1 /* 2 * Copyright (c) 2001, 2019, 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/classLoaderDataGraph.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsGCStats.hpp" 33 #include "gc/cms/cmsHeap.hpp" 34 #include "gc/cms/cmsOopClosures.inline.hpp" 35 #include "gc/cms/cmsVMOperations.hpp" 36 #include "gc/cms/compactibleFreeListSpace.hpp" 37 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 38 #include "gc/cms/concurrentMarkSweepThread.hpp" 39 #include "gc/cms/parNewGeneration.hpp" 40 #include "gc/cms/promotionInfo.inline.hpp" 41 #include "gc/serial/genMarkSweep.hpp" 42 #include "gc/serial/tenuredGeneration.hpp" 43 #include "gc/shared/adaptiveSizePolicy.hpp" 44 #include "gc/shared/cardGeneration.inline.hpp" 45 #include "gc/shared/cardTableRS.hpp" 46 #include "gc/shared/collectedHeap.inline.hpp" 47 #include "gc/shared/collectorCounters.hpp" 48 #include "gc/shared/collectorPolicy.hpp" 49 #include "gc/shared/gcLocker.hpp" 50 #include "gc/shared/gcPolicyCounters.hpp" 51 #include "gc/shared/gcTimer.hpp" 52 #include "gc/shared/gcTrace.hpp" 53 #include "gc/shared/gcTraceTime.inline.hpp" 54 #include "gc/shared/genCollectedHeap.hpp" 55 #include "gc/shared/genOopClosures.inline.hpp" 56 #include "gc/shared/isGCActiveMark.hpp" 57 #include "gc/shared/oopStorageParState.hpp" 58 #include "gc/shared/owstTaskTerminator.hpp" 59 #include "gc/shared/referencePolicy.hpp" 60 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 61 #include "gc/shared/space.inline.hpp" 62 #include "gc/shared/strongRootsScope.hpp" 63 #include "gc/shared/taskqueue.inline.hpp" 64 #include "gc/shared/weakProcessor.hpp" 65 #include "gc/shared/workerPolicy.hpp" 66 #include "logging/log.hpp" 67 #include "logging/logStream.hpp" 68 #include "memory/allocation.hpp" 69 #include "memory/binaryTreeDictionary.inline.hpp" 70 #include "memory/iterator.inline.hpp" 71 #include "memory/padded.hpp" 72 #include "memory/resourceArea.hpp" 73 #include "oops/access.inline.hpp" 74 #include "oops/oop.inline.hpp" 75 #include "prims/jvmtiExport.hpp" 76 #include "runtime/atomic.hpp" 77 #include "runtime/flags/flagSetting.hpp" 78 #include "runtime/globals_extension.hpp" 79 #include "runtime/handles.inline.hpp" 80 #include "runtime/java.hpp" 81 #include "runtime/orderAccess.hpp" 82 #include "runtime/timer.hpp" 83 #include "runtime/vmThread.hpp" 84 #include "services/memoryService.hpp" 85 #include "services/runtimeService.hpp" 86 #include "utilities/align.hpp" 87 #include "utilities/stack.inline.hpp" 88 #if INCLUDE_JVMCI 89 #include "jvmci/jvmci.hpp" 90 #endif 91 92 // statics 93 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 94 bool CMSCollector::_full_gc_requested = false; 95 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 96 97 ////////////////////////////////////////////////////////////////// 98 // In support of CMS/VM thread synchronization 99 ////////////////////////////////////////////////////////////////// 100 // We split use of the CGC_lock into 2 "levels". 101 // The low-level locking is of the usual CGC_lock monitor. We introduce 102 // a higher level "token" (hereafter "CMS token") built on top of the 103 // low level monitor (hereafter "CGC lock"). 104 // The token-passing protocol gives priority to the VM thread. The 105 // CMS-lock doesn't provide any fairness guarantees, but clients 106 // should ensure that it is only held for very short, bounded 107 // durations. 108 // 109 // When either of the CMS thread or the VM thread is involved in 110 // collection operations during which it does not want the other 111 // thread to interfere, it obtains the CMS token. 112 // 113 // If either thread tries to get the token while the other has 114 // it, that thread waits. However, if the VM thread and CMS thread 115 // both want the token, then the VM thread gets priority while the 116 // CMS thread waits. This ensures, for instance, that the "concurrent" 117 // phases of the CMS thread's work do not block out the VM thread 118 // for long periods of time as the CMS thread continues to hog 119 // the token. (See bug 4616232). 120 // 121 // The baton-passing functions are, however, controlled by the 122 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 123 // and here the low-level CMS lock, not the high level token, 124 // ensures mutual exclusion. 125 // 126 // Two important conditions that we have to satisfy: 127 // 1. if a thread does a low-level wait on the CMS lock, then it 128 // relinquishes the CMS token if it were holding that token 129 // when it acquired the low-level CMS lock. 130 // 2. any low-level notifications on the low-level lock 131 // should only be sent when a thread has relinquished the token. 132 // 133 // In the absence of either property, we'd have potential deadlock. 134 // 135 // We protect each of the CMS (concurrent and sequential) phases 136 // with the CMS _token_, not the CMS _lock_. 137 // 138 // The only code protected by CMS lock is the token acquisition code 139 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 140 // baton-passing code. 141 // 142 // Unfortunately, i couldn't come up with a good abstraction to factor and 143 // hide the naked CGC_lock manipulation in the baton-passing code 144 // further below. That's something we should try to do. Also, the proof 145 // of correctness of this 2-level locking scheme is far from obvious, 146 // and potentially quite slippery. We have an uneasy suspicion, for instance, 147 // that there may be a theoretical possibility of delay/starvation in the 148 // low-level lock/wait/notify scheme used for the baton-passing because of 149 // potential interference with the priority scheme embodied in the 150 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 151 // invocation further below and marked with "XXX 20011219YSR". 152 // Indeed, as we note elsewhere, this may become yet more slippery 153 // in the presence of multiple CMS and/or multiple VM threads. XXX 154 155 class CMSTokenSync: public StackObj { 156 private: 157 bool _is_cms_thread; 158 public: 159 CMSTokenSync(bool is_cms_thread): 160 _is_cms_thread(is_cms_thread) { 161 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 162 "Incorrect argument to constructor"); 163 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 164 } 165 166 ~CMSTokenSync() { 167 assert(_is_cms_thread ? 168 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 169 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 170 "Incorrect state"); 171 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 172 } 173 }; 174 175 // Convenience class that does a CMSTokenSync, and then acquires 176 // upto three locks. 177 class CMSTokenSyncWithLocks: public CMSTokenSync { 178 private: 179 // Note: locks are acquired in textual declaration order 180 // and released in the opposite order 181 MutexLockerEx _locker1, _locker2, _locker3; 182 public: 183 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 184 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 185 CMSTokenSync(is_cms_thread), 186 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 187 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 188 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 189 { } 190 }; 191 192 193 ////////////////////////////////////////////////////////////////// 194 // Concurrent Mark-Sweep Generation ///////////////////////////// 195 ////////////////////////////////////////////////////////////////// 196 197 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 198 199 // This struct contains per-thread things necessary to support parallel 200 // young-gen collection. 201 class CMSParGCThreadState: public CHeapObj<mtGC> { 202 public: 203 CompactibleFreeListSpaceLAB lab; 204 PromotionInfo promo; 205 206 // Constructor. 207 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 208 promo.setSpace(cfls); 209 } 210 }; 211 212 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 213 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 214 CardGeneration(rs, initial_byte_size, ct), 215 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 216 _did_compact(false) 217 { 218 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 219 HeapWord* end = (HeapWord*) _virtual_space.high(); 220 221 _direct_allocated_words = 0; 222 NOT_PRODUCT( 223 _numObjectsPromoted = 0; 224 _numWordsPromoted = 0; 225 _numObjectsAllocated = 0; 226 _numWordsAllocated = 0; 227 ) 228 229 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 230 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 231 _cmsSpace->_old_gen = this; 232 233 _gc_stats = new CMSGCStats(); 234 235 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 236 // offsets match. The ability to tell free chunks from objects 237 // depends on this property. 238 debug_only( 239 FreeChunk* junk = NULL; 240 assert(UseCompressedClassPointers || 241 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 242 "Offset of FreeChunk::_prev within FreeChunk must match" 243 " that of OopDesc::_klass within OopDesc"); 244 ) 245 246 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 247 for (uint i = 0; i < ParallelGCThreads; i++) { 248 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 249 } 250 251 _incremental_collection_failed = false; 252 // The "dilatation_factor" is the expansion that can occur on 253 // account of the fact that the minimum object size in the CMS 254 // generation may be larger than that in, say, a contiguous young 255 // generation. 256 // Ideally, in the calculation below, we'd compute the dilatation 257 // factor as: MinChunkSize/(promoting_gen's min object size) 258 // Since we do not have such a general query interface for the 259 // promoting generation, we'll instead just use the minimum 260 // object size (which today is a header's worth of space); 261 // note that all arithmetic is in units of HeapWords. 262 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 263 assert(_dilatation_factor >= 1.0, "from previous assert"); 264 } 265 266 267 // The field "_initiating_occupancy" represents the occupancy percentage 268 // at which we trigger a new collection cycle. Unless explicitly specified 269 // via CMSInitiatingOccupancyFraction (argument "io" below), it 270 // is calculated by: 271 // 272 // Let "f" be MinHeapFreeRatio in 273 // 274 // _initiating_occupancy = 100-f + 275 // f * (CMSTriggerRatio/100) 276 // where CMSTriggerRatio is the argument "tr" below. 277 // 278 // That is, if we assume the heap is at its desired maximum occupancy at the 279 // end of a collection, we let CMSTriggerRatio of the (purported) free 280 // space be allocated before initiating a new collection cycle. 281 // 282 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 283 assert(io <= 100 && tr <= 100, "Check the arguments"); 284 if (io >= 0) { 285 _initiating_occupancy = (double)io / 100.0; 286 } else { 287 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 288 (double)(tr * MinHeapFreeRatio) / 100.0) 289 / 100.0; 290 } 291 } 292 293 void ConcurrentMarkSweepGeneration::ref_processor_init() { 294 assert(collector() != NULL, "no collector"); 295 collector()->ref_processor_init(); 296 } 297 298 void CMSCollector::ref_processor_init() { 299 if (_ref_processor == NULL) { 300 // Allocate and initialize a reference processor 301 _ref_processor = 302 new ReferenceProcessor(&_span_based_discoverer, 303 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 304 ParallelGCThreads, // mt processing degree 305 _cmsGen->refs_discovery_is_mt(), // mt discovery 306 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 307 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 308 &_is_alive_closure, // closure for liveness info 309 false); // disable adjusting number of processing threads 310 // Initialize the _ref_processor field of CMSGen 311 _cmsGen->set_ref_processor(_ref_processor); 312 313 } 314 } 315 316 AdaptiveSizePolicy* CMSCollector::size_policy() { 317 return CMSHeap::heap()->size_policy(); 318 } 319 320 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 321 322 const char* gen_name = "old"; 323 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 324 // Generation Counters - generation 1, 1 subspace 325 _gen_counters = new GenerationCounters(gen_name, 1, 1, 326 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 327 328 _space_counters = new GSpaceCounters(gen_name, 0, 329 _virtual_space.reserved_size(), 330 this, _gen_counters); 331 } 332 333 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 334 _cms_gen(cms_gen) 335 { 336 assert(alpha <= 100, "bad value"); 337 _saved_alpha = alpha; 338 339 // Initialize the alphas to the bootstrap value of 100. 340 _gc0_alpha = _cms_alpha = 100; 341 342 _cms_begin_time.update(); 343 _cms_end_time.update(); 344 345 _gc0_duration = 0.0; 346 _gc0_period = 0.0; 347 _gc0_promoted = 0; 348 349 _cms_duration = 0.0; 350 _cms_period = 0.0; 351 _cms_allocated = 0; 352 353 _cms_used_at_gc0_begin = 0; 354 _cms_used_at_gc0_end = 0; 355 _allow_duty_cycle_reduction = false; 356 _valid_bits = 0; 357 } 358 359 double CMSStats::cms_free_adjustment_factor(size_t free) const { 360 // TBD: CR 6909490 361 return 1.0; 362 } 363 364 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 365 } 366 367 // If promotion failure handling is on use 368 // the padded average size of the promotion for each 369 // young generation collection. 370 double CMSStats::time_until_cms_gen_full() const { 371 size_t cms_free = _cms_gen->cmsSpace()->free(); 372 CMSHeap* heap = CMSHeap::heap(); 373 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 374 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 375 if (cms_free > expected_promotion) { 376 // Start a cms collection if there isn't enough space to promote 377 // for the next young collection. Use the padded average as 378 // a safety factor. 379 cms_free -= expected_promotion; 380 381 // Adjust by the safety factor. 382 double cms_free_dbl = (double)cms_free; 383 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 384 // Apply a further correction factor which tries to adjust 385 // for recent occurance of concurrent mode failures. 386 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 387 cms_free_dbl = cms_free_dbl * cms_adjustment; 388 389 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 390 cms_free, expected_promotion); 391 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 392 // Add 1 in case the consumption rate goes to zero. 393 return cms_free_dbl / (cms_consumption_rate() + 1.0); 394 } 395 return 0.0; 396 } 397 398 // Compare the duration of the cms collection to the 399 // time remaining before the cms generation is empty. 400 // Note that the time from the start of the cms collection 401 // to the start of the cms sweep (less than the total 402 // duration of the cms collection) can be used. This 403 // has been tried and some applications experienced 404 // promotion failures early in execution. This was 405 // possibly because the averages were not accurate 406 // enough at the beginning. 407 double CMSStats::time_until_cms_start() const { 408 // We add "gc0_period" to the "work" calculation 409 // below because this query is done (mostly) at the 410 // end of a scavenge, so we need to conservatively 411 // account for that much possible delay 412 // in the query so as to avoid concurrent mode failures 413 // due to starting the collection just a wee bit too 414 // late. 415 double work = cms_duration() + gc0_period(); 416 double deadline = time_until_cms_gen_full(); 417 // If a concurrent mode failure occurred recently, we want to be 418 // more conservative and halve our expected time_until_cms_gen_full() 419 if (work > deadline) { 420 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 421 cms_duration(), gc0_period(), time_until_cms_gen_full()); 422 return 0.0; 423 } 424 return work - deadline; 425 } 426 427 #ifndef PRODUCT 428 void CMSStats::print_on(outputStream *st) const { 429 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 430 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 431 gc0_duration(), gc0_period(), gc0_promoted()); 432 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 433 cms_duration(), cms_period(), cms_allocated()); 434 st->print(",cms_since_beg=%g,cms_since_end=%g", 435 cms_time_since_begin(), cms_time_since_end()); 436 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 437 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 438 439 if (valid()) { 440 st->print(",promo_rate=%g,cms_alloc_rate=%g", 441 promotion_rate(), cms_allocation_rate()); 442 st->print(",cms_consumption_rate=%g,time_until_full=%g", 443 cms_consumption_rate(), time_until_cms_gen_full()); 444 } 445 st->cr(); 446 } 447 #endif // #ifndef PRODUCT 448 449 CMSCollector::CollectorState CMSCollector::_collectorState = 450 CMSCollector::Idling; 451 bool CMSCollector::_foregroundGCIsActive = false; 452 bool CMSCollector::_foregroundGCShouldWait = false; 453 454 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 455 CardTableRS* ct, 456 ConcurrentMarkSweepPolicy* cp): 457 _overflow_list(NULL), 458 _conc_workers(NULL), // may be set later 459 _completed_initialization(false), 460 _collection_count_start(0), 461 _should_unload_classes(CMSClassUnloadingEnabled), 462 _concurrent_cycles_since_last_unload(0), 463 _roots_scanning_options(GenCollectedHeap::SO_None), 464 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 465 _verifying(false), 466 _collector_policy(cp), 467 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 468 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 469 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 470 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 471 _cms_start_registered(false), 472 _cmsGen(cmsGen), 473 // Adjust span to cover old (cms) gen 474 _span(cmsGen->reserved()), 475 _ct(ct), 476 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 477 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 478 -1 /* lock-free */, "No_lock" /* dummy */), 479 _restart_addr(NULL), 480 _ser_pmc_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_preclean_ovflw(0), 484 _ser_kac_ovflw(0), 485 _par_kac_ovflw(0), 486 #ifndef PRODUCT 487 _num_par_pushes(0), 488 #endif 489 _span_based_discoverer(_span), 490 _ref_processor(NULL), // will be set later 491 // Construct the is_alive_closure with _span & markBitMap 492 _is_alive_closure(_span, &_markBitMap), 493 _modUnionClosurePar(&_modUnionTable), 494 _between_prologue_and_epilogue(false), 495 _abort_preclean(false), 496 _start_sampling(false), 497 _stats(cmsGen), 498 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 499 //verify that this lock should be acquired with safepoint check. 500 Monitor::_safepoint_check_sometimes)), 501 _eden_chunk_array(NULL), // may be set in ctor body 502 _eden_chunk_index(0), // -- ditto -- 503 _eden_chunk_capacity(0), // -- ditto -- 504 _survivor_chunk_array(NULL), // -- ditto -- 505 _survivor_chunk_index(0), // -- ditto -- 506 _survivor_chunk_capacity(0), // -- ditto -- 507 _survivor_plab_array(NULL) // -- ditto -- 508 { 509 // Now expand the span and allocate the collection support structures 510 // (MUT, marking bit map etc.) to cover both generations subject to 511 // collection. 512 513 // For use by dirty card to oop closures. 514 _cmsGen->cmsSpace()->set_collector(this); 515 516 // Allocate MUT and marking bit map 517 { 518 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 519 if (!_markBitMap.allocate(_span)) { 520 log_warning(gc)("Failed to allocate CMS Bit Map"); 521 return; 522 } 523 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 524 } 525 { 526 _modUnionTable.allocate(_span); 527 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 528 } 529 530 if (!_markStack.allocate(MarkStackSize)) { 531 log_warning(gc)("Failed to allocate CMS Marking Stack"); 532 return; 533 } 534 535 // Support for multi-threaded concurrent phases 536 if (CMSConcurrentMTEnabled) { 537 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 538 // just for now 539 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 540 } 541 if (ConcGCThreads > 1) { 542 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 543 ConcGCThreads, true); 544 if (_conc_workers == NULL) { 545 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 546 CMSConcurrentMTEnabled = false; 547 } else { 548 _conc_workers->initialize_workers(); 549 } 550 } else { 551 CMSConcurrentMTEnabled = false; 552 } 553 } 554 if (!CMSConcurrentMTEnabled) { 555 ConcGCThreads = 0; 556 } else { 557 // Turn off CMSCleanOnEnter optimization temporarily for 558 // the MT case where it's not fixed yet; see 6178663. 559 CMSCleanOnEnter = false; 560 } 561 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 562 "Inconsistency"); 563 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 564 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 565 566 // Parallel task queues; these are shared for the 567 // concurrent and stop-world phases of CMS, but 568 // are not shared with parallel scavenge (ParNew). 569 { 570 uint i; 571 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 572 573 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 574 || ParallelRefProcEnabled) 575 && num_queues > 0) { 576 _task_queues = new OopTaskQueueSet(num_queues); 577 if (_task_queues == NULL) { 578 log_warning(gc)("task_queues allocation failure."); 579 return; 580 } 581 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 582 for (i = 0; i < num_queues; i++) { 583 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 584 if (q == NULL) { 585 log_warning(gc)("work_queue allocation failure."); 586 return; 587 } 588 _task_queues->register_queue(i, q); 589 } 590 for (i = 0; i < num_queues; i++) { 591 _task_queues->queue(i)->initialize(); 592 } 593 } 594 } 595 596 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 597 598 // Clip CMSBootstrapOccupancy between 0 and 100. 