1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardGeneration.inline.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/strongRootsScope.hpp" 57 #include "memory/tenuredGeneration.hpp" 58 #include "oops/oop.inline.hpp" 59 #include "prims/jvmtiExport.hpp" 60 #include "runtime/atomic.inline.hpp" 61 #include "runtime/globals_extension.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/orderAccess.inline.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "services/memoryService.hpp" 67 #include "services/runtimeService.hpp" 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->young_gen()->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 //verify that this lock should be acquired with safepoint check. 484 Monitor::_safepoint_check_sometimes)), 485 _eden_chunk_array(NULL), // may be set in ctor body 486 _eden_chunk_capacity(0), // -- ditto -- 487 _eden_chunk_index(0), // -- ditto -- 488 _survivor_plab_array(NULL), // -- ditto -- 489 _survivor_chunk_array(NULL), // -- ditto -- 490 _survivor_chunk_capacity(0), // -- ditto -- 491 _survivor_chunk_index(0), // -- ditto -- 492 _ser_pmc_preclean_ovflw(0), 493 _ser_kac_preclean_ovflw(0), 494 _ser_pmc_remark_ovflw(0), 495 _par_pmc_remark_ovflw(0), 496 _ser_kac_ovflw(0), 497 _par_kac_ovflw(0), 498 #ifndef PRODUCT 499 _num_par_pushes(0), 500 #endif 501 _collection_count_start(0), 502 _verifying(false), 503 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 504 _completed_initialization(false), 505 _collector_policy(cp), 506 _should_unload_classes(CMSClassUnloadingEnabled), 507 _concurrent_cycles_since_last_unload(0), 508 _roots_scanning_options(GenCollectedHeap::SO_None), 509 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 510 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 511 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 512 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 513 _cms_start_registered(false) 514 { 515 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 516 ExplicitGCInvokesConcurrent = true; 517 } 518 // Now expand the span and allocate the collection support structures 519 // (MUT, marking bit map etc.) to cover both generations subject to 520 // collection. 521 522 // For use by dirty card to oop closures. 523 _cmsGen->cmsSpace()->set_collector(this); 524 525 // Allocate MUT and marking bit map 526 { 527 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 528 if (!_markBitMap.allocate(_span)) { 529 warning("Failed to allocate CMS Bit Map"); 530 return; 531 } 532 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 533 } 534 { 535 _modUnionTable.allocate(_span); 536 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 537 } 538 539 if (!_markStack.allocate(MarkStackSize)) { 540 warning("Failed to allocate CMS Marking Stack"); 541 return; 542 } 543 544 // Support for multi-threaded concurrent phases 545 if (CMSConcurrentMTEnabled) { 546 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 547 // just for now 548 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 549 } 550 if (ConcGCThreads > 1) { 551 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 552 ConcGCThreads, true); 553 if (_conc_workers == NULL) { 554 warning("GC/CMS: _conc_workers allocation failure: " 555 "forcing -CMSConcurrentMTEnabled"); 556 CMSConcurrentMTEnabled = false; 557 } else { 558 _conc_workers->initialize_workers(); 559 } 560 } else { 561 CMSConcurrentMTEnabled = false; 562 } 563 } 564 if (!CMSConcurrentMTEnabled) { 565 ConcGCThreads = 0; 566 } else { 567 // Turn off CMSCleanOnEnter optimization temporarily for 568 // the MT case where it's not fixed yet; see 6178663. 569 CMSCleanOnEnter = false; 570 } 571 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 572 "Inconsistency"); 573 574 // Parallel task queues; these are shared for the 575 // concurrent and stop-world phases of CMS, but 576 // are not shared with parallel scavenge (ParNew). 577 { 578 uint i; 579 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 580 581 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 582 || ParallelRefProcEnabled) 583 && num_queues > 0) { 584 _task_queues = new OopTaskQueueSet(num_queues); 585 if (_task_queues == NULL) { 586 warning("task_queues allocation failure."); 587 return; 588 } 589 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 590 if (_hash_seed == NULL) { 591 warning("_hash_seed array allocation failure"); 592 return; 593 } 594 595 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 596 for (i = 0; i < num_queues; i++) { 597 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 598 if (q == NULL) { 599 warning("work_queue allocation failure."); 600 return; 601 } 602 _task_queues->register_queue(i, q); 603 } 604 for (i = 0; i < num_queues; i++) { 605 _task_queues->queue(i)->initialize(); 606 _hash_seed[i] = 17; // copied from ParNew 607 } 608 } 609 } 610 611 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 612 613 // Clip CMSBootstrapOccupancy between 0 and 100. 614 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 615 616 // Now tell CMS generations the identity of their collector 617 ConcurrentMarkSweepGeneration::set_collector(this); 618 619 // Create & start a CMS thread for this CMS collector 620 _cmsThread = ConcurrentMarkSweepThread::start(this); 621 assert(cmsThread() != NULL, "CMS Thread should have been created"); 622 assert(cmsThread()->collector() == this, 623 "CMS Thread should refer to this gen"); 624 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 625 626 // Support for parallelizing young gen rescan 627 GenCollectedHeap* gch = GenCollectedHeap::heap(); 628 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 629 _young_gen = (ParNewGeneration*)gch->young_gen(); 630 if (gch->supports_inline_contig_alloc()) { 631 _top_addr = gch->top_addr(); 632 _end_addr = gch->end_addr(); 633 assert(_young_gen != NULL, "no _young_gen"); 634 _eden_chunk_index = 0; 635 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 636 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 637 if (_eden_chunk_array == NULL) { 638 _eden_chunk_capacity = 0; 639 warning("GC/CMS: _eden_chunk_array allocation failure"); 640 } 641 } 642 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 643 644 // Support for parallelizing survivor space rescan 645 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 646 const size_t max_plab_samples = 647 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 648 649 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 650 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 651 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 652 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 653 || _cursor == NULL) { 654 warning("Failed to allocate survivor plab/chunk array"); 655 if (_survivor_plab_array != NULL) { 656 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 657 _survivor_plab_array = NULL; 658 } 659 if (_survivor_chunk_array != NULL) { 660 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 661 _survivor_chunk_array = NULL; 662 } 663 if (_cursor != NULL) { 664 FREE_C_HEAP_ARRAY(size_t, _cursor); 665 _cursor = NULL; 666 } 667 } else { 668 _survivor_chunk_capacity = 2*max_plab_samples; 669 for (uint i = 0; i < ParallelGCThreads; i++) { 670 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 671 if (vec == NULL) { 672 warning("Failed to allocate survivor plab array"); 673 for (int j = i; j > 0; j--) { 674 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 675 } 676 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 677 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 678 _survivor_plab_array = NULL; 679 _survivor_chunk_array = NULL; 680 _survivor_chunk_capacity = 0; 681 break; 682 } else { 683 ChunkArray* cur = 684 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 685 max_plab_samples); 686 assert(cur->end() == 0, "Should be 0"); 687 assert(cur->array() == vec, "Should be vec"); 688 assert(cur->capacity() == max_plab_samples, "Error"); 689 } 690 } 691 } 692 } 693 assert( ( _survivor_plab_array != NULL 694 && _survivor_chunk_array != NULL) 695 || ( _survivor_chunk_capacity == 0 696 && _survivor_chunk_index == 0), 697 "Error"); 698 699 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 700 _gc_counters = new CollectorCounters("CMS", 1); 701 _completed_initialization = true; 702 _inter_sweep_timer.start(); // start of time 703 } 704 705 size_t CMSCollector::plab_sample_minimum_size() { 706 // The default value of MinTLABSize is 2k, but there is 707 // no way to get the default value if the flag has been overridden. 708 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 709 } 710 711 const char* ConcurrentMarkSweepGeneration::name() const { 712 return "concurrent mark-sweep generation"; 713 } 714 void ConcurrentMarkSweepGeneration::update_counters() { 715 if (UsePerfData) { 716 _space_counters->update_all(); 717 _gen_counters->update_all(); 718 } 719 } 720 721 // this is an optimized version of update_counters(). it takes the 722 // used value as a parameter rather than computing it. 723 // 724 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 725 if (UsePerfData) { 726 _space_counters->update_used(used); 727 _space_counters->update_capacity(); 728 _gen_counters->update_all(); 729 } 730 } 731 732 void ConcurrentMarkSweepGeneration::print() const { 733 Generation::print(); 734 cmsSpace()->print(); 735 } 736 737 #ifndef PRODUCT 738 void ConcurrentMarkSweepGeneration::print_statistics() { 739 cmsSpace()->printFLCensus(0); 740 } 741 #endif 742 743 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 744 GenCollectedHeap* gch = GenCollectedHeap::heap(); 745 if (PrintGCDetails) { 746 if (Verbose) { 747 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 748 level(), short_name(), s, used(), capacity()); 749 } else { 750 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 751 level(), short_name(), s, used() / K, capacity() / K); 752 } 753 } 754 if (Verbose) { 755 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 756 gch->used(), gch->capacity()); 757 } else { 758 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 759 gch->used() / K, gch->capacity() / K); 760 } 761 } 762 763 size_t 764 ConcurrentMarkSweepGeneration::contiguous_available() const { 765 // dld proposes an improvement in precision here. If the committed 766 // part of the space ends in a free block we should add that to 767 // uncommitted size in the calculation below. Will make this 768 // change later, staying with the approximation below for the 769 // time being. -- ysr. 770 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 771 } 772 773 size_t 774 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 775 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 776 } 777 778 size_t ConcurrentMarkSweepGeneration::max_available() const { 779 return free() + _virtual_space.uncommitted_size(); 780 } 781 782 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 783 size_t available = max_available(); 784 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 785 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 786 if (Verbose && PrintGCDetails) { 787 gclog_or_tty->print_cr( 788 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 789 "max_promo("SIZE_FORMAT")", 790 res? "":" not", available, res? ">=":"<", 791 av_promo, max_promotion_in_bytes); 792 } 793 return res; 794 } 795 796 // At a promotion failure dump information on block layout in heap 797 // (cms old generation). 798 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 799 if (CMSDumpAtPromotionFailure) { 800 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 801 } 802 } 803 804 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 805 // Clear the promotion information. These pointers can be adjusted 806 // along with all the other pointers into the heap but 807 // compaction is expected to be a rare event with 808 // a heap using cms so don't do it without seeing the need. 809 for (uint i = 0; i < ParallelGCThreads; i++) { 810 _par_gc_thread_states[i]->promo.reset(); 811 } 812 } 813 814 void ConcurrentMarkSweepGeneration::compute_new_size() { 815 assert_locked_or_safepoint(Heap_lock); 816 817 // If incremental collection failed, we just want to expand 818 // to the limit. 819 if (incremental_collection_failed()) { 820 clear_incremental_collection_failed(); 821 grow_to_reserved(); 822 return; 823 } 824 825 // The heap has been compacted but not reset yet. 826 // Any metric such as free() or used() will be incorrect. 827 828 CardGeneration::compute_new_size(); 829 830 // Reset again after a possible resizing 831 if (did_compact()) { 832 cmsSpace()->reset_after_compaction(); 833 } 834 } 835 836 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 837 assert_locked_or_safepoint(Heap_lock); 838 839 // If incremental collection failed, we just want to expand 840 // to the limit. 841 if (incremental_collection_failed()) { 842 clear_incremental_collection_failed(); 843 grow_to_reserved(); 844 return; 845 } 846 847 double free_percentage = ((double) free()) / capacity(); 848 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 849 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 850 851 // compute expansion delta needed for reaching desired free percentage 852 if (free_percentage < desired_free_percentage) { 853 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 854 assert(desired_capacity >= capacity(), "invalid expansion size"); 855 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 856 if (PrintGCDetails && Verbose) { 857 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 858 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 859 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 860 gclog_or_tty->print_cr(" Desired free fraction %f", 861 desired_free_percentage); 862 gclog_or_tty->print_cr(" Maximum free fraction %f", 863 maximum_free_percentage); 864 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 865 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 866 desired_capacity/1000); 867 int prev_level = level() - 1; 868 if (prev_level >= 0) { 869 size_t prev_size = 0; 870 GenCollectedHeap* gch = GenCollectedHeap::heap(); 871 Generation* prev_gen = gch->young_gen(); 872 prev_size = prev_gen->capacity(); 873 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 874 prev_size/1000); 875 } 876 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 877 unsafe_max_alloc_nogc()/1000); 878 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 879 contiguous_available()/1000); 880 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 881 expand_bytes); 882 } 883 // safe if expansion fails 884 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 885 if (PrintGCDetails && Verbose) { 886 gclog_or_tty->print_cr(" Expanded free fraction %f", 887 ((double) free()) / capacity()); 888 } 889 } else { 890 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 891 assert(desired_capacity <= capacity(), "invalid expansion size"); 892 size_t shrink_bytes = capacity() - desired_capacity; 893 // Don't shrink unless the delta is greater than the minimum shrink we want 894 if (shrink_bytes >= MinHeapDeltaBytes) { 895 shrink_free_list_by(shrink_bytes); 896 } 897 } 898 } 899 900 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 901 return cmsSpace()->freelistLock(); 902 } 903 904 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 905 bool tlab) { 906 CMSSynchronousYieldRequest yr; 907 MutexLockerEx x(freelistLock(), 908 Mutex::_no_safepoint_check_flag); 909 return have_lock_and_allocate(size, tlab); 910 } 911 912 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 913 bool tlab /* ignored */) { 914 assert_lock_strong(freelistLock()); 915 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 916 HeapWord* res = cmsSpace()->allocate(adjustedSize); 917 // Allocate the object live (grey) if the background collector has 918 // started marking. This is necessary because the marker may 919 // have passed this address and consequently this object will 920 // not otherwise be greyed and would be incorrectly swept up. 921 // Note that if this object contains references, the writing 922 // of those references will dirty the card containing this object 923 // allowing the object to be blackened (and its references scanned) 924 // either during a preclean phase or at the final checkpoint. 925 if (res != NULL) { 926 // We may block here with an uninitialized object with 927 // its mark-bit or P-bits not yet set. Such objects need 928 // to be safely navigable by block_start(). 929 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 930 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 931 collector()->direct_allocated(res, adjustedSize); 932 _direct_allocated_words += adjustedSize; 933 // allocation counters 934 NOT_PRODUCT( 935 _numObjectsAllocated++; 936 _numWordsAllocated += (int)adjustedSize; 937 ) 938 } 939 return res; 940 } 941 942 // In the case of direct allocation by mutators in a generation that 943 // is being concurrently collected, the object must be allocated 944 // live (grey) if the background collector has started marking. 945 // This is necessary because the marker may 946 // have passed this address and consequently this object will 947 // not otherwise be greyed and would be incorrectly swept up. 948 // Note that if this object contains references, the writing 949 // of those references will dirty the card containing this object 950 // allowing the object to be blackened (and its references scanned) 951 // either during a preclean phase or at the final checkpoint. 952 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 953 assert(_markBitMap.covers(start, size), "Out of bounds"); 954 if (_collectorState >= Marking) { 955 MutexLockerEx y(_markBitMap.lock(), 956 Mutex::_no_safepoint_check_flag); 957 // [see comments preceding SweepClosure::do_blk() below for details] 958 // 959 // Can the P-bits be deleted now? JJJ 960 // 961 // 1. need to mark the object as live so it isn't collected 962 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 963 // 3. need to mark the end of the object so marking, precleaning or sweeping 964 // can skip over uninitialized or unparsable objects. An allocated 965 // object is considered uninitialized for our purposes as long as 966 // its klass word is NULL. All old gen objects are parsable 967 // as soon as they are initialized.) 968 _markBitMap.mark(start); // object is live 969 _markBitMap.mark(start + 1); // object is potentially uninitialized? 970 _markBitMap.mark(start + size - 1); 971 // mark end of object 972 } 973 // check that oop looks uninitialized 974 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 975 } 976 977 void CMSCollector::promoted(bool par, HeapWord* start, 978 bool is_obj_array, size_t obj_size) { 979 assert(_markBitMap.covers(start), "Out of bounds"); 980 // See comment in direct_allocated() about when objects should 981 // be allocated live. 982 if (_collectorState >= Marking) { 983 // we already hold the marking bit map lock, taken in 984 // the prologue 985 if (par) { 986 _markBitMap.par_mark(start); 987 } else { 988 _markBitMap.mark(start); 989 } 990 // We don't need to mark the object as uninitialized (as 991 // in direct_allocated above) because this is being done with the 992 // world stopped and the object will be initialized by the 993 // time the marking, precleaning or sweeping get to look at it. 994 // But see the code for copying objects into the CMS generation, 995 // where we need to ensure that concurrent readers of the 996 // block offset table are able to safely navigate a block that 997 // is in flux from being free to being allocated (and in 998 // transition while being copied into) and subsequently 999 // becoming a bona-fide object when the copy/promotion is complete. 1000 assert(SafepointSynchronize::is_at_safepoint(), 1001 "expect promotion only at safepoints"); 1002 1003 if (_collectorState < Sweeping) { 1004 // Mark the appropriate cards in the modUnionTable, so that 1005 // this object gets scanned before the sweep. If this is 1006 // not done, CMS generation references in the object might 1007 // not get marked. 1008 // For the case of arrays, which are otherwise precisely 1009 // marked, we need to dirty the entire array, not just its head. 1010 if (is_obj_array) { 1011 // The [par_]mark_range() method expects mr.end() below to 1012 // be aligned to the granularity of a bit's representation 1013 // in the heap. In the case of the MUT below, that's a 1014 // card size. 1015 MemRegion mr(start, 1016 (HeapWord*)round_to((intptr_t)(start + obj_size), 1017 CardTableModRefBS::card_size /* bytes */)); 1018 if (par) { 1019 _modUnionTable.par_mark_range(mr); 1020 } else { 1021 _modUnionTable.mark_range(mr); 1022 } 1023 } else { // not an obj array; we can just mark the head 1024 if (par) { 1025 _modUnionTable.par_mark(start); 1026 } else { 1027 _modUnionTable.mark(start); 1028 } 1029 } 1030 } 1031 } 1032 } 1033 1034 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1035 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1036 // allocate, copy and if necessary update promoinfo -- 1037 // delegate to underlying space. 1038 assert_lock_strong(freelistLock()); 1039 1040 #ifndef PRODUCT 1041 if (Universe::heap()->promotion_should_fail()) { 1042 return NULL; 1043 } 1044 #endif // #ifndef PRODUCT 1045 1046 oop res = _cmsSpace->promote(obj, obj_size); 1047 if (res == NULL) { 1048 // expand and retry 1049 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1050 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 1051 // Since this is the old generation, we don't try to promote 1052 // into a more senior generation. 1053 res = _cmsSpace->promote(obj, obj_size); 1054 } 1055 if (res != NULL) { 1056 // See comment in allocate() about when objects should 1057 // be allocated live. 1058 assert(obj->is_oop(), "Will dereference klass pointer below"); 1059 collector()->promoted(false, // Not parallel 1060 (HeapWord*)res, obj->is_objArray(), obj_size); 1061 // promotion counters 1062 NOT_PRODUCT( 1063 _numObjectsPromoted++; 1064 _numWordsPromoted += 1065 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1066 ) 1067 } 1068 return res; 1069 } 1070 1071 1072 // IMPORTANT: Notes on object size recognition in CMS. 1073 // --------------------------------------------------- 1074 // A block of storage in the CMS generation is always in 1075 // one of three states. A free block (FREE), an allocated 1076 // object (OBJECT) whose size() method reports the correct size, 1077 // and an intermediate state (TRANSIENT) in which its size cannot 1078 // be accurately determined. 1079 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1080 // ----------------------------------------------------- 1081 // FREE: klass_word & 1 == 1; mark_word holds block size 1082 // 1083 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1084 // obj->size() computes correct size 1085 // 1086 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1087 // 1088 // STATE IDENTIFICATION: (64 bit+COOPS) 1089 // ------------------------------------ 1090 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1091 // 1092 // OBJECT: klass_word installed; klass_word != 0; 1093 // obj->size() computes correct size 1094 // 1095 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1096 // 1097 // 1098 // STATE TRANSITION DIAGRAM 1099 // 1100 // mut / parnew mut / parnew 1101 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1102 // ^ | 1103 // |------------------------ DEAD <------------------------------------| 1104 // sweep mut 1105 // 1106 // While a block is in TRANSIENT state its size cannot be determined 1107 // so readers will either need to come back later or stall until 1108 // the size can be determined. Note that for the case of direct 1109 // allocation, P-bits, when available, may be used to determine the 1110 // size of an object that may not yet have been initialized. 1111 1112 // Things to support parallel young-gen collection. 1113 oop 1114 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1115 oop old, markOop m, 1116 size_t word_sz) { 1117 #ifndef PRODUCT 1118 if (Universe::heap()->promotion_should_fail()) { 1119 return NULL; 1120 } 1121 #endif // #ifndef PRODUCT 1122 1123 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1124 PromotionInfo* promoInfo = &ps->promo; 1125 // if we are tracking promotions, then first ensure space for 1126 // promotion (including spooling space for saving header if necessary). 1127 // then allocate and copy, then track promoted info if needed. 1128 // When tracking (see PromotionInfo::track()), the mark word may 1129 // be displaced and in this case restoration of the mark word 1130 // occurs in the (oop_since_save_marks_)iterate phase. 1131 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1132 // Out of space for allocating spooling buffers; 1133 // try expanding and allocating spooling buffers. 1134 if (!expand_and_ensure_spooling_space(promoInfo)) { 1135 return NULL; 1136 } 1137 } 1138 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1139 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1140 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1141 if (obj_ptr == NULL) { 1142 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1143 if (obj_ptr == NULL) { 1144 return NULL; 1145 } 1146 } 1147 oop obj = oop(obj_ptr); 1148 OrderAccess::storestore(); 1149 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1150 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1151 // IMPORTANT: See note on object initialization for CMS above. 1152 // Otherwise, copy the object. Here we must be careful to insert the 1153 // klass pointer last, since this marks the block as an allocated object. 1154 // Except with compressed oops it's the mark word. 1155 HeapWord* old_ptr = (HeapWord*)old; 1156 // Restore the mark word copied above. 1157 obj->set_mark(m); 1158 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1159 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1160 OrderAccess::storestore(); 1161 1162 if (UseCompressedClassPointers) { 1163 // Copy gap missed by (aligned) header size calculation below 1164 obj->set_klass_gap(old->klass_gap()); 1165 } 1166 if (word_sz > (size_t)oopDesc::header_size()) { 1167 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1168 obj_ptr + oopDesc::header_size(), 1169 word_sz - oopDesc::header_size()); 1170 } 1171 1172 // Now we can track the promoted object, if necessary. We take care 1173 // to delay the transition from uninitialized to full object 1174 // (i.e., insertion of klass pointer) until after, so that it 1175 // atomically becomes a promoted object. 1176 if (promoInfo->tracking()) { 1177 promoInfo->track((PromotedObject*)obj, old->klass()); 1178 } 1179 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1180 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1181 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1182 1183 // Finally, install the klass pointer (this should be volatile). 1184 OrderAccess::storestore(); 1185 obj->set_klass(old->klass()); 1186 // We should now be able to calculate the right size for this object 1187 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1188 1189 collector()->promoted(true, // parallel 1190 obj_ptr, old->is_objArray(), word_sz); 1191 1192 NOT_PRODUCT( 1193 Atomic::inc_ptr(&_numObjectsPromoted); 1194 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1195 ) 1196 1197 return obj; 1198 } 1199 1200 void 1201 ConcurrentMarkSweepGeneration:: 1202 par_promote_alloc_done(int thread_num) { 1203 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1204 ps->lab.retire(thread_num); 1205 } 1206 1207 void 1208 ConcurrentMarkSweepGeneration:: 1209 par_oop_since_save_marks_iterate_done(int thread_num) { 1210 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1211 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1212 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1213 } 1214 1215 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1216 size_t size, 1217 bool tlab) 1218 { 1219 // We allow a STW collection only if a full 1220 // collection was requested. 1221 return full || should_allocate(size, tlab); // FIX ME !!! 1222 // This and promotion failure handling are connected at the 1223 // hip and should be fixed by untying them. 1224 } 1225 1226 bool CMSCollector::shouldConcurrentCollect() { 1227 if (_full_gc_requested) { 1228 if (Verbose && PrintGCDetails) { 1229 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1230 " gc request (or gc_locker)"); 1231 } 1232 return true; 1233 } 1234 1235 FreelistLocker x(this); 1236 // ------------------------------------------------------------------ 1237 // Print out lots of information which affects the initiation of 1238 // a collection. 1239 if (PrintCMSInitiationStatistics && stats().valid()) { 1240 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1241 gclog_or_tty->stamp(); 1242 gclog_or_tty->cr(); 1243 stats().print_on(gclog_or_tty); 1244 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1245 stats().time_until_cms_gen_full()); 1246 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1247 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1248 _cmsGen->contiguous_available()); 1249 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1250 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1251 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1252 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1253 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1254 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1255 gclog_or_tty->print_cr("metadata initialized %d", 1256 MetaspaceGC::should_concurrent_collect()); 1257 } 1258 // ------------------------------------------------------------------ 1259 1260 // If the estimated time to complete a cms collection (cms_duration()) 1261 // is less than the estimated time remaining until the cms generation 1262 // is full, start a collection. 1263 if (!UseCMSInitiatingOccupancyOnly) { 1264 if (stats().valid()) { 1265 if (stats().time_until_cms_start() == 0.0) { 1266 return true; 1267 } 1268 } else { 1269 // We want to conservatively collect somewhat early in order 1270 // to try and "bootstrap" our CMS/promotion statistics; 1271 // this branch will not fire after the first successful CMS 1272 // collection because the stats should then be valid. 1273 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1274 if (Verbose && PrintGCDetails) { 1275 gclog_or_tty->print_cr( 1276 " CMSCollector: collect for bootstrapping statistics:" 1277 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1278 _bootstrap_occupancy); 1279 } 1280 return true; 1281 } 1282 } 1283 } 1284 1285 // Otherwise, we start a collection cycle if 1286 // old gen want a collection cycle started. Each may use 1287 // an appropriate criterion for making this decision. 1288 // XXX We need to make sure that the gen expansion 1289 // criterion dovetails well with this. XXX NEED TO FIX THIS 1290 if (_cmsGen->should_concurrent_collect()) { 1291 if (Verbose && PrintGCDetails) { 1292 gclog_or_tty->print_cr("CMS old gen initiated"); 1293 } 1294 return true; 1295 } 1296 1297 // We start a collection if we believe an incremental collection may fail; 1298 // this is not likely to be productive in practice because it's probably too 1299 // late anyway. 1300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1301 assert(gch->collector_policy()->is_generation_policy(), 1302 "You may want to check the correctness of the following"); 1303 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1304 if (Verbose && PrintGCDetails) { 1305 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1306 } 1307 return true; 1308 } 1309 1310 if (MetaspaceGC::should_concurrent_collect()) { 1311 if (Verbose && PrintGCDetails) { 1312 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1313 } 1314 return true; 1315 } 1316 1317 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1318 if (CMSTriggerInterval >= 0) { 1319 if (CMSTriggerInterval == 0) { 1320 // Trigger always 1321 return true; 1322 } 1323 1324 // Check the CMS time since begin (we do not check the stats validity 1325 // as we want to be able to trigger the first CMS cycle as well) 1326 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1327 if (Verbose && PrintGCDetails) { 1328 if (stats().valid()) { 1329 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1330 stats().cms_time_since_begin()); 1331 } else { 1332 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1333 } 1334 } 1335 return true; 1336 } 1337 } 1338 1339 return false; 1340 } 1341 1342 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1343 1344 // Clear _expansion_cause fields of constituent generations 1345 void CMSCollector::clear_expansion_cause() { 1346 _cmsGen->clear_expansion_cause(); 1347 } 1348 1349 // We should be conservative in starting a collection cycle. To 1350 // start too eagerly runs the risk of collecting too often in the 1351 // extreme. To collect too rarely falls back on full collections, 1352 // which works, even if not optimum in terms of concurrent work. 1353 // As a work around for too eagerly collecting, use the flag 1354 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1355 // giving the user an easily understandable way of controlling the 1356 // collections. 1357 // We want to start a new collection cycle if any of the following 1358 // conditions hold: 1359 // . our current occupancy exceeds the configured initiating occupancy 1360 // for this generation, or 1361 // . we recently needed to expand this space and have not, since that 1362 // expansion, done a collection of this generation, or 1363 // . the underlying space believes that it may be a good idea to initiate 1364 // a concurrent collection (this may be based on criteria such as the 1365 // following: the space uses linear allocation and linear allocation is 1366 // going to fail, or there is believed to be excessive fragmentation in 1367 // the generation, etc... or ... 1368 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1369 // the case of the old generation; see CR 6543076): 1370 // we may be approaching a point at which allocation requests may fail because 1371 // we will be out of sufficient free space given allocation rate estimates.] 1372 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1373 1374 assert_lock_strong(freelistLock()); 1375 if (occupancy() > initiating_occupancy()) { 1376 if (PrintGCDetails && Verbose) { 1377 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1378 short_name(), occupancy(), initiating_occupancy()); 1379 } 1380 return true; 1381 } 1382 if (UseCMSInitiatingOccupancyOnly) { 1383 return false; 1384 } 1385 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1386 if (PrintGCDetails && Verbose) { 1387 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1388 short_name()); 1389 } 1390 return true; 1391 } 1392 if (_cmsSpace->should_concurrent_collect()) { 1393 if (PrintGCDetails && Verbose) { 1394 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1395 short_name()); 1396 } 1397 return true; 1398 } 1399 return false; 1400 } 1401 1402 void ConcurrentMarkSweepGeneration::collect(bool full, 1403 bool clear_all_soft_refs, 1404 size_t size, 1405 bool tlab) 1406 { 1407 collector()->collect(full, clear_all_soft_refs, size, tlab); 1408 } 1409 1410 void CMSCollector::collect(bool full, 1411 bool clear_all_soft_refs, 1412 size_t size, 1413 bool tlab) 1414 { 1415 // The following "if" branch is present for defensive reasons. 1416 // In the current uses of this interface, it can be replaced with: 1417 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1418 // But I am not placing that assert here to allow future 1419 // generality in invoking this interface. 1420 if (GC_locker::is_active()) { 1421 // A consistency test for GC_locker 1422 assert(GC_locker::needs_gc(), "Should have been set already"); 1423 // Skip this foreground collection, instead 1424 // expanding the heap if necessary. 1425 // Need the free list locks for the call to free() in compute_new_size() 1426 compute_new_size(); 1427 return; 1428 } 1429 acquire_control_and_collect(full, clear_all_soft_refs); 1430 } 1431 1432 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1433 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1434 unsigned int gc_count = gch->total_full_collections(); 1435 if (gc_count == full_gc_count) { 1436 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1437 _full_gc_requested = true; 1438 _full_gc_cause = cause; 1439 CGC_lock->notify(); // nudge CMS thread 1440 } else { 1441 assert(gc_count > full_gc_count, "Error: causal loop"); 1442 } 1443 } 1444 1445 bool CMSCollector::is_external_interruption() { 1446 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1447 return GCCause::is_user_requested_gc(cause) || 1448 GCCause::is_serviceability_requested_gc(cause); 1449 } 1450 1451 void CMSCollector::report_concurrent_mode_interruption() { 1452 if (is_external_interruption()) { 1453 if (PrintGCDetails) { 1454 gclog_or_tty->print(" (concurrent mode interrupted)"); 1455 } 1456 } else { 1457 if (PrintGCDetails) { 1458 gclog_or_tty->print(" (concurrent mode failure)"); 1459 } 1460 _gc_tracer_cm->report_concurrent_mode_failure(); 1461 } 1462 } 1463 1464 1465 // The foreground and background collectors need to coordinate in order 1466 // to make sure that they do not mutually interfere with CMS collections. 1467 // When a background collection is active, 1468 // the foreground collector may need to take over (preempt) and 1469 // synchronously complete an ongoing collection. Depending on the 1470 // frequency of the background collections and the heap usage 1471 // of the application, this preemption can be seldom or frequent. 1472 // There are only certain 1473 // points in the background collection that the "collection-baton" 1474 // can be passed to the foreground collector. 1475 // 1476 // The foreground collector will wait for the baton before 1477 // starting any part of the collection. The foreground collector 1478 // will only wait at one location. 1479 // 1480 // The background collector will yield the baton before starting a new 1481 // phase of the collection (e.g., before initial marking, marking from roots, 1482 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1483 // of the loop which switches the phases. The background collector does some 1484 // of the phases (initial mark, final re-mark) with the world stopped. 1485 // Because of locking involved in stopping the world, 1486 // the foreground collector should not block waiting for the background 1487 // collector when it is doing a stop-the-world phase. The background 1488 // collector will yield the baton at an additional point just before 1489 // it enters a stop-the-world phase. Once the world is stopped, the 1490 // background collector checks the phase of the collection. If the 1491 // phase has not changed, it proceeds with the collection. If the 1492 // phase has changed, it skips that phase of the collection. See 1493 // the comments on the use of the Heap_lock in collect_in_background(). 