1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardGeneration.inline.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/tenuredGeneration.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "prims/jvmtiExport.hpp" 59 #include "runtime/atomic.inline.hpp" 60 #include "runtime/globals_extension.hpp" 61 #include "runtime/handles.inline.hpp" 62 #include "runtime/java.hpp" 63 #include "runtime/orderAccess.inline.hpp" 64 #include "runtime/vmThread.hpp" 65 #include "services/memoryService.hpp" 66 #include "services/runtimeService.hpp" 67 68 // statics 69 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 70 bool CMSCollector::_full_gc_requested = false; 71 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 72 73 ////////////////////////////////////////////////////////////////// 74 // In support of CMS/VM thread synchronization 75 ////////////////////////////////////////////////////////////////// 76 // We split use of the CGC_lock into 2 "levels". 77 // The low-level locking is of the usual CGC_lock monitor. We introduce 78 // a higher level "token" (hereafter "CMS token") built on top of the 79 // low level monitor (hereafter "CGC lock"). 80 // The token-passing protocol gives priority to the VM thread. The 81 // CMS-lock doesn't provide any fairness guarantees, but clients 82 // should ensure that it is only held for very short, bounded 83 // durations. 84 // 85 // When either of the CMS thread or the VM thread is involved in 86 // collection operations during which it does not want the other 87 // thread to interfere, it obtains the CMS token. 88 // 89 // If either thread tries to get the token while the other has 90 // it, that thread waits. However, if the VM thread and CMS thread 91 // both want the token, then the VM thread gets priority while the 92 // CMS thread waits. This ensures, for instance, that the "concurrent" 93 // phases of the CMS thread's work do not block out the VM thread 94 // for long periods of time as the CMS thread continues to hog 95 // the token. (See bug 4616232). 96 // 97 // The baton-passing functions are, however, controlled by the 98 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 99 // and here the low-level CMS lock, not the high level token, 100 // ensures mutual exclusion. 101 // 102 // Two important conditions that we have to satisfy: 103 // 1. if a thread does a low-level wait on the CMS lock, then it 104 // relinquishes the CMS token if it were holding that token 105 // when it acquired the low-level CMS lock. 106 // 2. any low-level notifications on the low-level lock 107 // should only be sent when a thread has relinquished the token. 108 // 109 // In the absence of either property, we'd have potential deadlock. 110 // 111 // We protect each of the CMS (concurrent and sequential) phases 112 // with the CMS _token_, not the CMS _lock_. 113 // 114 // The only code protected by CMS lock is the token acquisition code 115 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 116 // baton-passing code. 117 // 118 // Unfortunately, i couldn't come up with a good abstraction to factor and 119 // hide the naked CGC_lock manipulation in the baton-passing code 120 // further below. That's something we should try to do. Also, the proof 121 // of correctness of this 2-level locking scheme is far from obvious, 122 // and potentially quite slippery. We have an uneasy suspicion, for instance, 123 // that there may be a theoretical possibility of delay/starvation in the 124 // low-level lock/wait/notify scheme used for the baton-passing because of 125 // potential interference with the priority scheme embodied in the 126 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 127 // invocation further below and marked with "XXX 20011219YSR". 128 // Indeed, as we note elsewhere, this may become yet more slippery 129 // in the presence of multiple CMS and/or multiple VM threads. XXX 130 131 class CMSTokenSync: public StackObj { 132 private: 133 bool _is_cms_thread; 134 public: 135 CMSTokenSync(bool is_cms_thread): 136 _is_cms_thread(is_cms_thread) { 137 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 138 "Incorrect argument to constructor"); 139 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 140 } 141 142 ~CMSTokenSync() { 143 assert(_is_cms_thread ? 144 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 145 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 146 "Incorrect state"); 147 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 148 } 149 }; 150 151 // Convenience class that does a CMSTokenSync, and then acquires 152 // upto three locks. 153 class CMSTokenSyncWithLocks: public CMSTokenSync { 154 private: 155 // Note: locks are acquired in textual declaration order 156 // and released in the opposite order 157 MutexLockerEx _locker1, _locker2, _locker3; 158 public: 159 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 160 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 161 CMSTokenSync(is_cms_thread), 162 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 163 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 164 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 165 { } 166 }; 167 168 169 ////////////////////////////////////////////////////////////////// 170 // Concurrent Mark-Sweep Generation ///////////////////////////// 171 ////////////////////////////////////////////////////////////////// 172 173 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 174 175 // This struct contains per-thread things necessary to support parallel 176 // young-gen collection. 177 class CMSParGCThreadState: public CHeapObj<mtGC> { 178 public: 179 CFLS_LAB lab; 180 PromotionInfo promo; 181 182 // Constructor. 183 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 184 promo.setSpace(cfls); 185 } 186 }; 187 188 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 189 ReservedSpace rs, size_t initial_byte_size, int level, 190 CardTableRS* ct, bool use_adaptive_freelists, 191 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 192 CardGeneration(rs, initial_byte_size, level, ct), 193 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 194 _did_compact(false) 195 { 196 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 197 HeapWord* end = (HeapWord*) _virtual_space.high(); 198 199 _direct_allocated_words = 0; 200 NOT_PRODUCT( 201 _numObjectsPromoted = 0; 202 _numWordsPromoted = 0; 203 _numObjectsAllocated = 0; 204 _numWordsAllocated = 0; 205 ) 206 207 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 208 use_adaptive_freelists, 209 dictionaryChoice); 210 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 211 if (_cmsSpace == NULL) { 212 vm_exit_during_initialization( 213 "CompactibleFreeListSpace allocation failure"); 214 } 215 _cmsSpace->_gen = this; 216 217 _gc_stats = new CMSGCStats(); 218 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 220 // offsets match. The ability to tell free chunks from objects 221 // depends on this property. 222 debug_only( 223 FreeChunk* junk = NULL; 224 assert(UseCompressedClassPointers || 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 226 "Offset of FreeChunk::_prev within FreeChunk must match" 227 " that of OopDesc::_klass within OopDesc"); 228 ) 229 if (CollectedHeap::use_parallel_gc_threads()) { 230 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 231 _par_gc_thread_states = 232 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 233 if (_par_gc_thread_states == NULL) { 234 vm_exit_during_initialization("Could not allocate par gc structs"); 235 } 236 for (uint i = 0; i < ParallelGCThreads; i++) { 237 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 238 if (_par_gc_thread_states[i] == NULL) { 239 vm_exit_during_initialization("Could not allocate par gc structs"); 240 } 241 } 242 } else { 243 _par_gc_thread_states = NULL; 244 } 245 _incremental_collection_failed = false; 246 // The "dilatation_factor" is the expansion that can occur on 247 // account of the fact that the minimum object size in the CMS 248 // generation may be larger than that in, say, a contiguous young 249 // generation. 250 // Ideally, in the calculation below, we'd compute the dilatation 251 // factor as: MinChunkSize/(promoting_gen's min object size) 252 // Since we do not have such a general query interface for the 253 // promoting generation, we'll instead just use the minimum 254 // object size (which today is a header's worth of space); 255 // note that all arithmetic is in units of HeapWords. 256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 257 assert(_dilatation_factor >= 1.0, "from previous assert"); 258 } 259 260 261 // The field "_initiating_occupancy" represents the occupancy percentage 262 // at which we trigger a new collection cycle. Unless explicitly specified 263 // via CMSInitiatingOccupancyFraction (argument "io" below), it 264 // is calculated by: 265 // 266 // Let "f" be MinHeapFreeRatio in 267 // 268 // _initiating_occupancy = 100-f + 269 // f * (CMSTriggerRatio/100) 270 // where CMSTriggerRatio is the argument "tr" below. 271 // 272 // That is, if we assume the heap is at its desired maximum occupancy at the 273 // end of a collection, we let CMSTriggerRatio of the (purported) free 274 // space be allocated before initiating a new collection cycle. 275 // 276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 277 assert(io <= 100 && tr <= 100, "Check the arguments"); 278 if (io >= 0) { 279 _initiating_occupancy = (double)io / 100.0; 280 } else { 281 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 282 (double)(tr * MinHeapFreeRatio) / 100.0) 283 / 100.0; 284 } 285 } 286 287 void ConcurrentMarkSweepGeneration::ref_processor_init() { 288 assert(collector() != NULL, "no collector"); 289 collector()->ref_processor_init(); 290 } 291 292 void CMSCollector::ref_processor_init() { 293 if (_ref_processor == NULL) { 294 // Allocate and initialize a reference processor 295 _ref_processor = 296 new ReferenceProcessor(_span, // span 297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 298 (int) ParallelGCThreads, // mt processing degree 299 _cmsGen->refs_discovery_is_mt(), // mt discovery 300 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 302 &_is_alive_closure); // closure for liveness info 303 // Initialize the _ref_processor field of CMSGen 304 _cmsGen->set_ref_processor(_ref_processor); 305 306 } 307 } 308 309 AdaptiveSizePolicy* CMSCollector::size_policy() { 310 GenCollectedHeap* gch = GenCollectedHeap::heap(); 311 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 312 "Wrong type of heap"); 313 return gch->gen_policy()->size_policy(); 314 } 315 316 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 317 318 const char* gen_name = "old"; 319 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 320 321 // Generation Counters - generation 1, 1 subspace 322 _gen_counters = new GenerationCounters(gen_name, 1, 1, 323 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 324 325 _space_counters = new GSpaceCounters(gen_name, 0, 326 _virtual_space.reserved_size(), 327 this, _gen_counters); 328 } 329 330 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 331 _cms_gen(cms_gen) 332 { 333 assert(alpha <= 100, "bad value"); 334 _saved_alpha = alpha; 335 336 // Initialize the alphas to the bootstrap value of 100. 337 _gc0_alpha = _cms_alpha = 100; 338 339 _cms_begin_time.update(); 340 _cms_end_time.update(); 341 342 _gc0_duration = 0.0; 343 _gc0_period = 0.0; 344 _gc0_promoted = 0; 345 346 _cms_duration = 0.0; 347 _cms_period = 0.0; 348 _cms_allocated = 0; 349 350 _cms_used_at_gc0_begin = 0; 351 _cms_used_at_gc0_end = 0; 352 _allow_duty_cycle_reduction = false; 353 _valid_bits = 0; 354 } 355 356 double CMSStats::cms_free_adjustment_factor(size_t free) const { 357 // TBD: CR 6909490 358 return 1.0; 359 } 360 361 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 362 } 363 364 // If promotion failure handling is on use 365 // the padded average size of the promotion for each 366 // young generation collection. 367 double CMSStats::time_until_cms_gen_full() const { 368 size_t cms_free = _cms_gen->cmsSpace()->free(); 369 GenCollectedHeap* gch = GenCollectedHeap::heap(); 370 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 371 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 372 if (cms_free > expected_promotion) { 373 // Start a cms collection if there isn't enough space to promote 374 // for the next minor collection. Use the padded average as 375 // a safety factor. 376 cms_free -= expected_promotion; 377 378 // Adjust by the safety factor. 379 double cms_free_dbl = (double)cms_free; 380 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 381 // Apply a further correction factor which tries to adjust 382 // for recent occurance of concurrent mode failures. 383 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 384 cms_free_dbl = cms_free_dbl * cms_adjustment; 385 386 if (PrintGCDetails && Verbose) { 387 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 388 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 389 cms_free, expected_promotion); 390 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 391 cms_free_dbl, cms_consumption_rate() + 1.0); 392 } 393 // Add 1 in case the consumption rate goes to zero. 394 return cms_free_dbl / (cms_consumption_rate() + 1.0); 395 } 396 return 0.0; 397 } 398 399 // Compare the duration of the cms collection to the 400 // time remaining before the cms generation is empty. 401 // Note that the time from the start of the cms collection 402 // to the start of the cms sweep (less than the total 403 // duration of the cms collection) can be used. This 404 // has been tried and some applications experienced 405 // promotion failures early in execution. This was 406 // possibly because the averages were not accurate 407 // enough at the beginning. 408 double CMSStats::time_until_cms_start() const { 409 // We add "gc0_period" to the "work" calculation 410 // below because this query is done (mostly) at the 411 // end of a scavenge, so we need to conservatively 412 // account for that much possible delay 413 // in the query so as to avoid concurrent mode failures 414 // due to starting the collection just a wee bit too 415 // late. 416 double work = cms_duration() + gc0_period(); 417 double deadline = time_until_cms_gen_full(); 418 // If a concurrent mode failure occurred recently, we want to be 419 // more conservative and halve our expected time_until_cms_gen_full() 420 if (work > deadline) { 421 if (Verbose && PrintGCDetails) { 422 gclog_or_tty->print( 423 " CMSCollector: collect because of anticipated promotion " 424 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 425 gc0_period(), time_until_cms_gen_full()); 426 } 427 return 0.0; 428 } 429 return work - deadline; 430 } 431 432 #ifndef PRODUCT 433 void CMSStats::print_on(outputStream *st) const { 434 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 435 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 436 gc0_duration(), gc0_period(), gc0_promoted()); 437 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 438 cms_duration(), cms_period(), cms_allocated()); 439 st->print(",cms_since_beg=%g,cms_since_end=%g", 440 cms_time_since_begin(), cms_time_since_end()); 441 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 442 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 443 444 if (valid()) { 445 st->print(",promo_rate=%g,cms_alloc_rate=%g", 446 promotion_rate(), cms_allocation_rate()); 447 st->print(",cms_consumption_rate=%g,time_until_full=%g", 448 cms_consumption_rate(), time_until_cms_gen_full()); 449 } 450 st->print(" "); 451 } 452 #endif // #ifndef PRODUCT 453 454 CMSCollector::CollectorState CMSCollector::_collectorState = 455 CMSCollector::Idling; 456 bool CMSCollector::_foregroundGCIsActive = false; 457 bool CMSCollector::_foregroundGCShouldWait = false; 458 459 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 460 CardTableRS* ct, 461 ConcurrentMarkSweepPolicy* cp): 462 _cmsGen(cmsGen), 463 _ct(ct), 464 _ref_processor(NULL), // will be set later 465 _conc_workers(NULL), // may be set later 466 _abort_preclean(false), 467 _start_sampling(false), 468 _between_prologue_and_epilogue(false), 469 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 470 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 471 -1 /* lock-free */, "No_lock" /* dummy */), 472 _modUnionClosure(&_modUnionTable), 473 _modUnionClosurePar(&_modUnionTable), 474 // Adjust my span to cover old (cms) gen 475 _span(cmsGen->reserved()), 476 // Construct the is_alive_closure with _span & markBitMap 477 _is_alive_closure(_span, &_markBitMap), 478 _restart_addr(NULL), 479 _overflow_list(NULL), 480 _stats(cmsGen), 481 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 482 //verify that this lock should be acquired with safepoint check. 483 Monitor::_safepoint_check_sometimes)), 484 _eden_chunk_array(NULL), // may be set in ctor body 485 _eden_chunk_capacity(0), // -- ditto -- 486 _eden_chunk_index(0), // -- ditto -- 487 _survivor_plab_array(NULL), // -- ditto -- 488 _survivor_chunk_array(NULL), // -- ditto -- 489 _survivor_chunk_capacity(0), // -- ditto -- 490 _survivor_chunk_index(0), // -- ditto -- 491 _ser_pmc_preclean_ovflw(0), 492 _ser_kac_preclean_ovflw(0), 493 _ser_pmc_remark_ovflw(0), 494 _par_pmc_remark_ovflw(0), 495 _ser_kac_ovflw(0), 496 _par_kac_ovflw(0), 497 #ifndef PRODUCT 498 _num_par_pushes(0), 499 #endif 500 _collection_count_start(0), 501 _verifying(false), 502 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 503 _completed_initialization(false), 504 _collector_policy(cp), 505 _should_unload_classes(CMSClassUnloadingEnabled), 506 _concurrent_cycles_since_last_unload(0), 507 _roots_scanning_options(GenCollectedHeap::SO_None), 508 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 509 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 510 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 511 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 512 _cms_start_registered(false) 513 { 514 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 515 ExplicitGCInvokesConcurrent = true; 516 } 517 // Now expand the span and allocate the collection support structures 518 // (MUT, marking bit map etc.) to cover both generations subject to 519 // collection. 520 521 // For use by dirty card to oop closures. 522 _cmsGen->cmsSpace()->set_collector(this); 523 524 // Allocate MUT and marking bit map 525 { 526 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 527 if (!_markBitMap.allocate(_span)) { 528 warning("Failed to allocate CMS Bit Map"); 529 return; 530 } 531 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 532 } 533 { 534 _modUnionTable.allocate(_span); 535 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 536 } 537 538 if (!_markStack.allocate(MarkStackSize)) { 539 warning("Failed to allocate CMS Marking Stack"); 540 return; 541 } 542 543 // Support for multi-threaded concurrent phases 544 if (CMSConcurrentMTEnabled) { 545 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 546 // just for now 547 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 548 } 549 if (ConcGCThreads > 1) { 550 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 551 ConcGCThreads, true); 552 if (_conc_workers == NULL) { 553 warning("GC/CMS: _conc_workers allocation failure: " 554 "forcing -CMSConcurrentMTEnabled"); 555 CMSConcurrentMTEnabled = false; 556 } else { 557 _conc_workers->initialize_workers(); 558 } 559 } else { 560 CMSConcurrentMTEnabled = false; 561 } 562 } 563 if (!CMSConcurrentMTEnabled) { 564 ConcGCThreads = 0; 565 } else { 566 // Turn off CMSCleanOnEnter optimization temporarily for 567 // the MT case where it's not fixed yet; see 6178663. 568 CMSCleanOnEnter = false; 569 } 570 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 571 "Inconsistency"); 572 573 // Parallel task queues; these are shared for the 574 // concurrent and stop-world phases of CMS, but 575 // are not shared with parallel scavenge (ParNew). 576 { 577 uint i; 578 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 579 580 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 581 || ParallelRefProcEnabled) 582 && num_queues > 0) { 583 _task_queues = new OopTaskQueueSet(num_queues); 584 if (_task_queues == NULL) { 585 warning("task_queues allocation failure."); 586 return; 587 } 588 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 589 if (_hash_seed == NULL) { 590 warning("_hash_seed array allocation failure"); 591 return; 592 } 593 594 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 595 for (i = 0; i < num_queues; i++) { 596 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 597 if (q == NULL) { 598 warning("work_queue allocation failure."); 599 return; 600 } 601 _task_queues->register_queue(i, q); 602 } 603 for (i = 0; i < num_queues; i++) { 604 _task_queues->queue(i)->initialize(); 605 _hash_seed[i] = 17; // copied from ParNew 606 } 607 } 608 } 609 610 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 611 612 // Clip CMSBootstrapOccupancy between 0 and 100. 613 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 614 615 // Now tell CMS generations the identity of their collector 616 ConcurrentMarkSweepGeneration::set_collector(this); 617 618 // Create & start a CMS thread for this CMS collector 619 _cmsThread = ConcurrentMarkSweepThread::start(this); 620 assert(cmsThread() != NULL, "CMS Thread should have been created"); 621 assert(cmsThread()->collector() == this, 622 "CMS Thread should refer to this gen"); 623 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 624 625 // Support for parallelizing young gen rescan 626 GenCollectedHeap* gch = GenCollectedHeap::heap(); 627 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 628 _young_gen = (ParNewGeneration*)gch->young_gen(); 629 if (gch->supports_inline_contig_alloc()) { 630 _top_addr = gch->top_addr(); 631 _end_addr = gch->end_addr(); 632 assert(_young_gen != NULL, "no _young_gen"); 633 _eden_chunk_index = 0; 634 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 635 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 636 if (_eden_chunk_array == NULL) { 637 _eden_chunk_capacity = 0; 638 warning("GC/CMS: _eden_chunk_array allocation failure"); 639 } 640 } 641 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 642 643 // Support for parallelizing survivor space rescan 644 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 645 const size_t max_plab_samples = 646 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 647 648 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 649 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 650 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 651 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 652 || _cursor == NULL) { 653 warning("Failed to allocate survivor plab/chunk array"); 654 if (_survivor_plab_array != NULL) { 655 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 656 _survivor_plab_array = NULL; 657 } 658 if (_survivor_chunk_array != NULL) { 659 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 660 _survivor_chunk_array = NULL; 661 } 662 if (_cursor != NULL) { 663 FREE_C_HEAP_ARRAY(size_t, _cursor); 664 _cursor = NULL; 665 } 666 } else { 667 _survivor_chunk_capacity = 2*max_plab_samples; 668 for (uint i = 0; i < ParallelGCThreads; i++) { 669 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 670 if (vec == NULL) { 671 warning("Failed to allocate survivor plab array"); 672 for (int j = i; j > 0; j--) { 673 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 674 } 675 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 676 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 677 _survivor_plab_array = NULL; 678 _survivor_chunk_array = NULL; 679 _survivor_chunk_capacity = 0; 680 break; 681 } else { 682 ChunkArray* cur = 683 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 684 max_plab_samples); 685 assert(cur->end() == 0, "Should be 0"); 686 assert(cur->array() == vec, "Should be vec"); 687 assert(cur->capacity() == max_plab_samples, "Error"); 688 } 689 } 690 } 691 } 692 assert( ( _survivor_plab_array != NULL 693 && _survivor_chunk_array != NULL) 694 || ( _survivor_chunk_capacity == 0 695 && _survivor_chunk_index == 0), 696 "Error"); 697 698 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 699 _gc_counters = new CollectorCounters("CMS", 1); 700 _completed_initialization = true; 701 _inter_sweep_timer.start(); // start of time 702 } 703 704 size_t CMSCollector::plab_sample_minimum_size() { 705 // The default value of MinTLABSize is 2k, but there is 706 // no way to get the default value if the flag has been overridden. 707 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 708 } 709 710 const char* ConcurrentMarkSweepGeneration::name() const { 711 return "concurrent mark-sweep generation"; 712 } 713 void ConcurrentMarkSweepGeneration::update_counters() { 714 if (UsePerfData) { 715 _space_counters->update_all(); 716 _gen_counters->update_all(); 717 } 718 } 719 720 // this is an optimized version of update_counters(). it takes the 721 // used value as a parameter rather than computing it. 722 // 723 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 724 if (UsePerfData) { 725 _space_counters->update_used(used); 726 _space_counters->update_capacity(); 727 _gen_counters->update_all(); 728 } 729 } 730 731 void ConcurrentMarkSweepGeneration::print() const { 732 Generation::print(); 733 cmsSpace()->print(); 734 } 735 736 #ifndef PRODUCT 737 void ConcurrentMarkSweepGeneration::print_statistics() { 738 cmsSpace()->printFLCensus(0); 739 } 740 #endif 741 742 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 743 GenCollectedHeap* gch = GenCollectedHeap::heap(); 744 if (PrintGCDetails) { 745 if (Verbose) { 746 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 747 level(), short_name(), s, used(), capacity()); 748 } else { 749 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 750 level(), short_name(), s, used() / K, capacity() / K); 751 } 752 } 753 if (Verbose) { 754 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 755 gch->used(), gch->capacity()); 756 } else { 757 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 758 gch->used() / K, gch->capacity() / K); 759 } 760 } 761 762 size_t 763 ConcurrentMarkSweepGeneration::contiguous_available() const { 764 // dld proposes an improvement in precision here. If the committed 765 // part of the space ends in a free block we should add that to 766 // uncommitted size in the calculation below. Will make this 767 // change later, staying with the approximation below for the 768 // time being. -- ysr. 769 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 770 } 771 772 size_t 773 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 774 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 775 } 776 777 size_t ConcurrentMarkSweepGeneration::max_available() const { 778 return free() + _virtual_space.uncommitted_size(); 779 } 780 781 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 782 size_t available = max_available(); 783 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 784 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 785 if (Verbose && PrintGCDetails) { 786 gclog_or_tty->print_cr( 787 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 788 "max_promo("SIZE_FORMAT")", 789 res? "":" not", available, res? ">=":"<", 790 av_promo, max_promotion_in_bytes); 791 } 792 return res; 793 } 794 795 // At a promotion failure dump information on block layout in heap 796 // (cms old generation). 797 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 798 if (CMSDumpAtPromotionFailure) { 799 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 800 } 801 } 802 803 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 804 // Clear the promotion information. These pointers can be adjusted 805 // along with all the other pointers into the heap but 806 // compaction is expected to be a rare event with 807 // a heap using cms so don't do it without seeing the need. 808 for (uint i = 0; i < ParallelGCThreads; i++) { 809 _par_gc_thread_states[i]->promo.reset(); 810 } 811 } 812 813 void ConcurrentMarkSweepGeneration::compute_new_size() { 814 assert_locked_or_safepoint(Heap_lock); 815 816 // If incremental collection failed, we just want to expand 817 // to the limit. 818 if (incremental_collection_failed()) { 819 clear_incremental_collection_failed(); 820 grow_to_reserved(); 821 return; 822 } 823 824 // The heap has been compacted but not reset yet. 825 // Any metric such as free() or used() will be incorrect. 826 827 CardGeneration::compute_new_size(); 828 829 // Reset again after a possible resizing 830 if (did_compact()) { 831 cmsSpace()->reset_after_compaction(); 832 } 833 } 834 835 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 836 assert_locked_or_safepoint(Heap_lock); 837 838 // If incremental collection failed, we just want to expand 839 // to the limit. 840 if (incremental_collection_failed()) { 841 clear_incremental_collection_failed(); 842 grow_to_reserved(); 843 return; 844 } 845 846 double free_percentage = ((double) free()) / capacity(); 847 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 848 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 849 850 // compute expansion delta needed for reaching desired free percentage 851 if (free_percentage < desired_free_percentage) { 852 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 853 assert(desired_capacity >= capacity(), "invalid expansion size"); 854 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 855 if (PrintGCDetails && Verbose) { 856 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 857 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 858 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 859 gclog_or_tty->print_cr(" Desired free fraction %f", 860 desired_free_percentage); 861 gclog_or_tty->print_cr(" Maximum free fraction %f", 862 maximum_free_percentage); 863 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 864 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 865 desired_capacity/1000); 866 int prev_level = level() - 1; 867 if (prev_level >= 0) { 868 size_t prev_size = 0; 869 GenCollectedHeap* gch = GenCollectedHeap::heap(); 870 Generation* prev_gen = gch->young_gen(); 871 prev_size = prev_gen->capacity(); 872 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 873 prev_size/1000); 874 } 875 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 876 unsafe_max_alloc_nogc()/1000); 877 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 878 contiguous_available()/1000); 879 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 880 expand_bytes); 881 } 882 // safe if expansion fails 883 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 884 if (PrintGCDetails && Verbose) { 885 gclog_or_tty->print_cr(" Expanded free fraction %f", 886 ((double) free()) / capacity()); 887 } 888 } else { 889 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 890 assert(desired_capacity <= capacity(), "invalid expansion size"); 891 size_t shrink_bytes = capacity() - desired_capacity; 892 // Don't shrink unless the delta is greater than the minimum shrink we want 893 if (shrink_bytes >= MinHeapDeltaBytes) { 894 shrink_free_list_by(shrink_bytes); 895 } 896 } 897 } 898 899 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 900 return cmsSpace()->freelistLock(); 901 } 902 903 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 904 bool tlab) { 905 CMSSynchronousYieldRequest yr; 906 MutexLockerEx x(freelistLock(), 907 Mutex::_no_safepoint_check_flag); 908 return have_lock_and_allocate(size, tlab); 909 } 910 911 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 912 bool tlab /* ignored */) { 913 assert_lock_strong(freelistLock()); 914 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 915 HeapWord* res = cmsSpace()->allocate(adjustedSize); 916 // Allocate the object live (grey) if the background collector has 917 // started marking. This is necessary because the marker may 918 // have passed this address and consequently this object will 919 // not otherwise be greyed and would be incorrectly swept up. 920 // Note that if this object contains references, the writing 921 // of those references will dirty the card containing this object 922 // allowing the object to be blackened (and its references scanned) 923 // either during a preclean phase or at the final checkpoint. 924 if (res != NULL) { 925 // We may block here with an uninitialized object with 926 // its mark-bit or P-bits not yet set. Such objects need 927 // to be safely navigable by block_start(). 928 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 929 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 930 collector()->direct_allocated(res, adjustedSize); 931 _direct_allocated_words += adjustedSize; 932 // allocation counters 933 NOT_PRODUCT( 934 _numObjectsAllocated++; 935 _numWordsAllocated += (int)adjustedSize; 936 ) 937 } 938 return res; 939 } 940 941 // In the case of direct allocation by mutators in a generation that 942 // is being concurrently collected, the object must be allocated 943 // live (grey) if the background collector has started marking. 944 // This is necessary because the marker may 945 // have passed this address and consequently this object will 946 // not otherwise be greyed and would be incorrectly swept up. 947 // Note that if this object contains references, the writing 948 // of those references will dirty the card containing this object 949 // allowing the object to be blackened (and its references scanned) 950 // either during a preclean phase or at the final checkpoint. 951 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 952 assert(_markBitMap.covers(start, size), "Out of bounds"); 953 if (_collectorState >= Marking) { 954 MutexLockerEx y(_markBitMap.lock(), 955 Mutex::_no_safepoint_check_flag); 956 // [see comments preceding SweepClosure::do_blk() below for details] 957 // 958 // Can the P-bits be deleted now? JJJ 959 // 960 // 1. need to mark the object as live so it isn't collected 961 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 962 // 3. need to mark the end of the object so marking, precleaning or sweeping 963 // can skip over uninitialized or unparsable objects. An allocated 964 // object is considered uninitialized for our purposes as long as 965 // its klass word is NULL. All old gen objects are parsable 966 // as soon as they are initialized.) 967 _markBitMap.mark(start); // object is live 968 _markBitMap.mark(start + 1); // object is potentially uninitialized? 969 _markBitMap.mark(start + size - 1); 970 // mark end of object 971 } 972 // check that oop looks uninitialized 973 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 974 } 975 976 void CMSCollector::promoted(bool par, HeapWord* start, 977 bool is_obj_array, size_t obj_size) { 978 assert(_markBitMap.covers(start), "Out of bounds"); 979 // See comment in direct_allocated() about when objects should 980 // be allocated live. 981 if (_collectorState >= Marking) { 982 // we already hold the marking bit map lock, taken in 983 // the prologue 984 if (par) { 985 _markBitMap.par_mark(start); 986 } else { 987 _markBitMap.mark(start); 988 } 989 // We don't need to mark the object as uninitialized (as 990 // in direct_allocated above) because this is being done with the 991 // world stopped and the object will be initialized by the 992 // time the marking, precleaning or sweeping get to look at it. 993 // But see the code for copying objects into the CMS generation, 994 // where we need to ensure that concurrent readers of the 995 // block offset table are able to safely navigate a block that 996 // is in flux from being free to being allocated (and in 997 // transition while being copied into) and subsequently 998 // becoming a bona-fide object when the copy/promotion is complete. 999 assert(SafepointSynchronize::is_at_safepoint(), 1000 "expect promotion only at safepoints"); 1001 1002 if (_collectorState < Sweeping) { 1003 // Mark the appropriate cards in the modUnionTable, so that 1004 // this object gets scanned before the sweep. If this is 1005 // not done, CMS generation references in the object might 1006 // not get marked. 1007 // For the case of arrays, which are otherwise precisely 1008 // marked, we need to dirty the entire array, not just its head. 1009 if (is_obj_array) { 1010 // The [par_]mark_range() method expects mr.end() below to 1011 // be aligned to the granularity of a bit's representation 1012 // in the heap. In the case of the MUT below, that's a 1013 // card size. 1014 MemRegion mr(start, 1015 (HeapWord*)round_to((intptr_t)(start + obj_size), 1016 CardTableModRefBS::card_size /* bytes */)); 1017 if (par) { 1018 _modUnionTable.par_mark_range(mr); 1019 } else { 1020 _modUnionTable.mark_range(mr); 1021 } 1022 } else { // not an obj array; we can just mark the head 1023 if (par) { 1024 _modUnionTable.par_mark(start); 1025 } else { 1026 _modUnionTable.mark(start); 1027 } 1028 } 1029 } 1030 } 1031 } 1032 1033 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1034 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1035 // allocate, copy and if necessary update promoinfo -- 1036 // delegate to underlying space. 1037 assert_lock_strong(freelistLock()); 1038 1039 #ifndef PRODUCT 1040 if (Universe::heap()->promotion_should_fail()) { 1041 return NULL; 1042 } 1043 #endif // #ifndef PRODUCT 1044 1045 oop res = _cmsSpace->promote(obj, obj_size); 1046 if (res == NULL) { 1047 // expand and retry 1048 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1049 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 1050 // Since this is the old generation, we don't try to promote 1051 // into a more senior generation. 1052 res = _cmsSpace->promote(obj, obj_size); 1053 } 1054 if (res != NULL) { 1055 // See comment in allocate() about when objects should 1056 // be allocated live. 1057 assert(obj->is_oop(), "Will dereference klass pointer below"); 1058 collector()->promoted(false, // Not parallel 1059 (HeapWord*)res, obj->is_objArray(), obj_size); 1060 // promotion counters 1061 NOT_PRODUCT( 1062 _numObjectsPromoted++; 1063 _numWordsPromoted += 1064 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1065 ) 1066 } 1067 return res; 1068 } 1069 1070 1071 // IMPORTANT: Notes on object size recognition in CMS. 1072 // --------------------------------------------------- 1073 // A block of storage in the CMS generation is always in 1074 // one of three states. A free block (FREE), an allocated 1075 // object (OBJECT) whose size() method reports the correct size, 1076 // and an intermediate state (TRANSIENT) in which its size cannot 1077 // be accurately determined. 1078 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1079 // ----------------------------------------------------- 1080 // FREE: klass_word & 1 == 1; mark_word holds block size 1081 // 1082 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1083 // obj->size() computes correct size 1084 // 1085 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1086 // 1087 // STATE IDENTIFICATION: (64 bit+COOPS) 1088 // ------------------------------------ 1089 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1090 // 1091 // OBJECT: klass_word installed; klass_word != 0; 1092 // obj->size() computes correct size 1093 // 1094 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1095 // 1096 // 1097 // STATE TRANSITION DIAGRAM 1098 // 1099 // mut / parnew mut / parnew 1100 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1101 // ^ | 1102 // |------------------------ DEAD <------------------------------------| 1103 // sweep mut 1104 // 1105 // While a block is in TRANSIENT state its size cannot be determined 1106 // so readers will either need to come back later or stall until 1107 // the size can be determined. Note that for the case of direct 1108 // allocation, P-bits, when available, may be used to determine the 1109 // size of an object that may not yet have been initialized. 1110 1111 // Things to support parallel young-gen collection. 1112 oop 1113 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1114 oop old, markOop m, 1115 size_t word_sz) { 1116 #ifndef PRODUCT 1117 if (Universe::heap()->promotion_should_fail()) { 1118 return NULL; 1119 } 1120 #endif // #ifndef PRODUCT 1121 1122 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1123 PromotionInfo* promoInfo = &ps->promo; 1124 // if we are tracking promotions, then first ensure space for 1125 // promotion (including spooling space for saving header if necessary). 1126 // then allocate and copy, then track promoted info if needed. 1127 // When tracking (see PromotionInfo::track()), the mark word may 1128 // be displaced and in this case restoration of the mark word 1129 // occurs in the (oop_since_save_marks_)iterate phase. 1130 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1131 // Out of space for allocating spooling buffers; 1132 // try expanding and allocating spooling buffers. 