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 CompactibleSpace* 797 ConcurrentMarkSweepGeneration::first_compaction_space() const { 798 return _cmsSpace; 799 } 800 801 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 802 // Clear the promotion information. These pointers can be adjusted 803 // along with all the other pointers into the heap but 804 // compaction is expected to be a rare event with 805 // a heap using cms so don't do it without seeing the need. 806 for (uint i = 0; i < ParallelGCThreads; i++) { 807 _par_gc_thread_states[i]->promo.reset(); 808 } 809 } 810 811 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 812 blk->do_space(_cmsSpace); 813 } 814 815 void ConcurrentMarkSweepGeneration::compute_new_size() { 816 assert_locked_or_safepoint(Heap_lock); 817 818 // If incremental collection failed, we just want to expand 819 // to the limit. 820 if (incremental_collection_failed()) { 821 clear_incremental_collection_failed(); 822 grow_to_reserved(); 823 return; 824 } 825 826 // The heap has been compacted but not reset yet. 827 // Any metric such as free() or used() will be incorrect. 828 829 CardGeneration::compute_new_size(); 830 831 // Reset again after a possible resizing 832 if (did_compact()) { 833 cmsSpace()->reset_after_compaction(); 834 } 835 } 836 837 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 838 assert_locked_or_safepoint(Heap_lock); 839 840 // If incremental collection failed, we just want to expand 841 // to the limit. 842 if (incremental_collection_failed()) { 843 clear_incremental_collection_failed(); 844 grow_to_reserved(); 845 return; 846 } 847 848 double free_percentage = ((double) free()) / capacity(); 849 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 850 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 851 852 // compute expansion delta needed for reaching desired free percentage 853 if (free_percentage < desired_free_percentage) { 854 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 855 assert(desired_capacity >= capacity(), "invalid expansion size"); 856 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 857 if (PrintGCDetails && Verbose) { 858 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 859 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 860 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 861 gclog_or_tty->print_cr(" Desired free fraction %f", 862 desired_free_percentage); 863 gclog_or_tty->print_cr(" Maximum free fraction %f", 864 maximum_free_percentage); 865 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 866 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 867 desired_capacity/1000); 868 int prev_level = level() - 1; 869 if (prev_level >= 0) { 870 size_t prev_size = 0; 871 GenCollectedHeap* gch = GenCollectedHeap::heap(); 872 Generation* prev_gen = gch->_gens[prev_level]; 873 prev_size = prev_gen->capacity(); 874 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 875 prev_size/1000); 876 } 877 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 878 unsafe_max_alloc_nogc()/1000); 879 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 880 contiguous_available()/1000); 881 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 882 expand_bytes); 883 } 884 // safe if expansion fails 885 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 886 if (PrintGCDetails && Verbose) { 887 gclog_or_tty->print_cr(" Expanded free fraction %f", 888 ((double) free()) / capacity()); 889 } 890 } else { 891 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 892 assert(desired_capacity <= capacity(), "invalid expansion size"); 893 size_t shrink_bytes = capacity() - desired_capacity; 894 // Don't shrink unless the delta is greater than the minimum shrink we want 895 if (shrink_bytes >= MinHeapDeltaBytes) { 896 shrink_free_list_by(shrink_bytes); 897 } 898 } 899 } 900 901 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 902 return cmsSpace()->freelistLock(); 903 } 904 905 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 906 bool tlab) { 907 CMSSynchronousYieldRequest yr; 908 MutexLockerEx x(freelistLock(), 909 Mutex::_no_safepoint_check_flag); 910 return have_lock_and_allocate(size, tlab); 911 } 912 913 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 914 bool tlab /* ignored */) { 915 assert_lock_strong(freelistLock()); 916 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 917 HeapWord* res = cmsSpace()->allocate(adjustedSize); 918 // Allocate the object live (grey) if the background collector has 919 // started marking. This is necessary because the marker may 920 // have passed this address and consequently this object will 921 // not otherwise be greyed and would be incorrectly swept up. 922 // Note that if this object contains references, the writing 923 // of those references will dirty the card containing this object 924 // allowing the object to be blackened (and its references scanned) 925 // either during a preclean phase or at the final checkpoint. 926 if (res != NULL) { 927 // We may block here with an uninitialized object with 928 // its mark-bit or P-bits not yet set. Such objects need 929 // to be safely navigable by block_start(). 930 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 931 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 932 collector()->direct_allocated(res, adjustedSize); 933 _direct_allocated_words += adjustedSize; 934 // allocation counters 935 NOT_PRODUCT( 936 _numObjectsAllocated++; 937 _numWordsAllocated += (int)adjustedSize; 938 ) 939 } 940 return res; 941 } 942 943 // In the case of direct allocation by mutators in a generation that 944 // is being concurrently collected, the object must be allocated 945 // live (grey) if the background collector has started marking. 946 // This is necessary because the marker may 947 // have passed this address and consequently this object will 948 // not otherwise be greyed and would be incorrectly swept up. 949 // Note that if this object contains references, the writing 950 // of those references will dirty the card containing this object 951 // allowing the object to be blackened (and its references scanned) 952 // either during a preclean phase or at the final checkpoint. 953 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 954 assert(_markBitMap.covers(start, size), "Out of bounds"); 955 if (_collectorState >= Marking) { 956 MutexLockerEx y(_markBitMap.lock(), 957 Mutex::_no_safepoint_check_flag); 958 // [see comments preceding SweepClosure::do_blk() below for details] 959 // 960 // Can the P-bits be deleted now? JJJ 961 // 962 // 1. need to mark the object as live so it isn't collected 963 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 964 // 3. need to mark the end of the object so marking, precleaning or sweeping 965 // can skip over uninitialized or unparsable objects. An allocated 966 // object is considered uninitialized for our purposes as long as 967 // its klass word is NULL. All old gen objects are parsable 968 // as soon as they are initialized.) 969 _markBitMap.mark(start); // object is live 970 _markBitMap.mark(start + 1); // object is potentially uninitialized? 971 _markBitMap.mark(start + size - 1); 972 // mark end of object 973 } 974 // check that oop looks uninitialized 975 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 976 } 977 978 void CMSCollector::promoted(bool par, HeapWord* start, 979 bool is_obj_array, size_t obj_size) { 980 assert(_markBitMap.covers(start), "Out of bounds"); 981 // See comment in direct_allocated() about when objects should 982 // be allocated live. 983 if (_collectorState >= Marking) { 984 // we already hold the marking bit map lock, taken in 985 // the prologue 986 if (par) { 987 _markBitMap.par_mark(start); 988 } else { 989 _markBitMap.mark(start); 990 } 991 // We don't need to mark the object as uninitialized (as 992 // in direct_allocated above) because this is being done with the 993 // world stopped and the object will be initialized by the 994 // time the marking, precleaning or sweeping get to look at it. 995 // But see the code for copying objects into the CMS generation, 996 // where we need to ensure that concurrent readers of the 997 // block offset table are able to safely navigate a block that 998 // is in flux from being free to being allocated (and in 999 // transition while being copied into) and subsequently 1000 // becoming a bona-fide object when the copy/promotion is complete. 1001 assert(SafepointSynchronize::is_at_safepoint(), 1002 "expect promotion only at safepoints"); 1003 1004 if (_collectorState < Sweeping) { 1005 // Mark the appropriate cards in the modUnionTable, so that 1006 // this object gets scanned before the sweep. If this is 1007 // not done, CMS generation references in the object might 1008 // not get marked. 1009 // For the case of arrays, which are otherwise precisely 1010 // marked, we need to dirty the entire array, not just its head. 1011 if (is_obj_array) { 1012 // The [par_]mark_range() method expects mr.end() below to 1013 // be aligned to the granularity of a bit's representation 1014 // in the heap. In the case of the MUT below, that's a 1015 // card size. 1016 MemRegion mr(start, 1017 (HeapWord*)round_to((intptr_t)(start + obj_size), 1018 CardTableModRefBS::card_size /* bytes */)); 1019 if (par) { 1020 _modUnionTable.par_mark_range(mr); 1021 } else { 1022 _modUnionTable.mark_range(mr); 1023 } 1024 } else { // not an obj array; we can just mark the head 1025 if (par) { 1026 _modUnionTable.par_mark(start); 1027 } else { 1028 _modUnionTable.mark(start); 1029 } 1030 } 1031 } 1032 } 1033 } 1034 1035 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1036 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1037 // allocate, copy and if necessary update promoinfo -- 1038 // delegate to underlying space. 1039 assert_lock_strong(freelistLock()); 1040 1041 #ifndef PRODUCT 1042 if (Universe::heap()->promotion_should_fail()) { 1043 return NULL; 1044 } 1045 #endif // #ifndef PRODUCT 1046 1047 oop res = _cmsSpace->promote(obj, obj_size); 1048 if (res == NULL) { 1049 // expand and retry 1050 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1051 expand(s*HeapWordSize, MinHeapDeltaBytes, 1052 CMSExpansionCause::_satisfy_promotion); 1053 // Since there's currently no next generation, we don't try to promote 1054 // into a more senior generation. 1055 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1056 "is made to pass on a possibly failing " 1057 "promotion to next generation"); 1058 res = _cmsSpace->promote(obj, obj_size); 1059 } 1060 if (res != NULL) { 1061 // See comment in allocate() about when objects should 1062 // be allocated live. 1063 assert(obj->is_oop(), "Will dereference klass pointer below"); 1064 collector()->promoted(false, // Not parallel 1065 (HeapWord*)res, obj->is_objArray(), obj_size); 1066 // promotion counters 1067 NOT_PRODUCT( 1068 _numObjectsPromoted++; 1069 _numWordsPromoted += 1070 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1071 ) 1072 } 1073 return res; 1074 } 1075 1076 1077 // IMPORTANT: Notes on object size recognition in CMS. 1078 // --------------------------------------------------- 1079 // A block of storage in the CMS generation is always in 1080 // one of three states. A free block (FREE), an allocated 1081 // object (OBJECT) whose size() method reports the correct size, 1082 // and an intermediate state (TRANSIENT) in which its size cannot 1083 // be accurately determined. 1084 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1085 // ----------------------------------------------------- 1086 // FREE: klass_word & 1 == 1; mark_word holds block size 1087 // 1088 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1089 // obj->size() computes correct size 1090 // 1091 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1092 // 1093 // STATE IDENTIFICATION: (64 bit+COOPS) 1094 // ------------------------------------ 1095 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1096 // 1097 // OBJECT: klass_word installed; klass_word != 0; 1098 // obj->size() computes correct size 1099 // 1100 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1101 // 1102 // 1103 // STATE TRANSITION DIAGRAM 1104 // 1105 // mut / parnew mut / parnew 1106 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1107 // ^ | 1108 // |------------------------ DEAD <------------------------------------| 1109 // sweep mut 1110 // 1111 // While a block is in TRANSIENT state its size cannot be determined 1112 // so readers will either need to come back later or stall until 1113 // the size can be determined. Note that for the case of direct 1114 // allocation, P-bits, when available, may be used to determine the 1115 // size of an object that may not yet have been initialized. 1116 1117 // Things to support parallel young-gen collection. 1118 oop 1119 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1120 oop old, markOop m, 1121 size_t word_sz) { 1122 #ifndef PRODUCT 1123 if (Universe::heap()->promotion_should_fail()) { 1124 return NULL; 1125 } 1126 #endif // #ifndef PRODUCT 1127 1128 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1129 PromotionInfo* promoInfo = &ps->promo; 1130 // if we are tracking promotions, then first ensure space for 1131 // promotion (including spooling space for saving header if necessary). 1132 // then allocate and copy, then track promoted info if needed. 1133 // When tracking (see PromotionInfo::track()), the mark word may 1134 // be displaced and in this case restoration of the mark word 1135 // occurs in the (oop_since_save_marks_)iterate phase. 1136 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1137 // Out of space for allocating spooling buffers; 1138 // try expanding and allocating spooling buffers. 1139 if (!expand_and_ensure_spooling_space(promoInfo)) { 1140 return NULL; 1141 } 1142 } 1143 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1144 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1145 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1146 if (obj_ptr == NULL) { 1147 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1148 if (obj_ptr == NULL) { 1149 return NULL; 1150 } 1151 } 1152 oop obj = oop(obj_ptr); 1153 OrderAccess::storestore(); 1154 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1155 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1156 // IMPORTANT: See note on object initialization for CMS above. 1157 // Otherwise, copy the object. Here we must be careful to insert the 1158 // klass pointer last, since this marks the block as an allocated object. 1159 // Except with compressed oops it's the mark word. 1160 HeapWord* old_ptr = (HeapWord*)old; 1161 // Restore the mark word copied above. 1162 obj->set_mark(m); 1163 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1164 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1165 OrderAccess::storestore(); 1166 1167 if (UseCompressedClassPointers) { 1168 // Copy gap missed by (aligned) header size calculation below 1169 obj->set_klass_gap(old->klass_gap()); 1170 } 1171 if (word_sz > (size_t)oopDesc::header_size()) { 1172 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1173 obj_ptr + oopDesc::header_size(), 1174 word_sz - oopDesc::header_size()); 1175 } 1176 1177 // Now we can track the promoted object, if necessary. We take care 1178 // to delay the transition from uninitialized to full object 1179 // (i.e., insertion of klass pointer) until after, so that it 1180 // atomically becomes a promoted object. 1181 if (promoInfo->tracking()) { 1182 promoInfo->track((PromotedObject*)obj, old->klass()); 1183 } 1184 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1185 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1186 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1187 1188 // Finally, install the klass pointer (this should be volatile). 1189 OrderAccess::storestore(); 1190 obj->set_klass(old->klass()); 1191 // We should now be able to calculate the right size for this object 1192 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1193 1194 collector()->promoted(true, // parallel 1195 obj_ptr, old->is_objArray(), word_sz); 1196 1197 NOT_PRODUCT( 1198 Atomic::inc_ptr(&_numObjectsPromoted); 1199 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1200 ) 1201 1202 return obj; 1203 } 1204 1205 void 1206 ConcurrentMarkSweepGeneration:: 1207 par_promote_alloc_done(int thread_num) { 1208 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1209 ps->lab.retire(thread_num); 1210 } 1211 1212 void 1213 ConcurrentMarkSweepGeneration:: 1214 par_oop_since_save_marks_iterate_done(int thread_num) { 1215 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1216 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1217 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1218 } 1219 1220 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1221 size_t size, 1222 bool tlab) 1223 { 1224 // We allow a STW collection only if a full 1225 // collection was requested. 1226 return full || should_allocate(size, tlab); // FIX ME !!! 1227 // This and promotion failure handling are connected at the 1228 // hip and should be fixed by untying them. 1229 } 1230 1231 bool CMSCollector::shouldConcurrentCollect() { 1232 if (_full_gc_requested) { 1233 if (Verbose && PrintGCDetails) { 1234 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1235 " gc request (or gc_locker)"); 1236 } 1237 return true; 1238 } 1239 1240 FreelistLocker x(this); 1241 // ------------------------------------------------------------------ 1242 // Print out lots of information which affects the initiation of 1243 // a collection. 1244 if (PrintCMSInitiationStatistics && stats().valid()) { 1245 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1246 gclog_or_tty->stamp(); 1247 gclog_or_tty->cr(); 1248 stats().print_on(gclog_or_tty); 1249 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1250 stats().time_until_cms_gen_full()); 1251 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1252 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1253 _cmsGen->contiguous_available()); 1254 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1255 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1256 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1257 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1258 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1259 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1260 gclog_or_tty->print_cr("metadata initialized %d", 1261 MetaspaceGC::should_concurrent_collect()); 1262 } 1263 // ------------------------------------------------------------------ 1264 1265 // If the estimated time to complete a cms collection (cms_duration()) 1266 // is less than the estimated time remaining until the cms generation 1267 // is full, start a collection. 1268 if (!UseCMSInitiatingOccupancyOnly) { 1269 if (stats().valid()) { 1270 if (stats().time_until_cms_start() == 0.0) { 1271 return true; 1272 } 1273 } else { 1274 // We want to conservatively collect somewhat early in order 1275 // to try and "bootstrap" our CMS/promotion statistics; 1276 // this branch will not fire after the first successful CMS 1277 // collection because the stats should then be valid. 1278 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1279 if (Verbose && PrintGCDetails) { 1280 gclog_or_tty->print_cr( 1281 " CMSCollector: collect for bootstrapping statistics:" 1282 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1283 _bootstrap_occupancy); 1284 } 1285 return true; 1286 } 1287 } 1288 } 1289 1290 // Otherwise, we start a collection cycle if 1291 // old gen want a collection cycle started. Each may use 1292 // an appropriate criterion for making this decision. 1293 // XXX We need to make sure that the gen expansion 1294 // criterion dovetails well with this. XXX NEED TO FIX THIS 1295 if (_cmsGen->should_concurrent_collect()) { 1296 if (Verbose && PrintGCDetails) { 1297 gclog_or_tty->print_cr("CMS old gen initiated"); 1298 } 1299 return true; 1300 } 1301 1302 // We start a collection if we believe an incremental collection may fail; 1303 // this is not likely to be productive in practice because it's probably too 1304 // late anyway. 1305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1306 assert(gch->collector_policy()->is_generation_policy(), 1307 "You may want to check the correctness of the following"); 1308 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1309 if (Verbose && PrintGCDetails) { 1310 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1311 } 1312 return true; 1313 } 1314 1315 if (MetaspaceGC::should_concurrent_collect()) { 1316 if (Verbose && PrintGCDetails) { 1317 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1318 } 1319 return true; 1320 } 1321 1322 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1323 if (CMSTriggerInterval >= 0) { 1324 if (CMSTriggerInterval == 0) { 1325 // Trigger always 1326 return true; 1327 } 1328 1329 // Check the CMS time since begin (we do not check the stats validity 1330 // as we want to be able to trigger the first CMS cycle as well) 1331 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1332 if (Verbose && PrintGCDetails) { 1333 if (stats().valid()) { 1334 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1335 stats().cms_time_since_begin()); 1336 } else { 1337 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1338 } 1339 } 1340 return true; 1341 } 1342 } 1343 1344 return false; 1345 } 1346 1347 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1348 1349 // Clear _expansion_cause fields of constituent generations 1350 void CMSCollector::clear_expansion_cause() { 1351 _cmsGen->clear_expansion_cause(); 1352 } 1353 1354 // We should be conservative in starting a collection cycle. To 1355 // start too eagerly runs the risk of collecting too often in the 1356 // extreme. To collect too rarely falls back on full collections, 1357 // which works, even if not optimum in terms of concurrent work. 1358 // As a work around for too eagerly collecting, use the flag 1359 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1360 // giving the user an easily understandable way of controlling the 1361 // collections. 1362 // We want to start a new collection cycle if any of the following 1363 // conditions hold: 1364 // . our current occupancy exceeds the configured initiating occupancy 1365 // for this generation, or 1366 // . we recently needed to expand this space and have not, since that 1367 // expansion, done a collection of this generation, or 1368 // . the underlying space believes that it may be a good idea to initiate 1369 // a concurrent collection (this may be based on criteria such as the 1370 // following: the space uses linear allocation and linear allocation is 1371 // going to fail, or there is believed to be excessive fragmentation in 1372 // the generation, etc... or ... 1373 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1374 // the case of the old generation; see CR 6543076): 1375 // we may be approaching a point at which allocation requests may fail because 1376 // we will be out of sufficient free space given allocation rate estimates.] 1377 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1378 1379 assert_lock_strong(freelistLock()); 1380 if (occupancy() > initiating_occupancy()) { 1381 if (PrintGCDetails && Verbose) { 1382 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1383 short_name(), occupancy(), initiating_occupancy()); 1384 } 1385 return true; 1386 } 1387 if (UseCMSInitiatingOccupancyOnly) { 1388 return false; 1389 } 1390 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1391 if (PrintGCDetails && Verbose) { 1392 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1393 short_name()); 1394 } 1395 return true; 1396 } 1397 if (_cmsSpace->should_concurrent_collect()) { 1398 if (PrintGCDetails && Verbose) { 1399 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1400 short_name()); 1401 } 1402 return true; 1403 } 1404 return false; 1405 } 1406 1407 void ConcurrentMarkSweepGeneration::collect(bool full, 1408 bool clear_all_soft_refs, 1409 size_t size, 1410 bool tlab) 1411 { 1412 collector()->collect(full, clear_all_soft_refs, size, tlab); 1413 } 1414 1415 void CMSCollector::collect(bool full, 1416 bool clear_all_soft_refs, 1417 size_t size, 1418 bool tlab) 1419 { 1420 // The following "if" branch is present for defensive reasons. 1421 // In the current uses of this interface, it can be replaced with: 1422 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1423 // But I am not placing that assert here to allow future 1424 // generality in invoking this interface. 1425 if (GC_locker::is_active()) { 1426 // A consistency test for GC_locker 1427 assert(GC_locker::needs_gc(), "Should have been set already"); 1428 // Skip this foreground collection, instead 1429 // expanding the heap if necessary. 1430 // Need the free list locks for the call to free() in compute_new_size() 1431 compute_new_size(); 1432 return; 1433 } 1434 acquire_control_and_collect(full, clear_all_soft_refs); 1435 } 1436 1437 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1438 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1439 unsigned int gc_count = gch->total_full_collections(); 1440 if (gc_count == full_gc_count) { 1441 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1442 _full_gc_requested = true; 1443 _full_gc_cause = cause; 1444 CGC_lock->notify(); // nudge CMS thread 1445 } else { 1446 assert(gc_count > full_gc_count, "Error: causal loop"); 1447 } 1448 } 1449 1450 bool CMSCollector::is_external_interruption() { 1451 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1452 return GCCause::is_user_requested_gc(cause) || 1453 GCCause::is_serviceability_requested_gc(cause); 1454 } 1455 1456 void CMSCollector::report_concurrent_mode_interruption() { 1457 if (is_external_interruption()) { 1458 if (PrintGCDetails) { 1459 gclog_or_tty->print(" (concurrent mode interrupted)"); 1460 } 1461 } else { 1462 if (PrintGCDetails) { 1463 gclog_or_tty->print(" (concurrent mode failure)"); 1464 } 1465 _gc_tracer_cm->report_concurrent_mode_failure(); 1466 } 1467 } 1468 1469 1470 // The foreground and background collectors need to coordinate in order 1471 // to make sure that they do not mutually interfere with CMS collections. 1472 // When a background collection is active, 1473 // the foreground collector may need to take over (preempt) and 1474 // synchronously complete an ongoing collection. Depending on the 1475 // frequency of the background collections and the heap usage 1476 // of the application, this preemption can be seldom or frequent. 1477 // There are only certain 1478 // points in the background collection that the "collection-baton" 1479 // can be passed to the foreground collector. 1480 // 1481 // The foreground collector will wait for the baton before 1482 // starting any part of the collection. The foreground collector 1483 // will only wait at one location. 1484 // 1485 // The background collector will yield the baton before starting a new 1486 // phase of the collection (e.g., before initial marking, marking from roots, 1487 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1488 // of the loop which switches the phases. The background collector does some 1489 // of the phases (initial mark, final re-mark) with the world stopped. 1490 // Because of locking involved in stopping the world, 1491 // the foreground collector should not block waiting for the background 1492 // collector when it is doing a stop-the-world phase. The background 1493 // collector will yield the baton at an additional point just before 1494 // it enters a stop-the-world phase. Once the world is stopped, the 1495 // background collector checks the phase of the collection. If the 1496 // phase has not changed, it proceeds with the collection. If the 1497 // phase has changed, it skips that phase of the collection. See 1498 // the comments on the use of the Heap_lock in collect_in_background(). 1499 // 1500 // Variable used in baton passing. 1501 // _foregroundGCIsActive - Set to true by the foreground collector when 1502 // it wants the baton. The foreground clears it when it has finished 1503 // the collection. 1504 // _foregroundGCShouldWait - Set to true by the background collector 1505 // when it is running. The foreground collector waits while 1506 // _foregroundGCShouldWait is true. 1507 // CGC_lock - monitor used to protect access to the above variables 1508 // and to notify the foreground and background collectors. 1509 // _collectorState - current state of the CMS collection. 1510 // 1511 // The foreground collector 1512 // acquires the CGC_lock 1513 // sets _foregroundGCIsActive 1514 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1515 // various locks acquired in preparation for the collection 1516 // are released so as not to block the background collector 1517 // that is in the midst of a collection 1518 // proceeds with the collection 1519 // clears _foregroundGCIsActive 1520 // returns 1521 // 1522 // The background collector in a loop iterating on the phases of the 1523 // collection 1524 // acquires the CGC_lock 1525 // sets _foregroundGCShouldWait 1526 // if _foregroundGCIsActive is set 1527 // clears _foregroundGCShouldWait, notifies _CGC_lock 1528 // waits on _CGC_lock for _foregroundGCIsActive to become false 1529 // and exits the loop. 1530 // otherwise 1531 // proceed with that phase of the collection 1532 // if the phase is a stop-the-world phase, 1533 // yield the baton once more just before enqueueing 1534 // the stop-world CMS operation (executed by the VM thread). 1535 // returns after all phases of the collection are done 1536 // 1537 1538 void CMSCollector::acquire_control_and_collect(bool full, 1539 bool clear_all_soft_refs) { 1540 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1541 assert(!Thread::current()->is_ConcurrentGC_thread(), 1542 "shouldn't try to acquire control from self!"); 1543 1544 // Start the protocol for acquiring control of the 1545 // collection from the background collector (aka CMS thread). 1546 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1547 "VM thread should have CMS token"); 1548 // Remember the possibly interrupted state of an ongoing 1549 // concurrent collection 1550 CollectorState first_state = _collectorState; 1551 1552 // Signal to a possibly ongoing concurrent collection that 1553 // we want to do a foreground collection. 1554 _foregroundGCIsActive = true; 1555 1556 // release locks and wait for a notify from the background collector 1557 // releasing the locks in only necessary for phases which 1558 // do yields to improve the granularity of the collection. 1559 assert_lock_strong(bitMapLock()); 1560 // We need to lock the Free list lock for the space that we are 1561 // currently collecting. 1562 assert(haveFreelistLocks(), "Must be holding free list locks"); 1563 bitMapLock()->unlock(); 1564 releaseFreelistLocks(); 1565 { 1566 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1567 if (_foregroundGCShouldWait) { 1568 // We are going to be waiting for action for the CMS thread; 1569 // it had better not be gone (for instance at shutdown)! 1570 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1571 "CMS thread must be running"); 1572 // Wait here until the background collector gives us the go-ahead 1573 ConcurrentMarkSweepThread::clear_CMS_flag( 1574 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1575 // Get a possibly blocked CMS thread going: 1576 // Note that we set _foregroundGCIsActive true above, 1577 // without protection of the CGC_lock. 1578 CGC_lock->notify(); 1579 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1580 "Possible deadlock"); 1581 while (_foregroundGCShouldWait) { 1582 // wait for notification 1583 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1584 // Possibility of delay/starvation here, since CMS token does 1585 // not know to give priority to VM thread? Actually, i think 1586 // there wouldn't be any delay/starvation, but the proof of 1587 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1588 } 1589 ConcurrentMarkSweepThread::set_CMS_flag( 1590 ConcurrentMarkSweepThread::CMS_vm_has_token); 1591 } 1592 } 1593 // The CMS_token is already held. Get back the other locks. 1594 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1595 "VM thread should have CMS token"); 1596 getFreelistLocks(); 1597 bitMapLock()->lock_without_safepoint_check(); 1598 if (TraceCMSState) { 1599 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1600 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1601 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1602 } 1603 1604 // Inform cms gen if this was due to partial collection failing. 1605 // The CMS gen may use this fact to determine its expansion policy. 1606 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1607 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1608 assert(!_cmsGen->incremental_collection_failed(), 1609 "Should have been noticed, reacted to and cleared"); 1610 _cmsGen->set_incremental_collection_failed(); 1611 } 1612 1613 if (first_state > Idling) { 1614 report_concurrent_mode_interruption(); 1615 } 1616 1617 set_did_compact(true); 1618 1619 // If the collection is being acquired from the background 1620 // collector, there may be references on the discovered 1621 // references lists that have NULL referents (being those 1622 // that were concurrently cleared by a mutator) or 1623 // that are no longer active (having been enqueued concurrently 1624 // by the mutator). 1625 // Scrub the list of those references because Mark-Sweep-Compact 1626 // code assumes referents are not NULL and that all discovered 1627 // Reference objects are active. 1628 ref_processor()->clean_up_discovered_references(); 1629 1630 if (first_state > Idling) { 1631 save_heap_summary(); 1632 } 1633 1634 do_compaction_work(clear_all_soft_refs); 1635 1636 // Has the GC time limit been exceeded? 1637 size_t max_eden_size = _young_gen->max_capacity() - 1638 _young_gen->to()->capacity() - 1639 _young_gen->from()->capacity(); 1640 GCCause::Cause gc_cause = gch->gc_cause(); 1641 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1642 _young_gen->eden()->used(), 1643 _cmsGen->max_capacity(), 1644 max_eden_size, 1645 full, 1646 gc_cause, 1647 gch->collector_policy()); 1648 1649 // Reset the expansion cause, now that we just completed 1650 // a collection cycle. 1651 clear_expansion_cause(); 1652 _foregroundGCIsActive = false; 1653 return; 1654 } 1655 1656 // Resize the tenured generation 1657 // after obtaining the free list locks for the 1658 // two generations. 1659 void CMSCollector::compute_new_size() { 1660 assert_locked_or_safepoint(Heap_lock); 1661 FreelistLocker z(this); 1662 MetaspaceGC::compute_new_size(); 1663 _cmsGen->compute_new_size_free_list(); 1664 } 1665 1666 // A work method used by the foreground collector to do 1667 // a mark-sweep-compact. 1668 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1669 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1670 1671 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1672 gc_timer->register_gc_start(); 1673 1674 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1675 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1676 1677 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1678 1679 // Temporarily widen the span of the weak reference processing to 1680 // the entire heap. 1681 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1682 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1683 // Temporarily, clear the "is_alive_non_header" field of the 1684 // reference processor. 1685 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1686 // Temporarily make reference _processing_ single threaded (non-MT). 1687 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1688 // Temporarily make refs discovery atomic 1689 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1690 // Temporarily make reference _discovery_ single threaded (non-MT) 1691 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1692 1693 ref_processor()->set_enqueuing_is_done(false); 1694 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 1695 ref_processor()->setup_policy(clear_all_soft_refs); 1696 // If an asynchronous collection finishes, the _modUnionTable is 1697 // all clear. If we are assuming the collection from an asynchronous 1698 // collection, clear the _modUnionTable. 1699 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1700 "_modUnionTable should be clear if the baton was not passed"); 1701 _modUnionTable.clear_all(); 1702 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1703 "mod union for klasses should be clear if the baton was passed"); 1704 _ct->klass_rem_set()->clear_mod_union(); 1705 1706 // We must adjust the allocation statistics being maintained 1707 // in the free list space. We do so by reading and clearing 1708 // the sweep timer and updating the block flux rate estimates below. 1709 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1710 if (_inter_sweep_timer.is_active()) { 1711 _inter_sweep_timer.stop(); 1712 // Note that we do not use this sample to update the _inter_sweep_estimate. 1713 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1714 _inter_sweep_estimate.padded_average(), 1715 _intra_sweep_estimate.padded_average()); 1716 } 1717 1718 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1719 ref_processor(), clear_all_soft_refs); 1720 #ifdef ASSERT 1721 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1722 size_t free_size = cms_space->free(); 1723 assert(free_size == 1724 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1725 * HeapWordSize, 1726 "All the free space should be compacted into one chunk at top"); 1727 assert(cms_space->dictionary()->total_chunk_size( 1728 debug_only(cms_space->freelistLock())) == 0 || 1729 cms_space->totalSizeInIndexedFreeLists() == 0, 1730 "All the free space should be in a single chunk"); 1731 size_t num = cms_space->totalCount(); 1732 assert((free_size == 0 && num == 0) || 1733 (free_size > 0 && (num == 1 || num == 2)), 1734 "There should be at most 2 free chunks after compaction"); 1735 #endif // ASSERT 1736 _collectorState = Resetting; 1737 assert(_restart_addr == NULL, 1738 "Should have been NULL'd before baton was passed"); 1739 reset(false /* == !concurrent */); 1740 _cmsGen->reset_after_compaction(); 1741 _concurrent_cycles_since_last_unload = 0; 1742 1743 // Clear any data recorded in the PLAB chunk arrays. 