1494 // 1495 // Variable used in baton passing. 1496 // _foregroundGCIsActive - Set to true by the foreground collector when 1497 // it wants the baton. The foreground clears it when it has finished 1498 // the collection. 1499 // _foregroundGCShouldWait - Set to true by the background collector 1500 // when it is running. The foreground collector waits while 1501 // _foregroundGCShouldWait is true. 1502 // CGC_lock - monitor used to protect access to the above variables 1503 // and to notify the foreground and background collectors. 1504 // _collectorState - current state of the CMS collection. 1505 // 1506 // The foreground collector 1507 // acquires the CGC_lock 1508 // sets _foregroundGCIsActive 1509 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1510 // various locks acquired in preparation for the collection 1511 // are released so as not to block the background collector 1512 // that is in the midst of a collection 1513 // proceeds with the collection 1514 // clears _foregroundGCIsActive 1515 // returns 1516 // 1517 // The background collector in a loop iterating on the phases of the 1518 // collection 1519 // acquires the CGC_lock 1520 // sets _foregroundGCShouldWait 1521 // if _foregroundGCIsActive is set 1522 // clears _foregroundGCShouldWait, notifies _CGC_lock 1523 // waits on _CGC_lock for _foregroundGCIsActive to become false 1524 // and exits the loop. 1525 // otherwise 1526 // proceed with that phase of the collection 1527 // if the phase is a stop-the-world phase, 1528 // yield the baton once more just before enqueueing 1529 // the stop-world CMS operation (executed by the VM thread). 1530 // returns after all phases of the collection are done 1531 // 1532 1533 void CMSCollector::acquire_control_and_collect(bool full, 1534 bool clear_all_soft_refs) { 1535 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1536 assert(!Thread::current()->is_ConcurrentGC_thread(), 1537 "shouldn't try to acquire control from self!"); 1538 1539 // Start the protocol for acquiring control of the 1540 // collection from the background collector (aka CMS thread). 1541 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1542 "VM thread should have CMS token"); 1543 // Remember the possibly interrupted state of an ongoing 1544 // concurrent collection 1545 CollectorState first_state = _collectorState; 1546 1547 // Signal to a possibly ongoing concurrent collection that 1548 // we want to do a foreground collection. 1549 _foregroundGCIsActive = true; 1550 1551 // release locks and wait for a notify from the background collector 1552 // releasing the locks in only necessary for phases which 1553 // do yields to improve the granularity of the collection. 1554 assert_lock_strong(bitMapLock()); 1555 // We need to lock the Free list lock for the space that we are 1556 // currently collecting. 1557 assert(haveFreelistLocks(), "Must be holding free list locks"); 1558 bitMapLock()->unlock(); 1559 releaseFreelistLocks(); 1560 { 1561 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1562 if (_foregroundGCShouldWait) { 1563 // We are going to be waiting for action for the CMS thread; 1564 // it had better not be gone (for instance at shutdown)! 1565 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1566 "CMS thread must be running"); 1567 // Wait here until the background collector gives us the go-ahead 1568 ConcurrentMarkSweepThread::clear_CMS_flag( 1569 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1570 // Get a possibly blocked CMS thread going: 1571 // Note that we set _foregroundGCIsActive true above, 1572 // without protection of the CGC_lock. 1573 CGC_lock->notify(); 1574 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1575 "Possible deadlock"); 1576 while (_foregroundGCShouldWait) { 1577 // wait for notification 1578 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1579 // Possibility of delay/starvation here, since CMS token does 1580 // not know to give priority to VM thread? Actually, i think 1581 // there wouldn't be any delay/starvation, but the proof of 1582 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1583 } 1584 ConcurrentMarkSweepThread::set_CMS_flag( 1585 ConcurrentMarkSweepThread::CMS_vm_has_token); 1586 } 1587 } 1588 // The CMS_token is already held. Get back the other locks. 1589 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1590 "VM thread should have CMS token"); 1591 getFreelistLocks(); 1592 bitMapLock()->lock_without_safepoint_check(); 1593 if (TraceCMSState) { 1594 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1595 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1596 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1597 } 1598 1599 // Inform cms gen if this was due to partial collection failing. 1600 // The CMS gen may use this fact to determine its expansion policy. 1601 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1602 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1603 assert(!_cmsGen->incremental_collection_failed(), 1604 "Should have been noticed, reacted to and cleared"); 1605 _cmsGen->set_incremental_collection_failed(); 1606 } 1607 1608 if (first_state > Idling) { 1609 report_concurrent_mode_interruption(); 1610 } 1611 1612 set_did_compact(true); 1613 1614 // If the collection is being acquired from the background 1615 // collector, there may be references on the discovered 1616 // references lists. Abandon those references, since some 1617 // of them may have become unreachable after concurrent 1618 // discovery; the STW compacting collector will redo discovery 1619 // more precisely, without being subject to floating garbage. 1620 // Leaving otherwise unreachable references in the discovered 1621 // lists would require special handling. 1622 ref_processor()->disable_discovery(); 1623 ref_processor()->abandon_partial_discovery(); 1624 ref_processor()->verify_no_references_recorded(); 1625 1626 if (first_state > Idling) { 1627 save_heap_summary(); 1628 } 1629 1630 do_compaction_work(clear_all_soft_refs); 1631 1632 // Has the GC time limit been exceeded? 1633 size_t max_eden_size = _young_gen->max_capacity() - 1634 _young_gen->to()->capacity() - 1635 _young_gen->from()->capacity(); 1636 GCCause::Cause gc_cause = gch->gc_cause(); 1637 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1638 _young_gen->eden()->used(), 1639 _cmsGen->max_capacity(), 1640 max_eden_size, 1641 full, 1642 gc_cause, 1643 gch->collector_policy()); 1644 1645 // Reset the expansion cause, now that we just completed 1646 // a collection cycle. 1647 clear_expansion_cause(); 1648 _foregroundGCIsActive = false; 1649 return; 1650 } 1651 1652 // Resize the tenured generation 1653 // after obtaining the free list locks for the 1654 // two generations. 1655 void CMSCollector::compute_new_size() { 1656 assert_locked_or_safepoint(Heap_lock); 1657 FreelistLocker z(this); 1658 MetaspaceGC::compute_new_size(); 1659 _cmsGen->compute_new_size_free_list(); 1660 } 1661 1662 // A work method used by the foreground collector to do 1663 // a mark-sweep-compact. 1664 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1665 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1666 1667 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1668 gc_timer->register_gc_start(); 1669 1670 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1671 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1672 1673 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1674 1675 // Temporarily widen the span of the weak reference processing to 1676 // the entire heap. 1677 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1678 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1679 // Temporarily, clear the "is_alive_non_header" field of the 1680 // reference processor. 1681 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1682 // Temporarily make reference _processing_ single threaded (non-MT). 1683 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1684 // Temporarily make refs discovery atomic 1685 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1686 // Temporarily make reference _discovery_ single threaded (non-MT) 1687 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1688 1689 ref_processor()->set_enqueuing_is_done(false); 1690 ref_processor()->enable_discovery(); 1691 ref_processor()->setup_policy(clear_all_soft_refs); 1692 // If an asynchronous collection finishes, the _modUnionTable is 1693 // all clear. If we are assuming the collection from an asynchronous 1694 // collection, clear the _modUnionTable. 1695 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1696 "_modUnionTable should be clear if the baton was not passed"); 1697 _modUnionTable.clear_all(); 1698 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1699 "mod union for klasses should be clear if the baton was passed"); 1700 _ct->klass_rem_set()->clear_mod_union(); 1701 1702 // We must adjust the allocation statistics being maintained 1703 // in the free list space. We do so by reading and clearing 1704 // the sweep timer and updating the block flux rate estimates below. 1705 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1706 if (_inter_sweep_timer.is_active()) { 1707 _inter_sweep_timer.stop(); 1708 // Note that we do not use this sample to update the _inter_sweep_estimate. 1709 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1710 _inter_sweep_estimate.padded_average(), 1711 _intra_sweep_estimate.padded_average()); 1712 } 1713 1714 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1715 ref_processor(), clear_all_soft_refs); 1716 #ifdef ASSERT 1717 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1718 size_t free_size = cms_space->free(); 1719 assert(free_size == 1720 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1721 * HeapWordSize, 1722 "All the free space should be compacted into one chunk at top"); 1723 assert(cms_space->dictionary()->total_chunk_size( 1724 debug_only(cms_space->freelistLock())) == 0 || 1725 cms_space->totalSizeInIndexedFreeLists() == 0, 1726 "All the free space should be in a single chunk"); 1727 size_t num = cms_space->totalCount(); 1728 assert((free_size == 0 && num == 0) || 1729 (free_size > 0 && (num == 1 || num == 2)), 1730 "There should be at most 2 free chunks after compaction"); 1731 #endif // ASSERT 1732 _collectorState = Resetting; 1733 assert(_restart_addr == NULL, 1734 "Should have been NULL'd before baton was passed"); 1735 reset(false /* == !concurrent */); 1736 _cmsGen->reset_after_compaction(); 1737 _concurrent_cycles_since_last_unload = 0; 1738 1739 // Clear any data recorded in the PLAB chunk arrays. 1740 if (_survivor_plab_array != NULL) { 1741 reset_survivor_plab_arrays(); 1742 } 1743 1744 // Adjust the per-size allocation stats for the next epoch. 1745 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1746 // Restart the "inter sweep timer" for the next epoch. 1747 _inter_sweep_timer.reset(); 1748 _inter_sweep_timer.start(); 1749 1750 gc_timer->register_gc_end(); 1751 1752 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1753 1754 // For a mark-sweep-compact, compute_new_size() will be called 1755 // in the heap's do_collection() method. 1756 } 1757 1758 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1759 ContiguousSpace* eden_space = _young_gen->eden(); 1760 ContiguousSpace* from_space = _young_gen->from(); 1761 ContiguousSpace* to_space = _young_gen->to(); 1762 // Eden 1763 if (_eden_chunk_array != NULL) { 1764 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1765 p2i(eden_space->bottom()), p2i(eden_space->top()), 1766 p2i(eden_space->end()), eden_space->capacity()); 1767 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1768 "_eden_chunk_capacity=" SIZE_FORMAT, 1769 _eden_chunk_index, _eden_chunk_capacity); 1770 for (size_t i = 0; i < _eden_chunk_index; i++) { 1771 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1772 i, p2i(_eden_chunk_array[i])); 1773 } 1774 } 1775 // Survivor 1776 if (_survivor_chunk_array != NULL) { 1777 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1778 p2i(from_space->bottom()), p2i(from_space->top()), 1779 p2i(from_space->end()), from_space->capacity()); 1780 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1781 "_survivor_chunk_capacity=" SIZE_FORMAT, 1782 _survivor_chunk_index, _survivor_chunk_capacity); 1783 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1784 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1785 i, p2i(_survivor_chunk_array[i])); 1786 } 1787 } 1788 } 1789 1790 void CMSCollector::getFreelistLocks() const { 1791 // Get locks for all free lists in all generations that this 1792 // collector is responsible for 1793 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1794 } 1795 1796 void CMSCollector::releaseFreelistLocks() const { 1797 // Release locks for all free lists in all generations that this 1798 // collector is responsible for 1799 _cmsGen->freelistLock()->unlock(); 1800 } 1801 1802 bool CMSCollector::haveFreelistLocks() const { 1803 // Check locks for all free lists in all generations that this 1804 // collector is responsible for 1805 assert_lock_strong(_cmsGen->freelistLock()); 1806 PRODUCT_ONLY(ShouldNotReachHere()); 1807 return true; 1808 } 1809 1810 // A utility class that is used by the CMS collector to 1811 // temporarily "release" the foreground collector from its 1812 // usual obligation to wait for the background collector to 1813 // complete an ongoing phase before proceeding. 1814 class ReleaseForegroundGC: public StackObj { 1815 private: 1816 CMSCollector* _c; 1817 public: 1818 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1819 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1820 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1821 // allow a potentially blocked foreground collector to proceed 1822 _c->_foregroundGCShouldWait = false; 1823 if (_c->_foregroundGCIsActive) { 1824 CGC_lock->notify(); 1825 } 1826 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1827 "Possible deadlock"); 1828 } 1829 1830 ~ReleaseForegroundGC() { 1831 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1832 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1833 _c->_foregroundGCShouldWait = true; 1834 } 1835 }; 1836 1837 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1838 assert(Thread::current()->is_ConcurrentGC_thread(), 1839 "A CMS asynchronous collection is only allowed on a CMS thread."); 1840 1841 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1842 { 1843 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1844 MutexLockerEx hl(Heap_lock, safepoint_check); 1845 FreelistLocker fll(this); 1846 MutexLockerEx x(CGC_lock, safepoint_check); 1847 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1848 // The foreground collector is active or we're 1849 // not using asynchronous collections. Skip this 1850 // background collection. 1851 assert(!_foregroundGCShouldWait, "Should be clear"); 1852 return; 1853 } else { 1854 assert(_collectorState == Idling, "Should be idling before start."); 1855 _collectorState = InitialMarking; 1856 register_gc_start(cause); 1857 // Reset the expansion cause, now that we are about to begin 1858 // a new cycle. 1859 clear_expansion_cause(); 1860 1861 // Clear the MetaspaceGC flag since a concurrent collection 1862 // is starting but also clear it after the collection. 1863 MetaspaceGC::set_should_concurrent_collect(false); 1864 } 1865 // Decide if we want to enable class unloading as part of the 1866 // ensuing concurrent GC cycle. 1867 update_should_unload_classes(); 1868 _full_gc_requested = false; // acks all outstanding full gc requests 1869 _full_gc_cause = GCCause::_no_gc; 1870 // Signal that we are about to start a collection 1871 gch->increment_total_full_collections(); // ... starting a collection cycle 1872 _collection_count_start = gch->total_full_collections(); 1873 } 1874 1875 // Used for PrintGC 1876 size_t prev_used; 1877 if (PrintGC && Verbose) { 1878 prev_used = _cmsGen->used(); 1879 } 1880 1881 // The change of the collection state is normally done at this level; 1882 // the exceptions are phases that are executed while the world is 1883 // stopped. For those phases the change of state is done while the 1884 // world is stopped. For baton passing purposes this allows the 1885 // background collector to finish the phase and change state atomically. 1886 // The foreground collector cannot wait on a phase that is done 1887 // while the world is stopped because the foreground collector already 1888 // has the world stopped and would deadlock. 1889 while (_collectorState != Idling) { 1890 if (TraceCMSState) { 1891 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1892 p2i(Thread::current()), _collectorState); 1893 } 1894 // The foreground collector 1895 // holds the Heap_lock throughout its collection. 1896 // holds the CMS token (but not the lock) 1897 // except while it is waiting for the background collector to yield. 1898 // 1899 // The foreground collector should be blocked (not for long) 1900 // if the background collector is about to start a phase 1901 // executed with world stopped. If the background 1902 // collector has already started such a phase, the 1903 // foreground collector is blocked waiting for the 1904 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1905 // are executed in the VM thread. 1906 // 1907 // The locking order is 1908 // PendingListLock (PLL) -- if applicable (FinalMarking) 1909 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1910 // CMS token (claimed in 1911 // stop_world_and_do() --> 1912 // safepoint_synchronize() --> 1913 // CMSThread::synchronize()) 1914 1915 { 1916 // Check if the FG collector wants us to yield. 1917 CMSTokenSync x(true); // is cms thread 1918 if (waitForForegroundGC()) { 1919 // We yielded to a foreground GC, nothing more to be 1920 // done this round. 1921 assert(_foregroundGCShouldWait == false, "We set it to false in " 1922 "waitForForegroundGC()"); 1923 if (TraceCMSState) { 1924 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1925 " exiting collection CMS state %d", 1926 p2i(Thread::current()), _collectorState); 1927 } 1928 return; 1929 } else { 1930 // The background collector can run but check to see if the 1931 // foreground collector has done a collection while the 1932 // background collector was waiting to get the CGC_lock 1933 // above. If yes, break so that _foregroundGCShouldWait 1934 // is cleared before returning. 1935 if (_collectorState == Idling) { 1936 break; 1937 } 1938 } 1939 } 1940 1941 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1942 "should be waiting"); 1943 1944 switch (_collectorState) { 1945 case InitialMarking: 1946 { 1947 ReleaseForegroundGC x(this); 1948 stats().record_cms_begin(); 1949 VM_CMS_Initial_Mark initial_mark_op(this); 1950 VMThread::execute(&initial_mark_op); 1951 } 1952 // The collector state may be any legal state at this point 1953 // since the background collector may have yielded to the 1954 // foreground collector. 1955 break; 1956 case Marking: 1957 // initial marking in checkpointRootsInitialWork has been completed 1958 if (markFromRoots()) { // we were successful 1959 assert(_collectorState == Precleaning, "Collector state should " 1960 "have changed"); 1961 } else { 1962 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1963 } 1964 break; 1965 case Precleaning: 1966 // marking from roots in markFromRoots has been completed 1967 preclean(); 1968 assert(_collectorState == AbortablePreclean || 1969 _collectorState == FinalMarking, 1970 "Collector state should have changed"); 1971 break; 1972 case AbortablePreclean: 1973 abortable_preclean(); 1974 assert(_collectorState == FinalMarking, "Collector state should " 1975 "have changed"); 1976 break; 1977 case FinalMarking: 1978 { 1979 ReleaseForegroundGC x(this); 1980 1981 VM_CMS_Final_Remark final_remark_op(this); 1982 VMThread::execute(&final_remark_op); 1983 } 1984 assert(_foregroundGCShouldWait, "block post-condition"); 1985 break; 1986 case Sweeping: 1987 // final marking in checkpointRootsFinal has been completed 1988 sweep(); 1989 assert(_collectorState == Resizing, "Collector state change " 1990 "to Resizing must be done under the free_list_lock"); 1991 1992 case Resizing: { 1993 // Sweeping has been completed... 1994 // At this point the background collection has completed. 1995 // Don't move the call to compute_new_size() down 1996 // into code that might be executed if the background 1997 // collection was preempted. 1998 { 1999 ReleaseForegroundGC x(this); // unblock FG collection 2000 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2001 CMSTokenSync z(true); // not strictly needed. 2002 if (_collectorState == Resizing) { 2003 compute_new_size(); 2004 save_heap_summary(); 2005 _collectorState = Resetting; 2006 } else { 2007 assert(_collectorState == Idling, "The state should only change" 2008 " because the foreground collector has finished the collection"); 2009 } 2010 } 2011 break; 2012 } 2013 case Resetting: 2014 // CMS heap resizing has been completed 2015 reset(true); 2016 assert(_collectorState == Idling, "Collector state should " 2017 "have changed"); 2018 2019 MetaspaceGC::set_should_concurrent_collect(false); 2020 2021 stats().record_cms_end(); 2022 // Don't move the concurrent_phases_end() and compute_new_size() 2023 // calls to here because a preempted background collection 2024 // has it's state set to "Resetting". 2025 break; 2026 case Idling: 2027 default: 2028 ShouldNotReachHere(); 2029 break; 2030 } 2031 if (TraceCMSState) { 2032 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2033 p2i(Thread::current()), _collectorState); 2034 } 2035 assert(_foregroundGCShouldWait, "block post-condition"); 2036 } 2037 2038 // Should this be in gc_epilogue? 2039 collector_policy()->counters()->update_counters(); 2040 2041 { 2042 // Clear _foregroundGCShouldWait and, in the event that the 2043 // foreground collector is waiting, notify it, before 2044 // returning. 2045 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2046 _foregroundGCShouldWait = false; 2047 if (_foregroundGCIsActive) { 2048 CGC_lock->notify(); 2049 } 2050 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2051 "Possible deadlock"); 2052 } 2053 if (TraceCMSState) { 2054 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2055 " exiting collection CMS state %d", 2056 p2i(Thread::current()), _collectorState); 2057 } 2058 if (PrintGC && Verbose) { 2059 _cmsGen->print_heap_change(prev_used); 2060 } 2061 } 2062 2063 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2064 _cms_start_registered = true; 2065 _gc_timer_cm->register_gc_start(); 2066 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2067 } 2068 2069 void CMSCollector::register_gc_end() { 2070 if (_cms_start_registered) { 2071 report_heap_summary(GCWhen::AfterGC); 2072 2073 _gc_timer_cm->register_gc_end(); 2074 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2075 _cms_start_registered = false; 2076 } 2077 } 2078 2079 void CMSCollector::save_heap_summary() { 2080 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2081 _last_heap_summary = gch->create_heap_summary(); 2082 _last_metaspace_summary = gch->create_metaspace_summary(); 2083 } 2084 2085 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2086 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2087 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2088 } 2089 2090 bool CMSCollector::waitForForegroundGC() { 2091 bool res = false; 2092 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2093 "CMS thread should have CMS token"); 2094 // Block the foreground collector until the 2095 // background collectors decides whether to 2096 // yield. 2097 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2098 _foregroundGCShouldWait = true; 2099 if (_foregroundGCIsActive) { 2100 // The background collector yields to the 2101 // foreground collector and returns a value 2102 // indicating that it has yielded. The foreground 2103 // collector can proceed. 2104 res = true; 2105 _foregroundGCShouldWait = false; 2106 ConcurrentMarkSweepThread::clear_CMS_flag( 2107 ConcurrentMarkSweepThread::CMS_cms_has_token); 2108 ConcurrentMarkSweepThread::set_CMS_flag( 2109 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2110 // Get a possibly blocked foreground thread going 2111 CGC_lock->notify(); 2112 if (TraceCMSState) { 2113 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2114 p2i(Thread::current()), _collectorState); 2115 } 2116 while (_foregroundGCIsActive) { 2117 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2118 } 2119 ConcurrentMarkSweepThread::set_CMS_flag( 2120 ConcurrentMarkSweepThread::CMS_cms_has_token); 2121 ConcurrentMarkSweepThread::clear_CMS_flag( 2122 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2123 } 2124 if (TraceCMSState) { 2125 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2126 p2i(Thread::current()), _collectorState); 2127 } 2128 return res; 2129 } 2130 2131 // Because of the need to lock the free lists and other structures in 2132 // the collector, common to all the generations that the collector is 2133 // collecting, we need the gc_prologues of individual CMS generations 2134 // delegate to their collector. It may have been simpler had the 2135 // current infrastructure allowed one to call a prologue on a 2136 // collector. In the absence of that we have the generation's 2137 // prologue delegate to the collector, which delegates back 2138 // some "local" work to a worker method in the individual generations 2139 // that it's responsible for collecting, while itself doing any 2140 // work common to all generations it's responsible for. A similar 2141 // comment applies to the gc_epilogue()'s. 2142 // The role of the variable _between_prologue_and_epilogue is to 2143 // enforce the invocation protocol. 2144 void CMSCollector::gc_prologue(bool full) { 2145 // Call gc_prologue_work() for the CMSGen 2146 // we are responsible for. 2147 2148 // The following locking discipline assumes that we are only called 2149 // when the world is stopped. 2150 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2151 2152 // The CMSCollector prologue must call the gc_prologues for the 2153 // "generations" that it's responsible 2154 // for. 2155 2156 assert( Thread::current()->is_VM_thread() 2157 || ( CMSScavengeBeforeRemark 2158 && Thread::current()->is_ConcurrentGC_thread()), 2159 "Incorrect thread type for prologue execution"); 2160 2161 if (_between_prologue_and_epilogue) { 2162 // We have already been invoked; this is a gc_prologue delegation 2163 // from yet another CMS generation that we are responsible for, just 2164 // ignore it since all relevant work has already been done. 2165 return; 2166 } 2167 2168 // set a bit saying prologue has been called; cleared in epilogue 2169 _between_prologue_and_epilogue = true; 2170 // Claim locks for common data structures, then call gc_prologue_work() 2171 // for each CMSGen. 2172 2173 getFreelistLocks(); // gets free list locks on constituent spaces 2174 bitMapLock()->lock_without_safepoint_check(); 2175 2176 // Should call gc_prologue_work() for all cms gens we are responsible for 2177 bool duringMarking = _collectorState >= Marking 2178 && _collectorState < Sweeping; 2179 2180 // The young collections clear the modified oops state, which tells if 2181 // there are any modified oops in the class. The remark phase also needs 2182 // that information. Tell the young collection to save the union of all 2183 // modified klasses. 2184 if (duringMarking) { 2185 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2186 } 2187 2188 bool registerClosure = duringMarking; 2189 2190 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2191 &_modUnionClosurePar 2192 : &_modUnionClosure; 2193 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2194 2195 if (!full) { 2196 stats().record_gc0_begin(); 2197 } 2198 } 2199 2200 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2201 2202 _capacity_at_prologue = capacity(); 2203 _used_at_prologue = used(); 2204 2205 // Delegate to CMScollector which knows how to coordinate between 2206 // this and any other CMS generations that it is responsible for 2207 // collecting. 2208 collector()->gc_prologue(full); 2209 } 2210 2211 // This is a "private" interface for use by this generation's CMSCollector. 2212 // Not to be called directly by any other entity (for instance, 2213 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2214 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2215 bool registerClosure, ModUnionClosure* modUnionClosure) { 2216 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2217 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2218 "Should be NULL"); 2219 if (registerClosure) { 2220 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2221 } 2222 cmsSpace()->gc_prologue(); 2223 // Clear stat counters 2224 NOT_PRODUCT( 2225 assert(_numObjectsPromoted == 0, "check"); 2226 assert(_numWordsPromoted == 0, "check"); 2227 if (Verbose && PrintGC) { 2228 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2229 SIZE_FORMAT" bytes concurrently", 2230 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2231 } 2232 _numObjectsAllocated = 0; 2233 _numWordsAllocated = 0; 2234 ) 2235 } 2236 2237 void CMSCollector::gc_epilogue(bool full) { 2238 // The following locking discipline assumes that we are only called 2239 // when the world is stopped. 2240 assert(SafepointSynchronize::is_at_safepoint(), 2241 "world is stopped assumption"); 2242 2243 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2244 // if linear allocation blocks need to be appropriately marked to allow the 2245 // the blocks to be parsable. We also check here whether we need to nudge the 2246 // CMS collector thread to start a new cycle (if it's not already active). 2247 assert( Thread::current()->is_VM_thread() 2248 || ( CMSScavengeBeforeRemark 2249 && Thread::current()->is_ConcurrentGC_thread()), 2250 "Incorrect thread type for epilogue execution"); 2251 2252 if (!_between_prologue_and_epilogue) { 2253 // We have already been invoked; this is a gc_epilogue delegation 2254 // from yet another CMS generation that we are responsible for, just 2255 // ignore it since all relevant work has already been done. 2256 return; 2257 } 2258 assert(haveFreelistLocks(), "must have freelist locks"); 2259 assert_lock_strong(bitMapLock()); 2260 2261 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2262 2263 _cmsGen->gc_epilogue_work(full); 2264 2265 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2266 // in case sampling was not already enabled, enable it 2267 _start_sampling = true; 2268 } 2269 // reset _eden_chunk_array so sampling starts afresh 2270 _eden_chunk_index = 0; 2271 2272 size_t cms_used = _cmsGen->cmsSpace()->used(); 2273 2274 // update performance counters - this uses a special version of 2275 // update_counters() that allows the utilization to be passed as a 2276 // parameter, avoiding multiple calls to used(). 2277 // 2278 _cmsGen->update_counters(cms_used); 2279 2280 bitMapLock()->unlock(); 2281 releaseFreelistLocks(); 2282 2283 if (!CleanChunkPoolAsync) { 2284 Chunk::clean_chunk_pool(); 2285 } 2286 2287 set_did_compact(false); 2288 _between_prologue_and_epilogue = false; // ready for next cycle 2289 } 2290 2291 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2292 collector()->gc_epilogue(full); 2293 2294 // Also reset promotion tracking in par gc thread states. 2295 for (uint i = 0; i < ParallelGCThreads; i++) { 2296 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2297 } 2298 } 2299 2300 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2301 assert(!incremental_collection_failed(), "Should have been cleared"); 2302 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2303 cmsSpace()->gc_epilogue(); 2304 // Print stat counters 2305 NOT_PRODUCT( 2306 assert(_numObjectsAllocated == 0, "check"); 2307 assert(_numWordsAllocated == 0, "check"); 2308 if (Verbose && PrintGC) { 2309 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2310 SIZE_FORMAT" bytes", 2311 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2312 } 2313 _numObjectsPromoted = 0; 2314 _numWordsPromoted = 0; 2315 ) 2316 2317 if (PrintGC && Verbose) { 2318 // Call down the chain in contiguous_available needs the freelistLock 2319 // so print this out before releasing the freeListLock. 2320 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2321 contiguous_available()); 2322 } 2323 } 2324 2325 #ifndef PRODUCT 2326 bool CMSCollector::have_cms_token() { 2327 Thread* thr = Thread::current(); 2328 if (thr->is_VM_thread()) { 2329 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2330 } else if (thr->is_ConcurrentGC_thread()) { 2331 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2332 } else if (thr->is_GC_task_thread()) { 2333 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2334 ParGCRareEvent_lock->owned_by_self(); 2335 } 2336 return false; 2337 } 2338 #endif 2339 2340 // Check reachability of the given heap address in CMS generation, 2341 // treating all other generations as roots. 2342 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2343 // We could "guarantee" below, rather than assert, but I'll 2344 // leave these as "asserts" so that an adventurous debugger 2345 // could try this in the product build provided some subset of 2346 // the conditions were met, provided they were interested in the 2347 // results and knew that the computation below wouldn't interfere 2348 // with other concurrent computations mutating the structures 2349 // being read or written. 2350 assert(SafepointSynchronize::is_at_safepoint(), 2351 "Else mutations in object graph will make answer suspect"); 2352 assert(have_cms_token(), "Should hold cms token"); 2353 assert(haveFreelistLocks(), "must hold free list locks"); 2354 assert_lock_strong(bitMapLock()); 2355 2356 // Clear the marking bit map array before starting, but, just 2357 // for kicks, first report if the given address is already marked 2358 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2359 _markBitMap.isMarked(addr) ? "" : " not"); 2360 2361 if (verify_after_remark()) { 2362 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2363 bool result = verification_mark_bm()->isMarked(addr); 2364 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2365 result ? "IS" : "is NOT"); 2366 return result; 2367 } else { 2368 gclog_or_tty->print_cr("Could not compute result"); 2369 return false; 2370 } 2371 } 2372 2373 2374 void 2375 CMSCollector::print_on_error(outputStream* st) { 2376 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2377 if (collector != NULL) { 2378 CMSBitMap* bitmap = &collector->_markBitMap; 2379 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2380 bitmap->print_on_error(st, " Bits: "); 2381 2382 st->cr(); 2383 2384 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2385 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2386 mut_bitmap->print_on_error(st, " Bits: "); 2387 } 2388 } 2389 2390 //////////////////////////////////////////////////////// 2391 // CMS Verification Support 2392 //////////////////////////////////////////////////////// 2393 // Following the remark phase, the following invariant 2394 // should hold -- each object in the CMS heap which is 2395 // marked in markBitMap() should be marked in the verification_mark_bm(). 2396 2397 class VerifyMarkedClosure: public BitMapClosure { 2398 CMSBitMap* _marks; 2399 bool _failed; 2400 2401 public: 2402 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2403 2404 bool do_bit(size_t offset) { 2405 HeapWord* addr = _marks->offsetToHeapWord(offset); 2406 if (!_marks->isMarked(addr)) { 2407 oop(addr)->print_on(gclog_or_tty); 2408 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr)); 2409 _failed = true; 2410 } 2411 return true; 2412 } 2413 2414 bool failed() { return _failed; } 2415 }; 2416 2417 bool CMSCollector::verify_after_remark(bool silent) { 2418 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2419 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2420 static bool init = false; 2421 2422 assert(SafepointSynchronize::is_at_safepoint(), 2423 "Else mutations in object graph will make answer suspect"); 2424 assert(have_cms_token(), 2425 "Else there may be mutual interference in use of " 2426 " verification data structures"); 2427 assert(_collectorState > Marking && _collectorState <= Sweeping, 2428 "Else marking info checked here may be obsolete"); 2429 assert(haveFreelistLocks(), "must hold free list locks"); 2430 assert_lock_strong(bitMapLock()); 2431 2432 2433 // Allocate marking bit map if not already allocated 2434 if (!init) { // first time 2435 if (!verification_mark_bm()->allocate(_span)) { 2436 return false; 2437 } 2438 init = true; 2439 } 2440 2441 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2442 2443 // Turn off refs discovery -- so we will be tracing through refs. 2444 // This is as intended, because by this time 2445 // GC must already have cleared any refs that need to be cleared, 2446 // and traced those that need to be marked; moreover, 2447 // the marking done here is not going to interfere in any 2448 // way with the marking information used by GC. 2449 NoRefDiscovery no_discovery(ref_processor()); 2450 2451 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2452 2453 // Clear any marks from a previous round 2454 verification_mark_bm()->clear_all(); 2455 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2456 verify_work_stacks_empty(); 2457 2458 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2459 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2460 // Update the saved marks which may affect the root scans. 2461 gch->save_marks(); 2462 2463 if (CMSRemarkVerifyVariant == 1) { 2464 // In this first variant of verification, we complete 2465 // all marking, then check if the new marks-vector is 2466 // a subset of the CMS marks-vector. 2467 verify_after_remark_work_1(); 2468 } else if (CMSRemarkVerifyVariant == 2) { 2469 // In this second variant of verification, we flag an error 2470 // (i.e. an object reachable in the new marks-vector not reachable 2471 // in the CMS marks-vector) immediately, also indicating the 2472 // identify of an object (A) that references the unmarked object (B) -- 2473 // presumably, a mutation to A failed to be picked up by preclean/remark? 2474 verify_after_remark_work_2(); 2475 } else { 2476 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2477 CMSRemarkVerifyVariant); 2478 } 2479 if (!silent) gclog_or_tty->print(" done] "); 2480 return true; 2481 } 2482 2483 void CMSCollector::verify_after_remark_work_1() { 2484 ResourceMark rm; 2485 HandleMark hm; 2486 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2487 2488 // Get a clear set of claim bits for the roots processing to work with. 2489 ClassLoaderDataGraph::clear_claimed_marks(); 2490 2491 // Mark from roots one level into CMS 2492 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2493 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2494 2495 gch->gen_process_roots(_cmsGen->level(), 2496 true, // younger gens are roots 2497 true, // activate StrongRootsScope 2498 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2499 should_unload_classes(), 2500 ¬Older, 2501 NULL, 2502 NULL); // SSS: Provide correct closure 2503 2504 // Now mark from the roots 2505 MarkFromRootsClosure markFromRootsClosure(this, _span, 2506 verification_mark_bm(), verification_mark_stack(), 2507 false /* don't yield */, true /* verifying */); 2508 assert(_restart_addr == NULL, "Expected pre-condition"); 2509 verification_mark_bm()->iterate(&markFromRootsClosure); 2510 while (_restart_addr != NULL) { 2511 // Deal with stack overflow: by restarting at the indicated 2512 // address. 