1 /* 2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardTableRS.hpp" 46 #include "memory/collectorPolicy.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/genCollectedHeap.hpp" 49 #include "memory/genMarkSweep.hpp" 50 #include "memory/genOopClosures.inline.hpp" 51 #include "memory/iterator.inline.hpp" 52 #include "memory/padded.hpp" 53 #include "memory/referencePolicy.hpp" 54 #include "memory/resourceArea.hpp" 55 #include "memory/tenuredGeneration.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "prims/jvmtiExport.hpp" 58 #include "runtime/atomic.inline.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/orderAccess.inline.hpp" 63 #include "runtime/vmThread.hpp" 64 #include "services/memoryService.hpp" 65 #include "services/runtimeService.hpp" 66 67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 68 69 // statics 70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 71 bool CMSCollector::_full_gc_requested = false; 72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 73 74 ////////////////////////////////////////////////////////////////// 75 // In support of CMS/VM thread synchronization 76 ////////////////////////////////////////////////////////////////// 77 // We split use of the CGC_lock into 2 "levels". 78 // The low-level locking is of the usual CGC_lock monitor. We introduce 79 // a higher level "token" (hereafter "CMS token") built on top of the 80 // low level monitor (hereafter "CGC lock"). 81 // The token-passing protocol gives priority to the VM thread. The 82 // CMS-lock doesn't provide any fairness guarantees, but clients 83 // should ensure that it is only held for very short, bounded 84 // durations. 85 // 86 // When either of the CMS thread or the VM thread is involved in 87 // collection operations during which it does not want the other 88 // thread to interfere, it obtains the CMS token. 89 // 90 // If either thread tries to get the token while the other has 91 // it, that thread waits. However, if the VM thread and CMS thread 92 // both want the token, then the VM thread gets priority while the 93 // CMS thread waits. This ensures, for instance, that the "concurrent" 94 // phases of the CMS thread's work do not block out the VM thread 95 // for long periods of time as the CMS thread continues to hog 96 // the token. (See bug 4616232). 97 // 98 // The baton-passing functions are, however, controlled by the 99 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 100 // and here the low-level CMS lock, not the high level token, 101 // ensures mutual exclusion. 102 // 103 // Two important conditions that we have to satisfy: 104 // 1. if a thread does a low-level wait on the CMS lock, then it 105 // relinquishes the CMS token if it were holding that token 106 // when it acquired the low-level CMS lock. 107 // 2. any low-level notifications on the low-level lock 108 // should only be sent when a thread has relinquished the token. 109 // 110 // In the absence of either property, we'd have potential deadlock. 111 // 112 // We protect each of the CMS (concurrent and sequential) phases 113 // with the CMS _token_, not the CMS _lock_. 114 // 115 // The only code protected by CMS lock is the token acquisition code 116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 117 // baton-passing code. 118 // 119 // Unfortunately, i couldn't come up with a good abstraction to factor and 120 // hide the naked CGC_lock manipulation in the baton-passing code 121 // further below. That's something we should try to do. Also, the proof 122 // of correctness of this 2-level locking scheme is far from obvious, 123 // and potentially quite slippery. We have an uneasy suspicion, for instance, 124 // that there may be a theoretical possibility of delay/starvation in the 125 // low-level lock/wait/notify scheme used for the baton-passing because of 126 // potential interference with the priority scheme embodied in the 127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 128 // invocation further below and marked with "XXX 20011219YSR". 129 // Indeed, as we note elsewhere, this may become yet more slippery 130 // in the presence of multiple CMS and/or multiple VM threads. XXX 131 132 class CMSTokenSync: public StackObj { 133 private: 134 bool _is_cms_thread; 135 public: 136 CMSTokenSync(bool is_cms_thread): 137 _is_cms_thread(is_cms_thread) { 138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 139 "Incorrect argument to constructor"); 140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 141 } 142 143 ~CMSTokenSync() { 144 assert(_is_cms_thread ? 145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 147 "Incorrect state"); 148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 149 } 150 }; 151 152 // Convenience class that does a CMSTokenSync, and then acquires 153 // upto three locks. 154 class CMSTokenSyncWithLocks: public CMSTokenSync { 155 private: 156 // Note: locks are acquired in textual declaration order 157 // and released in the opposite order 158 MutexLockerEx _locker1, _locker2, _locker3; 159 public: 160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 162 CMSTokenSync(is_cms_thread), 163 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 164 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 165 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 166 { } 167 }; 168 169 170 ////////////////////////////////////////////////////////////////// 171 // Concurrent Mark-Sweep Generation ///////////////////////////// 172 ////////////////////////////////////////////////////////////////// 173 174 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 175 176 // This struct contains per-thread things necessary to support parallel 177 // young-gen collection. 178 class CMSParGCThreadState: public CHeapObj<mtGC> { 179 public: 180 CFLS_LAB lab; 181 PromotionInfo promo; 182 183 // Constructor. 184 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 185 promo.setSpace(cfls); 186 } 187 }; 188 189 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 190 ReservedSpace rs, size_t initial_byte_size, int level, 191 CardTableRS* ct, bool use_adaptive_freelists, 192 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 193 CardGeneration(rs, initial_byte_size, level, ct), 194 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 195 _debug_concurrent_cycle(true), 196 _did_compact(false) 197 { 198 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 199 HeapWord* end = (HeapWord*) _virtual_space.high(); 200 201 _direct_allocated_words = 0; 202 NOT_PRODUCT( 203 _numObjectsPromoted = 0; 204 _numWordsPromoted = 0; 205 _numObjectsAllocated = 0; 206 _numWordsAllocated = 0; 207 ) 208 209 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 210 use_adaptive_freelists, 211 dictionaryChoice); 212 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 213 if (_cmsSpace == NULL) { 214 vm_exit_during_initialization( 215 "CompactibleFreeListSpace allocation failure"); 216 } 217 _cmsSpace->_gen = this; 218 219 _gc_stats = new CMSGCStats(); 220 221 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 222 // offsets match. The ability to tell free chunks from objects 223 // depends on this property. 224 debug_only( 225 FreeChunk* junk = NULL; 226 assert(UseCompressedClassPointers || 227 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 228 "Offset of FreeChunk::_prev within FreeChunk must match" 229 " that of OopDesc::_klass within OopDesc"); 230 ) 231 if (CollectedHeap::use_parallel_gc_threads()) { 232 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 233 _par_gc_thread_states = 234 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 235 if (_par_gc_thread_states == NULL) { 236 vm_exit_during_initialization("Could not allocate par gc structs"); 237 } 238 for (uint i = 0; i < ParallelGCThreads; i++) { 239 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 240 if (_par_gc_thread_states[i] == NULL) { 241 vm_exit_during_initialization("Could not allocate par gc structs"); 242 } 243 } 244 } else { 245 _par_gc_thread_states = NULL; 246 } 247 _incremental_collection_failed = false; 248 // The "dilatation_factor" is the expansion that can occur on 249 // account of the fact that the minimum object size in the CMS 250 // generation may be larger than that in, say, a contiguous young 251 // generation. 252 // Ideally, in the calculation below, we'd compute the dilatation 253 // factor as: MinChunkSize/(promoting_gen's min object size) 254 // Since we do not have such a general query interface for the 255 // promoting generation, we'll instead just use the minimum 256 // object size (which today is a header's worth of space); 257 // note that all arithmetic is in units of HeapWords. 258 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 259 assert(_dilatation_factor >= 1.0, "from previous assert"); 260 } 261 262 263 // The field "_initiating_occupancy" represents the occupancy percentage 264 // at which we trigger a new collection cycle. Unless explicitly specified 265 // via CMSInitiatingOccupancyFraction (argument "io" below), it 266 // is calculated by: 267 // 268 // Let "f" be MinHeapFreeRatio in 269 // 270 // _initiating_occupancy = 100-f + 271 // f * (CMSTriggerRatio/100) 272 // where CMSTriggerRatio is the argument "tr" below. 273 // 274 // That is, if we assume the heap is at its desired maximum occupancy at the 275 // end of a collection, we let CMSTriggerRatio of the (purported) free 276 // space be allocated before initiating a new collection cycle. 277 // 278 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 279 assert(io <= 100 && tr <= 100, "Check the arguments"); 280 if (io >= 0) { 281 _initiating_occupancy = (double)io / 100.0; 282 } else { 283 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 284 (double)(tr * MinHeapFreeRatio) / 100.0) 285 / 100.0; 286 } 287 } 288 289 void ConcurrentMarkSweepGeneration::ref_processor_init() { 290 assert(collector() != NULL, "no collector"); 291 collector()->ref_processor_init(); 292 } 293 294 void CMSCollector::ref_processor_init() { 295 if (_ref_processor == NULL) { 296 // Allocate and initialize a reference processor 297 _ref_processor = 298 new ReferenceProcessor(_span, // span 299 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 300 (int) ParallelGCThreads, // mt processing degree 301 _cmsGen->refs_discovery_is_mt(), // mt discovery 302 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 303 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 304 &_is_alive_closure); // closure for liveness info 305 // Initialize the _ref_processor field of CMSGen 306 _cmsGen->set_ref_processor(_ref_processor); 307 308 } 309 } 310 311 AdaptiveSizePolicy* CMSCollector::size_policy() { 312 GenCollectedHeap* gch = GenCollectedHeap::heap(); 313 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 314 "Wrong type of heap"); 315 return gch->gen_policy()->size_policy(); 316 } 317 318 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 319 320 const char* gen_name = "old"; 321 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 322 323 // Generation Counters - generation 1, 1 subspace 324 _gen_counters = new GenerationCounters(gen_name, 1, 1, 325 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 326 327 _space_counters = new GSpaceCounters(gen_name, 0, 328 _virtual_space.reserved_size(), 329 this, _gen_counters); 330 } 331 332 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 333 _cms_gen(cms_gen) 334 { 335 assert(alpha <= 100, "bad value"); 336 _saved_alpha = alpha; 337 338 // Initialize the alphas to the bootstrap value of 100. 339 _gc0_alpha = _cms_alpha = 100; 340 341 _cms_begin_time.update(); 342 _cms_end_time.update(); 343 344 _gc0_duration = 0.0; 345 _gc0_period = 0.0; 346 _gc0_promoted = 0; 347 348 _cms_duration = 0.0; 349 _cms_period = 0.0; 350 _cms_allocated = 0; 351 352 _cms_used_at_gc0_begin = 0; 353 _cms_used_at_gc0_end = 0; 354 _allow_duty_cycle_reduction = false; 355 _valid_bits = 0; 356 } 357 358 double CMSStats::cms_free_adjustment_factor(size_t free) const { 359 // TBD: CR 6909490 360 return 1.0; 361 } 362 363 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 364 } 365 366 // If promotion failure handling is on use 367 // the padded average size of the promotion for each 368 // young generation collection. 369 double CMSStats::time_until_cms_gen_full() const { 370 size_t cms_free = _cms_gen->cmsSpace()->free(); 371 GenCollectedHeap* gch = GenCollectedHeap::heap(); 372 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 373 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 374 if (cms_free > expected_promotion) { 375 // Start a cms collection if there isn't enough space to promote 376 // for the next minor collection. Use the padded average as 377 // a safety factor. 378 cms_free -= expected_promotion; 379 380 // Adjust by the safety factor. 381 double cms_free_dbl = (double)cms_free; 382 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 383 // Apply a further correction factor which tries to adjust 384 // for recent occurance of concurrent mode failures. 385 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 386 cms_free_dbl = cms_free_dbl * cms_adjustment; 387 388 if (PrintGCDetails && Verbose) { 389 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 390 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 391 cms_free, expected_promotion); 392 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 393 cms_free_dbl, cms_consumption_rate() + 1.0); 394 } 395 // Add 1 in case the consumption rate goes to zero. 396 return cms_free_dbl / (cms_consumption_rate() + 1.0); 397 } 398 return 0.0; 399 } 400 401 // Compare the duration of the cms collection to the 402 // time remaining before the cms generation is empty. 403 // Note that the time from the start of the cms collection 404 // to the start of the cms sweep (less than the total 405 // duration of the cms collection) can be used. This 406 // has been tried and some applications experienced 407 // promotion failures early in execution. This was 408 // possibly because the averages were not accurate 409 // enough at the beginning. 410 double CMSStats::time_until_cms_start() const { 411 // We add "gc0_period" to the "work" calculation 412 // below because this query is done (mostly) at the 413 // end of a scavenge, so we need to conservatively 414 // account for that much possible delay 415 // in the query so as to avoid concurrent mode failures 416 // due to starting the collection just a wee bit too 417 // late. 418 double work = cms_duration() + gc0_period(); 419 double deadline = time_until_cms_gen_full(); 420 // If a concurrent mode failure occurred recently, we want to be 421 // more conservative and halve our expected time_until_cms_gen_full() 422 if (work > deadline) { 423 if (Verbose && PrintGCDetails) { 424 gclog_or_tty->print( 425 " CMSCollector: collect because of anticipated promotion " 426 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 427 gc0_period(), time_until_cms_gen_full()); 428 } 429 return 0.0; 430 } 431 return work - deadline; 432 } 433 434 #ifndef PRODUCT 435 void CMSStats::print_on(outputStream *st) const { 436 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 437 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 438 gc0_duration(), gc0_period(), gc0_promoted()); 439 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 440 cms_duration(), cms_period(), cms_allocated()); 441 st->print(",cms_since_beg=%g,cms_since_end=%g", 442 cms_time_since_begin(), cms_time_since_end()); 443 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 444 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 445 446 if (valid()) { 447 st->print(",promo_rate=%g,cms_alloc_rate=%g", 448 promotion_rate(), cms_allocation_rate()); 449 st->print(",cms_consumption_rate=%g,time_until_full=%g", 450 cms_consumption_rate(), time_until_cms_gen_full()); 451 } 452 st->print(" "); 453 } 454 #endif // #ifndef PRODUCT 455 456 CMSCollector::CollectorState CMSCollector::_collectorState = 457 CMSCollector::Idling; 458 bool CMSCollector::_foregroundGCIsActive = false; 459 bool CMSCollector::_foregroundGCShouldWait = false; 460 461 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 462 CardTableRS* ct, 463 ConcurrentMarkSweepPolicy* cp): 464 _cmsGen(cmsGen), 465 _ct(ct), 466 _ref_processor(NULL), // will be set later 467 _conc_workers(NULL), // may be set later 468 _abort_preclean(false), 469 _start_sampling(false), 470 _between_prologue_and_epilogue(false), 471 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 472 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 473 -1 /* lock-free */, "No_lock" /* dummy */), 474 _modUnionClosure(&_modUnionTable), 475 _modUnionClosurePar(&_modUnionTable), 476 // Adjust my span to cover old (cms) gen 477 _span(cmsGen->reserved()), 478 // Construct the is_alive_closure with _span & markBitMap 479 _is_alive_closure(_span, &_markBitMap), 480 _restart_addr(NULL), 481 _overflow_list(NULL), 482 _stats(cmsGen), 483 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 484 _eden_chunk_array(NULL), // may be set in ctor body 485 _eden_chunk_capacity(0), // -- ditto -- 486 _eden_chunk_index(0), // -- ditto -- 487 _survivor_plab_array(NULL), // -- ditto -- 488 _survivor_chunk_array(NULL), // -- ditto -- 489 _survivor_chunk_capacity(0), // -- ditto -- 490 _survivor_chunk_index(0), // -- ditto -- 491 _ser_pmc_preclean_ovflw(0), 492 _ser_kac_preclean_ovflw(0), 493 _ser_pmc_remark_ovflw(0), 494 _par_pmc_remark_ovflw(0), 495 _ser_kac_ovflw(0), 496 _par_kac_ovflw(0), 497 #ifndef PRODUCT 498 _num_par_pushes(0), 499 #endif 500 _collection_count_start(0), 501 _verifying(false), 502 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 503 _completed_initialization(false), 504 _collector_policy(cp), 505 _should_unload_classes(CMSClassUnloadingEnabled), 506 _concurrent_cycles_since_last_unload(0), 507 _roots_scanning_options(SharedHeap::SO_None), 508 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 509 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 510 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 511 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 512 _cms_start_registered(false) 513 { 514 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 515 ExplicitGCInvokesConcurrent = true; 516 } 517 // Now expand the span and allocate the collection support structures 518 // (MUT, marking bit map etc.) to cover both generations subject to 519 // collection. 520 521 // For use by dirty card to oop closures. 522 _cmsGen->cmsSpace()->set_collector(this); 523 524 // Allocate MUT and marking bit map 525 { 526 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 527 if (!_markBitMap.allocate(_span)) { 528 warning("Failed to allocate CMS Bit Map"); 529 return; 530 } 531 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 532 } 533 { 534 _modUnionTable.allocate(_span); 535 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 536 } 537 538 if (!_markStack.allocate(MarkStackSize)) { 539 warning("Failed to allocate CMS Marking Stack"); 540 return; 541 } 542 543 // Support for multi-threaded concurrent phases 544 if (CMSConcurrentMTEnabled) { 545 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 546 // just for now 547 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 548 } 549 if (ConcGCThreads > 1) { 550 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 551 ConcGCThreads, true); 552 if (_conc_workers == NULL) { 553 warning("GC/CMS: _conc_workers allocation failure: " 554 "forcing -CMSConcurrentMTEnabled"); 555 CMSConcurrentMTEnabled = false; 556 } else { 557 _conc_workers->initialize_workers(); 558 } 559 } else { 560 CMSConcurrentMTEnabled = false; 561 } 562 } 563 if (!CMSConcurrentMTEnabled) { 564 ConcGCThreads = 0; 565 } else { 566 // Turn off CMSCleanOnEnter optimization temporarily for 567 // the MT case where it's not fixed yet; see 6178663. 568 CMSCleanOnEnter = false; 569 } 570 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 571 "Inconsistency"); 572 573 // Parallel task queues; these are shared for the 574 // concurrent and stop-world phases of CMS, but 575 // are not shared with parallel scavenge (ParNew). 576 { 577 uint i; 578 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 579 580 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 581 || ParallelRefProcEnabled) 582 && num_queues > 0) { 583 _task_queues = new OopTaskQueueSet(num_queues); 584 if (_task_queues == NULL) { 585 warning("task_queues allocation failure."); 586 return; 587 } 588 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 589 if (_hash_seed == NULL) { 590 warning("_hash_seed array allocation failure"); 591 return; 592 } 593 594 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 595 for (i = 0; i < num_queues; i++) { 596 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 597 if (q == NULL) { 598 warning("work_queue allocation failure."); 599 return; 600 } 601 _task_queues->register_queue(i, q); 602 } 603 for (i = 0; i < num_queues; i++) { 604 _task_queues->queue(i)->initialize(); 605 _hash_seed[i] = 17; // copied from ParNew 606 } 607 } 608 } 609 610 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 611 612 // Clip CMSBootstrapOccupancy between 0 and 100. 613 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 614 615 // Now tell CMS generations the identity of their collector 616 ConcurrentMarkSweepGeneration::set_collector(this); 617 618 // Create & start a CMS thread for this CMS collector 619 _cmsThread = ConcurrentMarkSweepThread::start(this); 620 assert(cmsThread() != NULL, "CMS Thread should have been created"); 621 assert(cmsThread()->collector() == this, 622 "CMS Thread should refer to this gen"); 623 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 624 625 // Support for parallelizing young gen rescan 626 GenCollectedHeap* gch = GenCollectedHeap::heap(); 627 _young_gen = gch->prev_gen(_cmsGen); 628 if (gch->supports_inline_contig_alloc()) { 629 _top_addr = gch->top_addr(); 630 _end_addr = gch->end_addr(); 631 assert(_young_gen != NULL, "no _young_gen"); 632 _eden_chunk_index = 0; 633 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 634 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 635 if (_eden_chunk_array == NULL) { 636 _eden_chunk_capacity = 0; 637 warning("GC/CMS: _eden_chunk_array allocation failure"); 638 } 639 } 640 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 641 642 // Support for parallelizing survivor space rescan 643 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 644 const size_t max_plab_samples = 645 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 646 647 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 648 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 649 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 650 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 651 || _cursor == NULL) { 652 warning("Failed to allocate survivor plab/chunk array"); 653 if (_survivor_plab_array != NULL) { 654 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 655 _survivor_plab_array = NULL; 656 } 657 if (_survivor_chunk_array != NULL) { 658 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 659 _survivor_chunk_array = NULL; 660 } 661 if (_cursor != NULL) { 662 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 663 _cursor = NULL; 664 } 665 } else { 666 _survivor_chunk_capacity = 2*max_plab_samples; 667 for (uint i = 0; i < ParallelGCThreads; i++) { 668 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 669 if (vec == NULL) { 670 warning("Failed to allocate survivor plab array"); 671 for (int j = i; j > 0; j--) { 672 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 673 } 674 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 675 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 676 _survivor_plab_array = NULL; 677 _survivor_chunk_array = NULL; 678 _survivor_chunk_capacity = 0; 679 break; 680 } else { 681 ChunkArray* cur = 682 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 683 max_plab_samples); 684 assert(cur->end() == 0, "Should be 0"); 685 assert(cur->array() == vec, "Should be vec"); 686 assert(cur->capacity() == max_plab_samples, "Error"); 687 } 688 } 689 } 690 } 691 assert( ( _survivor_plab_array != NULL 692 && _survivor_chunk_array != NULL) 693 || ( _survivor_chunk_capacity == 0 694 && _survivor_chunk_index == 0), 695 "Error"); 696 697 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 698 _gc_counters = new CollectorCounters("CMS", 1); 699 _completed_initialization = true; 700 _inter_sweep_timer.start(); // start of time 701 } 702 703 const char* ConcurrentMarkSweepGeneration::name() const { 704 return "concurrent mark-sweep generation"; 705 } 706 void ConcurrentMarkSweepGeneration::update_counters() { 707 if (UsePerfData) { 708 _space_counters->update_all(); 709 _gen_counters->update_all(); 710 } 711 } 712 713 // this is an optimized version of update_counters(). it takes the 714 // used value as a parameter rather than computing it. 715 // 716 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 717 if (UsePerfData) { 718 _space_counters->update_used(used); 719 _space_counters->update_capacity(); 720 _gen_counters->update_all(); 721 } 722 } 723 724 void ConcurrentMarkSweepGeneration::print() const { 725 Generation::print(); 726 cmsSpace()->print(); 727 } 728 729 #ifndef PRODUCT 730 void ConcurrentMarkSweepGeneration::print_statistics() { 731 cmsSpace()->printFLCensus(0); 732 } 733 #endif 734 735 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 736 GenCollectedHeap* gch = GenCollectedHeap::heap(); 737 if (PrintGCDetails) { 738 if (Verbose) { 739 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 740 level(), short_name(), s, used(), capacity()); 741 } else { 742 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 743 level(), short_name(), s, used() / K, capacity() / K); 744 } 745 } 746 if (Verbose) { 747 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 748 gch->used(), gch->capacity()); 749 } else { 750 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 751 gch->used() / K, gch->capacity() / K); 752 } 753 } 754 755 size_t 756 ConcurrentMarkSweepGeneration::contiguous_available() const { 757 // dld proposes an improvement in precision here. If the committed 758 // part of the space ends in a free block we should add that to 759 // uncommitted size in the calculation below. Will make this 760 // change later, staying with the approximation below for the 761 // time being. -- ysr. 762 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 763 } 764 765 size_t 766 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 767 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 768 } 769 770 size_t ConcurrentMarkSweepGeneration::max_available() const { 771 return free() + _virtual_space.uncommitted_size(); 772 } 773 774 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 775 size_t available = max_available(); 776 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 777 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 778 if (Verbose && PrintGCDetails) { 779 gclog_or_tty->print_cr( 780 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 781 "max_promo("SIZE_FORMAT")", 782 res? "":" not", available, res? ">=":"<", 783 av_promo, max_promotion_in_bytes); 784 } 785 return res; 786 } 787 788 // At a promotion failure dump information on block layout in heap 789 // (cms old generation). 790 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 791 if (CMSDumpAtPromotionFailure) { 792 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 793 } 794 } 795 796 CompactibleSpace* 797 ConcurrentMarkSweepGeneration::first_compaction_space() const { 798 return _cmsSpace; 799 } 800 801 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 802 // Clear the promotion information. These pointers can be adjusted 803 // along with all the other pointers into the heap but 804 // compaction is expected to be a rare event with 805 // a heap using cms so don't do it without seeing the need. 806 for (uint i = 0; i < ParallelGCThreads; i++) { 807 _par_gc_thread_states[i]->promo.reset(); 808 } 809 } 810 811 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 812 blk->do_space(_cmsSpace); 813 } 814 815 void ConcurrentMarkSweepGeneration::compute_new_size() { 816 assert_locked_or_safepoint(Heap_lock); 817 818 // If incremental collection failed, we just want to expand 819 // to the limit. 820 if (incremental_collection_failed()) { 821 clear_incremental_collection_failed(); 822 grow_to_reserved(); 823 return; 824 } 825 826 // The heap has been compacted but not reset yet. 827 // Any metric such as free() or used() will be incorrect. 828 829 CardGeneration::compute_new_size(); 830 831 // Reset again after a possible resizing 832 if (did_compact()) { 833 cmsSpace()->reset_after_compaction(); 834 } 835 } 836 837 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 838 assert_locked_or_safepoint(Heap_lock); 839 840 // If incremental collection failed, we just want to expand 841 // to the limit. 842 if (incremental_collection_failed()) { 843 clear_incremental_collection_failed(); 844 grow_to_reserved(); 845 return; 846 } 847 848 double free_percentage = ((double) free()) / capacity(); 849 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 850 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 851 852 // compute expansion delta needed for reaching desired free percentage 853 if (free_percentage < desired_free_percentage) { 854 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 855 assert(desired_capacity >= capacity(), "invalid expansion size"); 856 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 857 if (PrintGCDetails && Verbose) { 858 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 859 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 860 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 861 gclog_or_tty->print_cr(" Desired free fraction %f", 862 desired_free_percentage); 863 gclog_or_tty->print_cr(" Maximum free fraction %f", 864 maximum_free_percentage); 865 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 866 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 867 desired_capacity/1000); 868 int prev_level = level() - 1; 869 if (prev_level >= 0) { 870 size_t prev_size = 0; 871 GenCollectedHeap* gch = GenCollectedHeap::heap(); 872 Generation* prev_gen = gch->_gens[prev_level]; 873 prev_size = prev_gen->capacity(); 874 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 875 prev_size/1000); 876 } 877 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 878 unsafe_max_alloc_nogc()/1000); 879 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 880 contiguous_available()/1000); 881 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 882 expand_bytes); 883 } 884 // safe if expansion fails 885 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 886 if (PrintGCDetails && Verbose) { 887 gclog_or_tty->print_cr(" Expanded free fraction %f", 888 ((double) free()) / capacity()); 889 } 890 } else { 891 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 892 assert(desired_capacity <= capacity(), "invalid expansion size"); 893 size_t shrink_bytes = capacity() - desired_capacity; 894 // Don't shrink unless the delta is greater than the minimum shrink we want 895 if (shrink_bytes >= MinHeapDeltaBytes) { 896 shrink_free_list_by(shrink_bytes); 897 } 898 } 899 } 900 901 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 902 return cmsSpace()->freelistLock(); 903 } 904 905 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 906 bool tlab) { 907 CMSSynchronousYieldRequest yr; 908 MutexLockerEx x(freelistLock(), 909 Mutex::_no_safepoint_check_flag); 910 return have_lock_and_allocate(size, tlab); 911 } 912 913 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 914 bool tlab /* ignored */) { 915 assert_lock_strong(freelistLock()); 916 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 917 HeapWord* res = cmsSpace()->allocate(adjustedSize); 918 // Allocate the object live (grey) if the background collector has 919 // started marking. This is necessary because the marker may 920 // have passed this address and consequently this object will 921 // not otherwise be greyed and would be incorrectly swept up. 922 // Note that if this object contains references, the writing 923 // of those references will dirty the card containing this object 924 // allowing the object to be blackened (and its references scanned) 925 // either during a preclean phase or at the final checkpoint. 926 if (res != NULL) { 927 // We may block here with an uninitialized object with 928 // its mark-bit or P-bits not yet set. Such objects need 929 // to be safely navigable by block_start(). 930 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 931 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 932 collector()->direct_allocated(res, adjustedSize); 933 _direct_allocated_words += adjustedSize; 934 // allocation counters 935 NOT_PRODUCT( 936 _numObjectsAllocated++; 937 _numWordsAllocated += (int)adjustedSize; 938 ) 939 } 940 return res; 941 } 942 943 // In the case of direct allocation by mutators in a generation that 944 // is being concurrently collected, the object must be allocated 945 // live (grey) if the background collector has started marking. 946 // This is necessary because the marker may 947 // have passed this address and consequently this object will 948 // not otherwise be greyed and would be incorrectly swept up. 949 // Note that if this object contains references, the writing 950 // of those references will dirty the card containing this object 951 // allowing the object to be blackened (and its references scanned) 952 // either during a preclean phase or at the final checkpoint. 953 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 954 assert(_markBitMap.covers(start, size), "Out of bounds"); 955 if (_collectorState >= Marking) { 956 MutexLockerEx y(_markBitMap.lock(), 957 Mutex::_no_safepoint_check_flag); 958 // [see comments preceding SweepClosure::do_blk() below for details] 959 // 960 // Can the P-bits be deleted now? JJJ 961 // 962 // 1. need to mark the object as live so it isn't collected 963 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 964 // 3. need to mark the end of the object so marking, precleaning or sweeping 965 // can skip over uninitialized or unparsable objects. An allocated 966 // object is considered uninitialized for our purposes as long as 967 // its klass word is NULL. All old gen objects are parsable 968 // as soon as they are initialized.) 969 _markBitMap.mark(start); // object is live 970 _markBitMap.mark(start + 1); // object is potentially uninitialized? 971 _markBitMap.mark(start + size - 1); 972 // mark end of object 973 } 974 // check that oop looks uninitialized 975 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 976 } 977 978 void CMSCollector::promoted(bool par, HeapWord* start, 979 bool is_obj_array, size_t obj_size) { 980 assert(_markBitMap.covers(start), "Out of bounds"); 981 // See comment in direct_allocated() about when objects should 982 // be allocated live. 983 if (_collectorState >= Marking) { 984 // we already hold the marking bit map lock, taken in 985 // the prologue 986 if (par) { 987 _markBitMap.par_mark(start); 988 } else { 989 _markBitMap.mark(start); 990 } 991 // We don't need to mark the object as uninitialized (as 992 // in direct_allocated above) because this is being done with the 993 // world stopped and the object will be initialized by the 994 // time the marking, precleaning or sweeping get to look at it. 995 // But see the code for copying objects into the CMS generation, 996 // where we need to ensure that concurrent readers of the 997 // block offset table are able to safely navigate a block that 998 // is in flux from being free to being allocated (and in 999 // transition while being copied into) and subsequently 1000 // becoming a bona-fide object when the copy/promotion is complete. 1001 assert(SafepointSynchronize::is_at_safepoint(), 1002 "expect promotion only at safepoints"); 1003 1004 if (_collectorState < Sweeping) { 1005 // Mark the appropriate cards in the modUnionTable, so that 1006 // this object gets scanned before the sweep. If this is 1007 // not done, CMS generation references in the object might 1008 // not get marked. 1009 // For the case of arrays, which are otherwise precisely 1010 // marked, we need to dirty the entire array, not just its head. 1011 if (is_obj_array) { 1012 // The [par_]mark_range() method expects mr.end() below to 1013 // be aligned to the granularity of a bit's representation 1014 // in the heap. In the case of the MUT below, that's a 1015 // card size. 1016 MemRegion mr(start, 1017 (HeapWord*)round_to((intptr_t)(start + obj_size), 1018 CardTableModRefBS::card_size /* bytes */)); 1019 if (par) { 1020 _modUnionTable.par_mark_range(mr); 1021 } else { 1022 _modUnionTable.mark_range(mr); 1023 } 1024 } else { // not an obj array; we can just mark the head 1025 if (par) { 1026 _modUnionTable.par_mark(start); 1027 } else { 1028 _modUnionTable.mark(start); 1029 } 1030 } 1031 } 1032 } 1033 } 1034 1035 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1036 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1037 // allocate, copy and if necessary update promoinfo -- 1038 // delegate to underlying space. 1039 assert_lock_strong(freelistLock()); 1040 1041 #ifndef PRODUCT 1042 if (Universe::heap()->promotion_should_fail()) { 1043 return NULL; 1044 } 1045 #endif // #ifndef PRODUCT 1046 1047 oop res = _cmsSpace->promote(obj, obj_size); 1048 if (res == NULL) { 1049 // expand and retry 1050 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1051 expand(s*HeapWordSize, MinHeapDeltaBytes, 1052 CMSExpansionCause::_satisfy_promotion); 1053 // Since there's currently no next generation, we don't try to promote 1054 // into a more senior generation. 1055 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1056 "is made to pass on a possibly failing " 1057 "promotion to next generation"); 1058 res = _cmsSpace->promote(obj, obj_size); 1059 } 1060 if (res != NULL) { 1061 // See comment in allocate() about when objects should 1062 // be allocated live. 1063 assert(obj->is_oop(), "Will dereference klass pointer below"); 1064 collector()->promoted(false, // Not parallel 1065 (HeapWord*)res, obj->is_objArray(), obj_size); 1066 // promotion counters 1067 NOT_PRODUCT( 1068 _numObjectsPromoted++; 1069 _numWordsPromoted += 1070 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1071 ) 1072 } 1073 return res; 1074 } 1075 1076 1077 // IMPORTANT: Notes on object size recognition in CMS. 1078 // --------------------------------------------------- 1079 // A block of storage in the CMS generation is always in 1080 // one of three states. A free block (FREE), an allocated 1081 // object (OBJECT) whose size() method reports the correct size, 1082 // and an intermediate state (TRANSIENT) in which its size cannot 1083 // be accurately determined. 1084 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1085 // ----------------------------------------------------- 1086 // FREE: klass_word & 1 == 1; mark_word holds block size 1087 // 1088 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1089 // obj->size() computes correct size 1090 // 1091 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1092 // 1093 // STATE IDENTIFICATION: (64 bit+COOPS) 1094 // ------------------------------------ 1095 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1096 // 1097 // OBJECT: klass_word installed; klass_word != 0; 1098 // obj->size() computes correct size 1099 // 1100 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1101 // 1102 // 1103 // STATE TRANSITION DIAGRAM 1104 // 1105 // mut / parnew mut / parnew 1106 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1107 // ^ | 1108 // |------------------------ DEAD <------------------------------------| 1109 // sweep mut 1110 // 1111 // While a block is in TRANSIENT state its size cannot be determined 1112 // so readers will either need to come back later or stall until 1113 // the size can be determined. Note that for the case of direct 1114 // allocation, P-bits, when available, may be used to determine the 1115 // size of an object that may not yet have been initialized. 1116 1117 // Things to support parallel young-gen collection. 1118 oop 1119 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1120 oop old, markOop m, 1121 size_t word_sz) { 1122 #ifndef PRODUCT 1123 if (Universe::heap()->promotion_should_fail()) { 1124 return NULL; 1125 } 1126 #endif // #ifndef PRODUCT 1127 1128 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1129 PromotionInfo* promoInfo = &ps->promo; 1130 // if we are tracking promotions, then first ensure space for 1131 // promotion (including spooling space for saving header if necessary). 1132 // then allocate and copy, then track promoted info if needed. 1133 // When tracking (see PromotionInfo::track()), the mark word may 1134 // be displaced and in this case restoration of the mark word 1135 // occurs in the (oop_since_save_marks_)iterate phase. 1136 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1137 // Out of space for allocating spooling buffers; 1138 // try expanding and allocating spooling buffers. 1139 if (!expand_and_ensure_spooling_space(promoInfo)) { 1140 return NULL; 1141 } 1142 } 1143 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1144 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1145 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1146 if (obj_ptr == NULL) { 1147 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1148 if (obj_ptr == NULL) { 1149 return NULL; 1150 } 1151 } 1152 oop obj = oop(obj_ptr); 1153 OrderAccess::storestore(); 1154 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1155 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1156 // IMPORTANT: See note on object initialization for CMS above. 