1133 if (!expand_and_ensure_spooling_space(promoInfo)) { 1134 return NULL; 1135 } 1136 } 1137 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1138 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1139 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1140 if (obj_ptr == NULL) { 1141 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1142 if (obj_ptr == NULL) { 1143 return NULL; 1144 } 1145 } 1146 oop obj = oop(obj_ptr); 1147 OrderAccess::storestore(); 1148 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1149 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1150 // IMPORTANT: See note on object initialization for CMS above. 1151 // Otherwise, copy the object. Here we must be careful to insert the 1152 // klass pointer last, since this marks the block as an allocated object. 1153 // Except with compressed oops it's the mark word. 1154 HeapWord* old_ptr = (HeapWord*)old; 1155 // Restore the mark word copied above. 1156 obj->set_mark(m); 1157 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1158 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1159 OrderAccess::storestore(); 1160 1161 if (UseCompressedClassPointers) { 1162 // Copy gap missed by (aligned) header size calculation below 1163 obj->set_klass_gap(old->klass_gap()); 1164 } 1165 if (word_sz > (size_t)oopDesc::header_size()) { 1166 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1167 obj_ptr + oopDesc::header_size(), 1168 word_sz - oopDesc::header_size()); 1169 } 1170 1171 // Now we can track the promoted object, if necessary. We take care 1172 // to delay the transition from uninitialized to full object 1173 // (i.e., insertion of klass pointer) until after, so that it 1174 // atomically becomes a promoted object. 1175 if (promoInfo->tracking()) { 1176 promoInfo->track((PromotedObject*)obj, old->klass()); 1177 } 1178 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1179 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1180 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1181 1182 // Finally, install the klass pointer (this should be volatile). 1183 OrderAccess::storestore(); 1184 obj->set_klass(old->klass()); 1185 // We should now be able to calculate the right size for this object 1186 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1187 1188 collector()->promoted(true, // parallel 1189 obj_ptr, old->is_objArray(), word_sz); 1190 1191 NOT_PRODUCT( 1192 Atomic::inc_ptr(&_numObjectsPromoted); 1193 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1194 ) 1195 1196 return obj; 1197 } 1198 1199 void 1200 ConcurrentMarkSweepGeneration:: 1201 par_promote_alloc_done(int thread_num) { 1202 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1203 ps->lab.retire(thread_num); 1204 } 1205 1206 void 1207 ConcurrentMarkSweepGeneration:: 1208 par_oop_since_save_marks_iterate_done(int thread_num) { 1209 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1210 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1211 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1212 } 1213 1214 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1215 size_t size, 1216 bool tlab) 1217 { 1218 // We allow a STW collection only if a full 1219 // collection was requested. 1220 return full || should_allocate(size, tlab); // FIX ME !!! 1221 // This and promotion failure handling are connected at the 1222 // hip and should be fixed by untying them. 1223 } 1224 1225 bool CMSCollector::shouldConcurrentCollect() { 1226 if (_full_gc_requested) { 1227 if (Verbose && PrintGCDetails) { 1228 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1229 " gc request (or gc_locker)"); 1230 } 1231 return true; 1232 } 1233 1234 FreelistLocker x(this); 1235 // ------------------------------------------------------------------ 1236 // Print out lots of information which affects the initiation of 1237 // a collection. 1238 if (PrintCMSInitiationStatistics && stats().valid()) { 1239 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1240 gclog_or_tty->stamp(); 1241 gclog_or_tty->cr(); 1242 stats().print_on(gclog_or_tty); 1243 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1244 stats().time_until_cms_gen_full()); 1245 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1246 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1247 _cmsGen->contiguous_available()); 1248 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1249 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1250 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1251 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1252 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1253 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1254 gclog_or_tty->print_cr("metadata initialized %d", 1255 MetaspaceGC::should_concurrent_collect()); 1256 } 1257 // ------------------------------------------------------------------ 1258 1259 // If the estimated time to complete a cms collection (cms_duration()) 1260 // is less than the estimated time remaining until the cms generation 1261 // is full, start a collection. 1262 if (!UseCMSInitiatingOccupancyOnly) { 1263 if (stats().valid()) { 1264 if (stats().time_until_cms_start() == 0.0) { 1265 return true; 1266 } 1267 } else { 1268 // We want to conservatively collect somewhat early in order 1269 // to try and "bootstrap" our CMS/promotion statistics; 1270 // this branch will not fire after the first successful CMS 1271 // collection because the stats should then be valid. 1272 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1273 if (Verbose && PrintGCDetails) { 1274 gclog_or_tty->print_cr( 1275 " CMSCollector: collect for bootstrapping statistics:" 1276 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1277 _bootstrap_occupancy); 1278 } 1279 return true; 1280 } 1281 } 1282 } 1283 1284 // Otherwise, we start a collection cycle if 1285 // old gen want a collection cycle started. Each may use 1286 // an appropriate criterion for making this decision. 1287 // XXX We need to make sure that the gen expansion 1288 // criterion dovetails well with this. XXX NEED TO FIX THIS 1289 if (_cmsGen->should_concurrent_collect()) { 1290 if (Verbose && PrintGCDetails) { 1291 gclog_or_tty->print_cr("CMS old gen initiated"); 1292 } 1293 return true; 1294 } 1295 1296 // We start a collection if we believe an incremental collection may fail; 1297 // this is not likely to be productive in practice because it's probably too 1298 // late anyway. 1299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1300 assert(gch->collector_policy()->is_generation_policy(), 1301 "You may want to check the correctness of the following"); 1302 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1303 if (Verbose && PrintGCDetails) { 1304 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1305 } 1306 return true; 1307 } 1308 1309 if (MetaspaceGC::should_concurrent_collect()) { 1310 if (Verbose && PrintGCDetails) { 1311 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1312 } 1313 return true; 1314 } 1315 1316 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1317 if (CMSTriggerInterval >= 0) { 1318 if (CMSTriggerInterval == 0) { 1319 // Trigger always 1320 return true; 1321 } 1322 1323 // Check the CMS time since begin (we do not check the stats validity 1324 // as we want to be able to trigger the first CMS cycle as well) 1325 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1326 if (Verbose && PrintGCDetails) { 1327 if (stats().valid()) { 1328 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1329 stats().cms_time_since_begin()); 1330 } else { 1331 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1332 } 1333 } 1334 return true; 1335 } 1336 } 1337 1338 return false; 1339 } 1340 1341 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1342 1343 // Clear _expansion_cause fields of constituent generations 1344 void CMSCollector::clear_expansion_cause() { 1345 _cmsGen->clear_expansion_cause(); 1346 } 1347 1348 // We should be conservative in starting a collection cycle. To 1349 // start too eagerly runs the risk of collecting too often in the 1350 // extreme. To collect too rarely falls back on full collections, 1351 // which works, even if not optimum in terms of concurrent work. 1352 // As a work around for too eagerly collecting, use the flag 1353 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1354 // giving the user an easily understandable way of controlling the 1355 // collections. 1356 // We want to start a new collection cycle if any of the following 1357 // conditions hold: 1358 // . our current occupancy exceeds the configured initiating occupancy 1359 // for this generation, or 1360 // . we recently needed to expand this space and have not, since that 1361 // expansion, done a collection of this generation, or 1362 // . the underlying space believes that it may be a good idea to initiate 1363 // a concurrent collection (this may be based on criteria such as the 1364 // following: the space uses linear allocation and linear allocation is 1365 // going to fail, or there is believed to be excessive fragmentation in 1366 // the generation, etc... or ... 1367 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1368 // the case of the old generation; see CR 6543076): 1369 // we may be approaching a point at which allocation requests may fail because 1370 // we will be out of sufficient free space given allocation rate estimates.] 1371 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1372 1373 assert_lock_strong(freelistLock()); 1374 if (occupancy() > initiating_occupancy()) { 1375 if (PrintGCDetails && Verbose) { 1376 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1377 short_name(), occupancy(), initiating_occupancy()); 1378 } 1379 return true; 1380 } 1381 if (UseCMSInitiatingOccupancyOnly) { 1382 return false; 1383 } 1384 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1385 if (PrintGCDetails && Verbose) { 1386 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1387 short_name()); 1388 } 1389 return true; 1390 } 1391 if (_cmsSpace->should_concurrent_collect()) { 1392 if (PrintGCDetails && Verbose) { 1393 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1394 short_name()); 1395 } 1396 return true; 1397 } 1398 return false; 1399 } 1400 1401 void ConcurrentMarkSweepGeneration::collect(bool full, 1402 bool clear_all_soft_refs, 1403 size_t size, 1404 bool tlab) 1405 { 1406 collector()->collect(full, clear_all_soft_refs, size, tlab); 1407 } 1408 1409 void CMSCollector::collect(bool full, 1410 bool clear_all_soft_refs, 1411 size_t size, 1412 bool tlab) 1413 { 1414 // The following "if" branch is present for defensive reasons. 1415 // In the current uses of this interface, it can be replaced with: 1416 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1417 // But I am not placing that assert here to allow future 1418 // generality in invoking this interface. 1419 if (GC_locker::is_active()) { 1420 // A consistency test for GC_locker 1421 assert(GC_locker::needs_gc(), "Should have been set already"); 1422 // Skip this foreground collection, instead 1423 // expanding the heap if necessary. 1424 // Need the free list locks for the call to free() in compute_new_size() 1425 compute_new_size(); 1426 return; 1427 } 1428 acquire_control_and_collect(full, clear_all_soft_refs); 1429 } 1430 1431 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1432 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1433 unsigned int gc_count = gch->total_full_collections(); 1434 if (gc_count == full_gc_count) { 1435 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1436 _full_gc_requested = true; 1437 _full_gc_cause = cause; 1438 CGC_lock->notify(); // nudge CMS thread 1439 } else { 1440 assert(gc_count > full_gc_count, "Error: causal loop"); 1441 } 1442 } 1443 1444 bool CMSCollector::is_external_interruption() { 1445 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1446 return GCCause::is_user_requested_gc(cause) || 1447 GCCause::is_serviceability_requested_gc(cause); 1448 } 1449 1450 void CMSCollector::report_concurrent_mode_interruption() { 1451 if (is_external_interruption()) { 1452 if (PrintGCDetails) { 1453 gclog_or_tty->print(" (concurrent mode interrupted)"); 1454 } 1455 } else { 1456 if (PrintGCDetails) { 1457 gclog_or_tty->print(" (concurrent mode failure)"); 1458 } 1459 _gc_tracer_cm->report_concurrent_mode_failure(); 1460 } 1461 } 1462 1463 1464 // The foreground and background collectors need to coordinate in order 1465 // to make sure that they do not mutually interfere with CMS collections. 1466 // When a background collection is active, 1467 // the foreground collector may need to take over (preempt) and 1468 // synchronously complete an ongoing collection. Depending on the 1469 // frequency of the background collections and the heap usage 1470 // of the application, this preemption can be seldom or frequent. 1471 // There are only certain 1472 // points in the background collection that the "collection-baton" 1473 // can be passed to the foreground collector. 1474 // 1475 // The foreground collector will wait for the baton before 1476 // starting any part of the collection. The foreground collector 1477 // will only wait at one location. 1478 // 1479 // The background collector will yield the baton before starting a new 1480 // phase of the collection (e.g., before initial marking, marking from roots, 1481 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1482 // of the loop which switches the phases. The background collector does some 1483 // of the phases (initial mark, final re-mark) with the world stopped. 1484 // Because of locking involved in stopping the world, 1485 // the foreground collector should not block waiting for the background 1486 // collector when it is doing a stop-the-world phase. The background 1487 // collector will yield the baton at an additional point just before 1488 // it enters a stop-the-world phase. Once the world is stopped, the 1489 // background collector checks the phase of the collection. If the 1490 // phase has not changed, it proceeds with the collection. If the 1491 // phase has changed, it skips that phase of the collection. See 1492 // the comments on the use of the Heap_lock in collect_in_background(). 1493 // 1494 // Variable used in baton passing. 1495 // _foregroundGCIsActive - Set to true by the foreground collector when 1496 // it wants the baton. The foreground clears it when it has finished 1497 // the collection. 1498 // _foregroundGCShouldWait - Set to true by the background collector 1499 // when it is running. The foreground collector waits while 1500 // _foregroundGCShouldWait is true. 1501 // CGC_lock - monitor used to protect access to the above variables 1502 // and to notify the foreground and background collectors. 1503 // _collectorState - current state of the CMS collection. 1504 // 1505 // The foreground collector 1506 // acquires the CGC_lock 1507 // sets _foregroundGCIsActive 1508 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1509 // various locks acquired in preparation for the collection 1510 // are released so as not to block the background collector 1511 // that is in the midst of a collection 1512 // proceeds with the collection 1513 // clears _foregroundGCIsActive 1514 // returns 1515 // 1516 // The background collector in a loop iterating on the phases of the 1517 // collection 1518 // acquires the CGC_lock 1519 // sets _foregroundGCShouldWait 1520 // if _foregroundGCIsActive is set 1521 // clears _foregroundGCShouldWait, notifies _CGC_lock 1522 // waits on _CGC_lock for _foregroundGCIsActive to become false 1523 // and exits the loop. 1524 // otherwise 1525 // proceed with that phase of the collection 1526 // if the phase is a stop-the-world phase, 1527 // yield the baton once more just before enqueueing 1528 // the stop-world CMS operation (executed by the VM thread). 1529 // returns after all phases of the collection are done 1530 // 1531 1532 void CMSCollector::acquire_control_and_collect(bool full, 1533 bool clear_all_soft_refs) { 1534 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1535 assert(!Thread::current()->is_ConcurrentGC_thread(), 1536 "shouldn't try to acquire control from self!"); 1537 1538 // Start the protocol for acquiring control of the 1539 // collection from the background collector (aka CMS thread). 1540 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1541 "VM thread should have CMS token"); 1542 // Remember the possibly interrupted state of an ongoing 1543 // concurrent collection 1544 CollectorState first_state = _collectorState; 1545 1546 // Signal to a possibly ongoing concurrent collection that 1547 // we want to do a foreground collection. 1548 _foregroundGCIsActive = true; 1549 1550 // release locks and wait for a notify from the background collector 1551 // releasing the locks in only necessary for phases which 1552 // do yields to improve the granularity of the collection. 1553 assert_lock_strong(bitMapLock()); 1554 // We need to lock the Free list lock for the space that we are 1555 // currently collecting. 1556 assert(haveFreelistLocks(), "Must be holding free list locks"); 1557 bitMapLock()->unlock(); 1558 releaseFreelistLocks(); 1559 { 1560 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1561 if (_foregroundGCShouldWait) { 1562 // We are going to be waiting for action for the CMS thread; 1563 // it had better not be gone (for instance at shutdown)! 1564 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1565 "CMS thread must be running"); 1566 // Wait here until the background collector gives us the go-ahead 1567 ConcurrentMarkSweepThread::clear_CMS_flag( 1568 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1569 // Get a possibly blocked CMS thread going: 1570 // Note that we set _foregroundGCIsActive true above, 1571 // without protection of the CGC_lock. 1572 CGC_lock->notify(); 1573 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1574 "Possible deadlock"); 1575 while (_foregroundGCShouldWait) { 1576 // wait for notification 1577 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1578 // Possibility of delay/starvation here, since CMS token does 1579 // not know to give priority to VM thread? Actually, i think 1580 // there wouldn't be any delay/starvation, but the proof of 1581 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1582 } 1583 ConcurrentMarkSweepThread::set_CMS_flag( 1584 ConcurrentMarkSweepThread::CMS_vm_has_token); 1585 } 1586 } 1587 // The CMS_token is already held. Get back the other locks. 1588 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1589 "VM thread should have CMS token"); 1590 getFreelistLocks(); 1591 bitMapLock()->lock_without_safepoint_check(); 1592 if (TraceCMSState) { 1593 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1594 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1595 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1596 } 1597 1598 // Inform cms gen if this was due to partial collection failing. 1599 // The CMS gen may use this fact to determine its expansion policy. 1600 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1601 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1602 assert(!_cmsGen->incremental_collection_failed(), 1603 "Should have been noticed, reacted to and cleared"); 1604 _cmsGen->set_incremental_collection_failed(); 1605 } 1606 1607 if (first_state > Idling) { 1608 report_concurrent_mode_interruption(); 1609 } 1610 1611 set_did_compact(true); 1612 1613 // If the collection is being acquired from the background 1614 // collector, there may be references on the discovered 1615 // references lists. Abandon those references, since some 1616 // of them may have become unreachable after concurrent 1617 // discovery; the STW compacting collector will redo discovery 1618 // more precisely, without being subject to floating garbage. 1619 // Leaving otherwise unreachable references in the discovered 1620 // lists would require special handling. 1621 ref_processor()->disable_discovery(); 1622 ref_processor()->abandon_partial_discovery(); 1623 ref_processor()->verify_no_references_recorded(); 1624 1625 if (first_state > Idling) { 1626 save_heap_summary(); 1627 } 1628 1629 do_compaction_work(clear_all_soft_refs); 1630 1631 // Has the GC time limit been exceeded? 1632 size_t max_eden_size = _young_gen->max_capacity() - 1633 _young_gen->to()->capacity() - 1634 _young_gen->from()->capacity(); 1635 GCCause::Cause gc_cause = gch->gc_cause(); 1636 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1637 _young_gen->eden()->used(), 1638 _cmsGen->max_capacity(), 1639 max_eden_size, 1640 full, 1641 gc_cause, 1642 gch->collector_policy()); 1643 1644 // Reset the expansion cause, now that we just completed 1645 // a collection cycle. 1646 clear_expansion_cause(); 1647 _foregroundGCIsActive = false; 1648 return; 1649 } 1650 1651 // Resize the tenured generation 1652 // after obtaining the free list locks for the 1653 // two generations. 1654 void CMSCollector::compute_new_size() { 1655 assert_locked_or_safepoint(Heap_lock); 1656 FreelistLocker z(this); 1657 MetaspaceGC::compute_new_size(); 1658 _cmsGen->compute_new_size_free_list(); 1659 } 1660 1661 // A work method used by the foreground collector to do 1662 // a mark-sweep-compact. 1663 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1664 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1665 1666 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1667 gc_timer->register_gc_start(); 1668 1669 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1670 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1671 1672 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1673 1674 // Temporarily widen the span of the weak reference processing to 1675 // the entire heap. 1676 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1677 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1678 // Temporarily, clear the "is_alive_non_header" field of the 1679 // reference processor. 1680 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1681 // Temporarily make reference _processing_ single threaded (non-MT). 1682 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1683 // Temporarily make refs discovery atomic 1684 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1685 // Temporarily make reference _discovery_ single threaded (non-MT) 1686 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1687 1688 ref_processor()->set_enqueuing_is_done(false); 1689 ref_processor()->enable_discovery(); 1690 ref_processor()->setup_policy(clear_all_soft_refs); 1691 // If an asynchronous collection finishes, the _modUnionTable is 1692 // all clear. If we are assuming the collection from an asynchronous 1693 // collection, clear the _modUnionTable. 1694 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1695 "_modUnionTable should be clear if the baton was not passed"); 1696 _modUnionTable.clear_all(); 1697 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1698 "mod union for klasses should be clear if the baton was passed"); 1699 _ct->klass_rem_set()->clear_mod_union(); 1700 1701 // We must adjust the allocation statistics being maintained 1702 // in the free list space. We do so by reading and clearing 1703 // the sweep timer and updating the block flux rate estimates below. 1704 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1705 if (_inter_sweep_timer.is_active()) { 1706 _inter_sweep_timer.stop(); 1707 // Note that we do not use this sample to update the _inter_sweep_estimate. 1708 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1709 _inter_sweep_estimate.padded_average(), 1710 _intra_sweep_estimate.padded_average()); 1711 } 1712 1713 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1714 ref_processor(), clear_all_soft_refs); 1715 #ifdef ASSERT 1716 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1717 size_t free_size = cms_space->free(); 1718 assert(free_size == 1719 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1720 * HeapWordSize, 1721 "All the free space should be compacted into one chunk at top"); 1722 assert(cms_space->dictionary()->total_chunk_size( 1723 debug_only(cms_space->freelistLock())) == 0 || 1724 cms_space->totalSizeInIndexedFreeLists() == 0, 1725 "All the free space should be in a single chunk"); 1726 size_t num = cms_space->totalCount(); 1727 assert((free_size == 0 && num == 0) || 1728 (free_size > 0 && (num == 1 || num == 2)), 1729 "There should be at most 2 free chunks after compaction"); 1730 #endif // ASSERT 1731 _collectorState = Resetting; 1732 assert(_restart_addr == NULL, 1733 "Should have been NULL'd before baton was passed"); 1734 reset(false /* == !concurrent */); 1735 _cmsGen->reset_after_compaction(); 1736 _concurrent_cycles_since_last_unload = 0; 1737 1738 // Clear any data recorded in the PLAB chunk arrays. 1739 if (_survivor_plab_array != NULL) { 1740 reset_survivor_plab_arrays(); 1741 } 1742 1743 // Adjust the per-size allocation stats for the next epoch. 1744 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1745 // Restart the "inter sweep timer" for the next epoch. 1746 _inter_sweep_timer.reset(); 1747 _inter_sweep_timer.start(); 1748 1749 gc_timer->register_gc_end(); 1750 1751 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1752 1753 // For a mark-sweep-compact, compute_new_size() will be called 1754 // in the heap's do_collection() method. 1755 } 1756 1757 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1758 ContiguousSpace* eden_space = _young_gen->eden(); 1759 ContiguousSpace* from_space = _young_gen->from(); 1760 ContiguousSpace* to_space = _young_gen->to(); 1761 // Eden 1762 if (_eden_chunk_array != NULL) { 1763 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1764 p2i(eden_space->bottom()), p2i(eden_space->top()), 1765 p2i(eden_space->end()), eden_space->capacity()); 1766 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1767 "_eden_chunk_capacity=" SIZE_FORMAT, 1768 _eden_chunk_index, _eden_chunk_capacity); 1769 for (size_t i = 0; i < _eden_chunk_index; i++) { 1770 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1771 i, p2i(_eden_chunk_array[i])); 1772 } 1773 } 1774 // Survivor 1775 if (_survivor_chunk_array != NULL) { 1776 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1777 p2i(from_space->bottom()), p2i(from_space->top()), 1778 p2i(from_space->end()), from_space->capacity()); 1779 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1780 "_survivor_chunk_capacity=" SIZE_FORMAT, 1781 _survivor_chunk_index, _survivor_chunk_capacity); 1782 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1783 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1784 i, p2i(_survivor_chunk_array[i])); 1785 } 1786 } 1787 } 1788 1789 void CMSCollector::getFreelistLocks() const { 1790 // Get locks for all free lists in all generations that this 1791 // collector is responsible for 1792 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1793 } 1794 1795 void CMSCollector::releaseFreelistLocks() const { 1796 // Release locks for all free lists in all generations that this 1797 // collector is responsible for 1798 _cmsGen->freelistLock()->unlock(); 1799 } 1800 1801 bool CMSCollector::haveFreelistLocks() const { 1802 // Check locks for all free lists in all generations that this 1803 // collector is responsible for 1804 assert_lock_strong(_cmsGen->freelistLock()); 1805 PRODUCT_ONLY(ShouldNotReachHere()); 1806 return true; 1807 } 1808 1809 // A utility class that is used by the CMS collector to 1810 // temporarily "release" the foreground collector from its 1811 // usual obligation to wait for the background collector to 1812 // complete an ongoing phase before proceeding. 1813 class ReleaseForegroundGC: public StackObj { 1814 private: 1815 CMSCollector* _c; 1816 public: 1817 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1818 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1819 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1820 // allow a potentially blocked foreground collector to proceed 1821 _c->_foregroundGCShouldWait = false; 1822 if (_c->_foregroundGCIsActive) { 1823 CGC_lock->notify(); 1824 } 1825 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1826 "Possible deadlock"); 1827 } 1828 1829 ~ReleaseForegroundGC() { 1830 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1831 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1832 _c->_foregroundGCShouldWait = true; 1833 } 1834 }; 1835 1836 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1837 assert(Thread::current()->is_ConcurrentGC_thread(), 1838 "A CMS asynchronous collection is only allowed on a CMS thread."); 1839 1840 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1841 { 1842 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1843 MutexLockerEx hl(Heap_lock, safepoint_check); 1844 FreelistLocker fll(this); 1845 MutexLockerEx x(CGC_lock, safepoint_check); 1846 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1847 // The foreground collector is active or we're 1848 // not using asynchronous collections. Skip this 1849 // background collection. 1850 assert(!_foregroundGCShouldWait, "Should be clear"); 1851 return; 1852 } else { 1853 assert(_collectorState == Idling, "Should be idling before start."); 1854 _collectorState = InitialMarking; 1855 register_gc_start(cause); 1856 // Reset the expansion cause, now that we are about to begin 1857 // a new cycle. 1858 clear_expansion_cause(); 1859 1860 // Clear the MetaspaceGC flag since a concurrent collection 1861 // is starting but also clear it after the collection. 1862 MetaspaceGC::set_should_concurrent_collect(false); 1863 } 1864 // Decide if we want to enable class unloading as part of the 1865 // ensuing concurrent GC cycle. 1866 update_should_unload_classes(); 1867 _full_gc_requested = false; // acks all outstanding full gc requests 1868 _full_gc_cause = GCCause::_no_gc; 1869 // Signal that we are about to start a collection 1870 gch->increment_total_full_collections(); // ... starting a collection cycle 1871 _collection_count_start = gch->total_full_collections(); 1872 } 1873 1874 // Used for PrintGC 1875 size_t prev_used; 1876 if (PrintGC && Verbose) { 1877 prev_used = _cmsGen->used(); 1878 } 1879 1880 // The change of the collection state is normally done at this level; 1881 // the exceptions are phases that are executed while the world is 1882 // stopped. For those phases the change of state is done while the 1883 // world is stopped. For baton passing purposes this allows the 1884 // background collector to finish the phase and change state atomically. 1885 // The foreground collector cannot wait on a phase that is done 1886 // while the world is stopped because the foreground collector already 1887 // has the world stopped and would deadlock. 1888 while (_collectorState != Idling) { 1889 if (TraceCMSState) { 1890 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1891 p2i(Thread::current()), _collectorState); 1892 } 1893 // The foreground collector 1894 // holds the Heap_lock throughout its collection. 1895 // holds the CMS token (but not the lock) 1896 // except while it is waiting for the background collector to yield. 1897 // 1898 // The foreground collector should be blocked (not for long) 1899 // if the background collector is about to start a phase 1900 // executed with world stopped. If the background 1901 // collector has already started such a phase, the 1902 // foreground collector is blocked waiting for the 1903 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1904 // are executed in the VM thread. 1905 // 1906 // The locking order is 1907 // PendingListLock (PLL) -- if applicable (FinalMarking) 1908 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1909 // CMS token (claimed in 1910 // stop_world_and_do() --> 1911 // safepoint_synchronize() --> 1912 // CMSThread::synchronize()) 1913 1914 { 1915 // Check if the FG collector wants us to yield. 1916 CMSTokenSync x(true); // is cms thread 1917 if (waitForForegroundGC()) { 1918 // We yielded to a foreground GC, nothing more to be 1919 // done this round. 1920 assert(_foregroundGCShouldWait == false, "We set it to false in " 1921 "waitForForegroundGC()"); 1922 if (TraceCMSState) { 1923 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1924 " exiting collection CMS state %d", 1925 p2i(Thread::current()), _collectorState); 1926 } 1927 return; 1928 } else { 1929 // The background collector can run but check to see if the 1930 // foreground collector has done a collection while the 1931 // background collector was waiting to get the CGC_lock 1932 // above. If yes, break so that _foregroundGCShouldWait 1933 // is cleared before returning. 1934 if (_collectorState == Idling) { 1935 break; 1936 } 1937 } 1938 } 1939 1940 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1941 "should be waiting"); 1942 1943 switch (_collectorState) { 1944 case InitialMarking: 1945 { 1946 ReleaseForegroundGC x(this); 1947 stats().record_cms_begin(); 1948 VM_CMS_Initial_Mark initial_mark_op(this); 1949 VMThread::execute(&initial_mark_op); 1950 } 1951 // The collector state may be any legal state at this point 1952 // since the background collector may have yielded to the 1953 // foreground collector. 1954 break; 1955 case Marking: 1956 // initial marking in checkpointRootsInitialWork has been completed 1957 if (markFromRoots()) { // we were successful 1958 assert(_collectorState == Precleaning, "Collector state should " 1959 "have changed"); 1960 } else { 1961 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1962 } 1963 break; 1964 case Precleaning: 1965 // marking from roots in markFromRoots has been completed 1966 preclean(); 1967 assert(_collectorState == AbortablePreclean || 1968 _collectorState == FinalMarking, 1969 "Collector state should have changed"); 1970 break; 1971 case AbortablePreclean: 1972 abortable_preclean(); 1973 assert(_collectorState == FinalMarking, "Collector state should " 1974 "have changed"); 1975 break; 1976 case FinalMarking: 1977 { 1978 ReleaseForegroundGC x(this); 1979 1980 VM_CMS_Final_Remark final_remark_op(this); 1981 VMThread::execute(&final_remark_op); 1982 } 1983 assert(_foregroundGCShouldWait, "block post-condition"); 1984 break; 1985 case Sweeping: 1986 // final marking in checkpointRootsFinal has been completed 1987 sweep(); 1988 assert(_collectorState == Resizing, "Collector state change " 1989 "to Resizing must be done under the free_list_lock"); 1990 1991 case Resizing: { 1992 // Sweeping has been completed... 1993 // At this point the background collection has completed. 1994 // Don't move the call to compute_new_size() down 1995 // into code that might be executed if the background 1996 // collection was preempted. 1997 { 1998 ReleaseForegroundGC x(this); // unblock FG collection 1999 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2000 CMSTokenSync z(true); // not strictly needed. 2001 if (_collectorState == Resizing) { 2002 compute_new_size(); 2003 save_heap_summary(); 2004 _collectorState = Resetting; 2005 } else { 2006 assert(_collectorState == Idling, "The state should only change" 2007 " because the foreground collector has finished the collection"); 2008 } 2009 } 2010 break; 2011 } 2012 case Resetting: 2013 // CMS heap resizing has been completed 2014 reset(true); 2015 assert(_collectorState == Idling, "Collector state should " 2016 "have changed"); 2017 2018 MetaspaceGC::set_should_concurrent_collect(false); 2019 2020 stats().record_cms_end(); 2021 // Don't move the concurrent_phases_end() and compute_new_size() 2022 // calls to here because a preempted background collection 2023 // has it's state set to "Resetting". 2024 break; 2025 case Idling: 2026 default: 2027 ShouldNotReachHere(); 2028 break; 2029 } 2030 if (TraceCMSState) { 2031 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2032 p2i(Thread::current()), _collectorState); 2033 } 2034 assert(_foregroundGCShouldWait, "block post-condition"); 2035 } 2036 2037 // Should this be in gc_epilogue? 2038 collector_policy()->counters()->update_counters(); 2039 2040 { 2041 // Clear _foregroundGCShouldWait and, in the event that the 2042 // foreground collector is waiting, notify it, before 2043 // returning. 2044 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2045 _foregroundGCShouldWait = false; 2046 if (_foregroundGCIsActive) { 2047 CGC_lock->notify(); 2048 } 2049 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2050 "Possible deadlock"); 2051 } 2052 if (TraceCMSState) { 2053 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2054 " exiting collection CMS state %d", 2055 p2i(Thread::current()), _collectorState); 2056 } 2057 if (PrintGC && Verbose) { 2058 _cmsGen->print_heap_change(prev_used); 2059 } 2060 } 2061 2062 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2063 _cms_start_registered = true; 2064 _gc_timer_cm->register_gc_start(); 2065 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2066 } 2067 2068 void CMSCollector::register_gc_end() { 2069 if (_cms_start_registered) { 2070 report_heap_summary(GCWhen::AfterGC); 2071 2072 _gc_timer_cm->register_gc_end(); 2073 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2074 _cms_start_registered = false; 2075 } 2076 } 2077 2078 void CMSCollector::save_heap_summary() { 2079 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2080 _last_heap_summary = gch->create_heap_summary(); 2081 _last_metaspace_summary = gch->create_metaspace_summary(); 2082 } 2083 2084 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2085 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2086 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2087 } 2088 2089 bool CMSCollector::waitForForegroundGC() { 2090 bool res = false; 2091 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2092 "CMS thread should have CMS token"); 2093 // Block the foreground collector until the 2094 // background collectors decides whether to 2095 // yield. 2096 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2097 _foregroundGCShouldWait = true; 2098 if (_foregroundGCIsActive) { 2099 // The background collector yields to the 2100 // foreground collector and returns a value 2101 // indicating that it has yielded. The foreground 2102 // collector can proceed. 2103 res = true; 2104 _foregroundGCShouldWait = false; 2105 ConcurrentMarkSweepThread::clear_CMS_flag( 2106 ConcurrentMarkSweepThread::CMS_cms_has_token); 2107 ConcurrentMarkSweepThread::set_CMS_flag( 2108 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2109 // Get a possibly blocked foreground thread going 2110 CGC_lock->notify(); 2111 if (TraceCMSState) { 2112 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2113 p2i(Thread::current()), _collectorState); 2114 } 2115 while (_foregroundGCIsActive) { 2116 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2117 } 2118 ConcurrentMarkSweepThread::set_CMS_flag( 2119 ConcurrentMarkSweepThread::CMS_cms_has_token); 2120 ConcurrentMarkSweepThread::clear_CMS_flag( 2121 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2122 } 2123 if (TraceCMSState) { 2124 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2125 p2i(Thread::current()), _collectorState); 2126 } 2127 return res; 2128 } 2129 2130 // Because of the need to lock the free lists and other structures in 2131 // the collector, common to all the generations that the collector is 2132 // collecting, we need the gc_prologues of individual CMS generations 2133 // delegate to their collector. It may have been simpler had the 2134 // current infrastructure allowed one to call a prologue on a 2135 // collector. In the absence of that we have the generation's 2136 // prologue delegate to the collector, which delegates back 2137 // some "local" work to a worker method in the individual generations 2138 // that it's responsible for collecting, while itself doing any 2139 // work common to all generations it's responsible for. A similar 2140 // comment applies to the gc_epilogue()'s. 2141 // The role of the variable _between_prologue_and_epilogue is to 2142 // enforce the invocation protocol. 2143 void CMSCollector::gc_prologue(bool full) { 2144 // Call gc_prologue_work() for the CMSGen 2145 // we are responsible for. 2146 2147 // The following locking discipline assumes that we are only called 2148 // when the world is stopped. 2149 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2150 2151 // The CMSCollector prologue must call the gc_prologues for the 2152 // "generations" that it's responsible 2153 // for. 2154 2155 assert( Thread::current()->is_VM_thread() 2156 || ( CMSScavengeBeforeRemark 2157 && Thread::current()->is_ConcurrentGC_thread()), 2158 "Incorrect thread type for prologue execution"); 2159 2160 if (_between_prologue_and_epilogue) { 2161 // We have already been invoked; this is a gc_prologue delegation 2162 // from yet another CMS generation that we are responsible for, just 2163 // ignore it since all relevant work has already been done. 2164 return; 2165 } 2166 2167 // set a bit saying prologue has been called; cleared in epilogue 2168 _between_prologue_and_epilogue = true; 2169 // Claim locks for common data structures, then call gc_prologue_work() 2170 // for each CMSGen. 