2513 HeapWord* ra = _restart_addr; 2514 markFromRootsClosure.reset(ra); 2515 _restart_addr = NULL; 2516 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2517 } 2518 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2519 verify_work_stacks_empty(); 2520 2521 // Marking completed -- now verify that each bit marked in 2522 // verification_mark_bm() is also marked in markBitMap(); flag all 2523 // errors by printing corresponding objects. 2524 VerifyMarkedClosure vcl(markBitMap()); 2525 verification_mark_bm()->iterate(&vcl); 2526 if (vcl.failed()) { 2527 gclog_or_tty->print("Verification failed"); 2528 Universe::heap()->print_on(gclog_or_tty); 2529 fatal("CMS: failed marking verification after remark"); 2530 } 2531 } 2532 2533 class VerifyKlassOopsKlassClosure : public KlassClosure { 2534 class VerifyKlassOopsClosure : public OopClosure { 2535 CMSBitMap* _bitmap; 2536 public: 2537 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2538 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2539 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2540 } _oop_closure; 2541 public: 2542 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2543 void do_klass(Klass* k) { 2544 k->oops_do(&_oop_closure); 2545 } 2546 }; 2547 2548 void CMSCollector::verify_after_remark_work_2() { 2549 ResourceMark rm; 2550 HandleMark hm; 2551 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2552 2553 // Get a clear set of claim bits for the roots processing to work with. 2554 ClassLoaderDataGraph::clear_claimed_marks(); 2555 2556 // Mark from roots one level into CMS 2557 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2558 markBitMap()); 2559 CLDToOopClosure cld_closure(¬Older, true); 2560 2561 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2562 2563 gch->gen_process_roots(_cmsGen->level(), 2564 true, // younger gens are roots 2565 true, // activate StrongRootsScope 2566 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2567 should_unload_classes(), 2568 ¬Older, 2569 NULL, 2570 &cld_closure); 2571 2572 // Now mark from the roots 2573 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2574 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2575 assert(_restart_addr == NULL, "Expected pre-condition"); 2576 verification_mark_bm()->iterate(&markFromRootsClosure); 2577 while (_restart_addr != NULL) { 2578 // Deal with stack overflow: by restarting at the indicated 2579 // address. 2580 HeapWord* ra = _restart_addr; 2581 markFromRootsClosure.reset(ra); 2582 _restart_addr = NULL; 2583 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2584 } 2585 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2586 verify_work_stacks_empty(); 2587 2588 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2589 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2590 2591 // Marking completed -- now verify that each bit marked in 2592 // verification_mark_bm() is also marked in markBitMap(); flag all 2593 // errors by printing corresponding objects. 2594 VerifyMarkedClosure vcl(markBitMap()); 2595 verification_mark_bm()->iterate(&vcl); 2596 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2597 } 2598 2599 void ConcurrentMarkSweepGeneration::save_marks() { 2600 // delegate to CMS space 2601 cmsSpace()->save_marks(); 2602 for (uint i = 0; i < ParallelGCThreads; i++) { 2603 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2604 } 2605 } 2606 2607 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2608 return cmsSpace()->no_allocs_since_save_marks(); 2609 } 2610 2611 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2612 \ 2613 void ConcurrentMarkSweepGeneration:: \ 2614 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2615 cl->set_generation(this); \ 2616 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2617 cl->reset_generation(); \ 2618 save_marks(); \ 2619 } 2620 2621 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2622 2623 void 2624 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2625 if (freelistLock()->owned_by_self()) { 2626 Generation::oop_iterate(cl); 2627 } else { 2628 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2629 Generation::oop_iterate(cl); 2630 } 2631 } 2632 2633 void 2634 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2635 if (freelistLock()->owned_by_self()) { 2636 Generation::object_iterate(cl); 2637 } else { 2638 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2639 Generation::object_iterate(cl); 2640 } 2641 } 2642 2643 void 2644 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2645 if (freelistLock()->owned_by_self()) { 2646 Generation::safe_object_iterate(cl); 2647 } else { 2648 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2649 Generation::safe_object_iterate(cl); 2650 } 2651 } 2652 2653 void 2654 ConcurrentMarkSweepGeneration::post_compact() { 2655 } 2656 2657 void 2658 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2659 // Fix the linear allocation blocks to look like free blocks. 2660 2661 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2662 // are not called when the heap is verified during universe initialization and 2663 // at vm shutdown. 2664 if (freelistLock()->owned_by_self()) { 2665 cmsSpace()->prepare_for_verify(); 2666 } else { 2667 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2668 cmsSpace()->prepare_for_verify(); 2669 } 2670 } 2671 2672 void 2673 ConcurrentMarkSweepGeneration::verify() { 2674 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2675 // are not called when the heap is verified during universe initialization and 2676 // at vm shutdown. 2677 if (freelistLock()->owned_by_self()) { 2678 cmsSpace()->verify(); 2679 } else { 2680 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2681 cmsSpace()->verify(); 2682 } 2683 } 2684 2685 void CMSCollector::verify() { 2686 _cmsGen->verify(); 2687 } 2688 2689 #ifndef PRODUCT 2690 bool CMSCollector::overflow_list_is_empty() const { 2691 assert(_num_par_pushes >= 0, "Inconsistency"); 2692 if (_overflow_list == NULL) { 2693 assert(_num_par_pushes == 0, "Inconsistency"); 2694 } 2695 return _overflow_list == NULL; 2696 } 2697 2698 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2699 // merely consolidate assertion checks that appear to occur together frequently. 2700 void CMSCollector::verify_work_stacks_empty() const { 2701 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2702 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2703 } 2704 2705 void CMSCollector::verify_overflow_empty() const { 2706 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2707 assert(no_preserved_marks(), "No preserved marks"); 2708 } 2709 #endif // PRODUCT 2710 2711 // Decide if we want to enable class unloading as part of the 2712 // ensuing concurrent GC cycle. We will collect and 2713 // unload classes if it's the case that: 2714 // (1) an explicit gc request has been made and the flag 2715 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2716 // (2) (a) class unloading is enabled at the command line, and 2717 // (b) old gen is getting really full 2718 // NOTE: Provided there is no change in the state of the heap between 2719 // calls to this method, it should have idempotent results. Moreover, 2720 // its results should be monotonically increasing (i.e. going from 0 to 1, 2721 // but not 1 to 0) between successive calls between which the heap was 2722 // not collected. For the implementation below, it must thus rely on 2723 // the property that concurrent_cycles_since_last_unload() 2724 // will not decrease unless a collection cycle happened and that 2725 // _cmsGen->is_too_full() are 2726 // themselves also monotonic in that sense. See check_monotonicity() 2727 // below. 2728 void CMSCollector::update_should_unload_classes() { 2729 _should_unload_classes = false; 2730 // Condition 1 above 2731 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2732 _should_unload_classes = true; 2733 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2734 // Disjuncts 2.b.(i,ii,iii) above 2735 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2736 CMSClassUnloadingMaxInterval) 2737 || _cmsGen->is_too_full(); 2738 } 2739 } 2740 2741 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2742 bool res = should_concurrent_collect(); 2743 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2744 return res; 2745 } 2746 2747 void CMSCollector::setup_cms_unloading_and_verification_state() { 2748 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2749 || VerifyBeforeExit; 2750 const int rso = GenCollectedHeap::SO_AllCodeCache; 2751 2752 // We set the proper root for this CMS cycle here. 2753 if (should_unload_classes()) { // Should unload classes this cycle 2754 remove_root_scanning_option(rso); // Shrink the root set appropriately 2755 set_verifying(should_verify); // Set verification state for this cycle 2756 return; // Nothing else needs to be done at this time 2757 } 2758 2759 // Not unloading classes this cycle 2760 assert(!should_unload_classes(), "Inconsistency!"); 2761 2762 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2763 // Include symbols, strings and code cache elements to prevent their resurrection. 2764 add_root_scanning_option(rso); 2765 set_verifying(true); 2766 } else if (verifying() && !should_verify) { 2767 // We were verifying, but some verification flags got disabled. 2768 set_verifying(false); 2769 // Exclude symbols, strings and code cache elements from root scanning to 2770 // reduce IM and RM pauses. 2771 remove_root_scanning_option(rso); 2772 } 2773 } 2774 2775 2776 #ifndef PRODUCT 2777 HeapWord* CMSCollector::block_start(const void* p) const { 2778 const HeapWord* addr = (HeapWord*)p; 2779 if (_span.contains(p)) { 2780 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2781 return _cmsGen->cmsSpace()->block_start(p); 2782 } 2783 } 2784 return NULL; 2785 } 2786 #endif 2787 2788 HeapWord* 2789 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2790 bool tlab, 2791 bool parallel) { 2792 CMSSynchronousYieldRequest yr; 2793 assert(!tlab, "Can't deal with TLAB allocation"); 2794 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2795 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2796 if (GCExpandToAllocateDelayMillis > 0) { 2797 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2798 } 2799 return have_lock_and_allocate(word_size, tlab); 2800 } 2801 2802 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2803 size_t bytes, 2804 size_t expand_bytes, 2805 CMSExpansionCause::Cause cause) 2806 { 2807 2808 bool success = expand(bytes, expand_bytes); 2809 2810 // remember why we expanded; this information is used 2811 // by shouldConcurrentCollect() when making decisions on whether to start 2812 // a new CMS cycle. 2813 if (success) { 2814 set_expansion_cause(cause); 2815 if (PrintGCDetails && Verbose) { 2816 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2817 CMSExpansionCause::to_string(cause)); 2818 } 2819 } 2820 } 2821 2822 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2823 HeapWord* res = NULL; 2824 MutexLocker x(ParGCRareEvent_lock); 2825 while (true) { 2826 // Expansion by some other thread might make alloc OK now: 2827 res = ps->lab.alloc(word_sz); 2828 if (res != NULL) return res; 2829 // If there's not enough expansion space available, give up. 2830 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2831 return NULL; 2832 } 2833 // Otherwise, we try expansion. 2834 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2835 // Now go around the loop and try alloc again; 2836 // A competing par_promote might beat us to the expansion space, 2837 // so we may go around the loop again if promotion fails again. 2838 if (GCExpandToAllocateDelayMillis > 0) { 2839 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2840 } 2841 } 2842 } 2843 2844 2845 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2846 PromotionInfo* promo) { 2847 MutexLocker x(ParGCRareEvent_lock); 2848 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2849 while (true) { 2850 // Expansion by some other thread might make alloc OK now: 2851 if (promo->ensure_spooling_space()) { 2852 assert(promo->has_spooling_space(), 2853 "Post-condition of successful ensure_spooling_space()"); 2854 return true; 2855 } 2856 // If there's not enough expansion space available, give up. 2857 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2858 return false; 2859 } 2860 // Otherwise, we try expansion. 2861 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2862 // Now go around the loop and try alloc again; 2863 // A competing allocation might beat us to the expansion space, 2864 // so we may go around the loop again if allocation fails again. 2865 if (GCExpandToAllocateDelayMillis > 0) { 2866 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2867 } 2868 } 2869 } 2870 2871 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2872 // Only shrink if a compaction was done so that all the free space 2873 // in the generation is in a contiguous block at the end. 2874 if (did_compact()) { 2875 CardGeneration::shrink(bytes); 2876 } 2877 } 2878 2879 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2880 assert_locked_or_safepoint(Heap_lock); 2881 } 2882 2883 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2884 assert_locked_or_safepoint(Heap_lock); 2885 assert_lock_strong(freelistLock()); 2886 if (PrintGCDetails && Verbose) { 2887 warning("Shrinking of CMS not yet implemented"); 2888 } 2889 return; 2890 } 2891 2892 2893 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2894 // phases. 2895 class CMSPhaseAccounting: public StackObj { 2896 public: 2897 CMSPhaseAccounting(CMSCollector *collector, 2898 const char *phase, 2899 const GCId gc_id, 2900 bool print_cr = true); 2901 ~CMSPhaseAccounting(); 2902 2903 private: 2904 CMSCollector *_collector; 2905 const char *_phase; 2906 elapsedTimer _wallclock; 2907 bool _print_cr; 2908 const GCId _gc_id; 2909 2910 public: 2911 // Not MT-safe; so do not pass around these StackObj's 2912 // where they may be accessed by other threads. 2913 jlong wallclock_millis() { 2914 assert(_wallclock.is_active(), "Wall clock should not stop"); 2915 _wallclock.stop(); // to record time 2916 jlong ret = _wallclock.milliseconds(); 2917 _wallclock.start(); // restart 2918 return ret; 2919 } 2920 }; 2921 2922 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2923 const char *phase, 2924 const GCId gc_id, 2925 bool print_cr) : 2926 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 2927 2928 if (PrintCMSStatistics != 0) { 2929 _collector->resetYields(); 2930 } 2931 if (PrintGCDetails) { 2932 gclog_or_tty->gclog_stamp(_gc_id); 2933 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2934 _collector->cmsGen()->short_name(), _phase); 2935 } 2936 _collector->resetTimer(); 2937 _wallclock.start(); 2938 _collector->startTimer(); 2939 } 2940 2941 CMSPhaseAccounting::~CMSPhaseAccounting() { 2942 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2943 _collector->stopTimer(); 2944 _wallclock.stop(); 2945 if (PrintGCDetails) { 2946 gclog_or_tty->gclog_stamp(_gc_id); 2947 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2948 _collector->cmsGen()->short_name(), 2949 _phase, _collector->timerValue(), _wallclock.seconds()); 2950 if (_print_cr) { 2951 gclog_or_tty->cr(); 2952 } 2953 if (PrintCMSStatistics != 0) { 2954 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2955 _collector->yields()); 2956 } 2957 } 2958 } 2959 2960 // CMS work 2961 2962 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2963 class CMSParMarkTask : public AbstractGangTask { 2964 protected: 2965 CMSCollector* _collector; 2966 int _n_workers; 2967 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 2968 AbstractGangTask(name), 2969 _collector(collector), 2970 _n_workers(n_workers) {} 2971 // Work method in support of parallel rescan ... of young gen spaces 2972 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2973 ContiguousSpace* space, 2974 HeapWord** chunk_array, size_t chunk_top); 2975 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2976 }; 2977 2978 // Parallel initial mark task 2979 class CMSParInitialMarkTask: public CMSParMarkTask { 2980 public: 2981 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 2982 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 2983 collector, n_workers) {} 2984 void work(uint worker_id); 2985 }; 2986 2987 // Checkpoint the roots into this generation from outside 2988 // this generation. [Note this initial checkpoint need only 2989 // be approximate -- we'll do a catch up phase subsequently.] 2990 void CMSCollector::checkpointRootsInitial() { 2991 assert(_collectorState == InitialMarking, "Wrong collector state"); 2992 check_correct_thread_executing(); 2993 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2994 2995 save_heap_summary(); 2996 report_heap_summary(GCWhen::BeforeGC); 2997 2998 ReferenceProcessor* rp = ref_processor(); 2999 assert(_restart_addr == NULL, "Control point invariant"); 3000 { 3001 // acquire locks for subsequent manipulations 3002 MutexLockerEx x(bitMapLock(), 3003 Mutex::_no_safepoint_check_flag); 3004 checkpointRootsInitialWork(); 3005 // enable ("weak") refs discovery 3006 rp->enable_discovery(); 3007 _collectorState = Marking; 3008 } 3009 } 3010 3011 void CMSCollector::checkpointRootsInitialWork() { 3012 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3013 assert(_collectorState == InitialMarking, "just checking"); 3014 3015 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3016 // precede our marking with a collection of all 3017 // younger generations to keep floating garbage to a minimum. 3018 // XXX: we won't do this for now -- it's an optimization to be done later. 3019 3020 // already have locks 3021 assert_lock_strong(bitMapLock()); 3022 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3023 3024 // Setup the verification and class unloading state for this 3025 // CMS collection cycle. 3026 setup_cms_unloading_and_verification_state(); 3027 3028 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3029 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3030 3031 // Reset all the PLAB chunk arrays if necessary. 3032 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3033 reset_survivor_plab_arrays(); 3034 } 3035 3036 ResourceMark rm; 3037 HandleMark hm; 3038 3039 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3040 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3041 3042 verify_work_stacks_empty(); 3043 verify_overflow_empty(); 3044 3045 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3046 // Update the saved marks which may affect the root scans. 3047 gch->save_marks(); 3048 3049 // weak reference processing has not started yet. 3050 ref_processor()->set_enqueuing_is_done(false); 3051 3052 // Need to remember all newly created CLDs, 3053 // so that we can guarantee that the remark finds them. 3054 ClassLoaderDataGraph::remember_new_clds(true); 3055 3056 // Whenever a CLD is found, it will be claimed before proceeding to mark 3057 // the klasses. The claimed marks need to be cleared before marking starts. 3058 ClassLoaderDataGraph::clear_claimed_marks(); 3059 3060 if (CMSPrintEdenSurvivorChunks) { 3061 print_eden_and_survivor_chunk_arrays(); 3062 } 3063 3064 { 3065 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3066 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3067 // The parallel version. 3068 FlexibleWorkGang* workers = gch->workers(); 3069 assert(workers != NULL, "Need parallel worker threads."); 3070 int n_workers = workers->active_workers(); 3071 CMSParInitialMarkTask tsk(this, n_workers); 3072 gch->set_par_threads(n_workers); 3073 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3074 if (n_workers > 1) { 3075 StrongRootsScope srs; 3076 workers->run_task(&tsk); 3077 } else { 3078 StrongRootsScope srs; 3079 tsk.work(0); 3080 } 3081 gch->set_par_threads(0); 3082 } else { 3083 // The serial version. 3084 CLDToOopClosure cld_closure(¬Older, true); 3085 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3086 gch->gen_process_roots(_cmsGen->level(), 3087 true, // younger gens are roots 3088 true, // activate StrongRootsScope 3089 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3090 should_unload_classes(), 3091 ¬Older, 3092 NULL, 3093 &cld_closure); 3094 } 3095 } 3096 3097 // Clear mod-union table; it will be dirtied in the prologue of 3098 // CMS generation per each younger generation collection. 3099 3100 assert(_modUnionTable.isAllClear(), 3101 "Was cleared in most recent final checkpoint phase" 3102 " or no bits are set in the gc_prologue before the start of the next " 3103 "subsequent marking phase."); 3104 3105 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3106 3107 // Save the end of the used_region of the constituent generations 3108 // to be used to limit the extent of sweep in each generation. 3109 save_sweep_limits(); 3110 verify_overflow_empty(); 3111 } 3112 3113 bool CMSCollector::markFromRoots() { 3114 // we might be tempted to assert that: 3115 // assert(!SafepointSynchronize::is_at_safepoint(), 3116 // "inconsistent argument?"); 3117 // However that wouldn't be right, because it's possible that 3118 // a safepoint is indeed in progress as a younger generation 3119 // stop-the-world GC happens even as we mark in this generation. 3120 assert(_collectorState == Marking, "inconsistent state?"); 3121 check_correct_thread_executing(); 3122 verify_overflow_empty(); 3123 3124 // Weak ref discovery note: We may be discovering weak 3125 // refs in this generation concurrent (but interleaved) with 3126 // weak ref discovery by a younger generation collector. 3127 3128 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3129 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3130 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3131 bool res = markFromRootsWork(); 3132 if (res) { 3133 _collectorState = Precleaning; 3134 } else { // We failed and a foreground collection wants to take over 3135 assert(_foregroundGCIsActive, "internal state inconsistency"); 3136 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3137 if (PrintGCDetails) { 3138 gclog_or_tty->print_cr("bailing out to foreground collection"); 3139 } 3140 } 3141 verify_overflow_empty(); 3142 return res; 3143 } 3144 3145 bool CMSCollector::markFromRootsWork() { 3146 // iterate over marked bits in bit map, doing a full scan and mark 3147 // from these roots using the following algorithm: 3148 // . if oop is to the right of the current scan pointer, 3149 // mark corresponding bit (we'll process it later) 3150 // . else (oop is to left of current scan pointer) 3151 // push oop on marking stack 3152 // . drain the marking stack 3153 3154 // Note that when we do a marking step we need to hold the 3155 // bit map lock -- recall that direct allocation (by mutators) 3156 // and promotion (by younger generation collectors) is also 3157 // marking the bit map. [the so-called allocate live policy.] 3158 // Because the implementation of bit map marking is not 3159 // robust wrt simultaneous marking of bits in the same word, 3160 // we need to make sure that there is no such interference 3161 // between concurrent such updates. 3162 3163 // already have locks 3164 assert_lock_strong(bitMapLock()); 3165 3166 verify_work_stacks_empty(); 3167 verify_overflow_empty(); 3168 bool result = false; 3169 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3170 result = do_marking_mt(); 3171 } else { 3172 result = do_marking_st(); 3173 } 3174 return result; 3175 } 3176 3177 // Forward decl 3178 class CMSConcMarkingTask; 3179 3180 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3181 CMSCollector* _collector; 3182 CMSConcMarkingTask* _task; 3183 public: 3184 virtual void yield(); 3185 3186 // "n_threads" is the number of threads to be terminated. 3187 // "queue_set" is a set of work queues of other threads. 3188 // "collector" is the CMS collector associated with this task terminator. 3189 // "yield" indicates whether we need the gang as a whole to yield. 3190 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3191 ParallelTaskTerminator(n_threads, queue_set), 3192 _collector(collector) { } 3193 3194 void set_task(CMSConcMarkingTask* task) { 3195 _task = task; 3196 } 3197 }; 3198 3199 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3200 CMSConcMarkingTask* _task; 3201 public: 3202 bool should_exit_termination(); 3203 void set_task(CMSConcMarkingTask* task) { 3204 _task = task; 3205 } 3206 }; 3207 3208 // MT Concurrent Marking Task 3209 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3210 CMSCollector* _collector; 3211 int _n_workers; // requested/desired # workers 3212 bool _result; 3213 CompactibleFreeListSpace* _cms_space; 3214 char _pad_front[64]; // padding to ... 3215 HeapWord* _global_finger; // ... avoid sharing cache line 3216 char _pad_back[64]; 3217 HeapWord* _restart_addr; 3218 3219 // Exposed here for yielding support 3220 Mutex* const _bit_map_lock; 3221 3222 // The per thread work queues, available here for stealing 3223 OopTaskQueueSet* _task_queues; 3224 3225 // Termination (and yielding) support 3226 CMSConcMarkingTerminator _term; 3227 CMSConcMarkingTerminatorTerminator _term_term; 3228 3229 public: 3230 CMSConcMarkingTask(CMSCollector* collector, 3231 CompactibleFreeListSpace* cms_space, 3232 YieldingFlexibleWorkGang* workers, 3233 OopTaskQueueSet* task_queues): 3234 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3235 _collector(collector), 3236 _cms_space(cms_space), 3237 _n_workers(0), _result(true), 3238 _task_queues(task_queues), 3239 _term(_n_workers, task_queues, _collector), 3240 _bit_map_lock(collector->bitMapLock()) 3241 { 3242 _requested_size = _n_workers; 3243 _term.set_task(this); 3244 _term_term.set_task(this); 3245 _restart_addr = _global_finger = _cms_space->bottom(); 3246 } 3247 3248 3249 OopTaskQueueSet* task_queues() { return _task_queues; } 3250 3251 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3252 3253 HeapWord** global_finger_addr() { return &_global_finger; } 3254 3255 CMSConcMarkingTerminator* terminator() { return &_term; } 3256 3257 virtual void set_for_termination(int active_workers) { 3258 terminator()->reset_for_reuse(active_workers); 3259 } 3260 3261 void work(uint worker_id); 3262 bool should_yield() { 3263 return ConcurrentMarkSweepThread::should_yield() 3264 && !_collector->foregroundGCIsActive(); 3265 } 3266 3267 virtual void coordinator_yield(); // stuff done by coordinator 3268 bool result() { return _result; } 3269 3270 void reset(HeapWord* ra) { 3271 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3272 _restart_addr = _global_finger = ra; 3273 _term.reset_for_reuse(); 3274 } 3275 3276 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3277 OopTaskQueue* work_q); 3278 3279 private: 3280 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3281 void do_work_steal(int i); 3282 void bump_global_finger(HeapWord* f); 3283 }; 3284 3285 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3286 assert(_task != NULL, "Error"); 3287 return _task->yielding(); 3288 // Note that we do not need the disjunct || _task->should_yield() above 3289 // because we want terminating threads to yield only if the task 3290 // is already in the midst of yielding, which happens only after at least one 3291 // thread has yielded. 3292 } 3293 3294 void CMSConcMarkingTerminator::yield() { 3295 if (_task->should_yield()) { 3296 _task->yield(); 3297 } else { 3298 ParallelTaskTerminator::yield(); 3299 } 3300 } 3301 3302 //////////////////////////////////////////////////////////////// 3303 // Concurrent Marking Algorithm Sketch 3304 //////////////////////////////////////////////////////////////// 3305 // Until all tasks exhausted (both spaces): 3306 // -- claim next available chunk 3307 // -- bump global finger via CAS 3308 // -- find first object that starts in this chunk 3309 // and start scanning bitmap from that position 3310 // -- scan marked objects for oops 3311 // -- CAS-mark target, and if successful: 3312 // . if target oop is above global finger (volatile read) 3313 // nothing to do 3314 // . if target oop is in chunk and above local finger 3315 // then nothing to do 3316 // . else push on work-queue 3317 // -- Deal with possible overflow issues: 3318 // . local work-queue overflow causes stuff to be pushed on 3319 // global (common) overflow queue 3320 // . always first empty local work queue 3321 // . then get a batch of oops from global work queue if any 3322 // . then do work stealing 3323 // -- When all tasks claimed (both spaces) 3324 // and local work queue empty, 3325 // then in a loop do: 3326 // . check global overflow stack; steal a batch of oops and trace 3327 // . try to steal from other threads oif GOS is empty 3328 // . if neither is available, offer termination 3329 // -- Terminate and return result 3330 // 3331 void CMSConcMarkingTask::work(uint worker_id) { 3332 elapsedTimer _timer; 3333 ResourceMark rm; 3334 HandleMark hm; 3335 3336 DEBUG_ONLY(_collector->verify_overflow_empty();) 3337 3338 // Before we begin work, our work queue should be empty 3339 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3340 // Scan the bitmap covering _cms_space, tracing through grey objects. 3341 _timer.start(); 3342 do_scan_and_mark(worker_id, _cms_space); 3343 _timer.stop(); 3344 if (PrintCMSStatistics != 0) { 3345 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3346 worker_id, _timer.seconds()); 3347 // XXX: need xxx/xxx type of notation, two timers 3348 } 3349 3350 // ... do work stealing 3351 _timer.reset(); 3352 _timer.start(); 3353 do_work_steal(worker_id); 3354 _timer.stop(); 3355 if (PrintCMSStatistics != 0) { 3356 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3357 worker_id, _timer.seconds()); 3358 // XXX: need xxx/xxx type of notation, two timers 3359 } 3360 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3361 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3362 // Note that under the current task protocol, the 3363 // following assertion is true even of the spaces 3364 // expanded since the completion of the concurrent 3365 // marking. XXX This will likely change under a strict 3366 // ABORT semantics. 3367 // After perm removal the comparison was changed to 3368 // greater than or equal to from strictly greater than. 3369 // Before perm removal the highest address sweep would 3370 // have been at the end of perm gen but now is at the 3371 // end of the tenured gen. 3372 assert(_global_finger >= _cms_space->end(), 3373 "All tasks have been completed"); 3374 DEBUG_ONLY(_collector->verify_overflow_empty();) 3375 } 3376 3377 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3378 HeapWord* read = _global_finger; 3379 HeapWord* cur = read; 3380 while (f > read) { 3381 cur = read; 3382 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3383 if (cur == read) { 3384 // our cas succeeded 3385 assert(_global_finger >= f, "protocol consistency"); 3386 break; 3387 } 3388 } 3389 } 3390 3391 // This is really inefficient, and should be redone by 3392 // using (not yet available) block-read and -write interfaces to the 3393 // stack and the work_queue. XXX FIX ME !!! 3394 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3395 OopTaskQueue* work_q) { 3396 // Fast lock-free check 3397 if (ovflw_stk->length() == 0) { 3398 return false; 3399 } 3400 assert(work_q->size() == 0, "Shouldn't steal"); 3401 MutexLockerEx ml(ovflw_stk->par_lock(), 3402 Mutex::_no_safepoint_check_flag); 3403 // Grab up to 1/4 the size of the work queue 3404 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3405 (size_t)ParGCDesiredObjsFromOverflowList); 3406 num = MIN2(num, ovflw_stk->length()); 3407 for (int i = (int) num; i > 0; i--) { 3408 oop cur = ovflw_stk->pop(); 3409 assert(cur != NULL, "Counted wrong?"); 3410 work_q->push(cur); 3411 } 3412 return num > 0; 3413 } 3414 3415 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3416 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3417 int n_tasks = pst->n_tasks(); 3418 // We allow that there may be no tasks to do here because 3419 // we are restarting after a stack overflow. 3420 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3421 uint nth_task = 0; 3422 3423 HeapWord* aligned_start = sp->bottom(); 3424 if (sp->used_region().contains(_restart_addr)) { 3425 // Align down to a card boundary for the start of 0th task 3426 // for this space. 3427 aligned_start = 3428 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3429 CardTableModRefBS::card_size); 3430 } 3431 3432 size_t chunk_size = sp->marking_task_size(); 3433 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3434 // Having claimed the nth task in this space, 3435 // compute the chunk that it corresponds to: 3436 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3437 aligned_start + (nth_task+1)*chunk_size); 3438 // Try and bump the global finger via a CAS; 3439 // note that we need to do the global finger bump 3440 // _before_ taking the intersection below, because 3441 // the task corresponding to that region will be 3442 // deemed done even if the used_region() expands 3443 // because of allocation -- as it almost certainly will 3444 // during start-up while the threads yield in the 3445 // closure below. 3446 HeapWord* finger = span.end(); 3447 bump_global_finger(finger); // atomically 3448 // There are null tasks here corresponding to chunks 3449 // beyond the "top" address of the space. 3450 span = span.intersection(sp->used_region()); 3451 if (!span.is_empty()) { // Non-null task 3452 HeapWord* prev_obj; 3453 assert(!span.contains(_restart_addr) || nth_task == 0, 3454 "Inconsistency"); 3455 if (nth_task == 0) { 3456 // For the 0th task, we'll not need to compute a block_start. 3457 if (span.contains(_restart_addr)) { 3458 // In the case of a restart because of stack overflow, 3459 // we might additionally skip a chunk prefix. 3460 prev_obj = _restart_addr; 3461 } else { 3462 prev_obj = span.start(); 3463 } 3464 } else { 3465 // We want to skip the first object because 3466 // the protocol is to scan any object in its entirety 3467 // that _starts_ in this span; a fortiori, any 3468 // object starting in an earlier span is scanned 3469 // as part of an earlier claimed task. 3470 // Below we use the "careful" version of block_start 3471 // so we do not try to navigate uninitialized objects. 3472 prev_obj = sp->block_start_careful(span.start()); 3473 // Below we use a variant of block_size that uses the 3474 // Printezis bits to avoid waiting for allocated 3475 // objects to become initialized/parsable. 3476 while (prev_obj < span.start()) { 3477 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3478 if (sz > 0) { 3479 prev_obj += sz; 3480 } else { 3481 // In this case we may end up doing a bit of redundant 3482 // scanning, but that appears unavoidable, short of 3483 // locking the free list locks; see bug 6324141. 3484 break; 3485 } 3486 } 3487 } 3488 if (prev_obj < span.end()) { 3489 MemRegion my_span = MemRegion(prev_obj, span.end()); 3490 // Do the marking work within a non-empty span -- 3491 // the last argument to the constructor indicates whether the 3492 // iteration should be incremental with periodic yields. 3493 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3494 &_collector->_markBitMap, 3495 work_queue(i), 3496 &_collector->_markStack); 3497 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3498 } // else nothing to do for this task 3499 } // else nothing to do for this task 3500 } 3501 // We'd be tempted to assert here that since there are no 3502 // more tasks left to claim in this space, the global_finger 3503 // must exceed space->top() and a fortiori space->end(). However, 3504 // that would not quite be correct because the bumping of 3505 // global_finger occurs strictly after the claiming of a task, 3506 // so by the time we reach here the global finger may not yet 3507 // have been bumped up by the thread that claimed the last 3508 // task. 3509 pst->all_tasks_completed(); 3510 } 3511 3512 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3513 private: 3514 CMSCollector* _collector; 3515 CMSConcMarkingTask* _task; 3516 MemRegion _span; 3517 CMSBitMap* _bit_map; 3518 CMSMarkStack* _overflow_stack; 3519 OopTaskQueue* _work_queue; 3520 protected: 3521 DO_OOP_WORK_DEFN 3522 public: 3523 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3524 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3525 MetadataAwareOopClosure(collector->ref_processor()), 3526 _collector(collector), 3527 _task(task), 3528 _span(collector->_span), 3529 _work_queue(work_queue), 3530 _bit_map(bit_map), 3531 _overflow_stack(overflow_stack) 3532 { } 3533 virtual void do_oop(oop* p); 3534 virtual void do_oop(narrowOop* p); 3535 3536 void trim_queue(size_t max); 3537 void handle_stack_overflow(HeapWord* lost); 3538 void do_yield_check() { 3539 if (_task->should_yield()) { 3540 _task->yield(); 3541 } 3542 } 3543 }; 3544 3545 // Grey object scanning during work stealing phase -- 3546 // the salient assumption here is that any references 3547 // that are in these stolen objects being scanned must 3548 // already have been initialized (else they would not have 3549 // been published), so we do not need to check for 3550 // uninitialized objects before pushing here. 3551 void Par_ConcMarkingClosure::do_oop(oop obj) { 3552 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3553 HeapWord* addr = (HeapWord*)obj; 3554 // Check if oop points into the CMS generation 3555 // and is not marked 3556 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3557 // a white object ... 3558 // If we manage to "claim" the object, by being the 3559 // first thread to mark it, then we push it on our 3560 // marking stack 3561 if (_bit_map->par_mark(addr)) { // ... now grey 3562 // push on work queue (grey set) 3563 bool simulate_overflow = false; 3564 NOT_PRODUCT( 3565 if (CMSMarkStackOverflowALot && 3566 _collector->simulate_overflow()) { 3567 // simulate a stack overflow 3568 simulate_overflow = true; 3569 } 3570 ) 3571 if (simulate_overflow || 3572 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3573 // stack overflow 3574 if (PrintCMSStatistics != 0) { 3575 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3576 SIZE_FORMAT, _overflow_stack->capacity()); 3577 } 3578 // We cannot assert that the overflow stack is full because 3579 // it may have been emptied since. 