1744 if (_survivor_plab_array != NULL) { 1745 reset_survivor_plab_arrays(); 1746 } 1747 1748 // Adjust the per-size allocation stats for the next epoch. 1749 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1750 // Restart the "inter sweep timer" for the next epoch. 1751 _inter_sweep_timer.reset(); 1752 _inter_sweep_timer.start(); 1753 1754 gc_timer->register_gc_end(); 1755 1756 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1757 1758 // For a mark-sweep-compact, compute_new_size() will be called 1759 // in the heap's do_collection() method. 1760 } 1761 1762 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1763 ContiguousSpace* eden_space = _young_gen->eden(); 1764 ContiguousSpace* from_space = _young_gen->from(); 1765 ContiguousSpace* to_space = _young_gen->to(); 1766 // Eden 1767 if (_eden_chunk_array != NULL) { 1768 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1769 eden_space->bottom(), eden_space->top(), 1770 eden_space->end(), eden_space->capacity()); 1771 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1772 "_eden_chunk_capacity=" SIZE_FORMAT, 1773 _eden_chunk_index, _eden_chunk_capacity); 1774 for (size_t i = 0; i < _eden_chunk_index; i++) { 1775 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1776 i, _eden_chunk_array[i]); 1777 } 1778 } 1779 // Survivor 1780 if (_survivor_chunk_array != NULL) { 1781 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1782 from_space->bottom(), from_space->top(), 1783 from_space->end(), from_space->capacity()); 1784 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1785 "_survivor_chunk_capacity=" SIZE_FORMAT, 1786 _survivor_chunk_index, _survivor_chunk_capacity); 1787 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1788 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1789 i, _survivor_chunk_array[i]); 1790 } 1791 } 1792 } 1793 1794 void CMSCollector::getFreelistLocks() const { 1795 // Get locks for all free lists in all generations that this 1796 // collector is responsible for 1797 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1798 } 1799 1800 void CMSCollector::releaseFreelistLocks() const { 1801 // Release locks for all free lists in all generations that this 1802 // collector is responsible for 1803 _cmsGen->freelistLock()->unlock(); 1804 } 1805 1806 bool CMSCollector::haveFreelistLocks() const { 1807 // Check locks for all free lists in all generations that this 1808 // collector is responsible for 1809 assert_lock_strong(_cmsGen->freelistLock()); 1810 PRODUCT_ONLY(ShouldNotReachHere()); 1811 return true; 1812 } 1813 1814 // A utility class that is used by the CMS collector to 1815 // temporarily "release" the foreground collector from its 1816 // usual obligation to wait for the background collector to 1817 // complete an ongoing phase before proceeding. 1818 class ReleaseForegroundGC: public StackObj { 1819 private: 1820 CMSCollector* _c; 1821 public: 1822 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1823 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1824 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1825 // allow a potentially blocked foreground collector to proceed 1826 _c->_foregroundGCShouldWait = false; 1827 if (_c->_foregroundGCIsActive) { 1828 CGC_lock->notify(); 1829 } 1830 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1831 "Possible deadlock"); 1832 } 1833 1834 ~ReleaseForegroundGC() { 1835 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1836 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1837 _c->_foregroundGCShouldWait = true; 1838 } 1839 }; 1840 1841 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1842 assert(Thread::current()->is_ConcurrentGC_thread(), 1843 "A CMS asynchronous collection is only allowed on a CMS thread."); 1844 1845 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1846 { 1847 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1848 MutexLockerEx hl(Heap_lock, safepoint_check); 1849 FreelistLocker fll(this); 1850 MutexLockerEx x(CGC_lock, safepoint_check); 1851 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1852 // The foreground collector is active or we're 1853 // not using asynchronous collections. Skip this 1854 // background collection. 1855 assert(!_foregroundGCShouldWait, "Should be clear"); 1856 return; 1857 } else { 1858 assert(_collectorState == Idling, "Should be idling before start."); 1859 _collectorState = InitialMarking; 1860 register_gc_start(cause); 1861 // Reset the expansion cause, now that we are about to begin 1862 // a new cycle. 1863 clear_expansion_cause(); 1864 1865 // Clear the MetaspaceGC flag since a concurrent collection 1866 // is starting but also clear it after the collection. 1867 MetaspaceGC::set_should_concurrent_collect(false); 1868 } 1869 // Decide if we want to enable class unloading as part of the 1870 // ensuing concurrent GC cycle. 1871 update_should_unload_classes(); 1872 _full_gc_requested = false; // acks all outstanding full gc requests 1873 _full_gc_cause = GCCause::_no_gc; 1874 // Signal that we are about to start a collection 1875 gch->increment_total_full_collections(); // ... starting a collection cycle 1876 _collection_count_start = gch->total_full_collections(); 1877 } 1878 1879 // Used for PrintGC 1880 size_t prev_used; 1881 if (PrintGC && Verbose) { 1882 prev_used = _cmsGen->used(); 1883 } 1884 1885 // The change of the collection state is normally done at this level; 1886 // the exceptions are phases that are executed while the world is 1887 // stopped. For those phases the change of state is done while the 1888 // world is stopped. For baton passing purposes this allows the 1889 // background collector to finish the phase and change state atomically. 1890 // The foreground collector cannot wait on a phase that is done 1891 // while the world is stopped because the foreground collector already 1892 // has the world stopped and would deadlock. 1893 while (_collectorState != Idling) { 1894 if (TraceCMSState) { 1895 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1896 Thread::current(), _collectorState); 1897 } 1898 // The foreground collector 1899 // holds the Heap_lock throughout its collection. 1900 // holds the CMS token (but not the lock) 1901 // except while it is waiting for the background collector to yield. 1902 // 1903 // The foreground collector should be blocked (not for long) 1904 // if the background collector is about to start a phase 1905 // executed with world stopped. If the background 1906 // collector has already started such a phase, the 1907 // foreground collector is blocked waiting for the 1908 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1909 // are executed in the VM thread. 1910 // 1911 // The locking order is 1912 // PendingListLock (PLL) -- if applicable (FinalMarking) 1913 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1914 // CMS token (claimed in 1915 // stop_world_and_do() --> 1916 // safepoint_synchronize() --> 1917 // CMSThread::synchronize()) 1918 1919 { 1920 // Check if the FG collector wants us to yield. 1921 CMSTokenSync x(true); // is cms thread 1922 if (waitForForegroundGC()) { 1923 // We yielded to a foreground GC, nothing more to be 1924 // done this round. 1925 assert(_foregroundGCShouldWait == false, "We set it to false in " 1926 "waitForForegroundGC()"); 1927 if (TraceCMSState) { 1928 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1929 " exiting collection CMS state %d", 1930 Thread::current(), _collectorState); 1931 } 1932 return; 1933 } else { 1934 // The background collector can run but check to see if the 1935 // foreground collector has done a collection while the 1936 // background collector was waiting to get the CGC_lock 1937 // above. If yes, break so that _foregroundGCShouldWait 1938 // is cleared before returning. 1939 if (_collectorState == Idling) { 1940 break; 1941 } 1942 } 1943 } 1944 1945 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1946 "should be waiting"); 1947 1948 switch (_collectorState) { 1949 case InitialMarking: 1950 { 1951 ReleaseForegroundGC x(this); 1952 stats().record_cms_begin(); 1953 VM_CMS_Initial_Mark initial_mark_op(this); 1954 VMThread::execute(&initial_mark_op); 1955 } 1956 // The collector state may be any legal state at this point 1957 // since the background collector may have yielded to the 1958 // foreground collector. 1959 break; 1960 case Marking: 1961 // initial marking in checkpointRootsInitialWork has been completed 1962 if (markFromRoots()) { // we were successful 1963 assert(_collectorState == Precleaning, "Collector state should " 1964 "have changed"); 1965 } else { 1966 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1967 } 1968 break; 1969 case Precleaning: 1970 // marking from roots in markFromRoots has been completed 1971 preclean(); 1972 assert(_collectorState == AbortablePreclean || 1973 _collectorState == FinalMarking, 1974 "Collector state should have changed"); 1975 break; 1976 case AbortablePreclean: 1977 abortable_preclean(); 1978 assert(_collectorState == FinalMarking, "Collector state should " 1979 "have changed"); 1980 break; 1981 case FinalMarking: 1982 { 1983 ReleaseForegroundGC x(this); 1984 1985 VM_CMS_Final_Remark final_remark_op(this); 1986 VMThread::execute(&final_remark_op); 1987 } 1988 assert(_foregroundGCShouldWait, "block post-condition"); 1989 break; 1990 case Sweeping: 1991 // final marking in checkpointRootsFinal has been completed 1992 sweep(); 1993 assert(_collectorState == Resizing, "Collector state change " 1994 "to Resizing must be done under the free_list_lock"); 1995 1996 case Resizing: { 1997 // Sweeping has been completed... 1998 // At this point the background collection has completed. 1999 // Don't move the call to compute_new_size() down 2000 // into code that might be executed if the background 2001 // collection was preempted. 2002 { 2003 ReleaseForegroundGC x(this); // unblock FG collection 2004 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2005 CMSTokenSync z(true); // not strictly needed. 2006 if (_collectorState == Resizing) { 2007 compute_new_size(); 2008 save_heap_summary(); 2009 _collectorState = Resetting; 2010 } else { 2011 assert(_collectorState == Idling, "The state should only change" 2012 " because the foreground collector has finished the collection"); 2013 } 2014 } 2015 break; 2016 } 2017 case Resetting: 2018 // CMS heap resizing has been completed 2019 reset(true); 2020 assert(_collectorState == Idling, "Collector state should " 2021 "have changed"); 2022 2023 MetaspaceGC::set_should_concurrent_collect(false); 2024 2025 stats().record_cms_end(); 2026 // Don't move the concurrent_phases_end() and compute_new_size() 2027 // calls to here because a preempted background collection 2028 // has it's state set to "Resetting". 2029 break; 2030 case Idling: 2031 default: 2032 ShouldNotReachHere(); 2033 break; 2034 } 2035 if (TraceCMSState) { 2036 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2037 Thread::current(), _collectorState); 2038 } 2039 assert(_foregroundGCShouldWait, "block post-condition"); 2040 } 2041 2042 // Should this be in gc_epilogue? 2043 collector_policy()->counters()->update_counters(); 2044 2045 { 2046 // Clear _foregroundGCShouldWait and, in the event that the 2047 // foreground collector is waiting, notify it, before 2048 // returning. 2049 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2050 _foregroundGCShouldWait = false; 2051 if (_foregroundGCIsActive) { 2052 CGC_lock->notify(); 2053 } 2054 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2055 "Possible deadlock"); 2056 } 2057 if (TraceCMSState) { 2058 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2059 " exiting collection CMS state %d", 2060 Thread::current(), _collectorState); 2061 } 2062 if (PrintGC && Verbose) { 2063 _cmsGen->print_heap_change(prev_used); 2064 } 2065 } 2066 2067 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2068 _cms_start_registered = true; 2069 _gc_timer_cm->register_gc_start(); 2070 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2071 } 2072 2073 void CMSCollector::register_gc_end() { 2074 if (_cms_start_registered) { 2075 report_heap_summary(GCWhen::AfterGC); 2076 2077 _gc_timer_cm->register_gc_end(); 2078 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2079 _cms_start_registered = false; 2080 } 2081 } 2082 2083 void CMSCollector::save_heap_summary() { 2084 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2085 _last_heap_summary = gch->create_heap_summary(); 2086 _last_metaspace_summary = gch->create_metaspace_summary(); 2087 } 2088 2089 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2090 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2091 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2092 } 2093 2094 bool CMSCollector::waitForForegroundGC() { 2095 bool res = false; 2096 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2097 "CMS thread should have CMS token"); 2098 // Block the foreground collector until the 2099 // background collectors decides whether to 2100 // yield. 2101 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2102 _foregroundGCShouldWait = true; 2103 if (_foregroundGCIsActive) { 2104 // The background collector yields to the 2105 // foreground collector and returns a value 2106 // indicating that it has yielded. The foreground 2107 // collector can proceed. 2108 res = true; 2109 _foregroundGCShouldWait = false; 2110 ConcurrentMarkSweepThread::clear_CMS_flag( 2111 ConcurrentMarkSweepThread::CMS_cms_has_token); 2112 ConcurrentMarkSweepThread::set_CMS_flag( 2113 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2114 // Get a possibly blocked foreground thread going 2115 CGC_lock->notify(); 2116 if (TraceCMSState) { 2117 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2118 Thread::current(), _collectorState); 2119 } 2120 while (_foregroundGCIsActive) { 2121 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2122 } 2123 ConcurrentMarkSweepThread::set_CMS_flag( 2124 ConcurrentMarkSweepThread::CMS_cms_has_token); 2125 ConcurrentMarkSweepThread::clear_CMS_flag( 2126 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2127 } 2128 if (TraceCMSState) { 2129 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2130 Thread::current(), _collectorState); 2131 } 2132 return res; 2133 } 2134 2135 // Because of the need to lock the free lists and other structures in 2136 // the collector, common to all the generations that the collector is 2137 // collecting, we need the gc_prologues of individual CMS generations 2138 // delegate to their collector. It may have been simpler had the 2139 // current infrastructure allowed one to call a prologue on a 2140 // collector. In the absence of that we have the generation's 2141 // prologue delegate to the collector, which delegates back 2142 // some "local" work to a worker method in the individual generations 2143 // that it's responsible for collecting, while itself doing any 2144 // work common to all generations it's responsible for. A similar 2145 // comment applies to the gc_epilogue()'s. 2146 // The role of the variable _between_prologue_and_epilogue is to 2147 // enforce the invocation protocol. 2148 void CMSCollector::gc_prologue(bool full) { 2149 // Call gc_prologue_work() for the CMSGen 2150 // we are responsible for. 2151 2152 // The following locking discipline assumes that we are only called 2153 // when the world is stopped. 2154 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2155 2156 // The CMSCollector prologue must call the gc_prologues for the 2157 // "generations" that it's responsible 2158 // for. 2159 2160 assert( Thread::current()->is_VM_thread() 2161 || ( CMSScavengeBeforeRemark 2162 && Thread::current()->is_ConcurrentGC_thread()), 2163 "Incorrect thread type for prologue execution"); 2164 2165 if (_between_prologue_and_epilogue) { 2166 // We have already been invoked; this is a gc_prologue delegation 2167 // from yet another CMS generation that we are responsible for, just 2168 // ignore it since all relevant work has already been done. 2169 return; 2170 } 2171 2172 // set a bit saying prologue has been called; cleared in epilogue 2173 _between_prologue_and_epilogue = true; 2174 // Claim locks for common data structures, then call gc_prologue_work() 2175 // for each CMSGen. 2176 2177 getFreelistLocks(); // gets free list locks on constituent spaces 2178 bitMapLock()->lock_without_safepoint_check(); 2179 2180 // Should call gc_prologue_work() for all cms gens we are responsible for 2181 bool duringMarking = _collectorState >= Marking 2182 && _collectorState < Sweeping; 2183 2184 // The young collections clear the modified oops state, which tells if 2185 // there are any modified oops in the class. The remark phase also needs 2186 // that information. Tell the young collection to save the union of all 2187 // modified klasses. 2188 if (duringMarking) { 2189 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2190 } 2191 2192 bool registerClosure = duringMarking; 2193 2194 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2195 &_modUnionClosurePar 2196 : &_modUnionClosure; 2197 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2198 2199 if (!full) { 2200 stats().record_gc0_begin(); 2201 } 2202 } 2203 2204 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2205 2206 _capacity_at_prologue = capacity(); 2207 _used_at_prologue = used(); 2208 2209 // Delegate to CMScollector which knows how to coordinate between 2210 // this and any other CMS generations that it is responsible for 2211 // collecting. 2212 collector()->gc_prologue(full); 2213 } 2214 2215 // This is a "private" interface for use by this generation's CMSCollector. 2216 // Not to be called directly by any other entity (for instance, 2217 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2218 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2219 bool registerClosure, ModUnionClosure* modUnionClosure) { 2220 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2221 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2222 "Should be NULL"); 2223 if (registerClosure) { 2224 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2225 } 2226 cmsSpace()->gc_prologue(); 2227 // Clear stat counters 2228 NOT_PRODUCT( 2229 assert(_numObjectsPromoted == 0, "check"); 2230 assert(_numWordsPromoted == 0, "check"); 2231 if (Verbose && PrintGC) { 2232 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2233 SIZE_FORMAT" bytes concurrently", 2234 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2235 } 2236 _numObjectsAllocated = 0; 2237 _numWordsAllocated = 0; 2238 ) 2239 } 2240 2241 void CMSCollector::gc_epilogue(bool full) { 2242 // The following locking discipline assumes that we are only called 2243 // when the world is stopped. 2244 assert(SafepointSynchronize::is_at_safepoint(), 2245 "world is stopped assumption"); 2246 2247 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2248 // if linear allocation blocks need to be appropriately marked to allow the 2249 // the blocks to be parsable. We also check here whether we need to nudge the 2250 // CMS collector thread to start a new cycle (if it's not already active). 2251 assert( Thread::current()->is_VM_thread() 2252 || ( CMSScavengeBeforeRemark 2253 && Thread::current()->is_ConcurrentGC_thread()), 2254 "Incorrect thread type for epilogue execution"); 2255 2256 if (!_between_prologue_and_epilogue) { 2257 // We have already been invoked; this is a gc_epilogue delegation 2258 // from yet another CMS generation that we are responsible for, just 2259 // ignore it since all relevant work has already been done. 2260 return; 2261 } 2262 assert(haveFreelistLocks(), "must have freelist locks"); 2263 assert_lock_strong(bitMapLock()); 2264 2265 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2266 2267 _cmsGen->gc_epilogue_work(full); 2268 2269 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2270 // in case sampling was not already enabled, enable it 2271 _start_sampling = true; 2272 } 2273 // reset _eden_chunk_array so sampling starts afresh 2274 _eden_chunk_index = 0; 2275 2276 size_t cms_used = _cmsGen->cmsSpace()->used(); 2277 2278 // update performance counters - this uses a special version of 2279 // update_counters() that allows the utilization to be passed as a 2280 // parameter, avoiding multiple calls to used(). 2281 // 2282 _cmsGen->update_counters(cms_used); 2283 2284 bitMapLock()->unlock(); 2285 releaseFreelistLocks(); 2286 2287 if (!CleanChunkPoolAsync) { 2288 Chunk::clean_chunk_pool(); 2289 } 2290 2291 set_did_compact(false); 2292 _between_prologue_and_epilogue = false; // ready for next cycle 2293 } 2294 2295 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2296 collector()->gc_epilogue(full); 2297 2298 // Also reset promotion tracking in par gc thread states. 2299 for (uint i = 0; i < ParallelGCThreads; i++) { 2300 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2301 } 2302 } 2303 2304 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2305 assert(!incremental_collection_failed(), "Should have been cleared"); 2306 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2307 cmsSpace()->gc_epilogue(); 2308 // Print stat counters 2309 NOT_PRODUCT( 2310 assert(_numObjectsAllocated == 0, "check"); 2311 assert(_numWordsAllocated == 0, "check"); 2312 if (Verbose && PrintGC) { 2313 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2314 SIZE_FORMAT" bytes", 2315 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2316 } 2317 _numObjectsPromoted = 0; 2318 _numWordsPromoted = 0; 2319 ) 2320 2321 if (PrintGC && Verbose) { 2322 // Call down the chain in contiguous_available needs the freelistLock 2323 // so print this out before releasing the freeListLock. 2324 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2325 contiguous_available()); 2326 } 2327 } 2328 2329 #ifndef PRODUCT 2330 bool CMSCollector::have_cms_token() { 2331 Thread* thr = Thread::current(); 2332 if (thr->is_VM_thread()) { 2333 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2334 } else if (thr->is_ConcurrentGC_thread()) { 2335 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2336 } else if (thr->is_GC_task_thread()) { 2337 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2338 ParGCRareEvent_lock->owned_by_self(); 2339 } 2340 return false; 2341 } 2342 #endif 2343 2344 // Check reachability of the given heap address in CMS generation, 2345 // treating all other generations as roots. 2346 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2347 // We could "guarantee" below, rather than assert, but I'll 2348 // leave these as "asserts" so that an adventurous debugger 2349 // could try this in the product build provided some subset of 2350 // the conditions were met, provided they were interested in the 2351 // results and knew that the computation below wouldn't interfere 2352 // with other concurrent computations mutating the structures 2353 // being read or written. 2354 assert(SafepointSynchronize::is_at_safepoint(), 2355 "Else mutations in object graph will make answer suspect"); 2356 assert(have_cms_token(), "Should hold cms token"); 2357 assert(haveFreelistLocks(), "must hold free list locks"); 2358 assert_lock_strong(bitMapLock()); 2359 2360 // Clear the marking bit map array before starting, but, just 2361 // for kicks, first report if the given address is already marked 2362 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2363 _markBitMap.isMarked(addr) ? "" : " not"); 2364 2365 if (verify_after_remark()) { 2366 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2367 bool result = verification_mark_bm()->isMarked(addr); 2368 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2369 result ? "IS" : "is NOT"); 2370 return result; 2371 } else { 2372 gclog_or_tty->print_cr("Could not compute result"); 2373 return false; 2374 } 2375 } 2376 2377 2378 void 2379 CMSCollector::print_on_error(outputStream* st) { 2380 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2381 if (collector != NULL) { 2382 CMSBitMap* bitmap = &collector->_markBitMap; 2383 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2384 bitmap->print_on_error(st, " Bits: "); 2385 2386 st->cr(); 2387 2388 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2389 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2390 mut_bitmap->print_on_error(st, " Bits: "); 2391 } 2392 } 2393 2394 //////////////////////////////////////////////////////// 2395 // CMS Verification Support 2396 //////////////////////////////////////////////////////// 2397 // Following the remark phase, the following invariant 2398 // should hold -- each object in the CMS heap which is 2399 // marked in markBitMap() should be marked in the verification_mark_bm(). 2400 2401 class VerifyMarkedClosure: public BitMapClosure { 2402 CMSBitMap* _marks; 2403 bool _failed; 2404 2405 public: 2406 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2407 2408 bool do_bit(size_t offset) { 2409 HeapWord* addr = _marks->offsetToHeapWord(offset); 2410 if (!_marks->isMarked(addr)) { 2411 oop(addr)->print_on(gclog_or_tty); 2412 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2413 _failed = true; 2414 } 2415 return true; 2416 } 2417 2418 bool failed() { return _failed; } 2419 }; 2420 2421 bool CMSCollector::verify_after_remark(bool silent) { 2422 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2423 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2424 static bool init = false; 2425 2426 assert(SafepointSynchronize::is_at_safepoint(), 2427 "Else mutations in object graph will make answer suspect"); 2428 assert(have_cms_token(), 2429 "Else there may be mutual interference in use of " 2430 " verification data structures"); 2431 assert(_collectorState > Marking && _collectorState <= Sweeping, 2432 "Else marking info checked here may be obsolete"); 2433 assert(haveFreelistLocks(), "must hold free list locks"); 2434 assert_lock_strong(bitMapLock()); 2435 2436 2437 // Allocate marking bit map if not already allocated 2438 if (!init) { // first time 2439 if (!verification_mark_bm()->allocate(_span)) { 2440 return false; 2441 } 2442 init = true; 2443 } 2444 2445 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2446 2447 // Turn off refs discovery -- so we will be tracing through refs. 2448 // This is as intended, because by this time 2449 // GC must already have cleared any refs that need to be cleared, 2450 // and traced those that need to be marked; moreover, 2451 // the marking done here is not going to interfere in any 2452 // way with the marking information used by GC. 2453 NoRefDiscovery no_discovery(ref_processor()); 2454 2455 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2456 2457 // Clear any marks from a previous round 2458 verification_mark_bm()->clear_all(); 2459 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2460 verify_work_stacks_empty(); 2461 2462 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2463 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2464 // Update the saved marks which may affect the root scans. 2465 gch->save_marks(); 2466 2467 if (CMSRemarkVerifyVariant == 1) { 2468 // In this first variant of verification, we complete 2469 // all marking, then check if the new marks-vector is 2470 // a subset of the CMS marks-vector. 2471 verify_after_remark_work_1(); 2472 } else if (CMSRemarkVerifyVariant == 2) { 2473 // In this second variant of verification, we flag an error 2474 // (i.e. an object reachable in the new marks-vector not reachable 2475 // in the CMS marks-vector) immediately, also indicating the 2476 // identify of an object (A) that references the unmarked object (B) -- 2477 // presumably, a mutation to A failed to be picked up by preclean/remark? 2478 verify_after_remark_work_2(); 2479 } else { 2480 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2481 CMSRemarkVerifyVariant); 2482 } 2483 if (!silent) gclog_or_tty->print(" done] "); 2484 return true; 2485 } 2486 2487 void CMSCollector::verify_after_remark_work_1() { 2488 ResourceMark rm; 2489 HandleMark hm; 2490 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2491 2492 // Get a clear set of claim bits for the roots processing to work with. 2493 ClassLoaderDataGraph::clear_claimed_marks(); 2494 2495 // Mark from roots one level into CMS 2496 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2497 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2498 2499 gch->gen_process_roots(_cmsGen->level(), 2500 true, // younger gens are roots 2501 true, // activate StrongRootsScope 2502 SharedHeap::ScanningOption(roots_scanning_options()), 2503 should_unload_classes(), 2504 ¬Older, 2505 NULL, 2506 NULL); // SSS: Provide correct closure 2507 2508 // Now mark from the roots 2509 MarkFromRootsClosure markFromRootsClosure(this, _span, 2510 verification_mark_bm(), verification_mark_stack(), 2511 false /* don't yield */, true /* verifying */); 2512 assert(_restart_addr == NULL, "Expected pre-condition"); 2513 verification_mark_bm()->iterate(&markFromRootsClosure); 2514 while (_restart_addr != NULL) { 2515 // Deal with stack overflow: by restarting at the indicated 2516 // address. 2517 HeapWord* ra = _restart_addr; 2518 markFromRootsClosure.reset(ra); 2519 _restart_addr = NULL; 2520 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2521 } 2522 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2523 verify_work_stacks_empty(); 2524 2525 // Marking completed -- now verify that each bit marked in 2526 // verification_mark_bm() is also marked in markBitMap(); flag all 2527 // errors by printing corresponding objects. 2528 VerifyMarkedClosure vcl(markBitMap()); 2529 verification_mark_bm()->iterate(&vcl); 2530 if (vcl.failed()) { 2531 gclog_or_tty->print("Verification failed"); 2532 Universe::heap()->print_on(gclog_or_tty); 2533 fatal("CMS: failed marking verification after remark"); 2534 } 2535 } 2536 2537 class VerifyKlassOopsKlassClosure : public KlassClosure { 2538 class VerifyKlassOopsClosure : public OopClosure { 2539 CMSBitMap* _bitmap; 2540 public: 2541 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2542 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2543 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2544 } _oop_closure; 2545 public: 2546 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2547 void do_klass(Klass* k) { 2548 k->oops_do(&_oop_closure); 2549 } 2550 }; 2551 2552 void CMSCollector::verify_after_remark_work_2() { 2553 ResourceMark rm; 2554 HandleMark hm; 2555 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2556 2557 // Get a clear set of claim bits for the roots processing to work with. 2558 ClassLoaderDataGraph::clear_claimed_marks(); 2559 2560 // Mark from roots one level into CMS 2561 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2562 markBitMap()); 2563 CLDToOopClosure cld_closure(¬Older, true); 2564 2565 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2566 2567 gch->gen_process_roots(_cmsGen->level(), 2568 true, // younger gens are roots 2569 true, // activate StrongRootsScope 2570 SharedHeap::ScanningOption(roots_scanning_options()), 2571 should_unload_classes(), 2572 ¬Older, 2573 NULL, 2574 &cld_closure); 2575 2576 // Now mark from the roots 2577 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2578 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2579 assert(_restart_addr == NULL, "Expected pre-condition"); 2580 verification_mark_bm()->iterate(&markFromRootsClosure); 2581 while (_restart_addr != NULL) { 2582 // Deal with stack overflow: by restarting at the indicated 2583 // address. 2584 HeapWord* ra = _restart_addr; 2585 markFromRootsClosure.reset(ra); 2586 _restart_addr = NULL; 2587 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2588 } 2589 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2590 verify_work_stacks_empty(); 2591 2592 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2593 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2594 2595 // Marking completed -- now verify that each bit marked in 2596 // verification_mark_bm() is also marked in markBitMap(); flag all 2597 // errors by printing corresponding objects. 2598 VerifyMarkedClosure vcl(markBitMap()); 2599 verification_mark_bm()->iterate(&vcl); 2600 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2601 } 2602 2603 void ConcurrentMarkSweepGeneration::save_marks() { 2604 // delegate to CMS space 2605 cmsSpace()->save_marks(); 2606 for (uint i = 0; i < ParallelGCThreads; i++) { 2607 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2608 } 2609 } 2610 2611 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2612 return cmsSpace()->no_allocs_since_save_marks(); 2613 } 2614 2615 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2616 \ 2617 void ConcurrentMarkSweepGeneration:: \ 2618 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2619 cl->set_generation(this); \ 2620 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2621 cl->reset_generation(); \ 2622 save_marks(); \ 2623 } 2624 2625 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2626 2627 void 2628 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2629 cl->set_generation(this); 2630 younger_refs_in_space_iterate(_cmsSpace, cl); 2631 cl->reset_generation(); 2632 } 2633 2634 void 2635 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2636 if (freelistLock()->owned_by_self()) { 2637 Generation::oop_iterate(cl); 2638 } else { 2639 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2640 Generation::oop_iterate(cl); 2641 } 2642 } 2643 2644 void 2645 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2646 if (freelistLock()->owned_by_self()) { 2647 Generation::object_iterate(cl); 2648 } else { 2649 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2650 Generation::object_iterate(cl); 2651 } 2652 } 2653 2654 void 2655 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2656 if (freelistLock()->owned_by_self()) { 2657 Generation::safe_object_iterate(cl); 2658 } else { 2659 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2660 Generation::safe_object_iterate(cl); 2661 } 2662 } 2663 2664 void 2665 ConcurrentMarkSweepGeneration::post_compact() { 2666 } 2667 2668 void 2669 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2670 // Fix the linear allocation blocks to look like free blocks. 2671 2672 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2673 // are not called when the heap is verified during universe initialization and 2674 // at vm shutdown. 2675 if (freelistLock()->owned_by_self()) { 2676 cmsSpace()->prepare_for_verify(); 2677 } else { 2678 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2679 cmsSpace()->prepare_for_verify(); 2680 } 2681 } 2682 2683 void 2684 ConcurrentMarkSweepGeneration::verify() { 2685 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2686 // are not called when the heap is verified during universe initialization and 2687 // at vm shutdown. 2688 if (freelistLock()->owned_by_self()) { 2689 cmsSpace()->verify(); 2690 } else { 2691 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2692 cmsSpace()->verify(); 2693 } 2694 } 2695 2696 void CMSCollector::verify() { 2697 _cmsGen->verify(); 2698 } 2699 2700 #ifndef PRODUCT 2701 bool CMSCollector::overflow_list_is_empty() const { 2702 assert(_num_par_pushes >= 0, "Inconsistency"); 2703 if (_overflow_list == NULL) { 2704 assert(_num_par_pushes == 0, "Inconsistency"); 2705 } 2706 return _overflow_list == NULL; 2707 } 2708 2709 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2710 // merely consolidate assertion checks that appear to occur together frequently. 2711 void CMSCollector::verify_work_stacks_empty() const { 2712 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2713 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2714 } 2715 2716 void CMSCollector::verify_overflow_empty() const { 2717 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2718 assert(no_preserved_marks(), "No preserved marks"); 2719 } 2720 #endif // PRODUCT 2721 2722 // Decide if we want to enable class unloading as part of the 2723 // ensuing concurrent GC cycle. We will collect and 2724 // unload classes if it's the case that: 2725 // (1) an explicit gc request has been made and the flag 2726 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2727 // (2) (a) class unloading is enabled at the command line, and 2728 // (b) old gen is getting really full 2729 // NOTE: Provided there is no change in the state of the heap between 2730 // calls to this method, it should have idempotent results. Moreover, 2731 // its results should be monotonically increasing (i.e. going from 0 to 1, 2732 // but not 1 to 0) between successive calls between which the heap was 2733 // not collected. For the implementation below, it must thus rely on 2734 // the property that concurrent_cycles_since_last_unload() 2735 // will not decrease unless a collection cycle happened and that 2736 // _cmsGen->is_too_full() are 2737 // themselves also monotonic in that sense. See check_monotonicity() 2738 // below. 2739 void CMSCollector::update_should_unload_classes() { 2740 _should_unload_classes = false; 2741 // Condition 1 above 2742 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2743 _should_unload_classes = true; 2744 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2745 // Disjuncts 2.b.