1157 // Otherwise, copy the object. Here we must be careful to insert the 1158 // klass pointer last, since this marks the block as an allocated object. 1159 // Except with compressed oops it's the mark word. 1160 HeapWord* old_ptr = (HeapWord*)old; 1161 // Restore the mark word copied above. 1162 obj->set_mark(m); 1163 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1164 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1165 OrderAccess::storestore(); 1166 1167 if (UseCompressedClassPointers) { 1168 // Copy gap missed by (aligned) header size calculation below 1169 obj->set_klass_gap(old->klass_gap()); 1170 } 1171 if (word_sz > (size_t)oopDesc::header_size()) { 1172 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1173 obj_ptr + oopDesc::header_size(), 1174 word_sz - oopDesc::header_size()); 1175 } 1176 1177 // Now we can track the promoted object, if necessary. We take care 1178 // to delay the transition from uninitialized to full object 1179 // (i.e., insertion of klass pointer) until after, so that it 1180 // atomically becomes a promoted object. 1181 if (promoInfo->tracking()) { 1182 promoInfo->track((PromotedObject*)obj, old->klass()); 1183 } 1184 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1185 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1186 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1187 1188 // Finally, install the klass pointer (this should be volatile). 1189 OrderAccess::storestore(); 1190 obj->set_klass(old->klass()); 1191 // We should now be able to calculate the right size for this object 1192 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1193 1194 collector()->promoted(true, // parallel 1195 obj_ptr, old->is_objArray(), word_sz); 1196 1197 NOT_PRODUCT( 1198 Atomic::inc_ptr(&_numObjectsPromoted); 1199 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1200 ) 1201 1202 return obj; 1203 } 1204 1205 void 1206 ConcurrentMarkSweepGeneration:: 1207 par_promote_alloc_undo(int thread_num, 1208 HeapWord* obj, size_t word_sz) { 1209 // CMS does not support promotion undo. 1210 ShouldNotReachHere(); 1211 } 1212 1213 void 1214 ConcurrentMarkSweepGeneration:: 1215 par_promote_alloc_done(int thread_num) { 1216 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1217 ps->lab.retire(thread_num); 1218 } 1219 1220 void 1221 ConcurrentMarkSweepGeneration:: 1222 par_oop_since_save_marks_iterate_done(int thread_num) { 1223 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1224 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1225 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1226 } 1227 1228 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1229 size_t size, 1230 bool tlab) 1231 { 1232 // We allow a STW collection only if a full 1233 // collection was requested. 1234 return full || should_allocate(size, tlab); // FIX ME !!! 1235 // This and promotion failure handling are connected at the 1236 // hip and should be fixed by untying them. 1237 } 1238 1239 bool CMSCollector::shouldConcurrentCollect() { 1240 if (_full_gc_requested) { 1241 if (Verbose && PrintGCDetails) { 1242 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1243 " gc request (or gc_locker)"); 1244 } 1245 return true; 1246 } 1247 1248 // For debugging purposes, change the type of collection. 1249 // Rotate between concurrent and stop-the-world full GCs. 1250 NOT_PRODUCT( 1251 if (RotateCMSCollectionTypes) { 1252 return _cmsGen->debug_concurrent_cycle(); 1253 } 1254 ) 1255 1256 FreelistLocker x(this); 1257 // ------------------------------------------------------------------ 1258 // Print out lots of information which affects the initiation of 1259 // a collection. 1260 if (PrintCMSInitiationStatistics && stats().valid()) { 1261 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1262 gclog_or_tty->stamp(); 1263 gclog_or_tty->cr(); 1264 stats().print_on(gclog_or_tty); 1265 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1266 stats().time_until_cms_gen_full()); 1267 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1268 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1269 _cmsGen->contiguous_available()); 1270 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1271 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1272 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1273 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1274 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1275 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1276 gclog_or_tty->print_cr("metadata initialized %d", 1277 MetaspaceGC::should_concurrent_collect()); 1278 } 1279 // ------------------------------------------------------------------ 1280 1281 // If the estimated time to complete a cms collection (cms_duration()) 1282 // is less than the estimated time remaining until the cms generation 1283 // is full, start a collection. 1284 if (!UseCMSInitiatingOccupancyOnly) { 1285 if (stats().valid()) { 1286 if (stats().time_until_cms_start() == 0.0) { 1287 return true; 1288 } 1289 } else { 1290 // We want to conservatively collect somewhat early in order 1291 // to try and "bootstrap" our CMS/promotion statistics; 1292 // this branch will not fire after the first successful CMS 1293 // collection because the stats should then be valid. 1294 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1295 if (Verbose && PrintGCDetails) { 1296 gclog_or_tty->print_cr( 1297 " CMSCollector: collect for bootstrapping statistics:" 1298 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1299 _bootstrap_occupancy); 1300 } 1301 return true; 1302 } 1303 } 1304 } 1305 1306 // Otherwise, we start a collection cycle if 1307 // old gen want a collection cycle started. Each may use 1308 // an appropriate criterion for making this decision. 1309 // XXX We need to make sure that the gen expansion 1310 // criterion dovetails well with this. XXX NEED TO FIX THIS 1311 if (_cmsGen->should_concurrent_collect()) { 1312 if (Verbose && PrintGCDetails) { 1313 gclog_or_tty->print_cr("CMS old gen initiated"); 1314 } 1315 return true; 1316 } 1317 1318 // We start a collection if we believe an incremental collection may fail; 1319 // this is not likely to be productive in practice because it's probably too 1320 // late anyway. 1321 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1322 assert(gch->collector_policy()->is_generation_policy(), 1323 "You may want to check the correctness of the following"); 1324 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1325 if (Verbose && PrintGCDetails) { 1326 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1327 } 1328 return true; 1329 } 1330 1331 if (MetaspaceGC::should_concurrent_collect()) { 1332 if (Verbose && PrintGCDetails) { 1333 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1334 } 1335 return true; 1336 } 1337 1338 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1339 if (CMSTriggerInterval >= 0) { 1340 if (CMSTriggerInterval == 0) { 1341 // Trigger always 1342 return true; 1343 } 1344 1345 // Check the CMS time since begin (we do not check the stats validity 1346 // as we want to be able to trigger the first CMS cycle as well) 1347 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1348 if (Verbose && PrintGCDetails) { 1349 if (stats().valid()) { 1350 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1351 stats().cms_time_since_begin()); 1352 } else { 1353 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1354 } 1355 } 1356 return true; 1357 } 1358 } 1359 1360 return false; 1361 } 1362 1363 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1364 1365 // Clear _expansion_cause fields of constituent generations 1366 void CMSCollector::clear_expansion_cause() { 1367 _cmsGen->clear_expansion_cause(); 1368 } 1369 1370 // We should be conservative in starting a collection cycle. To 1371 // start too eagerly runs the risk of collecting too often in the 1372 // extreme. To collect too rarely falls back on full collections, 1373 // which works, even if not optimum in terms of concurrent work. 1374 // As a work around for too eagerly collecting, use the flag 1375 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1376 // giving the user an easily understandable way of controlling the 1377 // collections. 1378 // We want to start a new collection cycle if any of the following 1379 // conditions hold: 1380 // . our current occupancy exceeds the configured initiating occupancy 1381 // for this generation, or 1382 // . we recently needed to expand this space and have not, since that 1383 // expansion, done a collection of this generation, or 1384 // . the underlying space believes that it may be a good idea to initiate 1385 // a concurrent collection (this may be based on criteria such as the 1386 // following: the space uses linear allocation and linear allocation is 1387 // going to fail, or there is believed to be excessive fragmentation in 1388 // the generation, etc... or ... 1389 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1390 // the case of the old generation; see CR 6543076): 1391 // we may be approaching a point at which allocation requests may fail because 1392 // we will be out of sufficient free space given allocation rate estimates.] 1393 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1394 1395 assert_lock_strong(freelistLock()); 1396 if (occupancy() > initiating_occupancy()) { 1397 if (PrintGCDetails && Verbose) { 1398 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1399 short_name(), occupancy(), initiating_occupancy()); 1400 } 1401 return true; 1402 } 1403 if (UseCMSInitiatingOccupancyOnly) { 1404 return false; 1405 } 1406 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1407 if (PrintGCDetails && Verbose) { 1408 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1409 short_name()); 1410 } 1411 return true; 1412 } 1413 if (_cmsSpace->should_concurrent_collect()) { 1414 if (PrintGCDetails && Verbose) { 1415 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1416 short_name()); 1417 } 1418 return true; 1419 } 1420 return false; 1421 } 1422 1423 void ConcurrentMarkSweepGeneration::collect(bool full, 1424 bool clear_all_soft_refs, 1425 size_t size, 1426 bool tlab) 1427 { 1428 collector()->collect(full, clear_all_soft_refs, size, tlab); 1429 } 1430 1431 void CMSCollector::collect(bool full, 1432 bool clear_all_soft_refs, 1433 size_t size, 1434 bool tlab) 1435 { 1436 // The following "if" branch is present for defensive reasons. 1437 // In the current uses of this interface, it can be replaced with: 1438 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1439 // But I am not placing that assert here to allow future 1440 // generality in invoking this interface. 1441 if (GC_locker::is_active()) { 1442 // A consistency test for GC_locker 1443 assert(GC_locker::needs_gc(), "Should have been set already"); 1444 // Skip this foreground collection, instead 1445 // expanding the heap if necessary. 1446 // Need the free list locks for the call to free() in compute_new_size() 1447 compute_new_size(); 1448 return; 1449 } 1450 acquire_control_and_collect(full, clear_all_soft_refs); 1451 } 1452 1453 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1454 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1455 unsigned int gc_count = gch->total_full_collections(); 1456 if (gc_count == full_gc_count) { 1457 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1458 _full_gc_requested = true; 1459 _full_gc_cause = cause; 1460 CGC_lock->notify(); // nudge CMS thread 1461 } else { 1462 assert(gc_count > full_gc_count, "Error: causal loop"); 1463 } 1464 } 1465 1466 bool CMSCollector::is_external_interruption() { 1467 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1468 return GCCause::is_user_requested_gc(cause) || 1469 GCCause::is_serviceability_requested_gc(cause); 1470 } 1471 1472 void CMSCollector::report_concurrent_mode_interruption() { 1473 if (is_external_interruption()) { 1474 if (PrintGCDetails) { 1475 gclog_or_tty->print(" (concurrent mode interrupted)"); 1476 } 1477 } else { 1478 if (PrintGCDetails) { 1479 gclog_or_tty->print(" (concurrent mode failure)"); 1480 } 1481 _gc_tracer_cm->report_concurrent_mode_failure(); 1482 } 1483 } 1484 1485 1486 // The foreground and background collectors need to coordinate in order 1487 // to make sure that they do not mutually interfere with CMS collections. 1488 // When a background collection is active, 1489 // the foreground collector may need to take over (preempt) and 1490 // synchronously complete an ongoing collection. Depending on the 1491 // frequency of the background collections and the heap usage 1492 // of the application, this preemption can be seldom or frequent. 1493 // There are only certain 1494 // points in the background collection that the "collection-baton" 1495 // can be passed to the foreground collector. 1496 // 1497 // The foreground collector will wait for the baton before 1498 // starting any part of the collection. The foreground collector 1499 // will only wait at one location. 1500 // 1501 // The background collector will yield the baton before starting a new 1502 // phase of the collection (e.g., before initial marking, marking from roots, 1503 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1504 // of the loop which switches the phases. The background collector does some 1505 // of the phases (initial mark, final re-mark) with the world stopped. 1506 // Because of locking involved in stopping the world, 1507 // the foreground collector should not block waiting for the background 1508 // collector when it is doing a stop-the-world phase. The background 1509 // collector will yield the baton at an additional point just before 1510 // it enters a stop-the-world phase. Once the world is stopped, the 1511 // background collector checks the phase of the collection. If the 1512 // phase has not changed, it proceeds with the collection. If the 1513 // phase has changed, it skips that phase of the collection. See 1514 // the comments on the use of the Heap_lock in collect_in_background(). 1515 // 1516 // Variable used in baton passing. 1517 // _foregroundGCIsActive - Set to true by the foreground collector when 1518 // it wants the baton. The foreground clears it when it has finished 1519 // the collection. 1520 // _foregroundGCShouldWait - Set to true by the background collector 1521 // when it is running. The foreground collector waits while 1522 // _foregroundGCShouldWait is true. 1523 // CGC_lock - monitor used to protect access to the above variables 1524 // and to notify the foreground and background collectors. 1525 // _collectorState - current state of the CMS collection. 1526 // 1527 // The foreground collector 1528 // acquires the CGC_lock 1529 // sets _foregroundGCIsActive 1530 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1531 // various locks acquired in preparation for the collection 1532 // are released so as not to block the background collector 1533 // that is in the midst of a collection 1534 // proceeds with the collection 1535 // clears _foregroundGCIsActive 1536 // returns 1537 // 1538 // The background collector in a loop iterating on the phases of the 1539 // collection 1540 // acquires the CGC_lock 1541 // sets _foregroundGCShouldWait 1542 // if _foregroundGCIsActive is set 1543 // clears _foregroundGCShouldWait, notifies _CGC_lock 1544 // waits on _CGC_lock for _foregroundGCIsActive to become false 1545 // and exits the loop. 1546 // otherwise 1547 // proceed with that phase of the collection 1548 // if the phase is a stop-the-world phase, 1549 // yield the baton once more just before enqueueing 1550 // the stop-world CMS operation (executed by the VM thread). 1551 // returns after all phases of the collection are done 1552 // 1553 1554 void CMSCollector::acquire_control_and_collect(bool full, 1555 bool clear_all_soft_refs) { 1556 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1557 assert(!Thread::current()->is_ConcurrentGC_thread(), 1558 "shouldn't try to acquire control from self!"); 1559 1560 // Start the protocol for acquiring control of the 1561 // collection from the background collector (aka CMS thread). 1562 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1563 "VM thread should have CMS token"); 1564 // Remember the possibly interrupted state of an ongoing 1565 // concurrent collection 1566 CollectorState first_state = _collectorState; 1567 1568 // Signal to a possibly ongoing concurrent collection that 1569 // we want to do a foreground collection. 1570 _foregroundGCIsActive = true; 1571 1572 // release locks and wait for a notify from the background collector 1573 // releasing the locks in only necessary for phases which 1574 // do yields to improve the granularity of the collection. 1575 assert_lock_strong(bitMapLock()); 1576 // We need to lock the Free list lock for the space that we are 1577 // currently collecting. 1578 assert(haveFreelistLocks(), "Must be holding free list locks"); 1579 bitMapLock()->unlock(); 1580 releaseFreelistLocks(); 1581 { 1582 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1583 if (_foregroundGCShouldWait) { 1584 // We are going to be waiting for action for the CMS thread; 1585 // it had better not be gone (for instance at shutdown)! 1586 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1587 "CMS thread must be running"); 1588 // Wait here until the background collector gives us the go-ahead 1589 ConcurrentMarkSweepThread::clear_CMS_flag( 1590 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1591 // Get a possibly blocked CMS thread going: 1592 // Note that we set _foregroundGCIsActive true above, 1593 // without protection of the CGC_lock. 1594 CGC_lock->notify(); 1595 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1596 "Possible deadlock"); 1597 while (_foregroundGCShouldWait) { 1598 // wait for notification 1599 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1600 // Possibility of delay/starvation here, since CMS token does 1601 // not know to give priority to VM thread? Actually, i think 1602 // there wouldn't be any delay/starvation, but the proof of 1603 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1604 } 1605 ConcurrentMarkSweepThread::set_CMS_flag( 1606 ConcurrentMarkSweepThread::CMS_vm_has_token); 1607 } 1608 } 1609 // The CMS_token is already held. Get back the other locks. 1610 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1611 "VM thread should have CMS token"); 1612 getFreelistLocks(); 1613 bitMapLock()->lock_without_safepoint_check(); 1614 if (TraceCMSState) { 1615 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1616 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1617 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1618 } 1619 1620 if (first_state > Idling) { 1621 report_concurrent_mode_interruption(); 1622 } 1623 1624 set_did_compact(true); 1625 1626 // If the collection is being acquired from the background 1627 // collector, there may be references on the discovered 1628 // references lists that have NULL referents (being those 1629 // that were concurrently cleared by a mutator) or 1630 // that are no longer active (having been enqueued concurrently 1631 // by the mutator). 1632 // Scrub the list of those references because Mark-Sweep-Compact 1633 // code assumes referents are not NULL and that all discovered 1634 // Reference objects are active. 1635 ref_processor()->clean_up_discovered_references(); 1636 1637 if (first_state > Idling) { 1638 save_heap_summary(); 1639 } 1640 1641 do_compaction_work(clear_all_soft_refs); 1642 1643 // Has the GC time limit been exceeded? 1644 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1645 size_t max_eden_size = young_gen->max_capacity() - 1646 young_gen->to()->capacity() - 1647 young_gen->from()->capacity(); 1648 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1649 GCCause::Cause gc_cause = gch->gc_cause(); 1650 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1651 young_gen->eden()->used(), 1652 _cmsGen->max_capacity(), 1653 max_eden_size, 1654 full, 1655 gc_cause, 1656 gch->collector_policy()); 1657 1658 // Reset the expansion cause, now that we just completed 1659 // a collection cycle. 1660 clear_expansion_cause(); 1661 _foregroundGCIsActive = false; 1662 return; 1663 } 1664 1665 // Resize the tenured generation 1666 // after obtaining the free list locks for the 1667 // two generations. 1668 void CMSCollector::compute_new_size() { 1669 assert_locked_or_safepoint(Heap_lock); 1670 FreelistLocker z(this); 1671 MetaspaceGC::compute_new_size(); 1672 _cmsGen->compute_new_size_free_list(); 1673 } 1674 1675 // A work method used by the foreground collector to do 1676 // a mark-sweep-compact. 1677 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1678 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1679 1680 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1681 gc_timer->register_gc_start(); 1682 1683 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1684 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1685 1686 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 1687 1688 // Temporarily widen the span of the weak reference processing to 1689 // the entire heap. 1690 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1691 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1692 // Temporarily, clear the "is_alive_non_header" field of the 1693 // reference processor. 1694 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1695 // Temporarily make reference _processing_ single threaded (non-MT). 1696 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1697 // Temporarily make refs discovery atomic 1698 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1699 // Temporarily make reference _discovery_ single threaded (non-MT) 1700 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1701 1702 ref_processor()->set_enqueuing_is_done(false); 1703 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 1704 ref_processor()->setup_policy(clear_all_soft_refs); 1705 // If an asynchronous collection finishes, the _modUnionTable is 1706 // all clear. If we are assuming the collection from an asynchronous 1707 // collection, clear the _modUnionTable. 1708 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1709 "_modUnionTable should be clear if the baton was not passed"); 1710 _modUnionTable.clear_all(); 1711 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1712 "mod union for klasses should be clear if the baton was passed"); 1713 _ct->klass_rem_set()->clear_mod_union(); 1714 1715 // We must adjust the allocation statistics being maintained 1716 // in the free list space. We do so by reading and clearing 1717 // the sweep timer and updating the block flux rate estimates below. 1718 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1719 if (_inter_sweep_timer.is_active()) { 1720 _inter_sweep_timer.stop(); 1721 // Note that we do not use this sample to update the _inter_sweep_estimate. 1722 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1723 _inter_sweep_estimate.padded_average(), 1724 _intra_sweep_estimate.padded_average()); 1725 } 1726 1727 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1728 ref_processor(), clear_all_soft_refs); 1729 #ifdef ASSERT 1730 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1731 size_t free_size = cms_space->free(); 1732 assert(free_size == 1733 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1734 * HeapWordSize, 1735 "All the free space should be compacted into one chunk at top"); 1736 assert(cms_space->dictionary()->total_chunk_size( 1737 debug_only(cms_space->freelistLock())) == 0 || 1738 cms_space->totalSizeInIndexedFreeLists() == 0, 1739 "All the free space should be in a single chunk"); 1740 size_t num = cms_space->totalCount(); 1741 assert((free_size == 0 && num == 0) || 1742 (free_size > 0 && (num == 1 || num == 2)), 1743 "There should be at most 2 free chunks after compaction"); 1744 #endif // ASSERT 1745 _collectorState = Resetting; 1746 assert(_restart_addr == NULL, 1747 "Should have been NULL'd before baton was passed"); 1748 reset(false /* == !concurrent */); 1749 _cmsGen->reset_after_compaction(); 1750 _concurrent_cycles_since_last_unload = 0; 1751 1752 // Clear any data recorded in the PLAB chunk arrays. 1753 if (_survivor_plab_array != NULL) { 1754 reset_survivor_plab_arrays(); 1755 } 1756 1757 // Adjust the per-size allocation stats for the next epoch. 1758 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1759 // Restart the "inter sweep timer" for the next epoch. 1760 _inter_sweep_timer.reset(); 1761 _inter_sweep_timer.start(); 1762 1763 gc_timer->register_gc_end(); 1764 1765 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1766 1767 // For a mark-sweep-compact, compute_new_size() will be called 1768 // in the heap's do_collection() method. 1769 } 1770 1771 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1772 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 1773 ContiguousSpace* eden_space = dng->eden(); 1774 ContiguousSpace* from_space = dng->from(); 1775 ContiguousSpace* to_space = dng->to(); 1776 // Eden 1777 if (_eden_chunk_array != NULL) { 1778 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1779 eden_space->bottom(), eden_space->top(), 1780 eden_space->end(), eden_space->capacity()); 1781 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1782 "_eden_chunk_capacity=" SIZE_FORMAT, 1783 _eden_chunk_index, _eden_chunk_capacity); 1784 for (size_t i = 0; i < _eden_chunk_index; i++) { 1785 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1786 i, _eden_chunk_array[i]); 1787 } 1788 } 1789 // Survivor 1790 if (_survivor_chunk_array != NULL) { 1791 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1792 from_space->bottom(), from_space->top(), 1793 from_space->end(), from_space->capacity()); 1794 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1795 "_survivor_chunk_capacity=" SIZE_FORMAT, 1796 _survivor_chunk_index, _survivor_chunk_capacity); 1797 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1798 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1799 i, _survivor_chunk_array[i]); 1800 } 1801 } 1802 } 1803 1804 void CMSCollector::getFreelistLocks() const { 1805 // Get locks for all free lists in all generations that this 1806 // collector is responsible for 1807 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1808 } 1809 1810 void CMSCollector::releaseFreelistLocks() const { 1811 // Release locks for all free lists in all generations that this 1812 // collector is responsible for 1813 _cmsGen->freelistLock()->unlock(); 1814 } 1815 1816 bool CMSCollector::haveFreelistLocks() const { 1817 // Check locks for all free lists in all generations that this 1818 // collector is responsible for 1819 assert_lock_strong(_cmsGen->freelistLock()); 1820 PRODUCT_ONLY(ShouldNotReachHere()); 1821 return true; 1822 } 1823 1824 // A utility class that is used by the CMS collector to 1825 // temporarily "release" the foreground collector from its 1826 // usual obligation to wait for the background collector to 1827 // complete an ongoing phase before proceeding. 1828 class ReleaseForegroundGC: public StackObj { 1829 private: 1830 CMSCollector* _c; 1831 public: 1832 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1833 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1834 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1835 // allow a potentially blocked foreground collector to proceed 1836 _c->_foregroundGCShouldWait = false; 1837 if (_c->_foregroundGCIsActive) { 1838 CGC_lock->notify(); 1839 } 1840 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1841 "Possible deadlock"); 1842 } 1843 1844 ~ReleaseForegroundGC() { 1845 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1846 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1847 _c->_foregroundGCShouldWait = true; 1848 } 1849 }; 1850 1851 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1852 assert(Thread::current()->is_ConcurrentGC_thread(), 1853 "A CMS asynchronous collection is only allowed on a CMS thread."); 1854 1855 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1856 { 1857 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1858 MutexLockerEx hl(Heap_lock, safepoint_check); 1859 FreelistLocker fll(this); 1860 MutexLockerEx x(CGC_lock, safepoint_check); 1861 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 1862 // The foreground collector is active or we're 1863 // not using asynchronous collections. Skip this 1864 // background collection. 1865 assert(!_foregroundGCShouldWait, "Should be clear"); 1866 return; 1867 } else { 1868 assert(_collectorState == Idling, "Should be idling before start."); 1869 _collectorState = InitialMarking; 1870 register_gc_start(cause); 1871 // Reset the expansion cause, now that we are about to begin 1872 // a new cycle. 1873 clear_expansion_cause(); 1874 1875 // Clear the MetaspaceGC flag since a concurrent collection 1876 // is starting but also clear it after the collection. 1877 MetaspaceGC::set_should_concurrent_collect(false); 1878 } 1879 // Decide if we want to enable class unloading as part of the 1880 // ensuing concurrent GC cycle. 1881 update_should_unload_classes(); 1882 _full_gc_requested = false; // acks all outstanding full gc requests 1883 _full_gc_cause = GCCause::_no_gc; 1884 // Signal that we are about to start a collection 1885 gch->increment_total_full_collections(); // ... starting a collection cycle 1886 _collection_count_start = gch->total_full_collections(); 1887 } 1888 1889 // Used for PrintGC 1890 size_t prev_used; 1891 if (PrintGC && Verbose) { 1892 prev_used = _cmsGen->used(); 1893 } 1894 1895 // The change of the collection state is normally done at this level; 1896 // the exceptions are phases that are executed while the world is 1897 // stopped. For those phases the change of state is done while the 1898 // world is stopped. For baton passing purposes this allows the 1899 // background collector to finish the phase and change state atomically. 1900 // The foreground collector cannot wait on a phase that is done 1901 // while the world is stopped because the foreground collector already 1902 // has the world stopped and would deadlock. 1903 while (_collectorState != Idling) { 1904 if (TraceCMSState) { 1905 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1906 Thread::current(), _collectorState); 1907 } 1908 // The foreground collector 1909 // holds the Heap_lock throughout its collection. 1910 // holds the CMS token (but not the lock) 1911 // except while it is waiting for the background collector to yield. 1912 // 1913 // The foreground collector should be blocked (not for long) 1914 // if the background collector is about to start a phase 1915 // executed with world stopped. If the background 1916 // collector has already started such a phase, the 1917 // foreground collector is blocked waiting for the 1918 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1919 // are executed in the VM thread. 1920 // 1921 // The locking order is 1922 // PendingListLock (PLL) -- if applicable (FinalMarking) 1923 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1924 // CMS token (claimed in 1925 // stop_world_and_do() --> 1926 // safepoint_synchronize() --> 1927 // CMSThread::synchronize()) 1928 1929 { 1930 // Check if the FG collector wants us to yield. 1931 CMSTokenSync x(true); // is cms thread 1932 if (waitForForegroundGC()) { 1933 // We yielded to a foreground GC, nothing more to be 1934 // done this round. 1935 assert(_foregroundGCShouldWait == false, "We set it to false in " 1936 "waitForForegroundGC()"); 1937 if (TraceCMSState) { 1938 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1939 " exiting collection CMS state %d", 1940 Thread::current(), _collectorState); 1941 } 1942 return; 1943 } else { 1944 // The background collector can run but check to see if the 1945 // foreground collector has done a collection while the 1946 // background collector was waiting to get the CGC_lock 1947 // above. If yes, break so that _foregroundGCShouldWait 1948 // is cleared before returning. 1949 if (_collectorState == Idling) { 1950 break; 1951 } 1952 } 1953 } 1954 1955 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1956 "should be waiting"); 1957 1958 switch (_collectorState) { 1959 case InitialMarking: 1960 { 1961 ReleaseForegroundGC x(this); 1962 stats().record_cms_begin(); 1963 VM_CMS_Initial_Mark initial_mark_op(this); 1964 VMThread::execute(&initial_mark_op); 1965 } 1966 // The collector state may be any legal state at this point 1967 // since the background collector may have yielded to the 1968 // foreground collector. 1969 break; 1970 case Marking: 1971 // initial marking in checkpointRootsInitialWork has been completed 1972 if (markFromRoots()) { // we were successful 1973 assert(_collectorState == Precleaning, "Collector state should " 1974 "have changed"); 1975 } else { 1976 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1977 } 1978 break; 1979 case Precleaning: 1980 // marking from roots in markFromRoots has been completed 1981 preclean(); 1982 assert(_collectorState == AbortablePreclean || 1983 _collectorState == FinalMarking, 1984 "Collector state should have changed"); 1985 break; 1986 case AbortablePreclean: 1987 abortable_preclean(); 1988 assert(_collectorState == FinalMarking, "Collector state should " 1989 "have changed"); 1990 break; 1991 case FinalMarking: 1992 { 1993 ReleaseForegroundGC x(this); 1994 1995 VM_CMS_Final_Remark final_remark_op(this); 1996 VMThread::execute(&final_remark_op); 1997 } 1998 assert(_foregroundGCShouldWait, "block post-condition"); 1999 break; 2000 case Sweeping: 2001 // final marking in checkpointRootsFinal has been completed 2002 sweep(); 2003 assert(_collectorState == Resizing, "Collector state change " 2004 "to Resizing must be done under the free_list_lock"); 2005 2006 case Resizing: { 2007 // Sweeping has been completed... 2008 // At this point the background collection has completed. 2009 // Don't move the call to compute_new_size() down 2010 // into code that might be executed if the background 2011 // collection was preempted. 2012 { 2013 ReleaseForegroundGC x(this); // unblock FG collection 2014 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2015 CMSTokenSync z(true); // not strictly needed. 2016 if (_collectorState == Resizing) { 2017 compute_new_size(); 2018 save_heap_summary(); 2019 _collectorState = Resetting; 2020 } else { 2021 assert(_collectorState == Idling, "The state should only change" 2022 " because the foreground collector has finished the collection"); 2023 } 2024 } 2025 break; 2026 } 2027 case Resetting: 2028 // CMS heap resizing has been completed 2029 reset(true); 2030 assert(_collectorState == Idling, "Collector state should " 2031 "have changed"); 2032 2033 MetaspaceGC::set_should_concurrent_collect(false); 2034 2035 stats().record_cms_end(); 2036 // Don't move the concurrent_phases_end() and compute_new_size() 2037 // calls to here because a preempted background collection 2038 // has it's state set to "Resetting". 2039 break; 2040 case Idling: 2041 default: 2042 ShouldNotReachHere(); 2043 break; 2044 } 2045 if (TraceCMSState) { 2046 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2047 Thread::current(), _collectorState); 2048 } 2049 assert(_foregroundGCShouldWait, "block post-condition"); 2050 } 2051 2052 // Should this be in gc_epilogue? 2053 collector_policy()->counters()->update_counters(); 2054 2055 { 2056 // Clear _foregroundGCShouldWait and, in the event that the 2057 // foreground collector is waiting, notify it, before 2058 // returning. 2059 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2060 _foregroundGCShouldWait = false; 2061 if (_foregroundGCIsActive) { 2062 CGC_lock->notify(); 2063 } 2064 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2065 "Possible deadlock"); 2066 } 2067 if (TraceCMSState) { 2068 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2069 " exiting collection CMS state %d", 2070 Thread::current(), _collectorState); 2071 } 2072 if (PrintGC && Verbose) { 2073 _cmsGen->print_heap_change(prev_used); 2074 } 2075 } 2076 2077 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2078 _cms_start_registered = true; 2079 _gc_timer_cm->register_gc_start(); 2080 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2081 } 2082 2083 void CMSCollector::register_gc_end() { 2084 if (_cms_start_registered) { 2085 report_heap_summary(GCWhen::AfterGC); 2086 2087 _gc_timer_cm->register_gc_end(); 2088 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2089 _cms_start_registered = false; 2090 } 2091 } 2092 2093 void CMSCollector::save_heap_summary() { 2094 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2095 _last_heap_summary = gch->create_heap_summary(); 2096 _last_metaspace_summary = gch->create_metaspace_summary(); 2097 } 2098 2099 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2100 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2101 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2102 } 2103 2104 bool CMSCollector::waitForForegroundGC() { 2105 bool res = false; 2106 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2107 "CMS thread should have CMS token"); 2108 // Block the foreground collector until the 2109 // background collectors decides whether to 2110 // yield. 2111 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2112 _foregroundGCShouldWait = true; 2113 if (_foregroundGCIsActive) { 2114 // The background collector yields to the 2115 // foreground collector and returns a value 2116 // indicating that it has yielded. The foreground 2117 // collector can proceed. 2118 res = true; 2119 _foregroundGCShouldWait = false; 2120 ConcurrentMarkSweepThread::clear_CMS_flag( 2121 ConcurrentMarkSweepThread::CMS_cms_has_token); 2122 ConcurrentMarkSweepThread::set_CMS_flag( 2123 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2124 // Get a possibly blocked foreground thread going 2125 CGC_lock->notify(); 2126 if (TraceCMSState) { 2127 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2128 Thread::current(), _collectorState); 2129 } 2130 while (_foregroundGCIsActive) { 2131 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2132 } 2133 ConcurrentMarkSweepThread::set_CMS_flag( 2134 ConcurrentMarkSweepThread::CMS_cms_has_token); 2135 ConcurrentMarkSweepThread::clear_CMS_flag( 2136 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2137 } 2138 if (TraceCMSState) { 2139 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2140 Thread::current(), _collectorState); 2141 } 2142 return res; 2143 } 2144 2145 // Because of the need to lock the free lists and other structures in 2146 // the collector, common to all the generations that the collector is 2147 // collecting, we need the gc_prologues of individual CMS generations 2148 // delegate to their collector. It may have been simpler had the 2149 // current infrastructure allowed one to call a prologue on a 2150 // collector. In the absence of that we have the generation's 2151 // prologue delegate to the collector, which delegates back 2152 // some "local" work to a worker method in the individual generations 2153 // that it's responsible for collecting, while itself doing any 2154 // work common to all generations it's responsible for. A similar 2155 // comment applies to the gc_epilogue()'s. 2156 // The role of the variable _between_prologue_and_epilogue is to 2157 // enforce the invocation protocol. 2158 void CMSCollector::gc_prologue(bool full) { 2159 // Call gc_prologue_work() for the CMSGen 2160 // we are responsible for. 2161 2162 // The following locking discipline assumes that we are only called 2163 // when the world is stopped. 2164 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2165 2166 // The CMSCollector prologue must call the gc_prologues for the 2167 // "generations" that it's responsible 2168 // for. 2169 2170 assert( Thread::current()->is_VM_thread() 2171 || ( CMSScavengeBeforeRemark 2172 && Thread::current()->is_ConcurrentGC_thread()), 2173 "Incorrect thread type for prologue execution"); 2174 2175 if (_between_prologue_and_epilogue) { 2176 // We have already been invoked; this is a gc_prologue delegation 2177 // from yet another CMS generation that we are responsible for, just 2178 // ignore it since all relevant work has already been done. 2179 return; 2180 } 2181 2182 // set a bit saying prologue has been called; cleared in epilogue 2183 _between_prologue_and_epilogue = true; 2184 // Claim locks for common data structures, then call gc_prologue_work() 2185 // for each CMSGen. 2186 2187 getFreelistLocks(); // gets free list locks on constituent spaces 2188 bitMapLock()->lock_without_safepoint_check(); 2189 2190 // Should call gc_prologue_work() for all cms gens we are responsible for 2191 bool duringMarking = _collectorState >= Marking 2192 && _collectorState < Sweeping; 2193 2194 // The young collections clear the modified oops state, which tells if 2195 // there are any modified oops in the class. The remark phase also needs 2196 // that information. Tell the young collection to save the union of all 2197 // modified klasses. 