3580 assert(simulate_overflow || 3581 _work_queue->size() == _work_queue->max_elems(), 3582 "Else push should have succeeded"); 3583 handle_stack_overflow(addr); 3584 } 3585 } // Else, some other thread got there first 3586 do_yield_check(); 3587 } 3588 } 3589 3590 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3591 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3592 3593 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3594 while (_work_queue->size() > max) { 3595 oop new_oop; 3596 if (_work_queue->pop_local(new_oop)) { 3597 assert(new_oop->is_oop(), "Should be an oop"); 3598 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3599 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3600 new_oop->oop_iterate(this); // do_oop() above 3601 do_yield_check(); 3602 } 3603 } 3604 } 3605 3606 // Upon stack overflow, we discard (part of) the stack, 3607 // remembering the least address amongst those discarded 3608 // in CMSCollector's _restart_address. 3609 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3610 // We need to do this under a mutex to prevent other 3611 // workers from interfering with the work done below. 3612 MutexLockerEx ml(_overflow_stack->par_lock(), 3613 Mutex::_no_safepoint_check_flag); 3614 // Remember the least grey address discarded 3615 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3616 _collector->lower_restart_addr(ra); 3617 _overflow_stack->reset(); // discard stack contents 3618 _overflow_stack->expand(); // expand the stack if possible 3619 } 3620 3621 3622 void CMSConcMarkingTask::do_work_steal(int i) { 3623 OopTaskQueue* work_q = work_queue(i); 3624 oop obj_to_scan; 3625 CMSBitMap* bm = &(_collector->_markBitMap); 3626 CMSMarkStack* ovflw = &(_collector->_markStack); 3627 int* seed = _collector->hash_seed(i); 3628 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3629 while (true) { 3630 cl.trim_queue(0); 3631 assert(work_q->size() == 0, "Should have been emptied above"); 3632 if (get_work_from_overflow_stack(ovflw, work_q)) { 3633 // Can't assert below because the work obtained from the 3634 // overflow stack may already have been stolen from us. 3635 // assert(work_q->size() > 0, "Work from overflow stack"); 3636 continue; 3637 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3638 assert(obj_to_scan->is_oop(), "Should be an oop"); 3639 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3640 obj_to_scan->oop_iterate(&cl); 3641 } else if (terminator()->offer_termination(&_term_term)) { 3642 assert(work_q->size() == 0, "Impossible!"); 3643 break; 3644 } else if (yielding() || should_yield()) { 3645 yield(); 3646 } 3647 } 3648 } 3649 3650 // This is run by the CMS (coordinator) thread. 3651 void CMSConcMarkingTask::coordinator_yield() { 3652 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3653 "CMS thread should hold CMS token"); 3654 // First give up the locks, then yield, then re-lock 3655 // We should probably use a constructor/destructor idiom to 3656 // do this unlock/lock or modify the MutexUnlocker class to 3657 // serve our purpose. XXX 3658 assert_lock_strong(_bit_map_lock); 3659 _bit_map_lock->unlock(); 3660 ConcurrentMarkSweepThread::desynchronize(true); 3661 _collector->stopTimer(); 3662 if (PrintCMSStatistics != 0) { 3663 _collector->incrementYields(); 3664 } 3665 3666 // It is possible for whichever thread initiated the yield request 3667 // not to get a chance to wake up and take the bitmap lock between 3668 // this thread releasing it and reacquiring it. So, while the 3669 // should_yield() flag is on, let's sleep for a bit to give the 3670 // other thread a chance to wake up. The limit imposed on the number 3671 // of iterations is defensive, to avoid any unforseen circumstances 3672 // putting us into an infinite loop. Since it's always been this 3673 // (coordinator_yield()) method that was observed to cause the 3674 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3675 // which is by default non-zero. For the other seven methods that 3676 // also perform the yield operation, as are using a different 3677 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3678 // can enable the sleeping for those methods too, if necessary. 3679 // See 6442774. 3680 // 3681 // We really need to reconsider the synchronization between the GC 3682 // thread and the yield-requesting threads in the future and we 3683 // should really use wait/notify, which is the recommended 3684 // way of doing this type of interaction. Additionally, we should 3685 // consolidate the eight methods that do the yield operation and they 3686 // are almost identical into one for better maintainability and 3687 // readability. See 6445193. 3688 // 3689 // Tony 2006.06.29 3690 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3691 ConcurrentMarkSweepThread::should_yield() && 3692 !CMSCollector::foregroundGCIsActive(); ++i) { 3693 os::sleep(Thread::current(), 1, false); 3694 } 3695 3696 ConcurrentMarkSweepThread::synchronize(true); 3697 _bit_map_lock->lock_without_safepoint_check(); 3698 _collector->startTimer(); 3699 } 3700 3701 bool CMSCollector::do_marking_mt() { 3702 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3703 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 3704 conc_workers()->total_workers(), 3705 conc_workers()->active_workers(), 3706 Threads::number_of_non_daemon_threads()); 3707 conc_workers()->set_active_workers(num_workers); 3708 3709 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3710 3711 CMSConcMarkingTask tsk(this, 3712 cms_space, 3713 conc_workers(), 3714 task_queues()); 3715 3716 // Since the actual number of workers we get may be different 3717 // from the number we requested above, do we need to do anything different 3718 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3719 // class?? XXX 3720 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3721 3722 // Refs discovery is already non-atomic. 3723 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3724 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3725 conc_workers()->start_task(&tsk); 3726 while (tsk.yielded()) { 3727 tsk.coordinator_yield(); 3728 conc_workers()->continue_task(&tsk); 3729 } 3730 // If the task was aborted, _restart_addr will be non-NULL 3731 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3732 while (_restart_addr != NULL) { 3733 // XXX For now we do not make use of ABORTED state and have not 3734 // yet implemented the right abort semantics (even in the original 3735 // single-threaded CMS case). That needs some more investigation 3736 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3737 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3738 // If _restart_addr is non-NULL, a marking stack overflow 3739 // occurred; we need to do a fresh marking iteration from the 3740 // indicated restart address. 3741 if (_foregroundGCIsActive) { 3742 // We may be running into repeated stack overflows, having 3743 // reached the limit of the stack size, while making very 3744 // slow forward progress. It may be best to bail out and 3745 // let the foreground collector do its job. 3746 // Clear _restart_addr, so that foreground GC 3747 // works from scratch. This avoids the headache of 3748 // a "rescan" which would otherwise be needed because 3749 // of the dirty mod union table & card table. 3750 _restart_addr = NULL; 3751 return false; 3752 } 3753 // Adjust the task to restart from _restart_addr 3754 tsk.reset(_restart_addr); 3755 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3756 _restart_addr); 3757 _restart_addr = NULL; 3758 // Get the workers going again 3759 conc_workers()->start_task(&tsk); 3760 while (tsk.yielded()) { 3761 tsk.coordinator_yield(); 3762 conc_workers()->continue_task(&tsk); 3763 } 3764 } 3765 assert(tsk.completed(), "Inconsistency"); 3766 assert(tsk.result() == true, "Inconsistency"); 3767 return true; 3768 } 3769 3770 bool CMSCollector::do_marking_st() { 3771 ResourceMark rm; 3772 HandleMark hm; 3773 3774 // Temporarily make refs discovery single threaded (non-MT) 3775 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3776 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3777 &_markStack, CMSYield); 3778 // the last argument to iterate indicates whether the iteration 3779 // should be incremental with periodic yields. 3780 _markBitMap.iterate(&markFromRootsClosure); 3781 // If _restart_addr is non-NULL, a marking stack overflow 3782 // occurred; we need to do a fresh iteration from the 3783 // indicated restart address. 3784 while (_restart_addr != NULL) { 3785 if (_foregroundGCIsActive) { 3786 // We may be running into repeated stack overflows, having 3787 // reached the limit of the stack size, while making very 3788 // slow forward progress. It may be best to bail out and 3789 // let the foreground collector do its job. 3790 // Clear _restart_addr, so that foreground GC 3791 // works from scratch. This avoids the headache of 3792 // a "rescan" which would otherwise be needed because 3793 // of the dirty mod union table & card table. 3794 _restart_addr = NULL; 3795 return false; // indicating failure to complete marking 3796 } 3797 // Deal with stack overflow: 3798 // we restart marking from _restart_addr 3799 HeapWord* ra = _restart_addr; 3800 markFromRootsClosure.reset(ra); 3801 _restart_addr = NULL; 3802 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3803 } 3804 return true; 3805 } 3806 3807 void CMSCollector::preclean() { 3808 check_correct_thread_executing(); 3809 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3810 verify_work_stacks_empty(); 3811 verify_overflow_empty(); 3812 _abort_preclean = false; 3813 if (CMSPrecleaningEnabled) { 3814 if (!CMSEdenChunksRecordAlways) { 3815 _eden_chunk_index = 0; 3816 } 3817 size_t used = get_eden_used(); 3818 size_t capacity = get_eden_capacity(); 3819 // Don't start sampling unless we will get sufficiently 3820 // many samples. 3821 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3822 * CMSScheduleRemarkEdenPenetration)) { 3823 _start_sampling = true; 3824 } else { 3825 _start_sampling = false; 3826 } 3827 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3828 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3829 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3830 } 3831 CMSTokenSync x(true); // is cms thread 3832 if (CMSPrecleaningEnabled) { 3833 sample_eden(); 3834 _collectorState = AbortablePreclean; 3835 } else { 3836 _collectorState = FinalMarking; 3837 } 3838 verify_work_stacks_empty(); 3839 verify_overflow_empty(); 3840 } 3841 3842 // Try and schedule the remark such that young gen 3843 // occupancy is CMSScheduleRemarkEdenPenetration %. 3844 void CMSCollector::abortable_preclean() { 3845 check_correct_thread_executing(); 3846 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3847 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3848 3849 // If Eden's current occupancy is below this threshold, 3850 // immediately schedule the remark; else preclean 3851 // past the next scavenge in an effort to 3852 // schedule the pause as described above. By choosing 3853 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3854 // we will never do an actual abortable preclean cycle. 3855 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3856 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3857 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3858 // We need more smarts in the abortable preclean 3859 // loop below to deal with cases where allocation 3860 // in young gen is very very slow, and our precleaning 3861 // is running a losing race against a horde of 3862 // mutators intent on flooding us with CMS updates 3863 // (dirty cards). 3864 // One, admittedly dumb, strategy is to give up 3865 // after a certain number of abortable precleaning loops 3866 // or after a certain maximum time. We want to make 3867 // this smarter in the next iteration. 3868 // XXX FIX ME!!! YSR 3869 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3870 while (!(should_abort_preclean() || 3871 ConcurrentMarkSweepThread::should_terminate())) { 3872 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3873 cumworkdone += workdone; 3874 loops++; 3875 // Voluntarily terminate abortable preclean phase if we have 3876 // been at it for too long. 3877 if ((CMSMaxAbortablePrecleanLoops != 0) && 3878 loops >= CMSMaxAbortablePrecleanLoops) { 3879 if (PrintGCDetails) { 3880 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3881 } 3882 break; 3883 } 3884 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3885 if (PrintGCDetails) { 3886 gclog_or_tty->print(" CMS: abort preclean due to time "); 3887 } 3888 break; 3889 } 3890 // If we are doing little work each iteration, we should 3891 // take a short break. 3892 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3893 // Sleep for some time, waiting for work to accumulate 3894 stopTimer(); 3895 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3896 startTimer(); 3897 waited++; 3898 } 3899 } 3900 if (PrintCMSStatistics > 0) { 3901 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3902 loops, waited, cumworkdone); 3903 } 3904 } 3905 CMSTokenSync x(true); // is cms thread 3906 if (_collectorState != Idling) { 3907 assert(_collectorState == AbortablePreclean, 3908 "Spontaneous state transition?"); 3909 _collectorState = FinalMarking; 3910 } // Else, a foreground collection completed this CMS cycle. 3911 return; 3912 } 3913 3914 // Respond to an Eden sampling opportunity 3915 void CMSCollector::sample_eden() { 3916 // Make sure a young gc cannot sneak in between our 3917 // reading and recording of a sample. 3918 assert(Thread::current()->is_ConcurrentGC_thread(), 3919 "Only the cms thread may collect Eden samples"); 3920 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3921 "Should collect samples while holding CMS token"); 3922 if (!_start_sampling) { 3923 return; 3924 } 3925 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3926 // is populated by the young generation. 3927 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3928 if (_eden_chunk_index < _eden_chunk_capacity) { 3929 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3930 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3931 "Unexpected state of Eden"); 3932 // We'd like to check that what we just sampled is an oop-start address; 3933 // however, we cannot do that here since the object may not yet have been 3934 // initialized. So we'll instead do the check when we _use_ this sample 3935 // later. 3936 if (_eden_chunk_index == 0 || 3937 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3938 _eden_chunk_array[_eden_chunk_index-1]) 3939 >= CMSSamplingGrain)) { 3940 _eden_chunk_index++; // commit sample 3941 } 3942 } 3943 } 3944 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3945 size_t used = get_eden_used(); 3946 size_t capacity = get_eden_capacity(); 3947 assert(used <= capacity, "Unexpected state of Eden"); 3948 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3949 _abort_preclean = true; 3950 } 3951 } 3952 } 3953 3954 3955 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3956 assert(_collectorState == Precleaning || 3957 _collectorState == AbortablePreclean, "incorrect state"); 3958 ResourceMark rm; 3959 HandleMark hm; 3960 3961 // Precleaning is currently not MT but the reference processor 3962 // may be set for MT. Disable it temporarily here. 3963 ReferenceProcessor* rp = ref_processor(); 3964 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3965 3966 // Do one pass of scrubbing the discovered reference lists 3967 // to remove any reference objects with strongly-reachable 3968 // referents. 3969 if (clean_refs) { 3970 CMSPrecleanRefsYieldClosure yield_cl(this); 3971 assert(rp->span().equals(_span), "Spans should be equal"); 3972 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3973 &_markStack, true /* preclean */); 3974 CMSDrainMarkingStackClosure complete_trace(this, 3975 _span, &_markBitMap, &_markStack, 3976 &keep_alive, true /* preclean */); 3977 3978 // We don't want this step to interfere with a young 3979 // collection because we don't want to take CPU 3980 // or memory bandwidth away from the young GC threads 3981 // (which may be as many as there are CPUs). 3982 // Note that we don't need to protect ourselves from 3983 // interference with mutators because they can't 3984 // manipulate the discovered reference lists nor affect 3985 // the computed reachability of the referents, the 3986 // only properties manipulated by the precleaning 3987 // of these reference lists. 3988 stopTimer(); 3989 CMSTokenSyncWithLocks x(true /* is cms thread */, 3990 bitMapLock()); 3991 startTimer(); 3992 sample_eden(); 3993 3994 // The following will yield to allow foreground 3995 // collection to proceed promptly. XXX YSR: 3996 // The code in this method may need further 3997 // tweaking for better performance and some restructuring 3998 // for cleaner interfaces. 3999 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4000 rp->preclean_discovered_references( 4001 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4002 gc_timer, _gc_tracer_cm->gc_id()); 4003 } 4004 4005 if (clean_survivor) { // preclean the active survivor space(s) 4006 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4007 &_markBitMap, &_modUnionTable, 4008 &_markStack, true /* precleaning phase */); 4009 stopTimer(); 4010 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4011 bitMapLock()); 4012 startTimer(); 4013 unsigned int before_count = 4014 GenCollectedHeap::heap()->total_collections(); 4015 SurvivorSpacePrecleanClosure 4016 sss_cl(this, _span, &_markBitMap, &_markStack, 4017 &pam_cl, before_count, CMSYield); 4018 _young_gen->from()->object_iterate_careful(&sss_cl); 4019 _young_gen->to()->object_iterate_careful(&sss_cl); 4020 } 4021 MarkRefsIntoAndScanClosure 4022 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4023 &_markStack, this, CMSYield, 4024 true /* precleaning phase */); 4025 // CAUTION: The following closure has persistent state that may need to 4026 // be reset upon a decrease in the sequence of addresses it 4027 // processes. 4028 ScanMarkedObjectsAgainCarefullyClosure 4029 smoac_cl(this, _span, 4030 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4031 4032 // Preclean dirty cards in ModUnionTable and CardTable using 4033 // appropriate convergence criterion; 4034 // repeat CMSPrecleanIter times unless we find that 4035 // we are losing. 4036 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4037 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4038 "Bad convergence multiplier"); 4039 assert(CMSPrecleanThreshold >= 100, 4040 "Unreasonably low CMSPrecleanThreshold"); 4041 4042 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4043 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4044 numIter < CMSPrecleanIter; 4045 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4046 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4047 if (Verbose && PrintGCDetails) { 4048 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4049 } 4050 // Either there are very few dirty cards, so re-mark 4051 // pause will be small anyway, or our pre-cleaning isn't 4052 // that much faster than the rate at which cards are being 4053 // dirtied, so we might as well stop and re-mark since 4054 // precleaning won't improve our re-mark time by much. 4055 if (curNumCards <= CMSPrecleanThreshold || 4056 (numIter > 0 && 4057 (curNumCards * CMSPrecleanDenominator > 4058 lastNumCards * CMSPrecleanNumerator))) { 4059 numIter++; 4060 cumNumCards += curNumCards; 4061 break; 4062 } 4063 } 4064 4065 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4066 4067 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4068 cumNumCards += curNumCards; 4069 if (PrintGCDetails && PrintCMSStatistics != 0) { 4070 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4071 curNumCards, cumNumCards, numIter); 4072 } 4073 return cumNumCards; // as a measure of useful work done 4074 } 4075 4076 // PRECLEANING NOTES: 4077 // Precleaning involves: 4078 // . reading the bits of the modUnionTable and clearing the set bits. 4079 // . For the cards corresponding to the set bits, we scan the 4080 // objects on those cards. This means we need the free_list_lock 4081 // so that we can safely iterate over the CMS space when scanning 4082 // for oops. 4083 // . When we scan the objects, we'll be both reading and setting 4084 // marks in the marking bit map, so we'll need the marking bit map. 4085 // . For protecting _collector_state transitions, we take the CGC_lock. 4086 // Note that any races in the reading of of card table entries by the 4087 // CMS thread on the one hand and the clearing of those entries by the 4088 // VM thread or the setting of those entries by the mutator threads on the 4089 // other are quite benign. However, for efficiency it makes sense to keep 4090 // the VM thread from racing with the CMS thread while the latter is 4091 // dirty card info to the modUnionTable. We therefore also use the 4092 // CGC_lock to protect the reading of the card table and the mod union 4093 // table by the CM thread. 4094 // . We run concurrently with mutator updates, so scanning 4095 // needs to be done carefully -- we should not try to scan 4096 // potentially uninitialized objects. 4097 // 4098 // Locking strategy: While holding the CGC_lock, we scan over and 4099 // reset a maximal dirty range of the mod union / card tables, then lock 4100 // the free_list_lock and bitmap lock to do a full marking, then 4101 // release these locks; and repeat the cycle. This allows for a 4102 // certain amount of fairness in the sharing of these locks between 4103 // the CMS collector on the one hand, and the VM thread and the 4104 // mutators on the other. 4105 4106 // NOTE: preclean_mod_union_table() and preclean_card_table() 4107 // further below are largely identical; if you need to modify 4108 // one of these methods, please check the other method too. 4109 4110 size_t CMSCollector::preclean_mod_union_table( 4111 ConcurrentMarkSweepGeneration* gen, 4112 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4113 verify_work_stacks_empty(); 4114 verify_overflow_empty(); 4115 4116 // strategy: starting with the first card, accumulate contiguous 4117 // ranges of dirty cards; clear these cards, then scan the region 4118 // covered by these cards. 4119 4120 // Since all of the MUT is committed ahead, we can just use 4121 // that, in case the generations expand while we are precleaning. 4122 // It might also be fine to just use the committed part of the 4123 // generation, but we might potentially miss cards when the 4124 // generation is rapidly expanding while we are in the midst 4125 // of precleaning. 4126 HeapWord* startAddr = gen->reserved().start(); 4127 HeapWord* endAddr = gen->reserved().end(); 4128 4129 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4130 4131 size_t numDirtyCards, cumNumDirtyCards; 4132 HeapWord *nextAddr, *lastAddr; 4133 for (cumNumDirtyCards = numDirtyCards = 0, 4134 nextAddr = lastAddr = startAddr; 4135 nextAddr < endAddr; 4136 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4137 4138 ResourceMark rm; 4139 HandleMark hm; 4140 4141 MemRegion dirtyRegion; 4142 { 4143 stopTimer(); 4144 // Potential yield point 4145 CMSTokenSync ts(true); 4146 startTimer(); 4147 sample_eden(); 4148 // Get dirty region starting at nextOffset (inclusive), 4149 // simultaneously clearing it. 4150 dirtyRegion = 4151 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4152 assert(dirtyRegion.start() >= nextAddr, 4153 "returned region inconsistent?"); 4154 } 4155 // Remember where the next search should begin. 4156 // The returned region (if non-empty) is a right open interval, 4157 // so lastOffset is obtained from the right end of that 4158 // interval. 4159 lastAddr = dirtyRegion.end(); 4160 // Should do something more transparent and less hacky XXX 4161 numDirtyCards = 4162 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4163 4164 // We'll scan the cards in the dirty region (with periodic 4165 // yields for foreground GC as needed). 4166 if (!dirtyRegion.is_empty()) { 4167 assert(numDirtyCards > 0, "consistency check"); 4168 HeapWord* stop_point = NULL; 4169 stopTimer(); 4170 // Potential yield point 4171 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4172 bitMapLock()); 4173 startTimer(); 4174 { 4175 verify_work_stacks_empty(); 4176 verify_overflow_empty(); 4177 sample_eden(); 4178 stop_point = 4179 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4180 } 4181 if (stop_point != NULL) { 4182 // The careful iteration stopped early either because it found an 4183 // uninitialized object, or because we were in the midst of an 4184 // "abortable preclean", which should now be aborted. Redirty 4185 // the bits corresponding to the partially-scanned or unscanned 4186 // cards. We'll either restart at the next block boundary or 4187 // abort the preclean. 4188 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4189 "Should only be AbortablePreclean."); 4190 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4191 if (should_abort_preclean()) { 4192 break; // out of preclean loop 4193 } else { 4194 // Compute the next address at which preclean should pick up; 4195 // might need bitMapLock in order to read P-bits. 4196 lastAddr = next_card_start_after_block(stop_point); 4197 } 4198 } 4199 } else { 4200 assert(lastAddr == endAddr, "consistency check"); 4201 assert(numDirtyCards == 0, "consistency check"); 4202 break; 4203 } 4204 } 4205 verify_work_stacks_empty(); 4206 verify_overflow_empty(); 4207 return cumNumDirtyCards; 4208 } 4209 4210 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4211 // below are largely identical; if you need to modify 4212 // one of these methods, please check the other method too. 4213 4214 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4215 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4216 // strategy: it's similar to precleamModUnionTable above, in that 4217 // we accumulate contiguous ranges of dirty cards, mark these cards 4218 // precleaned, then scan the region covered by these cards. 4219 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4220 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4221 4222 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4223 4224 size_t numDirtyCards, cumNumDirtyCards; 4225 HeapWord *lastAddr, *nextAddr; 4226 4227 for (cumNumDirtyCards = numDirtyCards = 0, 4228 nextAddr = lastAddr = startAddr; 4229 nextAddr < endAddr; 4230 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4231 4232 ResourceMark rm; 4233 HandleMark hm; 4234 4235 MemRegion dirtyRegion; 4236 { 4237 // See comments in "Precleaning notes" above on why we 4238 // do this locking. XXX Could the locking overheads be 4239 // too high when dirty cards are sparse? [I don't think so.] 4240 stopTimer(); 4241 CMSTokenSync x(true); // is cms thread 4242 startTimer(); 4243 sample_eden(); 4244 // Get and clear dirty region from card table 4245 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4246 MemRegion(nextAddr, endAddr), 4247 true, 4248 CardTableModRefBS::precleaned_card_val()); 4249 4250 assert(dirtyRegion.start() >= nextAddr, 4251 "returned region inconsistent?"); 4252 } 4253 lastAddr = dirtyRegion.end(); 4254 numDirtyCards = 4255 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4256 4257 if (!dirtyRegion.is_empty()) { 4258 stopTimer(); 4259 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4260 startTimer(); 4261 sample_eden(); 4262 verify_work_stacks_empty(); 4263 verify_overflow_empty(); 4264 HeapWord* stop_point = 4265 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4266 if (stop_point != NULL) { 4267 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4268 "Should only be AbortablePreclean."); 4269 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4270 if (should_abort_preclean()) { 4271 break; // out of preclean loop 4272 } else { 4273 // Compute the next address at which preclean should pick up. 4274 lastAddr = next_card_start_after_block(stop_point); 4275 } 4276 } 4277 } else { 4278 break; 4279 } 4280 } 4281 verify_work_stacks_empty(); 4282 verify_overflow_empty(); 4283 return cumNumDirtyCards; 4284 } 4285 4286 class PrecleanKlassClosure : public KlassClosure { 4287 KlassToOopClosure _cm_klass_closure; 4288 public: 4289 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4290 void do_klass(Klass* k) { 4291 if (k->has_accumulated_modified_oops()) { 4292 k->clear_accumulated_modified_oops(); 4293 4294 _cm_klass_closure.do_klass(k); 4295 } 4296 } 4297 }; 4298 4299 // The freelist lock is needed to prevent asserts, is it really needed? 4300 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4301 4302 cl->set_freelistLock(freelistLock); 4303 4304 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4305 4306 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4307 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4308 PrecleanKlassClosure preclean_klass_closure(cl); 4309 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4310 4311 verify_work_stacks_empty(); 4312 verify_overflow_empty(); 4313 } 4314 4315 void CMSCollector::checkpointRootsFinal() { 4316 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4317 check_correct_thread_executing(); 4318 // world is stopped at this checkpoint 4319 assert(SafepointSynchronize::is_at_safepoint(), 4320 "world should be stopped"); 4321 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4322 4323 verify_work_stacks_empty(); 4324 verify_overflow_empty(); 4325 4326 if (PrintGCDetails) { 4327 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4328 _young_gen->used() / K, 4329 _young_gen->capacity() / K); 4330 } 4331 { 4332 if (CMSScavengeBeforeRemark) { 4333 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4334 // Temporarily set flag to false, GCH->do_collection will 4335 // expect it to be false and set to true 4336 FlagSetting fl(gch->_is_gc_active, false); 4337 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4338 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4339 int level = _cmsGen->level() - 1; 4340 if (level >= 0) { 4341 gch->do_collection(true, // full (i.e. force, see below) 4342 false, // !clear_all_soft_refs 4343 0, // size 4344 false, // is_tlab 4345 level // max_level 4346 ); 4347 } 4348 } 4349 FreelistLocker x(this); 4350 MutexLockerEx y(bitMapLock(), 4351 Mutex::_no_safepoint_check_flag); 4352 checkpointRootsFinalWork(); 4353 } 4354 verify_work_stacks_empty(); 4355 verify_overflow_empty(); 4356 } 4357 4358 void CMSCollector::checkpointRootsFinalWork() { 4359 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4360 4361 assert(haveFreelistLocks(), "must have free list locks"); 4362 assert_lock_strong(bitMapLock()); 4363 4364 ResourceMark rm; 4365 HandleMark hm; 4366 4367 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4368 4369 if (should_unload_classes()) { 4370 CodeCache::gc_prologue(); 4371 } 4372 assert(haveFreelistLocks(), "must have free list locks"); 4373 assert_lock_strong(bitMapLock()); 4374 4375 // We might assume that we need not fill TLAB's when 4376 // CMSScavengeBeforeRemark is set, because we may have just done 4377 // a scavenge which would have filled all TLAB's -- and besides 4378 // Eden would be empty. This however may not always be the case -- 4379 // for instance although we asked for a scavenge, it may not have 4380 // happened because of a JNI critical section. We probably need 4381 // a policy for deciding whether we can in that case wait until 4382 // the critical section releases and then do the remark following 4383 // the scavenge, and skip it here. In the absence of that policy, 4384 // or of an indication of whether the scavenge did indeed occur, 4385 // we cannot rely on TLAB's having been filled and must do 4386 // so here just in case a scavenge did not happen. 4387 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4388 // Update the saved marks which may affect the root scans. 4389 gch->save_marks(); 4390 4391 if (CMSPrintEdenSurvivorChunks) { 4392 print_eden_and_survivor_chunk_arrays(); 4393 } 4394 4395 { 4396 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4397 4398 // Note on the role of the mod union table: 4399 // Since the marker in "markFromRoots" marks concurrently with 4400 // mutators, it is possible for some reachable objects not to have been 4401 // scanned. For instance, an only reference to an object A was 4402 // placed in object B after the marker scanned B. Unless B is rescanned, 4403 // A would be collected. Such updates to references in marked objects 4404 // are detected via the mod union table which is the set of all cards 4405 // dirtied since the first checkpoint in this GC cycle and prior to 4406 // the most recent young generation GC, minus those cleaned up by the 4407 // concurrent precleaning. 4408 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4409 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4410 do_remark_parallel(); 4411 } else { 4412 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4413 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4414 do_remark_non_parallel(); 4415 } 4416 } 4417 verify_work_stacks_empty(); 4418 verify_overflow_empty(); 4419 4420 { 4421 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4422 refProcessingWork(); 4423 } 4424 verify_work_stacks_empty(); 4425 verify_overflow_empty(); 4426 4427 if (should_unload_classes()) { 4428 CodeCache::gc_epilogue(); 4429 } 4430 JvmtiExport::gc_epilogue(); 4431 4432 // If we encountered any (marking stack / work queue) overflow 4433 // events during the current CMS cycle, take appropriate 4434 // remedial measures, where possible, so as to try and avoid 4435 // recurrence of that condition. 4436 assert(_markStack.isEmpty(), "No grey objects"); 4437 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4438 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4439 if (ser_ovflw > 0) { 4440 if (PrintCMSStatistics != 0) { 4441 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4442 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4443 ", kac_preclean="SIZE_FORMAT")", 4444 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4445 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4446 } 4447 _markStack.expand(); 4448 _ser_pmc_remark_ovflw = 0; 4449 _ser_pmc_preclean_ovflw = 0; 4450 _ser_kac_preclean_ovflw = 0; 4451 _ser_kac_ovflw = 0; 4452 } 4453 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4454 if (PrintCMSStatistics != 0) { 4455 gclog_or_tty->print_cr("Work queue overflow (benign) " 4456 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4457 _par_pmc_remark_ovflw, _par_kac_ovflw); 4458 } 4459 _par_pmc_remark_ovflw = 0; 4460 _par_kac_ovflw = 0; 4461 } 4462 if (PrintCMSStatistics != 0) { 4463 if (_markStack._hit_limit > 0) { 4464 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4465 _markStack._hit_limit); 4466 } 4467 if (_markStack._failed_double > 0) { 4468 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4469 " current capacity "SIZE_FORMAT, 4470 _markStack._failed_double, 4471 _markStack.capacity()); 4472 } 4473 } 4474 _markStack._hit_limit = 0; 4475 _markStack._failed_double = 0; 4476 4477 if ((VerifyAfterGC || VerifyDuringGC) && 4478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4479 verify_after_remark(); 4480 } 4481 4482 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4483 4484 // Change under the freelistLocks. 4485 _collectorState = Sweeping; 4486 // Call isAllClear() under bitMapLock 4487 assert(_modUnionTable.isAllClear(), 4488 "Should be clear by end of the final marking"); 4489 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4490 "Should be clear by end of the final marking"); 4491 } 4492 4493 void CMSParInitialMarkTask::work(uint worker_id) { 4494 elapsedTimer _timer; 4495 ResourceMark rm; 4496 HandleMark hm; 4497 4498 // ---------- scan from roots -------------- 4499 _timer.start(); 4500 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4501 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4502 4503 // ---------- young gen roots -------------- 4504 { 4505 work_on_young_gen_roots(worker_id, &par_mri_cl); 4506 _timer.stop(); 4507 if (PrintCMSStatistics != 0) { 4508 gclog_or_tty->print_cr( 4509 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4510 worker_id, _timer.seconds()); 4511 } 4512 } 4513 4514 // ---------- remaining roots -------------- 4515 _timer.reset(); 4516 _timer.start(); 4517 4518 CLDToOopClosure cld_closure(&par_mri_cl, true); 4519 4520 gch->gen_process_roots(_collector->_cmsGen->level(), 4521 false, // yg was scanned above 4522 false, // this is parallel code 4523 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4524 _collector->should_unload_classes(), 4525 &par_mri_cl, 4526 NULL, 4527 &cld_closure); 4528 assert(_collector->should_unload_classes() 4529 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4530 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4531 _timer.stop(); 4532 if (PrintCMSStatistics != 0) { 4533 gclog_or_tty->print_cr( 4534 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4535 worker_id, _timer.seconds()); 4536 } 4537 } 4538 4539 // Parallel remark task 4540 class CMSParRemarkTask: public CMSParMarkTask { 4541 CompactibleFreeListSpace* _cms_space; 4542 4543 // The per-thread work queues, available here for stealing. 4544 OopTaskQueueSet* _task_queues; 4545 ParallelTaskTerminator _term; 4546 4547 public: 4548 // A value of 0 passed to n_workers will cause the number of 4549 // workers to be taken from the active workers in the work gang. 4550 CMSParRemarkTask(CMSCollector* collector, 4551 CompactibleFreeListSpace* cms_space, 4552 int n_workers, FlexibleWorkGang* workers, 4553 OopTaskQueueSet* task_queues): 4554 CMSParMarkTask("Rescan roots and grey objects in parallel", 4555 collector, n_workers), 4556 _cms_space(cms_space), 4557 _task_queues(task_queues), 4558 _term(n_workers, task_queues) { } 4559 4560 OopTaskQueueSet* task_queues() { return _task_queues; } 4561 4562 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4563 4564 ParallelTaskTerminator* terminator() { return &_term; } 4565 int n_workers() { return _n_workers; } 4566 4567 void work(uint worker_id); 4568 4569 private: 4570 // ... of dirty cards in old space 4571 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4572 Par_MarkRefsIntoAndScanClosure* cl); 4573 4574 // ... work stealing for the above 4575 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4576 }; 4577 4578 class RemarkKlassClosure : public KlassClosure { 4579 KlassToOopClosure _cm_klass_closure; 4580 public: 4581 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4582 void do_klass(Klass* k) { 4583 // Check if we have modified any oops in the Klass during the concurrent marking. 4584 if (k->has_accumulated_modified_oops()) { 4585 k->clear_accumulated_modified_oops(); 4586 4587 // We could have transfered the current modified marks to the accumulated marks, 4588 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4589 } else if (k->has_modified_oops()) { 4590 // Don't clear anything, this info is needed by the next young collection. 