(i,ii,iii) above 2746 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2747 CMSClassUnloadingMaxInterval) 2748 || _cmsGen->is_too_full(); 2749 } 2750 } 2751 2752 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2753 bool res = should_concurrent_collect(); 2754 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2755 return res; 2756 } 2757 2758 void CMSCollector::setup_cms_unloading_and_verification_state() { 2759 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2760 || VerifyBeforeExit; 2761 const int rso = SharedHeap::SO_AllCodeCache; 2762 2763 // We set the proper root for this CMS cycle here. 2764 if (should_unload_classes()) { // Should unload classes this cycle 2765 remove_root_scanning_option(rso); // Shrink the root set appropriately 2766 set_verifying(should_verify); // Set verification state for this cycle 2767 return; // Nothing else needs to be done at this time 2768 } 2769 2770 // Not unloading classes this cycle 2771 assert(!should_unload_classes(), "Inconsistency!"); 2772 2773 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2774 // Include symbols, strings and code cache elements to prevent their resurrection. 2775 add_root_scanning_option(rso); 2776 set_verifying(true); 2777 } else if (verifying() && !should_verify) { 2778 // We were verifying, but some verification flags got disabled. 2779 set_verifying(false); 2780 // Exclude symbols, strings and code cache elements from root scanning to 2781 // reduce IM and RM pauses. 2782 remove_root_scanning_option(rso); 2783 } 2784 } 2785 2786 2787 #ifndef PRODUCT 2788 HeapWord* CMSCollector::block_start(const void* p) const { 2789 const HeapWord* addr = (HeapWord*)p; 2790 if (_span.contains(p)) { 2791 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2792 return _cmsGen->cmsSpace()->block_start(p); 2793 } 2794 } 2795 return NULL; 2796 } 2797 #endif 2798 2799 HeapWord* 2800 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2801 bool tlab, 2802 bool parallel) { 2803 CMSSynchronousYieldRequest yr; 2804 assert(!tlab, "Can't deal with TLAB allocation"); 2805 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2806 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 2807 CMSExpansionCause::_satisfy_allocation); 2808 if (GCExpandToAllocateDelayMillis > 0) { 2809 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2810 } 2811 return have_lock_and_allocate(word_size, tlab); 2812 } 2813 2814 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 2815 // TenuredGeneration, which makes me wonder if we should move this 2816 // to CardGeneration and share it... 2817 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 2818 return CardGeneration::expand(bytes, expand_bytes); 2819 } 2820 2821 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 2822 CMSExpansionCause::Cause cause) 2823 { 2824 2825 bool success = expand(bytes, expand_bytes); 2826 2827 // remember why we expanded; this information is used 2828 // by shouldConcurrentCollect() when making decisions on whether to start 2829 // a new CMS cycle. 2830 if (success) { 2831 set_expansion_cause(cause); 2832 if (PrintGCDetails && Verbose) { 2833 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2834 CMSExpansionCause::to_string(cause)); 2835 } 2836 } 2837 } 2838 2839 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2840 HeapWord* res = NULL; 2841 MutexLocker x(ParGCRareEvent_lock); 2842 while (true) { 2843 // Expansion by some other thread might make alloc OK now: 2844 res = ps->lab.alloc(word_sz); 2845 if (res != NULL) return res; 2846 // If there's not enough expansion space available, give up. 2847 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2848 return NULL; 2849 } 2850 // Otherwise, we try expansion. 2851 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 2852 CMSExpansionCause::_allocate_par_lab); 2853 // Now go around the loop and try alloc again; 2854 // A competing par_promote might beat us to the expansion space, 2855 // so we may go around the loop again if promotion fails again. 2856 if (GCExpandToAllocateDelayMillis > 0) { 2857 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2858 } 2859 } 2860 } 2861 2862 2863 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2864 PromotionInfo* promo) { 2865 MutexLocker x(ParGCRareEvent_lock); 2866 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2867 while (true) { 2868 // Expansion by some other thread might make alloc OK now: 2869 if (promo->ensure_spooling_space()) { 2870 assert(promo->has_spooling_space(), 2871 "Post-condition of successful ensure_spooling_space()"); 2872 return true; 2873 } 2874 // If there's not enough expansion space available, give up. 2875 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2876 return false; 2877 } 2878 // Otherwise, we try expansion. 2879 expand(refill_size_bytes, MinHeapDeltaBytes, 2880 CMSExpansionCause::_allocate_par_spooling_space); 2881 // Now go around the loop and try alloc again; 2882 // A competing allocation might beat us to the expansion space, 2883 // so we may go around the loop again if allocation fails again. 2884 if (GCExpandToAllocateDelayMillis > 0) { 2885 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2886 } 2887 } 2888 } 2889 2890 2891 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 2892 assert_locked_or_safepoint(ExpandHeap_lock); 2893 // Shrink committed space 2894 _virtual_space.shrink_by(bytes); 2895 // Shrink space; this also shrinks the space's BOT 2896 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 2897 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 2898 // Shrink the shared block offset array 2899 _bts->resize(new_word_size); 2900 MemRegion mr(_cmsSpace->bottom(), new_word_size); 2901 // Shrink the card table 2902 Universe::heap()->barrier_set()->resize_covered_region(mr); 2903 2904 if (Verbose && PrintGC) { 2905 size_t new_mem_size = _virtual_space.committed_size(); 2906 size_t old_mem_size = new_mem_size + bytes; 2907 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 2908 name(), old_mem_size/K, new_mem_size/K); 2909 } 2910 } 2911 2912 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2913 assert_locked_or_safepoint(Heap_lock); 2914 size_t size = ReservedSpace::page_align_size_down(bytes); 2915 // Only shrink if a compaction was done so that all the free space 2916 // in the generation is in a contiguous block at the end. 2917 if (size > 0 && did_compact()) { 2918 shrink_by(size); 2919 } 2920 } 2921 2922 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 2923 assert_locked_or_safepoint(Heap_lock); 2924 bool result = _virtual_space.expand_by(bytes); 2925 if (result) { 2926 size_t new_word_size = 2927 heap_word_size(_virtual_space.committed_size()); 2928 MemRegion mr(_cmsSpace->bottom(), new_word_size); 2929 _bts->resize(new_word_size); // resize the block offset shared array 2930 Universe::heap()->barrier_set()->resize_covered_region(mr); 2931 // Hmmmm... why doesn't CFLS::set_end verify locking? 2932 // This is quite ugly; FIX ME XXX 2933 _cmsSpace->assert_locked(freelistLock()); 2934 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 2935 2936 // update the space and generation capacity counters 2937 if (UsePerfData) { 2938 _space_counters->update_capacity(); 2939 _gen_counters->update_all(); 2940 } 2941 2942 if (Verbose && PrintGC) { 2943 size_t new_mem_size = _virtual_space.committed_size(); 2944 size_t old_mem_size = new_mem_size - bytes; 2945 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 2946 name(), old_mem_size/K, bytes/K, new_mem_size/K); 2947 } 2948 } 2949 return result; 2950 } 2951 2952 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 2953 assert_locked_or_safepoint(Heap_lock); 2954 bool success = true; 2955 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 2956 if (remaining_bytes > 0) { 2957 success = grow_by(remaining_bytes); 2958 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 2959 } 2960 return success; 2961 } 2962 2963 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2964 assert_locked_or_safepoint(Heap_lock); 2965 assert_lock_strong(freelistLock()); 2966 if (PrintGCDetails && Verbose) { 2967 warning("Shrinking of CMS not yet implemented"); 2968 } 2969 return; 2970 } 2971 2972 2973 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2974 // phases. 2975 class CMSPhaseAccounting: public StackObj { 2976 public: 2977 CMSPhaseAccounting(CMSCollector *collector, 2978 const char *phase, 2979 const GCId gc_id, 2980 bool print_cr = true); 2981 ~CMSPhaseAccounting(); 2982 2983 private: 2984 CMSCollector *_collector; 2985 const char *_phase; 2986 elapsedTimer _wallclock; 2987 bool _print_cr; 2988 const GCId _gc_id; 2989 2990 public: 2991 // Not MT-safe; so do not pass around these StackObj's 2992 // where they may be accessed by other threads. 2993 jlong wallclock_millis() { 2994 assert(_wallclock.is_active(), "Wall clock should not stop"); 2995 _wallclock.stop(); // to record time 2996 jlong ret = _wallclock.milliseconds(); 2997 _wallclock.start(); // restart 2998 return ret; 2999 } 3000 }; 3001 3002 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3003 const char *phase, 3004 const GCId gc_id, 3005 bool print_cr) : 3006 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 3007 3008 if (PrintCMSStatistics != 0) { 3009 _collector->resetYields(); 3010 } 3011 if (PrintGCDetails) { 3012 gclog_or_tty->gclog_stamp(_gc_id); 3013 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3014 _collector->cmsGen()->short_name(), _phase); 3015 } 3016 _collector->resetTimer(); 3017 _wallclock.start(); 3018 _collector->startTimer(); 3019 } 3020 3021 CMSPhaseAccounting::~CMSPhaseAccounting() { 3022 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3023 _collector->stopTimer(); 3024 _wallclock.stop(); 3025 if (PrintGCDetails) { 3026 gclog_or_tty->gclog_stamp(_gc_id); 3027 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3028 _collector->cmsGen()->short_name(), 3029 _phase, _collector->timerValue(), _wallclock.seconds()); 3030 if (_print_cr) { 3031 gclog_or_tty->cr(); 3032 } 3033 if (PrintCMSStatistics != 0) { 3034 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3035 _collector->yields()); 3036 } 3037 } 3038 } 3039 3040 // CMS work 3041 3042 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3043 class CMSParMarkTask : public AbstractGangTask { 3044 protected: 3045 CMSCollector* _collector; 3046 int _n_workers; 3047 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3048 AbstractGangTask(name), 3049 _collector(collector), 3050 _n_workers(n_workers) {} 3051 // Work method in support of parallel rescan ... of young gen spaces 3052 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3053 ContiguousSpace* space, 3054 HeapWord** chunk_array, size_t chunk_top); 3055 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3056 }; 3057 3058 // Parallel initial mark task 3059 class CMSParInitialMarkTask: public CMSParMarkTask { 3060 public: 3061 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3062 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3063 collector, n_workers) {} 3064 void work(uint worker_id); 3065 }; 3066 3067 // Checkpoint the roots into this generation from outside 3068 // this generation. [Note this initial checkpoint need only 3069 // be approximate -- we'll do a catch up phase subsequently.] 3070 void CMSCollector::checkpointRootsInitial() { 3071 assert(_collectorState == InitialMarking, "Wrong collector state"); 3072 check_correct_thread_executing(); 3073 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3074 3075 save_heap_summary(); 3076 report_heap_summary(GCWhen::BeforeGC); 3077 3078 ReferenceProcessor* rp = ref_processor(); 3079 SpecializationStats::clear(); 3080 assert(_restart_addr == NULL, "Control point invariant"); 3081 { 3082 // acquire locks for subsequent manipulations 3083 MutexLockerEx x(bitMapLock(), 3084 Mutex::_no_safepoint_check_flag); 3085 checkpointRootsInitialWork(); 3086 // enable ("weak") refs discovery 3087 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3088 _collectorState = Marking; 3089 } 3090 SpecializationStats::print(); 3091 } 3092 3093 void CMSCollector::checkpointRootsInitialWork() { 3094 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3095 assert(_collectorState == InitialMarking, "just checking"); 3096 3097 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3098 // precede our marking with a collection of all 3099 // younger generations to keep floating garbage to a minimum. 3100 // XXX: we won't do this for now -- it's an optimization to be done later. 3101 3102 // already have locks 3103 assert_lock_strong(bitMapLock()); 3104 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3105 3106 // Setup the verification and class unloading state for this 3107 // CMS collection cycle. 3108 setup_cms_unloading_and_verification_state(); 3109 3110 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3111 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3112 3113 // Reset all the PLAB chunk arrays if necessary. 3114 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3115 reset_survivor_plab_arrays(); 3116 } 3117 3118 ResourceMark rm; 3119 HandleMark hm; 3120 3121 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3122 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3123 3124 verify_work_stacks_empty(); 3125 verify_overflow_empty(); 3126 3127 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3128 // Update the saved marks which may affect the root scans. 3129 gch->save_marks(); 3130 3131 // weak reference processing has not started yet. 3132 ref_processor()->set_enqueuing_is_done(false); 3133 3134 // Need to remember all newly created CLDs, 3135 // so that we can guarantee that the remark finds them. 3136 ClassLoaderDataGraph::remember_new_clds(true); 3137 3138 // Whenever a CLD is found, it will be claimed before proceeding to mark 3139 // the klasses. The claimed marks need to be cleared before marking starts. 3140 ClassLoaderDataGraph::clear_claimed_marks(); 3141 3142 if (CMSPrintEdenSurvivorChunks) { 3143 print_eden_and_survivor_chunk_arrays(); 3144 } 3145 3146 { 3147 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3148 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3149 // The parallel version. 3150 FlexibleWorkGang* workers = gch->workers(); 3151 assert(workers != NULL, "Need parallel worker threads."); 3152 int n_workers = workers->active_workers(); 3153 CMSParInitialMarkTask tsk(this, n_workers); 3154 gch->set_par_threads(n_workers); 3155 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3156 if (n_workers > 1) { 3157 GenCollectedHeap::StrongRootsScope srs(gch); 3158 workers->run_task(&tsk); 3159 } else { 3160 GenCollectedHeap::StrongRootsScope srs(gch); 3161 tsk.work(0); 3162 } 3163 gch->set_par_threads(0); 3164 } else { 3165 // The serial version. 3166 CLDToOopClosure cld_closure(¬Older, true); 3167 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3168 gch->gen_process_roots(_cmsGen->level(), 3169 true, // younger gens are roots 3170 true, // activate StrongRootsScope 3171 SharedHeap::ScanningOption(roots_scanning_options()), 3172 should_unload_classes(), 3173 ¬Older, 3174 NULL, 3175 &cld_closure); 3176 } 3177 } 3178 3179 // Clear mod-union table; it will be dirtied in the prologue of 3180 // CMS generation per each younger generation collection. 3181 3182 assert(_modUnionTable.isAllClear(), 3183 "Was cleared in most recent final checkpoint phase" 3184 " or no bits are set in the gc_prologue before the start of the next " 3185 "subsequent marking phase."); 3186 3187 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3188 3189 // Save the end of the used_region of the constituent generations 3190 // to be used to limit the extent of sweep in each generation. 3191 save_sweep_limits(); 3192 verify_overflow_empty(); 3193 } 3194 3195 bool CMSCollector::markFromRoots() { 3196 // we might be tempted to assert that: 3197 // assert(!SafepointSynchronize::is_at_safepoint(), 3198 // "inconsistent argument?"); 3199 // However that wouldn't be right, because it's possible that 3200 // a safepoint is indeed in progress as a younger generation 3201 // stop-the-world GC happens even as we mark in this generation. 3202 assert(_collectorState == Marking, "inconsistent state?"); 3203 check_correct_thread_executing(); 3204 verify_overflow_empty(); 3205 3206 // Weak ref discovery note: We may be discovering weak 3207 // refs in this generation concurrent (but interleaved) with 3208 // weak ref discovery by a younger generation collector. 3209 3210 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3211 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3212 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3213 bool res = markFromRootsWork(); 3214 if (res) { 3215 _collectorState = Precleaning; 3216 } else { // We failed and a foreground collection wants to take over 3217 assert(_foregroundGCIsActive, "internal state inconsistency"); 3218 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3219 if (PrintGCDetails) { 3220 gclog_or_tty->print_cr("bailing out to foreground collection"); 3221 } 3222 } 3223 verify_overflow_empty(); 3224 return res; 3225 } 3226 3227 bool CMSCollector::markFromRootsWork() { 3228 // iterate over marked bits in bit map, doing a full scan and mark 3229 // from these roots using the following algorithm: 3230 // . if oop is to the right of the current scan pointer, 3231 // mark corresponding bit (we'll process it later) 3232 // . else (oop is to left of current scan pointer) 3233 // push oop on marking stack 3234 // . drain the marking stack 3235 3236 // Note that when we do a marking step we need to hold the 3237 // bit map lock -- recall that direct allocation (by mutators) 3238 // and promotion (by younger generation collectors) is also 3239 // marking the bit map. [the so-called allocate live policy.] 3240 // Because the implementation of bit map marking is not 3241 // robust wrt simultaneous marking of bits in the same word, 3242 // we need to make sure that there is no such interference 3243 // between concurrent such updates. 3244 3245 // already have locks 3246 assert_lock_strong(bitMapLock()); 3247 3248 verify_work_stacks_empty(); 3249 verify_overflow_empty(); 3250 bool result = false; 3251 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3252 result = do_marking_mt(); 3253 } else { 3254 result = do_marking_st(); 3255 } 3256 return result; 3257 } 3258 3259 // Forward decl 3260 class CMSConcMarkingTask; 3261 3262 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3263 CMSCollector* _collector; 3264 CMSConcMarkingTask* _task; 3265 public: 3266 virtual void yield(); 3267 3268 // "n_threads" is the number of threads to be terminated. 3269 // "queue_set" is a set of work queues of other threads. 3270 // "collector" is the CMS collector associated with this task terminator. 3271 // "yield" indicates whether we need the gang as a whole to yield. 3272 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3273 ParallelTaskTerminator(n_threads, queue_set), 3274 _collector(collector) { } 3275 3276 void set_task(CMSConcMarkingTask* task) { 3277 _task = task; 3278 } 3279 }; 3280 3281 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3282 CMSConcMarkingTask* _task; 3283 public: 3284 bool should_exit_termination(); 3285 void set_task(CMSConcMarkingTask* task) { 3286 _task = task; 3287 } 3288 }; 3289 3290 // MT Concurrent Marking Task 3291 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3292 CMSCollector* _collector; 3293 int _n_workers; // requested/desired # workers 3294 bool _result; 3295 CompactibleFreeListSpace* _cms_space; 3296 char _pad_front[64]; // padding to ... 3297 HeapWord* _global_finger; // ... avoid sharing cache line 3298 char _pad_back[64]; 3299 HeapWord* _restart_addr; 3300 3301 // Exposed here for yielding support 3302 Mutex* const _bit_map_lock; 3303 3304 // The per thread work queues, available here for stealing 3305 OopTaskQueueSet* _task_queues; 3306 3307 // Termination (and yielding) support 3308 CMSConcMarkingTerminator _term; 3309 CMSConcMarkingTerminatorTerminator _term_term; 3310 3311 public: 3312 CMSConcMarkingTask(CMSCollector* collector, 3313 CompactibleFreeListSpace* cms_space, 3314 YieldingFlexibleWorkGang* workers, 3315 OopTaskQueueSet* task_queues): 3316 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3317 _collector(collector), 3318 _cms_space(cms_space), 3319 _n_workers(0), _result(true), 3320 _task_queues(task_queues), 3321 _term(_n_workers, task_queues, _collector), 3322 _bit_map_lock(collector->bitMapLock()) 3323 { 3324 _requested_size = _n_workers; 3325 _term.set_task(this); 3326 _term_term.set_task(this); 3327 _restart_addr = _global_finger = _cms_space->bottom(); 3328 } 3329 3330 3331 OopTaskQueueSet* task_queues() { return _task_queues; } 3332 3333 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3334 3335 HeapWord** global_finger_addr() { return &_global_finger; } 3336 3337 CMSConcMarkingTerminator* terminator() { return &_term; } 3338 3339 virtual void set_for_termination(int active_workers) { 3340 terminator()->reset_for_reuse(active_workers); 3341 } 3342 3343 void work(uint worker_id); 3344 bool should_yield() { 3345 return ConcurrentMarkSweepThread::should_yield() 3346 && !_collector->foregroundGCIsActive(); 3347 } 3348 3349 virtual void coordinator_yield(); // stuff done by coordinator 3350 bool result() { return _result; } 3351 3352 void reset(HeapWord* ra) { 3353 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3354 _restart_addr = _global_finger = ra; 3355 _term.reset_for_reuse(); 3356 } 3357 3358 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3359 OopTaskQueue* work_q); 3360 3361 private: 3362 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3363 void do_work_steal(int i); 3364 void bump_global_finger(HeapWord* f); 3365 }; 3366 3367 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3368 assert(_task != NULL, "Error"); 3369 return _task->yielding(); 3370 // Note that we do not need the disjunct || _task->should_yield() above 3371 // because we want terminating threads to yield only if the task 3372 // is already in the midst of yielding, which happens only after at least one 3373 // thread has yielded. 3374 } 3375 3376 void CMSConcMarkingTerminator::yield() { 3377 if (_task->should_yield()) { 3378 _task->yield(); 3379 } else { 3380 ParallelTaskTerminator::yield(); 3381 } 3382 } 3383 3384 //////////////////////////////////////////////////////////////// 3385 // Concurrent Marking Algorithm Sketch 3386 //////////////////////////////////////////////////////////////// 3387 // Until all tasks exhausted (both spaces): 3388 // -- claim next available chunk 3389 // -- bump global finger via CAS 3390 // -- find first object that starts in this chunk 3391 // and start scanning bitmap from that position 3392 // -- scan marked objects for oops 3393 // -- CAS-mark target, and if successful: 3394 // . if target oop is above global finger (volatile read) 3395 // nothing to do 3396 // . if target oop is in chunk and above local finger 3397 // then nothing to do 3398 // . else push on work-queue 3399 // -- Deal with possible overflow issues: 3400 // . local work-queue overflow causes stuff to be pushed on 3401 // global (common) overflow queue 3402 // . always first empty local work queue 3403 // . then get a batch of oops from global work queue if any 3404 // . then do work stealing 3405 // -- When all tasks claimed (both spaces) 3406 // and local work queue empty, 3407 // then in a loop do: 3408 // . check global overflow stack; steal a batch of oops and trace 3409 // . try to steal from other threads oif GOS is empty 3410 // . if neither is available, offer termination 3411 // -- Terminate and return result 3412 // 3413 void CMSConcMarkingTask::work(uint worker_id) { 3414 elapsedTimer _timer; 3415 ResourceMark rm; 3416 HandleMark hm; 3417 3418 DEBUG_ONLY(_collector->verify_overflow_empty();) 3419 3420 // Before we begin work, our work queue should be empty 3421 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3422 // Scan the bitmap covering _cms_space, tracing through grey objects. 3423 _timer.start(); 3424 do_scan_and_mark(worker_id, _cms_space); 3425 _timer.stop(); 3426 if (PrintCMSStatistics != 0) { 3427 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3428 worker_id, _timer.seconds()); 3429 // XXX: need xxx/xxx type of notation, two timers 3430 } 3431 3432 // ... do work stealing 3433 _timer.reset(); 3434 _timer.start(); 3435 do_work_steal(worker_id); 3436 _timer.stop(); 3437 if (PrintCMSStatistics != 0) { 3438 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3439 worker_id, _timer.seconds()); 3440 // XXX: need xxx/xxx type of notation, two timers 3441 } 3442 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3443 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3444 // Note that under the current task protocol, the 3445 // following assertion is true even of the spaces 3446 // expanded since the completion of the concurrent 3447 // marking. XXX This will likely change under a strict 3448 // ABORT semantics. 3449 // After perm removal the comparison was changed to 3450 // greater than or equal to from strictly greater than. 3451 // Before perm removal the highest address sweep would 3452 // have been at the end of perm gen but now is at the 3453 // end of the tenured gen. 3454 assert(_global_finger >= _cms_space->end(), 3455 "All tasks have been completed"); 3456 DEBUG_ONLY(_collector->verify_overflow_empty();) 3457 } 3458 3459 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3460 HeapWord* read = _global_finger; 3461 HeapWord* cur = read; 3462 while (f > read) { 3463 cur = read; 3464 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3465 if (cur == read) { 3466 // our cas succeeded 3467 assert(_global_finger >= f, "protocol consistency"); 3468 break; 3469 } 3470 } 3471 } 3472 3473 // This is really inefficient, and should be redone by 3474 // using (not yet available) block-read and -write interfaces to the 3475 // stack and the work_queue. XXX FIX ME !!! 3476 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3477 OopTaskQueue* work_q) { 3478 // Fast lock-free check 3479 if (ovflw_stk->length() == 0) { 3480 return false; 3481 } 3482 assert(work_q->size() == 0, "Shouldn't steal"); 3483 MutexLockerEx ml(ovflw_stk->par_lock(), 3484 Mutex::_no_safepoint_check_flag); 3485 // Grab up to 1/4 the size of the work queue 3486 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3487 (size_t)ParGCDesiredObjsFromOverflowList); 3488 num = MIN2(num, ovflw_stk->length()); 3489 for (int i = (int) num; i > 0; i--) { 3490 oop cur = ovflw_stk->pop(); 3491 assert(cur != NULL, "Counted wrong?"); 3492 work_q->push(cur); 3493 } 3494 return num > 0; 3495 } 3496 3497 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3498 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3499 int n_tasks = pst->n_tasks(); 3500 // We allow that there may be no tasks to do here because 3501 // we are restarting after a stack overflow. 3502 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3503 uint nth_task = 0; 3504 3505 HeapWord* aligned_start = sp->bottom(); 3506 if (sp->used_region().contains(_restart_addr)) { 3507 // Align down to a card boundary for the start of 0th task 3508 // for this space. 3509 aligned_start = 3510 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3511 CardTableModRefBS::card_size); 3512 } 3513 3514 size_t chunk_size = sp->marking_task_size(); 3515 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3516 // Having claimed the nth task in this space, 3517 // compute the chunk that it corresponds to: 3518 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3519 aligned_start + (nth_task+1)*chunk_size); 3520 // Try and bump the global finger via a CAS; 3521 // note that we need to do the global finger bump 3522 // _before_ taking the intersection below, because 3523 // the task corresponding to that region will be 3524 // deemed done even if the used_region() expands 3525 // because of allocation -- as it almost certainly will 3526 // during start-up while the threads yield in the 3527 // closure below. 3528 HeapWord* finger = span.end(); 3529 bump_global_finger(finger); // atomically 3530 // There are null tasks here corresponding to chunks 3531 // beyond the "top" address of the space. 3532 span = span.intersection(sp->used_region()); 3533 if (!span.is_empty()) { // Non-null task 3534 HeapWord* prev_obj; 3535 assert(!span.contains(_restart_addr) || nth_task == 0, 3536 "Inconsistency"); 3537 if (nth_task == 0) { 3538 // For the 0th task, we'll not need to compute a block_start. 3539 if (span.contains(_restart_addr)) { 3540 // In the case of a restart because of stack overflow, 3541 // we might additionally skip a chunk prefix. 3542 prev_obj = _restart_addr; 3543 } else { 3544 prev_obj = span.start(); 3545 } 3546 } else { 3547 // We want to skip the first object because 3548 // the protocol is to scan any object in its entirety 3549 // that _starts_ in this span; a fortiori, any 3550 // object starting in an earlier span is scanned 3551 // as part of an earlier claimed task. 3552 // Below we use the "careful" version of block_start 3553 // so we do not try to navigate uninitialized objects. 3554 prev_obj = sp->block_start_careful(span.start()); 3555 // Below we use a variant of block_size that uses the 3556 // Printezis bits to avoid waiting for allocated 3557 // objects to become initialized/parsable. 3558 while (prev_obj < span.start()) { 3559 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3560 if (sz > 0) { 3561 prev_obj += sz; 3562 } else { 3563 // In this case we may end up doing a bit of redundant 3564 // scanning, but that appears unavoidable, short of 3565 // locking the free list locks; see bug 6324141. 3566 break; 3567 } 3568 } 3569 } 3570 if (prev_obj < span.end()) { 3571 MemRegion my_span = MemRegion(prev_obj, span.end()); 3572 // Do the marking work within a non-empty span -- 3573 // the last argument to the constructor indicates whether the 3574 // iteration should be incremental with periodic yields. 3575 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3576 &_collector->_markBitMap, 3577 work_queue(i), 3578 &_collector->_markStack); 3579 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3580 } // else nothing to do for this task 3581 } // else nothing to do for this task 3582 } 3583 // We'd be tempted to assert here that since there are no 3584 // more tasks left to claim in this space, the global_finger 3585 // must exceed space->top() and a fortiori space->end(). However, 3586 // that would not quite be correct because the bumping of 3587 // global_finger occurs strictly after the claiming of a task, 3588 // so by the time we reach here the global finger may not yet 3589 // have been bumped up by the thread that claimed the last 3590 // task. 3591 pst->all_tasks_completed(); 3592 } 3593 3594 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3595 private: 3596 CMSCollector* _collector; 3597 CMSConcMarkingTask* _task; 3598 MemRegion _span; 3599 CMSBitMap* _bit_map; 3600 CMSMarkStack* _overflow_stack; 3601 OopTaskQueue* _work_queue; 3602 protected: 3603 DO_OOP_WORK_DEFN 3604 public: 3605 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3606 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3607 MetadataAwareOopClosure(collector->ref_processor()), 3608 _collector(collector), 3609 _task(task), 3610 _span(collector->_span), 3611 _work_queue(work_queue), 3612 _bit_map(bit_map), 3613 _overflow_stack(overflow_stack) 3614 { } 3615 virtual void do_oop(oop* p); 3616 virtual void do_oop(narrowOop* p); 3617 3618 void trim_queue(size_t max); 3619 void handle_stack_overflow(HeapWord* lost); 3620 void do_yield_check() { 3621 if (_task->should_yield()) { 3622 _task->yield(); 3623 } 3624 } 3625 }; 3626 3627 // Grey object scanning during work stealing phase -- 3628 // the salient assumption here is that any references 3629 // that are in these stolen objects being scanned must 3630 // already have been initialized (else they would not have 3631 // been published), so we do not need to check for 3632 // uninitialized objects before pushing here. 3633 void Par_ConcMarkingClosure::do_oop(oop obj) { 3634 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3635 HeapWord* addr = (HeapWord*)obj; 3636 // Check if oop points into the CMS generation 3637 // and is not marked 3638 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3639 // a white object ... 3640 // If we manage to "claim" the object, by being the 3641 // first thread to mark it, then we push it on our 3642 // marking stack 3643 if (_bit_map->par_mark(addr)) { // ... now grey 3644 // push on work queue (grey set) 3645 bool simulate_overflow = false; 3646 NOT_PRODUCT( 3647 if (CMSMarkStackOverflowALot && 3648 _collector->simulate_overflow()) { 3649 // simulate a stack overflow 3650 simulate_overflow = true; 3651 } 3652 ) 3653 if (simulate_overflow || 3654 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3655 // stack overflow 3656 if (PrintCMSStatistics != 0) { 3657 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3658 SIZE_FORMAT, _overflow_stack->capacity()); 3659 } 3660 // We cannot assert that the overflow stack is full because 3661 // it may have been emptied since. 3662 assert(simulate_overflow || 3663 _work_queue->size() == _work_queue->max_elems(), 3664 "Else push should have succeeded"); 3665 handle_stack_overflow(addr); 3666 } 3667 } // Else, some other thread got there first 3668 do_yield_check(); 3669 } 3670 } 3671 3672 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3673 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3674 3675 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3676 while (_work_queue->size() > max) { 3677 oop new_oop; 3678 if (_work_queue->pop_local(new_oop)) { 3679 assert(new_oop->is_oop(), "Should be an oop"); 3680 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3681 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3682 new_oop->oop_iterate(this); // do_oop() above 3683 do_yield_check(); 3684 } 3685 } 3686 } 3687 3688 // Upon stack overflow, we discard (part of) the stack, 3689 // remembering the least address amongst those discarded 3690 // in CMSCollector's _restart_address. 3691 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3692 // We need to do this under a mutex to prevent other 3693 // workers from interfering with the work done below. 3694 MutexLockerEx ml(_overflow_stack->par_lock(), 3695 Mutex::_no_safepoint_check_flag); 3696 // Remember the least grey address discarded 3697 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3698 _collector->lower_restart_addr(ra); 3699 _overflow_stack->reset(); // discard stack contents 3700 _overflow_stack->expand(); // expand the stack if possible 3701 } 3702 3703 3704 void CMSConcMarkingTask::do_work_steal(int i) { 3705 OopTaskQueue* work_q = work_queue(i); 3706 oop obj_to_scan; 3707 CMSBitMap* bm = &(_collector->_markBitMap); 3708 CMSMarkStack* ovflw = &(_collector->_markStack); 3709 int* seed = _collector->hash_seed(i); 3710 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3711 while (true) { 3712 cl.trim_queue(0); 3713 assert(work_q->size() == 0, "Should have been emptied above"); 3714 if (get_work_from_overflow_stack(ovflw, work_q)) { 3715 // Can't assert below because the work obtained from the 3716 // overflow stack may already have been stolen from us. 3717 // assert(work_q->size() > 0, "Work from overflow stack"); 3718 continue; 3719 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3720 assert(obj_to_scan->is_oop(), "Should be an oop"); 3721 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3722 obj_to_scan->oop_iterate(&cl); 3723 } else if (terminator()->offer_termination(&_term_term)) { 3724 assert(work_q->size() == 0, "Impossible!"); 3725 break; 3726 } else if (yielding() || should_yield()) { 3727 yield(); 3728 } 3729 } 3730 } 3731 3732 // This is run by the CMS (coordinator) thread. 3733 void CMSConcMarkingTask::coordinator_yield() { 3734 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3735 "CMS thread should hold CMS token"); 3736 // First give up the locks, then yield, then re-lock 3737 // We should probably use a constructor/destructor idiom to 3738 // do this unlock/lock or modify the MutexUnlocker class to 3739 // serve our purpose. XXX 3740 assert_lock_strong(_bit_map_lock); 3741 _bit_map_lock->unlock(); 3742 ConcurrentMarkSweepThread::desynchronize(true); 3743 _collector->stopTimer(); 3744 if (PrintCMSStatistics != 0) { 3745 _collector->incrementYields(); 3746 } 3747 3748 // It is possible for whichever thread initiated the yield request 3749 // not to get a chance to wake up and take the bitmap lock between 3750 // this thread releasing it and reacquiring it. So, while the 3751 // should_yield() flag is on, let's sleep for a bit to give the 3752 // other thread a chance to wake up. The limit imposed on the number 3753 // of iterations is defensive, to avoid any unforseen circumstances 3754 // putting us into an infinite loop. Since it's always been this 3755 // (coordinator_yield()) method that was observed to cause the 3756 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3757 // which is by default non-zero. For the other seven methods that 3758 // also perform the yield operation, as are using a different 3759 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3760 // can enable the sleeping for those methods too, if necessary. 3761 // See 6442774. 3762 // 3763 // We really need to reconsider the synchronization between the GC 3764 // thread and the yield-requesting threads in the future and we 3765 // should really use wait/notify, which is the recommended 3766 // way of doing this type of interaction. Additionally, we should 3767 // consolidate the eight methods that do the yield operation and they 3768 // are almost identical into one for better maintainability and 3769 // readability. See 6445193. 