599 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 600 601 // Now tell CMS generations the identity of their collector 602 ConcurrentMarkSweepGeneration::set_collector(this); 603 604 // Create & start a CMS thread for this CMS collector 605 _cmsThread = ConcurrentMarkSweepThread::start(this); 606 assert(cmsThread() != NULL, "CMS Thread should have been created"); 607 assert(cmsThread()->collector() == this, 608 "CMS Thread should refer to this gen"); 609 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 610 611 // Support for parallelizing young gen rescan 612 CMSHeap* heap = CMSHeap::heap(); 613 _young_gen = heap->young_gen(); 614 if (heap->supports_inline_contig_alloc()) { 615 _top_addr = heap->top_addr(); 616 _end_addr = heap->end_addr(); 617 assert(_young_gen != NULL, "no _young_gen"); 618 _eden_chunk_index = 0; 619 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 620 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 621 } 622 623 // Support for parallelizing survivor space rescan 624 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 625 const size_t max_plab_samples = 626 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 627 628 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 629 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 631 _survivor_chunk_capacity = max_plab_samples; 632 for (uint i = 0; i < ParallelGCThreads; i++) { 633 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 634 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 635 assert(cur->end() == 0, "Should be 0"); 636 assert(cur->array() == vec, "Should be vec"); 637 assert(cur->capacity() == max_plab_samples, "Error"); 638 } 639 } 640 641 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 642 _gc_counters = new CollectorCounters("CMS full collection pauses", 1); 643 _cgc_counters = new CollectorCounters("CMS concurrent cycle pauses", 2); 644 _completed_initialization = true; 645 _inter_sweep_timer.start(); // start of time 646 } 647 648 const char* ConcurrentMarkSweepGeneration::name() const { 649 return "concurrent mark-sweep generation"; 650 } 651 void ConcurrentMarkSweepGeneration::update_counters() { 652 if (UsePerfData) { 653 _space_counters->update_all(); 654 _gen_counters->update_all(); 655 } 656 } 657 658 // this is an optimized version of update_counters(). it takes the 659 // used value as a parameter rather than computing it. 660 // 661 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 662 if (UsePerfData) { 663 _space_counters->update_used(used); 664 _space_counters->update_capacity(); 665 _gen_counters->update_all(); 666 } 667 } 668 669 void ConcurrentMarkSweepGeneration::print() const { 670 Generation::print(); 671 cmsSpace()->print(); 672 } 673 674 #ifndef PRODUCT 675 void ConcurrentMarkSweepGeneration::print_statistics() { 676 cmsSpace()->printFLCensus(0); 677 } 678 #endif 679 680 size_t 681 ConcurrentMarkSweepGeneration::contiguous_available() const { 682 // dld proposes an improvement in precision here. If the committed 683 // part of the space ends in a free block we should add that to 684 // uncommitted size in the calculation below. Will make this 685 // change later, staying with the approximation below for the 686 // time being. -- ysr. 687 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 688 } 689 690 size_t 691 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 692 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 693 } 694 695 size_t ConcurrentMarkSweepGeneration::max_available() const { 696 return free() + _virtual_space.uncommitted_size(); 697 } 698 699 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 700 size_t available = max_available(); 701 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 702 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 703 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 704 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 705 return res; 706 } 707 708 // At a promotion failure dump information on block layout in heap 709 // (cms old generation). 710 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 711 Log(gc, promotion) log; 712 if (log.is_trace()) { 713 LogStream ls(log.trace()); 714 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 715 } 716 } 717 718 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 719 // Clear the promotion information. These pointers can be adjusted 720 // along with all the other pointers into the heap but 721 // compaction is expected to be a rare event with 722 // a heap using cms so don't do it without seeing the need. 723 for (uint i = 0; i < ParallelGCThreads; i++) { 724 _par_gc_thread_states[i]->promo.reset(); 725 } 726 } 727 728 void ConcurrentMarkSweepGeneration::compute_new_size() { 729 assert_locked_or_safepoint(Heap_lock); 730 731 // If incremental collection failed, we just want to expand 732 // to the limit. 733 if (incremental_collection_failed()) { 734 clear_incremental_collection_failed(); 735 grow_to_reserved(); 736 return; 737 } 738 739 // The heap has been compacted but not reset yet. 740 // Any metric such as free() or used() will be incorrect. 741 742 CardGeneration::compute_new_size(); 743 744 // Reset again after a possible resizing 745 if (did_compact()) { 746 cmsSpace()->reset_after_compaction(); 747 } 748 } 749 750 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 751 assert_locked_or_safepoint(Heap_lock); 752 753 // If incremental collection failed, we just want to expand 754 // to the limit. 755 if (incremental_collection_failed()) { 756 clear_incremental_collection_failed(); 757 grow_to_reserved(); 758 return; 759 } 760 761 double free_percentage = ((double) free()) / capacity(); 762 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 763 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 764 765 // compute expansion delta needed for reaching desired free percentage 766 if (free_percentage < desired_free_percentage) { 767 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 768 assert(desired_capacity >= capacity(), "invalid expansion size"); 769 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 770 Log(gc) log; 771 if (log.is_trace()) { 772 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 773 log.trace("From compute_new_size: "); 774 log.trace(" Free fraction %f", free_percentage); 775 log.trace(" Desired free fraction %f", desired_free_percentage); 776 log.trace(" Maximum free fraction %f", maximum_free_percentage); 777 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 778 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 779 CMSHeap* heap = CMSHeap::heap(); 780 size_t young_size = heap->young_gen()->capacity(); 781 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 782 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 783 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 784 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 785 } 786 // safe if expansion fails 787 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 788 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 789 } else { 790 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 791 assert(desired_capacity <= capacity(), "invalid expansion size"); 792 size_t shrink_bytes = capacity() - desired_capacity; 793 // Don't shrink unless the delta is greater than the minimum shrink we want 794 if (shrink_bytes >= MinHeapDeltaBytes) { 795 shrink_free_list_by(shrink_bytes); 796 } 797 } 798 } 799 800 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 801 return cmsSpace()->freelistLock(); 802 } 803 804 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 805 CMSSynchronousYieldRequest yr; 806 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 807 return have_lock_and_allocate(size, tlab); 808 } 809 810 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 811 bool tlab /* ignored */) { 812 assert_lock_strong(freelistLock()); 813 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 814 HeapWord* res = cmsSpace()->allocate(adjustedSize); 815 // Allocate the object live (grey) if the background collector has 816 // started marking. This is necessary because the marker may 817 // have passed this address and consequently this object will 818 // not otherwise be greyed and would be incorrectly swept up. 819 // Note that if this object contains references, the writing 820 // of those references will dirty the card containing this object 821 // allowing the object to be blackened (and its references scanned) 822 // either during a preclean phase or at the final checkpoint. 823 if (res != NULL) { 824 // We may block here with an uninitialized object with 825 // its mark-bit or P-bits not yet set. Such objects need 826 // to be safely navigable by block_start(). 827 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 828 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 829 collector()->direct_allocated(res, adjustedSize); 830 _direct_allocated_words += adjustedSize; 831 // allocation counters 832 NOT_PRODUCT( 833 _numObjectsAllocated++; 834 _numWordsAllocated += (int)adjustedSize; 835 ) 836 } 837 return res; 838 } 839 840 // In the case of direct allocation by mutators in a generation that 841 // is being concurrently collected, the object must be allocated 842 // live (grey) if the background collector has started marking. 843 // This is necessary because the marker may 844 // have passed this address and consequently this object will 845 // not otherwise be greyed and would be incorrectly swept up. 846 // Note that if this object contains references, the writing 847 // of those references will dirty the card containing this object 848 // allowing the object to be blackened (and its references scanned) 849 // either during a preclean phase or at the final checkpoint. 850 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 851 assert(_markBitMap.covers(start, size), "Out of bounds"); 852 if (_collectorState >= Marking) { 853 MutexLockerEx y(_markBitMap.lock(), 854 Mutex::_no_safepoint_check_flag); 855 // [see comments preceding SweepClosure::do_blk() below for details] 856 // 857 // Can the P-bits be deleted now? JJJ 858 // 859 // 1. need to mark the object as live so it isn't collected 860 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 861 // 3. need to mark the end of the object so marking, precleaning or sweeping 862 // can skip over uninitialized or unparsable objects. An allocated 863 // object is considered uninitialized for our purposes as long as 864 // its klass word is NULL. All old gen objects are parsable 865 // as soon as they are initialized.) 866 _markBitMap.mark(start); // object is live 867 _markBitMap.mark(start + 1); // object is potentially uninitialized? 868 _markBitMap.mark(start + size - 1); 869 // mark end of object 870 } 871 // check that oop looks uninitialized 872 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 873 } 874 875 void CMSCollector::promoted(bool par, HeapWord* start, 876 bool is_obj_array, size_t obj_size) { 877 assert(_markBitMap.covers(start), "Out of bounds"); 878 // See comment in direct_allocated() about when objects should 879 // be allocated live. 880 if (_collectorState >= Marking) { 881 // we already hold the marking bit map lock, taken in 882 // the prologue 883 if (par) { 884 _markBitMap.par_mark(start); 885 } else { 886 _markBitMap.mark(start); 887 } 888 // We don't need to mark the object as uninitialized (as 889 // in direct_allocated above) because this is being done with the 890 // world stopped and the object will be initialized by the 891 // time the marking, precleaning or sweeping get to look at it. 892 // But see the code for copying objects into the CMS generation, 893 // where we need to ensure that concurrent readers of the 894 // block offset table are able to safely navigate a block that 895 // is in flux from being free to being allocated (and in 896 // transition while being copied into) and subsequently 897 // becoming a bona-fide object when the copy/promotion is complete. 898 assert(SafepointSynchronize::is_at_safepoint(), 899 "expect promotion only at safepoints"); 900 901 if (_collectorState < Sweeping) { 902 // Mark the appropriate cards in the modUnionTable, so that 903 // this object gets scanned before the sweep. If this is 904 // not done, CMS generation references in the object might 905 // not get marked. 906 // For the case of arrays, which are otherwise precisely 907 // marked, we need to dirty the entire array, not just its head. 908 if (is_obj_array) { 909 // The [par_]mark_range() method expects mr.end() below to 910 // be aligned to the granularity of a bit's representation 911 // in the heap. In the case of the MUT below, that's a 912 // card size. 913 MemRegion mr(start, 914 align_up(start + obj_size, 915 CardTable::card_size /* bytes */)); 916 if (par) { 917 _modUnionTable.par_mark_range(mr); 918 } else { 919 _modUnionTable.mark_range(mr); 920 } 921 } else { // not an obj array; we can just mark the head 922 if (par) { 923 _modUnionTable.par_mark(start); 924 } else { 925 _modUnionTable.mark(start); 926 } 927 } 928 } 929 } 930 } 931 932 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 933 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 934 // allocate, copy and if necessary update promoinfo -- 935 // delegate to underlying space. 936 assert_lock_strong(freelistLock()); 937 938 #ifndef PRODUCT 939 if (CMSHeap::heap()->promotion_should_fail()) { 940 return NULL; 941 } 942 #endif // #ifndef PRODUCT 943 944 oop res = _cmsSpace->promote(obj, obj_size); 945 if (res == NULL) { 946 // expand and retry 947 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 948 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 949 // Since this is the old generation, we don't try to promote 950 // into a more senior generation. 951 res = _cmsSpace->promote(obj, obj_size); 952 } 953 if (res != NULL) { 954 // See comment in allocate() about when objects should 955 // be allocated live. 956 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 957 collector()->promoted(false, // Not parallel 958 (HeapWord*)res, obj->is_objArray(), obj_size); 959 // promotion counters 960 NOT_PRODUCT( 961 _numObjectsPromoted++; 962 _numWordsPromoted += 963 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 964 ) 965 } 966 return res; 967 } 968 969 970 // IMPORTANT: Notes on object size recognition in CMS. 971 // --------------------------------------------------- 972 // A block of storage in the CMS generation is always in 973 // one of three states. A free block (FREE), an allocated 974 // object (OBJECT) whose size() method reports the correct size, 975 // and an intermediate state (TRANSIENT) in which its size cannot 976 // be accurately determined. 977 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 978 // ----------------------------------------------------- 979 // FREE: klass_word & 1 == 1; mark_word holds block size 980 // 981 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 982 // obj->size() computes correct size 983 // 984 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 985 // 986 // STATE IDENTIFICATION: (64 bit+COOPS) 987 // ------------------------------------ 988 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 989 // 990 // OBJECT: klass_word installed; klass_word != 0; 991 // obj->size() computes correct size 992 // 993 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 994 // 995 // 996 // STATE TRANSITION DIAGRAM 997 // 998 // mut / parnew mut / parnew 999 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1000 // ^ | 1001 // |------------------------ DEAD <------------------------------------| 1002 // sweep mut 1003 // 1004 // While a block is in TRANSIENT state its size cannot be determined 1005 // so readers will either need to come back later or stall until 1006 // the size can be determined. Note that for the case of direct 1007 // allocation, P-bits, when available, may be used to determine the 1008 // size of an object that may not yet have been initialized. 1009 1010 // Things to support parallel young-gen collection. 1011 oop 1012 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1013 oop old, markOop m, 1014 size_t word_sz) { 1015 #ifndef PRODUCT 1016 if (CMSHeap::heap()->promotion_should_fail()) { 1017 return NULL; 1018 } 1019 #endif // #ifndef PRODUCT 1020 1021 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1022 PromotionInfo* promoInfo = &ps->promo; 1023 // if we are tracking promotions, then first ensure space for 1024 // promotion (including spooling space for saving header if necessary). 1025 // then allocate and copy, then track promoted info if needed. 1026 // When tracking (see PromotionInfo::track()), the mark word may 1027 // be displaced and in this case restoration of the mark word 1028 // occurs in the (oop_since_save_marks_)iterate phase. 1029 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1030 // Out of space for allocating spooling buffers; 1031 // try expanding and allocating spooling buffers. 1032 if (!expand_and_ensure_spooling_space(promoInfo)) { 1033 return NULL; 1034 } 1035 } 1036 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1037 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1038 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1039 if (obj_ptr == NULL) { 1040 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1041 if (obj_ptr == NULL) { 1042 return NULL; 1043 } 1044 } 1045 oop obj = oop(obj_ptr); 1046 OrderAccess::storestore(); 1047 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1048 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1049 // IMPORTANT: See note on object initialization for CMS above. 1050 // Otherwise, copy the object. Here we must be careful to insert the 1051 // klass pointer last, since this marks the block as an allocated object. 1052 // Except with compressed oops it's the mark word. 1053 HeapWord* old_ptr = (HeapWord*)old; 1054 // Restore the mark word copied above. 1055 obj->set_mark_raw(m); 1056 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1057 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1058 OrderAccess::storestore(); 1059 1060 if (UseCompressedClassPointers) { 1061 // Copy gap missed by (aligned) header size calculation below 1062 obj->set_klass_gap(old->klass_gap()); 1063 } 1064 if (word_sz > (size_t)oopDesc::header_size()) { 1065 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1066 obj_ptr + oopDesc::header_size(), 1067 word_sz - oopDesc::header_size()); 1068 } 1069 1070 // Now we can track the promoted object, if necessary. We take care 1071 // to delay the transition from uninitialized to full object 1072 // (i.e., insertion of klass pointer) until after, so that it 1073 // atomically becomes a promoted object. 1074 if (promoInfo->tracking()) { 1075 promoInfo->track((PromotedObject*)obj, old->klass()); 1076 } 1077 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1078 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1079 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1080 1081 // Finally, install the klass pointer (this should be volatile). 1082 OrderAccess::storestore(); 1083 obj->set_klass(old->klass()); 1084 // We should now be able to calculate the right size for this object 1085 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1086 1087 collector()->promoted(true, // parallel 1088 obj_ptr, old->is_objArray(), word_sz); 1089 1090 NOT_PRODUCT( 1091 Atomic::inc(&_numObjectsPromoted); 1092 Atomic::add(alloc_sz, &_numWordsPromoted); 1093 ) 1094 1095 return obj; 1096 } 1097 1098 void 1099 ConcurrentMarkSweepGeneration:: 1100 par_promote_alloc_done(int thread_num) { 1101 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1102 ps->lab.retire(thread_num); 1103 } 1104 1105 void 1106 ConcurrentMarkSweepGeneration:: 1107 par_oop_since_save_marks_iterate_done(int thread_num) { 1108 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1109 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1110 ps->promo.promoted_oops_iterate(dummy_cl); 1111 1112 // Because card-scanning has been completed, subsequent phases 1113 // (e.g., reference processing) will not need to recognize which 1114 // objects have been promoted during this GC. So, we can now disable 1115 // promotion tracking. 1116 ps->promo.stopTrackingPromotions(); 1117 } 1118 1119 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1120 size_t size, 1121 bool tlab) 1122 { 1123 // We allow a STW collection only if a full 1124 // collection was requested. 1125 return full || should_allocate(size, tlab); // FIX ME !!! 1126 // This and promotion failure handling are connected at the 1127 // hip and should be fixed by untying them. 1128 } 1129 1130 bool CMSCollector::shouldConcurrentCollect() { 1131 LogTarget(Trace, gc) log; 1132 1133 if (_full_gc_requested) { 1134 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1135 return true; 1136 } 1137 1138 FreelistLocker x(this); 1139 // ------------------------------------------------------------------ 1140 // Print out lots of information which affects the initiation of 1141 // a collection. 1142 if (log.is_enabled() && stats().valid()) { 1143 log.print("CMSCollector shouldConcurrentCollect: "); 1144 1145 LogStream out(log); 1146 stats().print_on(&out); 1147 1148 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1149 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1150 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1151 log.print("promotion_rate=%g", stats().promotion_rate()); 1152 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1153 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1154 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1155 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1156 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1157 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1158 } 1159 // ------------------------------------------------------------------ 1160 1161 // If the estimated time to complete a cms collection (cms_duration()) 1162 // is less than the estimated time remaining until the cms generation 1163 // is full, start a collection. 1164 if (!UseCMSInitiatingOccupancyOnly) { 1165 if (stats().valid()) { 1166 if (stats().time_until_cms_start() == 0.0) { 1167 return true; 1168 } 1169 } else { 1170 // We want to conservatively collect somewhat early in order 1171 // to try and "bootstrap" our CMS/promotion statistics; 1172 // this branch will not fire after the first successful CMS 1173 // collection because the stats should then be valid. 1174 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1175 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1176 _cmsGen->occupancy(), _bootstrap_occupancy); 1177 return true; 1178 } 1179 } 1180 } 1181 1182 // Otherwise, we start a collection cycle if 1183 // old gen want a collection cycle started. Each may use 1184 // an appropriate criterion for making this decision. 1185 // XXX We need to make sure that the gen expansion 1186 // criterion dovetails well with this. XXX NEED TO FIX THIS 1187 if (_cmsGen->should_concurrent_collect()) { 1188 log.print("CMS old gen initiated"); 1189 return true; 1190 } 1191 1192 // We start a collection if we believe an incremental collection may fail; 1193 // this is not likely to be productive in practice because it's probably too 1194 // late anyway. 1195 CMSHeap* heap = CMSHeap::heap(); 1196 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1197 log.print("CMSCollector: collect because incremental collection will fail "); 1198 return true; 1199 } 1200 1201 if (MetaspaceGC::should_concurrent_collect()) { 1202 log.print("CMSCollector: collect for metadata allocation "); 1203 return true; 1204 } 1205 1206 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1207 if (CMSTriggerInterval >= 0) { 1208 if (CMSTriggerInterval == 0) { 1209 // Trigger always 1210 return true; 1211 } 1212 1213 // Check the CMS time since begin (we do not check the stats validity 1214 // as we want to be able to trigger the first CMS cycle as well) 1215 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1216 if (stats().valid()) { 1217 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1218 stats().cms_time_since_begin()); 1219 } else { 1220 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1221 } 1222 return true; 1223 } 1224 } 1225 1226 return false; 1227 } 1228 1229 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1230 1231 // Clear _expansion_cause fields of constituent generations 1232 void CMSCollector::clear_expansion_cause() { 1233 _cmsGen->clear_expansion_cause(); 1234 } 1235 1236 // We should be conservative in starting a collection cycle. To 1237 // start too eagerly runs the risk of collecting too often in the 1238 // extreme. To collect too rarely falls back on full collections, 1239 // which works, even if not optimum in terms of concurrent work. 1240 // As a work around for too eagerly collecting, use the flag 1241 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1242 // giving the user an easily understandable way of controlling the 1243 // collections. 