1 /* 2 * Copyright (c) 2001, 2014, 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/cardTableRS.hpp" 46 #include "memory/collectorPolicy.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/genCollectedHeap.hpp" 49 #include "memory/genMarkSweep.hpp" 50 #include "memory/genOopClosures.inline.hpp" 51 #include "memory/iterator.inline.hpp" 52 #include "memory/padded.hpp" 53 #include "memory/referencePolicy.hpp" 54 #include "memory/resourceArea.hpp" 55 #include "memory/tenuredGeneration.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "prims/jvmtiExport.hpp" 58 #include "runtime/atomic.inline.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/orderAccess.inline.hpp" 63 #include "runtime/vmThread.hpp" 64 #include "services/memoryService.hpp" 65 #include "services/runtimeService.hpp" 66 67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 68 69 // statics 70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 71 bool CMSCollector::_full_gc_requested = false; 72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 73 74 ////////////////////////////////////////////////////////////////// 75 // In support of CMS/VM thread synchronization 76 ////////////////////////////////////////////////////////////////// 77 // We split use of the CGC_lock into 2 "levels". 78 // The low-level locking is of the usual CGC_lock monitor. We introduce 79 // a higher level "token" (hereafter "CMS token") built on top of the 80 // low level monitor (hereafter "CGC lock"). 81 // The token-passing protocol gives priority to the VM thread. The 82 // CMS-lock doesn't provide any fairness guarantees, but clients 83 // should ensure that it is only held for very short, bounded 84 // durations. 85 // 86 // When either of the CMS thread or the VM thread is involved in 87 // collection operations during which it does not want the other 88 // thread to interfere, it obtains the CMS token. 89 // 90 // If either thread tries to get the token while the other has 91 // it, that thread waits. However, if the VM thread and CMS thread 92 // both want the token, then the VM thread gets priority while the 93 // CMS thread waits. This ensures, for instance, that the "concurrent" 94 // phases of the CMS thread's work do not block out the VM thread 95 // for long periods of time as the CMS thread continues to hog 96 // the token. (See bug 4616232). 97 // 98 // The baton-passing functions are, however, controlled by the 99 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 100 // and here the low-level CMS lock, not the high level token, 101 // ensures mutual exclusion. 102 // 103 // Two important conditions that we have to satisfy: 104 // 1. if a thread does a low-level wait on the CMS lock, then it 105 // relinquishes the CMS token if it were holding that token 106 // when it acquired the low-level CMS lock. 107 // 2. any low-level notifications on the low-level lock 108 // should only be sent when a thread has relinquished the token. 109 // 110 // In the absence of either property, we'd have potential deadlock. 111 // 112 // We protect each of the CMS (concurrent and sequential) phases 113 // with the CMS _token_, not the CMS _lock_. 114 // 115 // The only code protected by CMS lock is the token acquisition code 116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 117 // baton-passing code. 118 // 119 // Unfortunately, i couldn't come up with a good abstraction to factor and 120 // hide the naked CGC_lock manipulation in the baton-passing code 121 // further below. That's something we should try to do. Also, the proof 122 // of correctness of this 2-level locking scheme is far from obvious, 123 // and potentially quite slippery. We have an uneasy suspicion, for instance, 124 // that there may be a theoretical possibility of delay/starvation in the 125 // low-level lock/wait/notify scheme used for the baton-passing because of 126 // potential interference with the priority scheme embodied in the 127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 128 // invocation further below and marked with "XXX 20011219YSR". 129 // Indeed, as we note elsewhere, this may become yet more slippery 130 // in the presence of multiple CMS and/or multiple VM threads. XXX 131 132 class CMSTokenSync: public StackObj { 133 private: 134 bool _is_cms_thread; 135 public: 136 CMSTokenSync(bool is_cms_thread): 137 _is_cms_thread(is_cms_thread) { 138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 139 "Incorrect argument to constructor"); 140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 141 } 142 143 ~CMSTokenSync() { 144 assert(_is_cms_thread ? 145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 147 "Incorrect state"); 148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 149 } 150 }; 151 152 // Convenience class that does a CMSTokenSync, and then acquires 153 // upto three locks. 154 class CMSTokenSyncWithLocks: public CMSTokenSync { 155 private: 156 // Note: locks are acquired in textual declaration order 157 // and released in the opposite order 158 MutexLockerEx _locker1, _locker2, _locker3; 159 public: 160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 162 CMSTokenSync(is_cms_thread), 163 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 164 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 165 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 166 { } 167 }; 168 169 170 ////////////////////////////////////////////////////////////////// 171 // Concurrent Mark-Sweep Generation ///////////////////////////// 172 ////////////////////////////////////////////////////////////////// 173 174 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 175 176 // This struct contains per-thread things necessary to support parallel 177 // young-gen collection. 178 class CMSParGCThreadState: public CHeapObj<mtGC> { 179 public: 180 CFLS_LAB lab; 181 PromotionInfo promo; 182 183 // Constructor. 184 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 185 promo.setSpace(cfls); 186 } 187 }; 188 189 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 190 ReservedSpace rs, size_t initial_byte_size, int level, 191 CardTableRS* ct, bool use_adaptive_freelists, 192 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 193 CardGeneration(rs, initial_byte_size, level, ct), 194 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 195 _debug_collection_type(Concurrent_collection_type), 196 _did_compact(false) 197 { 198 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 199 HeapWord* end = (HeapWord*) _virtual_space.high(); 200 201 _direct_allocated_words = 0; 202 NOT_PRODUCT( 203 _numObjectsPromoted = 0; 204 _numWordsPromoted = 0; 205 _numObjectsAllocated = 0; 206 _numWordsAllocated = 0; 207 ) 208 209 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 210 use_adaptive_freelists, 211 dictionaryChoice); 212 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 213 if (_cmsSpace == NULL) { 214 vm_exit_during_initialization( 215 "CompactibleFreeListSpace allocation failure"); 216 } 217 _cmsSpace->_gen = this; 218 219 _gc_stats = new CMSGCStats(); 220 221 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 222 // offsets match. The ability to tell free chunks from objects 223 // depends on this property. 224 debug_only( 225 FreeChunk* junk = NULL; 226 assert(UseCompressedClassPointers || 227 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 228 "Offset of FreeChunk::_prev within FreeChunk must match" 229 " that of OopDesc::_klass within OopDesc"); 230 ) 231 if (CollectedHeap::use_parallel_gc_threads()) { 232 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 233 _par_gc_thread_states = 234 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 235 if (_par_gc_thread_states == NULL) { 236 vm_exit_during_initialization("Could not allocate par gc structs"); 237 } 238 for (uint i = 0; i < ParallelGCThreads; i++) { 239 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 240 if (_par_gc_thread_states[i] == NULL) { 241 vm_exit_during_initialization("Could not allocate par gc structs"); 242 } 243 } 244 } else { 245 _par_gc_thread_states = NULL; 246 } 247 _incremental_collection_failed = false; 248 // The "dilatation_factor" is the expansion that can occur on 249 // account of the fact that the minimum object size in the CMS 250 // generation may be larger than that in, say, a contiguous young 251 // generation. 252 // Ideally, in the calculation below, we'd compute the dilatation 253 // factor as: MinChunkSize/(promoting_gen's min object size) 254 // Since we do not have such a general query interface for the 255 // promoting generation, we'll instead just use the minimum 256 // object size (which today is a header's worth of space); 257 // note that all arithmetic is in units of HeapWords. 258 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 259 assert(_dilatation_factor >= 1.0, "from previous assert"); 260 } 261 262 263 // The field "_initiating_occupancy" represents the occupancy percentage 264 // at which we trigger a new collection cycle. Unless explicitly specified 265 // via CMSInitiatingOccupancyFraction (argument "io" below), it 266 // is calculated by: 267 // 268 // Let "f" be MinHeapFreeRatio in 269 // 270 // _initiating_occupancy = 100-f + 271 // f * (CMSTriggerRatio/100) 272 // where CMSTriggerRatio is the argument "tr" below. 273 // 274 // That is, if we assume the heap is at its desired maximum occupancy at the 275 // end of a collection, we let CMSTriggerRatio of the (purported) free 276 // space be allocated before initiating a new collection cycle. 277 // 278 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 279 assert(io <= 100 && tr <= 100, "Check the arguments"); 280 if (io >= 0) { 281 _initiating_occupancy = (double)io / 100.0; 282 } else { 283 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 284 (double)(tr * MinHeapFreeRatio) / 100.0) 285 / 100.0; 286 } 287 } 288 289 void ConcurrentMarkSweepGeneration::ref_processor_init() { 290 assert(collector() != NULL, "no collector"); 291 collector()->ref_processor_init(); 292 } 293 294 void CMSCollector::ref_processor_init() { 295 if (_ref_processor == NULL) { 296 // Allocate and initialize a reference processor 297 _ref_processor = 298 new ReferenceProcessor(_span, // span 299 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 300 (int) ParallelGCThreads, // mt processing degree 301 _cmsGen->refs_discovery_is_mt(), // mt discovery 302 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 303 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 304 &_is_alive_closure); // closure for liveness info 305 // Initialize the _ref_processor field of CMSGen 306 _cmsGen->set_ref_processor(_ref_processor); 307 308 } 309 } 310 311 AdaptiveSizePolicy* CMSCollector::size_policy() { 312 GenCollectedHeap* gch = GenCollectedHeap::heap(); 313 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 314 "Wrong type of heap"); 315 return gch->gen_policy()->size_policy(); 316 } 317 318 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 319 320 const char* gen_name = "old"; 321 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 322 323 // Generation Counters - generation 1, 1 subspace 324 _gen_counters = new GenerationCounters(gen_name, 1, 1, 325 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 326 327 _space_counters = new GSpaceCounters(gen_name, 0, 328 _virtual_space.reserved_size(), 329 this, _gen_counters); 330 } 331 332 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 333 _cms_gen(cms_gen) 334 { 335 assert(alpha <= 100, "bad value"); 336 _saved_alpha = alpha; 337 338 // Initialize the alphas to the bootstrap value of 100. 339 _gc0_alpha = _cms_alpha = 100; 340 341 _cms_begin_time.update(); 342 _cms_end_time.update(); 343 344 _gc0_duration = 0.0; 345 _gc0_period = 0.0; 346 _gc0_promoted = 0; 347 348 _cms_duration = 0.0; 349 _cms_period = 0.0; 350 _cms_allocated = 0; 351 352 _cms_used_at_gc0_begin = 0; 353 _cms_used_at_gc0_end = 0; 354 _allow_duty_cycle_reduction = false; 355 _valid_bits = 0; 356 } 357 358 double CMSStats::cms_free_adjustment_factor(size_t free) const { 359 // TBD: CR 6909490 360 return 1.0; 361 } 362 363 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 364 } 365 366 // If promotion failure handling is on use 367 // the padded average size of the promotion for each 368 // young generation collection. 369 double CMSStats::time_until_cms_gen_full() const { 370 size_t cms_free = _cms_gen->cmsSpace()->free(); 371 GenCollectedHeap* gch = GenCollectedHeap::heap(); 372 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 373 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 374 if (cms_free > expected_promotion) { 375 // Start a cms collection if there isn't enough space to promote 376 // for the next minor collection. Use the padded average as 377 // a safety factor. 378 cms_free -= expected_promotion; 379 380 // Adjust by the safety factor. 381 double cms_free_dbl = (double)cms_free; 382 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 383 // Apply a further correction factor which tries to adjust 384 // for recent occurance of concurrent mode failures. 385 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 386 cms_free_dbl = cms_free_dbl * cms_adjustment; 387 388 if (PrintGCDetails && Verbose) { 389 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 390 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 391 cms_free, expected_promotion); 392 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 393 cms_free_dbl, cms_consumption_rate() + 1.0); 394 } 395 // Add 1 in case the consumption rate goes to zero. 396 return cms_free_dbl / (cms_consumption_rate() + 1.0); 397 } 398 return 0.0; 399 } 400 401 // Compare the duration of the cms collection to the 402 // time remaining before the cms generation is empty. 403 // Note that the time from the start of the cms collection 404 // to the start of the cms sweep (less than the total 405 // duration of the cms collection) can be used. This 406 // has been tried and some applications experienced 407 // promotion failures early in execution. This was 408 // possibly because the averages were not accurate 409 // enough at the beginning. 410 double CMSStats::time_until_cms_start() const { 411 // We add "gc0_period" to the "work" calculation 412 // below because this query is done (mostly) at the 413 // end of a scavenge, so we need to conservatively 414 // account for that much possible delay 415 // in the query so as to avoid concurrent mode failures 416 // due to starting the collection just a wee bit too 417 // late. 418 double work = cms_duration() + gc0_period(); 419 double deadline = time_until_cms_gen_full(); 420 // If a concurrent mode failure occurred recently, we want to be 421 // more conservative and halve our expected time_until_cms_gen_full() 422 if (work > deadline) { 423 if (Verbose && PrintGCDetails) { 424 gclog_or_tty->print( 425 " CMSCollector: collect because of anticipated promotion " 426 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 427 gc0_period(), time_until_cms_gen_full()); 428 } 429 return 0.0; 430 } 431 return work - deadline; 432 } 433 434 #ifndef PRODUCT 435 void CMSStats::print_on(outputStream *st) const { 436 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 437 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 438 gc0_duration(), gc0_period(), gc0_promoted()); 439 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 440 cms_duration(), cms_period(), cms_allocated()); 441 st->print(",cms_since_beg=%g,cms_since_end=%g", 442 cms_time_since_begin(), cms_time_since_end()); 443 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 444 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 445 446 if (valid()) { 447 st->print(",promo_rate=%g,cms_alloc_rate=%g", 448 promotion_rate(), cms_allocation_rate()); 449 st->print(",cms_consumption_rate=%g,time_until_full=%g", 450 cms_consumption_rate(), time_until_cms_gen_full()); 451 } 452 st->print(" "); 453 } 454 #endif // #ifndef PRODUCT 455 456 CMSCollector::CollectorState CMSCollector::_collectorState = 457 CMSCollector::Idling; 458 bool CMSCollector::_foregroundGCIsActive = false; 459 bool CMSCollector::_foregroundGCShouldWait = false; 460 461 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 462 CardTableRS* ct, 463 ConcurrentMarkSweepPolicy* cp): 464 _cmsGen(cmsGen), 465 _ct(ct), 466 _ref_processor(NULL), // will be set later 467 _conc_workers(NULL), // may be set later 468 _abort_preclean(false), 469 _start_sampling(false), 470 _between_prologue_and_epilogue(false), 471 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 472 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 473 -1 /* lock-free */, "No_lock" /* dummy */), 474 _modUnionClosure(&_modUnionTable), 475 _modUnionClosurePar(&_modUnionTable), 476 // Adjust my span to cover old (cms) gen 477 _span(cmsGen->reserved()), 478 // Construct the is_alive_closure with _span & markBitMap 479 _is_alive_closure(_span, &_markBitMap), 480 _restart_addr(NULL), 481 _overflow_list(NULL), 482 _stats(cmsGen), 483 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 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(SharedHeap::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("Parallel CMS Threads", 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 _full_gcs_since_conc_gc = 0; 616 617 // Now tell CMS generations the identity of their collector 618 ConcurrentMarkSweepGeneration::set_collector(this); 619 620 // Create & start a CMS thread for this CMS collector 621 _cmsThread = ConcurrentMarkSweepThread::start(this); 622 assert(cmsThread() != NULL, "CMS Thread should have been created"); 623 assert(cmsThread()->collector() == this, 624 "CMS Thread should refer to this gen"); 625 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 626 627 // Support for parallelizing young gen rescan 628 GenCollectedHeap* gch = GenCollectedHeap::heap(); 629 _young_gen = gch->prev_gen(_cmsGen); 630 if (gch->supports_inline_contig_alloc()) { 631 _top_addr = gch->top_addr(); 632 _end_addr = gch->end_addr(); 633 assert(_young_gen != NULL, "no _young_gen"); 634 _eden_chunk_index = 0; 635 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 636 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 637 if (_eden_chunk_array == NULL) { 638 _eden_chunk_capacity = 0; 639 warning("GC/CMS: _eden_chunk_array allocation failure"); 640 } 641 } 642 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 643 644 // Support for parallelizing survivor space rescan 645 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 646 const size_t max_plab_samples = 647 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 648 649 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 650 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 651 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 652 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 653 || _cursor == NULL) { 654 warning("Failed to allocate survivor plab/chunk array"); 655 if (_survivor_plab_array != NULL) { 656 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 657 _survivor_plab_array = NULL; 658 } 659 if (_survivor_chunk_array != NULL) { 660 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 661 _survivor_chunk_array = NULL; 662 } 663 if (_cursor != NULL) { 664 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 665 _cursor = NULL; 666 } 667 } else { 668 _survivor_chunk_capacity = 2*max_plab_samples; 669 for (uint i = 0; i < ParallelGCThreads; i++) { 670 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 671 if (vec == NULL) { 672 warning("Failed to allocate survivor plab array"); 673 for (int j = i; j > 0; j--) { 674 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 675 } 676 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 677 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 678 _survivor_plab_array = NULL; 679 _survivor_chunk_array = NULL; 680 _survivor_chunk_capacity = 0; 681 break; 682 } else { 683 ChunkArray* cur = 684 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 685 max_plab_samples); 686 assert(cur->end() == 0, "Should be 0"); 687 assert(cur->array() == vec, "Should be vec"); 688 assert(cur->capacity() == max_plab_samples, "Error"); 689 } 690 } 691 } 692 } 693 assert( ( _survivor_plab_array != NULL 694 && _survivor_chunk_array != NULL) 695 || ( _survivor_chunk_capacity == 0 696 && _survivor_chunk_index == 0), 697 "Error"); 698 699 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 700 _gc_counters = new CollectorCounters("CMS", 1); 701 _completed_initialization = true; 702 _inter_sweep_timer.start(); // start of time 703 } 704 705 const char* ConcurrentMarkSweepGeneration::name() const { 706 return "concurrent mark-sweep generation"; 707 } 708 void ConcurrentMarkSweepGeneration::update_counters() { 709 if (UsePerfData) { 710 _space_counters->update_all(); 711 _gen_counters->update_all(); 712 } 713 } 714 715 // this is an optimized version of update_counters(). it takes the 716 // used value as a parameter rather than computing it. 717 // 718 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 719 if (UsePerfData) { 720 _space_counters->update_used(used); 721 _space_counters->update_capacity(); 722 _gen_counters->update_all(); 723 } 724 } 725 726 void ConcurrentMarkSweepGeneration::print() const { 727 Generation::print(); 728 cmsSpace()->print(); 729 } 730 731 #ifndef PRODUCT 732 void ConcurrentMarkSweepGeneration::print_statistics() { 733 cmsSpace()->printFLCensus(0); 734 } 735 #endif 736 737 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 738 GenCollectedHeap* gch = GenCollectedHeap::heap(); 739 if (PrintGCDetails) { 740 if (Verbose) { 741 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 742 level(), short_name(), s, used(), capacity()); 743 } else { 744 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 745 level(), short_name(), s, used() / K, capacity() / K); 746 } 747 } 748 if (Verbose) { 749 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 750 gch->used(), gch->capacity()); 751 } else { 752 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 753 gch->used() / K, gch->capacity() / K); 754 } 755 } 756 757 size_t 758 ConcurrentMarkSweepGeneration::contiguous_available() const { 759 // dld proposes an improvement in precision here. If the committed 760 // part of the space ends in a free block we should add that to 761 // uncommitted size in the calculation below. Will make this 762 // change later, staying with the approximation below for the 763 // time being. -- ysr. 764 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 765 } 766 767 size_t 768 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 769 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 770 } 771 772 size_t ConcurrentMarkSweepGeneration::max_available() const { 773 return free() + _virtual_space.uncommitted_size(); 774 } 775 776 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 777 size_t available = max_available(); 778 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 779 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 780 if (Verbose && PrintGCDetails) { 781 gclog_or_tty->print_cr( 782 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 783 "max_promo("SIZE_FORMAT")", 784 res? "":" not", available, res? ">=":"<", 785 av_promo, max_promotion_in_bytes); 786 } 787 return res; 788 } 789 790 // At a promotion failure dump information on block layout in heap 791 // (cms old generation). 792 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 793 if (CMSDumpAtPromotionFailure) { 794 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 795 } 796 } 797 798 CompactibleSpace* 799 ConcurrentMarkSweepGeneration::first_compaction_space() const { 800 return _cmsSpace; 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 if (CollectedHeap::use_parallel_gc_threads()) { 809 for (uint i = 0; i < ParallelGCThreads; i++) { 810 _par_gc_thread_states[i]->promo.reset(); 811 } 812 } 813 } 814 815 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 816 blk->do_space(_cmsSpace); 817 } 818 819 void ConcurrentMarkSweepGeneration::compute_new_size() { 820 assert_locked_or_safepoint(Heap_lock); 821 822 // If incremental collection failed, we just want to expand 823 // to the limit. 824 if (incremental_collection_failed()) { 825 clear_incremental_collection_failed(); 826 grow_to_reserved(); 827 return; 828 } 829 830 // The heap has been compacted but not reset yet. 831 // Any metric such as free() or used() will be incorrect. 832 833 CardGeneration::compute_new_size(); 834 835 // Reset again after a possible resizing 836 if (did_compact()) { 837 cmsSpace()->reset_after_compaction(); 838 } 839 } 840 841 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 842 assert_locked_or_safepoint(Heap_lock); 843 844 // If incremental collection failed, we just want to expand 845 // to the limit. 846 if (incremental_collection_failed()) { 847 clear_incremental_collection_failed(); 848 grow_to_reserved(); 849 return; 850 } 851 852 double free_percentage = ((double) free()) / capacity(); 853 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 854 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 855 856 // compute expansion delta needed for reaching desired free percentage 857 if (free_percentage < desired_free_percentage) { 858 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 859 assert(desired_capacity >= capacity(), "invalid expansion size"); 860 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 861 if (PrintGCDetails && Verbose) { 862 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 863 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 864 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 865 gclog_or_tty->print_cr(" Desired free fraction %f", 866 desired_free_percentage); 867 gclog_or_tty->print_cr(" Maximum free fraction %f", 868 maximum_free_percentage); 869 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 870 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 871 desired_capacity/1000); 872 int prev_level = level() - 1; 873 if (prev_level >= 0) { 874 size_t prev_size = 0; 875 GenCollectedHeap* gch = GenCollectedHeap::heap(); 876 Generation* prev_gen = gch->_gens[prev_level]; 877 prev_size = prev_gen->capacity(); 878 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 879 prev_size/1000); 880 } 881 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 882 unsafe_max_alloc_nogc()/1000); 883 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 884 contiguous_available()/1000); 885 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 886 expand_bytes); 887 } 888 // safe if expansion fails 889 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 890 if (PrintGCDetails && Verbose) { 891 gclog_or_tty->print_cr(" Expanded free fraction %f", 892 ((double) free()) / capacity()); 893 } 894 } else { 895 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 896 assert(desired_capacity <= capacity(), "invalid expansion size"); 897 size_t shrink_bytes = capacity() - desired_capacity; 898 // Don't shrink unless the delta is greater than the minimum shrink we want 899 if (shrink_bytes >= MinHeapDeltaBytes) { 900 shrink_free_list_by(shrink_bytes); 901 } 902 } 903 } 904 905 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 906 return cmsSpace()->freelistLock(); 907 } 908 909 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 910 bool tlab) { 911 CMSSynchronousYieldRequest yr; 912 MutexLockerEx x(freelistLock(), 913 Mutex::_no_safepoint_check_flag); 914 return have_lock_and_allocate(size, tlab); 915 } 916 917 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 918 bool tlab /* ignored */) { 919 assert_lock_strong(freelistLock()); 920 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 921 HeapWord* res = cmsSpace()->allocate(adjustedSize); 922 // Allocate the object live (grey) if the background collector has 923 // started marking. This is necessary because the marker may 924 // have passed this address and consequently this object will 925 // not otherwise be greyed and would be incorrectly swept up. 926 // Note that if this object contains references, the writing 927 // of those references will dirty the card containing this object 928 // allowing the object to be blackened (and its references scanned) 929 // either during a preclean phase or at the final checkpoint. 930 if (res != NULL) { 931 // We may block here with an uninitialized object with 932 // its mark-bit or P-bits not yet set. Such objects need 933 // to be safely navigable by block_start(). 934 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 935 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 936 collector()->direct_allocated(res, adjustedSize); 937 _direct_allocated_words += adjustedSize; 938 // allocation counters 939 NOT_PRODUCT( 940 _numObjectsAllocated++; 941 _numWordsAllocated += (int)adjustedSize; 942 ) 943 } 944 return res; 945 } 946 947 // In the case of direct allocation by mutators in a generation that 948 // is being concurrently collected, the object must be allocated 949 // live (grey) if the background collector has started marking. 950 // This is necessary because the marker may 951 // have passed this address and consequently this object will 952 // not otherwise be greyed and would be incorrectly swept up. 953 // Note that if this object contains references, the writing 954 // of those references will dirty the card containing this object 955 // allowing the object to be blackened (and its references scanned) 956 // either during a preclean phase or at the final checkpoint. 957 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 958 assert(_markBitMap.covers(start, size), "Out of bounds"); 959 if (_collectorState >= Marking) { 960 MutexLockerEx y(_markBitMap.lock(), 961 Mutex::_no_safepoint_check_flag); 962 // [see comments preceding SweepClosure::do_blk() below for details] 963 // 964 // Can the P-bits be deleted now? JJJ 965 // 966 // 1. need to mark the object as live so it isn't collected 967 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 968 // 3. need to mark the end of the object so marking, precleaning or sweeping 969 // can skip over uninitialized or unparsable objects. An allocated 970 // object is considered uninitialized for our purposes as long as 971 // its klass word is NULL. All old gen objects are parsable 972 // as soon as they are initialized.) 973 _markBitMap.mark(start); // object is live 974 _markBitMap.mark(start + 1); // object is potentially uninitialized? 975 _markBitMap.mark(start + size - 1); 976 // mark end of object 977 } 978 // check that oop looks uninitialized 979 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 980 } 981 982 void CMSCollector::promoted(bool par, HeapWord* start, 983 bool is_obj_array, size_t obj_size) { 984 assert(_markBitMap.covers(start), "Out of bounds"); 985 // See comment in direct_allocated() about when objects should 986 // be allocated live. 987 if (_collectorState >= Marking) { 988 // we already hold the marking bit map lock, taken in 989 // the prologue 990 if (par) { 991 _markBitMap.par_mark(start); 992 } else { 993 _markBitMap.mark(start); 994 } 995 // We don't need to mark the object as uninitialized (as 996 // in direct_allocated above) because this is being done with the 997 // world stopped and the object will be initialized by the 998 // time the marking, precleaning or sweeping get to look at it. 999 // But see the code for copying objects into the CMS generation, 1000 // where we need to ensure that concurrent readers of the 1001 // block offset table are able to safely navigate a block that 1002 // is in flux from being free to being allocated (and in 1003 // transition while being copied into) and subsequently 1004 // becoming a bona-fide object when the copy/promotion is complete. 1005 assert(SafepointSynchronize::is_at_safepoint(), 1006 "expect promotion only at safepoints"); 1007 1008 if (_collectorState < Sweeping) { 1009 // Mark the appropriate cards in the modUnionTable, so that 1010 // this object gets scanned before the sweep. If this is 1011 // not done, CMS generation references in the object might 1012 // not get marked. 1013 // For the case of arrays, which are otherwise precisely 1014 // marked, we need to dirty the entire array, not just its head. 1015 if (is_obj_array) { 1016 // The [par_]mark_range() method expects mr.end() below to 1017 // be aligned to the granularity of a bit's representation 1018 // in the heap. In the case of the MUT below, that's a 1019 // card size. 1020 MemRegion mr(start, 1021 (HeapWord*)round_to((intptr_t)(start + obj_size), 1022 CardTableModRefBS::card_size /* bytes */)); 1023 if (par) { 1024 _modUnionTable.par_mark_range(mr); 1025 } else { 1026 _modUnionTable.mark_range(mr); 1027 } 1028 } else { // not an obj array; we can just mark the head 1029 if (par) { 1030 _modUnionTable.par_mark(start); 1031 } else { 1032 _modUnionTable.mark(start); 1033 } 1034 } 1035 } 1036 } 1037 } 1038 1039 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1040 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1041 // allocate, copy and if necessary update promoinfo -- 1042 // delegate to underlying space. 1043 assert_lock_strong(freelistLock()); 1044 1045 #ifndef PRODUCT 1046 if (Universe::heap()->promotion_should_fail()) { 1047 return NULL; 1048 } 1049 #endif // #ifndef PRODUCT 1050 1051 oop res = _cmsSpace->promote(obj, obj_size); 1052 if (res == NULL) { 1053 // expand and retry 1054 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1055 expand(s*HeapWordSize, MinHeapDeltaBytes, 1056 CMSExpansionCause::_satisfy_promotion); 1057 // Since there's currently no next generation, we don't try to promote 1058 // into a more senior generation. 1059 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1060 "is made to pass on a possibly failing " 1061 "promotion to next generation"); 1062 res = _cmsSpace->promote(obj, obj_size); 1063 } 1064 if (res != NULL) { 1065 // See comment in allocate() about when objects should 1066 // be allocated live. 1067 assert(obj->is_oop(), "Will dereference klass pointer below"); 1068 collector()->promoted(false, // Not parallel 1069 (HeapWord*)res, obj->is_objArray(), obj_size); 1070 // promotion counters 1071 NOT_PRODUCT( 1072 _numObjectsPromoted++; 1073 _numWordsPromoted += 1074 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1075 ) 1076 } 1077 return res; 1078 } 1079 1080 1081 // IMPORTANT: Notes on object size recognition in CMS. 1082 // --------------------------------------------------- 1083 // A block of storage in the CMS generation is always in 1084 // one of three states. A free block (FREE), an allocated 1085 // object (OBJECT) whose size() method reports the correct size, 1086 // and an intermediate state (TRANSIENT) in which its size cannot 1087 // be accurately determined. 1088 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1089 // ----------------------------------------------------- 1090 // FREE: klass_word & 1 == 1; mark_word holds block size 1091 // 1092 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1093 // obj->size() computes correct size 1094 // 1095 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1096 // 1097 // STATE IDENTIFICATION: (64 bit+COOPS) 1098 // ------------------------------------ 1099 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1100 // 1101 // OBJECT: klass_word installed; klass_word != 0; 1102 // obj->size() computes correct size 1103 // 1104 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1105 // 1106 // 1107 // STATE TRANSITION DIAGRAM 1108 // 1109 // mut / parnew mut / parnew 1110 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1111 // ^ | 1112 // |------------------------ DEAD <------------------------------------| 1113 // sweep mut 1114 // 1115 // While a block is in TRANSIENT state its size cannot be determined 1116 // so readers will either need to come back later or stall until 1117 // the size can be determined. Note that for the case of direct 1118 // allocation, P-bits, when available, may be used to determine the 1119 // size of an object that may not yet have been initialized. 1120 1121 // Things to support parallel young-gen collection. 1122 oop 1123 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1124 oop old, markOop m, 1125 size_t word_sz) { 1126 #ifndef PRODUCT 1127 if (Universe::heap()->promotion_should_fail()) { 1128 return NULL; 1129 } 1130 #endif // #ifndef PRODUCT 1131 1132 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1133 PromotionInfo* promoInfo = &ps->promo; 1134 // if we are tracking promotions, then first ensure space for 1135 // promotion (including spooling space for saving header if necessary). 1136 // then allocate and copy, then track promoted info if needed. 1137 // When tracking (see PromotionInfo::track()), the mark word may 1138 // be displaced and in this case restoration of the mark word 1139 // occurs in the (oop_since_save_marks_)iterate phase. 1140 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1141 // Out of space for allocating spooling buffers; 1142 // try expanding and allocating spooling buffers. 1143 if (!expand_and_ensure_spooling_space(promoInfo)) { 1144 return NULL; 1145 } 1146 } 1147 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1148 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1149 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1150 if (obj_ptr == NULL) { 1151 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1152 if (obj_ptr == NULL) { 1153 return NULL; 1154 } 1155 } 1156 oop obj = oop(obj_ptr); 1157 OrderAccess::storestore(); 1158 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1159 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1160 // IMPORTANT: See note on object initialization for CMS above. 1161 // Otherwise, copy the object. Here we must be careful to insert the 1162 // klass pointer last, since this marks the block as an allocated object. 1163 // Except with compressed oops it's the mark word. 1164 HeapWord* old_ptr = (HeapWord*)old; 1165 // Restore the mark word copied above. 1166 obj->set_mark(m); 1167 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1168 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1169 OrderAccess::storestore(); 1170 1171 if (UseCompressedClassPointers) { 1172 // Copy gap missed by (aligned) header size calculation below 1173 obj->set_klass_gap(old->klass_gap()); 1174 } 1175 if (word_sz > (size_t)oopDesc::header_size()) { 1176 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1177 obj_ptr + oopDesc::header_size(), 1178 word_sz - oopDesc::header_size()); 1179 } 1180 1181 // Now we can track the promoted object, if necessary. We take care 1182 // to delay the transition from uninitialized to full object 1183 // (i.e., insertion of klass pointer) until after, so that it 1184 // atomically becomes a promoted object. 1185 if (promoInfo->tracking()) { 1186 promoInfo->track((PromotedObject*)obj, old->klass()); 1187 } 1188 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1189 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1190 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1191 1192 // Finally, install the klass pointer (this should be volatile). 1193 OrderAccess::storestore(); 1194 obj->set_klass(old->klass()); 1195 // We should now be able to calculate the right size for this object 1196 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1197 1198 collector()->promoted(true, // parallel 1199 obj_ptr, old->is_objArray(), word_sz); 1200 1201 NOT_PRODUCT( 1202 Atomic::inc_ptr(&_numObjectsPromoted); 1203 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1204 ) 1205 1206 return obj; 1207 } 1208 1209 void 1210 ConcurrentMarkSweepGeneration:: 1211 par_promote_alloc_undo(int thread_num, 1212 HeapWord* obj, size_t word_sz) { 1213 // CMS does not support promotion undo. 1214 ShouldNotReachHere(); 1215 } 1216 1217 void 1218 ConcurrentMarkSweepGeneration:: 1219 par_promote_alloc_done(int thread_num) { 1220 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1221 ps->lab.retire(thread_num); 1222 } 1223 1224 void 1225 ConcurrentMarkSweepGeneration:: 1226 par_oop_since_save_marks_iterate_done(int thread_num) { 1227 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1228 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1229 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1230 } 1231 1232 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1233 size_t size, 1234 bool tlab) 1235 { 1236 // We allow a STW collection only if a full 1237 // collection was requested. 1238 return full || should_allocate(size, tlab); // FIX ME !!! 1239 // This and promotion failure handling are connected at the 1240 // hip and should be fixed by untying them. 1241 } 1242 1243 bool CMSCollector::shouldConcurrentCollect() { 1244 if (_full_gc_requested) { 1245 if (Verbose && PrintGCDetails) { 1246 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1247 " gc request (or gc_locker)"); 1248 } 1249 return true; 1250 } 1251 1252 // For debugging purposes, change the type of collection. 1253 // If the rotation is not on the concurrent collection 1254 // type, don't start a concurrent collection. 1255 NOT_PRODUCT( 1256 if (RotateCMSCollectionTypes && 1257 (_cmsGen->debug_collection_type() != 1258 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1259 assert(_cmsGen->debug_collection_type() != 1260 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1261 "Bad cms collection type"); 1262 return false; 1263 } 1264 ) 1265 1266 FreelistLocker x(this); 1267 // ------------------------------------------------------------------ 1268 // Print out lots of information which affects the initiation of 1269 // a collection. 1270 if (PrintCMSInitiationStatistics && stats().valid()) { 1271 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1272 gclog_or_tty->stamp(); 1273 gclog_or_tty->cr(); 1274 stats().print_on(gclog_or_tty); 1275 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1276 stats().time_until_cms_gen_full()); 1277 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1278 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1279 _cmsGen->contiguous_available()); 1280 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1281 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1282 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1283 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1284 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1285 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1286 gclog_or_tty->print_cr("metadata initialized %d", 1287 MetaspaceGC::should_concurrent_collect()); 1288 } 1289 // ------------------------------------------------------------------ 1290 1291 // If the estimated time to complete a cms collection (cms_duration()) 1292 // is less than the estimated time remaining until the cms generation 1293 // is full, start a collection. 