3770 // 3771 // Tony 2006.06.29 3772 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3773 ConcurrentMarkSweepThread::should_yield() && 3774 !CMSCollector::foregroundGCIsActive(); ++i) { 3775 os::sleep(Thread::current(), 1, false); 3776 } 3777 3778 ConcurrentMarkSweepThread::synchronize(true); 3779 _bit_map_lock->lock_without_safepoint_check(); 3780 _collector->startTimer(); 3781 } 3782 3783 bool CMSCollector::do_marking_mt() { 3784 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3785 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 3786 conc_workers()->total_workers(), 3787 conc_workers()->active_workers(), 3788 Threads::number_of_non_daemon_threads()); 3789 conc_workers()->set_active_workers(num_workers); 3790 3791 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3792 3793 CMSConcMarkingTask tsk(this, 3794 cms_space, 3795 conc_workers(), 3796 task_queues()); 3797 3798 // Since the actual number of workers we get may be different 3799 // from the number we requested above, do we need to do anything different 3800 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3801 // class?? XXX 3802 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3803 3804 // Refs discovery is already non-atomic. 3805 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3806 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3807 conc_workers()->start_task(&tsk); 3808 while (tsk.yielded()) { 3809 tsk.coordinator_yield(); 3810 conc_workers()->continue_task(&tsk); 3811 } 3812 // If the task was aborted, _restart_addr will be non-NULL 3813 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3814 while (_restart_addr != NULL) { 3815 // XXX For now we do not make use of ABORTED state and have not 3816 // yet implemented the right abort semantics (even in the original 3817 // single-threaded CMS case). That needs some more investigation 3818 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3819 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3820 // If _restart_addr is non-NULL, a marking stack overflow 3821 // occurred; we need to do a fresh marking iteration from the 3822 // indicated restart address. 3823 if (_foregroundGCIsActive) { 3824 // We may be running into repeated stack overflows, having 3825 // reached the limit of the stack size, while making very 3826 // slow forward progress. It may be best to bail out and 3827 // let the foreground collector do its job. 3828 // Clear _restart_addr, so that foreground GC 3829 // works from scratch. This avoids the headache of 3830 // a "rescan" which would otherwise be needed because 3831 // of the dirty mod union table & card table. 3832 _restart_addr = NULL; 3833 return false; 3834 } 3835 // Adjust the task to restart from _restart_addr 3836 tsk.reset(_restart_addr); 3837 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3838 _restart_addr); 3839 _restart_addr = NULL; 3840 // Get the workers going again 3841 conc_workers()->start_task(&tsk); 3842 while (tsk.yielded()) { 3843 tsk.coordinator_yield(); 3844 conc_workers()->continue_task(&tsk); 3845 } 3846 } 3847 assert(tsk.completed(), "Inconsistency"); 3848 assert(tsk.result() == true, "Inconsistency"); 3849 return true; 3850 } 3851 3852 bool CMSCollector::do_marking_st() { 3853 ResourceMark rm; 3854 HandleMark hm; 3855 3856 // Temporarily make refs discovery single threaded (non-MT) 3857 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3858 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3859 &_markStack, CMSYield); 3860 // the last argument to iterate indicates whether the iteration 3861 // should be incremental with periodic yields. 3862 _markBitMap.iterate(&markFromRootsClosure); 3863 // If _restart_addr is non-NULL, a marking stack overflow 3864 // occurred; we need to do a fresh iteration from the 3865 // indicated restart address. 3866 while (_restart_addr != NULL) { 3867 if (_foregroundGCIsActive) { 3868 // We may be running into repeated stack overflows, having 3869 // reached the limit of the stack size, while making very 3870 // slow forward progress. It may be best to bail out and 3871 // let the foreground collector do its job. 3872 // Clear _restart_addr, so that foreground GC 3873 // works from scratch. This avoids the headache of 3874 // a "rescan" which would otherwise be needed because 3875 // of the dirty mod union table & card table. 3876 _restart_addr = NULL; 3877 return false; // indicating failure to complete marking 3878 } 3879 // Deal with stack overflow: 3880 // we restart marking from _restart_addr 3881 HeapWord* ra = _restart_addr; 3882 markFromRootsClosure.reset(ra); 3883 _restart_addr = NULL; 3884 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3885 } 3886 return true; 3887 } 3888 3889 void CMSCollector::preclean() { 3890 check_correct_thread_executing(); 3891 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3892 verify_work_stacks_empty(); 3893 verify_overflow_empty(); 3894 _abort_preclean = false; 3895 if (CMSPrecleaningEnabled) { 3896 if (!CMSEdenChunksRecordAlways) { 3897 _eden_chunk_index = 0; 3898 } 3899 size_t used = get_eden_used(); 3900 size_t capacity = get_eden_capacity(); 3901 // Don't start sampling unless we will get sufficiently 3902 // many samples. 3903 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3904 * CMSScheduleRemarkEdenPenetration)) { 3905 _start_sampling = true; 3906 } else { 3907 _start_sampling = false; 3908 } 3909 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3910 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3911 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3912 } 3913 CMSTokenSync x(true); // is cms thread 3914 if (CMSPrecleaningEnabled) { 3915 sample_eden(); 3916 _collectorState = AbortablePreclean; 3917 } else { 3918 _collectorState = FinalMarking; 3919 } 3920 verify_work_stacks_empty(); 3921 verify_overflow_empty(); 3922 } 3923 3924 // Try and schedule the remark such that young gen 3925 // occupancy is CMSScheduleRemarkEdenPenetration %. 3926 void CMSCollector::abortable_preclean() { 3927 check_correct_thread_executing(); 3928 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3929 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3930 3931 // If Eden's current occupancy is below this threshold, 3932 // immediately schedule the remark; else preclean 3933 // past the next scavenge in an effort to 3934 // schedule the pause as described above. By choosing 3935 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3936 // we will never do an actual abortable preclean cycle. 3937 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3938 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3939 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3940 // We need more smarts in the abortable preclean 3941 // loop below to deal with cases where allocation 3942 // in young gen is very very slow, and our precleaning 3943 // is running a losing race against a horde of 3944 // mutators intent on flooding us with CMS updates 3945 // (dirty cards). 3946 // One, admittedly dumb, strategy is to give up 3947 // after a certain number of abortable precleaning loops 3948 // or after a certain maximum time. We want to make 3949 // this smarter in the next iteration. 3950 // XXX FIX ME!!! YSR 3951 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3952 while (!(should_abort_preclean() || 3953 ConcurrentMarkSweepThread::should_terminate())) { 3954 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3955 cumworkdone += workdone; 3956 loops++; 3957 // Voluntarily terminate abortable preclean phase if we have 3958 // been at it for too long. 3959 if ((CMSMaxAbortablePrecleanLoops != 0) && 3960 loops >= CMSMaxAbortablePrecleanLoops) { 3961 if (PrintGCDetails) { 3962 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3963 } 3964 break; 3965 } 3966 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3967 if (PrintGCDetails) { 3968 gclog_or_tty->print(" CMS: abort preclean due to time "); 3969 } 3970 break; 3971 } 3972 // If we are doing little work each iteration, we should 3973 // take a short break. 3974 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3975 // Sleep for some time, waiting for work to accumulate 3976 stopTimer(); 3977 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3978 startTimer(); 3979 waited++; 3980 } 3981 } 3982 if (PrintCMSStatistics > 0) { 3983 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3984 loops, waited, cumworkdone); 3985 } 3986 } 3987 CMSTokenSync x(true); // is cms thread 3988 if (_collectorState != Idling) { 3989 assert(_collectorState == AbortablePreclean, 3990 "Spontaneous state transition?"); 3991 _collectorState = FinalMarking; 3992 } // Else, a foreground collection completed this CMS cycle. 3993 return; 3994 } 3995 3996 // Respond to an Eden sampling opportunity 3997 void CMSCollector::sample_eden() { 3998 // Make sure a young gc cannot sneak in between our 3999 // reading and recording of a sample. 4000 assert(Thread::current()->is_ConcurrentGC_thread(), 4001 "Only the cms thread may collect Eden samples"); 4002 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4003 "Should collect samples while holding CMS token"); 4004 if (!_start_sampling) { 4005 return; 4006 } 4007 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4008 // is populated by the young generation. 4009 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4010 if (_eden_chunk_index < _eden_chunk_capacity) { 4011 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4012 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4013 "Unexpected state of Eden"); 4014 // We'd like to check that what we just sampled is an oop-start address; 4015 // however, we cannot do that here since the object may not yet have been 4016 // initialized. So we'll instead do the check when we _use_ this sample 4017 // later. 4018 if (_eden_chunk_index == 0 || 4019 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4020 _eden_chunk_array[_eden_chunk_index-1]) 4021 >= CMSSamplingGrain)) { 4022 _eden_chunk_index++; // commit sample 4023 } 4024 } 4025 } 4026 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4027 size_t used = get_eden_used(); 4028 size_t capacity = get_eden_capacity(); 4029 assert(used <= capacity, "Unexpected state of Eden"); 4030 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4031 _abort_preclean = true; 4032 } 4033 } 4034 } 4035 4036 4037 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4038 assert(_collectorState == Precleaning || 4039 _collectorState == AbortablePreclean, "incorrect state"); 4040 ResourceMark rm; 4041 HandleMark hm; 4042 4043 // Precleaning is currently not MT but the reference processor 4044 // may be set for MT. Disable it temporarily here. 4045 ReferenceProcessor* rp = ref_processor(); 4046 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4047 4048 // Do one pass of scrubbing the discovered reference lists 4049 // to remove any reference objects with strongly-reachable 4050 // referents. 4051 if (clean_refs) { 4052 CMSPrecleanRefsYieldClosure yield_cl(this); 4053 assert(rp->span().equals(_span), "Spans should be equal"); 4054 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4055 &_markStack, true /* preclean */); 4056 CMSDrainMarkingStackClosure complete_trace(this, 4057 _span, &_markBitMap, &_markStack, 4058 &keep_alive, true /* preclean */); 4059 4060 // We don't want this step to interfere with a young 4061 // collection because we don't want to take CPU 4062 // or memory bandwidth away from the young GC threads 4063 // (which may be as many as there are CPUs). 4064 // Note that we don't need to protect ourselves from 4065 // interference with mutators because they can't 4066 // manipulate the discovered reference lists nor affect 4067 // the computed reachability of the referents, the 4068 // only properties manipulated by the precleaning 4069 // of these reference lists. 4070 stopTimer(); 4071 CMSTokenSyncWithLocks x(true /* is cms thread */, 4072 bitMapLock()); 4073 startTimer(); 4074 sample_eden(); 4075 4076 // The following will yield to allow foreground 4077 // collection to proceed promptly. XXX YSR: 4078 // The code in this method may need further 4079 // tweaking for better performance and some restructuring 4080 // for cleaner interfaces. 4081 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4082 rp->preclean_discovered_references( 4083 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4084 gc_timer, _gc_tracer_cm->gc_id()); 4085 } 4086 4087 if (clean_survivor) { // preclean the active survivor space(s) 4088 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4089 &_markBitMap, &_modUnionTable, 4090 &_markStack, true /* precleaning phase */); 4091 stopTimer(); 4092 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4093 bitMapLock()); 4094 startTimer(); 4095 unsigned int before_count = 4096 GenCollectedHeap::heap()->total_collections(); 4097 SurvivorSpacePrecleanClosure 4098 sss_cl(this, _span, &_markBitMap, &_markStack, 4099 &pam_cl, before_count, CMSYield); 4100 _young_gen->from()->object_iterate_careful(&sss_cl); 4101 _young_gen->to()->object_iterate_careful(&sss_cl); 4102 } 4103 MarkRefsIntoAndScanClosure 4104 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4105 &_markStack, this, CMSYield, 4106 true /* precleaning phase */); 4107 // CAUTION: The following closure has persistent state that may need to 4108 // be reset upon a decrease in the sequence of addresses it 4109 // processes. 4110 ScanMarkedObjectsAgainCarefullyClosure 4111 smoac_cl(this, _span, 4112 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4113 4114 // Preclean dirty cards in ModUnionTable and CardTable using 4115 // appropriate convergence criterion; 4116 // repeat CMSPrecleanIter times unless we find that 4117 // we are losing. 4118 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4119 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4120 "Bad convergence multiplier"); 4121 assert(CMSPrecleanThreshold >= 100, 4122 "Unreasonably low CMSPrecleanThreshold"); 4123 4124 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4125 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4126 numIter < CMSPrecleanIter; 4127 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4128 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4129 if (Verbose && PrintGCDetails) { 4130 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4131 } 4132 // Either there are very few dirty cards, so re-mark 4133 // pause will be small anyway, or our pre-cleaning isn't 4134 // that much faster than the rate at which cards are being 4135 // dirtied, so we might as well stop and re-mark since 4136 // precleaning won't improve our re-mark time by much. 4137 if (curNumCards <= CMSPrecleanThreshold || 4138 (numIter > 0 && 4139 (curNumCards * CMSPrecleanDenominator > 4140 lastNumCards * CMSPrecleanNumerator))) { 4141 numIter++; 4142 cumNumCards += curNumCards; 4143 break; 4144 } 4145 } 4146 4147 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4148 4149 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4150 cumNumCards += curNumCards; 4151 if (PrintGCDetails && PrintCMSStatistics != 0) { 4152 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4153 curNumCards, cumNumCards, numIter); 4154 } 4155 return cumNumCards; // as a measure of useful work done 4156 } 4157 4158 // PRECLEANING NOTES: 4159 // Precleaning involves: 4160 // . reading the bits of the modUnionTable and clearing the set bits. 4161 // . For the cards corresponding to the set bits, we scan the 4162 // objects on those cards. This means we need the free_list_lock 4163 // so that we can safely iterate over the CMS space when scanning 4164 // for oops. 4165 // . When we scan the objects, we'll be both reading and setting 4166 // marks in the marking bit map, so we'll need the marking bit map. 4167 // . For protecting _collector_state transitions, we take the CGC_lock. 4168 // Note that any races in the reading of of card table entries by the 4169 // CMS thread on the one hand and the clearing of those entries by the 4170 // VM thread or the setting of those entries by the mutator threads on the 4171 // other are quite benign. However, for efficiency it makes sense to keep 4172 // the VM thread from racing with the CMS thread while the latter is 4173 // dirty card info to the modUnionTable. We therefore also use the 4174 // CGC_lock to protect the reading of the card table and the mod union 4175 // table by the CM thread. 4176 // . We run concurrently with mutator updates, so scanning 4177 // needs to be done carefully -- we should not try to scan 4178 // potentially uninitialized objects. 4179 // 4180 // Locking strategy: While holding the CGC_lock, we scan over and 4181 // reset a maximal dirty range of the mod union / card tables, then lock 4182 // the free_list_lock and bitmap lock to do a full marking, then 4183 // release these locks; and repeat the cycle. This allows for a 4184 // certain amount of fairness in the sharing of these locks between 4185 // the CMS collector on the one hand, and the VM thread and the 4186 // mutators on the other. 4187 4188 // NOTE: preclean_mod_union_table() and preclean_card_table() 4189 // further below are largely identical; if you need to modify 4190 // one of these methods, please check the other method too. 4191 4192 size_t CMSCollector::preclean_mod_union_table( 4193 ConcurrentMarkSweepGeneration* gen, 4194 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4195 verify_work_stacks_empty(); 4196 verify_overflow_empty(); 4197 4198 // strategy: starting with the first card, accumulate contiguous 4199 // ranges of dirty cards; clear these cards, then scan the region 4200 // covered by these cards. 4201 4202 // Since all of the MUT is committed ahead, we can just use 4203 // that, in case the generations expand while we are precleaning. 4204 // It might also be fine to just use the committed part of the 4205 // generation, but we might potentially miss cards when the 4206 // generation is rapidly expanding while we are in the midst 4207 // of precleaning. 4208 HeapWord* startAddr = gen->reserved().start(); 4209 HeapWord* endAddr = gen->reserved().end(); 4210 4211 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4212 4213 size_t numDirtyCards, cumNumDirtyCards; 4214 HeapWord *nextAddr, *lastAddr; 4215 for (cumNumDirtyCards = numDirtyCards = 0, 4216 nextAddr = lastAddr = startAddr; 4217 nextAddr < endAddr; 4218 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4219 4220 ResourceMark rm; 4221 HandleMark hm; 4222 4223 MemRegion dirtyRegion; 4224 { 4225 stopTimer(); 4226 // Potential yield point 4227 CMSTokenSync ts(true); 4228 startTimer(); 4229 sample_eden(); 4230 // Get dirty region starting at nextOffset (inclusive), 4231 // simultaneously clearing it. 4232 dirtyRegion = 4233 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4234 assert(dirtyRegion.start() >= nextAddr, 4235 "returned region inconsistent?"); 4236 } 4237 // Remember where the next search should begin. 4238 // The returned region (if non-empty) is a right open interval, 4239 // so lastOffset is obtained from the right end of that 4240 // interval. 4241 lastAddr = dirtyRegion.end(); 4242 // Should do something more transparent and less hacky XXX 4243 numDirtyCards = 4244 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4245 4246 // We'll scan the cards in the dirty region (with periodic 4247 // yields for foreground GC as needed). 4248 if (!dirtyRegion.is_empty()) { 4249 assert(numDirtyCards > 0, "consistency check"); 4250 HeapWord* stop_point = NULL; 4251 stopTimer(); 4252 // Potential yield point 4253 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4254 bitMapLock()); 4255 startTimer(); 4256 { 4257 verify_work_stacks_empty(); 4258 verify_overflow_empty(); 4259 sample_eden(); 4260 stop_point = 4261 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4262 } 4263 if (stop_point != NULL) { 4264 // The careful iteration stopped early either because it found an 4265 // uninitialized object, or because we were in the midst of an 4266 // "abortable preclean", which should now be aborted. Redirty 4267 // the bits corresponding to the partially-scanned or unscanned 4268 // cards. We'll either restart at the next block boundary or 4269 // abort the preclean. 4270 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4271 "Should only be AbortablePreclean."); 4272 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4273 if (should_abort_preclean()) { 4274 break; // out of preclean loop 4275 } else { 4276 // Compute the next address at which preclean should pick up; 4277 // might need bitMapLock in order to read P-bits. 4278 lastAddr = next_card_start_after_block(stop_point); 4279 } 4280 } 4281 } else { 4282 assert(lastAddr == endAddr, "consistency check"); 4283 assert(numDirtyCards == 0, "consistency check"); 4284 break; 4285 } 4286 } 4287 verify_work_stacks_empty(); 4288 verify_overflow_empty(); 4289 return cumNumDirtyCards; 4290 } 4291 4292 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4293 // below are largely identical; if you need to modify 4294 // one of these methods, please check the other method too. 4295 4296 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4297 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4298 // strategy: it's similar to precleamModUnionTable above, in that 4299 // we accumulate contiguous ranges of dirty cards, mark these cards 4300 // precleaned, then scan the region covered by these cards. 4301 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4302 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4303 4304 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4305 4306 size_t numDirtyCards, cumNumDirtyCards; 4307 HeapWord *lastAddr, *nextAddr; 4308 4309 for (cumNumDirtyCards = numDirtyCards = 0, 4310 nextAddr = lastAddr = startAddr; 4311 nextAddr < endAddr; 4312 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4313 4314 ResourceMark rm; 4315 HandleMark hm; 4316 4317 MemRegion dirtyRegion; 4318 { 4319 // See comments in "Precleaning notes" above on why we 4320 // do this locking. XXX Could the locking overheads be 4321 // too high when dirty cards are sparse? [I don't think so.] 4322 stopTimer(); 4323 CMSTokenSync x(true); // is cms thread 4324 startTimer(); 4325 sample_eden(); 4326 // Get and clear dirty region from card table 4327 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4328 MemRegion(nextAddr, endAddr), 4329 true, 4330 CardTableModRefBS::precleaned_card_val()); 4331 4332 assert(dirtyRegion.start() >= nextAddr, 4333 "returned region inconsistent?"); 4334 } 4335 lastAddr = dirtyRegion.end(); 4336 numDirtyCards = 4337 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4338 4339 if (!dirtyRegion.is_empty()) { 4340 stopTimer(); 4341 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4342 startTimer(); 4343 sample_eden(); 4344 verify_work_stacks_empty(); 4345 verify_overflow_empty(); 4346 HeapWord* stop_point = 4347 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4348 if (stop_point != NULL) { 4349 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4350 "Should only be AbortablePreclean."); 4351 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4352 if (should_abort_preclean()) { 4353 break; // out of preclean loop 4354 } else { 4355 // Compute the next address at which preclean should pick up. 4356 lastAddr = next_card_start_after_block(stop_point); 4357 } 4358 } 4359 } else { 4360 break; 4361 } 4362 } 4363 verify_work_stacks_empty(); 4364 verify_overflow_empty(); 4365 return cumNumDirtyCards; 4366 } 4367 4368 class PrecleanKlassClosure : public KlassClosure { 4369 KlassToOopClosure _cm_klass_closure; 4370 public: 4371 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4372 void do_klass(Klass* k) { 4373 if (k->has_accumulated_modified_oops()) { 4374 k->clear_accumulated_modified_oops(); 4375 4376 _cm_klass_closure.do_klass(k); 4377 } 4378 } 4379 }; 4380 4381 // The freelist lock is needed to prevent asserts, is it really needed? 4382 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4383 4384 cl->set_freelistLock(freelistLock); 4385 4386 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4387 4388 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4389 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4390 PrecleanKlassClosure preclean_klass_closure(cl); 4391 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4392 4393 verify_work_stacks_empty(); 4394 verify_overflow_empty(); 4395 } 4396 4397 void CMSCollector::checkpointRootsFinal() { 4398 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4399 check_correct_thread_executing(); 4400 // world is stopped at this checkpoint 4401 assert(SafepointSynchronize::is_at_safepoint(), 4402 "world should be stopped"); 4403 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4404 4405 verify_work_stacks_empty(); 4406 verify_overflow_empty(); 4407 4408 SpecializationStats::clear(); 4409 if (PrintGCDetails) { 4410 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4411 _young_gen->used() / K, 4412 _young_gen->capacity() / K); 4413 } 4414 { 4415 if (CMSScavengeBeforeRemark) { 4416 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4417 // Temporarily set flag to false, GCH->do_collection will 4418 // expect it to be false and set to true 4419 FlagSetting fl(gch->_is_gc_active, false); 4420 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4421 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4422 int level = _cmsGen->level() - 1; 4423 if (level >= 0) { 4424 gch->do_collection(true, // full (i.e. force, see below) 4425 false, // !clear_all_soft_refs 4426 0, // size 4427 false, // is_tlab 4428 level // max_level 4429 ); 4430 } 4431 } 4432 FreelistLocker x(this); 4433 MutexLockerEx y(bitMapLock(), 4434 Mutex::_no_safepoint_check_flag); 4435 checkpointRootsFinalWork(); 4436 } 4437 verify_work_stacks_empty(); 4438 verify_overflow_empty(); 4439 SpecializationStats::print(); 4440 } 4441 4442 void CMSCollector::checkpointRootsFinalWork() { 4443 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4444 4445 assert(haveFreelistLocks(), "must have free list locks"); 4446 assert_lock_strong(bitMapLock()); 4447 4448 ResourceMark rm; 4449 HandleMark hm; 4450 4451 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4452 4453 if (should_unload_classes()) { 4454 CodeCache::gc_prologue(); 4455 } 4456 assert(haveFreelistLocks(), "must have free list locks"); 4457 assert_lock_strong(bitMapLock()); 4458 4459 // We might assume that we need not fill TLAB's when 4460 // CMSScavengeBeforeRemark is set, because we may have just done 4461 // a scavenge which would have filled all TLAB's -- and besides 4462 // Eden would be empty. This however may not always be the case -- 4463 // for instance although we asked for a scavenge, it may not have 4464 // happened because of a JNI critical section. We probably need 4465 // a policy for deciding whether we can in that case wait until 4466 // the critical section releases and then do the remark following 4467 // the scavenge, and skip it here. In the absence of that policy, 4468 // or of an indication of whether the scavenge did indeed occur, 4469 // we cannot rely on TLAB's having been filled and must do 4470 // so here just in case a scavenge did not happen. 4471 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4472 // Update the saved marks which may affect the root scans. 4473 gch->save_marks(); 4474 4475 if (CMSPrintEdenSurvivorChunks) { 4476 print_eden_and_survivor_chunk_arrays(); 4477 } 4478 4479 { 4480 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4481 4482 // Note on the role of the mod union table: 4483 // Since the marker in "markFromRoots" marks concurrently with 4484 // mutators, it is possible for some reachable objects not to have been 4485 // scanned. For instance, an only reference to an object A was 4486 // placed in object B after the marker scanned B. Unless B is rescanned, 4487 // A would be collected. Such updates to references in marked objects 4488 // are detected via the mod union table which is the set of all cards 4489 // dirtied since the first checkpoint in this GC cycle and prior to 4490 // the most recent young generation GC, minus those cleaned up by the 4491 // concurrent precleaning. 4492 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4493 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4494 do_remark_parallel(); 4495 } else { 4496 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4497 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4498 do_remark_non_parallel(); 4499 } 4500 } 4501 verify_work_stacks_empty(); 4502 verify_overflow_empty(); 4503 4504 { 4505 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4506 refProcessingWork(); 4507 } 4508 verify_work_stacks_empty(); 4509 verify_overflow_empty(); 4510 4511 if (should_unload_classes()) { 4512 CodeCache::gc_epilogue(); 4513 } 4514 JvmtiExport::gc_epilogue(); 4515 4516 // If we encountered any (marking stack / work queue) overflow 4517 // events during the current CMS cycle, take appropriate 4518 // remedial measures, where possible, so as to try and avoid 4519 // recurrence of that condition. 4520 assert(_markStack.isEmpty(), "No grey objects"); 4521 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4522 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4523 if (ser_ovflw > 0) { 4524 if (PrintCMSStatistics != 0) { 4525 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4526 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4527 ", kac_preclean="SIZE_FORMAT")", 4528 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4529 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4530 } 4531 _markStack.expand(); 4532 _ser_pmc_remark_ovflw = 0; 4533 _ser_pmc_preclean_ovflw = 0; 4534 _ser_kac_preclean_ovflw = 0; 4535 _ser_kac_ovflw = 0; 4536 } 4537 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4538 if (PrintCMSStatistics != 0) { 4539 gclog_or_tty->print_cr("Work queue overflow (benign) " 4540 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4541 _par_pmc_remark_ovflw, _par_kac_ovflw); 4542 } 4543 _par_pmc_remark_ovflw = 0; 4544 _par_kac_ovflw = 0; 4545 } 4546 if (PrintCMSStatistics != 0) { 4547 if (_markStack._hit_limit > 0) { 4548 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4549 _markStack._hit_limit); 4550 } 4551 if (_markStack._failed_double > 0) { 4552 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4553 " current capacity "SIZE_FORMAT, 4554 _markStack._failed_double, 4555 _markStack.capacity()); 4556 } 4557 } 4558 _markStack._hit_limit = 0; 4559 _markStack._failed_double = 0; 4560 4561 if ((VerifyAfterGC || VerifyDuringGC) && 4562 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4563 verify_after_remark(); 4564 } 4565 4566 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4567 4568 // Change under the freelistLocks. 4569 _collectorState = Sweeping; 4570 // Call isAllClear() under bitMapLock 4571 assert(_modUnionTable.isAllClear(), 4572 "Should be clear by end of the final marking"); 4573 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4574 "Should be clear by end of the final marking"); 4575 } 4576 4577 void CMSParInitialMarkTask::work(uint worker_id) { 4578 elapsedTimer _timer; 4579 ResourceMark rm; 4580 HandleMark hm; 4581 4582 // ---------- scan from roots -------------- 4583 _timer.start(); 4584 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4585 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4586 4587 // ---------- young gen roots -------------- 4588 { 4589 work_on_young_gen_roots(worker_id, &par_mri_cl); 4590 _timer.stop(); 4591 if (PrintCMSStatistics != 0) { 4592 gclog_or_tty->print_cr( 4593 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4594 worker_id, _timer.seconds()); 4595 } 4596 } 4597 4598 // ---------- remaining roots -------------- 4599 _timer.reset(); 4600 _timer.start(); 4601 4602 CLDToOopClosure cld_closure(&par_mri_cl, true); 4603 4604 gch->gen_process_roots(_collector->_cmsGen->level(), 4605 false, // yg was scanned above 4606 false, // this is parallel code 4607 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4608 _collector->should_unload_classes(), 4609 &par_mri_cl, 4610 NULL, 4611 &cld_closure); 4612 assert(_collector->should_unload_classes() 4613 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 4614 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4615 _timer.stop(); 4616 if (PrintCMSStatistics != 0) { 4617 gclog_or_tty->print_cr( 4618 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4619 worker_id, _timer.seconds()); 4620 } 4621 } 4622 4623 // Parallel remark task 4624 class CMSParRemarkTask: public CMSParMarkTask { 4625 CompactibleFreeListSpace* _cms_space; 4626 4627 // The per-thread work queues, available here for stealing. 4628 OopTaskQueueSet* _task_queues; 4629 ParallelTaskTerminator _term; 4630 4631 public: 4632 // A value of 0 passed to n_workers will cause the number of 4633 // workers to be taken from the active workers in the work gang. 4634 CMSParRemarkTask(CMSCollector* collector, 4635 CompactibleFreeListSpace* cms_space, 4636 int n_workers, FlexibleWorkGang* workers, 4637 OopTaskQueueSet* task_queues): 4638 CMSParMarkTask("Rescan roots and grey objects in parallel", 4639 collector, n_workers), 4640 _cms_space(cms_space), 4641 _task_queues(task_queues), 4642 _term(n_workers, task_queues) { } 4643 4644 OopTaskQueueSet* task_queues() { return _task_queues; } 4645 4646 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4647 4648 ParallelTaskTerminator* terminator() { return &_term; } 4649 int n_workers() { return _n_workers; } 4650 4651 void work(uint worker_id); 4652 4653 private: 4654 // ... of dirty cards in old space 4655 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4656 Par_MarkRefsIntoAndScanClosure* cl); 4657 4658 // ... work stealing for the above 4659 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4660 }; 4661 4662 class RemarkKlassClosure : public KlassClosure { 4663 KlassToOopClosure _cm_klass_closure; 4664 public: 4665 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4666 void do_klass(Klass* k) { 4667 // Check if we have modified any oops in the Klass during the concurrent marking. 4668 if (k->has_accumulated_modified_oops()) { 4669 k->clear_accumulated_modified_oops(); 4670 4671 // We could have transfered the current modified marks to the accumulated marks, 4672 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4673 } else if (k->has_modified_oops()) { 4674 // Don't clear anything, this info is needed by the next young collection. 4675 } else { 4676 // No modified oops in the Klass. 4677 return; 4678 } 4679 4680 // The klass has modified fields, need to scan the klass. 4681 _cm_klass_closure.do_klass(k); 4682 } 4683 }; 4684 4685 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4686 ParNewGeneration* young_gen = _collector->_young_gen; 4687 ContiguousSpace* eden_space = young_gen->eden(); 4688 ContiguousSpace* from_space = young_gen->from(); 4689 ContiguousSpace* to_space = young_gen->to(); 4690 4691 HeapWord** eca = _collector->_eden_chunk_array; 4692 size_t ect = _collector->_eden_chunk_index; 4693 HeapWord** sca = _collector->_survivor_chunk_array; 4694 size_t sct = _collector->_survivor_chunk_index; 4695 4696 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4697 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4698 4699 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4700 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4701 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4702 } 4703 4704 // work_queue(i) is passed to the closure 4705 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4706 // also is passed to do_dirty_card_rescan_tasks() and to 4707 // do_work_steal() to select the i-th task_queue. 4708 4709 void CMSParRemarkTask::work(uint worker_id) { 4710 elapsedTimer _timer; 4711 ResourceMark rm; 4712 HandleMark hm; 4713 4714 // ---------- rescan from roots -------------- 4715 _timer.start(); 4716 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4717 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4718 _collector->_span, _collector->ref_processor(), 4719 &(_collector->_markBitMap), 4720 work_queue(worker_id)); 4721 4722 // Rescan young gen roots first since these are likely 4723 // coarsely partitioned and may, on that account, constitute 4724 // the critical path; thus, it's best to start off that 4725 // work first. 4726 // ---------- young gen roots -------------- 4727 { 4728 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4729 _timer.stop(); 4730 if (PrintCMSStatistics != 0) { 4731 gclog_or_tty->print_cr( 4732 "Finished young gen rescan work in %dth thread: %3.3f sec", 4733 worker_id, _timer.seconds()); 4734 } 4735 } 4736 4737 // ---------- remaining roots -------------- 4738 _timer.reset(); 4739 _timer.start(); 4740 gch->gen_process_roots(_collector->_cmsGen->level(), 4741 false, // yg was scanned above 4742 false, // this is parallel code 4743 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4744 _collector->should_unload_classes(), 4745 &par_mrias_cl, 4746 NULL, 4747 NULL); // The dirty klasses will be handled below 4748 4749 assert(_collector->should_unload_classes() 4750 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 4751 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4752 _timer.stop(); 4753 if (PrintCMSStatistics != 0) { 4754 gclog_or_tty->print_cr( 4755 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4756 worker_id, _timer.seconds()); 4757 } 4758 4759 // ---------- unhandled CLD scanning ---------- 4760 if (worker_id == 0) { // Single threaded at the moment. 4761 _timer.reset(); 4762 _timer.start(); 4763 4764 // Scan all new class loader data objects and new dependencies that were 4765 // introduced during concurrent marking. 4766 ResourceMark rm; 4767 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4768 for (int i = 0; i < array->length(); i++) { 4769 par_mrias_cl.do_class_loader_data(array->at(i)); 4770 } 4771 4772 // We don't need to keep track of new CLDs anymore. 4773 ClassLoaderDataGraph::remember_new_clds(false); 4774 4775 _timer.stop(); 4776 if (PrintCMSStatistics != 0) { 4777 gclog_or_tty->print_cr( 4778 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4779 worker_id, _timer.seconds()); 4780 } 4781 } 4782 4783 // ---------- dirty klass scanning ---------- 4784 if (worker_id == 0) { // Single threaded at the moment. 4785 _timer.reset(); 4786 _timer.start(); 4787 4788 // Scan all classes that was dirtied during the concurrent marking phase. 