8111// stack. Return true if some objects were transferred. 8112bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8113 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8114 (size_t)ParGCDesiredObjsFromOverflowList); 8115 8116 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8117 assert(_collector->overflow_list_is_empty() || res, 8118 "If list is not empty, we should have taken something"); 8119 assert(!res || !_mark_stack->isEmpty(), 8120 "If we took something, it should now be on our stack"); 8121 return res; 8122} 8123 8124size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8125 size_t res = _sp->block_size_no_stall(addr, _collector); 8126 if (_sp->block_is_obj(addr)) { 8127 if (_live_bit_map->isMarked(addr)) { 8128 // It can't have been dead in a previous cycle 8129 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8130 } else { 8131 _dead_bit_map->mark(addr); // mark the dead object 8132 } 8133 } 8134 // Could be 0, if the block size could not be computed without stalling. 8135 return res; 8136} 8137 8138TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8139 8140 switch (phase) { 8141 case CMSCollector::InitialMarking: 8142 initialize(true /* fullGC */ , 8143 cause /* cause of the GC */, 8144 true /* recordGCBeginTime */, 8145 true /* recordPreGCUsage */, 8146 false /* recordPeakUsage */, 8147 false /* recordPostGCusage */, 8148 true /* recordAccumulatedGCTime */, 8149 false /* recordGCEndTime */, 8150 false /* countCollection */ ); 8151 break; 8152 8153 case CMSCollector::FinalMarking: 8154 initialize(true /* fullGC */ , 8155 cause /* cause of the GC */, 8156 false /* recordGCBeginTime */, 8157 false /* recordPreGCUsage */, 8158 false /* recordPeakUsage */, 8159 false /* recordPostGCusage */, 8160 true /* recordAccumulatedGCTime */, 8161 false /* recordGCEndTime */, 8162 false /* countCollection */ ); 8163 break; 8164 8165 case CMSCollector::Sweeping: 8166 initialize(true /* fullGC */ , 8167 cause /* cause of the GC */, 8168 false /* recordGCBeginTime */, 8169 false /* recordPreGCUsage */, 8170 true /* recordPeakUsage */, 8171 true /* recordPostGCusage */, 8172 false /* recordAccumulatedGCTime */, 8173 true /* recordGCEndTime */, 8174 true /* countCollection */ ); 8175 break; 8176 8177 default: 8178 ShouldNotReachHere(); 8179 } 8180} 8181