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