4789 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4790 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4791 4792 _timer.stop(); 4793 if (PrintCMSStatistics != 0) { 4794 gclog_or_tty->print_cr( 4795 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4796 worker_id, _timer.seconds()); 4797 } 4798 } 4799 4800 // We might have added oops to ClassLoaderData::_handles during the 4801 // concurrent marking phase. These oops point to newly allocated objects 4802 // that are guaranteed to be kept alive. Either by the direct allocation 4803 // code, or when the young collector processes the roots. Hence, 4804 // we don't have to revisit the _handles block during the remark phase. 4805 4806 // ---------- rescan dirty cards ------------ 4807 _timer.reset(); 4808 _timer.start(); 4809 4810 // Do the rescan tasks for each of the two spaces 4811 // (cms_space) in turn. 4812 // "worker_id" is passed to select the task_queue for "worker_id" 4813 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4814 _timer.stop(); 4815 if (PrintCMSStatistics != 0) { 4816 gclog_or_tty->print_cr( 4817 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4818 worker_id, _timer.seconds()); 4819 } 4820 4821 // ---------- steal work from other threads ... 4822 // ---------- ... and drain overflow list. 4823 _timer.reset(); 4824 _timer.start(); 4825 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4826 _timer.stop(); 4827 if (PrintCMSStatistics != 0) { 4828 gclog_or_tty->print_cr( 4829 "Finished work stealing in %dth thread: %3.3f sec", 4830 worker_id, _timer.seconds()); 4831 } 4832 } 4833 4834 // Note that parameter "i" is not used. 4835 void 4836 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4837 OopsInGenClosure* cl, ContiguousSpace* space, 4838 HeapWord** chunk_array, size_t chunk_top) { 4839 // Until all tasks completed: 4840 // . claim an unclaimed task 4841 // . compute region boundaries corresponding to task claimed 4842 // using chunk_array 4843 // . par_oop_iterate(cl) over that region 4844 4845 ResourceMark rm; 4846 HandleMark hm; 4847 4848 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4849 4850 uint nth_task = 0; 4851 uint n_tasks = pst->n_tasks(); 4852 4853 if (n_tasks > 0) { 4854 assert(pst->valid(), "Uninitialized use?"); 4855 HeapWord *start, *end; 4856 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4857 // We claimed task # nth_task; compute its boundaries. 4858 if (chunk_top == 0) { // no samples were taken 4859 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4860 start = space->bottom(); 4861 end = space->top(); 4862 } else if (nth_task == 0) { 4863 start = space->bottom(); 4864 end = chunk_array[nth_task]; 4865 } else if (nth_task < (uint)chunk_top) { 4866 assert(nth_task >= 1, "Control point invariant"); 4867 start = chunk_array[nth_task - 1]; 4868 end = chunk_array[nth_task]; 4869 } else { 4870 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4871 start = chunk_array[chunk_top - 1]; 4872 end = space->top(); 4873 } 4874 MemRegion mr(start, end); 4875 // Verify that mr is in space 4876 assert(mr.is_empty() || space->used_region().contains(mr), 4877 "Should be in space"); 4878 // Verify that "start" is an object boundary 4879 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4880 "Should be an oop"); 4881 space->par_oop_iterate(mr, cl); 4882 } 4883 pst->all_tasks_completed(); 4884 } 4885 } 4886 4887 void 4888 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4889 CompactibleFreeListSpace* sp, int i, 4890 Par_MarkRefsIntoAndScanClosure* cl) { 4891 // Until all tasks completed: 4892 // . claim an unclaimed task 4893 // . compute region boundaries corresponding to task claimed 4894 // . transfer dirty bits ct->mut for that region 4895 // . apply rescanclosure to dirty mut bits for that region 4896 4897 ResourceMark rm; 4898 HandleMark hm; 4899 4900 OopTaskQueue* work_q = work_queue(i); 4901 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4902 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4903 // CAUTION: This closure has state that persists across calls to 4904 // the work method dirty_range_iterate_clear() in that it has 4905 // embedded in it a (subtype of) UpwardsObjectClosure. The 4906 // use of that state in the embedded UpwardsObjectClosure instance 4907 // assumes that the cards are always iterated (even if in parallel 4908 // by several threads) in monotonically increasing order per each 4909 // thread. This is true of the implementation below which picks 4910 // card ranges (chunks) in monotonically increasing order globally 4911 // and, a-fortiori, in monotonically increasing order per thread 4912 // (the latter order being a subsequence of the former). 4913 // If the work code below is ever reorganized into a more chaotic 4914 // work-partitioning form than the current "sequential tasks" 4915 // paradigm, the use of that persistent state will have to be 4916 // revisited and modified appropriately. See also related 4917 // bug 4756801 work on which should examine this code to make 4918 // sure that the changes there do not run counter to the 4919 // assumptions made here and necessary for correctness and 4920 // efficiency. Note also that this code might yield inefficient 4921 // behavior in the case of very large objects that span one or 4922 // more work chunks. Such objects would potentially be scanned 4923 // several times redundantly. Work on 4756801 should try and 4924 // address that performance anomaly if at all possible. XXX 4925 MemRegion full_span = _collector->_span; 4926 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4927 MarkFromDirtyCardsClosure 4928 greyRescanClosure(_collector, full_span, // entire span of interest 4929 sp, bm, work_q, cl); 4930 4931 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4932 assert(pst->valid(), "Uninitialized use?"); 4933 uint nth_task = 0; 4934 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4935 MemRegion span = sp->used_region(); 4936 HeapWord* start_addr = span.start(); 4937 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4938 alignment); 4939 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4940 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4941 start_addr, "Check alignment"); 4942 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4943 chunk_size, "Check alignment"); 4944 4945 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4946 // Having claimed the nth_task, compute corresponding mem-region, 4947 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4948 // The alignment restriction ensures that we do not need any 4949 // synchronization with other gang-workers while setting or 4950 // clearing bits in thus chunk of the MUT. 4951 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4952 start_addr + (nth_task+1)*chunk_size); 4953 // The last chunk's end might be way beyond end of the 4954 // used region. In that case pull back appropriately. 4955 if (this_span.end() > end_addr) { 4956 this_span.set_end(end_addr); 4957 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4958 } 4959 // Iterate over the dirty cards covering this chunk, marking them 4960 // precleaned, and setting the corresponding bits in the mod union 4961 // table. Since we have been careful to partition at Card and MUT-word 4962 // boundaries no synchronization is needed between parallel threads. 4963 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4964 &modUnionClosure); 4965 4966 // Having transferred these marks into the modUnionTable, 4967 // rescan the marked objects on the dirty cards in the modUnionTable. 4968 // Even if this is at a synchronous collection, the initial marking 4969 // may have been done during an asynchronous collection so there 4970 // may be dirty bits in the mod-union table. 4971 _collector->_modUnionTable.dirty_range_iterate_clear( 4972 this_span, &greyRescanClosure); 4973 _collector->_modUnionTable.verifyNoOneBitsInRange( 4974 this_span.start(), 4975 this_span.end()); 4976 } 4977 pst->all_tasks_completed(); // declare that i am done 4978 } 4979 4980 // . see if we can share work_queues with ParNew? XXX 4981 void 4982 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4983 int* seed) { 4984 OopTaskQueue* work_q = work_queue(i); 4985 NOT_PRODUCT(int num_steals = 0;) 4986 oop obj_to_scan; 4987 CMSBitMap* bm = &(_collector->_markBitMap); 4988 4989 while (true) { 4990 // Completely finish any left over work from (an) earlier round(s) 4991 cl->trim_queue(0); 4992 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4993 (size_t)ParGCDesiredObjsFromOverflowList); 4994 // Now check if there's any work in the overflow list 4995 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4996 // only affects the number of attempts made to get work from the 4997 // overflow list and does not affect the number of workers. Just 4998 // pass ParallelGCThreads so this behavior is unchanged. 4999 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5000 work_q, 5001 ParallelGCThreads)) { 5002 // found something in global overflow list; 5003 // not yet ready to go stealing work from others. 5004 // We'd like to assert(work_q->size() != 0, ...) 5005 // because we just took work from the overflow list, 5006 // but of course we can't since all of that could have 5007 // been already stolen from us. 5008 // "He giveth and He taketh away." 5009 continue; 5010 } 5011 // Verify that we have no work before we resort to stealing 5012 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5013 // Try to steal from other queues that have work 5014 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5015 NOT_PRODUCT(num_steals++;) 5016 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5017 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5018 // Do scanning work 5019 obj_to_scan->oop_iterate(cl); 5020 // Loop around, finish this work, and try to steal some more 5021 } else if (terminator()->offer_termination()) { 5022 break; // nirvana from the infinite cycle 5023 } 5024 } 5025 NOT_PRODUCT( 5026 if (PrintCMSStatistics != 0) { 5027 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5028 } 5029 ) 5030 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5031 "Else our work is not yet done"); 5032 } 5033 5034 // Record object boundaries in _eden_chunk_array by sampling the eden 5035 // top in the slow-path eden object allocation code path and record 5036 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5037 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5038 // sampling in sample_eden() that activates during the part of the 5039 // preclean phase. 5040 void CMSCollector::sample_eden_chunk() { 5041 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5042 if (_eden_chunk_lock->try_lock()) { 5043 // Record a sample. This is the critical section. The contents 5044 // of the _eden_chunk_array have to be non-decreasing in the 5045 // address order. 5046 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5047 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5048 "Unexpected state of Eden"); 5049 if (_eden_chunk_index == 0 || 5050 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5051 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5052 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5053 _eden_chunk_index++; // commit sample 5054 } 5055 _eden_chunk_lock->unlock(); 5056 } 5057 } 5058 } 5059 5060 // Return a thread-local PLAB recording array, as appropriate. 5061 void* CMSCollector::get_data_recorder(int thr_num) { 5062 if (_survivor_plab_array != NULL && 5063 (CMSPLABRecordAlways || 5064 (_collectorState > Marking && _collectorState < FinalMarking))) { 5065 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5066 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5067 ca->reset(); // clear it so that fresh data is recorded 5068 return (void*) ca; 5069 } else { 5070 return NULL; 5071 } 5072 } 5073 5074 // Reset all the thread-local PLAB recording arrays 5075 void CMSCollector::reset_survivor_plab_arrays() { 5076 for (uint i = 0; i < ParallelGCThreads; i++) { 5077 _survivor_plab_array[i].reset(); 5078 } 5079 } 5080 5081 // Merge the per-thread plab arrays into the global survivor chunk 5082 // array which will provide the partitioning of the survivor space 5083 // for CMS initial scan and rescan. 5084 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5085 int no_of_gc_threads) { 5086 assert(_survivor_plab_array != NULL, "Error"); 5087 assert(_survivor_chunk_array != NULL, "Error"); 5088 assert(_collectorState == FinalMarking || 5089 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5090 for (int j = 0; j < no_of_gc_threads; j++) { 5091 _cursor[j] = 0; 5092 } 5093 HeapWord* top = surv->top(); 5094 size_t i; 5095 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5096 HeapWord* min_val = top; // Higher than any PLAB address 5097 uint min_tid = 0; // position of min_val this round 5098 for (int j = 0; j < no_of_gc_threads; j++) { 5099 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5100 if (_cursor[j] == cur_sca->end()) { 5101 continue; 5102 } 5103 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5104 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5105 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5106 if (cur_val < min_val) { 5107 min_tid = j; 5108 min_val = cur_val; 5109 } else { 5110 assert(cur_val < top, "All recorded addresses should be less"); 5111 } 5112 } 5113 // At this point min_val and min_tid are respectively 5114 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5115 // and the thread (j) that witnesses that address. 5116 // We record this address in the _survivor_chunk_array[i] 5117 // and increment _cursor[min_tid] prior to the next round i. 5118 if (min_val == top) { 5119 break; 5120 } 5121 _survivor_chunk_array[i] = min_val; 5122 _cursor[min_tid]++; 5123 } 5124 // We are all done; record the size of the _survivor_chunk_array 5125 _survivor_chunk_index = i; // exclusive: [0, i) 5126 if (PrintCMSStatistics > 0) { 5127 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5128 } 5129 // Verify that we used up all the recorded entries 5130 #ifdef ASSERT 5131 size_t total = 0; 5132 for (int j = 0; j < no_of_gc_threads; j++) { 5133 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5134 total += _cursor[j]; 5135 } 5136 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5137 // Check that the merged array is in sorted order 5138 if (total > 0) { 5139 for (size_t i = 0; i < total - 1; i++) { 5140 if (PrintCMSStatistics > 0) { 5141 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5142 i, _survivor_chunk_array[i]); 5143 } 5144 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5145 "Not sorted"); 5146 } 5147 } 5148 #endif // ASSERT 5149 } 5150 5151 // Set up the space's par_seq_tasks structure for work claiming 5152 // for parallel initial scan and rescan of young gen. 5153 // See ParRescanTask where this is currently used. 5154 void 5155 CMSCollector:: 5156 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5157 assert(n_threads > 0, "Unexpected n_threads argument"); 5158 5159 // Eden space 5160 if (!_young_gen->eden()->is_empty()) { 5161 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5162 assert(!pst->valid(), "Clobbering existing data?"); 5163 // Each valid entry in [0, _eden_chunk_index) represents a task. 5164 size_t n_tasks = _eden_chunk_index + 1; 5165 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5166 // Sets the condition for completion of the subtask (how many threads 5167 // need to finish in order to be done). 5168 pst->set_n_threads(n_threads); 5169 pst->set_n_tasks((int)n_tasks); 5170 } 5171 5172 // Merge the survivor plab arrays into _survivor_chunk_array 5173 if (_survivor_plab_array != NULL) { 5174 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5175 } else { 5176 assert(_survivor_chunk_index == 0, "Error"); 5177 } 5178 5179 // To space 5180 { 5181 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5182 assert(!pst->valid(), "Clobbering existing data?"); 5183 // Sets the condition for completion of the subtask (how many threads 5184 // need to finish in order to be done). 5185 pst->set_n_threads(n_threads); 5186 pst->set_n_tasks(1); 5187 assert(pst->valid(), "Error"); 5188 } 5189 5190 // From space 5191 { 5192 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5193 assert(!pst->valid(), "Clobbering existing data?"); 5194 size_t n_tasks = _survivor_chunk_index + 1; 5195 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5196 // Sets the condition for completion of the subtask (how many threads 5197 // need to finish in order to be done). 5198 pst->set_n_threads(n_threads); 5199 pst->set_n_tasks((int)n_tasks); 5200 assert(pst->valid(), "Error"); 5201 } 5202 } 5203 5204 // Parallel version of remark 5205 void CMSCollector::do_remark_parallel() { 5206 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5207 FlexibleWorkGang* workers = gch->workers(); 5208 assert(workers != NULL, "Need parallel worker threads."); 5209 // Choose to use the number of GC workers most recently set 5210 // into "active_workers". If active_workers is not set, set it 5211 // to ParallelGCThreads. 5212 int n_workers = workers->active_workers(); 5213 if (n_workers == 0) { 5214 assert(n_workers > 0, "Should have been set during scavenge"); 5215 n_workers = ParallelGCThreads; 5216 workers->set_active_workers(n_workers); 5217 } 5218 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5219 5220 CMSParRemarkTask tsk(this, 5221 cms_space, 5222 n_workers, workers, task_queues()); 5223 5224 // Set up for parallel process_roots work. 5225 gch->set_par_threads(n_workers); 5226 // We won't be iterating over the cards in the card table updating 5227 // the younger_gen cards, so we shouldn't call the following else 5228 // the verification code as well as subsequent younger_refs_iterate 5229 // code would get confused. XXX 5230 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5231 5232 // The young gen rescan work will not be done as part of 5233 // process_roots (which currently doesn't know how to 5234 // parallelize such a scan), but rather will be broken up into 5235 // a set of parallel tasks (via the sampling that the [abortable] 5236 // preclean phase did of eden, plus the [two] tasks of 5237 // scanning the [two] survivor spaces. Further fine-grain 5238 // parallelization of the scanning of the survivor spaces 5239 // themselves, and of precleaning of the younger gen itself 5240 // is deferred to the future. 5241 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5242 5243 // The dirty card rescan work is broken up into a "sequence" 5244 // of parallel tasks (per constituent space) that are dynamically 5245 // claimed by the parallel threads. 5246 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5247 5248 // It turns out that even when we're using 1 thread, doing the work in a 5249 // separate thread causes wide variance in run times. We can't help this 5250 // in the multi-threaded case, but we special-case n=1 here to get 5251 // repeatable measurements of the 1-thread overhead of the parallel code. 5252 if (n_workers > 1) { 5253 // Make refs discovery MT-safe, if it isn't already: it may not 5254 // necessarily be so, since it's possible that we are doing 5255 // ST marking. 5256 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5257 GenCollectedHeap::StrongRootsScope srs(gch); 5258 workers->run_task(&tsk); 5259 } else { 5260 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5261 GenCollectedHeap::StrongRootsScope srs(gch); 5262 tsk.work(0); 5263 } 5264 5265 gch->set_par_threads(0); // 0 ==> non-parallel. 5266 // restore, single-threaded for now, any preserved marks 5267 // as a result of work_q overflow 5268 restore_preserved_marks_if_any(); 5269 } 5270 5271 // Non-parallel version of remark 5272 void CMSCollector::do_remark_non_parallel() { 5273 ResourceMark rm; 5274 HandleMark hm; 5275 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5276 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5277 5278 MarkRefsIntoAndScanClosure 5279 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5280 &_markStack, this, 5281 false /* should_yield */, false /* not precleaning */); 5282 MarkFromDirtyCardsClosure 5283 markFromDirtyCardsClosure(this, _span, 5284 NULL, // space is set further below 5285 &_markBitMap, &_markStack, &mrias_cl); 5286 { 5287 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5288 // Iterate over the dirty cards, setting the corresponding bits in the 5289 // mod union table. 5290 { 5291 ModUnionClosure modUnionClosure(&_modUnionTable); 5292 _ct->ct_bs()->dirty_card_iterate( 5293 _cmsGen->used_region(), 5294 &modUnionClosure); 5295 } 5296 // Having transferred these marks into the modUnionTable, we just need 5297 // to rescan the marked objects on the dirty cards in the modUnionTable. 5298 // The initial marking may have been done during an asynchronous 5299 // collection so there may be dirty bits in the mod-union table. 5300 const int alignment = 5301 CardTableModRefBS::card_size * BitsPerWord; 5302 { 5303 // ... First handle dirty cards in CMS gen 5304 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5305 MemRegion ur = _cmsGen->used_region(); 5306 HeapWord* lb = ur.start(); 5307 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5308 MemRegion cms_span(lb, ub); 5309 _modUnionTable.dirty_range_iterate_clear(cms_span, 5310 &markFromDirtyCardsClosure); 5311 verify_work_stacks_empty(); 5312 if (PrintCMSStatistics != 0) { 5313 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5314 markFromDirtyCardsClosure.num_dirty_cards()); 5315 } 5316 } 5317 } 5318 if (VerifyDuringGC && 5319 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5320 HandleMark hm; // Discard invalid handles created during verification 5321 Universe::verify(); 5322 } 5323 { 5324 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5325 5326 verify_work_stacks_empty(); 5327 5328 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5329 GenCollectedHeap::StrongRootsScope srs(gch); 5330 5331 gch->gen_process_roots(_cmsGen->level(), 5332 true, // younger gens as roots 5333 false, // use the local StrongRootsScope 5334 SharedHeap::ScanningOption(roots_scanning_options()), 5335 should_unload_classes(), 5336 &mrias_cl, 5337 NULL, 5338 NULL); // The dirty klasses will be handled below 5339 5340 assert(should_unload_classes() 5341 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5342 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5343 } 5344 5345 { 5346 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5347 5348 verify_work_stacks_empty(); 5349 5350 // Scan all class loader data objects that might have been introduced 5351 // during concurrent marking. 5352 ResourceMark rm; 5353 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5354 for (int i = 0; i < array->length(); i++) { 5355 mrias_cl.do_class_loader_data(array->at(i)); 5356 } 5357 5358 // We don't need to keep track of new CLDs anymore. 5359 ClassLoaderDataGraph::remember_new_clds(false); 5360 5361 verify_work_stacks_empty(); 5362 } 5363 5364 { 5365 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5366 5367 verify_work_stacks_empty(); 5368 5369 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5370 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5371 5372 verify_work_stacks_empty(); 5373 } 5374 5375 // We might have added oops to ClassLoaderData::_handles during the 5376 // concurrent marking phase. These oops point to newly allocated objects 5377 // that are guaranteed to be kept alive. Either by the direct allocation 5378 // code, or when the young collector processes the roots. Hence, 5379 // we don't have to revisit the _handles block during the remark phase. 5380 5381 verify_work_stacks_empty(); 5382 // Restore evacuated mark words, if any, used for overflow list links 5383 if (!CMSOverflowEarlyRestoration) { 5384 restore_preserved_marks_if_any(); 5385 } 5386 verify_overflow_empty(); 5387 } 5388 5389 //////////////////////////////////////////////////////// 5390 // Parallel Reference Processing Task Proxy Class 5391 //////////////////////////////////////////////////////// 5392 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5393 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5394 CMSCollector* _collector; 5395 CMSBitMap* _mark_bit_map; 5396 const MemRegion _span; 5397 ProcessTask& _task; 5398 5399 public: 5400 CMSRefProcTaskProxy(ProcessTask& task, 5401 CMSCollector* collector, 5402 const MemRegion& span, 5403 CMSBitMap* mark_bit_map, 5404 AbstractWorkGang* workers, 5405 OopTaskQueueSet* task_queues): 5406 // XXX Should superclass AGTWOQ also know about AWG since it knows 5407 // about the task_queues used by the AWG? Then it could initialize 5408 // the terminator() object. See 6984287. The set_for_termination() 5409 // below is a temporary band-aid for the regression in 6984287. 5410 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5411 task_queues), 5412 _task(task), 5413 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5414 { 5415 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5416 "Inconsistency in _span"); 5417 set_for_termination(workers->active_workers()); 5418 } 5419 5420 OopTaskQueueSet* task_queues() { return queues(); } 5421 5422 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5423 5424 void do_work_steal(int i, 5425 CMSParDrainMarkingStackClosure* drain, 5426 CMSParKeepAliveClosure* keep_alive, 5427 int* seed); 5428 5429 virtual void work(uint worker_id); 5430 }; 5431 5432 void CMSRefProcTaskProxy::work(uint worker_id) { 5433 ResourceMark rm; 5434 HandleMark hm; 5435 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5436 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5437 _mark_bit_map, 5438 work_queue(worker_id)); 5439 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5440 _mark_bit_map, 5441 work_queue(worker_id)); 5442 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5443 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5444 if (_task.marks_oops_alive()) { 5445 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5446 _collector->hash_seed(worker_id)); 5447 } 5448 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5449 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5450 } 5451 5452 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5453 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5454 EnqueueTask& _task; 5455 5456 public: 5457 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5458 : AbstractGangTask("Enqueue reference objects in parallel"), 5459 _task(task) 5460 { } 5461 5462 virtual void work(uint worker_id) 5463 { 5464 _task.work(worker_id); 5465 } 5466 }; 5467 5468 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5469 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5470 _span(span), 5471 _bit_map(bit_map), 5472 _work_queue(work_queue), 5473 _mark_and_push(collector, span, bit_map, work_queue), 5474 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5475 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5476 { } 5477 5478 // . see if we can share work_queues with ParNew? XXX 5479 void CMSRefProcTaskProxy::do_work_steal(int i, 5480 CMSParDrainMarkingStackClosure* drain, 5481 CMSParKeepAliveClosure* keep_alive, 5482 int* seed) { 5483 OopTaskQueue* work_q = work_queue(i); 5484 NOT_PRODUCT(int num_steals = 0;) 5485 oop obj_to_scan; 5486 5487 while (true) { 5488 // Completely finish any left over work from (an) earlier round(s) 5489 drain->trim_queue(0); 5490 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5491 (size_t)ParGCDesiredObjsFromOverflowList); 5492 // Now check if there's any work in the overflow list 5493 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5494 // only affects the number of attempts made to get work from the 5495 // overflow list and does not affect the number of workers. Just 5496 // pass ParallelGCThreads so this behavior is unchanged. 5497 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5498 work_q, 5499 ParallelGCThreads)) { 5500 // Found something in global overflow list; 5501 // not yet ready to go stealing work from others. 5502 // We'd like to assert(work_q->size() != 0, ...) 5503 // because we just took work from the overflow list, 5504 // but of course we can't, since all of that might have 5505 // been already stolen from us. 5506 continue; 5507 } 5508 // Verify that we have no work before we resort to stealing 5509 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5510 // Try to steal from other queues that have work 5511 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5512 NOT_PRODUCT(num_steals++;) 5513 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5514 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5515 // Do scanning work 5516 obj_to_scan->oop_iterate(keep_alive); 5517 // Loop around, finish this work, and try to steal some more 5518 } else if (terminator()->offer_termination()) { 5519 break; // nirvana from the infinite cycle 5520 } 5521 } 5522 NOT_PRODUCT( 5523 if (PrintCMSStatistics != 0) { 5524 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5525 } 5526 ) 5527 } 5528 5529 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5530 { 5531 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5532 FlexibleWorkGang* workers = gch->workers(); 5533 assert(workers != NULL, "Need parallel worker threads."); 5534 CMSRefProcTaskProxy rp_task(task, &_collector, 5535 _collector.ref_processor()->span(), 5536 _collector.markBitMap(), 5537 workers, _collector.task_queues()); 5538 workers->run_task(&rp_task); 5539 } 5540 5541 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5542 { 5543 5544 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5545 FlexibleWorkGang* workers = gch->workers(); 5546 assert(workers != NULL, "Need parallel worker threads."); 5547 CMSRefEnqueueTaskProxy enq_task(task); 5548 workers->run_task(&enq_task); 5549 } 5550 5551 void CMSCollector::refProcessingWork() { 5552 ResourceMark rm; 5553 HandleMark hm; 5554 5555 ReferenceProcessor* rp = ref_processor(); 5556 assert(rp->span().equals(_span), "Spans should be equal"); 5557 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5558 // Process weak references. 5559 rp->setup_policy(false); 5560 verify_work_stacks_empty(); 5561 5562 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5563 &_markStack, false /* !preclean */); 5564 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5565 _span, &_markBitMap, &_markStack, 5566 &cmsKeepAliveClosure, false /* !preclean */); 5567 { 5568 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5569 5570 ReferenceProcessorStats stats; 5571 if (rp->processing_is_mt()) { 5572 // Set the degree of MT here. If the discovery is done MT, there 5573 // may have been a different number of threads doing the discovery 5574 // and a different number of discovered lists may have Ref objects. 5575 // That is OK as long as the Reference lists are balanced (see 5576 // balance_all_queues() and balance_queues()). 5577 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5578 int active_workers = ParallelGCThreads; 5579 FlexibleWorkGang* workers = gch->workers(); 5580 if (workers != NULL) { 5581 active_workers = workers->active_workers(); 5582 // The expectation is that active_workers will have already 5583 // been set to a reasonable value. If it has not been set, 5584 // investigate. 5585 assert(active_workers > 0, "Should have been set during scavenge"); 5586 } 5587 rp->set_active_mt_degree(active_workers); 5588 CMSRefProcTaskExecutor task_executor(*this); 5589 stats = rp->process_discovered_references(&_is_alive_closure, 5590 &cmsKeepAliveClosure, 5591 &cmsDrainMarkingStackClosure, 5592 &task_executor, 5593 _gc_timer_cm, 5594 _gc_tracer_cm->gc_id()); 5595 } else { 5596 stats = rp->process_discovered_references(&_is_alive_closure, 5597 &cmsKeepAliveClosure, 5598 &cmsDrainMarkingStackClosure, 5599 NULL, 5600 _gc_timer_cm, 5601 _gc_tracer_cm->gc_id()); 5602 } 5603 _gc_tracer_cm->report_gc_reference_stats(stats); 5604 5605 } 5606 5607 // This is the point where the entire marking should have completed. 5608 verify_work_stacks_empty(); 5609 5610 if (should_unload_classes()) { 5611 { 5612 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5613 5614 // Unload classes and purge the SystemDictionary. 5615 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5616 5617 // Unload nmethods. 5618 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5619 5620 // Prune dead klasses from subklass/sibling/implementor lists. 5621 Klass::clean_weak_klass_links(&_is_alive_closure); 5622 } 5623 5624 { 5625 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5626 // Clean up unreferenced symbols in symbol table. 5627 SymbolTable::unlink(); 5628 } 5629 5630 { 5631 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5632 // Delete entries for dead interned strings. 5633 StringTable::unlink(&_is_alive_closure); 5634 } 5635 } 5636 5637 5638 // Restore any preserved marks as a result of mark stack or 5639 // work queue overflow 5640 restore_preserved_marks_if_any(); // done single-threaded for now 5641 5642 rp->set_enqueuing_is_done(true); 5643 if (rp->processing_is_mt()) { 5644 rp->balance_all_queues(); 5645 CMSRefProcTaskExecutor task_executor(*this); 5646 rp->enqueue_discovered_references(&task_executor); 5647 } else { 5648 rp->enqueue_discovered_references(NULL); 5649 } 5650 rp->verify_no_references_recorded(); 5651 assert(!rp->discovery_enabled(), "should have been disabled"); 5652 } 5653 5654 #ifndef PRODUCT 5655 void CMSCollector::check_correct_thread_executing() { 5656 Thread* t = Thread::current(); 5657 // Only the VM thread or the CMS thread should be here. 5658 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5659 "Unexpected thread type"); 5660 // If this is the vm thread, the foreground process 5661 // should not be waiting. Note that _foregroundGCIsActive is 5662 // true while the foreground collector is waiting. 5663 if (_foregroundGCShouldWait) { 5664 // We cannot be the VM thread 5665 assert(t->is_ConcurrentGC_thread(), 5666 "Should be CMS thread"); 5667 } else { 5668 // We can be the CMS thread only if we are in a stop-world 5669 // phase of CMS collection. 5670 if (t->is_ConcurrentGC_thread()) { 5671 assert(_collectorState == InitialMarking || 5672 _collectorState == FinalMarking, 5673 "Should be a stop-world phase"); 5674 // The CMS thread should be holding the CMS_token. 5675 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5676 "Potential interference with concurrently " 5677 "executing VM thread"); 5678 } 5679 } 5680 } 5681 #endif 5682 5683 void CMSCollector::sweep() { 5684 assert(_collectorState == Sweeping, "just checking"); 5685 check_correct_thread_executing(); 5686 verify_work_stacks_empty(); 5687 verify_overflow_empty(); 5688 increment_sweep_count(); 5689 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5690 5691 _inter_sweep_timer.stop(); 5692 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5693 5694 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5695 _intra_sweep_timer.reset(); 5696 _intra_sweep_timer.start(); 5697 { 5698 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5699 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5700 // First sweep the old gen 5701 { 5702 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5703 bitMapLock()); 5704 sweepWork(_cmsGen); 5705 } 5706 5707 // Update Universe::_heap_*_at_gc figures. 5708 // We need all the free list locks to make the abstract state 5709 // transition from Sweeping to Resetting. See detailed note 5710 // further below. 5711 { 5712 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5713 // Update heap occupancy information which is used as 5714 // input to soft ref clearing policy at the next gc. 5715 Universe::update_heap_info_at_gc(); 5716 _collectorState = Resizing; 5717 } 5718 } 5719 verify_work_stacks_empty(); 5720 verify_overflow_empty(); 5721 5722 if (should_unload_classes()) { 5723 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5724 // requires that the virtual spaces are stable and not deleted. 5725 ClassLoaderDataGraph::set_should_purge(true); 5726 } 5727 5728 _intra_sweep_timer.stop(); 5729 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5730 5731 _inter_sweep_timer.reset(); 5732 _inter_sweep_timer.start(); 5733 5734 // We need to use a monotonically non-decreasing time in ms 5735 // or we will see time-warp warnings and os::javaTimeMillis() 5736 // does not guarantee monotonicity. 5737 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5738 update_time_of_last_gc(now); 5739 5740 // NOTE on abstract state transitions: 5741 // Mutators allocate-live and/or mark the mod-union table dirty 5742 // based on the state of the collection. The former is done in 5743 // the interval [Marking, Sweeping] and the latter in the interval 5744 // [Marking, Sweeping). Thus the transitions into the Marking state 5745 // and out of the Sweeping state must be synchronously visible 5746 // globally to the mutators. 5747 // The transition into the Marking state happens with the world 5748 // stopped so the mutators will globally see it. Sweeping is 5749 // done asynchronously by the background collector so the transition 5750 // from the Sweeping state to the Resizing state must be done 5751 // under the freelistLock (as is the check for whether to 5752 // allocate-live and whether to dirty the mod-union table). 5753 assert(_collectorState == Resizing, "Change of collector state to" 5754 " Resizing must be done under the freelistLocks (plural)"); 5755 5756 // Now that sweeping has been completed, we clear 5757 // the incremental_collection_failed flag, 5758 // thus inviting a younger gen collection to promote into 5759 // this generation. If such a promotion may still fail, 5760 // the flag will be set again when a young collection is 5761 // attempted. 