1244 // We want to start a new collection cycle if any of the following 1245 // conditions hold: 1246 // . our current occupancy exceeds the configured initiating occupancy 1247 // for this generation, or 1248 // . we recently needed to expand this space and have not, since that 1249 // expansion, done a collection of this generation, or 1250 // . the underlying space believes that it may be a good idea to initiate 1251 // a concurrent collection (this may be based on criteria such as the 1252 // following: the space uses linear allocation and linear allocation is 1253 // going to fail, or there is believed to be excessive fragmentation in 1254 // the generation, etc... or ... 1255 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1256 // the case of the old generation; see CR 6543076): 1257 // we may be approaching a point at which allocation requests may fail because 1258 // we will be out of sufficient free space given allocation rate estimates.] 1259 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1260 1261 assert_lock_strong(freelistLock()); 1262 if (occupancy() > initiating_occupancy()) { 1263 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1264 short_name(), occupancy(), initiating_occupancy()); 1265 return true; 1266 } 1267 if (UseCMSInitiatingOccupancyOnly) { 1268 return false; 1269 } 1270 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1271 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1272 return true; 1273 } 1274 return false; 1275 } 1276 1277 void ConcurrentMarkSweepGeneration::collect(bool full, 1278 bool clear_all_soft_refs, 1279 size_t size, 1280 bool tlab) 1281 { 1282 collector()->collect(full, clear_all_soft_refs, size, tlab); 1283 } 1284 1285 void CMSCollector::collect(bool full, 1286 bool clear_all_soft_refs, 1287 size_t size, 1288 bool tlab) 1289 { 1290 // The following "if" branch is present for defensive reasons. 1291 // In the current uses of this interface, it can be replaced with: 1292 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1293 // But I am not placing that assert here to allow future 1294 // generality in invoking this interface. 1295 if (GCLocker::is_active()) { 1296 // A consistency test for GCLocker 1297 assert(GCLocker::needs_gc(), "Should have been set already"); 1298 // Skip this foreground collection, instead 1299 // expanding the heap if necessary. 1300 // Need the free list locks for the call to free() in compute_new_size() 1301 compute_new_size(); 1302 return; 1303 } 1304 acquire_control_and_collect(full, clear_all_soft_refs); 1305 } 1306 1307 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1308 CMSHeap* heap = CMSHeap::heap(); 1309 unsigned int gc_count = heap->total_full_collections(); 1310 if (gc_count == full_gc_count) { 1311 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1312 _full_gc_requested = true; 1313 _full_gc_cause = cause; 1314 CGC_lock->notify(); // nudge CMS thread 1315 } else { 1316 assert(gc_count > full_gc_count, "Error: causal loop"); 1317 } 1318 } 1319 1320 bool CMSCollector::is_external_interruption() { 1321 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1322 return GCCause::is_user_requested_gc(cause) || 1323 GCCause::is_serviceability_requested_gc(cause); 1324 } 1325 1326 void CMSCollector::report_concurrent_mode_interruption() { 1327 if (is_external_interruption()) { 1328 log_debug(gc)("Concurrent mode interrupted"); 1329 } else { 1330 log_debug(gc)("Concurrent mode failure"); 1331 _gc_tracer_cm->report_concurrent_mode_failure(); 1332 } 1333 } 1334 1335 1336 // The foreground and background collectors need to coordinate in order 1337 // to make sure that they do not mutually interfere with CMS collections. 1338 // When a background collection is active, 1339 // the foreground collector may need to take over (preempt) and 1340 // synchronously complete an ongoing collection. Depending on the 1341 // frequency of the background collections and the heap usage 1342 // of the application, this preemption can be seldom or frequent. 1343 // There are only certain 1344 // points in the background collection that the "collection-baton" 1345 // can be passed to the foreground collector. 1346 // 1347 // The foreground collector will wait for the baton before 1348 // starting any part of the collection. The foreground collector 1349 // will only wait at one location. 1350 // 1351 // The background collector will yield the baton before starting a new 1352 // phase of the collection (e.g., before initial marking, marking from roots, 1353 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1354 // of the loop which switches the phases. The background collector does some 1355 // of the phases (initial mark, final re-mark) with the world stopped. 1356 // Because of locking involved in stopping the world, 1357 // the foreground collector should not block waiting for the background 1358 // collector when it is doing a stop-the-world phase. The background 1359 // collector will yield the baton at an additional point just before 1360 // it enters a stop-the-world phase. Once the world is stopped, the 1361 // background collector checks the phase of the collection. If the 1362 // phase has not changed, it proceeds with the collection. If the 1363 // phase has changed, it skips that phase of the collection. See 1364 // the comments on the use of the Heap_lock in collect_in_background(). 1365 // 1366 // Variable used in baton passing. 1367 // _foregroundGCIsActive - Set to true by the foreground collector when 1368 // it wants the baton. The foreground clears it when it has finished 1369 // the collection. 1370 // _foregroundGCShouldWait - Set to true by the background collector 1371 // when it is running. The foreground collector waits while 1372 // _foregroundGCShouldWait is true. 1373 // CGC_lock - monitor used to protect access to the above variables 1374 // and to notify the foreground and background collectors. 1375 // _collectorState - current state of the CMS collection. 1376 // 1377 // The foreground collector 1378 // acquires the CGC_lock 1379 // sets _foregroundGCIsActive 1380 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1381 // various locks acquired in preparation for the collection 1382 // are released so as not to block the background collector 1383 // that is in the midst of a collection 1384 // proceeds with the collection 1385 // clears _foregroundGCIsActive 1386 // returns 1387 // 1388 // The background collector in a loop iterating on the phases of the 1389 // collection 1390 // acquires the CGC_lock 1391 // sets _foregroundGCShouldWait 1392 // if _foregroundGCIsActive is set 1393 // clears _foregroundGCShouldWait, notifies _CGC_lock 1394 // waits on _CGC_lock for _foregroundGCIsActive to become false 1395 // and exits the loop. 1396 // otherwise 1397 // proceed with that phase of the collection 1398 // if the phase is a stop-the-world phase, 1399 // yield the baton once more just before enqueueing 1400 // the stop-world CMS operation (executed by the VM thread). 1401 // returns after all phases of the collection are done 1402 // 1403 1404 void CMSCollector::acquire_control_and_collect(bool full, 1405 bool clear_all_soft_refs) { 1406 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1407 assert(!Thread::current()->is_ConcurrentGC_thread(), 1408 "shouldn't try to acquire control from self!"); 1409 1410 // Start the protocol for acquiring control of the 1411 // collection from the background collector (aka CMS thread). 1412 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1413 "VM thread should have CMS token"); 1414 // Remember the possibly interrupted state of an ongoing 1415 // concurrent collection 1416 CollectorState first_state = _collectorState; 1417 1418 // Signal to a possibly ongoing concurrent collection that 1419 // we want to do a foreground collection. 1420 _foregroundGCIsActive = true; 1421 1422 // release locks and wait for a notify from the background collector 1423 // releasing the locks in only necessary for phases which 1424 // do yields to improve the granularity of the collection. 1425 assert_lock_strong(bitMapLock()); 1426 // We need to lock the Free list lock for the space that we are 1427 // currently collecting. 1428 assert(haveFreelistLocks(), "Must be holding free list locks"); 1429 bitMapLock()->unlock(); 1430 releaseFreelistLocks(); 1431 { 1432 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1433 if (_foregroundGCShouldWait) { 1434 // We are going to be waiting for action for the CMS thread; 1435 // it had better not be gone (for instance at shutdown)! 1436 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1437 "CMS thread must be running"); 1438 // Wait here until the background collector gives us the go-ahead 1439 ConcurrentMarkSweepThread::clear_CMS_flag( 1440 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1441 // Get a possibly blocked CMS thread going: 1442 // Note that we set _foregroundGCIsActive true above, 1443 // without protection of the CGC_lock. 1444 CGC_lock->notify(); 1445 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1446 "Possible deadlock"); 1447 while (_foregroundGCShouldWait) { 1448 // wait for notification 1449 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1450 // Possibility of delay/starvation here, since CMS token does 1451 // not know to give priority to VM thread? Actually, i think 1452 // there wouldn't be any delay/starvation, but the proof of 1453 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1454 } 1455 ConcurrentMarkSweepThread::set_CMS_flag( 1456 ConcurrentMarkSweepThread::CMS_vm_has_token); 1457 } 1458 } 1459 // The CMS_token is already held. Get back the other locks. 1460 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1461 "VM thread should have CMS token"); 1462 getFreelistLocks(); 1463 bitMapLock()->lock_without_safepoint_check(); 1464 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1465 p2i(Thread::current()), first_state); 1466 log_debug(gc, state)(" gets control with state %d", _collectorState); 1467 1468 // Inform cms gen if this was due to partial collection failing. 1469 // The CMS gen may use this fact to determine its expansion policy. 1470 CMSHeap* heap = CMSHeap::heap(); 1471 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1472 assert(!_cmsGen->incremental_collection_failed(), 1473 "Should have been noticed, reacted to and cleared"); 1474 _cmsGen->set_incremental_collection_failed(); 1475 } 1476 1477 if (first_state > Idling) { 1478 report_concurrent_mode_interruption(); 1479 } 1480 1481 set_did_compact(true); 1482 1483 // If the collection is being acquired from the background 1484 // collector, there may be references on the discovered 1485 // references lists. Abandon those references, since some 1486 // of them may have become unreachable after concurrent 1487 // discovery; the STW compacting collector will redo discovery 1488 // more precisely, without being subject to floating garbage. 1489 // Leaving otherwise unreachable references in the discovered 1490 // lists would require special handling. 1491 ref_processor()->disable_discovery(); 1492 ref_processor()->abandon_partial_discovery(); 1493 ref_processor()->verify_no_references_recorded(); 1494 1495 if (first_state > Idling) { 1496 save_heap_summary(); 1497 } 1498 1499 do_compaction_work(clear_all_soft_refs); 1500 1501 // Has the GC time limit been exceeded? 1502 size_t max_eden_size = _young_gen->max_eden_size(); 1503 GCCause::Cause gc_cause = heap->gc_cause(); 1504 size_policy()->check_gc_overhead_limit(_young_gen->eden()->used(), 1505 _cmsGen->max_capacity(), 1506 max_eden_size, 1507 full, 1508 gc_cause, 1509 heap->soft_ref_policy()); 1510 1511 // Reset the expansion cause, now that we just completed 1512 // a collection cycle. 1513 clear_expansion_cause(); 1514 _foregroundGCIsActive = false; 1515 return; 1516 } 1517 1518 // Resize the tenured generation 1519 // after obtaining the free list locks for the 1520 // two generations. 1521 void CMSCollector::compute_new_size() { 1522 assert_locked_or_safepoint(Heap_lock); 1523 FreelistLocker z(this); 1524 MetaspaceGC::compute_new_size(); 1525 _cmsGen->compute_new_size_free_list(); 1526 } 1527 1528 // A work method used by the foreground collector to do 1529 // a mark-sweep-compact. 1530 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1531 CMSHeap* heap = CMSHeap::heap(); 1532 1533 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1534 gc_timer->register_gc_start(); 1535 1536 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1537 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1538 1539 heap->pre_full_gc_dump(gc_timer); 1540 1541 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1542 1543 // Temporarily widen the span of the weak reference processing to 1544 // the entire heap. 1545 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1546 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1547 // Temporarily, clear the "is_alive_non_header" field of the 1548 // reference processor. 1549 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1550 // Temporarily make reference _processing_ single threaded (non-MT). 1551 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1552 // Temporarily make refs discovery atomic 1553 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1554 // Temporarily make reference _discovery_ single threaded (non-MT) 1555 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1556 1557 ref_processor()->set_enqueuing_is_done(false); 1558 ref_processor()->enable_discovery(); 1559 ref_processor()->setup_policy(clear_all_soft_refs); 1560 // If an asynchronous collection finishes, the _modUnionTable is 1561 // all clear. If we are assuming the collection from an asynchronous 1562 // collection, clear the _modUnionTable. 1563 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1564 "_modUnionTable should be clear if the baton was not passed"); 1565 _modUnionTable.clear_all(); 1566 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1567 "mod union for klasses should be clear if the baton was passed"); 1568 _ct->cld_rem_set()->clear_mod_union(); 1569 1570 1571 // We must adjust the allocation statistics being maintained 1572 // in the free list space. We do so by reading and clearing 1573 // the sweep timer and updating the block flux rate estimates below. 1574 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1575 if (_inter_sweep_timer.is_active()) { 1576 _inter_sweep_timer.stop(); 1577 // Note that we do not use this sample to update the _inter_sweep_estimate. 1578 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1579 _inter_sweep_estimate.padded_average(), 1580 _intra_sweep_estimate.padded_average()); 1581 } 1582 1583 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1584 #ifdef ASSERT 1585 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1586 size_t free_size = cms_space->free(); 1587 assert(free_size == 1588 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1589 * HeapWordSize, 1590 "All the free space should be compacted into one chunk at top"); 1591 assert(cms_space->dictionary()->total_chunk_size( 1592 debug_only(cms_space->freelistLock())) == 0 || 1593 cms_space->totalSizeInIndexedFreeLists() == 0, 1594 "All the free space should be in a single chunk"); 1595 size_t num = cms_space->totalCount(); 1596 assert((free_size == 0 && num == 0) || 1597 (free_size > 0 && (num == 1 || num == 2)), 1598 "There should be at most 2 free chunks after compaction"); 1599 #endif // ASSERT 1600 _collectorState = Resetting; 1601 assert(_restart_addr == NULL, 1602 "Should have been NULL'd before baton was passed"); 1603 reset_stw(); 1604 _cmsGen->reset_after_compaction(); 1605 _concurrent_cycles_since_last_unload = 0; 1606 1607 // Clear any data recorded in the PLAB chunk arrays. 1608 if (_survivor_plab_array != NULL) { 1609 reset_survivor_plab_arrays(); 1610 } 1611 1612 // Adjust the per-size allocation stats for the next epoch. 1613 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1614 // Restart the "inter sweep timer" for the next epoch. 1615 _inter_sweep_timer.reset(); 1616 _inter_sweep_timer.start(); 1617 1618 // No longer a need to do a concurrent collection for Metaspace. 1619 MetaspaceGC::set_should_concurrent_collect(false); 1620 1621 heap->post_full_gc_dump(gc_timer); 1622 1623 gc_timer->register_gc_end(); 1624 1625 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1626 1627 // For a mark-sweep-compact, compute_new_size() will be called 1628 // in the heap's do_collection() method. 1629 } 1630 1631 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1632 Log(gc, heap) log; 1633 if (!log.is_trace()) { 1634 return; 1635 } 1636 1637 ContiguousSpace* eden_space = _young_gen->eden(); 1638 ContiguousSpace* from_space = _young_gen->from(); 1639 ContiguousSpace* to_space = _young_gen->to(); 1640 // Eden 1641 if (_eden_chunk_array != NULL) { 1642 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1643 p2i(eden_space->bottom()), p2i(eden_space->top()), 1644 p2i(eden_space->end()), eden_space->capacity()); 1645 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1646 _eden_chunk_index, _eden_chunk_capacity); 1647 for (size_t i = 0; i < _eden_chunk_index; i++) { 1648 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1649 } 1650 } 1651 // Survivor 1652 if (_survivor_chunk_array != NULL) { 1653 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1654 p2i(from_space->bottom()), p2i(from_space->top()), 1655 p2i(from_space->end()), from_space->capacity()); 1656 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1657 _survivor_chunk_index, _survivor_chunk_capacity); 1658 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1659 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1660 } 1661 } 1662 } 1663 1664 void CMSCollector::getFreelistLocks() const { 1665 // Get locks for all free lists in all generations that this 1666 // collector is responsible for 1667 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1668 } 1669 1670 void CMSCollector::releaseFreelistLocks() const { 1671 // Release locks for all free lists in all generations that this 1672 // collector is responsible for 1673 _cmsGen->freelistLock()->unlock(); 1674 } 1675 1676 bool CMSCollector::haveFreelistLocks() const { 1677 // Check locks for all free lists in all generations that this 1678 // collector is responsible for 1679 assert_lock_strong(_cmsGen->freelistLock()); 1680 PRODUCT_ONLY(ShouldNotReachHere()); 1681 return true; 1682 } 1683 1684 // A utility class that is used by the CMS collector to 1685 // temporarily "release" the foreground collector from its 1686 // usual obligation to wait for the background collector to 1687 // complete an ongoing phase before proceeding. 1688 class ReleaseForegroundGC: public StackObj { 1689 private: 1690 CMSCollector* _c; 1691 public: 1692 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1693 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1694 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1695 // allow a potentially blocked foreground collector to proceed 1696 _c->_foregroundGCShouldWait = false; 1697 if (_c->_foregroundGCIsActive) { 1698 CGC_lock->notify(); 1699 } 1700 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1701 "Possible deadlock"); 1702 } 1703 1704 ~ReleaseForegroundGC() { 1705 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1706 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1707 _c->_foregroundGCShouldWait = true; 1708 } 1709 }; 1710 1711 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1712 assert(Thread::current()->is_ConcurrentGC_thread(), 1713 "A CMS asynchronous collection is only allowed on a CMS thread."); 1714 1715 CMSHeap* heap = CMSHeap::heap(); 1716 { 1717 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1718 MutexLockerEx hl(Heap_lock, safepoint_check); 1719 FreelistLocker fll(this); 1720 MutexLockerEx x(CGC_lock, safepoint_check); 1721 if (_foregroundGCIsActive) { 1722 // The foreground collector is. Skip this 1723 // background collection. 1724 assert(!_foregroundGCShouldWait, "Should be clear"); 1725 return; 1726 } else { 1727 assert(_collectorState == Idling, "Should be idling before start."); 1728 _collectorState = InitialMarking; 1729 register_gc_start(cause); 1730 // Reset the expansion cause, now that we are about to begin 1731 // a new cycle. 1732 clear_expansion_cause(); 1733 1734 // Clear the MetaspaceGC flag since a concurrent collection 1735 // is starting but also clear it after the collection. 1736 MetaspaceGC::set_should_concurrent_collect(false); 1737 } 1738 // Decide if we want to enable class unloading as part of the 1739 // ensuing concurrent GC cycle. 1740 update_should_unload_classes(); 1741 _full_gc_requested = false; // acks all outstanding full gc requests 1742 _full_gc_cause = GCCause::_no_gc; 1743 // Signal that we are about to start a collection 1744 heap->increment_total_full_collections(); // ... starting a collection cycle 1745 _collection_count_start = heap->total_full_collections(); 1746 } 1747 1748 size_t prev_used = _cmsGen->used(); 1749 1750 // The change of the collection state is normally done at this level; 1751 // the exceptions are phases that are executed while the world is 1752 // stopped. For those phases the change of state is done while the 1753 // world is stopped. For baton passing purposes this allows the 1754 // background collector to finish the phase and change state atomically. 1755 // The foreground collector cannot wait on a phase that is done 1756 // while the world is stopped because the foreground collector already 1757 // has the world stopped and would deadlock. 1758 while (_collectorState != Idling) { 1759 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1760 p2i(Thread::current()), _collectorState); 1761 // The foreground collector 1762 // holds the Heap_lock throughout its collection. 1763 // holds the CMS token (but not the lock) 1764 // except while it is waiting for the background collector to yield. 1765 // 1766 // The foreground collector should be blocked (not for long) 1767 // if the background collector is about to start a phase 1768 // executed with world stopped. If the background 1769 // collector has already started such a phase, the 1770 // foreground collector is blocked waiting for the 1771 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1772 // are executed in the VM thread. 1773 // 1774 // The locking order is 1775 // PendingListLock (PLL) -- if applicable (FinalMarking) 1776 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1777 // CMS token (claimed in 1778 // stop_world_and_do() --> 1779 // safepoint_synchronize() --> 1780 // CMSThread::synchronize()) 1781 1782 { 1783 // Check if the FG collector wants us to yield. 1784 CMSTokenSync x(true); // is cms thread 1785 if (waitForForegroundGC()) { 1786 // We yielded to a foreground GC, nothing more to be 1787 // done this round. 