2198 if (duringMarking) { 2199 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2200 } 2201 2202 bool registerClosure = duringMarking; 2203 2204 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2205 &_modUnionClosurePar 2206 : &_modUnionClosure; 2207 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2208 2209 if (!full) { 2210 stats().record_gc0_begin(); 2211 } 2212 } 2213 2214 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2215 2216 _capacity_at_prologue = capacity(); 2217 _used_at_prologue = used(); 2218 2219 // Delegate to CMScollector which knows how to coordinate between 2220 // this and any other CMS generations that it is responsible for 2221 // collecting. 2222 collector()->gc_prologue(full); 2223 } 2224 2225 // This is a "private" interface for use by this generation's CMSCollector. 2226 // Not to be called directly by any other entity (for instance, 2227 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2228 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2229 bool registerClosure, ModUnionClosure* modUnionClosure) { 2230 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2231 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2232 "Should be NULL"); 2233 if (registerClosure) { 2234 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2235 } 2236 cmsSpace()->gc_prologue(); 2237 // Clear stat counters 2238 NOT_PRODUCT( 2239 assert(_numObjectsPromoted == 0, "check"); 2240 assert(_numWordsPromoted == 0, "check"); 2241 if (Verbose && PrintGC) { 2242 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2243 SIZE_FORMAT" bytes concurrently", 2244 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2245 } 2246 _numObjectsAllocated = 0; 2247 _numWordsAllocated = 0; 2248 ) 2249 } 2250 2251 void CMSCollector::gc_epilogue(bool full) { 2252 // The following locking discipline assumes that we are only called 2253 // when the world is stopped. 2254 assert(SafepointSynchronize::is_at_safepoint(), 2255 "world is stopped assumption"); 2256 2257 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2258 // if linear allocation blocks need to be appropriately marked to allow the 2259 // the blocks to be parsable. We also check here whether we need to nudge the 2260 // CMS collector thread to start a new cycle (if it's not already active). 2261 assert( Thread::current()->is_VM_thread() 2262 || ( CMSScavengeBeforeRemark 2263 && Thread::current()->is_ConcurrentGC_thread()), 2264 "Incorrect thread type for epilogue execution"); 2265 2266 if (!_between_prologue_and_epilogue) { 2267 // We have already been invoked; this is a gc_epilogue delegation 2268 // from yet another CMS generation that we are responsible for, just 2269 // ignore it since all relevant work has already been done. 2270 return; 2271 } 2272 assert(haveFreelistLocks(), "must have freelist locks"); 2273 assert_lock_strong(bitMapLock()); 2274 2275 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2276 2277 _cmsGen->gc_epilogue_work(full); 2278 2279 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2280 // in case sampling was not already enabled, enable it 2281 _start_sampling = true; 2282 } 2283 // reset _eden_chunk_array so sampling starts afresh 2284 _eden_chunk_index = 0; 2285 2286 size_t cms_used = _cmsGen->cmsSpace()->used(); 2287 2288 // update performance counters - this uses a special version of 2289 // update_counters() that allows the utilization to be passed as a 2290 // parameter, avoiding multiple calls to used(). 2291 // 2292 _cmsGen->update_counters(cms_used); 2293 2294 bitMapLock()->unlock(); 2295 releaseFreelistLocks(); 2296 2297 if (!CleanChunkPoolAsync) { 2298 Chunk::clean_chunk_pool(); 2299 } 2300 2301 set_did_compact(false); 2302 _between_prologue_and_epilogue = false; // ready for next cycle 2303 } 2304 2305 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2306 collector()->gc_epilogue(full); 2307 2308 // Also reset promotion tracking in par gc thread states. 2309 for (uint i = 0; i < ParallelGCThreads; i++) { 2310 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2311 } 2312 } 2313 2314 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2315 assert(!incremental_collection_failed(), "Should have been cleared"); 2316 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2317 cmsSpace()->gc_epilogue(); 2318 // Print stat counters 2319 NOT_PRODUCT( 2320 assert(_numObjectsAllocated == 0, "check"); 2321 assert(_numWordsAllocated == 0, "check"); 2322 if (Verbose && PrintGC) { 2323 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2324 SIZE_FORMAT" bytes", 2325 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2326 } 2327 _numObjectsPromoted = 0; 2328 _numWordsPromoted = 0; 2329 ) 2330 2331 if (PrintGC && Verbose) { 2332 // Call down the chain in contiguous_available needs the freelistLock 2333 // so print this out before releasing the freeListLock. 2334 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2335 contiguous_available()); 2336 } 2337 } 2338 2339 #ifndef PRODUCT 2340 bool CMSCollector::have_cms_token() { 2341 Thread* thr = Thread::current(); 2342 if (thr->is_VM_thread()) { 2343 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2344 } else if (thr->is_ConcurrentGC_thread()) { 2345 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2346 } else if (thr->is_GC_task_thread()) { 2347 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2348 ParGCRareEvent_lock->owned_by_self(); 2349 } 2350 return false; 2351 } 2352 #endif 2353 2354 // Check reachability of the given heap address in CMS generation, 2355 // treating all other generations as roots. 2356 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2357 // We could "guarantee" below, rather than assert, but I'll 2358 // leave these as "asserts" so that an adventurous debugger 2359 // could try this in the product build provided some subset of 2360 // the conditions were met, provided they were interested in the 2361 // results and knew that the computation below wouldn't interfere 2362 // with other concurrent computations mutating the structures 2363 // being read or written. 2364 assert(SafepointSynchronize::is_at_safepoint(), 2365 "Else mutations in object graph will make answer suspect"); 2366 assert(have_cms_token(), "Should hold cms token"); 2367 assert(haveFreelistLocks(), "must hold free list locks"); 2368 assert_lock_strong(bitMapLock()); 2369 2370 // Clear the marking bit map array before starting, but, just 2371 // for kicks, first report if the given address is already marked 2372 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2373 _markBitMap.isMarked(addr) ? "" : " not"); 2374 2375 if (verify_after_remark()) { 2376 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2377 bool result = verification_mark_bm()->isMarked(addr); 2378 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2379 result ? "IS" : "is NOT"); 2380 return result; 2381 } else { 2382 gclog_or_tty->print_cr("Could not compute result"); 2383 return false; 2384 } 2385 } 2386 2387 2388 void 2389 CMSCollector::print_on_error(outputStream* st) { 2390 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2391 if (collector != NULL) { 2392 CMSBitMap* bitmap = &collector->_markBitMap; 2393 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2394 bitmap->print_on_error(st, " Bits: "); 2395 2396 st->cr(); 2397 2398 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2399 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2400 mut_bitmap->print_on_error(st, " Bits: "); 2401 } 2402 } 2403 2404 //////////////////////////////////////////////////////// 2405 // CMS Verification Support 2406 //////////////////////////////////////////////////////// 2407 // Following the remark phase, the following invariant 2408 // should hold -- each object in the CMS heap which is 2409 // marked in markBitMap() should be marked in the verification_mark_bm(). 2410 2411 class VerifyMarkedClosure: public BitMapClosure { 2412 CMSBitMap* _marks; 2413 bool _failed; 2414 2415 public: 2416 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2417 2418 bool do_bit(size_t offset) { 2419 HeapWord* addr = _marks->offsetToHeapWord(offset); 2420 if (!_marks->isMarked(addr)) { 2421 oop(addr)->print_on(gclog_or_tty); 2422 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2423 _failed = true; 2424 } 2425 return true; 2426 } 2427 2428 bool failed() { return _failed; } 2429 }; 2430 2431 bool CMSCollector::verify_after_remark(bool silent) { 2432 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2433 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2434 static bool init = false; 2435 2436 assert(SafepointSynchronize::is_at_safepoint(), 2437 "Else mutations in object graph will make answer suspect"); 2438 assert(have_cms_token(), 2439 "Else there may be mutual interference in use of " 2440 " verification data structures"); 2441 assert(_collectorState > Marking && _collectorState <= Sweeping, 2442 "Else marking info checked here may be obsolete"); 2443 assert(haveFreelistLocks(), "must hold free list locks"); 2444 assert_lock_strong(bitMapLock()); 2445 2446 2447 // Allocate marking bit map if not already allocated 2448 if (!init) { // first time 2449 if (!verification_mark_bm()->allocate(_span)) { 2450 return false; 2451 } 2452 init = true; 2453 } 2454 2455 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2456 2457 // Turn off refs discovery -- so we will be tracing through refs. 2458 // This is as intended, because by this time 2459 // GC must already have cleared any refs that need to be cleared, 2460 // and traced those that need to be marked; moreover, 2461 // the marking done here is not going to interfere in any 2462 // way with the marking information used by GC. 2463 NoRefDiscovery no_discovery(ref_processor()); 2464 2465 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2466 2467 // Clear any marks from a previous round 2468 verification_mark_bm()->clear_all(); 2469 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2470 verify_work_stacks_empty(); 2471 2472 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2473 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2474 // Update the saved marks which may affect the root scans. 2475 gch->save_marks(); 2476 2477 if (CMSRemarkVerifyVariant == 1) { 2478 // In this first variant of verification, we complete 2479 // all marking, then check if the new marks-vector is 2480 // a subset of the CMS marks-vector. 2481 verify_after_remark_work_1(); 2482 } else if (CMSRemarkVerifyVariant == 2) { 2483 // In this second variant of verification, we flag an error 2484 // (i.e. an object reachable in the new marks-vector not reachable 2485 // in the CMS marks-vector) immediately, also indicating the 2486 // identify of an object (A) that references the unmarked object (B) -- 2487 // presumably, a mutation to A failed to be picked up by preclean/remark? 2488 verify_after_remark_work_2(); 2489 } else { 2490 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2491 CMSRemarkVerifyVariant); 2492 } 2493 if (!silent) gclog_or_tty->print(" done] "); 2494 return true; 2495 } 2496 2497 void CMSCollector::verify_after_remark_work_1() { 2498 ResourceMark rm; 2499 HandleMark hm; 2500 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2501 2502 // Get a clear set of claim bits for the roots processing to work with. 2503 ClassLoaderDataGraph::clear_claimed_marks(); 2504 2505 // Mark from roots one level into CMS 2506 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2507 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2508 2509 gch->gen_process_roots(_cmsGen->level(), 2510 true, // younger gens are roots 2511 true, // activate StrongRootsScope 2512 SharedHeap::ScanningOption(roots_scanning_options()), 2513 should_unload_classes(), 2514 ¬Older, 2515 NULL, 2516 NULL); // SSS: Provide correct closure 2517 2518 // Now mark from the roots 2519 MarkFromRootsClosure markFromRootsClosure(this, _span, 2520 verification_mark_bm(), verification_mark_stack(), 2521 false /* don't yield */, true /* verifying */); 2522 assert(_restart_addr == NULL, "Expected pre-condition"); 2523 verification_mark_bm()->iterate(&markFromRootsClosure); 2524 while (_restart_addr != NULL) { 2525 // Deal with stack overflow: by restarting at the indicated 2526 // address. 2527 HeapWord* ra = _restart_addr; 2528 markFromRootsClosure.reset(ra); 2529 _restart_addr = NULL; 2530 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2531 } 2532 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2533 verify_work_stacks_empty(); 2534 2535 // Marking completed -- now verify that each bit marked in 2536 // verification_mark_bm() is also marked in markBitMap(); flag all 2537 // errors by printing corresponding objects. 2538 VerifyMarkedClosure vcl(markBitMap()); 2539 verification_mark_bm()->iterate(&vcl); 2540 if (vcl.failed()) { 2541 gclog_or_tty->print("Verification failed"); 2542 Universe::heap()->print_on(gclog_or_tty); 2543 fatal("CMS: failed marking verification after remark"); 2544 } 2545 } 2546 2547 class VerifyKlassOopsKlassClosure : public KlassClosure { 2548 class VerifyKlassOopsClosure : public OopClosure { 2549 CMSBitMap* _bitmap; 2550 public: 2551 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2552 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2553 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2554 } _oop_closure; 2555 public: 2556 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2557 void do_klass(Klass* k) { 2558 k->oops_do(&_oop_closure); 2559 } 2560 }; 2561 2562 void CMSCollector::verify_after_remark_work_2() { 2563 ResourceMark rm; 2564 HandleMark hm; 2565 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2566 2567 // Get a clear set of claim bits for the roots processing to work with. 2568 ClassLoaderDataGraph::clear_claimed_marks(); 2569 2570 // Mark from roots one level into CMS 2571 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2572 markBitMap()); 2573 CLDToOopClosure cld_closure(¬Older, true); 2574 2575 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2576 2577 gch->gen_process_roots(_cmsGen->level(), 2578 true, // younger gens are roots 2579 true, // activate StrongRootsScope 2580 SharedHeap::ScanningOption(roots_scanning_options()), 2581 should_unload_classes(), 2582 ¬Older, 2583 NULL, 2584 &cld_closure); 2585 2586 // Now mark from the roots 2587 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2588 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2589 assert(_restart_addr == NULL, "Expected pre-condition"); 2590 verification_mark_bm()->iterate(&markFromRootsClosure); 2591 while (_restart_addr != NULL) { 2592 // Deal with stack overflow: by restarting at the indicated 2593 // address. 2594 HeapWord* ra = _restart_addr; 2595 markFromRootsClosure.reset(ra); 2596 _restart_addr = NULL; 2597 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2598 } 2599 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2600 verify_work_stacks_empty(); 2601 2602 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2603 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2604 2605 // Marking completed -- now verify that each bit marked in 2606 // verification_mark_bm() is also marked in markBitMap(); flag all 2607 // errors by printing corresponding objects. 2608 VerifyMarkedClosure vcl(markBitMap()); 2609 verification_mark_bm()->iterate(&vcl); 2610 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2611 } 2612 2613 void ConcurrentMarkSweepGeneration::save_marks() { 2614 // delegate to CMS space 2615 cmsSpace()->save_marks(); 2616 for (uint i = 0; i < ParallelGCThreads; i++) { 2617 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2618 } 2619 } 2620 2621 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2622 return cmsSpace()->no_allocs_since_save_marks(); 2623 } 2624 2625 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2626 \ 2627 void ConcurrentMarkSweepGeneration:: \ 2628 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2629 cl->set_generation(this); \ 2630 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2631 cl->reset_generation(); \ 2632 save_marks(); \ 2633 } 2634 2635 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2636 2637 void 2638 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2639 cl->set_generation(this); 2640 younger_refs_in_space_iterate(_cmsSpace, cl); 2641 cl->reset_generation(); 2642 } 2643 2644 void 2645 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2646 if (freelistLock()->owned_by_self()) { 2647 Generation::oop_iterate(cl); 2648 } else { 2649 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2650 Generation::oop_iterate(cl); 2651 } 2652 } 2653 2654 void 2655 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2656 if (freelistLock()->owned_by_self()) { 2657 Generation::object_iterate(cl); 2658 } else { 2659 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2660 Generation::object_iterate(cl); 2661 } 2662 } 2663 2664 void 2665 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2666 if (freelistLock()->owned_by_self()) { 2667 Generation::safe_object_iterate(cl); 2668 } else { 2669 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2670 Generation::safe_object_iterate(cl); 2671 } 2672 } 2673 2674 void 2675 ConcurrentMarkSweepGeneration::post_compact() { 2676 } 2677 2678 void 2679 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2680 // Fix the linear allocation blocks to look like free blocks. 2681 2682 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2683 // are not called when the heap is verified during universe initialization and 2684 // at vm shutdown. 2685 if (freelistLock()->owned_by_self()) { 2686 cmsSpace()->prepare_for_verify(); 2687 } else { 2688 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2689 cmsSpace()->prepare_for_verify(); 2690 } 2691 } 2692 2693 void 2694 ConcurrentMarkSweepGeneration::verify() { 2695 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2696 // are not called when the heap is verified during universe initialization and 2697 // at vm shutdown. 2698 if (freelistLock()->owned_by_self()) { 2699 cmsSpace()->verify(); 2700 } else { 2701 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2702 cmsSpace()->verify(); 2703 } 2704 } 2705 2706 void CMSCollector::verify() { 2707 _cmsGen->verify(); 2708 } 2709 2710 #ifndef PRODUCT 2711 bool CMSCollector::overflow_list_is_empty() const { 2712 assert(_num_par_pushes >= 0, "Inconsistency"); 2713 if (_overflow_list == NULL) { 2714 assert(_num_par_pushes == 0, "Inconsistency"); 2715 } 2716 return _overflow_list == NULL; 2717 } 2718 2719 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2720 // merely consolidate assertion checks that appear to occur together frequently. 2721 void CMSCollector::verify_work_stacks_empty() const { 2722 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2723 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2724 } 2725 2726 void CMSCollector::verify_overflow_empty() const { 2727 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2728 assert(no_preserved_marks(), "No preserved marks"); 2729 } 2730 #endif // PRODUCT 2731 2732 // Decide if we want to enable class unloading as part of the 2733 // ensuing concurrent GC cycle. We will collect and 2734 // unload classes if it's the case that: 2735 // (1) an explicit gc request has been made and the flag 2736 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2737 // (2) (a) class unloading is enabled at the command line, and 2738 // (b) old gen is getting really full 2739 // NOTE: Provided there is no change in the state of the heap between 2740 // calls to this method, it should have idempotent results. Moreover, 2741 // its results should be monotonically increasing (i.e. going from 0 to 1, 2742 // but not 1 to 0) between successive calls between which the heap was 2743 // not collected. For the implementation below, it must thus rely on 2744 // the property that concurrent_cycles_since_last_unload() 2745 // will not decrease unless a collection cycle happened and that 2746 // _cmsGen->is_too_full() are 2747 // themselves also monotonic in that sense. See check_monotonicity() 2748 // below. 2749 void CMSCollector::update_should_unload_classes() { 2750 _should_unload_classes = false; 2751 // Condition 1 above 2752 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2753 _should_unload_classes = true; 2754 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2755 // Disjuncts 2.b.(i,ii,iii) above 2756 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2757 CMSClassUnloadingMaxInterval) 2758 || _cmsGen->is_too_full(); 2759 } 2760 } 2761 2762 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2763 bool res = should_concurrent_collect(); 2764 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2765 return res; 2766 } 2767 2768 void CMSCollector::setup_cms_unloading_and_verification_state() { 2769 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2770 || VerifyBeforeExit; 2771 const int rso = SharedHeap::SO_AllCodeCache; 2772 2773 // We set the proper root for this CMS cycle here. 2774 if (should_unload_classes()) { // Should unload classes this cycle 2775 remove_root_scanning_option(rso); // Shrink the root set appropriately 2776 set_verifying(should_verify); // Set verification state for this cycle 2777 return; // Nothing else needs to be done at this time 2778 } 2779 2780 // Not unloading classes this cycle 2781 assert(!should_unload_classes(), "Inconsistency!"); 2782 2783 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2784 // Include symbols, strings and code cache elements to prevent their resurrection. 2785 add_root_scanning_option(rso); 2786 set_verifying(true); 2787 } else if (verifying() && !should_verify) { 2788 // We were verifying, but some verification flags got disabled. 2789 set_verifying(false); 2790 // Exclude symbols, strings and code cache elements from root scanning to 2791 // reduce IM and RM pauses. 2792 remove_root_scanning_option(rso); 2793 } 2794 } 2795 2796 2797 #ifndef PRODUCT 2798 HeapWord* CMSCollector::block_start(const void* p) const { 2799 const HeapWord* addr = (HeapWord*)p; 2800 if (_span.contains(p)) { 2801 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2802 return _cmsGen->cmsSpace()->block_start(p); 2803 } 2804 } 2805 return NULL; 2806 } 2807 #endif 2808 2809 HeapWord* 2810 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2811 bool tlab, 2812 bool parallel) { 2813 CMSSynchronousYieldRequest yr; 2814 assert(!tlab, "Can't deal with TLAB allocation"); 2815 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2816 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 2817 CMSExpansionCause::_satisfy_allocation); 2818 if (GCExpandToAllocateDelayMillis > 0) { 2819 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2820 } 2821 return have_lock_and_allocate(word_size, tlab); 2822 } 2823 2824 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 2825 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 2826 // to CardGeneration and share it... 2827 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 2828 return CardGeneration::expand(bytes, expand_bytes); 2829 } 2830 2831 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 2832 CMSExpansionCause::Cause cause) 2833 { 2834 2835 bool success = expand(bytes, expand_bytes); 2836 2837 // remember why we expanded; this information is used 2838 // by shouldConcurrentCollect() when making decisions on whether to start 2839 // a new CMS cycle. 2840 if (success) { 2841 set_expansion_cause(cause); 2842 if (PrintGCDetails && Verbose) { 2843 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2844 CMSExpansionCause::to_string(cause)); 2845 } 2846 } 2847 } 2848 2849 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2850 HeapWord* res = NULL; 2851 MutexLocker x(ParGCRareEvent_lock); 2852 while (true) { 2853 // Expansion by some other thread might make alloc OK now: 2854 res = ps->lab.alloc(word_sz); 2855 if (res != NULL) return res; 2856 // If there's not enough expansion space available, give up. 2857 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2858 return NULL; 2859 } 2860 // Otherwise, we try expansion. 2861 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 2862 CMSExpansionCause::_allocate_par_lab); 2863 // Now go around the loop and try alloc again; 2864 // A competing par_promote might beat us to the expansion space, 2865 // so we may go around the loop again if promotion fails again. 2866 if (GCExpandToAllocateDelayMillis > 0) { 2867 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2868 } 2869 } 2870 } 2871 2872 2873 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2874 PromotionInfo* promo) { 2875 MutexLocker x(ParGCRareEvent_lock); 2876 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2877 while (true) { 2878 // Expansion by some other thread might make alloc OK now: 2879 if (promo->ensure_spooling_space()) { 2880 assert(promo->has_spooling_space(), 2881 "Post-condition of successful ensure_spooling_space()"); 2882 return true; 2883 } 2884 // If there's not enough expansion space available, give up. 2885 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2886 return false; 2887 } 2888 // Otherwise, we try expansion. 2889 expand(refill_size_bytes, MinHeapDeltaBytes, 2890 CMSExpansionCause::_allocate_par_spooling_space); 2891 // Now go around the loop and try alloc again; 2892 // A competing allocation might beat us to the expansion space, 2893 // so we may go around the loop again if allocation fails again. 2894 if (GCExpandToAllocateDelayMillis > 0) { 2895 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2896 } 2897 } 2898 } 2899 2900 2901 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 2902 assert_locked_or_safepoint(ExpandHeap_lock); 2903 // Shrink committed space 2904 _virtual_space.shrink_by(bytes); 2905 // Shrink space; this also shrinks the space's BOT 2906 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 2907 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 2908 // Shrink the shared block offset array 2909 _bts->resize(new_word_size); 2910 MemRegion mr(_cmsSpace->bottom(), new_word_size); 2911 // Shrink the card table 2912 Universe::heap()->barrier_set()->resize_covered_region(mr); 2913 2914 if (Verbose && PrintGC) { 2915 size_t new_mem_size = _virtual_space.committed_size(); 2916 size_t old_mem_size = new_mem_size + bytes; 2917 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 2918 name(), old_mem_size/K, new_mem_size/K); 2919 } 2920 } 2921 2922 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2923 assert_locked_or_safepoint(Heap_lock); 2924 size_t size = ReservedSpace::page_align_size_down(bytes); 2925 // Only shrink if a compaction was done so that all the free space 2926 // in the generation is in a contiguous block at the end. 2927 if (size > 0 && did_compact()) { 2928 shrink_by(size); 2929 } 2930 } 2931 2932 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 2933 assert_locked_or_safepoint(Heap_lock); 2934 bool result = _virtual_space.expand_by(bytes); 2935 if (result) { 2936 size_t new_word_size = 2937 heap_word_size(_virtual_space.committed_size()); 2938 MemRegion mr(_cmsSpace->bottom(), new_word_size); 2939 _bts->resize(new_word_size); // resize the block offset shared array 2940 Universe::heap()->barrier_set()->resize_covered_region(mr); 2941 // Hmmmm... why doesn't CFLS::set_end verify locking? 2942 // This is quite ugly; FIX ME XXX 2943 _cmsSpace->assert_locked(freelistLock()); 2944 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 2945 2946 // update the space and generation capacity counters 2947 if (UsePerfData) { 2948 _space_counters->update_capacity(); 2949 _gen_counters->update_all(); 2950 } 2951 2952 if (Verbose && PrintGC) { 2953 size_t new_mem_size = _virtual_space.committed_size(); 2954 size_t old_mem_size = new_mem_size - bytes; 2955 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 2956 name(), old_mem_size/K, bytes/K, new_mem_size/K); 2957 } 2958 } 2959 return result; 2960 } 2961 2962 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 2963 assert_locked_or_safepoint(Heap_lock); 2964 bool success = true; 2965 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 2966 if (remaining_bytes > 0) { 2967 success = grow_by(remaining_bytes); 2968 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 2969 } 2970 return success; 2971 } 2972 2973 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2974 assert_locked_or_safepoint(Heap_lock); 2975 assert_lock_strong(freelistLock()); 2976 if (PrintGCDetails && Verbose) { 2977 warning("Shrinking of CMS not yet implemented"); 2978 } 2979 return; 2980 } 2981 2982 2983 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2984 // phases. 2985 class CMSPhaseAccounting: public StackObj { 2986 public: 2987 CMSPhaseAccounting(CMSCollector *collector, 2988 const char *phase, 2989 const GCId gc_id, 2990 bool print_cr = true); 2991 ~CMSPhaseAccounting(); 2992 2993 private: 2994 CMSCollector *_collector; 2995 const char *_phase; 2996 elapsedTimer _wallclock; 2997 bool _print_cr; 2998 const GCId _gc_id; 2999 3000 public: 3001 // Not MT-safe; so do not pass around these StackObj's 3002 // where they may be accessed by other threads. 3003 jlong wallclock_millis() { 3004 assert(_wallclock.is_active(), "Wall clock should not stop"); 3005 _wallclock.stop(); // to record time 3006 jlong ret = _wallclock.milliseconds(); 3007 _wallclock.start(); // restart 3008 return ret; 3009 } 3010 }; 3011 3012 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3013 const char *phase, 3014 const GCId gc_id, 3015 bool print_cr) : 3016 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 3017 3018 if (PrintCMSStatistics != 0) { 3019 _collector->resetYields(); 3020 } 3021 if (PrintGCDetails) { 3022 gclog_or_tty->gclog_stamp(_gc_id); 3023 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3024 _collector->cmsGen()->short_name(), _phase); 3025 } 3026 _collector->resetTimer(); 3027 _wallclock.start(); 3028 _collector->startTimer(); 3029 } 3030 3031 CMSPhaseAccounting::~CMSPhaseAccounting() { 3032 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3033 _collector->stopTimer(); 3034 _wallclock.stop(); 3035 if (PrintGCDetails) { 3036 gclog_or_tty->gclog_stamp(_gc_id); 3037 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3038 _collector->cmsGen()->short_name(), 3039 _phase, _collector->timerValue(), _wallclock.seconds()); 3040 if (_print_cr) { 3041 gclog_or_tty->cr(); 3042 } 3043 if (PrintCMSStatistics != 0) { 3044 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3045 _collector->yields()); 3046 } 3047 } 3048 } 3049 3050 // CMS work 3051 3052 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3053 class CMSParMarkTask : public AbstractGangTask { 3054 protected: 3055 CMSCollector* _collector; 3056 int _n_workers; 3057 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3058 AbstractGangTask(name), 3059 _collector(collector), 3060 _n_workers(n_workers) {} 3061 // Work method in support of parallel rescan ... of young gen spaces 3062 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3063 ContiguousSpace* space, 3064 HeapWord** chunk_array, size_t chunk_top); 3065 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3066 }; 3067 3068 // Parallel initial mark task 3069 class CMSParInitialMarkTask: public CMSParMarkTask { 3070 public: 3071 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3072 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3073 collector, n_workers) {} 3074 void work(uint worker_id); 3075 }; 3076 3077 // Checkpoint the roots into this generation from outside 3078 // this generation. [Note this initial checkpoint need only 3079 // be approximate -- we'll do a catch up phase subsequently.] 3080 void CMSCollector::checkpointRootsInitial() { 3081 assert(_collectorState == InitialMarking, "Wrong collector state"); 3082 check_correct_thread_executing(); 3083 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3084 3085 save_heap_summary(); 3086 report_heap_summary(GCWhen::BeforeGC); 3087 3088 ReferenceProcessor* rp = ref_processor(); 3089 SpecializationStats::clear(); 3090 assert(_restart_addr == NULL, "Control point invariant"); 3091 { 3092 // acquire locks for subsequent manipulations 3093 MutexLockerEx x(bitMapLock(), 3094 Mutex::_no_safepoint_check_flag); 3095 checkpointRootsInitialWork(); 3096 // enable ("weak") refs discovery 3097 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3098 _collectorState = Marking; 3099 } 3100 SpecializationStats::print(); 3101 } 3102 3103 void CMSCollector::checkpointRootsInitialWork() { 3104 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3105 assert(_collectorState == InitialMarking, "just checking"); 3106 3107 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3108 // precede our marking with a collection of all 3109 // younger generations to keep floating garbage to a minimum. 3110 // XXX: we won't do this for now -- it's an optimization to be done later. 3111 3112 // already have locks 3113 assert_lock_strong(bitMapLock()); 3114 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3115 3116 // Setup the verification and class unloading state for this 3117 // CMS collection cycle. 3118 setup_cms_unloading_and_verification_state(); 3119 3120 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3121 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3122 3123 // Reset all the PLAB chunk arrays if necessary. 3124 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3125 reset_survivor_plab_arrays(); 3126 } 3127 3128 ResourceMark rm; 3129 HandleMark hm; 3130 3131 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3132 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3133 3134 verify_work_stacks_empty(); 3135 verify_overflow_empty(); 3136 3137 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3138 // Update the saved marks which may affect the root scans. 3139 gch->save_marks(); 3140 3141 // weak reference processing has not started yet. 3142 ref_processor()->set_enqueuing_is_done(false); 3143 3144 // Need to remember all newly created CLDs, 3145 // so that we can guarantee that the remark finds them. 3146 ClassLoaderDataGraph::remember_new_clds(true); 3147 3148 // Whenever a CLD is found, it will be claimed before proceeding to mark 3149 // the klasses. The claimed marks need to be cleared before marking starts. 3150 ClassLoaderDataGraph::clear_claimed_marks(); 3151 3152 if (CMSPrintEdenSurvivorChunks) { 3153 print_eden_and_survivor_chunk_arrays(); 3154 } 3155 3156 { 3157 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3158 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3159 // The parallel version. 3160 FlexibleWorkGang* workers = gch->workers(); 3161 assert(workers != NULL, "Need parallel worker threads."); 3162 int n_workers = workers->active_workers(); 3163 CMSParInitialMarkTask tsk(this, n_workers); 3164 gch->set_par_threads(n_workers); 3165 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3166 if (n_workers > 1) { 3167 GenCollectedHeap::StrongRootsScope srs(gch); 3168 workers->run_task(&tsk); 3169 } else { 3170 GenCollectedHeap::StrongRootsScope srs(gch); 3171 tsk.work(0); 3172 } 3173 gch->set_par_threads(0); 3174 } else { 3175 // The serial version. 3176 CLDToOopClosure cld_closure(¬Older, true); 3177 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3178 gch->gen_process_roots(_cmsGen->level(), 3179 true, // younger gens are roots 3180 true, // activate StrongRootsScope 3181 SharedHeap::ScanningOption(roots_scanning_options()), 3182 should_unload_classes(), 3183 ¬Older, 3184 NULL, 3185 &cld_closure); 3186 } 3187 } 3188 3189 // Clear mod-union table; it will be dirtied in the prologue of 3190 // CMS generation per each younger generation collection. 3191 3192 assert(_modUnionTable.isAllClear(), 3193 "Was cleared in most recent final checkpoint phase" 3194 " or no bits are set in the gc_prologue before the start of the next " 3195 "subsequent marking phase."); 3196 3197 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3198 3199 // Save the end of the used_region of the constituent generations 3200 // to be used to limit the extent of sweep in each generation. 3201 save_sweep_limits(); 3202 verify_overflow_empty(); 3203 } 3204 3205 bool CMSCollector::markFromRoots() { 3206 // we might be tempted to assert that: 3207 // assert(!SafepointSynchronize::is_at_safepoint(), 3208 // "inconsistent argument?"); 3209 // However that wouldn't be right, because it's possible that 3210 // a safepoint is indeed in progress as a younger generation 3211 // stop-the-world GC happens even as we mark in this generation. 3212 assert(_collectorState == Marking, "inconsistent state?"); 3213 check_correct_thread_executing(); 3214 verify_overflow_empty(); 3215 3216 // Weak ref discovery note: We may be discovering weak 3217 // refs in this generation concurrent (but interleaved) with 3218 // weak ref discovery by a younger generation collector. 3219 3220 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3221 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3222 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3223 bool res = markFromRootsWork(); 3224 if (res) { 3225 _collectorState = Precleaning; 3226 } else { // We failed and a foreground collection wants to take over 3227 assert(_foregroundGCIsActive, "internal state inconsistency"); 3228 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3229 if (PrintGCDetails) { 3230 gclog_or_tty->print_cr("bailing out to foreground collection"); 3231 } 3232 } 3233 verify_overflow_empty(); 3234 return res; 3235 } 3236 3237 bool CMSCollector::markFromRootsWork() { 3238 // iterate over marked bits in bit map, doing a full scan and mark 3239 // from these roots using the following algorithm: 3240 // . if oop is to the right of the current scan pointer, 3241 // mark corresponding bit (we'll process it later) 3242 // . else (oop is to left of current scan pointer) 3243 // push oop on marking stack 3244 // . drain the marking stack 3245 3246 // Note that when we do a marking step we need to hold the 3247 // bit map lock -- recall that direct allocation (by mutators) 3248 // and promotion (by younger generation collectors) is also 3249 // marking the bit map. [the so-called allocate live policy.] 3250 // Because the implementation of bit map marking is not 3251 // robust wrt simultaneous marking of bits in the same word, 3252 // we need to make sure that there is no such interference 3253 // between concurrent such updates. 3254 3255 // already have locks 3256 assert_lock_strong(bitMapLock()); 3257 3258 verify_work_stacks_empty(); 3259 verify_overflow_empty(); 3260 bool result = false; 3261 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3262 result = do_marking_mt(); 3263 } else { 3264 result = do_marking_st(); 3265 } 3266 return result; 3267 } 3268 3269 // Forward decl 3270 class CMSConcMarkingTask; 3271 3272 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3273 CMSCollector* _collector; 3274 CMSConcMarkingTask* _task; 3275 public: 3276 virtual void yield(); 3277 3278 // "n_threads" is the number of threads to be terminated. 3279 // "queue_set" is a set of work queues of other threads. 3280 // "collector" is the CMS collector associated with this task terminator. 3281 // "yield" indicates whether we need the gang as a whole to yield. 3282 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3283 ParallelTaskTerminator(n_threads, queue_set), 3284 _collector(collector) { } 3285 3286 void set_task(CMSConcMarkingTask* task) { 3287 _task = task; 3288 } 3289 }; 3290 3291 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3292 CMSConcMarkingTask* _task; 3293 public: 3294 bool should_exit_termination(); 3295 void set_task(CMSConcMarkingTask* task) { 3296 _task = task; 3297 } 3298 }; 3299 3300 // MT Concurrent Marking Task 3301 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3302 CMSCollector* _collector; 3303 int _n_workers; // requested/desired # workers 3304 bool _result; 3305 CompactibleFreeListSpace* _cms_space; 3306 char _pad_front[64]; // padding to ... 3307 HeapWord* _global_finger; // ... avoid sharing cache line 3308 char _pad_back[64]; 3309 HeapWord* _restart_addr; 3310 3311 // Exposed here for yielding support 3312 Mutex* const _bit_map_lock; 3313 3314 // The per thread work queues, available here for stealing 3315 OopTaskQueueSet* _task_queues; 3316 3317 // Termination (and yielding) support 3318 CMSConcMarkingTerminator _term; 3319 CMSConcMarkingTerminatorTerminator _term_term; 3320 3321 public: 3322 CMSConcMarkingTask(CMSCollector* collector, 3323 CompactibleFreeListSpace* cms_space, 3324 YieldingFlexibleWorkGang* workers, 3325 OopTaskQueueSet* task_queues): 3326 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3327 _collector(collector), 3328 _cms_space(cms_space), 3329 _n_workers(0), _result(true), 3330 _task_queues(task_queues), 3331 _term(_n_workers, task_queues, _collector), 3332 _bit_map_lock(collector->bitMapLock()) 3333 { 3334 _requested_size = _n_workers; 3335 _term.set_task(this); 3336 _term_term.set_task(this); 3337 _restart_addr = _global_finger = _cms_space->bottom(); 3338 } 3339 3340 3341 OopTaskQueueSet* task_queues() { return _task_queues; } 3342 3343 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3344 3345 HeapWord** global_finger_addr() { return &_global_finger; } 3346 3347 CMSConcMarkingTerminator* terminator() { return &_term; } 3348 3349 virtual void set_for_termination(int active_workers) { 3350 terminator()->reset_for_reuse(active_workers); 3351 } 3352 3353 void work(uint worker_id); 3354 bool should_yield() { 3355 return ConcurrentMarkSweepThread::should_yield() 3356 && !