2171 2172 getFreelistLocks(); // gets free list locks on constituent spaces 2173 bitMapLock()->lock_without_safepoint_check(); 2174 2175 // Should call gc_prologue_work() for all cms gens we are responsible for 2176 bool duringMarking = _collectorState >= Marking 2177 && _collectorState < Sweeping; 2178 2179 // The young collections clear the modified oops state, which tells if 2180 // there are any modified oops in the class. The remark phase also needs 2181 // that information. Tell the young collection to save the union of all 2182 // modified klasses. 2183 if (duringMarking) { 2184 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2185 } 2186 2187 bool registerClosure = duringMarking; 2188 2189 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2190 &_modUnionClosurePar 2191 : &_modUnionClosure; 2192 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2193 2194 if (!full) { 2195 stats().record_gc0_begin(); 2196 } 2197 } 2198 2199 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2200 2201 _capacity_at_prologue = capacity(); 2202 _used_at_prologue = used(); 2203 2204 // Delegate to CMScollector which knows how to coordinate between 2205 // this and any other CMS generations that it is responsible for 2206 // collecting. 2207 collector()->gc_prologue(full); 2208 } 2209 2210 // This is a "private" interface for use by this generation's CMSCollector. 2211 // Not to be called directly by any other entity (for instance, 2212 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2213 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2214 bool registerClosure, ModUnionClosure* modUnionClosure) { 2215 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2216 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2217 "Should be NULL"); 2218 if (registerClosure) { 2219 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2220 } 2221 cmsSpace()->gc_prologue(); 2222 // Clear stat counters 2223 NOT_PRODUCT( 2224 assert(_numObjectsPromoted == 0, "check"); 2225 assert(_numWordsPromoted == 0, "check"); 2226 if (Verbose && PrintGC) { 2227 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2228 SIZE_FORMAT" bytes concurrently", 2229 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2230 } 2231 _numObjectsAllocated = 0; 2232 _numWordsAllocated = 0; 2233 ) 2234 } 2235 2236 void CMSCollector::gc_epilogue(bool full) { 2237 // The following locking discipline assumes that we are only called 2238 // when the world is stopped. 2239 assert(SafepointSynchronize::is_at_safepoint(), 2240 "world is stopped assumption"); 2241 2242 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2243 // if linear allocation blocks need to be appropriately marked to allow the 2244 // the blocks to be parsable. We also check here whether we need to nudge the 2245 // CMS collector thread to start a new cycle (if it's not already active). 2246 assert( Thread::current()->is_VM_thread() 2247 || ( CMSScavengeBeforeRemark 2248 && Thread::current()->is_ConcurrentGC_thread()), 2249 "Incorrect thread type for epilogue execution"); 2250 2251 if (!_between_prologue_and_epilogue) { 2252 // We have already been invoked; this is a gc_epilogue delegation 2253 // from yet another CMS generation that we are responsible for, just 2254 // ignore it since all relevant work has already been done. 2255 return; 2256 } 2257 assert(haveFreelistLocks(), "must have freelist locks"); 2258 assert_lock_strong(bitMapLock()); 2259 2260 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2261 2262 _cmsGen->gc_epilogue_work(full); 2263 2264 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2265 // in case sampling was not already enabled, enable it 2266 _start_sampling = true; 2267 } 2268 // reset _eden_chunk_array so sampling starts afresh 2269 _eden_chunk_index = 0; 2270 2271 size_t cms_used = _cmsGen->cmsSpace()->used(); 2272 2273 // update performance counters - this uses a special version of 2274 // update_counters() that allows the utilization to be passed as a 2275 // parameter, avoiding multiple calls to used(). 2276 // 2277 _cmsGen->update_counters(cms_used); 2278 2279 bitMapLock()->unlock(); 2280 releaseFreelistLocks(); 2281 2282 if (!CleanChunkPoolAsync) { 2283 Chunk::clean_chunk_pool(); 2284 } 2285 2286 set_did_compact(false); 2287 _between_prologue_and_epilogue = false; // ready for next cycle 2288 } 2289 2290 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2291 collector()->gc_epilogue(full); 2292 2293 // Also reset promotion tracking in par gc thread states. 2294 for (uint i = 0; i < ParallelGCThreads; i++) { 2295 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2296 } 2297 } 2298 2299 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2300 assert(!incremental_collection_failed(), "Should have been cleared"); 2301 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2302 cmsSpace()->gc_epilogue(); 2303 // Print stat counters 2304 NOT_PRODUCT( 2305 assert(_numObjectsAllocated == 0, "check"); 2306 assert(_numWordsAllocated == 0, "check"); 2307 if (Verbose && PrintGC) { 2308 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2309 SIZE_FORMAT" bytes", 2310 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2311 } 2312 _numObjectsPromoted = 0; 2313 _numWordsPromoted = 0; 2314 ) 2315 2316 if (PrintGC && Verbose) { 2317 // Call down the chain in contiguous_available needs the freelistLock 2318 // so print this out before releasing the freeListLock. 2319 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2320 contiguous_available()); 2321 } 2322 } 2323 2324 #ifndef PRODUCT 2325 bool CMSCollector::have_cms_token() { 2326 Thread* thr = Thread::current(); 2327 if (thr->is_VM_thread()) { 2328 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2329 } else if (thr->is_ConcurrentGC_thread()) { 2330 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2331 } else if (thr->is_GC_task_thread()) { 2332 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2333 ParGCRareEvent_lock->owned_by_self(); 2334 } 2335 return false; 2336 } 2337 #endif 2338 2339 // Check reachability of the given heap address in CMS generation, 2340 // treating all other generations as roots. 2341 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2342 // We could "guarantee" below, rather than assert, but I'll 2343 // leave these as "asserts" so that an adventurous debugger 2344 // could try this in the product build provided some subset of 2345 // the conditions were met, provided they were interested in the 2346 // results and knew that the computation below wouldn't interfere 2347 // with other concurrent computations mutating the structures 2348 // being read or written. 2349 assert(SafepointSynchronize::is_at_safepoint(), 2350 "Else mutations in object graph will make answer suspect"); 2351 assert(have_cms_token(), "Should hold cms token"); 2352 assert(haveFreelistLocks(), "must hold free list locks"); 2353 assert_lock_strong(bitMapLock()); 2354 2355 // Clear the marking bit map array before starting, but, just 2356 // for kicks, first report if the given address is already marked 2357 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2358 _markBitMap.isMarked(addr) ? "" : " not"); 2359 2360 if (verify_after_remark()) { 2361 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2362 bool result = verification_mark_bm()->isMarked(addr); 2363 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2364 result ? "IS" : "is NOT"); 2365 return result; 2366 } else { 2367 gclog_or_tty->print_cr("Could not compute result"); 2368 return false; 2369 } 2370 } 2371 2372 2373 void 2374 CMSCollector::print_on_error(outputStream* st) { 2375 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2376 if (collector != NULL) { 2377 CMSBitMap* bitmap = &collector->_markBitMap; 2378 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2379 bitmap->print_on_error(st, " Bits: "); 2380 2381 st->cr(); 2382 2383 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2384 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2385 mut_bitmap->print_on_error(st, " Bits: "); 2386 } 2387 } 2388 2389 //////////////////////////////////////////////////////// 2390 // CMS Verification Support 2391 //////////////////////////////////////////////////////// 2392 // Following the remark phase, the following invariant 2393 // should hold -- each object in the CMS heap which is 2394 // marked in markBitMap() should be marked in the verification_mark_bm(). 2395 2396 class VerifyMarkedClosure: public BitMapClosure { 2397 CMSBitMap* _marks; 2398 bool _failed; 2399 2400 public: 2401 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2402 2403 bool do_bit(size_t offset) { 2404 HeapWord* addr = _marks->offsetToHeapWord(offset); 2405 if (!_marks->isMarked(addr)) { 2406 oop(addr)->print_on(gclog_or_tty); 2407 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2408 _failed = true; 2409 } 2410 return true; 2411 } 2412 2413 bool failed() { return _failed; } 2414 }; 2415 2416 bool CMSCollector::verify_after_remark(bool silent) { 2417 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2418 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2419 static bool init = false; 2420 2421 assert(SafepointSynchronize::is_at_safepoint(), 2422 "Else mutations in object graph will make answer suspect"); 2423 assert(have_cms_token(), 2424 "Else there may be mutual interference in use of " 2425 " verification data structures"); 2426 assert(_collectorState > Marking && _collectorState <= Sweeping, 2427 "Else marking info checked here may be obsolete"); 2428 assert(haveFreelistLocks(), "must hold free list locks"); 2429 assert_lock_strong(bitMapLock()); 2430 2431 2432 // Allocate marking bit map if not already allocated 2433 if (!init) { // first time 2434 if (!verification_mark_bm()->allocate(_span)) { 2435 return false; 2436 } 2437 init = true; 2438 } 2439 2440 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2441 2442 // Turn off refs discovery -- so we will be tracing through refs. 2443 // This is as intended, because by this time 2444 // GC must already have cleared any refs that need to be cleared, 2445 // and traced those that need to be marked; moreover, 2446 // the marking done here is not going to interfere in any 2447 // way with the marking information used by GC. 2448 NoRefDiscovery no_discovery(ref_processor()); 2449 2450 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2451 2452 // Clear any marks from a previous round 2453 verification_mark_bm()->clear_all(); 2454 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2455 verify_work_stacks_empty(); 2456 2457 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2458 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2459 // Update the saved marks which may affect the root scans. 2460 gch->save_marks(); 2461 2462 if (CMSRemarkVerifyVariant == 1) { 2463 // In this first variant of verification, we complete 2464 // all marking, then check if the new marks-vector is 2465 // a subset of the CMS marks-vector. 2466 verify_after_remark_work_1(); 2467 } else if (CMSRemarkVerifyVariant == 2) { 2468 // In this second variant of verification, we flag an error 2469 // (i.e. an object reachable in the new marks-vector not reachable 2470 // in the CMS marks-vector) immediately, also indicating the 2471 // identify of an object (A) that references the unmarked object (B) -- 2472 // presumably, a mutation to A failed to be picked up by preclean/remark? 2473 verify_after_remark_work_2(); 2474 } else { 2475 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2476 CMSRemarkVerifyVariant); 2477 } 2478 if (!silent) gclog_or_tty->print(" done] "); 2479 return true; 2480 } 2481 2482 void CMSCollector::verify_after_remark_work_1() { 2483 ResourceMark rm; 2484 HandleMark hm; 2485 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2486 2487 // Get a clear set of claim bits for the roots processing to work with. 2488 ClassLoaderDataGraph::clear_claimed_marks(); 2489 2490 // Mark from roots one level into CMS 2491 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2492 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2493 2494 gch->gen_process_roots(_cmsGen->level(), 2495 true, // younger gens are roots 2496 true, // activate StrongRootsScope 2497 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2498 should_unload_classes(), 2499 ¬Older, 2500 NULL, 2501 NULL); // SSS: Provide correct closure 2502 2503 // Now mark from the roots 2504 MarkFromRootsClosure markFromRootsClosure(this, _span, 2505 verification_mark_bm(), verification_mark_stack(), 2506 false /* don't yield */, true /* verifying */); 2507 assert(_restart_addr == NULL, "Expected pre-condition"); 2508 verification_mark_bm()->iterate(&markFromRootsClosure); 2509 while (_restart_addr != NULL) { 2510 // Deal with stack overflow: by restarting at the indicated 2511 // address. 2512 HeapWord* ra = _restart_addr; 2513 markFromRootsClosure.reset(ra); 2514 _restart_addr = NULL; 2515 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2516 } 2517 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2518 verify_work_stacks_empty(); 2519 2520 // Marking completed -- now verify that each bit marked in 2521 // verification_mark_bm() is also marked in markBitMap(); flag all 2522 // errors by printing corresponding objects. 2523 VerifyMarkedClosure vcl(markBitMap()); 2524 verification_mark_bm()->iterate(&vcl); 2525 if (vcl.failed()) { 2526 gclog_or_tty->print("Verification failed"); 2527 Universe::heap()->print_on(gclog_or_tty); 2528 fatal("CMS: failed marking verification after remark"); 2529 } 2530 } 2531 2532 class VerifyKlassOopsKlassClosure : public KlassClosure { 2533 class VerifyKlassOopsClosure : public OopClosure { 2534 CMSBitMap* _bitmap; 2535 public: 2536 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2537 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2538 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2539 } _oop_closure; 2540 public: 2541 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2542 void do_klass(Klass* k) { 2543 k->oops_do(&_oop_closure); 2544 } 2545 }; 2546 2547 void CMSCollector::verify_after_remark_work_2() { 2548 ResourceMark rm; 2549 HandleMark hm; 2550 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2551 2552 // Get a clear set of claim bits for the roots processing to work with. 2553 ClassLoaderDataGraph::clear_claimed_marks(); 2554 2555 // Mark from roots one level into CMS 2556 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2557 markBitMap()); 2558 CLDToOopClosure cld_closure(¬Older, true); 2559 2560 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2561 2562 gch->gen_process_roots(_cmsGen->level(), 2563 true, // younger gens are roots 2564 true, // activate StrongRootsScope 2565 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2566 should_unload_classes(), 2567 ¬Older, 2568 NULL, 2569 &cld_closure); 2570 2571 // Now mark from the roots 2572 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2573 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2574 assert(_restart_addr == NULL, "Expected pre-condition"); 2575 verification_mark_bm()->iterate(&markFromRootsClosure); 2576 while (_restart_addr != NULL) { 2577 // Deal with stack overflow: by restarting at the indicated 2578 // address. 2579 HeapWord* ra = _restart_addr; 2580 markFromRootsClosure.reset(ra); 2581 _restart_addr = NULL; 2582 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2583 } 2584 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2585 verify_work_stacks_empty(); 2586 2587 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2588 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2589 2590 // Marking completed -- now verify that each bit marked in 2591 // verification_mark_bm() is also marked in markBitMap(); flag all 2592 // errors by printing corresponding objects. 2593 VerifyMarkedClosure vcl(markBitMap()); 2594 verification_mark_bm()->iterate(&vcl); 2595 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2596 } 2597 2598 void ConcurrentMarkSweepGeneration::save_marks() { 2599 // delegate to CMS space 2600 cmsSpace()->save_marks(); 2601 for (uint i = 0; i < ParallelGCThreads; i++) { 2602 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2603 } 2604 } 2605 2606 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2607 return cmsSpace()->no_allocs_since_save_marks(); 2608 } 2609 2610 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2611 \ 2612 void ConcurrentMarkSweepGeneration:: \ 2613 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2614 cl->set_generation(this); \ 2615 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2616 cl->reset_generation(); \ 2617 save_marks(); \ 2618 } 2619 2620 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2621 2622 void 2623 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2624 if (freelistLock()->owned_by_self()) { 2625 Generation::oop_iterate(cl); 2626 } else { 2627 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2628 Generation::oop_iterate(cl); 2629 } 2630 } 2631 2632 void 2633 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2634 if (freelistLock()->owned_by_self()) { 2635 Generation::object_iterate(cl); 2636 } else { 2637 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2638 Generation::object_iterate(cl); 2639 } 2640 } 2641 2642 void 2643 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2644 if (freelistLock()->owned_by_self()) { 2645 Generation::safe_object_iterate(cl); 2646 } else { 2647 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2648 Generation::safe_object_iterate(cl); 2649 } 2650 } 2651 2652 void 2653 ConcurrentMarkSweepGeneration::post_compact() { 2654 } 2655 2656 void 2657 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2658 // Fix the linear allocation blocks to look like free blocks. 2659 2660 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2661 // are not called when the heap is verified during universe initialization and 2662 // at vm shutdown. 2663 if (freelistLock()->owned_by_self()) { 2664 cmsSpace()->prepare_for_verify(); 2665 } else { 2666 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2667 cmsSpace()->prepare_for_verify(); 2668 } 2669 } 2670 2671 void 2672 ConcurrentMarkSweepGeneration::verify() { 2673 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2674 // are not called when the heap is verified during universe initialization and 2675 // at vm shutdown. 2676 if (freelistLock()->owned_by_self()) { 2677 cmsSpace()->verify(); 2678 } else { 2679 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2680 cmsSpace()->verify(); 2681 } 2682 } 2683 2684 void CMSCollector::verify() { 2685 _cmsGen->verify(); 2686 } 2687 2688 #ifndef PRODUCT 2689 bool CMSCollector::overflow_list_is_empty() const { 2690 assert(_num_par_pushes >= 0, "Inconsistency"); 2691 if (_overflow_list == NULL) { 2692 assert(_num_par_pushes == 0, "Inconsistency"); 2693 } 2694 return _overflow_list == NULL; 2695 } 2696 2697 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2698 // merely consolidate assertion checks that appear to occur together frequently. 2699 void CMSCollector::verify_work_stacks_empty() const { 2700 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2701 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2702 } 2703 2704 void CMSCollector::verify_overflow_empty() const { 2705 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2706 assert(no_preserved_marks(), "No preserved marks"); 2707 } 2708 #endif // PRODUCT 2709 2710 // Decide if we want to enable class unloading as part of the 2711 // ensuing concurrent GC cycle. We will collect and 2712 // unload classes if it's the case that: 2713 // (1) an explicit gc request has been made and the flag 2714 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2715 // (2) (a) class unloading is enabled at the command line, and 2716 // (b) old gen is getting really full 2717 // NOTE: Provided there is no change in the state of the heap between 2718 // calls to this method, it should have idempotent results. Moreover, 2719 // its results should be monotonically increasing (i.e. going from 0 to 1, 2720 // but not 1 to 0) between successive calls between which the heap was 2721 // not collected. For the implementation below, it must thus rely on 2722 // the property that concurrent_cycles_since_last_unload() 2723 // will not decrease unless a collection cycle happened and that 2724 // _cmsGen->is_too_full() are 2725 // themselves also monotonic in that sense. See check_monotonicity() 2726 // below. 2727 void CMSCollector::update_should_unload_classes() { 2728 _should_unload_classes = false; 2729 // Condition 1 above 2730 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2731 _should_unload_classes = true; 2732 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2733 // Disjuncts 2.b.(i,ii,iii) above 2734 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2735 CMSClassUnloadingMaxInterval) 2736 || _cmsGen->is_too_full(); 2737 } 2738 } 2739 2740 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2741 bool res = should_concurrent_collect(); 2742 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2743 return res; 2744 } 2745 2746 void CMSCollector::setup_cms_unloading_and_verification_state() { 2747 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2748 || VerifyBeforeExit; 2749 const int rso = GenCollectedHeap::SO_AllCodeCache; 2750 2751 // We set the proper root for this CMS cycle here. 2752 if (should_unload_classes()) { // Should unload classes this cycle 2753 remove_root_scanning_option(rso); // Shrink the root set appropriately 2754 set_verifying(should_verify); // Set verification state for this cycle 2755 return; // Nothing else needs to be done at this time 2756 } 2757 2758 // Not unloading classes this cycle 2759 assert(!should_unload_classes(), "Inconsistency!"); 2760 2761 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2762 // Include symbols, strings and code cache elements to prevent their resurrection. 2763 add_root_scanning_option(rso); 2764 set_verifying(true); 2765 } else if (verifying() && !should_verify) { 2766 // We were verifying, but some verification flags got disabled. 2767 set_verifying(false); 2768 // Exclude symbols, strings and code cache elements from root scanning to 2769 // reduce IM and RM pauses. 2770 remove_root_scanning_option(rso); 2771 } 2772 } 2773 2774 2775 #ifndef PRODUCT 2776 HeapWord* CMSCollector::block_start(const void* p) const { 2777 const HeapWord* addr = (HeapWord*)p; 2778 if (_span.contains(p)) { 2779 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2780 return _cmsGen->cmsSpace()->block_start(p); 2781 } 2782 } 2783 return NULL; 2784 } 2785 #endif 2786 2787 HeapWord* 2788 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2789 bool tlab, 2790 bool parallel) { 2791 CMSSynchronousYieldRequest yr; 2792 assert(!tlab, "Can't deal with TLAB allocation"); 2793 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2794 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2795 if (GCExpandToAllocateDelayMillis > 0) { 2796 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2797 } 2798 return have_lock_and_allocate(word_size, tlab); 2799 } 2800 2801 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2802 size_t bytes, 2803 size_t expand_bytes, 2804 CMSExpansionCause::Cause cause) 2805 { 2806 2807 bool success = expand(bytes, expand_bytes); 2808 2809 // remember why we expanded; this information is used 2810 // by shouldConcurrentCollect() when making decisions on whether to start 2811 // a new CMS cycle. 2812 if (success) { 2813 set_expansion_cause(cause); 2814 if (PrintGCDetails && Verbose) { 2815 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2816 CMSExpansionCause::to_string(cause)); 2817 } 2818 } 2819 } 2820 2821 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2822 HeapWord* res = NULL; 2823 MutexLocker x(ParGCRareEvent_lock); 2824 while (true) { 2825 // Expansion by some other thread might make alloc OK now: 2826 res = ps->lab.alloc(word_sz); 2827 if (res != NULL) return res; 2828 // If there's not enough expansion space available, give up. 2829 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2830 return NULL; 2831 } 2832 // Otherwise, we try expansion. 2833 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2834 // Now go around the loop and try alloc again; 2835 // A competing par_promote might beat us to the expansion space, 2836 // so we may go around the loop again if promotion fails again. 2837 if (GCExpandToAllocateDelayMillis > 0) { 2838 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2839 } 2840 } 2841 } 2842 2843 2844 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2845 PromotionInfo* promo) { 2846 MutexLocker x(ParGCRareEvent_lock); 2847 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2848 while (true) { 2849 // Expansion by some other thread might make alloc OK now: 2850 if (promo->ensure_spooling_space()) { 2851 assert(promo->has_spooling_space(), 2852 "Post-condition of successful ensure_spooling_space()"); 2853 return true; 2854 } 2855 // If there's not enough expansion space available, give up. 2856 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2857 return false; 2858 } 2859 // Otherwise, we try expansion. 2860 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2861 // Now go around the loop and try alloc again; 2862 // A competing allocation might beat us to the expansion space, 2863 // so we may go around the loop again if allocation fails again. 2864 if (GCExpandToAllocateDelayMillis > 0) { 2865 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2866 } 2867 } 2868 } 2869 2870 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2871 // Only shrink if a compaction was done so that all the free space 2872 // in the generation is in a contiguous block at the end. 2873 if (did_compact()) { 2874 CardGeneration::shrink(bytes); 2875 } 2876 } 2877 2878 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2879 assert_locked_or_safepoint(Heap_lock); 2880 } 2881 2882 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2883 assert_locked_or_safepoint(Heap_lock); 2884 assert_lock_strong(freelistLock()); 2885 if (PrintGCDetails && Verbose) { 2886 warning("Shrinking of CMS not yet implemented"); 2887 } 2888 return; 2889 } 2890 2891 2892 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2893 // phases. 2894 class CMSPhaseAccounting: public StackObj { 2895 public: 2896 CMSPhaseAccounting(CMSCollector *collector, 2897 const char *phase, 2898 const GCId gc_id, 2899 bool print_cr = true); 2900 ~CMSPhaseAccounting(); 2901 2902 private: 2903 CMSCollector *_collector; 2904 const char *_phase; 2905 elapsedTimer _wallclock; 2906 bool _print_cr; 2907 const GCId _gc_id; 2908 2909 public: 2910 // Not MT-safe; so do not pass around these StackObj's 2911 // where they may be accessed by other threads. 2912 jlong wallclock_millis() { 2913 assert(_wallclock.is_active(), "Wall clock should not stop"); 2914 _wallclock.stop(); // to record time 2915 jlong ret = _wallclock.milliseconds(); 2916 _wallclock.start(); // restart 2917 return ret; 2918 } 2919 }; 2920 2921 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2922 const char *phase, 2923 const GCId gc_id, 2924 bool print_cr) : 2925 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2926 2927 if (PrintCMSStatistics != 0) { 2928 _collector->resetYields(); 2929 } 2930 if (PrintGCDetails) { 2931 gclog_or_tty->gclog_stamp(_gc_id); 2932 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2933 _collector->cmsGen()->short_name(), _phase); 2934 } 2935 _collector->resetTimer(); 2936 _wallclock.start(); 2937 _collector->startTimer(); 2938 } 2939 2940 CMSPhaseAccounting::~CMSPhaseAccounting() { 2941 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2942 _collector->stopTimer(); 2943 _wallclock.stop(); 2944 if (PrintGCDetails) { 2945 gclog_or_tty->gclog_stamp(_gc_id); 2946 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2947 _collector->cmsGen()->short_name(), 2948 _phase, _collector->timerValue(), _wallclock.seconds()); 2949 if (_print_cr) { 2950 gclog_or_tty->cr(); 2951 } 2952 if (PrintCMSStatistics != 0) { 2953 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2954 _collector->yields()); 2955 } 2956 } 2957 } 2958 2959 // CMS work 2960 2961 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2962 class CMSParMarkTask : public AbstractGangTask { 2963 protected: 2964 CMSCollector* _collector; 2965 int _n_workers; 2966 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 2967 AbstractGangTask(name), 2968 _collector(collector), 2969 _n_workers(n_workers) {} 2970 // Work method in support of parallel rescan ... of young gen spaces 2971 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2972 ContiguousSpace* space, 2973 HeapWord** chunk_array, size_t chunk_top); 2974 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2975 }; 2976 2977 // Parallel initial mark task 2978 class CMSParInitialMarkTask: public CMSParMarkTask { 2979 public: 2980 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 2981 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2982 collector, n_workers) {} 2983 void work(uint worker_id); 2984 }; 2985 2986 // Checkpoint the roots into this generation from outside 2987 // this generation. [Note this initial checkpoint need only 2988 // be approximate -- we'll do a catch up phase subsequently.] 2989 void CMSCollector::checkpointRootsInitial() { 2990 assert(_collectorState == InitialMarking, "Wrong collector state"); 2991 check_correct_thread_executing(); 2992 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2993 2994 save_heap_summary(); 2995 report_heap_summary(GCWhen::BeforeGC); 2996 2997 ReferenceProcessor* rp = ref_processor(); 2998 assert(_restart_addr == NULL, "Control point invariant"); 2999 { 3000 // acquire locks for subsequent manipulations 3001 MutexLockerEx x(bitMapLock(), 3002 Mutex::_no_safepoint_check_flag); 3003 checkpointRootsInitialWork(); 3004 // enable ("weak") refs discovery 3005 rp->enable_discovery(); 3006 _collectorState = Marking; 3007 } 3008 } 3009 3010 void CMSCollector::checkpointRootsInitialWork() { 3011 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3012 assert(_collectorState == InitialMarking, "just checking"); 3013 3014 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3015 // precede our marking with a collection of all 3016 // younger generations to keep floating garbage to a minimum. 3017 // XXX: we won't do this for now -- it's an optimization to be done later. 3018 3019 // already have locks 3020 assert_lock_strong(bitMapLock()); 3021 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3022 3023 // Setup the verification and class unloading state for this 3024 // CMS collection cycle. 3025 setup_cms_unloading_and_verification_state(); 3026 3027 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3028 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3029 3030 // Reset all the PLAB chunk arrays if necessary. 3031 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3032 reset_survivor_plab_arrays(); 3033 } 3034 3035 ResourceMark rm; 3036 HandleMark hm; 3037 3038 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3039 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3040 3041 verify_work_stacks_empty(); 3042 verify_overflow_empty(); 3043 3044 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3045 // Update the saved marks which may affect the root scans. 3046 gch->save_marks(); 3047 3048 // weak reference processing has not started yet. 3049 ref_processor()->set_enqueuing_is_done(false); 3050 3051 // Need to remember all newly created CLDs, 3052 // so that we can guarantee that the remark finds them. 3053 ClassLoaderDataGraph::remember_new_clds(true); 3054 3055 // Whenever a CLD is found, it will be claimed before proceeding to mark 3056 // the klasses. The claimed marks need to be cleared before marking starts. 3057 ClassLoaderDataGraph::clear_claimed_marks(); 3058 3059 if (CMSPrintEdenSurvivorChunks) { 3060 print_eden_and_survivor_chunk_arrays(); 3061 } 3062 3063 { 3064 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3065 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3066 // The parallel version. 3067 FlexibleWorkGang* workers = gch->workers(); 3068 assert(workers != NULL, "Need parallel worker threads."); 3069 int n_workers = workers->active_workers(); 3070 CMSParInitialMarkTask tsk(this, n_workers); 3071 gch->set_par_threads(n_workers); 3072 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3073 if (n_workers > 1) { 3074 GenCollectedHeap::StrongRootsScope srs; 3075 workers->run_task(&tsk); 3076 } else { 3077 GenCollectedHeap::StrongRootsScope srs; 3078 tsk.work(0); 3079 } 3080 gch->set_par_threads(0); 3081 } else { 3082 // The serial version. 3083 CLDToOopClosure cld_closure(¬Older, true); 3084 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3085 gch->gen_process_roots(_cmsGen->level(), 3086 true, // younger gens are roots 3087 true, // activate StrongRootsScope 3088 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3089 should_unload_classes(), 3090 ¬Older, 3091 NULL, 3092 &cld_closure); 3093 } 3094 } 3095 3096 // Clear mod-union table; it will be dirtied in the prologue of 3097 // CMS generation per each younger generation collection. 3098 3099 assert(_modUnionTable.isAllClear(), 3100 "Was cleared in most recent final checkpoint phase" 3101 " or no bits are set in the gc_prologue before the start of the next " 3102 "subsequent marking phase."); 3103 3104 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3105 3106 // Save the end of the used_region of the constituent generations 3107 // to be used to limit the extent of sweep in each generation. 3108 save_sweep_limits(); 3109 verify_overflow_empty(); 3110 } 3111 3112 bool CMSCollector::markFromRoots() { 3113 // we might be tempted to assert that: 3114 // assert(!SafepointSynchronize::is_at_safepoint(), 3115 // "inconsistent argument?"); 3116 // However that wouldn't be right, because it's possible that 3117 // a safepoint is indeed in progress as a younger generation 3118 // stop-the-world GC happens even as we mark in this generation. 3119 assert(_collectorState == Marking, "inconsistent state?"); 3120 check_correct_thread_executing(); 3121 verify_overflow_empty(); 3122 3123 // Weak ref discovery note: We may be discovering weak 3124 // refs in this generation concurrent (but interleaved) with 3125 // weak ref discovery by a younger generation collector. 3126 3127 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3128 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3129 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3130 bool res = markFromRootsWork(); 3131 if (res) { 3132 _collectorState = Precleaning; 3133 } else { // We failed and a foreground collection wants to take over 3134 assert(_foregroundGCIsActive, "internal state inconsistency"); 3135 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3136 if (PrintGCDetails) { 3137 gclog_or_tty->print_cr("bailing out to foreground collection"); 3138 } 3139 } 3140 verify_overflow_empty(); 3141 return res; 3142 } 3143 3144 bool CMSCollector::markFromRootsWork() { 3145 // iterate over marked bits in bit map, doing a full scan and mark 3146 // from these roots using the following algorithm: 3147 // . if oop is to the right of the current scan pointer, 3148 // mark corresponding bit (we'll process it later) 3149 // . else (oop is to left of current scan pointer) 3150 // push oop on marking stack 3151 // . drain the marking stack 3152 3153 // Note that when we do a marking step we need to hold the 3154 // bit map lock -- recall that direct allocation (by mutators) 3155 // and promotion (by younger generation collectors) is also 3156 // marking the bit map. [the so-called allocate live policy.] 3157 // Because the implementation of bit map marking is not 3158 // robust wrt simultaneous marking of bits in the same word, 3159 // we need to make sure that there is no such interference 3160 // between concurrent such updates. 3161 3162 // already have locks 3163 assert_lock_strong(bitMapLock()); 3164 3165 verify_work_stacks_empty(); 3166 verify_overflow_empty(); 3167 bool result = false; 3168 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3169 result = do_marking_mt(); 3170 } else { 3171 result = do_marking_st(); 3172 } 3173 return result; 3174 } 3175 3176 // Forward decl 3177 class CMSConcMarkingTask; 3178 3179 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3180 CMSCollector* _collector; 3181 CMSConcMarkingTask* _task; 3182 public: 3183 virtual void yield(); 3184 3185 // "n_threads" is the number of threads to be terminated. 3186 // "queue_set" is a set of work queues of other threads. 3187 // "collector" is the CMS collector associated with this task terminator. 3188 // "yield" indicates whether we need the gang as a whole to yield. 3189 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3190 ParallelTaskTerminator(n_threads, queue_set), 3191 _collector(collector) { } 3192 3193 void set_task(CMSConcMarkingTask* task) { 3194 _task = task; 3195 } 3196 }; 3197 3198 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3199 CMSConcMarkingTask* _task; 3200 public: 3201 bool should_exit_termination(); 3202 void set_task(CMSConcMarkingTask* task) { 3203 _task = task; 3204 } 3205 }; 3206 3207 // MT Concurrent Marking Task 3208 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3209 CMSCollector* _collector; 3210 int _n_workers; // requested/desired # workers 3211 bool _result; 3212 CompactibleFreeListSpace* _cms_space; 3213 char _pad_front[64]; // padding to ... 3214 HeapWord* _global_finger; // ... avoid sharing cache line 3215 char _pad_back[64]; 3216 HeapWord* _restart_addr; 3217 3218 // Exposed here for yielding support 3219 Mutex* const _bit_map_lock; 3220 3221 // The per thread work queues, available here for stealing 3222 OopTaskQueueSet* _task_queues; 3223 3224 // Termination (and yielding) support 3225 CMSConcMarkingTerminator _term; 3226 CMSConcMarkingTerminatorTerminator _term_term; 3227 3228 public: 3229 CMSConcMarkingTask(CMSCollector* collector, 3230 CompactibleFreeListSpace* cms_space, 3231 YieldingFlexibleWorkGang* workers, 3232 OopTaskQueueSet* task_queues): 3233 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3234 _collector(collector), 3235 _cms_space(cms_space), 3236 _n_workers(0), _result(true), 3237 _task_queues(task_queues), 3238 _term(_n_workers, task_queues, _collector), 3239 _bit_map_lock(collector->bitMapLock()) 3240 { 3241 _requested_size = _n_workers; 3242 _term.set_task(this); 3243 _term_term.set_task(this); 3244 _restart_addr = _global_finger = _cms_space->bottom(); 3245 } 3246 3247 3248 OopTaskQueueSet* task_queues() { return _task_queues; } 3249 3250 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3251 3252 HeapWord** global_finger_addr() { return &_global_finger; } 3253 3254 CMSConcMarkingTerminator* terminator() { return &_term; } 3255 3256 virtual void set_for_termination(int active_workers) { 3257 terminator()->reset_for_reuse(active_workers); 3258 } 3259 3260 void work(uint worker_id); 3261 bool should_yield() { 3262 return ConcurrentMarkSweepThread::should_yield() 3263 && !_collector->foregroundGCIsActive(); 3264 } 3265 3266 virtual void coordinator_yield(); // stuff done by coordinator 3267 bool result() { return _result; } 3268 3269 void reset(HeapWord* ra) { 3270 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3271 _restart_addr = _global_finger = ra; 3272 _term.reset_for_reuse(); 3273 } 3274 3275 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3276 OopTaskQueue* work_q); 3277 3278 private: 3279 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3280 void do_work_steal(int i); 3281 void bump_global_finger(HeapWord* f); 3282 }; 3283 3284 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3285 assert(_task != NULL, "Error"); 3286 return _task->yielding(); 3287 // Note that we do not need the disjunct || _task->should_yield() above 3288 // because we want terminating threads to yield only if the task 3289 // is already in the midst of yielding, which happens only after at least one 3290 // thread has yielded. 3291 } 3292 3293 void CMSConcMarkingTerminator::yield() { 3294 if (_task->should_yield()) { 3295 _task->yield(); 3296 } else { 3297 ParallelTaskTerminator::yield(); 3298 } 3299 } 3300 3301 //////////////////////////////////////////////////////////////// 3302 // Concurrent Marking Algorithm Sketch 3303 //////////////////////////////////////////////////////////////// 3304 // Until all tasks exhausted (both spaces): 3305 // -- claim next available chunk 3306 // -- bump global finger via CAS 3307 // -- find first object that starts in this chunk 3308 // and start scanning bitmap from that position 3309 // -- scan marked objects for oops 3310 // -- CAS-mark target, and if successful: 3311 // . if target oop is above global finger (volatile read) 3312 // nothing to do 3313 // . if target oop is in chunk and above local finger 3314 // then nothing to do 3315 // . else push on work-queue 3316 // -- Deal with possible overflow issues: 3317 // . local work-queue overflow causes stuff to be pushed on 3318 // global (common) overflow queue 3319 // . always first empty local work queue 3320 // . then get a batch of oops from global work queue if any 3321 // . then do work stealing 3322 // -- When all tasks claimed (both spaces) 3323 // and local work queue empty, 3324 // then in a loop do: 3325 // . check global overflow stack; steal a batch of oops and trace 3326 // . try to steal from other threads oif GOS is empty 3327 // . if neither is available, offer termination 3328 // -- Terminate and return result 3329 // 3330 void CMSConcMarkingTask::work(uint worker_id) { 3331 elapsedTimer _timer; 3332 ResourceMark rm; 3333 HandleMark hm; 3334 3335 DEBUG_ONLY(_collector->verify_overflow_empty();) 3336 3337 // Before we begin work, our work queue should be empty 3338 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3339 // Scan the bitmap covering _cms_space, tracing through grey objects. 