1788 assert(_foregroundGCShouldWait == false, "We set it to false in " 1789 "waitForForegroundGC()"); 1790 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1791 p2i(Thread::current()), _collectorState); 1792 return; 1793 } else { 1794 // The background collector can run but check to see if the 1795 // foreground collector has done a collection while the 1796 // background collector was waiting to get the CGC_lock 1797 // above. If yes, break so that _foregroundGCShouldWait 1798 // is cleared before returning. 1799 if (_collectorState == Idling) { 1800 break; 1801 } 1802 } 1803 } 1804 1805 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1806 "should be waiting"); 1807 1808 switch (_collectorState) { 1809 case InitialMarking: 1810 { 1811 ReleaseForegroundGC x(this); 1812 stats().record_cms_begin(); 1813 VM_CMS_Initial_Mark initial_mark_op(this); 1814 VMThread::execute(&initial_mark_op); 1815 } 1816 // The collector state may be any legal state at this point 1817 // since the background collector may have yielded to the 1818 // foreground collector. 1819 break; 1820 case Marking: 1821 // initial marking in checkpointRootsInitialWork has been completed 1822 if (markFromRoots()) { // we were successful 1823 assert(_collectorState == Precleaning, "Collector state should " 1824 "have changed"); 1825 } else { 1826 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1827 } 1828 break; 1829 case Precleaning: 1830 // marking from roots in markFromRoots has been completed 1831 preclean(); 1832 assert(_collectorState == AbortablePreclean || 1833 _collectorState == FinalMarking, 1834 "Collector state should have changed"); 1835 break; 1836 case AbortablePreclean: 1837 abortable_preclean(); 1838 assert(_collectorState == FinalMarking, "Collector state should " 1839 "have changed"); 1840 break; 1841 case FinalMarking: 1842 { 1843 ReleaseForegroundGC x(this); 1844 1845 VM_CMS_Final_Remark final_remark_op(this); 1846 VMThread::execute(&final_remark_op); 1847 } 1848 assert(_foregroundGCShouldWait, "block post-condition"); 1849 break; 1850 case Sweeping: 1851 // final marking in checkpointRootsFinal has been completed 1852 sweep(); 1853 assert(_collectorState == Resizing, "Collector state change " 1854 "to Resizing must be done under the free_list_lock"); 1855 1856 case Resizing: { 1857 // Sweeping has been completed... 1858 // At this point the background collection has completed. 1859 // Don't move the call to compute_new_size() down 1860 // into code that might be executed if the background 1861 // collection was preempted. 1862 { 1863 ReleaseForegroundGC x(this); // unblock FG collection 1864 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1865 CMSTokenSync z(true); // not strictly needed. 1866 if (_collectorState == Resizing) { 1867 compute_new_size(); 1868 save_heap_summary(); 1869 _collectorState = Resetting; 1870 } else { 1871 assert(_collectorState == Idling, "The state should only change" 1872 " because the foreground collector has finished the collection"); 1873 } 1874 } 1875 break; 1876 } 1877 case Resetting: 1878 // CMS heap resizing has been completed 1879 reset_concurrent(); 1880 assert(_collectorState == Idling, "Collector state should " 1881 "have changed"); 1882 1883 MetaspaceGC::set_should_concurrent_collect(false); 1884 1885 stats().record_cms_end(); 1886 // Don't move the concurrent_phases_end() and compute_new_size() 1887 // calls to here because a preempted background collection 1888 // has it's state set to "Resetting". 1889 break; 1890 case Idling: 1891 default: 1892 ShouldNotReachHere(); 1893 break; 1894 } 1895 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1896 p2i(Thread::current()), _collectorState); 1897 assert(_foregroundGCShouldWait, "block post-condition"); 1898 } 1899 1900 // Should this be in gc_epilogue? 1901 heap->counters()->update_counters(); 1902 1903 { 1904 // Clear _foregroundGCShouldWait and, in the event that the 1905 // foreground collector is waiting, notify it, before 1906 // returning. 1907 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1908 _foregroundGCShouldWait = false; 1909 if (_foregroundGCIsActive) { 1910 CGC_lock->notify(); 1911 } 1912 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1913 "Possible deadlock"); 1914 } 1915 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1916 p2i(Thread::current()), _collectorState); 1917 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1918 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1919 } 1920 1921 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1922 _cms_start_registered = true; 1923 _gc_timer_cm->register_gc_start(); 1924 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1925 } 1926 1927 void CMSCollector::register_gc_end() { 1928 if (_cms_start_registered) { 1929 report_heap_summary(GCWhen::AfterGC); 1930 1931 _gc_timer_cm->register_gc_end(); 1932 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1933 _cms_start_registered = false; 1934 } 1935 } 1936 1937 void CMSCollector::save_heap_summary() { 1938 CMSHeap* heap = CMSHeap::heap(); 1939 _last_heap_summary = heap->create_heap_summary(); 1940 _last_metaspace_summary = heap->create_metaspace_summary(); 1941 } 1942 1943 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1944 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1945 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1946 } 1947 1948 bool CMSCollector::waitForForegroundGC() { 1949 bool res = false; 1950 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1951 "CMS thread should have CMS token"); 1952 // Block the foreground collector until the 1953 // background collectors decides whether to 1954 // yield. 1955 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1956 _foregroundGCShouldWait = true; 1957 if (_foregroundGCIsActive) { 1958 // The background collector yields to the 1959 // foreground collector and returns a value 1960 // indicating that it has yielded. The foreground 1961 // collector can proceed. 1962 res = true; 1963 _foregroundGCShouldWait = false; 1964 ConcurrentMarkSweepThread::clear_CMS_flag( 1965 ConcurrentMarkSweepThread::CMS_cms_has_token); 1966 ConcurrentMarkSweepThread::set_CMS_flag( 1967 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1968 // Get a possibly blocked foreground thread going 1969 CGC_lock->notify(); 1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 while (_foregroundGCIsActive) { 1973 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1974 } 1975 ConcurrentMarkSweepThread::set_CMS_flag( 1976 ConcurrentMarkSweepThread::CMS_cms_has_token); 1977 ConcurrentMarkSweepThread::clear_CMS_flag( 1978 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1979 } 1980 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1981 p2i(Thread::current()), _collectorState); 1982 return res; 1983 } 1984 1985 // Because of the need to lock the free lists and other structures in 1986 // the collector, common to all the generations that the collector is 1987 // collecting, we need the gc_prologues of individual CMS generations 1988 // delegate to their collector. It may have been simpler had the 1989 // current infrastructure allowed one to call a prologue on a 1990 // collector. In the absence of that we have the generation's 1991 // prologue delegate to the collector, which delegates back 1992 // some "local" work to a worker method in the individual generations 1993 // that it's responsible for collecting, while itself doing any 1994 // work common to all generations it's responsible for. A similar 1995 // comment applies to the gc_epilogue()'s. 1996 // The role of the variable _between_prologue_and_epilogue is to 1997 // enforce the invocation protocol. 1998 void CMSCollector::gc_prologue(bool full) { 1999 // Call gc_prologue_work() for the CMSGen 2000 // we are responsible for. 2001 2002 // The following locking discipline assumes that we are only called 2003 // when the world is stopped. 2004 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2005 2006 // The CMSCollector prologue must call the gc_prologues for the 2007 // "generations" that it's responsible 2008 // for. 2009 2010 assert( Thread::current()->is_VM_thread() 2011 || ( CMSScavengeBeforeRemark 2012 && Thread::current()->is_ConcurrentGC_thread()), 2013 "Incorrect thread type for prologue execution"); 2014 2015 if (_between_prologue_and_epilogue) { 2016 // We have already been invoked; this is a gc_prologue delegation 2017 // from yet another CMS generation that we are responsible for, just 2018 // ignore it since all relevant work has already been done. 2019 return; 2020 } 2021 2022 // set a bit saying prologue has been called; cleared in epilogue 2023 _between_prologue_and_epilogue = true; 2024 // Claim locks for common data structures, then call gc_prologue_work() 2025 // for each CMSGen. 2026 2027 getFreelistLocks(); // gets free list locks on constituent spaces 2028 bitMapLock()->lock_without_safepoint_check(); 2029 2030 // Should call gc_prologue_work() for all cms gens we are responsible for 2031 bool duringMarking = _collectorState >= Marking 2032 && _collectorState < Sweeping; 2033 2034 // The young collections clear the modified oops state, which tells if 2035 // there are any modified oops in the class. The remark phase also needs 2036 // that information. Tell the young collection to save the union of all 2037 // modified klasses. 2038 if (duringMarking) { 2039 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2040 } 2041 2042 bool registerClosure = duringMarking; 2043 2044 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2045 2046 if (!full) { 2047 stats().record_gc0_begin(); 2048 } 2049 } 2050 2051 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2052 2053 _capacity_at_prologue = capacity(); 2054 _used_at_prologue = used(); 2055 2056 // We enable promotion tracking so that card-scanning can recognize 2057 // which objects have been promoted during this GC and skip them. 2058 for (uint i = 0; i < ParallelGCThreads; i++) { 2059 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2060 } 2061 2062 // Delegate to CMScollector which knows how to coordinate between 2063 // this and any other CMS generations that it is responsible for 2064 // collecting. 2065 collector()->gc_prologue(full); 2066 } 2067 2068 // This is a "private" interface for use by this generation's CMSCollector. 2069 // Not to be called directly by any other entity (for instance, 2070 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2071 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2072 bool registerClosure, ModUnionClosure* modUnionClosure) { 2073 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2074 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2075 "Should be NULL"); 2076 if (registerClosure) { 2077 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2078 } 2079 cmsSpace()->gc_prologue(); 2080 // Clear stat counters 2081 NOT_PRODUCT( 2082 assert(_numObjectsPromoted == 0, "check"); 2083 assert(_numWordsPromoted == 0, "check"); 2084 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2085 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2086 _numObjectsAllocated = 0; 2087 _numWordsAllocated = 0; 2088 ) 2089 } 2090 2091 void CMSCollector::gc_epilogue(bool full) { 2092 // The following locking discipline assumes that we are only called 2093 // when the world is stopped. 2094 assert(SafepointSynchronize::is_at_safepoint(), 2095 "world is stopped assumption"); 2096 2097 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2098 // if linear allocation blocks need to be appropriately marked to allow the 2099 // the blocks to be parsable. We also check here whether we need to nudge the 2100 // CMS collector thread to start a new cycle (if it's not already active). 2101 assert( Thread::current()->is_VM_thread() 2102 || ( CMSScavengeBeforeRemark 2103 && Thread::current()->is_ConcurrentGC_thread()), 2104 "Incorrect thread type for epilogue execution"); 2105 2106 if (!_between_prologue_and_epilogue) { 2107 // We have already been invoked; this is a gc_epilogue delegation 2108 // from yet another CMS generation that we are responsible for, just 2109 // ignore it since all relevant work has already been done. 2110 return; 2111 } 2112 assert(haveFreelistLocks(), "must have freelist locks"); 2113 assert_lock_strong(bitMapLock()); 2114 2115 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2116 2117 _cmsGen->gc_epilogue_work(full); 2118 2119 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2120 // in case sampling was not already enabled, enable it 2121 _start_sampling = true; 2122 } 2123 // reset _eden_chunk_array so sampling starts afresh 2124 _eden_chunk_index = 0; 2125 2126 size_t cms_used = _cmsGen->cmsSpace()->used(); 2127 2128 // update performance counters - this uses a special version of 2129 // update_counters() that allows the utilization to be passed as a 2130 // parameter, avoiding multiple calls to used(). 2131 // 2132 _cmsGen->update_counters(cms_used); 2133 2134 bitMapLock()->unlock(); 2135 releaseFreelistLocks(); 2136 2137 if (!CleanChunkPoolAsync) { 2138 Chunk::clean_chunk_pool(); 2139 } 2140 2141 set_did_compact(false); 2142 _between_prologue_and_epilogue = false; // ready for next cycle 2143 } 2144 2145 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2146 collector()->gc_epilogue(full); 2147 2148 // When using ParNew, promotion tracking should have already been 2149 // disabled. However, the prologue (which enables promotion 2150 // tracking) and epilogue are called irrespective of the type of 2151 // GC. So they will also be called before and after Full GCs, during 2152 // which promotion tracking will not be explicitly disabled. So, 2153 // it's safer to also disable it here too (to be symmetric with 2154 // enabling it in the prologue). 2155 for (uint i = 0; i < ParallelGCThreads; i++) { 2156 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2157 } 2158 } 2159 2160 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2161 assert(!incremental_collection_failed(), "Should have been cleared"); 2162 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2163 cmsSpace()->gc_epilogue(); 2164 // Print stat counters 2165 NOT_PRODUCT( 2166 assert(_numObjectsAllocated == 0, "check"); 2167 assert(_numWordsAllocated == 0, "check"); 2168 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2169 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2170 _numObjectsPromoted = 0; 2171 _numWordsPromoted = 0; 2172 ) 2173 2174 // Call down the chain in contiguous_available needs the freelistLock 2175 // so print this out before releasing the freeListLock. 2176 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2177 } 2178 2179 #ifndef PRODUCT 2180 bool CMSCollector::have_cms_token() { 2181 Thread* thr = Thread::current(); 2182 if (thr->is_VM_thread()) { 2183 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2184 } else if (thr->is_ConcurrentGC_thread()) { 2185 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2186 } else if (thr->is_GC_task_thread()) { 2187 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2188 ParGCRareEvent_lock->owned_by_self(); 2189 } 2190 return false; 2191 } 2192 2193 // Check reachability of the given heap address in CMS generation, 2194 // treating all other generations as roots. 2195 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2196 // We could "guarantee" below, rather than assert, but I'll 2197 // leave these as "asserts" so that an adventurous debugger 2198 // could try this in the product build provided some subset of 2199 // the conditions were met, provided they were interested in the 2200 // results and knew that the computation below wouldn't interfere 2201 // with other concurrent computations mutating the structures 2202 // being read or written. 2203 assert(SafepointSynchronize::is_at_safepoint(), 2204 "Else mutations in object graph will make answer suspect"); 2205 assert(have_cms_token(), "Should hold cms token"); 2206 assert(haveFreelistLocks(), "must hold free list locks"); 2207 assert_lock_strong(bitMapLock()); 2208 2209 // Clear the marking bit map array before starting, but, just 2210 // for kicks, first report if the given address is already marked 2211 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2212 _markBitMap.isMarked(addr) ? "" : " not"); 2213 2214 if (verify_after_remark()) { 2215 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2216 bool result = verification_mark_bm()->isMarked(addr); 2217 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2218 result ? "IS" : "is NOT"); 2219 return result; 2220 } else { 2221 tty->print_cr("Could not compute result"); 2222 return false; 2223 } 2224 } 2225 #endif 2226 2227 void 2228 CMSCollector::print_on_error(outputStream* st) { 2229 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2230 if (collector != NULL) { 2231 CMSBitMap* bitmap = &collector->_markBitMap; 2232 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2233 bitmap->print_on_error(st, " Bits: "); 2234 2235 st->cr(); 2236 2237 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2238 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2239 mut_bitmap->print_on_error(st, " Bits: "); 2240 } 2241 } 2242 2243 //////////////////////////////////////////////////////// 2244 // CMS Verification Support 2245 //////////////////////////////////////////////////////// 2246 // Following the remark phase, the following invariant 2247 // should hold -- each object in the CMS heap which is 2248 // marked in markBitMap() should be marked in the verification_mark_bm(). 2249 2250 class VerifyMarkedClosure: public BitMapClosure { 2251 CMSBitMap* _marks; 2252 bool _failed; 2253 2254 public: 2255 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2256 2257 bool do_bit(size_t offset) { 2258 HeapWord* addr = _marks->offsetToHeapWord(offset); 2259 if (!_marks->isMarked(addr)) { 2260 Log(gc, verify) log; 2261 ResourceMark rm; 2262 LogStream ls(log.error()); 2263 oop(addr)->print_on(&ls); 2264 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2265 _failed = true; 2266 } 2267 return true; 2268 } 2269 2270 bool failed() { return _failed; } 2271 }; 2272 2273 bool CMSCollector::verify_after_remark() { 2274 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2275 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2276 static bool init = false; 2277 2278 assert(SafepointSynchronize::is_at_safepoint(), 2279 "Else mutations in object graph will make answer suspect"); 2280 assert(have_cms_token(), 2281 "Else there may be mutual interference in use of " 2282 " verification data structures"); 2283 assert(_collectorState > Marking && _collectorState <= Sweeping, 2284 "Else marking info checked here may be obsolete"); 2285 assert(haveFreelistLocks(), "must hold free list locks"); 2286 assert_lock_strong(bitMapLock()); 2287 2288 2289 // Allocate marking bit map if not already allocated 2290 if (!init) { // first time 2291 if (!verification_mark_bm()->allocate(_span)) { 2292 return false; 2293 } 2294 init = true; 2295 } 2296 2297 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2298 2299 // Turn off refs discovery -- so we will be tracing through refs. 2300 // This is as intended, because by this time 2301 // GC must already have cleared any refs that need to be cleared, 2302 // and traced those that need to be marked; moreover, 2303 // the marking done here is not going to interfere in any 2304 // way with the marking information used by GC. 2305 NoRefDiscovery no_discovery(ref_processor()); 2306 2307 #if COMPILER2_OR_JVMCI 2308 DerivedPointerTableDeactivate dpt_deact; 2309 #endif 2310 2311 // Clear any marks from a previous round 2312 verification_mark_bm()->clear_all(); 2313 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2314 verify_work_stacks_empty(); 2315 2316 CMSHeap* heap = CMSHeap::heap(); 2317 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2318 // Update the saved marks which may affect the root scans. 2319 heap->save_marks(); 2320 2321 if (CMSRemarkVerifyVariant == 1) { 2322 // In this first variant of verification, we complete 2323 // all marking, then check if the new marks-vector is 2324 // a subset of the CMS marks-vector. 2325 verify_after_remark_work_1(); 2326 } else { 2327 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2328 // In this second variant of verification, we flag an error 2329 // (i.e. an object reachable in the new marks-vector not reachable 2330 // in the CMS marks-vector) immediately, also indicating the 2331 // identify of an object (A) that references the unmarked object (B) -- 2332 // presumably, a mutation to A failed to be picked up by preclean/remark? 2333 verify_after_remark_work_2(); 2334 } 2335 2336 return true; 2337 } 2338 2339 void CMSCollector::verify_after_remark_work_1() { 2340 ResourceMark rm; 2341 HandleMark hm; 2342 CMSHeap* heap = CMSHeap::heap(); 2343 2344 // Get a clear set of claim bits for the roots processing to work with. 2345 ClassLoaderDataGraph::clear_claimed_marks(); 2346 2347 // Mark from roots one level into CMS 2348 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2349 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2350 2351 { 2352 StrongRootsScope srs(1); 2353 2354 heap->cms_process_roots(&srs, 2355 true, // young gen as roots 2356 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2357 should_unload_classes(), 2358 ¬Older, 2359 NULL); 2360 } 2361 2362 // Now mark from the roots 2363 MarkFromRootsClosure markFromRootsClosure(this, _span, 2364 verification_mark_bm(), verification_mark_stack(), 2365 false /* don't yield */, true /* verifying */); 2366 assert(_restart_addr == NULL, "Expected pre-condition"); 2367 verification_mark_bm()->iterate(&markFromRootsClosure); 2368 while (_restart_addr != NULL) { 2369 // Deal with stack overflow: by restarting at the indicated 2370 // address. 2371 HeapWord* ra = _restart_addr; 2372 markFromRootsClosure.reset(ra); 2373 _restart_addr = NULL; 2374 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2375 } 2376 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2377 verify_work_stacks_empty(); 2378 2379 // Marking completed -- now verify that each bit marked in 2380 // verification_mark_bm() is also marked in markBitMap(); flag all 2381 // errors by printing corresponding objects. 2382 VerifyMarkedClosure vcl(markBitMap()); 2383 verification_mark_bm()->iterate(&vcl); 2384 if (vcl.failed()) { 2385 Log(gc, verify) log; 2386 log.error("Failed marking verification after remark"); 2387 ResourceMark rm; 2388 LogStream ls(log.error()); 2389 heap->print_on(&ls); 2390 fatal("CMS: failed marking verification after remark"); 2391 } 2392 } 2393 2394 class VerifyCLDOopsCLDClosure : public CLDClosure { 2395 class VerifyCLDOopsClosure : public OopClosure { 2396 CMSBitMap* _bitmap; 2397 public: 2398 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2399 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2400 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2401 } _oop_closure; 2402 public: 2403 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2404 void do_cld(ClassLoaderData* cld) { 2405 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2406 } 2407 }; 2408 2409 void CMSCollector::verify_after_remark_work_2() { 2410 ResourceMark rm; 2411 HandleMark hm; 2412 CMSHeap* heap = CMSHeap::heap(); 2413 2414 // Get a clear set of claim bits for the roots processing to work with. 2415 ClassLoaderDataGraph::clear_claimed_marks(); 2416 2417 // Mark from roots one level into CMS 2418 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2419 markBitMap()); 2420 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2421 2422 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2423 2424 { 2425 StrongRootsScope srs(1); 2426 2427 heap->cms_process_roots(&srs, 2428 true, // young gen as roots 2429 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2430 should_unload_classes(), 2431 ¬Older, 2432 &cld_closure); 2433 } 2434 2435 // Now mark from the roots 2436 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2437 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2438 assert(_restart_addr == NULL, "Expected pre-condition"); 2439 verification_mark_bm()->iterate(&markFromRootsClosure); 2440 while (_restart_addr != NULL) { 2441 // Deal with stack overflow: by restarting at the indicated 2442 // address. 