1294 if (!UseCMSInitiatingOccupancyOnly) { 1295 if (stats().valid()) { 1296 if (stats().time_until_cms_start() == 0.0) { 1297 return true; 1298 } 1299 } else { 1300 // We want to conservatively collect somewhat early in order 1301 // to try and "bootstrap" our CMS/promotion statistics; 1302 // this branch will not fire after the first successful CMS 1303 // collection because the stats should then be valid. 1304 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1305 if (Verbose && PrintGCDetails) { 1306 gclog_or_tty->print_cr( 1307 " CMSCollector: collect for bootstrapping statistics:" 1308 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1309 _bootstrap_occupancy); 1310 } 1311 return true; 1312 } 1313 } 1314 } 1315 1316 // Otherwise, we start a collection cycle if 1317 // old gen want a collection cycle started. Each may use 1318 // an appropriate criterion for making this decision. 1319 // XXX We need to make sure that the gen expansion 1320 // criterion dovetails well with this. XXX NEED TO FIX THIS 1321 if (_cmsGen->should_concurrent_collect()) { 1322 if (Verbose && PrintGCDetails) { 1323 gclog_or_tty->print_cr("CMS old gen initiated"); 1324 } 1325 return true; 1326 } 1327 1328 // We start a collection if we believe an incremental collection may fail; 1329 // this is not likely to be productive in practice because it's probably too 1330 // late anyway. 1331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1332 assert(gch->collector_policy()->is_generation_policy(), 1333 "You may want to check the correctness of the following"); 1334 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1335 if (Verbose && PrintGCDetails) { 1336 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1337 } 1338 return true; 1339 } 1340 1341 if (MetaspaceGC::should_concurrent_collect()) { 1342 if (Verbose && PrintGCDetails) { 1343 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1344 } 1345 return true; 1346 } 1347 1348 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1349 if (CMSTriggerInterval >= 0) { 1350 if (CMSTriggerInterval == 0) { 1351 // Trigger always 1352 return true; 1353 } 1354 1355 // Check the CMS time since begin (we do not check the stats validity 1356 // as we want to be able to trigger the first CMS cycle as well) 1357 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1358 if (Verbose && PrintGCDetails) { 1359 if (stats().valid()) { 1360 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1361 stats().cms_time_since_begin()); 1362 } else { 1363 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1364 } 1365 } 1366 return true; 1367 } 1368 } 1369 1370 return false; 1371 } 1372 1373 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1374 1375 // Clear _expansion_cause fields of constituent generations 1376 void CMSCollector::clear_expansion_cause() { 1377 _cmsGen->clear_expansion_cause(); 1378 } 1379 1380 // We should be conservative in starting a collection cycle. To 1381 // start too eagerly runs the risk of collecting too often in the 1382 // extreme. To collect too rarely falls back on full collections, 1383 // which works, even if not optimum in terms of concurrent work. 1384 // As a work around for too eagerly collecting, use the flag 1385 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1386 // giving the user an easily understandable way of controlling the 1387 // collections. 1388 // We want to start a new collection cycle if any of the following 1389 // conditions hold: 1390 // . our current occupancy exceeds the configured initiating occupancy 1391 // for this generation, or 1392 // . we recently needed to expand this space and have not, since that 1393 // expansion, done a collection of this generation, or 1394 // . the underlying space believes that it may be a good idea to initiate 1395 // a concurrent collection (this may be based on criteria such as the 1396 // following: the space uses linear allocation and linear allocation is 1397 // going to fail, or there is believed to be excessive fragmentation in 1398 // the generation, etc... or ... 1399 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1400 // the case of the old generation; see CR 6543076): 1401 // we may be approaching a point at which allocation requests may fail because 1402 // we will be out of sufficient free space given allocation rate estimates.] 1403 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1404 1405 assert_lock_strong(freelistLock()); 1406 if (occupancy() > initiating_occupancy()) { 1407 if (PrintGCDetails && Verbose) { 1408 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1409 short_name(), occupancy(), initiating_occupancy()); 1410 } 1411 return true; 1412 } 1413 if (UseCMSInitiatingOccupancyOnly) { 1414 return false; 1415 } 1416 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1417 if (PrintGCDetails && Verbose) { 1418 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1419 short_name()); 1420 } 1421 return true; 1422 } 1423 if (_cmsSpace->should_concurrent_collect()) { 1424 if (PrintGCDetails && Verbose) { 1425 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1426 short_name()); 1427 } 1428 return true; 1429 } 1430 return false; 1431 } 1432 1433 void ConcurrentMarkSweepGeneration::collect(bool full, 1434 bool clear_all_soft_refs, 1435 size_t size, 1436 bool tlab) 1437 { 1438 collector()->collect(full, clear_all_soft_refs, size, tlab); 1439 } 1440 1441 void CMSCollector::collect(bool full, 1442 bool clear_all_soft_refs, 1443 size_t size, 1444 bool tlab) 1445 { 1446 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1447 // For debugging purposes skip the collection if the state 1448 // is not currently idle 1449 if (TraceCMSState) { 1450 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1451 Thread::current(), full, _collectorState); 1452 } 1453 return; 1454 } 1455 1456 // The following "if" branch is present for defensive reasons. 1457 // In the current uses of this interface, it can be replaced with: 1458 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1459 // But I am not placing that assert here to allow future 1460 // generality in invoking this interface. 1461 if (GC_locker::is_active()) { 1462 // A consistency test for GC_locker 1463 assert(GC_locker::needs_gc(), "Should have been set already"); 1464 // Skip this foreground collection, instead 1465 // expanding the heap if necessary. 1466 // Need the free list locks for the call to free() in compute_new_size() 1467 compute_new_size(); 1468 return; 1469 } 1470 acquire_control_and_collect(full, clear_all_soft_refs); 1471 _full_gcs_since_conc_gc++; 1472 } 1473 1474 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1475 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1476 unsigned int gc_count = gch->total_full_collections(); 1477 if (gc_count == full_gc_count) { 1478 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1479 _full_gc_requested = true; 1480 _full_gc_cause = cause; 1481 CGC_lock->notify(); // nudge CMS thread 1482 } else { 1483 assert(gc_count > full_gc_count, "Error: causal loop"); 1484 } 1485 } 1486 1487 bool CMSCollector::is_external_interruption() { 1488 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1489 return GCCause::is_user_requested_gc(cause) || 1490 GCCause::is_serviceability_requested_gc(cause); 1491 } 1492 1493 void CMSCollector::report_concurrent_mode_interruption() { 1494 if (is_external_interruption()) { 1495 if (PrintGCDetails) { 1496 gclog_or_tty->print(" (concurrent mode interrupted)"); 1497 } 1498 } else { 1499 if (PrintGCDetails) { 1500 gclog_or_tty->print(" (concurrent mode failure)"); 1501 } 1502 _gc_tracer_cm->report_concurrent_mode_failure(); 1503 } 1504 } 1505 1506 1507 // The foreground and background collectors need to coordinate in order 1508 // to make sure that they do not mutually interfere with CMS collections. 1509 // When a background collection is active, 1510 // the foreground collector may need to take over (preempt) and 1511 // synchronously complete an ongoing collection. Depending on the 1512 // frequency of the background collections and the heap usage 1513 // of the application, this preemption can be seldom or frequent. 1514 // There are only certain 1515 // points in the background collection that the "collection-baton" 1516 // can be passed to the foreground collector. 1517 // 1518 // The foreground collector will wait for the baton before 1519 // starting any part of the collection. The foreground collector 1520 // will only wait at one location. 1521 // 1522 // The background collector will yield the baton before starting a new 1523 // phase of the collection (e.g., before initial marking, marking from roots, 1524 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1525 // of the loop which switches the phases. The background collector does some 1526 // of the phases (initial mark, final re-mark) with the world stopped. 1527 // Because of locking involved in stopping the world, 1528 // the foreground collector should not block waiting for the background 1529 // collector when it is doing a stop-the-world phase. The background 1530 // collector will yield the baton at an additional point just before 1531 // it enters a stop-the-world phase. Once the world is stopped, the 1532 // background collector checks the phase of the collection. If the 1533 // phase has not changed, it proceeds with the collection. If the 1534 // phase has changed, it skips that phase of the collection. See 1535 // the comments on the use of the Heap_lock in collect_in_background(). 1536 // 1537 // Variable used in baton passing. 1538 // _foregroundGCIsActive - Set to true by the foreground collector when 1539 // it wants the baton. The foreground clears it when it has finished 1540 // the collection. 1541 // _foregroundGCShouldWait - Set to true by the background collector 1542 // when it is running. The foreground collector waits while 1543 // _foregroundGCShouldWait is true. 1544 // CGC_lock - monitor used to protect access to the above variables 1545 // and to notify the foreground and background collectors. 1546 // _collectorState - current state of the CMS collection. 1547 // 1548 // The foreground collector 1549 // acquires the CGC_lock 1550 // sets _foregroundGCIsActive 1551 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1552 // various locks acquired in preparation for the collection 1553 // are released so as not to block the background collector 1554 // that is in the midst of a collection 1555 // proceeds with the collection 1556 // clears _foregroundGCIsActive 1557 // returns 1558 // 1559 // The background collector in a loop iterating on the phases of the 1560 // collection 1561 // acquires the CGC_lock 1562 // sets _foregroundGCShouldWait 1563 // if _foregroundGCIsActive is set 1564 // clears _foregroundGCShouldWait, notifies _CGC_lock 1565 // waits on _CGC_lock for _foregroundGCIsActive to become false 1566 // and exits the loop. 1567 // otherwise 1568 // proceed with that phase of the collection 1569 // if the phase is a stop-the-world phase, 1570 // yield the baton once more just before enqueueing 1571 // the stop-world CMS operation (executed by the VM thread). 1572 // returns after all phases of the collection are done 1573 // 1574 1575 void CMSCollector::acquire_control_and_collect(bool full, 1576 bool clear_all_soft_refs) { 1577 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1578 assert(!Thread::current()->is_ConcurrentGC_thread(), 1579 "shouldn't try to acquire control from self!"); 1580 1581 // Start the protocol for acquiring control of the 1582 // collection from the background collector (aka CMS thread). 1583 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1584 "VM thread should have CMS token"); 1585 // Remember the possibly interrupted state of an ongoing 1586 // concurrent collection 1587 CollectorState first_state = _collectorState; 1588 1589 // Signal to a possibly ongoing concurrent collection that 1590 // we want to do a foreground collection. 1591 _foregroundGCIsActive = true; 1592 1593 // release locks and wait for a notify from the background collector 1594 // releasing the locks in only necessary for phases which 1595 // do yields to improve the granularity of the collection. 1596 assert_lock_strong(bitMapLock()); 1597 // We need to lock the Free list lock for the space that we are 1598 // currently collecting. 1599 assert(haveFreelistLocks(), "Must be holding free list locks"); 1600 bitMapLock()->unlock(); 1601 releaseFreelistLocks(); 1602 { 1603 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1604 if (_foregroundGCShouldWait) { 1605 // We are going to be waiting for action for the CMS thread; 1606 // it had better not be gone (for instance at shutdown)! 1607 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1608 "CMS thread must be running"); 1609 // Wait here until the background collector gives us the go-ahead 1610 ConcurrentMarkSweepThread::clear_CMS_flag( 1611 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1612 // Get a possibly blocked CMS thread going: 1613 // Note that we set _foregroundGCIsActive true above, 1614 // without protection of the CGC_lock. 1615 CGC_lock->notify(); 1616 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1617 "Possible deadlock"); 1618 while (_foregroundGCShouldWait) { 1619 // wait for notification 1620 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1621 // Possibility of delay/starvation here, since CMS token does 1622 // not know to give priority to VM thread? Actually, i think 1623 // there wouldn't be any delay/starvation, but the proof of 1624 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1625 } 1626 ConcurrentMarkSweepThread::set_CMS_flag( 1627 ConcurrentMarkSweepThread::CMS_vm_has_token); 1628 } 1629 } 1630 // The CMS_token is already held. Get back the other locks. 1631 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1632 "VM thread should have CMS token"); 1633 getFreelistLocks(); 1634 bitMapLock()->lock_without_safepoint_check(); 1635 if (TraceCMSState) { 1636 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1637 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1638 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1639 } 1640 1641 // Check if we need to do a compaction, or if not, whether 1642 // we need to start the mark-sweep from scratch. 1643 bool should_compact = false; 1644 bool should_start_over = false; 1645 decide_foreground_collection_type(clear_all_soft_refs, 1646 &should_compact, &should_start_over); 1647 1648 NOT_PRODUCT( 1649 if (RotateCMSCollectionTypes) { 1650 if (_cmsGen->debug_collection_type() == 1651 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1652 should_compact = true; 1653 } else if (_cmsGen->debug_collection_type() == 1654 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1655 should_compact = false; 1656 } 1657 } 1658 ) 1659 1660 if (first_state > Idling) { 1661 report_concurrent_mode_interruption(); 1662 } 1663 1664 set_did_compact(should_compact); 1665 if (should_compact) { 1666 // If the collection is being acquired from the background 1667 // collector, there may be references on the discovered 1668 // references lists that have NULL referents (being those 1669 // that were concurrently cleared by a mutator) or 1670 // that are no longer active (having been enqueued concurrently 1671 // by the mutator). 1672 // Scrub the list of those references because Mark-Sweep-Compact 1673 // code assumes referents are not NULL and that all discovered 1674 // Reference objects are active. 1675 ref_processor()->clean_up_discovered_references(); 1676 1677 if (first_state > Idling) { 1678 save_heap_summary(); 1679 } 1680 1681 do_compaction_work(clear_all_soft_refs); 1682 1683 // Has the GC time limit been exceeded? 1684 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1685 size_t max_eden_size = young_gen->max_capacity() - 1686 young_gen->to()->capacity() - 1687 young_gen->from()->capacity(); 1688 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1689 GCCause::Cause gc_cause = gch->gc_cause(); 1690 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1691 young_gen->eden()->used(), 1692 _cmsGen->max_capacity(), 1693 max_eden_size, 1694 full, 1695 gc_cause, 1696 gch->collector_policy()); 1697 } else { 1698 do_mark_sweep_work(clear_all_soft_refs, first_state, 1699 should_start_over); 1700 } 1701 // Reset the expansion cause, now that we just completed 1702 // a collection cycle. 1703 clear_expansion_cause(); 1704 _foregroundGCIsActive = false; 1705 return; 1706 } 1707 1708 // Resize the tenured generation 1709 // after obtaining the free list locks for the 1710 // two generations. 1711 void CMSCollector::compute_new_size() { 1712 assert_locked_or_safepoint(Heap_lock); 1713 FreelistLocker z(this); 1714 MetaspaceGC::compute_new_size(); 1715 _cmsGen->compute_new_size_free_list(); 1716 } 1717 1718 // A work method used by foreground collection to determine 1719 // what type of collection (compacting or not, continuing or fresh) 1720 // it should do. 1721 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1722 // and CMSCompactWhenClearAllSoftRefs the default in the future 1723 // and do away with the flags after a suitable period. 1724 void CMSCollector::decide_foreground_collection_type( 1725 bool clear_all_soft_refs, bool* should_compact, 1726 bool* should_start_over) { 1727 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1728 // flag is set, and we have either requested a System.gc() or 1729 // the number of full gc's since the last concurrent cycle 1730 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1731 // or if an incremental collection has failed 1732 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1733 assert(gch->collector_policy()->is_generation_policy(), 1734 "You may want to check the correctness of the following"); 1735 // Inform cms gen if this was due to partial collection failing. 1736 // The CMS gen may use this fact to determine its expansion policy. 1737 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1738 assert(!_cmsGen->incremental_collection_failed(), 1739 "Should have been noticed, reacted to and cleared"); 1740 _cmsGen->set_incremental_collection_failed(); 1741 } 1742 *should_compact = 1743 UseCMSCompactAtFullCollection && 1744 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1745 GCCause::is_user_requested_gc(gch->gc_cause()) || 1746 gch->incremental_collection_will_fail(true /* consult_young */)); 1747 *should_start_over = false; 1748 if (clear_all_soft_refs && !*should_compact) { 1749 // We are about to do a last ditch collection attempt 1750 // so it would normally make sense to do a compaction 1751 // to reclaim as much space as possible. 1752 if (CMSCompactWhenClearAllSoftRefs) { 1753 // Default: The rationale is that in this case either 1754 // we are past the final marking phase, in which case 1755 // we'd have to start over, or so little has been done 1756 // that there's little point in saving that work. Compaction 1757 // appears to be the sensible choice in either case. 1758 *should_compact = true; 1759 } else { 1760 // We have been asked to clear all soft refs, but not to 1761 // compact. Make sure that we aren't past the final checkpoint 1762 // phase, for that is where we process soft refs. If we are already 1763 // past that phase, we'll need to redo the refs discovery phase and 1764 // if necessary clear soft refs that weren't previously 1765 // cleared. We do so by remembering the phase in which 1766 // we came in, and if we are past the refs processing 1767 // phase, we'll choose to just redo the mark-sweep 1768 // collection from scratch. 1769 if (_collectorState > FinalMarking) { 1770 // We are past the refs processing phase; 1771 // start over and do a fresh synchronous CMS cycle 1772 _collectorState = Resetting; // skip to reset to start new cycle 1773 reset(false /* == !asynch */); 1774 *should_start_over = true; 1775 } // else we can continue a possibly ongoing current cycle 1776 } 1777 } 1778 } 1779 1780 // A work method used by the foreground collector to do 1781 // a mark-sweep-compact. 1782 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1783 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1784 1785 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1786 gc_timer->register_gc_start(); 1787 1788 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1789 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1790 1791 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1792 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1793 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1794 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1795 } 1796 1797 // Temporarily widen the span of the weak reference processing to 1798 // the entire heap. 1799 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1800 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1801 // Temporarily, clear the "is_alive_non_header" field of the 1802 // reference processor. 1803 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1804 // Temporarily make reference _processing_ single threaded (non-MT). 1805 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1806 // Temporarily make refs discovery atomic 1807 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1808 // Temporarily make reference _discovery_ single threaded (non-MT) 1809 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1810 1811 ref_processor()->set_enqueuing_is_done(false); 1812 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 1813 ref_processor()->setup_policy(clear_all_soft_refs); 1814 // If an asynchronous collection finishes, the _modUnionTable is 1815 // all clear. If we are assuming the collection from an asynchronous 1816 // collection, clear the _modUnionTable. 1817 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1818 "_modUnionTable should be clear if the baton was not passed"); 1819 _modUnionTable.clear_all(); 1820 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1821 "mod union for klasses should be clear if the baton was passed"); 1822 _ct->klass_rem_set()->clear_mod_union(); 1823 1824 // We must adjust the allocation statistics being maintained 1825 // in the free list space. We do so by reading and clearing 1826 // the sweep timer and updating the block flux rate estimates below. 1827 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1828 if (_inter_sweep_timer.is_active()) { 1829 _inter_sweep_timer.stop(); 1830 // Note that we do not use this sample to update the _inter_sweep_estimate. 1831 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1832 _inter_sweep_estimate.padded_average(), 1833 _intra_sweep_estimate.padded_average()); 1834 } 1835 1836 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1837 ref_processor(), clear_all_soft_refs); 1838 #ifdef ASSERT 1839 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1840 size_t free_size = cms_space->free(); 1841 assert(free_size == 1842 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1843 * HeapWordSize, 1844 "All the free space should be compacted into one chunk at top"); 1845 assert(cms_space->dictionary()->total_chunk_size( 1846 debug_only(cms_space->freelistLock())) == 0 || 1847 cms_space->totalSizeInIndexedFreeLists() == 0, 1848 "All the free space should be in a single chunk"); 1849 size_t num = cms_space->totalCount(); 1850 assert((free_size == 0 && num == 0) || 1851 (free_size > 0 && (num == 1 || num == 2)), 1852 "There should be at most 2 free chunks after compaction"); 1853 #endif // ASSERT 1854 _collectorState = Resetting; 1855 assert(_restart_addr == NULL, 1856 "Should have been NULL'd before baton was passed"); 1857 reset(false /* == !asynch */); 1858 _cmsGen->reset_after_compaction(); 1859 _concurrent_cycles_since_last_unload = 0; 1860 1861 // Clear any data recorded in the PLAB chunk arrays. 1862 if (_survivor_plab_array != NULL) { 1863 reset_survivor_plab_arrays(); 1864 } 1865 1866 // Adjust the per-size allocation stats for the next epoch. 1867 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1868 // Restart the "inter sweep timer" for the next epoch. 1869 _inter_sweep_timer.reset(); 1870 _inter_sweep_timer.start(); 1871 1872 gc_timer->register_gc_end(); 1873 1874 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1875 1876 // For a mark-sweep-compact, compute_new_size() will be called 1877 // in the heap's do_collection() method. 1878 } 1879 1880 // A work method used by the foreground collector to do 1881 // a mark-sweep, after taking over from a possibly on-going 1882 // concurrent mark-sweep collection. 1883 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 1884 CollectorState first_state, bool should_start_over) { 1885 if (PrintGC && Verbose) { 1886 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 1887 "collector with count %d", 1888 _full_gcs_since_conc_gc); 1889 } 1890 switch (_collectorState) { 1891 case Idling: 1892 if (first_state == Idling || should_start_over) { 1893 // The background GC was not active, or should 1894 // restarted from scratch; start the cycle. 1895 _collectorState = InitialMarking; 1896 } 1897 // If first_state was not Idling, then a background GC 1898 // was in progress and has now finished. No need to do it 1899 // again. Leave the state as Idling. 1900 break; 1901 case Precleaning: 1902 // In the foreground case don't do the precleaning since 1903 // it is not done concurrently and there is extra work 1904 // required. 1905 _collectorState = FinalMarking; 1906 } 1907 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 1908 1909 // For a mark-sweep, compute_new_size() will be called 1910 // in the heap's do_collection() method. 1911 } 1912 1913 1914 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1915 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 1916 ContiguousSpace* eden_space = dng->eden(); 1917 ContiguousSpace* from_space = dng->from(); 1918 ContiguousSpace* to_space = dng->to(); 1919 // Eden 1920 if (_eden_chunk_array != NULL) { 1921 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1922 eden_space->bottom(), eden_space->top(), 1923 eden_space->end(), eden_space->capacity()); 1924 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1925 "_eden_chunk_capacity=" SIZE_FORMAT, 1926 _eden_chunk_index, _eden_chunk_capacity); 1927 for (size_t i = 0; i < _eden_chunk_index; i++) { 1928 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1929 i, _eden_chunk_array[i]); 1930 } 1931 } 1932 // Survivor 1933 if (_survivor_chunk_array != NULL) { 1934 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1935 from_space->bottom(), from_space->top(), 1936 from_space->end(), from_space->capacity()); 1937 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1938 "_survivor_chunk_capacity=" SIZE_FORMAT, 1939 _survivor_chunk_index, _survivor_chunk_capacity); 1940 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1941 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1942 i, _survivor_chunk_array[i]); 1943 } 1944 } 1945 } 1946 1947 void CMSCollector::getFreelistLocks() const { 1948 // Get locks for all free lists in all generations that this 1949 // collector is responsible for 1950 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1951 } 1952 1953 void CMSCollector::releaseFreelistLocks() const { 1954 // Release locks for all free lists in all generations that this 1955 // collector is responsible for 1956 _cmsGen->freelistLock()->unlock(); 1957 } 1958 1959 bool CMSCollector::haveFreelistLocks() const { 1960 // Check locks for all free lists in all generations that this 1961 // collector is responsible for 1962 assert_lock_strong(_cmsGen->freelistLock()); 1963 PRODUCT_ONLY(ShouldNotReachHere()); 1964 return true; 1965 } 1966 1967 // A utility class that is used by the CMS collector to 1968 // temporarily "release" the foreground collector from its 1969 // usual obligation to wait for the background collector to 1970 // complete an ongoing phase before proceeding. 1971 class ReleaseForegroundGC: public StackObj { 1972 private: 1973 CMSCollector* _c; 1974 public: 1975 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1976 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1977 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1978 // allow a potentially blocked foreground collector to proceed 1979 _c->_foregroundGCShouldWait = false; 1980 if (_c->_foregroundGCIsActive) { 1981 CGC_lock->notify(); 1982 } 1983 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1984 "Possible deadlock"); 1985 } 1986 1987 ~ReleaseForegroundGC() { 1988 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1989 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1990 _c->_foregroundGCShouldWait = true; 1991 } 1992 }; 1993 1994 // There are separate collect_in_background and collect_in_foreground because of 1995 // the different locking requirements of the background collector and the 1996 // foreground collector. There was originally an attempt to share 1997 // one "collect" method between the background collector and the foreground 1998 // collector but the if-then-else required made it cleaner to have 1999 // separate methods. 2000 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2001 assert(Thread::current()->is_ConcurrentGC_thread(), 2002 "A CMS asynchronous collection is only allowed on a CMS thread."); 2003 2004 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2005 { 2006 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2007 MutexLockerEx hl(Heap_lock, safepoint_check); 2008 FreelistLocker fll(this); 2009 MutexLockerEx x(CGC_lock, safepoint_check); 2010 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2011 // The foreground collector is active or we're 2012 // not using asynchronous collections. Skip this 2013 // background collection. 2014 assert(!_foregroundGCShouldWait, "Should be clear"); 2015 return; 2016 } else { 2017 assert(_collectorState == Idling, "Should be idling before start."); 2018 _collectorState = InitialMarking; 2019 register_gc_start(cause); 2020 // Reset the expansion cause, now that we are about to begin 2021 // a new cycle. 2022 clear_expansion_cause(); 2023 2024 // Clear the MetaspaceGC flag since a concurrent collection 2025 // is starting but also clear it after the collection. 2026 MetaspaceGC::set_should_concurrent_collect(false); 2027 } 2028 // Decide if we want to enable class unloading as part of the 2029 // ensuing concurrent GC cycle. 2030 update_should_unload_classes(); 2031 _full_gc_requested = false; // acks all outstanding full gc requests 2032 _full_gc_cause = GCCause::_no_gc; 2033 // Signal that we are about to start a collection 2034 gch->increment_total_full_collections(); // ... starting a collection cycle 2035 _collection_count_start = gch->total_full_collections(); 2036 } 2037 2038 // Used for PrintGC 2039 size_t prev_used; 2040 if (PrintGC && Verbose) { 2041 prev_used = _cmsGen->used(); // XXXPERM 2042 } 2043 2044 // The change of the collection state is normally done at this level; 2045 // the exceptions are phases that are executed while the world is 2046 // stopped. For those phases the change of state is done while the 2047 // world is stopped. For baton passing purposes this allows the 2048 // background collector to finish the phase and change state atomically. 2049 // The foreground collector cannot wait on a phase that is done 2050 // while the world is stopped because the foreground collector already 2051 // has the world stopped and would deadlock. 2052 while (_collectorState != Idling) { 2053 if (TraceCMSState) { 2054 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2055 Thread::current(), _collectorState); 2056 } 2057 // The foreground collector 2058 // holds the Heap_lock throughout its collection. 2059 // holds the CMS token (but not the lock) 2060 // except while it is waiting for the background collector to yield. 2061 // 2062 // The foreground collector should be blocked (not for long) 2063 // if the background collector is about to start a phase 2064 // executed with world stopped. If the background 2065 // collector has already started such a phase, the 2066 // foreground collector is blocked waiting for the 2067 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2068 // are executed in the VM thread. 2069 // 2070 // The locking order is 2071 // PendingListLock (PLL) -- if applicable (FinalMarking) 2072 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2073 // CMS token (claimed in 2074 // stop_world_and_do() --> 2075 // safepoint_synchronize() --> 2076 // CMSThread::synchronize()) 2077 2078 { 2079 // Check if the FG collector wants us to yield. 2080 CMSTokenSync x(true); // is cms thread 2081 if (waitForForegroundGC()) { 2082 // We yielded to a foreground GC, nothing more to be 2083 // done this round. 2084 assert(_foregroundGCShouldWait == false, "We set it to false in " 2085 "waitForForegroundGC()"); 2086 if (TraceCMSState) { 2087 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2088 " exiting collection CMS state %d", 2089 Thread::current(), _collectorState); 2090 } 2091 return; 2092 } else { 2093 // The background collector can run but check to see if the 2094 // foreground collector has done a collection while the 2095 // background collector was waiting to get the CGC_lock 2096 // above. If yes, break so that _foregroundGCShouldWait 2097 // is cleared before returning. 2098 if (_collectorState == Idling) { 2099 break; 2100 } 2101 } 2102 } 2103 2104 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2105 "should be waiting"); 2106 2107 switch (_collectorState) { 2108 case InitialMarking: 2109 { 2110 ReleaseForegroundGC x(this); 2111 stats().record_cms_begin(); 2112 VM_CMS_Initial_Mark initial_mark_op(this); 2113 VMThread::execute(&initial_mark_op); 2114 } 2115 // The collector state may be any legal state at this point 2116 // since the background collector may have yielded to the 2117 // foreground collector. 2118 break; 2119 case Marking: 2120 // initial marking in checkpointRootsInitialWork has been completed 2121 if (markFromRoots(true)) { // we were successful 2122 assert(_collectorState == Precleaning, "Collector state should " 2123 "have changed"); 2124 } else { 2125 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2126 } 2127 break; 2128 case Precleaning: 2129 // marking from roots in markFromRoots has been completed 2130 preclean(); 2131 assert(_collectorState == AbortablePreclean || 2132 _collectorState == FinalMarking, 2133 "Collector state should have changed"); 2134 break; 2135 case AbortablePreclean: 2136 abortable_preclean(); 2137 assert(_collectorState == FinalMarking, "Collector state should " 2138 "have changed"); 2139 break; 2140 case FinalMarking: 2141 { 2142 ReleaseForegroundGC x(this); 2143 2144 VM_CMS_Final_Remark final_remark_op(this); 2145 VMThread::execute(&final_remark_op); 2146 } 2147 assert(_foregroundGCShouldWait, "block post-condition"); 2148 break; 2149 case Sweeping: 2150 // final marking in checkpointRootsFinal has been completed 2151 sweep(true); 2152 assert(_collectorState == Resizing, "Collector state change " 2153 "to Resizing must be done under the free_list_lock"); 2154 _full_gcs_since_conc_gc = 0; 2155 2156 case Resizing: { 2157 // Sweeping has been completed... 2158 // At this point the background collection has completed. 2159 // Don't move the call to compute_new_size() down 2160 // into code that might be executed if the background 2161 // collection was preempted. 2162 { 2163 ReleaseForegroundGC x(this); // unblock FG collection 2164 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2165 CMSTokenSync z(true); // not strictly needed. 2166 if (_collectorState == Resizing) { 2167 compute_new_size(); 2168 save_heap_summary(); 2169 _collectorState = Resetting; 2170 } else { 2171 assert(_collectorState == Idling, "The state should only change" 2172 " because the foreground collector has finished the collection"); 2173 } 2174 } 2175 break; 2176 } 2177 case Resetting: 2178 // CMS heap resizing has been completed 2179 reset(true); 2180 assert(_collectorState == Idling, "Collector state should " 2181 "have changed"); 2182 2183 MetaspaceGC::set_should_concurrent_collect(false); 2184 2185 stats().record_cms_end(); 2186 // Don't move the concurrent_phases_end() and compute_new_size() 2187 // calls to here because a preempted background collection 2188 // has it's state set to "Resetting". 2189 break; 2190 case Idling: 2191 default: 2192 ShouldNotReachHere(); 2193 break; 2194 } 2195 if (TraceCMSState) { 2196 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2197 Thread::current(), _collectorState); 2198 } 2199 assert(_foregroundGCShouldWait, "block post-condition"); 2200 } 2201 2202 // Should this be in gc_epilogue? 2203 collector_policy()->counters()->update_counters(); 2204 2205 { 2206 // Clear _foregroundGCShouldWait and, in the event that the 2207 // foreground collector is waiting, notify it, before 2208 // returning. 2209 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2210 _foregroundGCShouldWait = false; 2211 if (_foregroundGCIsActive) { 2212 CGC_lock->notify(); 2213 } 2214 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2215 "Possible deadlock"); 2216 } 2217 if (TraceCMSState) { 2218 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2219 " exiting collection CMS state %d", 2220 Thread::current(), _collectorState); 2221 } 2222 if (PrintGC && Verbose) { 2223 _cmsGen->print_heap_change(prev_used); 2224 } 2225 } 2226 2227 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2228 if (!_cms_start_registered) { 2229 register_gc_start(cause); 2230 } 2231 } 2232 2233 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2234 _cms_start_registered = true; 2235 _gc_timer_cm->register_gc_start(); 2236 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2237 } 2238 2239 void CMSCollector::register_gc_end() { 2240 if (_cms_start_registered) { 2241 report_heap_summary(GCWhen::AfterGC); 2242 2243 _gc_timer_cm->register_gc_end(); 2244 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2245 _cms_start_registered = false; 2246 } 2247 } 2248 2249 void CMSCollector::save_heap_summary() { 2250 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2251 _last_heap_summary = gch->create_heap_summary(); 2252 _last_metaspace_summary = gch->create_metaspace_summary(); 2253 } 2254 2255 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2256 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2257 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2258 } 2259 2260 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2261 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2262 "Foreground collector should be waiting, not executing"); 2263 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2264 "may only be done by the VM Thread with the world stopped"); 2265 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2266 "VM thread should have CMS token"); 2267 2268 // The gc id is created in register_foreground_gc_start if this collection is synchronous 2269 const GCId gc_id = _collectorState == InitialMarking ? GCId::peek() : _gc_tracer_cm->gc_id(); 2270 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2271 true, NULL, gc_id);) 2272 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2273 2274 HandleMark hm; // Discard invalid handles created during verification 2275 2276 if (VerifyBeforeGC && 2277 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2278 Universe::verify(); 2279 } 2280 2281 // Snapshot the soft reference policy to be used in this collection cycle. 2282 ref_processor()->setup_policy(clear_all_soft_refs); 2283 2284 // Decide if class unloading should be done 2285 update_should_unload_classes(); 2286 2287 bool init_mark_was_synchronous = false; // until proven otherwise 2288 while (_collectorState != Idling) { 2289 if (TraceCMSState) { 2290 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2291 Thread::current(), _collectorState); 2292 } 2293 switch (_collectorState) { 2294 case InitialMarking: 2295 register_foreground_gc_start(cause); 2296 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2297 checkpointRootsInitial(false); 2298 assert(_collectorState == Marking, "Collector state should have changed" 2299 " within checkpointRootsInitial()"); 2300 break; 2301 case Marking: 2302 // initial marking in checkpointRootsInitialWork has been completed 2303 if (VerifyDuringGC && 2304 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2305 Universe::verify("Verify before initial mark: "); 2306 } 2307 { 2308 bool res = markFromRoots(false); 2309 assert(res && _collectorState == FinalMarking, "Collector state should " 2310 "have changed"); 2311 break; 2312 } 2313 case FinalMarking: 2314 if (VerifyDuringGC && 2315 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2316 Universe::verify("Verify before re-mark: "); 2317 } 2318 checkpointRootsFinal(false, clear_all_soft_refs, 2319 init_mark_was_synchronous); 2320 assert(_collectorState == Sweeping, "Collector state should not " 2321 "have changed within checkpointRootsFinal()"); 2322 break; 2323 case Sweeping: 2324 // final marking in checkpointRootsFinal has been completed 2325 if (VerifyDuringGC && 2326 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2327 Universe::verify("Verify before sweep: "); 2328 } 2329 sweep(false); 2330 assert(_collectorState == Resizing, "Incorrect state"); 2331 break; 2332 case Resizing: { 2333 // Sweeping has been completed; the actual resize in this case 2334 // is done separately; nothing to be done in this state. 2335 _collectorState = Resetting; 2336 break; 2337 } 2338 case Resetting: 2339 // The heap has been resized. 2340 if (VerifyDuringGC && 2341 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2342 Universe::verify("Verify before reset: "); 2343 } 2344 save_heap_summary(); 2345 reset(false); 2346 assert(_collectorState == Idling, "Collector state should " 2347 "have changed"); 2348 break; 2349 case Precleaning: 2350 case AbortablePreclean: 2351 // Elide the preclean phase 2352 _collectorState = FinalMarking; 2353 break; 2354 default: 2355 ShouldNotReachHere(); 2356 } 2357 if (TraceCMSState) { 2358 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2359 Thread::current(), _collectorState); 2360 } 2361 } 2362 2363 if (VerifyAfterGC && 2364 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2365 Universe::verify(); 2366 } 2367 if (TraceCMSState) { 2368 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2369 " exiting collection CMS state %d", 2370 Thread::current(), _collectorState); 2371 } 2372 } 2373 2374 bool CMSCollector::waitForForegroundGC() { 2375 bool res = false; 2376 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2377 "CMS thread should have CMS token"); 2378 // Block the foreground collector until the 2379 // background collectors decides whether to 2380 // yield. 