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