2443 HeapWord* ra = _restart_addr; 2444 markFromRootsClosure.reset(ra); 2445 _restart_addr = NULL; 2446 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2447 } 2448 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2449 verify_work_stacks_empty(); 2450 2451 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2452 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2453 2454 // Marking completed -- now verify that each bit marked in 2455 // verification_mark_bm() is also marked in markBitMap(); flag all 2456 // errors by printing corresponding objects. 2457 VerifyMarkedClosure vcl(markBitMap()); 2458 verification_mark_bm()->iterate(&vcl); 2459 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2460 } 2461 2462 void ConcurrentMarkSweepGeneration::save_marks() { 2463 // delegate to CMS space 2464 cmsSpace()->save_marks(); 2465 } 2466 2467 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2468 return cmsSpace()->no_allocs_since_save_marks(); 2469 } 2470 2471 void 2472 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2473 if (freelistLock()->owned_by_self()) { 2474 Generation::oop_iterate(cl); 2475 } else { 2476 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2477 Generation::oop_iterate(cl); 2478 } 2479 } 2480 2481 void 2482 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2483 if (freelistLock()->owned_by_self()) { 2484 Generation::object_iterate(cl); 2485 } else { 2486 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2487 Generation::object_iterate(cl); 2488 } 2489 } 2490 2491 void 2492 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2493 if (freelistLock()->owned_by_self()) { 2494 Generation::safe_object_iterate(cl); 2495 } else { 2496 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2497 Generation::safe_object_iterate(cl); 2498 } 2499 } 2500 2501 void 2502 ConcurrentMarkSweepGeneration::post_compact() { 2503 } 2504 2505 void 2506 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2507 // Fix the linear allocation blocks to look like free blocks. 2508 2509 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2510 // are not called when the heap is verified during universe initialization and 2511 // at vm shutdown. 2512 if (freelistLock()->owned_by_self()) { 2513 cmsSpace()->prepare_for_verify(); 2514 } else { 2515 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2516 cmsSpace()->prepare_for_verify(); 2517 } 2518 } 2519 2520 void 2521 ConcurrentMarkSweepGeneration::verify() { 2522 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2523 // are not called when the heap is verified during universe initialization and 2524 // at vm shutdown. 2525 if (freelistLock()->owned_by_self()) { 2526 cmsSpace()->verify(); 2527 } else { 2528 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2529 cmsSpace()->verify(); 2530 } 2531 } 2532 2533 void CMSCollector::verify() { 2534 _cmsGen->verify(); 2535 } 2536 2537 #ifndef PRODUCT 2538 bool CMSCollector::overflow_list_is_empty() const { 2539 assert(_num_par_pushes >= 0, "Inconsistency"); 2540 if (_overflow_list == NULL) { 2541 assert(_num_par_pushes == 0, "Inconsistency"); 2542 } 2543 return _overflow_list == NULL; 2544 } 2545 2546 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2547 // merely consolidate assertion checks that appear to occur together frequently. 2548 void CMSCollector::verify_work_stacks_empty() const { 2549 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2550 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2551 } 2552 2553 void CMSCollector::verify_overflow_empty() const { 2554 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2555 assert(no_preserved_marks(), "No preserved marks"); 2556 } 2557 #endif // PRODUCT 2558 2559 // Decide if we want to enable class unloading as part of the 2560 // ensuing concurrent GC cycle. We will collect and 2561 // unload classes if it's the case that: 2562 // (a) class unloading is enabled at the command line, and 2563 // (b) old gen is getting really full 2564 // NOTE: Provided there is no change in the state of the heap between 2565 // calls to this method, it should have idempotent results. Moreover, 2566 // its results should be monotonically increasing (i.e. going from 0 to 1, 2567 // but not 1 to 0) between successive calls between which the heap was 2568 // not collected. For the implementation below, it must thus rely on 2569 // the property that concurrent_cycles_since_last_unload() 2570 // will not decrease unless a collection cycle happened and that 2571 // _cmsGen->is_too_full() are 2572 // themselves also monotonic in that sense. See check_monotonicity() 2573 // below. 2574 void CMSCollector::update_should_unload_classes() { 2575 _should_unload_classes = false; 2576 if (CMSClassUnloadingEnabled) { 2577 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2578 CMSClassUnloadingMaxInterval) 2579 || _cmsGen->is_too_full(); 2580 } 2581 } 2582 2583 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2584 bool res = should_concurrent_collect(); 2585 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2586 return res; 2587 } 2588 2589 void CMSCollector::setup_cms_unloading_and_verification_state() { 2590 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2591 || VerifyBeforeExit; 2592 const int rso = GenCollectedHeap::SO_AllCodeCache; 2593 2594 // We set the proper root for this CMS cycle here. 2595 if (should_unload_classes()) { // Should unload classes this cycle 2596 remove_root_scanning_option(rso); // Shrink the root set appropriately 2597 set_verifying(should_verify); // Set verification state for this cycle 2598 return; // Nothing else needs to be done at this time 2599 } 2600 2601 // Not unloading classes this cycle 2602 assert(!should_unload_classes(), "Inconsistency!"); 2603 2604 // If we are not unloading classes then add SO_AllCodeCache to root 2605 // scanning options. 2606 add_root_scanning_option(rso); 2607 2608 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2609 set_verifying(true); 2610 } else if (verifying() && !should_verify) { 2611 // We were verifying, but some verification flags got disabled. 2612 set_verifying(false); 2613 // Exclude symbols, strings and code cache elements from root scanning to 2614 // reduce IM and RM pauses. 2615 remove_root_scanning_option(rso); 2616 } 2617 } 2618 2619 2620 #ifndef PRODUCT 2621 HeapWord* CMSCollector::block_start(const void* p) const { 2622 const HeapWord* addr = (HeapWord*)p; 2623 if (_span.contains(p)) { 2624 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2625 return _cmsGen->cmsSpace()->block_start(p); 2626 } 2627 } 2628 return NULL; 2629 } 2630 #endif 2631 2632 HeapWord* 2633 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2634 bool tlab, 2635 bool parallel) { 2636 CMSSynchronousYieldRequest yr; 2637 assert(!tlab, "Can't deal with TLAB allocation"); 2638 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2639 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2640 if (GCExpandToAllocateDelayMillis > 0) { 2641 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2642 } 2643 return have_lock_and_allocate(word_size, tlab); 2644 } 2645 2646 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2647 size_t bytes, 2648 size_t expand_bytes, 2649 CMSExpansionCause::Cause cause) 2650 { 2651 2652 bool success = expand(bytes, expand_bytes); 2653 2654 // remember why we expanded; this information is used 2655 // by shouldConcurrentCollect() when making decisions on whether to start 2656 // a new CMS cycle. 2657 if (success) { 2658 set_expansion_cause(cause); 2659 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2660 } 2661 } 2662 2663 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2664 HeapWord* res = NULL; 2665 MutexLocker x(ParGCRareEvent_lock); 2666 while (true) { 2667 // Expansion by some other thread might make alloc OK now: 2668 res = ps->lab.alloc(word_sz); 2669 if (res != NULL) return res; 2670 // If there's not enough expansion space available, give up. 2671 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2672 return NULL; 2673 } 2674 // Otherwise, we try expansion. 2675 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2676 // Now go around the loop and try alloc again; 2677 // A competing par_promote might beat us to the expansion space, 2678 // so we may go around the loop again if promotion fails again. 2679 if (GCExpandToAllocateDelayMillis > 0) { 2680 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2681 } 2682 } 2683 } 2684 2685 2686 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2687 PromotionInfo* promo) { 2688 MutexLocker x(ParGCRareEvent_lock); 2689 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2690 while (true) { 2691 // Expansion by some other thread might make alloc OK now: 2692 if (promo->ensure_spooling_space()) { 2693 assert(promo->has_spooling_space(), 2694 "Post-condition of successful ensure_spooling_space()"); 2695 return true; 2696 } 2697 // If there's not enough expansion space available, give up. 2698 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2699 return false; 2700 } 2701 // Otherwise, we try expansion. 2702 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2703 // Now go around the loop and try alloc again; 2704 // A competing allocation might beat us to the expansion space, 2705 // so we may go around the loop again if allocation fails again. 2706 if (GCExpandToAllocateDelayMillis > 0) { 2707 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2708 } 2709 } 2710 } 2711 2712 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2713 // Only shrink if a compaction was done so that all the free space 2714 // in the generation is in a contiguous block at the end. 2715 if (did_compact()) { 2716 CardGeneration::shrink(bytes); 2717 } 2718 } 2719 2720 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2721 assert_locked_or_safepoint(Heap_lock); 2722 } 2723 2724 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2725 assert_locked_or_safepoint(Heap_lock); 2726 assert_lock_strong(freelistLock()); 2727 log_trace(gc)("Shrinking of CMS not yet implemented"); 2728 return; 2729 } 2730 2731 2732 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2733 // phases. 2734 class CMSPhaseAccounting: public StackObj { 2735 public: 2736 CMSPhaseAccounting(CMSCollector *collector, 2737 const char *title); 2738 ~CMSPhaseAccounting(); 2739 2740 private: 2741 CMSCollector *_collector; 2742 const char *_title; 2743 GCTraceConcTime(Info, gc) _trace_time; 2744 2745 public: 2746 // Not MT-safe; so do not pass around these StackObj's 2747 // where they may be accessed by other threads. 2748 double wallclock_millis() { 2749 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2750 } 2751 }; 2752 2753 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2754 const char *title) : 2755 _collector(collector), _title(title), _trace_time(title) { 2756 2757 _collector->resetYields(); 2758 _collector->resetTimer(); 2759 _collector->startTimer(); 2760 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2761 } 2762 2763 CMSPhaseAccounting::~CMSPhaseAccounting() { 2764 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2765 _collector->stopTimer(); 2766 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_millis(_collector->timerTicks())); 2767 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2768 } 2769 2770 // CMS work 2771 2772 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2773 class CMSParMarkTask : public AbstractGangTask { 2774 protected: 2775 CMSCollector* _collector; 2776 uint _n_workers; 2777 OopStorage::ParState<false, false> _par_state_string; 2778 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2779 AbstractGangTask(name), 2780 _collector(collector), 2781 _n_workers(n_workers), 2782 _par_state_string(StringTable::weak_storage()) {} 2783 // Work method in support of parallel rescan ... of young gen spaces 2784 void do_young_space_rescan(OopsInGenClosure* cl, 2785 ContiguousSpace* space, 2786 HeapWord** chunk_array, size_t chunk_top); 2787 void work_on_young_gen_roots(OopsInGenClosure* cl); 2788 }; 2789 2790 // Parallel initial mark task 2791 class CMSParInitialMarkTask: public CMSParMarkTask { 2792 StrongRootsScope* _strong_roots_scope; 2793 public: 2794 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2795 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2796 _strong_roots_scope(strong_roots_scope) {} 2797 void work(uint worker_id); 2798 }; 2799 2800 // Checkpoint the roots into this generation from outside 2801 // this generation. [Note this initial checkpoint need only 2802 // be approximate -- we'll do a catch up phase subsequently.] 2803 void CMSCollector::checkpointRootsInitial() { 2804 assert(_collectorState == InitialMarking, "Wrong collector state"); 2805 check_correct_thread_executing(); 2806 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2807 2808 save_heap_summary(); 2809 report_heap_summary(GCWhen::BeforeGC); 2810 2811 ReferenceProcessor* rp = ref_processor(); 2812 assert(_restart_addr == NULL, "Control point invariant"); 2813 { 2814 // acquire locks for subsequent manipulations 2815 MutexLockerEx x(bitMapLock(), 2816 Mutex::_no_safepoint_check_flag); 2817 checkpointRootsInitialWork(); 2818 // enable ("weak") refs discovery 2819 rp->enable_discovery(); 2820 _collectorState = Marking; 2821 } 2822 } 2823 2824 void CMSCollector::checkpointRootsInitialWork() { 2825 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2826 assert(_collectorState == InitialMarking, "just checking"); 2827 2828 // Already have locks. 2829 assert_lock_strong(bitMapLock()); 2830 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2831 2832 // Setup the verification and class unloading state for this 2833 // CMS collection cycle. 2834 setup_cms_unloading_and_verification_state(); 2835 2836 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2837 2838 // Reset all the PLAB chunk arrays if necessary. 2839 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2840 reset_survivor_plab_arrays(); 2841 } 2842 2843 ResourceMark rm; 2844 HandleMark hm; 2845 2846 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2847 CMSHeap* heap = CMSHeap::heap(); 2848 2849 verify_work_stacks_empty(); 2850 verify_overflow_empty(); 2851 2852 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2853 // Update the saved marks which may affect the root scans. 2854 heap->save_marks(); 2855 2856 // weak reference processing has not started yet. 2857 ref_processor()->set_enqueuing_is_done(false); 2858 2859 // Need to remember all newly created CLDs, 2860 // so that we can guarantee that the remark finds them. 2861 ClassLoaderDataGraph::remember_new_clds(true); 2862 2863 // Whenever a CLD is found, it will be claimed before proceeding to mark 2864 // the klasses. The claimed marks need to be cleared before marking starts. 2865 ClassLoaderDataGraph::clear_claimed_marks(); 2866 2867 print_eden_and_survivor_chunk_arrays(); 2868 2869 { 2870 #if COMPILER2_OR_JVMCI 2871 DerivedPointerTableDeactivate dpt_deact; 2872 #endif 2873 if (CMSParallelInitialMarkEnabled) { 2874 // The parallel version. 2875 WorkGang* workers = heap->workers(); 2876 assert(workers != NULL, "Need parallel worker threads."); 2877 uint n_workers = workers->active_workers(); 2878 2879 StrongRootsScope srs(n_workers); 2880 2881 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2882 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2883 // If the total workers is greater than 1, then multiple workers 2884 // may be used at some time and the initialization has been set 2885 // such that the single threaded path cannot be used. 2886 if (workers->total_workers() > 1) { 2887 workers->run_task(&tsk); 2888 } else { 2889 tsk.work(0); 2890 } 2891 } else { 2892 // The serial version. 2893 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2894 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2895 2896 StrongRootsScope srs(1); 2897 2898 heap->cms_process_roots(&srs, 2899 true, // young gen as roots 2900 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2901 should_unload_classes(), 2902 ¬Older, 2903 &cld_closure); 2904 } 2905 } 2906 2907 // Clear mod-union table; it will be dirtied in the prologue of 2908 // CMS generation per each young generation collection. 2909 2910 assert(_modUnionTable.isAllClear(), 2911 "Was cleared in most recent final checkpoint phase" 2912 " or no bits are set in the gc_prologue before the start of the next " 2913 "subsequent marking phase."); 2914 2915 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2916 2917 // Save the end of the used_region of the constituent generations 2918 // to be used to limit the extent of sweep in each generation. 2919 save_sweep_limits(); 2920 verify_overflow_empty(); 2921 } 2922 2923 bool CMSCollector::markFromRoots() { 2924 // we might be tempted to assert that: 2925 // assert(!SafepointSynchronize::is_at_safepoint(), 2926 // "inconsistent argument?"); 2927 // However that wouldn't be right, because it's possible that 2928 // a safepoint is indeed in progress as a young generation 2929 // stop-the-world GC happens even as we mark in this generation. 2930 assert(_collectorState == Marking, "inconsistent state?"); 2931 check_correct_thread_executing(); 2932 verify_overflow_empty(); 2933 2934 // Weak ref discovery note: We may be discovering weak 2935 // refs in this generation concurrent (but interleaved) with 2936 // weak ref discovery by the young generation collector. 2937 2938 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2939 GCTraceCPUTime tcpu; 2940 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2941 bool res = markFromRootsWork(); 2942 if (res) { 2943 _collectorState = Precleaning; 2944 } else { // We failed and a foreground collection wants to take over 2945 assert(_foregroundGCIsActive, "internal state inconsistency"); 2946 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2947 log_debug(gc)("bailing out to foreground collection"); 2948 } 2949 verify_overflow_empty(); 2950 return res; 2951 } 2952 2953 bool CMSCollector::markFromRootsWork() { 2954 // iterate over marked bits in bit map, doing a full scan and mark 2955 // from these roots using the following algorithm: 2956 // . if oop is to the right of the current scan pointer, 2957 // mark corresponding bit (we'll process it later) 2958 // . else (oop is to left of current scan pointer) 2959 // push oop on marking stack 2960 // . drain the marking stack 2961 2962 // Note that when we do a marking step we need to hold the 2963 // bit map lock -- recall that direct allocation (by mutators) 2964 // and promotion (by the young generation collector) is also 2965 // marking the bit map. [the so-called allocate live policy.] 2966 // Because the implementation of bit map marking is not 2967 // robust wrt simultaneous marking of bits in the same word, 2968 // we need to make sure that there is no such interference 2969 // between concurrent such updates. 2970 2971 // already have locks 2972 assert_lock_strong(bitMapLock()); 2973 2974 verify_work_stacks_empty(); 2975 verify_overflow_empty(); 2976 bool result = false; 2977 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2978 result = do_marking_mt(); 2979 } else { 2980 result = do_marking_st(); 2981 } 2982 return result; 2983 } 2984 2985 // Forward decl 2986 class CMSConcMarkingTask; 2987 2988 class CMSConcMarkingParallelTerminator: public ParallelTaskTerminator { 2989 CMSCollector* _collector; 2990 CMSConcMarkingTask* _task; 2991 public: 2992 virtual void yield(); 2993 2994 // "n_threads" is the number of threads to be terminated. 2995 // "queue_set" is a set of work queues of other threads. 2996 // "collector" is the CMS collector associated with this task terminator. 2997 // "yield" indicates whether we need the gang as a whole to yield. 2998 CMSConcMarkingParallelTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2999 ParallelTaskTerminator(n_threads, queue_set), 3000 _collector(collector) { } 3001 3002 void set_task(CMSConcMarkingTask* task) { 3003 _task = task; 3004 } 3005 }; 3006 3007 class CMSConcMarkingOWSTTerminator: public OWSTTaskTerminator { 3008 CMSCollector* _collector; 3009 CMSConcMarkingTask* _task; 3010 public: 3011 virtual void yield(); 3012 3013 // "n_threads" is the number of threads to be terminated. 3014 // "queue_set" is a set of work queues of other threads. 3015 // "collector" is the CMS collector associated with this task terminator. 3016 // "yield" indicates whether we need the gang as a whole to yield. 3017 CMSConcMarkingOWSTTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3018 OWSTTaskTerminator(n_threads, queue_set), 3019 _collector(collector) { } 3020 3021 void set_task(CMSConcMarkingTask* task) { 3022 _task = task; 3023 } 3024 }; 3025 3026 class CMSConcMarkingTaskTerminator { 3027 private: 3028 ParallelTaskTerminator* _term; 3029 public: 3030 CMSConcMarkingTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) { 3031 if (UseOWSTTaskTerminator) { 3032 _term = new CMSConcMarkingOWSTTerminator(n_threads, queue_set, collector); 3033 } else { 3034 _term = new CMSConcMarkingParallelTerminator(n_threads, queue_set, collector); 3035 } 3036 } 3037 ~CMSConcMarkingTaskTerminator() { 3038 assert(_term != NULL, "Must not be NULL"); 3039 delete _term; 3040 } 3041 3042 void set_task(CMSConcMarkingTask* task); 3043 ParallelTaskTerminator* terminator() const { return _term; } 3044 }; 3045 3046 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3047 CMSConcMarkingTask* _task; 3048 public: 3049 bool should_exit_termination(); 3050 void set_task(CMSConcMarkingTask* task) { 3051 _task = task; 3052 } 3053 }; 3054 3055 // MT Concurrent Marking Task 3056 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3057 CMSCollector* _collector; 3058 uint _n_workers; // requested/desired # workers 3059 bool _result; 3060 CompactibleFreeListSpace* _cms_space; 3061 char _pad_front[64]; // padding to ... 