4591 } else { 4592 // No modified oops in the Klass. 4593 return; 4594 } 4595 4596 // The klass has modified fields, need to scan the klass. 4597 _cm_klass_closure.do_klass(k); 4598 } 4599 }; 4600 4601 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4602 ParNewGeneration* young_gen = _collector->_young_gen; 4603 ContiguousSpace* eden_space = young_gen->eden(); 4604 ContiguousSpace* from_space = young_gen->from(); 4605 ContiguousSpace* to_space = young_gen->to(); 4606 4607 HeapWord** eca = _collector->_eden_chunk_array; 4608 size_t ect = _collector->_eden_chunk_index; 4609 HeapWord** sca = _collector->_survivor_chunk_array; 4610 size_t sct = _collector->_survivor_chunk_index; 4611 4612 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4613 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4614 4615 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4616 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4617 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4618 } 4619 4620 // work_queue(i) is passed to the closure 4621 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4622 // also is passed to do_dirty_card_rescan_tasks() and to 4623 // do_work_steal() to select the i-th task_queue. 4624 4625 void CMSParRemarkTask::work(uint worker_id) { 4626 elapsedTimer _timer; 4627 ResourceMark rm; 4628 HandleMark hm; 4629 4630 // ---------- rescan from roots -------------- 4631 _timer.start(); 4632 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4633 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4634 _collector->_span, _collector->ref_processor(), 4635 &(_collector->_markBitMap), 4636 work_queue(worker_id)); 4637 4638 // Rescan young gen roots first since these are likely 4639 // coarsely partitioned and may, on that account, constitute 4640 // the critical path; thus, it's best to start off that 4641 // work first. 4642 // ---------- young gen roots -------------- 4643 { 4644 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4645 _timer.stop(); 4646 if (PrintCMSStatistics != 0) { 4647 gclog_or_tty->print_cr( 4648 "Finished young gen rescan work in %dth thread: %3.3f sec", 4649 worker_id, _timer.seconds()); 4650 } 4651 } 4652 4653 // ---------- remaining roots -------------- 4654 _timer.reset(); 4655 _timer.start(); 4656 gch->gen_process_roots(_collector->_cmsGen->level(), 4657 false, // yg was scanned above 4658 false, // this is parallel code 4659 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4660 _collector->should_unload_classes(), 4661 &par_mrias_cl, 4662 NULL, 4663 NULL); // The dirty klasses will be handled below 4664 4665 assert(_collector->should_unload_classes() 4666 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4667 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4668 _timer.stop(); 4669 if (PrintCMSStatistics != 0) { 4670 gclog_or_tty->print_cr( 4671 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4672 worker_id, _timer.seconds()); 4673 } 4674 4675 // ---------- unhandled CLD scanning ---------- 4676 if (worker_id == 0) { // Single threaded at the moment. 4677 _timer.reset(); 4678 _timer.start(); 4679 4680 // Scan all new class loader data objects and new dependencies that were 4681 // introduced during concurrent marking. 4682 ResourceMark rm; 4683 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4684 for (int i = 0; i < array->length(); i++) { 4685 par_mrias_cl.do_class_loader_data(array->at(i)); 4686 } 4687 4688 // We don't need to keep track of new CLDs anymore. 4689 ClassLoaderDataGraph::remember_new_clds(false); 4690 4691 _timer.stop(); 4692 if (PrintCMSStatistics != 0) { 4693 gclog_or_tty->print_cr( 4694 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4695 worker_id, _timer.seconds()); 4696 } 4697 } 4698 4699 // ---------- dirty klass scanning ---------- 4700 if (worker_id == 0) { // Single threaded at the moment. 4701 _timer.reset(); 4702 _timer.start(); 4703 4704 // Scan all classes that was dirtied during the concurrent marking phase. 4705 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4706 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4707 4708 _timer.stop(); 4709 if (PrintCMSStatistics != 0) { 4710 gclog_or_tty->print_cr( 4711 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4712 worker_id, _timer.seconds()); 4713 } 4714 } 4715 4716 // We might have added oops to ClassLoaderData::_handles during the 4717 // concurrent marking phase. These oops point to newly allocated objects 4718 // that are guaranteed to be kept alive. Either by the direct allocation 4719 // code, or when the young collector processes the roots. Hence, 4720 // we don't have to revisit the _handles block during the remark phase. 4721 4722 // ---------- rescan dirty cards ------------ 4723 _timer.reset(); 4724 _timer.start(); 4725 4726 // Do the rescan tasks for each of the two spaces 4727 // (cms_space) in turn. 4728 // "worker_id" is passed to select the task_queue for "worker_id" 4729 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4730 _timer.stop(); 4731 if (PrintCMSStatistics != 0) { 4732 gclog_or_tty->print_cr( 4733 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4734 worker_id, _timer.seconds()); 4735 } 4736 4737 // ---------- steal work from other threads ... 4738 // ---------- ... and drain overflow list. 4739 _timer.reset(); 4740 _timer.start(); 4741 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4742 _timer.stop(); 4743 if (PrintCMSStatistics != 0) { 4744 gclog_or_tty->print_cr( 4745 "Finished work stealing in %dth thread: %3.3f sec", 4746 worker_id, _timer.seconds()); 4747 } 4748 } 4749 4750 // Note that parameter "i" is not used. 4751 void 4752 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4753 OopsInGenClosure* cl, ContiguousSpace* space, 4754 HeapWord** chunk_array, size_t chunk_top) { 4755 // Until all tasks completed: 4756 // . claim an unclaimed task 4757 // . compute region boundaries corresponding to task claimed 4758 // using chunk_array 4759 // . par_oop_iterate(cl) over that region 4760 4761 ResourceMark rm; 4762 HandleMark hm; 4763 4764 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4765 4766 uint nth_task = 0; 4767 uint n_tasks = pst->n_tasks(); 4768 4769 if (n_tasks > 0) { 4770 assert(pst->valid(), "Uninitialized use?"); 4771 HeapWord *start, *end; 4772 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4773 // We claimed task # nth_task; compute its boundaries. 4774 if (chunk_top == 0) { // no samples were taken 4775 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4776 start = space->bottom(); 4777 end = space->top(); 4778 } else if (nth_task == 0) { 4779 start = space->bottom(); 4780 end = chunk_array[nth_task]; 4781 } else if (nth_task < (uint)chunk_top) { 4782 assert(nth_task >= 1, "Control point invariant"); 4783 start = chunk_array[nth_task - 1]; 4784 end = chunk_array[nth_task]; 4785 } else { 4786 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4787 start = chunk_array[chunk_top - 1]; 4788 end = space->top(); 4789 } 4790 MemRegion mr(start, end); 4791 // Verify that mr is in space 4792 assert(mr.is_empty() || space->used_region().contains(mr), 4793 "Should be in space"); 4794 // Verify that "start" is an object boundary 4795 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4796 "Should be an oop"); 4797 space->par_oop_iterate(mr, cl); 4798 } 4799 pst->all_tasks_completed(); 4800 } 4801 } 4802 4803 void 4804 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4805 CompactibleFreeListSpace* sp, int i, 4806 Par_MarkRefsIntoAndScanClosure* cl) { 4807 // Until all tasks completed: 4808 // . claim an unclaimed task 4809 // . compute region boundaries corresponding to task claimed 4810 // . transfer dirty bits ct->mut for that region 4811 // . apply rescanclosure to dirty mut bits for that region 4812 4813 ResourceMark rm; 4814 HandleMark hm; 4815 4816 OopTaskQueue* work_q = work_queue(i); 4817 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4818 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4819 // CAUTION: This closure has state that persists across calls to 4820 // the work method dirty_range_iterate_clear() in that it has 4821 // embedded in it a (subtype of) UpwardsObjectClosure. The 4822 // use of that state in the embedded UpwardsObjectClosure instance 4823 // assumes that the cards are always iterated (even if in parallel 4824 // by several threads) in monotonically increasing order per each 4825 // thread. This is true of the implementation below which picks 4826 // card ranges (chunks) in monotonically increasing order globally 4827 // and, a-fortiori, in monotonically increasing order per thread 4828 // (the latter order being a subsequence of the former). 4829 // If the work code below is ever reorganized into a more chaotic 4830 // work-partitioning form than the current "sequential tasks" 4831 // paradigm, the use of that persistent state will have to be 4832 // revisited and modified appropriately. See also related 4833 // bug 4756801 work on which should examine this code to make 4834 // sure that the changes there do not run counter to the 4835 // assumptions made here and necessary for correctness and 4836 // efficiency. Note also that this code might yield inefficient 4837 // behavior in the case of very large objects that span one or 4838 // more work chunks. Such objects would potentially be scanned 4839 // several times redundantly. Work on 4756801 should try and 4840 // address that performance anomaly if at all possible. XXX 4841 MemRegion full_span = _collector->_span; 4842 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4843 MarkFromDirtyCardsClosure 4844 greyRescanClosure(_collector, full_span, // entire span of interest 4845 sp, bm, work_q, cl); 4846 4847 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4848 assert(pst->valid(), "Uninitialized use?"); 4849 uint nth_task = 0; 4850 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4851 MemRegion span = sp->used_region(); 4852 HeapWord* start_addr = span.start(); 4853 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4854 alignment); 4855 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4856 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4857 start_addr, "Check alignment"); 4858 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4859 chunk_size, "Check alignment"); 4860 4861 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4862 // Having claimed the nth_task, compute corresponding mem-region, 4863 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4864 // The alignment restriction ensures that we do not need any 4865 // synchronization with other gang-workers while setting or 4866 // clearing bits in thus chunk of the MUT. 4867 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4868 start_addr + (nth_task+1)*chunk_size); 4869 // The last chunk's end might be way beyond end of the 4870 // used region. In that case pull back appropriately. 4871 if (this_span.end() > end_addr) { 4872 this_span.set_end(end_addr); 4873 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4874 } 4875 // Iterate over the dirty cards covering this chunk, marking them 4876 // precleaned, and setting the corresponding bits in the mod union 4877 // table. Since we have been careful to partition at Card and MUT-word 4878 // boundaries no synchronization is needed between parallel threads. 4879 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4880 &modUnionClosure); 4881 4882 // Having transferred these marks into the modUnionTable, 4883 // rescan the marked objects on the dirty cards in the modUnionTable. 4884 // Even if this is at a synchronous collection, the initial marking 4885 // may have been done during an asynchronous collection so there 4886 // may be dirty bits in the mod-union table. 4887 _collector->_modUnionTable.dirty_range_iterate_clear( 4888 this_span, &greyRescanClosure); 4889 _collector->_modUnionTable.verifyNoOneBitsInRange( 4890 this_span.start(), 4891 this_span.end()); 4892 } 4893 pst->all_tasks_completed(); // declare that i am done 4894 } 4895 4896 // . see if we can share work_queues with ParNew? XXX 4897 void 4898 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4899 int* seed) { 4900 OopTaskQueue* work_q = work_queue(i); 4901 NOT_PRODUCT(int num_steals = 0;) 4902 oop obj_to_scan; 4903 CMSBitMap* bm = &(_collector->_markBitMap); 4904 4905 while (true) { 4906 // Completely finish any left over work from (an) earlier round(s) 4907 cl->trim_queue(0); 4908 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4909 (size_t)ParGCDesiredObjsFromOverflowList); 4910 // Now check if there's any work in the overflow list 4911 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4912 // only affects the number of attempts made to get work from the 4913 // overflow list and does not affect the number of workers. Just 4914 // pass ParallelGCThreads so this behavior is unchanged. 4915 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4916 work_q, 4917 ParallelGCThreads)) { 4918 // found something in global overflow list; 4919 // not yet ready to go stealing work from others. 4920 // We'd like to assert(work_q->size() != 0, ...) 4921 // because we just took work from the overflow list, 4922 // but of course we can't since all of that could have 4923 // been already stolen from us. 4924 // "He giveth and He taketh away." 4925 continue; 4926 } 4927 // Verify that we have no work before we resort to stealing 4928 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4929 // Try to steal from other queues that have work 4930 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4931 NOT_PRODUCT(num_steals++;) 4932 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4933 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4934 // Do scanning work 4935 obj_to_scan->oop_iterate(cl); 4936 // Loop around, finish this work, and try to steal some more 4937 } else if (terminator()->offer_termination()) { 4938 break; // nirvana from the infinite cycle 4939 } 4940 } 4941 NOT_PRODUCT( 4942 if (PrintCMSStatistics != 0) { 4943 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4944 } 4945 ) 4946 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4947 "Else our work is not yet done"); 4948 } 4949 4950 // Record object boundaries in _eden_chunk_array by sampling the eden 4951 // top in the slow-path eden object allocation code path and record 4952 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4953 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4954 // sampling in sample_eden() that activates during the part of the 4955 // preclean phase. 4956 void CMSCollector::sample_eden_chunk() { 4957 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4958 if (_eden_chunk_lock->try_lock()) { 4959 // Record a sample. This is the critical section. The contents 4960 // of the _eden_chunk_array have to be non-decreasing in the 4961 // address order. 4962 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4963 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4964 "Unexpected state of Eden"); 4965 if (_eden_chunk_index == 0 || 4966 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4967 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4968 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4969 _eden_chunk_index++; // commit sample 4970 } 4971 _eden_chunk_lock->unlock(); 4972 } 4973 } 4974 } 4975 4976 // Return a thread-local PLAB recording array, as appropriate. 4977 void* CMSCollector::get_data_recorder(int thr_num) { 4978 if (_survivor_plab_array != NULL && 4979 (CMSPLABRecordAlways || 4980 (_collectorState > Marking && _collectorState < FinalMarking))) { 4981 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4982 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4983 ca->reset(); // clear it so that fresh data is recorded 4984 return (void*) ca; 4985 } else { 4986 return NULL; 4987 } 4988 } 4989 4990 // Reset all the thread-local PLAB recording arrays 4991 void CMSCollector::reset_survivor_plab_arrays() { 4992 for (uint i = 0; i < ParallelGCThreads; i++) { 4993 _survivor_plab_array[i].reset(); 4994 } 4995 } 4996 4997 // Merge the per-thread plab arrays into the global survivor chunk 4998 // array which will provide the partitioning of the survivor space 4999 // for CMS initial scan and rescan. 5000 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5001 int no_of_gc_threads) { 5002 assert(_survivor_plab_array != NULL, "Error"); 5003 assert(_survivor_chunk_array != NULL, "Error"); 5004 assert(_collectorState == FinalMarking || 5005 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5006 for (int j = 0; j < no_of_gc_threads; j++) { 5007 _cursor[j] = 0; 5008 } 5009 HeapWord* top = surv->top(); 5010 size_t i; 5011 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5012 HeapWord* min_val = top; // Higher than any PLAB address 5013 uint min_tid = 0; // position of min_val this round 5014 for (int j = 0; j < no_of_gc_threads; j++) { 5015 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5016 if (_cursor[j] == cur_sca->end()) { 5017 continue; 5018 } 5019 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5020 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5021 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5022 if (cur_val < min_val) { 5023 min_tid = j; 5024 min_val = cur_val; 5025 } else { 5026 assert(cur_val < top, "All recorded addresses should be less"); 5027 } 5028 } 5029 // At this point min_val and min_tid are respectively 5030 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5031 // and the thread (j) that witnesses that address. 5032 // We record this address in the _survivor_chunk_array[i] 5033 // and increment _cursor[min_tid] prior to the next round i. 5034 if (min_val == top) { 5035 break; 5036 } 5037 _survivor_chunk_array[i] = min_val; 5038 _cursor[min_tid]++; 5039 } 5040 // We are all done; record the size of the _survivor_chunk_array 5041 _survivor_chunk_index = i; // exclusive: [0, i) 5042 if (PrintCMSStatistics > 0) { 5043 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5044 } 5045 // Verify that we used up all the recorded entries 5046 #ifdef ASSERT 5047 size_t total = 0; 5048 for (int j = 0; j < no_of_gc_threads; j++) { 5049 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5050 total += _cursor[j]; 5051 } 5052 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5053 // Check that the merged array is in sorted order 5054 if (total > 0) { 5055 for (size_t i = 0; i < total - 1; i++) { 5056 if (PrintCMSStatistics > 0) { 5057 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5058 i, p2i(_survivor_chunk_array[i])); 5059 } 5060 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5061 "Not sorted"); 5062 } 5063 } 5064 #endif // ASSERT 5065 } 5066 5067 // Set up the space's par_seq_tasks structure for work claiming 5068 // for parallel initial scan and rescan of young gen. 5069 // See ParRescanTask where this is currently used. 5070 void 5071 CMSCollector:: 5072 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5073 assert(n_threads > 0, "Unexpected n_threads argument"); 5074 5075 // Eden space 5076 if (!_young_gen->eden()->is_empty()) { 5077 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 5078 assert(!pst->valid(), "Clobbering existing data?"); 5079 // Each valid entry in [0, _eden_chunk_index) represents a task. 5080 size_t n_tasks = _eden_chunk_index + 1; 5081 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5082 // Sets the condition for completion of the subtask (how many threads 5083 // need to finish in order to be done). 5084 pst->set_n_threads(n_threads); 5085 pst->set_n_tasks((int)n_tasks); 5086 } 5087 5088 // Merge the survivor plab arrays into _survivor_chunk_array 5089 if (_survivor_plab_array != NULL) { 5090 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5091 } else { 5092 assert(_survivor_chunk_index == 0, "Error"); 5093 } 5094 5095 // To space 5096 { 5097 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5098 assert(!pst->valid(), "Clobbering existing data?"); 5099 // Sets the condition for completion of the subtask (how many threads 5100 // need to finish in order to be done). 5101 pst->set_n_threads(n_threads); 5102 pst->set_n_tasks(1); 5103 assert(pst->valid(), "Error"); 5104 } 5105 5106 // From space 5107 { 5108 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5109 assert(!pst->valid(), "Clobbering existing data?"); 5110 size_t n_tasks = _survivor_chunk_index + 1; 5111 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5112 // Sets the condition for completion of the subtask (how many threads 5113 // need to finish in order to be done). 5114 pst->set_n_threads(n_threads); 5115 pst->set_n_tasks((int)n_tasks); 5116 assert(pst->valid(), "Error"); 5117 } 5118 } 5119 5120 // Parallel version of remark 5121 void CMSCollector::do_remark_parallel() { 5122 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5123 FlexibleWorkGang* workers = gch->workers(); 5124 assert(workers != NULL, "Need parallel worker threads."); 5125 // Choose to use the number of GC workers most recently set 5126 // into "active_workers". If active_workers is not set, set it 5127 // to ParallelGCThreads. 5128 int n_workers = workers->active_workers(); 5129 if (n_workers == 0) { 5130 assert(n_workers > 0, "Should have been set during scavenge"); 5131 n_workers = ParallelGCThreads; 5132 workers->set_active_workers(n_workers); 5133 } 5134 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5135 5136 CMSParRemarkTask tsk(this, 5137 cms_space, 5138 n_workers, workers, task_queues()); 5139 5140 // Set up for parallel process_roots work. 5141 gch->set_par_threads(n_workers); 5142 // We won't be iterating over the cards in the card table updating 5143 // the younger_gen cards, so we shouldn't call the following else 5144 // the verification code as well as subsequent younger_refs_iterate 5145 // code would get confused. XXX 5146 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5147 5148 // The young gen rescan work will not be done as part of 5149 // process_roots (which currently doesn't know how to 5150 // parallelize such a scan), but rather will be broken up into 5151 // a set of parallel tasks (via the sampling that the [abortable] 5152 // preclean phase did of eden, plus the [two] tasks of 5153 // scanning the [two] survivor spaces. Further fine-grain 5154 // parallelization of the scanning of the survivor spaces 5155 // themselves, and of precleaning of the younger gen itself 5156 // is deferred to the future. 5157 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5158 5159 // The dirty card rescan work is broken up into a "sequence" 5160 // of parallel tasks (per constituent space) that are dynamically 5161 // claimed by the parallel threads. 5162 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5163 5164 // It turns out that even when we're using 1 thread, doing the work in a 5165 // separate thread causes wide variance in run times. We can't help this 5166 // in the multi-threaded case, but we special-case n=1 here to get 5167 // repeatable measurements of the 1-thread overhead of the parallel code. 5168 if (n_workers > 1) { 5169 // Make refs discovery MT-safe, if it isn't already: it may not 5170 // necessarily be so, since it's possible that we are doing 5171 // ST marking. 5172 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5173 StrongRootsScope srs; 5174 workers->run_task(&tsk); 5175 } else { 5176 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5177 StrongRootsScope srs; 5178 tsk.work(0); 5179 } 5180 5181 gch->set_par_threads(0); // 0 ==> non-parallel. 5182 // restore, single-threaded for now, any preserved marks 5183 // as a result of work_q overflow 5184 restore_preserved_marks_if_any(); 5185 } 5186 5187 // Non-parallel version of remark 5188 void CMSCollector::do_remark_non_parallel() { 5189 ResourceMark rm; 5190 HandleMark hm; 5191 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5192 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5193 5194 MarkRefsIntoAndScanClosure 5195 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5196 &_markStack, this, 5197 false /* should_yield */, false /* not precleaning */); 5198 MarkFromDirtyCardsClosure 5199 markFromDirtyCardsClosure(this, _span, 5200 NULL, // space is set further below 5201 &_markBitMap, &_markStack, &mrias_cl); 5202 { 5203 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5204 // Iterate over the dirty cards, setting the corresponding bits in the 5205 // mod union table. 5206 { 5207 ModUnionClosure modUnionClosure(&_modUnionTable); 5208 _ct->ct_bs()->dirty_card_iterate( 5209 _cmsGen->used_region(), 5210 &modUnionClosure); 5211 } 5212 // Having transferred these marks into the modUnionTable, we just need 5213 // to rescan the marked objects on the dirty cards in the modUnionTable. 5214 // The initial marking may have been done during an asynchronous 5215 // collection so there may be dirty bits in the mod-union table. 5216 const int alignment = 5217 CardTableModRefBS::card_size * BitsPerWord; 5218 { 5219 // ... First handle dirty cards in CMS gen 5220 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5221 MemRegion ur = _cmsGen->used_region(); 5222 HeapWord* lb = ur.start(); 5223 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5224 MemRegion cms_span(lb, ub); 5225 _modUnionTable.dirty_range_iterate_clear(cms_span, 5226 &markFromDirtyCardsClosure); 5227 verify_work_stacks_empty(); 5228 if (PrintCMSStatistics != 0) { 5229 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5230 markFromDirtyCardsClosure.num_dirty_cards()); 5231 } 5232 } 5233 } 5234 if (VerifyDuringGC && 5235 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5236 HandleMark hm; // Discard invalid handles created during verification 5237 Universe::verify(); 5238 } 5239 { 5240 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5241 5242 verify_work_stacks_empty(); 5243 5244 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5245 StrongRootsScope srs; 5246 5247 gch->gen_process_roots(_cmsGen->level(), 5248 true, // younger gens as roots 5249 false, // use the local StrongRootsScope 5250 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5251 should_unload_classes(), 5252 &mrias_cl, 5253 NULL, 5254 NULL); // The dirty klasses will be handled below 5255 5256 assert(should_unload_classes() 5257 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5258 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5259 } 5260 5261 { 5262 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5263 5264 verify_work_stacks_empty(); 5265 5266 // Scan all class loader data objects that might have been introduced 5267 // during concurrent marking. 5268 ResourceMark rm; 5269 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5270 for (int i = 0; i < array->length(); i++) { 5271 mrias_cl.do_class_loader_data(array->at(i)); 5272 } 5273 5274 // We don't need to keep track of new CLDs anymore. 5275 ClassLoaderDataGraph::remember_new_clds(false); 5276 5277 verify_work_stacks_empty(); 5278 } 5279 5280 { 5281 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5282 5283 verify_work_stacks_empty(); 5284 5285 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5286 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5287 5288 verify_work_stacks_empty(); 5289 } 5290 5291 // We might have added oops to ClassLoaderData::_handles during the 5292 // concurrent marking phase. These oops point to newly allocated objects 5293 // that are guaranteed to be kept alive. Either by the direct allocation 5294 // code, or when the young collector processes the roots. Hence, 5295 // we don't have to revisit the _handles block during the remark phase. 5296 5297 verify_work_stacks_empty(); 5298 // Restore evacuated mark words, if any, used for overflow list links 5299 if (!CMSOverflowEarlyRestoration) { 5300 restore_preserved_marks_if_any(); 5301 } 5302 verify_overflow_empty(); 5303 } 5304 5305 //////////////////////////////////////////////////////// 5306 // Parallel Reference Processing Task Proxy Class 5307 //////////////////////////////////////////////////////// 5308 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5309 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5310 CMSCollector* _collector; 5311 CMSBitMap* _mark_bit_map; 5312 const MemRegion _span; 5313 ProcessTask& _task; 5314 5315 public: 5316 CMSRefProcTaskProxy(ProcessTask& task, 5317 CMSCollector* collector, 5318 const MemRegion& span, 5319 CMSBitMap* mark_bit_map, 5320 AbstractWorkGang* workers, 5321 OopTaskQueueSet* task_queues): 5322 // XXX Should superclass AGTWOQ also know about AWG since it knows 5323 // about the task_queues used by the AWG? Then it could initialize 5324 // the terminator() object. See 6984287. The set_for_termination() 5325 // below is a temporary band-aid for the regression in 6984287. 5326 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5327 task_queues), 5328 _task(task), 5329 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5330 { 5331 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5332 "Inconsistency in _span"); 5333 set_for_termination(workers->active_workers()); 5334 } 5335 5336 OopTaskQueueSet* task_queues() { return queues(); } 5337 5338 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5339 5340 void do_work_steal(int i, 5341 CMSParDrainMarkingStackClosure* drain, 5342 CMSParKeepAliveClosure* keep_alive, 5343 int* seed); 5344 5345 virtual void work(uint worker_id); 5346 }; 5347 5348 void CMSRefProcTaskProxy::work(uint worker_id) { 5349 ResourceMark rm; 5350 HandleMark hm; 5351 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5352 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5353 _mark_bit_map, 5354 work_queue(worker_id)); 5355 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5356 _mark_bit_map, 5357 work_queue(worker_id)); 5358 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5359 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5360 if (_task.marks_oops_alive()) { 5361 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5362 _collector->hash_seed(worker_id)); 5363 } 5364 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5365 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5366 } 5367 5368 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5369 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5370 EnqueueTask& _task; 5371 5372 public: 5373 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5374 : AbstractGangTask("Enqueue reference objects in parallel"), 5375 _task(task) 5376 { } 5377 5378 virtual void work(uint worker_id) 5379 { 5380 _task.work(worker_id); 5381 } 5382 }; 5383 5384 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5385 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5386 _span(span), 5387 _bit_map(bit_map), 5388 _work_queue(work_queue), 5389 _mark_and_push(collector, span, bit_map, work_queue), 5390 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5391 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5392 { } 5393 5394 // . see if we can share work_queues with ParNew? XXX 5395 void CMSRefProcTaskProxy::do_work_steal(int i, 5396 CMSParDrainMarkingStackClosure* drain, 5397 CMSParKeepAliveClosure* keep_alive, 5398 int* seed) { 5399 OopTaskQueue* work_q = work_queue(i); 5400 NOT_PRODUCT(int num_steals = 0;) 5401 oop obj_to_scan; 5402 5403 while (true) { 5404 // Completely finish any left over work from (an) earlier round(s) 5405 drain->trim_queue(0); 5406 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5407 (size_t)ParGCDesiredObjsFromOverflowList); 5408 // Now check if there's any work in the overflow list 5409 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5410 // only affects the number of attempts made to get work from the 5411 // overflow list and does not affect the number of workers. Just 5412 // pass ParallelGCThreads so this behavior is unchanged. 5413 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5414 work_q, 5415 ParallelGCThreads)) { 5416 // Found something in global overflow list; 5417 // not yet ready to go stealing work from others. 5418 // We'd like to assert(work_q->size() != 0, ...) 5419 // because we just took work from the overflow list, 5420 // but of course we can't, since all of that might have 5421 // been already stolen from us. 5422 continue; 5423 } 5424 // Verify that we have no work before we resort to stealing 5425 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5426 // Try to steal from other queues that have work 5427 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5428 NOT_PRODUCT(num_steals++;) 5429 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5430 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5431 // Do scanning work 5432 obj_to_scan->oop_iterate(keep_alive); 5433 // Loop around, finish this work, and try to steal some more 5434 } else if (terminator()->offer_termination()) { 5435 break; // nirvana from the infinite cycle 5436 } 5437 } 5438 NOT_PRODUCT( 5439 if (PrintCMSStatistics != 0) { 5440 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5441 } 5442 ) 5443 } 5444 5445 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5446 { 5447 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5448 FlexibleWorkGang* workers = gch->workers(); 5449 assert(workers != NULL, "Need parallel worker threads."); 5450 CMSRefProcTaskProxy rp_task(task, &_collector, 5451 _collector.ref_processor()->span(), 5452 _collector.markBitMap(), 5453 workers, _collector.task_queues()); 5454 workers->run_task(&rp_task); 5455 } 5456 5457 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5458 { 5459 5460 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5461 FlexibleWorkGang* workers = gch->workers(); 5462 assert(workers != NULL, "Need parallel worker threads."); 5463 CMSRefEnqueueTaskProxy enq_task(task); 5464 workers->run_task(&enq_task); 5465 } 5466 5467 void CMSCollector::refProcessingWork() { 5468 ResourceMark rm; 5469 HandleMark hm; 5470 5471 ReferenceProcessor* rp = ref_processor(); 5472 assert(rp->span().equals(_span), "Spans should be equal"); 5473 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5474 // Process weak references. 5475 rp->setup_policy(false); 5476 verify_work_stacks_empty(); 5477 5478 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5479 &_markStack, false /* !preclean */); 5480 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5481 _span, &_markBitMap, &_markStack, 5482 &cmsKeepAliveClosure, false /* !preclean */); 5483 { 5484 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5485 5486 ReferenceProcessorStats stats; 5487 if (rp->processing_is_mt()) { 5488 // Set the degree of MT here. If the discovery is done MT, there 5489 // may have been a different number of threads doing the discovery 5490 // and a different number of discovered lists may have Ref objects. 5491 // That is OK as long as the Reference lists are balanced (see 5492 // balance_all_queues() and balance_queues()). 5493 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5494 int active_workers = ParallelGCThreads; 5495 FlexibleWorkGang* workers = gch->workers(); 5496 if (workers != NULL) { 5497 active_workers = workers->active_workers(); 5498 // The expectation is that active_workers will have already 5499 // been set to a reasonable value. If it has not been set, 5500 // investigate. 5501 assert(active_workers > 0, "Should have been set during scavenge"); 5502 } 5503 rp->set_active_mt_degree(active_workers); 5504 CMSRefProcTaskExecutor task_executor(*this); 5505 stats = rp->process_discovered_references(&_is_alive_closure, 5506 &cmsKeepAliveClosure, 5507 &cmsDrainMarkingStackClosure, 5508 &task_executor, 5509 _gc_timer_cm, 5510 _gc_tracer_cm->gc_id()); 5511 } else { 5512 stats = rp->process_discovered_references(&_is_alive_closure, 5513 &cmsKeepAliveClosure, 5514 &cmsDrainMarkingStackClosure, 5515 NULL, 5516 _gc_timer_cm, 5517 _gc_tracer_cm->gc_id()); 5518 } 5519 _gc_tracer_cm->report_gc_reference_stats(stats); 5520 5521 } 5522 5523 // This is the point where the entire marking should have completed. 5524 verify_work_stacks_empty(); 5525 5526 if (should_unload_classes()) { 5527 { 5528 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5529 5530 // Unload classes and purge the SystemDictionary. 5531 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5532 5533 // Unload nmethods. 5534 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5535 5536 // Prune dead klasses from subklass/sibling/implementor lists. 5537 Klass::clean_weak_klass_links(&_is_alive_closure); 5538 } 5539 5540 { 5541 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5542 // Clean up unreferenced symbols in symbol table. 5543 SymbolTable::unlink(); 5544 } 5545 5546 { 5547 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5548 // Delete entries for dead interned strings. 5549 StringTable::unlink(&_is_alive_closure); 5550 } 5551 } 5552 5553 5554 // Restore any preserved marks as a result of mark stack or 5555 // work queue overflow 5556 restore_preserved_marks_if_any(); // done single-threaded for now 5557 5558 rp->set_enqueuing_is_done(true); 5559 if (rp->processing_is_mt()) { 5560 rp->balance_all_queues(); 5561 CMSRefProcTaskExecutor task_executor(*this); 5562 rp->enqueue_discovered_references(&task_executor); 5563 } else { 5564 rp->enqueue_discovered_references(NULL); 5565 } 5566 rp->verify_no_references_recorded(); 5567 assert(!rp->discovery_enabled(), "should have been disabled"); 5568 } 5569 5570 #ifndef PRODUCT 5571 void CMSCollector::check_correct_thread_executing() { 5572 Thread* t = Thread::current(); 5573 // Only the VM thread or the CMS thread should be here. 5574 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5575 "Unexpected thread type"); 5576 // If this is the vm thread, the foreground process 5577 // should not be waiting. Note that _foregroundGCIsActive is 5578 // true while the foreground collector is waiting. 5579 if (_foregroundGCShouldWait) { 5580 // We cannot be the VM thread 5581 assert(t->is_ConcurrentGC_thread(), 5582 "Should be CMS thread"); 5583 } else { 5584 // We can be the CMS thread only if we are in a stop-world 5585 // phase of CMS collection. 5586 if (t->is_ConcurrentGC_thread()) { 5587 assert(_collectorState == InitialMarking || 5588 _collectorState == FinalMarking, 5589 "Should be a stop-world phase"); 5590 // The CMS thread should be holding the CMS_token. 5591 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5592 "Potential interference with concurrently " 5593 "executing VM thread"); 5594 } 5595 } 5596 } 5597 #endif 5598 5599 void CMSCollector::sweep() { 5600 assert(_collectorState == Sweeping, "just checking"); 5601 check_correct_thread_executing(); 5602 verify_work_stacks_empty(); 5603 verify_overflow_empty(); 5604 increment_sweep_count(); 5605 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5606 5607 _inter_sweep_timer.stop(); 5608 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5609 5610 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5611 _intra_sweep_timer.reset(); 5612 _intra_sweep_timer.start(); 5613 { 5614 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5615 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5616 // First sweep the old gen 5617 { 5618 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5619 bitMapLock()); 5620 sweepWork(_cmsGen); 5621 } 5622 5623 // Update Universe::_heap_*_at_gc figures. 