3340 _timer.start(); 3341 do_scan_and_mark(worker_id, _cms_space); 3342 _timer.stop(); 3343 if (PrintCMSStatistics != 0) { 3344 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3345 worker_id, _timer.seconds()); 3346 // XXX: need xxx/xxx type of notation, two timers 3347 } 3348 3349 // ... do work stealing 3350 _timer.reset(); 3351 _timer.start(); 3352 do_work_steal(worker_id); 3353 _timer.stop(); 3354 if (PrintCMSStatistics != 0) { 3355 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3356 worker_id, _timer.seconds()); 3357 // XXX: need xxx/xxx type of notation, two timers 3358 } 3359 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3360 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3361 // Note that under the current task protocol, the 3362 // following assertion is true even of the spaces 3363 // expanded since the completion of the concurrent 3364 // marking. XXX This will likely change under a strict 3365 // ABORT semantics. 3366 // After perm removal the comparison was changed to 3367 // greater than or equal to from strictly greater than. 3368 // Before perm removal the highest address sweep would 3369 // have been at the end of perm gen but now is at the 3370 // end of the tenured gen. 3371 assert(_global_finger >= _cms_space->end(), 3372 "All tasks have been completed"); 3373 DEBUG_ONLY(_collector->verify_overflow_empty();) 3374 } 3375 3376 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3377 HeapWord* read = _global_finger; 3378 HeapWord* cur = read; 3379 while (f > read) { 3380 cur = read; 3381 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3382 if (cur == read) { 3383 // our cas succeeded 3384 assert(_global_finger >= f, "protocol consistency"); 3385 break; 3386 } 3387 } 3388 } 3389 3390 // This is really inefficient, and should be redone by 3391 // using (not yet available) block-read and -write interfaces to the 3392 // stack and the work_queue. XXX FIX ME !!! 3393 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3394 OopTaskQueue* work_q) { 3395 // Fast lock-free check 3396 if (ovflw_stk->length() == 0) { 3397 return false; 3398 } 3399 assert(work_q->size() == 0, "Shouldn't steal"); 3400 MutexLockerEx ml(ovflw_stk->par_lock(), 3401 Mutex::_no_safepoint_check_flag); 3402 // Grab up to 1/4 the size of the work queue 3403 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3404 (size_t)ParGCDesiredObjsFromOverflowList); 3405 num = MIN2(num, ovflw_stk->length()); 3406 for (int i = (int) num; i > 0; i--) { 3407 oop cur = ovflw_stk->pop(); 3408 assert(cur != NULL, "Counted wrong?"); 3409 work_q->push(cur); 3410 } 3411 return num > 0; 3412 } 3413 3414 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3415 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3416 int n_tasks = pst->n_tasks(); 3417 // We allow that there may be no tasks to do here because 3418 // we are restarting after a stack overflow. 3419 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3420 uint nth_task = 0; 3421 3422 HeapWord* aligned_start = sp->bottom(); 3423 if (sp->used_region().contains(_restart_addr)) { 3424 // Align down to a card boundary for the start of 0th task 3425 // for this space. 3426 aligned_start = 3427 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3428 CardTableModRefBS::card_size); 3429 } 3430 3431 size_t chunk_size = sp->marking_task_size(); 3432 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3433 // Having claimed the nth task in this space, 3434 // compute the chunk that it corresponds to: 3435 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3436 aligned_start + (nth_task+1)*chunk_size); 3437 // Try and bump the global finger via a CAS; 3438 // note that we need to do the global finger bump 3439 // _before_ taking the intersection below, because 3440 // the task corresponding to that region will be 3441 // deemed done even if the used_region() expands 3442 // because of allocation -- as it almost certainly will 3443 // during start-up while the threads yield in the 3444 // closure below. 3445 HeapWord* finger = span.end(); 3446 bump_global_finger(finger); // atomically 3447 // There are null tasks here corresponding to chunks 3448 // beyond the "top" address of the space. 3449 span = span.intersection(sp->used_region()); 3450 if (!span.is_empty()) { // Non-null task 3451 HeapWord* prev_obj; 3452 assert(!span.contains(_restart_addr) || nth_task == 0, 3453 "Inconsistency"); 3454 if (nth_task == 0) { 3455 // For the 0th task, we'll not need to compute a block_start. 3456 if (span.contains(_restart_addr)) { 3457 // In the case of a restart because of stack overflow, 3458 // we might additionally skip a chunk prefix. 3459 prev_obj = _restart_addr; 3460 } else { 3461 prev_obj = span.start(); 3462 } 3463 } else { 3464 // We want to skip the first object because 3465 // the protocol is to scan any object in its entirety 3466 // that _starts_ in this span; a fortiori, any 3467 // object starting in an earlier span is scanned 3468 // as part of an earlier claimed task. 3469 // Below we use the "careful" version of block_start 3470 // so we do not try to navigate uninitialized objects. 3471 prev_obj = sp->block_start_careful(span.start()); 3472 // Below we use a variant of block_size that uses the 3473 // Printezis bits to avoid waiting for allocated 3474 // objects to become initialized/parsable. 3475 while (prev_obj < span.start()) { 3476 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3477 if (sz > 0) { 3478 prev_obj += sz; 3479 } else { 3480 // In this case we may end up doing a bit of redundant 3481 // scanning, but that appears unavoidable, short of 3482 // locking the free list locks; see bug 6324141. 3483 break; 3484 } 3485 } 3486 } 3487 if (prev_obj < span.end()) { 3488 MemRegion my_span = MemRegion(prev_obj, span.end()); 3489 // Do the marking work within a non-empty span -- 3490 // the last argument to the constructor indicates whether the 3491 // iteration should be incremental with periodic yields. 3492 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3493 &_collector->_markBitMap, 3494 work_queue(i), 3495 &_collector->_markStack); 3496 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3497 } // else nothing to do for this task 3498 } // else nothing to do for this task 3499 } 3500 // We'd be tempted to assert here that since there are no 3501 // more tasks left to claim in this space, the global_finger 3502 // must exceed space->top() and a fortiori space->end(). However, 3503 // that would not quite be correct because the bumping of 3504 // global_finger occurs strictly after the claiming of a task, 3505 // so by the time we reach here the global finger may not yet 3506 // have been bumped up by the thread that claimed the last 3507 // task. 3508 pst->all_tasks_completed(); 3509 } 3510 3511 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3512 private: 3513 CMSCollector* _collector; 3514 CMSConcMarkingTask* _task; 3515 MemRegion _span; 3516 CMSBitMap* _bit_map; 3517 CMSMarkStack* _overflow_stack; 3518 OopTaskQueue* _work_queue; 3519 protected: 3520 DO_OOP_WORK_DEFN 3521 public: 3522 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3523 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3524 MetadataAwareOopClosure(collector->ref_processor()), 3525 _collector(collector), 3526 _task(task), 3527 _span(collector->_span), 3528 _work_queue(work_queue), 3529 _bit_map(bit_map), 3530 _overflow_stack(overflow_stack) 3531 { } 3532 virtual void do_oop(oop* p); 3533 virtual void do_oop(narrowOop* p); 3534 3535 void trim_queue(size_t max); 3536 void handle_stack_overflow(HeapWord* lost); 3537 void do_yield_check() { 3538 if (_task->should_yield()) { 3539 _task->yield(); 3540 } 3541 } 3542 }; 3543 3544 // Grey object scanning during work stealing phase -- 3545 // the salient assumption here is that any references 3546 // that are in these stolen objects being scanned must 3547 // already have been initialized (else they would not have 3548 // been published), so we do not need to check for 3549 // uninitialized objects before pushing here. 3550 void Par_ConcMarkingClosure::do_oop(oop obj) { 3551 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3552 HeapWord* addr = (HeapWord*)obj; 3553 // Check if oop points into the CMS generation 3554 // and is not marked 3555 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3556 // a white object ... 3557 // If we manage to "claim" the object, by being the 3558 // first thread to mark it, then we push it on our 3559 // marking stack 3560 if (_bit_map->par_mark(addr)) { // ... now grey 3561 // push on work queue (grey set) 3562 bool simulate_overflow = false; 3563 NOT_PRODUCT( 3564 if (CMSMarkStackOverflowALot && 3565 _collector->simulate_overflow()) { 3566 // simulate a stack overflow 3567 simulate_overflow = true; 3568 } 3569 ) 3570 if (simulate_overflow || 3571 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3572 // stack overflow 3573 if (PrintCMSStatistics != 0) { 3574 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3575 SIZE_FORMAT, _overflow_stack->capacity()); 3576 } 3577 // We cannot assert that the overflow stack is full because 3578 // it may have been emptied since. 3579 assert(simulate_overflow || 3580 _work_queue->size() == _work_queue->max_elems(), 3581 "Else push should have succeeded"); 3582 handle_stack_overflow(addr); 3583 } 3584 } // Else, some other thread got there first 3585 do_yield_check(); 3586 } 3587 } 3588 3589 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3590 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3591 3592 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3593 while (_work_queue->size() > max) { 3594 oop new_oop; 3595 if (_work_queue->pop_local(new_oop)) { 3596 assert(new_oop->is_oop(), "Should be an oop"); 3597 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3598 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3599 new_oop->oop_iterate(this); // do_oop() above 3600 do_yield_check(); 3601 } 3602 } 3603 } 3604 3605 // Upon stack overflow, we discard (part of) the stack, 3606 // remembering the least address amongst those discarded 3607 // in CMSCollector's _restart_address. 3608 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3609 // We need to do this under a mutex to prevent other 3610 // workers from interfering with the work done below. 3611 MutexLockerEx ml(_overflow_stack->par_lock(), 3612 Mutex::_no_safepoint_check_flag); 3613 // Remember the least grey address discarded 3614 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3615 _collector->lower_restart_addr(ra); 3616 _overflow_stack->reset(); // discard stack contents 3617 _overflow_stack->expand(); // expand the stack if possible 3618 } 3619 3620 3621 void CMSConcMarkingTask::do_work_steal(int i) { 3622 OopTaskQueue* work_q = work_queue(i); 3623 oop obj_to_scan; 3624 CMSBitMap* bm = &(_collector->_markBitMap); 3625 CMSMarkStack* ovflw = &(_collector->_markStack); 3626 int* seed = _collector->hash_seed(i); 3627 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3628 while (true) { 3629 cl.trim_queue(0); 3630 assert(work_q->size() == 0, "Should have been emptied above"); 3631 if (get_work_from_overflow_stack(ovflw, work_q)) { 3632 // Can't assert below because the work obtained from the 3633 // overflow stack may already have been stolen from us. 3634 // assert(work_q->size() > 0, "Work from overflow stack"); 3635 continue; 3636 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3637 assert(obj_to_scan->is_oop(), "Should be an oop"); 3638 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3639 obj_to_scan->oop_iterate(&cl); 3640 } else if (terminator()->offer_termination(&_term_term)) { 3641 assert(work_q->size() == 0, "Impossible!"); 3642 break; 3643 } else if (yielding() || should_yield()) { 3644 yield(); 3645 } 3646 } 3647 } 3648 3649 // This is run by the CMS (coordinator) thread. 3650 void CMSConcMarkingTask::coordinator_yield() { 3651 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3652 "CMS thread should hold CMS token"); 3653 // First give up the locks, then yield, then re-lock 3654 // We should probably use a constructor/destructor idiom to 3655 // do this unlock/lock or modify the MutexUnlocker class to 3656 // serve our purpose. XXX 3657 assert_lock_strong(_bit_map_lock); 3658 _bit_map_lock->unlock(); 3659 ConcurrentMarkSweepThread::desynchronize(true); 3660 _collector->stopTimer(); 3661 if (PrintCMSStatistics != 0) { 3662 _collector->incrementYields(); 3663 } 3664 3665 // It is possible for whichever thread initiated the yield request 3666 // not to get a chance to wake up and take the bitmap lock between 3667 // this thread releasing it and reacquiring it. So, while the 3668 // should_yield() flag is on, let's sleep for a bit to give the 3669 // other thread a chance to wake up. The limit imposed on the number 3670 // of iterations is defensive, to avoid any unforseen circumstances 3671 // putting us into an infinite loop. Since it's always been this 3672 // (coordinator_yield()) method that was observed to cause the 3673 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3674 // which is by default non-zero. For the other seven methods that 3675 // also perform the yield operation, as are using a different 3676 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3677 // can enable the sleeping for those methods too, if necessary. 3678 // See 6442774. 3679 // 3680 // We really need to reconsider the synchronization between the GC 3681 // thread and the yield-requesting threads in the future and we 3682 // should really use wait/notify, which is the recommended 3683 // way of doing this type of interaction. Additionally, we should 3684 // consolidate the eight methods that do the yield operation and they 3685 // are almost identical into one for better maintainability and 3686 // readability. See 6445193. 3687 // 3688 // Tony 2006.06.29 3689 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3690 ConcurrentMarkSweepThread::should_yield() && 3691 !CMSCollector::foregroundGCIsActive(); ++i) { 3692 os::sleep(Thread::current(), 1, false); 3693 } 3694 3695 ConcurrentMarkSweepThread::synchronize(true); 3696 _bit_map_lock->lock_without_safepoint_check(); 3697 _collector->startTimer(); 3698 } 3699 3700 bool CMSCollector::do_marking_mt() { 3701 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3702 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 3703 conc_workers()->total_workers(), 3704 conc_workers()->active_workers(), 3705 Threads::number_of_non_daemon_threads()); 3706 conc_workers()->set_active_workers(num_workers); 3707 3708 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3709 3710 CMSConcMarkingTask tsk(this, 3711 cms_space, 3712 conc_workers(), 3713 task_queues()); 3714 3715 // Since the actual number of workers we get may be different 3716 // from the number we requested above, do we need to do anything different 3717 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3718 // class?? XXX 3719 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3720 3721 // Refs discovery is already non-atomic. 3722 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3723 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3724 conc_workers()->start_task(&tsk); 3725 while (tsk.yielded()) { 3726 tsk.coordinator_yield(); 3727 conc_workers()->continue_task(&tsk); 3728 } 3729 // If the task was aborted, _restart_addr will be non-NULL 3730 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3731 while (_restart_addr != NULL) { 3732 // XXX For now we do not make use of ABORTED state and have not 3733 // yet implemented the right abort semantics (even in the original 3734 // single-threaded CMS case). That needs some more investigation 3735 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3736 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3737 // If _restart_addr is non-NULL, a marking stack overflow 3738 // occurred; we need to do a fresh marking iteration from the 3739 // indicated restart address. 3740 if (_foregroundGCIsActive) { 3741 // We may be running into repeated stack overflows, having 3742 // reached the limit of the stack size, while making very 3743 // slow forward progress. It may be best to bail out and 3744 // let the foreground collector do its job. 3745 // Clear _restart_addr, so that foreground GC 3746 // works from scratch. This avoids the headache of 3747 // a "rescan" which would otherwise be needed because 3748 // of the dirty mod union table & card table. 3749 _restart_addr = NULL; 3750 return false; 3751 } 3752 // Adjust the task to restart from _restart_addr 3753 tsk.reset(_restart_addr); 3754 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3755 _restart_addr); 3756 _restart_addr = NULL; 3757 // Get the workers going again 3758 conc_workers()->start_task(&tsk); 3759 while (tsk.yielded()) { 3760 tsk.coordinator_yield(); 3761 conc_workers()->continue_task(&tsk); 3762 } 3763 } 3764 assert(tsk.completed(), "Inconsistency"); 3765 assert(tsk.result() == true, "Inconsistency"); 3766 return true; 3767 } 3768 3769 bool CMSCollector::do_marking_st() { 3770 ResourceMark rm; 3771 HandleMark hm; 3772 3773 // Temporarily make refs discovery single threaded (non-MT) 3774 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3775 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3776 &_markStack, CMSYield); 3777 // the last argument to iterate indicates whether the iteration 3778 // should be incremental with periodic yields. 3779 _markBitMap.iterate(&markFromRootsClosure); 3780 // If _restart_addr is non-NULL, a marking stack overflow 3781 // occurred; we need to do a fresh iteration from the 3782 // indicated restart address. 3783 while (_restart_addr != NULL) { 3784 if (_foregroundGCIsActive) { 3785 // We may be running into repeated stack overflows, having 3786 // reached the limit of the stack size, while making very 3787 // slow forward progress. It may be best to bail out and 3788 // let the foreground collector do its job. 3789 // Clear _restart_addr, so that foreground GC 3790 // works from scratch. This avoids the headache of 3791 // a "rescan" which would otherwise be needed because 3792 // of the dirty mod union table & card table. 3793 _restart_addr = NULL; 3794 return false; // indicating failure to complete marking 3795 } 3796 // Deal with stack overflow: 3797 // we restart marking from _restart_addr 3798 HeapWord* ra = _restart_addr; 3799 markFromRootsClosure.reset(ra); 3800 _restart_addr = NULL; 3801 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3802 } 3803 return true; 3804 } 3805 3806 void CMSCollector::preclean() { 3807 check_correct_thread_executing(); 3808 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3809 verify_work_stacks_empty(); 3810 verify_overflow_empty(); 3811 _abort_preclean = false; 3812 if (CMSPrecleaningEnabled) { 3813 if (!CMSEdenChunksRecordAlways) { 3814 _eden_chunk_index = 0; 3815 } 3816 size_t used = get_eden_used(); 3817 size_t capacity = get_eden_capacity(); 3818 // Don't start sampling unless we will get sufficiently 3819 // many samples. 3820 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3821 * CMSScheduleRemarkEdenPenetration)) { 3822 _start_sampling = true; 3823 } else { 3824 _start_sampling = false; 3825 } 3826 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3827 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3828 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3829 } 3830 CMSTokenSync x(true); // is cms thread 3831 if (CMSPrecleaningEnabled) { 3832 sample_eden(); 3833 _collectorState = AbortablePreclean; 3834 } else { 3835 _collectorState = FinalMarking; 3836 } 3837 verify_work_stacks_empty(); 3838 verify_overflow_empty(); 3839 } 3840 3841 // Try and schedule the remark such that young gen 3842 // occupancy is CMSScheduleRemarkEdenPenetration %. 3843 void CMSCollector::abortable_preclean() { 3844 check_correct_thread_executing(); 3845 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3846 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3847 3848 // If Eden's current occupancy is below this threshold, 3849 // immediately schedule the remark; else preclean 3850 // past the next scavenge in an effort to 3851 // schedule the pause as described above. By choosing 3852 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3853 // we will never do an actual abortable preclean cycle. 3854 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3855 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3856 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3857 // We need more smarts in the abortable preclean 3858 // loop below to deal with cases where allocation 3859 // in young gen is very very slow, and our precleaning 3860 // is running a losing race against a horde of 3861 // mutators intent on flooding us with CMS updates 3862 // (dirty cards). 3863 // One, admittedly dumb, strategy is to give up 3864 // after a certain number of abortable precleaning loops 3865 // or after a certain maximum time. We want to make 3866 // this smarter in the next iteration. 3867 // XXX FIX ME!!! YSR 3868 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3869 while (!(should_abort_preclean() || 3870 ConcurrentMarkSweepThread::should_terminate())) { 3871 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3872 cumworkdone += workdone; 3873 loops++; 3874 // Voluntarily terminate abortable preclean phase if we have 3875 // been at it for too long. 3876 if ((CMSMaxAbortablePrecleanLoops != 0) && 3877 loops >= CMSMaxAbortablePrecleanLoops) { 3878 if (PrintGCDetails) { 3879 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3880 } 3881 break; 3882 } 3883 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3884 if (PrintGCDetails) { 3885 gclog_or_tty->print(" CMS: abort preclean due to time "); 3886 } 3887 break; 3888 } 3889 // If we are doing little work each iteration, we should 3890 // take a short break. 3891 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3892 // Sleep for some time, waiting for work to accumulate 3893 stopTimer(); 3894 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3895 startTimer(); 3896 waited++; 3897 } 3898 } 3899 if (PrintCMSStatistics > 0) { 3900 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3901 loops, waited, cumworkdone); 3902 } 3903 } 3904 CMSTokenSync x(true); // is cms thread 3905 if (_collectorState != Idling) { 3906 assert(_collectorState == AbortablePreclean, 3907 "Spontaneous state transition?"); 3908 _collectorState = FinalMarking; 3909 } // Else, a foreground collection completed this CMS cycle. 3910 return; 3911 } 3912 3913 // Respond to an Eden sampling opportunity 3914 void CMSCollector::sample_eden() { 3915 // Make sure a young gc cannot sneak in between our 3916 // reading and recording of a sample. 3917 assert(Thread::current()->is_ConcurrentGC_thread(), 3918 "Only the cms thread may collect Eden samples"); 3919 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3920 "Should collect samples while holding CMS token"); 3921 if (!_start_sampling) { 3922 return; 3923 } 3924 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3925 // is populated by the young generation. 3926 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3927 if (_eden_chunk_index < _eden_chunk_capacity) { 3928 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3929 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3930 "Unexpected state of Eden"); 3931 // We'd like to check that what we just sampled is an oop-start address; 3932 // however, we cannot do that here since the object may not yet have been 3933 // initialized. So we'll instead do the check when we _use_ this sample 3934 // later. 3935 if (_eden_chunk_index == 0 || 3936 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3937 _eden_chunk_array[_eden_chunk_index-1]) 3938 >= CMSSamplingGrain)) { 3939 _eden_chunk_index++; // commit sample 3940 } 3941 } 3942 } 3943 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3944 size_t used = get_eden_used(); 3945 size_t capacity = get_eden_capacity(); 3946 assert(used <= capacity, "Unexpected state of Eden"); 3947 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3948 _abort_preclean = true; 3949 } 3950 } 3951 } 3952 3953 3954 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3955 assert(_collectorState == Precleaning || 3956 _collectorState == AbortablePreclean, "incorrect state"); 3957 ResourceMark rm; 3958 HandleMark hm; 3959 3960 // Precleaning is currently not MT but the reference processor 3961 // may be set for MT. Disable it temporarily here. 3962 ReferenceProcessor* rp = ref_processor(); 3963 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3964 3965 // Do one pass of scrubbing the discovered reference lists 3966 // to remove any reference objects with strongly-reachable 3967 // referents. 3968 if (clean_refs) { 3969 CMSPrecleanRefsYieldClosure yield_cl(this); 3970 assert(rp->span().equals(_span), "Spans should be equal"); 3971 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3972 &_markStack, true /* preclean */); 3973 CMSDrainMarkingStackClosure complete_trace(this, 3974 _span, &_markBitMap, &_markStack, 3975 &keep_alive, true /* preclean */); 3976 3977 // We don't want this step to interfere with a young 3978 // collection because we don't want to take CPU 3979 // or memory bandwidth away from the young GC threads 3980 // (which may be as many as there are CPUs). 3981 // Note that we don't need to protect ourselves from 3982 // interference with mutators because they can't 3983 // manipulate the discovered reference lists nor affect 3984 // the computed reachability of the referents, the 3985 // only properties manipulated by the precleaning 3986 // of these reference lists. 3987 stopTimer(); 3988 CMSTokenSyncWithLocks x(true /* is cms thread */, 3989 bitMapLock()); 3990 startTimer(); 3991 sample_eden(); 3992 3993 // The following will yield to allow foreground 3994 // collection to proceed promptly. XXX YSR: 3995 // The code in this method may need further 3996 // tweaking for better performance and some restructuring 3997 // for cleaner interfaces. 3998 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3999 rp->preclean_discovered_references( 4000 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4001 gc_timer, _gc_tracer_cm->gc_id()); 4002 } 4003 4004 if (clean_survivor) { // preclean the active survivor space(s) 4005 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4006 &_markBitMap, &_modUnionTable, 4007 &_markStack, true /* precleaning phase */); 4008 stopTimer(); 4009 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4010 bitMapLock()); 4011 startTimer(); 4012 unsigned int before_count = 4013 GenCollectedHeap::heap()->total_collections(); 4014 SurvivorSpacePrecleanClosure 4015 sss_cl(this, _span, &_markBitMap, &_markStack, 4016 &pam_cl, before_count, CMSYield); 4017 _young_gen->from()->object_iterate_careful(&sss_cl); 4018 _young_gen->to()->object_iterate_careful(&sss_cl); 4019 } 4020 MarkRefsIntoAndScanClosure 4021 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4022 &_markStack, this, CMSYield, 4023 true /* precleaning phase */); 4024 // CAUTION: The following closure has persistent state that may need to 4025 // be reset upon a decrease in the sequence of addresses it 4026 // processes. 4027 ScanMarkedObjectsAgainCarefullyClosure 4028 smoac_cl(this, _span, 4029 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4030 4031 // Preclean dirty cards in ModUnionTable and CardTable using 4032 // appropriate convergence criterion; 4033 // repeat CMSPrecleanIter times unless we find that 4034 // we are losing. 4035 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4036 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4037 "Bad convergence multiplier"); 4038 assert(CMSPrecleanThreshold >= 100, 4039 "Unreasonably low CMSPrecleanThreshold"); 4040 4041 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4042 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4043 numIter < CMSPrecleanIter; 4044 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4045 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4046 if (Verbose && PrintGCDetails) { 4047 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4048 } 4049 // Either there are very few dirty cards, so re-mark 4050 // pause will be small anyway, or our pre-cleaning isn't 4051 // that much faster than the rate at which cards are being 4052 // dirtied, so we might as well stop and re-mark since 4053 // precleaning won't improve our re-mark time by much. 4054 if (curNumCards <= CMSPrecleanThreshold || 4055 (numIter > 0 && 4056 (curNumCards * CMSPrecleanDenominator > 4057 lastNumCards * CMSPrecleanNumerator))) { 4058 numIter++; 4059 cumNumCards += curNumCards; 4060 break; 4061 } 4062 } 4063 4064 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4065 4066 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4067 cumNumCards += curNumCards; 4068 if (PrintGCDetails && PrintCMSStatistics != 0) { 4069 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4070 curNumCards, cumNumCards, numIter); 4071 } 4072 return cumNumCards; // as a measure of useful work done 4073 } 4074 4075 // PRECLEANING NOTES: 4076 // Precleaning involves: 4077 // . reading the bits of the modUnionTable and clearing the set bits. 4078 // . For the cards corresponding to the set bits, we scan the 4079 // objects on those cards. This means we need the free_list_lock 4080 // so that we can safely iterate over the CMS space when scanning 4081 // for oops. 4082 // . When we scan the objects, we'll be both reading and setting 4083 // marks in the marking bit map, so we'll need the marking bit map. 4084 // . For protecting _collector_state transitions, we take the CGC_lock. 4085 // Note that any races in the reading of of card table entries by the 4086 // CMS thread on the one hand and the clearing of those entries by the 4087 // VM thread or the setting of those entries by the mutator threads on the 4088 // other are quite benign. However, for efficiency it makes sense to keep 4089 // the VM thread from racing with the CMS thread while the latter is 4090 // dirty card info to the modUnionTable. We therefore also use the 4091 // CGC_lock to protect the reading of the card table and the mod union 4092 // table by the CM thread. 4093 // . We run concurrently with mutator updates, so scanning 4094 // needs to be done carefully -- we should not try to scan 4095 // potentially uninitialized objects. 4096 // 4097 // Locking strategy: While holding the CGC_lock, we scan over and 4098 // reset a maximal dirty range of the mod union / card tables, then lock 4099 // the free_list_lock and bitmap lock to do a full marking, then 4100 // release these locks; and repeat the cycle. This allows for a 4101 // certain amount of fairness in the sharing of these locks between 4102 // the CMS collector on the one hand, and the VM thread and the 4103 // mutators on the other. 4104 4105 // NOTE: preclean_mod_union_table() and preclean_card_table() 4106 // further below are largely identical; if you need to modify 4107 // one of these methods, please check the other method too. 4108 4109 size_t CMSCollector::preclean_mod_union_table( 4110 ConcurrentMarkSweepGeneration* gen, 4111 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4112 verify_work_stacks_empty(); 4113 verify_overflow_empty(); 4114 4115 // strategy: starting with the first card, accumulate contiguous 4116 // ranges of dirty cards; clear these cards, then scan the region 4117 // covered by these cards. 4118 4119 // Since all of the MUT is committed ahead, we can just use 4120 // that, in case the generations expand while we are precleaning. 4121 // It might also be fine to just use the committed part of the 4122 // generation, but we might potentially miss cards when the 4123 // generation is rapidly expanding while we are in the midst 4124 // of precleaning. 4125 HeapWord* startAddr = gen->reserved().start(); 4126 HeapWord* endAddr = gen->reserved().end(); 4127 4128 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4129 4130 size_t numDirtyCards, cumNumDirtyCards; 4131 HeapWord *nextAddr, *lastAddr; 4132 for (cumNumDirtyCards = numDirtyCards = 0, 4133 nextAddr = lastAddr = startAddr; 4134 nextAddr < endAddr; 4135 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4136 4137 ResourceMark rm; 4138 HandleMark hm; 4139 4140 MemRegion dirtyRegion; 4141 { 4142 stopTimer(); 4143 // Potential yield point 4144 CMSTokenSync ts(true); 4145 startTimer(); 4146 sample_eden(); 4147 // Get dirty region starting at nextOffset (inclusive), 4148 // simultaneously clearing it. 4149 dirtyRegion = 4150 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4151 assert(dirtyRegion.start() >= nextAddr, 4152 "returned region inconsistent?"); 4153 } 4154 // Remember where the next search should begin. 4155 // The returned region (if non-empty) is a right open interval, 4156 // so lastOffset is obtained from the right end of that 4157 // interval. 4158 lastAddr = dirtyRegion.end(); 4159 // Should do something more transparent and less hacky XXX 4160 numDirtyCards = 4161 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4162 4163 // We'll scan the cards in the dirty region (with periodic 4164 // yields for foreground GC as needed). 4165 if (!dirtyRegion.is_empty()) { 4166 assert(numDirtyCards > 0, "consistency check"); 4167 HeapWord* stop_point = NULL; 4168 stopTimer(); 4169 // Potential yield point 4170 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4171 bitMapLock()); 4172 startTimer(); 4173 { 4174 verify_work_stacks_empty(); 4175 verify_overflow_empty(); 4176 sample_eden(); 4177 stop_point = 4178 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4179 } 4180 if (stop_point != NULL) { 4181 // The careful iteration stopped early either because it found an 4182 // uninitialized object, or because we were in the midst of an 4183 // "abortable preclean", which should now be aborted. Redirty 4184 // the bits corresponding to the partially-scanned or unscanned 4185 // cards. We'll either restart at the next block boundary or 4186 // abort the preclean. 4187 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4188 "Should only be AbortablePreclean."); 4189 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4190 if (should_abort_preclean()) { 4191 break; // out of preclean loop 4192 } else { 4193 // Compute the next address at which preclean should pick up; 4194 // might need bitMapLock in order to read P-bits. 4195 lastAddr = next_card_start_after_block(stop_point); 4196 } 4197 } 4198 } else { 4199 assert(lastAddr == endAddr, "consistency check"); 4200 assert(numDirtyCards == 0, "consistency check"); 4201 break; 4202 } 4203 } 4204 verify_work_stacks_empty(); 4205 verify_overflow_empty(); 4206 return cumNumDirtyCards; 4207 } 4208 4209 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4210 // below are largely identical; if you need to modify 4211 // one of these methods, please check the other method too. 4212 4213 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4214 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4215 // strategy: it's similar to precleamModUnionTable above, in that 4216 // we accumulate contiguous ranges of dirty cards, mark these cards 4217 // precleaned, then scan the region covered by these cards. 4218 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4219 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4220 4221 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4222 4223 size_t numDirtyCards, cumNumDirtyCards; 4224 HeapWord *lastAddr, *nextAddr; 4225 4226 for (cumNumDirtyCards = numDirtyCards = 0, 4227 nextAddr = lastAddr = startAddr; 4228 nextAddr < endAddr; 4229 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4230 4231 ResourceMark rm; 4232 HandleMark hm; 4233 4234 MemRegion dirtyRegion; 4235 { 4236 // See comments in "Precleaning notes" above on why we 4237 // do this locking. XXX Could the locking overheads be 4238 // too high when dirty cards are sparse? [I don't think so.] 4239 stopTimer(); 4240 CMSTokenSync x(true); // is cms thread 4241 startTimer(); 4242 sample_eden(); 4243 // Get and clear dirty region from card table 4244 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4245 MemRegion(nextAddr, endAddr), 4246 true, 4247 CardTableModRefBS::precleaned_card_val()); 4248 4249 assert(dirtyRegion.start() >= nextAddr, 4250 "returned region inconsistent?"); 4251 } 4252 lastAddr = dirtyRegion.end(); 4253 numDirtyCards = 4254 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4255 4256 if (!dirtyRegion.is_empty()) { 4257 stopTimer(); 4258 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4259 startTimer(); 4260 sample_eden(); 4261 verify_work_stacks_empty(); 4262 verify_overflow_empty(); 4263 HeapWord* stop_point = 4264 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4265 if (stop_point != NULL) { 4266 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4267 "Should only be AbortablePreclean."); 4268 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4269 if (should_abort_preclean()) { 4270 break; // out of preclean loop 4271 } else { 4272 // Compute the next address at which preclean should pick up. 4273 lastAddr = next_card_start_after_block(stop_point); 4274 } 4275 } 4276 } else { 4277 break; 4278 } 4279 } 4280 verify_work_stacks_empty(); 4281 verify_overflow_empty(); 4282 return cumNumDirtyCards; 4283 } 4284 4285 class PrecleanKlassClosure : public KlassClosure { 4286 KlassToOopClosure _cm_klass_closure; 4287 public: 4288 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4289 void do_klass(Klass* k) { 4290 if (k->has_accumulated_modified_oops()) { 4291 k->clear_accumulated_modified_oops(); 4292 4293 _cm_klass_closure.do_klass(k); 4294 } 4295 } 4296 }; 4297 4298 // The freelist lock is needed to prevent asserts, is it really needed? 4299 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4300 4301 cl->set_freelistLock(freelistLock); 4302 4303 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4304 4305 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4306 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4307 PrecleanKlassClosure preclean_klass_closure(cl); 4308 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4309 4310 verify_work_stacks_empty(); 4311 verify_overflow_empty(); 4312 } 4313 4314 void CMSCollector::checkpointRootsFinal() { 4315 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4316 check_correct_thread_executing(); 4317 // world is stopped at this checkpoint 4318 assert(SafepointSynchronize::is_at_safepoint(), 4319 "world should be stopped"); 4320 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4321 4322 verify_work_stacks_empty(); 4323 verify_overflow_empty(); 4324 4325 if (PrintGCDetails) { 4326 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4327 _young_gen->used() / K, 4328 _young_gen->capacity() / K); 4329 } 4330 { 4331 if (CMSScavengeBeforeRemark) { 4332 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4333 // Temporarily set flag to false, GCH->do_collection will 4334 // expect it to be false and set to true 4335 FlagSetting fl(gch->_is_gc_active, false); 4336 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4337 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4338 int level = _cmsGen->level() - 1; 4339 if (level >= 0) { 4340 gch->do_collection(true, // full (i.e. force, see below) 4341 false, // !clear_all_soft_refs 4342 0, // size 4343 false, // is_tlab 4344 level // max_level 4345 ); 4346 } 4347 } 4348 FreelistLocker x(this); 4349 MutexLockerEx y(bitMapLock(), 4350 Mutex::_no_safepoint_check_flag); 4351 checkpointRootsFinalWork(); 4352 } 4353 verify_work_stacks_empty(); 4354 verify_overflow_empty(); 4355 } 4356 4357 void CMSCollector::checkpointRootsFinalWork() { 4358 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4359 4360 assert(haveFreelistLocks(), "must have free list locks"); 4361 assert_lock_strong(bitMapLock()); 4362 4363 ResourceMark rm; 4364 HandleMark hm; 4365 4366 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4367 4368 if (should_unload_classes()) { 4369 CodeCache::gc_prologue(); 4370 } 4371 assert(haveFreelistLocks(), "must have free list locks"); 4372 assert_lock_strong(bitMapLock()); 4373 4374 // We might assume that we need not fill TLAB's when 4375 // CMSScavengeBeforeRemark is set, because we may have just done 4376 // a scavenge which would have filled all TLAB's -- and besides 4377 // Eden would be empty. This however may not always be the case -- 4378 // for instance although we asked for a scavenge, it may not have 4379 // happened because of a JNI critical section. We probably need 4380 // a policy for deciding whether we can in that case wait until 4381 // the critical section releases and then do the remark following 4382 // the scavenge, and skip it here. In the absence of that policy, 4383 // or of an indication of whether the scavenge did indeed occur, 4384 // we cannot rely on TLAB's having been filled and must do 4385 // so here just in case a scavenge did not happen. 