2381 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2382 _foregroundGCShouldWait = true; 2383 if (_foregroundGCIsActive) { 2384 // The background collector yields to the 2385 // foreground collector and returns a value 2386 // indicating that it has yielded. The foreground 2387 // collector can proceed. 2388 res = true; 2389 _foregroundGCShouldWait = false; 2390 ConcurrentMarkSweepThread::clear_CMS_flag( 2391 ConcurrentMarkSweepThread::CMS_cms_has_token); 2392 ConcurrentMarkSweepThread::set_CMS_flag( 2393 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2394 // Get a possibly blocked foreground thread going 2395 CGC_lock->notify(); 2396 if (TraceCMSState) { 2397 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2398 Thread::current(), _collectorState); 2399 } 2400 while (_foregroundGCIsActive) { 2401 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2402 } 2403 ConcurrentMarkSweepThread::set_CMS_flag( 2404 ConcurrentMarkSweepThread::CMS_cms_has_token); 2405 ConcurrentMarkSweepThread::clear_CMS_flag( 2406 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2407 } 2408 if (TraceCMSState) { 2409 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2410 Thread::current(), _collectorState); 2411 } 2412 return res; 2413 } 2414 2415 // Because of the need to lock the free lists and other structures in 2416 // the collector, common to all the generations that the collector is 2417 // collecting, we need the gc_prologues of individual CMS generations 2418 // delegate to their collector. It may have been simpler had the 2419 // current infrastructure allowed one to call a prologue on a 2420 // collector. In the absence of that we have the generation's 2421 // prologue delegate to the collector, which delegates back 2422 // some "local" work to a worker method in the individual generations 2423 // that it's responsible for collecting, while itself doing any 2424 // work common to all generations it's responsible for. A similar 2425 // comment applies to the gc_epilogue()'s. 2426 // The role of the variable _between_prologue_and_epilogue is to 2427 // enforce the invocation protocol. 2428 void CMSCollector::gc_prologue(bool full) { 2429 // Call gc_prologue_work() for the CMSGen 2430 // we are responsible for. 2431 2432 // The following locking discipline assumes that we are only called 2433 // when the world is stopped. 2434 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2435 2436 // The CMSCollector prologue must call the gc_prologues for the 2437 // "generations" that it's responsible 2438 // for. 2439 2440 assert( Thread::current()->is_VM_thread() 2441 || ( CMSScavengeBeforeRemark 2442 && Thread::current()->is_ConcurrentGC_thread()), 2443 "Incorrect thread type for prologue execution"); 2444 2445 if (_between_prologue_and_epilogue) { 2446 // We have already been invoked; this is a gc_prologue delegation 2447 // from yet another CMS generation that we are responsible for, just 2448 // ignore it since all relevant work has already been done. 2449 return; 2450 } 2451 2452 // set a bit saying prologue has been called; cleared in epilogue 2453 _between_prologue_and_epilogue = true; 2454 // Claim locks for common data structures, then call gc_prologue_work() 2455 // for each CMSGen. 2456 2457 getFreelistLocks(); // gets free list locks on constituent spaces 2458 bitMapLock()->lock_without_safepoint_check(); 2459 2460 // Should call gc_prologue_work() for all cms gens we are responsible for 2461 bool duringMarking = _collectorState >= Marking 2462 && _collectorState < Sweeping; 2463 2464 // The young collections clear the modified oops state, which tells if 2465 // there are any modified oops in the class. The remark phase also needs 2466 // that information. Tell the young collection to save the union of all 2467 // modified klasses. 2468 if (duringMarking) { 2469 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2470 } 2471 2472 bool registerClosure = duringMarking; 2473 2474 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2475 &_modUnionClosurePar 2476 : &_modUnionClosure; 2477 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2478 2479 if (!full) { 2480 stats().record_gc0_begin(); 2481 } 2482 } 2483 2484 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2485 2486 _capacity_at_prologue = capacity(); 2487 _used_at_prologue = used(); 2488 2489 // Delegate to CMScollector which knows how to coordinate between 2490 // this and any other CMS generations that it is responsible for 2491 // collecting. 2492 collector()->gc_prologue(full); 2493 } 2494 2495 // This is a "private" interface for use by this generation's CMSCollector. 2496 // Not to be called directly by any other entity (for instance, 2497 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2498 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2499 bool registerClosure, ModUnionClosure* modUnionClosure) { 2500 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2501 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2502 "Should be NULL"); 2503 if (registerClosure) { 2504 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2505 } 2506 cmsSpace()->gc_prologue(); 2507 // Clear stat counters 2508 NOT_PRODUCT( 2509 assert(_numObjectsPromoted == 0, "check"); 2510 assert(_numWordsPromoted == 0, "check"); 2511 if (Verbose && PrintGC) { 2512 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2513 SIZE_FORMAT" bytes concurrently", 2514 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2515 } 2516 _numObjectsAllocated = 0; 2517 _numWordsAllocated = 0; 2518 ) 2519 } 2520 2521 void CMSCollector::gc_epilogue(bool full) { 2522 // The following locking discipline assumes that we are only called 2523 // when the world is stopped. 2524 assert(SafepointSynchronize::is_at_safepoint(), 2525 "world is stopped assumption"); 2526 2527 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2528 // if linear allocation blocks need to be appropriately marked to allow the 2529 // the blocks to be parsable. We also check here whether we need to nudge the 2530 // CMS collector thread to start a new cycle (if it's not already active). 2531 assert( Thread::current()->is_VM_thread() 2532 || ( CMSScavengeBeforeRemark 2533 && Thread::current()->is_ConcurrentGC_thread()), 2534 "Incorrect thread type for epilogue execution"); 2535 2536 if (!_between_prologue_and_epilogue) { 2537 // We have already been invoked; this is a gc_epilogue delegation 2538 // from yet another CMS generation that we are responsible for, just 2539 // ignore it since all relevant work has already been done. 2540 return; 2541 } 2542 assert(haveFreelistLocks(), "must have freelist locks"); 2543 assert_lock_strong(bitMapLock()); 2544 2545 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2546 2547 _cmsGen->gc_epilogue_work(full); 2548 2549 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2550 // in case sampling was not already enabled, enable it 2551 _start_sampling = true; 2552 } 2553 // reset _eden_chunk_array so sampling starts afresh 2554 _eden_chunk_index = 0; 2555 2556 size_t cms_used = _cmsGen->cmsSpace()->used(); 2557 2558 // update performance counters - this uses a special version of 2559 // update_counters() that allows the utilization to be passed as a 2560 // parameter, avoiding multiple calls to used(). 2561 // 2562 _cmsGen->update_counters(cms_used); 2563 2564 bitMapLock()->unlock(); 2565 releaseFreelistLocks(); 2566 2567 if (!CleanChunkPoolAsync) { 2568 Chunk::clean_chunk_pool(); 2569 } 2570 2571 set_did_compact(false); 2572 _between_prologue_and_epilogue = false; // ready for next cycle 2573 } 2574 2575 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2576 collector()->gc_epilogue(full); 2577 2578 // Also reset promotion tracking in par gc thread states. 2579 if (CollectedHeap::use_parallel_gc_threads()) { 2580 for (uint i = 0; i < ParallelGCThreads; i++) { 2581 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2582 } 2583 } 2584 } 2585 2586 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2587 assert(!incremental_collection_failed(), "Should have been cleared"); 2588 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2589 cmsSpace()->gc_epilogue(); 2590 // Print stat counters 2591 NOT_PRODUCT( 2592 assert(_numObjectsAllocated == 0, "check"); 2593 assert(_numWordsAllocated == 0, "check"); 2594 if (Verbose && PrintGC) { 2595 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2596 SIZE_FORMAT" bytes", 2597 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2598 } 2599 _numObjectsPromoted = 0; 2600 _numWordsPromoted = 0; 2601 ) 2602 2603 if (PrintGC && Verbose) { 2604 // Call down the chain in contiguous_available needs the freelistLock 2605 // so print this out before releasing the freeListLock. 2606 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2607 contiguous_available()); 2608 } 2609 } 2610 2611 #ifndef PRODUCT 2612 bool CMSCollector::have_cms_token() { 2613 Thread* thr = Thread::current(); 2614 if (thr->is_VM_thread()) { 2615 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2616 } else if (thr->is_ConcurrentGC_thread()) { 2617 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2618 } else if (thr->is_GC_task_thread()) { 2619 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2620 ParGCRareEvent_lock->owned_by_self(); 2621 } 2622 return false; 2623 } 2624 #endif 2625 2626 // Check reachability of the given heap address in CMS generation, 2627 // treating all other generations as roots. 2628 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2629 // We could "guarantee" below, rather than assert, but I'll 2630 // leave these as "asserts" so that an adventurous debugger 2631 // could try this in the product build provided some subset of 2632 // the conditions were met, provided they were interested in the 2633 // results and knew that the computation below wouldn't interfere 2634 // with other concurrent computations mutating the structures 2635 // being read or written. 2636 assert(SafepointSynchronize::is_at_safepoint(), 2637 "Else mutations in object graph will make answer suspect"); 2638 assert(have_cms_token(), "Should hold cms token"); 2639 assert(haveFreelistLocks(), "must hold free list locks"); 2640 assert_lock_strong(bitMapLock()); 2641 2642 // Clear the marking bit map array before starting, but, just 2643 // for kicks, first report if the given address is already marked 2644 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2645 _markBitMap.isMarked(addr) ? "" : " not"); 2646 2647 if (verify_after_remark()) { 2648 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2649 bool result = verification_mark_bm()->isMarked(addr); 2650 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2651 result ? "IS" : "is NOT"); 2652 return result; 2653 } else { 2654 gclog_or_tty->print_cr("Could not compute result"); 2655 return false; 2656 } 2657 } 2658 2659 2660 void 2661 CMSCollector::print_on_error(outputStream* st) { 2662 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2663 if (collector != NULL) { 2664 CMSBitMap* bitmap = &collector->_markBitMap; 2665 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2666 bitmap->print_on_error(st, " Bits: "); 2667 2668 st->cr(); 2669 2670 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2671 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2672 mut_bitmap->print_on_error(st, " Bits: "); 2673 } 2674 } 2675 2676 //////////////////////////////////////////////////////// 2677 // CMS Verification Support 2678 //////////////////////////////////////////////////////// 2679 // Following the remark phase, the following invariant 2680 // should hold -- each object in the CMS heap which is 2681 // marked in markBitMap() should be marked in the verification_mark_bm(). 2682 2683 class VerifyMarkedClosure: public BitMapClosure { 2684 CMSBitMap* _marks; 2685 bool _failed; 2686 2687 public: 2688 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2689 2690 bool do_bit(size_t offset) { 2691 HeapWord* addr = _marks->offsetToHeapWord(offset); 2692 if (!_marks->isMarked(addr)) { 2693 oop(addr)->print_on(gclog_or_tty); 2694 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2695 _failed = true; 2696 } 2697 return true; 2698 } 2699 2700 bool failed() { return _failed; } 2701 }; 2702 2703 bool CMSCollector::verify_after_remark(bool silent) { 2704 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2705 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2706 static bool init = false; 2707 2708 assert(SafepointSynchronize::is_at_safepoint(), 2709 "Else mutations in object graph will make answer suspect"); 2710 assert(have_cms_token(), 2711 "Else there may be mutual interference in use of " 2712 " verification data structures"); 2713 assert(_collectorState > Marking && _collectorState <= Sweeping, 2714 "Else marking info checked here may be obsolete"); 2715 assert(haveFreelistLocks(), "must hold free list locks"); 2716 assert_lock_strong(bitMapLock()); 2717 2718 2719 // Allocate marking bit map if not already allocated 2720 if (!init) { // first time 2721 if (!verification_mark_bm()->allocate(_span)) { 2722 return false; 2723 } 2724 init = true; 2725 } 2726 2727 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2728 2729 // Turn off refs discovery -- so we will be tracing through refs. 2730 // This is as intended, because by this time 2731 // GC must already have cleared any refs that need to be cleared, 2732 // and traced those that need to be marked; moreover, 2733 // the marking done here is not going to interfere in any 2734 // way with the marking information used by GC. 2735 NoRefDiscovery no_discovery(ref_processor()); 2736 2737 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2738 2739 // Clear any marks from a previous round 2740 verification_mark_bm()->clear_all(); 2741 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2742 verify_work_stacks_empty(); 2743 2744 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2745 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2746 // Update the saved marks which may affect the root scans. 2747 gch->save_marks(); 2748 2749 if (CMSRemarkVerifyVariant == 1) { 2750 // In this first variant of verification, we complete 2751 // all marking, then check if the new marks-vector is 2752 // a subset of the CMS marks-vector. 2753 verify_after_remark_work_1(); 2754 } else if (CMSRemarkVerifyVariant == 2) { 2755 // In this second variant of verification, we flag an error 2756 // (i.e. an object reachable in the new marks-vector not reachable 2757 // in the CMS marks-vector) immediately, also indicating the 2758 // identify of an object (A) that references the unmarked object (B) -- 2759 // presumably, a mutation to A failed to be picked up by preclean/remark? 2760 verify_after_remark_work_2(); 2761 } else { 2762 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2763 CMSRemarkVerifyVariant); 2764 } 2765 if (!silent) gclog_or_tty->print(" done] "); 2766 return true; 2767 } 2768 2769 void CMSCollector::verify_after_remark_work_1() { 2770 ResourceMark rm; 2771 HandleMark hm; 2772 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2773 2774 // Get a clear set of claim bits for the roots processing to work with. 2775 ClassLoaderDataGraph::clear_claimed_marks(); 2776 2777 // Mark from roots one level into CMS 2778 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2779 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2780 2781 gch->gen_process_roots(_cmsGen->level(), 2782 true, // younger gens are roots 2783 true, // activate StrongRootsScope 2784 SharedHeap::ScanningOption(roots_scanning_options()), 2785 should_unload_classes(), 2786 ¬Older, 2787 NULL, 2788 NULL); // SSS: Provide correct closure 2789 2790 // Now mark from the roots 2791 MarkFromRootsClosure markFromRootsClosure(this, _span, 2792 verification_mark_bm(), verification_mark_stack(), 2793 false /* don't yield */, true /* verifying */); 2794 assert(_restart_addr == NULL, "Expected pre-condition"); 2795 verification_mark_bm()->iterate(&markFromRootsClosure); 2796 while (_restart_addr != NULL) { 2797 // Deal with stack overflow: by restarting at the indicated 2798 // address. 2799 HeapWord* ra = _restart_addr; 2800 markFromRootsClosure.reset(ra); 2801 _restart_addr = NULL; 2802 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2803 } 2804 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2805 verify_work_stacks_empty(); 2806 2807 // Marking completed -- now verify that each bit marked in 2808 // verification_mark_bm() is also marked in markBitMap(); flag all 2809 // errors by printing corresponding objects. 2810 VerifyMarkedClosure vcl(markBitMap()); 2811 verification_mark_bm()->iterate(&vcl); 2812 if (vcl.failed()) { 2813 gclog_or_tty->print("Verification failed"); 2814 Universe::heap()->print_on(gclog_or_tty); 2815 fatal("CMS: failed marking verification after remark"); 2816 } 2817 } 2818 2819 class VerifyKlassOopsKlassClosure : public KlassClosure { 2820 class VerifyKlassOopsClosure : public OopClosure { 2821 CMSBitMap* _bitmap; 2822 public: 2823 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2824 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2825 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2826 } _oop_closure; 2827 public: 2828 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2829 void do_klass(Klass* k) { 2830 k->oops_do(&_oop_closure); 2831 } 2832 }; 2833 2834 void CMSCollector::verify_after_remark_work_2() { 2835 ResourceMark rm; 2836 HandleMark hm; 2837 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2838 2839 // Get a clear set of claim bits for the roots processing to work with. 2840 ClassLoaderDataGraph::clear_claimed_marks(); 2841 2842 // Mark from roots one level into CMS 2843 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2844 markBitMap()); 2845 CLDToOopClosure cld_closure(¬Older, true); 2846 2847 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2848 2849 gch->gen_process_roots(_cmsGen->level(), 2850 true, // younger gens are roots 2851 true, // activate StrongRootsScope 2852 SharedHeap::ScanningOption(roots_scanning_options()), 2853 should_unload_classes(), 2854 ¬Older, 2855 NULL, 2856 &cld_closure); 2857 2858 // Now mark from the roots 2859 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2860 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2861 assert(_restart_addr == NULL, "Expected pre-condition"); 2862 verification_mark_bm()->iterate(&markFromRootsClosure); 2863 while (_restart_addr != NULL) { 2864 // Deal with stack overflow: by restarting at the indicated 2865 // address. 2866 HeapWord* ra = _restart_addr; 2867 markFromRootsClosure.reset(ra); 2868 _restart_addr = NULL; 2869 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2870 } 2871 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2872 verify_work_stacks_empty(); 2873 2874 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2875 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2876 2877 // Marking completed -- now verify that each bit marked in 2878 // verification_mark_bm() is also marked in markBitMap(); flag all 2879 // errors by printing corresponding objects. 2880 VerifyMarkedClosure vcl(markBitMap()); 2881 verification_mark_bm()->iterate(&vcl); 2882 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2883 } 2884 2885 void ConcurrentMarkSweepGeneration::save_marks() { 2886 // delegate to CMS space 2887 cmsSpace()->save_marks(); 2888 for (uint i = 0; i < ParallelGCThreads; i++) { 2889 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2890 } 2891 } 2892 2893 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2894 return cmsSpace()->no_allocs_since_save_marks(); 2895 } 2896 2897 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2898 \ 2899 void ConcurrentMarkSweepGeneration:: \ 2900 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2901 cl->set_generation(this); \ 2902 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2903 cl->reset_generation(); \ 2904 save_marks(); \ 2905 } 2906 2907 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2908 2909 void 2910 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2911 cl->set_generation(this); 2912 younger_refs_in_space_iterate(_cmsSpace, cl); 2913 cl->reset_generation(); 2914 } 2915 2916 void 2917 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2918 if (freelistLock()->owned_by_self()) { 2919 Generation::oop_iterate(cl); 2920 } else { 2921 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2922 Generation::oop_iterate(cl); 2923 } 2924 } 2925 2926 void 2927 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2928 if (freelistLock()->owned_by_self()) { 2929 Generation::object_iterate(cl); 2930 } else { 2931 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2932 Generation::object_iterate(cl); 2933 } 2934 } 2935 2936 void 2937 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2938 if (freelistLock()->owned_by_self()) { 2939 Generation::safe_object_iterate(cl); 2940 } else { 2941 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2942 Generation::safe_object_iterate(cl); 2943 } 2944 } 2945 2946 void 2947 ConcurrentMarkSweepGeneration::post_compact() { 2948 } 2949 2950 void 2951 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2952 // Fix the linear allocation blocks to look like free blocks. 2953 2954 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2955 // are not called when the heap is verified during universe initialization and 2956 // at vm shutdown. 2957 if (freelistLock()->owned_by_self()) { 2958 cmsSpace()->prepare_for_verify(); 2959 } else { 2960 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2961 cmsSpace()->prepare_for_verify(); 2962 } 2963 } 2964 2965 void 2966 ConcurrentMarkSweepGeneration::verify() { 2967 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2968 // are not called when the heap is verified during universe initialization and 2969 // at vm shutdown. 2970 if (freelistLock()->owned_by_self()) { 2971 cmsSpace()->verify(); 2972 } else { 2973 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2974 cmsSpace()->verify(); 2975 } 2976 } 2977 2978 void CMSCollector::verify() { 2979 _cmsGen->verify(); 2980 } 2981 2982 #ifndef PRODUCT 2983 bool CMSCollector::overflow_list_is_empty() const { 2984 assert(_num_par_pushes >= 0, "Inconsistency"); 2985 if (_overflow_list == NULL) { 2986 assert(_num_par_pushes == 0, "Inconsistency"); 2987 } 2988 return _overflow_list == NULL; 2989 } 2990 2991 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2992 // merely consolidate assertion checks that appear to occur together frequently. 2993 void CMSCollector::verify_work_stacks_empty() const { 2994 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2995 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2996 } 2997 2998 void CMSCollector::verify_overflow_empty() const { 2999 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3000 assert(no_preserved_marks(), "No preserved marks"); 3001 } 3002 #endif // PRODUCT 3003 3004 // Decide if we want to enable class unloading as part of the 3005 // ensuing concurrent GC cycle. We will collect and 3006 // unload classes if it's the case that: 3007 // (1) an explicit gc request has been made and the flag 3008 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3009 // (2) (a) class unloading is enabled at the command line, and 3010 // (b) old gen is getting really full 3011 // NOTE: Provided there is no change in the state of the heap between 3012 // calls to this method, it should have idempotent results. Moreover, 3013 // its results should be monotonically increasing (i.e. going from 0 to 1, 3014 // but not 1 to 0) between successive calls between which the heap was 3015 // not collected. For the implementation below, it must thus rely on 3016 // the property that concurrent_cycles_since_last_unload() 3017 // will not decrease unless a collection cycle happened and that 3018 // _cmsGen->is_too_full() are 3019 // themselves also monotonic in that sense. See check_monotonicity() 3020 // below. 3021 void CMSCollector::update_should_unload_classes() { 3022 _should_unload_classes = false; 3023 // Condition 1 above 3024 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3025 _should_unload_classes = true; 3026 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3027 // Disjuncts 2.b.(i,ii,iii) above 3028 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3029 CMSClassUnloadingMaxInterval) 3030 || _cmsGen->is_too_full(); 3031 } 3032 } 3033 3034 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3035 bool res = should_concurrent_collect(); 3036 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3037 return res; 3038 } 3039 3040 void CMSCollector::setup_cms_unloading_and_verification_state() { 3041 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3042 || VerifyBeforeExit; 3043 const int rso = SharedHeap::SO_AllCodeCache; 3044 3045 // We set the proper root for this CMS cycle here. 3046 if (should_unload_classes()) { // Should unload classes this cycle 3047 remove_root_scanning_option(rso); // Shrink the root set appropriately 3048 set_verifying(should_verify); // Set verification state for this cycle 3049 return; // Nothing else needs to be done at this time 3050 } 3051 3052 // Not unloading classes this cycle 3053 assert(!should_unload_classes(), "Inconsistency!"); 3054 3055 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3056 // Include symbols, strings and code cache elements to prevent their resurrection. 3057 add_root_scanning_option(rso); 3058 set_verifying(true); 3059 } else if (verifying() && !should_verify) { 3060 // We were verifying, but some verification flags got disabled. 3061 set_verifying(false); 3062 // Exclude symbols, strings and code cache elements from root scanning to 3063 // reduce IM and RM pauses. 3064 remove_root_scanning_option(rso); 3065 } 3066 } 3067 3068 3069 #ifndef PRODUCT 3070 HeapWord* CMSCollector::block_start(const void* p) const { 3071 const HeapWord* addr = (HeapWord*)p; 3072 if (_span.contains(p)) { 3073 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3074 return _cmsGen->cmsSpace()->block_start(p); 3075 } 3076 } 3077 return NULL; 3078 } 3079 #endif 3080 3081 HeapWord* 3082 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3083 bool tlab, 3084 bool parallel) { 3085 CMSSynchronousYieldRequest yr; 3086 assert(!tlab, "Can't deal with TLAB allocation"); 3087 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3088 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3089 CMSExpansionCause::_satisfy_allocation); 3090 if (GCExpandToAllocateDelayMillis > 0) { 3091 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3092 } 3093 return have_lock_and_allocate(word_size, tlab); 3094 } 3095 3096 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3097 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3098 // to CardGeneration and share it... 3099 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3100 return CardGeneration::expand(bytes, expand_bytes); 3101 } 3102 3103 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3104 CMSExpansionCause::Cause cause) 3105 { 3106 3107 bool success = expand(bytes, expand_bytes); 3108 3109 // remember why we expanded; this information is used 3110 // by shouldConcurrentCollect() when making decisions on whether to start 3111 // a new CMS cycle. 3112 if (success) { 3113 set_expansion_cause(cause); 3114 if (PrintGCDetails && Verbose) { 3115 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3116 CMSExpansionCause::to_string(cause)); 3117 } 3118 } 3119 } 3120 3121 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3122 HeapWord* res = NULL; 3123 MutexLocker x(ParGCRareEvent_lock); 3124 while (true) { 3125 // Expansion by some other thread might make alloc OK now: 3126 res = ps->lab.alloc(word_sz); 3127 if (res != NULL) return res; 3128 // If there's not enough expansion space available, give up. 3129 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3130 return NULL; 3131 } 3132 // Otherwise, we try expansion. 3133 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3134 CMSExpansionCause::_allocate_par_lab); 3135 // Now go around the loop and try alloc again; 3136 // A competing par_promote might beat us to the expansion space, 3137 // so we may go around the loop again if promotion fails again. 3138 if (GCExpandToAllocateDelayMillis > 0) { 3139 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3140 } 3141 } 3142 } 3143 3144 3145 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3146 PromotionInfo* promo) { 3147 MutexLocker x(ParGCRareEvent_lock); 3148 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3149 while (true) { 3150 // Expansion by some other thread might make alloc OK now: 3151 if (promo->ensure_spooling_space()) { 3152 assert(promo->has_spooling_space(), 3153 "Post-condition of successful ensure_spooling_space()"); 3154 return true; 3155 } 3156 // If there's not enough expansion space available, give up. 3157 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3158 return false; 3159 } 3160 // Otherwise, we try expansion. 3161 expand(refill_size_bytes, MinHeapDeltaBytes, 3162 CMSExpansionCause::_allocate_par_spooling_space); 3163 // Now go around the loop and try alloc again; 3164 // A competing allocation might beat us to the expansion space, 3165 // so we may go around the loop again if allocation fails again. 3166 if (GCExpandToAllocateDelayMillis > 0) { 3167 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3168 } 3169 } 3170 } 3171 3172 3173 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3174 assert_locked_or_safepoint(ExpandHeap_lock); 3175 // Shrink committed space 3176 _virtual_space.shrink_by(bytes); 3177 // Shrink space; this also shrinks the space's BOT 3178 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3179 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3180 // Shrink the shared block offset array 3181 _bts->resize(new_word_size); 3182 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3183 // Shrink the card table 3184 Universe::heap()->barrier_set()->resize_covered_region(mr); 3185 3186 if (Verbose && PrintGC) { 3187 size_t new_mem_size = _virtual_space.committed_size(); 3188 size_t old_mem_size = new_mem_size + bytes; 3189 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3190 name(), old_mem_size/K, new_mem_size/K); 3191 } 3192 } 3193 3194 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3195 assert_locked_or_safepoint(Heap_lock); 3196 size_t size = ReservedSpace::page_align_size_down(bytes); 3197 // Only shrink if a compaction was done so that all the free space 3198 // in the generation is in a contiguous block at the end. 3199 if (size > 0 && did_compact()) { 3200 shrink_by(size); 3201 } 3202 } 3203 3204 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3205 assert_locked_or_safepoint(Heap_lock); 3206 bool result = _virtual_space.expand_by(bytes); 3207 if (result) { 3208 size_t new_word_size = 3209 heap_word_size(_virtual_space.committed_size()); 3210 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3211 _bts->resize(new_word_size); // resize the block offset shared array 3212 Universe::heap()->barrier_set()->resize_covered_region(mr); 3213 // Hmmmm... why doesn't CFLS::set_end verify locking? 3214 // This is quite ugly; FIX ME XXX 3215 _cmsSpace->assert_locked(freelistLock()); 3216 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3217 3218 // update the space and generation capacity counters 3219 if (UsePerfData) { 3220 _space_counters->update_capacity(); 3221 _gen_counters->update_all(); 3222 } 3223 3224 if (Verbose && PrintGC) { 3225 size_t new_mem_size = _virtual_space.committed_size(); 3226 size_t old_mem_size = new_mem_size - bytes; 3227 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3228 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3229 } 3230 } 3231 return result; 3232 } 3233 3234 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3235 assert_locked_or_safepoint(Heap_lock); 3236 bool success = true; 3237 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3238 if (remaining_bytes > 0) { 3239 success = grow_by(remaining_bytes); 3240 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3241 } 3242 return success; 3243 } 3244 3245 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3246 assert_locked_or_safepoint(Heap_lock); 3247 assert_lock_strong(freelistLock()); 3248 if (PrintGCDetails && Verbose) { 3249 warning("Shrinking of CMS not yet implemented"); 3250 } 3251 return; 3252 } 3253 3254 3255 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3256 // phases. 3257 class CMSPhaseAccounting: public StackObj { 3258 public: 3259 CMSPhaseAccounting(CMSCollector *collector, 3260 const char *phase, 3261 const GCId gc_id, 3262 bool print_cr = true); 3263 ~CMSPhaseAccounting(); 3264 3265 private: 3266 CMSCollector *_collector; 3267 const char *_phase; 3268 elapsedTimer _wallclock; 3269 bool _print_cr; 3270 const GCId _gc_id; 3271 3272 public: 3273 // Not MT-safe; so do not pass around these StackObj's 3274 // where they may be accessed by other threads. 3275 jlong wallclock_millis() { 3276 assert(_wallclock.is_active(), "Wall clock should not stop"); 3277 _wallclock.stop(); // to record time 3278 jlong ret = _wallclock.milliseconds(); 3279 _wallclock.start(); // restart 3280 return ret; 3281 } 3282 }; 3283 3284 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3285 const char *phase, 3286 const GCId gc_id, 3287 bool print_cr) : 3288 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 3289 3290 if (PrintCMSStatistics != 0) { 3291 _collector->resetYields(); 3292 } 3293 if (PrintGCDetails) { 3294 gclog_or_tty->gclog_stamp(_gc_id); 3295 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3296 _collector->cmsGen()->short_name(), _phase); 3297 } 3298 _collector->resetTimer(); 3299 _wallclock.start(); 3300 _collector->startTimer(); 3301 } 3302 3303 CMSPhaseAccounting::~CMSPhaseAccounting() { 3304 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3305 _collector->stopTimer(); 3306 _wallclock.stop(); 3307 if (PrintGCDetails) { 3308 gclog_or_tty->gclog_stamp(_gc_id); 3309 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3310 _collector->cmsGen()->short_name(), 3311 _phase, _collector->timerValue(), _wallclock.seconds()); 3312 if (_print_cr) { 3313 gclog_or_tty->cr(); 3314 } 3315 if (PrintCMSStatistics != 0) { 3316 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3317 _collector->yields()); 3318 } 3319 } 3320 } 3321 3322 // CMS work 3323 3324 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3325 class CMSParMarkTask : public AbstractGangTask { 3326 protected: 3327 CMSCollector* _collector; 3328 int _n_workers; 3329 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3330 AbstractGangTask(name), 3331 _collector(collector), 3332 _n_workers(n_workers) {} 3333 // Work method in support of parallel rescan ... of young gen spaces 3334 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3335 ContiguousSpace* space, 3336 HeapWord** chunk_array, size_t chunk_top); 3337 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3338 }; 3339 3340 // Parallel initial mark task 3341 class CMSParInitialMarkTask: public CMSParMarkTask { 3342 public: 3343 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3344 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3345 collector, n_workers) {} 3346 void work(uint worker_id); 3347 }; 3348 3349 // Checkpoint the roots into this generation from outside 3350 // this generation. [Note this initial checkpoint need only 3351 // be approximate -- we'll do a catch up phase subsequently.] 3352 void CMSCollector::checkpointRootsInitial(bool asynch) { 3353 assert(_collectorState == InitialMarking, "Wrong collector state"); 3354 check_correct_thread_executing(); 3355 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3356 3357 save_heap_summary(); 3358 report_heap_summary(GCWhen::BeforeGC); 3359 3360 ReferenceProcessor* rp = ref_processor(); 3361 SpecializationStats::clear(); 3362 assert(_restart_addr == NULL, "Control point invariant"); 3363 if (asynch) { 3364 // acquire locks for subsequent manipulations 3365 MutexLockerEx x(bitMapLock(), 3366 Mutex::_no_safepoint_check_flag); 3367 checkpointRootsInitialWork(asynch); 3368 // enable ("weak") refs discovery 3369 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3370 _collectorState = Marking; 3371 } else { 3372 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3373 // which recognizes if we are a CMS generation, and doesn't try to turn on 3374 // discovery; verify that they aren't meddling. 3375 assert(!rp->discovery_is_atomic(), 3376 "incorrect setting of discovery predicate"); 3377 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3378 "ref discovery for this generation kind"); 3379 // already have locks 3380 checkpointRootsInitialWork(asynch); 3381 // now enable ("weak") refs discovery 3382 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3383 _collectorState = Marking; 3384 } 3385 SpecializationStats::print(); 3386 } 3387 3388 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3389 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3390 assert(_collectorState == InitialMarking, "just checking"); 3391 3392 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3393 // precede our marking with a collection of all 3394 // younger generations to keep floating garbage to a minimum. 3395 // XXX: we won't do this for now -- it's an optimization to be done later. 3396 3397 // already have locks 3398 assert_lock_strong(bitMapLock()); 3399 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3400 3401 // Setup the verification and class unloading state for this 3402 // CMS collection cycle. 3403 setup_cms_unloading_and_verification_state(); 3404 3405 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3406 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3407 3408 // Reset all the PLAB chunk arrays if necessary. 3409 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3410 reset_survivor_plab_arrays(); 3411 } 3412 3413 ResourceMark rm; 3414 HandleMark hm; 3415 3416 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3417 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3418 3419 verify_work_stacks_empty(); 3420 verify_overflow_empty(); 3421 3422 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3423 // Update the saved marks which may affect the root scans. 3424 gch->save_marks(); 3425 3426 // weak reference processing has not started yet. 3427 ref_processor()->set_enqueuing_is_done(false); 3428 3429 // Need to remember all newly created CLDs, 3430 // so that we can guarantee that the remark finds them. 3431 ClassLoaderDataGraph::remember_new_clds(true); 3432 3433 // Whenever a CLD is found, it will be claimed before proceeding to mark 3434 // the klasses. The claimed marks need to be cleared before marking starts. 3435 ClassLoaderDataGraph::clear_claimed_marks(); 3436 3437 if (CMSPrintEdenSurvivorChunks) { 3438 print_eden_and_survivor_chunk_arrays(); 3439 } 3440 3441 { 3442 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3443 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3444 // The parallel version. 3445 FlexibleWorkGang* workers = gch->workers(); 3446 assert(workers != NULL, "Need parallel worker threads."); 3447 int n_workers = workers->active_workers(); 3448 CMSParInitialMarkTask tsk(this, n_workers); 3449 gch->set_par_threads(n_workers); 3450 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3451 if (n_workers > 1) { 3452 GenCollectedHeap::StrongRootsScope srs(gch); 3453 workers->run_task(&tsk); 3454 } else { 3455 GenCollectedHeap::StrongRootsScope srs(gch); 3456 tsk.work(0); 3457 } 3458 gch->set_par_threads(0); 3459 } else { 3460 // The serial version. 3461 CLDToOopClosure cld_closure(¬Older, true); 3462 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3463 gch->gen_process_roots(_cmsGen->level(), 3464 true, // younger gens are roots 3465 true, // activate StrongRootsScope 3466 SharedHeap::ScanningOption(roots_scanning_options()), 3467 should_unload_classes(), 3468 ¬Older, 3469 NULL, 3470 &cld_closure); 3471 } 3472 } 3473 3474 // Clear mod-union table; it will be dirtied in the prologue of 3475 // CMS generation per each younger generation collection. 3476 3477 assert(_modUnionTable.isAllClear(), 3478 "Was cleared in most recent final checkpoint phase" 3479 " or no bits are set in the gc_prologue before the start of the next " 3480 "subsequent marking phase."); 3481 3482 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3483 3484 // Save the end of the used_region of the constituent generations 3485 // to be used to limit the extent of sweep in each generation. 3486 save_sweep_limits(); 3487 verify_overflow_empty(); 3488 } 3489 3490 bool CMSCollector::markFromRoots(bool asynch) { 3491 // we might be tempted to assert that: 3492 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3493 // "inconsistent argument?"); 3494 // However that wouldn't be right, because it's possible that 3495 // a safepoint is indeed in progress as a younger generation 3496 // stop-the-world GC happens even as we mark in this generation. 3497 assert(_collectorState == Marking, "inconsistent state?"); 3498 check_correct_thread_executing(); 3499 verify_overflow_empty(); 3500 3501 bool res; 3502 if (asynch) { 3503 // Weak ref discovery note: We may be discovering weak 3504 // refs in this generation concurrent (but interleaved) with 3505 // weak ref discovery by a younger generation collector. 3506 3507 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3508 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3509 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3510 res = markFromRootsWork(asynch); 3511 if (res) { 3512 _collectorState = Precleaning; 3513 } else { // We failed and a foreground collection wants to take over 3514 assert(_foregroundGCIsActive, "internal state inconsistency"); 3515 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3516 if (PrintGCDetails) { 3517 gclog_or_tty->print_cr("bailing out to foreground collection"); 3518 } 3519 } 3520 } else { 3521 assert(SafepointSynchronize::is_at_safepoint(), 3522 "inconsistent with asynch == false"); 3523 // already have locks 3524 res = markFromRootsWork(asynch); 3525 _collectorState = FinalMarking; 3526 } 3527 verify_overflow_empty(); 3528 return res; 3529 } 3530 3531 bool CMSCollector::markFromRootsWork(bool asynch) { 3532 // iterate over marked bits in bit map, doing a full scan and mark 3533 // from these roots using the following algorithm: 3534 // . if oop is to the right of the current scan pointer, 3535 // mark corresponding bit (we'll process it later) 3536 // . else (oop is to left of current scan pointer) 3537 // push oop on marking stack 3538 // . drain the marking stack 3539 3540 // Note that when we do a marking step we need to hold the 3541 // bit map lock -- recall that direct allocation (by mutators) 3542 // and promotion (by younger generation collectors) is also 3543 // marking the bit map. [the so-called allocate live policy.] 