_collector->foregroundGCIsActive(); 3357 } 3358 3359 virtual void coordinator_yield(); // stuff done by coordinator 3360 bool result() { return _result; } 3361 3362 void reset(HeapWord* ra) { 3363 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3364 _restart_addr = _global_finger = ra; 3365 _term.reset_for_reuse(); 3366 } 3367 3368 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3369 OopTaskQueue* work_q); 3370 3371 private: 3372 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3373 void do_work_steal(int i); 3374 void bump_global_finger(HeapWord* f); 3375 }; 3376 3377 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3378 assert(_task != NULL, "Error"); 3379 return _task->yielding(); 3380 // Note that we do not need the disjunct || _task->should_yield() above 3381 // because we want terminating threads to yield only if the task 3382 // is already in the midst of yielding, which happens only after at least one 3383 // thread has yielded. 3384 } 3385 3386 void CMSConcMarkingTerminator::yield() { 3387 if (_task->should_yield()) { 3388 _task->yield(); 3389 } else { 3390 ParallelTaskTerminator::yield(); 3391 } 3392 } 3393 3394 //////////////////////////////////////////////////////////////// 3395 // Concurrent Marking Algorithm Sketch 3396 //////////////////////////////////////////////////////////////// 3397 // Until all tasks exhausted (both spaces): 3398 // -- claim next available chunk 3399 // -- bump global finger via CAS 3400 // -- find first object that starts in this chunk 3401 // and start scanning bitmap from that position 3402 // -- scan marked objects for oops 3403 // -- CAS-mark target, and if successful: 3404 // . if target oop is above global finger (volatile read) 3405 // nothing to do 3406 // . if target oop is in chunk and above local finger 3407 // then nothing to do 3408 // . else push on work-queue 3409 // -- Deal with possible overflow issues: 3410 // . local work-queue overflow causes stuff to be pushed on 3411 // global (common) overflow queue 3412 // . always first empty local work queue 3413 // . then get a batch of oops from global work queue if any 3414 // . then do work stealing 3415 // -- When all tasks claimed (both spaces) 3416 // and local work queue empty, 3417 // then in a loop do: 3418 // . check global overflow stack; steal a batch of oops and trace 3419 // . try to steal from other threads oif GOS is empty 3420 // . if neither is available, offer termination 3421 // -- Terminate and return result 3422 // 3423 void CMSConcMarkingTask::work(uint worker_id) { 3424 elapsedTimer _timer; 3425 ResourceMark rm; 3426 HandleMark hm; 3427 3428 DEBUG_ONLY(_collector->verify_overflow_empty();) 3429 3430 // Before we begin work, our work queue should be empty 3431 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3432 // Scan the bitmap covering _cms_space, tracing through grey objects. 3433 _timer.start(); 3434 do_scan_and_mark(worker_id, _cms_space); 3435 _timer.stop(); 3436 if (PrintCMSStatistics != 0) { 3437 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3438 worker_id, _timer.seconds()); 3439 // XXX: need xxx/xxx type of notation, two timers 3440 } 3441 3442 // ... do work stealing 3443 _timer.reset(); 3444 _timer.start(); 3445 do_work_steal(worker_id); 3446 _timer.stop(); 3447 if (PrintCMSStatistics != 0) { 3448 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3449 worker_id, _timer.seconds()); 3450 // XXX: need xxx/xxx type of notation, two timers 3451 } 3452 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3453 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3454 // Note that under the current task protocol, the 3455 // following assertion is true even of the spaces 3456 // expanded since the completion of the concurrent 3457 // marking. XXX This will likely change under a strict 3458 // ABORT semantics. 3459 // After perm removal the comparison was changed to 3460 // greater than or equal to from strictly greater than. 3461 // Before perm removal the highest address sweep would 3462 // have been at the end of perm gen but now is at the 3463 // end of the tenured gen. 3464 assert(_global_finger >= _cms_space->end(), 3465 "All tasks have been completed"); 3466 DEBUG_ONLY(_collector->verify_overflow_empty();) 3467 } 3468 3469 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3470 HeapWord* read = _global_finger; 3471 HeapWord* cur = read; 3472 while (f > read) { 3473 cur = read; 3474 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3475 if (cur == read) { 3476 // our cas succeeded 3477 assert(_global_finger >= f, "protocol consistency"); 3478 break; 3479 } 3480 } 3481 } 3482 3483 // This is really inefficient, and should be redone by 3484 // using (not yet available) block-read and -write interfaces to the 3485 // stack and the work_queue. XXX FIX ME !!! 3486 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3487 OopTaskQueue* work_q) { 3488 // Fast lock-free check 3489 if (ovflw_stk->length() == 0) { 3490 return false; 3491 } 3492 assert(work_q->size() == 0, "Shouldn't steal"); 3493 MutexLockerEx ml(ovflw_stk->par_lock(), 3494 Mutex::_no_safepoint_check_flag); 3495 // Grab up to 1/4 the size of the work queue 3496 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3497 (size_t)ParGCDesiredObjsFromOverflowList); 3498 num = MIN2(num, ovflw_stk->length()); 3499 for (int i = (int) num; i > 0; i--) { 3500 oop cur = ovflw_stk->pop(); 3501 assert(cur != NULL, "Counted wrong?"); 3502 work_q->push(cur); 3503 } 3504 return num > 0; 3505 } 3506 3507 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3508 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3509 int n_tasks = pst->n_tasks(); 3510 // We allow that there may be no tasks to do here because 3511 // we are restarting after a stack overflow. 3512 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3513 uint nth_task = 0; 3514 3515 HeapWord* aligned_start = sp->bottom(); 3516 if (sp->used_region().contains(_restart_addr)) { 3517 // Align down to a card boundary for the start of 0th task 3518 // for this space. 3519 aligned_start = 3520 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3521 CardTableModRefBS::card_size); 3522 } 3523 3524 size_t chunk_size = sp->marking_task_size(); 3525 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3526 // Having claimed the nth task in this space, 3527 // compute the chunk that it corresponds to: 3528 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3529 aligned_start + (nth_task+1)*chunk_size); 3530 // Try and bump the global finger via a CAS; 3531 // note that we need to do the global finger bump 3532 // _before_ taking the intersection below, because 3533 // the task corresponding to that region will be 3534 // deemed done even if the used_region() expands 3535 // because of allocation -- as it almost certainly will 3536 // during start-up while the threads yield in the 3537 // closure below. 3538 HeapWord* finger = span.end(); 3539 bump_global_finger(finger); // atomically 3540 // There are null tasks here corresponding to chunks 3541 // beyond the "top" address of the space. 3542 span = span.intersection(sp->used_region()); 3543 if (!span.is_empty()) { // Non-null task 3544 HeapWord* prev_obj; 3545 assert(!span.contains(_restart_addr) || nth_task == 0, 3546 "Inconsistency"); 3547 if (nth_task == 0) { 3548 // For the 0th task, we'll not need to compute a block_start. 3549 if (span.contains(_restart_addr)) { 3550 // In the case of a restart because of stack overflow, 3551 // we might additionally skip a chunk prefix. 3552 prev_obj = _restart_addr; 3553 } else { 3554 prev_obj = span.start(); 3555 } 3556 } else { 3557 // We want to skip the first object because 3558 // the protocol is to scan any object in its entirety 3559 // that _starts_ in this span; a fortiori, any 3560 // object starting in an earlier span is scanned 3561 // as part of an earlier claimed task. 3562 // Below we use the "careful" version of block_start 3563 // so we do not try to navigate uninitialized objects. 3564 prev_obj = sp->block_start_careful(span.start()); 3565 // Below we use a variant of block_size that uses the 3566 // Printezis bits to avoid waiting for allocated 3567 // objects to become initialized/parsable. 3568 while (prev_obj < span.start()) { 3569 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3570 if (sz > 0) { 3571 prev_obj += sz; 3572 } else { 3573 // In this case we may end up doing a bit of redundant 3574 // scanning, but that appears unavoidable, short of 3575 // locking the free list locks; see bug 6324141. 3576 break; 3577 } 3578 } 3579 } 3580 if (prev_obj < span.end()) { 3581 MemRegion my_span = MemRegion(prev_obj, span.end()); 3582 // Do the marking work within a non-empty span -- 3583 // the last argument to the constructor indicates whether the 3584 // iteration should be incremental with periodic yields. 3585 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3586 &_collector->_markBitMap, 3587 work_queue(i), 3588 &_collector->_markStack); 3589 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3590 } // else nothing to do for this task 3591 } // else nothing to do for this task 3592 } 3593 // We'd be tempted to assert here that since there are no 3594 // more tasks left to claim in this space, the global_finger 3595 // must exceed space->top() and a fortiori space->end(). However, 3596 // that would not quite be correct because the bumping of 3597 // global_finger occurs strictly after the claiming of a task, 3598 // so by the time we reach here the global finger may not yet 3599 // have been bumped up by the thread that claimed the last 3600 // task. 3601 pst->all_tasks_completed(); 3602 } 3603 3604 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3605 private: 3606 CMSCollector* _collector; 3607 CMSConcMarkingTask* _task; 3608 MemRegion _span; 3609 CMSBitMap* _bit_map; 3610 CMSMarkStack* _overflow_stack; 3611 OopTaskQueue* _work_queue; 3612 protected: 3613 DO_OOP_WORK_DEFN 3614 public: 3615 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3616 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3617 MetadataAwareOopClosure(collector->ref_processor()), 3618 _collector(collector), 3619 _task(task), 3620 _span(collector->_span), 3621 _work_queue(work_queue), 3622 _bit_map(bit_map), 3623 _overflow_stack(overflow_stack) 3624 { } 3625 virtual void do_oop(oop* p); 3626 virtual void do_oop(narrowOop* p); 3627 3628 void trim_queue(size_t max); 3629 void handle_stack_overflow(HeapWord* lost); 3630 void do_yield_check() { 3631 if (_task->should_yield()) { 3632 _task->yield(); 3633 } 3634 } 3635 }; 3636 3637 // Grey object scanning during work stealing phase -- 3638 // the salient assumption here is that any references 3639 // that are in these stolen objects being scanned must 3640 // already have been initialized (else they would not have 3641 // been published), so we do not need to check for 3642 // uninitialized objects before pushing here. 3643 void Par_ConcMarkingClosure::do_oop(oop obj) { 3644 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 3645 HeapWord* addr = (HeapWord*)obj; 3646 // Check if oop points into the CMS generation 3647 // and is not marked 3648 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3649 // a white object ... 3650 // If we manage to "claim" the object, by being the 3651 // first thread to mark it, then we push it on our 3652 // marking stack 3653 if (_bit_map->par_mark(addr)) { // ... now grey 3654 // push on work queue (grey set) 3655 bool simulate_overflow = false; 3656 NOT_PRODUCT( 3657 if (CMSMarkStackOverflowALot && 3658 _collector->simulate_overflow()) { 3659 // simulate a stack overflow 3660 simulate_overflow = true; 3661 } 3662 ) 3663 if (simulate_overflow || 3664 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3665 // stack overflow 3666 if (PrintCMSStatistics != 0) { 3667 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3668 SIZE_FORMAT, _overflow_stack->capacity()); 3669 } 3670 // We cannot assert that the overflow stack is full because 3671 // it may have been emptied since. 3672 assert(simulate_overflow || 3673 _work_queue->size() == _work_queue->max_elems(), 3674 "Else push should have succeeded"); 3675 handle_stack_overflow(addr); 3676 } 3677 } // Else, some other thread got there first 3678 do_yield_check(); 3679 } 3680 } 3681 3682 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3683 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3684 3685 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3686 while (_work_queue->size() > max) { 3687 oop new_oop; 3688 if (_work_queue->pop_local(new_oop)) { 3689 assert(new_oop->is_oop(), "Should be an oop"); 3690 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3691 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3692 new_oop->oop_iterate(this); // do_oop() above 3693 do_yield_check(); 3694 } 3695 } 3696 } 3697 3698 // Upon stack overflow, we discard (part of) the stack, 3699 // remembering the least address amongst those discarded 3700 // in CMSCollector's _restart_address. 3701 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3702 // We need to do this under a mutex to prevent other 3703 // workers from interfering with the work done below. 3704 MutexLockerEx ml(_overflow_stack->par_lock(), 3705 Mutex::_no_safepoint_check_flag); 3706 // Remember the least grey address discarded 3707 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3708 _collector->lower_restart_addr(ra); 3709 _overflow_stack->reset(); // discard stack contents 3710 _overflow_stack->expand(); // expand the stack if possible 3711 } 3712 3713 3714 void CMSConcMarkingTask::do_work_steal(int i) { 3715 OopTaskQueue* work_q = work_queue(i); 3716 oop obj_to_scan; 3717 CMSBitMap* bm = &(_collector->_markBitMap); 3718 CMSMarkStack* ovflw = &(_collector->_markStack); 3719 int* seed = _collector->hash_seed(i); 3720 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3721 while (true) { 3722 cl.trim_queue(0); 3723 assert(work_q->size() == 0, "Should have been emptied above"); 3724 if (get_work_from_overflow_stack(ovflw, work_q)) { 3725 // Can't assert below because the work obtained from the 3726 // overflow stack may already have been stolen from us. 3727 // assert(work_q->size() > 0, "Work from overflow stack"); 3728 continue; 3729 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3730 assert(obj_to_scan->is_oop(), "Should be an oop"); 3731 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3732 obj_to_scan->oop_iterate(&cl); 3733 } else if (terminator()->offer_termination(&_term_term)) { 3734 assert(work_q->size() == 0, "Impossible!"); 3735 break; 3736 } else if (yielding() || should_yield()) { 3737 yield(); 3738 } 3739 } 3740 } 3741 3742 // This is run by the CMS (coordinator) thread. 3743 void CMSConcMarkingTask::coordinator_yield() { 3744 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3745 "CMS thread should hold CMS token"); 3746 // First give up the locks, then yield, then re-lock 3747 // We should probably use a constructor/destructor idiom to 3748 // do this unlock/lock or modify the MutexUnlocker class to 3749 // serve our purpose. XXX 3750 assert_lock_strong(_bit_map_lock); 3751 _bit_map_lock->unlock(); 3752 ConcurrentMarkSweepThread::desynchronize(true); 3753 _collector->stopTimer(); 3754 if (PrintCMSStatistics != 0) { 3755 _collector->incrementYields(); 3756 } 3757 3758 // It is possible for whichever thread initiated the yield request 3759 // not to get a chance to wake up and take the bitmap lock between 3760 // this thread releasing it and reacquiring it. So, while the 3761 // should_yield() flag is on, let's sleep for a bit to give the 3762 // other thread a chance to wake up. The limit imposed on the number 3763 // of iterations is defensive, to avoid any unforseen circumstances 3764 // putting us into an infinite loop. Since it's always been this 3765 // (coordinator_yield()) method that was observed to cause the 3766 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3767 // which is by default non-zero. For the other seven methods that 3768 // also perform the yield operation, as are using a different 3769 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3770 // can enable the sleeping for those methods too, if necessary. 3771 // See 6442774. 3772 // 3773 // We really need to reconsider the synchronization between the GC 3774 // thread and the yield-requesting threads in the future and we 3775 // should really use wait/notify, which is the recommended 3776 // way of doing this type of interaction. Additionally, we should 3777 // consolidate the eight methods that do the yield operation and they 3778 // are almost identical into one for better maintainability and 3779 // readability. See 6445193. 3780 // 3781 // Tony 2006.06.29 3782 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3783 ConcurrentMarkSweepThread::should_yield() && 3784 !CMSCollector::foregroundGCIsActive(); ++i) { 3785 os::sleep(Thread::current(), 1, false); 3786 } 3787 3788 ConcurrentMarkSweepThread::synchronize(true); 3789 _bit_map_lock->lock_without_safepoint_check(); 3790 _collector->startTimer(); 3791 } 3792 3793 bool CMSCollector::do_marking_mt() { 3794 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3795 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 3796 conc_workers()->total_workers(), 3797 conc_workers()->active_workers(), 3798 Threads::number_of_non_daemon_threads()); 3799 conc_workers()->set_active_workers(num_workers); 3800 3801 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3802 3803 CMSConcMarkingTask tsk(this, 3804 cms_space, 3805 conc_workers(), 3806 task_queues()); 3807 3808 // Since the actual number of workers we get may be different 3809 // from the number we requested above, do we need to do anything different 3810 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3811 // class?? XXX 3812 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3813 3814 // Refs discovery is already non-atomic. 3815 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3816 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3817 conc_workers()->start_task(&tsk); 3818 while (tsk.yielded()) { 3819 tsk.coordinator_yield(); 3820 conc_workers()->continue_task(&tsk); 3821 } 3822 // If the task was aborted, _restart_addr will be non-NULL 3823 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3824 while (_restart_addr != NULL) { 3825 // XXX For now we do not make use of ABORTED state and have not 3826 // yet implemented the right abort semantics (even in the original 3827 // single-threaded CMS case). That needs some more investigation 3828 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3829 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3830 // If _restart_addr is non-NULL, a marking stack overflow 3831 // occurred; we need to do a fresh marking iteration from the 3832 // indicated restart address. 3833 if (_foregroundGCIsActive) { 3834 // We may be running into repeated stack overflows, having 3835 // reached the limit of the stack size, while making very 3836 // slow forward progress. It may be best to bail out and 3837 // let the foreground collector do its job. 3838 // Clear _restart_addr, so that foreground GC 3839 // works from scratch. This avoids the headache of 3840 // a "rescan" which would otherwise be needed because 3841 // of the dirty mod union table & card table. 3842 _restart_addr = NULL; 3843 return false; 3844 } 3845 // Adjust the task to restart from _restart_addr 3846 tsk.reset(_restart_addr); 3847 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3848 _restart_addr); 3849 _restart_addr = NULL; 3850 // Get the workers going again 3851 conc_workers()->start_task(&tsk); 3852 while (tsk.yielded()) { 3853 tsk.coordinator_yield(); 3854 conc_workers()->continue_task(&tsk); 3855 } 3856 } 3857 assert(tsk.completed(), "Inconsistency"); 3858 assert(tsk.result() == true, "Inconsistency"); 3859 return true; 3860 } 3861 3862 bool CMSCollector::do_marking_st() { 3863 ResourceMark rm; 3864 HandleMark hm; 3865 3866 // Temporarily make refs discovery single threaded (non-MT) 3867 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3868 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3869 &_markStack, CMSYield); 3870 // the last argument to iterate indicates whether the iteration 3871 // should be incremental with periodic yields. 3872 _markBitMap.iterate(&markFromRootsClosure); 3873 // If _restart_addr is non-NULL, a marking stack overflow 3874 // occurred; we need to do a fresh iteration from the 3875 // indicated restart address. 3876 while (_restart_addr != NULL) { 3877 if (_foregroundGCIsActive) { 3878 // We may be running into repeated stack overflows, having 3879 // reached the limit of the stack size, while making very 3880 // slow forward progress. It may be best to bail out and 3881 // let the foreground collector do its job. 3882 // Clear _restart_addr, so that foreground GC 3883 // works from scratch. This avoids the headache of 3884 // a "rescan" which would otherwise be needed because 3885 // of the dirty mod union table & card table. 3886 _restart_addr = NULL; 3887 return false; // indicating failure to complete marking 3888 } 3889 // Deal with stack overflow: 3890 // we restart marking from _restart_addr 3891 HeapWord* ra = _restart_addr; 3892 markFromRootsClosure.reset(ra); 3893 _restart_addr = NULL; 3894 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3895 } 3896 return true; 3897 } 3898 3899 void CMSCollector::preclean() { 3900 check_correct_thread_executing(); 3901 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3902 verify_work_stacks_empty(); 3903 verify_overflow_empty(); 3904 _abort_preclean = false; 3905 if (CMSPrecleaningEnabled) { 3906 if (!CMSEdenChunksRecordAlways) { 3907 _eden_chunk_index = 0; 3908 } 3909 size_t used = get_eden_used(); 3910 size_t capacity = get_eden_capacity(); 3911 // Don't start sampling unless we will get sufficiently 3912 // many samples. 3913 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3914 * CMSScheduleRemarkEdenPenetration)) { 3915 _start_sampling = true; 3916 } else { 3917 _start_sampling = false; 3918 } 3919 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3920 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3921 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3922 } 3923 CMSTokenSync x(true); // is cms thread 3924 if (CMSPrecleaningEnabled) { 3925 sample_eden(); 3926 _collectorState = AbortablePreclean; 3927 } else { 3928 _collectorState = FinalMarking; 3929 } 3930 verify_work_stacks_empty(); 3931 verify_overflow_empty(); 3932 } 3933 3934 // Try and schedule the remark such that young gen 3935 // occupancy is CMSScheduleRemarkEdenPenetration %. 3936 void CMSCollector::abortable_preclean() { 3937 check_correct_thread_executing(); 3938 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3939 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3940 3941 // If Eden's current occupancy is below this threshold, 3942 // immediately schedule the remark; else preclean 3943 // past the next scavenge in an effort to 3944 // schedule the pause as described above. By choosing 3945 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3946 // we will never do an actual abortable preclean cycle. 3947 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3948 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3949 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3950 // We need more smarts in the abortable preclean 3951 // loop below to deal with cases where allocation 3952 // in young gen is very very slow, and our precleaning 3953 // is running a losing race against a horde of 3954 // mutators intent on flooding us with CMS updates 3955 // (dirty cards). 3956 // One, admittedly dumb, strategy is to give up 3957 // after a certain number of abortable precleaning loops 3958 // or after a certain maximum time. We want to make 3959 // this smarter in the next iteration. 3960 // XXX FIX ME!!! YSR 3961 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3962 while (!(should_abort_preclean() || 3963 ConcurrentMarkSweepThread::should_terminate())) { 3964 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3965 cumworkdone += workdone; 3966 loops++; 3967 // Voluntarily terminate abortable preclean phase if we have 3968 // been at it for too long. 3969 if ((CMSMaxAbortablePrecleanLoops != 0) && 3970 loops >= CMSMaxAbortablePrecleanLoops) { 3971 if (PrintGCDetails) { 3972 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3973 } 3974 break; 3975 } 3976 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3977 if (PrintGCDetails) { 3978 gclog_or_tty->print(" CMS: abort preclean due to time "); 3979 } 3980 break; 3981 } 3982 // If we are doing little work each iteration, we should 3983 // take a short break. 3984 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3985 // Sleep for some time, waiting for work to accumulate 3986 stopTimer(); 3987 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3988 startTimer(); 3989 waited++; 3990 } 3991 } 3992 if (PrintCMSStatistics > 0) { 3993 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3994 loops, waited, cumworkdone); 3995 } 3996 } 3997 CMSTokenSync x(true); // is cms thread 3998 if (_collectorState != Idling) { 3999 assert(_collectorState == AbortablePreclean, 4000 "Spontaneous state transition?"); 4001 _collectorState = FinalMarking; 4002 } // Else, a foreground collection completed this CMS cycle. 4003 return; 4004 } 4005 4006 // Respond to an Eden sampling opportunity 4007 void CMSCollector::sample_eden() { 4008 // Make sure a young gc cannot sneak in between our 4009 // reading and recording of a sample. 4010 assert(Thread::current()->is_ConcurrentGC_thread(), 4011 "Only the cms thread may collect Eden samples"); 4012 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4013 "Should collect samples while holding CMS token"); 4014 if (!_start_sampling) { 4015 return; 4016 } 4017 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4018 // is populated by the young generation. 4019 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4020 if (_eden_chunk_index < _eden_chunk_capacity) { 4021 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4022 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4023 "Unexpected state of Eden"); 4024 // We'd like to check that what we just sampled is an oop-start address; 4025 // however, we cannot do that here since the object may not yet have been 4026 // initialized. So we'll instead do the check when we _use_ this sample 4027 // later. 4028 if (_eden_chunk_index == 0 || 4029 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4030 _eden_chunk_array[_eden_chunk_index-1]) 4031 >= CMSSamplingGrain)) { 4032 _eden_chunk_index++; // commit sample 4033 } 4034 } 4035 } 4036 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4037 size_t used = get_eden_used(); 4038 size_t capacity = get_eden_capacity(); 4039 assert(used <= capacity, "Unexpected state of Eden"); 4040 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4041 _abort_preclean = true; 4042 } 4043 } 4044 } 4045 4046 4047 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4048 assert(_collectorState == Precleaning || 4049 _collectorState == AbortablePreclean, "incorrect state"); 4050 ResourceMark rm; 4051 HandleMark hm; 4052 4053 // Precleaning is currently not MT but the reference processor 4054 // may be set for MT. Disable it temporarily here. 4055 ReferenceProcessor* rp = ref_processor(); 4056 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4057 4058 // Do one pass of scrubbing the discovered reference lists 4059 // to remove any reference objects with strongly-reachable 4060 // referents. 4061 if (clean_refs) { 4062 CMSPrecleanRefsYieldClosure yield_cl(this); 4063 assert(rp->span().equals(_span), "Spans should be equal"); 4064 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4065 &_markStack, true /* preclean */); 4066 CMSDrainMarkingStackClosure complete_trace(this, 4067 _span, &_markBitMap, &_markStack, 4068 &keep_alive, true /* preclean */); 4069 4070 // We don't want this step to interfere with a young 4071 // collection because we don't want to take CPU 4072 // or memory bandwidth away from the young GC threads 4073 // (which may be as many as there are CPUs). 4074 // Note that we don't need to protect ourselves from 4075 // interference with mutators because they can't 4076 // manipulate the discovered reference lists nor affect 4077 // the computed reachability of the referents, the 4078 // only properties manipulated by the precleaning 4079 // of these reference lists. 4080 stopTimer(); 4081 CMSTokenSyncWithLocks x(true /* is cms thread */, 4082 bitMapLock()); 4083 startTimer(); 4084 sample_eden(); 4085 4086 // The following will yield to allow foreground 4087 // collection to proceed promptly. XXX YSR: 4088 // The code in this method may need further 4089 // tweaking for better performance and some restructuring 4090 // for cleaner interfaces. 4091 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4092 rp->preclean_discovered_references( 4093 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4094 gc_timer, _gc_tracer_cm->gc_id()); 4095 } 4096 4097 if (clean_survivor) { // preclean the active survivor space(s) 4098 assert(_young_gen->kind() == Generation::DefNew || 4099 _young_gen->kind() == Generation::ParNew, 4100 "incorrect type for cast"); 4101 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4102 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4103 &_markBitMap, &_modUnionTable, 4104 &_markStack, true /* precleaning phase */); 4105 stopTimer(); 4106 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4107 bitMapLock()); 4108 startTimer(); 4109 unsigned int before_count = 4110 GenCollectedHeap::heap()->total_collections(); 4111 SurvivorSpacePrecleanClosure 4112 sss_cl(this, _span, &_markBitMap, &_markStack, 4113 &pam_cl, before_count, CMSYield); 4114 dng->from()->object_iterate_careful(&sss_cl); 4115 dng->to()->object_iterate_careful(&sss_cl); 4116 } 4117 MarkRefsIntoAndScanClosure 4118 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4119 &_markStack, this, CMSYield, 4120 true /* precleaning phase */); 4121 // CAUTION: The following closure has persistent state that may need to 4122 // be reset upon a decrease in the sequence of addresses it 4123 // processes. 4124 ScanMarkedObjectsAgainCarefullyClosure 4125 smoac_cl(this, _span, 4126 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4127 4128 // Preclean dirty cards in ModUnionTable and CardTable using 4129 // appropriate convergence criterion; 4130 // repeat CMSPrecleanIter times unless we find that 4131 // we are losing. 4132 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4133 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4134 "Bad convergence multiplier"); 4135 assert(CMSPrecleanThreshold >= 100, 4136 "Unreasonably low CMSPrecleanThreshold"); 4137 4138 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4139 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4140 numIter < CMSPrecleanIter; 4141 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4142 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4143 if (Verbose && PrintGCDetails) { 4144 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4145 } 4146 // Either there are very few dirty cards, so re-mark 4147 // pause will be small anyway, or our pre-cleaning isn't 4148 // that much faster than the rate at which cards are being 4149 // dirtied, so we might as well stop and re-mark since 4150 // precleaning won't improve our re-mark time by much. 4151 if (curNumCards <= CMSPrecleanThreshold || 4152 (numIter > 0 && 4153 (curNumCards * CMSPrecleanDenominator > 4154 lastNumCards * CMSPrecleanNumerator))) { 4155 numIter++; 4156 cumNumCards += curNumCards; 4157 break; 4158 } 4159 } 4160 4161 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4162 4163 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4164 cumNumCards += curNumCards; 4165 if (PrintGCDetails && PrintCMSStatistics != 0) { 4166 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4167 curNumCards, cumNumCards, numIter); 4168 } 4169 return cumNumCards; // as a measure of useful work done 4170 } 4171 4172 // PRECLEANING NOTES: 4173 // Precleaning involves: 4174 // . reading the bits of the modUnionTable and clearing the set bits. 4175 // . For the cards corresponding to the set bits, we scan the 4176 // objects on those cards. This means we need the free_list_lock 4177 // so that we can safely iterate over the CMS space when scanning 4178 // for oops. 4179 // . When we scan the objects, we'll be both reading and setting 4180 // marks in the marking bit map, so we'll need the marking bit map. 4181 // . For protecting _collector_state transitions, we take the CGC_lock. 4182 // Note that any races in the reading of of card table entries by the 4183 // CMS thread on the one hand and the clearing of those entries by the 4184 // VM thread or the setting of those entries by the mutator threads on the 4185 // other are quite benign. However, for efficiency it makes sense to keep 4186 // the VM thread from racing with the CMS thread while the latter is 4187 // dirty card info to the modUnionTable. We therefore also use the 4188 // CGC_lock to protect the reading of the card table and the mod union 4189 // table by the CM thread. 4190 // . We run concurrently with mutator updates, so scanning 4191 // needs to be done carefully -- we should not try to scan 4192 // potentially uninitialized objects. 4193 // 4194 // Locking strategy: While holding the CGC_lock, we scan over and 4195 // reset a maximal dirty range of the mod union / card tables, then lock 4196 // the free_list_lock and bitmap lock to do a full marking, then 4197 // release these locks; and repeat the cycle. This allows for a 4198 // certain amount of fairness in the sharing of these locks between 4199 // the CMS collector on the one hand, and the VM thread and the 4200 // mutators on the other. 4201 4202 // NOTE: preclean_mod_union_table() and preclean_card_table() 4203 // further below are largely identical; if you need to modify 4204 // one of these methods, please check the other method too. 4205 4206 size_t CMSCollector::preclean_mod_union_table( 4207 ConcurrentMarkSweepGeneration* gen, 4208 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4209 verify_work_stacks_empty(); 4210 verify_overflow_empty(); 4211 4212 // strategy: starting with the first card, accumulate contiguous 4213 // ranges of dirty cards; clear these cards, then scan the region 4214 // covered by these cards. 4215 4216 // Since all of the MUT is committed ahead, we can just use 4217 // that, in case the generations expand while we are precleaning. 4218 // It might also be fine to just use the committed part of the 4219 // generation, but we might potentially miss cards when the 4220 // generation is rapidly expanding while we are in the midst 4221 // of precleaning. 4222 HeapWord* startAddr = gen->reserved().start(); 4223 HeapWord* endAddr = gen->reserved().end(); 4224 4225 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4226 4227 size_t numDirtyCards, cumNumDirtyCards; 4228 HeapWord *nextAddr, *lastAddr; 4229 for (cumNumDirtyCards = numDirtyCards = 0, 4230 nextAddr = lastAddr = startAddr; 4231 nextAddr < endAddr; 4232 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4233 4234 ResourceMark rm; 4235 HandleMark hm; 4236 4237 MemRegion dirtyRegion; 4238 { 4239 stopTimer(); 4240 // Potential yield point 4241 CMSTokenSync ts(true); 4242 startTimer(); 4243 sample_eden(); 4244 // Get dirty region starting at nextOffset (inclusive), 4245 // simultaneously clearing it. 4246 dirtyRegion = 4247 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4248 assert(dirtyRegion.start() >= nextAddr, 4249 "returned region inconsistent?"); 4250 } 4251 // Remember where the next search should begin. 4252 // The returned region (if non-empty) is a right open interval, 4253 // so lastOffset is obtained from the right end of that 4254 // interval. 4255 lastAddr = dirtyRegion.end(); 4256 // Should do something more transparent and less hacky XXX 4257 numDirtyCards = 4258 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4259 4260 // We'll scan the cards in the dirty region (with periodic 4261 // yields for foreground GC as needed). 4262 if (!dirtyRegion.is_empty()) { 4263 assert(numDirtyCards > 0, "consistency check"); 4264 HeapWord* stop_point = NULL; 4265 stopTimer(); 4266 // Potential yield point 4267 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4268 bitMapLock()); 4269 startTimer(); 4270 { 4271 verify_work_stacks_empty(); 4272 verify_overflow_empty(); 4273 sample_eden(); 4274 stop_point = 4275 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4276 } 4277 if (stop_point != NULL) { 4278 // The careful iteration stopped early either because it found an 4279 // uninitialized object, or because we were in the midst of an 4280 // "abortable preclean", which should now be aborted. Redirty 4281 // the bits corresponding to the partially-scanned or unscanned 4282 // cards. We'll either restart at the next block boundary or 4283 // abort the preclean. 4284 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4285 "Should only be AbortablePreclean."); 4286 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4287 if (should_abort_preclean()) { 4288 break; // out of preclean loop 4289 } else { 4290 // Compute the next address at which preclean should pick up; 4291 // might need bitMapLock in order to read P-bits. 4292 lastAddr = next_card_start_after_block(stop_point); 4293 } 4294 } 4295 } else { 4296 assert(lastAddr == endAddr, "consistency check"); 4297 assert(numDirtyCards == 0, "consistency check"); 4298 break; 4299 } 4300 } 4301 verify_work_stacks_empty(); 4302 verify_overflow_empty(); 4303 return cumNumDirtyCards; 4304 } 4305 4306 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4307 // below are largely identical; if you need to modify 4308 // one of these methods, please check the other method too. 4309 4310 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4311 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4312 // strategy: it's similar to precleamModUnionTable above, in that 4313 // we accumulate contiguous ranges of dirty cards, mark these cards 4314 // precleaned, then scan the region covered by these cards. 4315 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4316 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4317 4318 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4319 4320 size_t numDirtyCards, cumNumDirtyCards; 4321 HeapWord *lastAddr, *nextAddr; 4322 4323 for (cumNumDirtyCards = numDirtyCards = 0, 4324 nextAddr = lastAddr = startAddr; 4325 nextAddr < endAddr; 4326 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4327 4328 ResourceMark rm; 4329 HandleMark hm; 4330 4331 MemRegion dirtyRegion; 4332 { 4333 // See comments in "Precleaning notes" above on why we 4334 // do this locking. XXX Could the locking overheads be 4335 // too high when dirty cards are sparse? [I don't think so.] 4336 stopTimer(); 4337 CMSTokenSync x(true); // is cms thread 4338 startTimer(); 4339 sample_eden(); 4340 // Get and clear dirty region from card table 4341 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4342 MemRegion(nextAddr, endAddr), 4343 true, 4344 CardTableModRefBS::precleaned_card_val()); 4345 4346 assert(dirtyRegion.start() >= nextAddr, 4347 "returned region inconsistent?"); 4348 } 4349 lastAddr = dirtyRegion.end(); 4350 numDirtyCards = 4351 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4352 4353 if (!dirtyRegion.is_empty()) { 4354 stopTimer(); 4355 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4356 startTimer(); 4357 sample_eden(); 4358 verify_work_stacks_empty(); 4359 verify_overflow_empty(); 4360 HeapWord* stop_point = 4361 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4362 if (stop_point != NULL) { 4363 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4364 "Should only be AbortablePreclean."); 4365 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4366 if (should_abort_preclean()) { 4367 break; // out of preclean loop 4368 } else { 4369 // Compute the next address at which preclean should pick up. 4370 lastAddr = next_card_start_after_block(stop_point); 4371 } 4372 } 4373 } else { 4374 break; 4375 } 4376 } 4377 verify_work_stacks_empty(); 4378 verify_overflow_empty(); 4379 return cumNumDirtyCards; 4380 } 4381 4382 class PrecleanKlassClosure : public KlassClosure { 4383 KlassToOopClosure _cm_klass_closure; 4384 public: 4385 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4386 void do_klass(Klass* k) { 4387 if (k->has_accumulated_modified_oops()) { 4388 k->clear_accumulated_modified_oops(); 4389 4390 _cm_klass_closure.do_klass(k); 4391 } 4392 } 4393 }; 4394 4395 // The freelist lock is needed to prevent asserts, is it really needed? 