4386 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4387 // Update the saved marks which may affect the root scans. 4388 gch->save_marks(); 4389 4390 if (CMSPrintEdenSurvivorChunks) { 4391 print_eden_and_survivor_chunk_arrays(); 4392 } 4393 4394 { 4395 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4396 4397 // Note on the role of the mod union table: 4398 // Since the marker in "markFromRoots" marks concurrently with 4399 // mutators, it is possible for some reachable objects not to have been 4400 // scanned. For instance, an only reference to an object A was 4401 // placed in object B after the marker scanned B. Unless B is rescanned, 4402 // A would be collected. Such updates to references in marked objects 4403 // are detected via the mod union table which is the set of all cards 4404 // dirtied since the first checkpoint in this GC cycle and prior to 4405 // the most recent young generation GC, minus those cleaned up by the 4406 // concurrent precleaning. 4407 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4408 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4409 do_remark_parallel(); 4410 } else { 4411 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4412 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4413 do_remark_non_parallel(); 4414 } 4415 } 4416 verify_work_stacks_empty(); 4417 verify_overflow_empty(); 4418 4419 { 4420 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4421 refProcessingWork(); 4422 } 4423 verify_work_stacks_empty(); 4424 verify_overflow_empty(); 4425 4426 if (should_unload_classes()) { 4427 CodeCache::gc_epilogue(); 4428 } 4429 JvmtiExport::gc_epilogue(); 4430 4431 // If we encountered any (marking stack / work queue) overflow 4432 // events during the current CMS cycle, take appropriate 4433 // remedial measures, where possible, so as to try and avoid 4434 // recurrence of that condition. 4435 assert(_markStack.isEmpty(), "No grey objects"); 4436 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4437 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4438 if (ser_ovflw > 0) { 4439 if (PrintCMSStatistics != 0) { 4440 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4441 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4442 ", kac_preclean="SIZE_FORMAT")", 4443 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4444 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4445 } 4446 _markStack.expand(); 4447 _ser_pmc_remark_ovflw = 0; 4448 _ser_pmc_preclean_ovflw = 0; 4449 _ser_kac_preclean_ovflw = 0; 4450 _ser_kac_ovflw = 0; 4451 } 4452 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4453 if (PrintCMSStatistics != 0) { 4454 gclog_or_tty->print_cr("Work queue overflow (benign) " 4455 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4456 _par_pmc_remark_ovflw, _par_kac_ovflw); 4457 } 4458 _par_pmc_remark_ovflw = 0; 4459 _par_kac_ovflw = 0; 4460 } 4461 if (PrintCMSStatistics != 0) { 4462 if (_markStack._hit_limit > 0) { 4463 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4464 _markStack._hit_limit); 4465 } 4466 if (_markStack._failed_double > 0) { 4467 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4468 " current capacity "SIZE_FORMAT, 4469 _markStack._failed_double, 4470 _markStack.capacity()); 4471 } 4472 } 4473 _markStack._hit_limit = 0; 4474 _markStack._failed_double = 0; 4475 4476 if ((VerifyAfterGC || VerifyDuringGC) && 4477 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4478 verify_after_remark(); 4479 } 4480 4481 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4482 4483 // Change under the freelistLocks. 4484 _collectorState = Sweeping; 4485 // Call isAllClear() under bitMapLock 4486 assert(_modUnionTable.isAllClear(), 4487 "Should be clear by end of the final marking"); 4488 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4489 "Should be clear by end of the final marking"); 4490 } 4491 4492 void CMSParInitialMarkTask::work(uint worker_id) { 4493 elapsedTimer _timer; 4494 ResourceMark rm; 4495 HandleMark hm; 4496 4497 // ---------- scan from roots -------------- 4498 _timer.start(); 4499 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4500 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4501 4502 // ---------- young gen roots -------------- 4503 { 4504 work_on_young_gen_roots(worker_id, &par_mri_cl); 4505 _timer.stop(); 4506 if (PrintCMSStatistics != 0) { 4507 gclog_or_tty->print_cr( 4508 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4509 worker_id, _timer.seconds()); 4510 } 4511 } 4512 4513 // ---------- remaining roots -------------- 4514 _timer.reset(); 4515 _timer.start(); 4516 4517 CLDToOopClosure cld_closure(&par_mri_cl, true); 4518 4519 gch->gen_process_roots(_collector->_cmsGen->level(), 4520 false, // yg was scanned above 4521 false, // this is parallel code 4522 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4523 _collector->should_unload_classes(), 4524 &par_mri_cl, 4525 NULL, 4526 &cld_closure); 4527 assert(_collector->should_unload_classes() 4528 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4529 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4530 _timer.stop(); 4531 if (PrintCMSStatistics != 0) { 4532 gclog_or_tty->print_cr( 4533 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4534 worker_id, _timer.seconds()); 4535 } 4536 } 4537 4538 // Parallel remark task 4539 class CMSParRemarkTask: public CMSParMarkTask { 4540 CompactibleFreeListSpace* _cms_space; 4541 4542 // The per-thread work queues, available here for stealing. 4543 OopTaskQueueSet* _task_queues; 4544 ParallelTaskTerminator _term; 4545 4546 public: 4547 // A value of 0 passed to n_workers will cause the number of 4548 // workers to be taken from the active workers in the work gang. 4549 CMSParRemarkTask(CMSCollector* collector, 4550 CompactibleFreeListSpace* cms_space, 4551 int n_workers, FlexibleWorkGang* workers, 4552 OopTaskQueueSet* task_queues): 4553 CMSParMarkTask("Rescan roots and grey objects in parallel", 4554 collector, n_workers), 4555 _cms_space(cms_space), 4556 _task_queues(task_queues), 4557 _term(n_workers, task_queues) { } 4558 4559 OopTaskQueueSet* task_queues() { return _task_queues; } 4560 4561 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4562 4563 ParallelTaskTerminator* terminator() { return &_term; } 4564 int n_workers() { return _n_workers; } 4565 4566 void work(uint worker_id); 4567 4568 private: 4569 // ... of dirty cards in old space 4570 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4571 Par_MarkRefsIntoAndScanClosure* cl); 4572 4573 // ... work stealing for the above 4574 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4575 }; 4576 4577 class RemarkKlassClosure : public KlassClosure { 4578 KlassToOopClosure _cm_klass_closure; 4579 public: 4580 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4581 void do_klass(Klass* k) { 4582 // Check if we have modified any oops in the Klass during the concurrent marking. 4583 if (k->has_accumulated_modified_oops()) { 4584 k->clear_accumulated_modified_oops(); 4585 4586 // We could have transfered the current modified marks to the accumulated marks, 4587 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4588 } else if (k->has_modified_oops()) { 4589 // Don't clear anything, this info is needed by the next young collection. 4590 } else { 4591 // No modified oops in the Klass. 4592 return; 4593 } 4594 4595 // The klass has modified fields, need to scan the klass. 4596 _cm_klass_closure.do_klass(k); 4597 } 4598 }; 4599 4600 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4601 ParNewGeneration* young_gen = _collector->_young_gen; 4602 ContiguousSpace* eden_space = young_gen->eden(); 4603 ContiguousSpace* from_space = young_gen->from(); 4604 ContiguousSpace* to_space = young_gen->to(); 4605 4606 HeapWord** eca = _collector->_eden_chunk_array; 4607 size_t ect = _collector->_eden_chunk_index; 4608 HeapWord** sca = _collector->_survivor_chunk_array; 4609 size_t sct = _collector->_survivor_chunk_index; 4610 4611 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4612 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4613 4614 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4615 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4616 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4617 } 4618 4619 // work_queue(i) is passed to the closure 4620 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4621 // also is passed to do_dirty_card_rescan_tasks() and to 4622 // do_work_steal() to select the i-th task_queue. 4623 4624 void CMSParRemarkTask::work(uint worker_id) { 4625 elapsedTimer _timer; 4626 ResourceMark rm; 4627 HandleMark hm; 4628 4629 // ---------- rescan from roots -------------- 4630 _timer.start(); 4631 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4632 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4633 _collector->_span, _collector->ref_processor(), 4634 &(_collector->_markBitMap), 4635 work_queue(worker_id)); 4636 4637 // Rescan young gen roots first since these are likely 4638 // coarsely partitioned and may, on that account, constitute 4639 // the critical path; thus, it's best to start off that 4640 // work first. 4641 // ---------- young gen roots -------------- 4642 { 4643 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4644 _timer.stop(); 4645 if (PrintCMSStatistics != 0) { 4646 gclog_or_tty->print_cr( 4647 "Finished young gen rescan work in %dth thread: %3.3f sec", 4648 worker_id, _timer.seconds()); 4649 } 4650 } 4651 4652 // ---------- remaining roots -------------- 4653 _timer.reset(); 4654 _timer.start(); 4655 gch->gen_process_roots(_collector->_cmsGen->level(), 4656 false, // yg was scanned above 4657 false, // this is parallel code 4658 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4659 _collector->should_unload_classes(), 4660 &par_mrias_cl, 4661 NULL, 4662 NULL); // The dirty klasses will be handled below 4663 4664 assert(_collector->should_unload_classes() 4665 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4666 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4667 _timer.stop(); 4668 if (PrintCMSStatistics != 0) { 4669 gclog_or_tty->print_cr( 4670 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4671 worker_id, _timer.seconds()); 4672 } 4673 4674 // ---------- unhandled CLD scanning ---------- 4675 if (worker_id == 0) { // Single threaded at the moment. 4676 _timer.reset(); 4677 _timer.start(); 4678 4679 // Scan all new class loader data objects and new dependencies that were 4680 // introduced during concurrent marking. 4681 ResourceMark rm; 4682 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4683 for (int i = 0; i < array->length(); i++) { 4684 par_mrias_cl.do_class_loader_data(array->at(i)); 4685 } 4686 4687 // We don't need to keep track of new CLDs anymore. 4688 ClassLoaderDataGraph::remember_new_clds(false); 4689 4690 _timer.stop(); 4691 if (PrintCMSStatistics != 0) { 4692 gclog_or_tty->print_cr( 4693 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4694 worker_id, _timer.seconds()); 4695 } 4696 } 4697 4698 // ---------- dirty klass scanning ---------- 4699 if (worker_id == 0) { // Single threaded at the moment. 4700 _timer.reset(); 4701 _timer.start(); 4702 4703 // Scan all classes that was dirtied during the concurrent marking phase. 4704 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4705 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4706 4707 _timer.stop(); 4708 if (PrintCMSStatistics != 0) { 4709 gclog_or_tty->print_cr( 4710 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4711 worker_id, _timer.seconds()); 4712 } 4713 } 4714 4715 // We might have added oops to ClassLoaderData::_handles during the 4716 // concurrent marking phase. These oops point to newly allocated objects 4717 // that are guaranteed to be kept alive. Either by the direct allocation 4718 // code, or when the young collector processes the roots. Hence, 4719 // we don't have to revisit the _handles block during the remark phase. 4720 4721 // ---------- rescan dirty cards ------------ 4722 _timer.reset(); 4723 _timer.start(); 4724 4725 // Do the rescan tasks for each of the two spaces 4726 // (cms_space) in turn. 4727 // "worker_id" is passed to select the task_queue for "worker_id" 4728 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4729 _timer.stop(); 4730 if (PrintCMSStatistics != 0) { 4731 gclog_or_tty->print_cr( 4732 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4733 worker_id, _timer.seconds()); 4734 } 4735 4736 // ---------- steal work from other threads ... 4737 // ---------- ... and drain overflow list. 4738 _timer.reset(); 4739 _timer.start(); 4740 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4741 _timer.stop(); 4742 if (PrintCMSStatistics != 0) { 4743 gclog_or_tty->print_cr( 4744 "Finished work stealing in %dth thread: %3.3f sec", 4745 worker_id, _timer.seconds()); 4746 } 4747 } 4748 4749 // Note that parameter "i" is not used. 4750 void 4751 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4752 OopsInGenClosure* cl, ContiguousSpace* space, 4753 HeapWord** chunk_array, size_t chunk_top) { 4754 // Until all tasks completed: 4755 // . claim an unclaimed task 4756 // . compute region boundaries corresponding to task claimed 4757 // using chunk_array 4758 // . par_oop_iterate(cl) over that region 4759 4760 ResourceMark rm; 4761 HandleMark hm; 4762 4763 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4764 4765 uint nth_task = 0; 4766 uint n_tasks = pst->n_tasks(); 4767 4768 if (n_tasks > 0) { 4769 assert(pst->valid(), "Uninitialized use?"); 4770 HeapWord *start, *end; 4771 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4772 // We claimed task # nth_task; compute its boundaries. 4773 if (chunk_top == 0) { // no samples were taken 4774 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4775 start = space->bottom(); 4776 end = space->top(); 4777 } else if (nth_task == 0) { 4778 start = space->bottom(); 4779 end = chunk_array[nth_task]; 4780 } else if (nth_task < (uint)chunk_top) { 4781 assert(nth_task >= 1, "Control point invariant"); 4782 start = chunk_array[nth_task - 1]; 4783 end = chunk_array[nth_task]; 4784 } else { 4785 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4786 start = chunk_array[chunk_top - 1]; 4787 end = space->top(); 4788 } 4789 MemRegion mr(start, end); 4790 // Verify that mr is in space 4791 assert(mr.is_empty() || space->used_region().contains(mr), 4792 "Should be in space"); 4793 // Verify that "start" is an object boundary 4794 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4795 "Should be an oop"); 4796 space->par_oop_iterate(mr, cl); 4797 } 4798 pst->all_tasks_completed(); 4799 } 4800 } 4801 4802 void 4803 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4804 CompactibleFreeListSpace* sp, int i, 4805 Par_MarkRefsIntoAndScanClosure* cl) { 4806 // Until all tasks completed: 4807 // . claim an unclaimed task 4808 // . compute region boundaries corresponding to task claimed 4809 // . transfer dirty bits ct->mut for that region 4810 // . apply rescanclosure to dirty mut bits for that region 4811 4812 ResourceMark rm; 4813 HandleMark hm; 4814 4815 OopTaskQueue* work_q = work_queue(i); 4816 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4817 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4818 // CAUTION: This closure has state that persists across calls to 4819 // the work method dirty_range_iterate_clear() in that it has 4820 // embedded in it a (subtype of) UpwardsObjectClosure. The 4821 // use of that state in the embedded UpwardsObjectClosure instance 4822 // assumes that the cards are always iterated (even if in parallel 4823 // by several threads) in monotonically increasing order per each 4824 // thread. This is true of the implementation below which picks 4825 // card ranges (chunks) in monotonically increasing order globally 4826 // and, a-fortiori, in monotonically increasing order per thread 4827 // (the latter order being a subsequence of the former). 4828 // If the work code below is ever reorganized into a more chaotic 4829 // work-partitioning form than the current "sequential tasks" 4830 // paradigm, the use of that persistent state will have to be 4831 // revisited and modified appropriately. See also related 4832 // bug 4756801 work on which should examine this code to make 4833 // sure that the changes there do not run counter to the 4834 // assumptions made here and necessary for correctness and 4835 // efficiency. Note also that this code might yield inefficient 4836 // behavior in the case of very large objects that span one or 4837 // more work chunks. Such objects would potentially be scanned 4838 // several times redundantly. Work on 4756801 should try and 4839 // address that performance anomaly if at all possible. XXX 4840 MemRegion full_span = _collector->_span; 4841 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4842 MarkFromDirtyCardsClosure 4843 greyRescanClosure(_collector, full_span, // entire span of interest 4844 sp, bm, work_q, cl); 4845 4846 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4847 assert(pst->valid(), "Uninitialized use?"); 4848 uint nth_task = 0; 4849 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4850 MemRegion span = sp->used_region(); 4851 HeapWord* start_addr = span.start(); 4852 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4853 alignment); 4854 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4855 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4856 start_addr, "Check alignment"); 4857 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4858 chunk_size, "Check alignment"); 4859 4860 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4861 // Having claimed the nth_task, compute corresponding mem-region, 4862 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4863 // The alignment restriction ensures that we do not need any 4864 // synchronization with other gang-workers while setting or 4865 // clearing bits in thus chunk of the MUT. 4866 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4867 start_addr + (nth_task+1)*chunk_size); 4868 // The last chunk's end might be way beyond end of the 4869 // used region. In that case pull back appropriately. 4870 if (this_span.end() > end_addr) { 4871 this_span.set_end(end_addr); 4872 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4873 } 4874 // Iterate over the dirty cards covering this chunk, marking them 4875 // precleaned, and setting the corresponding bits in the mod union 4876 // table. Since we have been careful to partition at Card and MUT-word 4877 // boundaries no synchronization is needed between parallel threads. 4878 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4879 &modUnionClosure); 4880 4881 // Having transferred these marks into the modUnionTable, 4882 // rescan the marked objects on the dirty cards in the modUnionTable. 4883 // Even if this is at a synchronous collection, the initial marking 4884 // may have been done during an asynchronous collection so there 4885 // may be dirty bits in the mod-union table. 4886 _collector->_modUnionTable.dirty_range_iterate_clear( 4887 this_span, &greyRescanClosure); 4888 _collector->_modUnionTable.verifyNoOneBitsInRange( 4889 this_span.start(), 4890 this_span.end()); 4891 } 4892 pst->all_tasks_completed(); // declare that i am done 4893 } 4894 4895 // . see if we can share work_queues with ParNew? XXX 4896 void 4897 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4898 int* seed) { 4899 OopTaskQueue* work_q = work_queue(i); 4900 NOT_PRODUCT(int num_steals = 0;) 4901 oop obj_to_scan; 4902 CMSBitMap* bm = &(_collector->_markBitMap); 4903 4904 while (true) { 4905 // Completely finish any left over work from (an) earlier round(s) 4906 cl->trim_queue(0); 4907 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4908 (size_t)ParGCDesiredObjsFromOverflowList); 4909 // Now check if there's any work in the overflow list 4910 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4911 // only affects the number of attempts made to get work from the 4912 // overflow list and does not affect the number of workers. Just 4913 // pass ParallelGCThreads so this behavior is unchanged. 4914 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4915 work_q, 4916 ParallelGCThreads)) { 4917 // found something in global overflow list; 4918 // not yet ready to go stealing work from others. 4919 // We'd like to assert(work_q->size() != 0, ...) 4920 // because we just took work from the overflow list, 4921 // but of course we can't since all of that could have 4922 // been already stolen from us. 4923 // "He giveth and He taketh away." 4924 continue; 4925 } 4926 // Verify that we have no work before we resort to stealing 4927 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4928 // Try to steal from other queues that have work 4929 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4930 NOT_PRODUCT(num_steals++;) 4931 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4932 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4933 // Do scanning work 4934 obj_to_scan->oop_iterate(cl); 4935 // Loop around, finish this work, and try to steal some more 4936 } else if (terminator()->offer_termination()) { 4937 break; // nirvana from the infinite cycle 4938 } 4939 } 4940 NOT_PRODUCT( 4941 if (PrintCMSStatistics != 0) { 4942 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4943 } 4944 ) 4945 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4946 "Else our work is not yet done"); 4947 } 4948 4949 // Record object boundaries in _eden_chunk_array by sampling the eden 4950 // top in the slow-path eden object allocation code path and record 4951 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4952 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4953 // sampling in sample_eden() that activates during the part of the 4954 // preclean phase. 4955 void CMSCollector::sample_eden_chunk() { 4956 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4957 if (_eden_chunk_lock->try_lock()) { 4958 // Record a sample. This is the critical section. The contents 4959 // of the _eden_chunk_array have to be non-decreasing in the 4960 // address order. 4961 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4962 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4963 "Unexpected state of Eden"); 4964 if (_eden_chunk_index == 0 || 4965 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4966 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4967 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4968 _eden_chunk_index++; // commit sample 4969 } 4970 _eden_chunk_lock->unlock(); 4971 } 4972 } 4973 } 4974 4975 // Return a thread-local PLAB recording array, as appropriate. 4976 void* CMSCollector::get_data_recorder(int thr_num) { 4977 if (_survivor_plab_array != NULL && 4978 (CMSPLABRecordAlways || 4979 (_collectorState > Marking && _collectorState < FinalMarking))) { 4980 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4981 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4982 ca->reset(); // clear it so that fresh data is recorded 4983 return (void*) ca; 4984 } else { 4985 return NULL; 4986 } 4987 } 4988 4989 // Reset all the thread-local PLAB recording arrays 4990 void CMSCollector::reset_survivor_plab_arrays() { 4991 for (uint i = 0; i < ParallelGCThreads; i++) { 4992 _survivor_plab_array[i].reset(); 4993 } 4994 } 4995 4996 // Merge the per-thread plab arrays into the global survivor chunk 4997 // array which will provide the partitioning of the survivor space 4998 // for CMS initial scan and rescan. 4999 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5000 int no_of_gc_threads) { 5001 assert(_survivor_plab_array != NULL, "Error"); 5002 assert(_survivor_chunk_array != NULL, "Error"); 5003 assert(_collectorState == FinalMarking || 5004 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5005 for (int j = 0; j < no_of_gc_threads; j++) { 5006 _cursor[j] = 0; 5007 } 5008 HeapWord* top = surv->top(); 5009 size_t i; 5010 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5011 HeapWord* min_val = top; // Higher than any PLAB address 5012 uint min_tid = 0; // position of min_val this round 5013 for (int j = 0; j < no_of_gc_threads; j++) { 5014 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5015 if (_cursor[j] == cur_sca->end()) { 5016 continue; 5017 } 5018 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5019 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5020 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5021 if (cur_val < min_val) { 5022 min_tid = j; 5023 min_val = cur_val; 5024 } else { 5025 assert(cur_val < top, "All recorded addresses should be less"); 5026 } 5027 } 5028 // At this point min_val and min_tid are respectively 5029 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5030 // and the thread (j) that witnesses that address. 5031 // We record this address in the _survivor_chunk_array[i] 5032 // and increment _cursor[min_tid] prior to the next round i. 5033 if (min_val == top) { 5034 break; 5035 } 5036 _survivor_chunk_array[i] = min_val; 5037 _cursor[min_tid]++; 5038 } 5039 // We are all done; record the size of the _survivor_chunk_array 5040 _survivor_chunk_index = i; // exclusive: [0, i) 5041 if (PrintCMSStatistics > 0) { 5042 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5043 } 5044 // Verify that we used up all the recorded entries 5045 #ifdef ASSERT 5046 size_t total = 0; 5047 for (int j = 0; j < no_of_gc_threads; j++) { 5048 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5049 total += _cursor[j]; 5050 } 5051 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5052 // Check that the merged array is in sorted order 5053 if (total > 0) { 5054 for (size_t i = 0; i < total - 1; i++) { 5055 if (PrintCMSStatistics > 0) { 5056 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5057 i, p2i(_survivor_chunk_array[i])); 5058 } 5059 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5060 "Not sorted"); 5061 } 5062 } 5063 #endif // ASSERT 5064 } 5065 5066 // Set up the space's par_seq_tasks structure for work claiming 5067 // for parallel initial scan and rescan of young gen. 5068 // See ParRescanTask where this is currently used. 5069 void 5070 CMSCollector:: 5071 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5072 assert(n_threads > 0, "Unexpected n_threads argument"); 5073 5074 // Eden space 5075 if (!_young_gen->eden()->is_empty()) { 5076 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5077 assert(!pst->valid(), "Clobbering existing data?"); 5078 // Each valid entry in [0, _eden_chunk_index) represents a task. 5079 size_t n_tasks = _eden_chunk_index + 1; 5080 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5081 // Sets the condition for completion of the subtask (how many threads 5082 // need to finish in order to be done). 5083 pst->set_n_threads(n_threads); 5084 pst->set_n_tasks((int)n_tasks); 5085 } 5086 5087 // Merge the survivor plab arrays into _survivor_chunk_array 5088 if (_survivor_plab_array != NULL) { 5089 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5090 } else { 5091 assert(_survivor_chunk_index == 0, "Error"); 5092 } 5093 5094 // To space 5095 { 5096 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5097 assert(!pst->valid(), "Clobbering existing data?"); 5098 // Sets the condition for completion of the subtask (how many threads 5099 // need to finish in order to be done). 5100 pst->set_n_threads(n_threads); 5101 pst->set_n_tasks(1); 5102 assert(pst->valid(), "Error"); 5103 } 5104 5105 // From space 5106 { 5107 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5108 assert(!pst->valid(), "Clobbering existing data?"); 5109 size_t n_tasks = _survivor_chunk_index + 1; 5110 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5111 // Sets the condition for completion of the subtask (how many threads 5112 // need to finish in order to be done). 5113 pst->set_n_threads(n_threads); 5114 pst->set_n_tasks((int)n_tasks); 5115 assert(pst->valid(), "Error"); 5116 } 5117 } 5118 5119 // Parallel version of remark 5120 void CMSCollector::do_remark_parallel() { 5121 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5122 FlexibleWorkGang* workers = gch->workers(); 5123 assert(workers != NULL, "Need parallel worker threads."); 5124 // Choose to use the number of GC workers most recently set 5125 // into "active_workers". If active_workers is not set, set it 5126 // to ParallelGCThreads. 5127 int n_workers = workers->active_workers(); 5128 if (n_workers == 0) { 5129 assert(n_workers > 0, "Should have been set during scavenge"); 5130 n_workers = ParallelGCThreads; 5131 workers->set_active_workers(n_workers); 5132 } 5133 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5134 5135 CMSParRemarkTask tsk(this, 5136 cms_space, 5137 n_workers, workers, task_queues()); 5138 5139 // Set up for parallel process_roots work. 5140 gch->set_par_threads(n_workers); 5141 // We won't be iterating over the cards in the card table updating 5142 // the younger_gen cards, so we shouldn't call the following else 5143 // the verification code as well as subsequent younger_refs_iterate 5144 // code would get confused. XXX 5145 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5146 5147 // The young gen rescan work will not be done as part of 5148 // process_roots (which currently doesn't know how to 5149 // parallelize such a scan), but rather will be broken up into 5150 // a set of parallel tasks (via the sampling that the [abortable] 5151 // preclean phase did of eden, plus the [two] tasks of 5152 // scanning the [two] survivor spaces. Further fine-grain 5153 // parallelization of the scanning of the survivor spaces 5154 // themselves, and of precleaning of the younger gen itself 5155 // is deferred to the future. 5156 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5157 5158 // The dirty card rescan work is broken up into a "sequence" 5159 // of parallel tasks (per constituent space) that are dynamically 5160 // claimed by the parallel threads. 5161 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5162 5163 // It turns out that even when we're using 1 thread, doing the work in a 5164 // separate thread causes wide variance in run times. We can't help this 5165 // in the multi-threaded case, but we special-case n=1 here to get 5166 // repeatable measurements of the 1-thread overhead of the parallel code. 5167 if (n_workers > 1) { 5168 // Make refs discovery MT-safe, if it isn't already: it may not 5169 // necessarily be so, since it's possible that we are doing 5170 // ST marking. 5171 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5172 GenCollectedHeap::StrongRootsScope srs; 5173 workers->run_task(&tsk); 5174 } else { 5175 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5176 GenCollectedHeap::StrongRootsScope srs; 5177 tsk.work(0); 5178 } 5179 5180 gch->set_par_threads(0); // 0 ==> non-parallel. 5181 // restore, single-threaded for now, any preserved marks 5182 // as a result of work_q overflow 5183 restore_preserved_marks_if_any(); 5184 } 5185 5186 // Non-parallel version of remark 5187 void CMSCollector::do_remark_non_parallel() { 5188 ResourceMark rm; 5189 HandleMark hm; 5190 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5191 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5192 5193 MarkRefsIntoAndScanClosure 5194 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5195 &_markStack, this, 5196 false /* should_yield */, false /* not precleaning */); 5197 MarkFromDirtyCardsClosure 5198 markFromDirtyCardsClosure(this, _span, 5199 NULL, // space is set further below 5200 &_markBitMap, &_markStack, &mrias_cl); 5201 { 5202 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5203 // Iterate over the dirty cards, setting the corresponding bits in the 5204 // mod union table. 5205 { 5206 ModUnionClosure modUnionClosure(&_modUnionTable); 5207 _ct->ct_bs()->dirty_card_iterate( 5208 _cmsGen->used_region(), 5209 &modUnionClosure); 5210 } 5211 // Having transferred these marks into the modUnionTable, we just need 5212 // to rescan the marked objects on the dirty cards in the modUnionTable. 5213 // The initial marking may have been done during an asynchronous 5214 // collection so there may be dirty bits in the mod-union table. 5215 const int alignment = 5216 CardTableModRefBS::card_size * BitsPerWord; 5217 { 5218 // ... First handle dirty cards in CMS gen 5219 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5220 MemRegion ur = _cmsGen->used_region(); 5221 HeapWord* lb = ur.start(); 5222 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5223 MemRegion cms_span(lb, ub); 5224 _modUnionTable.dirty_range_iterate_clear(cms_span, 5225 &markFromDirtyCardsClosure); 5226 verify_work_stacks_empty(); 5227 if (PrintCMSStatistics != 0) { 5228 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5229 markFromDirtyCardsClosure.num_dirty_cards()); 5230 } 5231 } 5232 } 5233 if (VerifyDuringGC && 5234 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5235 HandleMark hm; // Discard invalid handles created during verification 5236 Universe::verify(); 5237 } 5238 { 5239 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5240 5241 verify_work_stacks_empty(); 5242 5243 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5244 GenCollectedHeap::StrongRootsScope srs; 5245 5246 gch->gen_process_roots(_cmsGen->level(), 5247 true, // younger gens as roots 5248 false, // use the local StrongRootsScope 5249 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5250 should_unload_classes(), 5251 &mrias_cl, 5252 NULL, 5253 NULL); // The dirty klasses will be handled below 5254 5255 assert(should_unload_classes() 5256 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5257 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5258 } 5259 5260 { 5261 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5262 5263 verify_work_stacks_empty(); 5264 5265 // Scan all class loader data objects that might have been introduced 5266 // during concurrent marking. 5267 ResourceMark rm; 5268 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5269 for (int i = 0; i < array->length(); i++) { 5270 mrias_cl.do_class_loader_data(array->at(i)); 5271 } 5272 5273 // We don't need to keep track of new CLDs anymore. 5274 ClassLoaderDataGraph::remember_new_clds(false); 5275 5276 verify_work_stacks_empty(); 5277 } 5278 5279 { 5280 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5281 5282 verify_work_stacks_empty(); 5283 5284 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5285 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5286 5287 verify_work_stacks_empty(); 5288 } 5289 5290 // We might have added oops to ClassLoaderData::_handles during the 5291 // concurrent marking phase. These oops point to newly allocated objects 5292 // that are guaranteed to be kept alive. Either by the direct allocation 5293 // code, or when the young collector processes the roots. Hence, 5294 // we don't have to revisit the _handles block during the remark phase. 5295 5296 verify_work_stacks_empty(); 5297 // Restore evacuated mark words, if any, used for overflow list links 5298 if (!CMSOverflowEarlyRestoration) { 5299 restore_preserved_marks_if_any(); 5300 } 5301 verify_overflow_empty(); 5302 } 5303 5304 //////////////////////////////////////////////////////// 5305 // Parallel Reference Processing Task Proxy Class 5306 //////////////////////////////////////////////////////// 5307 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5308 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5309 CMSCollector* _collector; 5310 CMSBitMap* _mark_bit_map; 5311 const MemRegion _span; 5312 ProcessTask& _task; 5313 5314 public: 5315 CMSRefProcTaskProxy(ProcessTask& task, 5316 CMSCollector* collector, 5317 const MemRegion& span, 5318 CMSBitMap* mark_bit_map, 5319 AbstractWorkGang* workers, 5320 OopTaskQueueSet* task_queues): 5321 // XXX Should superclass AGTWOQ also know about AWG since it knows 5322 // about the task_queues used by the AWG? Then it could initialize 5323 // the terminator() object. See 6984287. The set_for_termination() 5324 // below is a temporary band-aid for the regression in 6984287. 5325 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5326 task_queues), 5327 _task(task), 5328 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5329 { 5330 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5331 "Inconsistency in _span"); 5332 set_for_termination(workers->active_workers()); 5333 } 5334 5335 OopTaskQueueSet* task_queues() { return queues(); } 5336 5337 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5338 5339 void do_work_steal(int i, 5340 CMSParDrainMarkingStackClosure* drain, 5341 CMSParKeepAliveClosure* keep_alive, 5342 int* seed); 5343 5344 virtual void work(uint worker_id); 5345 }; 5346 5347 void CMSRefProcTaskProxy::work(uint worker_id) { 5348 ResourceMark rm; 5349 HandleMark hm; 5350 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5351 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5352 _mark_bit_map, 5353 work_queue(worker_id)); 5354 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5355 _mark_bit_map, 5356 work_queue(worker_id)); 5357 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5358 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5359 if (_task.marks_oops_alive()) { 5360 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5361 _collector->hash_seed(worker_id)); 5362 } 5363 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5364 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5365 } 5366 5367 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5368 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5369 EnqueueTask& _task; 5370 5371 public: 5372 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5373 : AbstractGangTask("Enqueue reference objects in parallel"), 5374 _task(task) 5375 { } 5376 5377 virtual void work(uint worker_id) 5378 { 5379 _task.work(worker_id); 5380 } 5381 }; 5382 5383 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5384 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5385 _span(span), 5386 _bit_map(bit_map), 5387 _work_queue(work_queue), 5388 _mark_and_push(collector, span, bit_map, work_queue), 5389 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5390 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5391 { } 5392 5393 // . see if we can share work_queues with ParNew? XXX 5394 void CMSRefProcTaskProxy::do_work_steal(int i, 5395 CMSParDrainMarkingStackClosure* drain, 5396 CMSParKeepAliveClosure* keep_alive, 5397 int* seed) { 5398 OopTaskQueue* work_q = work_queue(i); 5399 NOT_PRODUCT(int num_steals = 0;) 5400 oop obj_to_scan; 5401 5402 while (true) { 5403 // Completely finish any left over work from (an) earlier round(s) 5404 drain->trim_queue(0); 5405 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5406 (size_t)ParGCDesiredObjsFromOverflowList); 5407 // Now check if there's any work in the overflow list 5408 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5409 // only affects the number of attempts made to get work from the 5410 // overflow list and does not affect the number of workers. Just 5411 // pass ParallelGCThreads so this behavior is unchanged. 5412 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5413 work_q, 5414 ParallelGCThreads)) { 5415 // Found something in global overflow list; 5416 // not yet ready to go stealing work from others. 5417 // We'd like to assert(work_q->size() != 0, ...) 