3544 // Because the implementation of bit map marking is not 3545 // robust wrt simultaneous marking of bits in the same word, 3546 // we need to make sure that there is no such interference 3547 // between concurrent such updates. 3548 3549 // already have locks 3550 assert_lock_strong(bitMapLock()); 3551 3552 verify_work_stacks_empty(); 3553 verify_overflow_empty(); 3554 bool result = false; 3555 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3556 result = do_marking_mt(asynch); 3557 } else { 3558 result = do_marking_st(asynch); 3559 } 3560 return result; 3561 } 3562 3563 // Forward decl 3564 class CMSConcMarkingTask; 3565 3566 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3567 CMSCollector* _collector; 3568 CMSConcMarkingTask* _task; 3569 public: 3570 virtual void yield(); 3571 3572 // "n_threads" is the number of threads to be terminated. 3573 // "queue_set" is a set of work queues of other threads. 3574 // "collector" is the CMS collector associated with this task terminator. 3575 // "yield" indicates whether we need the gang as a whole to yield. 3576 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3577 ParallelTaskTerminator(n_threads, queue_set), 3578 _collector(collector) { } 3579 3580 void set_task(CMSConcMarkingTask* task) { 3581 _task = task; 3582 } 3583 }; 3584 3585 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3586 CMSConcMarkingTask* _task; 3587 public: 3588 bool should_exit_termination(); 3589 void set_task(CMSConcMarkingTask* task) { 3590 _task = task; 3591 } 3592 }; 3593 3594 // MT Concurrent Marking Task 3595 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3596 CMSCollector* _collector; 3597 int _n_workers; // requested/desired # workers 3598 bool _asynch; 3599 bool _result; 3600 CompactibleFreeListSpace* _cms_space; 3601 char _pad_front[64]; // padding to ... 3602 HeapWord* _global_finger; // ... avoid sharing cache line 3603 char _pad_back[64]; 3604 HeapWord* _restart_addr; 3605 3606 // Exposed here for yielding support 3607 Mutex* const _bit_map_lock; 3608 3609 // The per thread work queues, available here for stealing 3610 OopTaskQueueSet* _task_queues; 3611 3612 // Termination (and yielding) support 3613 CMSConcMarkingTerminator _term; 3614 CMSConcMarkingTerminatorTerminator _term_term; 3615 3616 public: 3617 CMSConcMarkingTask(CMSCollector* collector, 3618 CompactibleFreeListSpace* cms_space, 3619 bool asynch, 3620 YieldingFlexibleWorkGang* workers, 3621 OopTaskQueueSet* task_queues): 3622 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3623 _collector(collector), 3624 _cms_space(cms_space), 3625 _asynch(asynch), _n_workers(0), _result(true), 3626 _task_queues(task_queues), 3627 _term(_n_workers, task_queues, _collector), 3628 _bit_map_lock(collector->bitMapLock()) 3629 { 3630 _requested_size = _n_workers; 3631 _term.set_task(this); 3632 _term_term.set_task(this); 3633 _restart_addr = _global_finger = _cms_space->bottom(); 3634 } 3635 3636 3637 OopTaskQueueSet* task_queues() { return _task_queues; } 3638 3639 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3640 3641 HeapWord** global_finger_addr() { return &_global_finger; } 3642 3643 CMSConcMarkingTerminator* terminator() { return &_term; } 3644 3645 virtual void set_for_termination(int active_workers) { 3646 terminator()->reset_for_reuse(active_workers); 3647 } 3648 3649 void work(uint worker_id); 3650 bool should_yield() { 3651 return ConcurrentMarkSweepThread::should_yield() 3652 && !_collector->foregroundGCIsActive() 3653 && _asynch; 3654 } 3655 3656 virtual void coordinator_yield(); // stuff done by coordinator 3657 bool result() { return _result; } 3658 3659 void reset(HeapWord* ra) { 3660 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3661 _restart_addr = _global_finger = ra; 3662 _term.reset_for_reuse(); 3663 } 3664 3665 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3666 OopTaskQueue* work_q); 3667 3668 private: 3669 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3670 void do_work_steal(int i); 3671 void bump_global_finger(HeapWord* f); 3672 }; 3673 3674 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3675 assert(_task != NULL, "Error"); 3676 return _task->yielding(); 3677 // Note that we do not need the disjunct || _task->should_yield() above 3678 // because we want terminating threads to yield only if the task 3679 // is already in the midst of yielding, which happens only after at least one 3680 // thread has yielded. 3681 } 3682 3683 void CMSConcMarkingTerminator::yield() { 3684 if (_task->should_yield()) { 3685 _task->yield(); 3686 } else { 3687 ParallelTaskTerminator::yield(); 3688 } 3689 } 3690 3691 //////////////////////////////////////////////////////////////// 3692 // Concurrent Marking Algorithm Sketch 3693 //////////////////////////////////////////////////////////////// 3694 // Until all tasks exhausted (both spaces): 3695 // -- claim next available chunk 3696 // -- bump global finger via CAS 3697 // -- find first object that starts in this chunk 3698 // and start scanning bitmap from that position 3699 // -- scan marked objects for oops 3700 // -- CAS-mark target, and if successful: 3701 // . if target oop is above global finger (volatile read) 3702 // nothing to do 3703 // . if target oop is in chunk and above local finger 3704 // then nothing to do 3705 // . else push on work-queue 3706 // -- Deal with possible overflow issues: 3707 // . local work-queue overflow causes stuff to be pushed on 3708 // global (common) overflow queue 3709 // . always first empty local work queue 3710 // . then get a batch of oops from global work queue if any 3711 // . then do work stealing 3712 // -- When all tasks claimed (both spaces) 3713 // and local work queue empty, 3714 // then in a loop do: 3715 // . check global overflow stack; steal a batch of oops and trace 3716 // . try to steal from other threads oif GOS is empty 3717 // . if neither is available, offer termination 3718 // -- Terminate and return result 3719 // 3720 void CMSConcMarkingTask::work(uint worker_id) { 3721 elapsedTimer _timer; 3722 ResourceMark rm; 3723 HandleMark hm; 3724 3725 DEBUG_ONLY(_collector->verify_overflow_empty();) 3726 3727 // Before we begin work, our work queue should be empty 3728 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3729 // Scan the bitmap covering _cms_space, tracing through grey objects. 3730 _timer.start(); 3731 do_scan_and_mark(worker_id, _cms_space); 3732 _timer.stop(); 3733 if (PrintCMSStatistics != 0) { 3734 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3735 worker_id, _timer.seconds()); 3736 // XXX: need xxx/xxx type of notation, two timers 3737 } 3738 3739 // ... do work stealing 3740 _timer.reset(); 3741 _timer.start(); 3742 do_work_steal(worker_id); 3743 _timer.stop(); 3744 if (PrintCMSStatistics != 0) { 3745 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3746 worker_id, _timer.seconds()); 3747 // XXX: need xxx/xxx type of notation, two timers 3748 } 3749 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3750 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3751 // Note that under the current task protocol, the 3752 // following assertion is true even of the spaces 3753 // expanded since the completion of the concurrent 3754 // marking. XXX This will likely change under a strict 3755 // ABORT semantics. 3756 // After perm removal the comparison was changed to 3757 // greater than or equal to from strictly greater than. 3758 // Before perm removal the highest address sweep would 3759 // have been at the end of perm gen but now is at the 3760 // end of the tenured gen. 3761 assert(_global_finger >= _cms_space->end(), 3762 "All tasks have been completed"); 3763 DEBUG_ONLY(_collector->verify_overflow_empty();) 3764 } 3765 3766 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3767 HeapWord* read = _global_finger; 3768 HeapWord* cur = read; 3769 while (f > read) { 3770 cur = read; 3771 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3772 if (cur == read) { 3773 // our cas succeeded 3774 assert(_global_finger >= f, "protocol consistency"); 3775 break; 3776 } 3777 } 3778 } 3779 3780 // This is really inefficient, and should be redone by 3781 // using (not yet available) block-read and -write interfaces to the 3782 // stack and the work_queue. XXX FIX ME !!! 3783 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3784 OopTaskQueue* work_q) { 3785 // Fast lock-free check 3786 if (ovflw_stk->length() == 0) { 3787 return false; 3788 } 3789 assert(work_q->size() == 0, "Shouldn't steal"); 3790 MutexLockerEx ml(ovflw_stk->par_lock(), 3791 Mutex::_no_safepoint_check_flag); 3792 // Grab up to 1/4 the size of the work queue 3793 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3794 (size_t)ParGCDesiredObjsFromOverflowList); 3795 num = MIN2(num, ovflw_stk->length()); 3796 for (int i = (int) num; i > 0; i--) { 3797 oop cur = ovflw_stk->pop(); 3798 assert(cur != NULL, "Counted wrong?"); 3799 work_q->push(cur); 3800 } 3801 return num > 0; 3802 } 3803 3804 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3805 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3806 int n_tasks = pst->n_tasks(); 3807 // We allow that there may be no tasks to do here because 3808 // we are restarting after a stack overflow. 3809 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3810 uint nth_task = 0; 3811 3812 HeapWord* aligned_start = sp->bottom(); 3813 if (sp->used_region().contains(_restart_addr)) { 3814 // Align down to a card boundary for the start of 0th task 3815 // for this space. 3816 aligned_start = 3817 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3818 CardTableModRefBS::card_size); 3819 } 3820 3821 size_t chunk_size = sp->marking_task_size(); 3822 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3823 // Having claimed the nth task in this space, 3824 // compute the chunk that it corresponds to: 3825 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3826 aligned_start + (nth_task+1)*chunk_size); 3827 // Try and bump the global finger via a CAS; 3828 // note that we need to do the global finger bump 3829 // _before_ taking the intersection below, because 3830 // the task corresponding to that region will be 3831 // deemed done even if the used_region() expands 3832 // because of allocation -- as it almost certainly will 3833 // during start-up while the threads yield in the 3834 // closure below. 3835 HeapWord* finger = span.end(); 3836 bump_global_finger(finger); // atomically 3837 // There are null tasks here corresponding to chunks 3838 // beyond the "top" address of the space. 3839 span = span.intersection(sp->used_region()); 3840 if (!span.is_empty()) { // Non-null task 3841 HeapWord* prev_obj; 3842 assert(!span.contains(_restart_addr) || nth_task == 0, 3843 "Inconsistency"); 3844 if (nth_task == 0) { 3845 // For the 0th task, we'll not need to compute a block_start. 3846 if (span.contains(_restart_addr)) { 3847 // In the case of a restart because of stack overflow, 3848 // we might additionally skip a chunk prefix. 3849 prev_obj = _restart_addr; 3850 } else { 3851 prev_obj = span.start(); 3852 } 3853 } else { 3854 // We want to skip the first object because 3855 // the protocol is to scan any object in its entirety 3856 // that _starts_ in this span; a fortiori, any 3857 // object starting in an earlier span is scanned 3858 // as part of an earlier claimed task. 3859 // Below we use the "careful" version of block_start 3860 // so we do not try to navigate uninitialized objects. 3861 prev_obj = sp->block_start_careful(span.start()); 3862 // Below we use a variant of block_size that uses the 3863 // Printezis bits to avoid waiting for allocated 3864 // objects to become initialized/parsable. 3865 while (prev_obj < span.start()) { 3866 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3867 if (sz > 0) { 3868 prev_obj += sz; 3869 } else { 3870 // In this case we may end up doing a bit of redundant 3871 // scanning, but that appears unavoidable, short of 3872 // locking the free list locks; see bug 6324141. 3873 break; 3874 } 3875 } 3876 } 3877 if (prev_obj < span.end()) { 3878 MemRegion my_span = MemRegion(prev_obj, span.end()); 3879 // Do the marking work within a non-empty span -- 3880 // the last argument to the constructor indicates whether the 3881 // iteration should be incremental with periodic yields. 3882 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3883 &_collector->_markBitMap, 3884 work_queue(i), 3885 &_collector->_markStack, 3886 _asynch); 3887 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3888 } // else nothing to do for this task 3889 } // else nothing to do for this task 3890 } 3891 // We'd be tempted to assert here that since there are no 3892 // more tasks left to claim in this space, the global_finger 3893 // must exceed space->top() and a fortiori space->end(). However, 3894 // that would not quite be correct because the bumping of 3895 // global_finger occurs strictly after the claiming of a task, 3896 // so by the time we reach here the global finger may not yet 3897 // have been bumped up by the thread that claimed the last 3898 // task. 3899 pst->all_tasks_completed(); 3900 } 3901 3902 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3903 private: 3904 CMSCollector* _collector; 3905 CMSConcMarkingTask* _task; 3906 MemRegion _span; 3907 CMSBitMap* _bit_map; 3908 CMSMarkStack* _overflow_stack; 3909 OopTaskQueue* _work_queue; 3910 protected: 3911 DO_OOP_WORK_DEFN 3912 public: 3913 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3914 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3915 MetadataAwareOopClosure(collector->ref_processor()), 3916 _collector(collector), 3917 _task(task), 3918 _span(collector->_span), 3919 _work_queue(work_queue), 3920 _bit_map(bit_map), 3921 _overflow_stack(overflow_stack) 3922 { } 3923 virtual void do_oop(oop* p); 3924 virtual void do_oop(narrowOop* p); 3925 3926 void trim_queue(size_t max); 3927 void handle_stack_overflow(HeapWord* lost); 3928 void do_yield_check() { 3929 if (_task->should_yield()) { 3930 _task->yield(); 3931 } 3932 } 3933 }; 3934 3935 // Grey object scanning during work stealing phase -- 3936 // the salient assumption here is that any references 3937 // that are in these stolen objects being scanned must 3938 // already have been initialized (else they would not have 3939 // been published), so we do not need to check for 3940 // uninitialized objects before pushing here. 3941 void Par_ConcMarkingClosure::do_oop(oop obj) { 3942 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3943 HeapWord* addr = (HeapWord*)obj; 3944 // Check if oop points into the CMS generation 3945 // and is not marked 3946 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3947 // a white object ... 3948 // If we manage to "claim" the object, by being the 3949 // first thread to mark it, then we push it on our 3950 // marking stack 3951 if (_bit_map->par_mark(addr)) { // ... now grey 3952 // push on work queue (grey set) 3953 bool simulate_overflow = false; 3954 NOT_PRODUCT( 3955 if (CMSMarkStackOverflowALot && 3956 _collector->simulate_overflow()) { 3957 // simulate a stack overflow 3958 simulate_overflow = true; 3959 } 3960 ) 3961 if (simulate_overflow || 3962 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3963 // stack overflow 3964 if (PrintCMSStatistics != 0) { 3965 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3966 SIZE_FORMAT, _overflow_stack->capacity()); 3967 } 3968 // We cannot assert that the overflow stack is full because 3969 // it may have been emptied since. 3970 assert(simulate_overflow || 3971 _work_queue->size() == _work_queue->max_elems(), 3972 "Else push should have succeeded"); 3973 handle_stack_overflow(addr); 3974 } 3975 } // Else, some other thread got there first 3976 do_yield_check(); 3977 } 3978 } 3979 3980 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3981 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3982 3983 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3984 while (_work_queue->size() > max) { 3985 oop new_oop; 3986 if (_work_queue->pop_local(new_oop)) { 3987 assert(new_oop->is_oop(), "Should be an oop"); 3988 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3989 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3990 new_oop->oop_iterate(this); // do_oop() above 3991 do_yield_check(); 3992 } 3993 } 3994 } 3995 3996 // Upon stack overflow, we discard (part of) the stack, 3997 // remembering the least address amongst those discarded 3998 // in CMSCollector's _restart_address. 3999 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4000 // We need to do this under a mutex to prevent other 4001 // workers from interfering with the work done below. 4002 MutexLockerEx ml(_overflow_stack->par_lock(), 4003 Mutex::_no_safepoint_check_flag); 4004 // Remember the least grey address discarded 4005 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4006 _collector->lower_restart_addr(ra); 4007 _overflow_stack->reset(); // discard stack contents 4008 _overflow_stack->expand(); // expand the stack if possible 4009 } 4010 4011 4012 void CMSConcMarkingTask::do_work_steal(int i) { 4013 OopTaskQueue* work_q = work_queue(i); 4014 oop obj_to_scan; 4015 CMSBitMap* bm = &(_collector->_markBitMap); 4016 CMSMarkStack* ovflw = &(_collector->_markStack); 4017 int* seed = _collector->hash_seed(i); 4018 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4019 while (true) { 4020 cl.trim_queue(0); 4021 assert(work_q->size() == 0, "Should have been emptied above"); 4022 if (get_work_from_overflow_stack(ovflw, work_q)) { 4023 // Can't assert below because the work obtained from the 4024 // overflow stack may already have been stolen from us. 4025 // assert(work_q->size() > 0, "Work from overflow stack"); 4026 continue; 4027 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4028 assert(obj_to_scan->is_oop(), "Should be an oop"); 4029 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4030 obj_to_scan->oop_iterate(&cl); 4031 } else if (terminator()->offer_termination(&_term_term)) { 4032 assert(work_q->size() == 0, "Impossible!"); 4033 break; 4034 } else if (yielding() || should_yield()) { 4035 yield(); 4036 } 4037 } 4038 } 4039 4040 // This is run by the CMS (coordinator) thread. 4041 void CMSConcMarkingTask::coordinator_yield() { 4042 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4043 "CMS thread should hold CMS token"); 4044 // First give up the locks, then yield, then re-lock 4045 // We should probably use a constructor/destructor idiom to 4046 // do this unlock/lock or modify the MutexUnlocker class to 4047 // serve our purpose. XXX 4048 assert_lock_strong(_bit_map_lock); 4049 _bit_map_lock->unlock(); 4050 ConcurrentMarkSweepThread::desynchronize(true); 4051 _collector->stopTimer(); 4052 if (PrintCMSStatistics != 0) { 4053 _collector->incrementYields(); 4054 } 4055 4056 // It is possible for whichever thread initiated the yield request 4057 // not to get a chance to wake up and take the bitmap lock between 4058 // this thread releasing it and reacquiring it. So, while the 4059 // should_yield() flag is on, let's sleep for a bit to give the 4060 // other thread a chance to wake up. The limit imposed on the number 4061 // of iterations is defensive, to avoid any unforseen circumstances 4062 // putting us into an infinite loop. Since it's always been this 4063 // (coordinator_yield()) method that was observed to cause the 4064 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4065 // which is by default non-zero. For the other seven methods that 4066 // also perform the yield operation, as are using a different 4067 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4068 // can enable the sleeping for those methods too, if necessary. 4069 // See 6442774. 4070 // 4071 // We really need to reconsider the synchronization between the GC 4072 // thread and the yield-requesting threads in the future and we 4073 // should really use wait/notify, which is the recommended 4074 // way of doing this type of interaction. Additionally, we should 4075 // consolidate the eight methods that do the yield operation and they 4076 // are almost identical into one for better maintainability and 4077 // readability. See 6445193. 4078 // 4079 // Tony 2006.06.29 4080 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4081 ConcurrentMarkSweepThread::should_yield() && 4082 !CMSCollector::foregroundGCIsActive(); ++i) { 4083 os::sleep(Thread::current(), 1, false); 4084 } 4085 4086 ConcurrentMarkSweepThread::synchronize(true); 4087 _bit_map_lock->lock_without_safepoint_check(); 4088 _collector->startTimer(); 4089 } 4090 4091 bool CMSCollector::do_marking_mt(bool asynch) { 4092 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4093 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4094 conc_workers()->total_workers(), 4095 conc_workers()->active_workers(), 4096 Threads::number_of_non_daemon_threads()); 4097 conc_workers()->set_active_workers(num_workers); 4098 4099 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4100 4101 CMSConcMarkingTask tsk(this, 4102 cms_space, 4103 asynch, 4104 conc_workers(), 4105 task_queues()); 4106 4107 // Since the actual number of workers we get may be different 4108 // from the number we requested above, do we need to do anything different 4109 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4110 // class?? XXX 4111 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4112 4113 // Refs discovery is already non-atomic. 4114 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4115 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4116 conc_workers()->start_task(&tsk); 4117 while (tsk.yielded()) { 4118 tsk.coordinator_yield(); 4119 conc_workers()->continue_task(&tsk); 4120 } 4121 // If the task was aborted, _restart_addr will be non-NULL 4122 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4123 while (_restart_addr != NULL) { 4124 // XXX For now we do not make use of ABORTED state and have not 4125 // yet implemented the right abort semantics (even in the original 4126 // single-threaded CMS case). That needs some more investigation 4127 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4128 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4129 // If _restart_addr is non-NULL, a marking stack overflow 4130 // occurred; we need to do a fresh marking iteration from the 4131 // indicated restart address. 4132 if (_foregroundGCIsActive && asynch) { 4133 // We may be running into repeated stack overflows, having 4134 // reached the limit of the stack size, while making very 4135 // slow forward progress. It may be best to bail out and 4136 // let the foreground collector do its job. 4137 // Clear _restart_addr, so that foreground GC 4138 // works from scratch. This avoids the headache of 4139 // a "rescan" which would otherwise be needed because 4140 // of the dirty mod union table & card table. 4141 _restart_addr = NULL; 4142 return false; 4143 } 4144 // Adjust the task to restart from _restart_addr 4145 tsk.reset(_restart_addr); 4146 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4147 _restart_addr); 4148 _restart_addr = NULL; 4149 // Get the workers going again 4150 conc_workers()->start_task(&tsk); 4151 while (tsk.yielded()) { 4152 tsk.coordinator_yield(); 4153 conc_workers()->continue_task(&tsk); 4154 } 4155 } 4156 assert(tsk.completed(), "Inconsistency"); 4157 assert(tsk.result() == true, "Inconsistency"); 4158 return true; 4159 } 4160 4161 bool CMSCollector::do_marking_st(bool asynch) { 4162 ResourceMark rm; 4163 HandleMark hm; 4164 4165 // Temporarily make refs discovery single threaded (non-MT) 4166 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4167 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4168 &_markStack, CMSYield && asynch); 4169 // the last argument to iterate indicates whether the iteration 4170 // should be incremental with periodic yields. 4171 _markBitMap.iterate(&markFromRootsClosure); 4172 // If _restart_addr is non-NULL, a marking stack overflow 4173 // occurred; we need to do a fresh iteration from the 4174 // indicated restart address. 4175 while (_restart_addr != NULL) { 4176 if (_foregroundGCIsActive && asynch) { 4177 // We may be running into repeated stack overflows, having 4178 // reached the limit of the stack size, while making very 4179 // slow forward progress. It may be best to bail out and 4180 // let the foreground collector do its job. 4181 // Clear _restart_addr, so that foreground GC 4182 // works from scratch. This avoids the headache of 4183 // a "rescan" which would otherwise be needed because 4184 // of the dirty mod union table & card table. 4185 _restart_addr = NULL; 4186 return false; // indicating failure to complete marking 4187 } 4188 // Deal with stack overflow: 4189 // we restart marking from _restart_addr 4190 HeapWord* ra = _restart_addr; 4191 markFromRootsClosure.reset(ra); 4192 _restart_addr = NULL; 4193 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4194 } 4195 return true; 4196 } 4197 4198 void CMSCollector::preclean() { 4199 check_correct_thread_executing(); 4200 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4201 verify_work_stacks_empty(); 4202 verify_overflow_empty(); 4203 _abort_preclean = false; 4204 if (CMSPrecleaningEnabled) { 4205 if (!CMSEdenChunksRecordAlways) { 4206 _eden_chunk_index = 0; 4207 } 4208 size_t used = get_eden_used(); 4209 size_t capacity = get_eden_capacity(); 4210 // Don't start sampling unless we will get sufficiently 4211 // many samples. 4212 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4213 * CMSScheduleRemarkEdenPenetration)) { 4214 _start_sampling = true; 4215 } else { 4216 _start_sampling = false; 4217 } 4218 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4219 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4220 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4221 } 4222 CMSTokenSync x(true); // is cms thread 4223 if (CMSPrecleaningEnabled) { 4224 sample_eden(); 4225 _collectorState = AbortablePreclean; 4226 } else { 4227 _collectorState = FinalMarking; 4228 } 4229 verify_work_stacks_empty(); 4230 verify_overflow_empty(); 4231 } 4232 4233 // Try and schedule the remark such that young gen 4234 // occupancy is CMSScheduleRemarkEdenPenetration %. 4235 void CMSCollector::abortable_preclean() { 4236 check_correct_thread_executing(); 4237 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4238 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4239 4240 // If Eden's current occupancy is below this threshold, 4241 // immediately schedule the remark; else preclean 4242 // past the next scavenge in an effort to 4243 // schedule the pause as described above. By choosing 4244 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4245 // we will never do an actual abortable preclean cycle. 4246 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4247 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4248 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4249 // We need more smarts in the abortable preclean 4250 // loop below to deal with cases where allocation 4251 // in young gen is very very slow, and our precleaning 4252 // is running a losing race against a horde of 4253 // mutators intent on flooding us with CMS updates 4254 // (dirty cards). 4255 // One, admittedly dumb, strategy is to give up 4256 // after a certain number of abortable precleaning loops 4257 // or after a certain maximum time. We want to make 4258 // this smarter in the next iteration. 4259 // XXX FIX ME!!! YSR 4260 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4261 while (!(should_abort_preclean() || 4262 ConcurrentMarkSweepThread::should_terminate())) { 4263 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4264 cumworkdone += workdone; 4265 loops++; 4266 // Voluntarily terminate abortable preclean phase if we have 4267 // been at it for too long. 4268 if ((CMSMaxAbortablePrecleanLoops != 0) && 4269 loops >= CMSMaxAbortablePrecleanLoops) { 4270 if (PrintGCDetails) { 4271 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4272 } 4273 break; 4274 } 4275 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4276 if (PrintGCDetails) { 4277 gclog_or_tty->print(" CMS: abort preclean due to time "); 4278 } 4279 break; 4280 } 4281 // If we are doing little work each iteration, we should 4282 // take a short break. 4283 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4284 // Sleep for some time, waiting for work to accumulate 4285 stopTimer(); 4286 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4287 startTimer(); 4288 waited++; 4289 } 4290 } 4291 if (PrintCMSStatistics > 0) { 4292 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 4293 loops, waited, cumworkdone); 4294 } 4295 } 4296 CMSTokenSync x(true); // is cms thread 4297 if (_collectorState != Idling) { 4298 assert(_collectorState == AbortablePreclean, 4299 "Spontaneous state transition?"); 4300 _collectorState = FinalMarking; 4301 } // Else, a foreground collection completed this CMS cycle. 4302 return; 4303 } 4304 4305 // Respond to an Eden sampling opportunity 4306 void CMSCollector::sample_eden() { 4307 // Make sure a young gc cannot sneak in between our 4308 // reading and recording of a sample. 4309 assert(Thread::current()->is_ConcurrentGC_thread(), 4310 "Only the cms thread may collect Eden samples"); 4311 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4312 "Should collect samples while holding CMS token"); 4313 if (!_start_sampling) { 4314 return; 4315 } 4316 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4317 // is populated by the young generation. 4318 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4319 if (_eden_chunk_index < _eden_chunk_capacity) { 4320 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4321 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4322 "Unexpected state of Eden"); 4323 // We'd like to check that what we just sampled is an oop-start address; 4324 // however, we cannot do that here since the object may not yet have been 4325 // initialized. So we'll instead do the check when we _use_ this sample 4326 // later. 4327 if (_eden_chunk_index == 0 || 4328 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4329 _eden_chunk_array[_eden_chunk_index-1]) 4330 >= CMSSamplingGrain)) { 4331 _eden_chunk_index++; // commit sample 4332 } 4333 } 4334 } 4335 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4336 size_t used = get_eden_used(); 4337 size_t capacity = get_eden_capacity(); 4338 assert(used <= capacity, "Unexpected state of Eden"); 4339 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4340 _abort_preclean = true; 4341 } 4342 } 4343 } 4344 4345 4346 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4347 assert(_collectorState == Precleaning || 4348 _collectorState == AbortablePreclean, "incorrect state"); 4349 ResourceMark rm; 4350 HandleMark hm; 4351 4352 // Precleaning is currently not MT but the reference processor 4353 // may be set for MT. Disable it temporarily here. 4354 ReferenceProcessor* rp = ref_processor(); 4355 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4356 4357 // Do one pass of scrubbing the discovered reference lists 4358 // to remove any reference objects with strongly-reachable 4359 // referents. 4360 if (clean_refs) { 4361 CMSPrecleanRefsYieldClosure yield_cl(this); 4362 assert(rp->span().equals(_span), "Spans should be equal"); 4363 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4364 &_markStack, true /* preclean */); 4365 CMSDrainMarkingStackClosure complete_trace(this, 4366 _span, &_markBitMap, &_markStack, 4367 &keep_alive, true /* preclean */); 4368 4369 // We don't want this step to interfere with a young 4370 // collection because we don't want to take CPU 4371 // or memory bandwidth away from the young GC threads 4372 // (which may be as many as there are CPUs). 4373 // Note that we don't need to protect ourselves from 4374 // interference with mutators because they can't 4375 // manipulate the discovered reference lists nor affect 4376 // the computed reachability of the referents, the 4377 // only properties manipulated by the precleaning 4378 // of these reference lists. 4379 stopTimer(); 4380 CMSTokenSyncWithLocks x(true /* is cms thread */, 4381 bitMapLock()); 4382 startTimer(); 4383 sample_eden(); 4384 4385 // The following will yield to allow foreground 4386 // collection to proceed promptly. XXX YSR: 4387 // The code in this method may need further 4388 // tweaking for better performance and some restructuring 4389 // for cleaner interfaces. 4390 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4391 rp->preclean_discovered_references( 4392 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4393 gc_timer, _gc_tracer_cm->gc_id()); 4394 } 4395 4396 if (clean_survivor) { // preclean the active survivor space(s) 4397 assert(_young_gen->kind() == Generation::DefNew || 4398 _young_gen->kind() == Generation::ParNew, 4399 "incorrect type for cast"); 4400 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4401 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4402 &_markBitMap, &_modUnionTable, 4403 &_markStack, true /* precleaning phase */); 4404 stopTimer(); 4405 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4406 bitMapLock()); 4407 startTimer(); 4408 unsigned int before_count = 4409 GenCollectedHeap::heap()->total_collections(); 4410 SurvivorSpacePrecleanClosure 4411 sss_cl(this, _span, &_markBitMap, &_markStack, 4412 &pam_cl, before_count, CMSYield); 4413 dng->from()->object_iterate_careful(&sss_cl); 4414 dng->to()->object_iterate_careful(&sss_cl); 4415 } 4416 MarkRefsIntoAndScanClosure 4417 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4418 &_markStack, this, CMSYield, 4419 true /* precleaning phase */); 4420 // CAUTION: The following closure has persistent state that may need to 4421 // be reset upon a decrease in the sequence of addresses it 4422 // processes. 4423 ScanMarkedObjectsAgainCarefullyClosure 4424 smoac_cl(this, _span, 4425 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4426 4427 // Preclean dirty cards in ModUnionTable and CardTable using 4428 // appropriate convergence criterion; 4429 // repeat CMSPrecleanIter times unless we find that 4430 // we are losing. 4431 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4432 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4433 "Bad convergence multiplier"); 4434 assert(CMSPrecleanThreshold >= 100, 4435 "Unreasonably low CMSPrecleanThreshold"); 4436 4437 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4438 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4439 numIter < CMSPrecleanIter; 4440 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4441 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4442 if (Verbose && PrintGCDetails) { 4443 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4444 } 4445 // Either there are very few dirty cards, so re-mark 4446 // pause will be small anyway, or our pre-cleaning isn't 4447 // that much faster than the rate at which cards are being 4448 // dirtied, so we might as well stop and re-mark since 4449 // precleaning won't improve our re-mark time by much. 4450 if (curNumCards <= CMSPrecleanThreshold || 4451 (numIter > 0 && 4452 (curNumCards * CMSPrecleanDenominator > 4453 lastNumCards * CMSPrecleanNumerator))) { 4454 numIter++; 4455 cumNumCards += curNumCards; 4456 break; 4457 } 4458 } 4459 4460 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4461 4462 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4463 cumNumCards += curNumCards; 4464 if (PrintGCDetails && PrintCMSStatistics != 0) { 4465 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4466 curNumCards, cumNumCards, numIter); 4467 } 4468 return cumNumCards; // as a measure of useful work done 4469 } 4470 4471 // PRECLEANING NOTES: 4472 // Precleaning involves: 4473 // . reading the bits of the modUnionTable and clearing the set bits. 4474 // . For the cards corresponding to the set bits, we scan the 4475 // objects on those cards. This means we need the free_list_lock 4476 // so that we can safely iterate over the CMS space when scanning 4477 // for oops. 4478 // . When we scan the objects, we'll be both reading and setting 4479 // marks in the marking bit map, so we'll need the marking bit map. 4480 // . For protecting _collector_state transitions, we take the CGC_lock. 4481 // Note that any races in the reading of of card table entries by the 4482 // CMS thread on the one hand and the clearing of those entries by the 4483 // VM thread or the setting of those entries by the mutator threads on the 4484 // other are quite benign. However, for efficiency it makes sense to keep 4485 // the VM thread from racing with the CMS thread while the latter is 4486 // dirty card info to the modUnionTable. We therefore also use the 4487 // CGC_lock to protect the reading of the card table and the mod union 4488 // table by the CM thread. 4489 // . We run concurrently with mutator updates, so scanning 4490 // needs to be done carefully -- we should not try to scan 4491 // potentially uninitialized objects. 4492 // 4493 // Locking strategy: While holding the CGC_lock, we scan over and 4494 // reset a maximal dirty range of the mod union / card tables, then lock 4495 // the free_list_lock and bitmap lock to do a full marking, then 4496 // release these locks; and repeat the cycle. This allows for a 4497 // certain amount of fairness in the sharing of these locks between 4498 // the CMS collector on the one hand, and the VM thread and the 4499 // mutators on the other. 4500 4501 // NOTE: preclean_mod_union_table() and preclean_card_table() 4502 // further below are largely identical; if you need to modify 4503 // one of these methods, please check the other method too. 4504 4505 size_t CMSCollector::preclean_mod_union_table( 4506 ConcurrentMarkSweepGeneration* gen, 4507 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4508 verify_work_stacks_empty(); 4509 verify_overflow_empty(); 4510 4511 // strategy: starting with the first card, accumulate contiguous 4512 // ranges of dirty cards; clear these cards, then scan the region 4513 // covered by these cards. 4514 4515 // Since all of the MUT is committed ahead, we can just use 4516 // that, in case the generations expand while we are precleaning. 4517 // It might also be fine to just use the committed part of the 4518 // generation, but we might potentially miss cards when the 4519 // generation is rapidly expanding while we are in the midst 4520 // of precleaning. 4521 HeapWord* startAddr = gen->reserved().start(); 4522 HeapWord* endAddr = gen->reserved().end(); 4523 4524 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4525 4526 size_t numDirtyCards, cumNumDirtyCards; 4527 HeapWord *nextAddr, *lastAddr; 4528 for (cumNumDirtyCards = numDirtyCards = 0, 4529 nextAddr = lastAddr = startAddr; 4530 nextAddr < endAddr; 4531 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4532 4533 ResourceMark rm; 4534 HandleMark hm; 4535 4536 MemRegion dirtyRegion; 4537 { 4538 stopTimer(); 4539 // Potential yield point 4540 CMSTokenSync ts(true); 4541 startTimer(); 4542 sample_eden(); 4543 // Get dirty region starting at nextOffset (inclusive), 4544 // simultaneously clearing it. 4545 dirtyRegion = 4546 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4547 assert(dirtyRegion.start() >= nextAddr, 4548 "returned region inconsistent?"); 4549 } 4550 // Remember where the next search should begin. 4551 // The returned region (if non-empty) is a right open interval, 4552 // so lastOffset is obtained from the right end of that 4553 // interval. 4554 lastAddr = dirtyRegion.end(); 4555 // Should do something more transparent and less hacky XXX 4556 numDirtyCards = 4557 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4558 4559 // We'll scan the cards in the dirty region (with periodic 4560 // yields for foreground GC as needed). 4561 if (!dirtyRegion.is_empty()) { 4562 assert(numDirtyCards > 0, "consistency check"); 4563 HeapWord* stop_point = NULL; 4564 stopTimer(); 4565 // Potential yield point 4566 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4567 bitMapLock()); 4568 startTimer(); 4569 { 4570 verify_work_stacks_empty(); 4571 verify_overflow_empty(); 4572 sample_eden(); 4573 stop_point = 4574 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4575 } 4576 if (stop_point != NULL) { 4577 // The careful iteration stopped early either because it found an 4578 // uninitialized object, or because we were in the midst of an 4579 // "abortable preclean", which should now be aborted. Redirty 4580 // the bits corresponding to the partially-scanned or unscanned 4581 // cards. We'll either restart at the next block boundary or 4582 // abort the preclean. 4583 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4584 "Should only be AbortablePreclean."); 4585 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4586 if (should_abort_preclean()) { 4587 break; // out of preclean loop 4588 } else { 4589 // Compute the next address at which preclean should pick up; 4590 // might need bitMapLock in order to read P-bits. 4591 lastAddr = next_card_start_after_block(stop_point); 4592 } 4593 } 4594 } else { 4595 assert(lastAddr == endAddr, "consistency check"); 4596 assert(numDirtyCards == 0, "consistency check"); 4597 break; 4598 } 4599 } 4600 verify_work_stacks_empty(); 4601 verify_overflow_empty(); 4602 return cumNumDirtyCards; 4603 } 4604 4605 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4606 // below are largely identical; if you need to modify 4607 // one of these methods, please check the other method too. 4608 4609 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4610 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4611 // strategy: it's similar to precleamModUnionTable above, in that 4612 // we accumulate contiguous ranges of dirty cards, mark these cards 4613 // precleaned, then scan the region covered by these cards. 