5762 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5763 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5764 gch->update_full_collections_completed(_collection_count_start); 5765 } 5766 5767 // FIX ME!!! Looks like this belongs in CFLSpace, with 5768 // CMSGen merely delegating to it. 5769 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5770 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5771 HeapWord* minAddr = _cmsSpace->bottom(); 5772 HeapWord* largestAddr = 5773 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5774 if (largestAddr == NULL) { 5775 // The dictionary appears to be empty. In this case 5776 // try to coalesce at the end of the heap. 5777 largestAddr = _cmsSpace->end(); 5778 } 5779 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5780 size_t nearLargestOffset = 5781 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5782 if (PrintFLSStatistics != 0) { 5783 gclog_or_tty->print_cr( 5784 "CMS: Large Block: " PTR_FORMAT ";" 5785 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5786 largestAddr, 5787 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 5788 } 5789 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5790 } 5791 5792 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5793 return addr >= _cmsSpace->nearLargestChunk(); 5794 } 5795 5796 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5797 return _cmsSpace->find_chunk_at_end(); 5798 } 5799 5800 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5801 bool full) { 5802 // The next lower level has been collected. Gather any statistics 5803 // that are of interest at this point. 5804 if (!full && (current_level + 1) == level()) { 5805 // Gather statistics on the young generation collection. 5806 collector()->stats().record_gc0_end(used()); 5807 } 5808 } 5809 5810 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5811 // We iterate over the space(s) underlying this generation, 5812 // checking the mark bit map to see if the bits corresponding 5813 // to specific blocks are marked or not. Blocks that are 5814 // marked are live and are not swept up. All remaining blocks 5815 // are swept up, with coalescing on-the-fly as we sweep up 5816 // contiguous free and/or garbage blocks: 5817 // We need to ensure that the sweeper synchronizes with allocators 5818 // and stop-the-world collectors. In particular, the following 5819 // locks are used: 5820 // . CMS token: if this is held, a stop the world collection cannot occur 5821 // . freelistLock: if this is held no allocation can occur from this 5822 // generation by another thread 5823 // . bitMapLock: if this is held, no other thread can access or update 5824 // 5825 5826 // Note that we need to hold the freelistLock if we use 5827 // block iterate below; else the iterator might go awry if 5828 // a mutator (or promotion) causes block contents to change 5829 // (for instance if the allocator divvies up a block). 5830 // If we hold the free list lock, for all practical purposes 5831 // young generation GC's can't occur (they'll usually need to 5832 // promote), so we might as well prevent all young generation 5833 // GC's while we do a sweeping step. For the same reason, we might 5834 // as well take the bit map lock for the entire duration 5835 5836 // check that we hold the requisite locks 5837 assert(have_cms_token(), "Should hold cms token"); 5838 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5839 assert_lock_strong(gen->freelistLock()); 5840 assert_lock_strong(bitMapLock()); 5841 5842 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5843 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5844 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5845 _inter_sweep_estimate.padded_average(), 5846 _intra_sweep_estimate.padded_average()); 5847 gen->setNearLargestChunk(); 5848 5849 { 5850 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5851 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5852 // We need to free-up/coalesce garbage/blocks from a 5853 // co-terminal free run. This is done in the SweepClosure 5854 // destructor; so, do not remove this scope, else the 5855 // end-of-sweep-census below will be off by a little bit. 5856 } 5857 gen->cmsSpace()->sweep_completed(); 5858 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5859 if (should_unload_classes()) { // unloaded classes this cycle, 5860 _concurrent_cycles_since_last_unload = 0; // ... reset count 5861 } else { // did not unload classes, 5862 _concurrent_cycles_since_last_unload++; // ... increment count 5863 } 5864 } 5865 5866 // Reset CMS data structures (for now just the marking bit map) 5867 // preparatory for the next cycle. 5868 void CMSCollector::reset(bool concurrent) { 5869 if (concurrent) { 5870 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5871 5872 // If the state is not "Resetting", the foreground thread 5873 // has done a collection and the resetting. 5874 if (_collectorState != Resetting) { 5875 assert(_collectorState == Idling, "The state should only change" 5876 " because the foreground collector has finished the collection"); 5877 return; 5878 } 5879 5880 // Clear the mark bitmap (no grey objects to start with) 5881 // for the next cycle. 5882 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5883 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5884 5885 HeapWord* curAddr = _markBitMap.startWord(); 5886 while (curAddr < _markBitMap.endWord()) { 5887 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5888 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5889 _markBitMap.clear_large_range(chunk); 5890 if (ConcurrentMarkSweepThread::should_yield() && 5891 !foregroundGCIsActive() && 5892 CMSYield) { 5893 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5894 "CMS thread should hold CMS token"); 5895 assert_lock_strong(bitMapLock()); 5896 bitMapLock()->unlock(); 5897 ConcurrentMarkSweepThread::desynchronize(true); 5898 stopTimer(); 5899 if (PrintCMSStatistics != 0) { 5900 incrementYields(); 5901 } 5902 5903 // See the comment in coordinator_yield() 5904 for (unsigned i = 0; i < CMSYieldSleepCount && 5905 ConcurrentMarkSweepThread::should_yield() && 5906 !CMSCollector::foregroundGCIsActive(); ++i) { 5907 os::sleep(Thread::current(), 1, false); 5908 } 5909 5910 ConcurrentMarkSweepThread::synchronize(true); 5911 bitMapLock()->lock_without_safepoint_check(); 5912 startTimer(); 5913 } 5914 curAddr = chunk.end(); 5915 } 5916 // A successful mostly concurrent collection has been done. 5917 // Because only the full (i.e., concurrent mode failure) collections 5918 // are being measured for gc overhead limits, clean the "near" flag 5919 // and count. 5920 size_policy()->reset_gc_overhead_limit_count(); 5921 _collectorState = Idling; 5922 } else { 5923 // already have the lock 5924 assert(_collectorState == Resetting, "just checking"); 5925 assert_lock_strong(bitMapLock()); 5926 _markBitMap.clear_all(); 5927 _collectorState = Idling; 5928 } 5929 5930 register_gc_end(); 5931 } 5932 5933 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5934 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5935 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5936 TraceCollectorStats tcs(counters()); 5937 5938 switch (op) { 5939 case CMS_op_checkpointRootsInitial: { 5940 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5941 checkpointRootsInitial(); 5942 if (PrintGC) { 5943 _cmsGen->printOccupancy("initial-mark"); 5944 } 5945 break; 5946 } 5947 case CMS_op_checkpointRootsFinal: { 5948 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5949 checkpointRootsFinal(); 5950 if (PrintGC) { 5951 _cmsGen->printOccupancy("remark"); 5952 } 5953 break; 5954 } 5955 default: 5956 fatal("No such CMS_op"); 5957 } 5958 } 5959 5960 #ifndef PRODUCT 5961 size_t const CMSCollector::skip_header_HeapWords() { 5962 return FreeChunk::header_size(); 5963 } 5964 5965 // Try and collect here conditions that should hold when 5966 // CMS thread is exiting. The idea is that the foreground GC 5967 // thread should not be blocked if it wants to terminate 5968 // the CMS thread and yet continue to run the VM for a while 5969 // after that. 5970 void CMSCollector::verify_ok_to_terminate() const { 5971 assert(Thread::current()->is_ConcurrentGC_thread(), 5972 "should be called by CMS thread"); 5973 assert(!_foregroundGCShouldWait, "should be false"); 5974 // We could check here that all the various low-level locks 5975 // are not held by the CMS thread, but that is overkill; see 5976 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5977 // is checked. 5978 } 5979 #endif 5980 5981 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5982 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5983 "missing Printezis mark?"); 5984 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5985 size_t size = pointer_delta(nextOneAddr + 1, addr); 5986 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5987 "alignment problem"); 5988 assert(size >= 3, "Necessary for Printezis marks to work"); 5989 return size; 5990 } 5991 5992 // A variant of the above (block_size_using_printezis_bits()) except 5993 // that we return 0 if the P-bits are not yet set. 5994 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5995 if (_markBitMap.isMarked(addr + 1)) { 5996 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5997 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5998 size_t size = pointer_delta(nextOneAddr + 1, addr); 5999 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6000 "alignment problem"); 6001 assert(size >= 3, "Necessary for Printezis marks to work"); 6002 return size; 6003 } 6004 return 0; 6005 } 6006 6007 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6008 size_t sz = 0; 6009 oop p = (oop)addr; 6010 if (p->klass_or_null() != NULL) { 6011 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6012 } else { 6013 sz = block_size_using_printezis_bits(addr); 6014 } 6015 assert(sz > 0, "size must be nonzero"); 6016 HeapWord* next_block = addr + sz; 6017 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6018 CardTableModRefBS::card_size); 6019 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6020 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6021 "must be different cards"); 6022 return next_card; 6023 } 6024 6025 6026 // CMS Bit Map Wrapper ///////////////////////////////////////// 6027 6028 // Construct a CMS bit map infrastructure, but don't create the 6029 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6030 // further below. 6031 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6032 _bm(), 6033 _shifter(shifter), 6034 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6035 { 6036 _bmStartWord = 0; 6037 _bmWordSize = 0; 6038 } 6039 6040 bool CMSBitMap::allocate(MemRegion mr) { 6041 _bmStartWord = mr.start(); 6042 _bmWordSize = mr.word_size(); 6043 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6044 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6045 if (!brs.is_reserved()) { 6046 warning("CMS bit map allocation failure"); 6047 return false; 6048 } 6049 // For now we'll just commit all of the bit map up front. 6050 // Later on we'll try to be more parsimonious with swap. 6051 if (!_virtual_space.initialize(brs, brs.size())) { 6052 warning("CMS bit map backing store failure"); 6053 return false; 6054 } 6055 assert(_virtual_space.committed_size() == brs.size(), 6056 "didn't reserve backing store for all of CMS bit map?"); 6057 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6058 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6059 _bmWordSize, "inconsistency in bit map sizing"); 6060 _bm.set_size(_bmWordSize >> _shifter); 6061 6062 // bm.clear(); // can we rely on getting zero'd memory? verify below 6063 assert(isAllClear(), 6064 "Expected zero'd memory from ReservedSpace constructor"); 6065 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6066 "consistency check"); 6067 return true; 6068 } 6069 6070 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6071 HeapWord *next_addr, *end_addr, *last_addr; 6072 assert_locked(); 6073 assert(covers(mr), "out-of-range error"); 6074 // XXX assert that start and end are appropriately aligned 6075 for (next_addr = mr.start(), end_addr = mr.end(); 6076 next_addr < end_addr; next_addr = last_addr) { 6077 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6078 last_addr = dirty_region.end(); 6079 if (!dirty_region.is_empty()) { 6080 cl->do_MemRegion(dirty_region); 6081 } else { 6082 assert(last_addr == end_addr, "program logic"); 6083 return; 6084 } 6085 } 6086 } 6087 6088 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6089 _bm.print_on_error(st, prefix); 6090 } 6091 6092 #ifndef PRODUCT 6093 void CMSBitMap::assert_locked() const { 6094 CMSLockVerifier::assert_locked(lock()); 6095 } 6096 6097 bool CMSBitMap::covers(MemRegion mr) const { 6098 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6099 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6100 "size inconsistency"); 6101 return (mr.start() >= _bmStartWord) && 6102 (mr.end() <= endWord()); 6103 } 6104 6105 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6106 return (start >= _bmStartWord && (start + size) <= endWord()); 6107 } 6108 6109 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6110 // verify that there are no 1 bits in the interval [left, right) 6111 FalseBitMapClosure falseBitMapClosure; 6112 iterate(&falseBitMapClosure, left, right); 6113 } 6114 6115 void CMSBitMap::region_invariant(MemRegion mr) 6116 { 6117 assert_locked(); 6118 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6119 assert(!mr.is_empty(), "unexpected empty region"); 6120 assert(covers(mr), "mr should be covered by bit map"); 6121 // convert address range into offset range 6122 size_t start_ofs = heapWordToOffset(mr.start()); 6123 // Make sure that end() is appropriately aligned 6124 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6125 (1 << (_shifter+LogHeapWordSize))), 6126 "Misaligned mr.end()"); 6127 size_t end_ofs = heapWordToOffset(mr.end()); 6128 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6129 } 6130 6131 #endif 6132 6133 bool CMSMarkStack::allocate(size_t size) { 6134 // allocate a stack of the requisite depth 6135 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6136 size * sizeof(oop))); 6137 if (!rs.is_reserved()) { 6138 warning("CMSMarkStack allocation failure"); 6139 return false; 6140 } 6141 if (!_virtual_space.initialize(rs, rs.size())) { 6142 warning("CMSMarkStack backing store failure"); 6143 return false; 6144 } 6145 assert(_virtual_space.committed_size() == rs.size(), 6146 "didn't reserve backing store for all of CMS stack?"); 6147 _base = (oop*)(_virtual_space.low()); 6148 _index = 0; 6149 _capacity = size; 6150 NOT_PRODUCT(_max_depth = 0); 6151 return true; 6152 } 6153 6154 // XXX FIX ME !!! In the MT case we come in here holding a 6155 // leaf lock. For printing we need to take a further lock 6156 // which has lower rank. We need to recalibrate the two 6157 // lock-ranks involved in order to be able to print the 6158 // messages below. (Or defer the printing to the caller. 6159 // For now we take the expedient path of just disabling the 6160 // messages for the problematic case.) 6161 void CMSMarkStack::expand() { 6162 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6163 if (_capacity == MarkStackSizeMax) { 6164 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6165 // We print a warning message only once per CMS cycle. 6166 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6167 } 6168 return; 6169 } 6170 // Double capacity if possible 6171 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6172 // Do not give up existing stack until we have managed to 6173 // get the double capacity that we desired. 6174 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6175 new_capacity * sizeof(oop))); 6176 if (rs.is_reserved()) { 6177 // Release the backing store associated with old stack 6178 _virtual_space.release(); 6179 // Reinitialize virtual space for new stack 6180 if (!_virtual_space.initialize(rs, rs.size())) { 6181 fatal("Not enough swap for expanded marking stack"); 6182 } 6183 _base = (oop*)(_virtual_space.low()); 6184 _index = 0; 6185 _capacity = new_capacity; 6186 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6187 // Failed to double capacity, continue; 6188 // we print a detail message only once per CMS cycle. 6189 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6190 SIZE_FORMAT"K", 6191 _capacity / K, new_capacity / K); 6192 } 6193 } 6194 6195 6196 // Closures 6197 // XXX: there seems to be a lot of code duplication here; 6198 // should refactor and consolidate common code. 6199 6200 // This closure is used to mark refs into the CMS generation in 6201 // the CMS bit map. Called at the first checkpoint. This closure 6202 // assumes that we do not need to re-mark dirty cards; if the CMS 6203 // generation on which this is used is not an oldest 6204 // generation then this will lose younger_gen cards! 6205 6206 MarkRefsIntoClosure::MarkRefsIntoClosure( 6207 MemRegion span, CMSBitMap* bitMap): 6208 _span(span), 6209 _bitMap(bitMap) 6210 { 6211 assert(_ref_processor == NULL, "deliberately left NULL"); 6212 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6213 } 6214 6215 void MarkRefsIntoClosure::do_oop(oop obj) { 6216 // if p points into _span, then mark corresponding bit in _markBitMap 6217 assert(obj->is_oop(), "expected an oop"); 6218 HeapWord* addr = (HeapWord*)obj; 6219 if (_span.contains(addr)) { 6220 // this should be made more efficient 6221 _bitMap->mark(addr); 6222 } 6223 } 6224 6225 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6226 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6227 6228 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6229 MemRegion span, CMSBitMap* bitMap): 6230 _span(span), 6231 _bitMap(bitMap) 6232 { 6233 assert(_ref_processor == NULL, "deliberately left NULL"); 6234 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6235 } 6236 6237 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6238 // if p points into _span, then mark corresponding bit in _markBitMap 6239 assert(obj->is_oop(), "expected an oop"); 6240 HeapWord* addr = (HeapWord*)obj; 6241 if (_span.contains(addr)) { 6242 // this should be made more efficient 6243 _bitMap->par_mark(addr); 6244 } 6245 } 6246 6247 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6248 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6249 6250 // A variant of the above, used for CMS marking verification. 6251 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6252 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6253 _span(span), 6254 _verification_bm(verification_bm), 6255 _cms_bm(cms_bm) 6256 { 6257 assert(_ref_processor == NULL, "deliberately left NULL"); 6258 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6259 } 6260 6261 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6262 // if p points into _span, then mark corresponding bit in _markBitMap 6263 assert(obj->is_oop(), "expected an oop"); 6264 HeapWord* addr = (HeapWord*)obj; 6265 if (_span.contains(addr)) { 6266 _verification_bm->mark(addr); 6267 if (!_cms_bm->isMarked(addr)) { 6268 oop(addr)->print(); 6269 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6270 fatal("... aborting"); 6271 } 6272 } 6273 } 6274 6275 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6276 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6277 6278 ////////////////////////////////////////////////// 6279 // MarkRefsIntoAndScanClosure 6280 ////////////////////////////////////////////////// 6281 6282 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6283 ReferenceProcessor* rp, 6284 CMSBitMap* bit_map, 6285 CMSBitMap* mod_union_table, 6286 CMSMarkStack* mark_stack, 6287 CMSCollector* collector, 6288 bool should_yield, 6289 bool concurrent_precleaning): 6290 _collector(collector), 6291 _span(span), 6292 _bit_map(bit_map), 6293 _mark_stack(mark_stack), 6294 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6295 mark_stack, concurrent_precleaning), 6296 _yield(should_yield), 6297 _concurrent_precleaning(concurrent_precleaning), 6298 _freelistLock(NULL) 6299 { 6300 _ref_processor = rp; 6301 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6302 } 6303 6304 // This closure is used to mark refs into the CMS generation at the 6305 // second (final) checkpoint, and to scan and transitively follow 6306 // the unmarked oops. It is also used during the concurrent precleaning 6307 // phase while scanning objects on dirty cards in the CMS generation. 6308 // The marks are made in the marking bit map and the marking stack is 6309 // used for keeping the (newly) grey objects during the scan. 6310 // The parallel version (Par_...) appears further below. 6311 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6312 if (obj != NULL) { 6313 assert(obj->is_oop(), "expected an oop"); 6314 HeapWord* addr = (HeapWord*)obj; 6315 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6316 assert(_collector->overflow_list_is_empty(), 6317 "overflow list should be empty"); 6318 if (_span.contains(addr) && 6319 !_bit_map->isMarked(addr)) { 6320 // mark bit map (object is now grey) 6321 _bit_map->mark(addr); 6322 // push on marking stack (stack should be empty), and drain the 6323 // stack by applying this closure to the oops in the oops popped 6324 // from the stack (i.e. blacken the grey objects) 6325 bool res = _mark_stack->push(obj); 6326 assert(res, "Should have space to push on empty stack"); 6327 do { 6328 oop new_oop = _mark_stack->pop(); 6329 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6330 assert(_bit_map->isMarked((HeapWord*)new_oop), 6331 "only grey objects on this stack"); 6332 // iterate over the oops in this oop, marking and pushing 6333 // the ones in CMS heap (i.e. in _span). 6334 new_oop->oop_iterate(&_pushAndMarkClosure); 6335 // check if it's time to yield 6336 do_yield_check(); 6337 } while (!_mark_stack->isEmpty() || 6338 (!_concurrent_precleaning && take_from_overflow_list())); 6339 // if marking stack is empty, and we are not doing this 6340 // during precleaning, then check the overflow list 6341 } 6342 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6343 assert(_collector->overflow_list_is_empty(), 6344 "overflow list was drained above"); 6345 // We could restore evacuated mark words, if any, used for 6346 // overflow list links here because the overflow list is 6347 // provably empty here. That would reduce the maximum 6348 // size requirements for preserved_{oop,mark}_stack. 6349 // But we'll just postpone it until we are all done 6350 // so we can just stream through. 6351 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6352 _collector->restore_preserved_marks_if_any(); 6353 assert(_collector->no_preserved_marks(), "No preserved marks"); 6354 } 6355 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6356 "All preserved marks should have been restored above"); 6357 } 6358 } 6359 6360 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6361 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6362 6363 void MarkRefsIntoAndScanClosure::do_yield_work() { 6364 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6365 "CMS thread should hold CMS token"); 6366 assert_lock_strong(_freelistLock); 6367 assert_lock_strong(_bit_map->lock()); 6368 // relinquish the free_list_lock and bitMaplock() 6369 _bit_map->lock()->unlock(); 6370 _freelistLock->unlock(); 6371 ConcurrentMarkSweepThread::desynchronize(true); 6372 _collector->stopTimer(); 6373 if (PrintCMSStatistics != 0) { 6374 _collector->incrementYields(); 6375 } 6376 6377 // See the comment in coordinator_yield() 6378 for (unsigned i = 0; 6379 i < CMSYieldSleepCount && 6380 ConcurrentMarkSweepThread::should_yield() && 6381 !CMSCollector::foregroundGCIsActive(); 6382 ++i) { 6383 os::sleep(Thread::current(), 1, false); 6384 } 6385 6386 ConcurrentMarkSweepThread::synchronize(true); 6387 _freelistLock->lock_without_safepoint_check(); 6388 _bit_map->lock()->lock_without_safepoint_check(); 6389 _collector->startTimer(); 6390 } 6391 6392 /////////////////////////////////////////////////////////// 6393 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6394 // MarkRefsIntoAndScanClosure 6395 /////////////////////////////////////////////////////////// 6396 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6397 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6398 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6399 _span(span), 6400 _bit_map(bit_map), 6401 _work_queue(work_queue), 6402 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6403 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6404 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6405 { 6406 _ref_processor = rp; 6407 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6408 } 6409 6410 // This closure is used to mark refs into the CMS generation at the 6411 // second (final) checkpoint, and to scan and transitively follow 6412 // the unmarked oops. The marks are made in the marking bit map and 6413 // the work_queue is used for keeping the (newly) grey objects during 6414 // the scan phase whence they are also available for stealing by parallel 6415 // threads. Since the marking bit map is shared, updates are 6416 // synchronized (via CAS). 6417 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6418 if (obj != NULL) { 6419 // Ignore mark word because this could be an already marked oop 6420 // that may be chained at the end of the overflow list. 6421 assert(obj->is_oop(true), "expected an oop"); 6422 HeapWord* addr = (HeapWord*)obj; 6423 if (_span.contains(addr) && 6424 !_bit_map->isMarked(addr)) { 6425 // mark bit map (object will become grey): 6426 // It is possible for several threads to be 6427 // trying to "claim" this object concurrently; 6428 // the unique thread that succeeds in marking the 6429 // object first will do the subsequent push on 6430 // to the work queue (or overflow list). 6431 if (_bit_map->par_mark(addr)) { 6432 // push on work_queue (which may not be empty), and trim the 6433 // queue to an appropriate length by applying this closure to 6434 // the oops in the oops popped from the stack (i.e. blacken the 6435 // grey objects) 6436 bool res = _work_queue->push(obj); 6437 assert(res, "Low water mark should be less than capacity?"); 6438 trim_queue(_low_water_mark); 6439 } // Else, another thread claimed the object 6440 } 6441 } 6442 } 6443 6444 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6445 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6446 6447 // This closure is used to rescan the marked objects on the dirty cards 6448 // in the mod union table and the card table proper. 6449 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6450 oop p, MemRegion mr) { 6451 6452 size_t size = 0; 6453 HeapWord* addr = (HeapWord*)p; 6454 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6455 assert(_span.contains(addr), "we are scanning the CMS generation"); 6456 // check if it's time to yield 6457 if (do_yield_check()) { 6458 // We yielded for some foreground stop-world work, 6459 // and we have been asked to abort this ongoing preclean cycle. 6460 return 0; 6461 } 6462 if (_bitMap->isMarked(addr)) { 6463 // it's marked; is it potentially uninitialized? 6464 if (p->klass_or_null() != NULL) { 6465 // an initialized object; ignore mark word in verification below 6466 // since we are running concurrent with mutators 6467 assert(p->is_oop(true), "should be an oop"); 6468 if (p->is_objArray()) { 6469 // objArrays are precisely marked; restrict scanning 6470 // to dirty cards only. 6471 size = CompactibleFreeListSpace::adjustObjectSize( 6472 p->oop_iterate(_scanningClosure, mr)); 6473 } else { 6474 // A non-array may have been imprecisely marked; we need 6475 // to scan object in its entirety. 6476 size = CompactibleFreeListSpace::adjustObjectSize( 6477 p->oop_iterate(_scanningClosure)); 6478 } 6479 #ifdef ASSERT 6480 size_t direct_size = 6481 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6482 assert(size == direct_size, "Inconsistency in size"); 6483 assert(size >= 3, "Necessary for Printezis marks to work"); 6484 if (!_bitMap->isMarked(addr+1)) { 6485 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6486 } else { 6487 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6488 assert(_bitMap->isMarked(addr+size-1), 6489 "inconsistent Printezis mark"); 6490 } 6491 #endif // ASSERT 6492 } else { 6493 // An uninitialized object. 6494 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6495 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6496 size = pointer_delta(nextOneAddr + 1, addr); 6497 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6498 "alignment problem"); 6499 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6500 // will dirty the card when the klass pointer is installed in the 6501 // object (signaling the completion of initialization). 6502 } 6503 } else { 6504 // Either a not yet marked object or an uninitialized object 6505 if (p->klass_or_null() == NULL) { 6506 // An uninitialized object, skip to the next card, since 6507 // we may not be able to read its P-bits yet. 6508 assert(size == 0, "Initial value"); 6509 } else { 6510 // An object not (yet) reached by marking: we merely need to 6511 // compute its size so as to go look at the next block. 6512 assert(p->is_oop(true), "should be an oop"); 6513 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6514 } 6515 } 6516 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6517 return size; 6518 } 6519 6520 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6522 "CMS thread should hold CMS token"); 6523 assert_lock_strong(_freelistLock); 6524 assert_lock_strong(_bitMap->lock()); 6525 // relinquish the free_list_lock and bitMaplock() 6526 _bitMap->lock()->unlock(); 6527 _freelistLock->unlock(); 6528 ConcurrentMarkSweepThread::desynchronize(true); 6529 _collector->stopTimer(); 6530 if (PrintCMSStatistics != 0) { 6531 _collector->incrementYields(); 6532 } 6533 6534 // See the comment in coordinator_yield() 6535 for (unsigned i = 0; i < CMSYieldSleepCount && 6536 ConcurrentMarkSweepThread::should_yield() && 6537 !CMSCollector::foregroundGCIsActive(); ++i) { 6538 os::sleep(Thread::current(), 1, false); 6539 } 6540 6541 ConcurrentMarkSweepThread::synchronize(true); 6542 _freelistLock->lock_without_safepoint_check(); 6543 _bitMap->lock()->lock_without_safepoint_check(); 6544 _collector->startTimer(); 6545 } 6546 6547 6548 ////////////////////////////////////////////////////////////////// 6549 // SurvivorSpacePrecleanClosure 6550 ////////////////////////////////////////////////////////////////// 6551 // This (single-threaded) closure is used to preclean the oops in 6552 // the survivor spaces. 6553 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6554 6555 HeapWord* addr = (HeapWord*)p; 6556 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6557 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6558 assert(p->klass_or_null() != NULL, "object should be initialized"); 6559 // an initialized object; ignore mark word in verification below 6560 // since we are running concurrent with mutators 6561 assert(p->is_oop(true), "should be an oop"); 6562 // Note that we do not yield while we iterate over 6563 // the interior oops of p, pushing the relevant ones 6564 // on our marking stack. 6565 size_t size = p->oop_iterate(_scanning_closure); 6566 do_yield_check(); 6567 // Observe that below, we do not abandon the preclean 6568 // phase as soon as we should; rather we empty the 6569 // marking stack before returning. This is to satisfy 6570 // some existing assertions. In general, it may be a 6571 // good idea to abort immediately and complete the marking 6572 // from the grey objects at a later time. 6573 while (!_mark_stack->isEmpty()) { 6574 oop new_oop = _mark_stack->pop(); 6575 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6576 assert(_bit_map->isMarked((HeapWord*)new_oop), 6577 "only grey objects on this stack"); 6578 // iterate over the oops in this oop, marking and pushing 6579 // the ones in CMS heap (i.e. in _span). 6580 new_oop->oop_iterate(_scanning_closure); 6581 // check if it's time to yield 6582 do_yield_check(); 6583 } 6584 unsigned int after_count = 6585 GenCollectedHeap::heap()->total_collections(); 6586 bool abort = (_before_count != after_count) || 6587 _collector->should_abort_preclean(); 6588 return abort ? 0 : size; 6589 } 6590 6591 void SurvivorSpacePrecleanClosure::do_yield_work() { 6592 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6593 "CMS thread should hold CMS token"); 6594 assert_lock_strong(_bit_map->lock()); 6595 // Relinquish the bit map lock 6596 _bit_map->lock()->unlock(); 6597 ConcurrentMarkSweepThread::desynchronize(true); 6598 _collector->stopTimer(); 6599 if (PrintCMSStatistics != 0) { 6600 _collector->incrementYields(); 6601 } 6602 6603 // See the comment in coordinator_yield() 6604 for (unsigned i = 0; i < CMSYieldSleepCount && 6605 ConcurrentMarkSweepThread::should_yield() && 6606 !CMSCollector::foregroundGCIsActive(); ++i) { 6607 os::sleep(Thread::current(), 1, false); 6608 } 6609 6610 ConcurrentMarkSweepThread::synchronize(true); 6611 _bit_map->lock()->lock_without_safepoint_check(); 6612 _collector->startTimer(); 6613 } 6614 6615 // This closure is used to rescan the marked objects on the dirty cards 6616 // in the mod union table and the card table proper. In the parallel 6617 // case, although the bitMap is shared, we do a single read so the 6618 // isMarked() query is "safe". 6619 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6620 // Ignore mark word because we are running concurrent with mutators 6621 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6622 HeapWord* addr = (HeapWord*)p; 6623 assert(_span.contains(addr), "we are scanning the CMS generation"); 6624 bool is_obj_array = false; 6625 #ifdef ASSERT 6626 if (!_parallel) { 6627 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6628 assert(_collector->overflow_list_is_empty(), 6629 "overflow list should be empty"); 6630 6631 } 6632 #endif // ASSERT 6633 if (_bit_map->isMarked(addr)) { 6634 // Obj arrays are precisely marked, non-arrays are not; 6635 // so we scan objArrays precisely and non-arrays in their 6636 // entirety. 6637 if (p->is_objArray()) { 6638 is_obj_array = true; 6639 if (_parallel) { 6640 p->oop_iterate(_par_scan_closure, mr); 6641 } else { 6642 p->oop_iterate(_scan_closure, mr); 6643 } 6644 } else { 6645 if (_parallel) { 6646 p->oop_iterate(_par_scan_closure); 6647 } else { 6648 p->oop_iterate(_scan_closure); 6649 } 6650 } 6651 } 6652 #ifdef ASSERT 6653 if (!_parallel) { 6654 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6655 assert(_collector->overflow_list_is_empty(), 6656 "overflow list should be empty"); 6657 6658 } 6659 #endif // ASSERT 6660 return is_obj_array; 6661 } 6662 6663 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6664 MemRegion span, 6665 CMSBitMap* bitMap, CMSMarkStack* markStack, 6666 bool should_yield, bool verifying): 6667 _collector(collector), 6668 _span(span), 6669 _bitMap(bitMap), 6670 _mut(&collector->_modUnionTable), 6671 _markStack(markStack), 6672 _yield(should_yield), 6673 _skipBits(0) 6674 { 6675 assert(_markStack->isEmpty(), "stack should be empty"); 6676 _finger = _bitMap->startWord(); 6677 _threshold = _finger; 6678 assert(_collector->_restart_addr == NULL, "Sanity check"); 6679 assert(_span.contains(_finger), "Out of bounds _finger?"); 6680 DEBUG_ONLY(_verifying = verifying;) 6681 } 6682 6683 void MarkFromRootsClosure::reset(HeapWord* addr) { 6684 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6685 assert(_span.contains(addr), "Out of bounds _finger?"); 6686 _finger = addr; 6687 _threshold = (HeapWord*)round_to( 6688 (intptr_t)_finger, CardTableModRefBS::card_size); 6689 } 6690 6691 // Should revisit to see if this should be restructured for 6692 // greater efficiency. 6693 bool MarkFromRootsClosure::do_bit(size_t offset) { 6694 if (_skipBits > 0) { 6695 _skipBits--; 6696 return true; 6697 } 6698 // convert offset into a HeapWord* 6699 HeapWord* addr = _bitMap->startWord() + offset; 6700 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6701 "address out of range"); 6702 assert(_bitMap->isMarked(addr), "tautology"); 6703 if (_bitMap->isMarked(addr+1)) { 6704 // this is an allocated but not yet initialized object 6705 assert(_skipBits == 0, "tautology"); 6706 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6707 oop p = oop(addr); 6708 if (p->klass_or_null() == NULL) { 6709 DEBUG_ONLY(if (!_verifying) {) 6710 // We re-dirty the cards on which this object lies and increase 6711 // the _threshold so that we'll come back to scan this object 6712 // during the preclean or remark phase. (CMSCleanOnEnter) 6713 if (CMSCleanOnEnter) { 6714 size_t sz = _collector->block_size_using_printezis_bits(addr); 6715 HeapWord* end_card_addr = (HeapWord*)round_to( 6716 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6717 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6718 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6719 // Bump _threshold to end_card_addr; note that 6720 // _threshold cannot possibly exceed end_card_addr, anyhow. 6721 // This prevents future clearing of the card as the scan proceeds 6722 // to the right. 6723 assert(_threshold <= end_card_addr, 6724 "Because we are just scanning into this object"); 6725 if (_threshold < end_card_addr) { 6726 _threshold = end_card_addr; 6727 } 6728 if (p->klass_or_null() != NULL) { 6729 // Redirty the range of cards... 6730 _mut->mark_range(redirty_range); 6731 } // ...else the setting of klass will dirty the card anyway. 6732 } 6733 DEBUG_ONLY(}) 6734 return true; 6735 } 6736 } 6737 scanOopsInOop(addr); 6738 return true; 6739 } 6740 6741 // We take a break if we've been at this for a while, 6742 // so as to avoid monopolizing the locks involved. 