4396 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4397 4398 cl->set_freelistLock(freelistLock); 4399 4400 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4401 4402 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4403 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4404 PrecleanKlassClosure preclean_klass_closure(cl); 4405 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4406 4407 verify_work_stacks_empty(); 4408 verify_overflow_empty(); 4409 } 4410 4411 void CMSCollector::checkpointRootsFinal() { 4412 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4413 check_correct_thread_executing(); 4414 // world is stopped at this checkpoint 4415 assert(SafepointSynchronize::is_at_safepoint(), 4416 "world should be stopped"); 4417 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4418 4419 verify_work_stacks_empty(); 4420 verify_overflow_empty(); 4421 4422 SpecializationStats::clear(); 4423 if (PrintGCDetails) { 4424 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4425 _young_gen->used() / K, 4426 _young_gen->capacity() / K); 4427 } 4428 { 4429 if (CMSScavengeBeforeRemark) { 4430 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4431 // Temporarily set flag to false, GCH->do_collection will 4432 // expect it to be false and set to true 4433 FlagSetting fl(gch->_is_gc_active, false); 4434 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4435 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4436 int level = _cmsGen->level() - 1; 4437 if (level >= 0) { 4438 gch->do_collection(true, // full (i.e. force, see below) 4439 false, // !clear_all_soft_refs 4440 0, // size 4441 false, // is_tlab 4442 level // max_level 4443 ); 4444 } 4445 } 4446 FreelistLocker x(this); 4447 MutexLockerEx y(bitMapLock(), 4448 Mutex::_no_safepoint_check_flag); 4449 checkpointRootsFinalWork(); 4450 } 4451 verify_work_stacks_empty(); 4452 verify_overflow_empty(); 4453 SpecializationStats::print(); 4454 } 4455 4456 void CMSCollector::checkpointRootsFinalWork() { 4457 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4458 4459 assert(haveFreelistLocks(), "must have free list locks"); 4460 assert_lock_strong(bitMapLock()); 4461 4462 ResourceMark rm; 4463 HandleMark hm; 4464 4465 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4466 4467 if (should_unload_classes()) { 4468 CodeCache::gc_prologue(); 4469 } 4470 assert(haveFreelistLocks(), "must have free list locks"); 4471 assert_lock_strong(bitMapLock()); 4472 4473 // We might assume that we need not fill TLAB's when 4474 // CMSScavengeBeforeRemark is set, because we may have just done 4475 // a scavenge which would have filled all TLAB's -- and besides 4476 // Eden would be empty. This however may not always be the case -- 4477 // for instance although we asked for a scavenge, it may not have 4478 // happened because of a JNI critical section. We probably need 4479 // a policy for deciding whether we can in that case wait until 4480 // the critical section releases and then do the remark following 4481 // the scavenge, and skip it here. In the absence of that policy, 4482 // or of an indication of whether the scavenge did indeed occur, 4483 // we cannot rely on TLAB's having been filled and must do 4484 // so here just in case a scavenge did not happen. 4485 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4486 // Update the saved marks which may affect the root scans. 4487 gch->save_marks(); 4488 4489 if (CMSPrintEdenSurvivorChunks) { 4490 print_eden_and_survivor_chunk_arrays(); 4491 } 4492 4493 { 4494 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4495 4496 // Note on the role of the mod union table: 4497 // Since the marker in "markFromRoots" marks concurrently with 4498 // mutators, it is possible for some reachable objects not to have been 4499 // scanned. For instance, an only reference to an object A was 4500 // placed in object B after the marker scanned B. Unless B is rescanned, 4501 // A would be collected. Such updates to references in marked objects 4502 // are detected via the mod union table which is the set of all cards 4503 // dirtied since the first checkpoint in this GC cycle and prior to 4504 // the most recent young generation GC, minus those cleaned up by the 4505 // concurrent precleaning. 4506 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4507 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 4508 do_remark_parallel(); 4509 } else { 4510 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4511 _gc_timer_cm, _gc_tracer_cm->gc_id()); 4512 do_remark_non_parallel(); 4513 } 4514 } 4515 verify_work_stacks_empty(); 4516 verify_overflow_empty(); 4517 4518 { 4519 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4520 refProcessingWork(); 4521 } 4522 verify_work_stacks_empty(); 4523 verify_overflow_empty(); 4524 4525 if (should_unload_classes()) { 4526 CodeCache::gc_epilogue(); 4527 } 4528 JvmtiExport::gc_epilogue(); 4529 4530 // If we encountered any (marking stack / work queue) overflow 4531 // events during the current CMS cycle, take appropriate 4532 // remedial measures, where possible, so as to try and avoid 4533 // recurrence of that condition. 4534 assert(_markStack.isEmpty(), "No grey objects"); 4535 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4536 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4537 if (ser_ovflw > 0) { 4538 if (PrintCMSStatistics != 0) { 4539 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4540 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4541 ", kac_preclean="SIZE_FORMAT")", 4542 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4543 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4544 } 4545 _markStack.expand(); 4546 _ser_pmc_remark_ovflw = 0; 4547 _ser_pmc_preclean_ovflw = 0; 4548 _ser_kac_preclean_ovflw = 0; 4549 _ser_kac_ovflw = 0; 4550 } 4551 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4552 if (PrintCMSStatistics != 0) { 4553 gclog_or_tty->print_cr("Work queue overflow (benign) " 4554 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4555 _par_pmc_remark_ovflw, _par_kac_ovflw); 4556 } 4557 _par_pmc_remark_ovflw = 0; 4558 _par_kac_ovflw = 0; 4559 } 4560 if (PrintCMSStatistics != 0) { 4561 if (_markStack._hit_limit > 0) { 4562 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4563 _markStack._hit_limit); 4564 } 4565 if (_markStack._failed_double > 0) { 4566 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4567 " current capacity "SIZE_FORMAT, 4568 _markStack._failed_double, 4569 _markStack.capacity()); 4570 } 4571 } 4572 _markStack._hit_limit = 0; 4573 _markStack._failed_double = 0; 4574 4575 if ((VerifyAfterGC || VerifyDuringGC) && 4576 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4577 verify_after_remark(); 4578 } 4579 4580 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4581 4582 // Change under the freelistLocks. 4583 _collectorState = Sweeping; 4584 // Call isAllClear() under bitMapLock 4585 assert(_modUnionTable.isAllClear(), 4586 "Should be clear by end of the final marking"); 4587 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4588 "Should be clear by end of the final marking"); 4589 } 4590 4591 void CMSParInitialMarkTask::work(uint worker_id) { 4592 elapsedTimer _timer; 4593 ResourceMark rm; 4594 HandleMark hm; 4595 4596 // ---------- scan from roots -------------- 4597 _timer.start(); 4598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4599 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4600 4601 // ---------- young gen roots -------------- 4602 { 4603 work_on_young_gen_roots(worker_id, &par_mri_cl); 4604 _timer.stop(); 4605 if (PrintCMSStatistics != 0) { 4606 gclog_or_tty->print_cr( 4607 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4608 worker_id, _timer.seconds()); 4609 } 4610 } 4611 4612 // ---------- remaining roots -------------- 4613 _timer.reset(); 4614 _timer.start(); 4615 4616 CLDToOopClosure cld_closure(&par_mri_cl, true); 4617 4618 gch->gen_process_roots(_collector->_cmsGen->level(), 4619 false, // yg was scanned above 4620 false, // this is parallel code 4621 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4622 _collector->should_unload_classes(), 4623 &par_mri_cl, 4624 NULL, 4625 &cld_closure); 4626 assert(_collector->should_unload_classes() 4627 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 4628 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4629 _timer.stop(); 4630 if (PrintCMSStatistics != 0) { 4631 gclog_or_tty->print_cr( 4632 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4633 worker_id, _timer.seconds()); 4634 } 4635 } 4636 4637 // Parallel remark task 4638 class CMSParRemarkTask: public CMSParMarkTask { 4639 CompactibleFreeListSpace* _cms_space; 4640 4641 // The per-thread work queues, available here for stealing. 4642 OopTaskQueueSet* _task_queues; 4643 ParallelTaskTerminator _term; 4644 4645 public: 4646 // A value of 0 passed to n_workers will cause the number of 4647 // workers to be taken from the active workers in the work gang. 4648 CMSParRemarkTask(CMSCollector* collector, 4649 CompactibleFreeListSpace* cms_space, 4650 int n_workers, FlexibleWorkGang* workers, 4651 OopTaskQueueSet* task_queues): 4652 CMSParMarkTask("Rescan roots and grey objects in parallel", 4653 collector, n_workers), 4654 _cms_space(cms_space), 4655 _task_queues(task_queues), 4656 _term(n_workers, task_queues) { } 4657 4658 OopTaskQueueSet* task_queues() { return _task_queues; } 4659 4660 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4661 4662 ParallelTaskTerminator* terminator() { return &_term; } 4663 int n_workers() { return _n_workers; } 4664 4665 void work(uint worker_id); 4666 4667 private: 4668 // ... of dirty cards in old space 4669 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4670 Par_MarkRefsIntoAndScanClosure* cl); 4671 4672 // ... work stealing for the above 4673 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4674 }; 4675 4676 class RemarkKlassClosure : public KlassClosure { 4677 KlassToOopClosure _cm_klass_closure; 4678 public: 4679 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4680 void do_klass(Klass* k) { 4681 // Check if we have modified any oops in the Klass during the concurrent marking. 4682 if (k->has_accumulated_modified_oops()) { 4683 k->clear_accumulated_modified_oops(); 4684 4685 // We could have transfered the current modified marks to the accumulated marks, 4686 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4687 } else if (k->has_modified_oops()) { 4688 // Don't clear anything, this info is needed by the next young collection. 4689 } else { 4690 // No modified oops in the Klass. 4691 return; 4692 } 4693 4694 // The klass has modified fields, need to scan the klass. 4695 _cm_klass_closure.do_klass(k); 4696 } 4697 }; 4698 4699 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4700 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 4701 ContiguousSpace* eden_space = dng->eden(); 4702 ContiguousSpace* from_space = dng->from(); 4703 ContiguousSpace* to_space = dng->to(); 4704 4705 HeapWord** eca = _collector->_eden_chunk_array; 4706 size_t ect = _collector->_eden_chunk_index; 4707 HeapWord** sca = _collector->_survivor_chunk_array; 4708 size_t sct = _collector->_survivor_chunk_index; 4709 4710 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4711 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4712 4713 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4714 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4715 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4716 } 4717 4718 // work_queue(i) is passed to the closure 4719 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4720 // also is passed to do_dirty_card_rescan_tasks() and to 4721 // do_work_steal() to select the i-th task_queue. 4722 4723 void CMSParRemarkTask::work(uint worker_id) { 4724 elapsedTimer _timer; 4725 ResourceMark rm; 4726 HandleMark hm; 4727 4728 // ---------- rescan from roots -------------- 4729 _timer.start(); 4730 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4731 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4732 _collector->_span, _collector->ref_processor(), 4733 &(_collector->_markBitMap), 4734 work_queue(worker_id)); 4735 4736 // Rescan young gen roots first since these are likely 4737 // coarsely partitioned and may, on that account, constitute 4738 // the critical path; thus, it's best to start off that 4739 // work first. 4740 // ---------- young gen roots -------------- 4741 { 4742 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4743 _timer.stop(); 4744 if (PrintCMSStatistics != 0) { 4745 gclog_or_tty->print_cr( 4746 "Finished young gen rescan work in %dth thread: %3.3f sec", 4747 worker_id, _timer.seconds()); 4748 } 4749 } 4750 4751 // ---------- remaining roots -------------- 4752 _timer.reset(); 4753 _timer.start(); 4754 gch->gen_process_roots(_collector->_cmsGen->level(), 4755 false, // yg was scanned above 4756 false, // this is parallel code 4757 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4758 _collector->should_unload_classes(), 4759 &par_mrias_cl, 4760 NULL, 4761 NULL); // The dirty klasses will be handled below 4762 4763 assert(_collector->should_unload_classes() 4764 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 4765 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4766 _timer.stop(); 4767 if (PrintCMSStatistics != 0) { 4768 gclog_or_tty->print_cr( 4769 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4770 worker_id, _timer.seconds()); 4771 } 4772 4773 // ---------- unhandled CLD scanning ---------- 4774 if (worker_id == 0) { // Single threaded at the moment. 4775 _timer.reset(); 4776 _timer.start(); 4777 4778 // Scan all new class loader data objects and new dependencies that were 4779 // introduced during concurrent marking. 4780 ResourceMark rm; 4781 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4782 for (int i = 0; i < array->length(); i++) { 4783 par_mrias_cl.do_class_loader_data(array->at(i)); 4784 } 4785 4786 // We don't need to keep track of new CLDs anymore. 4787 ClassLoaderDataGraph::remember_new_clds(false); 4788 4789 _timer.stop(); 4790 if (PrintCMSStatistics != 0) { 4791 gclog_or_tty->print_cr( 4792 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4793 worker_id, _timer.seconds()); 4794 } 4795 } 4796 4797 // ---------- dirty klass scanning ---------- 4798 if (worker_id == 0) { // Single threaded at the moment. 4799 _timer.reset(); 4800 _timer.start(); 4801 4802 // Scan all classes that was dirtied during the concurrent marking phase. 4803 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4804 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4805 4806 _timer.stop(); 4807 if (PrintCMSStatistics != 0) { 4808 gclog_or_tty->print_cr( 4809 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4810 worker_id, _timer.seconds()); 4811 } 4812 } 4813 4814 // We might have added oops to ClassLoaderData::_handles during the 4815 // concurrent marking phase. These oops point to newly allocated objects 4816 // that are guaranteed to be kept alive. Either by the direct allocation 4817 // code, or when the young collector processes the roots. Hence, 4818 // we don't have to revisit the _handles block during the remark phase. 4819 4820 // ---------- rescan dirty cards ------------ 4821 _timer.reset(); 4822 _timer.start(); 4823 4824 // Do the rescan tasks for each of the two spaces 4825 // (cms_space) in turn. 4826 // "worker_id" is passed to select the task_queue for "worker_id" 4827 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4828 _timer.stop(); 4829 if (PrintCMSStatistics != 0) { 4830 gclog_or_tty->print_cr( 4831 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4832 worker_id, _timer.seconds()); 4833 } 4834 4835 // ---------- steal work from other threads ... 4836 // ---------- ... and drain overflow list. 4837 _timer.reset(); 4838 _timer.start(); 4839 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4840 _timer.stop(); 4841 if (PrintCMSStatistics != 0) { 4842 gclog_or_tty->print_cr( 4843 "Finished work stealing in %dth thread: %3.3f sec", 4844 worker_id, _timer.seconds()); 4845 } 4846 } 4847 4848 // Note that parameter "i" is not used. 4849 void 4850 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4851 OopsInGenClosure* cl, ContiguousSpace* space, 4852 HeapWord** chunk_array, size_t chunk_top) { 4853 // Until all tasks completed: 4854 // . claim an unclaimed task 4855 // . compute region boundaries corresponding to task claimed 4856 // using chunk_array 4857 // . par_oop_iterate(cl) over that region 4858 4859 ResourceMark rm; 4860 HandleMark hm; 4861 4862 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4863 4864 uint nth_task = 0; 4865 uint n_tasks = pst->n_tasks(); 4866 4867 if (n_tasks > 0) { 4868 assert(pst->valid(), "Uninitialized use?"); 4869 HeapWord *start, *end; 4870 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4871 // We claimed task # nth_task; compute its boundaries. 4872 if (chunk_top == 0) { // no samples were taken 4873 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4874 start = space->bottom(); 4875 end = space->top(); 4876 } else if (nth_task == 0) { 4877 start = space->bottom(); 4878 end = chunk_array[nth_task]; 4879 } else if (nth_task < (uint)chunk_top) { 4880 assert(nth_task >= 1, "Control point invariant"); 4881 start = chunk_array[nth_task - 1]; 4882 end = chunk_array[nth_task]; 4883 } else { 4884 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4885 start = chunk_array[chunk_top - 1]; 4886 end = space->top(); 4887 } 4888 MemRegion mr(start, end); 4889 // Verify that mr is in space 4890 assert(mr.is_empty() || space->used_region().contains(mr), 4891 "Should be in space"); 4892 // Verify that "start" is an object boundary 4893 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4894 "Should be an oop"); 4895 space->par_oop_iterate(mr, cl); 4896 } 4897 pst->all_tasks_completed(); 4898 } 4899 } 4900 4901 void 4902 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4903 CompactibleFreeListSpace* sp, int i, 4904 Par_MarkRefsIntoAndScanClosure* cl) { 4905 // Until all tasks completed: 4906 // . claim an unclaimed task 4907 // . compute region boundaries corresponding to task claimed 4908 // . transfer dirty bits ct->mut for that region 4909 // . apply rescanclosure to dirty mut bits for that region 4910 4911 ResourceMark rm; 4912 HandleMark hm; 4913 4914 OopTaskQueue* work_q = work_queue(i); 4915 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4916 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4917 // CAUTION: This closure has state that persists across calls to 4918 // the work method dirty_range_iterate_clear() in that it has 4919 // embedded in it a (subtype of) UpwardsObjectClosure. The 4920 // use of that state in the embedded UpwardsObjectClosure instance 4921 // assumes that the cards are always iterated (even if in parallel 4922 // by several threads) in monotonically increasing order per each 4923 // thread. This is true of the implementation below which picks 4924 // card ranges (chunks) in monotonically increasing order globally 4925 // and, a-fortiori, in monotonically increasing order per thread 4926 // (the latter order being a subsequence of the former). 4927 // If the work code below is ever reorganized into a more chaotic 4928 // work-partitioning form than the current "sequential tasks" 4929 // paradigm, the use of that persistent state will have to be 4930 // revisited and modified appropriately. See also related 4931 // bug 4756801 work on which should examine this code to make 4932 // sure that the changes there do not run counter to the 4933 // assumptions made here and necessary for correctness and 4934 // efficiency. Note also that this code might yield inefficient 4935 // behavior in the case of very large objects that span one or 4936 // more work chunks. Such objects would potentially be scanned 4937 // several times redundantly. Work on 4756801 should try and 4938 // address that performance anomaly if at all possible. XXX 4939 MemRegion full_span = _collector->_span; 4940 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4941 MarkFromDirtyCardsClosure 4942 greyRescanClosure(_collector, full_span, // entire span of interest 4943 sp, bm, work_q, cl); 4944 4945 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4946 assert(pst->valid(), "Uninitialized use?"); 4947 uint nth_task = 0; 4948 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4949 MemRegion span = sp->used_region(); 4950 HeapWord* start_addr = span.start(); 4951 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4952 alignment); 4953 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4954 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4955 start_addr, "Check alignment"); 4956 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4957 chunk_size, "Check alignment"); 4958 4959 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4960 // Having claimed the nth_task, compute corresponding mem-region, 4961 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4962 // The alignment restriction ensures that we do not need any 4963 // synchronization with other gang-workers while setting or 4964 // clearing bits in thus chunk of the MUT. 4965 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4966 start_addr + (nth_task+1)*chunk_size); 4967 // The last chunk's end might be way beyond end of the 4968 // used region. In that case pull back appropriately. 4969 if (this_span.end() > end_addr) { 4970 this_span.set_end(end_addr); 4971 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4972 } 4973 // Iterate over the dirty cards covering this chunk, marking them 4974 // precleaned, and setting the corresponding bits in the mod union 4975 // table. Since we have been careful to partition at Card and MUT-word 4976 // boundaries no synchronization is needed between parallel threads. 4977 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4978 &modUnionClosure); 4979 4980 // Having transferred these marks into the modUnionTable, 4981 // rescan the marked objects on the dirty cards in the modUnionTable. 4982 // Even if this is at a synchronous collection, the initial marking 4983 // may have been done during an asynchronous collection so there 4984 // may be dirty bits in the mod-union table. 4985 _collector->_modUnionTable.dirty_range_iterate_clear( 4986 this_span, &greyRescanClosure); 4987 _collector->_modUnionTable.verifyNoOneBitsInRange( 4988 this_span.start(), 4989 this_span.end()); 4990 } 4991 pst->all_tasks_completed(); // declare that i am done 4992 } 4993 4994 // . see if we can share work_queues with ParNew? XXX 4995 void 4996 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4997 int* seed) { 4998 OopTaskQueue* work_q = work_queue(i); 4999 NOT_PRODUCT(int num_steals = 0;) 5000 oop obj_to_scan; 5001 CMSBitMap* bm = &(_collector->_markBitMap); 5002 5003 while (true) { 5004 // Completely finish any left over work from (an) earlier round(s) 5005 cl->trim_queue(0); 5006 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5007 (size_t)ParGCDesiredObjsFromOverflowList); 5008 // Now check if there's any work in the overflow list 5009 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5010 // only affects the number of attempts made to get work from the 5011 // overflow list and does not affect the number of workers. Just 5012 // pass ParallelGCThreads so this behavior is unchanged. 5013 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5014 work_q, 5015 ParallelGCThreads)) { 5016 // found something in global overflow list; 5017 // not yet ready to go stealing work from others. 5018 // We'd like to assert(work_q->size() != 0, ...) 5019 // because we just took work from the overflow list, 5020 // but of course we can't since all of that could have 5021 // been already stolen from us. 5022 // "He giveth and He taketh away." 5023 continue; 5024 } 5025 // Verify that we have no work before we resort to stealing 5026 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5027 // Try to steal from other queues that have work 5028 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5029 NOT_PRODUCT(num_steals++;) 5030 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5031 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5032 // Do scanning work 5033 obj_to_scan->oop_iterate(cl); 5034 // Loop around, finish this work, and try to steal some more 5035 } else if (terminator()->offer_termination()) { 5036 break; // nirvana from the infinite cycle 5037 } 5038 } 5039 NOT_PRODUCT( 5040 if (PrintCMSStatistics != 0) { 5041 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5042 } 5043 ) 5044 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5045 "Else our work is not yet done"); 5046 } 5047 5048 // Record object boundaries in _eden_chunk_array by sampling the eden 5049 // top in the slow-path eden object allocation code path and record 5050 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5051 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5052 // sampling in sample_eden() that activates during the part of the 5053 // preclean phase. 5054 void CMSCollector::sample_eden_chunk() { 5055 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5056 if (_eden_chunk_lock->try_lock()) { 5057 // Record a sample. This is the critical section. The contents 5058 // of the _eden_chunk_array have to be non-decreasing in the 5059 // address order. 5060 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5061 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5062 "Unexpected state of Eden"); 5063 if (_eden_chunk_index == 0 || 5064 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5065 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5066 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5067 _eden_chunk_index++; // commit sample 5068 } 5069 _eden_chunk_lock->unlock(); 5070 } 5071 } 5072 } 5073 5074 // Return a thread-local PLAB recording array, as appropriate. 5075 void* CMSCollector::get_data_recorder(int thr_num) { 5076 if (_survivor_plab_array != NULL && 5077 (CMSPLABRecordAlways || 5078 (_collectorState > Marking && _collectorState < FinalMarking))) { 5079 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5080 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5081 ca->reset(); // clear it so that fresh data is recorded 5082 return (void*) ca; 5083 } else { 5084 return NULL; 5085 } 5086 } 5087 5088 // Reset all the thread-local PLAB recording arrays 5089 void CMSCollector::reset_survivor_plab_arrays() { 5090 for (uint i = 0; i < ParallelGCThreads; i++) { 5091 _survivor_plab_array[i].reset(); 5092 } 5093 } 5094 5095 // Merge the per-thread plab arrays into the global survivor chunk 5096 // array which will provide the partitioning of the survivor space 5097 // for CMS initial scan and rescan. 5098 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5099 int no_of_gc_threads) { 5100 assert(_survivor_plab_array != NULL, "Error"); 5101 assert(_survivor_chunk_array != NULL, "Error"); 5102 assert(_collectorState == FinalMarking || 5103 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5104 for (int j = 0; j < no_of_gc_threads; j++) { 5105 _cursor[j] = 0; 5106 } 5107 HeapWord* top = surv->top(); 5108 size_t i; 5109 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5110 HeapWord* min_val = top; // Higher than any PLAB address 5111 uint min_tid = 0; // position of min_val this round 5112 for (int j = 0; j < no_of_gc_threads; j++) { 5113 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5114 if (_cursor[j] == cur_sca->end()) { 5115 continue; 5116 } 5117 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5118 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5119 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5120 if (cur_val < min_val) { 5121 min_tid = j; 5122 min_val = cur_val; 5123 } else { 5124 assert(cur_val < top, "All recorded addresses should be less"); 5125 } 5126 } 5127 // At this point min_val and min_tid are respectively 5128 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5129 // and the thread (j) that witnesses that address. 5130 // We record this address in the _survivor_chunk_array[i] 5131 // and increment _cursor[min_tid] prior to the next round i. 5132 if (min_val == top) { 5133 break; 5134 } 5135 _survivor_chunk_array[i] = min_val; 5136 _cursor[min_tid]++; 5137 } 5138 // We are all done; record the size of the _survivor_chunk_array 5139 _survivor_chunk_index = i; // exclusive: [0, i) 5140 if (PrintCMSStatistics > 0) { 5141 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5142 } 5143 // Verify that we used up all the recorded entries 5144 #ifdef ASSERT 5145 size_t total = 0; 5146 for (int j = 0; j < no_of_gc_threads; j++) { 5147 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5148 total += _cursor[j]; 5149 } 5150 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5151 // Check that the merged array is in sorted order 5152 if (total > 0) { 5153 for (size_t i = 0; i < total - 1; i++) { 5154 if (PrintCMSStatistics > 0) { 5155 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5156 i, _survivor_chunk_array[i]); 5157 } 5158 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5159 "Not sorted"); 5160 } 5161 } 5162 #endif // ASSERT 5163 } 5164 5165 // Set up the space's par_seq_tasks structure for work claiming 5166 // for parallel initial scan and rescan of young gen. 5167 // See ParRescanTask where this is currently used. 5168 void 5169 CMSCollector:: 5170 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5171 assert(n_threads > 0, "Unexpected n_threads argument"); 5172 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5173 5174 // Eden space 5175 if (!dng->eden()->is_empty()) { 5176 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5177 assert(!pst->valid(), "Clobbering existing data?"); 5178 // Each valid entry in [0, _eden_chunk_index) represents a task. 5179 size_t n_tasks = _eden_chunk_index + 1; 5180 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5181 // Sets the condition for completion of the subtask (how many threads 5182 // need to finish in order to be done). 5183 pst->set_n_threads(n_threads); 5184 pst->set_n_tasks((int)n_tasks); 5185 } 5186 5187 // Merge the survivor plab arrays into _survivor_chunk_array 5188 if (_survivor_plab_array != NULL) { 5189 merge_survivor_plab_arrays(dng->from(), n_threads); 5190 } else { 5191 assert(_survivor_chunk_index == 0, "Error"); 5192 } 5193 5194 // To space 5195 { 5196 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5197 assert(!pst->valid(), "Clobbering existing data?"); 5198 // Sets the condition for completion of the subtask (how many threads 5199 // need to finish in order to be done). 5200 pst->set_n_threads(n_threads); 5201 pst->set_n_tasks(1); 5202 assert(pst->valid(), "Error"); 5203 } 5204 5205 // From space 5206 { 5207 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5208 assert(!pst->valid(), "Clobbering existing data?"); 5209 size_t n_tasks = _survivor_chunk_index + 1; 5210 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5211 // Sets the condition for completion of the subtask (how many threads 5212 // need to finish in order to be done). 5213 pst->set_n_threads(n_threads); 5214 pst->set_n_tasks((int)n_tasks); 5215 assert(pst->valid(), "Error"); 5216 } 5217 } 5218 5219 // Parallel version of remark 5220 void CMSCollector::do_remark_parallel() { 5221 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5222 FlexibleWorkGang* workers = gch->workers(); 5223 assert(workers != NULL, "Need parallel worker threads."); 5224 // Choose to use the number of GC workers most recently set 5225 // into "active_workers". If active_workers is not set, set it 5226 // to ParallelGCThreads. 5227 int n_workers = workers->active_workers(); 5228 if (n_workers == 0) { 5229 assert(n_workers > 0, "Should have been set during scavenge"); 5230 n_workers = ParallelGCThreads; 5231 workers->set_active_workers(n_workers); 5232 } 5233 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5234 5235 CMSParRemarkTask tsk(this, 5236 cms_space, 5237 n_workers, workers, task_queues()); 5238 5239 // Set up for parallel process_roots work. 5240 gch->set_par_threads(n_workers); 5241 // We won't be iterating over the cards in the card table updating 5242 // the younger_gen cards, so we shouldn't call the following else 5243 // the verification code as well as subsequent younger_refs_iterate 5244 // code would get confused. XXX 5245 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5246 5247 // The young gen rescan work will not be done as part of 5248 // process_roots (which currently doesn't know how to 5249 // parallelize such a scan), but rather will be broken up into 5250 // a set of parallel tasks (via the sampling that the [abortable] 5251 // preclean phase did of eden, plus the [two] tasks of 5252 // scanning the [two] survivor spaces. Further fine-grain 5253 // parallelization of the scanning of the survivor spaces 5254 // themselves, and of precleaning of the younger gen itself 5255 // is deferred to the future. 5256 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5257 5258 // The dirty card rescan work is broken up into a "sequence" 5259 // of parallel tasks (per constituent space) that are dynamically 5260 // claimed by the parallel threads. 5261 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5262 5263 // It turns out that even when we're using 1 thread, doing the work in a 5264 // separate thread causes wide variance in run times. We can't help this 5265 // in the multi-threaded case, but we special-case n=1 here to get 5266 // repeatable measurements of the 1-thread overhead of the parallel code. 5267 if (n_workers > 1) { 5268 // Make refs discovery MT-safe, if it isn't already: it may not 5269 // necessarily be so, since it's possible that we are doing 5270 // ST marking. 5271 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5272 GenCollectedHeap::StrongRootsScope srs(gch); 5273 workers->run_task(&tsk); 5274 } else { 5275 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5276 GenCollectedHeap::StrongRootsScope srs(gch); 5277 tsk.work(0); 5278 } 5279 5280 gch->set_par_threads(0); // 0 ==> non-parallel. 5281 // restore, single-threaded for now, any preserved marks 5282 // as a result of work_q overflow 5283 restore_preserved_marks_if_any(); 5284 } 5285 5286 // Non-parallel version of remark 5287 void CMSCollector::do_remark_non_parallel() { 5288 ResourceMark rm; 5289 HandleMark hm; 5290 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5291 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5292 5293 MarkRefsIntoAndScanClosure 5294 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5295 &_markStack, this, 5296 false /* should_yield */, false /* not precleaning */); 5297 MarkFromDirtyCardsClosure 5298 markFromDirtyCardsClosure(this, _span, 5299 NULL, // space is set further below 5300 &_markBitMap, &_markStack, &mrias_cl); 5301 { 5302 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5303 // Iterate over the dirty cards, setting the corresponding bits in the 5304 // mod union table. 5305 { 5306 ModUnionClosure modUnionClosure(&_modUnionTable); 5307 _ct->ct_bs()->dirty_card_iterate( 5308 _cmsGen->used_region(), 5309 &modUnionClosure); 5310 } 5311 // Having transferred these marks into the modUnionTable, we just need 5312 // to rescan the marked objects on the dirty cards in the modUnionTable. 5313 // The initial marking may have been done during an asynchronous 5314 // collection so there may be dirty bits in the mod-union table. 5315 const int alignment = 5316 CardTableModRefBS::card_size * BitsPerWord; 5317 { 5318 // ... First handle dirty cards in CMS gen 5319 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5320 MemRegion ur = _cmsGen->used_region(); 5321 HeapWord* lb = ur.start(); 5322 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5323 MemRegion cms_span(lb, ub); 5324 _modUnionTable.dirty_range_iterate_clear(cms_span, 5325 &markFromDirtyCardsClosure); 5326 verify_work_stacks_empty(); 5327 if (PrintCMSStatistics != 0) { 5328 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5329 markFromDirtyCardsClosure.num_dirty_cards()); 5330 } 5331 } 5332 } 5333 if (VerifyDuringGC && 5334 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5335 HandleMark hm; // Discard invalid handles created during verification 5336 Universe::verify(); 5337 } 5338 { 5339 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5340 5341 verify_work_stacks_empty(); 5342 5343 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5344 GenCollectedHeap::StrongRootsScope srs(gch); 5345 5346 gch->gen_process_roots(_cmsGen->level(), 5347 true, // younger gens as roots 5348 false, // use the local StrongRootsScope 5349 SharedHeap::ScanningOption(roots_scanning_options()), 5350 should_unload_classes(), 5351 &mrias_cl, 5352 NULL, 5353 NULL); // The dirty klasses will be handled below 5354 5355 assert(should_unload_classes() 5356 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5357 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5358 } 5359 5360 { 5361 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5362 5363 verify_work_stacks_empty(); 5364 5365 // Scan all class loader data objects that might have been introduced 5366 // during concurrent marking. 5367 ResourceMark rm; 5368 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5369 for (int i = 0; i < array->length(); i++) { 5370 mrias_cl.do_class_loader_data(array->at(i)); 5371 } 5372 5373 // We don't need to keep track of new CLDs anymore. 5374 ClassLoaderDataGraph::remember_new_clds(false); 5375 5376 verify_work_stacks_empty(); 5377 } 5378 5379 { 5380 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5381 5382 verify_work_stacks_empty(); 5383 5384 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5385 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5386 5387 verify_work_stacks_empty(); 5388 } 5389 5390 // We might have added oops to ClassLoaderData::_handles during the 5391 // concurrent marking phase. These oops point to newly allocated objects 5392 // that are guaranteed to be kept alive. Either by the direct allocation 5393 // code, or when the young collector processes the roots. Hence, 5394 // we don't have to revisit the _handles block during the remark phase. 5395 5396 verify_work_stacks_empty(); 5397 // Restore evacuated mark words, if any, used for overflow list links 5398 if (!CMSOverflowEarlyRestoration) { 5399 restore_preserved_marks_if_any(); 5400 } 5401 verify_overflow_empty(); 5402 } 5403 5404 //////////////////////////////////////////////////////// 5405 // Parallel Reference Processing Task Proxy Class 5406 //////////////////////////////////////////////////////// 5407 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5408 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5409 CMSCollector* _collector; 5410 CMSBitMap* _mark_bit_map; 5411 const MemRegion _span; 5412 ProcessTask& _task; 5413 5414 public: 5415 CMSRefProcTaskProxy(ProcessTask& task, 5416 CMSCollector* collector, 5417 const MemRegion& span, 5418 CMSBitMap* mark_bit_map, 5419 AbstractWorkGang* workers, 5420 OopTaskQueueSet* task_queues): 5421 // XXX Should superclass AGTWOQ also know about AWG since it knows 5422 // about the task_queues used by the AWG? Then it could initialize 5423 // the terminator() object. See 6984287. The set_for_termination() 5424 // below is a temporary band-aid for the regression in 6984287. 5425 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5426 task_queues), 5427 _task(task), 5428 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5429 { 5430 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5431 "Inconsistency in _span"); 5432 set_for_termination(workers->active_workers()); 5433 } 5434 5435 OopTaskQueueSet* task_queues() { return queues(); } 5436 5437 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5438 5439 void do_work_steal(int i, 5440 CMSParDrainMarkingStackClosure* drain, 5441 CMSParKeepAliveClosure* keep_alive, 5442 int* seed); 5443 5444 virtual void work(uint worker_id); 5445 }; 5446 5447 void CMSRefProcTaskProxy::work(uint worker_id) { 5448 ResourceMark rm; 5449 HandleMark hm; 5450 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5451 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5452 _mark_bit_map, 5453 work_queue(worker_id)); 5454 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5455 _mark_bit_map, 5456 work_queue(worker_id)); 5457 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5458 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5459 if (_task.marks_oops_alive()) { 5460 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5461 _collector->hash_seed(worker_id)); 5462 } 5463 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5464 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5465 } 5466 5467 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5468 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5469 EnqueueTask& _task; 5470 5471 public: 5472 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5473 : AbstractGangTask("Enqueue reference objects in parallel"), 5474 _task(task) 5475 { } 5476 5477 virtual void work(uint worker_id) 5478 { 5479 _task.work(worker_id); 5480 } 5481 }; 5482 5483 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5484 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5485 _span(span), 5486 _bit_map(bit_map), 5487 _work_queue(work_queue), 5488 _mark_and_push(collector, span, bit_map, work_queue), 5489 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5490 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5491 { } 5492 5493 // . see if we can share work_queues with ParNew? XXX 5494 void CMSRefProcTaskProxy::do_work_steal(int i, 5495 CMSParDrainMarkingStackClosure* drain, 5496 CMSParKeepAliveClosure* keep_alive, 5497 int* seed) { 5498 OopTaskQueue* work_q = work_queue(i); 5499 NOT_PRODUCT(int num_steals = 0;) 5500 oop obj_to_scan; 5501 5502 while (true) { 5503 // Completely finish any left over work from (an) earlier round(s) 5504 drain->trim_queue(0); 5505 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5506 (size_t)ParGCDesiredObjsFromOverflowList); 5507 // Now check if there's any work in the overflow list 5508 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5509 // only affects the number of attempts made to get work from the 5510 // overflow list and does not affect the number of workers. Just 5511 // pass ParallelGCThreads so this behavior is unchanged. 5512 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5513 work_q, 5514 ParallelGCThreads)) { 5515 // Found something in global overflow list; 5516 // not yet ready to go stealing work from others. 