5418 // because we just took work from the overflow list, 5419 // but of course we can't, since all of that might have 5420 // been already stolen from us. 5421 continue; 5422 } 5423 // Verify that we have no work before we resort to stealing 5424 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5425 // Try to steal from other queues that have work 5426 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5427 NOT_PRODUCT(num_steals++;) 5428 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5429 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5430 // Do scanning work 5431 obj_to_scan->oop_iterate(keep_alive); 5432 // Loop around, finish this work, and try to steal some more 5433 } else if (terminator()->offer_termination()) { 5434 break; // nirvana from the infinite cycle 5435 } 5436 } 5437 NOT_PRODUCT( 5438 if (PrintCMSStatistics != 0) { 5439 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5440 } 5441 ) 5442 } 5443 5444 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5445 { 5446 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5447 FlexibleWorkGang* workers = gch->workers(); 5448 assert(workers != NULL, "Need parallel worker threads."); 5449 CMSRefProcTaskProxy rp_task(task, &_collector, 5450 _collector.ref_processor()->span(), 5451 _collector.markBitMap(), 5452 workers, _collector.task_queues()); 5453 workers->run_task(&rp_task); 5454 } 5455 5456 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5457 { 5458 5459 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5460 FlexibleWorkGang* workers = gch->workers(); 5461 assert(workers != NULL, "Need parallel worker threads."); 5462 CMSRefEnqueueTaskProxy enq_task(task); 5463 workers->run_task(&enq_task); 5464 } 5465 5466 void CMSCollector::refProcessingWork() { 5467 ResourceMark rm; 5468 HandleMark hm; 5469 5470 ReferenceProcessor* rp = ref_processor(); 5471 assert(rp->span().equals(_span), "Spans should be equal"); 5472 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5473 // Process weak references. 5474 rp->setup_policy(false); 5475 verify_work_stacks_empty(); 5476 5477 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5478 &_markStack, false /* !preclean */); 5479 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5480 _span, &_markBitMap, &_markStack, 5481 &cmsKeepAliveClosure, false /* !preclean */); 5482 { 5483 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5484 5485 ReferenceProcessorStats stats; 5486 if (rp->processing_is_mt()) { 5487 // Set the degree of MT here. If the discovery is done MT, there 5488 // may have been a different number of threads doing the discovery 5489 // and a different number of discovered lists may have Ref objects. 5490 // That is OK as long as the Reference lists are balanced (see 5491 // balance_all_queues() and balance_queues()). 5492 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5493 int active_workers = ParallelGCThreads; 5494 FlexibleWorkGang* workers = gch->workers(); 5495 if (workers != NULL) { 5496 active_workers = workers->active_workers(); 5497 // The expectation is that active_workers will have already 5498 // been set to a reasonable value. If it has not been set, 5499 // investigate. 5500 assert(active_workers > 0, "Should have been set during scavenge"); 5501 } 5502 rp->set_active_mt_degree(active_workers); 5503 CMSRefProcTaskExecutor task_executor(*this); 5504 stats = rp->process_discovered_references(&_is_alive_closure, 5505 &cmsKeepAliveClosure, 5506 &cmsDrainMarkingStackClosure, 5507 &task_executor, 5508 _gc_timer_cm, 5509 _gc_tracer_cm->gc_id()); 5510 } else { 5511 stats = rp->process_discovered_references(&_is_alive_closure, 5512 &cmsKeepAliveClosure, 5513 &cmsDrainMarkingStackClosure, 5514 NULL, 5515 _gc_timer_cm, 5516 _gc_tracer_cm->gc_id()); 5517 } 5518 _gc_tracer_cm->report_gc_reference_stats(stats); 5519 5520 } 5521 5522 // This is the point where the entire marking should have completed. 5523 verify_work_stacks_empty(); 5524 5525 if (should_unload_classes()) { 5526 { 5527 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5528 5529 // Unload classes and purge the SystemDictionary. 5530 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5531 5532 // Unload nmethods. 5533 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5534 5535 // Prune dead klasses from subklass/sibling/implementor lists. 5536 Klass::clean_weak_klass_links(&_is_alive_closure); 5537 } 5538 5539 { 5540 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5541 // Clean up unreferenced symbols in symbol table. 5542 SymbolTable::unlink(); 5543 } 5544 5545 { 5546 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5547 // Delete entries for dead interned strings. 5548 StringTable::unlink(&_is_alive_closure); 5549 } 5550 } 5551 5552 5553 // Restore any preserved marks as a result of mark stack or 5554 // work queue overflow 5555 restore_preserved_marks_if_any(); // done single-threaded for now 5556 5557 rp->set_enqueuing_is_done(true); 5558 if (rp->processing_is_mt()) { 5559 rp->balance_all_queues(); 5560 CMSRefProcTaskExecutor task_executor(*this); 5561 rp->enqueue_discovered_references(&task_executor); 5562 } else { 5563 rp->enqueue_discovered_references(NULL); 5564 } 5565 rp->verify_no_references_recorded(); 5566 assert(!rp->discovery_enabled(), "should have been disabled"); 5567 } 5568 5569 #ifndef PRODUCT 5570 void CMSCollector::check_correct_thread_executing() { 5571 Thread* t = Thread::current(); 5572 // Only the VM thread or the CMS thread should be here. 5573 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5574 "Unexpected thread type"); 5575 // If this is the vm thread, the foreground process 5576 // should not be waiting. Note that _foregroundGCIsActive is 5577 // true while the foreground collector is waiting. 5578 if (_foregroundGCShouldWait) { 5579 // We cannot be the VM thread 5580 assert(t->is_ConcurrentGC_thread(), 5581 "Should be CMS thread"); 5582 } else { 5583 // We can be the CMS thread only if we are in a stop-world 5584 // phase of CMS collection. 5585 if (t->is_ConcurrentGC_thread()) { 5586 assert(_collectorState == InitialMarking || 5587 _collectorState == FinalMarking, 5588 "Should be a stop-world phase"); 5589 // The CMS thread should be holding the CMS_token. 5590 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5591 "Potential interference with concurrently " 5592 "executing VM thread"); 5593 } 5594 } 5595 } 5596 #endif 5597 5598 void CMSCollector::sweep() { 5599 assert(_collectorState == Sweeping, "just checking"); 5600 check_correct_thread_executing(); 5601 verify_work_stacks_empty(); 5602 verify_overflow_empty(); 5603 increment_sweep_count(); 5604 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5605 5606 _inter_sweep_timer.stop(); 5607 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5608 5609 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5610 _intra_sweep_timer.reset(); 5611 _intra_sweep_timer.start(); 5612 { 5613 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5614 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5615 // First sweep the old gen 5616 { 5617 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5618 bitMapLock()); 5619 sweepWork(_cmsGen); 5620 } 5621 5622 // Update Universe::_heap_*_at_gc figures. 5623 // We need all the free list locks to make the abstract state 5624 // transition from Sweeping to Resetting. See detailed note 5625 // further below. 5626 { 5627 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5628 // Update heap occupancy information which is used as 5629 // input to soft ref clearing policy at the next gc. 5630 Universe::update_heap_info_at_gc(); 5631 _collectorState = Resizing; 5632 } 5633 } 5634 verify_work_stacks_empty(); 5635 verify_overflow_empty(); 5636 5637 if (should_unload_classes()) { 5638 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5639 // requires that the virtual spaces are stable and not deleted. 5640 ClassLoaderDataGraph::set_should_purge(true); 5641 } 5642 5643 _intra_sweep_timer.stop(); 5644 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5645 5646 _inter_sweep_timer.reset(); 5647 _inter_sweep_timer.start(); 5648 5649 // We need to use a monotonically non-decreasing time in ms 5650 // or we will see time-warp warnings and os::javaTimeMillis() 5651 // does not guarantee monotonicity. 5652 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5653 update_time_of_last_gc(now); 5654 5655 // NOTE on abstract state transitions: 5656 // Mutators allocate-live and/or mark the mod-union table dirty 5657 // based on the state of the collection. The former is done in 5658 // the interval [Marking, Sweeping] and the latter in the interval 5659 // [Marking, Sweeping). Thus the transitions into the Marking state 5660 // and out of the Sweeping state must be synchronously visible 5661 // globally to the mutators. 5662 // The transition into the Marking state happens with the world 5663 // stopped so the mutators will globally see it. Sweeping is 5664 // done asynchronously by the background collector so the transition 5665 // from the Sweeping state to the Resizing state must be done 5666 // under the freelistLock (as is the check for whether to 5667 // allocate-live and whether to dirty the mod-union table). 5668 assert(_collectorState == Resizing, "Change of collector state to" 5669 " Resizing must be done under the freelistLocks (plural)"); 5670 5671 // Now that sweeping has been completed, we clear 5672 // the incremental_collection_failed flag, 5673 // thus inviting a younger gen collection to promote into 5674 // this generation. If such a promotion may still fail, 5675 // the flag will be set again when a young collection is 5676 // attempted. 5677 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5678 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5679 gch->update_full_collections_completed(_collection_count_start); 5680 } 5681 5682 // FIX ME!!! Looks like this belongs in CFLSpace, with 5683 // CMSGen merely delegating to it. 5684 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5685 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5686 HeapWord* minAddr = _cmsSpace->bottom(); 5687 HeapWord* largestAddr = 5688 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5689 if (largestAddr == NULL) { 5690 // The dictionary appears to be empty. In this case 5691 // try to coalesce at the end of the heap. 5692 largestAddr = _cmsSpace->end(); 5693 } 5694 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5695 size_t nearLargestOffset = 5696 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5697 if (PrintFLSStatistics != 0) { 5698 gclog_or_tty->print_cr( 5699 "CMS: Large Block: " PTR_FORMAT ";" 5700 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5701 p2i(largestAddr), 5702 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5703 } 5704 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5705 } 5706 5707 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5708 return addr >= _cmsSpace->nearLargestChunk(); 5709 } 5710 5711 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5712 return _cmsSpace->find_chunk_at_end(); 5713 } 5714 5715 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5716 bool full) { 5717 // The next lower level has been collected. Gather any statistics 5718 // that are of interest at this point. 5719 if (!full && (current_level + 1) == level()) { 5720 // Gather statistics on the young generation collection. 5721 collector()->stats().record_gc0_end(used()); 5722 } 5723 } 5724 5725 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5726 // We iterate over the space(s) underlying this generation, 5727 // checking the mark bit map to see if the bits corresponding 5728 // to specific blocks are marked or not. Blocks that are 5729 // marked are live and are not swept up. All remaining blocks 5730 // are swept up, with coalescing on-the-fly as we sweep up 5731 // contiguous free and/or garbage blocks: 5732 // We need to ensure that the sweeper synchronizes with allocators 5733 // and stop-the-world collectors. In particular, the following 5734 // locks are used: 5735 // . CMS token: if this is held, a stop the world collection cannot occur 5736 // . freelistLock: if this is held no allocation can occur from this 5737 // generation by another thread 5738 // . bitMapLock: if this is held, no other thread can access or update 5739 // 5740 5741 // Note that we need to hold the freelistLock if we use 5742 // block iterate below; else the iterator might go awry if 5743 // a mutator (or promotion) causes block contents to change 5744 // (for instance if the allocator divvies up a block). 5745 // If we hold the free list lock, for all practical purposes 5746 // young generation GC's can't occur (they'll usually need to 5747 // promote), so we might as well prevent all young generation 5748 // GC's while we do a sweeping step. For the same reason, we might 5749 // as well take the bit map lock for the entire duration 5750 5751 // check that we hold the requisite locks 5752 assert(have_cms_token(), "Should hold cms token"); 5753 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5754 assert_lock_strong(gen->freelistLock()); 5755 assert_lock_strong(bitMapLock()); 5756 5757 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5758 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5759 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5760 _inter_sweep_estimate.padded_average(), 5761 _intra_sweep_estimate.padded_average()); 5762 gen->setNearLargestChunk(); 5763 5764 { 5765 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5766 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5767 // We need to free-up/coalesce garbage/blocks from a 5768 // co-terminal free run. This is done in the SweepClosure 5769 // destructor; so, do not remove this scope, else the 5770 // end-of-sweep-census below will be off by a little bit. 5771 } 5772 gen->cmsSpace()->sweep_completed(); 5773 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5774 if (should_unload_classes()) { // unloaded classes this cycle, 5775 _concurrent_cycles_since_last_unload = 0; // ... reset count 5776 } else { // did not unload classes, 5777 _concurrent_cycles_since_last_unload++; // ... increment count 5778 } 5779 } 5780 5781 // Reset CMS data structures (for now just the marking bit map) 5782 // preparatory for the next cycle. 5783 void CMSCollector::reset(bool concurrent) { 5784 if (concurrent) { 5785 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5786 5787 // If the state is not "Resetting", the foreground thread 5788 // has done a collection and the resetting. 5789 if (_collectorState != Resetting) { 5790 assert(_collectorState == Idling, "The state should only change" 5791 " because the foreground collector has finished the collection"); 5792 return; 5793 } 5794 5795 // Clear the mark bitmap (no grey objects to start with) 5796 // for the next cycle. 5797 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5798 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5799 5800 HeapWord* curAddr = _markBitMap.startWord(); 5801 while (curAddr < _markBitMap.endWord()) { 5802 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5803 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5804 _markBitMap.clear_large_range(chunk); 5805 if (ConcurrentMarkSweepThread::should_yield() && 5806 !foregroundGCIsActive() && 5807 CMSYield) { 5808 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5809 "CMS thread should hold CMS token"); 5810 assert_lock_strong(bitMapLock()); 5811 bitMapLock()->unlock(); 5812 ConcurrentMarkSweepThread::desynchronize(true); 5813 stopTimer(); 5814 if (PrintCMSStatistics != 0) { 5815 incrementYields(); 5816 } 5817 5818 // See the comment in coordinator_yield() 5819 for (unsigned i = 0; i < CMSYieldSleepCount && 5820 ConcurrentMarkSweepThread::should_yield() && 5821 !CMSCollector::foregroundGCIsActive(); ++i) { 5822 os::sleep(Thread::current(), 1, false); 5823 } 5824 5825 ConcurrentMarkSweepThread::synchronize(true); 5826 bitMapLock()->lock_without_safepoint_check(); 5827 startTimer(); 5828 } 5829 curAddr = chunk.end(); 5830 } 5831 // A successful mostly concurrent collection has been done. 5832 // Because only the full (i.e., concurrent mode failure) collections 5833 // are being measured for gc overhead limits, clean the "near" flag 5834 // and count. 5835 size_policy()->reset_gc_overhead_limit_count(); 5836 _collectorState = Idling; 5837 } else { 5838 // already have the lock 5839 assert(_collectorState == Resetting, "just checking"); 5840 assert_lock_strong(bitMapLock()); 5841 _markBitMap.clear_all(); 5842 _collectorState = Idling; 5843 } 5844 5845 register_gc_end(); 5846 } 5847 5848 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5849 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5850 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5851 TraceCollectorStats tcs(counters()); 5852 5853 switch (op) { 5854 case CMS_op_checkpointRootsInitial: { 5855 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5856 checkpointRootsInitial(); 5857 if (PrintGC) { 5858 _cmsGen->printOccupancy("initial-mark"); 5859 } 5860 break; 5861 } 5862 case CMS_op_checkpointRootsFinal: { 5863 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5864 checkpointRootsFinal(); 5865 if (PrintGC) { 5866 _cmsGen->printOccupancy("remark"); 5867 } 5868 break; 5869 } 5870 default: 5871 fatal("No such CMS_op"); 5872 } 5873 } 5874 5875 #ifndef PRODUCT 5876 size_t const CMSCollector::skip_header_HeapWords() { 5877 return FreeChunk::header_size(); 5878 } 5879 5880 // Try and collect here conditions that should hold when 5881 // CMS thread is exiting. The idea is that the foreground GC 5882 // thread should not be blocked if it wants to terminate 5883 // the CMS thread and yet continue to run the VM for a while 5884 // after that. 5885 void CMSCollector::verify_ok_to_terminate() const { 5886 assert(Thread::current()->is_ConcurrentGC_thread(), 5887 "should be called by CMS thread"); 5888 assert(!_foregroundGCShouldWait, "should be false"); 5889 // We could check here that all the various low-level locks 5890 // are not held by the CMS thread, but that is overkill; see 5891 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5892 // is checked. 5893 } 5894 #endif 5895 5896 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5897 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5898 "missing Printezis mark?"); 5899 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5900 size_t size = pointer_delta(nextOneAddr + 1, addr); 5901 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5902 "alignment problem"); 5903 assert(size >= 3, "Necessary for Printezis marks to work"); 5904 return size; 5905 } 5906 5907 // A variant of the above (block_size_using_printezis_bits()) except 5908 // that we return 0 if the P-bits are not yet set. 5909 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5910 if (_markBitMap.isMarked(addr + 1)) { 5911 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5912 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5913 size_t size = pointer_delta(nextOneAddr + 1, addr); 5914 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5915 "alignment problem"); 5916 assert(size >= 3, "Necessary for Printezis marks to work"); 5917 return size; 5918 } 5919 return 0; 5920 } 5921 5922 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5923 size_t sz = 0; 5924 oop p = (oop)addr; 5925 if (p->klass_or_null() != NULL) { 5926 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5927 } else { 5928 sz = block_size_using_printezis_bits(addr); 5929 } 5930 assert(sz > 0, "size must be nonzero"); 5931 HeapWord* next_block = addr + sz; 5932 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5933 CardTableModRefBS::card_size); 5934 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5935 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5936 "must be different cards"); 5937 return next_card; 5938 } 5939 5940 5941 // CMS Bit Map Wrapper ///////////////////////////////////////// 5942 5943 // Construct a CMS bit map infrastructure, but don't create the 5944 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5945 // further below. 5946 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5947 _bm(), 5948 _shifter(shifter), 5949 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5950 Monitor::_safepoint_check_sometimes) : NULL) 5951 { 5952 _bmStartWord = 0; 5953 _bmWordSize = 0; 5954 } 5955 5956 bool CMSBitMap::allocate(MemRegion mr) { 5957 _bmStartWord = mr.start(); 5958 _bmWordSize = mr.word_size(); 5959 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5960 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5961 if (!brs.is_reserved()) { 5962 warning("CMS bit map allocation failure"); 5963 return false; 5964 } 5965 // For now we'll just commit all of the bit map up front. 5966 // Later on we'll try to be more parsimonious with swap. 5967 if (!_virtual_space.initialize(brs, brs.size())) { 5968 warning("CMS bit map backing store failure"); 5969 return false; 5970 } 5971 assert(_virtual_space.committed_size() == brs.size(), 5972 "didn't reserve backing store for all of CMS bit map?"); 5973 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5974 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5975 _bmWordSize, "inconsistency in bit map sizing"); 5976 _bm.set_size(_bmWordSize >> _shifter); 5977 5978 // bm.clear(); // can we rely on getting zero'd memory? verify below 5979 assert(isAllClear(), 5980 "Expected zero'd memory from ReservedSpace constructor"); 5981 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5982 "consistency check"); 5983 return true; 5984 } 5985 5986 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5987 HeapWord *next_addr, *end_addr, *last_addr; 5988 assert_locked(); 5989 assert(covers(mr), "out-of-range error"); 5990 // XXX assert that start and end are appropriately aligned 5991 for (next_addr = mr.start(), end_addr = mr.end(); 5992 next_addr < end_addr; next_addr = last_addr) { 5993 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5994 last_addr = dirty_region.end(); 5995 if (!dirty_region.is_empty()) { 5996 cl->do_MemRegion(dirty_region); 5997 } else { 5998 assert(last_addr == end_addr, "program logic"); 5999 return; 6000 } 6001 } 6002 } 6003 6004 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6005 _bm.print_on_error(st, prefix); 6006 } 6007 6008 #ifndef PRODUCT 6009 void CMSBitMap::assert_locked() const { 6010 CMSLockVerifier::assert_locked(lock()); 6011 } 6012 6013 bool CMSBitMap::covers(MemRegion mr) const { 6014 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6015 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6016 "size inconsistency"); 6017 return (mr.start() >= _bmStartWord) && 6018 (mr.end() <= endWord()); 6019 } 6020 6021 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6022 return (start >= _bmStartWord && (start + size) <= endWord()); 6023 } 6024 6025 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6026 // verify that there are no 1 bits in the interval [left, right) 6027 FalseBitMapClosure falseBitMapClosure; 6028 iterate(&falseBitMapClosure, left, right); 6029 } 6030 6031 void CMSBitMap::region_invariant(MemRegion mr) 6032 { 6033 assert_locked(); 6034 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6035 assert(!mr.is_empty(), "unexpected empty region"); 6036 assert(covers(mr), "mr should be covered by bit map"); 6037 // convert address range into offset range 6038 size_t start_ofs = heapWordToOffset(mr.start()); 6039 // Make sure that end() is appropriately aligned 6040 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6041 (1 << (_shifter+LogHeapWordSize))), 6042 "Misaligned mr.end()"); 6043 size_t end_ofs = heapWordToOffset(mr.end()); 6044 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6045 } 6046 6047 #endif 6048 6049 bool CMSMarkStack::allocate(size_t size) { 6050 // allocate a stack of the requisite depth 6051 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6052 size * sizeof(oop))); 6053 if (!rs.is_reserved()) { 6054 warning("CMSMarkStack allocation failure"); 6055 return false; 6056 } 6057 if (!_virtual_space.initialize(rs, rs.size())) { 6058 warning("CMSMarkStack backing store failure"); 6059 return false; 6060 } 6061 assert(_virtual_space.committed_size() == rs.size(), 6062 "didn't reserve backing store for all of CMS stack?"); 6063 _base = (oop*)(_virtual_space.low()); 6064 _index = 0; 6065 _capacity = size; 6066 NOT_PRODUCT(_max_depth = 0); 6067 return true; 6068 } 6069 6070 // XXX FIX ME !!! In the MT case we come in here holding a 6071 // leaf lock. For printing we need to take a further lock 6072 // which has lower rank. We need to recalibrate the two 6073 // lock-ranks involved in order to be able to print the 6074 // messages below. (Or defer the printing to the caller. 6075 // For now we take the expedient path of just disabling the 6076 // messages for the problematic case.) 6077 void CMSMarkStack::expand() { 6078 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6079 if (_capacity == MarkStackSizeMax) { 6080 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6081 // We print a warning message only once per CMS cycle. 6082 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6083 } 6084 return; 6085 } 6086 // Double capacity if possible 6087 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6088 // Do not give up existing stack until we have managed to 6089 // get the double capacity that we desired. 6090 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6091 new_capacity * sizeof(oop))); 6092 if (rs.is_reserved()) { 6093 // Release the backing store associated with old stack 6094 _virtual_space.release(); 6095 // Reinitialize virtual space for new stack 6096 if (!_virtual_space.initialize(rs, rs.size())) { 6097 fatal("Not enough swap for expanded marking stack"); 6098 } 6099 _base = (oop*)(_virtual_space.low()); 6100 _index = 0; 6101 _capacity = new_capacity; 6102 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6103 // Failed to double capacity, continue; 6104 // we print a detail message only once per CMS cycle. 6105 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6106 SIZE_FORMAT"K", 6107 _capacity / K, new_capacity / K); 6108 } 6109 } 6110 6111 6112 // Closures 6113 // XXX: there seems to be a lot of code duplication here; 6114 // should refactor and consolidate common code. 6115 6116 // This closure is used to mark refs into the CMS generation in 6117 // the CMS bit map. Called at the first checkpoint. This closure 6118 // assumes that we do not need to re-mark dirty cards; if the CMS 6119 // generation on which this is used is not an oldest 6120 // generation then this will lose younger_gen cards! 6121 6122 MarkRefsIntoClosure::MarkRefsIntoClosure( 6123 MemRegion span, CMSBitMap* bitMap): 6124 _span(span), 6125 _bitMap(bitMap) 6126 { 6127 assert(_ref_processor == NULL, "deliberately left NULL"); 6128 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6129 } 6130 6131 void MarkRefsIntoClosure::do_oop(oop obj) { 6132 // if p points into _span, then mark corresponding bit in _markBitMap 6133 assert(obj->is_oop(), "expected an oop"); 6134 HeapWord* addr = (HeapWord*)obj; 6135 if (_span.contains(addr)) { 6136 // this should be made more efficient 6137 _bitMap->mark(addr); 6138 } 6139 } 6140 6141 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6142 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6143 6144 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6145 MemRegion span, CMSBitMap* bitMap): 6146 _span(span), 6147 _bitMap(bitMap) 6148 { 6149 assert(_ref_processor == NULL, "deliberately left NULL"); 6150 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6151 } 6152 6153 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6154 // if p points into _span, then mark corresponding bit in _markBitMap 6155 assert(obj->is_oop(), "expected an oop"); 6156 HeapWord* addr = (HeapWord*)obj; 6157 if (_span.contains(addr)) { 6158 // this should be made more efficient 6159 _bitMap->par_mark(addr); 6160 } 6161 } 6162 6163 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6164 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6165 6166 // A variant of the above, used for CMS marking verification. 6167 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6168 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6169 _span(span), 6170 _verification_bm(verification_bm), 6171 _cms_bm(cms_bm) 6172 { 6173 assert(_ref_processor == NULL, "deliberately left NULL"); 6174 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6175 } 6176 6177 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6178 // if p points into _span, then mark corresponding bit in _markBitMap 6179 assert(obj->is_oop(), "expected an oop"); 6180 HeapWord* addr = (HeapWord*)obj; 6181 if (_span.contains(addr)) { 6182 _verification_bm->mark(addr); 6183 if (!_cms_bm->isMarked(addr)) { 6184 oop(addr)->print(); 6185 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6186 fatal("... aborting"); 6187 } 6188 } 6189 } 6190 6191 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6192 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6193 6194 ////////////////////////////////////////////////// 6195 // MarkRefsIntoAndScanClosure 6196 ////////////////////////////////////////////////// 6197 6198 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6199 ReferenceProcessor* rp, 6200 CMSBitMap* bit_map, 6201 CMSBitMap* mod_union_table, 6202 CMSMarkStack* mark_stack, 6203 CMSCollector* collector, 6204 bool should_yield, 6205 bool concurrent_precleaning): 6206 _collector(collector), 6207 _span(span), 6208 _bit_map(bit_map), 6209 _mark_stack(mark_stack), 6210 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6211 mark_stack, concurrent_precleaning), 6212 _yield(should_yield), 6213 _concurrent_precleaning(concurrent_precleaning), 6214 _freelistLock(NULL) 6215 { 6216 _ref_processor = rp; 6217 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6218 } 6219 6220 // This closure is used to mark refs into the CMS generation at the 6221 // second (final) checkpoint, and to scan and transitively follow 6222 // the unmarked oops. It is also used during the concurrent precleaning 6223 // phase while scanning objects on dirty cards in the CMS generation. 6224 // The marks are made in the marking bit map and the marking stack is 6225 // used for keeping the (newly) grey objects during the scan. 6226 // The parallel version (Par_...) appears further below. 6227 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6228 if (obj != NULL) { 6229 assert(obj->is_oop(), "expected an oop"); 6230 HeapWord* addr = (HeapWord*)obj; 6231 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6232 assert(_collector->overflow_list_is_empty(), 6233 "overflow list should be empty"); 6234 if (_span.contains(addr) && 6235 !_bit_map->isMarked(addr)) { 6236 // mark bit map (object is now grey) 6237 _bit_map->mark(addr); 6238 // push on marking stack (stack should be empty), and drain the 6239 // stack by applying this closure to the oops in the oops popped 6240 // from the stack (i.e. blacken the grey objects) 6241 bool res = _mark_stack->push(obj); 6242 assert(res, "Should have space to push on empty stack"); 6243 do { 6244 oop new_oop = _mark_stack->pop(); 6245 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6246 assert(_bit_map->isMarked((HeapWord*)new_oop), 6247 "only grey objects on this stack"); 6248 // iterate over the oops in this oop, marking and pushing 6249 // the ones in CMS heap (i.e. in _span). 6250 new_oop->oop_iterate(&_pushAndMarkClosure); 6251 // check if it's time to yield 6252 do_yield_check(); 6253 } while (!_mark_stack->isEmpty() || 6254 (!_concurrent_precleaning && take_from_overflow_list())); 6255 // if marking stack is empty, and we are not doing this 6256 // during precleaning, then check the overflow list 6257 } 6258 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6259 assert(_collector->overflow_list_is_empty(), 6260 "overflow list was drained above"); 6261 // We could restore evacuated mark words, if any, used for 6262 // overflow list links here because the overflow list is 6263 // provably empty here. That would reduce the maximum 6264 // size requirements for preserved_{oop,mark}_stack. 6265 // But we'll just postpone it until we are all done 6266 // so we can just stream through. 6267 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6268 _collector->restore_preserved_marks_if_any(); 6269 assert(_collector->no_preserved_marks(), "No preserved marks"); 6270 } 6271 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6272 "All preserved marks should have been restored above"); 6273 } 6274 } 6275 6276 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6277 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6278 6279 void MarkRefsIntoAndScanClosure::do_yield_work() { 6280 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6281 "CMS thread should hold CMS token"); 6282 assert_lock_strong(_freelistLock); 6283 assert_lock_strong(_bit_map->lock()); 6284 // relinquish the free_list_lock and bitMaplock() 6285 _bit_map->lock()->unlock(); 6286 _freelistLock->unlock(); 6287 ConcurrentMarkSweepThread::desynchronize(true); 6288 _collector->stopTimer(); 6289 if (PrintCMSStatistics != 0) { 6290 _collector->incrementYields(); 6291 } 6292 6293 // See the comment in coordinator_yield() 6294 for (unsigned i = 0; 6295 i < CMSYieldSleepCount && 6296 ConcurrentMarkSweepThread::should_yield() && 6297 !CMSCollector::foregroundGCIsActive(); 6298 ++i) { 6299 os::sleep(Thread::current(), 1, false); 6300 } 6301 6302 ConcurrentMarkSweepThread::synchronize(true); 6303 _freelistLock->lock_without_safepoint_check(); 6304 _bit_map->lock()->lock_without_safepoint_check(); 6305 _collector->startTimer(); 6306 } 6307 6308 /////////////////////////////////////////////////////////// 6309 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6310 // MarkRefsIntoAndScanClosure 6311 /////////////////////////////////////////////////////////// 6312 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6313 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6314 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6315 _span(span), 6316 _bit_map(bit_map), 6317 _work_queue(work_queue), 6318 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6319 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6320 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6321 { 6322 _ref_processor = rp; 6323 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6324 } 6325 6326 // This closure is used to mark refs into the CMS generation at the 6327 // second (final) checkpoint, and to scan and transitively follow 6328 // the unmarked oops. The marks are made in the marking bit map and 6329 // the work_queue is used for keeping the (newly) grey objects during 6330 // the scan phase whence they are also available for stealing by parallel 6331 // threads. Since the marking bit map is shared, updates are 6332 // synchronized (via CAS). 6333 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6334 if (obj != NULL) { 6335 // Ignore mark word because this could be an already marked oop 6336 // that may be chained at the end of the overflow list. 6337 assert(obj->is_oop(true), "expected an oop"); 6338 HeapWord* addr = (HeapWord*)obj; 6339 if (_span.contains(addr) && 6340 !_bit_map->isMarked(addr)) { 6341 // mark bit map (object will become grey): 6342 // It is possible for several threads to be 6343 // trying to "claim" this object concurrently; 6344 // the unique thread that succeeds in marking the 6345 // object first will do the subsequent push on 6346 // to the work queue (or overflow list). 6347 if (_bit_map->par_mark(addr)) { 6348 // push on work_queue (which may not be empty), and trim the 6349 // queue to an appropriate length by applying this closure to 6350 // the oops in the oops popped from the stack (i.e. blacken the 6351 // grey objects) 6352 bool res = _work_queue->push(obj); 6353 assert(res, "Low water mark should be less than capacity?"); 6354 trim_queue(_low_water_mark); 6355 } // Else, another thread claimed the object 6356 } 6357 } 6358 } 6359 6360 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6361 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6362 6363 // This closure is used to rescan the marked objects on the dirty cards 6364 // in the mod union table and the card table proper. 6365 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6366 oop p, MemRegion mr) { 6367 6368 size_t size = 0; 6369 HeapWord* addr = (HeapWord*)p; 6370 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6371 assert(_span.contains(addr), "we are scanning the CMS generation"); 6372 // check if it's time to yield 6373 if (do_yield_check()) { 6374 // We yielded for some foreground stop-world work, 6375 // and we have been asked to abort this ongoing preclean cycle. 6376 return 0; 6377 } 6378 if (_bitMap->isMarked(addr)) { 6379 // it's marked; is it potentially uninitialized? 6380 if (p->klass_or_null() != NULL) { 6381 // an initialized object; ignore mark word in verification below 6382 // since we are running concurrent with mutators 6383 assert(p->is_oop(true), "should be an oop"); 6384 if (p->is_objArray()) { 6385 // objArrays are precisely marked; restrict scanning 6386 // to dirty cards only. 6387 size = CompactibleFreeListSpace::adjustObjectSize( 6388 p->oop_iterate(_scanningClosure, mr)); 6389 } else { 6390 // A non-array may have been imprecisely marked; we need 6391 // to scan object in its entirety. 6392 size = CompactibleFreeListSpace::adjustObjectSize( 6393 p->oop_iterate(_scanningClosure)); 6394 } 6395 #ifdef ASSERT 6396 size_t direct_size = 6397 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6398 assert(size == direct_size, "Inconsistency in size"); 6399 assert(size >= 3, "Necessary for Printezis marks to work"); 6400 if (!_bitMap->isMarked(addr+1)) { 6401 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6402 } else { 6403 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6404 assert(_bitMap->isMarked(addr+size-1), 6405 "inconsistent Printezis mark"); 6406 } 6407 #endif // ASSERT 6408 } else { 6409 // An uninitialized object. 6410 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6411 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6412 size = pointer_delta(nextOneAddr + 1, addr); 6413 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6414 "alignment problem"); 6415 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6416 // will dirty the card when the klass pointer is installed in the 6417 // object (signaling the completion of initialization). 6418 } 6419 } else { 6420 // Either a not yet marked object or an uninitialized object 6421 if (p->klass_or_null() == NULL) { 6422 // An uninitialized object, skip to the next card, since 6423 // we may not be able to read its P-bits yet. 6424 assert(size == 0, "Initial value"); 6425 } else { 6426 // An object not (yet) reached by marking: we merely need to 6427 // compute its size so as to go look at the next block. 6428 assert(p->is_oop(true), "should be an oop"); 6429 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6430 } 6431 } 6432 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6433 return size; 6434 } 6435 6436 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6437 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6438 "CMS thread should hold CMS token"); 6439 assert_lock_strong(_freelistLock); 6440 assert_lock_strong(_bitMap->lock()); 6441 // relinquish the free_list_lock and bitMaplock() 6442 _bitMap->lock()->unlock(); 6443 _freelistLock->unlock(); 6444 ConcurrentMarkSweepThread::desynchronize(true); 6445 _collector->stopTimer(); 6446 if (PrintCMSStatistics != 0) { 6447 _collector->incrementYields(); 6448 } 6449 6450 // See the comment in coordinator_yield() 6451 for (unsigned i = 0; i < CMSYieldSleepCount && 6452 ConcurrentMarkSweepThread::should_yield() && 6453 !CMSCollector::foregroundGCIsActive(); ++i) { 6454 os::sleep(Thread::current(), 1, false); 6455 } 6456 6457 ConcurrentMarkSweepThread::synchronize(true); 6458 _freelistLock->lock_without_safepoint_check(); 6459 _bitMap->lock()->lock_without_safepoint_check(); 6460 _collector->startTimer(); 6461 } 6462 6463 6464 ////////////////////////////////////////////////////////////////// 6465 // SurvivorSpacePrecleanClosure 6466 ////////////////////////////////////////////////////////////////// 6467 // This (single-threaded) closure is used to preclean the oops in 6468 // the survivor spaces. 6469 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6470 6471 HeapWord* addr = (HeapWord*)p; 6472 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6473 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6474 assert(p->klass_or_null() != NULL, "object should be initialized"); 6475 // an initialized object; ignore mark word in verification below 6476 // since we are running concurrent with mutators 6477 assert(p->is_oop(true), "should be an oop"); 6478 // Note that we do not yield while we iterate over 6479 // the interior oops of p, pushing the relevant ones 6480 // on our marking stack. 6481 size_t size = p->oop_iterate(_scanning_closure); 6482 do_yield_check(); 6483 // Observe that below, we do not abandon the preclean 6484 // phase as soon as we should; rather we empty the 6485 // marking stack before returning. This is to satisfy 6486 // some existing assertions. In general, it may be a 6487 // good idea to abort immediately and complete the marking 6488 // from the grey objects at a later time. 6489 while (!_mark_stack->isEmpty()) { 6490 oop new_oop = _mark_stack->pop(); 6491 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6492 assert(_bit_map->isMarked((HeapWord*)new_oop), 6493 "only grey objects on this stack"); 6494 // iterate over the oops in this oop, marking and pushing 6495 // the ones in CMS heap (i.e. in _span). 6496 new_oop->oop_iterate(_scanning_closure); 6497 // check if it's time to yield 6498 do_yield_check(); 6499 } 6500 unsigned int after_count = 6501 GenCollectedHeap::heap()->total_collections(); 6502 bool abort = (_before_count != after_count) || 6503 _collector->should_abort_preclean(); 6504 return abort ? 