4614 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4615 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4616 4617 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4618 4619 size_t numDirtyCards, cumNumDirtyCards; 4620 HeapWord *lastAddr, *nextAddr; 4621 4622 for (cumNumDirtyCards = numDirtyCards = 0, 4623 nextAddr = lastAddr = startAddr; 4624 nextAddr < endAddr; 4625 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4626 4627 ResourceMark rm; 4628 HandleMark hm; 4629 4630 MemRegion dirtyRegion; 4631 { 4632 // See comments in "Precleaning notes" above on why we 4633 // do this locking. XXX Could the locking overheads be 4634 // too high when dirty cards are sparse? [I don't think so.] 4635 stopTimer(); 4636 CMSTokenSync x(true); // is cms thread 4637 startTimer(); 4638 sample_eden(); 4639 // Get and clear dirty region from card table 4640 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4641 MemRegion(nextAddr, endAddr), 4642 true, 4643 CardTableModRefBS::precleaned_card_val()); 4644 4645 assert(dirtyRegion.start() >= nextAddr, 4646 "returned region inconsistent?"); 4647 } 4648 lastAddr = dirtyRegion.end(); 4649 numDirtyCards = 4650 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4651 4652 if (!dirtyRegion.is_empty()) { 4653 stopTimer(); 4654 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4655 startTimer(); 4656 sample_eden(); 4657 verify_work_stacks_empty(); 4658 verify_overflow_empty(); 4659 HeapWord* stop_point = 4660 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4661 if (stop_point != NULL) { 4662 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4663 "Should only be AbortablePreclean."); 4664 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4665 if (should_abort_preclean()) { 4666 break; // out of preclean loop 4667 } else { 4668 // Compute the next address at which preclean should pick up. 4669 lastAddr = next_card_start_after_block(stop_point); 4670 } 4671 } 4672 } else { 4673 break; 4674 } 4675 } 4676 verify_work_stacks_empty(); 4677 verify_overflow_empty(); 4678 return cumNumDirtyCards; 4679 } 4680 4681 class PrecleanKlassClosure : public KlassClosure { 4682 KlassToOopClosure _cm_klass_closure; 4683 public: 4684 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4685 void do_klass(Klass* k) { 4686 if (k->has_accumulated_modified_oops()) { 4687 k->clear_accumulated_modified_oops(); 4688 4689 _cm_klass_closure.do_klass(k); 4690 } 4691 } 4692 }; 4693 4694 // The freelist lock is needed to prevent asserts, is it really needed? 4695 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4696 4697 cl->set_freelistLock(freelistLock); 4698 4699 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4700 4701 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4702 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4703 PrecleanKlassClosure preclean_klass_closure(cl); 4704 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4705 4706 verify_work_stacks_empty(); 4707 verify_overflow_empty(); 4708 } 4709 4710 void CMSCollector::checkpointRootsFinal(bool asynch, 4711 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4712 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4713 check_correct_thread_executing(); 4714 // world is stopped at this checkpoint 4715 assert(SafepointSynchronize::is_at_safepoint(), 4716 "world should be stopped"); 4717 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4718 4719 verify_work_stacks_empty(); 4720 verify_overflow_empty(); 4721 4722 SpecializationStats::clear(); 4723 if (PrintGCDetails) { 4724 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4725 _young_gen->used() / K, 4726 _young_gen->capacity() / K); 4727 } 4728 if (asynch) { 4729 if (CMSScavengeBeforeRemark) { 4730 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4731 // Temporarily set flag to false, GCH->do_collection will 4732 // expect it to be false and set to true 4733 FlagSetting fl(gch->_is_gc_active, false); 4734 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4735 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4736 int level = _cmsGen->level() - 1; 4737 if (level >= 0) { 4738 gch->do_collection(true, // full (i.e. force, see below) 4739 false, // !clear_all_soft_refs 4740 0, // size 4741 false, // is_tlab 4742 level // max_level 4743 ); 4744 } 4745 } 4746 FreelistLocker x(this); 4747 MutexLockerEx y(bitMapLock(), 4748 Mutex::_no_safepoint_check_flag); 4749 assert(!init_mark_was_synchronous, "but that's impossible!"); 4750 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4751 } else { 4752 // already have all the locks 4753 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4754 init_mark_was_synchronous); 4755 } 4756 verify_work_stacks_empty(); 4757 verify_overflow_empty(); 4758 SpecializationStats::print(); 4759 } 4760 4761 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4762 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4763 4764 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4765 4766 assert(haveFreelistLocks(), "must have free list locks"); 4767 assert_lock_strong(bitMapLock()); 4768 4769 ResourceMark rm; 4770 HandleMark hm; 4771 4772 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4773 4774 if (should_unload_classes()) { 4775 CodeCache::gc_prologue(); 4776 } 4777 assert(haveFreelistLocks(), "must have free list locks"); 4778 assert_lock_strong(bitMapLock()); 4779 4780 if (!init_mark_was_synchronous) { 4781 // We might assume that we need not fill TLAB's when 4782 // CMSScavengeBeforeRemark is set, because we may have just done 4783 // a scavenge which would have filled all TLAB's -- and besides 4784 // Eden would be empty. This however may not always be the case -- 4785 // for instance although we asked for a scavenge, it may not have 4786 // happened because of a JNI critical section. We probably need 4787 // a policy for deciding whether we can in that case wait until 4788 // the critical section releases and then do the remark following 4789 // the scavenge, and skip it here. In the absence of that policy, 4790 // or of an indication of whether the scavenge did indeed occur, 4791 // we cannot rely on TLAB's having been filled and must do 4792 // so here just in case a scavenge did not happen. 4793 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4794 // Update the saved marks which may affect the root scans. 4795 gch->save_marks(); 4796 4797 if (CMSPrintEdenSurvivorChunks) { 4798 print_eden_and_survivor_chunk_arrays(); 4799 } 4800 4801 { 4802 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4803 4804 // Note on the role of the mod union table: 4805 // Since the marker in "markFromRoots" marks concurrently with 4806 // mutators, it is possible for some reachable objects not to have been 4807 // scanned. For instance, an only reference to an object A was 4808 // placed in object B after the marker scanned B. Unless B is rescanned, 4809 // A would be collected. Such updates to references in marked objects 4810 // are detected via the mod union table which is the set of all cards 4811 // dirtied since the first checkpoint in this GC cycle and prior to 4812 // the most recent young generation GC, minus those cleaned up by the 4813 // concurrent precleaning. 4814 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4815 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4816 do_remark_parallel(); 4817 } else { 4818 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4819 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4820 do_remark_non_parallel(); 4821 } 4822 } 4823 } else { 4824 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4825 // The initial mark was stop-world, so there's no rescanning to 4826 // do; go straight on to the next step below. 4827 } 4828 verify_work_stacks_empty(); 4829 verify_overflow_empty(); 4830 4831 { 4832 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4833 refProcessingWork(asynch, clear_all_soft_refs); 4834 } 4835 verify_work_stacks_empty(); 4836 verify_overflow_empty(); 4837 4838 if (should_unload_classes()) { 4839 CodeCache::gc_epilogue(); 4840 } 4841 JvmtiExport::gc_epilogue(); 4842 4843 // If we encountered any (marking stack / work queue) overflow 4844 // events during the current CMS cycle, take appropriate 4845 // remedial measures, where possible, so as to try and avoid 4846 // recurrence of that condition. 4847 assert(_markStack.isEmpty(), "No grey objects"); 4848 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4849 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4850 if (ser_ovflw > 0) { 4851 if (PrintCMSStatistics != 0) { 4852 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4853 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4854 ", kac_preclean="SIZE_FORMAT")", 4855 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4856 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4857 } 4858 _markStack.expand(); 4859 _ser_pmc_remark_ovflw = 0; 4860 _ser_pmc_preclean_ovflw = 0; 4861 _ser_kac_preclean_ovflw = 0; 4862 _ser_kac_ovflw = 0; 4863 } 4864 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4865 if (PrintCMSStatistics != 0) { 4866 gclog_or_tty->print_cr("Work queue overflow (benign) " 4867 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4868 _par_pmc_remark_ovflw, _par_kac_ovflw); 4869 } 4870 _par_pmc_remark_ovflw = 0; 4871 _par_kac_ovflw = 0; 4872 } 4873 if (PrintCMSStatistics != 0) { 4874 if (_markStack._hit_limit > 0) { 4875 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4876 _markStack._hit_limit); 4877 } 4878 if (_markStack._failed_double > 0) { 4879 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4880 " current capacity "SIZE_FORMAT, 4881 _markStack._failed_double, 4882 _markStack.capacity()); 4883 } 4884 } 4885 _markStack._hit_limit = 0; 4886 _markStack._failed_double = 0; 4887 4888 if ((VerifyAfterGC || VerifyDuringGC) && 4889 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4890 verify_after_remark(); 4891 } 4892 4893 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4894 4895 // Change under the freelistLocks. 4896 _collectorState = Sweeping; 4897 // Call isAllClear() under bitMapLock 4898 assert(_modUnionTable.isAllClear(), 4899 "Should be clear by end of the final marking"); 4900 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4901 "Should be clear by end of the final marking"); 4902 } 4903 4904 void CMSParInitialMarkTask::work(uint worker_id) { 4905 elapsedTimer _timer; 4906 ResourceMark rm; 4907 HandleMark hm; 4908 4909 // ---------- scan from roots -------------- 4910 _timer.start(); 4911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4912 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4913 4914 // ---------- young gen roots -------------- 4915 { 4916 work_on_young_gen_roots(worker_id, &par_mri_cl); 4917 _timer.stop(); 4918 if (PrintCMSStatistics != 0) { 4919 gclog_or_tty->print_cr( 4920 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4921 worker_id, _timer.seconds()); 4922 } 4923 } 4924 4925 // ---------- remaining roots -------------- 4926 _timer.reset(); 4927 _timer.start(); 4928 4929 CLDToOopClosure cld_closure(&par_mri_cl, true); 4930 4931 gch->gen_process_roots(_collector->_cmsGen->level(), 4932 false, // yg was scanned above 4933 false, // this is parallel code 4934 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4935 _collector->should_unload_classes(), 4936 &par_mri_cl, 4937 NULL, 4938 &cld_closure); 4939 assert(_collector->should_unload_classes() 4940 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 4941 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4942 _timer.stop(); 4943 if (PrintCMSStatistics != 0) { 4944 gclog_or_tty->print_cr( 4945 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4946 worker_id, _timer.seconds()); 4947 } 4948 } 4949 4950 // Parallel remark task 4951 class CMSParRemarkTask: public CMSParMarkTask { 4952 CompactibleFreeListSpace* _cms_space; 4953 4954 // The per-thread work queues, available here for stealing. 4955 OopTaskQueueSet* _task_queues; 4956 ParallelTaskTerminator _term; 4957 4958 public: 4959 // A value of 0 passed to n_workers will cause the number of 4960 // workers to be taken from the active workers in the work gang. 4961 CMSParRemarkTask(CMSCollector* collector, 4962 CompactibleFreeListSpace* cms_space, 4963 int n_workers, FlexibleWorkGang* workers, 4964 OopTaskQueueSet* task_queues): 4965 CMSParMarkTask("Rescan roots and grey objects in parallel", 4966 collector, n_workers), 4967 _cms_space(cms_space), 4968 _task_queues(task_queues), 4969 _term(n_workers, task_queues) { } 4970 4971 OopTaskQueueSet* task_queues() { return _task_queues; } 4972 4973 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4974 4975 ParallelTaskTerminator* terminator() { return &_term; } 4976 int n_workers() { return _n_workers; } 4977 4978 void work(uint worker_id); 4979 4980 private: 4981 // ... of dirty cards in old space 4982 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4983 Par_MarkRefsIntoAndScanClosure* cl); 4984 4985 // ... work stealing for the above 4986 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4987 }; 4988 4989 class RemarkKlassClosure : public KlassClosure { 4990 KlassToOopClosure _cm_klass_closure; 4991 public: 4992 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4993 void do_klass(Klass* k) { 4994 // Check if we have modified any oops in the Klass during the concurrent marking. 4995 if (k->has_accumulated_modified_oops()) { 4996 k->clear_accumulated_modified_oops(); 4997 4998 // We could have transfered the current modified marks to the accumulated marks, 4999 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5000 } else if (k->has_modified_oops()) { 5001 // Don't clear anything, this info is needed by the next young collection. 5002 } else { 5003 // No modified oops in the Klass. 5004 return; 5005 } 5006 5007 // The klass has modified fields, need to scan the klass. 5008 _cm_klass_closure.do_klass(k); 5009 } 5010 }; 5011 5012 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5013 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5014 ContiguousSpace* eden_space = dng->eden(); 5015 ContiguousSpace* from_space = dng->from(); 5016 ContiguousSpace* to_space = dng->to(); 5017 5018 HeapWord** eca = _collector->_eden_chunk_array; 5019 size_t ect = _collector->_eden_chunk_index; 5020 HeapWord** sca = _collector->_survivor_chunk_array; 5021 size_t sct = _collector->_survivor_chunk_index; 5022 5023 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5024 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5025 5026 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5027 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5028 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5029 } 5030 5031 // work_queue(i) is passed to the closure 5032 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5033 // also is passed to do_dirty_card_rescan_tasks() and to 5034 // do_work_steal() to select the i-th task_queue. 5035 5036 void CMSParRemarkTask::work(uint worker_id) { 5037 elapsedTimer _timer; 5038 ResourceMark rm; 5039 HandleMark hm; 5040 5041 // ---------- rescan from roots -------------- 5042 _timer.start(); 5043 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5044 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5045 _collector->_span, _collector->ref_processor(), 5046 &(_collector->_markBitMap), 5047 work_queue(worker_id)); 5048 5049 // Rescan young gen roots first since these are likely 5050 // coarsely partitioned and may, on that account, constitute 5051 // the critical path; thus, it's best to start off that 5052 // work first. 5053 // ---------- young gen roots -------------- 5054 { 5055 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5056 _timer.stop(); 5057 if (PrintCMSStatistics != 0) { 5058 gclog_or_tty->print_cr( 5059 "Finished young gen rescan work in %dth thread: %3.3f sec", 5060 worker_id, _timer.seconds()); 5061 } 5062 } 5063 5064 // ---------- remaining roots -------------- 5065 _timer.reset(); 5066 _timer.start(); 5067 gch->gen_process_roots(_collector->_cmsGen->level(), 5068 false, // yg was scanned above 5069 false, // this is parallel code 5070 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5071 _collector->should_unload_classes(), 5072 &par_mrias_cl, 5073 NULL, 5074 NULL); // The dirty klasses will be handled below 5075 5076 assert(_collector->should_unload_classes() 5077 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5078 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5079 _timer.stop(); 5080 if (PrintCMSStatistics != 0) { 5081 gclog_or_tty->print_cr( 5082 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5083 worker_id, _timer.seconds()); 5084 } 5085 5086 // ---------- unhandled CLD scanning ---------- 5087 if (worker_id == 0) { // Single threaded at the moment. 5088 _timer.reset(); 5089 _timer.start(); 5090 5091 // Scan all new class loader data objects and new dependencies that were 5092 // introduced during concurrent marking. 5093 ResourceMark rm; 5094 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5095 for (int i = 0; i < array->length(); i++) { 5096 par_mrias_cl.do_class_loader_data(array->at(i)); 5097 } 5098 5099 // We don't need to keep track of new CLDs anymore. 5100 ClassLoaderDataGraph::remember_new_clds(false); 5101 5102 _timer.stop(); 5103 if (PrintCMSStatistics != 0) { 5104 gclog_or_tty->print_cr( 5105 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5106 worker_id, _timer.seconds()); 5107 } 5108 } 5109 5110 // ---------- dirty klass scanning ---------- 5111 if (worker_id == 0) { // Single threaded at the moment. 5112 _timer.reset(); 5113 _timer.start(); 5114 5115 // Scan all classes that was dirtied during the concurrent marking phase. 5116 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5117 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5118 5119 _timer.stop(); 5120 if (PrintCMSStatistics != 0) { 5121 gclog_or_tty->print_cr( 5122 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5123 worker_id, _timer.seconds()); 5124 } 5125 } 5126 5127 // We might have added oops to ClassLoaderData::_handles during the 5128 // concurrent marking phase. These oops point to newly allocated objects 5129 // that are guaranteed to be kept alive. Either by the direct allocation 5130 // code, or when the young collector processes the roots. Hence, 5131 // we don't have to revisit the _handles block during the remark phase. 5132 5133 // ---------- rescan dirty cards ------------ 5134 _timer.reset(); 5135 _timer.start(); 5136 5137 // Do the rescan tasks for each of the two spaces 5138 // (cms_space) in turn. 5139 // "worker_id" is passed to select the task_queue for "worker_id" 5140 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5141 _timer.stop(); 5142 if (PrintCMSStatistics != 0) { 5143 gclog_or_tty->print_cr( 5144 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5145 worker_id, _timer.seconds()); 5146 } 5147 5148 // ---------- steal work from other threads ... 5149 // ---------- ... and drain overflow list. 5150 _timer.reset(); 5151 _timer.start(); 5152 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5153 _timer.stop(); 5154 if (PrintCMSStatistics != 0) { 5155 gclog_or_tty->print_cr( 5156 "Finished work stealing in %dth thread: %3.3f sec", 5157 worker_id, _timer.seconds()); 5158 } 5159 } 5160 5161 // Note that parameter "i" is not used. 5162 void 5163 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5164 OopsInGenClosure* cl, ContiguousSpace* space, 5165 HeapWord** chunk_array, size_t chunk_top) { 5166 // Until all tasks completed: 5167 // . claim an unclaimed task 5168 // . compute region boundaries corresponding to task claimed 5169 // using chunk_array 5170 // . par_oop_iterate(cl) over that region 5171 5172 ResourceMark rm; 5173 HandleMark hm; 5174 5175 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5176 5177 uint nth_task = 0; 5178 uint n_tasks = pst->n_tasks(); 5179 5180 if (n_tasks > 0) { 5181 assert(pst->valid(), "Uninitialized use?"); 5182 HeapWord *start, *end; 5183 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5184 // We claimed task # nth_task; compute its boundaries. 5185 if (chunk_top == 0) { // no samples were taken 5186 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 5187 start = space->bottom(); 5188 end = space->top(); 5189 } else if (nth_task == 0) { 5190 start = space->bottom(); 5191 end = chunk_array[nth_task]; 5192 } else if (nth_task < (uint)chunk_top) { 5193 assert(nth_task >= 1, "Control point invariant"); 5194 start = chunk_array[nth_task - 1]; 5195 end = chunk_array[nth_task]; 5196 } else { 5197 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5198 start = chunk_array[chunk_top - 1]; 5199 end = space->top(); 5200 } 5201 MemRegion mr(start, end); 5202 // Verify that mr is in space 5203 assert(mr.is_empty() || space->used_region().contains(mr), 5204 "Should be in space"); 5205 // Verify that "start" is an object boundary 5206 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5207 "Should be an oop"); 5208 space->par_oop_iterate(mr, cl); 5209 } 5210 pst->all_tasks_completed(); 5211 } 5212 } 5213 5214 void 5215 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5216 CompactibleFreeListSpace* sp, int i, 5217 Par_MarkRefsIntoAndScanClosure* cl) { 5218 // Until all tasks completed: 5219 // . claim an unclaimed task 5220 // . compute region boundaries corresponding to task claimed 5221 // . transfer dirty bits ct->mut for that region 5222 // . apply rescanclosure to dirty mut bits for that region 5223 5224 ResourceMark rm; 5225 HandleMark hm; 5226 5227 OopTaskQueue* work_q = work_queue(i); 5228 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5229 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5230 // CAUTION: This closure has state that persists across calls to 5231 // the work method dirty_range_iterate_clear() in that it has 5232 // embedded in it a (subtype of) UpwardsObjectClosure. The 5233 // use of that state in the embedded UpwardsObjectClosure instance 5234 // assumes that the cards are always iterated (even if in parallel 5235 // by several threads) in monotonically increasing order per each 5236 // thread. This is true of the implementation below which picks 5237 // card ranges (chunks) in monotonically increasing order globally 5238 // and, a-fortiori, in monotonically increasing order per thread 5239 // (the latter order being a subsequence of the former). 5240 // If the work code below is ever reorganized into a more chaotic 5241 // work-partitioning form than the current "sequential tasks" 5242 // paradigm, the use of that persistent state will have to be 5243 // revisited and modified appropriately. See also related 5244 // bug 4756801 work on which should examine this code to make 5245 // sure that the changes there do not run counter to the 5246 // assumptions made here and necessary for correctness and 5247 // efficiency. Note also that this code might yield inefficient 5248 // behavior in the case of very large objects that span one or 5249 // more work chunks. Such objects would potentially be scanned 5250 // several times redundantly. Work on 4756801 should try and 5251 // address that performance anomaly if at all possible. XXX 5252 MemRegion full_span = _collector->_span; 5253 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5254 MarkFromDirtyCardsClosure 5255 greyRescanClosure(_collector, full_span, // entire span of interest 5256 sp, bm, work_q, cl); 5257 5258 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5259 assert(pst->valid(), "Uninitialized use?"); 5260 uint nth_task = 0; 5261 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5262 MemRegion span = sp->used_region(); 5263 HeapWord* start_addr = span.start(); 5264 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5265 alignment); 5266 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5267 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5268 start_addr, "Check alignment"); 5269 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5270 chunk_size, "Check alignment"); 5271 5272 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5273 // Having claimed the nth_task, compute corresponding mem-region, 5274 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 5275 // The alignment restriction ensures that we do not need any 5276 // synchronization with other gang-workers while setting or 5277 // clearing bits in thus chunk of the MUT. 5278 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5279 start_addr + (nth_task+1)*chunk_size); 5280 // The last chunk's end might be way beyond end of the 5281 // used region. In that case pull back appropriately. 5282 if (this_span.end() > end_addr) { 5283 this_span.set_end(end_addr); 5284 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5285 } 5286 // Iterate over the dirty cards covering this chunk, marking them 5287 // precleaned, and setting the corresponding bits in the mod union 5288 // table. Since we have been careful to partition at Card and MUT-word 5289 // boundaries no synchronization is needed between parallel threads. 5290 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5291 &modUnionClosure); 5292 5293 // Having transferred these marks into the modUnionTable, 5294 // rescan the marked objects on the dirty cards in the modUnionTable. 5295 // Even if this is at a synchronous collection, the initial marking 5296 // may have been done during an asynchronous collection so there 5297 // may be dirty bits in the mod-union table. 5298 _collector->_modUnionTable.dirty_range_iterate_clear( 5299 this_span, &greyRescanClosure); 5300 _collector->_modUnionTable.verifyNoOneBitsInRange( 5301 this_span.start(), 5302 this_span.end()); 5303 } 5304 pst->all_tasks_completed(); // declare that i am done 5305 } 5306 5307 // . see if we can share work_queues with ParNew? XXX 5308 void 5309 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5310 int* seed) { 5311 OopTaskQueue* work_q = work_queue(i); 5312 NOT_PRODUCT(int num_steals = 0;) 5313 oop obj_to_scan; 5314 CMSBitMap* bm = &(_collector->_markBitMap); 5315 5316 while (true) { 5317 // Completely finish any left over work from (an) earlier round(s) 5318 cl->trim_queue(0); 5319 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5320 (size_t)ParGCDesiredObjsFromOverflowList); 5321 // Now check if there's any work in the overflow list 5322 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5323 // only affects the number of attempts made to get work from the 5324 // overflow list and does not affect the number of workers. Just 5325 // pass ParallelGCThreads so this behavior is unchanged. 5326 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5327 work_q, 5328 ParallelGCThreads)) { 5329 // found something in global overflow list; 5330 // not yet ready to go stealing work from others. 5331 // We'd like to assert(work_q->size() != 0, ...) 5332 // because we just took work from the overflow list, 5333 // but of course we can't since all of that could have 5334 // been already stolen from us. 5335 // "He giveth and He taketh away." 5336 continue; 5337 } 5338 // Verify that we have no work before we resort to stealing 5339 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5340 // Try to steal from other queues that have work 5341 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5342 NOT_PRODUCT(num_steals++;) 5343 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5344 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5345 // Do scanning work 5346 obj_to_scan->oop_iterate(cl); 5347 // Loop around, finish this work, and try to steal some more 5348 } else if (terminator()->offer_termination()) { 5349 break; // nirvana from the infinite cycle 5350 } 5351 } 5352 NOT_PRODUCT( 5353 if (PrintCMSStatistics != 0) { 5354 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5355 } 5356 ) 5357 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5358 "Else our work is not yet done"); 5359 } 5360 5361 // Record object boundaries in _eden_chunk_array by sampling the eden 5362 // top in the slow-path eden object allocation code path and record 5363 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5364 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5365 // sampling in sample_eden() that activates during the part of the 5366 // preclean phase. 5367 void CMSCollector::sample_eden_chunk() { 5368 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5369 if (_eden_chunk_lock->try_lock()) { 5370 // Record a sample. This is the critical section. The contents 5371 // of the _eden_chunk_array have to be non-decreasing in the 5372 // address order. 5373 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5374 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5375 "Unexpected state of Eden"); 5376 if (_eden_chunk_index == 0 || 5377 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5378 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5379 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5380 _eden_chunk_index++; // commit sample 5381 } 5382 _eden_chunk_lock->unlock(); 5383 } 5384 } 5385 } 5386 5387 // Return a thread-local PLAB recording array, as appropriate. 5388 void* CMSCollector::get_data_recorder(int thr_num) { 5389 if (_survivor_plab_array != NULL && 5390 (CMSPLABRecordAlways || 5391 (_collectorState > Marking && _collectorState < FinalMarking))) { 5392 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5393 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5394 ca->reset(); // clear it so that fresh data is recorded 5395 return (void*) ca; 5396 } else { 5397 return NULL; 5398 } 5399 } 5400 5401 // Reset all the thread-local PLAB recording arrays 5402 void CMSCollector::reset_survivor_plab_arrays() { 5403 for (uint i = 0; i < ParallelGCThreads; i++) { 5404 _survivor_plab_array[i].reset(); 5405 } 5406 } 5407 5408 // Merge the per-thread plab arrays into the global survivor chunk 5409 // array which will provide the partitioning of the survivor space 5410 // for CMS initial scan and rescan. 5411 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5412 int no_of_gc_threads) { 5413 assert(_survivor_plab_array != NULL, "Error"); 5414 assert(_survivor_chunk_array != NULL, "Error"); 5415 assert(_collectorState == FinalMarking || 5416 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5417 for (int j = 0; j < no_of_gc_threads; j++) { 5418 _cursor[j] = 0; 5419 } 5420 HeapWord* top = surv->top(); 5421 size_t i; 5422 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5423 HeapWord* min_val = top; // Higher than any PLAB address 5424 uint min_tid = 0; // position of min_val this round 5425 for (int j = 0; j < no_of_gc_threads; j++) { 5426 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5427 if (_cursor[j] == cur_sca->end()) { 5428 continue; 5429 } 5430 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5431 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5432 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5433 if (cur_val < min_val) { 5434 min_tid = j; 5435 min_val = cur_val; 5436 } else { 5437 assert(cur_val < top, "All recorded addresses should be less"); 5438 } 5439 } 5440 // At this point min_val and min_tid are respectively 5441 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5442 // and the thread (j) that witnesses that address. 5443 // We record this address in the _survivor_chunk_array[i] 5444 // and increment _cursor[min_tid] prior to the next round i. 5445 if (min_val == top) { 5446 break; 5447 } 5448 _survivor_chunk_array[i] = min_val; 5449 _cursor[min_tid]++; 5450 } 5451 // We are all done; record the size of the _survivor_chunk_array 5452 _survivor_chunk_index = i; // exclusive: [0, i) 5453 if (PrintCMSStatistics > 0) { 5454 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5455 } 5456 // Verify that we used up all the recorded entries 5457 #ifdef ASSERT 5458 size_t total = 0; 5459 for (int j = 0; j < no_of_gc_threads; j++) { 5460 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5461 total += _cursor[j]; 5462 } 5463 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5464 // Check that the merged array is in sorted order 5465 if (total > 0) { 5466 for (size_t i = 0; i < total - 1; i++) { 5467 if (PrintCMSStatistics > 0) { 5468 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5469 i, _survivor_chunk_array[i]); 5470 } 5471 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5472 "Not sorted"); 5473 } 5474 } 5475 #endif // ASSERT 5476 } 5477 5478 // Set up the space's par_seq_tasks structure for work claiming 5479 // for parallel initial scan and rescan of young gen. 5480 // See ParRescanTask where this is currently used. 5481 void 5482 CMSCollector:: 5483 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5484 assert(n_threads > 0, "Unexpected n_threads argument"); 5485 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5486 5487 // Eden space 5488 if (!dng->eden()->is_empty()) { 5489 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5490 assert(!pst->valid(), "Clobbering existing data?"); 5491 // Each valid entry in [0, _eden_chunk_index) represents a task. 5492 size_t n_tasks = _eden_chunk_index + 1; 5493 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5494 // Sets the condition for completion of the subtask (how many threads 5495 // need to finish in order to be done). 5496 pst->set_n_threads(n_threads); 5497 pst->set_n_tasks((int)n_tasks); 5498 } 5499 5500 // Merge the survivor plab arrays into _survivor_chunk_array 5501 if (_survivor_plab_array != NULL) { 5502 merge_survivor_plab_arrays(dng->from(), n_threads); 5503 } else { 5504 assert(_survivor_chunk_index == 0, "Error"); 5505 } 5506 5507 // To space 5508 { 5509 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5510 assert(!pst->valid(), "Clobbering existing data?"); 5511 // Sets the condition for completion of the subtask (how many threads 5512 // need to finish in order to be done). 5513 pst->set_n_threads(n_threads); 5514 pst->set_n_tasks(1); 5515 assert(pst->valid(), "Error"); 5516 } 5517 5518 // From space 5519 { 5520 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5521 assert(!pst->valid(), "Clobbering existing data?"); 5522 size_t n_tasks = _survivor_chunk_index + 1; 5523 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5524 // Sets the condition for completion of the subtask (how many threads 5525 // need to finish in order to be done). 5526 pst->set_n_threads(n_threads); 5527 pst->set_n_tasks((int)n_tasks); 5528 assert(pst->valid(), "Error"); 5529 } 5530 } 5531 5532 // Parallel version of remark 5533 void CMSCollector::do_remark_parallel() { 5534 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5535 FlexibleWorkGang* workers = gch->workers(); 5536 assert(workers != NULL, "Need parallel worker threads."); 5537 // Choose to use the number of GC workers most recently set 5538 // into "active_workers". If active_workers is not set, set it 5539 // to ParallelGCThreads. 5540 int n_workers = workers->active_workers(); 5541 if (n_workers == 0) { 5542 assert(n_workers > 0, "Should have been set during scavenge"); 5543 n_workers = ParallelGCThreads; 5544 workers->set_active_workers(n_workers); 5545 } 5546 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5547 5548 CMSParRemarkTask tsk(this, 5549 cms_space, 5550 n_workers, workers, task_queues()); 5551 5552 // Set up for parallel process_roots work. 5553 gch->set_par_threads(n_workers); 5554 // We won't be iterating over the cards in the card table updating 5555 // the younger_gen cards, so we shouldn't call the following else 5556 // the verification code as well as subsequent younger_refs_iterate 5557 // code would get confused. XXX 5558 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5559 5560 // The young gen rescan work will not be done as part of 5561 // process_roots (which currently doesn't know how to 5562 // parallelize such a scan), but rather will be broken up into 5563 // a set of parallel tasks (via the sampling that the [abortable] 5564 // preclean phase did of eden, plus the [two] tasks of 5565 // scanning the [two] survivor spaces. Further fine-grain 5566 // parallelization of the scanning of the survivor spaces 5567 // themselves, and of precleaning of the younger gen itself 5568 // is deferred to the future. 5569 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5570 5571 // The dirty card rescan work is broken up into a "sequence" 5572 // of parallel tasks (per constituent space) that are dynamically 5573 // claimed by the parallel threads. 5574 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5575 5576 // It turns out that even when we're using 1 thread, doing the work in a 5577 // separate thread causes wide variance in run times. We can't help this 5578 // in the multi-threaded case, but we special-case n=1 here to get 5579 // repeatable measurements of the 1-thread overhead of the parallel code. 5580 if (n_workers > 1) { 5581 // Make refs discovery MT-safe, if it isn't already: it may not 5582 // necessarily be so, since it's possible that we are doing 5583 // ST marking. 5584 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5585 GenCollectedHeap::StrongRootsScope srs(gch); 5586 workers->run_task(&tsk); 5587 } else { 5588 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5589 GenCollectedHeap::StrongRootsScope srs(gch); 5590 tsk.work(0); 5591 } 5592 5593 gch->set_par_threads(0); // 0 ==> non-parallel. 5594 // restore, single-threaded for now, any preserved marks 5595 // as a result of work_q overflow 5596 restore_preserved_marks_if_any(); 5597 } 5598 5599 // Non-parallel version of remark 5600 void CMSCollector::do_remark_non_parallel() { 5601 ResourceMark rm; 5602 HandleMark hm; 5603 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5604 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5605 5606 MarkRefsIntoAndScanClosure 5607 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5608 &_markStack, this, 5609 false /* should_yield */, false /* not precleaning */); 5610 MarkFromDirtyCardsClosure 5611 markFromDirtyCardsClosure(this, _span, 5612 NULL, // space is set further below 5613 &_markBitMap, &_markStack, &mrias_cl); 5614 { 5615 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5616 // Iterate over the dirty cards, setting the corresponding bits in the 5617 // mod union table. 5618 { 5619 ModUnionClosure modUnionClosure(&_modUnionTable); 5620 _ct->ct_bs()->dirty_card_iterate( 5621 _cmsGen->used_region(), 5622 &modUnionClosure); 5623 } 5624 // Having transferred these marks into the modUnionTable, we just need 5625 // to rescan the marked objects on the dirty cards in the modUnionTable. 5626 // The initial marking may have been done during an asynchronous 5627 // collection so there may be dirty bits in the mod-union table. 5628 const int alignment = 5629 CardTableModRefBS::card_size * BitsPerWord; 5630 { 5631 // ... First handle dirty cards in CMS gen 5632 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5633 MemRegion ur = _cmsGen->used_region(); 5634 HeapWord* lb = ur.start(); 5635 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5636 MemRegion cms_span(lb, ub); 5637 _modUnionTable.dirty_range_iterate_clear(cms_span, 5638 &markFromDirtyCardsClosure); 5639 verify_work_stacks_empty(); 5640 if (PrintCMSStatistics != 0) { 5641 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5642 markFromDirtyCardsClosure.num_dirty_cards()); 5643 } 5644 } 5645 } 5646 if (VerifyDuringGC && 5647 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5648 HandleMark hm; // Discard invalid handles created during verification 5649 Universe::verify(); 5650 } 5651 { 5652 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5653 5654 verify_work_stacks_empty(); 5655 5656 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5657 GenCollectedHeap::StrongRootsScope srs(gch); 5658 5659 gch->gen_process_roots(_cmsGen->level(), 5660 true, // younger gens as roots 5661 false, // use the local StrongRootsScope 5662 SharedHeap::ScanningOption(roots_scanning_options()), 5663 should_unload_classes(), 5664 &mrias_cl, 5665 NULL, 5666 NULL); // The dirty klasses will be handled below 5667 5668 assert(should_unload_classes() 5669 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5670 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5671 } 5672 5673 { 5674 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5675 5676 verify_work_stacks_empty(); 5677 5678 // Scan all class loader data objects that might have been introduced 5679 // during concurrent marking. 5680 ResourceMark rm; 5681 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5682 for (int i = 0; i < array->length(); i++) { 5683 mrias_cl.do_class_loader_data(array->at(i)); 5684 } 5685 5686 // We don't need to keep track of new CLDs anymore. 5687 ClassLoaderDataGraph::remember_new_clds(false); 5688 5689 verify_work_stacks_empty(); 5690 } 5691 5692 { 5693 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5694 5695 verify_work_stacks_empty(); 5696 5697 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5698 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5699 5700 verify_work_stacks_empty(); 5701 } 5702 5703 // We might have added oops to ClassLoaderData::_handles during the 5704 // concurrent marking phase. These oops point to newly allocated objects 5705 // that are guaranteed to be kept alive. Either by the direct allocation 5706 // code, or when the young collector processes the roots. Hence, 5707 // we don't have to revisit the _handles block during the remark phase. 5708 5709 verify_work_stacks_empty(); 5710 // Restore evacuated mark words, if any, used for overflow list links 5711 if (!CMSOverflowEarlyRestoration) { 5712 restore_preserved_marks_if_any(); 5713 } 5714 verify_overflow_empty(); 5715 } 5716 5717 //////////////////////////////////////////////////////// 5718 // Parallel Reference Processing Task Proxy Class 5719 //////////////////////////////////////////////////////// 5720 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5721 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5722 CMSCollector* _collector; 5723 CMSBitMap* _mark_bit_map; 5724 const MemRegion _span; 5725 ProcessTask& _task; 5726 5727 public: 5728 CMSRefProcTaskProxy(ProcessTask& task, 5729 CMSCollector* collector, 5730 const MemRegion& span, 5731 CMSBitMap* mark_bit_map, 5732 AbstractWorkGang* workers, 5733 OopTaskQueueSet* task_queues): 5734 // XXX Should superclass AGTWOQ also know about AWG since it knows 5735 // about the task_queues used by the AWG? Then it could initialize 5736 // the terminator() object. See 6984287. The set_for_termination() 5737 // below is a temporary band-aid for the regression in 6984287. 