6743 void MarkFromRootsClosure::do_yield_work() { 6744 // First give up the locks, then yield, then re-lock 6745 // We should probably use a constructor/destructor idiom to 6746 // do this unlock/lock or modify the MutexUnlocker class to 6747 // serve our purpose. XXX 6748 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6749 "CMS thread should hold CMS token"); 6750 assert_lock_strong(_bitMap->lock()); 6751 _bitMap->lock()->unlock(); 6752 ConcurrentMarkSweepThread::desynchronize(true); 6753 _collector->stopTimer(); 6754 if (PrintCMSStatistics != 0) { 6755 _collector->incrementYields(); 6756 } 6757 6758 // See the comment in coordinator_yield() 6759 for (unsigned i = 0; i < CMSYieldSleepCount && 6760 ConcurrentMarkSweepThread::should_yield() && 6761 !CMSCollector::foregroundGCIsActive(); ++i) { 6762 os::sleep(Thread::current(), 1, false); 6763 } 6764 6765 ConcurrentMarkSweepThread::synchronize(true); 6766 _bitMap->lock()->lock_without_safepoint_check(); 6767 _collector->startTimer(); 6768 } 6769 6770 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6771 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6772 assert(_markStack->isEmpty(), 6773 "should drain stack to limit stack usage"); 6774 // convert ptr to an oop preparatory to scanning 6775 oop obj = oop(ptr); 6776 // Ignore mark word in verification below, since we 6777 // may be running concurrent with mutators. 6778 assert(obj->is_oop(true), "should be an oop"); 6779 assert(_finger <= ptr, "_finger runneth ahead"); 6780 // advance the finger to right end of this object 6781 _finger = ptr + obj->size(); 6782 assert(_finger > ptr, "we just incremented it above"); 6783 // On large heaps, it may take us some time to get through 6784 // the marking phase. During 6785 // this time it's possible that a lot of mutations have 6786 // accumulated in the card table and the mod union table -- 6787 // these mutation records are redundant until we have 6788 // actually traced into the corresponding card. 6789 // Here, we check whether advancing the finger would make 6790 // us cross into a new card, and if so clear corresponding 6791 // cards in the MUT (preclean them in the card-table in the 6792 // future). 6793 6794 DEBUG_ONLY(if (!_verifying) {) 6795 // The clean-on-enter optimization is disabled by default, 6796 // until we fix 6178663. 6797 if (CMSCleanOnEnter && (_finger > _threshold)) { 6798 // [_threshold, _finger) represents the interval 6799 // of cards to be cleared in MUT (or precleaned in card table). 6800 // The set of cards to be cleared is all those that overlap 6801 // with the interval [_threshold, _finger); note that 6802 // _threshold is always kept card-aligned but _finger isn't 6803 // always card-aligned. 6804 HeapWord* old_threshold = _threshold; 6805 assert(old_threshold == (HeapWord*)round_to( 6806 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6807 "_threshold should always be card-aligned"); 6808 _threshold = (HeapWord*)round_to( 6809 (intptr_t)_finger, CardTableModRefBS::card_size); 6810 MemRegion mr(old_threshold, _threshold); 6811 assert(!mr.is_empty(), "Control point invariant"); 6812 assert(_span.contains(mr), "Should clear within span"); 6813 _mut->clear_range(mr); 6814 } 6815 DEBUG_ONLY(}) 6816 // Note: the finger doesn't advance while we drain 6817 // the stack below. 6818 PushOrMarkClosure pushOrMarkClosure(_collector, 6819 _span, _bitMap, _markStack, 6820 _finger, this); 6821 bool res = _markStack->push(obj); 6822 assert(res, "Empty non-zero size stack should have space for single push"); 6823 while (!_markStack->isEmpty()) { 6824 oop new_oop = _markStack->pop(); 6825 // Skip verifying header mark word below because we are 6826 // running concurrent with mutators. 6827 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6828 // now scan this oop's oops 6829 new_oop->oop_iterate(&pushOrMarkClosure); 6830 do_yield_check(); 6831 } 6832 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6833 } 6834 6835 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6836 CMSCollector* collector, MemRegion span, 6837 CMSBitMap* bit_map, 6838 OopTaskQueue* work_queue, 6839 CMSMarkStack* overflow_stack): 6840 _collector(collector), 6841 _whole_span(collector->_span), 6842 _span(span), 6843 _bit_map(bit_map), 6844 _mut(&collector->_modUnionTable), 6845 _work_queue(work_queue), 6846 _overflow_stack(overflow_stack), 6847 _skip_bits(0), 6848 _task(task) 6849 { 6850 assert(_work_queue->size() == 0, "work_queue should be empty"); 6851 _finger = span.start(); 6852 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6853 assert(_span.contains(_finger), "Out of bounds _finger?"); 6854 } 6855 6856 // Should revisit to see if this should be restructured for 6857 // greater efficiency. 6858 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6859 if (_skip_bits > 0) { 6860 _skip_bits--; 6861 return true; 6862 } 6863 // convert offset into a HeapWord* 6864 HeapWord* addr = _bit_map->startWord() + offset; 6865 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6866 "address out of range"); 6867 assert(_bit_map->isMarked(addr), "tautology"); 6868 if (_bit_map->isMarked(addr+1)) { 6869 // this is an allocated object that might not yet be initialized 6870 assert(_skip_bits == 0, "tautology"); 6871 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6872 oop p = oop(addr); 6873 if (p->klass_or_null() == NULL) { 6874 // in the case of Clean-on-Enter optimization, redirty card 6875 // and avoid clearing card by increasing the threshold. 6876 return true; 6877 } 6878 } 6879 scan_oops_in_oop(addr); 6880 return true; 6881 } 6882 6883 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6884 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6885 // Should we assert that our work queue is empty or 6886 // below some drain limit? 6887 assert(_work_queue->size() == 0, 6888 "should drain stack to limit stack usage"); 6889 // convert ptr to an oop preparatory to scanning 6890 oop obj = oop(ptr); 6891 // Ignore mark word in verification below, since we 6892 // may be running concurrent with mutators. 6893 assert(obj->is_oop(true), "should be an oop"); 6894 assert(_finger <= ptr, "_finger runneth ahead"); 6895 // advance the finger to right end of this object 6896 _finger = ptr + obj->size(); 6897 assert(_finger > ptr, "we just incremented it above"); 6898 // On large heaps, it may take us some time to get through 6899 // the marking phase. During 6900 // this time it's possible that a lot of mutations have 6901 // accumulated in the card table and the mod union table -- 6902 // these mutation records are redundant until we have 6903 // actually traced into the corresponding card. 6904 // Here, we check whether advancing the finger would make 6905 // us cross into a new card, and if so clear corresponding 6906 // cards in the MUT (preclean them in the card-table in the 6907 // future). 6908 6909 // The clean-on-enter optimization is disabled by default, 6910 // until we fix 6178663. 6911 if (CMSCleanOnEnter && (_finger > _threshold)) { 6912 // [_threshold, _finger) represents the interval 6913 // of cards to be cleared in MUT (or precleaned in card table). 6914 // The set of cards to be cleared is all those that overlap 6915 // with the interval [_threshold, _finger); note that 6916 // _threshold is always kept card-aligned but _finger isn't 6917 // always card-aligned. 6918 HeapWord* old_threshold = _threshold; 6919 assert(old_threshold == (HeapWord*)round_to( 6920 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6921 "_threshold should always be card-aligned"); 6922 _threshold = (HeapWord*)round_to( 6923 (intptr_t)_finger, CardTableModRefBS::card_size); 6924 MemRegion mr(old_threshold, _threshold); 6925 assert(!mr.is_empty(), "Control point invariant"); 6926 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6927 _mut->clear_range(mr); 6928 } 6929 6930 // Note: the local finger doesn't advance while we drain 6931 // the stack below, but the global finger sure can and will. 6932 HeapWord** gfa = _task->global_finger_addr(); 6933 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6934 _span, _bit_map, 6935 _work_queue, 6936 _overflow_stack, 6937 _finger, 6938 gfa, this); 6939 bool res = _work_queue->push(obj); // overflow could occur here 6940 assert(res, "Will hold once we use workqueues"); 6941 while (true) { 6942 oop new_oop; 6943 if (!_work_queue->pop_local(new_oop)) { 6944 // We emptied our work_queue; check if there's stuff that can 6945 // be gotten from the overflow stack. 6946 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6947 _overflow_stack, _work_queue)) { 6948 do_yield_check(); 6949 continue; 6950 } else { // done 6951 break; 6952 } 6953 } 6954 // Skip verifying header mark word below because we are 6955 // running concurrent with mutators. 6956 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6957 // now scan this oop's oops 6958 new_oop->oop_iterate(&pushOrMarkClosure); 6959 do_yield_check(); 6960 } 6961 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6962 } 6963 6964 // Yield in response to a request from VM Thread or 6965 // from mutators. 6966 void Par_MarkFromRootsClosure::do_yield_work() { 6967 assert(_task != NULL, "sanity"); 6968 _task->yield(); 6969 } 6970 6971 // A variant of the above used for verifying CMS marking work. 6972 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6973 MemRegion span, 6974 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6975 CMSMarkStack* mark_stack): 6976 _collector(collector), 6977 _span(span), 6978 _verification_bm(verification_bm), 6979 _cms_bm(cms_bm), 6980 _mark_stack(mark_stack), 6981 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6982 mark_stack) 6983 { 6984 assert(_mark_stack->isEmpty(), "stack should be empty"); 6985 _finger = _verification_bm->startWord(); 6986 assert(_collector->_restart_addr == NULL, "Sanity check"); 6987 assert(_span.contains(_finger), "Out of bounds _finger?"); 6988 } 6989 6990 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6991 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6992 assert(_span.contains(addr), "Out of bounds _finger?"); 6993 _finger = addr; 6994 } 6995 6996 // Should revisit to see if this should be restructured for 6997 // greater efficiency. 6998 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6999 // convert offset into a HeapWord* 7000 HeapWord* addr = _verification_bm->startWord() + offset; 7001 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7002 "address out of range"); 7003 assert(_verification_bm->isMarked(addr), "tautology"); 7004 assert(_cms_bm->isMarked(addr), "tautology"); 7005 7006 assert(_mark_stack->isEmpty(), 7007 "should drain stack to limit stack usage"); 7008 // convert addr to an oop preparatory to scanning 7009 oop obj = oop(addr); 7010 assert(obj->is_oop(), "should be an oop"); 7011 assert(_finger <= addr, "_finger runneth ahead"); 7012 // advance the finger to right end of this object 7013 _finger = addr + obj->size(); 7014 assert(_finger > addr, "we just incremented it above"); 7015 // Note: the finger doesn't advance while we drain 7016 // the stack below. 7017 bool res = _mark_stack->push(obj); 7018 assert(res, "Empty non-zero size stack should have space for single push"); 7019 while (!_mark_stack->isEmpty()) { 7020 oop new_oop = _mark_stack->pop(); 7021 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7022 // now scan this oop's oops 7023 new_oop->oop_iterate(&_pam_verify_closure); 7024 } 7025 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7026 return true; 7027 } 7028 7029 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7030 CMSCollector* collector, MemRegion span, 7031 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7032 CMSMarkStack* mark_stack): 7033 MetadataAwareOopClosure(collector->ref_processor()), 7034 _collector(collector), 7035 _span(span), 7036 _verification_bm(verification_bm), 7037 _cms_bm(cms_bm), 7038 _mark_stack(mark_stack) 7039 { } 7040 7041 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7042 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7043 7044 // Upon stack overflow, we discard (part of) the stack, 7045 // remembering the least address amongst those discarded 7046 // in CMSCollector's _restart_address. 7047 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7048 // Remember the least grey address discarded 7049 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7050 _collector->lower_restart_addr(ra); 7051 _mark_stack->reset(); // discard stack contents 7052 _mark_stack->expand(); // expand the stack if possible 7053 } 7054 7055 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7056 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7057 HeapWord* addr = (HeapWord*)obj; 7058 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7059 // Oop lies in _span and isn't yet grey or black 7060 _verification_bm->mark(addr); // now grey 7061 if (!_cms_bm->isMarked(addr)) { 7062 oop(addr)->print(); 7063 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7064 addr); 7065 fatal("... aborting"); 7066 } 7067 7068 if (!_mark_stack->push(obj)) { // stack overflow 7069 if (PrintCMSStatistics != 0) { 7070 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7071 SIZE_FORMAT, _mark_stack->capacity()); 7072 } 7073 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7074 handle_stack_overflow(addr); 7075 } 7076 // anything including and to the right of _finger 7077 // will be scanned as we iterate over the remainder of the 7078 // bit map 7079 } 7080 } 7081 7082 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7083 MemRegion span, 7084 CMSBitMap* bitMap, CMSMarkStack* markStack, 7085 HeapWord* finger, MarkFromRootsClosure* parent) : 7086 MetadataAwareOopClosure(collector->ref_processor()), 7087 _collector(collector), 7088 _span(span), 7089 _bitMap(bitMap), 7090 _markStack(markStack), 7091 _finger(finger), 7092 _parent(parent) 7093 { } 7094 7095 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7096 MemRegion span, 7097 CMSBitMap* bit_map, 7098 OopTaskQueue* work_queue, 7099 CMSMarkStack* overflow_stack, 7100 HeapWord* finger, 7101 HeapWord** global_finger_addr, 7102 Par_MarkFromRootsClosure* parent) : 7103 MetadataAwareOopClosure(collector->ref_processor()), 7104 _collector(collector), 7105 _whole_span(collector->_span), 7106 _span(span), 7107 _bit_map(bit_map), 7108 _work_queue(work_queue), 7109 _overflow_stack(overflow_stack), 7110 _finger(finger), 7111 _global_finger_addr(global_finger_addr), 7112 _parent(parent) 7113 { } 7114 7115 // Assumes thread-safe access by callers, who are 7116 // responsible for mutual exclusion. 7117 void CMSCollector::lower_restart_addr(HeapWord* low) { 7118 assert(_span.contains(low), "Out of bounds addr"); 7119 if (_restart_addr == NULL) { 7120 _restart_addr = low; 7121 } else { 7122 _restart_addr = MIN2(_restart_addr, low); 7123 } 7124 } 7125 7126 // Upon stack overflow, we discard (part of) the stack, 7127 // remembering the least address amongst those discarded 7128 // in CMSCollector's _restart_address. 7129 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7130 // Remember the least grey address discarded 7131 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7132 _collector->lower_restart_addr(ra); 7133 _markStack->reset(); // discard stack contents 7134 _markStack->expand(); // expand the stack if possible 7135 } 7136 7137 // Upon stack overflow, we discard (part of) the stack, 7138 // remembering the least address amongst those discarded 7139 // in CMSCollector's _restart_address. 7140 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7141 // We need to do this under a mutex to prevent other 7142 // workers from interfering with the work done below. 7143 MutexLockerEx ml(_overflow_stack->par_lock(), 7144 Mutex::_no_safepoint_check_flag); 7145 // Remember the least grey address discarded 7146 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7147 _collector->lower_restart_addr(ra); 7148 _overflow_stack->reset(); // discard stack contents 7149 _overflow_stack->expand(); // expand the stack if possible 7150 } 7151 7152 void PushOrMarkClosure::do_oop(oop obj) { 7153 // Ignore mark word because we are running concurrent with mutators. 7154 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7155 HeapWord* addr = (HeapWord*)obj; 7156 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7157 // Oop lies in _span and isn't yet grey or black 7158 _bitMap->mark(addr); // now grey 7159 if (addr < _finger) { 7160 // the bit map iteration has already either passed, or 7161 // sampled, this bit in the bit map; we'll need to 7162 // use the marking stack to scan this oop's oops. 7163 bool simulate_overflow = false; 7164 NOT_PRODUCT( 7165 if (CMSMarkStackOverflowALot && 7166 _collector->simulate_overflow()) { 7167 // simulate a stack overflow 7168 simulate_overflow = true; 7169 } 7170 ) 7171 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7172 if (PrintCMSStatistics != 0) { 7173 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7174 SIZE_FORMAT, _markStack->capacity()); 7175 } 7176 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7177 handle_stack_overflow(addr); 7178 } 7179 } 7180 // anything including and to the right of _finger 7181 // will be scanned as we iterate over the remainder of the 7182 // bit map 7183 do_yield_check(); 7184 } 7185 } 7186 7187 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7188 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7189 7190 void Par_PushOrMarkClosure::do_oop(oop obj) { 7191 // Ignore mark word because we are running concurrent with mutators. 7192 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7193 HeapWord* addr = (HeapWord*)obj; 7194 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7195 // Oop lies in _span and isn't yet grey or black 7196 // We read the global_finger (volatile read) strictly after marking oop 7197 bool res = _bit_map->par_mark(addr); // now grey 7198 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7199 // Should we push this marked oop on our stack? 7200 // -- if someone else marked it, nothing to do 7201 // -- if target oop is above global finger nothing to do 7202 // -- if target oop is in chunk and above local finger 7203 // then nothing to do 7204 // -- else push on work queue 7205 if ( !res // someone else marked it, they will deal with it 7206 || (addr >= *gfa) // will be scanned in a later task 7207 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7208 return; 7209 } 7210 // the bit map iteration has already either passed, or 7211 // sampled, this bit in the bit map; we'll need to 7212 // use the marking stack to scan this oop's oops. 7213 bool simulate_overflow = false; 7214 NOT_PRODUCT( 7215 if (CMSMarkStackOverflowALot && 7216 _collector->simulate_overflow()) { 7217 // simulate a stack overflow 7218 simulate_overflow = true; 7219 } 7220 ) 7221 if (simulate_overflow || 7222 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7223 // stack overflow 7224 if (PrintCMSStatistics != 0) { 7225 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7226 SIZE_FORMAT, _overflow_stack->capacity()); 7227 } 7228 // We cannot assert that the overflow stack is full because 7229 // it may have been emptied since. 7230 assert(simulate_overflow || 7231 _work_queue->size() == _work_queue->max_elems(), 7232 "Else push should have succeeded"); 7233 handle_stack_overflow(addr); 7234 } 7235 do_yield_check(); 7236 } 7237 } 7238 7239 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7240 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7241 7242 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7243 MemRegion span, 7244 ReferenceProcessor* rp, 7245 CMSBitMap* bit_map, 7246 CMSBitMap* mod_union_table, 7247 CMSMarkStack* mark_stack, 7248 bool concurrent_precleaning): 7249 MetadataAwareOopClosure(rp), 7250 _collector(collector), 7251 _span(span), 7252 _bit_map(bit_map), 7253 _mod_union_table(mod_union_table), 7254 _mark_stack(mark_stack), 7255 _concurrent_precleaning(concurrent_precleaning) 7256 { 7257 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7258 } 7259 7260 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7261 // the non-parallel version (the parallel version appears further below.) 7262 void PushAndMarkClosure::do_oop(oop obj) { 7263 // Ignore mark word verification. If during concurrent precleaning, 7264 // the object monitor may be locked. If during the checkpoint 7265 // phases, the object may already have been reached by a different 7266 // path and may be at the end of the global overflow list (so 7267 // the mark word may be NULL). 7268 assert(obj->is_oop_or_null(true /* ignore mark word */), 7269 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7270 HeapWord* addr = (HeapWord*)obj; 7271 // Check if oop points into the CMS generation 7272 // and is not marked 7273 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7274 // a white object ... 7275 _bit_map->mark(addr); // ... now grey 7276 // push on the marking stack (grey set) 7277 bool simulate_overflow = false; 7278 NOT_PRODUCT( 7279 if (CMSMarkStackOverflowALot && 7280 _collector->simulate_overflow()) { 7281 // simulate a stack overflow 7282 simulate_overflow = true; 7283 } 7284 ) 7285 if (simulate_overflow || !_mark_stack->push(obj)) { 7286 if (_concurrent_precleaning) { 7287 // During precleaning we can just dirty the appropriate card(s) 7288 // in the mod union table, thus ensuring that the object remains 7289 // in the grey set and continue. In the case of object arrays 7290 // we need to dirty all of the cards that the object spans, 7291 // since the rescan of object arrays will be limited to the 7292 // dirty cards. 7293 // Note that no one can be interfering with us in this action 7294 // of dirtying the mod union table, so no locking or atomics 7295 // are required. 7296 if (obj->is_objArray()) { 7297 size_t sz = obj->size(); 7298 HeapWord* end_card_addr = (HeapWord*)round_to( 7299 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7300 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7301 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7302 _mod_union_table->mark_range(redirty_range); 7303 } else { 7304 _mod_union_table->mark(addr); 7305 } 7306 _collector->_ser_pmc_preclean_ovflw++; 7307 } else { 7308 // During the remark phase, we need to remember this oop 7309 // in the overflow list. 7310 _collector->push_on_overflow_list(obj); 7311 _collector->_ser_pmc_remark_ovflw++; 7312 } 7313 } 7314 } 7315 } 7316 7317 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7318 MemRegion span, 7319 ReferenceProcessor* rp, 7320 CMSBitMap* bit_map, 7321 OopTaskQueue* work_queue): 7322 MetadataAwareOopClosure(rp), 7323 _collector(collector), 7324 _span(span), 7325 _bit_map(bit_map), 7326 _work_queue(work_queue) 7327 { 7328 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7329 } 7330 7331 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7332 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7333 7334 // Grey object rescan during second checkpoint phase -- 7335 // the parallel version. 7336 void Par_PushAndMarkClosure::do_oop(oop obj) { 7337 // In the assert below, we ignore the mark word because 7338 // this oop may point to an already visited object that is 7339 // on the overflow stack (in which case the mark word has 7340 // been hijacked for chaining into the overflow stack -- 7341 // if this is the last object in the overflow stack then 7342 // its mark word will be NULL). Because this object may 7343 // have been subsequently popped off the global overflow 7344 // stack, and the mark word possibly restored to the prototypical 7345 // value, by the time we get to examined this failing assert in 7346 // the debugger, is_oop_or_null(false) may subsequently start 7347 // to hold. 7348 assert(obj->is_oop_or_null(true), 7349 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7350 HeapWord* addr = (HeapWord*)obj; 7351 // Check if oop points into the CMS generation 7352 // and is not marked 7353 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7354 // a white object ... 7355 // If we manage to "claim" the object, by being the 7356 // first thread to mark it, then we push it on our 7357 // marking stack 7358 if (_bit_map->par_mark(addr)) { // ... now grey 7359 // push on work queue (grey set) 7360 bool simulate_overflow = false; 7361 NOT_PRODUCT( 7362 if (CMSMarkStackOverflowALot && 7363 _collector->par_simulate_overflow()) { 7364 // simulate a stack overflow 7365 simulate_overflow = true; 7366 } 7367 ) 7368 if (simulate_overflow || !_work_queue->push(obj)) { 7369 _collector->par_push_on_overflow_list(obj); 7370 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7371 } 7372 } // Else, some other thread got there first 7373 } 7374 } 7375 7376 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7377 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7378 7379 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7380 Mutex* bml = _collector->bitMapLock(); 7381 assert_lock_strong(bml); 7382 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7383 "CMS thread should hold CMS token"); 7384 7385 bml->unlock(); 7386 ConcurrentMarkSweepThread::desynchronize(true); 7387 7388 _collector->stopTimer(); 7389 if (PrintCMSStatistics != 0) { 7390 _collector->incrementYields(); 7391 } 7392 7393 // See the comment in coordinator_yield() 7394 for (unsigned i = 0; i < CMSYieldSleepCount && 7395 ConcurrentMarkSweepThread::should_yield() && 7396 !CMSCollector::foregroundGCIsActive(); ++i) { 7397 os::sleep(Thread::current(), 1, false); 7398 } 7399 7400 ConcurrentMarkSweepThread::synchronize(true); 7401 bml->lock(); 7402 7403 _collector->startTimer(); 7404 } 7405 7406 bool CMSPrecleanRefsYieldClosure::should_return() { 7407 if (ConcurrentMarkSweepThread::should_yield()) { 7408 do_yield_work(); 7409 } 7410 return _collector->foregroundGCIsActive(); 7411 } 7412 7413 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7414 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7415 "mr should be aligned to start at a card boundary"); 7416 // We'd like to assert: 7417 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7418 // "mr should be a range of cards"); 7419 // However, that would be too strong in one case -- the last 7420 // partition ends at _unallocated_block which, in general, can be 7421 // an arbitrary boundary, not necessarily card aligned. 7422 if (PrintCMSStatistics != 0) { 7423 _num_dirty_cards += 7424 mr.word_size()/CardTableModRefBS::card_size_in_words; 7425 } 7426 _space->object_iterate_mem(mr, &_scan_cl); 7427 } 7428 7429 SweepClosure::SweepClosure(CMSCollector* collector, 7430 ConcurrentMarkSweepGeneration* g, 7431 CMSBitMap* bitMap, bool should_yield) : 7432 _collector(collector), 7433 _g(g), 7434 _sp(g->cmsSpace()), 7435 _limit(_sp->sweep_limit()), 7436 _freelistLock(_sp->freelistLock()), 7437 _bitMap(bitMap), 7438 _yield(should_yield), 7439 _inFreeRange(false), // No free range at beginning of sweep 7440 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7441 _lastFreeRangeCoalesced(false), 7442 _freeFinger(g->used_region().start()) 7443 { 7444 NOT_PRODUCT( 7445 _numObjectsFreed = 0; 7446 _numWordsFreed = 0; 7447 _numObjectsLive = 0; 7448 _numWordsLive = 0; 7449 _numObjectsAlreadyFree = 0; 7450 _numWordsAlreadyFree = 0; 7451 _last_fc = NULL; 7452 7453 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7454 _sp->dictionary()->initialize_dict_returned_bytes(); 7455 ) 7456 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7457 "sweep _limit out of bounds"); 7458 if (CMSTraceSweeper) { 7459 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7460 _limit); 7461 } 7462 } 7463 7464 void SweepClosure::print_on(outputStream* st) const { 7465 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7466 _sp->bottom(), _sp->end()); 7467 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7468 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7469 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7470 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7471 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7472 } 7473 7474 #ifndef PRODUCT 7475 // Assertion checking only: no useful work in product mode -- 7476 // however, if any of the flags below become product flags, 7477 // you may need to review this code to see if it needs to be 7478 // enabled in product mode. 7479 SweepClosure::~SweepClosure() { 7480 assert_lock_strong(_freelistLock); 7481 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7482 "sweep _limit out of bounds"); 7483 if (inFreeRange()) { 7484 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7485 print(); 7486 ShouldNotReachHere(); 7487 } 7488 if (Verbose && PrintGC) { 7489 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7490 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7491 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7492 SIZE_FORMAT" bytes " 7493 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7494 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7495 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7496 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7497 * sizeof(HeapWord); 7498 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7499 7500 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7501 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7502 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7503 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7504 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7505 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7506 indexListReturnedBytes); 7507 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7508 dict_returned_bytes); 7509 } 7510 } 7511 if (CMSTraceSweeper) { 7512 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7513 _limit); 7514 } 7515 } 7516 #endif // PRODUCT 7517 7518 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7519 bool freeRangeInFreeLists) { 7520 if (CMSTraceSweeper) { 7521 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7522 freeFinger, freeRangeInFreeLists); 7523 } 7524 assert(!inFreeRange(), "Trampling existing free range"); 7525 set_inFreeRange(true); 7526 set_lastFreeRangeCoalesced(false); 7527 7528 set_freeFinger(freeFinger); 7529 set_freeRangeInFreeLists(freeRangeInFreeLists); 7530 if (CMSTestInFreeList) { 7531 if (freeRangeInFreeLists) { 7532 FreeChunk* fc = (FreeChunk*) freeFinger; 7533 assert(fc->is_free(), "A chunk on the free list should be free."); 7534 assert(fc->size() > 0, "Free range should have a size"); 7535 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7536 } 7537 } 7538 } 7539 7540 // Note that the sweeper runs concurrently with mutators. Thus, 7541 // it is possible for direct allocation in this generation to happen 7542 // in the middle of the sweep. Note that the sweeper also coalesces 7543 // contiguous free blocks. Thus, unless the sweeper and the allocator 7544 // synchronize appropriately freshly allocated blocks may get swept up. 7545 // This is accomplished by the sweeper locking the free lists while 7546 // it is sweeping. Thus blocks that are determined to be free are 7547 // indeed free. There is however one additional complication: 7548 // blocks that have been allocated since the final checkpoint and 7549 // mark, will not have been marked and so would be treated as 7550 // unreachable and swept up. To prevent this, the allocator marks 7551 // the bit map when allocating during the sweep phase. This leads, 7552 // however, to a further complication -- objects may have been allocated 7553 // but not yet initialized -- in the sense that the header isn't yet 7554 // installed. The sweeper can not then determine the size of the block 7555 // in order to skip over it. To deal with this case, we use a technique 7556 // (due to Printezis) to encode such uninitialized block sizes in the 7557 // bit map. Since the bit map uses a bit per every HeapWord, but the 7558 // CMS generation has a minimum object size of 3 HeapWords, it follows 7559 // that "normal marks" won't be adjacent in the bit map (there will 7560 // always be at least two 0 bits between successive 1 bits). We make use 7561 // of these "unused" bits to represent uninitialized blocks -- the bit 7562 // corresponding to the start of the uninitialized object and the next 7563 // bit are both set. Finally, a 1 bit marks the end of the object that 7564 // started with the two consecutive 1 bits to indicate its potentially 7565 // uninitialized state. 7566 7567 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7568 FreeChunk* fc = (FreeChunk*)addr; 7569 size_t res; 7570 7571 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7572 // than "addr == _limit" because although _limit was a block boundary when 7573 // we started the sweep, it may no longer be one because heap expansion 7574 // may have caused us to coalesce the block ending at the address _limit 7575 // with a newly expanded chunk (this happens when _limit was set to the 7576 // previous _end of the space), so we may have stepped past _limit: 7577 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7578 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7579 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7580 "sweep _limit out of bounds"); 7581 assert(addr < _sp->end(), "addr out of bounds"); 7582 // Flush any free range we might be holding as a single 7583 // coalesced chunk to the appropriate free list. 7584 if (inFreeRange()) { 7585 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7586 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 7587 flush_cur_free_chunk(freeFinger(), 7588 pointer_delta(addr, freeFinger())); 7589 if (CMSTraceSweeper) { 7590 gclog_or_tty->print("Sweep: last chunk: "); 7591 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7592 "[coalesced:%d]\n", 7593 freeFinger(), pointer_delta(addr, freeFinger()), 7594 lastFreeRangeCoalesced() ? 1 : 0); 7595 } 7596 } 7597 7598 // help the iterator loop finish 7599 return pointer_delta(_sp->end(), addr); 7600 } 7601 7602 assert(addr < _limit, "sweep invariant"); 7603 // check if we should yield 7604 do_yield_check(addr); 7605 if (fc->is_free()) { 7606 // Chunk that is already free 7607 res = fc->size(); 7608 do_already_free_chunk(fc); 7609 debug_only(_sp->verifyFreeLists()); 7610 // If we flush the chunk at hand in lookahead_and_flush() 7611 // and it's coalesced with a preceding chunk, then the 7612 // process of "mangling" the payload of the coalesced block 7613 // will cause erasure of the size information from the 7614 // (erstwhile) header of all the coalesced blocks but the 7615 // first, so the first disjunct in the assert will not hold 7616 // in that specific case (in which case the second disjunct 7617 // will hold). 7618 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7619 "Otherwise the size info doesn't change at this step"); 7620 NOT_PRODUCT( 7621 _numObjectsAlreadyFree++; 7622 _numWordsAlreadyFree += res; 7623 ) 7624 NOT_PRODUCT(_last_fc = fc;) 7625 } else if (!_bitMap->isMarked(addr)) { 7626 // Chunk is fresh garbage 7627 res = do_garbage_chunk(fc); 7628 debug_only(_sp->verifyFreeLists()); 7629 NOT_PRODUCT( 7630 _numObjectsFreed++; 7631 _numWordsFreed += res; 7632 ) 7633 } else { 7634 // Chunk that is alive. 7635 res = do_live_chunk(fc); 7636 debug_only(_sp->verifyFreeLists()); 7637 NOT_PRODUCT( 7638 _numObjectsLive++; 7639 _numWordsLive += res; 7640 ) 7641 } 7642 return res; 7643 } 7644 7645 // For the smart allocation, record following 7646 // split deaths - a free chunk is removed from its free list because 7647 // it is being split into two or more chunks. 7648 // split birth - a free chunk is being added to its free list because 7649 // a larger free chunk has been split and resulted in this free chunk. 7650 // coal death - a free chunk is being removed from its free list because 7651 // it is being coalesced into a large free chunk. 7652 // coal birth - a free chunk is being added to its free list because 7653 // it was created when two or more free chunks where coalesced into 7654 // this free chunk. 7655 // 7656 // These statistics are used to determine the desired number of free 7657 // chunks of a given size. The desired number is chosen to be relative 7658 // to the end of a CMS sweep. The desired number at the end of a sweep 7659 // is the 7660 // count-at-end-of-previous-sweep (an amount that was enough) 7661 // - count-at-beginning-of-current-sweep (the excess) 7662 // + split-births (gains in this size during interval) 7663 // - split-deaths (demands on this size during interval) 7664 // where the interval is from the end of one sweep to the end of the 7665 // next. 7666 // 7667 // When sweeping the sweeper maintains an accumulated chunk which is 7668 // the chunk that is made up of chunks that have been coalesced. That 7669 // will be termed the left-hand chunk. A new chunk of garbage that 7670 // is being considered for coalescing will be referred to as the 7671 // right-hand chunk. 7672 // 7673 // When making a decision on whether to coalesce a right-hand chunk with 7674 // the current left-hand chunk, the current count vs. the desired count 7675 // of the left-hand chunk is considered. Also if the right-hand chunk 7676 // is near the large chunk at the end of the heap (see 7677 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7678 // left-hand chunk is coalesced. 7679 // 7680 // When making a decision about whether to split a chunk, the desired count 7681 // vs. the current count of the candidate to be split is also considered. 7682 // If the candidate is underpopulated (currently fewer chunks than desired) 7683 // a chunk of an overpopulated (currently more chunks than desired) size may 7684 // be chosen. The "hint" associated with a free list, if non-null, points 7685 // to a free list which may be overpopulated. 7686 // 7687 7688 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7689 const size_t size = fc->size(); 7690 // Chunks that cannot be coalesced are not in the 7691 // free lists. 7692 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7693 assert(_sp->verify_chunk_in_free_list(fc), 7694 "free chunk should be in free lists"); 7695 } 7696 // a chunk that is already free, should not have been 7697 // marked in the bit map 7698 HeapWord* const addr = (HeapWord*) fc; 7699 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7700 // Verify that the bit map has no bits marked between 7701 // addr and purported end of this block. 