0 : size; 6505 } 6506 6507 void SurvivorSpacePrecleanClosure::do_yield_work() { 6508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6509 "CMS thread should hold CMS token"); 6510 assert_lock_strong(_bit_map->lock()); 6511 // Relinquish the bit map lock 6512 _bit_map->lock()->unlock(); 6513 ConcurrentMarkSweepThread::desynchronize(true); 6514 _collector->stopTimer(); 6515 if (PrintCMSStatistics != 0) { 6516 _collector->incrementYields(); 6517 } 6518 6519 // See the comment in coordinator_yield() 6520 for (unsigned i = 0; i < CMSYieldSleepCount && 6521 ConcurrentMarkSweepThread::should_yield() && 6522 !CMSCollector::foregroundGCIsActive(); ++i) { 6523 os::sleep(Thread::current(), 1, false); 6524 } 6525 6526 ConcurrentMarkSweepThread::synchronize(true); 6527 _bit_map->lock()->lock_without_safepoint_check(); 6528 _collector->startTimer(); 6529 } 6530 6531 // This closure is used to rescan the marked objects on the dirty cards 6532 // in the mod union table and the card table proper. In the parallel 6533 // case, although the bitMap is shared, we do a single read so the 6534 // isMarked() query is "safe". 6535 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6536 // Ignore mark word because we are running concurrent with mutators 6537 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6538 HeapWord* addr = (HeapWord*)p; 6539 assert(_span.contains(addr), "we are scanning the CMS generation"); 6540 bool is_obj_array = false; 6541 #ifdef ASSERT 6542 if (!_parallel) { 6543 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6544 assert(_collector->overflow_list_is_empty(), 6545 "overflow list should be empty"); 6546 6547 } 6548 #endif // ASSERT 6549 if (_bit_map->isMarked(addr)) { 6550 // Obj arrays are precisely marked, non-arrays are not; 6551 // so we scan objArrays precisely and non-arrays in their 6552 // entirety. 6553 if (p->is_objArray()) { 6554 is_obj_array = true; 6555 if (_parallel) { 6556 p->oop_iterate(_par_scan_closure, mr); 6557 } else { 6558 p->oop_iterate(_scan_closure, mr); 6559 } 6560 } else { 6561 if (_parallel) { 6562 p->oop_iterate(_par_scan_closure); 6563 } else { 6564 p->oop_iterate(_scan_closure); 6565 } 6566 } 6567 } 6568 #ifdef ASSERT 6569 if (!_parallel) { 6570 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6571 assert(_collector->overflow_list_is_empty(), 6572 "overflow list should be empty"); 6573 6574 } 6575 #endif // ASSERT 6576 return is_obj_array; 6577 } 6578 6579 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6580 MemRegion span, 6581 CMSBitMap* bitMap, CMSMarkStack* markStack, 6582 bool should_yield, bool verifying): 6583 _collector(collector), 6584 _span(span), 6585 _bitMap(bitMap), 6586 _mut(&collector->_modUnionTable), 6587 _markStack(markStack), 6588 _yield(should_yield), 6589 _skipBits(0) 6590 { 6591 assert(_markStack->isEmpty(), "stack should be empty"); 6592 _finger = _bitMap->startWord(); 6593 _threshold = _finger; 6594 assert(_collector->_restart_addr == NULL, "Sanity check"); 6595 assert(_span.contains(_finger), "Out of bounds _finger?"); 6596 DEBUG_ONLY(_verifying = verifying;) 6597 } 6598 6599 void MarkFromRootsClosure::reset(HeapWord* addr) { 6600 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6601 assert(_span.contains(addr), "Out of bounds _finger?"); 6602 _finger = addr; 6603 _threshold = (HeapWord*)round_to( 6604 (intptr_t)_finger, CardTableModRefBS::card_size); 6605 } 6606 6607 // Should revisit to see if this should be restructured for 6608 // greater efficiency. 6609 bool MarkFromRootsClosure::do_bit(size_t offset) { 6610 if (_skipBits > 0) { 6611 _skipBits--; 6612 return true; 6613 } 6614 // convert offset into a HeapWord* 6615 HeapWord* addr = _bitMap->startWord() + offset; 6616 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6617 "address out of range"); 6618 assert(_bitMap->isMarked(addr), "tautology"); 6619 if (_bitMap->isMarked(addr+1)) { 6620 // this is an allocated but not yet initialized object 6621 assert(_skipBits == 0, "tautology"); 6622 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6623 oop p = oop(addr); 6624 if (p->klass_or_null() == NULL) { 6625 DEBUG_ONLY(if (!_verifying) {) 6626 // We re-dirty the cards on which this object lies and increase 6627 // the _threshold so that we'll come back to scan this object 6628 // during the preclean or remark phase. (CMSCleanOnEnter) 6629 if (CMSCleanOnEnter) { 6630 size_t sz = _collector->block_size_using_printezis_bits(addr); 6631 HeapWord* end_card_addr = (HeapWord*)round_to( 6632 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6633 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6634 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6635 // Bump _threshold to end_card_addr; note that 6636 // _threshold cannot possibly exceed end_card_addr, anyhow. 6637 // This prevents future clearing of the card as the scan proceeds 6638 // to the right. 6639 assert(_threshold <= end_card_addr, 6640 "Because we are just scanning into this object"); 6641 if (_threshold < end_card_addr) { 6642 _threshold = end_card_addr; 6643 } 6644 if (p->klass_or_null() != NULL) { 6645 // Redirty the range of cards... 6646 _mut->mark_range(redirty_range); 6647 } // ...else the setting of klass will dirty the card anyway. 6648 } 6649 DEBUG_ONLY(}) 6650 return true; 6651 } 6652 } 6653 scanOopsInOop(addr); 6654 return true; 6655 } 6656 6657 // We take a break if we've been at this for a while, 6658 // so as to avoid monopolizing the locks involved. 6659 void MarkFromRootsClosure::do_yield_work() { 6660 // First give up the locks, then yield, then re-lock 6661 // We should probably use a constructor/destructor idiom to 6662 // do this unlock/lock or modify the MutexUnlocker class to 6663 // serve our purpose. XXX 6664 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6665 "CMS thread should hold CMS token"); 6666 assert_lock_strong(_bitMap->lock()); 6667 _bitMap->lock()->unlock(); 6668 ConcurrentMarkSweepThread::desynchronize(true); 6669 _collector->stopTimer(); 6670 if (PrintCMSStatistics != 0) { 6671 _collector->incrementYields(); 6672 } 6673 6674 // See the comment in coordinator_yield() 6675 for (unsigned i = 0; i < CMSYieldSleepCount && 6676 ConcurrentMarkSweepThread::should_yield() && 6677 !CMSCollector::foregroundGCIsActive(); ++i) { 6678 os::sleep(Thread::current(), 1, false); 6679 } 6680 6681 ConcurrentMarkSweepThread::synchronize(true); 6682 _bitMap->lock()->lock_without_safepoint_check(); 6683 _collector->startTimer(); 6684 } 6685 6686 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6687 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6688 assert(_markStack->isEmpty(), 6689 "should drain stack to limit stack usage"); 6690 // convert ptr to an oop preparatory to scanning 6691 oop obj = oop(ptr); 6692 // Ignore mark word in verification below, since we 6693 // may be running concurrent with mutators. 6694 assert(obj->is_oop(true), "should be an oop"); 6695 assert(_finger <= ptr, "_finger runneth ahead"); 6696 // advance the finger to right end of this object 6697 _finger = ptr + obj->size(); 6698 assert(_finger > ptr, "we just incremented it above"); 6699 // On large heaps, it may take us some time to get through 6700 // the marking phase. During 6701 // this time it's possible that a lot of mutations have 6702 // accumulated in the card table and the mod union table -- 6703 // these mutation records are redundant until we have 6704 // actually traced into the corresponding card. 6705 // Here, we check whether advancing the finger would make 6706 // us cross into a new card, and if so clear corresponding 6707 // cards in the MUT (preclean them in the card-table in the 6708 // future). 6709 6710 DEBUG_ONLY(if (!_verifying) {) 6711 // The clean-on-enter optimization is disabled by default, 6712 // until we fix 6178663. 6713 if (CMSCleanOnEnter && (_finger > _threshold)) { 6714 // [_threshold, _finger) represents the interval 6715 // of cards to be cleared in MUT (or precleaned in card table). 6716 // The set of cards to be cleared is all those that overlap 6717 // with the interval [_threshold, _finger); note that 6718 // _threshold is always kept card-aligned but _finger isn't 6719 // always card-aligned. 6720 HeapWord* old_threshold = _threshold; 6721 assert(old_threshold == (HeapWord*)round_to( 6722 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6723 "_threshold should always be card-aligned"); 6724 _threshold = (HeapWord*)round_to( 6725 (intptr_t)_finger, CardTableModRefBS::card_size); 6726 MemRegion mr(old_threshold, _threshold); 6727 assert(!mr.is_empty(), "Control point invariant"); 6728 assert(_span.contains(mr), "Should clear within span"); 6729 _mut->clear_range(mr); 6730 } 6731 DEBUG_ONLY(}) 6732 // Note: the finger doesn't advance while we drain 6733 // the stack below. 6734 PushOrMarkClosure pushOrMarkClosure(_collector, 6735 _span, _bitMap, _markStack, 6736 _finger, this); 6737 bool res = _markStack->push(obj); 6738 assert(res, "Empty non-zero size stack should have space for single push"); 6739 while (!_markStack->isEmpty()) { 6740 oop new_oop = _markStack->pop(); 6741 // Skip verifying header mark word below because we are 6742 // running concurrent with mutators. 6743 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6744 // now scan this oop's oops 6745 new_oop->oop_iterate(&pushOrMarkClosure); 6746 do_yield_check(); 6747 } 6748 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6749 } 6750 6751 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6752 CMSCollector* collector, MemRegion span, 6753 CMSBitMap* bit_map, 6754 OopTaskQueue* work_queue, 6755 CMSMarkStack* overflow_stack): 6756 _collector(collector), 6757 _whole_span(collector->_span), 6758 _span(span), 6759 _bit_map(bit_map), 6760 _mut(&collector->_modUnionTable), 6761 _work_queue(work_queue), 6762 _overflow_stack(overflow_stack), 6763 _skip_bits(0), 6764 _task(task) 6765 { 6766 assert(_work_queue->size() == 0, "work_queue should be empty"); 6767 _finger = span.start(); 6768 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6769 assert(_span.contains(_finger), "Out of bounds _finger?"); 6770 } 6771 6772 // Should revisit to see if this should be restructured for 6773 // greater efficiency. 6774 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6775 if (_skip_bits > 0) { 6776 _skip_bits--; 6777 return true; 6778 } 6779 // convert offset into a HeapWord* 6780 HeapWord* addr = _bit_map->startWord() + offset; 6781 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6782 "address out of range"); 6783 assert(_bit_map->isMarked(addr), "tautology"); 6784 if (_bit_map->isMarked(addr+1)) { 6785 // this is an allocated object that might not yet be initialized 6786 assert(_skip_bits == 0, "tautology"); 6787 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6788 oop p = oop(addr); 6789 if (p->klass_or_null() == NULL) { 6790 // in the case of Clean-on-Enter optimization, redirty card 6791 // and avoid clearing card by increasing the threshold. 6792 return true; 6793 } 6794 } 6795 scan_oops_in_oop(addr); 6796 return true; 6797 } 6798 6799 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6800 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6801 // Should we assert that our work queue is empty or 6802 // below some drain limit? 6803 assert(_work_queue->size() == 0, 6804 "should drain stack to limit stack usage"); 6805 // convert ptr to an oop preparatory to scanning 6806 oop obj = oop(ptr); 6807 // Ignore mark word in verification below, since we 6808 // may be running concurrent with mutators. 6809 assert(obj->is_oop(true), "should be an oop"); 6810 assert(_finger <= ptr, "_finger runneth ahead"); 6811 // advance the finger to right end of this object 6812 _finger = ptr + obj->size(); 6813 assert(_finger > ptr, "we just incremented it above"); 6814 // On large heaps, it may take us some time to get through 6815 // the marking phase. During 6816 // this time it's possible that a lot of mutations have 6817 // accumulated in the card table and the mod union table -- 6818 // these mutation records are redundant until we have 6819 // actually traced into the corresponding card. 6820 // Here, we check whether advancing the finger would make 6821 // us cross into a new card, and if so clear corresponding 6822 // cards in the MUT (preclean them in the card-table in the 6823 // future). 6824 6825 // The clean-on-enter optimization is disabled by default, 6826 // until we fix 6178663. 6827 if (CMSCleanOnEnter && (_finger > _threshold)) { 6828 // [_threshold, _finger) represents the interval 6829 // of cards to be cleared in MUT (or precleaned in card table). 6830 // The set of cards to be cleared is all those that overlap 6831 // with the interval [_threshold, _finger); note that 6832 // _threshold is always kept card-aligned but _finger isn't 6833 // always card-aligned. 6834 HeapWord* old_threshold = _threshold; 6835 assert(old_threshold == (HeapWord*)round_to( 6836 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6837 "_threshold should always be card-aligned"); 6838 _threshold = (HeapWord*)round_to( 6839 (intptr_t)_finger, CardTableModRefBS::card_size); 6840 MemRegion mr(old_threshold, _threshold); 6841 assert(!mr.is_empty(), "Control point invariant"); 6842 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6843 _mut->clear_range(mr); 6844 } 6845 6846 // Note: the local finger doesn't advance while we drain 6847 // the stack below, but the global finger sure can and will. 6848 HeapWord** gfa = _task->global_finger_addr(); 6849 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6850 _span, _bit_map, 6851 _work_queue, 6852 _overflow_stack, 6853 _finger, 6854 gfa, this); 6855 bool res = _work_queue->push(obj); // overflow could occur here 6856 assert(res, "Will hold once we use workqueues"); 6857 while (true) { 6858 oop new_oop; 6859 if (!_work_queue->pop_local(new_oop)) { 6860 // We emptied our work_queue; check if there's stuff that can 6861 // be gotten from the overflow stack. 6862 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6863 _overflow_stack, _work_queue)) { 6864 do_yield_check(); 6865 continue; 6866 } else { // done 6867 break; 6868 } 6869 } 6870 // Skip verifying header mark word below because we are 6871 // running concurrent with mutators. 6872 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6873 // now scan this oop's oops 6874 new_oop->oop_iterate(&pushOrMarkClosure); 6875 do_yield_check(); 6876 } 6877 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6878 } 6879 6880 // Yield in response to a request from VM Thread or 6881 // from mutators. 6882 void Par_MarkFromRootsClosure::do_yield_work() { 6883 assert(_task != NULL, "sanity"); 6884 _task->yield(); 6885 } 6886 6887 // A variant of the above used for verifying CMS marking work. 6888 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6889 MemRegion span, 6890 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6891 CMSMarkStack* mark_stack): 6892 _collector(collector), 6893 _span(span), 6894 _verification_bm(verification_bm), 6895 _cms_bm(cms_bm), 6896 _mark_stack(mark_stack), 6897 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6898 mark_stack) 6899 { 6900 assert(_mark_stack->isEmpty(), "stack should be empty"); 6901 _finger = _verification_bm->startWord(); 6902 assert(_collector->_restart_addr == NULL, "Sanity check"); 6903 assert(_span.contains(_finger), "Out of bounds _finger?"); 6904 } 6905 6906 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6907 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6908 assert(_span.contains(addr), "Out of bounds _finger?"); 6909 _finger = addr; 6910 } 6911 6912 // Should revisit to see if this should be restructured for 6913 // greater efficiency. 6914 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6915 // convert offset into a HeapWord* 6916 HeapWord* addr = _verification_bm->startWord() + offset; 6917 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6918 "address out of range"); 6919 assert(_verification_bm->isMarked(addr), "tautology"); 6920 assert(_cms_bm->isMarked(addr), "tautology"); 6921 6922 assert(_mark_stack->isEmpty(), 6923 "should drain stack to limit stack usage"); 6924 // convert addr to an oop preparatory to scanning 6925 oop obj = oop(addr); 6926 assert(obj->is_oop(), "should be an oop"); 6927 assert(_finger <= addr, "_finger runneth ahead"); 6928 // advance the finger to right end of this object 6929 _finger = addr + obj->size(); 6930 assert(_finger > addr, "we just incremented it above"); 6931 // Note: the finger doesn't advance while we drain 6932 // the stack below. 6933 bool res = _mark_stack->push(obj); 6934 assert(res, "Empty non-zero size stack should have space for single push"); 6935 while (!_mark_stack->isEmpty()) { 6936 oop new_oop = _mark_stack->pop(); 6937 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6938 // now scan this oop's oops 6939 new_oop->oop_iterate(&_pam_verify_closure); 6940 } 6941 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6942 return true; 6943 } 6944 6945 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6946 CMSCollector* collector, MemRegion span, 6947 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6948 CMSMarkStack* mark_stack): 6949 MetadataAwareOopClosure(collector->ref_processor()), 6950 _collector(collector), 6951 _span(span), 6952 _verification_bm(verification_bm), 6953 _cms_bm(cms_bm), 6954 _mark_stack(mark_stack) 6955 { } 6956 6957 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6958 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6959 6960 // Upon stack overflow, we discard (part of) the stack, 6961 // remembering the least address amongst those discarded 6962 // in CMSCollector's _restart_address. 6963 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6964 // Remember the least grey address discarded 6965 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6966 _collector->lower_restart_addr(ra); 6967 _mark_stack->reset(); // discard stack contents 6968 _mark_stack->expand(); // expand the stack if possible 6969 } 6970 6971 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6972 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6973 HeapWord* addr = (HeapWord*)obj; 6974 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6975 // Oop lies in _span and isn't yet grey or black 6976 _verification_bm->mark(addr); // now grey 6977 if (!_cms_bm->isMarked(addr)) { 6978 oop(addr)->print(); 6979 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6980 p2i(addr)); 6981 fatal("... aborting"); 6982 } 6983 6984 if (!_mark_stack->push(obj)) { // stack overflow 6985 if (PrintCMSStatistics != 0) { 6986 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6987 SIZE_FORMAT, _mark_stack->capacity()); 6988 } 6989 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6990 handle_stack_overflow(addr); 6991 } 6992 // anything including and to the right of _finger 6993 // will be scanned as we iterate over the remainder of the 6994 // bit map 6995 } 6996 } 6997 6998 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6999 MemRegion span, 7000 CMSBitMap* bitMap, CMSMarkStack* markStack, 7001 HeapWord* finger, MarkFromRootsClosure* parent) : 7002 MetadataAwareOopClosure(collector->ref_processor()), 7003 _collector(collector), 7004 _span(span), 7005 _bitMap(bitMap), 7006 _markStack(markStack), 7007 _finger(finger), 7008 _parent(parent) 7009 { } 7010 7011 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7012 MemRegion span, 7013 CMSBitMap* bit_map, 7014 OopTaskQueue* work_queue, 7015 CMSMarkStack* overflow_stack, 7016 HeapWord* finger, 7017 HeapWord** global_finger_addr, 7018 Par_MarkFromRootsClosure* parent) : 7019 MetadataAwareOopClosure(collector->ref_processor()), 7020 _collector(collector), 7021 _whole_span(collector->_span), 7022 _span(span), 7023 _bit_map(bit_map), 7024 _work_queue(work_queue), 7025 _overflow_stack(overflow_stack), 7026 _finger(finger), 7027 _global_finger_addr(global_finger_addr), 7028 _parent(parent) 7029 { } 7030 7031 // Assumes thread-safe access by callers, who are 7032 // responsible for mutual exclusion. 7033 void CMSCollector::lower_restart_addr(HeapWord* low) { 7034 assert(_span.contains(low), "Out of bounds addr"); 7035 if (_restart_addr == NULL) { 7036 _restart_addr = low; 7037 } else { 7038 _restart_addr = MIN2(_restart_addr, low); 7039 } 7040 } 7041 7042 // Upon stack overflow, we discard (part of) the stack, 7043 // remembering the least address amongst those discarded 7044 // in CMSCollector's _restart_address. 7045 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7046 // Remember the least grey address discarded 7047 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7048 _collector->lower_restart_addr(ra); 7049 _markStack->reset(); // discard stack contents 7050 _markStack->expand(); // expand the stack if possible 7051 } 7052 7053 // Upon stack overflow, we discard (part of) the stack, 7054 // remembering the least address amongst those discarded 7055 // in CMSCollector's _restart_address. 7056 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7057 // We need to do this under a mutex to prevent other 7058 // workers from interfering with the work done below. 7059 MutexLockerEx ml(_overflow_stack->par_lock(), 7060 Mutex::_no_safepoint_check_flag); 7061 // Remember the least grey address discarded 7062 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7063 _collector->lower_restart_addr(ra); 7064 _overflow_stack->reset(); // discard stack contents 7065 _overflow_stack->expand(); // expand the stack if possible 7066 } 7067 7068 void PushOrMarkClosure::do_oop(oop obj) { 7069 // Ignore mark word because we are running concurrent with mutators. 7070 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7071 HeapWord* addr = (HeapWord*)obj; 7072 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7073 // Oop lies in _span and isn't yet grey or black 7074 _bitMap->mark(addr); // now grey 7075 if (addr < _finger) { 7076 // the bit map iteration has already either passed, or 7077 // sampled, this bit in the bit map; we'll need to 7078 // use the marking stack to scan this oop's oops. 7079 bool simulate_overflow = false; 7080 NOT_PRODUCT( 7081 if (CMSMarkStackOverflowALot && 7082 _collector->simulate_overflow()) { 7083 // simulate a stack overflow 7084 simulate_overflow = true; 7085 } 7086 ) 7087 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7088 if (PrintCMSStatistics != 0) { 7089 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7090 SIZE_FORMAT, _markStack->capacity()); 7091 } 7092 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7093 handle_stack_overflow(addr); 7094 } 7095 } 7096 // anything including and to the right of _finger 7097 // will be scanned as we iterate over the remainder of the 7098 // bit map 7099 do_yield_check(); 7100 } 7101 } 7102 7103 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7104 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7105 7106 void Par_PushOrMarkClosure::do_oop(oop obj) { 7107 // Ignore mark word because we are running concurrent with mutators. 7108 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7109 HeapWord* addr = (HeapWord*)obj; 7110 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7111 // Oop lies in _span and isn't yet grey or black 7112 // We read the global_finger (volatile read) strictly after marking oop 7113 bool res = _bit_map->par_mark(addr); // now grey 7114 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7115 // Should we push this marked oop on our stack? 7116 // -- if someone else marked it, nothing to do 7117 // -- if target oop is above global finger nothing to do 7118 // -- if target oop is in chunk and above local finger 7119 // then nothing to do 7120 // -- else push on work queue 7121 if ( !res // someone else marked it, they will deal with it 7122 || (addr >= *gfa) // will be scanned in a later task 7123 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7124 return; 7125 } 7126 // the bit map iteration has already either passed, or 7127 // sampled, this bit in the bit map; we'll need to 7128 // use the marking stack to scan this oop's oops. 7129 bool simulate_overflow = false; 7130 NOT_PRODUCT( 7131 if (CMSMarkStackOverflowALot && 7132 _collector->simulate_overflow()) { 7133 // simulate a stack overflow 7134 simulate_overflow = true; 7135 } 7136 ) 7137 if (simulate_overflow || 7138 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7139 // stack overflow 7140 if (PrintCMSStatistics != 0) { 7141 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7142 SIZE_FORMAT, _overflow_stack->capacity()); 7143 } 7144 // We cannot assert that the overflow stack is full because 7145 // it may have been emptied since. 7146 assert(simulate_overflow || 7147 _work_queue->size() == _work_queue->max_elems(), 7148 "Else push should have succeeded"); 7149 handle_stack_overflow(addr); 7150 } 7151 do_yield_check(); 7152 } 7153 } 7154 7155 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7156 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7157 7158 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7159 MemRegion span, 7160 ReferenceProcessor* rp, 7161 CMSBitMap* bit_map, 7162 CMSBitMap* mod_union_table, 7163 CMSMarkStack* mark_stack, 7164 bool concurrent_precleaning): 7165 MetadataAwareOopClosure(rp), 7166 _collector(collector), 7167 _span(span), 7168 _bit_map(bit_map), 7169 _mod_union_table(mod_union_table), 7170 _mark_stack(mark_stack), 7171 _concurrent_precleaning(concurrent_precleaning) 7172 { 7173 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7174 } 7175 7176 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7177 // the non-parallel version (the parallel version appears further below.) 7178 void PushAndMarkClosure::do_oop(oop obj) { 7179 // Ignore mark word verification. If during concurrent precleaning, 7180 // the object monitor may be locked. If during the checkpoint 7181 // phases, the object may already have been reached by a different 7182 // path and may be at the end of the global overflow list (so 7183 // the mark word may be NULL). 7184 assert(obj->is_oop_or_null(true /* ignore mark word */), 7185 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7186 HeapWord* addr = (HeapWord*)obj; 7187 // Check if oop points into the CMS generation 7188 // and is not marked 7189 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7190 // a white object ... 7191 _bit_map->mark(addr); // ... now grey 7192 // push on the marking stack (grey set) 7193 bool simulate_overflow = false; 7194 NOT_PRODUCT( 7195 if (CMSMarkStackOverflowALot && 7196 _collector->simulate_overflow()) { 7197 // simulate a stack overflow 7198 simulate_overflow = true; 7199 } 7200 ) 7201 if (simulate_overflow || !_mark_stack->push(obj)) { 7202 if (_concurrent_precleaning) { 7203 // During precleaning we can just dirty the appropriate card(s) 7204 // in the mod union table, thus ensuring that the object remains 7205 // in the grey set and continue. In the case of object arrays 7206 // we need to dirty all of the cards that the object spans, 7207 // since the rescan of object arrays will be limited to the 7208 // dirty cards. 7209 // Note that no one can be interfering with us in this action 7210 // of dirtying the mod union table, so no locking or atomics 7211 // are required. 7212 if (obj->is_objArray()) { 7213 size_t sz = obj->size(); 7214 HeapWord* end_card_addr = (HeapWord*)round_to( 7215 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7216 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7217 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7218 _mod_union_table->mark_range(redirty_range); 7219 } else { 7220 _mod_union_table->mark(addr); 7221 } 7222 _collector->_ser_pmc_preclean_ovflw++; 7223 } else { 7224 // During the remark phase, we need to remember this oop 7225 // in the overflow list. 7226 _collector->push_on_overflow_list(obj); 7227 _collector->_ser_pmc_remark_ovflw++; 7228 } 7229 } 7230 } 7231 } 7232 7233 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7234 MemRegion span, 7235 ReferenceProcessor* rp, 7236 CMSBitMap* bit_map, 7237 OopTaskQueue* work_queue): 7238 MetadataAwareOopClosure(rp), 7239 _collector(collector), 7240 _span(span), 7241 _bit_map(bit_map), 7242 _work_queue(work_queue) 7243 { 7244 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7245 } 7246 7247 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7248 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7249 7250 // Grey object rescan during second checkpoint phase -- 7251 // the parallel version. 7252 void Par_PushAndMarkClosure::do_oop(oop obj) { 7253 // In the assert below, we ignore the mark word because 7254 // this oop may point to an already visited object that is 7255 // on the overflow stack (in which case the mark word has 7256 // been hijacked for chaining into the overflow stack -- 7257 // if this is the last object in the overflow stack then 7258 // its mark word will be NULL). Because this object may 7259 // have been subsequently popped off the global overflow 7260 // stack, and the mark word possibly restored to the prototypical 7261 // value, by the time we get to examined this failing assert in 7262 // the debugger, is_oop_or_null(false) may subsequently start 7263 // to hold. 7264 assert(obj->is_oop_or_null(true), 7265 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7266 HeapWord* addr = (HeapWord*)obj; 7267 // Check if oop points into the CMS generation 7268 // and is not marked 7269 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7270 // a white object ... 7271 // If we manage to "claim" the object, by being the 7272 // first thread to mark it, then we push it on our 7273 // marking stack 7274 if (_bit_map->par_mark(addr)) { // ... now grey 7275 // push on work queue (grey set) 7276 bool simulate_overflow = false; 7277 NOT_PRODUCT( 7278 if (CMSMarkStackOverflowALot && 7279 _collector->par_simulate_overflow()) { 7280 // simulate a stack overflow 7281 simulate_overflow = true; 7282 } 7283 ) 7284 if (simulate_overflow || !_work_queue->push(obj)) { 7285 _collector->par_push_on_overflow_list(obj); 7286 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7287 } 7288 } // Else, some other thread got there first 7289 } 7290 } 7291 7292 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7293 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7294 7295 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7296 Mutex* bml = _collector->bitMapLock(); 7297 assert_lock_strong(bml); 7298 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7299 "CMS thread should hold CMS token"); 7300 7301 bml->unlock(); 7302 ConcurrentMarkSweepThread::desynchronize(true); 7303 7304 _collector->stopTimer(); 7305 if (PrintCMSStatistics != 0) { 7306 _collector->incrementYields(); 7307 } 7308 7309 // See the comment in coordinator_yield() 7310 for (unsigned i = 0; i < CMSYieldSleepCount && 7311 ConcurrentMarkSweepThread::should_yield() && 7312 !CMSCollector::foregroundGCIsActive(); ++i) { 7313 os::sleep(Thread::current(), 1, false); 7314 } 7315 7316 ConcurrentMarkSweepThread::synchronize(true); 7317 bml->lock(); 7318 7319 _collector->startTimer(); 7320 } 7321 7322 bool CMSPrecleanRefsYieldClosure::should_return() { 7323 if (ConcurrentMarkSweepThread::should_yield()) { 7324 do_yield_work(); 7325 } 7326 return _collector->foregroundGCIsActive(); 7327 } 7328 7329 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7330 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7331 "mr should be aligned to start at a card boundary"); 7332 // We'd like to assert: 7333 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7334 // "mr should be a range of cards"); 7335 // However, that would be too strong in one case -- the last 7336 // partition ends at _unallocated_block which, in general, can be 7337 // an arbitrary boundary, not necessarily card aligned. 7338 if (PrintCMSStatistics != 0) { 7339 _num_dirty_cards += 7340 mr.word_size()/CardTableModRefBS::card_size_in_words; 7341 } 7342 _space->object_iterate_mem(mr, &_scan_cl); 7343 } 7344 7345 SweepClosure::SweepClosure(CMSCollector* collector, 7346 ConcurrentMarkSweepGeneration* g, 7347 CMSBitMap* bitMap, bool should_yield) : 7348 _collector(collector), 7349 _g(g), 7350 _sp(g->cmsSpace()), 7351 _limit(_sp->sweep_limit()), 7352 _freelistLock(_sp->freelistLock()), 7353 _bitMap(bitMap), 7354 _yield(should_yield), 7355 _inFreeRange(false), // No free range at beginning of sweep 7356 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7357 _lastFreeRangeCoalesced(false), 7358 _freeFinger(g->used_region().start()) 7359 { 7360 NOT_PRODUCT( 7361 _numObjectsFreed = 0; 7362 _numWordsFreed = 0; 7363 _numObjectsLive = 0; 7364 _numWordsLive = 0; 7365 _numObjectsAlreadyFree = 0; 7366 _numWordsAlreadyFree = 0; 7367 _last_fc = NULL; 7368 7369 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7370 _sp->dictionary()->initialize_dict_returned_bytes(); 7371 ) 7372 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7373 "sweep _limit out of bounds"); 7374 if (CMSTraceSweeper) { 7375 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7376 p2i(_limit)); 7377 } 7378 } 7379 7380 void SweepClosure::print_on(outputStream* st) const { 7381 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7382 p2i(_sp->bottom()), p2i(_sp->end())); 7383 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7384 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7385 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7386 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7387 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7388 } 7389 7390 #ifndef PRODUCT 7391 // Assertion checking only: no useful work in product mode -- 7392 // however, if any of the flags below become product flags, 7393 // you may need to review this code to see if it needs to be 7394 // enabled in product mode. 7395 SweepClosure::~SweepClosure() { 7396 assert_lock_strong(_freelistLock); 7397 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7398 "sweep _limit out of bounds"); 7399 if (inFreeRange()) { 7400 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7401 print(); 7402 ShouldNotReachHere(); 7403 } 7404 if (Verbose && PrintGC) { 7405 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7406 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7407 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7408 SIZE_FORMAT" bytes " 7409 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7410 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7411 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7412 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7413 * sizeof(HeapWord); 7414 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7415 7416 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7417 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7418 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7419 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7420 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7421 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7422 indexListReturnedBytes); 7423 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7424 dict_returned_bytes); 7425 } 7426 } 7427 if (CMSTraceSweeper) { 7428 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7429 p2i(_limit)); 7430 } 7431 } 7432 #endif // PRODUCT 7433 7434 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7435 bool freeRangeInFreeLists) { 7436 if (CMSTraceSweeper) { 7437 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7438 p2i(freeFinger), freeRangeInFreeLists); 7439 } 7440 assert(!inFreeRange(), "Trampling existing free range"); 7441 set_inFreeRange(true); 7442 set_lastFreeRangeCoalesced(false); 7443 7444 set_freeFinger(freeFinger); 7445 set_freeRangeInFreeLists(freeRangeInFreeLists); 7446 if (CMSTestInFreeList) { 7447 if (freeRangeInFreeLists) { 7448 FreeChunk* fc = (FreeChunk*) freeFinger; 7449 assert(fc->is_free(), "A chunk on the free list should be free."); 7450 assert(fc->size() > 0, "Free range should have a size"); 7451 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7452 } 7453 } 7454 } 7455 7456 // Note that the sweeper runs concurrently with mutators. Thus, 7457 // it is possible for direct allocation in this generation to happen 7458 // in the middle of the sweep. Note that the sweeper also coalesces 7459 // contiguous free blocks. Thus, unless the sweeper and the allocator 7460 // synchronize appropriately freshly allocated blocks may get swept up. 7461 // This is accomplished by the sweeper locking the free lists while 7462 // it is sweeping. Thus blocks that are determined to be free are 7463 // indeed free. There is however one additional complication: 7464 // blocks that have been allocated since the final checkpoint and 7465 // mark, will not have been marked and so would be treated as 7466 // unreachable and swept up. To prevent this, the allocator marks 7467 // the bit map when allocating during the sweep phase. This leads, 7468 // however, to a further complication -- objects may have been allocated 7469 // but not yet initialized -- in the sense that the header isn't yet 7470 // installed. The sweeper can not then determine the size of the block 7471 // in order to skip over it. To deal with this case, we use a technique 7472 // (due to Printezis) to encode such uninitialized block sizes in the 7473 // bit map. Since the bit map uses a bit per every HeapWord, but the 7474 // CMS generation has a minimum object size of 3 HeapWords, it follows 7475 // that "normal marks" won't be adjacent in the bit map (there will 7476 // always be at least two 0 bits between successive 1 bits). We make use 7477 // of these "unused" bits to represent uninitialized blocks -- the bit 7478 // corresponding to the start of the uninitialized object and the next 7479 // bit are both set. Finally, a 1 bit marks the end of the object that 7480 // started with the two consecutive 1 bits to indicate its potentially 7481 // uninitialized state. 7482 7483 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7484 FreeChunk* fc = (FreeChunk*)addr; 7485 size_t res; 7486 7487 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7488 // than "addr == _limit" because although _limit was a block boundary when 7489 // we started the sweep, it may no longer be one because heap expansion 7490 // may have caused us to coalesce the block ending at the address _limit 7491 // with a newly expanded chunk (this happens when _limit was set to the 7492 // previous _end of the space), so we may have stepped past _limit: 7493 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7494 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7495 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7496 "sweep _limit out of bounds"); 7497 assert(addr < _sp->end(), "addr out of bounds"); 7498 // Flush any free range we might be holding as a single 7499 // coalesced chunk to the appropriate free list. 7500 if (inFreeRange()) { 7501 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7502 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7503 flush_cur_free_chunk(freeFinger(), 7504 pointer_delta(addr, freeFinger())); 7505 if (CMSTraceSweeper) { 7506 gclog_or_tty->print("Sweep: last chunk: "); 7507 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7508 "[coalesced:%d]\n", 7509 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7510 lastFreeRangeCoalesced() ? 1 : 0); 7511 } 7512 } 7513 7514 // help the iterator loop finish 7515 return pointer_delta(_sp->end(), addr); 7516 } 7517 7518 assert(addr < _limit, "sweep invariant"); 7519 // check if we should yield 7520 do_yield_check(addr); 7521 if (fc->is_free()) { 7522 // Chunk that is already free 7523 res = fc->size(); 7524 do_already_free_chunk(fc); 7525 debug_only(_sp->verifyFreeLists()); 7526 // If we flush the chunk at hand in lookahead_and_flush() 7527 // and it's coalesced with a preceding chunk, then the 7528 // process of "mangling" the payload of the coalesced block 7529 // will cause erasure of the size information from the 7530 // (erstwhile) header of all the coalesced blocks but the 7531 // first, so the first disjunct in the assert will not hold 7532 // in that specific case (in which case the second disjunct 7533 // will hold). 7534 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7535 "Otherwise the size info doesn't change at this step"); 7536 NOT_PRODUCT( 7537 _numObjectsAlreadyFree++; 7538 _numWordsAlreadyFree += res; 7539 ) 7540 NOT_PRODUCT(_last_fc = fc;) 7541 } else if (!_bitMap->isMarked(addr)) { 7542 // Chunk is fresh garbage 7543 res = do_garbage_chunk(fc); 7544 debug_only(_sp->verifyFreeLists()); 7545 NOT_PRODUCT( 7546 _numObjectsFreed++; 7547 _numWordsFreed += res; 7548 ) 7549 } else { 7550 // Chunk that is alive. 7551 res = do_live_chunk(fc); 7552 debug_only(_sp->verifyFreeLists()); 7553 NOT_PRODUCT( 7554 _numObjectsLive++; 7555 _numWordsLive += res; 7556 ) 7557 } 7558 return res; 7559 } 7560 7561 // For the smart allocation, record following 7562 // split deaths - a free chunk is removed from its free list because 7563 // it is being split into two or more chunks. 