5624 // We need all the free list locks to make the abstract state 5625 // transition from Sweeping to Resetting. See detailed note 5626 // further below. 5627 { 5628 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5629 // Update heap occupancy information which is used as 5630 // input to soft ref clearing policy at the next gc. 5631 Universe::update_heap_info_at_gc(); 5632 _collectorState = Resizing; 5633 } 5634 } 5635 verify_work_stacks_empty(); 5636 verify_overflow_empty(); 5637 5638 if (should_unload_classes()) { 5639 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5640 // requires that the virtual spaces are stable and not deleted. 5641 ClassLoaderDataGraph::set_should_purge(true); 5642 } 5643 5644 _intra_sweep_timer.stop(); 5645 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5646 5647 _inter_sweep_timer.reset(); 5648 _inter_sweep_timer.start(); 5649 5650 // We need to use a monotonically non-decreasing time in ms 5651 // or we will see time-warp warnings and os::javaTimeMillis() 5652 // does not guarantee monotonicity. 5653 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5654 update_time_of_last_gc(now); 5655 5656 // NOTE on abstract state transitions: 5657 // Mutators allocate-live and/or mark the mod-union table dirty 5658 // based on the state of the collection. The former is done in 5659 // the interval [Marking, Sweeping] and the latter in the interval 5660 // [Marking, Sweeping). Thus the transitions into the Marking state 5661 // and out of the Sweeping state must be synchronously visible 5662 // globally to the mutators. 5663 // The transition into the Marking state happens with the world 5664 // stopped so the mutators will globally see it. Sweeping is 5665 // done asynchronously by the background collector so the transition 5666 // from the Sweeping state to the Resizing state must be done 5667 // under the freelistLock (as is the check for whether to 5668 // allocate-live and whether to dirty the mod-union table). 5669 assert(_collectorState == Resizing, "Change of collector state to" 5670 " Resizing must be done under the freelistLocks (plural)"); 5671 5672 // Now that sweeping has been completed, we clear 5673 // the incremental_collection_failed flag, 5674 // thus inviting a younger gen collection to promote into 5675 // this generation. If such a promotion may still fail, 5676 // the flag will be set again when a young collection is 5677 // attempted. 5678 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5679 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5680 gch->update_full_collections_completed(_collection_count_start); 5681 } 5682 5683 // FIX ME!!! Looks like this belongs in CFLSpace, with 5684 // CMSGen merely delegating to it. 5685 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5686 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5687 HeapWord* minAddr = _cmsSpace->bottom(); 5688 HeapWord* largestAddr = 5689 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5690 if (largestAddr == NULL) { 5691 // The dictionary appears to be empty. In this case 5692 // try to coalesce at the end of the heap. 5693 largestAddr = _cmsSpace->end(); 5694 } 5695 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5696 size_t nearLargestOffset = 5697 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5698 if (PrintFLSStatistics != 0) { 5699 gclog_or_tty->print_cr( 5700 "CMS: Large Block: " PTR_FORMAT ";" 5701 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5702 p2i(largestAddr), 5703 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5704 } 5705 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5706 } 5707 5708 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5709 return addr >= _cmsSpace->nearLargestChunk(); 5710 } 5711 5712 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5713 return _cmsSpace->find_chunk_at_end(); 5714 } 5715 5716 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5717 bool full) { 5718 // The next lower level has been collected. Gather any statistics 5719 // that are of interest at this point. 5720 if (!full && (current_level + 1) == level()) { 5721 // Gather statistics on the young generation collection. 5722 collector()->stats().record_gc0_end(used()); 5723 } 5724 } 5725 5726 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5727 // We iterate over the space(s) underlying this generation, 5728 // checking the mark bit map to see if the bits corresponding 5729 // to specific blocks are marked or not. Blocks that are 5730 // marked are live and are not swept up. All remaining blocks 5731 // are swept up, with coalescing on-the-fly as we sweep up 5732 // contiguous free and/or garbage blocks: 5733 // We need to ensure that the sweeper synchronizes with allocators 5734 // and stop-the-world collectors. In particular, the following 5735 // locks are used: 5736 // . CMS token: if this is held, a stop the world collection cannot occur 5737 // . freelistLock: if this is held no allocation can occur from this 5738 // generation by another thread 5739 // . bitMapLock: if this is held, no other thread can access or update 5740 // 5741 5742 // Note that we need to hold the freelistLock if we use 5743 // block iterate below; else the iterator might go awry if 5744 // a mutator (or promotion) causes block contents to change 5745 // (for instance if the allocator divvies up a block). 5746 // If we hold the free list lock, for all practical purposes 5747 // young generation GC's can't occur (they'll usually need to 5748 // promote), so we might as well prevent all young generation 5749 // GC's while we do a sweeping step. For the same reason, we might 5750 // as well take the bit map lock for the entire duration 5751 5752 // check that we hold the requisite locks 5753 assert(have_cms_token(), "Should hold cms token"); 5754 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5755 assert_lock_strong(gen->freelistLock()); 5756 assert_lock_strong(bitMapLock()); 5757 5758 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5759 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5760 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5761 _inter_sweep_estimate.padded_average(), 5762 _intra_sweep_estimate.padded_average()); 5763 gen->setNearLargestChunk(); 5764 5765 { 5766 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5767 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5768 // We need to free-up/coalesce garbage/blocks from a 5769 // co-terminal free run. This is done in the SweepClosure 5770 // destructor; so, do not remove this scope, else the 5771 // end-of-sweep-census below will be off by a little bit. 5772 } 5773 gen->cmsSpace()->sweep_completed(); 5774 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5775 if (should_unload_classes()) { // unloaded classes this cycle, 5776 _concurrent_cycles_since_last_unload = 0; // ... reset count 5777 } else { // did not unload classes, 5778 _concurrent_cycles_since_last_unload++; // ... increment count 5779 } 5780 } 5781 5782 // Reset CMS data structures (for now just the marking bit map) 5783 // preparatory for the next cycle. 5784 void CMSCollector::reset(bool concurrent) { 5785 if (concurrent) { 5786 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5787 5788 // If the state is not "Resetting", the foreground thread 5789 // has done a collection and the resetting. 5790 if (_collectorState != Resetting) { 5791 assert(_collectorState == Idling, "The state should only change" 5792 " because the foreground collector has finished the collection"); 5793 return; 5794 } 5795 5796 // Clear the mark bitmap (no grey objects to start with) 5797 // for the next cycle. 5798 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5799 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5800 5801 HeapWord* curAddr = _markBitMap.startWord(); 5802 while (curAddr < _markBitMap.endWord()) { 5803 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5804 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5805 _markBitMap.clear_large_range(chunk); 5806 if (ConcurrentMarkSweepThread::should_yield() && 5807 !foregroundGCIsActive() && 5808 CMSYield) { 5809 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5810 "CMS thread should hold CMS token"); 5811 assert_lock_strong(bitMapLock()); 5812 bitMapLock()->unlock(); 5813 ConcurrentMarkSweepThread::desynchronize(true); 5814 stopTimer(); 5815 if (PrintCMSStatistics != 0) { 5816 incrementYields(); 5817 } 5818 5819 // See the comment in coordinator_yield() 5820 for (unsigned i = 0; i < CMSYieldSleepCount && 5821 ConcurrentMarkSweepThread::should_yield() && 5822 !CMSCollector::foregroundGCIsActive(); ++i) { 5823 os::sleep(Thread::current(), 1, false); 5824 } 5825 5826 ConcurrentMarkSweepThread::synchronize(true); 5827 bitMapLock()->lock_without_safepoint_check(); 5828 startTimer(); 5829 } 5830 curAddr = chunk.end(); 5831 } 5832 // A successful mostly concurrent collection has been done. 5833 // Because only the full (i.e., concurrent mode failure) collections 5834 // are being measured for gc overhead limits, clean the "near" flag 5835 // and count. 5836 size_policy()->reset_gc_overhead_limit_count(); 5837 _collectorState = Idling; 5838 } else { 5839 // already have the lock 5840 assert(_collectorState == Resetting, "just checking"); 5841 assert_lock_strong(bitMapLock()); 5842 _markBitMap.clear_all(); 5843 _collectorState = Idling; 5844 } 5845 5846 register_gc_end(); 5847 } 5848 5849 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5850 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5851 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5852 TraceCollectorStats tcs(counters()); 5853 5854 switch (op) { 5855 case CMS_op_checkpointRootsInitial: { 5856 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5857 checkpointRootsInitial(); 5858 if (PrintGC) { 5859 _cmsGen->printOccupancy("initial-mark"); 5860 } 5861 break; 5862 } 5863 case CMS_op_checkpointRootsFinal: { 5864 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5865 checkpointRootsFinal(); 5866 if (PrintGC) { 5867 _cmsGen->printOccupancy("remark"); 5868 } 5869 break; 5870 } 5871 default: 5872 fatal("No such CMS_op"); 5873 } 5874 } 5875 5876 #ifndef PRODUCT 5877 size_t const CMSCollector::skip_header_HeapWords() { 5878 return FreeChunk::header_size(); 5879 } 5880 5881 // Try and collect here conditions that should hold when 5882 // CMS thread is exiting. The idea is that the foreground GC 5883 // thread should not be blocked if it wants to terminate 5884 // the CMS thread and yet continue to run the VM for a while 5885 // after that. 5886 void CMSCollector::verify_ok_to_terminate() const { 5887 assert(Thread::current()->is_ConcurrentGC_thread(), 5888 "should be called by CMS thread"); 5889 assert(!_foregroundGCShouldWait, "should be false"); 5890 // We could check here that all the various low-level locks 5891 // are not held by the CMS thread, but that is overkill; see 5892 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5893 // is checked. 5894 } 5895 #endif 5896 5897 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5898 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5899 "missing Printezis mark?"); 5900 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5901 size_t size = pointer_delta(nextOneAddr + 1, addr); 5902 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5903 "alignment problem"); 5904 assert(size >= 3, "Necessary for Printezis marks to work"); 5905 return size; 5906 } 5907 5908 // A variant of the above (block_size_using_printezis_bits()) except 5909 // that we return 0 if the P-bits are not yet set. 5910 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5911 if (_markBitMap.isMarked(addr + 1)) { 5912 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5913 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5914 size_t size = pointer_delta(nextOneAddr + 1, addr); 5915 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5916 "alignment problem"); 5917 assert(size >= 3, "Necessary for Printezis marks to work"); 5918 return size; 5919 } 5920 return 0; 5921 } 5922 5923 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5924 size_t sz = 0; 5925 oop p = (oop)addr; 5926 if (p->klass_or_null() != NULL) { 5927 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5928 } else { 5929 sz = block_size_using_printezis_bits(addr); 5930 } 5931 assert(sz > 0, "size must be nonzero"); 5932 HeapWord* next_block = addr + sz; 5933 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5934 CardTableModRefBS::card_size); 5935 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5936 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5937 "must be different cards"); 5938 return next_card; 5939 } 5940 5941 5942 // CMS Bit Map Wrapper ///////////////////////////////////////// 5943 5944 // Construct a CMS bit map infrastructure, but don't create the 5945 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5946 // further below. 5947 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5948 _bm(), 5949 _shifter(shifter), 5950 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5951 Monitor::_safepoint_check_sometimes) : NULL) 5952 { 5953 _bmStartWord = 0; 5954 _bmWordSize = 0; 5955 } 5956 5957 bool CMSBitMap::allocate(MemRegion mr) { 5958 _bmStartWord = mr.start(); 5959 _bmWordSize = mr.word_size(); 5960 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5961 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5962 if (!brs.is_reserved()) { 5963 warning("CMS bit map allocation failure"); 5964 return false; 5965 } 5966 // For now we'll just commit all of the bit map up front. 5967 // Later on we'll try to be more parsimonious with swap. 5968 if (!_virtual_space.initialize(brs, brs.size())) { 5969 warning("CMS bit map backing store failure"); 5970 return false; 5971 } 5972 assert(_virtual_space.committed_size() == brs.size(), 5973 "didn't reserve backing store for all of CMS bit map?"); 5974 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5975 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5976 _bmWordSize, "inconsistency in bit map sizing"); 5977 _bm.set_size(_bmWordSize >> _shifter); 5978 5979 // bm.clear(); // can we rely on getting zero'd memory? verify below 5980 assert(isAllClear(), 5981 "Expected zero'd memory from ReservedSpace constructor"); 5982 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5983 "consistency check"); 5984 return true; 5985 } 5986 5987 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5988 HeapWord *next_addr, *end_addr, *last_addr; 5989 assert_locked(); 5990 assert(covers(mr), "out-of-range error"); 5991 // XXX assert that start and end are appropriately aligned 5992 for (next_addr = mr.start(), end_addr = mr.end(); 5993 next_addr < end_addr; next_addr = last_addr) { 5994 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5995 last_addr = dirty_region.end(); 5996 if (!dirty_region.is_empty()) { 5997 cl->do_MemRegion(dirty_region); 5998 } else { 5999 assert(last_addr == end_addr, "program logic"); 6000 return; 6001 } 6002 } 6003 } 6004 6005 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6006 _bm.print_on_error(st, prefix); 6007 } 6008 6009 #ifndef PRODUCT 6010 void CMSBitMap::assert_locked() const { 6011 CMSLockVerifier::assert_locked(lock()); 6012 } 6013 6014 bool CMSBitMap::covers(MemRegion mr) const { 6015 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6016 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6017 "size inconsistency"); 6018 return (mr.start() >= _bmStartWord) && 6019 (mr.end() <= endWord()); 6020 } 6021 6022 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6023 return (start >= _bmStartWord && (start + size) <= endWord()); 6024 } 6025 6026 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6027 // verify that there are no 1 bits in the interval [left, right) 6028 FalseBitMapClosure falseBitMapClosure; 6029 iterate(&falseBitMapClosure, left, right); 6030 } 6031 6032 void CMSBitMap::region_invariant(MemRegion mr) 6033 { 6034 assert_locked(); 6035 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6036 assert(!mr.is_empty(), "unexpected empty region"); 6037 assert(covers(mr), "mr should be covered by bit map"); 6038 // convert address range into offset range 6039 size_t start_ofs = heapWordToOffset(mr.start()); 6040 // Make sure that end() is appropriately aligned 6041 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6042 (1 << (_shifter+LogHeapWordSize))), 6043 "Misaligned mr.end()"); 6044 size_t end_ofs = heapWordToOffset(mr.end()); 6045 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6046 } 6047 6048 #endif 6049 6050 bool CMSMarkStack::allocate(size_t size) { 6051 // allocate a stack of the requisite depth 6052 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6053 size * sizeof(oop))); 6054 if (!rs.is_reserved()) { 6055 warning("CMSMarkStack allocation failure"); 6056 return false; 6057 } 6058 if (!_virtual_space.initialize(rs, rs.size())) { 6059 warning("CMSMarkStack backing store failure"); 6060 return false; 6061 } 6062 assert(_virtual_space.committed_size() == rs.size(), 6063 "didn't reserve backing store for all of CMS stack?"); 6064 _base = (oop*)(_virtual_space.low()); 6065 _index = 0; 6066 _capacity = size; 6067 NOT_PRODUCT(_max_depth = 0); 6068 return true; 6069 } 6070 6071 // XXX FIX ME !!! In the MT case we come in here holding a 6072 // leaf lock. For printing we need to take a further lock 6073 // which has lower rank. We need to recalibrate the two 6074 // lock-ranks involved in order to be able to print the 6075 // messages below. (Or defer the printing to the caller. 6076 // For now we take the expedient path of just disabling the 6077 // messages for the problematic case.) 6078 void CMSMarkStack::expand() { 6079 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6080 if (_capacity == MarkStackSizeMax) { 6081 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6082 // We print a warning message only once per CMS cycle. 6083 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6084 } 6085 return; 6086 } 6087 // Double capacity if possible 6088 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6089 // Do not give up existing stack until we have managed to 6090 // get the double capacity that we desired. 6091 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6092 new_capacity * sizeof(oop))); 6093 if (rs.is_reserved()) { 6094 // Release the backing store associated with old stack 6095 _virtual_space.release(); 6096 // Reinitialize virtual space for new stack 6097 if (!_virtual_space.initialize(rs, rs.size())) { 6098 fatal("Not enough swap for expanded marking stack"); 6099 } 6100 _base = (oop*)(_virtual_space.low()); 6101 _index = 0; 6102 _capacity = new_capacity; 6103 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6104 // Failed to double capacity, continue; 6105 // we print a detail message only once per CMS cycle. 6106 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6107 SIZE_FORMAT"K", 6108 _capacity / K, new_capacity / K); 6109 } 6110 } 6111 6112 6113 // Closures 6114 // XXX: there seems to be a lot of code duplication here; 6115 // should refactor and consolidate common code. 6116 6117 // This closure is used to mark refs into the CMS generation in 6118 // the CMS bit map. Called at the first checkpoint. This closure 6119 // assumes that we do not need to re-mark dirty cards; if the CMS 6120 // generation on which this is used is not an oldest 6121 // generation then this will lose younger_gen cards! 6122 6123 MarkRefsIntoClosure::MarkRefsIntoClosure( 6124 MemRegion span, CMSBitMap* bitMap): 6125 _span(span), 6126 _bitMap(bitMap) 6127 { 6128 assert(_ref_processor == NULL, "deliberately left NULL"); 6129 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6130 } 6131 6132 void MarkRefsIntoClosure::do_oop(oop obj) { 6133 // if p points into _span, then mark corresponding bit in _markBitMap 6134 assert(obj->is_oop(), "expected an oop"); 6135 HeapWord* addr = (HeapWord*)obj; 6136 if (_span.contains(addr)) { 6137 // this should be made more efficient 6138 _bitMap->mark(addr); 6139 } 6140 } 6141 6142 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6143 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6144 6145 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6146 MemRegion span, CMSBitMap* bitMap): 6147 _span(span), 6148 _bitMap(bitMap) 6149 { 6150 assert(_ref_processor == NULL, "deliberately left NULL"); 6151 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6152 } 6153 6154 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6155 // if p points into _span, then mark corresponding bit in _markBitMap 6156 assert(obj->is_oop(), "expected an oop"); 6157 HeapWord* addr = (HeapWord*)obj; 6158 if (_span.contains(addr)) { 6159 // this should be made more efficient 6160 _bitMap->par_mark(addr); 6161 } 6162 } 6163 6164 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6165 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6166 6167 // A variant of the above, used for CMS marking verification. 6168 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6169 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6170 _span(span), 6171 _verification_bm(verification_bm), 6172 _cms_bm(cms_bm) 6173 { 6174 assert(_ref_processor == NULL, "deliberately left NULL"); 6175 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6176 } 6177 6178 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6179 // if p points into _span, then mark corresponding bit in _markBitMap 6180 assert(obj->is_oop(), "expected an oop"); 6181 HeapWord* addr = (HeapWord*)obj; 6182 if (_span.contains(addr)) { 6183 _verification_bm->mark(addr); 6184 if (!_cms_bm->isMarked(addr)) { 6185 oop(addr)->print(); 6186 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6187 fatal("... aborting"); 6188 } 6189 } 6190 } 6191 6192 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6193 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6194 6195 ////////////////////////////////////////////////// 6196 // MarkRefsIntoAndScanClosure 6197 ////////////////////////////////////////////////// 6198 6199 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6200 ReferenceProcessor* rp, 6201 CMSBitMap* bit_map, 6202 CMSBitMap* mod_union_table, 6203 CMSMarkStack* mark_stack, 6204 CMSCollector* collector, 6205 bool should_yield, 6206 bool concurrent_precleaning): 6207 _collector(collector), 6208 _span(span), 6209 _bit_map(bit_map), 6210 _mark_stack(mark_stack), 6211 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6212 mark_stack, concurrent_precleaning), 6213 _yield(should_yield), 6214 _concurrent_precleaning(concurrent_precleaning), 6215 _freelistLock(NULL) 6216 { 6217 _ref_processor = rp; 6218 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6219 } 6220 6221 // This closure is used to mark refs into the CMS generation at the 6222 // second (final) checkpoint, and to scan and transitively follow 6223 // the unmarked oops. It is also used during the concurrent precleaning 6224 // phase while scanning objects on dirty cards in the CMS generation. 6225 // The marks are made in the marking bit map and the marking stack is 6226 // used for keeping the (newly) grey objects during the scan. 6227 // The parallel version (Par_...) appears further below. 6228 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6229 if (obj != NULL) { 6230 assert(obj->is_oop(), "expected an oop"); 6231 HeapWord* addr = (HeapWord*)obj; 6232 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6233 assert(_collector->overflow_list_is_empty(), 6234 "overflow list should be empty"); 6235 if (_span.contains(addr) && 6236 !_bit_map->isMarked(addr)) { 6237 // mark bit map (object is now grey) 6238 _bit_map->mark(addr); 6239 // push on marking stack (stack should be empty), and drain the 6240 // stack by applying this closure to the oops in the oops popped 6241 // from the stack (i.e. blacken the grey objects) 6242 bool res = _mark_stack->push(obj); 6243 assert(res, "Should have space to push on empty stack"); 6244 do { 6245 oop new_oop = _mark_stack->pop(); 6246 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6247 assert(_bit_map->isMarked((HeapWord*)new_oop), 6248 "only grey objects on this stack"); 6249 // iterate over the oops in this oop, marking and pushing 6250 // the ones in CMS heap (i.e. in _span). 6251 new_oop->oop_iterate(&_pushAndMarkClosure); 6252 // check if it's time to yield 6253 do_yield_check(); 6254 } while (!_mark_stack->isEmpty() || 6255 (!_concurrent_precleaning && take_from_overflow_list())); 6256 // if marking stack is empty, and we are not doing this 6257 // during precleaning, then check the overflow list 6258 } 6259 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6260 assert(_collector->overflow_list_is_empty(), 6261 "overflow list was drained above"); 6262 // We could restore evacuated mark words, if any, used for 6263 // overflow list links here because the overflow list is 6264 // provably empty here. That would reduce the maximum 6265 // size requirements for preserved_{oop,mark}_stack. 6266 // But we'll just postpone it until we are all done 6267 // so we can just stream through. 6268 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6269 _collector->restore_preserved_marks_if_any(); 6270 assert(_collector->no_preserved_marks(), "No preserved marks"); 6271 } 6272 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6273 "All preserved marks should have been restored above"); 6274 } 6275 } 6276 6277 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6278 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6279 6280 void MarkRefsIntoAndScanClosure::do_yield_work() { 6281 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6282 "CMS thread should hold CMS token"); 6283 assert_lock_strong(_freelistLock); 6284 assert_lock_strong(_bit_map->lock()); 6285 // relinquish the free_list_lock and bitMaplock() 6286 _bit_map->lock()->unlock(); 6287 _freelistLock->unlock(); 6288 ConcurrentMarkSweepThread::desynchronize(true); 6289 _collector->stopTimer(); 6290 if (PrintCMSStatistics != 0) { 6291 _collector->incrementYields(); 6292 } 6293 6294 // See the comment in coordinator_yield() 6295 for (unsigned i = 0; 6296 i < CMSYieldSleepCount && 6297 ConcurrentMarkSweepThread::should_yield() && 6298 !CMSCollector::foregroundGCIsActive(); 6299 ++i) { 6300 os::sleep(Thread::current(), 1, false); 6301 } 6302 6303 ConcurrentMarkSweepThread::synchronize(true); 6304 _freelistLock->lock_without_safepoint_check(); 6305 _bit_map->lock()->lock_without_safepoint_check(); 6306 _collector->startTimer(); 6307 } 6308 6309 /////////////////////////////////////////////////////////// 6310 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6311 // MarkRefsIntoAndScanClosure 6312 /////////////////////////////////////////////////////////// 6313 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6314 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6315 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6316 _span(span), 6317 _bit_map(bit_map), 6318 _work_queue(work_queue), 6319 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6320 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6321 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6322 { 6323 _ref_processor = rp; 6324 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6325 } 6326 6327 // This closure is used to mark refs into the CMS generation at the 6328 // second (final) checkpoint, and to scan and transitively follow 6329 // the unmarked oops. The marks are made in the marking bit map and 6330 // the work_queue is used for keeping the (newly) grey objects during 6331 // the scan phase whence they are also available for stealing by parallel 6332 // threads. Since the marking bit map is shared, updates are 6333 // synchronized (via CAS). 6334 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6335 if (obj != NULL) { 6336 // Ignore mark word because this could be an already marked oop 6337 // that may be chained at the end of the overflow list. 6338 assert(obj->is_oop(true), "expected an oop"); 6339 HeapWord* addr = (HeapWord*)obj; 6340 if (_span.contains(addr) && 6341 !_bit_map->isMarked(addr)) { 6342 // mark bit map (object will become grey): 6343 // It is possible for several threads to be 6344 // trying to "claim" this object concurrently; 6345 // the unique thread that succeeds in marking the 6346 // object first will do the subsequent push on 6347 // to the work queue (or overflow list). 6348 if (_bit_map->par_mark(addr)) { 6349 // push on work_queue (which may not be empty), and trim the 6350 // queue to an appropriate length by applying this closure to 6351 // the oops in the oops popped from the stack (i.e. blacken the 6352 // grey objects) 6353 bool res = _work_queue->push(obj); 6354 assert(res, "Low water mark should be less than capacity?"); 6355 trim_queue(_low_water_mark); 6356 } // Else, another thread claimed the object 6357 } 6358 } 6359 } 6360 6361 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6362 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6363 6364 // This closure is used to rescan the marked objects on the dirty cards 6365 // in the mod union table and the card table proper. 6366 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6367 oop p, MemRegion mr) { 6368 6369 size_t size = 0; 6370 HeapWord* addr = (HeapWord*)p; 6371 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6372 assert(_span.contains(addr), "we are scanning the CMS generation"); 6373 // check if it's time to yield 6374 if (do_yield_check()) { 6375 // We yielded for some foreground stop-world work, 6376 // and we have been asked to abort this ongoing preclean cycle. 6377 return 0; 6378 } 6379 if (_bitMap->isMarked(addr)) { 6380 // it's marked; is it potentially uninitialized? 6381 if (p->klass_or_null() != NULL) { 6382 // an initialized object; ignore mark word in verification below 6383 // since we are running concurrent with mutators 6384 assert(p->is_oop(true), "should be an oop"); 6385 if (p->is_objArray()) { 6386 // objArrays are precisely marked; restrict scanning 6387 // to dirty cards only. 6388 size = CompactibleFreeListSpace::adjustObjectSize( 6389 p->oop_iterate(_scanningClosure, mr)); 6390 } else { 6391 // A non-array may have been imprecisely marked; we need 6392 // to scan object in its entirety. 6393 size = CompactibleFreeListSpace::adjustObjectSize( 6394 p->oop_iterate(_scanningClosure)); 6395 } 6396 #ifdef ASSERT 6397 size_t direct_size = 6398 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6399 assert(size == direct_size, "Inconsistency in size"); 6400 assert(size >= 3, "Necessary for Printezis marks to work"); 6401 if (!_bitMap->isMarked(addr+1)) { 6402 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6403 } else { 6404 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6405 assert(_bitMap->isMarked(addr+size-1), 6406 "inconsistent Printezis mark"); 6407 } 6408 #endif // ASSERT 6409 } else { 6410 // An uninitialized object. 6411 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6412 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6413 size = pointer_delta(nextOneAddr + 1, addr); 6414 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6415 "alignment problem"); 6416 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6417 // will dirty the card when the klass pointer is installed in the 6418 // object (signaling the completion of initialization). 6419 } 6420 } else { 6421 // Either a not yet marked object or an uninitialized object 6422 if (p->klass_or_null() == NULL) { 6423 // An uninitialized object, skip to the next card, since 6424 // we may not be able to read its P-bits yet. 6425 assert(size == 0, "Initial value"); 6426 } else { 6427 // An object not (yet) reached by marking: we merely need to 6428 // compute its size so as to go look at the next block. 6429 assert(p->is_oop(true), "should be an oop"); 6430 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6431 } 6432 } 6433 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6434 return size; 6435 } 6436 6437 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6438 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6439 "CMS thread should hold CMS token"); 6440 assert_lock_strong(_freelistLock); 6441 assert_lock_strong(_bitMap->lock()); 6442 // relinquish the free_list_lock and bitMaplock() 6443 _bitMap->lock()->unlock(); 6444 _freelistLock->unlock(); 6445 ConcurrentMarkSweepThread::desynchronize(true); 6446 _collector->stopTimer(); 6447 if (PrintCMSStatistics != 0) { 6448 _collector->incrementYields(); 6449 } 6450 6451 // See the comment in coordinator_yield() 6452 for (unsigned i = 0; i < CMSYieldSleepCount && 6453 ConcurrentMarkSweepThread::should_yield() && 6454 !CMSCollector::foregroundGCIsActive(); ++i) { 6455 os::sleep(Thread::current(), 1, false); 6456 } 6457 6458 ConcurrentMarkSweepThread::synchronize(true); 6459 _freelistLock->lock_without_safepoint_check(); 6460 _bitMap->lock()->lock_without_safepoint_check(); 6461 _collector->startTimer(); 6462 } 6463 6464 6465 ////////////////////////////////////////////////////////////////// 6466 // SurvivorSpacePrecleanClosure 6467 ////////////////////////////////////////////////////////////////// 6468 // This (single-threaded) closure is used to preclean the oops in 6469 // the survivor spaces. 6470 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6471 6472 HeapWord* addr = (HeapWord*)p; 6473 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6474 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6475 assert(p->klass_or_null() != NULL, "object should be initialized"); 6476 // an initialized object; ignore mark word in verification below 6477 // since we are running concurrent with mutators 6478 assert(p->is_oop(true), "should be an oop"); 6479 // Note that we do not yield while we iterate over 6480 // the interior oops of p, pushing the relevant ones 6481 // on our marking stack. 6482 size_t size = p->oop_iterate(_scanning_closure); 6483 do_yield_check(); 6484 // Observe that below, we do not abandon the preclean 6485 // phase as soon as we should; rather we empty the 6486 // marking stack before returning. This is to satisfy 6487 // some existing assertions. In general, it may be a 6488 // good idea to abort immediately and complete the marking 6489 // from the grey objects at a later time. 6490 while (!_mark_stack->isEmpty()) { 6491 oop new_oop = _mark_stack->pop(); 6492 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6493 assert(_bit_map->isMarked((HeapWord*)new_oop), 6494 "only grey objects on this stack"); 6495 // iterate over the oops in this oop, marking and pushing 6496 // the ones in CMS heap (i.e. in _span). 6497 new_oop->oop_iterate(_scanning_closure); 6498 // check if it's time to yield 6499 do_yield_check(); 6500 } 6501 unsigned int after_count = 6502 GenCollectedHeap::heap()->total_collections(); 6503 bool abort = (_before_count != after_count) || 6504 _collector->should_abort_preclean(); 6505 return abort ? 0 : size; 6506 } 6507 6508 void SurvivorSpacePrecleanClosure::do_yield_work() { 6509 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6510 "CMS thread should hold CMS token"); 6511 assert_lock_strong(_bit_map->lock()); 6512 // Relinquish the bit map lock 6513 _bit_map->lock()->unlock(); 6514 ConcurrentMarkSweepThread::desynchronize(true); 6515 _collector->stopTimer(); 6516 if (PrintCMSStatistics != 0) { 6517 _collector->incrementYields(); 6518 } 6519 6520 // See the comment in coordinator_yield() 6521 for (unsigned i = 0; i < CMSYieldSleepCount && 6522 ConcurrentMarkSweepThread::should_yield() && 6523 !CMSCollector::foregroundGCIsActive(); ++i) { 6524 os::sleep(Thread::current(), 1, false); 6525 } 6526 6527 ConcurrentMarkSweepThread::synchronize(true); 6528 _bit_map->lock()->lock_without_safepoint_check(); 6529 _collector->startTimer(); 6530 } 6531 6532 // This closure is used to rescan the marked objects on the dirty cards 6533 // in the mod union table and the card table proper. In the parallel 6534 // case, although the bitMap is shared, we do a single read so the 6535 // isMarked() query is "safe". 6536 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6537 // Ignore mark word because we are running concurrent with mutators 6538 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6539 HeapWord* addr = (HeapWord*)p; 6540 assert(_span.contains(addr), "we are scanning the CMS generation"); 6541 bool is_obj_array = false; 6542 #ifdef ASSERT 6543 if (!_parallel) { 6544 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6545 assert(_collector->overflow_list_is_empty(), 6546 "overflow list should be empty"); 6547 6548 } 6549 #endif // ASSERT 6550 if (_bit_map->isMarked(addr)) { 6551 // Obj arrays are precisely marked, non-arrays are not; 6552 // so we scan objArrays precisely and non-arrays in their 6553 // entirety. 6554 if (p->is_objArray()) { 6555 is_obj_array = true; 6556 if (_parallel) { 6557 p->oop_iterate(_par_scan_closure, mr); 6558 } else { 6559 p->oop_iterate(_scan_closure, mr); 6560 } 6561 } else { 6562 if (_parallel) { 6563 p->oop_iterate(_par_scan_closure); 6564 } else { 6565 p->oop_iterate(_scan_closure); 6566 } 6567 } 6568 } 6569 #ifdef ASSERT 6570 if (!_parallel) { 6571 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6572 assert(_collector->overflow_list_is_empty(), 6573 "overflow list should be empty"); 6574 6575 } 6576 #endif // ASSERT 6577 return is_obj_array; 6578 } 6579 6580 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6581 MemRegion span, 6582 CMSBitMap* bitMap, CMSMarkStack* markStack, 6583 bool should_yield, bool verifying): 6584 _collector(collector), 6585 _span(span), 6586 _bitMap(bitMap), 6587 _mut(&collector->_modUnionTable), 6588 _markStack(markStack), 6589 _yield(should_yield), 6590 _skipBits(0) 6591 { 6592 assert(_markStack->isEmpty(), "stack should be empty"); 6593 _finger = _bitMap->startWord(); 6594 _threshold = _finger; 6595 assert(_collector->_restart_addr == NULL, "Sanity check"); 6596 assert(_span.contains(_finger), "Out of bounds _finger?"); 6597 DEBUG_ONLY(_verifying = verifying;) 6598 } 6599 6600 void MarkFromRootsClosure::reset(HeapWord* addr) { 6601 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6602 assert(_span.contains(addr), "Out of bounds _finger?"); 6603 _finger = addr; 6604 _threshold = (HeapWord*)round_to( 6605 (intptr_t)_finger, CardTableModRefBS::card_size); 6606 } 6607 6608 // Should revisit to see if this should be restructured for 6609 // greater efficiency. 6610 bool MarkFromRootsClosure::do_bit(size_t offset) { 6611 if (_skipBits > 0) { 6612 _skipBits--; 6613 return true; 6614 } 6615 // convert offset into a HeapWord* 6616 HeapWord* addr = _bitMap->startWord() + offset; 6617 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6618 "address out of range"); 6619 assert(_bitMap->isMarked(addr), "tautology"); 6620 if (_bitMap->isMarked(addr+1)) { 6621 // this is an allocated but not yet initialized object 6622 assert(_skipBits == 0, "tautology"); 6623 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6624 oop p = oop(addr); 6625 if (p->klass_or_null() == NULL) { 6626 DEBUG_ONLY(if (!