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