7702 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7703 7704 // Some chunks cannot be coalesced under any circumstances. 7705 // See the definition of cantCoalesce(). 7706 if (!fc->cantCoalesce()) { 7707 // This chunk can potentially be coalesced. 7708 if (_sp->adaptive_freelists()) { 7709 // All the work is done in 7710 do_post_free_or_garbage_chunk(fc, size); 7711 } else { // Not adaptive free lists 7712 // this is a free chunk that can potentially be coalesced by the sweeper; 7713 if (!inFreeRange()) { 7714 // if the next chunk is a free block that can't be coalesced 7715 // it doesn't make sense to remove this chunk from the free lists 7716 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7717 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7718 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7719 nextChunk->is_free() && // ... which is free... 7720 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7721 // nothing to do 7722 } else { 7723 // Potentially the start of a new free range: 7724 // Don't eagerly remove it from the free lists. 7725 // No need to remove it if it will just be put 7726 // back again. (Also from a pragmatic point of view 7727 // if it is a free block in a region that is beyond 7728 // any allocated blocks, an assertion will fail) 7729 // Remember the start of a free run. 7730 initialize_free_range(addr, true); 7731 // end - can coalesce with next chunk 7732 } 7733 } else { 7734 // the midst of a free range, we are coalescing 7735 print_free_block_coalesced(fc); 7736 if (CMSTraceSweeper) { 7737 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 7738 } 7739 // remove it from the free lists 7740 _sp->removeFreeChunkFromFreeLists(fc); 7741 set_lastFreeRangeCoalesced(true); 7742 // If the chunk is being coalesced and the current free range is 7743 // in the free lists, remove the current free range so that it 7744 // will be returned to the free lists in its entirety - all 7745 // the coalesced pieces included. 7746 if (freeRangeInFreeLists()) { 7747 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7748 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7749 "Size of free range is inconsistent with chunk size."); 7750 if (CMSTestInFreeList) { 7751 assert(_sp->verify_chunk_in_free_list(ffc), 7752 "free range is not in free lists"); 7753 } 7754 _sp->removeFreeChunkFromFreeLists(ffc); 7755 set_freeRangeInFreeLists(false); 7756 } 7757 } 7758 } 7759 // Note that if the chunk is not coalescable (the else arm 7760 // below), we unconditionally flush, without needing to do 7761 // a "lookahead," as we do below. 7762 if (inFreeRange()) lookahead_and_flush(fc, size); 7763 } else { 7764 // Code path common to both original and adaptive free lists. 7765 7766 // cant coalesce with previous block; this should be treated 7767 // as the end of a free run if any 7768 if (inFreeRange()) { 7769 // we kicked some butt; time to pick up the garbage 7770 assert(freeFinger() < addr, "freeFinger points too high"); 7771 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7772 } 7773 // else, nothing to do, just continue 7774 } 7775 } 7776 7777 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7778 // This is a chunk of garbage. It is not in any free list. 7779 // Add it to a free list or let it possibly be coalesced into 7780 // a larger chunk. 7781 HeapWord* const addr = (HeapWord*) fc; 7782 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7783 7784 if (_sp->adaptive_freelists()) { 7785 // Verify that the bit map has no bits marked between 7786 // addr and purported end of just dead object. 7787 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7788 7789 do_post_free_or_garbage_chunk(fc, size); 7790 } else { 7791 if (!inFreeRange()) { 7792 // start of a new free range 7793 assert(size > 0, "A free range should have a size"); 7794 initialize_free_range(addr, false); 7795 } else { 7796 // this will be swept up when we hit the end of the 7797 // free range 7798 if (CMSTraceSweeper) { 7799 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 7800 } 7801 // If the chunk is being coalesced and the current free range is 7802 // in the free lists, remove the current free range so that it 7803 // will be returned to the free lists in its entirety - all 7804 // the coalesced pieces included. 7805 if (freeRangeInFreeLists()) { 7806 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7807 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7808 "Size of free range is inconsistent with chunk size."); 7809 if (CMSTestInFreeList) { 7810 assert(_sp->verify_chunk_in_free_list(ffc), 7811 "free range is not in free lists"); 7812 } 7813 _sp->removeFreeChunkFromFreeLists(ffc); 7814 set_freeRangeInFreeLists(false); 7815 } 7816 set_lastFreeRangeCoalesced(true); 7817 } 7818 // this will be swept up when we hit the end of the free range 7819 7820 // Verify that the bit map has no bits marked between 7821 // addr and purported end of just dead object. 7822 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7823 } 7824 assert(_limit >= addr + size, 7825 "A freshly garbage chunk can't possibly straddle over _limit"); 7826 if (inFreeRange()) lookahead_and_flush(fc, size); 7827 return size; 7828 } 7829 7830 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7831 HeapWord* addr = (HeapWord*) fc; 7832 // The sweeper has just found a live object. Return any accumulated 7833 // left hand chunk to the free lists. 7834 if (inFreeRange()) { 7835 assert(freeFinger() < addr, "freeFinger points too high"); 7836 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7837 } 7838 7839 // This object is live: we'd normally expect this to be 7840 // an oop, and like to assert the following: 7841 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7842 // However, as we commented above, this may be an object whose 7843 // header hasn't yet been initialized. 7844 size_t size; 7845 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7846 if (_bitMap->isMarked(addr + 1)) { 7847 // Determine the size from the bit map, rather than trying to 7848 // compute it from the object header. 7849 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7850 size = pointer_delta(nextOneAddr + 1, addr); 7851 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7852 "alignment problem"); 7853 7854 #ifdef ASSERT 7855 if (oop(addr)->klass_or_null() != NULL) { 7856 // Ignore mark word because we are running concurrent with mutators 7857 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7858 assert(size == 7859 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7860 "P-mark and computed size do not agree"); 7861 } 7862 #endif 7863 7864 } else { 7865 // This should be an initialized object that's alive. 7866 assert(oop(addr)->klass_or_null() != NULL, 7867 "Should be an initialized object"); 7868 // Ignore mark word because we are running concurrent with mutators 7869 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7870 // Verify that the bit map has no bits marked between 7871 // addr and purported end of this block. 7872 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7873 assert(size >= 3, "Necessary for Printezis marks to work"); 7874 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7875 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7876 } 7877 return size; 7878 } 7879 7880 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7881 size_t chunkSize) { 7882 // do_post_free_or_garbage_chunk() should only be called in the case 7883 // of the adaptive free list allocator. 7884 const bool fcInFreeLists = fc->is_free(); 7885 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7886 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7887 if (CMSTestInFreeList && fcInFreeLists) { 7888 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7889 } 7890 7891 if (CMSTraceSweeper) { 7892 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 7893 } 7894 7895 HeapWord* const fc_addr = (HeapWord*) fc; 7896 7897 bool coalesce; 7898 const size_t left = pointer_delta(fc_addr, freeFinger()); 7899 const size_t right = chunkSize; 7900 switch (FLSCoalescePolicy) { 7901 // numeric value forms a coalition aggressiveness metric 7902 case 0: { // never coalesce 7903 coalesce = false; 7904 break; 7905 } 7906 case 1: { // coalesce if left & right chunks on overpopulated lists 7907 coalesce = _sp->coalOverPopulated(left) && 7908 _sp->coalOverPopulated(right); 7909 break; 7910 } 7911 case 2: { // coalesce if left chunk on overpopulated list (default) 7912 coalesce = _sp->coalOverPopulated(left); 7913 break; 7914 } 7915 case 3: { // coalesce if left OR right chunk on overpopulated list 7916 coalesce = _sp->coalOverPopulated(left) || 7917 _sp->coalOverPopulated(right); 7918 break; 7919 } 7920 case 4: { // always coalesce 7921 coalesce = true; 7922 break; 7923 } 7924 default: 7925 ShouldNotReachHere(); 7926 } 7927 7928 // Should the current free range be coalesced? 7929 // If the chunk is in a free range and either we decided to coalesce above 7930 // or the chunk is near the large block at the end of the heap 7931 // (isNearLargestChunk() returns true), then coalesce this chunk. 7932 const bool doCoalesce = inFreeRange() 7933 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7934 if (doCoalesce) { 7935 // Coalesce the current free range on the left with the new 7936 // chunk on the right. If either is on a free list, 7937 // it must be removed from the list and stashed in the closure. 7938 if (freeRangeInFreeLists()) { 7939 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7940 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7941 "Size of free range is inconsistent with chunk size."); 7942 if (CMSTestInFreeList) { 7943 assert(_sp->verify_chunk_in_free_list(ffc), 7944 "Chunk is not in free lists"); 7945 } 7946 _sp->coalDeath(ffc->size()); 7947 _sp->removeFreeChunkFromFreeLists(ffc); 7948 set_freeRangeInFreeLists(false); 7949 } 7950 if (fcInFreeLists) { 7951 _sp->coalDeath(chunkSize); 7952 assert(fc->size() == chunkSize, 7953 "The chunk has the wrong size or is not in the free lists"); 7954 _sp->removeFreeChunkFromFreeLists(fc); 7955 } 7956 set_lastFreeRangeCoalesced(true); 7957 print_free_block_coalesced(fc); 7958 } else { // not in a free range and/or should not coalesce 7959 // Return the current free range and start a new one. 7960 if (inFreeRange()) { 7961 // In a free range but cannot coalesce with the right hand chunk. 7962 // Put the current free range into the free lists. 7963 flush_cur_free_chunk(freeFinger(), 7964 pointer_delta(fc_addr, freeFinger())); 7965 } 7966 // Set up for new free range. Pass along whether the right hand 7967 // chunk is in the free lists. 7968 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7969 } 7970 } 7971 7972 // Lookahead flush: 7973 // If we are tracking a free range, and this is the last chunk that 7974 // we'll look at because its end crosses past _limit, we'll preemptively 7975 // flush it along with any free range we may be holding on to. Note that 7976 // this can be the case only for an already free or freshly garbage 7977 // chunk. If this block is an object, it can never straddle 7978 // over _limit. The "straddling" occurs when _limit is set at 7979 // the previous end of the space when this cycle started, and 7980 // a subsequent heap expansion caused the previously co-terminal 7981 // free block to be coalesced with the newly expanded portion, 7982 // thus rendering _limit a non-block-boundary making it dangerous 7983 // for the sweeper to step over and examine. 7984 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7985 assert(inFreeRange(), "Should only be called if currently in a free range."); 7986 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7987 assert(_sp->used_region().contains(eob - 1), 7988 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7989 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7990 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7991 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 7992 if (eob >= _limit) { 7993 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7994 if (CMSTraceSweeper) { 7995 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7996 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7997 "[" PTR_FORMAT "," PTR_FORMAT ")", 7998 _limit, fc, eob, _sp->bottom(), _sp->end()); 7999 } 8000 // Return the storage we are tracking back into the free lists. 8001 if (CMSTraceSweeper) { 8002 gclog_or_tty->print_cr("Flushing ... "); 8003 } 8004 assert(freeFinger() < eob, "Error"); 8005 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8006 } 8007 } 8008 8009 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8010 assert(inFreeRange(), "Should only be called if currently in a free range."); 8011 assert(size > 0, 8012 "A zero sized chunk cannot be added to the free lists."); 8013 if (!freeRangeInFreeLists()) { 8014 if (CMSTestInFreeList) { 8015 FreeChunk* fc = (FreeChunk*) chunk; 8016 fc->set_size(size); 8017 assert(!_sp->verify_chunk_in_free_list(fc), 8018 "chunk should not be in free lists yet"); 8019 } 8020 if (CMSTraceSweeper) { 8021 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8022 chunk, size); 8023 } 8024 // A new free range is going to be starting. The current 8025 // free range has not been added to the free lists yet or 8026 // was removed so add it back. 8027 // If the current free range was coalesced, then the death 8028 // of the free range was recorded. Record a birth now. 8029 if (lastFreeRangeCoalesced()) { 8030 _sp->coalBirth(size); 8031 } 8032 _sp->addChunkAndRepairOffsetTable(chunk, size, 8033 lastFreeRangeCoalesced()); 8034 } else if (CMSTraceSweeper) { 8035 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8036 } 8037 set_inFreeRange(false); 8038 set_freeRangeInFreeLists(false); 8039 } 8040 8041 // We take a break if we've been at this for a while, 8042 // so as to avoid monopolizing the locks involved. 8043 void SweepClosure::do_yield_work(HeapWord* addr) { 8044 // Return current free chunk being used for coalescing (if any) 8045 // to the appropriate freelist. After yielding, the next 8046 // free block encountered will start a coalescing range of 8047 // free blocks. If the next free block is adjacent to the 8048 // chunk just flushed, they will need to wait for the next 8049 // sweep to be coalesced. 8050 if (inFreeRange()) { 8051 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8052 } 8053 8054 // First give up the locks, then yield, then re-lock. 8055 // We should probably use a constructor/destructor idiom to 8056 // do this unlock/lock or modify the MutexUnlocker class to 8057 // serve our purpose. XXX 8058 assert_lock_strong(_bitMap->lock()); 8059 assert_lock_strong(_freelistLock); 8060 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8061 "CMS thread should hold CMS token"); 8062 _bitMap->lock()->unlock(); 8063 _freelistLock->unlock(); 8064 ConcurrentMarkSweepThread::desynchronize(true); 8065 _collector->stopTimer(); 8066 if (PrintCMSStatistics != 0) { 8067 _collector->incrementYields(); 8068 } 8069 8070 // See the comment in coordinator_yield() 8071 for (unsigned i = 0; i < CMSYieldSleepCount && 8072 ConcurrentMarkSweepThread::should_yield() && 8073 !CMSCollector::foregroundGCIsActive(); ++i) { 8074 os::sleep(Thread::current(), 1, false); 8075 } 8076 8077 ConcurrentMarkSweepThread::synchronize(true); 8078 _freelistLock->lock(); 8079 _bitMap->lock()->lock_without_safepoint_check(); 8080 _collector->startTimer(); 8081 } 8082 8083 #ifndef PRODUCT 8084 // This is actually very useful in a product build if it can 8085 // be called from the debugger. Compile it into the product 8086 // as needed. 8087 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8088 return debug_cms_space->verify_chunk_in_free_list(fc); 8089 } 8090 #endif 8091 8092 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8093 if (CMSTraceSweeper) { 8094 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8095 fc, fc->size()); 8096 } 8097 } 8098 8099 // CMSIsAliveClosure 8100 bool CMSIsAliveClosure::do_object_b(oop obj) { 8101 HeapWord* addr = (HeapWord*)obj; 8102 return addr != NULL && 8103 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8104 } 8105 8106 8107 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8108 MemRegion span, 8109 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8110 bool cpc): 8111 _collector(collector), 8112 _span(span), 8113 _bit_map(bit_map), 8114 _mark_stack(mark_stack), 8115 _concurrent_precleaning(cpc) { 8116 assert(!_span.is_empty(), "Empty span could spell trouble"); 8117 } 8118 8119 8120 // CMSKeepAliveClosure: the serial version 8121 void CMSKeepAliveClosure::do_oop(oop obj) { 8122 HeapWord* addr = (HeapWord*)obj; 8123 if (_span.contains(addr) && 8124 !_bit_map->isMarked(addr)) { 8125 _bit_map->mark(addr); 8126 bool simulate_overflow = false; 8127 NOT_PRODUCT( 8128 if (CMSMarkStackOverflowALot && 8129 _collector->simulate_overflow()) { 8130 // simulate a stack overflow 8131 simulate_overflow = true; 8132 } 8133 ) 8134 if (simulate_overflow || !_mark_stack->push(obj)) { 8135 if (_concurrent_precleaning) { 8136 // We dirty the overflown object and let the remark 8137 // phase deal with it. 8138 assert(_collector->overflow_list_is_empty(), "Error"); 8139 // In the case of object arrays, we need to dirty all of 8140 // the cards that the object spans. No locking or atomics 8141 // are needed since no one else can be mutating the mod union 8142 // table. 8143 if (obj->is_objArray()) { 8144 size_t sz = obj->size(); 8145 HeapWord* end_card_addr = 8146 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8147 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8148 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8149 _collector->_modUnionTable.mark_range(redirty_range); 8150 } else { 8151 _collector->_modUnionTable.mark(addr); 8152 } 8153 _collector->_ser_kac_preclean_ovflw++; 8154 } else { 8155 _collector->push_on_overflow_list(obj); 8156 _collector->_ser_kac_ovflw++; 8157 } 8158 } 8159 } 8160 } 8161 8162 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8163 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8164 8165 // CMSParKeepAliveClosure: a parallel version of the above. 8166 // The work queues are private to each closure (thread), 8167 // but (may be) available for stealing by other threads. 8168 void CMSParKeepAliveClosure::do_oop(oop obj) { 8169 HeapWord* addr = (HeapWord*)obj; 8170 if (_span.contains(addr) && 8171 !_bit_map->isMarked(addr)) { 8172 // In general, during recursive tracing, several threads 8173 // may be concurrently getting here; the first one to 8174 // "tag" it, claims it. 8175 if (_bit_map->par_mark(addr)) { 8176 bool res = _work_queue->push(obj); 8177 assert(res, "Low water mark should be much less than capacity"); 8178 // Do a recursive trim in the hope that this will keep 8179 // stack usage lower, but leave some oops for potential stealers 8180 trim_queue(_low_water_mark); 8181 } // Else, another thread got there first 8182 } 8183 } 8184 8185 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8186 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8187 8188 void CMSParKeepAliveClosure::trim_queue(uint max) { 8189 while (_work_queue->size() > max) { 8190 oop new_oop; 8191 if (_work_queue->pop_local(new_oop)) { 8192 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8193 assert(_bit_map->isMarked((HeapWord*)new_oop), 8194 "no white objects on this stack!"); 8195 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8196 // iterate over the oops in this oop, marking and pushing 8197 // the ones in CMS heap (i.e. in _span). 8198 new_oop->oop_iterate(&_mark_and_push); 8199 } 8200 } 8201 } 8202 8203 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8204 CMSCollector* collector, 8205 MemRegion span, CMSBitMap* bit_map, 8206 OopTaskQueue* work_queue): 8207 _collector(collector), 8208 _span(span), 8209 _bit_map(bit_map), 8210 _work_queue(work_queue) { } 8211 8212 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8213 HeapWord* addr = (HeapWord*)obj; 8214 if (_span.contains(addr) && 8215 !_bit_map->isMarked(addr)) { 8216 if (_bit_map->par_mark(addr)) { 8217 bool simulate_overflow = false; 8218 NOT_PRODUCT( 8219 if (CMSMarkStackOverflowALot && 8220 _collector->par_simulate_overflow()) { 8221 // simulate a stack overflow 8222 simulate_overflow = true; 8223 } 8224 ) 8225 if (simulate_overflow || !_work_queue->push(obj)) { 8226 _collector->par_push_on_overflow_list(obj); 8227 _collector->_par_kac_ovflw++; 8228 } 8229 } // Else another thread got there already 8230 } 8231 } 8232 8233 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8234 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8235 8236 ////////////////////////////////////////////////////////////////// 8237 // CMSExpansionCause ///////////////////////////// 8238 ////////////////////////////////////////////////////////////////// 8239 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8240 switch (cause) { 8241 case _no_expansion: 8242 return "No expansion"; 8243 case _satisfy_free_ratio: 8244 return "Free ratio"; 8245 case _satisfy_promotion: 8246 return "Satisfy promotion"; 8247 case _satisfy_allocation: 8248 return "allocation"; 8249 case _allocate_par_lab: 8250 return "Par LAB"; 8251 case _allocate_par_spooling_space: 8252 return "Par Spooling Space"; 8253 case _adaptive_size_policy: 8254 return "Ergonomics"; 8255 default: 8256 return "unknown"; 8257 } 8258 } 8259 8260 void CMSDrainMarkingStackClosure::do_void() { 8261 // the max number to take from overflow list at a time 8262 const size_t num = _mark_stack->capacity()/4; 8263 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8264 "Overflow list should be NULL during concurrent phases"); 8265 while (!_mark_stack->isEmpty() || 8266 // if stack is empty, check the overflow list 8267 _collector->take_from_overflow_list(num, _mark_stack)) { 8268 oop obj = _mark_stack->pop(); 8269 HeapWord* addr = (HeapWord*)obj; 8270 assert(_span.contains(addr), "Should be within span"); 8271 assert(_bit_map->isMarked(addr), "Should be marked"); 8272 assert(obj->is_oop(), "Should be an oop"); 8273 obj->oop_iterate(_keep_alive); 8274 } 8275 } 8276 8277 void CMSParDrainMarkingStackClosure::do_void() { 8278 // drain queue 8279 trim_queue(0); 8280 } 8281 8282 // Trim our work_queue so its length is below max at return 8283 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8284 while (_work_queue->size() > max) { 8285 oop new_oop; 8286 if (_work_queue->pop_local(new_oop)) { 8287 assert(new_oop->is_oop(), "Expected an oop"); 8288 assert(_bit_map->isMarked((HeapWord*)new_oop), 8289 "no white objects on this stack!"); 8290 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8291 // iterate over the oops in this oop, marking and pushing 8292 // the ones in CMS heap (i.e. in _span). 8293 new_oop->oop_iterate(&_mark_and_push); 8294 } 8295 } 8296 } 8297 8298 //////////////////////////////////////////////////////////////////// 8299 // Support for Marking Stack Overflow list handling and related code 8300 //////////////////////////////////////////////////////////////////// 8301 // Much of the following code is similar in shape and spirit to the 8302 // code used in ParNewGC. We should try and share that code 8303 // as much as possible in the future. 8304 8305 #ifndef PRODUCT 8306 // Debugging support for CMSStackOverflowALot 8307 8308 // It's OK to call this multi-threaded; the worst thing 8309 // that can happen is that we'll get a bunch of closely 8310 // spaced simulated overflows, but that's OK, in fact 8311 // probably good as it would exercise the overflow code 8312 // under contention. 8313 bool CMSCollector::simulate_overflow() { 8314 if (_overflow_counter-- <= 0) { // just being defensive 8315 _overflow_counter = CMSMarkStackOverflowInterval; 8316 return true; 8317 } else { 8318 return false; 8319 } 8320 } 8321 8322 bool CMSCollector::par_simulate_overflow() { 8323 return simulate_overflow(); 8324 } 8325 #endif 8326 8327 // Single-threaded 8328 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8329 assert(stack->isEmpty(), "Expected precondition"); 8330 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8331 size_t i = num; 8332 oop cur = _overflow_list; 8333 const markOop proto = markOopDesc::prototype(); 8334 NOT_PRODUCT(ssize_t n = 0;) 8335 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8336 next = oop(cur->mark()); 8337 cur->set_mark(proto); // until proven otherwise 8338 assert(cur->is_oop(), "Should be an oop"); 8339 bool res = stack->push(cur); 8340 assert(res, "Bit off more than can chew?"); 8341 NOT_PRODUCT(n++;) 8342 } 8343 _overflow_list = cur; 8344 #ifndef PRODUCT 8345 assert(_num_par_pushes >= n, "Too many pops?"); 8346 _num_par_pushes -=n; 8347 #endif 8348 return !stack->isEmpty(); 8349 } 8350 8351 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8352 // (MT-safe) Get a prefix of at most "num" from the list. 8353 // The overflow list is chained through the mark word of 8354 // each object in the list. We fetch the entire list, 8355 // break off a prefix of the right size and return the 8356 // remainder. If other threads try to take objects from 8357 // the overflow list at that time, they will wait for 8358 // some time to see if data becomes available. If (and 8359 // only if) another thread places one or more object(s) 8360 // on the global list before we have returned the suffix 8361 // to the global list, we will walk down our local list 8362 // to find its end and append the global list to 8363 // our suffix before returning it. This suffix walk can 8364 // prove to be expensive (quadratic in the amount of traffic) 8365 // when there are many objects in the overflow list and 8366 // there is much producer-consumer contention on the list. 8367 // *NOTE*: The overflow list manipulation code here and 8368 // in ParNewGeneration:: are very similar in shape, 8369 // except that in the ParNew case we use the old (from/eden) 8370 // copy of the object to thread the list via its klass word. 8371 // Because of the common code, if you make any changes in 8372 // the code below, please check the ParNew version to see if 8373 // similar changes might be needed. 8374 // CR 6797058 has been filed to consolidate the common code. 8375 bool CMSCollector::par_take_from_overflow_list(size_t num, 8376 OopTaskQueue* work_q, 8377 int no_of_gc_threads) { 8378 assert(work_q->size() == 0, "First empty local work queue"); 8379 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8380 if (_overflow_list == NULL) { 8381 return false; 8382 } 8383 // Grab the entire list; we'll put back a suffix 8384 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8385 Thread* tid = Thread::current(); 8386 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8387 // set to ParallelGCThreads. 8388 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8389 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8390 // If the list is busy, we spin for a short while, 8391 // sleeping between attempts to get the list. 8392 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8393 os::sleep(tid, sleep_time_millis, false); 8394 if (_overflow_list == NULL) { 8395 // Nothing left to take 8396 return false; 8397 } else if (_overflow_list != BUSY) { 8398 // Try and grab the prefix 8399 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8400 } 8401 } 8402 // If the list was found to be empty, or we spun long 8403 // enough, we give up and return empty-handed. If we leave 8404 // the list in the BUSY state below, it must be the case that 8405 // some other thread holds the overflow list and will set it 8406 // to a non-BUSY state in the future. 8407 if (prefix == NULL || prefix == BUSY) { 8408 // Nothing to take or waited long enough 8409 if (prefix == NULL) { 8410 // Write back the NULL in case we overwrote it with BUSY above 8411 // and it is still the same value. 8412 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8413 } 8414 return false; 8415 } 8416 assert(prefix != NULL && prefix != BUSY, "Error"); 8417 size_t i = num; 8418 oop cur = prefix; 8419 // Walk down the first "num" objects, unless we reach the end. 8420 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8421 if (cur->mark() == NULL) { 8422 // We have "num" or fewer elements in the list, so there 8423 // is nothing to return to the global list. 8424 // Write back the NULL in lieu of the BUSY we wrote 8425 // above, if it is still the same value. 8426 if (_overflow_list == BUSY) { 8427 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8428 } 8429 } else { 8430 // Chop off the suffix and return it to the global list. 8431 assert(cur->mark() != BUSY, "Error"); 8432 oop suffix_head = cur->mark(); // suffix will be put back on global list 8433 cur->set_mark(NULL); // break off suffix 8434 // It's possible that the list is still in the empty(busy) state 8435 // we left it in a short while ago; in that case we may be 8436 // able to place back the suffix without incurring the cost 8437 // of a walk down the list. 8438 oop observed_overflow_list = _overflow_list; 8439 oop cur_overflow_list = observed_overflow_list; 8440 bool attached = false; 8441 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8442 observed_overflow_list = 8443 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8444 if (cur_overflow_list == observed_overflow_list) { 8445 attached = true; 8446 break; 8447 } else cur_overflow_list = observed_overflow_list; 8448 } 8449 if (!attached) { 8450 // Too bad, someone else sneaked in (at least) an element; we'll need 8451 // to do a splice. Find tail of suffix so we can prepend suffix to global 8452 // list. 8453 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8454 oop suffix_tail = cur; 8455 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8456 "Tautology"); 8457 observed_overflow_list = _overflow_list; 8458 do { 8459 cur_overflow_list = observed_overflow_list; 8460 if (cur_overflow_list != BUSY) { 8461 // Do the splice ... 8462 suffix_tail->set_mark(markOop(cur_overflow_list)); 8463 } else { // cur_overflow_list == BUSY 8464 suffix_tail->set_mark(NULL); 8465 } 8466 // ... and try to place spliced list back on overflow_list ... 8467 observed_overflow_list = 8468 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8469 } while (cur_overflow_list != observed_overflow_list); 8470 // ... until we have succeeded in doing so. 8471 } 8472 } 8473 8474 // Push the prefix elements on work_q 8475 assert(prefix != NULL, "control point invariant"); 8476 const markOop proto = markOopDesc::prototype(); 8477 oop next; 8478 NOT_PRODUCT(ssize_t n = 0;) 8479 for (cur = prefix; cur != NULL; cur = next) { 8480 next = oop(cur->mark()); 8481 cur->set_mark(proto); // until proven otherwise 8482 assert(cur->is_oop(), "Should be an oop"); 8483 bool res = work_q->push(cur); 8484 assert(res, "Bit off more than we can chew?"); 8485 NOT_PRODUCT(n++;) 8486 } 8487 #ifndef PRODUCT 8488 assert(_num_par_pushes >= n, "Too many pops?"); 8489 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8490 #endif 8491 return true; 8492 } 8493 8494 // Single-threaded 8495 void CMSCollector::push_on_overflow_list(oop p) { 8496 NOT_PRODUCT(_num_par_pushes++;) 8497 assert(p->is_oop(), "Not an oop"); 8498 preserve_mark_if_necessary(p); 8499 p->set_mark((markOop)_overflow_list); 8500 _overflow_list = p; 8501 } 8502 8503 // Multi-threaded; use CAS to prepend to overflow list 8504 void CMSCollector::par_push_on_overflow_list(oop p) { 8505 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8506 assert(p->is_oop(), "Not an oop"); 8507 par_preserve_mark_if_necessary(p); 8508 oop observed_overflow_list = _overflow_list; 8509 oop cur_overflow_list; 8510 do { 8511 cur_overflow_list = observed_overflow_list; 8512 if (cur_overflow_list != BUSY) { 8513 p->set_mark(markOop(cur_overflow_list)); 8514 } else { 8515 p->set_mark(NULL); 8516 } 8517 observed_overflow_list = 8518 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8519 } while (cur_overflow_list != observed_overflow_list); 8520 } 8521 #undef BUSY 8522 8523 // Single threaded 8524 // General Note on GrowableArray: pushes may silently fail 8525 // because we are (temporarily) out of C-heap for expanding 8526 // the stack. The problem is quite ubiquitous and affects 8527 // a lot of code in the JVM. The prudent thing for GrowableArray 8528 // to do (for now) is to exit with an error. However, that may 8529 // be too draconian in some cases because the caller may be 8530 // able to recover without much harm. For such cases, we 8531 // should probably introduce a "soft_push" method which returns 8532 // an indication of success or failure with the assumption that 8533 // the caller may be able to recover from a failure; code in 8534 // the VM can then be changed, incrementally, to deal with such 8535 // failures where possible, thus, incrementally hardening the VM 8536 // in such low resource situations. 8537 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8538 _preserved_oop_stack.push(p); 8539 _preserved_mark_stack.push(m); 8540 assert(m == p->mark(), "Mark word changed"); 8541 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8542 "bijection"); 8543 } 8544 8545 // Single threaded 8546 void CMSCollector::preserve_mark_if_necessary(oop p) { 8547 markOop m = p->mark(); 8548 if (m->must_be_preserved(p)) { 8549 preserve_mark_work(p, m); 8550 } 8551 } 8552 8553 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8554 markOop m = p->mark(); 8555 if (m->must_be_preserved(p)) { 8556 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8557 // Even though we read the mark word without holding 8558 // the lock, we are assured that it will not change 8559 // because we "own" this oop, so no other thread can 8560 // be trying to push it on the overflow list; see 8561 // the assertion in preserve_mark_work() that checks 8562 // that m == p->mark(). 8563 preserve_mark_work(p, m); 8564 } 8565 } 8566 8567 // We should be able to do this multi-threaded, 8568 // a chunk of stack being a task (this is 8569 // correct because each oop only ever appears 8570 // once in the overflow list. However, it's 8571 // not very easy to completely overlap this with 8572 // other operations, so will generally not be done 8573 // until all work's been completed. Because we 8574 // expect the preserved oop stack (set) to be small, 8575 // it's probably fine to do this single-threaded. 8576 // We can explore cleverer concurrent/overlapped/parallel 8577 // processing of preserved marks if we feel the 8578 // need for this in the future. Stack overflow should 8579 // be so rare in practice and, when it happens, its 8580 // effect on performance so great that this will 8581 // likely just be in the noise anyway. 8582 void CMSCollector::restore_preserved_marks_if_any() { 8583 assert(SafepointSynchronize::is_at_safepoint(), 8584 "world should be stopped"); 8585 assert(Thread::current()->is_ConcurrentGC_thread() || 8586 Thread::current()->is_VM_thread(), 8587 "should be single-threaded"); 8588 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8589 "bijection"); 8590 8591 while (!_preserved_oop_stack.is_empty()) { 8592 oop p = _preserved_oop_stack.pop(); 8593 assert(p->is_oop(), "Should be an oop"); 8594 assert(_span.contains(p), "oop should be in _span"); 8595 assert(p->mark() == markOopDesc::prototype(), 8596 "Set when taken from overflow list"); 8597 markOop m = _preserved_mark_stack.pop(); 8598 p->set_mark(m); 8599 } 8600 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8601 "stacks were cleared above"); 8602 } 8603 8604 #ifndef PRODUCT 8605 bool CMSCollector::no_preserved_marks() const { 8606 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8607 } 8608 #endif 8609 8610 // Transfer some number of overflown objects to usual marking 8611 // stack. Return true if some objects were transferred. 8612 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8613 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8614 (size_t)ParGCDesiredObjsFromOverflowList); 8615 8616 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8617 assert(_collector->overflow_list_is_empty() || res, 8618 "If list is not empty, we should have taken something"); 8619 assert(!res || !_mark_stack->isEmpty(), 8620 "If we took something, it should now be on our stack"); 8621 return res; 8622 } 8623 8624 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8625 size_t res = _sp->block_size_no_stall(addr, _collector); 8626 if (_sp->block_is_obj(addr)) { 8627 if (_live_bit_map->isMarked(addr)) { 8628 // It can't have been dead in a previous cycle 8629 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8630 } else { 8631 _dead_bit_map->mark(addr); // mark the dead object 8632 } 8633 } 8634 // Could be 0, if the block size could not be computed without stalling. 8635 return res; 8636 } 8637 8638 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8639 8640 switch (phase) { 8641 case CMSCollector::InitialMarking: 8642 initialize(true /* fullGC */ , 8643 cause /* cause of the GC */, 8644 true /* recordGCBeginTime */, 8645 true /* recordPreGCUsage */, 8646 false /* recordPeakUsage */, 8647 false /* recordPostGCusage */, 8648 true /* recordAccumulatedGCTime */, 8649 false /* recordGCEndTime */, 8650 false /* countCollection */ ); 8651 break; 8652 8653 case CMSCollector::FinalMarking: 8654 initialize(true /* fullGC */ , 8655 cause /* cause of the GC */, 8656 false /* recordGCBeginTime */, 8657 false /* recordPreGCUsage */, 8658 false /* recordPeakUsage */, 8659 false /* recordPostGCusage */, 8660 true /* recordAccumulatedGCTime */, 8661 false /* recordGCEndTime */, 8662 false /* countCollection */ ); 8663 break; 8664 8665 case CMSCollector::Sweeping: 8666 initialize(true /* fullGC */ , 8667 cause /* cause of the GC */, 8668 false /* recordGCBeginTime */, 8669 false /* recordPreGCUsage */, 8670 true /* recordPeakUsage */, 8671 true /* recordPostGCusage */, 8672 false /* recordAccumulatedGCTime */, 8673 true /* recordGCEndTime */, 8674 true /* countCollection */ ); 8675 break; 8676 8677 default: 8678 ShouldNotReachHere(); 8679 } 8680 }