5517 // We'd like to assert(work_q->size() != 0, ...) 5518 // because we just took work from the overflow list, 5519 // but of course we can't, since all of that might have 5520 // been already stolen from us. 5521 continue; 5522 } 5523 // Verify that we have no work before we resort to stealing 5524 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5525 // Try to steal from other queues that have work 5526 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5527 NOT_PRODUCT(num_steals++;) 5528 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5529 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5530 // Do scanning work 5531 obj_to_scan->oop_iterate(keep_alive); 5532 // Loop around, finish this work, and try to steal some more 5533 } else if (terminator()->offer_termination()) { 5534 break; // nirvana from the infinite cycle 5535 } 5536 } 5537 NOT_PRODUCT( 5538 if (PrintCMSStatistics != 0) { 5539 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5540 } 5541 ) 5542 } 5543 5544 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5545 { 5546 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5547 FlexibleWorkGang* workers = gch->workers(); 5548 assert(workers != NULL, "Need parallel worker threads."); 5549 CMSRefProcTaskProxy rp_task(task, &_collector, 5550 _collector.ref_processor()->span(), 5551 _collector.markBitMap(), 5552 workers, _collector.task_queues()); 5553 workers->run_task(&rp_task); 5554 } 5555 5556 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5557 { 5558 5559 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5560 FlexibleWorkGang* workers = gch->workers(); 5561 assert(workers != NULL, "Need parallel worker threads."); 5562 CMSRefEnqueueTaskProxy enq_task(task); 5563 workers->run_task(&enq_task); 5564 } 5565 5566 void CMSCollector::refProcessingWork() { 5567 ResourceMark rm; 5568 HandleMark hm; 5569 5570 ReferenceProcessor* rp = ref_processor(); 5571 assert(rp->span().equals(_span), "Spans should be equal"); 5572 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5573 // Process weak references. 5574 rp->setup_policy(false); 5575 verify_work_stacks_empty(); 5576 5577 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5578 &_markStack, false /* !preclean */); 5579 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5580 _span, &_markBitMap, &_markStack, 5581 &cmsKeepAliveClosure, false /* !preclean */); 5582 { 5583 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5584 5585 ReferenceProcessorStats stats; 5586 if (rp->processing_is_mt()) { 5587 // Set the degree of MT here. If the discovery is done MT, there 5588 // may have been a different number of threads doing the discovery 5589 // and a different number of discovered lists may have Ref objects. 5590 // That is OK as long as the Reference lists are balanced (see 5591 // balance_all_queues() and balance_queues()). 5592 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5593 int active_workers = ParallelGCThreads; 5594 FlexibleWorkGang* workers = gch->workers(); 5595 if (workers != NULL) { 5596 active_workers = workers->active_workers(); 5597 // The expectation is that active_workers will have already 5598 // been set to a reasonable value. If it has not been set, 5599 // investigate. 5600 assert(active_workers > 0, "Should have been set during scavenge"); 5601 } 5602 rp->set_active_mt_degree(active_workers); 5603 CMSRefProcTaskExecutor task_executor(*this); 5604 stats = rp->process_discovered_references(&_is_alive_closure, 5605 &cmsKeepAliveClosure, 5606 &cmsDrainMarkingStackClosure, 5607 &task_executor, 5608 _gc_timer_cm, 5609 _gc_tracer_cm->gc_id()); 5610 } else { 5611 stats = rp->process_discovered_references(&_is_alive_closure, 5612 &cmsKeepAliveClosure, 5613 &cmsDrainMarkingStackClosure, 5614 NULL, 5615 _gc_timer_cm, 5616 _gc_tracer_cm->gc_id()); 5617 } 5618 _gc_tracer_cm->report_gc_reference_stats(stats); 5619 5620 } 5621 5622 // This is the point where the entire marking should have completed. 5623 verify_work_stacks_empty(); 5624 5625 if (should_unload_classes()) { 5626 { 5627 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5628 5629 // Unload classes and purge the SystemDictionary. 5630 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5631 5632 // Unload nmethods. 5633 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5634 5635 // Prune dead klasses from subklass/sibling/implementor lists. 5636 Klass::clean_weak_klass_links(&_is_alive_closure); 5637 } 5638 5639 { 5640 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5641 // Clean up unreferenced symbols in symbol table. 5642 SymbolTable::unlink(); 5643 } 5644 5645 { 5646 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5647 // Delete entries for dead interned strings. 5648 StringTable::unlink(&_is_alive_closure); 5649 } 5650 } 5651 5652 5653 // Restore any preserved marks as a result of mark stack or 5654 // work queue overflow 5655 restore_preserved_marks_if_any(); // done single-threaded for now 5656 5657 rp->set_enqueuing_is_done(true); 5658 if (rp->processing_is_mt()) { 5659 rp->balance_all_queues(); 5660 CMSRefProcTaskExecutor task_executor(*this); 5661 rp->enqueue_discovered_references(&task_executor); 5662 } else { 5663 rp->enqueue_discovered_references(NULL); 5664 } 5665 rp->verify_no_references_recorded(); 5666 assert(!rp->discovery_enabled(), "should have been disabled"); 5667 } 5668 5669 #ifndef PRODUCT 5670 void CMSCollector::check_correct_thread_executing() { 5671 Thread* t = Thread::current(); 5672 // Only the VM thread or the CMS thread should be here. 5673 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5674 "Unexpected thread type"); 5675 // If this is the vm thread, the foreground process 5676 // should not be waiting. Note that _foregroundGCIsActive is 5677 // true while the foreground collector is waiting. 5678 if (_foregroundGCShouldWait) { 5679 // We cannot be the VM thread 5680 assert(t->is_ConcurrentGC_thread(), 5681 "Should be CMS thread"); 5682 } else { 5683 // We can be the CMS thread only if we are in a stop-world 5684 // phase of CMS collection. 5685 if (t->is_ConcurrentGC_thread()) { 5686 assert(_collectorState == InitialMarking || 5687 _collectorState == FinalMarking, 5688 "Should be a stop-world phase"); 5689 // The CMS thread should be holding the CMS_token. 5690 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5691 "Potential interference with concurrently " 5692 "executing VM thread"); 5693 } 5694 } 5695 } 5696 #endif 5697 5698 void CMSCollector::sweep() { 5699 assert(_collectorState == Sweeping, "just checking"); 5700 check_correct_thread_executing(); 5701 verify_work_stacks_empty(); 5702 verify_overflow_empty(); 5703 increment_sweep_count(); 5704 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5705 5706 _inter_sweep_timer.stop(); 5707 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5708 5709 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5710 _intra_sweep_timer.reset(); 5711 _intra_sweep_timer.start(); 5712 { 5713 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5714 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5715 // First sweep the old gen 5716 { 5717 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5718 bitMapLock()); 5719 sweepWork(_cmsGen); 5720 } 5721 5722 // Update Universe::_heap_*_at_gc figures. 5723 // We need all the free list locks to make the abstract state 5724 // transition from Sweeping to Resetting. See detailed note 5725 // further below. 5726 { 5727 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5728 // Update heap occupancy information which is used as 5729 // input to soft ref clearing policy at the next gc. 5730 Universe::update_heap_info_at_gc(); 5731 _collectorState = Resizing; 5732 } 5733 } 5734 verify_work_stacks_empty(); 5735 verify_overflow_empty(); 5736 5737 if (should_unload_classes()) { 5738 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5739 // requires that the virtual spaces are stable and not deleted. 5740 ClassLoaderDataGraph::set_should_purge(true); 5741 } 5742 5743 _intra_sweep_timer.stop(); 5744 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5745 5746 _inter_sweep_timer.reset(); 5747 _inter_sweep_timer.start(); 5748 5749 // We need to use a monotonically non-decreasing time in ms 5750 // or we will see time-warp warnings and os::javaTimeMillis() 5751 // does not guarantee monotonicity. 5752 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5753 update_time_of_last_gc(now); 5754 5755 // NOTE on abstract state transitions: 5756 // Mutators allocate-live and/or mark the mod-union table dirty 5757 // based on the state of the collection. The former is done in 5758 // the interval [Marking, Sweeping] and the latter in the interval 5759 // [Marking, Sweeping). Thus the transitions into the Marking state 5760 // and out of the Sweeping state must be synchronously visible 5761 // globally to the mutators. 5762 // The transition into the Marking state happens with the world 5763 // stopped so the mutators will globally see it. Sweeping is 5764 // done asynchronously by the background collector so the transition 5765 // from the Sweeping state to the Resizing state must be done 5766 // under the freelistLock (as is the check for whether to 5767 // allocate-live and whether to dirty the mod-union table). 5768 assert(_collectorState == Resizing, "Change of collector state to" 5769 " Resizing must be done under the freelistLocks (plural)"); 5770 5771 // Now that sweeping has been completed, we clear 5772 // the incremental_collection_failed flag, 5773 // thus inviting a younger gen collection to promote into 5774 // this generation. If such a promotion may still fail, 5775 // the flag will be set again when a young collection is 5776 // attempted. 5777 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5778 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5779 gch->update_full_collections_completed(_collection_count_start); 5780 } 5781 5782 // FIX ME!!! Looks like this belongs in CFLSpace, with 5783 // CMSGen merely delegating to it. 5784 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5785 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5786 HeapWord* minAddr = _cmsSpace->bottom(); 5787 HeapWord* largestAddr = 5788 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5789 if (largestAddr == NULL) { 5790 // The dictionary appears to be empty. In this case 5791 // try to coalesce at the end of the heap. 5792 largestAddr = _cmsSpace->end(); 5793 } 5794 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5795 size_t nearLargestOffset = 5796 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5797 if (PrintFLSStatistics != 0) { 5798 gclog_or_tty->print_cr( 5799 "CMS: Large Block: " PTR_FORMAT ";" 5800 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5801 largestAddr, 5802 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 5803 } 5804 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5805 } 5806 5807 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5808 return addr >= _cmsSpace->nearLargestChunk(); 5809 } 5810 5811 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5812 return _cmsSpace->find_chunk_at_end(); 5813 } 5814 5815 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5816 bool full) { 5817 // The next lower level has been collected. Gather any statistics 5818 // that are of interest at this point. 5819 if (!full && (current_level + 1) == level()) { 5820 // Gather statistics on the young generation collection. 5821 collector()->stats().record_gc0_end(used()); 5822 } 5823 } 5824 5825 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 5826 if (PrintGCDetails && Verbose) { 5827 if (_debug_concurrent_cycle) { 5828 gclog_or_tty->print_cr("Rotate from concurrent to STW collections"); 5829 } else { 5830 gclog_or_tty->print_cr("Rotate from STW to concurrent collections"); 5831 } 5832 } 5833 _debug_concurrent_cycle = !_debug_concurrent_cycle; 5834 } 5835 5836 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) { 5837 // We iterate over the space(s) underlying this generation, 5838 // checking the mark bit map to see if the bits corresponding 5839 // to specific blocks are marked or not. Blocks that are 5840 // marked are live and are not swept up. All remaining blocks 5841 // are swept up, with coalescing on-the-fly as we sweep up 5842 // contiguous free and/or garbage blocks: 5843 // We need to ensure that the sweeper synchronizes with allocators 5844 // and stop-the-world collectors. In particular, the following 5845 // locks are used: 5846 // . CMS token: if this is held, a stop the world collection cannot occur 5847 // . freelistLock: if this is held no allocation can occur from this 5848 // generation by another thread 5849 // . bitMapLock: if this is held, no other thread can access or update 5850 // 5851 5852 // Note that we need to hold the freelistLock if we use 5853 // block iterate below; else the iterator might go awry if 5854 // a mutator (or promotion) causes block contents to change 5855 // (for instance if the allocator divvies up a block). 5856 // If we hold the free list lock, for all practical purposes 5857 // young generation GC's can't occur (they'll usually need to 5858 // promote), so we might as well prevent all young generation 5859 // GC's while we do a sweeping step. For the same reason, we might 5860 // as well take the bit map lock for the entire duration 5861 5862 // check that we hold the requisite locks 5863 assert(have_cms_token(), "Should hold cms token"); 5864 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5865 assert_lock_strong(gen->freelistLock()); 5866 assert_lock_strong(bitMapLock()); 5867 5868 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5869 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5870 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5871 _inter_sweep_estimate.padded_average(), 5872 _intra_sweep_estimate.padded_average()); 5873 gen->setNearLargestChunk(); 5874 5875 { 5876 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield); 5877 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5878 // We need to free-up/coalesce garbage/blocks from a 5879 // co-terminal free run. This is done in the SweepClosure 5880 // destructor; so, do not remove this scope, else the 5881 // end-of-sweep-census below will be off by a little bit. 5882 } 5883 gen->cmsSpace()->sweep_completed(); 5884 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5885 if (should_unload_classes()) { // unloaded classes this cycle, 5886 _concurrent_cycles_since_last_unload = 0; // ... reset count 5887 } else { // did not unload classes, 5888 _concurrent_cycles_since_last_unload++; // ... increment count 5889 } 5890 } 5891 5892 // Reset CMS data structures (for now just the marking bit map) 5893 // preparatory for the next cycle. 5894 void CMSCollector::reset(bool concurrent) { 5895 if (concurrent) { 5896 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5897 5898 // If the state is not "Resetting", the foreground thread 5899 // has done a collection and the resetting. 5900 if (_collectorState != Resetting) { 5901 assert(_collectorState == Idling, "The state should only change" 5902 " because the foreground collector has finished the collection"); 5903 return; 5904 } 5905 5906 // Clear the mark bitmap (no grey objects to start with) 5907 // for the next cycle. 5908 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5909 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 5910 5911 HeapWord* curAddr = _markBitMap.startWord(); 5912 while (curAddr < _markBitMap.endWord()) { 5913 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5914 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5915 _markBitMap.clear_large_range(chunk); 5916 if (ConcurrentMarkSweepThread::should_yield() && 5917 !foregroundGCIsActive() && 5918 CMSYield) { 5919 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5920 "CMS thread should hold CMS token"); 5921 assert_lock_strong(bitMapLock()); 5922 bitMapLock()->unlock(); 5923 ConcurrentMarkSweepThread::desynchronize(true); 5924 stopTimer(); 5925 if (PrintCMSStatistics != 0) { 5926 incrementYields(); 5927 } 5928 5929 // See the comment in coordinator_yield() 5930 for (unsigned i = 0; i < CMSYieldSleepCount && 5931 ConcurrentMarkSweepThread::should_yield() && 5932 !CMSCollector::foregroundGCIsActive(); ++i) { 5933 os::sleep(Thread::current(), 1, false); 5934 } 5935 5936 ConcurrentMarkSweepThread::synchronize(true); 5937 bitMapLock()->lock_without_safepoint_check(); 5938 startTimer(); 5939 } 5940 curAddr = chunk.end(); 5941 } 5942 // A successful mostly concurrent collection has been done. 5943 // Because only the full (i.e., concurrent mode failure) collections 5944 // are being measured for gc overhead limits, clean the "near" flag 5945 // and count. 5946 size_policy()->reset_gc_overhead_limit_count(); 5947 _collectorState = Idling; 5948 } else { 5949 // already have the lock 5950 assert(_collectorState == Resetting, "just checking"); 5951 assert_lock_strong(bitMapLock()); 5952 _markBitMap.clear_all(); 5953 _collectorState = Idling; 5954 } 5955 5956 NOT_PRODUCT( 5957 if (RotateCMSCollectionTypes) { 5958 _cmsGen->rotate_debug_collection_type(); 5959 } 5960 ) 5961 5962 register_gc_end(); 5963 } 5964 5965 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5966 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 5967 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5968 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 5969 TraceCollectorStats tcs(counters()); 5970 5971 switch (op) { 5972 case CMS_op_checkpointRootsInitial: { 5973 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5974 checkpointRootsInitial(); 5975 if (PrintGC) { 5976 _cmsGen->printOccupancy("initial-mark"); 5977 } 5978 break; 5979 } 5980 case CMS_op_checkpointRootsFinal: { 5981 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5982 checkpointRootsFinal(); 5983 if (PrintGC) { 5984 _cmsGen->printOccupancy("remark"); 5985 } 5986 break; 5987 } 5988 default: 5989 fatal("No such CMS_op"); 5990 } 5991 } 5992 5993 #ifndef PRODUCT 5994 size_t const CMSCollector::skip_header_HeapWords() { 5995 return FreeChunk::header_size(); 5996 } 5997 5998 // Try and collect here conditions that should hold when 5999 // CMS thread is exiting. The idea is that the foreground GC 6000 // thread should not be blocked if it wants to terminate 6001 // the CMS thread and yet continue to run the VM for a while 6002 // after that. 6003 void CMSCollector::verify_ok_to_terminate() const { 6004 assert(Thread::current()->is_ConcurrentGC_thread(), 6005 "should be called by CMS thread"); 6006 assert(!_foregroundGCShouldWait, "should be false"); 6007 // We could check here that all the various low-level locks 6008 // are not held by the CMS thread, but that is overkill; see 6009 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6010 // is checked. 6011 } 6012 #endif 6013 6014 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6015 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6016 "missing Printezis mark?"); 6017 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6018 size_t size = pointer_delta(nextOneAddr + 1, addr); 6019 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6020 "alignment problem"); 6021 assert(size >= 3, "Necessary for Printezis marks to work"); 6022 return size; 6023 } 6024 6025 // A variant of the above (block_size_using_printezis_bits()) except 6026 // that we return 0 if the P-bits are not yet set. 6027 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6028 if (_markBitMap.isMarked(addr + 1)) { 6029 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6030 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6031 size_t size = pointer_delta(nextOneAddr + 1, addr); 6032 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6033 "alignment problem"); 6034 assert(size >= 3, "Necessary for Printezis marks to work"); 6035 return size; 6036 } 6037 return 0; 6038 } 6039 6040 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6041 size_t sz = 0; 6042 oop p = (oop)addr; 6043 if (p->klass_or_null() != NULL) { 6044 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6045 } else { 6046 sz = block_size_using_printezis_bits(addr); 6047 } 6048 assert(sz > 0, "size must be nonzero"); 6049 HeapWord* next_block = addr + sz; 6050 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6051 CardTableModRefBS::card_size); 6052 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6053 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6054 "must be different cards"); 6055 return next_card; 6056 } 6057 6058 6059 // CMS Bit Map Wrapper ///////////////////////////////////////// 6060 6061 // Construct a CMS bit map infrastructure, but don't create the 6062 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6063 // further below. 6064 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6065 _bm(), 6066 _shifter(shifter), 6067 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6068 { 6069 _bmStartWord = 0; 6070 _bmWordSize = 0; 6071 } 6072 6073 bool CMSBitMap::allocate(MemRegion mr) { 6074 _bmStartWord = mr.start(); 6075 _bmWordSize = mr.word_size(); 6076 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6077 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6078 if (!brs.is_reserved()) { 6079 warning("CMS bit map allocation failure"); 6080 return false; 6081 } 6082 // For now we'll just commit all of the bit map up front. 6083 // Later on we'll try to be more parsimonious with swap. 6084 if (!_virtual_space.initialize(brs, brs.size())) { 6085 warning("CMS bit map backing store failure"); 6086 return false; 6087 } 6088 assert(_virtual_space.committed_size() == brs.size(), 6089 "didn't reserve backing store for all of CMS bit map?"); 6090 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6091 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6092 _bmWordSize, "inconsistency in bit map sizing"); 6093 _bm.set_size(_bmWordSize >> _shifter); 6094 6095 // bm.clear(); // can we rely on getting zero'd memory? verify below 6096 assert(isAllClear(), 6097 "Expected zero'd memory from ReservedSpace constructor"); 6098 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6099 "consistency check"); 6100 return true; 6101 } 6102 6103 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6104 HeapWord *next_addr, *end_addr, *last_addr; 6105 assert_locked(); 6106 assert(covers(mr), "out-of-range error"); 6107 // XXX assert that start and end are appropriately aligned 6108 for (next_addr = mr.start(), end_addr = mr.end(); 6109 next_addr < end_addr; next_addr = last_addr) { 6110 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6111 last_addr = dirty_region.end(); 6112 if (!dirty_region.is_empty()) { 6113 cl->do_MemRegion(dirty_region); 6114 } else { 6115 assert(last_addr == end_addr, "program logic"); 6116 return; 6117 } 6118 } 6119 } 6120 6121 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6122 _bm.print_on_error(st, prefix); 6123 } 6124 6125 #ifndef PRODUCT 6126 void CMSBitMap::assert_locked() const { 6127 CMSLockVerifier::assert_locked(lock()); 6128 } 6129 6130 bool CMSBitMap::covers(MemRegion mr) const { 6131 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6132 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6133 "size inconsistency"); 6134 return (mr.start() >= _bmStartWord) && 6135 (mr.end() <= endWord()); 6136 } 6137 6138 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6139 return (start >= _bmStartWord && (start + size) <= endWord()); 6140 } 6141 6142 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6143 // verify that there are no 1 bits in the interval [left, right) 6144 FalseBitMapClosure falseBitMapClosure; 6145 iterate(&falseBitMapClosure, left, right); 6146 } 6147 6148 void CMSBitMap::region_invariant(MemRegion mr) 6149 { 6150 assert_locked(); 6151 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6152 assert(!mr.is_empty(), "unexpected empty region"); 6153 assert(covers(mr), "mr should be covered by bit map"); 6154 // convert address range into offset range 6155 size_t start_ofs = heapWordToOffset(mr.start()); 6156 // Make sure that end() is appropriately aligned 6157 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6158 (1 << (_shifter+LogHeapWordSize))), 6159 "Misaligned mr.end()"); 6160 size_t end_ofs = heapWordToOffset(mr.end()); 6161 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6162 } 6163 6164 #endif 6165 6166 bool CMSMarkStack::allocate(size_t size) { 6167 // allocate a stack of the requisite depth 6168 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6169 size * sizeof(oop))); 6170 if (!rs.is_reserved()) { 6171 warning("CMSMarkStack allocation failure"); 6172 return false; 6173 } 6174 if (!_virtual_space.initialize(rs, rs.size())) { 6175 warning("CMSMarkStack backing store failure"); 6176 return false; 6177 } 6178 assert(_virtual_space.committed_size() == rs.size(), 6179 "didn't reserve backing store for all of CMS stack?"); 6180 _base = (oop*)(_virtual_space.low()); 6181 _index = 0; 6182 _capacity = size; 6183 NOT_PRODUCT(_max_depth = 0); 6184 return true; 6185 } 6186 6187 // XXX FIX ME !!! In the MT case we come in here holding a 6188 // leaf lock. For printing we need to take a further lock 6189 // which has lower rank. We need to recalibrate the two 6190 // lock-ranks involved in order to be able to print the 6191 // messages below. (Or defer the printing to the caller. 6192 // For now we take the expedient path of just disabling the 6193 // messages for the problematic case.) 6194 void CMSMarkStack::expand() { 6195 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6196 if (_capacity == MarkStackSizeMax) { 6197 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6198 // We print a warning message only once per CMS cycle. 6199 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6200 } 6201 return; 6202 } 6203 // Double capacity if possible 6204 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6205 // Do not give up existing stack until we have managed to 6206 // get the double capacity that we desired. 6207 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6208 new_capacity * sizeof(oop))); 6209 if (rs.is_reserved()) { 6210 // Release the backing store associated with old stack 6211 _virtual_space.release(); 6212 // Reinitialize virtual space for new stack 6213 if (!_virtual_space.initialize(rs, rs.size())) { 6214 fatal("Not enough swap for expanded marking stack"); 6215 } 6216 _base = (oop*)(_virtual_space.low()); 6217 _index = 0; 6218 _capacity = new_capacity; 6219 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6220 // Failed to double capacity, continue; 6221 // we print a detail message only once per CMS cycle. 6222 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6223 SIZE_FORMAT"K", 6224 _capacity / K, new_capacity / K); 6225 } 6226 } 6227 6228 6229 // Closures 6230 // XXX: there seems to be a lot of code duplication here; 6231 // should refactor and consolidate common code. 6232 6233 // This closure is used to mark refs into the CMS generation in 6234 // the CMS bit map. Called at the first checkpoint. This closure 6235 // assumes that we do not need to re-mark dirty cards; if the CMS 6236 // generation on which this is used is not an oldest 6237 // generation then this will lose younger_gen cards! 6238 6239 MarkRefsIntoClosure::MarkRefsIntoClosure( 6240 MemRegion span, CMSBitMap* bitMap): 6241 _span(span), 6242 _bitMap(bitMap) 6243 { 6244 assert(_ref_processor == NULL, "deliberately left NULL"); 6245 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6246 } 6247 6248 void MarkRefsIntoClosure::do_oop(oop obj) { 6249 // if p points into _span, then mark corresponding bit in _markBitMap 6250 assert(obj->is_oop(), "expected an oop"); 6251 HeapWord* addr = (HeapWord*)obj; 6252 if (_span.contains(addr)) { 6253 // this should be made more efficient 6254 _bitMap->mark(addr); 6255 } 6256 } 6257 6258 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6259 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6260 6261 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6262 MemRegion span, CMSBitMap* bitMap): 6263 _span(span), 6264 _bitMap(bitMap) 6265 { 6266 assert(_ref_processor == NULL, "deliberately left NULL"); 6267 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6268 } 6269 6270 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6271 // if p points into _span, then mark corresponding bit in _markBitMap 6272 assert(obj->is_oop(), "expected an oop"); 6273 HeapWord* addr = (HeapWord*)obj; 6274 if (_span.contains(addr)) { 6275 // this should be made more efficient 6276 _bitMap->par_mark(addr); 6277 } 6278 } 6279 6280 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6281 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6282 6283 // A variant of the above, used for CMS marking verification. 6284 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6285 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6286 _span(span), 6287 _verification_bm(verification_bm), 6288 _cms_bm(cms_bm) 6289 { 6290 assert(_ref_processor == NULL, "deliberately left NULL"); 6291 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6292 } 6293 6294 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6295 // if p points into _span, then mark corresponding bit in _markBitMap 6296 assert(obj->is_oop(), "expected an oop"); 6297 HeapWord* addr = (HeapWord*)obj; 6298 if (_span.contains(addr)) { 6299 _verification_bm->mark(addr); 6300 if (!_cms_bm->isMarked(addr)) { 6301 oop(addr)->print(); 6302 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6303 fatal("... aborting"); 6304 } 6305 } 6306 } 6307 6308 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6309 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6310 6311 ////////////////////////////////////////////////// 6312 // MarkRefsIntoAndScanClosure 6313 ////////////////////////////////////////////////// 6314 6315 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6316 ReferenceProcessor* rp, 6317 CMSBitMap* bit_map, 6318 CMSBitMap* mod_union_table, 6319 CMSMarkStack* mark_stack, 6320 CMSCollector* collector, 6321 bool should_yield, 6322 bool concurrent_precleaning): 6323 _collector(collector), 6324 _span(span), 6325 _bit_map(bit_map), 6326 _mark_stack(mark_stack), 6327 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6328 mark_stack, concurrent_precleaning), 6329 _yield(should_yield), 6330 _concurrent_precleaning(concurrent_precleaning), 6331 _freelistLock(NULL) 6332 { 6333 _ref_processor = rp; 6334 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6335 } 6336 6337 // This closure is used to mark refs into the CMS generation at the 6338 // second (final) checkpoint, and to scan and transitively follow 6339 // the unmarked oops. It is also used during the concurrent precleaning 6340 // phase while scanning objects on dirty cards in the CMS generation. 6341 // The marks are made in the marking bit map and the marking stack is 6342 // used for keeping the (newly) grey objects during the scan. 6343 // The parallel version (Par_...) appears further below. 6344 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6345 if (obj != NULL) { 6346 assert(obj->is_oop(), "expected an oop"); 6347 HeapWord* addr = (HeapWord*)obj; 6348 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6349 assert(_collector->overflow_list_is_empty(), 6350 "overflow list should be empty"); 6351 if (_span.contains(addr) && 6352 !_bit_map->isMarked(addr)) { 6353 // mark bit map (object is now grey) 6354 _bit_map->mark(addr); 6355 // push on marking stack (stack should be empty), and drain the 6356 // stack by applying this closure to the oops in the oops popped 6357 // from the stack (i.e. blacken the grey objects) 6358 bool res = _mark_stack->push(obj); 6359 assert(res, "Should have space to push on empty stack"); 6360 do { 6361 oop new_oop = _mark_stack->pop(); 6362 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6363 assert(_bit_map->isMarked((HeapWord*)new_oop), 6364 "only grey objects on this stack"); 6365 // iterate over the oops in this oop, marking and pushing 6366 // the ones in CMS heap (i.e. in _span). 6367 new_oop->oop_iterate(&_pushAndMarkClosure); 6368 // check if it's time to yield 6369 do_yield_check(); 6370 } while (!_mark_stack->isEmpty() || 6371 (!_concurrent_precleaning && take_from_overflow_list())); 6372 // if marking stack is empty, and we are not doing this 6373 // during precleaning, then check the overflow list 6374 } 6375 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6376 assert(_collector->overflow_list_is_empty(), 6377 "overflow list was drained above"); 6378 // We could restore evacuated mark words, if any, used for 6379 // overflow list links here because the overflow list is 6380 // provably empty here. That would reduce the maximum 6381 // size requirements for preserved_{oop,mark}_stack. 6382 // But we'll just postpone it until we are all done 6383 // so we can just stream through. 6384 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6385 _collector->restore_preserved_marks_if_any(); 6386 assert(_collector->no_preserved_marks(), "No preserved marks"); 6387 } 6388 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6389 "All preserved marks should have been restored above"); 6390 } 6391 } 6392 6393 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6394 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6395 6396 void MarkRefsIntoAndScanClosure::do_yield_work() { 6397 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6398 "CMS thread should hold CMS token"); 6399 assert_lock_strong(_freelistLock); 6400 assert_lock_strong(_bit_map->lock()); 6401 // relinquish the free_list_lock and bitMaplock() 6402 _bit_map->lock()->unlock(); 6403 _freelistLock->unlock(); 6404 ConcurrentMarkSweepThread::desynchronize(true); 6405 _collector->stopTimer(); 6406 if (PrintCMSStatistics != 0) { 6407 _collector->incrementYields(); 6408 } 6409 6410 // See the comment in coordinator_yield() 6411 for (unsigned i = 0; 6412 i < CMSYieldSleepCount && 6413 ConcurrentMarkSweepThread::should_yield() && 6414 !CMSCollector::foregroundGCIsActive(); 6415 ++i) { 6416 os::sleep(Thread::current(), 1, false); 6417 } 6418 6419 ConcurrentMarkSweepThread::synchronize(true); 6420 _freelistLock->lock_without_safepoint_check(); 6421 _bit_map->lock()->lock_without_safepoint_check(); 6422 _collector->startTimer(); 6423 } 6424 6425 /////////////////////////////////////////////////////////// 6426 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6427 // MarkRefsIntoAndScanClosure 6428 /////////////////////////////////////////////////////////// 6429 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6430 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6431 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6432 _span(span), 6433 _bit_map(bit_map), 6434 _work_queue(work_queue), 6435 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6436 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6437 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6438 { 6439 _ref_processor = rp; 6440 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6441 } 6442 6443 // This closure is used to mark refs into the CMS generation at the 6444 // second (final) checkpoint, and to scan and transitively follow 6445 // the unmarked oops. The marks are made in the marking bit map and 6446 // the work_queue is used for keeping the (newly) grey objects during 6447 // the scan phase whence they are also available for stealing by parallel 6448 // threads. Since the marking bit map is shared, updates are 6449 // synchronized (via CAS). 6450 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6451 if (obj != NULL) { 6452 // Ignore mark word because this could be an already marked oop 6453 // that may be chained at the end of the overflow list. 6454 assert(obj->is_oop(true), "expected an oop"); 6455 HeapWord* addr = (HeapWord*)obj; 6456 if (_span.contains(addr) && 6457 !_bit_map->isMarked(addr)) { 6458 // mark bit map (object will become grey): 6459 // It is possible for several threads to be 6460 // trying to "claim" this object concurrently; 6461 // the unique thread that succeeds in marking the 6462 // object first will do the subsequent push on 6463 // to the work queue (or overflow list). 6464 if (_bit_map->par_mark(addr)) { 6465 // push on work_queue (which may not be empty), and trim the 6466 // queue to an appropriate length by applying this closure to 6467 // the oops in the oops popped from the stack (i.e. blacken the 6468 // grey objects) 6469 bool res = _work_queue->push(obj); 6470 assert(res, "Low water mark should be less than capacity?"); 6471 trim_queue(_low_water_mark); 6472 } // Else, another thread claimed the object 6473 } 6474 } 6475 } 6476 6477 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6478 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6479 6480 // This closure is used to rescan the marked objects on the dirty cards 6481 // in the mod union table and the card table proper. 6482 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6483 oop p, MemRegion mr) { 6484 6485 size_t size = 0; 6486 HeapWord* addr = (HeapWord*)p; 6487 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6488 assert(_span.contains(addr), "we are scanning the CMS generation"); 6489 // check if it's time to yield 6490 if (do_yield_check()) { 6491 // We yielded for some foreground stop-world work, 6492 // and we have been asked to abort this ongoing preclean cycle. 6493 return 0; 6494 } 6495 if (_bitMap->isMarked(addr)) { 6496 // it's marked; is it potentially uninitialized? 6497 if (p->klass_or_null() != NULL) { 6498 // an initialized object; ignore mark word in verification below 6499 // since we are running concurrent with mutators 6500 assert(p->is_oop(true), "should be an oop"); 6501 if (p->is_objArray()) { 6502 // objArrays are precisely marked; restrict scanning 6503 // to dirty cards only. 6504 size = CompactibleFreeListSpace::adjustObjectSize( 6505 p->oop_iterate(_scanningClosure, mr)); 6506 } else { 6507 // A non-array may have been imprecisely marked; we need 6508 // to scan object in its entirety. 6509 size = CompactibleFreeListSpace::adjustObjectSize( 6510 p->oop_iterate(_scanningClosure)); 6511 } 6512 #ifdef ASSERT 6513 size_t direct_size = 6514 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6515 assert(size == direct_size, "Inconsistency in size"); 6516 assert(size >= 3, "Necessary for Printezis marks to work"); 6517 if (!_bitMap->isMarked(addr+1)) { 6518 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6519 } else { 6520 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6521 assert(_bitMap->isMarked(addr+size-1), 6522 "inconsistent Printezis mark"); 6523 } 6524 #endif // ASSERT 6525 } else { 6526 // An uninitialized object. 6527 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6528 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6529 size = pointer_delta(nextOneAddr + 1, addr); 6530 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6531 "alignment problem"); 6532 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6533 // will dirty the card when the klass pointer is installed in the 6534 // object (signaling the completion of initialization). 6535 } 6536 } else { 6537 // Either a not yet marked object or an uninitialized object 6538 if (p->klass_or_null() == NULL) { 6539 // An uninitialized object, skip to the next card, since 6540 // we may not be able to read its P-bits yet. 6541 assert(size == 0, "Initial value"); 6542 } else { 6543 // An object not (yet) reached by marking: we merely need to 6544 // compute its size so as to go look at the next block. 6545 assert(p->is_oop(true), "should be an oop"); 6546 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6547 } 6548 } 6549 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6550 return size; 6551 } 6552 6553 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6554 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6555 "CMS thread should hold CMS token"); 6556 assert_lock_strong(_freelistLock); 6557 assert_lock_strong(_bitMap->lock()); 6558 // relinquish the free_list_lock and bitMaplock() 6559 _bitMap->lock()->unlock(); 6560 _freelistLock->unlock(); 6561 ConcurrentMarkSweepThread::desynchronize(true); 6562 _collector->stopTimer(); 6563 if (PrintCMSStatistics != 0) { 6564 _collector->incrementYields(); 6565 } 6566 6567 // See the comment in coordinator_yield() 6568 for (unsigned i = 0; i < CMSYieldSleepCount && 6569 ConcurrentMarkSweepThread::should_yield() && 6570 !CMSCollector::foregroundGCIsActive(); ++i) { 6571 os::sleep(Thread::current(), 1, false); 6572 } 6573 6574 ConcurrentMarkSweepThread::synchronize(true); 6575 _freelistLock->lock_without_safepoint_check(); 6576 _bitMap->lock()->lock_without_safepoint_check(); 6577 _collector->startTimer(); 6578 } 6579 6580 6581 ////////////////////////////////////////////////////////////////// 6582 // SurvivorSpacePrecleanClosure 6583 ////////////////////////////////////////////////////////////////// 6584 // This (single-threaded) closure is used to preclean the oops in 6585 // the survivor spaces. 6586 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6587 6588 HeapWord* addr = (HeapWord*)p; 6589 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6590 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6591 assert(p->klass_or_null() != NULL, "object should be initialized"); 6592 // an initialized object; ignore mark word in verification below 6593 // since we are running concurrent with mutators 6594 assert(p->is_oop(true), "should be an oop"); 6595 // Note that we do not yield while we iterate over 6596 // the interior oops of p, pushing the relevant ones 6597 // on our marking stack. 