5738 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5739 task_queues), 5740 _task(task), 5741 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5742 { 5743 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5744 "Inconsistency in _span"); 5745 set_for_termination(workers->active_workers()); 5746 } 5747 5748 OopTaskQueueSet* task_queues() { return queues(); } 5749 5750 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5751 5752 void do_work_steal(int i, 5753 CMSParDrainMarkingStackClosure* drain, 5754 CMSParKeepAliveClosure* keep_alive, 5755 int* seed); 5756 5757 virtual void work(uint worker_id); 5758 }; 5759 5760 void CMSRefProcTaskProxy::work(uint worker_id) { 5761 ResourceMark rm; 5762 HandleMark hm; 5763 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5764 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5765 _mark_bit_map, 5766 work_queue(worker_id)); 5767 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5768 _mark_bit_map, 5769 work_queue(worker_id)); 5770 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5771 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5772 if (_task.marks_oops_alive()) { 5773 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5774 _collector->hash_seed(worker_id)); 5775 } 5776 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5777 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5778 } 5779 5780 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5781 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5782 EnqueueTask& _task; 5783 5784 public: 5785 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5786 : AbstractGangTask("Enqueue reference objects in parallel"), 5787 _task(task) 5788 { } 5789 5790 virtual void work(uint worker_id) 5791 { 5792 _task.work(worker_id); 5793 } 5794 }; 5795 5796 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5797 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5798 _span(span), 5799 _bit_map(bit_map), 5800 _work_queue(work_queue), 5801 _mark_and_push(collector, span, bit_map, work_queue), 5802 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5803 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5804 { } 5805 5806 // . see if we can share work_queues with ParNew? XXX 5807 void CMSRefProcTaskProxy::do_work_steal(int i, 5808 CMSParDrainMarkingStackClosure* drain, 5809 CMSParKeepAliveClosure* keep_alive, 5810 int* seed) { 5811 OopTaskQueue* work_q = work_queue(i); 5812 NOT_PRODUCT(int num_steals = 0;) 5813 oop obj_to_scan; 5814 5815 while (true) { 5816 // Completely finish any left over work from (an) earlier round(s) 5817 drain->trim_queue(0); 5818 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5819 (size_t)ParGCDesiredObjsFromOverflowList); 5820 // Now check if there's any work in the overflow list 5821 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5822 // only affects the number of attempts made to get work from the 5823 // overflow list and does not affect the number of workers. Just 5824 // pass ParallelGCThreads so this behavior is unchanged. 5825 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5826 work_q, 5827 ParallelGCThreads)) { 5828 // Found something in global overflow list; 5829 // not yet ready to go stealing work from others. 5830 // We'd like to assert(work_q->size() != 0, ...) 5831 // because we just took work from the overflow list, 5832 // but of course we can't, since all of that might have 5833 // been already stolen from us. 5834 continue; 5835 } 5836 // Verify that we have no work before we resort to stealing 5837 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5838 // Try to steal from other queues that have work 5839 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5840 NOT_PRODUCT(num_steals++;) 5841 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5842 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5843 // Do scanning work 5844 obj_to_scan->oop_iterate(keep_alive); 5845 // Loop around, finish this work, and try to steal some more 5846 } else if (terminator()->offer_termination()) { 5847 break; // nirvana from the infinite cycle 5848 } 5849 } 5850 NOT_PRODUCT( 5851 if (PrintCMSStatistics != 0) { 5852 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5853 } 5854 ) 5855 } 5856 5857 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5858 { 5859 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5860 FlexibleWorkGang* workers = gch->workers(); 5861 assert(workers != NULL, "Need parallel worker threads."); 5862 CMSRefProcTaskProxy rp_task(task, &_collector, 5863 _collector.ref_processor()->span(), 5864 _collector.markBitMap(), 5865 workers, _collector.task_queues()); 5866 workers->run_task(&rp_task); 5867 } 5868 5869 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5870 { 5871 5872 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5873 FlexibleWorkGang* workers = gch->workers(); 5874 assert(workers != NULL, "Need parallel worker threads."); 5875 CMSRefEnqueueTaskProxy enq_task(task); 5876 workers->run_task(&enq_task); 5877 } 5878 5879 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5880 5881 ResourceMark rm; 5882 HandleMark hm; 5883 5884 ReferenceProcessor* rp = ref_processor(); 5885 assert(rp->span().equals(_span), "Spans should be equal"); 5886 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5887 // Process weak references. 5888 rp->setup_policy(clear_all_soft_refs); 5889 verify_work_stacks_empty(); 5890 5891 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5892 &_markStack, false /* !preclean */); 5893 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5894 _span, &_markBitMap, &_markStack, 5895 &cmsKeepAliveClosure, false /* !preclean */); 5896 { 5897 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5898 5899 ReferenceProcessorStats stats; 5900 if (rp->processing_is_mt()) { 5901 // Set the degree of MT here. If the discovery is done MT, there 5902 // may have been a different number of threads doing the discovery 5903 // and a different number of discovered lists may have Ref objects. 5904 // That is OK as long as the Reference lists are balanced (see 5905 // balance_all_queues() and balance_queues()). 5906 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5907 int active_workers = ParallelGCThreads; 5908 FlexibleWorkGang* workers = gch->workers(); 5909 if (workers != NULL) { 5910 active_workers = workers->active_workers(); 5911 // The expectation is that active_workers will have already 5912 // been set to a reasonable value. If it has not been set, 5913 // investigate. 5914 assert(active_workers > 0, "Should have been set during scavenge"); 5915 } 5916 rp->set_active_mt_degree(active_workers); 5917 CMSRefProcTaskExecutor task_executor(*this); 5918 stats = rp->process_discovered_references(&_is_alive_closure, 5919 &cmsKeepAliveClosure, 5920 &cmsDrainMarkingStackClosure, 5921 &task_executor, 5922 _gc_timer_cm, 5923 _gc_tracer_cm->gc_id()); 5924 } else { 5925 stats = rp->process_discovered_references(&_is_alive_closure, 5926 &cmsKeepAliveClosure, 5927 &cmsDrainMarkingStackClosure, 5928 NULL, 5929 _gc_timer_cm, 5930 _gc_tracer_cm->gc_id()); 5931 } 5932 _gc_tracer_cm->report_gc_reference_stats(stats); 5933 5934 } 5935 5936 // This is the point where the entire marking should have completed. 5937 verify_work_stacks_empty(); 5938 5939 if (should_unload_classes()) { 5940 { 5941 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5942 5943 // Unload classes and purge the SystemDictionary. 5944 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5945 5946 // Unload nmethods. 5947 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5948 5949 // Prune dead klasses from subklass/sibling/implementor lists. 5950 Klass::clean_weak_klass_links(&_is_alive_closure); 5951 } 5952 5953 { 5954 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5955 // Clean up unreferenced symbols in symbol table. 5956 SymbolTable::unlink(); 5957 } 5958 5959 { 5960 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5961 // Delete entries for dead interned strings. 5962 StringTable::unlink(&_is_alive_closure); 5963 } 5964 } 5965 5966 5967 // Restore any preserved marks as a result of mark stack or 5968 // work queue overflow 5969 restore_preserved_marks_if_any(); // done single-threaded for now 5970 5971 rp->set_enqueuing_is_done(true); 5972 if (rp->processing_is_mt()) { 5973 rp->balance_all_queues(); 5974 CMSRefProcTaskExecutor task_executor(*this); 5975 rp->enqueue_discovered_references(&task_executor); 5976 } else { 5977 rp->enqueue_discovered_references(NULL); 5978 } 5979 rp->verify_no_references_recorded(); 5980 assert(!rp->discovery_enabled(), "should have been disabled"); 5981 } 5982 5983 #ifndef PRODUCT 5984 void CMSCollector::check_correct_thread_executing() { 5985 Thread* t = Thread::current(); 5986 // Only the VM thread or the CMS thread should be here. 5987 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5988 "Unexpected thread type"); 5989 // If this is the vm thread, the foreground process 5990 // should not be waiting. Note that _foregroundGCIsActive is 5991 // true while the foreground collector is waiting. 5992 if (_foregroundGCShouldWait) { 5993 // We cannot be the VM thread 5994 assert(t->is_ConcurrentGC_thread(), 5995 "Should be CMS thread"); 5996 } else { 5997 // We can be the CMS thread only if we are in a stop-world 5998 // phase of CMS collection. 5999 if (t->is_ConcurrentGC_thread()) { 6000 assert(_collectorState == InitialMarking || 6001 _collectorState == FinalMarking, 6002 "Should be a stop-world phase"); 6003 // The CMS thread should be holding the CMS_token. 6004 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6005 "Potential interference with concurrently " 6006 "executing VM thread"); 6007 } 6008 } 6009 } 6010 #endif 6011 6012 void CMSCollector::sweep(bool asynch) { 6013 assert(_collectorState == Sweeping, "just checking"); 6014 check_correct_thread_executing(); 6015 verify_work_stacks_empty(); 6016 verify_overflow_empty(); 6017 increment_sweep_count(); 6018 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6019 6020 _inter_sweep_timer.stop(); 6021 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6022 6023 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6024 _intra_sweep_timer.reset(); 6025 _intra_sweep_timer.start(); 6026 if (asynch) { 6027 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6028 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6029 // First sweep the old gen 6030 { 6031 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6032 bitMapLock()); 6033 sweepWork(_cmsGen, asynch); 6034 } 6035 6036 // Update Universe::_heap_*_at_gc figures. 6037 // We need all the free list locks to make the abstract state 6038 // transition from Sweeping to Resetting. See detailed note 6039 // further below. 6040 { 6041 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6042 // Update heap occupancy information which is used as 6043 // input to soft ref clearing policy at the next gc. 6044 Universe::update_heap_info_at_gc(); 6045 _collectorState = Resizing; 6046 } 6047 } else { 6048 // already have needed locks 6049 sweepWork(_cmsGen, asynch); 6050 // Update heap occupancy information which is used as 6051 // input to soft ref clearing policy at the next gc. 6052 Universe::update_heap_info_at_gc(); 6053 _collectorState = Resizing; 6054 } 6055 verify_work_stacks_empty(); 6056 verify_overflow_empty(); 6057 6058 if (should_unload_classes()) { 6059 // Delay purge to the beginning of the next safepoint. Metaspace::contains 6060 // requires that the virtual spaces are stable and not deleted. 6061 ClassLoaderDataGraph::set_should_purge(true); 6062 } 6063 6064 _intra_sweep_timer.stop(); 6065 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6066 6067 _inter_sweep_timer.reset(); 6068 _inter_sweep_timer.start(); 6069 6070 // We need to use a monotonically non-decreasing time in ms 6071 // or we will see time-warp warnings and os::javaTimeMillis() 6072 // does not guarantee monotonicity. 6073 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6074 update_time_of_last_gc(now); 6075 6076 // NOTE on abstract state transitions: 6077 // Mutators allocate-live and/or mark the mod-union table dirty 6078 // based on the state of the collection. The former is done in 6079 // the interval [Marking, Sweeping] and the latter in the interval 6080 // [Marking, Sweeping). Thus the transitions into the Marking state 6081 // and out of the Sweeping state must be synchronously visible 6082 // globally to the mutators. 6083 // The transition into the Marking state happens with the world 6084 // stopped so the mutators will globally see it. Sweeping is 6085 // done asynchronously by the background collector so the transition 6086 // from the Sweeping state to the Resizing state must be done 6087 // under the freelistLock (as is the check for whether to 6088 // allocate-live and whether to dirty the mod-union table). 6089 assert(_collectorState == Resizing, "Change of collector state to" 6090 " Resizing must be done under the freelistLocks (plural)"); 6091 6092 // Now that sweeping has been completed, we clear 6093 // the incremental_collection_failed flag, 6094 // thus inviting a younger gen collection to promote into 6095 // this generation. If such a promotion may still fail, 6096 // the flag will be set again when a young collection is 6097 // attempted. 6098 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6099 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6100 gch->update_full_collections_completed(_collection_count_start); 6101 } 6102 6103 // FIX ME!!! Looks like this belongs in CFLSpace, with 6104 // CMSGen merely delegating to it. 6105 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6106 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6107 HeapWord* minAddr = _cmsSpace->bottom(); 6108 HeapWord* largestAddr = 6109 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6110 if (largestAddr == NULL) { 6111 // The dictionary appears to be empty. In this case 6112 // try to coalesce at the end of the heap. 6113 largestAddr = _cmsSpace->end(); 6114 } 6115 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6116 size_t nearLargestOffset = 6117 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6118 if (PrintFLSStatistics != 0) { 6119 gclog_or_tty->print_cr( 6120 "CMS: Large Block: " PTR_FORMAT ";" 6121 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6122 largestAddr, 6123 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6124 } 6125 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6126 } 6127 6128 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6129 return addr >= _cmsSpace->nearLargestChunk(); 6130 } 6131 6132 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6133 return _cmsSpace->find_chunk_at_end(); 6134 } 6135 6136 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6137 bool full) { 6138 // The next lower level has been collected. Gather any statistics 6139 // that are of interest at this point. 6140 if (!full && (current_level + 1) == level()) { 6141 // Gather statistics on the young generation collection. 6142 collector()->stats().record_gc0_end(used()); 6143 } 6144 } 6145 6146 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6147 if (PrintGCDetails && Verbose) { 6148 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6149 } 6150 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6151 _debug_collection_type = 6152 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6153 if (PrintGCDetails && Verbose) { 6154 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6155 } 6156 } 6157 6158 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6159 bool asynch) { 6160 // We iterate over the space(s) underlying this generation, 6161 // checking the mark bit map to see if the bits corresponding 6162 // to specific blocks are marked or not. Blocks that are 6163 // marked are live and are not swept up. All remaining blocks 6164 // are swept up, with coalescing on-the-fly as we sweep up 6165 // contiguous free and/or garbage blocks: 6166 // We need to ensure that the sweeper synchronizes with allocators 6167 // and stop-the-world collectors. In particular, the following 6168 // locks are used: 6169 // . CMS token: if this is held, a stop the world collection cannot occur 6170 // . freelistLock: if this is held no allocation can occur from this 6171 // generation by another thread 6172 // . bitMapLock: if this is held, no other thread can access or update 6173 // 6174 6175 // Note that we need to hold the freelistLock if we use 6176 // block iterate below; else the iterator might go awry if 6177 // a mutator (or promotion) causes block contents to change 6178 // (for instance if the allocator divvies up a block). 6179 // If we hold the free list lock, for all practical purposes 6180 // young generation GC's can't occur (they'll usually need to 6181 // promote), so we might as well prevent all young generation 6182 // GC's while we do a sweeping step. For the same reason, we might 6183 // as well take the bit map lock for the entire duration 6184 6185 // check that we hold the requisite locks 6186 assert(have_cms_token(), "Should hold cms token"); 6187 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6188 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6189 "Should possess CMS token to sweep"); 6190 assert_lock_strong(gen->freelistLock()); 6191 assert_lock_strong(bitMapLock()); 6192 6193 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6194 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6195 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6196 _inter_sweep_estimate.padded_average(), 6197 _intra_sweep_estimate.padded_average()); 6198 gen->setNearLargestChunk(); 6199 6200 { 6201 SweepClosure sweepClosure(this, gen, &_markBitMap, 6202 CMSYield && asynch); 6203 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6204 // We need to free-up/coalesce garbage/blocks from a 6205 // co-terminal free run. This is done in the SweepClosure 6206 // destructor; so, do not remove this scope, else the 6207 // end-of-sweep-census below will be off by a little bit. 6208 } 6209 gen->cmsSpace()->sweep_completed(); 6210 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6211 if (should_unload_classes()) { // unloaded classes this cycle, 6212 _concurrent_cycles_since_last_unload = 0; // ... reset count 6213 } else { // did not unload classes, 6214 _concurrent_cycles_since_last_unload++; // ... increment count 6215 } 6216 } 6217 6218 // Reset CMS data structures (for now just the marking bit map) 6219 // preparatory for the next cycle. 6220 void CMSCollector::reset(bool asynch) { 6221 if (asynch) { 6222 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6223 6224 // If the state is not "Resetting", the foreground thread 6225 // has done a collection and the resetting. 6226 if (_collectorState != Resetting) { 6227 assert(_collectorState == Idling, "The state should only change" 6228 " because the foreground collector has finished the collection"); 6229 return; 6230 } 6231 6232 // Clear the mark bitmap (no grey objects to start with) 6233 // for the next cycle. 6234 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6235 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6236 6237 HeapWord* curAddr = _markBitMap.startWord(); 6238 while (curAddr < _markBitMap.endWord()) { 6239 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6240 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6241 _markBitMap.clear_large_range(chunk); 6242 if (ConcurrentMarkSweepThread::should_yield() && 6243 !foregroundGCIsActive() && 6244 CMSYield) { 6245 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6246 "CMS thread should hold CMS token"); 6247 assert_lock_strong(bitMapLock()); 6248 bitMapLock()->unlock(); 6249 ConcurrentMarkSweepThread::desynchronize(true); 6250 stopTimer(); 6251 if (PrintCMSStatistics != 0) { 6252 incrementYields(); 6253 } 6254 6255 // See the comment in coordinator_yield() 6256 for (unsigned i = 0; i < CMSYieldSleepCount && 6257 ConcurrentMarkSweepThread::should_yield() && 6258 !CMSCollector::foregroundGCIsActive(); ++i) { 6259 os::sleep(Thread::current(), 1, false); 6260 } 6261 6262 ConcurrentMarkSweepThread::synchronize(true); 6263 bitMapLock()->lock_without_safepoint_check(); 6264 startTimer(); 6265 } 6266 curAddr = chunk.end(); 6267 } 6268 // A successful mostly concurrent collection has been done. 6269 // Because only the full (i.e., concurrent mode failure) collections 6270 // are being measured for gc overhead limits, clean the "near" flag 6271 // and count. 6272 size_policy()->reset_gc_overhead_limit_count(); 6273 _collectorState = Idling; 6274 } else { 6275 // already have the lock 6276 assert(_collectorState == Resetting, "just checking"); 6277 assert_lock_strong(bitMapLock()); 6278 _markBitMap.clear_all(); 6279 _collectorState = Idling; 6280 } 6281 6282 NOT_PRODUCT( 6283 if (RotateCMSCollectionTypes) { 6284 _cmsGen->rotate_debug_collection_type(); 6285 } 6286 ) 6287 6288 register_gc_end(); 6289 } 6290 6291 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6292 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6293 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6294 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 6295 TraceCollectorStats tcs(counters()); 6296 6297 switch (op) { 6298 case CMS_op_checkpointRootsInitial: { 6299 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6300 checkpointRootsInitial(true); // asynch 6301 if (PrintGC) { 6302 _cmsGen->printOccupancy("initial-mark"); 6303 } 6304 break; 6305 } 6306 case CMS_op_checkpointRootsFinal: { 6307 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6308 checkpointRootsFinal(true, // asynch 6309 false, // !clear_all_soft_refs 6310 false); // !init_mark_was_synchronous 6311 if (PrintGC) { 6312 _cmsGen->printOccupancy("remark"); 6313 } 6314 break; 6315 } 6316 default: 6317 fatal("No such CMS_op"); 6318 } 6319 } 6320 6321 #ifndef PRODUCT 6322 size_t const CMSCollector::skip_header_HeapWords() { 6323 return FreeChunk::header_size(); 6324 } 6325 6326 // Try and collect here conditions that should hold when 6327 // CMS thread is exiting. The idea is that the foreground GC 6328 // thread should not be blocked if it wants to terminate 6329 // the CMS thread and yet continue to run the VM for a while 6330 // after that. 6331 void CMSCollector::verify_ok_to_terminate() const { 6332 assert(Thread::current()->is_ConcurrentGC_thread(), 6333 "should be called by CMS thread"); 6334 assert(!_foregroundGCShouldWait, "should be false"); 6335 // We could check here that all the various low-level locks 6336 // are not held by the CMS thread, but that is overkill; see 6337 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6338 // is checked. 6339 } 6340 #endif 6341 6342 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6343 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6344 "missing Printezis mark?"); 6345 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6346 size_t size = pointer_delta(nextOneAddr + 1, addr); 6347 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6348 "alignment problem"); 6349 assert(size >= 3, "Necessary for Printezis marks to work"); 6350 return size; 6351 } 6352 6353 // A variant of the above (block_size_using_printezis_bits()) except 6354 // that we return 0 if the P-bits are not yet set. 6355 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6356 if (_markBitMap.isMarked(addr + 1)) { 6357 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6358 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6359 size_t size = pointer_delta(nextOneAddr + 1, addr); 6360 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6361 "alignment problem"); 6362 assert(size >= 3, "Necessary for Printezis marks to work"); 6363 return size; 6364 } 6365 return 0; 6366 } 6367 6368 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6369 size_t sz = 0; 6370 oop p = (oop)addr; 6371 if (p->klass_or_null() != NULL) { 6372 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6373 } else { 6374 sz = block_size_using_printezis_bits(addr); 6375 } 6376 assert(sz > 0, "size must be nonzero"); 6377 HeapWord* next_block = addr + sz; 6378 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6379 CardTableModRefBS::card_size); 6380 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6381 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6382 "must be different cards"); 6383 return next_card; 6384 } 6385 6386 6387 // CMS Bit Map Wrapper ///////////////////////////////////////// 6388 6389 // Construct a CMS bit map infrastructure, but don't create the 6390 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6391 // further below. 6392 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6393 _bm(), 6394 _shifter(shifter), 6395 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6396 { 6397 _bmStartWord = 0; 6398 _bmWordSize = 0; 6399 } 6400 6401 bool CMSBitMap::allocate(MemRegion mr) { 6402 _bmStartWord = mr.start(); 6403 _bmWordSize = mr.word_size(); 6404 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6405 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6406 if (!brs.is_reserved()) { 6407 warning("CMS bit map allocation failure"); 6408 return false; 6409 } 6410 // For now we'll just commit all of the bit map up front. 6411 // Later on we'll try to be more parsimonious with swap. 6412 if (!_virtual_space.initialize(brs, brs.size())) { 6413 warning("CMS bit map backing store failure"); 6414 return false; 6415 } 6416 assert(_virtual_space.committed_size() == brs.size(), 6417 "didn't reserve backing store for all of CMS bit map?"); 6418 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6419 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6420 _bmWordSize, "inconsistency in bit map sizing"); 6421 _bm.set_size(_bmWordSize >> _shifter); 6422 6423 // bm.clear(); // can we rely on getting zero'd memory? verify below 6424 assert(isAllClear(), 6425 "Expected zero'd memory from ReservedSpace constructor"); 6426 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6427 "consistency check"); 6428 return true; 6429 } 6430 6431 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6432 HeapWord *next_addr, *end_addr, *last_addr; 6433 assert_locked(); 6434 assert(covers(mr), "out-of-range error"); 6435 // XXX assert that start and end are appropriately aligned 6436 for (next_addr = mr.start(), end_addr = mr.end(); 6437 next_addr < end_addr; next_addr = last_addr) { 6438 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6439 last_addr = dirty_region.end(); 6440 if (!dirty_region.is_empty()) { 6441 cl->do_MemRegion(dirty_region); 6442 } else { 6443 assert(last_addr == end_addr, "program logic"); 6444 return; 6445 } 6446 } 6447 } 6448 6449 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6450 _bm.print_on_error(st, prefix); 6451 } 6452 6453 #ifndef PRODUCT 6454 void CMSBitMap::assert_locked() const { 6455 CMSLockVerifier::assert_locked(lock()); 6456 } 6457 6458 bool CMSBitMap::covers(MemRegion mr) const { 6459 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6460 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6461 "size inconsistency"); 6462 return (mr.start() >= _bmStartWord) && 6463 (mr.end() <= endWord()); 6464 } 6465 6466 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6467 return (start >= _bmStartWord && (start + size) <= endWord()); 6468 } 6469 6470 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6471 // verify that there are no 1 bits in the interval [left, right) 6472 FalseBitMapClosure falseBitMapClosure; 6473 iterate(&falseBitMapClosure, left, right); 6474 } 6475 6476 void CMSBitMap::region_invariant(MemRegion mr) 6477 { 6478 assert_locked(); 6479 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6480 assert(!mr.is_empty(), "unexpected empty region"); 6481 assert(covers(mr), "mr should be covered by bit map"); 6482 // convert address range into offset range 6483 size_t start_ofs = heapWordToOffset(mr.start()); 6484 // Make sure that end() is appropriately aligned 6485 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6486 (1 << (_shifter+LogHeapWordSize))), 6487 "Misaligned mr.end()"); 6488 size_t end_ofs = heapWordToOffset(mr.end()); 6489 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6490 } 6491 6492 #endif 6493 6494 bool CMSMarkStack::allocate(size_t size) { 6495 // allocate a stack of the requisite depth 6496 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6497 size * sizeof(oop))); 6498 if (!rs.is_reserved()) { 6499 warning("CMSMarkStack allocation failure"); 6500 return false; 6501 } 6502 if (!_virtual_space.initialize(rs, rs.size())) { 6503 warning("CMSMarkStack backing store failure"); 6504 return false; 6505 } 6506 assert(_virtual_space.committed_size() == rs.size(), 6507 "didn't reserve backing store for all of CMS stack?"); 6508 _base = (oop*)(_virtual_space.low()); 6509 _index = 0; 6510 _capacity = size; 6511 NOT_PRODUCT(_max_depth = 0); 6512 return true; 6513 } 6514 6515 // XXX FIX ME !!! In the MT case we come in here holding a 6516 // leaf lock. For printing we need to take a further lock 6517 // which has lower rank. We need to recalibrate the two 6518 // lock-ranks involved in order to be able to print the 6519 // messages below. (Or defer the printing to the caller. 6520 // For now we take the expedient path of just disabling the 6521 // messages for the problematic case.) 6522 void CMSMarkStack::expand() { 6523 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6524 if (_capacity == MarkStackSizeMax) { 6525 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6526 // We print a warning message only once per CMS cycle. 6527 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6528 } 6529 return; 6530 } 6531 // Double capacity if possible 6532 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6533 // Do not give up existing stack until we have managed to 6534 // get the double capacity that we desired. 6535 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6536 new_capacity * sizeof(oop))); 6537 if (rs.is_reserved()) { 6538 // Release the backing store associated with old stack 6539 _virtual_space.release(); 6540 // Reinitialize virtual space for new stack 6541 if (!_virtual_space.initialize(rs, rs.size())) { 6542 fatal("Not enough swap for expanded marking stack"); 6543 } 6544 _base = (oop*)(_virtual_space.low()); 6545 _index = 0; 6546 _capacity = new_capacity; 6547 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6548 // Failed to double capacity, continue; 6549 // we print a detail message only once per CMS cycle. 6550 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6551 SIZE_FORMAT"K", 6552 _capacity / K, new_capacity / K); 6553 } 6554 } 6555 6556 6557 // Closures 6558 // XXX: there seems to be a lot of code duplication here; 6559 // should refactor and consolidate common code. 6560 6561 // This closure is used to mark refs into the CMS generation in 6562 // the CMS bit map. Called at the first checkpoint. This closure 6563 // assumes that we do not need to re-mark dirty cards; if the CMS 6564 // generation on which this is used is not an oldest 6565 // generation then this will lose younger_gen cards! 6566 6567 MarkRefsIntoClosure::MarkRefsIntoClosure( 6568 MemRegion span, CMSBitMap* bitMap): 6569 _span(span), 6570 _bitMap(bitMap) 6571 { 6572 assert(_ref_processor == NULL, "deliberately left NULL"); 6573 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6574 } 6575 6576 void MarkRefsIntoClosure::do_oop(oop obj) { 6577 // if p points into _span, then mark corresponding bit in _markBitMap 6578 assert(obj->is_oop(), "expected an oop"); 6579 HeapWord* addr = (HeapWord*)obj; 6580 if (_span.contains(addr)) { 6581 // this should be made more efficient 6582 _bitMap->mark(addr); 6583 } 6584 } 6585 6586 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6587 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6588 6589 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6590 MemRegion span, CMSBitMap* bitMap): 6591 _span(span), 6592 _bitMap(bitMap) 6593 { 6594 assert(_ref_processor == NULL, "deliberately left NULL"); 6595 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6596 } 6597 6598 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6599 // if p points into _span, then mark corresponding bit in _markBitMap 6600 assert(obj->is_oop(), "expected an oop"); 6601 HeapWord* addr = (HeapWord*)obj; 6602 if (_span.contains(addr)) { 6603 // this should be made more efficient 6604 _bitMap->par_mark(addr); 6605 } 6606 } 6607 6608 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6609 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6610 6611 // A variant of the above, used for CMS marking verification. 6612 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6613 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6614 _span(span), 6615 _verification_bm(verification_bm), 6616 _cms_bm(cms_bm) 6617 { 6618 assert(_ref_processor == NULL, "deliberately left NULL"); 6619 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6620 } 6621 6622 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6623 // if p points into _span, then mark corresponding bit in _markBitMap 6624 assert(obj->is_oop(), "expected an oop"); 6625 HeapWord* addr = (HeapWord*)obj; 6626 if (_span.contains(addr)) { 6627 _verification_bm->mark(addr); 6628 if (!_cms_bm->isMarked(addr)) { 6629 oop(addr)->print(); 6630 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6631 fatal("... aborting"); 6632 } 6633 } 6634 } 6635 6636 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6637 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6638 6639 ////////////////////////////////////////////////// 6640 // MarkRefsIntoAndScanClosure 6641 ////////////////////////////////////////////////// 6642 6643 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6644 ReferenceProcessor* rp, 6645 CMSBitMap* bit_map, 6646 CMSBitMap* mod_union_table, 6647 CMSMarkStack* mark_stack, 6648 CMSCollector* collector, 6649 bool should_yield, 6650 bool concurrent_precleaning): 6651 _collector(collector), 6652 _span(span), 6653 _bit_map(bit_map), 6654 _mark_stack(mark_stack), 6655 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6656 mark_stack, concurrent_precleaning), 6657 _yield(should_yield), 6658 _concurrent_precleaning(concurrent_precleaning), 6659 _freelistLock(NULL) 6660 { 6661 _ref_processor = rp; 6662 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6663 } 6664 6665 // This closure is used to mark refs into the CMS generation at the 6666 // second (final) checkpoint, and to scan and transitively follow 6667 // the unmarked oops. It is also used during the concurrent precleaning 6668 // phase while scanning objects on dirty cards in the CMS generation. 6669 // The marks are made in the marking bit map and the marking stack is 6670 // used for keeping the (newly) grey objects during the scan. 6671 // The parallel version (Par_...) appears further below. 6672 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6673 if (obj != NULL) { 6674 assert(obj->is_oop(), "expected an oop"); 6675 HeapWord* addr = (HeapWord*)obj; 6676 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6677 assert(_collector->overflow_list_is_empty(), 6678 "overflow list should be empty"); 6679 if (_span.contains(addr) && 6680 !_bit_map->isMarked(addr)) { 6681 // mark bit map (object is now grey) 6682 _bit_map->mark(addr); 6683 // push on marking stack (stack should be empty), and drain the 6684 // stack by applying this closure to the oops in the oops popped 6685 // from the stack (i.e. blacken the grey objects) 6686 bool res = _mark_stack->push(obj); 6687 assert(res, "Should have space to push on empty stack"); 6688 do { 6689 oop new_oop = _mark_stack->pop(); 6690 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6691 assert(_bit_map->isMarked((HeapWord*)new_oop), 6692 "only grey objects on this stack"); 6693 // iterate over the oops in this oop, marking and pushing 6694 // the ones in CMS heap (i.e. in _span). 6695 new_oop->oop_iterate(&_pushAndMarkClosure); 6696 // check if it's time to yield 6697 do_yield_check(); 6698 } while (!_mark_stack->isEmpty() || 6699 (!_concurrent_precleaning && take_from_overflow_list())); 6700 // if marking stack is empty, and we are not doing this 6701 // during precleaning, then check the overflow list 6702 } 6703 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6704 assert(_collector->overflow_list_is_empty(), 6705 "overflow list was drained above"); 6706 // We could restore evacuated mark words, if any, used for 6707 // overflow list links here because the overflow list is 6708 // provably empty here. That would reduce the maximum 6709 // size requirements for preserved_{oop,mark}_stack. 6710 // But we'll just postpone it until we are all done 6711 // so we can just stream through. 6712 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6713 _collector->restore_preserved_marks_if_any(); 6714 assert(_collector->no_preserved_marks(), "No preserved marks"); 6715 } 6716 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6717 "All preserved marks should have been restored above"); 6718 } 6719 } 6720 6721 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6722 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6723 6724 void MarkRefsIntoAndScanClosure::do_yield_work() { 6725 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6726 "CMS thread should hold CMS token"); 6727 assert_lock_strong(_freelistLock); 6728 assert_lock_strong(_bit_map->lock()); 6729 // relinquish the free_list_lock and bitMaplock() 6730 _bit_map->lock()->unlock(); 6731 _freelistLock->unlock(); 6732 ConcurrentMarkSweepThread::desynchronize(true); 6733 _collector->stopTimer(); 6734 if (PrintCMSStatistics != 0) { 6735 _collector->incrementYields(); 6736 } 6737 6738 // See the comment in coordinator_yield() 6739 for (unsigned i = 0; 6740 i < CMSYieldSleepCount && 6741 ConcurrentMarkSweepThread::should_yield() && 6742 !CMSCollector::foregroundGCIsActive(); 6743 ++i) { 6744 os::sleep(Thread::current(), 1, false); 6745 } 6746 6747 ConcurrentMarkSweepThread::synchronize(true); 6748 _freelistLock->lock_without_safepoint_check(); 6749 _bit_map->lock()->lock_without_safepoint_check(); 6750 _collector->startTimer(); 6751 } 6752 6753 /////////////////////////////////////////////////////////// 6754 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6755 // MarkRefsIntoAndScanClosure 6756 /////////////////////////////////////////////////////////// 6757 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6758 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6759 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6760 _span(span), 6761 _bit_map(bit_map), 6762 _work_queue(work_queue), 6763 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6764 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6765 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6766 { 6767 _ref_processor = rp; 6768 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6769 } 6770 6771 // This closure is used to mark refs into the CMS generation at the 6772 // second (final) checkpoint, and to scan and transitively follow 6773 // the unmarked oops. The marks are made in the marking bit map and 6774 // the work_queue is used for keeping the (newly) grey objects during 6775 // the scan phase whence they are also available for stealing by parallel 6776 // threads. Since the marking bit map is shared, updates are 6777 // synchronized (via CAS). 6778 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6779 if (obj != NULL) { 6780 // Ignore mark word because this could be an already marked oop 6781 // that may be chained at the end of the overflow list. 6782 assert(obj->is_oop(true), "expected an oop"); 6783 HeapWord* addr = (HeapWord*)obj; 6784 if (_span.contains(addr) && 6785 !_bit_map->isMarked(addr)) { 6786 // mark bit map (object will become grey): 6787 // It is possible for several threads to be 6788 // trying to "claim" this object concurrently; 6789 // the unique thread that succeeds in marking the 6790 // object first will do the subsequent push on 6791 // to the work queue (or overflow list). 6792 if (_bit_map->par_mark(addr)) { 6793 // push on work_queue (which may not be empty), and trim the 6794 // queue to an appropriate length by applying this closure to 6795 // the oops in the oops popped from the stack (i.e. blacken the 6796 // grey objects) 6797 bool res = _work_queue->push(obj); 6798 assert(res, "Low water mark should be less than capacity?"); 6799 trim_queue(_low_water_mark); 6800 } // Else, another thread claimed the object 6801 } 6802 } 6803 } 6804 6805 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6806 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6807 6808 // This closure is used to rescan the marked objects on the dirty cards 6809 // in the mod union table and the card table proper. 6810 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6811 oop p, MemRegion mr) { 6812 6813 size_t size = 0; 6814 HeapWord* addr = (HeapWord*)p; 6815 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6816 assert(_span.contains(addr), "we are scanning the CMS generation"); 6817 // check if it's time to yield 6818 if (do_yield_check()) { 6819 // We yielded for some foreground stop-world work, 6820 // and we have been asked to abort this ongoing preclean cycle. 6821 return 0; 6822 } 6823 if (_bitMap->isMarked(addr)) { 6824 // it's marked; is it potentially uninitialized? 6825 if (p->klass_or_null() != NULL) { 6826 // an initialized object; ignore mark word in verification below 6827 // since we are running concurrent with mutators 6828 assert(p->is_oop(true), "should be an oop"); 6829 if (p->is_objArray()) { 6830 // objArrays are precisely marked; restrict scanning 6831 // to dirty cards only. 6832 size = CompactibleFreeListSpace::adjustObjectSize( 6833 p->oop_iterate(_scanningClosure, mr)); 6834 } else { 6835 // A non-array may have been imprecisely marked; we need 6836 // to scan object in its entirety. 6837 size = CompactibleFreeListSpace::adjustObjectSize( 6838 p->oop_iterate(_scanningClosure)); 6839 } 6840 #ifdef ASSERT 6841 size_t direct_size = 6842 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6843 assert(size == direct_size, "Inconsistency in size"); 6844 assert(size >= 3, "Necessary for Printezis marks to work"); 6845 if (!_bitMap->isMarked(addr+1)) { 6846 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6847 } else { 6848 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6849 assert(_bitMap->isMarked(addr+size-1), 6850 "inconsistent Printezis mark"); 6851 } 6852 #endif // ASSERT 6853 } else { 6854 // An uninitialized object. 6855 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6856 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6857 size = pointer_delta(nextOneAddr + 1, addr); 6858 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6859 "alignment problem"); 6860 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6861 // will dirty the card when the klass pointer is installed in the 6862 // object (signaling the completion of initialization). 