_verifying) {) 6627 // We re-dirty the cards on which this object lies and increase 6628 // the _threshold so that we'll come back to scan this object 6629 // during the preclean or remark phase. (CMSCleanOnEnter) 6630 if (CMSCleanOnEnter) { 6631 size_t sz = _collector->block_size_using_printezis_bits(addr); 6632 HeapWord* end_card_addr = (HeapWord*)round_to( 6633 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6634 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6635 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6636 // Bump _threshold to end_card_addr; note that 6637 // _threshold cannot possibly exceed end_card_addr, anyhow. 6638 // This prevents future clearing of the card as the scan proceeds 6639 // to the right. 6640 assert(_threshold <= end_card_addr, 6641 "Because we are just scanning into this object"); 6642 if (_threshold < end_card_addr) { 6643 _threshold = end_card_addr; 6644 } 6645 if (p->klass_or_null() != NULL) { 6646 // Redirty the range of cards... 6647 _mut->mark_range(redirty_range); 6648 } // ...else the setting of klass will dirty the card anyway. 6649 } 6650 DEBUG_ONLY(}) 6651 return true; 6652 } 6653 } 6654 scanOopsInOop(addr); 6655 return true; 6656 } 6657 6658 // We take a break if we've been at this for a while, 6659 // so as to avoid monopolizing the locks involved. 6660 void MarkFromRootsClosure::do_yield_work() { 6661 // First give up the locks, then yield, then re-lock 6662 // We should probably use a constructor/destructor idiom to 6663 // do this unlock/lock or modify the MutexUnlocker class to 6664 // serve our purpose. XXX 6665 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6666 "CMS thread should hold CMS token"); 6667 assert_lock_strong(_bitMap->lock()); 6668 _bitMap->lock()->unlock(); 6669 ConcurrentMarkSweepThread::desynchronize(true); 6670 _collector->stopTimer(); 6671 if (PrintCMSStatistics != 0) { 6672 _collector->incrementYields(); 6673 } 6674 6675 // See the comment in coordinator_yield() 6676 for (unsigned i = 0; i < CMSYieldSleepCount && 6677 ConcurrentMarkSweepThread::should_yield() && 6678 !CMSCollector::foregroundGCIsActive(); ++i) { 6679 os::sleep(Thread::current(), 1, false); 6680 } 6681 6682 ConcurrentMarkSweepThread::synchronize(true); 6683 _bitMap->lock()->lock_without_safepoint_check(); 6684 _collector->startTimer(); 6685 } 6686 6687 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6688 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6689 assert(_markStack->isEmpty(), 6690 "should drain stack to limit stack usage"); 6691 // convert ptr to an oop preparatory to scanning 6692 oop obj = oop(ptr); 6693 // Ignore mark word in verification below, since we 6694 // may be running concurrent with mutators. 6695 assert(obj->is_oop(true), "should be an oop"); 6696 assert(_finger <= ptr, "_finger runneth ahead"); 6697 // advance the finger to right end of this object 6698 _finger = ptr + obj->size(); 6699 assert(_finger > ptr, "we just incremented it above"); 6700 // On large heaps, it may take us some time to get through 6701 // the marking phase. During 6702 // this time it's possible that a lot of mutations have 6703 // accumulated in the card table and the mod union table -- 6704 // these mutation records are redundant until we have 6705 // actually traced into the corresponding card. 6706 // Here, we check whether advancing the finger would make 6707 // us cross into a new card, and if so clear corresponding 6708 // cards in the MUT (preclean them in the card-table in the 6709 // future). 6710 6711 DEBUG_ONLY(if (!_verifying) {) 6712 // The clean-on-enter optimization is disabled by default, 6713 // until we fix 6178663. 6714 if (CMSCleanOnEnter && (_finger > _threshold)) { 6715 // [_threshold, _finger) represents the interval 6716 // of cards to be cleared in MUT (or precleaned in card table). 6717 // The set of cards to be cleared is all those that overlap 6718 // with the interval [_threshold, _finger); note that 6719 // _threshold is always kept card-aligned but _finger isn't 6720 // always card-aligned. 6721 HeapWord* old_threshold = _threshold; 6722 assert(old_threshold == (HeapWord*)round_to( 6723 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6724 "_threshold should always be card-aligned"); 6725 _threshold = (HeapWord*)round_to( 6726 (intptr_t)_finger, CardTableModRefBS::card_size); 6727 MemRegion mr(old_threshold, _threshold); 6728 assert(!mr.is_empty(), "Control point invariant"); 6729 assert(_span.contains(mr), "Should clear within span"); 6730 _mut->clear_range(mr); 6731 } 6732 DEBUG_ONLY(}) 6733 // Note: the finger doesn't advance while we drain 6734 // the stack below. 6735 PushOrMarkClosure pushOrMarkClosure(_collector, 6736 _span, _bitMap, _markStack, 6737 _finger, this); 6738 bool res = _markStack->push(obj); 6739 assert(res, "Empty non-zero size stack should have space for single push"); 6740 while (!_markStack->isEmpty()) { 6741 oop new_oop = _markStack->pop(); 6742 // Skip verifying header mark word below because we are 6743 // running concurrent with mutators. 6744 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6745 // now scan this oop's oops 6746 new_oop->oop_iterate(&pushOrMarkClosure); 6747 do_yield_check(); 6748 } 6749 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6750 } 6751 6752 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6753 CMSCollector* collector, MemRegion span, 6754 CMSBitMap* bit_map, 6755 OopTaskQueue* work_queue, 6756 CMSMarkStack* overflow_stack): 6757 _collector(collector), 6758 _whole_span(collector->_span), 6759 _span(span), 6760 _bit_map(bit_map), 6761 _mut(&collector->_modUnionTable), 6762 _work_queue(work_queue), 6763 _overflow_stack(overflow_stack), 6764 _skip_bits(0), 6765 _task(task) 6766 { 6767 assert(_work_queue->size() == 0, "work_queue should be empty"); 6768 _finger = span.start(); 6769 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6770 assert(_span.contains(_finger), "Out of bounds _finger?"); 6771 } 6772 6773 // Should revisit to see if this should be restructured for 6774 // greater efficiency. 6775 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6776 if (_skip_bits > 0) { 6777 _skip_bits--; 6778 return true; 6779 } 6780 // convert offset into a HeapWord* 6781 HeapWord* addr = _bit_map->startWord() + offset; 6782 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6783 "address out of range"); 6784 assert(_bit_map->isMarked(addr), "tautology"); 6785 if (_bit_map->isMarked(addr+1)) { 6786 // this is an allocated object that might not yet be initialized 6787 assert(_skip_bits == 0, "tautology"); 6788 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6789 oop p = oop(addr); 6790 if (p->klass_or_null() == NULL) { 6791 // in the case of Clean-on-Enter optimization, redirty card 6792 // and avoid clearing card by increasing the threshold. 6793 return true; 6794 } 6795 } 6796 scan_oops_in_oop(addr); 6797 return true; 6798 } 6799 6800 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6801 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6802 // Should we assert that our work queue is empty or 6803 // below some drain limit? 6804 assert(_work_queue->size() == 0, 6805 "should drain stack to limit stack usage"); 6806 // convert ptr to an oop preparatory to scanning 6807 oop obj = oop(ptr); 6808 // Ignore mark word in verification below, since we 6809 // may be running concurrent with mutators. 6810 assert(obj->is_oop(true), "should be an oop"); 6811 assert(_finger <= ptr, "_finger runneth ahead"); 6812 // advance the finger to right end of this object 6813 _finger = ptr + obj->size(); 6814 assert(_finger > ptr, "we just incremented it above"); 6815 // On large heaps, it may take us some time to get through 6816 // the marking phase. During 6817 // this time it's possible that a lot of mutations have 6818 // accumulated in the card table and the mod union table -- 6819 // these mutation records are redundant until we have 6820 // actually traced into the corresponding card. 6821 // Here, we check whether advancing the finger would make 6822 // us cross into a new card, and if so clear corresponding 6823 // cards in the MUT (preclean them in the card-table in the 6824 // future). 6825 6826 // The clean-on-enter optimization is disabled by default, 6827 // until we fix 6178663. 6828 if (CMSCleanOnEnter && (_finger > _threshold)) { 6829 // [_threshold, _finger) represents the interval 6830 // of cards to be cleared in MUT (or precleaned in card table). 6831 // The set of cards to be cleared is all those that overlap 6832 // with the interval [_threshold, _finger); note that 6833 // _threshold is always kept card-aligned but _finger isn't 6834 // always card-aligned. 6835 HeapWord* old_threshold = _threshold; 6836 assert(old_threshold == (HeapWord*)round_to( 6837 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6838 "_threshold should always be card-aligned"); 6839 _threshold = (HeapWord*)round_to( 6840 (intptr_t)_finger, CardTableModRefBS::card_size); 6841 MemRegion mr(old_threshold, _threshold); 6842 assert(!mr.is_empty(), "Control point invariant"); 6843 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6844 _mut->clear_range(mr); 6845 } 6846 6847 // Note: the local finger doesn't advance while we drain 6848 // the stack below, but the global finger sure can and will. 6849 HeapWord** gfa = _task->global_finger_addr(); 6850 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6851 _span, _bit_map, 6852 _work_queue, 6853 _overflow_stack, 6854 _finger, 6855 gfa, this); 6856 bool res = _work_queue->push(obj); // overflow could occur here 6857 assert(res, "Will hold once we use workqueues"); 6858 while (true) { 6859 oop new_oop; 6860 if (!_work_queue->pop_local(new_oop)) { 6861 // We emptied our work_queue; check if there's stuff that can 6862 // be gotten from the overflow stack. 6863 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6864 _overflow_stack, _work_queue)) { 6865 do_yield_check(); 6866 continue; 6867 } else { // done 6868 break; 6869 } 6870 } 6871 // Skip verifying header mark word below because we are 6872 // running concurrent with mutators. 6873 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6874 // now scan this oop's oops 6875 new_oop->oop_iterate(&pushOrMarkClosure); 6876 do_yield_check(); 6877 } 6878 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6879 } 6880 6881 // Yield in response to a request from VM Thread or 6882 // from mutators. 6883 void Par_MarkFromRootsClosure::do_yield_work() { 6884 assert(_task != NULL, "sanity"); 6885 _task->yield(); 6886 } 6887 6888 // A variant of the above used for verifying CMS marking work. 6889 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6890 MemRegion span, 6891 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6892 CMSMarkStack* mark_stack): 6893 _collector(collector), 6894 _span(span), 6895 _verification_bm(verification_bm), 6896 _cms_bm(cms_bm), 6897 _mark_stack(mark_stack), 6898 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6899 mark_stack) 6900 { 6901 assert(_mark_stack->isEmpty(), "stack should be empty"); 6902 _finger = _verification_bm->startWord(); 6903 assert(_collector->_restart_addr == NULL, "Sanity check"); 6904 assert(_span.contains(_finger), "Out of bounds _finger?"); 6905 } 6906 6907 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6908 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6909 assert(_span.contains(addr), "Out of bounds _finger?"); 6910 _finger = addr; 6911 } 6912 6913 // Should revisit to see if this should be restructured for 6914 // greater efficiency. 6915 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6916 // convert offset into a HeapWord* 6917 HeapWord* addr = _verification_bm->startWord() + offset; 6918 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6919 "address out of range"); 6920 assert(_verification_bm->isMarked(addr), "tautology"); 6921 assert(_cms_bm->isMarked(addr), "tautology"); 6922 6923 assert(_mark_stack->isEmpty(), 6924 "should drain stack to limit stack usage"); 6925 // convert addr to an oop preparatory to scanning 6926 oop obj = oop(addr); 6927 assert(obj->is_oop(), "should be an oop"); 6928 assert(_finger <= addr, "_finger runneth ahead"); 6929 // advance the finger to right end of this object 6930 _finger = addr + obj->size(); 6931 assert(_finger > addr, "we just incremented it above"); 6932 // Note: the finger doesn't advance while we drain 6933 // the stack below. 6934 bool res = _mark_stack->push(obj); 6935 assert(res, "Empty non-zero size stack should have space for single push"); 6936 while (!_mark_stack->isEmpty()) { 6937 oop new_oop = _mark_stack->pop(); 6938 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6939 // now scan this oop's oops 6940 new_oop->oop_iterate(&_pam_verify_closure); 6941 } 6942 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6943 return true; 6944 } 6945 6946 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6947 CMSCollector* collector, MemRegion span, 6948 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6949 CMSMarkStack* mark_stack): 6950 MetadataAwareOopClosure(collector->ref_processor()), 6951 _collector(collector), 6952 _span(span), 6953 _verification_bm(verification_bm), 6954 _cms_bm(cms_bm), 6955 _mark_stack(mark_stack) 6956 { } 6957 6958 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6959 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6960 6961 // Upon stack overflow, we discard (part of) the stack, 6962 // remembering the least address amongst those discarded 6963 // in CMSCollector's _restart_address. 6964 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6965 // Remember the least grey address discarded 6966 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6967 _collector->lower_restart_addr(ra); 6968 _mark_stack->reset(); // discard stack contents 6969 _mark_stack->expand(); // expand the stack if possible 6970 } 6971 6972 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6973 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 6974 HeapWord* addr = (HeapWord*)obj; 6975 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6976 // Oop lies in _span and isn't yet grey or black 6977 _verification_bm->mark(addr); // now grey 6978 if (!_cms_bm->isMarked(addr)) { 6979 oop(addr)->print(); 6980 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6981 p2i(addr)); 6982 fatal("... aborting"); 6983 } 6984 6985 if (!_mark_stack->push(obj)) { // stack overflow 6986 if (PrintCMSStatistics != 0) { 6987 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6988 SIZE_FORMAT, _mark_stack->capacity()); 6989 } 6990 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6991 handle_stack_overflow(addr); 6992 } 6993 // anything including and to the right of _finger 6994 // will be scanned as we iterate over the remainder of the 6995 // bit map 6996 } 6997 } 6998 6999 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7000 MemRegion span, 7001 CMSBitMap* bitMap, CMSMarkStack* markStack, 7002 HeapWord* finger, MarkFromRootsClosure* parent) : 7003 MetadataAwareOopClosure(collector->ref_processor()), 7004 _collector(collector), 7005 _span(span), 7006 _bitMap(bitMap), 7007 _markStack(markStack), 7008 _finger(finger), 7009 _parent(parent) 7010 { } 7011 7012 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7013 MemRegion span, 7014 CMSBitMap* bit_map, 7015 OopTaskQueue* work_queue, 7016 CMSMarkStack* overflow_stack, 7017 HeapWord* finger, 7018 HeapWord** global_finger_addr, 7019 Par_MarkFromRootsClosure* parent) : 7020 MetadataAwareOopClosure(collector->ref_processor()), 7021 _collector(collector), 7022 _whole_span(collector->_span), 7023 _span(span), 7024 _bit_map(bit_map), 7025 _work_queue(work_queue), 7026 _overflow_stack(overflow_stack), 7027 _finger(finger), 7028 _global_finger_addr(global_finger_addr), 7029 _parent(parent) 7030 { } 7031 7032 // Assumes thread-safe access by callers, who are 7033 // responsible for mutual exclusion. 7034 void CMSCollector::lower_restart_addr(HeapWord* low) { 7035 assert(_span.contains(low), "Out of bounds addr"); 7036 if (_restart_addr == NULL) { 7037 _restart_addr = low; 7038 } else { 7039 _restart_addr = MIN2(_restart_addr, low); 7040 } 7041 } 7042 7043 // Upon stack overflow, we discard (part of) the stack, 7044 // remembering the least address amongst those discarded 7045 // in CMSCollector's _restart_address. 7046 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7047 // Remember the least grey address discarded 7048 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7049 _collector->lower_restart_addr(ra); 7050 _markStack->reset(); // discard stack contents 7051 _markStack->expand(); // expand the stack if possible 7052 } 7053 7054 // Upon stack overflow, we discard (part of) the stack, 7055 // remembering the least address amongst those discarded 7056 // in CMSCollector's _restart_address. 7057 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7058 // We need to do this under a mutex to prevent other 7059 // workers from interfering with the work done below. 7060 MutexLockerEx ml(_overflow_stack->par_lock(), 7061 Mutex::_no_safepoint_check_flag); 7062 // Remember the least grey address discarded 7063 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7064 _collector->lower_restart_addr(ra); 7065 _overflow_stack->reset(); // discard stack contents 7066 _overflow_stack->expand(); // expand the stack if possible 7067 } 7068 7069 void PushOrMarkClosure::do_oop(oop obj) { 7070 // Ignore mark word because we are running concurrent with mutators. 7071 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7072 HeapWord* addr = (HeapWord*)obj; 7073 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7074 // Oop lies in _span and isn't yet grey or black 7075 _bitMap->mark(addr); // now grey 7076 if (addr < _finger) { 7077 // the bit map iteration has already either passed, or 7078 // sampled, this bit in the bit map; we'll need to 7079 // use the marking stack to scan this oop's oops. 7080 bool simulate_overflow = false; 7081 NOT_PRODUCT( 7082 if (CMSMarkStackOverflowALot && 7083 _collector->simulate_overflow()) { 7084 // simulate a stack overflow 7085 simulate_overflow = true; 7086 } 7087 ) 7088 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7089 if (PrintCMSStatistics != 0) { 7090 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7091 SIZE_FORMAT, _markStack->capacity()); 7092 } 7093 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7094 handle_stack_overflow(addr); 7095 } 7096 } 7097 // anything including and to the right of _finger 7098 // will be scanned as we iterate over the remainder of the 7099 // bit map 7100 do_yield_check(); 7101 } 7102 } 7103 7104 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7105 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7106 7107 void Par_PushOrMarkClosure::do_oop(oop obj) { 7108 // Ignore mark word because we are running concurrent with mutators. 7109 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7110 HeapWord* addr = (HeapWord*)obj; 7111 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7112 // Oop lies in _span and isn't yet grey or black 7113 // We read the global_finger (volatile read) strictly after marking oop 7114 bool res = _bit_map->par_mark(addr); // now grey 7115 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7116 // Should we push this marked oop on our stack? 7117 // -- if someone else marked it, nothing to do 7118 // -- if target oop is above global finger nothing to do 7119 // -- if target oop is in chunk and above local finger 7120 // then nothing to do 7121 // -- else push on work queue 7122 if ( !res // someone else marked it, they will deal with it 7123 || (addr >= *gfa) // will be scanned in a later task 7124 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7125 return; 7126 } 7127 // the bit map iteration has already either passed, or 7128 // sampled, this bit in the bit map; we'll need to 7129 // use the marking stack to scan this oop's oops. 7130 bool simulate_overflow = false; 7131 NOT_PRODUCT( 7132 if (CMSMarkStackOverflowALot && 7133 _collector->simulate_overflow()) { 7134 // simulate a stack overflow 7135 simulate_overflow = true; 7136 } 7137 ) 7138 if (simulate_overflow || 7139 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7140 // stack overflow 7141 if (PrintCMSStatistics != 0) { 7142 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7143 SIZE_FORMAT, _overflow_stack->capacity()); 7144 } 7145 // We cannot assert that the overflow stack is full because 7146 // it may have been emptied since. 7147 assert(simulate_overflow || 7148 _work_queue->size() == _work_queue->max_elems(), 7149 "Else push should have succeeded"); 7150 handle_stack_overflow(addr); 7151 } 7152 do_yield_check(); 7153 } 7154 } 7155 7156 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7157 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7158 7159 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7160 MemRegion span, 7161 ReferenceProcessor* rp, 7162 CMSBitMap* bit_map, 7163 CMSBitMap* mod_union_table, 7164 CMSMarkStack* mark_stack, 7165 bool concurrent_precleaning): 7166 MetadataAwareOopClosure(rp), 7167 _collector(collector), 7168 _span(span), 7169 _bit_map(bit_map), 7170 _mod_union_table(mod_union_table), 7171 _mark_stack(mark_stack), 7172 _concurrent_precleaning(concurrent_precleaning) 7173 { 7174 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7175 } 7176 7177 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7178 // the non-parallel version (the parallel version appears further below.) 7179 void PushAndMarkClosure::do_oop(oop obj) { 7180 // Ignore mark word verification. If during concurrent precleaning, 7181 // the object monitor may be locked. If during the checkpoint 7182 // phases, the object may already have been reached by a different 7183 // path and may be at the end of the global overflow list (so 7184 // the mark word may be NULL). 7185 assert(obj->is_oop_or_null(true /* ignore mark word */), 7186 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7187 HeapWord* addr = (HeapWord*)obj; 7188 // Check if oop points into the CMS generation 7189 // and is not marked 7190 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7191 // a white object ... 7192 _bit_map->mark(addr); // ... now grey 7193 // push on the marking stack (grey set) 7194 bool simulate_overflow = false; 7195 NOT_PRODUCT( 7196 if (CMSMarkStackOverflowALot && 7197 _collector->simulate_overflow()) { 7198 // simulate a stack overflow 7199 simulate_overflow = true; 7200 } 7201 ) 7202 if (simulate_overflow || !_mark_stack->push(obj)) { 7203 if (_concurrent_precleaning) { 7204 // During precleaning we can just dirty the appropriate card(s) 7205 // in the mod union table, thus ensuring that the object remains 7206 // in the grey set and continue. In the case of object arrays 7207 // we need to dirty all of the cards that the object spans, 7208 // since the rescan of object arrays will be limited to the 7209 // dirty cards. 7210 // Note that no one can be interfering with us in this action 7211 // of dirtying the mod union table, so no locking or atomics 7212 // are required. 7213 if (obj->is_objArray()) { 7214 size_t sz = obj->size(); 7215 HeapWord* end_card_addr = (HeapWord*)round_to( 7216 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7217 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7218 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7219 _mod_union_table->mark_range(redirty_range); 7220 } else { 7221 _mod_union_table->mark(addr); 7222 } 7223 _collector->_ser_pmc_preclean_ovflw++; 7224 } else { 7225 // During the remark phase, we need to remember this oop 7226 // in the overflow list. 7227 _collector->push_on_overflow_list(obj); 7228 _collector->_ser_pmc_remark_ovflw++; 7229 } 7230 } 7231 } 7232 } 7233 7234 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7235 MemRegion span, 7236 ReferenceProcessor* rp, 7237 CMSBitMap* bit_map, 7238 OopTaskQueue* work_queue): 7239 MetadataAwareOopClosure(rp), 7240 _collector(collector), 7241 _span(span), 7242 _bit_map(bit_map), 7243 _work_queue(work_queue) 7244 { 7245 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7246 } 7247 7248 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7249 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7250 7251 // Grey object rescan during second checkpoint phase -- 7252 // the parallel version. 7253 void Par_PushAndMarkClosure::do_oop(oop obj) { 7254 // In the assert below, we ignore the mark word because 7255 // this oop may point to an already visited object that is 7256 // on the overflow stack (in which case the mark word has 7257 // been hijacked for chaining into the overflow stack -- 7258 // if this is the last object in the overflow stack then 7259 // its mark word will be NULL). Because this object may 7260 // have been subsequently popped off the global overflow 7261 // stack, and the mark word possibly restored to the prototypical 7262 // value, by the time we get to examined this failing assert in 7263 // the debugger, is_oop_or_null(false) may subsequently start 7264 // to hold. 7265 assert(obj->is_oop_or_null(true), 7266 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7267 HeapWord* addr = (HeapWord*)obj; 7268 // Check if oop points into the CMS generation 7269 // and is not marked 7270 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7271 // a white object ... 7272 // If we manage to "claim" the object, by being the 7273 // first thread to mark it, then we push it on our 7274 // marking stack 7275 if (_bit_map->par_mark(addr)) { // ... now grey 7276 // push on work queue (grey set) 7277 bool simulate_overflow = false; 7278 NOT_PRODUCT( 7279 if (CMSMarkStackOverflowALot && 7280 _collector->par_simulate_overflow()) { 7281 // simulate a stack overflow 7282 simulate_overflow = true; 7283 } 7284 ) 7285 if (simulate_overflow || !_work_queue->push(obj)) { 7286 _collector->par_push_on_overflow_list(obj); 7287 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7288 } 7289 } // Else, some other thread got there first 7290 } 7291 } 7292 7293 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7294 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7295 7296 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7297 Mutex* bml = _collector->bitMapLock(); 7298 assert_lock_strong(bml); 7299 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7300 "CMS thread should hold CMS token"); 7301 7302 bml->unlock(); 7303 ConcurrentMarkSweepThread::desynchronize(true); 7304 7305 _collector->stopTimer(); 7306 if (PrintCMSStatistics != 0) { 7307 _collector->incrementYields(); 7308 } 7309 7310 // See the comment in coordinator_yield() 7311 for (unsigned i = 0; i < CMSYieldSleepCount && 7312 ConcurrentMarkSweepThread::should_yield() && 7313 !CMSCollector::foregroundGCIsActive(); ++i) { 7314 os::sleep(Thread::current(), 1, false); 7315 } 7316 7317 ConcurrentMarkSweepThread::synchronize(true); 7318 bml->lock(); 7319 7320 _collector->startTimer(); 7321 } 7322 7323 bool CMSPrecleanRefsYieldClosure::should_return() { 7324 if (ConcurrentMarkSweepThread::should_yield()) { 7325 do_yield_work(); 7326 } 7327 return _collector->foregroundGCIsActive(); 7328 } 7329 7330 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7331 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7332 "mr should be aligned to start at a card boundary"); 7333 // We'd like to assert: 7334 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7335 // "mr should be a range of cards"); 7336 // However, that would be too strong in one case -- the last 7337 // partition ends at _unallocated_block which, in general, can be 7338 // an arbitrary boundary, not necessarily card aligned. 7339 if (PrintCMSStatistics != 0) { 7340 _num_dirty_cards += 7341 mr.word_size()/CardTableModRefBS::card_size_in_words; 7342 } 7343 _space->object_iterate_mem(mr, &_scan_cl); 7344 } 7345 7346 SweepClosure::SweepClosure(CMSCollector* collector, 7347 ConcurrentMarkSweepGeneration* g, 7348 CMSBitMap* bitMap, bool should_yield) : 7349 _collector(collector), 7350 _g(g), 7351 _sp(g->cmsSpace()), 7352 _limit(_sp->sweep_limit()), 7353 _freelistLock(_sp->freelistLock()), 7354 _bitMap(bitMap), 7355 _yield(should_yield), 7356 _inFreeRange(false), // No free range at beginning of sweep 7357 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7358 _lastFreeRangeCoalesced(false), 7359 _freeFinger(g->used_region().start()) 7360 { 7361 NOT_PRODUCT( 7362 _numObjectsFreed = 0; 7363 _numWordsFreed = 0; 7364 _numObjectsLive = 0; 7365 _numWordsLive = 0; 7366 _numObjectsAlreadyFree = 0; 7367 _numWordsAlreadyFree = 0; 7368 _last_fc = NULL; 7369 7370 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7371 _sp->dictionary()->initialize_dict_returned_bytes(); 7372 ) 7373 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7374 "sweep _limit out of bounds"); 7375 if (CMSTraceSweeper) { 7376 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7377 p2i(_limit)); 7378 } 7379 } 7380 7381 void SweepClosure::print_on(outputStream* st) const { 7382 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7383 p2i(_sp->bottom()), p2i(_sp->end())); 7384 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7385 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7386 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7387 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7388 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7389 } 7390 7391 #ifndef PRODUCT 7392 // Assertion checking only: no useful work in product mode -- 7393 // however, if any of the flags below become product flags, 7394 // you may need to review this code to see if it needs to be 7395 // enabled in product mode. 7396 SweepClosure::~SweepClosure() { 7397 assert_lock_strong(_freelistLock); 7398 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7399 "sweep _limit out of bounds"); 7400 if (inFreeRange()) { 7401 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7402 print(); 7403 ShouldNotReachHere(); 7404 } 7405 if (Verbose && PrintGC) { 7406 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7407 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7408 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7409 SIZE_FORMAT" bytes " 7410 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7411 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7412 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7413 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7414 * sizeof(HeapWord); 7415 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7416 7417 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7418 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7419 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7420 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7421 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7422 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7423 indexListReturnedBytes); 7424 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7425 dict_returned_bytes); 7426 } 7427 } 7428 if (CMSTraceSweeper) { 7429 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7430 p2i(_limit)); 7431 } 7432 } 7433 #endif // PRODUCT 7434 7435 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7436 bool freeRangeInFreeLists) { 7437 if (CMSTraceSweeper) { 7438 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7439 p2i(freeFinger), freeRangeInFreeLists); 7440 } 7441 assert(!inFreeRange(), "Trampling existing free range"); 7442 set_inFreeRange(true); 7443 set_lastFreeRangeCoalesced(false); 7444 7445 set_freeFinger(freeFinger); 7446 set_freeRangeInFreeLists(freeRangeInFreeLists); 7447 if (CMSTestInFreeList) { 7448 if (freeRangeInFreeLists) { 7449 FreeChunk* fc = (FreeChunk*) freeFinger; 7450 assert(fc->is_free(), "A chunk on the free list should be free."); 7451 assert(fc->size() > 0, "Free range should have a size"); 7452 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7453 } 7454 } 7455 } 7456 7457 // Note that the sweeper runs concurrently with mutators. Thus, 7458 // it is possible for direct allocation in this generation to happen 7459 // in the middle of the sweep. Note that the sweeper also coalesces 7460 // contiguous free blocks. Thus, unless the sweeper and the allocator 7461 // synchronize appropriately freshly allocated blocks may get swept up. 7462 // This is accomplished by the sweeper locking the free lists while 7463 // it is sweeping. Thus blocks that are determined to be free are 7464 // indeed free. There is however one additional complication: 7465 // blocks that have been allocated since the final checkpoint and 7466 // mark, will not have been marked and so would be treated as 7467 // unreachable and swept up. To prevent this, the allocator marks 7468 // the bit map when allocating during the sweep phase. This leads, 7469 // however, to a further complication -- objects may have been allocated 7470 // but not yet initialized -- in the sense that the header isn't yet 7471 // installed. The sweeper can not then determine the size of the block 7472 // in order to skip over it. To deal with this case, we use a technique 7473 // (due to Printezis) to encode such uninitialized block sizes in the 7474 // bit map. Since the bit map uses a bit per every HeapWord, but the 7475 // CMS generation has a minimum object size of 3 HeapWords, it follows 7476 // that "normal marks" won't be adjacent in the bit map (there will 7477 // always be at least two 0 bits between successive 1 bits). We make use 7478 // of these "unused" bits to represent uninitialized blocks -- the bit 7479 // corresponding to the start of the uninitialized object and the next 7480 // bit are both set. Finally, a 1 bit marks the end of the object that 7481 // started with the two consecutive 1 bits to indicate its potentially 7482 // uninitialized state. 7483 7484 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7485 FreeChunk* fc = (FreeChunk*)addr; 7486 size_t res; 7487 7488 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7489 // than "addr == _limit" because although _limit was a block boundary when 7490 // we started the sweep, it may no longer be one because heap expansion 7491 // may have caused us to coalesce the block ending at the address _limit 7492 // with a newly expanded chunk (this happens when _limit was set to the 7493 // previous _end of the space), so we may have stepped past _limit: 7494 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7495 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7496 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7497 "sweep _limit out of bounds"); 7498 assert(addr < _sp->end(), "addr out of bounds"); 7499 // Flush any free range we might be holding as a single 7500 // coalesced chunk to the appropriate free list. 7501 if (inFreeRange()) { 7502 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7503 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger()))); 7504 flush_cur_free_chunk(freeFinger(), 7505 pointer_delta(addr, freeFinger())); 7506 if (CMSTraceSweeper) { 7507 gclog_or_tty->print("Sweep: last chunk: "); 7508 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7509 "[coalesced:%d]\n", 7510 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7511 lastFreeRangeCoalesced() ? 1 : 0); 7512 } 7513 } 7514 7515 // help the iterator loop finish 7516 return pointer_delta(_sp->end(), addr); 7517 } 7518 7519 assert(addr < _limit, "sweep invariant"); 7520 // check if we should yield 7521 do_yield_check(addr); 7522 if (fc->is_free()) { 7523 // Chunk that is already free 7524 res = fc->size(); 7525 do_already_free_chunk(fc); 7526 debug_only(_sp->verifyFreeLists()); 7527 // If we flush the chunk at hand in lookahead_and_flush() 7528 // and it's coalesced with a preceding chunk, then the 7529 // process of "mangling" the payload of the coalesced block 7530 // will cause erasure of the size information from the 7531 // (erstwhile) header of all the coalesced blocks but the 7532 // first, so the first disjunct in the assert will not hold 7533 // in that specific case (in which case the second disjunct 7534 // will hold). 7535 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7536 "Otherwise the size info doesn't change at this step"); 7537 NOT_PRODUCT( 7538 _numObjectsAlreadyFree++; 7539 _numWordsAlreadyFree += res; 7540 ) 7541 NOT_PRODUCT(_last_fc = fc;) 7542 } else if (!_bitMap->isMarked(addr)) { 7543 // Chunk is fresh garbage 7544 res = do_garbage_chunk(fc); 7545 debug_only(_sp->verifyFreeLists()); 7546 NOT_PRODUCT( 7547 _numObjectsFreed++; 7548 _numWordsFreed += res; 7549 ) 7550 } else { 7551 // Chunk that is alive. 7552 res = do_live_chunk(fc); 7553 debug_only(_sp->verifyFreeLists()); 7554 NOT_PRODUCT( 7555 _numObjectsLive++; 7556 _numWordsLive += res; 7557 ) 7558 } 7559 return res; 7560 } 7561 7562 // For the smart allocation, record following 7563 // split deaths - a free chunk is removed from its free list because 7564 // it is being split into two or more chunks. 7565 // split birth - a free chunk is being added to its free list because 7566 // a larger free chunk has been split and resulted in this free chunk. 7567 // coal death - a free chunk is being removed from its free list because 7568 // it is being coalesced into a large free chunk. 7569 // coal birth - a free chunk is being added to its free list because 7570 // it was created when two or more free chunks where coalesced into 7571 // this free chunk. 7572 // 7573 // These statistics are used to determine the desired number of free 7574 // chunks of a given size. The desired number is chosen to be relative 7575 // to the end of a CMS sweep. The desired number at the end of a sweep 7576 // is the 7577 // count-at-end-of-previous-sweep (an amount that was enough) 7578 // - count-at-beginning-of-current-sweep (the excess) 7579 // + split-births (gains in this size during interval) 7580 // - split-deaths (demands on this size during interval) 7581 // where the interval is from the end of one sweep to the end of the 7582 // next. 7583 // 7584 // When sweeping the sweeper maintains an accumulated chunk which is 7585 // the chunk that is made up of chunks that have been coalesced. That 7586 // will be termed the left-hand chunk. A new chunk of garbage that 7587 // is being considered for coalescing will be referred to as the 7588 // right-hand chunk. 7589 // 7590 // When making a decision on whether to coalesce a right-hand chunk with 7591 // the current left-hand chunk, the current count vs. the desired count 7592 // of the left-hand chunk is considered. Also if the right-hand chunk 7593 // is near the large chunk at the end of the heap (see 7594 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7595 // left-hand chunk is coalesced. 7596 // 7597 // When making a decision about whether to split a chunk, the desired count 7598 // vs. the current count of the candidate to be split is also considered. 