6598 size_t size = p->oop_iterate(_scanning_closure); 6599 do_yield_check(); 6600 // Observe that below, we do not abandon the preclean 6601 // phase as soon as we should; rather we empty the 6602 // marking stack before returning. This is to satisfy 6603 // some existing assertions. In general, it may be a 6604 // good idea to abort immediately and complete the marking 6605 // from the grey objects at a later time. 6606 while (!_mark_stack->isEmpty()) { 6607 oop new_oop = _mark_stack->pop(); 6608 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6609 assert(_bit_map->isMarked((HeapWord*)new_oop), 6610 "only grey objects on this stack"); 6611 // iterate over the oops in this oop, marking and pushing 6612 // the ones in CMS heap (i.e. in _span). 6613 new_oop->oop_iterate(_scanning_closure); 6614 // check if it's time to yield 6615 do_yield_check(); 6616 } 6617 unsigned int after_count = 6618 GenCollectedHeap::heap()->total_collections(); 6619 bool abort = (_before_count != after_count) || 6620 _collector->should_abort_preclean(); 6621 return abort ? 0 : size; 6622 } 6623 6624 void SurvivorSpacePrecleanClosure::do_yield_work() { 6625 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6626 "CMS thread should hold CMS token"); 6627 assert_lock_strong(_bit_map->lock()); 6628 // Relinquish the bit map lock 6629 _bit_map->lock()->unlock(); 6630 ConcurrentMarkSweepThread::desynchronize(true); 6631 _collector->stopTimer(); 6632 if (PrintCMSStatistics != 0) { 6633 _collector->incrementYields(); 6634 } 6635 6636 // See the comment in coordinator_yield() 6637 for (unsigned i = 0; i < CMSYieldSleepCount && 6638 ConcurrentMarkSweepThread::should_yield() && 6639 !CMSCollector::foregroundGCIsActive(); ++i) { 6640 os::sleep(Thread::current(), 1, false); 6641 } 6642 6643 ConcurrentMarkSweepThread::synchronize(true); 6644 _bit_map->lock()->lock_without_safepoint_check(); 6645 _collector->startTimer(); 6646 } 6647 6648 // This closure is used to rescan the marked objects on the dirty cards 6649 // in the mod union table and the card table proper. In the parallel 6650 // case, although the bitMap is shared, we do a single read so the 6651 // isMarked() query is "safe". 6652 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6653 // Ignore mark word because we are running concurrent with mutators 6654 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 6655 HeapWord* addr = (HeapWord*)p; 6656 assert(_span.contains(addr), "we are scanning the CMS generation"); 6657 bool is_obj_array = false; 6658 #ifdef ASSERT 6659 if (!_parallel) { 6660 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6661 assert(_collector->overflow_list_is_empty(), 6662 "overflow list should be empty"); 6663 6664 } 6665 #endif // ASSERT 6666 if (_bit_map->isMarked(addr)) { 6667 // Obj arrays are precisely marked, non-arrays are not; 6668 // so we scan objArrays precisely and non-arrays in their 6669 // entirety. 6670 if (p->is_objArray()) { 6671 is_obj_array = true; 6672 if (_parallel) { 6673 p->oop_iterate(_par_scan_closure, mr); 6674 } else { 6675 p->oop_iterate(_scan_closure, mr); 6676 } 6677 } else { 6678 if (_parallel) { 6679 p->oop_iterate(_par_scan_closure); 6680 } else { 6681 p->oop_iterate(_scan_closure); 6682 } 6683 } 6684 } 6685 #ifdef ASSERT 6686 if (!_parallel) { 6687 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6688 assert(_collector->overflow_list_is_empty(), 6689 "overflow list should be empty"); 6690 6691 } 6692 #endif // ASSERT 6693 return is_obj_array; 6694 } 6695 6696 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6697 MemRegion span, 6698 CMSBitMap* bitMap, CMSMarkStack* markStack, 6699 bool should_yield, bool verifying): 6700 _collector(collector), 6701 _span(span), 6702 _bitMap(bitMap), 6703 _mut(&collector->_modUnionTable), 6704 _markStack(markStack), 6705 _yield(should_yield), 6706 _skipBits(0) 6707 { 6708 assert(_markStack->isEmpty(), "stack should be empty"); 6709 _finger = _bitMap->startWord(); 6710 _threshold = _finger; 6711 assert(_collector->_restart_addr == NULL, "Sanity check"); 6712 assert(_span.contains(_finger), "Out of bounds _finger?"); 6713 DEBUG_ONLY(_verifying = verifying;) 6714 } 6715 6716 void MarkFromRootsClosure::reset(HeapWord* addr) { 6717 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6718 assert(_span.contains(addr), "Out of bounds _finger?"); 6719 _finger = addr; 6720 _threshold = (HeapWord*)round_to( 6721 (intptr_t)_finger, CardTableModRefBS::card_size); 6722 } 6723 6724 // Should revisit to see if this should be restructured for 6725 // greater efficiency. 6726 bool MarkFromRootsClosure::do_bit(size_t offset) { 6727 if (_skipBits > 0) { 6728 _skipBits--; 6729 return true; 6730 } 6731 // convert offset into a HeapWord* 6732 HeapWord* addr = _bitMap->startWord() + offset; 6733 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6734 "address out of range"); 6735 assert(_bitMap->isMarked(addr), "tautology"); 6736 if (_bitMap->isMarked(addr+1)) { 6737 // this is an allocated but not yet initialized object 6738 assert(_skipBits == 0, "tautology"); 6739 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6740 oop p = oop(addr); 6741 if (p->klass_or_null() == NULL) { 6742 DEBUG_ONLY(if (!_verifying) {) 6743 // We re-dirty the cards on which this object lies and increase 6744 // the _threshold so that we'll come back to scan this object 6745 // during the preclean or remark phase. (CMSCleanOnEnter) 6746 if (CMSCleanOnEnter) { 6747 size_t sz = _collector->block_size_using_printezis_bits(addr); 6748 HeapWord* end_card_addr = (HeapWord*)round_to( 6749 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6750 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6751 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6752 // Bump _threshold to end_card_addr; note that 6753 // _threshold cannot possibly exceed end_card_addr, anyhow. 6754 // This prevents future clearing of the card as the scan proceeds 6755 // to the right. 6756 assert(_threshold <= end_card_addr, 6757 "Because we are just scanning into this object"); 6758 if (_threshold < end_card_addr) { 6759 _threshold = end_card_addr; 6760 } 6761 if (p->klass_or_null() != NULL) { 6762 // Redirty the range of cards... 6763 _mut->mark_range(redirty_range); 6764 } // ...else the setting of klass will dirty the card anyway. 6765 } 6766 DEBUG_ONLY(}) 6767 return true; 6768 } 6769 } 6770 scanOopsInOop(addr); 6771 return true; 6772 } 6773 6774 // We take a break if we've been at this for a while, 6775 // so as to avoid monopolizing the locks involved. 6776 void MarkFromRootsClosure::do_yield_work() { 6777 // First give up the locks, then yield, then re-lock 6778 // We should probably use a constructor/destructor idiom to 6779 // do this unlock/lock or modify the MutexUnlocker class to 6780 // serve our purpose. XXX 6781 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6782 "CMS thread should hold CMS token"); 6783 assert_lock_strong(_bitMap->lock()); 6784 _bitMap->lock()->unlock(); 6785 ConcurrentMarkSweepThread::desynchronize(true); 6786 _collector->stopTimer(); 6787 if (PrintCMSStatistics != 0) { 6788 _collector->incrementYields(); 6789 } 6790 6791 // See the comment in coordinator_yield() 6792 for (unsigned i = 0; i < CMSYieldSleepCount && 6793 ConcurrentMarkSweepThread::should_yield() && 6794 !CMSCollector::foregroundGCIsActive(); ++i) { 6795 os::sleep(Thread::current(), 1, false); 6796 } 6797 6798 ConcurrentMarkSweepThread::synchronize(true); 6799 _bitMap->lock()->lock_without_safepoint_check(); 6800 _collector->startTimer(); 6801 } 6802 6803 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6804 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6805 assert(_markStack->isEmpty(), 6806 "should drain stack to limit stack usage"); 6807 // convert ptr to an oop preparatory to scanning 6808 oop obj = oop(ptr); 6809 // Ignore mark word in verification below, since we 6810 // may be running concurrent with mutators. 6811 assert(obj->is_oop(true), "should be an oop"); 6812 assert(_finger <= ptr, "_finger runneth ahead"); 6813 // advance the finger to right end of this object 6814 _finger = ptr + obj->size(); 6815 assert(_finger > ptr, "we just incremented it above"); 6816 // On large heaps, it may take us some time to get through 6817 // the marking phase. During 6818 // this time it's possible that a lot of mutations have 6819 // accumulated in the card table and the mod union table -- 6820 // these mutation records are redundant until we have 6821 // actually traced into the corresponding card. 6822 // Here, we check whether advancing the finger would make 6823 // us cross into a new card, and if so clear corresponding 6824 // cards in the MUT (preclean them in the card-table in the 6825 // future). 6826 6827 DEBUG_ONLY(if (!_verifying) {) 6828 // The clean-on-enter optimization is disabled by default, 6829 // until we fix 6178663. 6830 if (CMSCleanOnEnter && (_finger > _threshold)) { 6831 // [_threshold, _finger) represents the interval 6832 // of cards to be cleared in MUT (or precleaned in card table). 6833 // The set of cards to be cleared is all those that overlap 6834 // with the interval [_threshold, _finger); note that 6835 // _threshold is always kept card-aligned but _finger isn't 6836 // always card-aligned. 6837 HeapWord* old_threshold = _threshold; 6838 assert(old_threshold == (HeapWord*)round_to( 6839 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6840 "_threshold should always be card-aligned"); 6841 _threshold = (HeapWord*)round_to( 6842 (intptr_t)_finger, CardTableModRefBS::card_size); 6843 MemRegion mr(old_threshold, _threshold); 6844 assert(!mr.is_empty(), "Control point invariant"); 6845 assert(_span.contains(mr), "Should clear within span"); 6846 _mut->clear_range(mr); 6847 } 6848 DEBUG_ONLY(}) 6849 // Note: the finger doesn't advance while we drain 6850 // the stack below. 6851 PushOrMarkClosure pushOrMarkClosure(_collector, 6852 _span, _bitMap, _markStack, 6853 _finger, this); 6854 bool res = _markStack->push(obj); 6855 assert(res, "Empty non-zero size stack should have space for single push"); 6856 while (!_markStack->isEmpty()) { 6857 oop new_oop = _markStack->pop(); 6858 // Skip verifying header mark word below because we are 6859 // running concurrent with mutators. 6860 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6861 // now scan this oop's oops 6862 new_oop->oop_iterate(&pushOrMarkClosure); 6863 do_yield_check(); 6864 } 6865 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6866 } 6867 6868 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6869 CMSCollector* collector, MemRegion span, 6870 CMSBitMap* bit_map, 6871 OopTaskQueue* work_queue, 6872 CMSMarkStack* overflow_stack): 6873 _collector(collector), 6874 _whole_span(collector->_span), 6875 _span(span), 6876 _bit_map(bit_map), 6877 _mut(&collector->_modUnionTable), 6878 _work_queue(work_queue), 6879 _overflow_stack(overflow_stack), 6880 _skip_bits(0), 6881 _task(task) 6882 { 6883 assert(_work_queue->size() == 0, "work_queue should be empty"); 6884 _finger = span.start(); 6885 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6886 assert(_span.contains(_finger), "Out of bounds _finger?"); 6887 } 6888 6889 // Should revisit to see if this should be restructured for 6890 // greater efficiency. 6891 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6892 if (_skip_bits > 0) { 6893 _skip_bits--; 6894 return true; 6895 } 6896 // convert offset into a HeapWord* 6897 HeapWord* addr = _bit_map->startWord() + offset; 6898 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6899 "address out of range"); 6900 assert(_bit_map->isMarked(addr), "tautology"); 6901 if (_bit_map->isMarked(addr+1)) { 6902 // this is an allocated object that might not yet be initialized 6903 assert(_skip_bits == 0, "tautology"); 6904 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6905 oop p = oop(addr); 6906 if (p->klass_or_null() == NULL) { 6907 // in the case of Clean-on-Enter optimization, redirty card 6908 // and avoid clearing card by increasing the threshold. 6909 return true; 6910 } 6911 } 6912 scan_oops_in_oop(addr); 6913 return true; 6914 } 6915 6916 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6917 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6918 // Should we assert that our work queue is empty or 6919 // below some drain limit? 6920 assert(_work_queue->size() == 0, 6921 "should drain stack to limit stack usage"); 6922 // convert ptr to an oop preparatory to scanning 6923 oop obj = oop(ptr); 6924 // Ignore mark word in verification below, since we 6925 // may be running concurrent with mutators. 6926 assert(obj->is_oop(true), "should be an oop"); 6927 assert(_finger <= ptr, "_finger runneth ahead"); 6928 // advance the finger to right end of this object 6929 _finger = ptr + obj->size(); 6930 assert(_finger > ptr, "we just incremented it above"); 6931 // On large heaps, it may take us some time to get through 6932 // the marking phase. During 6933 // this time it's possible that a lot of mutations have 6934 // accumulated in the card table and the mod union table -- 6935 // these mutation records are redundant until we have 6936 // actually traced into the corresponding card. 6937 // Here, we check whether advancing the finger would make 6938 // us cross into a new card, and if so clear corresponding 6939 // cards in the MUT (preclean them in the card-table in the 6940 // future). 6941 6942 // The clean-on-enter optimization is disabled by default, 6943 // until we fix 6178663. 6944 if (CMSCleanOnEnter && (_finger > _threshold)) { 6945 // [_threshold, _finger) represents the interval 6946 // of cards to be cleared in MUT (or precleaned in card table). 6947 // The set of cards to be cleared is all those that overlap 6948 // with the interval [_threshold, _finger); note that 6949 // _threshold is always kept card-aligned but _finger isn't 6950 // always card-aligned. 6951 HeapWord* old_threshold = _threshold; 6952 assert(old_threshold == (HeapWord*)round_to( 6953 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6954 "_threshold should always be card-aligned"); 6955 _threshold = (HeapWord*)round_to( 6956 (intptr_t)_finger, CardTableModRefBS::card_size); 6957 MemRegion mr(old_threshold, _threshold); 6958 assert(!mr.is_empty(), "Control point invariant"); 6959 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6960 _mut->clear_range(mr); 6961 } 6962 6963 // Note: the local finger doesn't advance while we drain 6964 // the stack below, but the global finger sure can and will. 6965 HeapWord** gfa = _task->global_finger_addr(); 6966 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6967 _span, _bit_map, 6968 _work_queue, 6969 _overflow_stack, 6970 _finger, 6971 gfa, this); 6972 bool res = _work_queue->push(obj); // overflow could occur here 6973 assert(res, "Will hold once we use workqueues"); 6974 while (true) { 6975 oop new_oop; 6976 if (!_work_queue->pop_local(new_oop)) { 6977 // We emptied our work_queue; check if there's stuff that can 6978 // be gotten from the overflow stack. 6979 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6980 _overflow_stack, _work_queue)) { 6981 do_yield_check(); 6982 continue; 6983 } else { // done 6984 break; 6985 } 6986 } 6987 // Skip verifying header mark word below because we are 6988 // running concurrent with mutators. 6989 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6990 // now scan this oop's oops 6991 new_oop->oop_iterate(&pushOrMarkClosure); 6992 do_yield_check(); 6993 } 6994 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6995 } 6996 6997 // Yield in response to a request from VM Thread or 6998 // from mutators. 6999 void Par_MarkFromRootsClosure::do_yield_work() { 7000 assert(_task != NULL, "sanity"); 7001 _task->yield(); 7002 } 7003 7004 // A variant of the above used for verifying CMS marking work. 7005 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7006 MemRegion span, 7007 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7008 CMSMarkStack* mark_stack): 7009 _collector(collector), 7010 _span(span), 7011 _verification_bm(verification_bm), 7012 _cms_bm(cms_bm), 7013 _mark_stack(mark_stack), 7014 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7015 mark_stack) 7016 { 7017 assert(_mark_stack->isEmpty(), "stack should be empty"); 7018 _finger = _verification_bm->startWord(); 7019 assert(_collector->_restart_addr == NULL, "Sanity check"); 7020 assert(_span.contains(_finger), "Out of bounds _finger?"); 7021 } 7022 7023 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7024 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7025 assert(_span.contains(addr), "Out of bounds _finger?"); 7026 _finger = addr; 7027 } 7028 7029 // Should revisit to see if this should be restructured for 7030 // greater efficiency. 7031 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7032 // convert offset into a HeapWord* 7033 HeapWord* addr = _verification_bm->startWord() + offset; 7034 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7035 "address out of range"); 7036 assert(_verification_bm->isMarked(addr), "tautology"); 7037 assert(_cms_bm->isMarked(addr), "tautology"); 7038 7039 assert(_mark_stack->isEmpty(), 7040 "should drain stack to limit stack usage"); 7041 // convert addr to an oop preparatory to scanning 7042 oop obj = oop(addr); 7043 assert(obj->is_oop(), "should be an oop"); 7044 assert(_finger <= addr, "_finger runneth ahead"); 7045 // advance the finger to right end of this object 7046 _finger = addr + obj->size(); 7047 assert(_finger > addr, "we just incremented it above"); 7048 // Note: the finger doesn't advance while we drain 7049 // the stack below. 7050 bool res = _mark_stack->push(obj); 7051 assert(res, "Empty non-zero size stack should have space for single push"); 7052 while (!_mark_stack->isEmpty()) { 7053 oop new_oop = _mark_stack->pop(); 7054 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7055 // now scan this oop's oops 7056 new_oop->oop_iterate(&_pam_verify_closure); 7057 } 7058 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7059 return true; 7060 } 7061 7062 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7063 CMSCollector* collector, MemRegion span, 7064 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7065 CMSMarkStack* mark_stack): 7066 MetadataAwareOopClosure(collector->ref_processor()), 7067 _collector(collector), 7068 _span(span), 7069 _verification_bm(verification_bm), 7070 _cms_bm(cms_bm), 7071 _mark_stack(mark_stack) 7072 { } 7073 7074 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7075 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7076 7077 // Upon stack overflow, we discard (part of) the stack, 7078 // remembering the least address amongst those discarded 7079 // in CMSCollector's _restart_address. 7080 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7081 // Remember the least grey address discarded 7082 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7083 _collector->lower_restart_addr(ra); 7084 _mark_stack->reset(); // discard stack contents 7085 _mark_stack->expand(); // expand the stack if possible 7086 } 7087 7088 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7089 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7090 HeapWord* addr = (HeapWord*)obj; 7091 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7092 // Oop lies in _span and isn't yet grey or black 7093 _verification_bm->mark(addr); // now grey 7094 if (!_cms_bm->isMarked(addr)) { 7095 oop(addr)->print(); 7096 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7097 addr); 7098 fatal("... aborting"); 7099 } 7100 7101 if (!_mark_stack->push(obj)) { // stack overflow 7102 if (PrintCMSStatistics != 0) { 7103 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7104 SIZE_FORMAT, _mark_stack->capacity()); 7105 } 7106 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7107 handle_stack_overflow(addr); 7108 } 7109 // anything including and to the right of _finger 7110 // will be scanned as we iterate over the remainder of the 7111 // bit map 7112 } 7113 } 7114 7115 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7116 MemRegion span, 7117 CMSBitMap* bitMap, CMSMarkStack* markStack, 7118 HeapWord* finger, MarkFromRootsClosure* parent) : 7119 MetadataAwareOopClosure(collector->ref_processor()), 7120 _collector(collector), 7121 _span(span), 7122 _bitMap(bitMap), 7123 _markStack(markStack), 7124 _finger(finger), 7125 _parent(parent) 7126 { } 7127 7128 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7129 MemRegion span, 7130 CMSBitMap* bit_map, 7131 OopTaskQueue* work_queue, 7132 CMSMarkStack* overflow_stack, 7133 HeapWord* finger, 7134 HeapWord** global_finger_addr, 7135 Par_MarkFromRootsClosure* parent) : 7136 MetadataAwareOopClosure(collector->ref_processor()), 7137 _collector(collector), 7138 _whole_span(collector->_span), 7139 _span(span), 7140 _bit_map(bit_map), 7141 _work_queue(work_queue), 7142 _overflow_stack(overflow_stack), 7143 _finger(finger), 7144 _global_finger_addr(global_finger_addr), 7145 _parent(parent) 7146 { } 7147 7148 // Assumes thread-safe access by callers, who are 7149 // responsible for mutual exclusion. 7150 void CMSCollector::lower_restart_addr(HeapWord* low) { 7151 assert(_span.contains(low), "Out of bounds addr"); 7152 if (_restart_addr == NULL) { 7153 _restart_addr = low; 7154 } else { 7155 _restart_addr = MIN2(_restart_addr, low); 7156 } 7157 } 7158 7159 // Upon stack overflow, we discard (part of) the stack, 7160 // remembering the least address amongst those discarded 7161 // in CMSCollector's _restart_address. 7162 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7163 // Remember the least grey address discarded 7164 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7165 _collector->lower_restart_addr(ra); 7166 _markStack->reset(); // discard stack contents 7167 _markStack->expand(); // expand the stack if possible 7168 } 7169 7170 // Upon stack overflow, we discard (part of) the stack, 7171 // remembering the least address amongst those discarded 7172 // in CMSCollector's _restart_address. 7173 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7174 // We need to do this under a mutex to prevent other 7175 // workers from interfering with the work done below. 7176 MutexLockerEx ml(_overflow_stack->par_lock(), 7177 Mutex::_no_safepoint_check_flag); 7178 // Remember the least grey address discarded 7179 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7180 _collector->lower_restart_addr(ra); 7181 _overflow_stack->reset(); // discard stack contents 7182 _overflow_stack->expand(); // expand the stack if possible 7183 } 7184 7185 void PushOrMarkClosure::do_oop(oop obj) { 7186 // Ignore mark word because we are running concurrent with mutators. 7187 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7188 HeapWord* addr = (HeapWord*)obj; 7189 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7190 // Oop lies in _span and isn't yet grey or black 7191 _bitMap->mark(addr); // now grey 7192 if (addr < _finger) { 7193 // the bit map iteration has already either passed, or 7194 // sampled, this bit in the bit map; we'll need to 7195 // use the marking stack to scan this oop's oops. 7196 bool simulate_overflow = false; 7197 NOT_PRODUCT( 7198 if (CMSMarkStackOverflowALot && 7199 _collector->simulate_overflow()) { 7200 // simulate a stack overflow 7201 simulate_overflow = true; 7202 } 7203 ) 7204 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7205 if (PrintCMSStatistics != 0) { 7206 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7207 SIZE_FORMAT, _markStack->capacity()); 7208 } 7209 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7210 handle_stack_overflow(addr); 7211 } 7212 } 7213 // anything including and to the right of _finger 7214 // will be scanned as we iterate over the remainder of the 7215 // bit map 7216 do_yield_check(); 7217 } 7218 } 7219 7220 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7221 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7222 7223 void Par_PushOrMarkClosure::do_oop(oop obj) { 7224 // Ignore mark word because we are running concurrent with mutators. 7225 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7226 HeapWord* addr = (HeapWord*)obj; 7227 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7228 // Oop lies in _span and isn't yet grey or black 7229 // We read the global_finger (volatile read) strictly after marking oop 7230 bool res = _bit_map->par_mark(addr); // now grey 7231 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7232 // Should we push this marked oop on our stack? 7233 // -- if someone else marked it, nothing to do 7234 // -- if target oop is above global finger nothing to do 7235 // -- if target oop is in chunk and above local finger 7236 // then nothing to do 7237 // -- else push on work queue 7238 if ( !res // someone else marked it, they will deal with it 7239 || (addr >= *gfa) // will be scanned in a later task 7240 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7241 return; 7242 } 7243 // the bit map iteration has already either passed, or 7244 // sampled, this bit in the bit map; we'll need to 7245 // use the marking stack to scan this oop's oops. 7246 bool simulate_overflow = false; 7247 NOT_PRODUCT( 7248 if (CMSMarkStackOverflowALot && 7249 _collector->simulate_overflow()) { 7250 // simulate a stack overflow 7251 simulate_overflow = true; 7252 } 7253 ) 7254 if (simulate_overflow || 7255 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7256 // stack overflow 7257 if (PrintCMSStatistics != 0) { 7258 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7259 SIZE_FORMAT, _overflow_stack->capacity()); 7260 } 7261 // We cannot assert that the overflow stack is full because 7262 // it may have been emptied since. 7263 assert(simulate_overflow || 7264 _work_queue->size() == _work_queue->max_elems(), 7265 "Else push should have succeeded"); 7266 handle_stack_overflow(addr); 7267 } 7268 do_yield_check(); 7269 } 7270 } 7271 7272 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7273 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7274 7275 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7276 MemRegion span, 7277 ReferenceProcessor* rp, 7278 CMSBitMap* bit_map, 7279 CMSBitMap* mod_union_table, 7280 CMSMarkStack* mark_stack, 7281 bool concurrent_precleaning): 7282 MetadataAwareOopClosure(rp), 7283 _collector(collector), 7284 _span(span), 7285 _bit_map(bit_map), 7286 _mod_union_table(mod_union_table), 7287 _mark_stack(mark_stack), 7288 _concurrent_precleaning(concurrent_precleaning) 7289 { 7290 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7291 } 7292 7293 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7294 // the non-parallel version (the parallel version appears further below.) 7295 void PushAndMarkClosure::do_oop(oop obj) { 7296 // Ignore mark word verification. If during concurrent precleaning, 7297 // the object monitor may be locked. If during the checkpoint 7298 // phases, the object may already have been reached by a different 7299 // path and may be at the end of the global overflow list (so 7300 // the mark word may be NULL). 7301 assert(obj->is_oop_or_null(true /* ignore mark word */), 7302 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7303 HeapWord* addr = (HeapWord*)obj; 7304 // Check if oop points into the CMS generation 7305 // and is not marked 7306 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7307 // a white object ... 7308 _bit_map->mark(addr); // ... now grey 7309 // push on the marking stack (grey set) 7310 bool simulate_overflow = false; 7311 NOT_PRODUCT( 7312 if (CMSMarkStackOverflowALot && 7313 _collector->simulate_overflow()) { 7314 // simulate a stack overflow 7315 simulate_overflow = true; 7316 } 7317 ) 7318 if (simulate_overflow || !_mark_stack->push(obj)) { 7319 if (_concurrent_precleaning) { 7320 // During precleaning we can just dirty the appropriate card(s) 7321 // in the mod union table, thus ensuring that the object remains 7322 // in the grey set and continue. In the case of object arrays 7323 // we need to dirty all of the cards that the object spans, 7324 // since the rescan of object arrays will be limited to the 7325 // dirty cards. 7326 // Note that no one can be interfering with us in this action 7327 // of dirtying the mod union table, so no locking or atomics 7328 // are required. 7329 if (obj->is_objArray()) { 7330 size_t sz = obj->size(); 7331 HeapWord* end_card_addr = (HeapWord*)round_to( 7332 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7333 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7334 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7335 _mod_union_table->mark_range(redirty_range); 7336 } else { 7337 _mod_union_table->mark(addr); 7338 } 7339 _collector->_ser_pmc_preclean_ovflw++; 7340 } else { 7341 // During the remark phase, we need to remember this oop 7342 // in the overflow list. 7343 _collector->push_on_overflow_list(obj); 7344 _collector->_ser_pmc_remark_ovflw++; 7345 } 7346 } 7347 } 7348 } 7349 7350 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7351 MemRegion span, 7352 ReferenceProcessor* rp, 7353 CMSBitMap* bit_map, 7354 OopTaskQueue* work_queue): 7355 MetadataAwareOopClosure(rp), 7356 _collector(collector), 7357 _span(span), 7358 _bit_map(bit_map), 7359 _work_queue(work_queue) 7360 { 7361 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7362 } 7363 7364 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7365 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7366 7367 // Grey object rescan during second checkpoint phase -- 7368 // the parallel version. 7369 void Par_PushAndMarkClosure::do_oop(oop obj) { 7370 // In the assert below, we ignore the mark word because 7371 // this oop may point to an already visited object that is 7372 // on the overflow stack (in which case the mark word has 7373 // been hijacked for chaining into the overflow stack -- 7374 // if this is the last object in the overflow stack then 7375 // its mark word will be NULL). Because this object may 7376 // have been subsequently popped off the global overflow 7377 // stack, and the mark word possibly restored to the prototypical 7378 // value, by the time we get to examined this failing assert in 7379 // the debugger, is_oop_or_null(false) may subsequently start 7380 // to hold. 7381 assert(obj->is_oop_or_null(true), 7382 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7383 HeapWord* addr = (HeapWord*)obj; 7384 // Check if oop points into the CMS generation 7385 // and is not marked 7386 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7387 // a white object ... 7388 // If we manage to "claim" the object, by being the 7389 // first thread to mark it, then we push it on our 7390 // marking stack 7391 if (_bit_map->par_mark(addr)) { // ... now grey 7392 // push on work queue (grey set) 7393 bool simulate_overflow = false; 7394 NOT_PRODUCT( 7395 if (CMSMarkStackOverflowALot && 7396 _collector->par_simulate_overflow()) { 7397 // simulate a stack overflow 7398 simulate_overflow = true; 7399 } 7400 ) 7401 if (simulate_overflow || !_work_queue->push(obj)) { 7402 _collector->par_push_on_overflow_list(obj); 7403 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7404 } 7405 } // Else, some other thread got there first 7406 } 7407 } 7408 7409 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7410 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7411 7412 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7413 Mutex* bml = _collector->bitMapLock(); 7414 assert_lock_strong(bml); 7415 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7416 "CMS thread should hold CMS token"); 7417 7418 bml->unlock(); 7419 ConcurrentMarkSweepThread::desynchronize(true); 7420 7421 _collector->stopTimer(); 7422 if (PrintCMSStatistics != 0) { 7423 _collector->incrementYields(); 7424 } 7425 7426 // See the comment in coordinator_yield() 7427 for (unsigned i = 0; i < CMSYieldSleepCount && 7428 ConcurrentMarkSweepThread::should_yield() && 7429 !CMSCollector::foregroundGCIsActive(); ++i) { 7430 os::sleep(Thread::current(), 1, false); 7431 } 7432 7433 ConcurrentMarkSweepThread::synchronize(true); 7434 bml->lock(); 7435 7436 _collector->startTimer(); 7437 } 7438 7439 bool CMSPrecleanRefsYieldClosure::should_return() { 7440 if (ConcurrentMarkSweepThread::should_yield()) { 7441 do_yield_work(); 7442 } 7443 return _collector->foregroundGCIsActive(); 7444 } 7445 7446 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7447 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7448 "mr should be aligned to start at a card boundary"); 7449 // We'd like to assert: 7450 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7451 // "mr should be a range of cards"); 7452 // However, that would be too strong in one case -- the last 7453 // partition ends at _unallocated_block which, in general, can be 7454 // an arbitrary boundary, not necessarily card aligned. 7455 if (PrintCMSStatistics != 0) { 7456 _num_dirty_cards += 7457 mr.word_size()/CardTableModRefBS::card_size_in_words; 7458 } 7459 _space->object_iterate_mem(mr, &_scan_cl); 7460 } 7461 7462 SweepClosure::SweepClosure(CMSCollector* collector, 7463 ConcurrentMarkSweepGeneration* g, 7464 CMSBitMap* bitMap, bool should_yield) : 7465 _collector(collector), 7466 _g(g), 7467 _sp(g->cmsSpace()), 7468 _limit(_sp->sweep_limit()), 7469 _freelistLock(_sp->freelistLock()), 7470 _bitMap(bitMap), 7471 _yield(should_yield), 7472 _inFreeRange(false), // No free range at beginning of sweep 7473 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7474 _lastFreeRangeCoalesced(false), 7475 _freeFinger(g->used_region().start()) 7476 { 7477 NOT_PRODUCT( 7478 _numObjectsFreed = 0; 7479 _numWordsFreed = 0; 7480 _numObjectsLive = 0; 7481 _numWordsLive = 0; 7482 _numObjectsAlreadyFree = 0; 7483 _numWordsAlreadyFree = 0; 7484 _last_fc = NULL; 7485 7486 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7487 _sp->dictionary()->initialize_dict_returned_bytes(); 7488 ) 7489 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7490 "sweep _limit out of bounds"); 7491 if (CMSTraceSweeper) { 7492 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7493 _limit); 7494 } 7495 } 7496 7497 void SweepClosure::print_on(outputStream* st) const { 7498 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7499 _sp->bottom(), _sp->end()); 7500 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7501 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7502 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7503 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7504 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7505 } 7506 7507 #ifndef PRODUCT 7508 // Assertion checking only: no useful work in product mode -- 7509 // however, if any of the flags below become product flags, 7510 // you may need to review this code to see if it needs to be 7511 // enabled in product mode. 7512 SweepClosure::~SweepClosure() { 7513 assert_lock_strong(_freelistLock); 7514 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7515 "sweep _limit out of bounds"); 7516 if (inFreeRange()) { 7517 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7518 print(); 7519 ShouldNotReachHere(); 7520 } 7521 if (Verbose && PrintGC) { 7522 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7523 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7524 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7525 SIZE_FORMAT" bytes " 7526 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7527 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7528 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7529 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7530 * sizeof(HeapWord); 7531 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7532 7533 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7534 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7535 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7536 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7537 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7538 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7539 indexListReturnedBytes); 7540 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7541 dict_returned_bytes); 7542 } 7543 } 7544 if (CMSTraceSweeper) { 7545 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7546 _limit); 7547 } 7548 } 7549 #endif // PRODUCT 7550 7551 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7552 bool freeRangeInFreeLists) { 7553 if (CMSTraceSweeper) { 7554 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7555 freeFinger, freeRangeInFreeLists); 7556 } 7557 assert(!inFreeRange(), "Trampling existing free range"); 7558 set_inFreeRange(true); 7559 set_lastFreeRangeCoalesced(false); 7560 7561 set_freeFinger(freeFinger); 7562 set_freeRangeInFreeLists(freeRangeInFreeLists); 7563 if (CMSTestInFreeList) { 7564 if (freeRangeInFreeLists) { 7565 FreeChunk* fc = (FreeChunk*) freeFinger; 7566 assert(fc->is_free(), "A chunk on the free list should be free."); 7567 assert(fc->size() > 0, "Free range should have a size"); 7568 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7569 } 7570 } 7571 } 7572 7573 // Note that the sweeper runs concurrently with mutators. Thus, 7574 // it is possible for direct allocation in this generation to happen 7575 // in the middle of the sweep. Note that the sweeper also coalesces 7576 // contiguous free blocks. Thus, unless the sweeper and the allocator 7577 // synchronize appropriately freshly allocated blocks may get swept up. 7578 // This is accomplished by the sweeper locking the free lists while 7579 // it is sweeping. Thus blocks that are determined to be free are 7580 // indeed free. There is however one additional complication: 7581 // blocks that have been allocated since the final checkpoint and 7582 // mark, will not have been marked and so would be treated as 7583 // unreachable and swept up. To prevent this, the allocator marks 7584 // the bit map when allocating during the sweep phase. This leads, 7585 // however, to a further complication -- objects may have been allocated 7586 // but not yet initialized -- in the sense that the header isn't yet 7587 // installed. The sweeper can not then determine the size of the block 7588 // in order to skip over it. To deal with this case, we use a technique 7589 // (due to Printezis) to encode such uninitialized block sizes in the 7590 // bit map. Since the bit map uses a bit per every HeapWord, but the 7591 // CMS generation has a minimum object size of 3 HeapWords, it follows 7592 // that "normal marks" won't be adjacent in the bit map (there will 7593 // always be at least two 0 bits between successive 1 bits). We make use 7594 // of these "unused" bits to represent uninitialized blocks -- the bit 7595 // corresponding to the start of the uninitialized object and the next 7596 // bit are both set. Finally, a 1 bit marks the end of the object that 7597 // started with the two consecutive 1 bits to indicate its potentially 7598 // uninitialized state. 7599 7600 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7601 FreeChunk* fc = (FreeChunk*)addr; 7602 size_t res; 7603 7604 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7605 // than "addr == _limit" because although _limit was a block boundary when 7606 // we started the sweep, it may no longer be one because heap expansion 7607 // may have caused us to coalesce the block ending at the address _limit 7608 // with a newly expanded chunk (this happens when _limit was set to the 7609 // previous _end of the space), so we may have stepped past _limit: 7610 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7611 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7612 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7613 "sweep _limit out of bounds"); 7614 assert(addr < _sp->end(), "addr out of bounds"); 7615 // Flush any free range we might be holding as a single 7616 // coalesced chunk to the appropriate free list. 7617 if (inFreeRange()) { 7618 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7619 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 7620 flush_cur_free_chunk(freeFinger(), 7621 pointer_delta(addr, freeFinger())); 7622 if (CMSTraceSweeper) { 7623 gclog_or_tty->print("Sweep: last chunk: "); 7624 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 7625 "[coalesced:%d]\n", 7626 freeFinger(), pointer_delta(addr, freeFinger()), 7627 lastFreeRangeCoalesced() ? 1 : 0); 7628 } 7629 } 7630 7631 // help the iterator loop finish 7632 return pointer_delta(_sp->end(), addr); 7633 } 7634 7635 assert(addr < _limit, "sweep invariant"); 7636 // check if we should yield 7637 do_yield_check(addr); 7638 if (fc->is_free()) { 7639 // Chunk that is already free 7640 res = fc->size(); 7641 do_already_free_chunk(fc); 7642 debug_only(_sp->verifyFreeLists()); 7643 // If we flush the chunk at hand in lookahead_and_flush() 7644 // and it's coalesced with a preceding chunk, then the 7645 // process of "mangling" the payload of the coalesced block 7646 // will cause erasure of the size information from the 7647 // (erstwhile) header of all the coalesced blocks but the 7648 // first, so the first disjunct in the assert will not hold 7649 // in that specific case (in which case the second disjunct 7650 // will hold). 7651 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7652 "Otherwise the size info doesn't change at this step"); 7653 NOT_PRODUCT( 7654 _numObjectsAlreadyFree++; 7655 _numWordsAlreadyFree += res; 7656 ) 7657 NOT_PRODUCT(_last_fc = fc;) 7658 } else if (!