3062 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3063 char _pad_back[64]; 3064 HeapWord* _restart_addr; 3065 3066 // Exposed here for yielding support 3067 Mutex* const _bit_map_lock; 3068 3069 // The per thread work queues, available here for stealing 3070 OopTaskQueueSet* _task_queues; 3071 3072 // Termination (and yielding) support 3073 CMSConcMarkingTaskTerminator _term; 3074 CMSConcMarkingTerminatorTerminator _term_term; 3075 3076 public: 3077 CMSConcMarkingTask(CMSCollector* collector, 3078 CompactibleFreeListSpace* cms_space, 3079 YieldingFlexibleWorkGang* workers, 3080 OopTaskQueueSet* task_queues): 3081 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3082 _collector(collector), 3083 _n_workers(0), 3084 _result(true), 3085 _cms_space(cms_space), 3086 _bit_map_lock(collector->bitMapLock()), 3087 _task_queues(task_queues), 3088 _term(_n_workers, task_queues, _collector) 3089 { 3090 _requested_size = _n_workers; 3091 _term.set_task(this); 3092 _term_term.set_task(this); 3093 _restart_addr = _global_finger = _cms_space->bottom(); 3094 } 3095 3096 3097 OopTaskQueueSet* task_queues() { return _task_queues; } 3098 3099 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3100 3101 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3102 3103 ParallelTaskTerminator* terminator() { return _term.terminator(); } 3104 3105 virtual void set_for_termination(uint active_workers) { 3106 terminator()->reset_for_reuse(active_workers); 3107 } 3108 3109 void work(uint worker_id); 3110 bool should_yield() { 3111 return ConcurrentMarkSweepThread::should_yield() 3112 && !_collector->foregroundGCIsActive(); 3113 } 3114 3115 virtual void coordinator_yield(); // stuff done by coordinator 3116 bool result() { return _result; } 3117 3118 void reset(HeapWord* ra) { 3119 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3120 _restart_addr = _global_finger = ra; 3121 _term.terminator()->reset_for_reuse(); 3122 } 3123 3124 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3125 OopTaskQueue* work_q); 3126 3127 private: 3128 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3129 void do_work_steal(int i); 3130 void bump_global_finger(HeapWord* f); 3131 }; 3132 3133 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3134 assert(_task != NULL, "Error"); 3135 return _task->yielding(); 3136 // Note that we do not need the disjunct || _task->should_yield() above 3137 // because we want terminating threads to yield only if the task 3138 // is already in the midst of yielding, which happens only after at least one 3139 // thread has yielded. 3140 } 3141 3142 void CMSConcMarkingParallelTerminator::yield() { 3143 if (_task->should_yield()) { 3144 _task->yield(); 3145 } else { 3146 ParallelTaskTerminator::yield(); 3147 } 3148 } 3149 3150 void CMSConcMarkingOWSTTerminator::yield() { 3151 if (_task->should_yield()) { 3152 _task->yield(); 3153 } else { 3154 OWSTTaskTerminator::yield(); 3155 } 3156 } 3157 3158 void CMSConcMarkingTaskTerminator::set_task(CMSConcMarkingTask* task) { 3159 if (UseOWSTTaskTerminator) { 3160 ((CMSConcMarkingOWSTTerminator*)_term)->set_task(task); 3161 } else { 3162 ((CMSConcMarkingParallelTerminator*)_term)->set_task(task); 3163 } 3164 } 3165 3166 //////////////////////////////////////////////////////////////// 3167 // Concurrent Marking Algorithm Sketch 3168 //////////////////////////////////////////////////////////////// 3169 // Until all tasks exhausted (both spaces): 3170 // -- claim next available chunk 3171 // -- bump global finger via CAS 3172 // -- find first object that starts in this chunk 3173 // and start scanning bitmap from that position 3174 // -- scan marked objects for oops 3175 // -- CAS-mark target, and if successful: 3176 // . if target oop is above global finger (volatile read) 3177 // nothing to do 3178 // . if target oop is in chunk and above local finger 3179 // then nothing to do 3180 // . else push on work-queue 3181 // -- Deal with possible overflow issues: 3182 // . local work-queue overflow causes stuff to be pushed on 3183 // global (common) overflow queue 3184 // . always first empty local work queue 3185 // . then get a batch of oops from global work queue if any 3186 // . then do work stealing 3187 // -- When all tasks claimed (both spaces) 3188 // and local work queue empty, 3189 // then in a loop do: 3190 // . check global overflow stack; steal a batch of oops and trace 3191 // . try to steal from other threads oif GOS is empty 3192 // . if neither is available, offer termination 3193 // -- Terminate and return result 3194 // 3195 void CMSConcMarkingTask::work(uint worker_id) { 3196 elapsedTimer _timer; 3197 ResourceMark rm; 3198 HandleMark hm; 3199 3200 DEBUG_ONLY(_collector->verify_overflow_empty();) 3201 3202 // Before we begin work, our work queue should be empty 3203 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3204 // Scan the bitmap covering _cms_space, tracing through grey objects. 3205 _timer.start(); 3206 do_scan_and_mark(worker_id, _cms_space); 3207 _timer.stop(); 3208 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3209 3210 // ... do work stealing 3211 _timer.reset(); 3212 _timer.start(); 3213 do_work_steal(worker_id); 3214 _timer.stop(); 3215 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3216 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3217 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3218 // Note that under the current task protocol, the 3219 // following assertion is true even of the spaces 3220 // expanded since the completion of the concurrent 3221 // marking. XXX This will likely change under a strict 3222 // ABORT semantics. 3223 // After perm removal the comparison was changed to 3224 // greater than or equal to from strictly greater than. 3225 // Before perm removal the highest address sweep would 3226 // have been at the end of perm gen but now is at the 3227 // end of the tenured gen. 3228 assert(_global_finger >= _cms_space->end(), 3229 "All tasks have been completed"); 3230 DEBUG_ONLY(_collector->verify_overflow_empty();) 3231 } 3232 3233 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3234 HeapWord* read = _global_finger; 3235 HeapWord* cur = read; 3236 while (f > read) { 3237 cur = read; 3238 read = Atomic::cmpxchg(f, &_global_finger, cur); 3239 if (cur == read) { 3240 // our cas succeeded 3241 assert(_global_finger >= f, "protocol consistency"); 3242 break; 3243 } 3244 } 3245 } 3246 3247 // This is really inefficient, and should be redone by 3248 // using (not yet available) block-read and -write interfaces to the 3249 // stack and the work_queue. XXX FIX ME !!! 3250 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3251 OopTaskQueue* work_q) { 3252 // Fast lock-free check 3253 if (ovflw_stk->length() == 0) { 3254 return false; 3255 } 3256 assert(work_q->size() == 0, "Shouldn't steal"); 3257 MutexLockerEx ml(ovflw_stk->par_lock(), 3258 Mutex::_no_safepoint_check_flag); 3259 // Grab up to 1/4 the size of the work queue 3260 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3261 (size_t)ParGCDesiredObjsFromOverflowList); 3262 num = MIN2(num, ovflw_stk->length()); 3263 for (int i = (int) num; i > 0; i--) { 3264 oop cur = ovflw_stk->pop(); 3265 assert(cur != NULL, "Counted wrong?"); 3266 work_q->push(cur); 3267 } 3268 return num > 0; 3269 } 3270 3271 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3272 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3273 int n_tasks = pst->n_tasks(); 3274 // We allow that there may be no tasks to do here because 3275 // we are restarting after a stack overflow. 3276 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3277 uint nth_task = 0; 3278 3279 HeapWord* aligned_start = sp->bottom(); 3280 if (sp->used_region().contains(_restart_addr)) { 3281 // Align down to a card boundary for the start of 0th task 3282 // for this space. 3283 aligned_start = align_down(_restart_addr, CardTable::card_size); 3284 } 3285 3286 size_t chunk_size = sp->marking_task_size(); 3287 while (pst->try_claim_task(/* reference */ nth_task)) { 3288 // Having claimed the nth task in this space, 3289 // compute the chunk that it corresponds to: 3290 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3291 aligned_start + (nth_task+1)*chunk_size); 3292 // Try and bump the global finger via a CAS; 3293 // note that we need to do the global finger bump 3294 // _before_ taking the intersection below, because 3295 // the task corresponding to that region will be 3296 // deemed done even if the used_region() expands 3297 // because of allocation -- as it almost certainly will 3298 // during start-up while the threads yield in the 3299 // closure below. 3300 HeapWord* finger = span.end(); 3301 bump_global_finger(finger); // atomically 3302 // There are null tasks here corresponding to chunks 3303 // beyond the "top" address of the space. 3304 span = span.intersection(sp->used_region()); 3305 if (!span.is_empty()) { // Non-null task 3306 HeapWord* prev_obj; 3307 assert(!span.contains(_restart_addr) || nth_task == 0, 3308 "Inconsistency"); 3309 if (nth_task == 0) { 3310 // For the 0th task, we'll not need to compute a block_start. 3311 if (span.contains(_restart_addr)) { 3312 // In the case of a restart because of stack overflow, 3313 // we might additionally skip a chunk prefix. 3314 prev_obj = _restart_addr; 3315 } else { 3316 prev_obj = span.start(); 3317 } 3318 } else { 3319 // We want to skip the first object because 3320 // the protocol is to scan any object in its entirety 3321 // that _starts_ in this span; a fortiori, any 3322 // object starting in an earlier span is scanned 3323 // as part of an earlier claimed task. 3324 // Below we use the "careful" version of block_start 3325 // so we do not try to navigate uninitialized objects. 3326 prev_obj = sp->block_start_careful(span.start()); 3327 // Below we use a variant of block_size that uses the 3328 // Printezis bits to avoid waiting for allocated 3329 // objects to become initialized/parsable. 3330 while (prev_obj < span.start()) { 3331 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3332 if (sz > 0) { 3333 prev_obj += sz; 3334 } else { 3335 // In this case we may end up doing a bit of redundant 3336 // scanning, but that appears unavoidable, short of 3337 // locking the free list locks; see bug 6324141. 3338 break; 3339 } 3340 } 3341 } 3342 if (prev_obj < span.end()) { 3343 MemRegion my_span = MemRegion(prev_obj, span.end()); 3344 // Do the marking work within a non-empty span -- 3345 // the last argument to the constructor indicates whether the 3346 // iteration should be incremental with periodic yields. 3347 ParMarkFromRootsClosure cl(this, _collector, my_span, 3348 &_collector->_markBitMap, 3349 work_queue(i), 3350 &_collector->_markStack); 3351 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3352 } // else nothing to do for this task 3353 } // else nothing to do for this task 3354 } 3355 // We'd be tempted to assert here that since there are no 3356 // more tasks left to claim in this space, the global_finger 3357 // must exceed space->top() and a fortiori space->end(). However, 3358 // that would not quite be correct because the bumping of 3359 // global_finger occurs strictly after the claiming of a task, 3360 // so by the time we reach here the global finger may not yet 3361 // have been bumped up by the thread that claimed the last 3362 // task. 3363 pst->all_tasks_completed(); 3364 } 3365 3366 class ParConcMarkingClosure: public MetadataVisitingOopIterateClosure { 3367 private: 3368 CMSCollector* _collector; 3369 CMSConcMarkingTask* _task; 3370 MemRegion _span; 3371 CMSBitMap* _bit_map; 3372 CMSMarkStack* _overflow_stack; 3373 OopTaskQueue* _work_queue; 3374 protected: 3375 DO_OOP_WORK_DEFN 3376 public: 3377 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3378 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3379 MetadataVisitingOopIterateClosure(collector->ref_processor()), 3380 _collector(collector), 3381 _task(task), 3382 _span(collector->_span), 3383 _bit_map(bit_map), 3384 _overflow_stack(overflow_stack), 3385 _work_queue(work_queue) 3386 { } 3387 virtual void do_oop(oop* p); 3388 virtual void do_oop(narrowOop* p); 3389 3390 void trim_queue(size_t max); 3391 void handle_stack_overflow(HeapWord* lost); 3392 void do_yield_check() { 3393 if (_task->should_yield()) { 3394 _task->yield(); 3395 } 3396 } 3397 }; 3398 3399 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3400 3401 // Grey object scanning during work stealing phase -- 3402 // the salient assumption here is that any references 3403 // that are in these stolen objects being scanned must 3404 // already have been initialized (else they would not have 3405 // been published), so we do not need to check for 3406 // uninitialized objects before pushing here. 3407 void ParConcMarkingClosure::do_oop(oop obj) { 3408 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3409 HeapWord* addr = (HeapWord*)obj; 3410 // Check if oop points into the CMS generation 3411 // and is not marked 3412 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3413 // a white object ... 3414 // If we manage to "claim" the object, by being the 3415 // first thread to mark it, then we push it on our 3416 // marking stack 3417 if (_bit_map->par_mark(addr)) { // ... now grey 3418 // push on work queue (grey set) 3419 bool simulate_overflow = false; 3420 NOT_PRODUCT( 3421 if (CMSMarkStackOverflowALot && 3422 _collector->simulate_overflow()) { 3423 // simulate a stack overflow 3424 simulate_overflow = true; 3425 } 3426 ) 3427 if (simulate_overflow || 3428 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3429 // stack overflow 3430 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3431 // We cannot assert that the overflow stack is full because 3432 // it may have been emptied since. 3433 assert(simulate_overflow || 3434 _work_queue->size() == _work_queue->max_elems(), 3435 "Else push should have succeeded"); 3436 handle_stack_overflow(addr); 3437 } 3438 } // Else, some other thread got there first 3439 do_yield_check(); 3440 } 3441 } 3442 3443 void ParConcMarkingClosure::trim_queue(size_t max) { 3444 while (_work_queue->size() > max) { 3445 oop new_oop; 3446 if (_work_queue->pop_local(new_oop)) { 3447 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3448 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3449 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3450 new_oop->oop_iterate(this); // do_oop() above 3451 do_yield_check(); 3452 } 3453 } 3454 } 3455 3456 // Upon stack overflow, we discard (part of) the stack, 3457 // remembering the least address amongst those discarded 3458 // in CMSCollector's _restart_address. 3459 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3460 // We need to do this under a mutex to prevent other 3461 // workers from interfering with the work done below. 3462 MutexLockerEx ml(_overflow_stack->par_lock(), 3463 Mutex::_no_safepoint_check_flag); 3464 // Remember the least grey address discarded 3465 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3466 _collector->lower_restart_addr(ra); 3467 _overflow_stack->reset(); // discard stack contents 3468 _overflow_stack->expand(); // expand the stack if possible 3469 } 3470 3471 3472 void CMSConcMarkingTask::do_work_steal(int i) { 3473 OopTaskQueue* work_q = work_queue(i); 3474 oop obj_to_scan; 3475 CMSBitMap* bm = &(_collector->_markBitMap); 3476 CMSMarkStack* ovflw = &(_collector->_markStack); 3477 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3478 while (true) { 3479 cl.trim_queue(0); 3480 assert(work_q->size() == 0, "Should have been emptied above"); 3481 if (get_work_from_overflow_stack(ovflw, work_q)) { 3482 // Can't assert below because the work obtained from the 3483 // overflow stack may already have been stolen from us. 3484 // assert(work_q->size() > 0, "Work from overflow stack"); 3485 continue; 3486 } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 3487 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3488 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3489 obj_to_scan->oop_iterate(&cl); 3490 } else if (terminator()->offer_termination(&_term_term)) { 3491 assert(work_q->size() == 0, "Impossible!"); 3492 break; 3493 } else if (yielding() || should_yield()) { 3494 yield(); 3495 } 3496 } 3497 } 3498 3499 // This is run by the CMS (coordinator) thread. 3500 void CMSConcMarkingTask::coordinator_yield() { 3501 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3502 "CMS thread should hold CMS token"); 3503 // First give up the locks, then yield, then re-lock 3504 // We should probably use a constructor/destructor idiom to 3505 // do this unlock/lock or modify the MutexUnlocker class to 3506 // serve our purpose. XXX 3507 assert_lock_strong(_bit_map_lock); 3508 _bit_map_lock->unlock(); 3509 ConcurrentMarkSweepThread::desynchronize(true); 3510 _collector->stopTimer(); 3511 _collector->incrementYields(); 3512 3513 // It is possible for whichever thread initiated the yield request 3514 // not to get a chance to wake up and take the bitmap lock between 3515 // this thread releasing it and reacquiring it. So, while the 3516 // should_yield() flag is on, let's sleep for a bit to give the 3517 // other thread a chance to wake up. The limit imposed on the number 3518 // of iterations is defensive, to avoid any unforseen circumstances 3519 // putting us into an infinite loop. Since it's always been this 3520 // (coordinator_yield()) method that was observed to cause the 3521 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3522 // which is by default non-zero. For the other seven methods that 3523 // also perform the yield operation, as are using a different 3524 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3525 // can enable the sleeping for those methods too, if necessary. 3526 // See 6442774. 3527 // 3528 // We really need to reconsider the synchronization between the GC 3529 // thread and the yield-requesting threads in the future and we 3530 // should really use wait/notify, which is the recommended 3531 // way of doing this type of interaction. Additionally, we should 3532 // consolidate the eight methods that do the yield operation and they 3533 // are almost identical into one for better maintainability and 3534 // readability. See 6445193. 3535 // 3536 // Tony 2006.06.29 3537 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3538 ConcurrentMarkSweepThread::should_yield() && 3539 !CMSCollector::foregroundGCIsActive(); ++i) { 3540 os::sleep(Thread::current(), 1, false); 3541 } 3542 3543 ConcurrentMarkSweepThread::synchronize(true); 3544 _bit_map_lock->lock_without_safepoint_check(); 3545 _collector->startTimer(); 3546 } 3547 3548 bool CMSCollector::do_marking_mt() { 3549 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3550 uint num_workers = WorkerPolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3551 conc_workers()->active_workers(), 3552 Threads::number_of_non_daemon_threads()); 3553 num_workers = conc_workers()->update_active_workers(num_workers); 3554 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3555 3556 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3557 3558 CMSConcMarkingTask tsk(this, 3559 cms_space, 3560 conc_workers(), 3561 task_queues()); 3562 3563 // Since the actual number of workers we get may be different 3564 // from the number we requested above, do we need to do anything different 3565 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3566 // class?? XXX 3567 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3568 3569 // Refs discovery is already non-atomic. 3570 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3571 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3572 conc_workers()->start_task(&tsk); 3573 while (tsk.yielded()) { 3574 tsk.coordinator_yield(); 3575 conc_workers()->continue_task(&tsk); 3576 } 3577 // If the task was aborted, _restart_addr will be non-NULL 3578 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3579 while (_restart_addr != NULL) { 3580 // XXX For now we do not make use of ABORTED state and have not 3581 // yet implemented the right abort semantics (even in the original 3582 // single-threaded CMS case). That needs some more investigation 3583 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3584 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3585 // If _restart_addr is non-NULL, a marking stack overflow 3586 // occurred; we need to do a fresh marking iteration from the 3587 // indicated restart address. 3588 if (_foregroundGCIsActive) { 3589 // We may be running into repeated stack overflows, having 3590 // reached the limit of the stack size, while making very 3591 // slow forward progress. It may be best to bail out and 3592 // let the foreground collector do its job. 3593 // Clear _restart_addr, so that foreground GC 3594 // works from scratch. This avoids the headache of 3595 // a "rescan" which would otherwise be needed because 3596 // of the dirty mod union table & card table. 3597 _restart_addr = NULL; 3598 return false; 3599 } 3600 // Adjust the task to restart from _restart_addr 3601 tsk.reset(_restart_addr); 3602 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3603 _restart_addr); 3604 _restart_addr = NULL; 3605 // Get the workers going again 3606 conc_workers()->start_task(&tsk); 3607 while (tsk.yielded()) { 3608 tsk.coordinator_yield(); 3609 conc_workers()->continue_task(&tsk); 3610 } 3611 } 3612 assert(tsk.completed(), "Inconsistency"); 3613 assert(tsk.result() == true, "Inconsistency"); 3614 return true; 3615 } 3616 3617 bool CMSCollector::do_marking_st() { 3618 ResourceMark rm; 3619 HandleMark hm; 3620 3621 // Temporarily make refs discovery single threaded (non-MT) 3622 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3623 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3624 &_markStack, CMSYield); 3625 // the last argument to iterate indicates whether the iteration 3626 // should be incremental with periodic yields. 3627 _markBitMap.iterate(&markFromRootsClosure); 3628 // If _restart_addr is non-NULL, a marking stack overflow 3629 // occurred; we need to do a fresh iteration from the 3630 // indicated restart address. 3631 while (_restart_addr != NULL) { 3632 if (_foregroundGCIsActive) { 3633 // We may be running into repeated stack overflows, having 3634 // reached the limit of the stack size, while making very 3635 // slow forward progress. It may be best to bail out and 3636 // let the foreground collector do its job. 3637 // Clear _restart_addr, so that foreground GC 3638 // works from scratch. This avoids the headache of 3639 // a "rescan" which would otherwise be needed because 3640 // of the dirty mod union table & card table. 3641 _restart_addr = NULL; 3642 return false; // indicating failure to complete marking 3643 } 3644 // Deal with stack overflow: 3645 // we restart marking from _restart_addr 3646 HeapWord* ra = _restart_addr; 3647 markFromRootsClosure.reset(ra); 3648 _restart_addr = NULL; 3649 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3650 } 3651 return true; 3652 } 3653 3654 void CMSCollector::preclean() { 3655 check_correct_thread_executing(); 3656 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3657 verify_work_stacks_empty(); 3658 verify_overflow_empty(); 3659 _abort_preclean = false; 3660 if (CMSPrecleaningEnabled) { 3661 if (!CMSEdenChunksRecordAlways) { 3662 _eden_chunk_index = 0; 3663 } 3664 size_t used = get_eden_used(); 3665 size_t capacity = get_eden_capacity(); 3666 // Don't start sampling unless we will get sufficiently 3667 // many samples. 3668 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3669 * CMSScheduleRemarkEdenPenetration)) { 3670 _start_sampling = true; 3671 } else { 3672 _start_sampling = false; 3673 } 3674 GCTraceCPUTime tcpu; 3675 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3676 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3677 } 3678 CMSTokenSync x(true); // is cms thread 3679 if (CMSPrecleaningEnabled) { 3680 sample_eden(); 3681 _collectorState = AbortablePreclean; 3682 } else { 3683 _collectorState = FinalMarking; 3684 } 3685 verify_work_stacks_empty(); 3686 verify_overflow_empty(); 3687 } 3688 3689 // Try and schedule the remark such that young gen 3690 // occupancy is CMSScheduleRemarkEdenPenetration %. 