7599 // If the candidate is underpopulated (currently fewer chunks than desired) 7600 // a chunk of an overpopulated (currently more chunks than desired) size may 7601 // be chosen. The "hint" associated with a free list, if non-null, points 7602 // to a free list which may be overpopulated. 7603 // 7604 7605 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7606 const size_t size = fc->size(); 7607 // Chunks that cannot be coalesced are not in the 7608 // free lists. 7609 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7610 assert(_sp->verify_chunk_in_free_list(fc), 7611 "free chunk should be in free lists"); 7612 } 7613 // a chunk that is already free, should not have been 7614 // marked in the bit map 7615 HeapWord* const addr = (HeapWord*) fc; 7616 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7617 // Verify that the bit map has no bits marked between 7618 // addr and purported end of this block. 7619 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7620 7621 // Some chunks cannot be coalesced under any circumstances. 7622 // See the definition of cantCoalesce(). 7623 if (!fc->cantCoalesce()) { 7624 // This chunk can potentially be coalesced. 7625 if (_sp->adaptive_freelists()) { 7626 // All the work is done in 7627 do_post_free_or_garbage_chunk(fc, size); 7628 } else { // Not adaptive free lists 7629 // this is a free chunk that can potentially be coalesced by the sweeper; 7630 if (!inFreeRange()) { 7631 // if the next chunk is a free block that can't be coalesced 7632 // it doesn't make sense to remove this chunk from the free lists 7633 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7634 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7635 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7636 nextChunk->is_free() && // ... which is free... 7637 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7638 // nothing to do 7639 } else { 7640 // Potentially the start of a new free range: 7641 // Don't eagerly remove it from the free lists. 7642 // No need to remove it if it will just be put 7643 // back again. (Also from a pragmatic point of view 7644 // if it is a free block in a region that is beyond 7645 // any allocated blocks, an assertion will fail) 7646 // Remember the start of a free run. 7647 initialize_free_range(addr, true); 7648 // end - can coalesce with next chunk 7649 } 7650 } else { 7651 // the midst of a free range, we are coalescing 7652 print_free_block_coalesced(fc); 7653 if (CMSTraceSweeper) { 7654 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7655 } 7656 // remove it from the free lists 7657 _sp->removeFreeChunkFromFreeLists(fc); 7658 set_lastFreeRangeCoalesced(true); 7659 // If the chunk is being coalesced and the current free range is 7660 // in the free lists, remove the current free range so that it 7661 // will be returned to the free lists in its entirety - all 7662 // the coalesced pieces included. 7663 if (freeRangeInFreeLists()) { 7664 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7665 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7666 "Size of free range is inconsistent with chunk size."); 7667 if (CMSTestInFreeList) { 7668 assert(_sp->verify_chunk_in_free_list(ffc), 7669 "free range is not in free lists"); 7670 } 7671 _sp->removeFreeChunkFromFreeLists(ffc); 7672 set_freeRangeInFreeLists(false); 7673 } 7674 } 7675 } 7676 // Note that if the chunk is not coalescable (the else arm 7677 // below), we unconditionally flush, without needing to do 7678 // a "lookahead," as we do below. 7679 if (inFreeRange()) lookahead_and_flush(fc, size); 7680 } else { 7681 // Code path common to both original and adaptive free lists. 7682 7683 // cant coalesce with previous block; this should be treated 7684 // as the end of a free run if any 7685 if (inFreeRange()) { 7686 // we kicked some butt; time to pick up the garbage 7687 assert(freeFinger() < addr, "freeFinger points too high"); 7688 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7689 } 7690 // else, nothing to do, just continue 7691 } 7692 } 7693 7694 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7695 // This is a chunk of garbage. It is not in any free list. 7696 // Add it to a free list or let it possibly be coalesced into 7697 // a larger chunk. 7698 HeapWord* const addr = (HeapWord*) fc; 7699 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7700 7701 if (_sp->adaptive_freelists()) { 7702 // Verify that the bit map has no bits marked between 7703 // addr and purported end of just dead object. 7704 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7705 7706 do_post_free_or_garbage_chunk(fc, size); 7707 } else { 7708 if (!inFreeRange()) { 7709 // start of a new free range 7710 assert(size > 0, "A free range should have a size"); 7711 initialize_free_range(addr, false); 7712 } else { 7713 // this will be swept up when we hit the end of the 7714 // free range 7715 if (CMSTraceSweeper) { 7716 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size); 7717 } 7718 // If the chunk is being coalesced and the current free range is 7719 // in the free lists, remove the current free range so that it 7720 // will be returned to the free lists in its entirety - all 7721 // the coalesced pieces included. 7722 if (freeRangeInFreeLists()) { 7723 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7724 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7725 "Size of free range is inconsistent with chunk size."); 7726 if (CMSTestInFreeList) { 7727 assert(_sp->verify_chunk_in_free_list(ffc), 7728 "free range is not in free lists"); 7729 } 7730 _sp->removeFreeChunkFromFreeLists(ffc); 7731 set_freeRangeInFreeLists(false); 7732 } 7733 set_lastFreeRangeCoalesced(true); 7734 } 7735 // this will be swept up when we hit the end of the free range 7736 7737 // Verify that the bit map has no bits marked between 7738 // addr and purported end of just dead object. 7739 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7740 } 7741 assert(_limit >= addr + size, 7742 "A freshly garbage chunk can't possibly straddle over _limit"); 7743 if (inFreeRange()) lookahead_and_flush(fc, size); 7744 return size; 7745 } 7746 7747 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7748 HeapWord* addr = (HeapWord*) fc; 7749 // The sweeper has just found a live object. Return any accumulated 7750 // left hand chunk to the free lists. 7751 if (inFreeRange()) { 7752 assert(freeFinger() < addr, "freeFinger points too high"); 7753 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7754 } 7755 7756 // This object is live: we'd normally expect this to be 7757 // an oop, and like to assert the following: 7758 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7759 // However, as we commented above, this may be an object whose 7760 // header hasn't yet been initialized. 7761 size_t size; 7762 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7763 if (_bitMap->isMarked(addr + 1)) { 7764 // Determine the size from the bit map, rather than trying to 7765 // compute it from the object header. 7766 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7767 size = pointer_delta(nextOneAddr + 1, addr); 7768 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7769 "alignment problem"); 7770 7771 #ifdef ASSERT 7772 if (oop(addr)->klass_or_null() != NULL) { 7773 // Ignore mark word because we are running concurrent with mutators 7774 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7775 assert(size == 7776 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7777 "P-mark and computed size do not agree"); 7778 } 7779 #endif 7780 7781 } else { 7782 // This should be an initialized object that's alive. 7783 assert(oop(addr)->klass_or_null() != NULL, 7784 "Should be an initialized object"); 7785 // Ignore mark word because we are running concurrent with mutators 7786 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7787 // Verify that the bit map has no bits marked between 7788 // addr and purported end of this block. 7789 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7790 assert(size >= 3, "Necessary for Printezis marks to work"); 7791 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7792 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7793 } 7794 return size; 7795 } 7796 7797 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7798 size_t chunkSize) { 7799 // do_post_free_or_garbage_chunk() should only be called in the case 7800 // of the adaptive free list allocator. 7801 const bool fcInFreeLists = fc->is_free(); 7802 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7803 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7804 if (CMSTestInFreeList && fcInFreeLists) { 7805 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7806 } 7807 7808 if (CMSTraceSweeper) { 7809 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7810 } 7811 7812 HeapWord* const fc_addr = (HeapWord*) fc; 7813 7814 bool coalesce; 7815 const size_t left = pointer_delta(fc_addr, freeFinger()); 7816 const size_t right = chunkSize; 7817 switch (FLSCoalescePolicy) { 7818 // numeric value forms a coalition aggressiveness metric 7819 case 0: { // never coalesce 7820 coalesce = false; 7821 break; 7822 } 7823 case 1: { // coalesce if left & right chunks on overpopulated lists 7824 coalesce = _sp->coalOverPopulated(left) && 7825 _sp->coalOverPopulated(right); 7826 break; 7827 } 7828 case 2: { // coalesce if left chunk on overpopulated list (default) 7829 coalesce = _sp->coalOverPopulated(left); 7830 break; 7831 } 7832 case 3: { // coalesce if left OR right chunk on overpopulated list 7833 coalesce = _sp->coalOverPopulated(left) || 7834 _sp->coalOverPopulated(right); 7835 break; 7836 } 7837 case 4: { // always coalesce 7838 coalesce = true; 7839 break; 7840 } 7841 default: 7842 ShouldNotReachHere(); 7843 } 7844 7845 // Should the current free range be coalesced? 7846 // If the chunk is in a free range and either we decided to coalesce above 7847 // or the chunk is near the large block at the end of the heap 7848 // (isNearLargestChunk() returns true), then coalesce this chunk. 7849 const bool doCoalesce = inFreeRange() 7850 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7851 if (doCoalesce) { 7852 // Coalesce the current free range on the left with the new 7853 // chunk on the right. If either is on a free list, 7854 // it must be removed from the list and stashed in the closure. 7855 if (freeRangeInFreeLists()) { 7856 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7857 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7858 "Size of free range is inconsistent with chunk size."); 7859 if (CMSTestInFreeList) { 7860 assert(_sp->verify_chunk_in_free_list(ffc), 7861 "Chunk is not in free lists"); 7862 } 7863 _sp->coalDeath(ffc->size()); 7864 _sp->removeFreeChunkFromFreeLists(ffc); 7865 set_freeRangeInFreeLists(false); 7866 } 7867 if (fcInFreeLists) { 7868 _sp->coalDeath(chunkSize); 7869 assert(fc->size() == chunkSize, 7870 "The chunk has the wrong size or is not in the free lists"); 7871 _sp->removeFreeChunkFromFreeLists(fc); 7872 } 7873 set_lastFreeRangeCoalesced(true); 7874 print_free_block_coalesced(fc); 7875 } else { // not in a free range and/or should not coalesce 7876 // Return the current free range and start a new one. 7877 if (inFreeRange()) { 7878 // In a free range but cannot coalesce with the right hand chunk. 7879 // Put the current free range into the free lists. 7880 flush_cur_free_chunk(freeFinger(), 7881 pointer_delta(fc_addr, freeFinger())); 7882 } 7883 // Set up for new free range. Pass along whether the right hand 7884 // chunk is in the free lists. 7885 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7886 } 7887 } 7888 7889 // Lookahead flush: 7890 // If we are tracking a free range, and this is the last chunk that 7891 // we'll look at because its end crosses past _limit, we'll preemptively 7892 // flush it along with any free range we may be holding on to. Note that 7893 // this can be the case only for an already free or freshly garbage 7894 // chunk. If this block is an object, it can never straddle 7895 // over _limit. The "straddling" occurs when _limit is set at 7896 // the previous end of the space when this cycle started, and 7897 // a subsequent heap expansion caused the previously co-terminal 7898 // free block to be coalesced with the newly expanded portion, 7899 // thus rendering _limit a non-block-boundary making it dangerous 7900 // for the sweeper to step over and examine. 7901 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7902 assert(inFreeRange(), "Should only be called if currently in a free range."); 7903 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7904 assert(_sp->used_region().contains(eob - 1), 7905 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7906 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7907 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7908 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size)); 7909 if (eob >= _limit) { 7910 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7911 if (CMSTraceSweeper) { 7912 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7913 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7914 "[" PTR_FORMAT "," PTR_FORMAT ")", 7915 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7916 } 7917 // Return the storage we are tracking back into the free lists. 7918 if (CMSTraceSweeper) { 7919 gclog_or_tty->print_cr("Flushing ... "); 7920 } 7921 assert(freeFinger() < eob, "Error"); 7922 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7923 } 7924 } 7925 7926 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7927 assert(inFreeRange(), "Should only be called if currently in a free range."); 7928 assert(size > 0, 7929 "A zero sized chunk cannot be added to the free lists."); 7930 if (!freeRangeInFreeLists()) { 7931 if (CMSTestInFreeList) { 7932 FreeChunk* fc = (FreeChunk*) chunk; 7933 fc->set_size(size); 7934 assert(!_sp->verify_chunk_in_free_list(fc), 7935 "chunk should not be in free lists yet"); 7936 } 7937 if (CMSTraceSweeper) { 7938 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7939 p2i(chunk), size); 7940 } 7941 // A new free range is going to be starting. The current 7942 // free range has not been added to the free lists yet or 7943 // was removed so add it back. 7944 // If the current free range was coalesced, then the death 7945 // of the free range was recorded. Record a birth now. 7946 if (lastFreeRangeCoalesced()) { 7947 _sp->coalBirth(size); 7948 } 7949 _sp->addChunkAndRepairOffsetTable(chunk, size, 7950 lastFreeRangeCoalesced()); 7951 } else if (CMSTraceSweeper) { 7952 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7953 } 7954 set_inFreeRange(false); 7955 set_freeRangeInFreeLists(false); 7956 } 7957 7958 // We take a break if we've been at this for a while, 7959 // so as to avoid monopolizing the locks involved. 7960 void SweepClosure::do_yield_work(HeapWord* addr) { 7961 // Return current free chunk being used for coalescing (if any) 7962 // to the appropriate freelist. After yielding, the next 7963 // free block encountered will start a coalescing range of 7964 // free blocks. If the next free block is adjacent to the 7965 // chunk just flushed, they will need to wait for the next 7966 // sweep to be coalesced. 7967 if (inFreeRange()) { 7968 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7969 } 7970 7971 // First give up the locks, then yield, then re-lock. 7972 // We should probably use a constructor/destructor idiom to 7973 // do this unlock/lock or modify the MutexUnlocker class to 7974 // serve our purpose. XXX 7975 assert_lock_strong(_bitMap->lock()); 7976 assert_lock_strong(_freelistLock); 7977 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7978 "CMS thread should hold CMS token"); 7979 _bitMap->lock()->unlock(); 7980 _freelistLock->unlock(); 7981 ConcurrentMarkSweepThread::desynchronize(true); 7982 _collector->stopTimer(); 7983 if (PrintCMSStatistics != 0) { 7984 _collector->incrementYields(); 7985 } 7986 7987 // See the comment in coordinator_yield() 7988 for (unsigned i = 0; i < CMSYieldSleepCount && 7989 ConcurrentMarkSweepThread::should_yield() && 7990 !CMSCollector::foregroundGCIsActive(); ++i) { 7991 os::sleep(Thread::current(), 1, false); 7992 } 7993 7994 ConcurrentMarkSweepThread::synchronize(true); 7995 _freelistLock->lock(); 7996 _bitMap->lock()->lock_without_safepoint_check(); 7997 _collector->startTimer(); 7998 } 7999 8000 #ifndef PRODUCT 8001 // This is actually very useful in a product build if it can 8002 // be called from the debugger. Compile it into the product 8003 // as needed. 8004 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8005 return debug_cms_space->verify_chunk_in_free_list(fc); 8006 } 8007 #endif 8008 8009 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8010 if (CMSTraceSweeper) { 8011 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8012 p2i(fc), fc->size()); 8013 } 8014 } 8015 8016 // CMSIsAliveClosure 8017 bool CMSIsAliveClosure::do_object_b(oop obj) { 8018 HeapWord* addr = (HeapWord*)obj; 8019 return addr != NULL && 8020 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8021 } 8022 8023 8024 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8025 MemRegion span, 8026 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8027 bool cpc): 8028 _collector(collector), 8029 _span(span), 8030 _bit_map(bit_map), 8031 _mark_stack(mark_stack), 8032 _concurrent_precleaning(cpc) { 8033 assert(!_span.is_empty(), "Empty span could spell trouble"); 8034 } 8035 8036 8037 // CMSKeepAliveClosure: the serial version 8038 void CMSKeepAliveClosure::do_oop(oop obj) { 8039 HeapWord* addr = (HeapWord*)obj; 8040 if (_span.contains(addr) && 8041 !_bit_map->isMarked(addr)) { 8042 _bit_map->mark(addr); 8043 bool simulate_overflow = false; 8044 NOT_PRODUCT( 8045 if (CMSMarkStackOverflowALot && 8046 _collector->simulate_overflow()) { 8047 // simulate a stack overflow 8048 simulate_overflow = true; 8049 } 8050 ) 8051 if (simulate_overflow || !_mark_stack->push(obj)) { 8052 if (_concurrent_precleaning) { 8053 // We dirty the overflown object and let the remark 8054 // phase deal with it. 8055 assert(_collector->overflow_list_is_empty(), "Error"); 8056 // In the case of object arrays, we need to dirty all of 8057 // the cards that the object spans. No locking or atomics 8058 // are needed since no one else can be mutating the mod union 8059 // table. 8060 if (obj->is_objArray()) { 8061 size_t sz = obj->size(); 8062 HeapWord* end_card_addr = 8063 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8064 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8065 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8066 _collector->_modUnionTable.mark_range(redirty_range); 8067 } else { 8068 _collector->_modUnionTable.mark(addr); 8069 } 8070 _collector->_ser_kac_preclean_ovflw++; 8071 } else { 8072 _collector->push_on_overflow_list(obj); 8073 _collector->_ser_kac_ovflw++; 8074 } 8075 } 8076 } 8077 } 8078 8079 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8080 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8081 8082 // CMSParKeepAliveClosure: a parallel version of the above. 8083 // The work queues are private to each closure (thread), 8084 // but (may be) available for stealing by other threads. 8085 void CMSParKeepAliveClosure::do_oop(oop obj) { 8086 HeapWord* addr = (HeapWord*)obj; 8087 if (_span.contains(addr) && 8088 !_bit_map->isMarked(addr)) { 8089 // In general, during recursive tracing, several threads 8090 // may be concurrently getting here; the first one to 8091 // "tag" it, claims it. 8092 if (_bit_map->par_mark(addr)) { 8093 bool res = _work_queue->push(obj); 8094 assert(res, "Low water mark should be much less than capacity"); 8095 // Do a recursive trim in the hope that this will keep 8096 // stack usage lower, but leave some oops for potential stealers 8097 trim_queue(_low_water_mark); 8098 } // Else, another thread got there first 8099 } 8100 } 8101 8102 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8103 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8104 8105 void CMSParKeepAliveClosure::trim_queue(uint max) { 8106 while (_work_queue->size() > max) { 8107 oop new_oop; 8108 if (_work_queue->pop_local(new_oop)) { 8109 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8110 assert(_bit_map->isMarked((HeapWord*)new_oop), 8111 "no white objects on this stack!"); 8112 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8113 // iterate over the oops in this oop, marking and pushing 8114 // the ones in CMS heap (i.e. in _span). 8115 new_oop->oop_iterate(&_mark_and_push); 8116 } 8117 } 8118 } 8119 8120 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8121 CMSCollector* collector, 8122 MemRegion span, CMSBitMap* bit_map, 8123 OopTaskQueue* work_queue): 8124 _collector(collector), 8125 _span(span), 8126 _bit_map(bit_map), 8127 _work_queue(work_queue) { } 8128 8129 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8130 HeapWord* addr = (HeapWord*)obj; 8131 if (_span.contains(addr) && 8132 !_bit_map->isMarked(addr)) { 8133 if (_bit_map->par_mark(addr)) { 8134 bool simulate_overflow = false; 8135 NOT_PRODUCT( 8136 if (CMSMarkStackOverflowALot && 8137 _collector->par_simulate_overflow()) { 8138 // simulate a stack overflow 8139 simulate_overflow = true; 8140 } 8141 ) 8142 if (simulate_overflow || !_work_queue->push(obj)) { 8143 _collector->par_push_on_overflow_list(obj); 8144 _collector->_par_kac_ovflw++; 8145 } 8146 } // Else another thread got there already 8147 } 8148 } 8149 8150 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8151 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8152 8153 ////////////////////////////////////////////////////////////////// 8154 // CMSExpansionCause ///////////////////////////// 8155 ////////////////////////////////////////////////////////////////// 8156 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8157 switch (cause) { 8158 case _no_expansion: 8159 return "No expansion"; 8160 case _satisfy_free_ratio: 8161 return "Free ratio"; 8162 case _satisfy_promotion: 8163 return "Satisfy promotion"; 8164 case _satisfy_allocation: 8165 return "allocation"; 8166 case _allocate_par_lab: 8167 return "Par LAB"; 8168 case _allocate_par_spooling_space: 8169 return "Par Spooling Space"; 8170 case _adaptive_size_policy: 8171 return "Ergonomics"; 8172 default: 8173 return "unknown"; 8174 } 8175 } 8176 8177 void CMSDrainMarkingStackClosure::do_void() { 8178 // the max number to take from overflow list at a time 8179 const size_t num = _mark_stack->capacity()/4; 8180 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8181 "Overflow list should be NULL during concurrent phases"); 8182 while (!_mark_stack->isEmpty() || 8183 // if stack is empty, check the overflow list 8184 _collector->take_from_overflow_list(num, _mark_stack)) { 8185 oop obj = _mark_stack->pop(); 8186 HeapWord* addr = (HeapWord*)obj; 8187 assert(_span.contains(addr), "Should be within span"); 8188 assert(_bit_map->isMarked(addr), "Should be marked"); 8189 assert(obj->is_oop(), "Should be an oop"); 8190 obj->oop_iterate(_keep_alive); 8191 } 8192 } 8193 8194 void CMSParDrainMarkingStackClosure::do_void() { 8195 // drain queue 8196 trim_queue(0); 8197 } 8198 8199 // Trim our work_queue so its length is below max at return 8200 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8201 while (_work_queue->size() > max) { 8202 oop new_oop; 8203 if (_work_queue->pop_local(new_oop)) { 8204 assert(new_oop->is_oop(), "Expected an oop"); 8205 assert(_bit_map->isMarked((HeapWord*)new_oop), 8206 "no white objects on this stack!"); 8207 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8208 // iterate over the oops in this oop, marking and pushing 8209 // the ones in CMS heap (i.e. in _span). 8210 new_oop->oop_iterate(&_mark_and_push); 8211 } 8212 } 8213 } 8214 8215 //////////////////////////////////////////////////////////////////// 8216 // Support for Marking Stack Overflow list handling and related code 8217 //////////////////////////////////////////////////////////////////// 8218 // Much of the following code is similar in shape and spirit to the 8219 // code used in ParNewGC. We should try and share that code 8220 // as much as possible in the future. 8221 8222 #ifndef PRODUCT 8223 // Debugging support for CMSStackOverflowALot 8224 8225 // It's OK to call this multi-threaded; the worst thing 8226 // that can happen is that we'll get a bunch of closely 8227 // spaced simulated overflows, but that's OK, in fact 8228 // probably good as it would exercise the overflow code 8229 // under contention. 8230 bool CMSCollector::simulate_overflow() { 8231 if (_overflow_counter-- <= 0) { // just being defensive 8232 _overflow_counter = CMSMarkStackOverflowInterval; 8233 return true; 8234 } else { 8235 return false; 8236 } 8237 } 8238 8239 bool CMSCollector::par_simulate_overflow() { 8240 return simulate_overflow(); 8241 } 8242 #endif 8243 8244 // Single-threaded 8245 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8246 assert(stack->isEmpty(), "Expected precondition"); 8247 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8248 size_t i = num; 8249 oop cur = _overflow_list; 8250 const markOop proto = markOopDesc::prototype(); 8251 NOT_PRODUCT(ssize_t n = 0;) 8252 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8253 next = oop(cur->mark()); 8254 cur->set_mark(proto); // until proven otherwise 8255 assert(cur->is_oop(), "Should be an oop"); 8256 bool res = stack->push(cur); 8257 assert(res, "Bit off more than can chew?"); 8258 NOT_PRODUCT(n++;) 8259 } 8260 _overflow_list = cur; 8261 #ifndef PRODUCT 8262 assert(_num_par_pushes >= n, "Too many pops?"); 8263 _num_par_pushes -=n; 8264 #endif 8265 return !stack->isEmpty(); 8266 } 8267 8268 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8269 // (MT-safe) Get a prefix of at most "num" from the list. 8270 // The overflow list is chained through the mark word of 8271 // each object in the list. We fetch the entire list, 8272 // break off a prefix of the right size and return the 8273 // remainder. If other threads try to take objects from 8274 // the overflow list at that time, they will wait for 8275 // some time to see if data becomes available. If (and 8276 // only if) another thread places one or more object(s) 8277 // on the global list before we have returned the suffix 8278 // to the global list, we will walk down our local list 8279 // to find its end and append the global list to 8280 // our suffix before returning it. This suffix walk can 8281 // prove to be expensive (quadratic in the amount of traffic) 8282 // when there are many objects in the overflow list and 8283 // there is much producer-consumer contention on the list. 8284 // *NOTE*: The overflow list manipulation code here and 8285 // in ParNewGeneration:: are very similar in shape, 8286 // except that in the ParNew case we use the old (from/eden) 8287 // copy of the object to thread the list via its klass word. 8288 // Because of the common code, if you make any changes in 8289 // the code below, please check the ParNew version to see if 8290 // similar changes might be needed. 8291 // CR 6797058 has been filed to consolidate the common code. 8292 bool CMSCollector::par_take_from_overflow_list(size_t num, 8293 OopTaskQueue* work_q, 8294 int no_of_gc_threads) { 8295 assert(work_q->size() == 0, "First empty local work queue"); 8296 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8297 if (_overflow_list == NULL) { 8298 return false; 8299 } 8300 // Grab the entire list; we'll put back a suffix 8301 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8302 Thread* tid = Thread::current(); 8303 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8304 // set to ParallelGCThreads. 8305 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8306 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8307 // If the list is busy, we spin for a short while, 8308 // sleeping between attempts to get the list. 8309 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8310 os::sleep(tid, sleep_time_millis, false); 8311 if (_overflow_list == NULL) { 8312 // Nothing left to take 8313 return false; 8314 } else if (_overflow_list != BUSY) { 8315 // Try and grab the prefix 8316 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8317 } 8318 } 8319 // If the list was found to be empty, or we spun long 8320 // enough, we give up and return empty-handed. If we leave 8321 // the list in the BUSY state below, it must be the case that 8322 // some other thread holds the overflow list and will set it 8323 // to a non-BUSY state in the future. 8324 if (prefix == NULL || prefix == BUSY) { 8325 // Nothing to take or waited long enough 8326 if (prefix == NULL) { 8327 // Write back the NULL in case we overwrote it with BUSY above 8328 // and it is still the same value. 8329 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8330 } 8331 return false; 8332 } 8333 assert(prefix != NULL && prefix != BUSY, "Error"); 8334 size_t i = num; 8335 oop cur = prefix; 8336 // Walk down the first "num" objects, unless we reach the end. 8337 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8338 if (cur->mark() == NULL) { 8339 // We have "num" or fewer elements in the list, so there 8340 // is nothing to return to the global list. 8341 // Write back the NULL in lieu of the BUSY we wrote 8342 // above, if it is still the same value. 8343 if (_overflow_list == BUSY) { 8344 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8345 } 8346 } else { 8347 // Chop off the suffix and return it to the global list. 8348 assert(cur->mark() != BUSY, "Error"); 8349 oop suffix_head = cur->mark(); // suffix will be put back on global list 8350 cur->set_mark(NULL); // break off suffix 8351 // It's possible that the list is still in the empty(busy) state 8352 // we left it in a short while ago; in that case we may be 8353 // able to place back the suffix without incurring the cost 8354 // of a walk down the list. 8355 oop observed_overflow_list = _overflow_list; 8356 oop cur_overflow_list = observed_overflow_list; 8357 bool attached = false; 8358 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8359 observed_overflow_list = 8360 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8361 if (cur_overflow_list == observed_overflow_list) { 8362 attached = true; 8363 break; 8364 } else cur_overflow_list = observed_overflow_list; 8365 } 8366 if (!attached) { 8367 // Too bad, someone else sneaked in (at least) an element; we'll need 8368 // to do a splice. Find tail of suffix so we can prepend suffix to global 8369 // list. 8370 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8371 oop suffix_tail = cur; 8372 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8373 "Tautology"); 8374 observed_overflow_list = _overflow_list; 8375 do { 8376 cur_overflow_list = observed_overflow_list; 8377 if (cur_overflow_list != BUSY) { 8378 // Do the splice ... 8379 suffix_tail->set_mark(markOop(cur_overflow_list)); 8380 } else { // cur_overflow_list == BUSY 8381 suffix_tail->set_mark(NULL); 8382 } 8383 // ... and try to place spliced list back on overflow_list ... 8384 observed_overflow_list = 8385 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8386 } while (cur_overflow_list != observed_overflow_list); 8387 // ... until we have succeeded in doing so. 8388 } 8389 } 8390 8391 // Push the prefix elements on work_q 8392 assert(prefix != NULL, "control point invariant"); 8393 const markOop proto = markOopDesc::prototype(); 8394 oop next; 8395 NOT_PRODUCT(ssize_t n = 0;) 8396 for (cur = prefix; cur != NULL; cur = next) { 8397 next = oop(cur->mark()); 8398 cur->set_mark(proto); // until proven otherwise 8399 assert(cur->is_oop(), "Should be an oop"); 8400 bool res = work_q->push(cur); 8401 assert(res, "Bit off more than we can chew?"); 8402 NOT_PRODUCT(n++;) 8403 } 8404 #ifndef PRODUCT 8405 assert(_num_par_pushes >= n, "Too many pops?"); 8406 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8407 #endif 8408 return true; 8409 } 8410 8411 // Single-threaded 8412 void CMSCollector::push_on_overflow_list(oop p) { 8413 NOT_PRODUCT(_num_par_pushes++;) 8414 assert(p->is_oop(), "Not an oop"); 8415 preserve_mark_if_necessary(p); 8416 p->set_mark((markOop)_overflow_list); 8417 _overflow_list = p; 8418 } 8419 8420 // Multi-threaded; use CAS to prepend to overflow list 8421 void CMSCollector::par_push_on_overflow_list(oop p) { 8422 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8423 assert(p->is_oop(), "Not an oop"); 8424 par_preserve_mark_if_necessary(p); 8425 oop observed_overflow_list = _overflow_list; 8426 oop cur_overflow_list; 8427 do { 8428 cur_overflow_list = observed_overflow_list; 8429 if (cur_overflow_list != BUSY) { 8430 p->set_mark(markOop(cur_overflow_list)); 8431 } else { 8432 p->set_mark(NULL); 8433 } 8434 observed_overflow_list = 8435 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8436 } while (cur_overflow_list != observed_overflow_list); 8437 } 8438 #undef BUSY 8439 8440 // Single threaded 8441 // General Note on GrowableArray: pushes may silently fail 8442 // because we are (temporarily) out of C-heap for expanding 8443 // the stack. The problem is quite ubiquitous and affects 8444 // a lot of code in the JVM. The prudent thing for GrowableArray 8445 // to do (for now) is to exit with an error. However, that may 8446 // be too draconian in some cases because the caller may be 8447 // able to recover without much harm. For such cases, we 8448 // should probably introduce a "soft_push" method which returns 8449 // an indication of success or failure with the assumption that 8450 // the caller may be able to recover from a failure; code in 8451 // the VM can then be changed, incrementally, to deal with such 8452 // failures where possible, thus, incrementally hardening the VM 8453 // in such low resource situations. 8454 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8455 _preserved_oop_stack.push(p); 8456 _preserved_mark_stack.push(m); 8457 assert(m == p->mark(), "Mark word changed"); 8458 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8459 "bijection"); 8460 } 8461 8462 // Single threaded 8463 void CMSCollector::preserve_mark_if_necessary(oop p) { 8464 markOop m = p->mark(); 8465 if (m->must_be_preserved(p)) { 8466 preserve_mark_work(p, m); 8467 } 8468 } 8469 8470 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8471 markOop m = p->mark(); 8472 if (m->must_be_preserved(p)) { 8473 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8474 // Even though we read the mark word without holding 8475 // the lock, we are assured that it will not change 8476 // because we "own" this oop, so no other thread can 8477 // be trying to push it on the overflow list; see 8478 // the assertion in preserve_mark_work() that checks 8479 // that m == p->mark(). 8480 preserve_mark_work(p, m); 8481 } 8482 } 8483 8484 // We should be able to do this multi-threaded, 8485 // a chunk of stack being a task (this is 8486 // correct because each oop only ever appears 8487 // once in the overflow list. However, it's 8488 // not very easy to completely overlap this with 8489 // other operations, so will generally not be done 8490 // until all work's been completed. Because we 8491 // expect the preserved oop stack (set) to be small, 8492 // it's probably fine to do this single-threaded. 8493 // We can explore cleverer concurrent/overlapped/parallel 8494 // processing of preserved marks if we feel the 8495 // need for this in the future. Stack overflow should 8496 // be so rare in practice and, when it happens, its 8497 // effect on performance so great that this will 8498 // likely just be in the noise anyway. 8499 void CMSCollector::restore_preserved_marks_if_any() { 8500 assert(SafepointSynchronize::is_at_safepoint(), 8501 "world should be stopped"); 8502 assert(Thread::current()->is_ConcurrentGC_thread() || 8503 Thread::current()->is_VM_thread(), 8504 "should be single-threaded"); 8505 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8506 "bijection"); 8507 8508 while (!_preserved_oop_stack.is_empty()) { 8509 oop p = _preserved_oop_stack.pop(); 8510 assert(p->is_oop(), "Should be an oop"); 8511 assert(_span.contains(p), "oop should be in _span"); 8512 assert(p->mark() == markOopDesc::prototype(), 8513 "Set when taken from overflow list"); 8514 markOop m = _preserved_mark_stack.pop(); 8515 p->set_mark(m); 8516 } 8517 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8518 "stacks were cleared above"); 8519 } 8520 8521 #ifndef PRODUCT 8522 bool CMSCollector::no_preserved_marks() const { 8523 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8524 } 8525 #endif 8526 8527 // Transfer some number of overflown objects to usual marking 8528 // stack. Return true if some objects were transferred. 8529 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8530 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8531 (size_t)ParGCDesiredObjsFromOverflowList); 8532 8533 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8534 assert(_collector->overflow_list_is_empty() || res, 8535 "If list is not empty, we should have taken something"); 8536 assert(!res || !_mark_stack->isEmpty(), 8537 "If we took something, it should now be on our stack"); 8538 return res; 8539 } 8540 8541 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8542 size_t res = _sp->block_size_no_stall(addr, _collector); 8543 if (_sp->block_is_obj(addr)) { 8544 if (_live_bit_map->isMarked(addr)) { 8545 // It can't have been dead in a previous cycle 8546 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8547 } else { 8548 _dead_bit_map->mark(addr); // mark the dead object 8549 } 8550 } 8551 // Could be 0, if the block size could not be computed without stalling. 8552 return res; 8553 } 8554 8555 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8556 8557 switch (phase) { 8558 case CMSCollector::InitialMarking: 8559 initialize(true /* fullGC */ , 8560 cause /* cause of the GC */, 8561 true /* recordGCBeginTime */, 8562 true /* recordPreGCUsage */, 8563 false /* recordPeakUsage */, 8564 false /* recordPostGCusage */, 8565 true /* recordAccumulatedGCTime */, 8566 false /* recordGCEndTime */, 8567 false /* countCollection */ ); 8568 break; 8569 8570 case CMSCollector::FinalMarking: 8571 initialize(true /* fullGC */ , 8572 cause /* cause of the GC */, 8573 false /* recordGCBeginTime */, 8574 false /* recordPreGCUsage */, 8575 false /* recordPeakUsage */, 8576 false /* recordPostGCusage */, 8577 true /* recordAccumulatedGCTime */, 8578 false /* recordGCEndTime */, 8579 false /* countCollection */ ); 8580 break; 8581 8582 case CMSCollector::Sweeping: 8583 initialize(true /* fullGC */ , 8584 cause /* cause of the GC */, 8585 false /* recordGCBeginTime */, 8586 false /* recordPreGCUsage */, 8587 true /* recordPeakUsage */, 8588 true /* recordPostGCusage */, 8589 false /* recordAccumulatedGCTime */, 8590 true /* recordGCEndTime */, 8591 true /* countCollection */ ); 8592 break; 8593 8594 default: 8595 ShouldNotReachHere(); 8596 } 8597 }