_bitMap->isMarked(addr)) { 7659 // Chunk is fresh garbage 7660 res = do_garbage_chunk(fc); 7661 debug_only(_sp->verifyFreeLists()); 7662 NOT_PRODUCT( 7663 _numObjectsFreed++; 7664 _numWordsFreed += res; 7665 ) 7666 } else { 7667 // Chunk that is alive. 7668 res = do_live_chunk(fc); 7669 debug_only(_sp->verifyFreeLists()); 7670 NOT_PRODUCT( 7671 _numObjectsLive++; 7672 _numWordsLive += res; 7673 ) 7674 } 7675 return res; 7676 } 7677 7678 // For the smart allocation, record following 7679 // split deaths - a free chunk is removed from its free list because 7680 // it is being split into two or more chunks. 7681 // split birth - a free chunk is being added to its free list because 7682 // a larger free chunk has been split and resulted in this free chunk. 7683 // coal death - a free chunk is being removed from its free list because 7684 // it is being coalesced into a large free chunk. 7685 // coal birth - a free chunk is being added to its free list because 7686 // it was created when two or more free chunks where coalesced into 7687 // this free chunk. 7688 // 7689 // These statistics are used to determine the desired number of free 7690 // chunks of a given size. The desired number is chosen to be relative 7691 // to the end of a CMS sweep. The desired number at the end of a sweep 7692 // is the 7693 // count-at-end-of-previous-sweep (an amount that was enough) 7694 // - count-at-beginning-of-current-sweep (the excess) 7695 // + split-births (gains in this size during interval) 7696 // - split-deaths (demands on this size during interval) 7697 // where the interval is from the end of one sweep to the end of the 7698 // next. 7699 // 7700 // When sweeping the sweeper maintains an accumulated chunk which is 7701 // the chunk that is made up of chunks that have been coalesced. That 7702 // will be termed the left-hand chunk. A new chunk of garbage that 7703 // is being considered for coalescing will be referred to as the 7704 // right-hand chunk. 7705 // 7706 // When making a decision on whether to coalesce a right-hand chunk with 7707 // the current left-hand chunk, the current count vs. the desired count 7708 // of the left-hand chunk is considered. Also if the right-hand chunk 7709 // is near the large chunk at the end of the heap (see 7710 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7711 // left-hand chunk is coalesced. 7712 // 7713 // When making a decision about whether to split a chunk, the desired count 7714 // vs. the current count of the candidate to be split is also considered. 7715 // If the candidate is underpopulated (currently fewer chunks than desired) 7716 // a chunk of an overpopulated (currently more chunks than desired) size may 7717 // be chosen. The "hint" associated with a free list, if non-null, points 7718 // to a free list which may be overpopulated. 7719 // 7720 7721 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7722 const size_t size = fc->size(); 7723 // Chunks that cannot be coalesced are not in the 7724 // free lists. 7725 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7726 assert(_sp->verify_chunk_in_free_list(fc), 7727 "free chunk should be in free lists"); 7728 } 7729 // a chunk that is already free, should not have been 7730 // marked in the bit map 7731 HeapWord* const addr = (HeapWord*) fc; 7732 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7733 // Verify that the bit map has no bits marked between 7734 // addr and purported end of this block. 7735 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7736 7737 // Some chunks cannot be coalesced under any circumstances. 7738 // See the definition of cantCoalesce(). 7739 if (!fc->cantCoalesce()) { 7740 // This chunk can potentially be coalesced. 7741 if (_sp->adaptive_freelists()) { 7742 // All the work is done in 7743 do_post_free_or_garbage_chunk(fc, size); 7744 } else { // Not adaptive free lists 7745 // this is a free chunk that can potentially be coalesced by the sweeper; 7746 if (!inFreeRange()) { 7747 // if the next chunk is a free block that can't be coalesced 7748 // it doesn't make sense to remove this chunk from the free lists 7749 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7750 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 7751 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 7752 nextChunk->is_free() && // ... which is free... 7753 nextChunk->cantCoalesce()) { // ... but can't be coalesced 7754 // nothing to do 7755 } else { 7756 // Potentially the start of a new free range: 7757 // Don't eagerly remove it from the free lists. 7758 // No need to remove it if it will just be put 7759 // back again. (Also from a pragmatic point of view 7760 // if it is a free block in a region that is beyond 7761 // any allocated blocks, an assertion will fail) 7762 // Remember the start of a free run. 7763 initialize_free_range(addr, true); 7764 // end - can coalesce with next chunk 7765 } 7766 } else { 7767 // the midst of a free range, we are coalescing 7768 print_free_block_coalesced(fc); 7769 if (CMSTraceSweeper) { 7770 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 7771 } 7772 // remove it from the free lists 7773 _sp->removeFreeChunkFromFreeLists(fc); 7774 set_lastFreeRangeCoalesced(true); 7775 // If the chunk is being coalesced and the current free range is 7776 // in the free lists, remove the current free range so that it 7777 // will be returned to the free lists in its entirety - all 7778 // the coalesced pieces included. 7779 if (freeRangeInFreeLists()) { 7780 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7781 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7782 "Size of free range is inconsistent with chunk size."); 7783 if (CMSTestInFreeList) { 7784 assert(_sp->verify_chunk_in_free_list(ffc), 7785 "free range is not in free lists"); 7786 } 7787 _sp->removeFreeChunkFromFreeLists(ffc); 7788 set_freeRangeInFreeLists(false); 7789 } 7790 } 7791 } 7792 // Note that if the chunk is not coalescable (the else arm 7793 // below), we unconditionally flush, without needing to do 7794 // a "lookahead," as we do below. 7795 if (inFreeRange()) lookahead_and_flush(fc, size); 7796 } else { 7797 // Code path common to both original and adaptive free lists. 7798 7799 // cant coalesce with previous block; this should be treated 7800 // as the end of a free run if any 7801 if (inFreeRange()) { 7802 // we kicked some butt; time to pick up the garbage 7803 assert(freeFinger() < addr, "freeFinger points too high"); 7804 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7805 } 7806 // else, nothing to do, just continue 7807 } 7808 } 7809 7810 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7811 // This is a chunk of garbage. It is not in any free list. 7812 // Add it to a free list or let it possibly be coalesced into 7813 // a larger chunk. 7814 HeapWord* const addr = (HeapWord*) fc; 7815 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7816 7817 if (_sp->adaptive_freelists()) { 7818 // Verify that the bit map has no bits marked between 7819 // addr and purported end of just dead object. 7820 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7821 7822 do_post_free_or_garbage_chunk(fc, size); 7823 } else { 7824 if (!inFreeRange()) { 7825 // start of a new free range 7826 assert(size > 0, "A free range should have a size"); 7827 initialize_free_range(addr, false); 7828 } else { 7829 // this will be swept up when we hit the end of the 7830 // free range 7831 if (CMSTraceSweeper) { 7832 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 7833 } 7834 // If the chunk is being coalesced and the current free range is 7835 // in the free lists, remove the current free range so that it 7836 // will be returned to the free lists in its entirety - all 7837 // the coalesced pieces included. 7838 if (freeRangeInFreeLists()) { 7839 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7840 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7841 "Size of free range is inconsistent with chunk size."); 7842 if (CMSTestInFreeList) { 7843 assert(_sp->verify_chunk_in_free_list(ffc), 7844 "free range is not in free lists"); 7845 } 7846 _sp->removeFreeChunkFromFreeLists(ffc); 7847 set_freeRangeInFreeLists(false); 7848 } 7849 set_lastFreeRangeCoalesced(true); 7850 } 7851 // this will be swept up when we hit the end of the free range 7852 7853 // Verify that the bit map has no bits marked between 7854 // addr and purported end of just dead object. 7855 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7856 } 7857 assert(_limit >= addr + size, 7858 "A freshly garbage chunk can't possibly straddle over _limit"); 7859 if (inFreeRange()) lookahead_and_flush(fc, size); 7860 return size; 7861 } 7862 7863 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7864 HeapWord* addr = (HeapWord*) fc; 7865 // The sweeper has just found a live object. Return any accumulated 7866 // left hand chunk to the free lists. 7867 if (inFreeRange()) { 7868 assert(freeFinger() < addr, "freeFinger points too high"); 7869 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7870 } 7871 7872 // This object is live: we'd normally expect this to be 7873 // an oop, and like to assert the following: 7874 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7875 // However, as we commented above, this may be an object whose 7876 // header hasn't yet been initialized. 7877 size_t size; 7878 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7879 if (_bitMap->isMarked(addr + 1)) { 7880 // Determine the size from the bit map, rather than trying to 7881 // compute it from the object header. 7882 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7883 size = pointer_delta(nextOneAddr + 1, addr); 7884 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7885 "alignment problem"); 7886 7887 #ifdef ASSERT 7888 if (oop(addr)->klass_or_null() != NULL) { 7889 // Ignore mark word because we are running concurrent with mutators 7890 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7891 assert(size == 7892 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7893 "P-mark and computed size do not agree"); 7894 } 7895 #endif 7896 7897 } else { 7898 // This should be an initialized object that's alive. 7899 assert(oop(addr)->klass_or_null() != NULL, 7900 "Should be an initialized object"); 7901 // Ignore mark word because we are running concurrent with mutators 7902 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7903 // Verify that the bit map has no bits marked between 7904 // addr and purported end of this block. 7905 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7906 assert(size >= 3, "Necessary for Printezis marks to work"); 7907 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7908 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7909 } 7910 return size; 7911 } 7912 7913 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7914 size_t chunkSize) { 7915 // do_post_free_or_garbage_chunk() should only be called in the case 7916 // of the adaptive free list allocator. 7917 const bool fcInFreeLists = fc->is_free(); 7918 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7919 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7920 if (CMSTestInFreeList && fcInFreeLists) { 7921 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7922 } 7923 7924 if (CMSTraceSweeper) { 7925 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 7926 } 7927 7928 HeapWord* const fc_addr = (HeapWord*) fc; 7929 7930 bool coalesce; 7931 const size_t left = pointer_delta(fc_addr, freeFinger()); 7932 const size_t right = chunkSize; 7933 switch (FLSCoalescePolicy) { 7934 // numeric value forms a coalition aggressiveness metric 7935 case 0: { // never coalesce 7936 coalesce = false; 7937 break; 7938 } 7939 case 1: { // coalesce if left & right chunks on overpopulated lists 7940 coalesce = _sp->coalOverPopulated(left) && 7941 _sp->coalOverPopulated(right); 7942 break; 7943 } 7944 case 2: { // coalesce if left chunk on overpopulated list (default) 7945 coalesce = _sp->coalOverPopulated(left); 7946 break; 7947 } 7948 case 3: { // coalesce if left OR right chunk on overpopulated list 7949 coalesce = _sp->coalOverPopulated(left) || 7950 _sp->coalOverPopulated(right); 7951 break; 7952 } 7953 case 4: { // always coalesce 7954 coalesce = true; 7955 break; 7956 } 7957 default: 7958 ShouldNotReachHere(); 7959 } 7960 7961 // Should the current free range be coalesced? 7962 // If the chunk is in a free range and either we decided to coalesce above 7963 // or the chunk is near the large block at the end of the heap 7964 // (isNearLargestChunk() returns true), then coalesce this chunk. 7965 const bool doCoalesce = inFreeRange() 7966 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7967 if (doCoalesce) { 7968 // Coalesce the current free range on the left with the new 7969 // chunk on the right. If either is on a free list, 7970 // it must be removed from the list and stashed in the closure. 7971 if (freeRangeInFreeLists()) { 7972 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7973 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7974 "Size of free range is inconsistent with chunk size."); 7975 if (CMSTestInFreeList) { 7976 assert(_sp->verify_chunk_in_free_list(ffc), 7977 "Chunk is not in free lists"); 7978 } 7979 _sp->coalDeath(ffc->size()); 7980 _sp->removeFreeChunkFromFreeLists(ffc); 7981 set_freeRangeInFreeLists(false); 7982 } 7983 if (fcInFreeLists) { 7984 _sp->coalDeath(chunkSize); 7985 assert(fc->size() == chunkSize, 7986 "The chunk has the wrong size or is not in the free lists"); 7987 _sp->removeFreeChunkFromFreeLists(fc); 7988 } 7989 set_lastFreeRangeCoalesced(true); 7990 print_free_block_coalesced(fc); 7991 } else { // not in a free range and/or should not coalesce 7992 // Return the current free range and start a new one. 7993 if (inFreeRange()) { 7994 // In a free range but cannot coalesce with the right hand chunk. 7995 // Put the current free range into the free lists. 7996 flush_cur_free_chunk(freeFinger(), 7997 pointer_delta(fc_addr, freeFinger())); 7998 } 7999 // Set up for new free range. Pass along whether the right hand 8000 // chunk is in the free lists. 8001 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8002 } 8003 } 8004 8005 // Lookahead flush: 8006 // If we are tracking a free range, and this is the last chunk that 8007 // we'll look at because its end crosses past _limit, we'll preemptively 8008 // flush it along with any free range we may be holding on to. Note that 8009 // this can be the case only for an already free or freshly garbage 8010 // chunk. If this block is an object, it can never straddle 8011 // over _limit. The "straddling" occurs when _limit is set at 8012 // the previous end of the space when this cycle started, and 8013 // a subsequent heap expansion caused the previously co-terminal 8014 // free block to be coalesced with the newly expanded portion, 8015 // thus rendering _limit a non-block-boundary making it dangerous 8016 // for the sweeper to step over and examine. 8017 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8018 assert(inFreeRange(), "Should only be called if currently in a free range."); 8019 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8020 assert(_sp->used_region().contains(eob - 1), 8021 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8022 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8023 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8024 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8025 if (eob >= _limit) { 8026 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8027 if (CMSTraceSweeper) { 8028 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8029 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8030 "[" PTR_FORMAT "," PTR_FORMAT ")", 8031 _limit, fc, eob, _sp->bottom(), _sp->end()); 8032 } 8033 // Return the storage we are tracking back into the free lists. 8034 if (CMSTraceSweeper) { 8035 gclog_or_tty->print_cr("Flushing ... "); 8036 } 8037 assert(freeFinger() < eob, "Error"); 8038 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8039 } 8040 } 8041 8042 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8043 assert(inFreeRange(), "Should only be called if currently in a free range."); 8044 assert(size > 0, 8045 "A zero sized chunk cannot be added to the free lists."); 8046 if (!freeRangeInFreeLists()) { 8047 if (CMSTestInFreeList) { 8048 FreeChunk* fc = (FreeChunk*) chunk; 8049 fc->set_size(size); 8050 assert(!_sp->verify_chunk_in_free_list(fc), 8051 "chunk should not be in free lists yet"); 8052 } 8053 if (CMSTraceSweeper) { 8054 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8055 chunk, size); 8056 } 8057 // A new free range is going to be starting. The current 8058 // free range has not been added to the free lists yet or 8059 // was removed so add it back. 8060 // If the current free range was coalesced, then the death 8061 // of the free range was recorded. Record a birth now. 8062 if (lastFreeRangeCoalesced()) { 8063 _sp->coalBirth(size); 8064 } 8065 _sp->addChunkAndRepairOffsetTable(chunk, size, 8066 lastFreeRangeCoalesced()); 8067 } else if (CMSTraceSweeper) { 8068 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8069 } 8070 set_inFreeRange(false); 8071 set_freeRangeInFreeLists(false); 8072 } 8073 8074 // We take a break if we've been at this for a while, 8075 // so as to avoid monopolizing the locks involved. 8076 void SweepClosure::do_yield_work(HeapWord* addr) { 8077 // Return current free chunk being used for coalescing (if any) 8078 // to the appropriate freelist. After yielding, the next 8079 // free block encountered will start a coalescing range of 8080 // free blocks. If the next free block is adjacent to the 8081 // chunk just flushed, they will need to wait for the next 8082 // sweep to be coalesced. 8083 if (inFreeRange()) { 8084 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8085 } 8086 8087 // First give up the locks, then yield, then re-lock. 8088 // We should probably use a constructor/destructor idiom to 8089 // do this unlock/lock or modify the MutexUnlocker class to 8090 // serve our purpose. XXX 8091 assert_lock_strong(_bitMap->lock()); 8092 assert_lock_strong(_freelistLock); 8093 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8094 "CMS thread should hold CMS token"); 8095 _bitMap->lock()->unlock(); 8096 _freelistLock->unlock(); 8097 ConcurrentMarkSweepThread::desynchronize(true); 8098 _collector->stopTimer(); 8099 if (PrintCMSStatistics != 0) { 8100 _collector->incrementYields(); 8101 } 8102 8103 // See the comment in coordinator_yield() 8104 for (unsigned i = 0; i < CMSYieldSleepCount && 8105 ConcurrentMarkSweepThread::should_yield() && 8106 !CMSCollector::foregroundGCIsActive(); ++i) { 8107 os::sleep(Thread::current(), 1, false); 8108 } 8109 8110 ConcurrentMarkSweepThread::synchronize(true); 8111 _freelistLock->lock(); 8112 _bitMap->lock()->lock_without_safepoint_check(); 8113 _collector->startTimer(); 8114 } 8115 8116 #ifndef PRODUCT 8117 // This is actually very useful in a product build if it can 8118 // be called from the debugger. Compile it into the product 8119 // as needed. 8120 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8121 return debug_cms_space->verify_chunk_in_free_list(fc); 8122 } 8123 #endif 8124 8125 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8126 if (CMSTraceSweeper) { 8127 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8128 fc, fc->size()); 8129 } 8130 } 8131 8132 // CMSIsAliveClosure 8133 bool CMSIsAliveClosure::do_object_b(oop obj) { 8134 HeapWord* addr = (HeapWord*)obj; 8135 return addr != NULL && 8136 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8137 } 8138 8139 8140 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8141 MemRegion span, 8142 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8143 bool cpc): 8144 _collector(collector), 8145 _span(span), 8146 _bit_map(bit_map), 8147 _mark_stack(mark_stack), 8148 _concurrent_precleaning(cpc) { 8149 assert(!_span.is_empty(), "Empty span could spell trouble"); 8150 } 8151 8152 8153 // CMSKeepAliveClosure: the serial version 8154 void CMSKeepAliveClosure::do_oop(oop obj) { 8155 HeapWord* addr = (HeapWord*)obj; 8156 if (_span.contains(addr) && 8157 !_bit_map->isMarked(addr)) { 8158 _bit_map->mark(addr); 8159 bool simulate_overflow = false; 8160 NOT_PRODUCT( 8161 if (CMSMarkStackOverflowALot && 8162 _collector->simulate_overflow()) { 8163 // simulate a stack overflow 8164 simulate_overflow = true; 8165 } 8166 ) 8167 if (simulate_overflow || !_mark_stack->push(obj)) { 8168 if (_concurrent_precleaning) { 8169 // We dirty the overflown object and let the remark 8170 // phase deal with it. 8171 assert(_collector->overflow_list_is_empty(), "Error"); 8172 // In the case of object arrays, we need to dirty all of 8173 // the cards that the object spans. No locking or atomics 8174 // are needed since no one else can be mutating the mod union 8175 // table. 8176 if (obj->is_objArray()) { 8177 size_t sz = obj->size(); 8178 HeapWord* end_card_addr = 8179 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8180 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8181 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8182 _collector->_modUnionTable.mark_range(redirty_range); 8183 } else { 8184 _collector->_modUnionTable.mark(addr); 8185 } 8186 _collector->_ser_kac_preclean_ovflw++; 8187 } else { 8188 _collector->push_on_overflow_list(obj); 8189 _collector->_ser_kac_ovflw++; 8190 } 8191 } 8192 } 8193 } 8194 8195 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8196 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8197 8198 // CMSParKeepAliveClosure: a parallel version of the above. 8199 // The work queues are private to each closure (thread), 8200 // but (may be) available for stealing by other threads. 8201 void CMSParKeepAliveClosure::do_oop(oop obj) { 8202 HeapWord* addr = (HeapWord*)obj; 8203 if (_span.contains(addr) && 8204 !_bit_map->isMarked(addr)) { 8205 // In general, during recursive tracing, several threads 8206 // may be concurrently getting here; the first one to 8207 // "tag" it, claims it. 8208 if (_bit_map->par_mark(addr)) { 8209 bool res = _work_queue->push(obj); 8210 assert(res, "Low water mark should be much less than capacity"); 8211 // Do a recursive trim in the hope that this will keep 8212 // stack usage lower, but leave some oops for potential stealers 8213 trim_queue(_low_water_mark); 8214 } // Else, another thread got there first 8215 } 8216 } 8217 8218 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8219 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8220 8221 void CMSParKeepAliveClosure::trim_queue(uint max) { 8222 while (_work_queue->size() > max) { 8223 oop new_oop; 8224 if (_work_queue->pop_local(new_oop)) { 8225 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8226 assert(_bit_map->isMarked((HeapWord*)new_oop), 8227 "no white objects on this stack!"); 8228 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8229 // iterate over the oops in this oop, marking and pushing 8230 // the ones in CMS heap (i.e. in _span). 8231 new_oop->oop_iterate(&_mark_and_push); 8232 } 8233 } 8234 } 8235 8236 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8237 CMSCollector* collector, 8238 MemRegion span, CMSBitMap* bit_map, 8239 OopTaskQueue* work_queue): 8240 _collector(collector), 8241 _span(span), 8242 _bit_map(bit_map), 8243 _work_queue(work_queue) { } 8244 8245 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8246 HeapWord* addr = (HeapWord*)obj; 8247 if (_span.contains(addr) && 8248 !_bit_map->isMarked(addr)) { 8249 if (_bit_map->par_mark(addr)) { 8250 bool simulate_overflow = false; 8251 NOT_PRODUCT( 8252 if (CMSMarkStackOverflowALot && 8253 _collector->par_simulate_overflow()) { 8254 // simulate a stack overflow 8255 simulate_overflow = true; 8256 } 8257 ) 8258 if (simulate_overflow || !_work_queue->push(obj)) { 8259 _collector->par_push_on_overflow_list(obj); 8260 _collector->_par_kac_ovflw++; 8261 } 8262 } // Else another thread got there already 8263 } 8264 } 8265 8266 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8267 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8268 8269 ////////////////////////////////////////////////////////////////// 8270 // CMSExpansionCause ///////////////////////////// 8271 ////////////////////////////////////////////////////////////////// 8272 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8273 switch (cause) { 8274 case _no_expansion: 8275 return "No expansion"; 8276 case _satisfy_free_ratio: 8277 return "Free ratio"; 8278 case _satisfy_promotion: 8279 return "Satisfy promotion"; 8280 case _satisfy_allocation: 8281 return "allocation"; 8282 case _allocate_par_lab: 8283 return "Par LAB"; 8284 case _allocate_par_spooling_space: 8285 return "Par Spooling Space"; 8286 case _adaptive_size_policy: 8287 return "Ergonomics"; 8288 default: 8289 return "unknown"; 8290 } 8291 } 8292 8293 void CMSDrainMarkingStackClosure::do_void() { 8294 // the max number to take from overflow list at a time 8295 const size_t num = _mark_stack->capacity()/4; 8296 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8297 "Overflow list should be NULL during concurrent phases"); 8298 while (!_mark_stack->isEmpty() || 8299 // if stack is empty, check the overflow list 8300 _collector->take_from_overflow_list(num, _mark_stack)) { 8301 oop obj = _mark_stack->pop(); 8302 HeapWord* addr = (HeapWord*)obj; 8303 assert(_span.contains(addr), "Should be within span"); 8304 assert(_bit_map->isMarked(addr), "Should be marked"); 8305 assert(obj->is_oop(), "Should be an oop"); 8306 obj->oop_iterate(_keep_alive); 8307 } 8308 } 8309 8310 void CMSParDrainMarkingStackClosure::do_void() { 8311 // drain queue 8312 trim_queue(0); 8313 } 8314 8315 // Trim our work_queue so its length is below max at return 8316 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8317 while (_work_queue->size() > max) { 8318 oop new_oop; 8319 if (_work_queue->pop_local(new_oop)) { 8320 assert(new_oop->is_oop(), "Expected an oop"); 8321 assert(_bit_map->isMarked((HeapWord*)new_oop), 8322 "no white objects on this stack!"); 8323 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8324 // iterate over the oops in this oop, marking and pushing 8325 // the ones in CMS heap (i.e. in _span). 8326 new_oop->oop_iterate(&_mark_and_push); 8327 } 8328 } 8329 } 8330 8331 //////////////////////////////////////////////////////////////////// 8332 // Support for Marking Stack Overflow list handling and related code 8333 //////////////////////////////////////////////////////////////////// 8334 // Much of the following code is similar in shape and spirit to the 8335 // code used in ParNewGC. We should try and share that code 8336 // as much as possible in the future. 8337 8338 #ifndef PRODUCT 8339 // Debugging support for CMSStackOverflowALot 8340 8341 // It's OK to call this multi-threaded; the worst thing 8342 // that can happen is that we'll get a bunch of closely 8343 // spaced simulated overflows, but that's OK, in fact 8344 // probably good as it would exercise the overflow code 8345 // under contention. 8346 bool CMSCollector::simulate_overflow() { 8347 if (_overflow_counter-- <= 0) { // just being defensive 8348 _overflow_counter = CMSMarkStackOverflowInterval; 8349 return true; 8350 } else { 8351 return false; 8352 } 8353 } 8354 8355 bool CMSCollector::par_simulate_overflow() { 8356 return simulate_overflow(); 8357 } 8358 #endif 8359 8360 // Single-threaded 8361 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8362 assert(stack->isEmpty(), "Expected precondition"); 8363 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8364 size_t i = num; 8365 oop cur = _overflow_list; 8366 const markOop proto = markOopDesc::prototype(); 8367 NOT_PRODUCT(ssize_t n = 0;) 8368 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8369 next = oop(cur->mark()); 8370 cur->set_mark(proto); // until proven otherwise 8371 assert(cur->is_oop(), "Should be an oop"); 8372 bool res = stack->push(cur); 8373 assert(res, "Bit off more than can chew?"); 8374 NOT_PRODUCT(n++;) 8375 } 8376 _overflow_list = cur; 8377 #ifndef PRODUCT 8378 assert(_num_par_pushes >= n, "Too many pops?"); 8379 _num_par_pushes -=n; 8380 #endif 8381 return !stack->isEmpty(); 8382 } 8383 8384 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8385 // (MT-safe) Get a prefix of at most "num" from the list. 8386 // The overflow list is chained through the mark word of 8387 // each object in the list. We fetch the entire list, 8388 // break off a prefix of the right size and return the 8389 // remainder. If other threads try to take objects from 8390 // the overflow list at that time, they will wait for 8391 // some time to see if data becomes available. If (and 8392 // only if) another thread places one or more object(s) 8393 // on the global list before we have returned the suffix 8394 // to the global list, we will walk down our local list 8395 // to find its end and append the global list to 8396 // our suffix before returning it. This suffix walk can 8397 // prove to be expensive (quadratic in the amount of traffic) 8398 // when there are many objects in the overflow list and 8399 // there is much producer-consumer contention on the list. 8400 // *NOTE*: The overflow list manipulation code here and 8401 // in ParNewGeneration:: are very similar in shape, 8402 // except that in the ParNew case we use the old (from/eden) 8403 // copy of the object to thread the list via its klass word. 8404 // Because of the common code, if you make any changes in 8405 // the code below, please check the ParNew version to see if 8406 // similar changes might be needed. 8407 // CR 6797058 has been filed to consolidate the common code. 8408 bool CMSCollector::par_take_from_overflow_list(size_t num, 8409 OopTaskQueue* work_q, 8410 int no_of_gc_threads) { 8411 assert(work_q->size() == 0, "First empty local work queue"); 8412 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8413 if (_overflow_list == NULL) { 8414 return false; 8415 } 8416 // Grab the entire list; we'll put back a suffix 8417 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8418 Thread* tid = Thread::current(); 8419 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8420 // set to ParallelGCThreads. 8421 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8422 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8423 // If the list is busy, we spin for a short while, 8424 // sleeping between attempts to get the list. 8425 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8426 os::sleep(tid, sleep_time_millis, false); 8427 if (_overflow_list == NULL) { 8428 // Nothing left to take 8429 return false; 8430 } else if (_overflow_list != BUSY) { 8431 // Try and grab the prefix 8432 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8433 } 8434 } 8435 // If the list was found to be empty, or we spun long 8436 // enough, we give up and return empty-handed. If we leave 8437 // the list in the BUSY state below, it must be the case that 8438 // some other thread holds the overflow list and will set it 8439 // to a non-BUSY state in the future. 8440 if (prefix == NULL || prefix == BUSY) { 8441 // Nothing to take or waited long enough 8442 if (prefix == NULL) { 8443 // Write back the NULL in case we overwrote it with BUSY above 8444 // and it is still the same value. 8445 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8446 } 8447 return false; 8448 } 8449 assert(prefix != NULL && prefix != BUSY, "Error"); 8450 size_t i = num; 8451 oop cur = prefix; 8452 // Walk down the first "num" objects, unless we reach the end. 8453 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8454 if (cur->mark() == NULL) { 8455 // We have "num" or fewer elements in the list, so there 8456 // is nothing to return to the global list. 8457 // Write back the NULL in lieu of the BUSY we wrote 8458 // above, if it is still the same value. 8459 if (_overflow_list == BUSY) { 8460 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8461 } 8462 } else { 8463 // Chop off the suffix and return it to the global list. 8464 assert(cur->mark() != BUSY, "Error"); 8465 oop suffix_head = cur->mark(); // suffix will be put back on global list 8466 cur->set_mark(NULL); // break off suffix 8467 // It's possible that the list is still in the empty(busy) state 8468 // we left it in a short while ago; in that case we may be 8469 // able to place back the suffix without incurring the cost 8470 // of a walk down the list. 8471 oop observed_overflow_list = _overflow_list; 8472 oop cur_overflow_list = observed_overflow_list; 8473 bool attached = false; 8474 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8475 observed_overflow_list = 8476 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8477 if (cur_overflow_list == observed_overflow_list) { 8478 attached = true; 8479 break; 8480 } else cur_overflow_list = observed_overflow_list; 8481 } 8482 if (!attached) { 8483 // Too bad, someone else sneaked in (at least) an element; we'll need 8484 // to do a splice. Find tail of suffix so we can prepend suffix to global 8485 // list. 8486 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8487 oop suffix_tail = cur; 8488 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8489 "Tautology"); 8490 observed_overflow_list = _overflow_list; 8491 do { 8492 cur_overflow_list = observed_overflow_list; 8493 if (cur_overflow_list != BUSY) { 8494 // Do the splice ... 8495 suffix_tail->set_mark(markOop(cur_overflow_list)); 8496 } else { // cur_overflow_list == BUSY 8497 suffix_tail->set_mark(NULL); 8498 } 8499 // ... and try to place spliced list back on overflow_list ... 8500 observed_overflow_list = 8501 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8502 } while (cur_overflow_list != observed_overflow_list); 8503 // ... until we have succeeded in doing so. 8504 } 8505 } 8506 8507 // Push the prefix elements on work_q 8508 assert(prefix != NULL, "control point invariant"); 8509 const markOop proto = markOopDesc::prototype(); 8510 oop next; 8511 NOT_PRODUCT(ssize_t n = 0;) 8512 for (cur = prefix; cur != NULL; cur = next) { 8513 next = oop(cur->mark()); 8514 cur->set_mark(proto); // until proven otherwise 8515 assert(cur->is_oop(), "Should be an oop"); 8516 bool res = work_q->push(cur); 8517 assert(res, "Bit off more than we can chew?"); 8518 NOT_PRODUCT(n++;) 8519 } 8520 #ifndef PRODUCT 8521 assert(_num_par_pushes >= n, "Too many pops?"); 8522 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8523 #endif 8524 return true; 8525 } 8526 8527 // Single-threaded 8528 void CMSCollector::push_on_overflow_list(oop p) { 8529 NOT_PRODUCT(_num_par_pushes++;) 8530 assert(p->is_oop(), "Not an oop"); 8531 preserve_mark_if_necessary(p); 8532 p->set_mark((markOop)_overflow_list); 8533 _overflow_list = p; 8534 } 8535 8536 // Multi-threaded; use CAS to prepend to overflow list 8537 void CMSCollector::par_push_on_overflow_list(oop p) { 8538 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8539 assert(p->is_oop(), "Not an oop"); 8540 par_preserve_mark_if_necessary(p); 8541 oop observed_overflow_list = _overflow_list; 8542 oop cur_overflow_list; 8543 do { 8544 cur_overflow_list = observed_overflow_list; 8545 if (cur_overflow_list != BUSY) { 8546 p->set_mark(markOop(cur_overflow_list)); 8547 } else { 8548 p->set_mark(NULL); 8549 } 8550 observed_overflow_list = 8551 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8552 } while (cur_overflow_list != observed_overflow_list); 8553 } 8554 #undef BUSY 8555 8556 // Single threaded 8557 // General Note on GrowableArray: pushes may silently fail 8558 // because we are (temporarily) out of C-heap for expanding 8559 // the stack. The problem is quite ubiquitous and affects 8560 // a lot of code in the JVM. The prudent thing for GrowableArray 8561 // to do (for now) is to exit with an error. However, that may 8562 // be too draconian in some cases because the caller may be 8563 // able to recover without much harm. For such cases, we 8564 // should probably introduce a "soft_push" method which returns 8565 // an indication of success or failure with the assumption that 8566 // the caller may be able to recover from a failure; code in 8567 // the VM can then be changed, incrementally, to deal with such 8568 // failures where possible, thus, incrementally hardening the VM 8569 // in such low resource situations. 8570 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8571 _preserved_oop_stack.push(p); 8572 _preserved_mark_stack.push(m); 8573 assert(m == p->mark(), "Mark word changed"); 8574 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8575 "bijection"); 8576 } 8577 8578 // Single threaded 8579 void CMSCollector::preserve_mark_if_necessary(oop p) { 8580 markOop m = p->mark(); 8581 if (m->must_be_preserved(p)) { 8582 preserve_mark_work(p, m); 8583 } 8584 } 8585 8586 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8587 markOop m = p->mark(); 8588 if (m->must_be_preserved(p)) { 8589 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8590 // Even though we read the mark word without holding 8591 // the lock, we are assured that it will not change 8592 // because we "own" this oop, so no other thread can 8593 // be trying to push it on the overflow list; see 8594 // the assertion in preserve_mark_work() that checks 8595 // that m == p->mark(). 8596 preserve_mark_work(p, m); 8597 } 8598 } 8599 8600 // We should be able to do this multi-threaded, 8601 // a chunk of stack being a task (this is 8602 // correct because each oop only ever appears 8603 // once in the overflow list. However, it's 8604 // not very easy to completely overlap this with 8605 // other operations, so will generally not be done 8606 // until all work's been completed. Because we 8607 // expect the preserved oop stack (set) to be small, 8608 // it's probably fine to do this single-threaded. 8609 // We can explore cleverer concurrent/overlapped/parallel 8610 // processing of preserved marks if we feel the 8611 // need for this in the future. Stack overflow should 8612 // be so rare in practice and, when it happens, its 8613 // effect on performance so great that this will 8614 // likely just be in the noise anyway. 8615 void CMSCollector::restore_preserved_marks_if_any() { 8616 assert(SafepointSynchronize::is_at_safepoint(), 8617 "world should be stopped"); 8618 assert(Thread::current()->is_ConcurrentGC_thread() || 8619 Thread::current()->is_VM_thread(), 8620 "should be single-threaded"); 8621 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8622 "bijection"); 8623 8624 while (!_preserved_oop_stack.is_empty()) { 8625 oop p = _preserved_oop_stack.pop(); 8626 assert(p->is_oop(), "Should be an oop"); 8627 assert(_span.contains(p), "oop should be in _span"); 8628 assert(p->mark() == markOopDesc::prototype(), 8629 "Set when taken from overflow list"); 8630 markOop m = _preserved_mark_stack.pop(); 8631 p->set_mark(m); 8632 } 8633 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8634 "stacks were cleared above"); 8635 } 8636 8637 #ifndef PRODUCT 8638 bool CMSCollector::no_preserved_marks() const { 8639 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8640 } 8641 #endif 8642 8643 // Transfer some number of overflown objects to usual marking 8644 // stack. Return true if some objects were transferred. 8645 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8646 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8647 (size_t)ParGCDesiredObjsFromOverflowList); 8648 8649 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8650 assert(_collector->overflow_list_is_empty() || res, 8651 "If list is not empty, we should have taken something"); 8652 assert(!res || !_mark_stack->isEmpty(), 8653 "If we took something, it should now be on our stack"); 8654 return res; 8655 } 8656 8657 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8658 size_t res = _sp->block_size_no_stall(addr, _collector); 8659 if (_sp->block_is_obj(addr)) { 8660 if (_live_bit_map->isMarked(addr)) { 8661 // It can't have been dead in a previous cycle 8662 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8663 } else { 8664 _dead_bit_map->mark(addr); // mark the dead object 8665 } 8666 } 8667 // Could be 0, if the block size could not be computed without stalling. 8668 return res; 8669 } 8670 8671 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8672 8673 switch (phase) { 8674 case CMSCollector::InitialMarking: 8675 initialize(true /* fullGC */ , 8676 cause /* cause of the GC */, 8677 true /* recordGCBeginTime */, 8678 true /* recordPreGCUsage */, 8679 false /* recordPeakUsage */, 8680 false /* recordPostGCusage */, 8681 true /* recordAccumulatedGCTime */, 8682 false /* recordGCEndTime */, 8683 false /* countCollection */ ); 8684 break; 8685 8686 case CMSCollector::FinalMarking: 8687 initialize(true /* fullGC */ , 8688 cause /* cause of the GC */, 8689 false /* recordGCBeginTime */, 8690 false /* recordPreGCUsage */, 8691 false /* recordPeakUsage */, 8692 false /* recordPostGCusage */, 8693 true /* recordAccumulatedGCTime */, 8694 false /* recordGCEndTime */, 8695 false /* countCollection */ ); 8696 break; 8697 8698 case CMSCollector::Sweeping: 8699 initialize(true /* fullGC */ , 8700 cause /* cause of the GC */, 8701 false /* recordGCBeginTime */, 8702 false /* recordPreGCUsage */, 8703 true /* recordPeakUsage */, 8704 true /* recordPostGCusage */, 8705 false /* recordAccumulatedGCTime */, 8706 true /* recordGCEndTime */, 8707 true /* countCollection */ ); 8708 break; 8709 8710 default: 8711 ShouldNotReachHere(); 8712 } 8713 }