3691 void CMSCollector::abortable_preclean() { 3692 check_correct_thread_executing(); 3693 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3694 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3695 3696 // If Eden's current occupancy is below this threshold, 3697 // immediately schedule the remark; else preclean 3698 // past the next scavenge in an effort to 3699 // schedule the pause as described above. By choosing 3700 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3701 // we will never do an actual abortable preclean cycle. 3702 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3703 GCTraceCPUTime tcpu; 3704 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3705 // We need more smarts in the abortable preclean 3706 // loop below to deal with cases where allocation 3707 // in young gen is very very slow, and our precleaning 3708 // is running a losing race against a horde of 3709 // mutators intent on flooding us with CMS updates 3710 // (dirty cards). 3711 // One, admittedly dumb, strategy is to give up 3712 // after a certain number of abortable precleaning loops 3713 // or after a certain maximum time. We want to make 3714 // this smarter in the next iteration. 3715 // XXX FIX ME!!! YSR 3716 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3717 while (!(should_abort_preclean() || 3718 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3719 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3720 cumworkdone += workdone; 3721 loops++; 3722 // Voluntarily terminate abortable preclean phase if we have 3723 // been at it for too long. 3724 if ((CMSMaxAbortablePrecleanLoops != 0) && 3725 loops >= CMSMaxAbortablePrecleanLoops) { 3726 log_debug(gc)(" CMS: abort preclean due to loops "); 3727 break; 3728 } 3729 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3730 log_debug(gc)(" CMS: abort preclean due to time "); 3731 break; 3732 } 3733 // If we are doing little work each iteration, we should 3734 // take a short break. 3735 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3736 // Sleep for some time, waiting for work to accumulate 3737 stopTimer(); 3738 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3739 startTimer(); 3740 waited++; 3741 } 3742 } 3743 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3744 loops, waited, cumworkdone); 3745 } 3746 CMSTokenSync x(true); // is cms thread 3747 if (_collectorState != Idling) { 3748 assert(_collectorState == AbortablePreclean, 3749 "Spontaneous state transition?"); 3750 _collectorState = FinalMarking; 3751 } // Else, a foreground collection completed this CMS cycle. 3752 return; 3753 } 3754 3755 // Respond to an Eden sampling opportunity 3756 void CMSCollector::sample_eden() { 3757 // Make sure a young gc cannot sneak in between our 3758 // reading and recording of a sample. 3759 assert(Thread::current()->is_ConcurrentGC_thread(), 3760 "Only the cms thread may collect Eden samples"); 3761 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3762 "Should collect samples while holding CMS token"); 3763 if (!_start_sampling) { 3764 return; 3765 } 3766 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3767 // is populated by the young generation. 3768 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3769 if (_eden_chunk_index < _eden_chunk_capacity) { 3770 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3771 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3772 "Unexpected state of Eden"); 3773 // We'd like to check that what we just sampled is an oop-start address; 3774 // however, we cannot do that here since the object may not yet have been 3775 // initialized. So we'll instead do the check when we _use_ this sample 3776 // later. 3777 if (_eden_chunk_index == 0 || 3778 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3779 _eden_chunk_array[_eden_chunk_index-1]) 3780 >= CMSSamplingGrain)) { 3781 _eden_chunk_index++; // commit sample 3782 } 3783 } 3784 } 3785 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3786 size_t used = get_eden_used(); 3787 size_t capacity = get_eden_capacity(); 3788 assert(used <= capacity, "Unexpected state of Eden"); 3789 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3790 _abort_preclean = true; 3791 } 3792 } 3793 } 3794 3795 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3796 assert(_collectorState == Precleaning || 3797 _collectorState == AbortablePreclean, "incorrect state"); 3798 ResourceMark rm; 3799 HandleMark hm; 3800 3801 // Precleaning is currently not MT but the reference processor 3802 // may be set for MT. Disable it temporarily here. 3803 ReferenceProcessor* rp = ref_processor(); 3804 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3805 3806 // Do one pass of scrubbing the discovered reference lists 3807 // to remove any reference objects with strongly-reachable 3808 // referents. 3809 if (clean_refs) { 3810 CMSPrecleanRefsYieldClosure yield_cl(this); 3811 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3812 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3813 &_markStack, true /* preclean */); 3814 CMSDrainMarkingStackClosure complete_trace(this, 3815 _span, &_markBitMap, &_markStack, 3816 &keep_alive, true /* preclean */); 3817 3818 // We don't want this step to interfere with a young 3819 // collection because we don't want to take CPU 3820 // or memory bandwidth away from the young GC threads 3821 // (which may be as many as there are CPUs). 3822 // Note that we don't need to protect ourselves from 3823 // interference with mutators because they can't 3824 // manipulate the discovered reference lists nor affect 3825 // the computed reachability of the referents, the 3826 // only properties manipulated by the precleaning 3827 // of these reference lists. 3828 stopTimer(); 3829 CMSTokenSyncWithLocks x(true /* is cms thread */, 3830 bitMapLock()); 3831 startTimer(); 3832 sample_eden(); 3833 3834 // The following will yield to allow foreground 3835 // collection to proceed promptly. XXX YSR: 3836 // The code in this method may need further 3837 // tweaking for better performance and some restructuring 3838 // for cleaner interfaces. 3839 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3840 rp->preclean_discovered_references( 3841 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3842 gc_timer); 3843 } 3844 3845 if (clean_survivor) { // preclean the active survivor space(s) 3846 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3847 &_markBitMap, &_modUnionTable, 3848 &_markStack, true /* precleaning phase */); 3849 stopTimer(); 3850 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3851 bitMapLock()); 3852 startTimer(); 3853 unsigned int before_count = 3854 CMSHeap::heap()->total_collections(); 3855 SurvivorSpacePrecleanClosure 3856 sss_cl(this, _span, &_markBitMap, &_markStack, 3857 &pam_cl, before_count, CMSYield); 3858 _young_gen->from()->object_iterate_careful(&sss_cl); 3859 _young_gen->to()->object_iterate_careful(&sss_cl); 3860 } 3861 MarkRefsIntoAndScanClosure 3862 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3863 &_markStack, this, CMSYield, 3864 true /* precleaning phase */); 3865 // CAUTION: The following closure has persistent state that may need to 3866 // be reset upon a decrease in the sequence of addresses it 3867 // processes. 3868 ScanMarkedObjectsAgainCarefullyClosure 3869 smoac_cl(this, _span, 3870 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3871 3872 // Preclean dirty cards in ModUnionTable and CardTable using 3873 // appropriate convergence criterion; 3874 // repeat CMSPrecleanIter times unless we find that 3875 // we are losing. 3876 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3877 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3878 "Bad convergence multiplier"); 3879 assert(CMSPrecleanThreshold >= 100, 3880 "Unreasonably low CMSPrecleanThreshold"); 3881 3882 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3883 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3884 numIter < CMSPrecleanIter; 3885 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3886 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3887 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3888 // Either there are very few dirty cards, so re-mark 3889 // pause will be small anyway, or our pre-cleaning isn't 3890 // that much faster than the rate at which cards are being 3891 // dirtied, so we might as well stop and re-mark since 3892 // precleaning won't improve our re-mark time by much. 3893 if (curNumCards <= CMSPrecleanThreshold || 3894 (numIter > 0 && 3895 (curNumCards * CMSPrecleanDenominator > 3896 lastNumCards * CMSPrecleanNumerator))) { 3897 numIter++; 3898 cumNumCards += curNumCards; 3899 break; 3900 } 3901 } 3902 3903 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3904 3905 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3906 cumNumCards += curNumCards; 3907 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3908 curNumCards, cumNumCards, numIter); 3909 return cumNumCards; // as a measure of useful work done 3910 } 3911 3912 // PRECLEANING NOTES: 3913 // Precleaning involves: 3914 // . reading the bits of the modUnionTable and clearing the set bits. 3915 // . For the cards corresponding to the set bits, we scan the 3916 // objects on those cards. This means we need the free_list_lock 3917 // so that we can safely iterate over the CMS space when scanning 3918 // for oops. 3919 // . When we scan the objects, we'll be both reading and setting 3920 // marks in the marking bit map, so we'll need the marking bit map. 3921 // . For protecting _collector_state transitions, we take the CGC_lock. 3922 // Note that any races in the reading of of card table entries by the 3923 // CMS thread on the one hand and the clearing of those entries by the 3924 // VM thread or the setting of those entries by the mutator threads on the 3925 // other are quite benign. However, for efficiency it makes sense to keep 3926 // the VM thread from racing with the CMS thread while the latter is 3927 // dirty card info to the modUnionTable. We therefore also use the 3928 // CGC_lock to protect the reading of the card table and the mod union 3929 // table by the CM thread. 3930 // . We run concurrently with mutator updates, so scanning 3931 // needs to be done carefully -- we should not try to scan 3932 // potentially uninitialized objects. 3933 // 3934 // Locking strategy: While holding the CGC_lock, we scan over and 3935 // reset a maximal dirty range of the mod union / card tables, then lock 3936 // the free_list_lock and bitmap lock to do a full marking, then 3937 // release these locks; and repeat the cycle. This allows for a 3938 // certain amount of fairness in the sharing of these locks between 3939 // the CMS collector on the one hand, and the VM thread and the 3940 // mutators on the other. 3941 3942 // NOTE: preclean_mod_union_table() and preclean_card_table() 3943 // further below are largely identical; if you need to modify 3944 // one of these methods, please check the other method too. 3945 3946 size_t CMSCollector::preclean_mod_union_table( 3947 ConcurrentMarkSweepGeneration* old_gen, 3948 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3949 verify_work_stacks_empty(); 3950 verify_overflow_empty(); 3951 3952 // strategy: starting with the first card, accumulate contiguous 3953 // ranges of dirty cards; clear these cards, then scan the region 3954 // covered by these cards. 3955 3956 // Since all of the MUT is committed ahead, we can just use 3957 // that, in case the generations expand while we are precleaning. 3958 // It might also be fine to just use the committed part of the 3959 // generation, but we might potentially miss cards when the 3960 // generation is rapidly expanding while we are in the midst 3961 // of precleaning. 3962 HeapWord* startAddr = old_gen->reserved().start(); 3963 HeapWord* endAddr = old_gen->reserved().end(); 3964 3965 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3966 3967 size_t numDirtyCards, cumNumDirtyCards; 3968 HeapWord *nextAddr, *lastAddr; 3969 for (cumNumDirtyCards = numDirtyCards = 0, 3970 nextAddr = lastAddr = startAddr; 3971 nextAddr < endAddr; 3972 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3973 3974 ResourceMark rm; 3975 HandleMark hm; 3976 3977 MemRegion dirtyRegion; 3978 { 3979 stopTimer(); 3980 // Potential yield point 3981 CMSTokenSync ts(true); 3982 startTimer(); 3983 sample_eden(); 3984 // Get dirty region starting at nextOffset (inclusive), 3985 // simultaneously clearing it. 3986 dirtyRegion = 3987 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3988 assert(dirtyRegion.start() >= nextAddr, 3989 "returned region inconsistent?"); 3990 } 3991 // Remember where the next search should begin. 3992 // The returned region (if non-empty) is a right open interval, 3993 // so lastOffset is obtained from the right end of that 3994 // interval. 3995 lastAddr = dirtyRegion.end(); 3996 // Should do something more transparent and less hacky XXX 3997 numDirtyCards = 3998 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3999 4000 // We'll scan the cards in the dirty region (with periodic 4001 // yields for foreground GC as needed). 4002 if (!dirtyRegion.is_empty()) { 4003 assert(numDirtyCards > 0, "consistency check"); 4004 HeapWord* stop_point = NULL; 4005 stopTimer(); 4006 // Potential yield point 4007 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4008 bitMapLock()); 4009 startTimer(); 4010 { 4011 verify_work_stacks_empty(); 4012 verify_overflow_empty(); 4013 sample_eden(); 4014 stop_point = 4015 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4016 } 4017 if (stop_point != NULL) { 4018 // The careful iteration stopped early either because it found an 4019 // uninitialized object, or because we were in the midst of an 4020 // "abortable preclean", which should now be aborted. Redirty 4021 // the bits corresponding to the partially-scanned or unscanned 4022 // cards. We'll either restart at the next block boundary or 4023 // abort the preclean. 4024 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4025 "Should only be AbortablePreclean."); 4026 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4027 if (should_abort_preclean()) { 4028 break; // out of preclean loop 4029 } else { 4030 // Compute the next address at which preclean should pick up; 4031 // might need bitMapLock in order to read P-bits. 4032 lastAddr = next_card_start_after_block(stop_point); 4033 } 4034 } 4035 } else { 4036 assert(lastAddr == endAddr, "consistency check"); 4037 assert(numDirtyCards == 0, "consistency check"); 4038 break; 4039 } 4040 } 4041 verify_work_stacks_empty(); 4042 verify_overflow_empty(); 4043 return cumNumDirtyCards; 4044 } 4045 4046 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4047 // below are largely identical; if you need to modify 4048 // one of these methods, please check the other method too. 4049 4050 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4051 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4052 // strategy: it's similar to precleamModUnionTable above, in that 4053 // we accumulate contiguous ranges of dirty cards, mark these cards 4054 // precleaned, then scan the region covered by these cards. 4055 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4056 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4057 4058 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4059 4060 size_t numDirtyCards, cumNumDirtyCards; 4061 HeapWord *lastAddr, *nextAddr; 4062 4063 for (cumNumDirtyCards = numDirtyCards = 0, 4064 nextAddr = lastAddr = startAddr; 4065 nextAddr < endAddr; 4066 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4067 4068 ResourceMark rm; 4069 HandleMark hm; 4070 4071 MemRegion dirtyRegion; 4072 { 4073 // See comments in "Precleaning notes" above on why we 4074 // do this locking. XXX Could the locking overheads be 4075 // too high when dirty cards are sparse? [I don't think so.] 4076 stopTimer(); 4077 CMSTokenSync x(true); // is cms thread 4078 startTimer(); 4079 sample_eden(); 4080 // Get and clear dirty region from card table 4081 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4082 true, 4083 CardTable::precleaned_card_val()); 4084 4085 assert(dirtyRegion.start() >= nextAddr, 4086 "returned region inconsistent?"); 4087 } 4088 lastAddr = dirtyRegion.end(); 4089 numDirtyCards = 4090 dirtyRegion.word_size()/CardTable::card_size_in_words; 4091 4092 if (!dirtyRegion.is_empty()) { 4093 stopTimer(); 4094 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4095 startTimer(); 4096 sample_eden(); 4097 verify_work_stacks_empty(); 4098 verify_overflow_empty(); 4099 HeapWord* stop_point = 4100 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4101 if (stop_point != NULL) { 4102 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4103 "Should only be AbortablePreclean."); 4104 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4105 if (should_abort_preclean()) { 4106 break; // out of preclean loop 4107 } else { 4108 // Compute the next address at which preclean should pick up. 4109 lastAddr = next_card_start_after_block(stop_point); 4110 } 4111 } 4112 } else { 4113 break; 4114 } 4115 } 4116 verify_work_stacks_empty(); 4117 verify_overflow_empty(); 4118 return cumNumDirtyCards; 4119 } 4120 4121 class PrecleanCLDClosure : public CLDClosure { 4122 MetadataVisitingOopsInGenClosure* _cm_closure; 4123 public: 4124 PrecleanCLDClosure(MetadataVisitingOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4125 void do_cld(ClassLoaderData* cld) { 4126 if (cld->has_accumulated_modified_oops()) { 4127 cld->clear_accumulated_modified_oops(); 4128 4129 _cm_closure->do_cld(cld); 4130 } 4131 } 4132 }; 4133 4134 // The freelist lock is needed to prevent asserts, is it really needed? 4135 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4136 // Needed to walk CLDG 4137 MutexLocker ml(ClassLoaderDataGraph_lock); 4138 4139 cl->set_freelistLock(freelistLock); 4140 4141 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4142 4143 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4144 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4145 PrecleanCLDClosure preclean_closure(cl); 4146 ClassLoaderDataGraph::cld_do(&preclean_closure); 4147 4148 verify_work_stacks_empty(); 4149 verify_overflow_empty(); 4150 } 4151 4152 void CMSCollector::checkpointRootsFinal() { 4153 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4154 check_correct_thread_executing(); 4155 // world is stopped at this checkpoint 4156 assert(SafepointSynchronize::is_at_safepoint(), 4157 "world should be stopped"); 4158 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4159 4160 verify_work_stacks_empty(); 4161 verify_overflow_empty(); 4162 4163 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4164 _young_gen->used() / K, _young_gen->capacity() / K); 4165 { 4166 if (CMSScavengeBeforeRemark) { 4167 CMSHeap* heap = CMSHeap::heap(); 4168 // Temporarily set flag to false, GCH->do_collection will 4169 // expect it to be false and set to true 4170 FlagSetting fl(heap->_is_gc_active, false); 4171 4172 heap->do_collection(true, // full (i.e. force, see below) 4173 false, // !clear_all_soft_refs 4174 0, // size 4175 false, // is_tlab 4176 GenCollectedHeap::YoungGen // type 4177 ); 4178 } 4179 FreelistLocker x(this); 4180 MutexLockerEx y(bitMapLock(), 4181 Mutex::_no_safepoint_check_flag); 4182 checkpointRootsFinalWork(); 4183 } 4184 verify_work_stacks_empty(); 4185 verify_overflow_empty(); 4186 } 4187 4188 void CMSCollector::checkpointRootsFinalWork() { 4189 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4190 4191 assert(haveFreelistLocks(), "must have free list locks"); 4192 assert_lock_strong(bitMapLock()); 4193 4194 ResourceMark rm; 4195 HandleMark hm; 4196 4197 CMSHeap* heap = CMSHeap::heap(); 4198 4199 assert(haveFreelistLocks(), "must have free list locks"); 4200 assert_lock_strong(bitMapLock()); 4201 4202 // We might assume that we need not fill TLAB's when 4203 // CMSScavengeBeforeRemark is set, because we may have just done 4204 // a scavenge which would have filled all TLAB's -- and besides 4205 // Eden would be empty. This however may not always be the case -- 4206 // for instance although we asked for a scavenge, it may not have 4207 // happened because of a JNI critical section. We probably need 4208 // a policy for deciding whether we can in that case wait until 4209 // the critical section releases and then do the remark following 4210 // the scavenge, and skip it here. In the absence of that policy, 4211 // or of an indication of whether the scavenge did indeed occur, 4212 // we cannot rely on TLAB's having been filled and must do 4213 // so here just in case a scavenge did not happen. 4214 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4215 // Update the saved marks which may affect the root scans. 4216 heap->save_marks(); 4217 4218 print_eden_and_survivor_chunk_arrays(); 4219 4220 { 4221 #if COMPILER2_OR_JVMCI 4222 DerivedPointerTableDeactivate dpt_deact; 4223 #endif 4224 4225 // Note on the role of the mod union table: 4226 // Since the marker in "markFromRoots" marks concurrently with 4227 // mutators, it is possible for some reachable objects not to have been 4228 // scanned. For instance, an only reference to an object A was 4229 // placed in object B after the marker scanned B. Unless B is rescanned, 4230 // A would be collected. Such updates to references in marked objects 4231 // are detected via the mod union table which is the set of all cards 4232 // dirtied since the first checkpoint in this GC cycle and prior to 4233 // the most recent young generation GC, minus those cleaned up by the 4234 // concurrent precleaning. 4235 if (CMSParallelRemarkEnabled) { 4236 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4237 do_remark_parallel(); 4238 } else { 4239 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4240 do_remark_non_parallel(); 4241 } 4242 } 4243 verify_work_stacks_empty(); 4244 verify_overflow_empty(); 4245 4246 { 4247 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4248 refProcessingWork(); 4249 } 4250 verify_work_stacks_empty(); 4251 verify_overflow_empty(); 4252 4253 if (should_unload_classes()) { 4254 heap->prune_scavengable_nmethods(); 4255 } 4256 JvmtiExport::gc_epilogue(); 4257 4258 // If we encountered any (marking stack / work queue) overflow 4259 // events during the current CMS cycle, take appropriate 4260 // remedial measures, where possible, so as to try and avoid 4261 // recurrence of that condition. 