6863 } 6864 } else { 6865 // Either a not yet marked object or an uninitialized object 6866 if (p->klass_or_null() == NULL) { 6867 // An uninitialized object, skip to the next card, since 6868 // we may not be able to read its P-bits yet. 6869 assert(size == 0, "Initial value"); 6870 } else { 6871 // An object not (yet) reached by marking: we merely need to 6872 // compute its size so as to go look at the next block. 6873 assert(p->is_oop(true), "should be an oop"); 6874 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6875 } 6876 } 6877 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6878 return size; 6879 } 6880 6881 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6882 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6883 "CMS thread should hold CMS token"); 6884 assert_lock_strong(_freelistLock); 6885 assert_lock_strong(_bitMap->lock()); 6886 // relinquish the free_list_lock and bitMaplock() 6887 _bitMap->lock()->unlock(); 6888 _freelistLock->unlock(); 6889 ConcurrentMarkSweepThread::desynchronize(true); 6890 _collector->stopTimer(); 6891 if (PrintCMSStatistics != 0) { 6892 _collector->incrementYields(); 6893 } 6894 6895 // See the comment in coordinator_yield() 6896 for (unsigned i = 0; i < CMSYieldSleepCount && 6897 ConcurrentMarkSweepThread::should_yield() && 6898 !CMSCollector::foregroundGCIsActive(); ++i) { 6899 os::sleep(Thread::current(), 1, false); 6900 } 6901 6902 ConcurrentMarkSweepThread::synchronize(true); 6903 _freelistLock->lock_without_safepoint_check(); 6904 _bitMap->lock()->lock_without_safepoint_check(); 6905 _collector->startTimer(); 6906 } 6907 6908 6909 ////////////////////////////////////////////////////////////////// 6910 // SurvivorSpacePrecleanClosure 6911 ////////////////////////////////////////////////////////////////// 6912 // This (single-threaded) closure is used to preclean the oops in 6913 // the survivor spaces. 6914 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6915 6916 HeapWord* addr = (HeapWord*)p; 6917 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6918 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6919 assert(p->klass_or_null() != NULL, "object should be initialized"); 6920 // an initialized object; ignore mark word in verification below 6921 // since we are running concurrent with mutators 6922 assert(p->is_oop(true), "should be an oop"); 6923 // Note that we do not yield while we iterate over 6924 // the interior oops of p, pushing the relevant ones 6925 // on our marking stack. 6926 size_t size = p->oop_iterate(_scanning_closure); 6927 do_yield_check(); 6928 // Observe that below, we do not abandon the preclean 6929 // phase as soon as we should; rather we empty the 6930 // marking stack before returning. This is to satisfy 6931 // some existing assertions. In general, it may be a 6932 // good idea to abort immediately and complete the marking 6933 // from the grey objects at a later time. 6934 while (!_mark_stack->isEmpty()) { 6935 oop new_oop = _mark_stack->pop(); 6936 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6937 assert(_bit_map->isMarked((HeapWord*)new_oop), 6938 "only grey objects on this stack"); 6939 // iterate over the oops in this oop, marking and pushing 6940 // the ones in CMS heap (i.e. in _span). 6941 new_oop->oop_iterate(_scanning_closure); 6942 // check if it's time to yield 6943 do_yield_check(); 6944 } 6945 unsigned int after_count = 6946 GenCollectedHeap::heap()->total_collections(); 6947 bool abort = (_before_count != after_count) || 6948 _collector->should_abort_preclean(); 6949 return abort ? 0 : size; 6950 } 6951 6952 void SurvivorSpacePrecleanClosure::do_yield_work() { 6953 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6954 "CMS thread should hold CMS token"); 6955 assert_lock_strong(_bit_map->lock()); 6956 // Relinquish the bit map lock 6957 _bit_map->lock()->unlock(); 6958 ConcurrentMarkSweepThread::desynchronize(true); 6959 _collector->stopTimer(); 6960 if (PrintCMSStatistics != 0) { 6961 _collector->incrementYields(); 6962 } 6963 6964 // See the comment in coordinator_yield() 6965 for (unsigned i = 0; i < CMSYieldSleepCount && 6966 ConcurrentMarkSweepThread::should_yield() && 6967 !CMSCollector::foregroundGCIsActive(); ++i) { 6968 os::sleep(Thread::current(), 1, false); 6969 } 6970 6971 ConcurrentMarkSweepThread::synchronize(true); 6972 _bit_map->lock()->lock_without_safepoint_check(); 6973 _collector->startTimer(); 6974 } 6975 6976 // This closure is used to rescan the marked objects on the dirty cards 6977 // in the mod union table and the card table proper. In the parallel 6978 // case, although the bitMap is shared, we do a single read so the 6979 // isMarked() query is "safe". 6980 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6981 // Ignore mark word because we are running concurrent with mutators 6982 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6983 HeapWord* addr = (HeapWord*)p; 6984 assert(_span.contains(addr), "we are scanning the CMS generation"); 6985 bool is_obj_array = false; 6986 #ifdef ASSERT 6987 if (!_parallel) { 6988 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6989 assert(_collector->overflow_list_is_empty(), 6990 "overflow list should be empty"); 6991 6992 } 6993 #endif // ASSERT 6994 if (_bit_map->isMarked(addr)) { 6995 // Obj arrays are precisely marked, non-arrays are not; 6996 // so we scan objArrays precisely and non-arrays in their 6997 // entirety. 6998 if (p->is_objArray()) { 6999 is_obj_array = true; 7000 if (_parallel) { 7001 p->oop_iterate(_par_scan_closure, mr); 7002 } else { 7003 p->oop_iterate(_scan_closure, mr); 7004 } 7005 } else { 7006 if (_parallel) { 7007 p->oop_iterate(_par_scan_closure); 7008 } else { 7009 p->oop_iterate(_scan_closure); 7010 } 7011 } 7012 } 7013 #ifdef ASSERT 7014 if (!_parallel) { 7015 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7016 assert(_collector->overflow_list_is_empty(), 7017 "overflow list should be empty"); 7018 7019 } 7020 #endif // ASSERT 7021 return is_obj_array; 7022 } 7023 7024 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7025 MemRegion span, 7026 CMSBitMap* bitMap, CMSMarkStack* markStack, 7027 bool should_yield, bool verifying): 7028 _collector(collector), 7029 _span(span), 7030 _bitMap(bitMap), 7031 _mut(&collector->_modUnionTable), 7032 _markStack(markStack), 7033 _yield(should_yield), 7034 _skipBits(0) 7035 { 7036 assert(_markStack->isEmpty(), "stack should be empty"); 7037 _finger = _bitMap->startWord(); 7038 _threshold = _finger; 7039 assert(_collector->_restart_addr == NULL, "Sanity check"); 7040 assert(_span.contains(_finger), "Out of bounds _finger?"); 7041 DEBUG_ONLY(_verifying = verifying;) 7042 } 7043 7044 void MarkFromRootsClosure::reset(HeapWord* addr) { 7045 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7046 assert(_span.contains(addr), "Out of bounds _finger?"); 7047 _finger = addr; 7048 _threshold = (HeapWord*)round_to( 7049 (intptr_t)_finger, CardTableModRefBS::card_size); 7050 } 7051 7052 // Should revisit to see if this should be restructured for 7053 // greater efficiency. 7054 bool MarkFromRootsClosure::do_bit(size_t offset) { 7055 if (_skipBits > 0) { 7056 _skipBits--; 7057 return true; 7058 } 7059 // convert offset into a HeapWord* 7060 HeapWord* addr = _bitMap->startWord() + offset; 7061 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7062 "address out of range"); 7063 assert(_bitMap->isMarked(addr), "tautology"); 7064 if (_bitMap->isMarked(addr+1)) { 7065 // this is an allocated but not yet initialized object 7066 assert(_skipBits == 0, "tautology"); 7067 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7068 oop p = oop(addr); 7069 if (p->klass_or_null() == NULL) { 7070 DEBUG_ONLY(if (!_verifying) {) 7071 // We re-dirty the cards on which this object lies and increase 7072 // the _threshold so that we'll come back to scan this object 7073 // during the preclean or remark phase. (CMSCleanOnEnter) 7074 if (CMSCleanOnEnter) { 7075 size_t sz = _collector->block_size_using_printezis_bits(addr); 7076 HeapWord* end_card_addr = (HeapWord*)round_to( 7077 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7078 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7079 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7080 // Bump _threshold to end_card_addr; note that 7081 // _threshold cannot possibly exceed end_card_addr, anyhow. 7082 // This prevents future clearing of the card as the scan proceeds 7083 // to the right. 7084 assert(_threshold <= end_card_addr, 7085 "Because we are just scanning into this object"); 7086 if (_threshold < end_card_addr) { 7087 _threshold = end_card_addr; 7088 } 7089 if (p->klass_or_null() != NULL) { 7090 // Redirty the range of cards... 7091 _mut->mark_range(redirty_range); 7092 } // ...else the setting of klass will dirty the card anyway. 7093 } 7094 DEBUG_ONLY(}) 7095 return true; 7096 } 7097 } 7098 scanOopsInOop(addr); 7099 return true; 7100 } 7101 7102 // We take a break if we've been at this for a while, 7103 // so as to avoid monopolizing the locks involved. 7104 void MarkFromRootsClosure::do_yield_work() { 7105 // First give up the locks, then yield, then re-lock 7106 // We should probably use a constructor/destructor idiom to 7107 // do this unlock/lock or modify the MutexUnlocker class to 7108 // serve our purpose. XXX 7109 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7110 "CMS thread should hold CMS token"); 7111 assert_lock_strong(_bitMap->lock()); 7112 _bitMap->lock()->unlock(); 7113 ConcurrentMarkSweepThread::desynchronize(true); 7114 _collector->stopTimer(); 7115 if (PrintCMSStatistics != 0) { 7116 _collector->incrementYields(); 7117 } 7118 7119 // See the comment in coordinator_yield() 7120 for (unsigned i = 0; i < CMSYieldSleepCount && 7121 ConcurrentMarkSweepThread::should_yield() && 7122 !CMSCollector::foregroundGCIsActive(); ++i) { 7123 os::sleep(Thread::current(), 1, false); 7124 } 7125 7126 ConcurrentMarkSweepThread::synchronize(true); 7127 _bitMap->lock()->lock_without_safepoint_check(); 7128 _collector->startTimer(); 7129 } 7130 7131 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7132 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7133 assert(_markStack->isEmpty(), 7134 "should drain stack to limit stack usage"); 7135 // convert ptr to an oop preparatory to scanning 7136 oop obj = oop(ptr); 7137 // Ignore mark word in verification below, since we 7138 // may be running concurrent with mutators. 7139 assert(obj->is_oop(true), "should be an oop"); 7140 assert(_finger <= ptr, "_finger runneth ahead"); 7141 // advance the finger to right end of this object 7142 _finger = ptr + obj->size(); 7143 assert(_finger > ptr, "we just incremented it above"); 7144 // On large heaps, it may take us some time to get through 7145 // the marking phase. During 7146 // this time it's possible that a lot of mutations have 7147 // accumulated in the card table and the mod union table -- 7148 // these mutation records are redundant until we have 7149 // actually traced into the corresponding card. 7150 // Here, we check whether advancing the finger would make 7151 // us cross into a new card, and if so clear corresponding 7152 // cards in the MUT (preclean them in the card-table in the 7153 // future). 7154 7155 DEBUG_ONLY(if (!_verifying) {) 7156 // The clean-on-enter optimization is disabled by default, 7157 // until we fix 6178663. 7158 if (CMSCleanOnEnter && (_finger > _threshold)) { 7159 // [_threshold, _finger) represents the interval 7160 // of cards to be cleared in MUT (or precleaned in card table). 7161 // The set of cards to be cleared is all those that overlap 7162 // with the interval [_threshold, _finger); note that 7163 // _threshold is always kept card-aligned but _finger isn't 7164 // always card-aligned. 7165 HeapWord* old_threshold = _threshold; 7166 assert(old_threshold == (HeapWord*)round_to( 7167 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7168 "_threshold should always be card-aligned"); 7169 _threshold = (HeapWord*)round_to( 7170 (intptr_t)_finger, CardTableModRefBS::card_size); 7171 MemRegion mr(old_threshold, _threshold); 7172 assert(!mr.is_empty(), "Control point invariant"); 7173 assert(_span.contains(mr), "Should clear within span"); 7174 _mut->clear_range(mr); 7175 } 7176 DEBUG_ONLY(}) 7177 // Note: the finger doesn't advance while we drain 7178 // the stack below. 7179 PushOrMarkClosure pushOrMarkClosure(_collector, 7180 _span, _bitMap, _markStack, 7181 _finger, this); 7182 bool res = _markStack->push(obj); 7183 assert(res, "Empty non-zero size stack should have space for single push"); 7184 while (!_markStack->isEmpty()) { 7185 oop new_oop = _markStack->pop(); 7186 // Skip verifying header mark word below because we are 7187 // running concurrent with mutators. 7188 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7189 // now scan this oop's oops 7190 new_oop->oop_iterate(&pushOrMarkClosure); 7191 do_yield_check(); 7192 } 7193 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7194 } 7195 7196 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7197 CMSCollector* collector, MemRegion span, 7198 CMSBitMap* bit_map, 7199 OopTaskQueue* work_queue, 7200 CMSMarkStack* overflow_stack, 7201 bool should_yield): 7202 _collector(collector), 7203 _whole_span(collector->_span), 7204 _span(span), 7205 _bit_map(bit_map), 7206 _mut(&collector->_modUnionTable), 7207 _work_queue(work_queue), 7208 _overflow_stack(overflow_stack), 7209 _yield(should_yield), 7210 _skip_bits(0), 7211 _task(task) 7212 { 7213 assert(_work_queue->size() == 0, "work_queue should be empty"); 7214 _finger = span.start(); 7215 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7216 assert(_span.contains(_finger), "Out of bounds _finger?"); 7217 } 7218 7219 // Should revisit to see if this should be restructured for 7220 // greater efficiency. 7221 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7222 if (_skip_bits > 0) { 7223 _skip_bits--; 7224 return true; 7225 } 7226 // convert offset into a HeapWord* 7227 HeapWord* addr = _bit_map->startWord() + offset; 7228 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7229 "address out of range"); 7230 assert(_bit_map->isMarked(addr), "tautology"); 7231 if (_bit_map->isMarked(addr+1)) { 7232 // this is an allocated object that might not yet be initialized 7233 assert(_skip_bits == 0, "tautology"); 7234 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7235 oop p = oop(addr); 7236 if (p->klass_or_null() == NULL) { 7237 // in the case of Clean-on-Enter optimization, redirty card 7238 // and avoid clearing card by increasing the threshold. 7239 return true; 7240 } 7241 } 7242 scan_oops_in_oop(addr); 7243 return true; 7244 } 7245 7246 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7247 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7248 // Should we assert that our work queue is empty or 7249 // below some drain limit? 7250 assert(_work_queue->size() == 0, 7251 "should drain stack to limit stack usage"); 7252 // convert ptr to an oop preparatory to scanning 7253 oop obj = oop(ptr); 7254 // Ignore mark word in verification below, since we 7255 // may be running concurrent with mutators. 7256 assert(obj->is_oop(true), "should be an oop"); 7257 assert(_finger <= ptr, "_finger runneth ahead"); 7258 // advance the finger to right end of this object 7259 _finger = ptr + obj->size(); 7260 assert(_finger > ptr, "we just incremented it above"); 7261 // On large heaps, it may take us some time to get through 7262 // the marking phase. During 7263 // this time it's possible that a lot of mutations have 7264 // accumulated in the card table and the mod union table -- 7265 // these mutation records are redundant until we have 7266 // actually traced into the corresponding card. 7267 // Here, we check whether advancing the finger would make 7268 // us cross into a new card, and if so clear corresponding 7269 // cards in the MUT (preclean them in the card-table in the 7270 // future). 7271 7272 // The clean-on-enter optimization is disabled by default, 7273 // until we fix 6178663. 7274 if (CMSCleanOnEnter && (_finger > _threshold)) { 7275 // [_threshold, _finger) represents the interval 7276 // of cards to be cleared in MUT (or precleaned in card table). 7277 // The set of cards to be cleared is all those that overlap 7278 // with the interval [_threshold, _finger); note that 7279 // _threshold is always kept card-aligned but _finger isn't 7280 // always card-aligned. 7281 HeapWord* old_threshold = _threshold; 7282 assert(old_threshold == (HeapWord*)round_to( 7283 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7284 "_threshold should always be card-aligned"); 7285 _threshold = (HeapWord*)round_to( 7286 (intptr_t)_finger, CardTableModRefBS::card_size); 7287 MemRegion mr(old_threshold, _threshold); 7288 assert(!mr.is_empty(), "Control point invariant"); 7289 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7290 _mut->clear_range(mr); 7291 } 7292 7293 // Note: the local finger doesn't advance while we drain 7294 // the stack below, but the global finger sure can and will. 7295 HeapWord** gfa = _task->global_finger_addr(); 7296 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7297 _span, _bit_map, 7298 _work_queue, 7299 _overflow_stack, 7300 _finger, 7301 gfa, this); 7302 bool res = _work_queue->push(obj); // overflow could occur here 7303 assert(res, "Will hold once we use workqueues"); 7304 while (true) { 7305 oop new_oop; 7306 if (!_work_queue->pop_local(new_oop)) { 7307 // We emptied our work_queue; check if there's stuff that can 7308 // be gotten from the overflow stack. 7309 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7310 _overflow_stack, _work_queue)) { 7311 do_yield_check(); 7312 continue; 7313 } else { // done 7314 break; 7315 } 7316 } 7317 // Skip verifying header mark word below because we are 7318 // running concurrent with mutators. 7319 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7320 // now scan this oop's oops 7321 new_oop->oop_iterate(&pushOrMarkClosure); 7322 do_yield_check(); 7323 } 7324 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7325 } 7326 7327 // Yield in response to a request from VM Thread or 7328 // from mutators. 7329 void Par_MarkFromRootsClosure::do_yield_work() { 7330 assert(_task != NULL, "sanity"); 7331 _task->yield(); 7332 } 7333 7334 // A variant of the above used for verifying CMS marking work. 7335 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7336 MemRegion span, 7337 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7338 CMSMarkStack* mark_stack): 7339 _collector(collector), 7340 _span(span), 7341 _verification_bm(verification_bm), 7342 _cms_bm(cms_bm), 7343 _mark_stack(mark_stack), 7344 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7345 mark_stack) 7346 { 7347 assert(_mark_stack->isEmpty(), "stack should be empty"); 7348 _finger = _verification_bm->startWord(); 7349 assert(_collector->_restart_addr == NULL, "Sanity check"); 7350 assert(_span.contains(_finger), "Out of bounds _finger?"); 7351 } 7352 7353 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7354 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7355 assert(_span.contains(addr), "Out of bounds _finger?"); 7356 _finger = addr; 7357 } 7358 7359 // Should revisit to see if this should be restructured for 7360 // greater efficiency. 7361 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7362 // convert offset into a HeapWord* 7363 HeapWord* addr = _verification_bm->startWord() + offset; 7364 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7365 "address out of range"); 7366 assert(_verification_bm->isMarked(addr), "tautology"); 7367 assert(_cms_bm->isMarked(addr), "tautology"); 7368 7369 assert(_mark_stack->isEmpty(), 7370 "should drain stack to limit stack usage"); 7371 // convert addr to an oop preparatory to scanning 7372 oop obj = oop(addr); 7373 assert(obj->is_oop(), "should be an oop"); 7374 assert(_finger <= addr, "_finger runneth ahead"); 7375 // advance the finger to right end of this object 7376 _finger = addr + obj->size(); 7377 assert(_finger > addr, "we just incremented it above"); 7378 // Note: the finger doesn't advance while we drain 7379 // the stack below. 7380 bool res = _mark_stack->push(obj); 7381 assert(res, "Empty non-zero size stack should have space for single push"); 7382 while (!_mark_stack->isEmpty()) { 7383 oop new_oop = _mark_stack->pop(); 7384 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7385 // now scan this oop's oops 7386 new_oop->oop_iterate(&_pam_verify_closure); 7387 } 7388 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7389 return true; 7390 } 7391 7392 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7393 CMSCollector* collector, MemRegion span, 7394 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7395 CMSMarkStack* mark_stack): 7396 MetadataAwareOopClosure(collector->ref_processor()), 7397 _collector(collector), 7398 _span(span), 7399 _verification_bm(verification_bm), 7400 _cms_bm(cms_bm), 7401 _mark_stack(mark_stack) 7402 { } 7403 7404 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7405 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7406 7407 // Upon stack overflow, we discard (part of) the stack, 7408 // remembering the least address amongst those discarded 7409 // in CMSCollector's _restart_address. 7410 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7411 // Remember the least grey address discarded 7412 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7413 _collector->lower_restart_addr(ra); 7414 _mark_stack->reset(); // discard stack contents 7415 _mark_stack->expand(); // expand the stack if possible 7416 } 7417 7418 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7419 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7420 HeapWord* addr = (HeapWord*)obj; 7421 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7422 // Oop lies in _span and isn't yet grey or black 7423 _verification_bm->mark(addr); // now grey 7424 if (!_cms_bm->isMarked(addr)) { 7425 oop(addr)->print(); 7426 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7427 addr); 7428 fatal("... aborting"); 7429 } 7430 7431 if (!_mark_stack->push(obj)) { // stack overflow 7432 if (PrintCMSStatistics != 0) { 7433 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7434 SIZE_FORMAT, _mark_stack->capacity()); 7435 } 7436 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7437 handle_stack_overflow(addr); 7438 } 7439 // anything including and to the right of _finger 7440 // will be scanned as we iterate over the remainder of the 7441 // bit map 7442 } 7443 } 7444 7445 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7446 MemRegion span, 7447 CMSBitMap* bitMap, CMSMarkStack* markStack, 7448 HeapWord* finger, MarkFromRootsClosure* parent) : 7449 MetadataAwareOopClosure(collector->ref_processor()), 7450 _collector(collector), 7451 _span(span), 7452 _bitMap(bitMap), 7453 _markStack(markStack), 7454 _finger(finger), 7455 _parent(parent) 7456 { } 7457 7458 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7459 MemRegion span, 7460 CMSBitMap* bit_map, 7461 OopTaskQueue* work_queue, 7462 CMSMarkStack* overflow_stack, 7463 HeapWord* finger, 7464 HeapWord** global_finger_addr, 7465 Par_MarkFromRootsClosure* parent) : 7466 MetadataAwareOopClosure(collector->ref_processor()), 7467 _collector(collector), 7468 _whole_span(collector->_span), 7469 _span(span), 7470 _bit_map(bit_map), 7471 _work_queue(work_queue), 7472 _overflow_stack(overflow_stack), 7473 _finger(finger), 7474 _global_finger_addr(global_finger_addr), 7475 _parent(parent) 7476 { } 7477 7478 // Assumes thread-safe access by callers, who are 7479 // responsible for mutual exclusion. 7480 void CMSCollector::lower_restart_addr(HeapWord* low) { 7481 assert(_span.contains(low), "Out of bounds addr"); 7482 if (_restart_addr == NULL) { 7483 _restart_addr = low; 7484 } else { 7485 _restart_addr = MIN2(_restart_addr, low); 7486 } 7487 } 7488 7489 // Upon stack overflow, we discard (part of) the stack, 7490 // remembering the least address amongst those discarded 7491 // in CMSCollector's _restart_address. 7492 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7493 // Remember the least grey address discarded 7494 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7495 _collector->lower_restart_addr(ra); 7496 _markStack->reset(); // discard stack contents 7497 _markStack->expand(); // expand the stack if possible 7498 } 7499 7500 // Upon stack overflow, we discard (part of) the stack, 7501 // remembering the least address amongst those discarded 7502 // in CMSCollector's _restart_address. 7503 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7504 // We need to do this under a mutex to prevent other 7505 // workers from interfering with the work done below. 7506 MutexLockerEx ml(_overflow_stack->par_lock(), 7507 Mutex::_no_safepoint_check_flag); 7508 // Remember the least grey address discarded 7509 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7510 _collector->lower_restart_addr(ra); 7511 _overflow_stack->reset(); // discard stack contents 7512 _overflow_stack->expand(); // expand the stack if possible 7513 } 7514 7515 void PushOrMarkClosure::do_oop(oop obj) { 7516 // Ignore mark word because we are running concurrent with mutators. 7517 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7518 HeapWord* addr = (HeapWord*)obj; 7519 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7520 // Oop lies in _span and isn't yet grey or black 7521 _bitMap->mark(addr); // now grey 7522 if (addr < _finger) { 7523 // the bit map iteration has already either passed, or 7524 // sampled, this bit in the bit map; we'll need to 7525 // use the marking stack to scan this oop's oops. 7526 bool simulate_overflow = false; 7527 NOT_PRODUCT( 7528 if (CMSMarkStackOverflowALot && 7529 _collector->simulate_overflow()) { 7530 // simulate a stack overflow 7531 simulate_overflow = true; 7532 } 7533 ) 7534 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7535 if (PrintCMSStatistics != 0) { 7536 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7537 SIZE_FORMAT, _markStack->capacity()); 7538 } 7539 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7540 handle_stack_overflow(addr); 7541 } 7542 } 7543 // anything including and to the right of _finger 7544 // will be scanned as we iterate over the remainder of the 7545 // bit map 7546 do_yield_check(); 7547 } 7548 } 7549 7550 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7551 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7552 7553 void Par_PushOrMarkClosure::do_oop(oop obj) { 7554 // Ignore mark word because we are running concurrent with mutators. 7555 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7556 HeapWord* addr = (HeapWord*)obj; 7557 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7558 // Oop lies in _span and isn't yet grey or black 7559 // We read the global_finger (volatile read) strictly after marking oop 7560 bool res = _bit_map->par_mark(addr); // now grey 7561 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7562 // Should we push this marked oop on our stack? 7563 // -- if someone else marked it, nothing to do 7564 // -- if target oop is above global finger nothing to do 7565 // -- if target oop is in chunk and above local finger 7566 // then nothing to do 7567 // -- else push on work queue 7568 if ( !res // someone else marked it, they will deal with it 7569 || (addr >= *gfa) // will be scanned in a later task 7570 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7571 return; 7572 } 7573 // the bit map iteration has already either passed, or 7574 // sampled, this bit in the bit map; we'll need to 7575 // use the marking stack to scan this oop's oops. 7576 bool simulate_overflow = false; 7577 NOT_PRODUCT( 7578 if (CMSMarkStackOverflowALot && 7579 _collector->simulate_overflow()) { 7580 // simulate a stack overflow 7581 simulate_overflow = true; 7582 } 7583 ) 7584 if (simulate_overflow || 7585 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7586 // stack overflow 7587 if (PrintCMSStatistics != 0) { 7588 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7589 SIZE_FORMAT, _overflow_stack->capacity()); 7590 } 7591 // We cannot assert that the overflow stack is full because 7592 // it may have been emptied since. 7593 assert(simulate_overflow || 7594 _work_queue->size() == _work_queue->max_elems(), 7595 "Else push should have succeeded"); 7596 handle_stack_overflow(addr); 7597 } 7598 do_yield_check(); 7599 } 7600 } 7601 7602 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7603 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7604 7605 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7606 MemRegion span, 7607 ReferenceProcessor* rp, 7608 CMSBitMap* bit_map, 7609 CMSBitMap* mod_union_table, 7610 CMSMarkStack* mark_stack, 7611 bool concurrent_precleaning): 7612 MetadataAwareOopClosure(rp), 7613 _collector(collector), 7614 _span(span), 7615 _bit_map(bit_map), 7616 _mod_union_table(mod_union_table), 7617 _mark_stack(mark_stack), 7618 _concurrent_precleaning(concurrent_precleaning) 7619 { 7620 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7621 } 7622 7623 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7624 // the non-parallel version (the parallel version appears further below.) 7625 void PushAndMarkClosure::do_oop(oop obj) { 7626 // Ignore mark word verification. If during concurrent precleaning, 7627 // the object monitor may be locked. If during the checkpoint 7628 // phases, the object may already have been reached by a different 7629 // path and may be at the end of the global overflow list (so 7630 // the mark word may be NULL). 7631 assert(obj->is_oop_or_null(true /* ignore mark word */), 7632 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7633 HeapWord* addr = (HeapWord*)obj; 7634 // Check if oop points into the CMS generation 7635 // and is not marked 7636 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7637 // a white object ... 7638 _bit_map->mark(addr); // ... now grey 7639 // push on the marking stack (grey set) 7640 bool simulate_overflow = false; 7641 NOT_PRODUCT( 7642 if (CMSMarkStackOverflowALot && 7643 _collector->simulate_overflow()) { 7644 // simulate a stack overflow 7645 simulate_overflow = true; 7646 } 7647 ) 7648 if (simulate_overflow || !_mark_stack->push(obj)) { 7649 if (_concurrent_precleaning) { 7650 // During precleaning we can just dirty the appropriate card(s) 7651 // in the mod union table, thus ensuring that the object remains 7652 // in the grey set and continue. In the case of object arrays 7653 // we need to dirty all of the cards that the object spans, 7654 // since the rescan of object arrays will be limited to the 7655 // dirty cards. 7656 // Note that no one can be interfering with us in this action 7657 // of dirtying the mod union table, so no locking or atomics 7658 // are required. 7659 if (obj->is_objArray()) { 7660 size_t sz = obj->size(); 7661 HeapWord* end_card_addr = (HeapWord*)round_to( 7662 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7663 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7664 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7665 _mod_union_table->mark_range(redirty_range); 7666 } else { 7667 _mod_union_table->mark(addr); 7668 } 7669 _collector->_ser_pmc_preclean_ovflw++; 7670 } else { 7671 // During the remark phase, we need to remember this oop 7672 // in the overflow list. 7673 _collector->push_on_overflow_list(obj); 7674 _collector->_ser_pmc_remark_ovflw++; 7675 } 7676 } 7677 } 7678 } 7679 7680 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7681 MemRegion span, 7682 ReferenceProcessor* rp, 7683 CMSBitMap* bit_map, 7684 OopTaskQueue* work_queue): 7685 MetadataAwareOopClosure(rp), 7686 _collector(collector), 7687 _span(span), 7688 _bit_map(bit_map), 7689 _work_queue(work_queue) 7690 { 7691 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7692 } 7693 7694 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7695 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7696 7697 // Grey object rescan during second checkpoint phase -- 7698 // the parallel version. 7699 void Par_PushAndMarkClosure::do_oop(oop obj) { 7700 // In the assert below, we ignore the mark word because 7701 // this oop may point to an already visited object that is 7702 // on the overflow stack (in which case the mark word has 7703 // been hijacked for chaining into the overflow stack -- 7704 // if this is the last object in the overflow stack then 7705 // its mark word will be NULL). Because this object may 7706 // have been subsequently popped off the global overflow 7707 // stack, and the mark word possibly restored to the prototypical 7708 // value, by the time we get to examined this failing assert in 7709 // the debugger, is_oop_or_null(false) may subsequently start 7710 // to hold. 7711 assert(obj->is_oop_or_null(true), 7712 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7713 HeapWord* addr = (HeapWord*)obj; 7714 // Check if oop points into the CMS generation 7715 // and is not marked 7716 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7717 // a white object ... 7718 // If we manage to "claim" the object, by being the 7719 // first thread to mark it, then we push it on our 7720 // marking stack 7721 if (_bit_map->par_mark(addr)) { // ... now grey 7722 // push on work queue (grey set) 7723 bool simulate_overflow = false; 7724 NOT_PRODUCT( 7725 if (CMSMarkStackOverflowALot && 7726 _collector->par_simulate_overflow()) { 7727 // simulate a stack overflow 7728 simulate_overflow = true; 7729 } 7730 ) 7731 if (simulate_overflow || !_work_queue->push(obj)) { 7732 _collector->par_push_on_overflow_list(obj); 7733 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7734 } 7735 } // Else, some other thread got there first 7736 } 7737 } 7738 7739 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7740 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7741 7742 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7743 Mutex* bml = _collector->bitMapLock(); 7744 assert_lock_strong(bml); 7745 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7746 "CMS thread should hold CMS token"); 7747 7748 bml->unlock(); 7749 ConcurrentMarkSweepThread::desynchronize(true); 7750 7751 _collector->stopTimer(); 7752 if (PrintCMSStatistics != 0) { 7753 _collector->incrementYields(); 7754 } 7755 7756 // See the comment in coordinator_yield() 7757 for (unsigned i = 0; i < CMSYieldSleepCount && 7758 ConcurrentMarkSweepThread::should_yield() && 7759 !CMSCollector::foregroundGCIsActive(); ++i) { 7760 os::sleep(Thread::current(), 1, false); 7761 } 7762 7763 ConcurrentMarkSweepThread::synchronize(true); 7764 bml->lock(); 7765 7766 _collector->startTimer(); 7767 } 7768 7769 bool CMSPrecleanRefsYieldClosure::should_return() { 7770 if (ConcurrentMarkSweepThread::should_yield()) { 7771 do_yield_work(); 7772 } 7773 return _collector->foregroundGCIsActive(); 7774 } 7775 7776 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7777 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7778 "mr should be aligned to start at a card boundary"); 7779 // We'd like to assert: 7780 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7781 // "mr should be a range of cards"); 7782 // However, that would be too strong in one case -- the last 7783 // partition ends at _unallocated_block which, in general, can be 7784 // an arbitrary boundary, not necessarily card aligned. 7785 if (PrintCMSStatistics != 0) { 7786 _num_dirty_cards += 7787 mr.word_size()/CardTableModRefBS::card_size_in_words; 7788 } 7789 _space->object_iterate_mem(mr, &_scan_cl); 7790 } 7791 7792 SweepClosure::SweepClosure(CMSCollector* collector, 7793 ConcurrentMarkSweepGeneration* g, 7794 CMSBitMap* bitMap, bool should_yield) : 7795 _collector(collector), 7796 _g(g), 7797 _sp(g->cmsSpace()), 7798 _limit(_sp->sweep_limit()), 7799 _freelistLock(_sp->freelistLock()), 7800 _bitMap(bitMap), 7801 _yield(should_yield), 7802 _inFreeRange(false), // No free range at beginning of sweep 7803 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7804 _lastFreeRangeCoalesced(false), 7805 _freeFinger(g->used_region().start()) 7806 { 7807 NOT_PRODUCT( 7808 _numObjectsFreed = 0; 7809 _numWordsFreed = 0; 7810 _numObjectsLive = 0; 7811 _numWordsLive = 0; 7812 _numObjectsAlreadyFree = 0; 7813 _numWordsAlreadyFree = 0; 7814 _last_fc = NULL; 7815 7816 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7817 _sp->dictionary()->initialize_dict_returned_bytes(); 7818 ) 7819 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7820 "sweep _limit out of bounds"); 7821 if (CMSTraceSweeper) { 7822 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7823 _limit); 7824 } 7825 } 7826 7827 void SweepClosure::print_on(outputStream* st) const { 7828 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7829 _sp->bottom(), _sp->end()); 7830 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7831 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7832 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7833 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7834 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7835 } 7836 7837 #ifndef PRODUCT 7838 // Assertion checking only: no useful work in product mode -- 7839 // however, if any of the flags below become product flags, 7840 // you may need to review this code to see if it needs to be 7841 // enabled in product mode. 7842 SweepClosure::~SweepClosure() { 7843 assert_lock_strong(_freelistLock); 7844 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7845 "sweep _limit out of bounds"); 7846 if (inFreeRange()) { 7847 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7848 print(); 7849 ShouldNotReachHere(); 7850 } 7851 if (Verbose && PrintGC) { 7852 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7853 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7854 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7855 SIZE_FORMAT" bytes " 7856 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7857 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7858 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7859 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7860 * sizeof(HeapWord); 7861 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7862 7863 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7864 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7865 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7866 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7867 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7868 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7869 indexListReturnedBytes); 7870 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7871 dict_returned_bytes); 7872 } 7873 } 7874 if (CMSTraceSweeper) { 7875 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7876 _limit); 7877 } 7878 } 7879 #endif // PRODUCT 7880 7881 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7882 bool freeRangeInFreeLists) { 7883 if (CMSTraceSweeper) { 7884 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7885 freeFinger, freeRangeInFreeLists); 7886 } 7887 assert(!inFreeRange(), "Trampling existing free range"); 7888 set_inFreeRange(true); 7889 set_lastFreeRangeCoalesced(false); 7890 7891 set_freeFinger(freeFinger); 7892 set_freeRangeInFreeLists(freeRangeInFreeLists); 7893 if (CMSTestInFreeList) { 7894 if (freeRangeInFreeLists) { 7895 FreeChunk* fc = (FreeChunk*) freeFinger; 7896 assert(fc->is_free(), "A chunk on the free list should be free."); 7897 assert(fc->size() > 0, "Free range should have a size"); 7898 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7899 } 7900 } 7901 } 7902 7903 // Note that the sweeper runs concurrently with mutators. Thus, 7904 // it is possible for direct allocation in this generation to happen 7905 // in the middle of the sweep. Note that the sweeper also coalesces 7906 // contiguous free blocks. Thus, unless the sweeper and the allocator 7907 // synchronize appropriately freshly allocated blocks may get swept up. 7908 // This is accomplished by the sweeper locking the free lists while 7909 // it is sweeping. Thus blocks that are determined to be free are 7910 // indeed free. There is however one additional complication: 7911 // blocks that have been allocated since the final checkpoint and 7912 // mark, will not have been marked and so would be treated as 7913 // unreachable and swept up. To prevent this, the allocator marks 7914 // the bit map when allocating during the sweep phase. This leads, 7915 // however, to a further complication -- objects may have been allocated 7916 // but not yet initialized -- in the sense that the header isn't yet 7917 // installed. The sweeper can not then determine the size of the block 7918 // in order to skip over it. To deal with this case, we use a technique 7919 // (due to Printezis) to encode such uninitialized block sizes in the 7920 // bit map. Since the bit map uses a bit per every HeapWord, but the 7921 // CMS generation has a minimum object size of 3 HeapWords, it follows 7922 // that "normal marks" won't be adjacent in the bit map (there will 7923 // always be at least two 0 bits between successive 1 bits). We make use 7924 // of these "unused" bits to represent uninitialized blocks -- the bit 7925 // corresponding to the start of the uninitialized object and the next 7926 // bit are both set. Finally, a 1 bit marks the end of the object that 7927 // started with the two consecutive 1 bits to indicate its potentially 7928 // uninitialized state. 7929 7930 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7931 FreeChunk* fc = (FreeChunk*)addr; 7932 size_t res; 7933 7934 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7935 // than "addr == _limit" because although _limit was a block boundary when 7936 // we started the sweep, it may no longer be one because heap expansion 7937 // may have caused us to coalesce the block ending at the address _limit 7938 // with a newly expanded chunk (this happens when _limit was set to the 7939 // previous _end of the space), so we may have stepped past _limit: 7940 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7941 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7942 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7943 "sweep _limit out of bounds"); 7944 assert(addr < _sp->end(), "addr out of bounds"); 7945 // Flush any free range we might be holding as a single 7946 // coalesced chunk to the appropriate free list. 7947 if (inFreeRange()) { 7948 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7949 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 7950 flush_cur_free_chunk(freeFinger(), 7951 pointer_delta(addr, freeFinger())); 7952 if (CMSTraceSweeper) { 7953 gclog_or_tty->print("Sweep: last chunk: "); 7954 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7955 "[coalesced:%d]\n", 7956 freeFinger(), pointer_delta(addr, freeFinger()), 7957 lastFreeRangeCoalesced() ? 