7564 // split birth - a free chunk is being added to its free list because 7565 // a larger free chunk has been split and resulted in this free chunk. 7566 // coal death - a free chunk is being removed from its free list because 7567 // it is being coalesced into a large free chunk. 7568 // coal birth - a free chunk is being added to its free list because 7569 // it was created when two or more free chunks where coalesced into 7570 // this free chunk. 7571 // 7572 // These statistics are used to determine the desired number of free 7573 // chunks of a given size. The desired number is chosen to be relative 7574 // to the end of a CMS sweep. The desired number at the end of a sweep 7575 // is the 7576 // count-at-end-of-previous-sweep (an amount that was enough) 7577 // - count-at-beginning-of-current-sweep (the excess) 7578 // + split-births (gains in this size during interval) 7579 // - split-deaths (demands on this size during interval) 7580 // where the interval is from the end of one sweep to the end of the 7581 // next. 7582 // 7583 // When sweeping the sweeper maintains an accumulated chunk which is 7584 // the chunk that is made up of chunks that have been coalesced. That 7585 // will be termed the left-hand chunk. A new chunk of garbage that 7586 // is being considered for coalescing will be referred to as the 7587 // right-hand chunk. 7588 // 7589 // When making a decision on whether to coalesce a right-hand chunk with 7590 // the current left-hand chunk, the current count vs. the desired count 7591 // of the left-hand chunk is considered. Also if the right-hand chunk 7592 // is near the large chunk at the end of the heap (see 7593 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7594 // left-hand chunk is coalesced. 7595 // 7596 // When making a decision about whether to split a chunk, the desired count 7597 // vs. the current count of the candidate to be split is also considered. 7598 // If the candidate is underpopulated (currently fewer chunks than desired) 7599 // a chunk of an overpopulated (currently more chunks than desired) size may 7600 // be chosen. The "hint" associated with a free list, if non-null, points 7601 // to a free list which may be overpopulated. 7602 // 7603 7604 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7605 const size_t size = fc->size(); 7606 // Chunks that cannot be coalesced are not in the 7607 // free lists. 7608 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7609 assert(_sp->verify_chunk_in_free_list(fc), 7610 "free chunk should be in free lists"); 7611 } 7612 // a chunk that is already free, should not have been 7613 // marked in the bit map 7614 HeapWord* const addr = (HeapWord*) fc; 7615 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7616 // Verify that the bit map has no bits marked between 7617 // addr and purported end of this block. 7618 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7619 7620 // Some chunks cannot be coalesced under any circumstances. 7621 // See the definition of cantCoalesce(). 7622 if (!fc->cantCoalesce()) { 7623 // This chunk can potentially be coalesced. 7624 if (_sp->adaptive_freelists()) { 7625 // All the work is done in 7626 do_post_free_or_garbage_chunk(fc, size); 7627 } else { // Not adaptive free lists 7628 // this is a free chunk that can potentially be coalesced by the sweeper; 7629 if (!inFreeRange()) { 7630 // if the next chunk is a free block that can't be coalesced 7631 // it doesn't make sense to remove this chunk from the free lists 7632 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7633 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7634 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7635 nextChunk->is_free() && // ... which is free... 7636 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7637 // nothing to do 7638 } else { 7639 // Potentially the start of a new free range: 7640 // Don't eagerly remove it from the free lists. 7641 // No need to remove it if it will just be put 7642 // back again. (Also from a pragmatic point of view 7643 // if it is a free block in a region that is beyond 7644 // any allocated blocks, an assertion will fail) 7645 // Remember the start of a free run. 7646 initialize_free_range(addr, true); 7647 // end - can coalesce with next chunk 7648 } 7649 } else { 7650 // the midst of a free range, we are coalescing 7651 print_free_block_coalesced(fc); 7652 if (CMSTraceSweeper) { 7653 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7654 } 7655 // remove it from the free lists 7656 _sp->removeFreeChunkFromFreeLists(fc); 7657 set_lastFreeRangeCoalesced(true); 7658 // If the chunk is being coalesced and the current free range is 7659 // in the free lists, remove the current free range so that it 7660 // will be returned to the free lists in its entirety - all 7661 // the coalesced pieces included. 7662 if (freeRangeInFreeLists()) { 7663 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7664 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7665 "Size of free range is inconsistent with chunk size."); 7666 if (CMSTestInFreeList) { 7667 assert(_sp->verify_chunk_in_free_list(ffc), 7668 "free range is not in free lists"); 7669 } 7670 _sp->removeFreeChunkFromFreeLists(ffc); 7671 set_freeRangeInFreeLists(false); 7672 } 7673 } 7674 } 7675 // Note that if the chunk is not coalescable (the else arm 7676 // below), we unconditionally flush, without needing to do 7677 // a "lookahead," as we do below. 7678 if (inFreeRange()) lookahead_and_flush(fc, size); 7679 } else { 7680 // Code path common to both original and adaptive free lists. 7681 7682 // cant coalesce with previous block; this should be treated 7683 // as the end of a free run if any 7684 if (inFreeRange()) { 7685 // we kicked some butt; time to pick up the garbage 7686 assert(freeFinger() < addr, "freeFinger points too high"); 7687 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7688 } 7689 // else, nothing to do, just continue 7690 } 7691 } 7692 7693 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7694 // This is a chunk of garbage. It is not in any free list. 7695 // Add it to a free list or let it possibly be coalesced into 7696 // a larger chunk. 7697 HeapWord* const addr = (HeapWord*) fc; 7698 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7699 7700 if (_sp->adaptive_freelists()) { 7701 // Verify that the bit map has no bits marked between 7702 // addr and purported end of just dead object. 7703 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7704 7705 do_post_free_or_garbage_chunk(fc, size); 7706 } else { 7707 if (!inFreeRange()) { 7708 // start of a new free range 7709 assert(size > 0, "A free range should have a size"); 7710 initialize_free_range(addr, false); 7711 } else { 7712 // this will be swept up when we hit the end of the 7713 // free range 7714 if (CMSTraceSweeper) { 7715 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7716 } 7717 // If the chunk is being coalesced and the current free range is 7718 // in the free lists, remove the current free range so that it 7719 // will be returned to the free lists in its entirety - all 7720 // the coalesced pieces included. 7721 if (freeRangeInFreeLists()) { 7722 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7723 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7724 "Size of free range is inconsistent with chunk size."); 7725 if (CMSTestInFreeList) { 7726 assert(_sp->verify_chunk_in_free_list(ffc), 7727 "free range is not in free lists"); 7728 } 7729 _sp->removeFreeChunkFromFreeLists(ffc); 7730 set_freeRangeInFreeLists(false); 7731 } 7732 set_lastFreeRangeCoalesced(true); 7733 } 7734 // this will be swept up when we hit the end of the free range 7735 7736 // Verify that the bit map has no bits marked between 7737 // addr and purported end of just dead object. 7738 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7739 } 7740 assert(_limit >= addr + size, 7741 "A freshly garbage chunk can't possibly straddle over _limit"); 7742 if (inFreeRange()) lookahead_and_flush(fc, size); 7743 return size; 7744 } 7745 7746 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7747 HeapWord* addr = (HeapWord*) fc; 7748 // The sweeper has just found a live object. Return any accumulated 7749 // left hand chunk to the free lists. 7750 if (inFreeRange()) { 7751 assert(freeFinger() < addr, "freeFinger points too high"); 7752 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7753 } 7754 7755 // This object is live: we'd normally expect this to be 7756 // an oop, and like to assert the following: 7757 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7758 // However, as we commented above, this may be an object whose 7759 // header hasn't yet been initialized. 7760 size_t size; 7761 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7762 if (_bitMap->isMarked(addr + 1)) { 7763 // Determine the size from the bit map, rather than trying to 7764 // compute it from the object header. 7765 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7766 size = pointer_delta(nextOneAddr + 1, addr); 7767 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7768 "alignment problem"); 7769 7770 #ifdef ASSERT 7771 if (oop(addr)->klass_or_null() != NULL) { 7772 // Ignore mark word because we are running concurrent with mutators 7773 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7774 assert(size == 7775 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7776 "P-mark and computed size do not agree"); 7777 } 7778 #endif 7779 7780 } else { 7781 // This should be an initialized object that's alive. 7782 assert(oop(addr)->klass_or_null() != NULL, 7783 "Should be an initialized object"); 7784 // Ignore mark word because we are running concurrent with mutators 7785 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7786 // Verify that the bit map has no bits marked between 7787 // addr and purported end of this block. 7788 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7789 assert(size >= 3, "Necessary for Printezis marks to work"); 7790 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7791 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7792 } 7793 return size; 7794 } 7795 7796 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7797 size_t chunkSize) { 7798 // do_post_free_or_garbage_chunk() should only be called in the case 7799 // of the adaptive free list allocator. 7800 const bool fcInFreeLists = fc->is_free(); 7801 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7802 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7803 if (CMSTestInFreeList && fcInFreeLists) { 7804 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7805 } 7806 7807 if (CMSTraceSweeper) { 7808 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7809 } 7810 7811 HeapWord* const fc_addr = (HeapWord*) fc; 7812 7813 bool coalesce; 7814 const size_t left = pointer_delta(fc_addr, freeFinger()); 7815 const size_t right = chunkSize; 7816 switch (FLSCoalescePolicy) { 7817 // numeric value forms a coalition aggressiveness metric 7818 case 0: { // never coalesce 7819 coalesce = false; 7820 break; 7821 } 7822 case 1: { // coalesce if left & right chunks on overpopulated lists 7823 coalesce = _sp->coalOverPopulated(left) && 7824 _sp->coalOverPopulated(right); 7825 break; 7826 } 7827 case 2: { // coalesce if left chunk on overpopulated list (default) 7828 coalesce = _sp->coalOverPopulated(left); 7829 break; 7830 } 7831 case 3: { // coalesce if left OR right chunk on overpopulated list 7832 coalesce = _sp->coalOverPopulated(left) || 7833 _sp->coalOverPopulated(right); 7834 break; 7835 } 7836 case 4: { // always coalesce 7837 coalesce = true; 7838 break; 7839 } 7840 default: 7841 ShouldNotReachHere(); 7842 } 7843 7844 // Should the current free range be coalesced? 7845 // If the chunk is in a free range and either we decided to coalesce above 7846 // or the chunk is near the large block at the end of the heap 7847 // (isNearLargestChunk() returns true), then coalesce this chunk. 7848 const bool doCoalesce = inFreeRange() 7849 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7850 if (doCoalesce) { 7851 // Coalesce the current free range on the left with the new 7852 // chunk on the right. If either is on a free list, 7853 // it must be removed from the list and stashed in the closure. 7854 if (freeRangeInFreeLists()) { 7855 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7856 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7857 "Size of free range is inconsistent with chunk size."); 7858 if (CMSTestInFreeList) { 7859 assert(_sp->verify_chunk_in_free_list(ffc), 7860 "Chunk is not in free lists"); 7861 } 7862 _sp->coalDeath(ffc->size()); 7863 _sp->removeFreeChunkFromFreeLists(ffc); 7864 set_freeRangeInFreeLists(false); 7865 } 7866 if (fcInFreeLists) { 7867 _sp->coalDeath(chunkSize); 7868 assert(fc->size() == chunkSize, 7869 "The chunk has the wrong size or is not in the free lists"); 7870 _sp->removeFreeChunkFromFreeLists(fc); 7871 } 7872 set_lastFreeRangeCoalesced(true); 7873 print_free_block_coalesced(fc); 7874 } else { // not in a free range and/or should not coalesce 7875 // Return the current free range and start a new one. 7876 if (inFreeRange()) { 7877 // In a free range but cannot coalesce with the right hand chunk. 7878 // Put the current free range into the free lists. 7879 flush_cur_free_chunk(freeFinger(), 7880 pointer_delta(fc_addr, freeFinger())); 7881 } 7882 // Set up for new free range. Pass along whether the right hand 7883 // chunk is in the free lists. 7884 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7885 } 7886 } 7887 7888 // Lookahead flush: 7889 // If we are tracking a free range, and this is the last chunk that 7890 // we'll look at because its end crosses past _limit, we'll preemptively 7891 // flush it along with any free range we may be holding on to. Note that 7892 // this can be the case only for an already free or freshly garbage 7893 // chunk. If this block is an object, it can never straddle 7894 // over _limit. The "straddling" occurs when _limit is set at 7895 // the previous end of the space when this cycle started, and 7896 // a subsequent heap expansion caused the previously co-terminal 7897 // free block to be coalesced with the newly expanded portion, 7898 // thus rendering _limit a non-block-boundary making it dangerous 7899 // for the sweeper to step over and examine. 7900 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7901 assert(inFreeRange(), "Should only be called if currently in a free range."); 7902 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7903 assert(_sp->used_region().contains(eob - 1), 7904 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7905 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7906 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7907 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7908 if (eob >= _limit) { 7909 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7910 if (CMSTraceSweeper) { 7911 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7912 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7913 "[" PTR_FORMAT "," PTR_FORMAT ")", 7914 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7915 } 7916 // Return the storage we are tracking back into the free lists. 7917 if (CMSTraceSweeper) { 7918 gclog_or_tty->print_cr("Flushing ... "); 7919 } 7920 assert(freeFinger() < eob, "Error"); 7921 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7922 } 7923 } 7924 7925 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7926 assert(inFreeRange(), "Should only be called if currently in a free range."); 7927 assert(size > 0, 7928 "A zero sized chunk cannot be added to the free lists."); 7929 if (!freeRangeInFreeLists()) { 7930 if (CMSTestInFreeList) { 7931 FreeChunk* fc = (FreeChunk*) chunk; 7932 fc->set_size(size); 7933 assert(!_sp->verify_chunk_in_free_list(fc), 7934 "chunk should not be in free lists yet"); 7935 } 7936 if (CMSTraceSweeper) { 7937 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7938 p2i(chunk), size); 7939 } 7940 // A new free range is going to be starting. The current 7941 // free range has not been added to the free lists yet or 7942 // was removed so add it back. 7943 // If the current free range was coalesced, then the death 7944 // of the free range was recorded. Record a birth now. 7945 if (lastFreeRangeCoalesced()) { 7946 _sp->coalBirth(size); 7947 } 7948 _sp->addChunkAndRepairOffsetTable(chunk, size, 7949 lastFreeRangeCoalesced()); 7950 } else if (CMSTraceSweeper) { 7951 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7952 } 7953 set_inFreeRange(false); 7954 set_freeRangeInFreeLists(false); 7955 } 7956 7957 // We take a break if we've been at this for a while, 7958 // so as to avoid monopolizing the locks involved. 7959 void SweepClosure::do_yield_work(HeapWord* addr) { 7960 // Return current free chunk being used for coalescing (if any) 7961 // to the appropriate freelist. After yielding, the next 7962 // free block encountered will start a coalescing range of 7963 // free blocks. If the next free block is adjacent to the 7964 // chunk just flushed, they will need to wait for the next 7965 // sweep to be coalesced. 7966 if (inFreeRange()) { 7967 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7968 } 7969 7970 // First give up the locks, then yield, then re-lock. 7971 // We should probably use a constructor/destructor idiom to 7972 // do this unlock/lock or modify the MutexUnlocker class to 7973 // serve our purpose. XXX 7974 assert_lock_strong(_bitMap->lock()); 7975 assert_lock_strong(_freelistLock); 7976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7977 "CMS thread should hold CMS token"); 7978 _bitMap->lock()->unlock(); 7979 _freelistLock->unlock(); 7980 ConcurrentMarkSweepThread::desynchronize(true); 7981 _collector->stopTimer(); 7982 if (PrintCMSStatistics != 0) { 7983 _collector->incrementYields(); 7984 } 7985 7986 // See the comment in coordinator_yield() 7987 for (unsigned i = 0; i < CMSYieldSleepCount && 7988 ConcurrentMarkSweepThread::should_yield() && 7989 !CMSCollector::foregroundGCIsActive(); ++i) { 7990 os::sleep(Thread::current(), 1, false); 7991 } 7992 7993 ConcurrentMarkSweepThread::synchronize(true); 7994 _freelistLock->lock(); 7995 _bitMap->lock()->lock_without_safepoint_check(); 7996 _collector->startTimer(); 7997 } 7998 7999 #ifndef PRODUCT 8000 // This is actually very useful in a product build if it can 8001 // be called from the debugger. Compile it into the product 8002 // as needed. 8003 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8004 return debug_cms_space->verify_chunk_in_free_list(fc); 8005 } 8006 #endif 8007 8008 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8009 if (CMSTraceSweeper) { 8010 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8011 p2i(fc), fc->size()); 8012 } 8013 } 8014 8015 // CMSIsAliveClosure 8016 bool CMSIsAliveClosure::do_object_b(oop obj) { 8017 HeapWord* addr = (HeapWord*)obj; 8018 return addr != NULL && 8019 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8020 } 8021 8022 8023 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8024 MemRegion span, 8025 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8026 bool cpc): 8027 _collector(collector), 8028 _span(span), 8029 _bit_map(bit_map), 8030 _mark_stack(mark_stack), 8031 _concurrent_precleaning(cpc) { 8032 assert(!_span.is_empty(), "Empty span could spell trouble"); 8033 } 8034 8035 8036 // CMSKeepAliveClosure: the serial version 8037 void CMSKeepAliveClosure::do_oop(oop obj) { 8038 HeapWord* addr = (HeapWord*)obj; 8039 if (_span.contains(addr) && 8040 !_bit_map->isMarked(addr)) { 8041 _bit_map->mark(addr); 8042 bool simulate_overflow = false; 8043 NOT_PRODUCT( 8044 if (CMSMarkStackOverflowALot && 8045 _collector->simulate_overflow()) { 8046 // simulate a stack overflow 8047 simulate_overflow = true; 8048 } 8049 ) 8050 if (simulate_overflow || !_mark_stack->push(obj)) { 8051 if (_concurrent_precleaning) { 8052 // We dirty the overflown object and let the remark 8053 // phase deal with it. 8054 assert(_collector->overflow_list_is_empty(), "Error"); 8055 // In the case of object arrays, we need to dirty all of 8056 // the cards that the object spans. No locking or atomics 8057 // are needed since no one else can be mutating the mod union 8058 // table. 8059 if (obj->is_objArray()) { 8060 size_t sz = obj->size(); 8061 HeapWord* end_card_addr = 8062 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8063 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8064 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8065 _collector->_modUnionTable.mark_range(redirty_range); 8066 } else { 8067 _collector->_modUnionTable.mark(addr); 8068 } 8069 _collector->_ser_kac_preclean_ovflw++; 8070 } else { 8071 _collector->push_on_overflow_list(obj); 8072 _collector->_ser_kac_ovflw++; 8073 } 8074 } 8075 } 8076 } 8077 8078 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8079 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8080 8081 // CMSParKeepAliveClosure: a parallel version of the above. 8082 // The work queues are private to each closure (thread), 8083 // but (may be) available for stealing by other threads. 8084 void CMSParKeepAliveClosure::do_oop(oop obj) { 8085 HeapWord* addr = (HeapWord*)obj; 8086 if (_span.contains(addr) && 8087 !_bit_map->isMarked(addr)) { 8088 // In general, during recursive tracing, several threads 8089 // may be concurrently getting here; the first one to 8090 // "tag" it, claims it. 8091 if (_bit_map->par_mark(addr)) { 8092 bool res = _work_queue->push(obj); 8093 assert(res, "Low water mark should be much less than capacity"); 8094 // Do a recursive trim in the hope that this will keep 8095 // stack usage lower, but leave some oops for potential stealers 8096 trim_queue(_low_water_mark); 8097 } // Else, another thread got there first 8098 } 8099 } 8100 8101 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8102 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8103 8104 void CMSParKeepAliveClosure::trim_queue(uint max) { 8105 while (_work_queue->size() > max) { 8106 oop new_oop; 8107 if (_work_queue->pop_local(new_oop)) { 8108 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8109 assert(_bit_map->isMarked((HeapWord*)new_oop), 8110 "no white objects on this stack!"); 8111 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8112 // iterate over the oops in this oop, marking and pushing 8113 // the ones in CMS heap (i.e. in _span). 8114 new_oop->oop_iterate(&_mark_and_push); 8115 } 8116 } 8117 } 8118 8119 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8120 CMSCollector* collector, 8121 MemRegion span, CMSBitMap* bit_map, 8122 OopTaskQueue* work_queue): 8123 _collector(collector), 8124 _span(span), 8125 _bit_map(bit_map), 8126 _work_queue(work_queue) { } 8127 8128 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8129 HeapWord* addr = (HeapWord*)obj; 8130 if (_span.contains(addr) && 8131 !_bit_map->isMarked(addr)) { 8132 if (_bit_map->par_mark(addr)) { 8133 bool simulate_overflow = false; 8134 NOT_PRODUCT( 8135 if (CMSMarkStackOverflowALot && 8136 _collector->par_simulate_overflow()) { 8137 // simulate a stack overflow 8138 simulate_overflow = true; 8139 } 8140 ) 8141 if (simulate_overflow || !_work_queue->push(obj)) { 8142 _collector->par_push_on_overflow_list(obj); 8143 _collector->_par_kac_ovflw++; 8144 } 8145 } // Else another thread got there already 8146 } 8147 } 8148 8149 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8150 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8151 8152 ////////////////////////////////////////////////////////////////// 8153 // CMSExpansionCause ///////////////////////////// 8154 ////////////////////////////////////////////////////////////////// 8155 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8156 switch (cause) { 8157 case _no_expansion: 8158 return "No expansion"; 8159 case _satisfy_free_ratio: 8160 return "Free ratio"; 8161 case _satisfy_promotion: 8162 return "Satisfy promotion"; 8163 case _satisfy_allocation: 8164 return "allocation"; 8165 case _allocate_par_lab: 8166 return "Par LAB"; 8167 case _allocate_par_spooling_space: 8168 return "Par Spooling Space"; 8169 case _adaptive_size_policy: 8170 return "Ergonomics"; 8171 default: 8172 return "unknown"; 8173 } 8174 } 8175 8176 void CMSDrainMarkingStackClosure::do_void() { 8177 // the max number to take from overflow list at a time 8178 const size_t num = _mark_stack->capacity()/4; 8179 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8180 "Overflow list should be NULL during concurrent phases"); 8181 while (!_mark_stack->isEmpty() || 8182 // if stack is empty, check the overflow list 8183 _collector->take_from_overflow_list(num, _mark_stack)) { 8184 oop obj = _mark_stack->pop(); 8185 HeapWord* addr = (HeapWord*)obj; 8186 assert(_span.contains(addr), "Should be within span"); 8187 assert(_bit_map->isMarked(addr), "Should be marked"); 8188 assert(obj->is_oop(), "Should be an oop"); 8189 obj->oop_iterate(_keep_alive); 8190 } 8191 } 8192 8193 void CMSParDrainMarkingStackClosure::do_void() { 8194 // drain queue 8195 trim_queue(0); 8196 } 8197 8198 // Trim our work_queue so its length is below max at return 8199 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8200 while (_work_queue->size() > max) { 8201 oop new_oop; 8202 if (_work_queue->pop_local(new_oop)) { 8203 assert(new_oop->is_oop(), "Expected an oop"); 8204 assert(_bit_map->isMarked((HeapWord*)new_oop), 8205 "no white objects on this stack!"); 8206 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8207 // iterate over the oops in this oop, marking and pushing 8208 // the ones in CMS heap (i.e. in _span). 8209 new_oop->oop_iterate(&_mark_and_push); 8210 } 8211 } 8212 } 8213 8214 //////////////////////////////////////////////////////////////////// 8215 // Support for Marking Stack Overflow list handling and related code 8216 //////////////////////////////////////////////////////////////////// 8217 // Much of the following code is similar in shape and spirit to the 8218 // code used in ParNewGC. We should try and share that code 8219 // as much as possible in the future. 8220 8221 #ifndef PRODUCT 8222 // Debugging support for CMSStackOverflowALot 8223 8224 // It's OK to call this multi-threaded; the worst thing 8225 // that can happen is that we'll get a bunch of closely 8226 // spaced simulated overflows, but that's OK, in fact 8227 // probably good as it would exercise the overflow code 8228 // under contention. 8229 bool CMSCollector::simulate_overflow() { 8230 if (_overflow_counter-- <= 0) { // just being defensive 8231 _overflow_counter = CMSMarkStackOverflowInterval; 8232 return true; 8233 } else { 8234 return false; 8235 } 8236 } 8237 8238 bool CMSCollector::par_simulate_overflow() { 8239 return simulate_overflow(); 8240 } 8241 #endif 8242 8243 // Single-threaded 8244 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8245 assert(stack->isEmpty(), "Expected precondition"); 8246 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8247 size_t i = num; 8248 oop cur = _overflow_list; 8249 const markOop proto = markOopDesc::prototype(); 8250 NOT_PRODUCT(ssize_t n = 0;) 8251 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8252 next = oop(cur->mark()); 8253 cur->set_mark(proto); // until proven otherwise 8254 assert(cur->is_oop(), "Should be an oop"); 8255 bool res = stack->push(cur); 8256 assert(res, "Bit off more than can chew?"); 8257 NOT_PRODUCT(n++;) 8258 } 8259 _overflow_list = cur; 8260 #ifndef PRODUCT 8261 assert(_num_par_pushes >= n, "Too many pops?"); 8262 _num_par_pushes -=n; 8263 #endif 8264 return !stack->isEmpty(); 8265 } 8266 8267 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8268 // (MT-safe) Get a prefix of at most "num" from the list. 8269 // The overflow list is chained through the mark word of 8270 // each object in the list. We fetch the entire list, 8271 // break off a prefix of the right size and return the 8272 // remainder. If other threads try to take objects from 8273 // the overflow list at that time, they will wait for 8274 // some time to see if data becomes available. If (and 8275 // only if) another thread places one or more object(s) 8276 // on the global list before we have returned the suffix 8277 // to the global list, we will walk down our local list 8278 // to find its end and append the global list to 8279 // our suffix before returning it. This suffix walk can 8280 // prove to be expensive (quadratic in the amount of traffic) 8281 // when there are many objects in the overflow list and 8282 // there is much producer-consumer contention on the list. 8283 // *NOTE*: The overflow list manipulation code here and 8284 // in ParNewGeneration:: are very similar in shape, 8285 // except that in the ParNew case we use the old (from/eden) 8286 // copy of the object to thread the list via its klass word. 8287 // Because of the common code, if you make any changes in 8288 // the code below, please check the ParNew version to see if 8289 // similar changes might be needed. 8290 // CR 6797058 has been filed to consolidate the common code. 8291 bool CMSCollector::par_take_from_overflow_list(size_t num, 8292 OopTaskQueue* work_q, 8293 int no_of_gc_threads) { 8294 assert(work_q->size() == 0, "First empty local work queue"); 8295 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8296 if (_overflow_list == NULL) { 8297 return false; 8298 } 8299 // Grab the entire list; we'll put back a suffix 8300 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8301 Thread* tid = Thread::current(); 8302 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8303 // set to ParallelGCThreads. 8304 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8305 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8306 // If the list is busy, we spin for a short while, 8307 // sleeping between attempts to get the list. 8308 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8309 os::sleep(tid, sleep_time_millis, false); 8310 if (_overflow_list == NULL) { 8311 // Nothing left to take 8312 return false; 8313 } else if (_overflow_list != BUSY) { 8314 // Try and grab the prefix 8315 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8316 } 8317 } 8318 // If the list was found to be empty, or we spun long 8319 // enough, we give up and return empty-handed. If we leave 8320 // the list in the BUSY state below, it must be the case that 8321 // some other thread holds the overflow list and will set it 8322 // to a non-BUSY state in the future. 8323 if (prefix == NULL || prefix == BUSY) { 8324 // Nothing to take or waited long enough 8325 if (prefix == NULL) { 8326 // Write back the NULL in case we overwrote it with BUSY above 8327 // and it is still the same value. 8328 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8329 } 8330 return false; 8331 } 8332 assert(prefix != NULL && prefix != BUSY, "Error"); 8333 size_t i = num; 8334 oop cur = prefix; 8335 // Walk down the first "num" objects, unless we reach the end. 8336 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8337 if (cur->mark() == NULL) { 8338 // We have "num" or fewer elements in the list, so there 8339 // is nothing to return to the global list. 8340 // Write back the NULL in lieu of the BUSY we wrote 8341 // above, if it is still the same value. 8342 if (_overflow_list == BUSY) { 8343 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8344 } 8345 } else { 8346 // Chop off the suffix and return it to the global list. 8347 assert(cur->mark() != BUSY, "Error"); 8348 oop suffix_head = cur->mark(); // suffix will be put back on global list 8349 cur->set_mark(NULL); // break off suffix 8350 // It's possible that the list is still in the empty(busy) state 8351 // we left it in a short while ago; in that case we may be 8352 // able to place back the suffix without incurring the cost 8353 // of a walk down the list. 8354 oop observed_overflow_list = _overflow_list; 8355 oop cur_overflow_list = observed_overflow_list; 8356 bool attached = false; 8357 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8358 observed_overflow_list = 8359 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8360 if (cur_overflow_list == observed_overflow_list) { 8361 attached = true; 8362 break; 8363 } else cur_overflow_list = observed_overflow_list; 8364 } 8365 if (!attached) { 8366 // Too bad, someone else sneaked in (at least) an element; we'll need 8367 // to do a splice. Find tail of suffix so we can prepend suffix to global 8368 // list. 8369 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8370 oop suffix_tail = cur; 8371 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8372 "Tautology"); 8373 observed_overflow_list = _overflow_list; 8374 do { 8375 cur_overflow_list = observed_overflow_list; 8376 if (cur_overflow_list != BUSY) { 8377 // Do the splice ... 8378 suffix_tail->set_mark(markOop(cur_overflow_list)); 8379 } else { // cur_overflow_list == BUSY 8380 suffix_tail->set_mark(NULL); 8381 } 8382 // ... and try to place spliced list back on overflow_list ... 8383 observed_overflow_list = 8384 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8385 } while (cur_overflow_list != observed_overflow_list); 8386 // ... until we have succeeded in doing so. 8387 } 8388 } 8389 8390 // Push the prefix elements on work_q 8391 assert(prefix != NULL, "control point invariant"); 8392 const markOop proto = markOopDesc::prototype(); 8393 oop next; 8394 NOT_PRODUCT(ssize_t n = 0;) 8395 for (cur = prefix; cur != NULL; cur = next) { 8396 next = oop(cur->mark()); 8397 cur->set_mark(proto); // until proven otherwise 8398 assert(cur->is_oop(), "Should be an oop"); 8399 bool res = work_q->push(cur); 8400 assert(res, "Bit off more than we can chew?"); 8401 NOT_PRODUCT(n++;) 8402 } 8403 #ifndef PRODUCT 8404 assert(_num_par_pushes >= n, "Too many pops?"); 8405 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8406 #endif 8407 return true; 8408 } 8409 8410 // Single-threaded 8411 void CMSCollector::push_on_overflow_list(oop p) { 8412 NOT_PRODUCT(_num_par_pushes++;) 8413 assert(p->is_oop(), "Not an oop"); 8414 preserve_mark_if_necessary(p); 8415 p->set_mark((markOop)_overflow_list); 8416 _overflow_list = p; 8417 } 8418 8419 // Multi-threaded; use CAS to prepend to overflow list 8420 void CMSCollector::par_push_on_overflow_list(oop p) { 8421 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8422 assert(p->is_oop(), "Not an oop"); 8423 par_preserve_mark_if_necessary(p); 8424 oop observed_overflow_list = _overflow_list; 8425 oop cur_overflow_list; 8426 do { 8427 cur_overflow_list = observed_overflow_list; 8428 if (cur_overflow_list != BUSY) { 8429 p->set_mark(markOop(cur_overflow_list)); 8430 } else { 8431 p->set_mark(NULL); 8432 } 8433 observed_overflow_list = 8434 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8435 } while (cur_overflow_list != observed_overflow_list); 8436 } 8437 #undef BUSY 8438 8439 // Single threaded 8440 // General Note on GrowableArray: pushes may silently fail 8441 // because we are (temporarily) out of C-heap for expanding 8442 // the stack. The problem is quite ubiquitous and affects 8443 // a lot of code in the JVM. The prudent thing for GrowableArray 8444 // to do (for now) is to exit with an error. However, that may 8445 // be too draconian in some cases because the caller may be 8446 // able to recover without much harm. For such cases, we 8447 // should probably introduce a "soft_push" method which returns 8448 // an indication of success or failure with the assumption that 8449 // the caller may be able to recover from a failure; code in 8450 // the VM can then be changed, incrementally, to deal with such 8451 // failures where possible, thus, incrementally hardening the VM 8452 // in such low resource situations. 8453 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8454 _preserved_oop_stack.push(p); 8455 _preserved_mark_stack.push(m); 8456 assert(m == p->mark(), "Mark word changed"); 8457 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8458 "bijection"); 8459 } 8460 8461 // Single threaded 8462 void CMSCollector::preserve_mark_if_necessary(oop p) { 8463 markOop m = p->mark(); 8464 if (m->must_be_preserved(p)) { 8465 preserve_mark_work(p, m); 8466 } 8467 } 8468 8469 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8470 markOop m = p->mark(); 8471 if (m->must_be_preserved(p)) { 8472 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8473 // Even though we read the mark word without holding 8474 // the lock, we are assured that it will not change 8475 // because we "own" this oop, so no other thread can 8476 // be trying to push it on the overflow list; see 8477 // the assertion in preserve_mark_work() that checks 8478 // that m == p->mark(). 8479 preserve_mark_work(p, m); 8480 } 8481 } 8482 8483 // We should be able to do this multi-threaded, 8484 // a chunk of stack being a task (this is 8485 // correct because each oop only ever appears 8486 // once in the overflow list. However, it's 8487 // not very easy to completely overlap this with 8488 // other operations, so will generally not be done 8489 // until all work's been completed. Because we 8490 // expect the preserved oop stack (set) to be small, 8491 // it's probably fine to do this single-threaded. 8492 // We can explore cleverer concurrent/overlapped/parallel 8493 // processing of preserved marks if we feel the 8494 // need for this in the future. Stack overflow should 8495 // be so rare in practice and, when it happens, its 8496 // effect on performance so great that this will 8497 // likely just be in the noise anyway. 8498 void CMSCollector::restore_preserved_marks_if_any() { 8499 assert(SafepointSynchronize::is_at_safepoint(), 8500 "world should be stopped"); 8501 assert(Thread::current()->is_ConcurrentGC_thread() || 8502 Thread::current()->is_VM_thread(), 8503 "should be single-threaded"); 8504 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8505 "bijection"); 8506 8507 while (!_preserved_oop_stack.is_empty()) { 8508 oop p = _preserved_oop_stack.pop(); 8509 assert(p->is_oop(), "Should be an oop"); 8510 assert(_span.contains(p), "oop should be in _span"); 8511 assert(p->mark() == markOopDesc::prototype(), 8512 "Set when taken from overflow list"); 8513 markOop m = _preserved_mark_stack.pop(); 8514 p->set_mark(m); 8515 } 8516 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8517 "stacks were cleared above"); 8518 } 8519 8520 #ifndef PRODUCT 8521 bool CMSCollector::no_preserved_marks() const { 8522 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8523 } 8524 #endif 8525 8526 // Transfer some number of overflown objects to usual marking 8527 // stack. Return true if some objects were transferred. 8528 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8529 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8530 (size_t)ParGCDesiredObjsFromOverflowList); 8531 8532 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8533 assert(_collector->overflow_list_is_empty() || res, 8534 "If list is not empty, we should have taken something"); 8535 assert(!res || !_mark_stack->isEmpty(), 8536 "If we took something, it should now be on our stack"); 8537 return res; 8538 } 8539 8540 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8541 size_t res = _sp->block_size_no_stall(addr, _collector); 8542 if (_sp->block_is_obj(addr)) { 8543 if (_live_bit_map->isMarked(addr)) { 8544 // It can't have been dead in a previous cycle 8545 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8546 } else { 8547 _dead_bit_map->mark(addr); // mark the dead object 8548 } 8549 } 8550 // Could be 0, if the block size could not be computed without stalling. 8551 return res; 8552 } 8553 8554 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8555 8556 switch (phase) { 8557 case CMSCollector::InitialMarking: 8558 initialize(true /* fullGC */ , 8559 cause /* cause of the GC */, 8560 true /* recordGCBeginTime */, 8561 true /* recordPreGCUsage */, 8562 false /* recordPeakUsage */, 8563 false /* recordPostGCusage */, 8564 true /* recordAccumulatedGCTime */, 8565 false /* recordGCEndTime */, 8566 false /* countCollection */ ); 8567 break; 8568 8569 case CMSCollector::FinalMarking: 8570 initialize(true /* fullGC */ , 8571 cause /* cause of the GC */, 8572 false /* recordGCBeginTime */, 8573 false /* recordPreGCUsage */, 8574 false /* recordPeakUsage */, 8575 false /* recordPostGCusage */, 8576 true /* recordAccumulatedGCTime */, 8577 false /* recordGCEndTime */, 8578 false /* countCollection */ ); 8579 break; 8580 8581 case CMSCollector::Sweeping: 8582 initialize(true /* fullGC */ , 8583 cause /* cause of the GC */, 8584 false /* recordGCBeginTime */, 8585 false /* recordPreGCUsage */, 8586 true /* recordPeakUsage */, 8587 true /* recordPostGCusage */, 8588 false /* recordAccumulatedGCTime */, 8589 true /* recordGCEndTime */, 8590 true /* countCollection */ ); 8591 break; 8592 8593 default: 8594 ShouldNotReachHere(); 8595 } 8596 }