1 : 0); 7958 } 7959 } 7960 7961 // help the iterator loop finish 7962 return pointer_delta(_sp->end(), addr); 7963 } 7964 7965 assert(addr < _limit, "sweep invariant"); 7966 // check if we should yield 7967 do_yield_check(addr); 7968 if (fc->is_free()) { 7969 // Chunk that is already free 7970 res = fc->size(); 7971 do_already_free_chunk(fc); 7972 debug_only(_sp->verifyFreeLists()); 7973 // If we flush the chunk at hand in lookahead_and_flush() 7974 // and it's coalesced with a preceding chunk, then the 7975 // process of "mangling" the payload of the coalesced block 7976 // will cause erasure of the size information from the 7977 // (erstwhile) header of all the coalesced blocks but the 7978 // first, so the first disjunct in the assert will not hold 7979 // in that specific case (in which case the second disjunct 7980 // will hold). 7981 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7982 "Otherwise the size info doesn't change at this step"); 7983 NOT_PRODUCT( 7984 _numObjectsAlreadyFree++; 7985 _numWordsAlreadyFree += res; 7986 ) 7987 NOT_PRODUCT(_last_fc = fc;) 7988 } else if (!_bitMap->isMarked(addr)) { 7989 // Chunk is fresh garbage 7990 res = do_garbage_chunk(fc); 7991 debug_only(_sp->verifyFreeLists()); 7992 NOT_PRODUCT( 7993 _numObjectsFreed++; 7994 _numWordsFreed += res; 7995 ) 7996 } else { 7997 // Chunk that is alive. 7998 res = do_live_chunk(fc); 7999 debug_only(_sp->verifyFreeLists()); 8000 NOT_PRODUCT( 8001 _numObjectsLive++; 8002 _numWordsLive += res; 8003 ) 8004 } 8005 return res; 8006 } 8007 8008 // For the smart allocation, record following 8009 // split deaths - a free chunk is removed from its free list because 8010 // it is being split into two or more chunks. 8011 // split birth - a free chunk is being added to its free list because 8012 // a larger free chunk has been split and resulted in this free chunk. 8013 // coal death - a free chunk is being removed from its free list because 8014 // it is being coalesced into a large free chunk. 8015 // coal birth - a free chunk is being added to its free list because 8016 // it was created when two or more free chunks where coalesced into 8017 // this free chunk. 8018 // 8019 // These statistics are used to determine the desired number of free 8020 // chunks of a given size. The desired number is chosen to be relative 8021 // to the end of a CMS sweep. The desired number at the end of a sweep 8022 // is the 8023 // count-at-end-of-previous-sweep (an amount that was enough) 8024 // - count-at-beginning-of-current-sweep (the excess) 8025 // + split-births (gains in this size during interval) 8026 // - split-deaths (demands on this size during interval) 8027 // where the interval is from the end of one sweep to the end of the 8028 // next. 8029 // 8030 // When sweeping the sweeper maintains an accumulated chunk which is 8031 // the chunk that is made up of chunks that have been coalesced. That 8032 // will be termed the left-hand chunk. A new chunk of garbage that 8033 // is being considered for coalescing will be referred to as the 8034 // right-hand chunk. 8035 // 8036 // When making a decision on whether to coalesce a right-hand chunk with 8037 // the current left-hand chunk, the current count vs. the desired count 8038 // of the left-hand chunk is considered. Also if the right-hand chunk 8039 // is near the large chunk at the end of the heap (see 8040 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8041 // left-hand chunk is coalesced. 8042 // 8043 // When making a decision about whether to split a chunk, the desired count 8044 // vs. the current count of the candidate to be split is also considered. 8045 // If the candidate is underpopulated (currently fewer chunks than desired) 8046 // a chunk of an overpopulated (currently more chunks than desired) size may 8047 // be chosen. The "hint" associated with a free list, if non-null, points 8048 // to a free list which may be overpopulated. 8049 // 8050 8051 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8052 const size_t size = fc->size(); 8053 // Chunks that cannot be coalesced are not in the 8054 // free lists. 8055 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8056 assert(_sp->verify_chunk_in_free_list(fc), 8057 "free chunk should be in free lists"); 8058 } 8059 // a chunk that is already free, should not have been 8060 // marked in the bit map 8061 HeapWord* const addr = (HeapWord*) fc; 8062 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8063 // Verify that the bit map has no bits marked between 8064 // addr and purported end of this block. 8065 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8066 8067 // Some chunks cannot be coalesced under any circumstances. 8068 // See the definition of cantCoalesce(). 8069 if (!fc->cantCoalesce()) { 8070 // This chunk can potentially be coalesced. 8071 if (_sp->adaptive_freelists()) { 8072 // All the work is done in 8073 do_post_free_or_garbage_chunk(fc, size); 8074 } else { // Not adaptive free lists 8075 // this is a free chunk that can potentially be coalesced by the sweeper; 8076 if (!inFreeRange()) { 8077 // if the next chunk is a free block that can't be coalesced 8078 // it doesn't make sense to remove this chunk from the free lists 8079 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8080 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8081 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8082 nextChunk->is_free() && // ... which is free... 8083 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8084 // nothing to do 8085 } else { 8086 // Potentially the start of a new free range: 8087 // Don't eagerly remove it from the free lists. 8088 // No need to remove it if it will just be put 8089 // back again. (Also from a pragmatic point of view 8090 // if it is a free block in a region that is beyond 8091 // any allocated blocks, an assertion will fail) 8092 // Remember the start of a free run. 8093 initialize_free_range(addr, true); 8094 // end - can coalesce with next chunk 8095 } 8096 } else { 8097 // the midst of a free range, we are coalescing 8098 print_free_block_coalesced(fc); 8099 if (CMSTraceSweeper) { 8100 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8101 } 8102 // remove it from the free lists 8103 _sp->removeFreeChunkFromFreeLists(fc); 8104 set_lastFreeRangeCoalesced(true); 8105 // If the chunk is being coalesced and the current free range is 8106 // in the free lists, remove the current free range so that it 8107 // will be returned to the free lists in its entirety - all 8108 // the coalesced pieces included. 8109 if (freeRangeInFreeLists()) { 8110 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8111 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8112 "Size of free range is inconsistent with chunk size."); 8113 if (CMSTestInFreeList) { 8114 assert(_sp->verify_chunk_in_free_list(ffc), 8115 "free range is not in free lists"); 8116 } 8117 _sp->removeFreeChunkFromFreeLists(ffc); 8118 set_freeRangeInFreeLists(false); 8119 } 8120 } 8121 } 8122 // Note that if the chunk is not coalescable (the else arm 8123 // below), we unconditionally flush, without needing to do 8124 // a "lookahead," as we do below. 8125 if (inFreeRange()) lookahead_and_flush(fc, size); 8126 } else { 8127 // Code path common to both original and adaptive free lists. 8128 8129 // cant coalesce with previous block; this should be treated 8130 // as the end of a free run if any 8131 if (inFreeRange()) { 8132 // we kicked some butt; time to pick up the garbage 8133 assert(freeFinger() < addr, "freeFinger points too high"); 8134 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8135 } 8136 // else, nothing to do, just continue 8137 } 8138 } 8139 8140 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8141 // This is a chunk of garbage. It is not in any free list. 8142 // Add it to a free list or let it possibly be coalesced into 8143 // a larger chunk. 8144 HeapWord* const addr = (HeapWord*) fc; 8145 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8146 8147 if (_sp->adaptive_freelists()) { 8148 // Verify that the bit map has no bits marked between 8149 // addr and purported end of just dead object. 8150 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8151 8152 do_post_free_or_garbage_chunk(fc, size); 8153 } else { 8154 if (!inFreeRange()) { 8155 // start of a new free range 8156 assert(size > 0, "A free range should have a size"); 8157 initialize_free_range(addr, false); 8158 } else { 8159 // this will be swept up when we hit the end of the 8160 // free range 8161 if (CMSTraceSweeper) { 8162 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8163 } 8164 // If the chunk is being coalesced and the current free range is 8165 // in the free lists, remove the current free range so that it 8166 // will be returned to the free lists in its entirety - all 8167 // the coalesced pieces included. 8168 if (freeRangeInFreeLists()) { 8169 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8170 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8171 "Size of free range is inconsistent with chunk size."); 8172 if (CMSTestInFreeList) { 8173 assert(_sp->verify_chunk_in_free_list(ffc), 8174 "free range is not in free lists"); 8175 } 8176 _sp->removeFreeChunkFromFreeLists(ffc); 8177 set_freeRangeInFreeLists(false); 8178 } 8179 set_lastFreeRangeCoalesced(true); 8180 } 8181 // this will be swept up when we hit the end of the free range 8182 8183 // Verify that the bit map has no bits marked between 8184 // addr and purported end of just dead object. 8185 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8186 } 8187 assert(_limit >= addr + size, 8188 "A freshly garbage chunk can't possibly straddle over _limit"); 8189 if (inFreeRange()) lookahead_and_flush(fc, size); 8190 return size; 8191 } 8192 8193 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8194 HeapWord* addr = (HeapWord*) fc; 8195 // The sweeper has just found a live object. Return any accumulated 8196 // left hand chunk to the free lists. 8197 if (inFreeRange()) { 8198 assert(freeFinger() < addr, "freeFinger points too high"); 8199 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8200 } 8201 8202 // This object is live: we'd normally expect this to be 8203 // an oop, and like to assert the following: 8204 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8205 // However, as we commented above, this may be an object whose 8206 // header hasn't yet been initialized. 8207 size_t size; 8208 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8209 if (_bitMap->isMarked(addr + 1)) { 8210 // Determine the size from the bit map, rather than trying to 8211 // compute it from the object header. 8212 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8213 size = pointer_delta(nextOneAddr + 1, addr); 8214 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8215 "alignment problem"); 8216 8217 #ifdef ASSERT 8218 if (oop(addr)->klass_or_null() != NULL) { 8219 // Ignore mark word because we are running concurrent with mutators 8220 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8221 assert(size == 8222 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8223 "P-mark and computed size do not agree"); 8224 } 8225 #endif 8226 8227 } else { 8228 // This should be an initialized object that's alive. 8229 assert(oop(addr)->klass_or_null() != NULL, 8230 "Should be an initialized object"); 8231 // Ignore mark word because we are running concurrent with mutators 8232 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8233 // Verify that the bit map has no bits marked between 8234 // addr and purported end of this block. 8235 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8236 assert(size >= 3, "Necessary for Printezis marks to work"); 8237 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8238 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8239 } 8240 return size; 8241 } 8242 8243 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8244 size_t chunkSize) { 8245 // do_post_free_or_garbage_chunk() should only be called in the case 8246 // of the adaptive free list allocator. 8247 const bool fcInFreeLists = fc->is_free(); 8248 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8249 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8250 if (CMSTestInFreeList && fcInFreeLists) { 8251 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8252 } 8253 8254 if (CMSTraceSweeper) { 8255 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 8256 } 8257 8258 HeapWord* const fc_addr = (HeapWord*) fc; 8259 8260 bool coalesce; 8261 const size_t left = pointer_delta(fc_addr, freeFinger()); 8262 const size_t right = chunkSize; 8263 switch (FLSCoalescePolicy) { 8264 // numeric value forms a coalition aggressiveness metric 8265 case 0: { // never coalesce 8266 coalesce = false; 8267 break; 8268 } 8269 case 1: { // coalesce if left & right chunks on overpopulated lists 8270 coalesce = _sp->coalOverPopulated(left) && 8271 _sp->coalOverPopulated(right); 8272 break; 8273 } 8274 case 2: { // coalesce if left chunk on overpopulated list (default) 8275 coalesce = _sp->coalOverPopulated(left); 8276 break; 8277 } 8278 case 3: { // coalesce if left OR right chunk on overpopulated list 8279 coalesce = _sp->coalOverPopulated(left) || 8280 _sp->coalOverPopulated(right); 8281 break; 8282 } 8283 case 4: { // always coalesce 8284 coalesce = true; 8285 break; 8286 } 8287 default: 8288 ShouldNotReachHere(); 8289 } 8290 8291 // Should the current free range be coalesced? 8292 // If the chunk is in a free range and either we decided to coalesce above 8293 // or the chunk is near the large block at the end of the heap 8294 // (isNearLargestChunk() returns true), then coalesce this chunk. 8295 const bool doCoalesce = inFreeRange() 8296 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8297 if (doCoalesce) { 8298 // Coalesce the current free range on the left with the new 8299 // chunk on the right. If either is on a free list, 8300 // it must be removed from the list and stashed in the closure. 8301 if (freeRangeInFreeLists()) { 8302 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8303 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8304 "Size of free range is inconsistent with chunk size."); 8305 if (CMSTestInFreeList) { 8306 assert(_sp->verify_chunk_in_free_list(ffc), 8307 "Chunk is not in free lists"); 8308 } 8309 _sp->coalDeath(ffc->size()); 8310 _sp->removeFreeChunkFromFreeLists(ffc); 8311 set_freeRangeInFreeLists(false); 8312 } 8313 if (fcInFreeLists) { 8314 _sp->coalDeath(chunkSize); 8315 assert(fc->size() == chunkSize, 8316 "The chunk has the wrong size or is not in the free lists"); 8317 _sp->removeFreeChunkFromFreeLists(fc); 8318 } 8319 set_lastFreeRangeCoalesced(true); 8320 print_free_block_coalesced(fc); 8321 } else { // not in a free range and/or should not coalesce 8322 // Return the current free range and start a new one. 8323 if (inFreeRange()) { 8324 // In a free range but cannot coalesce with the right hand chunk. 8325 // Put the current free range into the free lists. 8326 flush_cur_free_chunk(freeFinger(), 8327 pointer_delta(fc_addr, freeFinger())); 8328 } 8329 // Set up for new free range. Pass along whether the right hand 8330 // chunk is in the free lists. 8331 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8332 } 8333 } 8334 8335 // Lookahead flush: 8336 // If we are tracking a free range, and this is the last chunk that 8337 // we'll look at because its end crosses past _limit, we'll preemptively 8338 // flush it along with any free range we may be holding on to. Note that 8339 // this can be the case only for an already free or freshly garbage 8340 // chunk. If this block is an object, it can never straddle 8341 // over _limit. The "straddling" occurs when _limit is set at 8342 // the previous end of the space when this cycle started, and 8343 // a subsequent heap expansion caused the previously co-terminal 8344 // free block to be coalesced with the newly expanded portion, 8345 // thus rendering _limit a non-block-boundary making it dangerous 8346 // for the sweeper to step over and examine. 8347 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8348 assert(inFreeRange(), "Should only be called if currently in a free range."); 8349 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8350 assert(_sp->used_region().contains(eob - 1), 8351 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8352 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8353 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8354 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8355 if (eob >= _limit) { 8356 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8357 if (CMSTraceSweeper) { 8358 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8359 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8360 "[" PTR_FORMAT "," PTR_FORMAT ")", 8361 _limit, fc, eob, _sp->bottom(), _sp->end()); 8362 } 8363 // Return the storage we are tracking back into the free lists. 8364 if (CMSTraceSweeper) { 8365 gclog_or_tty->print_cr("Flushing ... "); 8366 } 8367 assert(freeFinger() < eob, "Error"); 8368 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8369 } 8370 } 8371 8372 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8373 assert(inFreeRange(), "Should only be called if currently in a free range."); 8374 assert(size > 0, 8375 "A zero sized chunk cannot be added to the free lists."); 8376 if (!freeRangeInFreeLists()) { 8377 if (CMSTestInFreeList) { 8378 FreeChunk* fc = (FreeChunk*) chunk; 8379 fc->set_size(size); 8380 assert(!_sp->verify_chunk_in_free_list(fc), 8381 "chunk should not be in free lists yet"); 8382 } 8383 if (CMSTraceSweeper) { 8384 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8385 chunk, size); 8386 } 8387 // A new free range is going to be starting. The current 8388 // free range has not been added to the free lists yet or 8389 // was removed so add it back. 8390 // If the current free range was coalesced, then the death 8391 // of the free range was recorded. Record a birth now. 8392 if (lastFreeRangeCoalesced()) { 8393 _sp->coalBirth(size); 8394 } 8395 _sp->addChunkAndRepairOffsetTable(chunk, size, 8396 lastFreeRangeCoalesced()); 8397 } else if (CMSTraceSweeper) { 8398 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8399 } 8400 set_inFreeRange(false); 8401 set_freeRangeInFreeLists(false); 8402 } 8403 8404 // We take a break if we've been at this for a while, 8405 // so as to avoid monopolizing the locks involved. 8406 void SweepClosure::do_yield_work(HeapWord* addr) { 8407 // Return current free chunk being used for coalescing (if any) 8408 // to the appropriate freelist. After yielding, the next 8409 // free block encountered will start a coalescing range of 8410 // free blocks. If the next free block is adjacent to the 8411 // chunk just flushed, they will need to wait for the next 8412 // sweep to be coalesced. 8413 if (inFreeRange()) { 8414 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8415 } 8416 8417 // First give up the locks, then yield, then re-lock. 8418 // We should probably use a constructor/destructor idiom to 8419 // do this unlock/lock or modify the MutexUnlocker class to 8420 // serve our purpose. XXX 8421 assert_lock_strong(_bitMap->lock()); 8422 assert_lock_strong(_freelistLock); 8423 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8424 "CMS thread should hold CMS token"); 8425 _bitMap->lock()->unlock(); 8426 _freelistLock->unlock(); 8427 ConcurrentMarkSweepThread::desynchronize(true); 8428 _collector->stopTimer(); 8429 if (PrintCMSStatistics != 0) { 8430 _collector->incrementYields(); 8431 } 8432 8433 // See the comment in coordinator_yield() 8434 for (unsigned i = 0; i < CMSYieldSleepCount && 8435 ConcurrentMarkSweepThread::should_yield() && 8436 !CMSCollector::foregroundGCIsActive(); ++i) { 8437 os::sleep(Thread::current(), 1, false); 8438 } 8439 8440 ConcurrentMarkSweepThread::synchronize(true); 8441 _freelistLock->lock(); 8442 _bitMap->lock()->lock_without_safepoint_check(); 8443 _collector->startTimer(); 8444 } 8445 8446 #ifndef PRODUCT 8447 // This is actually very useful in a product build if it can 8448 // be called from the debugger. Compile it into the product 8449 // as needed. 8450 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8451 return debug_cms_space->verify_chunk_in_free_list(fc); 8452 } 8453 #endif 8454 8455 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8456 if (CMSTraceSweeper) { 8457 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8458 fc, fc->size()); 8459 } 8460 } 8461 8462 // CMSIsAliveClosure 8463 bool CMSIsAliveClosure::do_object_b(oop obj) { 8464 HeapWord* addr = (HeapWord*)obj; 8465 return addr != NULL && 8466 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8467 } 8468 8469 8470 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8471 MemRegion span, 8472 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8473 bool cpc): 8474 _collector(collector), 8475 _span(span), 8476 _bit_map(bit_map), 8477 _mark_stack(mark_stack), 8478 _concurrent_precleaning(cpc) { 8479 assert(!_span.is_empty(), "Empty span could spell trouble"); 8480 } 8481 8482 8483 // CMSKeepAliveClosure: the serial version 8484 void CMSKeepAliveClosure::do_oop(oop obj) { 8485 HeapWord* addr = (HeapWord*)obj; 8486 if (_span.contains(addr) && 8487 !_bit_map->isMarked(addr)) { 8488 _bit_map->mark(addr); 8489 bool simulate_overflow = false; 8490 NOT_PRODUCT( 8491 if (CMSMarkStackOverflowALot && 8492 _collector->simulate_overflow()) { 8493 // simulate a stack overflow 8494 simulate_overflow = true; 8495 } 8496 ) 8497 if (simulate_overflow || !_mark_stack->push(obj)) { 8498 if (_concurrent_precleaning) { 8499 // We dirty the overflown object and let the remark 8500 // phase deal with it. 8501 assert(_collector->overflow_list_is_empty(), "Error"); 8502 // In the case of object arrays, we need to dirty all of 8503 // the cards that the object spans. No locking or atomics 8504 // are needed since no one else can be mutating the mod union 8505 // table. 8506 if (obj->is_objArray()) { 8507 size_t sz = obj->size(); 8508 HeapWord* end_card_addr = 8509 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8510 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8511 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8512 _collector->_modUnionTable.mark_range(redirty_range); 8513 } else { 8514 _collector->_modUnionTable.mark(addr); 8515 } 8516 _collector->_ser_kac_preclean_ovflw++; 8517 } else { 8518 _collector->push_on_overflow_list(obj); 8519 _collector->_ser_kac_ovflw++; 8520 } 8521 } 8522 } 8523 } 8524 8525 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8526 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8527 8528 // CMSParKeepAliveClosure: a parallel version of the above. 8529 // The work queues are private to each closure (thread), 8530 // but (may be) available for stealing by other threads. 8531 void CMSParKeepAliveClosure::do_oop(oop obj) { 8532 HeapWord* addr = (HeapWord*)obj; 8533 if (_span.contains(addr) && 8534 !_bit_map->isMarked(addr)) { 8535 // In general, during recursive tracing, several threads 8536 // may be concurrently getting here; the first one to 8537 // "tag" it, claims it. 8538 if (_bit_map->par_mark(addr)) { 8539 bool res = _work_queue->push(obj); 8540 assert(res, "Low water mark should be much less than capacity"); 8541 // Do a recursive trim in the hope that this will keep 8542 // stack usage lower, but leave some oops for potential stealers 8543 trim_queue(_low_water_mark); 8544 } // Else, another thread got there first 8545 } 8546 } 8547 8548 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8549 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8550 8551 void CMSParKeepAliveClosure::trim_queue(uint max) { 8552 while (_work_queue->size() > max) { 8553 oop new_oop; 8554 if (_work_queue->pop_local(new_oop)) { 8555 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8556 assert(_bit_map->isMarked((HeapWord*)new_oop), 8557 "no white objects on this stack!"); 8558 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8559 // iterate over the oops in this oop, marking and pushing 8560 // the ones in CMS heap (i.e. in _span). 8561 new_oop->oop_iterate(&_mark_and_push); 8562 } 8563 } 8564 } 8565 8566 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8567 CMSCollector* collector, 8568 MemRegion span, CMSBitMap* bit_map, 8569 OopTaskQueue* work_queue): 8570 _collector(collector), 8571 _span(span), 8572 _bit_map(bit_map), 8573 _work_queue(work_queue) { } 8574 8575 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8576 HeapWord* addr = (HeapWord*)obj; 8577 if (_span.contains(addr) && 8578 !_bit_map->isMarked(addr)) { 8579 if (_bit_map->par_mark(addr)) { 8580 bool simulate_overflow = false; 8581 NOT_PRODUCT( 8582 if (CMSMarkStackOverflowALot && 8583 _collector->par_simulate_overflow()) { 8584 // simulate a stack overflow 8585 simulate_overflow = true; 8586 } 8587 ) 8588 if (simulate_overflow || !_work_queue->push(obj)) { 8589 _collector->par_push_on_overflow_list(obj); 8590 _collector->_par_kac_ovflw++; 8591 } 8592 } // Else another thread got there already 8593 } 8594 } 8595 8596 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8597 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8598 8599 ////////////////////////////////////////////////////////////////// 8600 // CMSExpansionCause ///////////////////////////// 8601 ////////////////////////////////////////////////////////////////// 8602 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8603 switch (cause) { 8604 case _no_expansion: 8605 return "No expansion"; 8606 case _satisfy_free_ratio: 8607 return "Free ratio"; 8608 case _satisfy_promotion: 8609 return "Satisfy promotion"; 8610 case _satisfy_allocation: 8611 return "allocation"; 8612 case _allocate_par_lab: 8613 return "Par LAB"; 8614 case _allocate_par_spooling_space: 8615 return "Par Spooling Space"; 8616 case _adaptive_size_policy: 8617 return "Ergonomics"; 8618 default: 8619 return "unknown"; 8620 } 8621 } 8622 8623 void CMSDrainMarkingStackClosure::do_void() { 8624 // the max number to take from overflow list at a time 8625 const size_t num = _mark_stack->capacity()/4; 8626 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8627 "Overflow list should be NULL during concurrent phases"); 8628 while (!_mark_stack->isEmpty() || 8629 // if stack is empty, check the overflow list 8630 _collector->take_from_overflow_list(num, _mark_stack)) { 8631 oop obj = _mark_stack->pop(); 8632 HeapWord* addr = (HeapWord*)obj; 8633 assert(_span.contains(addr), "Should be within span"); 8634 assert(_bit_map->isMarked(addr), "Should be marked"); 8635 assert(obj->is_oop(), "Should be an oop"); 8636 obj->oop_iterate(_keep_alive); 8637 } 8638 } 8639 8640 void CMSParDrainMarkingStackClosure::do_void() { 8641 // drain queue 8642 trim_queue(0); 8643 } 8644 8645 // Trim our work_queue so its length is below max at return 8646 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8647 while (_work_queue->size() > max) { 8648 oop new_oop; 8649 if (_work_queue->pop_local(new_oop)) { 8650 assert(new_oop->is_oop(), "Expected an oop"); 8651 assert(_bit_map->isMarked((HeapWord*)new_oop), 8652 "no white objects on this stack!"); 8653 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8654 // iterate over the oops in this oop, marking and pushing 8655 // the ones in CMS heap (i.e. in _span). 8656 new_oop->oop_iterate(&_mark_and_push); 8657 } 8658 } 8659 } 8660 8661 //////////////////////////////////////////////////////////////////// 8662 // Support for Marking Stack Overflow list handling and related code 8663 //////////////////////////////////////////////////////////////////// 8664 // Much of the following code is similar in shape and spirit to the 8665 // code used in ParNewGC. We should try and share that code 8666 // as much as possible in the future. 8667 8668 #ifndef PRODUCT 8669 // Debugging support for CMSStackOverflowALot 8670 8671 // It's OK to call this multi-threaded; the worst thing 8672 // that can happen is that we'll get a bunch of closely 8673 // spaced simulated overflows, but that's OK, in fact 8674 // probably good as it would exercise the overflow code 8675 // under contention. 8676 bool CMSCollector::simulate_overflow() { 8677 if (_overflow_counter-- <= 0) { // just being defensive 8678 _overflow_counter = CMSMarkStackOverflowInterval; 8679 return true; 8680 } else { 8681 return false; 8682 } 8683 } 8684 8685 bool CMSCollector::par_simulate_overflow() { 8686 return simulate_overflow(); 8687 } 8688 #endif 8689 8690 // Single-threaded 8691 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8692 assert(stack->isEmpty(), "Expected precondition"); 8693 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8694 size_t i = num; 8695 oop cur = _overflow_list; 8696 const markOop proto = markOopDesc::prototype(); 8697 NOT_PRODUCT(ssize_t n = 0;) 8698 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8699 next = oop(cur->mark()); 8700 cur->set_mark(proto); // until proven otherwise 8701 assert(cur->is_oop(), "Should be an oop"); 8702 bool res = stack->push(cur); 8703 assert(res, "Bit off more than can chew?"); 8704 NOT_PRODUCT(n++;) 8705 } 8706 _overflow_list = cur; 8707 #ifndef PRODUCT 8708 assert(_num_par_pushes >= n, "Too many pops?"); 8709 _num_par_pushes -=n; 8710 #endif 8711 return !stack->isEmpty(); 8712 } 8713 8714 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8715 // (MT-safe) Get a prefix of at most "num" from the list. 8716 // The overflow list is chained through the mark word of 8717 // each object in the list. We fetch the entire list, 8718 // break off a prefix of the right size and return the 8719 // remainder. If other threads try to take objects from 8720 // the overflow list at that time, they will wait for 8721 // some time to see if data becomes available. If (and 8722 // only if) another thread places one or more object(s) 8723 // on the global list before we have returned the suffix 8724 // to the global list, we will walk down our local list 8725 // to find its end and append the global list to 8726 // our suffix before returning it. This suffix walk can 8727 // prove to be expensive (quadratic in the amount of traffic) 8728 // when there are many objects in the overflow list and 8729 // there is much producer-consumer contention on the list. 8730 // *NOTE*: The overflow list manipulation code here and 8731 // in ParNewGeneration:: are very similar in shape, 8732 // except that in the ParNew case we use the old (from/eden) 8733 // copy of the object to thread the list via its klass word. 8734 // Because of the common code, if you make any changes in 8735 // the code below, please check the ParNew version to see if 8736 // similar changes might be needed. 8737 // CR 6797058 has been filed to consolidate the common code. 8738 bool CMSCollector::par_take_from_overflow_list(size_t num, 8739 OopTaskQueue* work_q, 8740 int no_of_gc_threads) { 8741 assert(work_q->size() == 0, "First empty local work queue"); 8742 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8743 if (_overflow_list == NULL) { 8744 return false; 8745 } 8746 // Grab the entire list; we'll put back a suffix 8747 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8748 Thread* tid = Thread::current(); 8749 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8750 // set to ParallelGCThreads. 8751 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8752 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8753 // If the list is busy, we spin for a short while, 8754 // sleeping between attempts to get the list. 8755 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8756 os::sleep(tid, sleep_time_millis, false); 8757 if (_overflow_list == NULL) { 8758 // Nothing left to take 8759 return false; 8760 } else if (_overflow_list != BUSY) { 8761 // Try and grab the prefix 8762 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8763 } 8764 } 8765 // If the list was found to be empty, or we spun long 8766 // enough, we give up and return empty-handed. If we leave 8767 // the list in the BUSY state below, it must be the case that 8768 // some other thread holds the overflow list and will set it 8769 // to a non-BUSY state in the future. 8770 if (prefix == NULL || prefix == BUSY) { 8771 // Nothing to take or waited long enough 8772 if (prefix == NULL) { 8773 // Write back the NULL in case we overwrote it with BUSY above 8774 // and it is still the same value. 8775 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8776 } 8777 return false; 8778 } 8779 assert(prefix != NULL && prefix != BUSY, "Error"); 8780 size_t i = num; 8781 oop cur = prefix; 8782 // Walk down the first "num" objects, unless we reach the end. 8783 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8784 if (cur->mark() == NULL) { 8785 // We have "num" or fewer elements in the list, so there 8786 // is nothing to return to the global list. 8787 // Write back the NULL in lieu of the BUSY we wrote 8788 // above, if it is still the same value. 8789 if (_overflow_list == BUSY) { 8790 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8791 } 8792 } else { 8793 // Chop off the suffix and return it to the global list. 8794 assert(cur->mark() != BUSY, "Error"); 8795 oop suffix_head = cur->mark(); // suffix will be put back on global list 8796 cur->set_mark(NULL); // break off suffix 8797 // It's possible that the list is still in the empty(busy) state 8798 // we left it in a short while ago; in that case we may be 8799 // able to place back the suffix without incurring the cost 8800 // of a walk down the list. 8801 oop observed_overflow_list = _overflow_list; 8802 oop cur_overflow_list = observed_overflow_list; 8803 bool attached = false; 8804 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8805 observed_overflow_list = 8806 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8807 if (cur_overflow_list == observed_overflow_list) { 8808 attached = true; 8809 break; 8810 } else cur_overflow_list = observed_overflow_list; 8811 } 8812 if (!attached) { 8813 // Too bad, someone else sneaked in (at least) an element; we'll need 8814 // to do a splice. Find tail of suffix so we can prepend suffix to global 8815 // list. 8816 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8817 oop suffix_tail = cur; 8818 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8819 "Tautology"); 8820 observed_overflow_list = _overflow_list; 8821 do { 8822 cur_overflow_list = observed_overflow_list; 8823 if (cur_overflow_list != BUSY) { 8824 // Do the splice ... 8825 suffix_tail->set_mark(markOop(cur_overflow_list)); 8826 } else { // cur_overflow_list == BUSY 8827 suffix_tail->set_mark(NULL); 8828 } 8829 // ... and try to place spliced list back on overflow_list ... 8830 observed_overflow_list = 8831 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8832 } while (cur_overflow_list != observed_overflow_list); 8833 // ... until we have succeeded in doing so. 8834 } 8835 } 8836 8837 // Push the prefix elements on work_q 8838 assert(prefix != NULL, "control point invariant"); 8839 const markOop proto = markOopDesc::prototype(); 8840 oop next; 8841 NOT_PRODUCT(ssize_t n = 0;) 8842 for (cur = prefix; cur != NULL; cur = next) { 8843 next = oop(cur->mark()); 8844 cur->set_mark(proto); // until proven otherwise 8845 assert(cur->is_oop(), "Should be an oop"); 8846 bool res = work_q->push(cur); 8847 assert(res, "Bit off more than we can chew?"); 8848 NOT_PRODUCT(n++;) 8849 } 8850 #ifndef PRODUCT 8851 assert(_num_par_pushes >= n, "Too many pops?"); 8852 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8853 #endif 8854 return true; 8855 } 8856 8857 // Single-threaded 8858 void CMSCollector::push_on_overflow_list(oop p) { 8859 NOT_PRODUCT(_num_par_pushes++;) 8860 assert(p->is_oop(), "Not an oop"); 8861 preserve_mark_if_necessary(p); 8862 p->set_mark((markOop)_overflow_list); 8863 _overflow_list = p; 8864 } 8865 8866 // Multi-threaded; use CAS to prepend to overflow list 8867 void CMSCollector::par_push_on_overflow_list(oop p) { 8868 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8869 assert(p->is_oop(), "Not an oop"); 8870 par_preserve_mark_if_necessary(p); 8871 oop observed_overflow_list = _overflow_list; 8872 oop cur_overflow_list; 8873 do { 8874 cur_overflow_list = observed_overflow_list; 8875 if (cur_overflow_list != BUSY) { 8876 p->set_mark(markOop(cur_overflow_list)); 8877 } else { 8878 p->set_mark(NULL); 8879 } 8880 observed_overflow_list = 8881 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8882 } while (cur_overflow_list != observed_overflow_list); 8883 } 8884 #undef BUSY 8885 8886 // Single threaded 8887 // General Note on GrowableArray: pushes may silently fail 8888 // because we are (temporarily) out of C-heap for expanding 8889 // the stack. The problem is quite ubiquitous and affects 8890 // a lot of code in the JVM. The prudent thing for GrowableArray 8891 // to do (for now) is to exit with an error. However, that may 8892 // be too draconian in some cases because the caller may be 8893 // able to recover without much harm. For such cases, we 8894 // should probably introduce a "soft_push" method which returns 8895 // an indication of success or failure with the assumption that 8896 // the caller may be able to recover from a failure; code in 8897 // the VM can then be changed, incrementally, to deal with such 8898 // failures where possible, thus, incrementally hardening the VM 8899 // in such low resource situations. 8900 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8901 _preserved_oop_stack.push(p); 8902 _preserved_mark_stack.push(m); 8903 assert(m == p->mark(), "Mark word changed"); 8904 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8905 "bijection"); 8906 } 8907 8908 // Single threaded 8909 void CMSCollector::preserve_mark_if_necessary(oop p) { 8910 markOop m = p->mark(); 8911 if (m->must_be_preserved(p)) { 8912 preserve_mark_work(p, m); 8913 } 8914 } 8915 8916 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8917 markOop m = p->mark(); 8918 if (m->must_be_preserved(p)) { 8919 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8920 // Even though we read the mark word without holding 8921 // the lock, we are assured that it will not change 8922 // because we "own" this oop, so no other thread can 8923 // be trying to push it on the overflow list; see 8924 // the assertion in preserve_mark_work() that checks 8925 // that m == p->mark(). 8926 preserve_mark_work(p, m); 8927 } 8928 } 8929 8930 // We should be able to do this multi-threaded, 8931 // a chunk of stack being a task (this is 8932 // correct because each oop only ever appears 8933 // once in the overflow list. However, it's 8934 // not very easy to completely overlap this with 8935 // other operations, so will generally not be done 8936 // until all work's been completed. Because we 8937 // expect the preserved oop stack (set) to be small, 8938 // it's probably fine to do this single-threaded. 8939 // We can explore cleverer concurrent/overlapped/parallel 8940 // processing of preserved marks if we feel the 8941 // need for this in the future. Stack overflow should 8942 // be so rare in practice and, when it happens, its 8943 // effect on performance so great that this will 8944 // likely just be in the noise anyway. 8945 void CMSCollector::restore_preserved_marks_if_any() { 8946 assert(SafepointSynchronize::is_at_safepoint(), 8947 "world should be stopped"); 8948 assert(Thread::current()->is_ConcurrentGC_thread() || 8949 Thread::current()->is_VM_thread(), 8950 "should be single-threaded"); 8951 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8952 "bijection"); 8953 8954 while (!_preserved_oop_stack.is_empty()) { 8955 oop p = _preserved_oop_stack.pop(); 8956 assert(p->is_oop(), "Should be an oop"); 8957 assert(_span.contains(p), "oop should be in _span"); 8958 assert(p->mark() == markOopDesc::prototype(), 8959 "Set when taken from overflow list"); 8960 markOop m = _preserved_mark_stack.pop(); 8961 p->set_mark(m); 8962 } 8963 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8964 "stacks were cleared above"); 8965 } 8966 8967 #ifndef PRODUCT 8968 bool CMSCollector::no_preserved_marks() const { 8969 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8970 } 8971 #endif 8972 8973 // Transfer some number of overflown objects to usual marking 8974 // stack. Return true if some objects were transferred. 8975 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8976 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8977 (size_t)ParGCDesiredObjsFromOverflowList); 8978 8979 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8980 assert(_collector->overflow_list_is_empty() || res, 8981 "If list is not empty, we should have taken something"); 8982 assert(!res || !_mark_stack->isEmpty(), 8983 "If we took something, it should now be on our stack"); 8984 return res; 8985 } 8986 8987 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8988 size_t res = _sp->block_size_no_stall(addr, _collector); 8989 if (_sp->block_is_obj(addr)) { 8990 if (_live_bit_map->isMarked(addr)) { 8991 // It can't have been dead in a previous cycle 8992 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8993 } else { 8994 _dead_bit_map->mark(addr); // mark the dead object 8995 } 8996 } 8997 // Could be 0, if the block size could not be computed without stalling. 8998 return res; 8999 } 9000 9001 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9002 9003 switch (phase) { 9004 case CMSCollector::InitialMarking: 9005 initialize(true /* fullGC */ , 9006 cause /* cause of the GC */, 9007 true /* recordGCBeginTime */, 9008 true /* recordPreGCUsage */, 9009 false /* recordPeakUsage */, 9010 false /* recordPostGCusage */, 9011 true /* recordAccumulatedGCTime */, 9012 false /* recordGCEndTime */, 9013 false /* countCollection */ ); 9014 break; 9015 9016 case CMSCollector::FinalMarking: 9017 initialize(true /* fullGC */ , 9018 cause /* cause of the GC */, 9019 false /* recordGCBeginTime */, 9020 false /* recordPreGCUsage */, 9021 false /* recordPeakUsage */, 9022 false /* recordPostGCusage */, 9023 true /* recordAccumulatedGCTime */, 9024 false /* recordGCEndTime */, 9025 false /* countCollection */ ); 9026 break; 9027 9028 case CMSCollector::Sweeping: 9029 initialize(true /* fullGC */ , 9030 cause /* cause of the GC */, 9031 false /* recordGCBeginTime */, 9032 false /* recordPreGCUsage */, 9033 true /* recordPeakUsage */, 9034 true /* recordPostGCusage */, 9035 false /* recordAccumulatedGCTime */, 9036 true /* recordGCEndTime */, 9037 true /* countCollection */ ); 9038 break; 9039 9040 default: 9041 ShouldNotReachHere(); 9042 } 9043 }