4262 assert(_markStack.isEmpty(), "No grey objects"); 4263 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4264 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4265 if (ser_ovflw > 0) { 4266 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4267 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4268 _markStack.expand(); 4269 _ser_pmc_remark_ovflw = 0; 4270 _ser_pmc_preclean_ovflw = 0; 4271 _ser_kac_preclean_ovflw = 0; 4272 _ser_kac_ovflw = 0; 4273 } 4274 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4275 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4276 _par_pmc_remark_ovflw, _par_kac_ovflw); 4277 _par_pmc_remark_ovflw = 0; 4278 _par_kac_ovflw = 0; 4279 } 4280 if (_markStack._hit_limit > 0) { 4281 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4282 _markStack._hit_limit); 4283 } 4284 if (_markStack._failed_double > 0) { 4285 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4286 _markStack._failed_double, _markStack.capacity()); 4287 } 4288 _markStack._hit_limit = 0; 4289 _markStack._failed_double = 0; 4290 4291 if ((VerifyAfterGC || VerifyDuringGC) && 4292 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4293 verify_after_remark(); 4294 } 4295 4296 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4297 4298 // Change under the freelistLocks. 4299 _collectorState = Sweeping; 4300 // Call isAllClear() under bitMapLock 4301 assert(_modUnionTable.isAllClear(), 4302 "Should be clear by end of the final marking"); 4303 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4304 "Should be clear by end of the final marking"); 4305 } 4306 4307 void CMSParInitialMarkTask::work(uint worker_id) { 4308 elapsedTimer _timer; 4309 ResourceMark rm; 4310 HandleMark hm; 4311 4312 // ---------- scan from roots -------------- 4313 _timer.start(); 4314 CMSHeap* heap = CMSHeap::heap(); 4315 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4316 4317 // ---------- young gen roots -------------- 4318 { 4319 work_on_young_gen_roots(&par_mri_cl); 4320 _timer.stop(); 4321 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4322 } 4323 4324 // ---------- remaining roots -------------- 4325 _timer.reset(); 4326 _timer.start(); 4327 4328 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4329 4330 heap->cms_process_roots(_strong_roots_scope, 4331 false, // yg was scanned above 4332 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4333 _collector->should_unload_classes(), 4334 &par_mri_cl, 4335 &cld_closure); 4336 4337 assert(_collector->should_unload_classes() 4338 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4339 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4340 _timer.stop(); 4341 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4342 } 4343 4344 // Parallel remark task 4345 class CMSParRemarkTask: public CMSParMarkTask { 4346 CompactibleFreeListSpace* _cms_space; 4347 4348 // The per-thread work queues, available here for stealing. 4349 OopTaskQueueSet* _task_queues; 4350 TaskTerminator _term; 4351 StrongRootsScope* _strong_roots_scope; 4352 4353 public: 4354 // A value of 0 passed to n_workers will cause the number of 4355 // workers to be taken from the active workers in the work gang. 4356 CMSParRemarkTask(CMSCollector* collector, 4357 CompactibleFreeListSpace* cms_space, 4358 uint n_workers, WorkGang* workers, 4359 OopTaskQueueSet* task_queues, 4360 StrongRootsScope* strong_roots_scope): 4361 CMSParMarkTask("Rescan roots and grey objects in parallel", 4362 collector, n_workers), 4363 _cms_space(cms_space), 4364 _task_queues(task_queues), 4365 _term(n_workers, task_queues), 4366 _strong_roots_scope(strong_roots_scope) { } 4367 4368 OopTaskQueueSet* task_queues() { return _task_queues; } 4369 4370 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4371 4372 ParallelTaskTerminator* terminator() { return _term.terminator(); } 4373 uint n_workers() { return _n_workers; } 4374 4375 void work(uint worker_id); 4376 4377 private: 4378 // ... of dirty cards in old space 4379 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4380 ParMarkRefsIntoAndScanClosure* cl); 4381 4382 // ... work stealing for the above 4383 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4384 }; 4385 4386 class RemarkCLDClosure : public CLDClosure { 4387 CLDToOopClosure _cm_closure; 4388 public: 4389 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4390 void do_cld(ClassLoaderData* cld) { 4391 // Check if we have modified any oops in the CLD during the concurrent marking. 4392 if (cld->has_accumulated_modified_oops()) { 4393 cld->clear_accumulated_modified_oops(); 4394 4395 // We could have transfered the current modified marks to the accumulated marks, 4396 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4397 } else if (cld->has_modified_oops()) { 4398 // Don't clear anything, this info is needed by the next young collection. 4399 } else { 4400 // No modified oops in the ClassLoaderData. 4401 return; 4402 } 4403 4404 // The klass has modified fields, need to scan the klass. 4405 _cm_closure.do_cld(cld); 4406 } 4407 }; 4408 4409 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4410 ParNewGeneration* young_gen = _collector->_young_gen; 4411 ContiguousSpace* eden_space = young_gen->eden(); 4412 ContiguousSpace* from_space = young_gen->from(); 4413 ContiguousSpace* to_space = young_gen->to(); 4414 4415 HeapWord** eca = _collector->_eden_chunk_array; 4416 size_t ect = _collector->_eden_chunk_index; 4417 HeapWord** sca = _collector->_survivor_chunk_array; 4418 size_t sct = _collector->_survivor_chunk_index; 4419 4420 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4421 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4422 4423 do_young_space_rescan(cl, to_space, NULL, 0); 4424 do_young_space_rescan(cl, from_space, sca, sct); 4425 do_young_space_rescan(cl, eden_space, eca, ect); 4426 } 4427 4428 // work_queue(i) is passed to the closure 4429 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4430 // also is passed to do_dirty_card_rescan_tasks() and to 4431 // do_work_steal() to select the i-th task_queue. 4432 4433 void CMSParRemarkTask::work(uint worker_id) { 4434 elapsedTimer _timer; 4435 ResourceMark rm; 4436 HandleMark hm; 4437 4438 // ---------- rescan from roots -------------- 4439 _timer.start(); 4440 CMSHeap* heap = CMSHeap::heap(); 4441 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4442 _collector->_span, _collector->ref_processor(), 4443 &(_collector->_markBitMap), 4444 work_queue(worker_id)); 4445 4446 // Rescan young gen roots first since these are likely 4447 // coarsely partitioned and may, on that account, constitute 4448 // the critical path; thus, it's best to start off that 4449 // work first. 4450 // ---------- young gen roots -------------- 4451 { 4452 work_on_young_gen_roots(&par_mrias_cl); 4453 _timer.stop(); 4454 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4455 } 4456 4457 // ---------- remaining roots -------------- 4458 _timer.reset(); 4459 _timer.start(); 4460 heap->cms_process_roots(_strong_roots_scope, 4461 false, // yg was scanned above 4462 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4463 _collector->should_unload_classes(), 4464 &par_mrias_cl, 4465 NULL); // The dirty klasses will be handled below 4466 4467 assert(_collector->should_unload_classes() 4468 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4469 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4470 _timer.stop(); 4471 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4472 4473 // ---------- unhandled CLD scanning ---------- 4474 if (worker_id == 0) { // Single threaded at the moment. 4475 _timer.reset(); 4476 _timer.start(); 4477 4478 // Scan all new class loader data objects and new dependencies that were 4479 // introduced during concurrent marking. 4480 ResourceMark rm; 4481 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4482 for (int i = 0; i < array->length(); i++) { 4483 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4484 } 4485 4486 // We don't need to keep track of new CLDs anymore. 4487 ClassLoaderDataGraph::remember_new_clds(false); 4488 4489 _timer.stop(); 4490 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4491 } 4492 4493 // We might have added oops to ClassLoaderData::_handles during the 4494 // concurrent marking phase. These oops do not always point to newly allocated objects 4495 // that are guaranteed to be kept alive. Hence, 4496 // we do have to revisit the _handles block during the remark phase. 4497 4498 // ---------- dirty CLD scanning ---------- 4499 if (worker_id == 0) { // Single threaded at the moment. 4500 _timer.reset(); 4501 _timer.start(); 4502 4503 // Scan all classes that was dirtied during the concurrent marking phase. 4504 RemarkCLDClosure remark_closure(&par_mrias_cl); 4505 ClassLoaderDataGraph::cld_do(&remark_closure); 4506 4507 _timer.stop(); 4508 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4509 } 4510 4511 4512 // ---------- rescan dirty cards ------------ 4513 _timer.reset(); 4514 _timer.start(); 4515 4516 // Do the rescan tasks for each of the two spaces 4517 // (cms_space) in turn. 4518 // "worker_id" is passed to select the task_queue for "worker_id" 4519 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4520 _timer.stop(); 4521 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4522 4523 // ---------- steal work from other threads ... 4524 // ---------- ... and drain overflow list. 4525 _timer.reset(); 4526 _timer.start(); 4527 do_work_steal(worker_id, &par_mrias_cl); 4528 _timer.stop(); 4529 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4530 } 4531 4532 void 4533 CMSParMarkTask::do_young_space_rescan( 4534 OopsInGenClosure* cl, ContiguousSpace* space, 4535 HeapWord** chunk_array, size_t chunk_top) { 4536 // Until all tasks completed: 4537 // . claim an unclaimed task 4538 // . compute region boundaries corresponding to task claimed 4539 // using chunk_array 4540 // . par_oop_iterate(cl) over that region 4541 4542 ResourceMark rm; 4543 HandleMark hm; 4544 4545 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4546 4547 uint nth_task = 0; 4548 uint n_tasks = pst->n_tasks(); 4549 4550 if (n_tasks > 0) { 4551 assert(pst->valid(), "Uninitialized use?"); 4552 HeapWord *start, *end; 4553 while (pst->try_claim_task(/* reference */ nth_task)) { 4554 // We claimed task # nth_task; compute its boundaries. 4555 if (chunk_top == 0) { // no samples were taken 4556 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4557 start = space->bottom(); 4558 end = space->top(); 4559 } else if (nth_task == 0) { 4560 start = space->bottom(); 4561 end = chunk_array[nth_task]; 4562 } else if (nth_task < (uint)chunk_top) { 4563 assert(nth_task >= 1, "Control point invariant"); 4564 start = chunk_array[nth_task - 1]; 4565 end = chunk_array[nth_task]; 4566 } else { 4567 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4568 start = chunk_array[chunk_top - 1]; 4569 end = space->top(); 4570 } 4571 MemRegion mr(start, end); 4572 // Verify that mr is in space 4573 assert(mr.is_empty() || space->used_region().contains(mr), 4574 "Should be in space"); 4575 // Verify that "start" is an object boundary 4576 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4577 "Should be an oop"); 4578 space->par_oop_iterate(mr, cl); 4579 } 4580 pst->all_tasks_completed(); 4581 } 4582 } 4583 4584 void 4585 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4586 CompactibleFreeListSpace* sp, int i, 4587 ParMarkRefsIntoAndScanClosure* cl) { 4588 // Until all tasks completed: 4589 // . claim an unclaimed task 4590 // . compute region boundaries corresponding to task claimed 4591 // . transfer dirty bits ct->mut for that region 4592 // . apply rescanclosure to dirty mut bits for that region 4593 4594 ResourceMark rm; 4595 HandleMark hm; 4596 4597 OopTaskQueue* work_q = work_queue(i); 4598 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4599 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4600 // CAUTION: This closure has state that persists across calls to 4601 // the work method dirty_range_iterate_clear() in that it has 4602 // embedded in it a (subtype of) UpwardsObjectClosure. The 4603 // use of that state in the embedded UpwardsObjectClosure instance 4604 // assumes that the cards are always iterated (even if in parallel 4605 // by several threads) in monotonically increasing order per each 4606 // thread. This is true of the implementation below which picks 4607 // card ranges (chunks) in monotonically increasing order globally 4608 // and, a-fortiori, in monotonically increasing order per thread 4609 // (the latter order being a subsequence of the former). 4610 // If the work code below is ever reorganized into a more chaotic 4611 // work-partitioning form than the current "sequential tasks" 4612 // paradigm, the use of that persistent state will have to be 4613 // revisited and modified appropriately. See also related 4614 // bug 4756801 work on which should examine this code to make 4615 // sure that the changes there do not run counter to the 4616 // assumptions made here and necessary for correctness and 4617 // efficiency. Note also that this code might yield inefficient 4618 // behavior in the case of very large objects that span one or 4619 // more work chunks. Such objects would potentially be scanned 4620 // several times redundantly. Work on 4756801 should try and 4621 // address that performance anomaly if at all possible. XXX 4622 MemRegion full_span = _collector->_span; 4623 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4624 MarkFromDirtyCardsClosure 4625 greyRescanClosure(_collector, full_span, // entire span of interest 4626 sp, bm, work_q, cl); 4627 4628 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4629 assert(pst->valid(), "Uninitialized use?"); 4630 uint nth_task = 0; 4631 const int alignment = CardTable::card_size * BitsPerWord; 4632 MemRegion span = sp->used_region(); 4633 HeapWord* start_addr = span.start(); 4634 HeapWord* end_addr = align_up(span.end(), alignment); 4635 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4636 assert(is_aligned(start_addr, alignment), "Check alignment"); 4637 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4638 4639 while (pst->try_claim_task(/* reference */ nth_task)) { 4640 // Having claimed the nth_task, compute corresponding mem-region, 4641 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4642 // The alignment restriction ensures that we do not need any 4643 // synchronization with other gang-workers while setting or 4644 // clearing bits in thus chunk of the MUT. 4645 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4646 start_addr + (nth_task+1)*chunk_size); 4647 // The last chunk's end might be way beyond end of the 4648 // used region. In that case pull back appropriately. 4649 if (this_span.end() > end_addr) { 4650 this_span.set_end(end_addr); 4651 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4652 } 4653 // Iterate over the dirty cards covering this chunk, marking them 4654 // precleaned, and setting the corresponding bits in the mod union 4655 // table. Since we have been careful to partition at Card and MUT-word 4656 // boundaries no synchronization is needed between parallel threads. 4657 _collector->_ct->dirty_card_iterate(this_span, 4658 &modUnionClosure); 4659 4660 // Having transferred these marks into the modUnionTable, 4661 // rescan the marked objects on the dirty cards in the modUnionTable. 4662 // Even if this is at a synchronous collection, the initial marking 4663 // may have been done during an asynchronous collection so there 4664 // may be dirty bits in the mod-union table. 4665 _collector->_modUnionTable.dirty_range_iterate_clear( 4666 this_span, &greyRescanClosure); 4667 _collector->_modUnionTable.verifyNoOneBitsInRange( 4668 this_span.start(), 4669 this_span.end()); 4670 } 4671 pst->all_tasks_completed(); // declare that i am done 4672 } 4673 4674 // . see if we can share work_queues with ParNew? XXX 4675 void 4676 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4677 OopTaskQueue* work_q = work_queue(i); 4678 NOT_PRODUCT(int num_steals = 0;) 4679 oop obj_to_scan; 4680 CMSBitMap* bm = &(_collector->_markBitMap); 4681 4682 while (true) { 4683 // Completely finish any left over work from (an) earlier round(s) 4684 cl->trim_queue(0); 4685 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4686 (size_t)ParGCDesiredObjsFromOverflowList); 4687 // Now check if there's any work in the overflow list 4688 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4689 // only affects the number of attempts made to get work from the 4690 // overflow list and does not affect the number of workers. Just 4691 // pass ParallelGCThreads so this behavior is unchanged. 4692 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4693 work_q, 4694 ParallelGCThreads)) { 4695 // found something in global overflow list; 4696 // not yet ready to go stealing work from others. 4697 // We'd like to assert(work_q->size() != 0, ...) 4698 // because we just took work from the overflow list, 4699 // but of course we can't since all of that could have 4700 // been already stolen from us. 4701 // "He giveth and He taketh away." 4702 continue; 4703 } 4704 // Verify that we have no work before we resort to stealing 4705 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4706 // Try to steal from other queues that have work 4707 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4708 NOT_PRODUCT(num_steals++;) 4709 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4710 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4711 // Do scanning work 4712 obj_to_scan->oop_iterate(cl); 4713 // Loop around, finish this work, and try to steal some more 4714 } else if (terminator()->offer_termination()) { 4715 break; // nirvana from the infinite cycle 4716 } 4717 } 4718 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4719 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4720 "Else our work is not yet done"); 4721 } 4722 4723 // Record object boundaries in _eden_chunk_array by sampling the eden 4724 // top in the slow-path eden object allocation code path and record 4725 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4726 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4727 // sampling in sample_eden() that activates during the part of the 4728 // preclean phase. 4729 void CMSCollector::sample_eden_chunk() { 4730 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4731 if (_eden_chunk_lock->try_lock()) { 4732 // Record a sample. This is the critical section. The contents 4733 // of the _eden_chunk_array have to be non-decreasing in the 4734 // address order. 4735 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4736 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4737 "Unexpected state of Eden"); 4738 if (_eden_chunk_index == 0 || 4739 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4740 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4741 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4742 _eden_chunk_index++; // commit sample 4743 } 4744 _eden_chunk_lock->unlock(); 4745 } 4746 } 4747 } 4748 4749 // Return a thread-local PLAB recording array, as appropriate. 4750 void* CMSCollector::get_data_recorder(int thr_num) { 4751 if (_survivor_plab_array != NULL && 4752 (CMSPLABRecordAlways || 4753 (_collectorState > Marking && _collectorState < FinalMarking))) { 4754 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4755 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4756 ca->reset(); // clear it so that fresh data is recorded 4757 return (void*) ca; 4758 } else { 4759 return NULL; 4760 } 4761 } 4762 4763 // Reset all the thread-local PLAB recording arrays 4764 void CMSCollector::reset_survivor_plab_arrays() { 4765 for (uint i = 0; i < ParallelGCThreads; i++) { 4766 _survivor_plab_array[i].reset(); 4767 } 4768 } 4769 4770 // Merge the per-thread plab arrays into the global survivor chunk 4771 // array which will provide the partitioning of the survivor space 4772 // for CMS initial scan and rescan. 4773 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4774 int no_of_gc_threads) { 4775 assert(_survivor_plab_array != NULL, "Error"); 4776 assert(_survivor_chunk_array != NULL, "Error"); 4777 assert(_collectorState == FinalMarking || 4778 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4779 for (int j = 0; j < no_of_gc_threads; j++) { 4780 _cursor[j] = 0; 4781 } 4782 HeapWord* top = surv->top(); 4783 size_t i; 4784 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4785 HeapWord* min_val = top; // Higher than any PLAB address 4786 uint min_tid = 0; // position of min_val this round 4787 for (int j = 0; j < no_of_gc_threads; j++) { 4788 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4789 if (_cursor[j] == cur_sca->end()) { 4790 continue; 4791 } 4792 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4793 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4794 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4795 if (cur_val < min_val) { 4796 min_tid = j; 4797 min_val = cur_val; 4798 } else { 4799 assert(cur_val < top, "All recorded addresses should be less"); 4800 } 4801 } 4802 // At this point min_val and min_tid are respectively 4803 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4804 // and the thread (j) that witnesses that address. 4805 // We record this address in the _survivor_chunk_array[i] 4806 // and increment _cursor[min_tid] prior to the next round i. 4807 if (min_val == top) { 4808 break; 4809 } 4810 _survivor_chunk_array[i] = min_val; 4811 _cursor[min_tid]++; 4812 } 4813 // We are all done; record the size of the _survivor_chunk_array 4814 _survivor_chunk_index = i; // exclusive: [0, i) 4815 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4816 // Verify that we used up all the recorded entries 4817 #ifdef ASSERT 4818 size_t total = 0; 4819 for (int j = 0; j < no_of_gc_threads; j++) { 4820 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4821 total += _cursor[j]; 4822 } 4823 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4824 // Check that the merged array is in sorted order 4825 if (total > 0) { 4826 for (size_t i = 0; i < total - 1; i++) { 4827 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4828 i, p2i(_survivor_chunk_array[i])); 4829 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4830 "Not sorted"); 4831 } 4832 } 4833 #endif // ASSERT 4834 } 4835 4836 // Set up the space's par_seq_tasks structure for work claiming 4837 // for parallel initial scan and rescan of young gen. 4838 // See ParRescanTask where this is currently used. 4839 void 4840 CMSCollector:: 4841 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4842 assert(n_threads > 0, "Unexpected n_threads argument"); 4843 4844 // Eden space 4845 if (!_young_gen->eden()->is_empty()) { 4846 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4847 assert(!pst->valid(), "Clobbering existing data?"); 4848 // Each valid entry in [0, _eden_chunk_index) represents a task. 4849 size_t n_tasks = _eden_chunk_index + 1; 4850 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4851 // Sets the condition for completion of the subtask (how many threads 4852 // need to finish in order to be done). 4853 pst->set_n_threads(n_threads); 4854 pst->set_n_tasks((int)n_tasks); 4855 } 4856 4857 // Merge the survivor plab arrays into _survivor_chunk_array 4858 if (_survivor_plab_array != NULL) { 4859 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4860 } else { 4861 assert(_survivor_chunk_index == 0, "Error"); 4862 } 4863 4864 // To space 4865 { 4866 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4867 assert(!pst->valid(), "Clobbering existing data?"); 4868 // Sets the condition for completion of the subtask (how many threads 4869 // need to finish in order to be done). 4870 pst->set_n_threads(n_threads); 4871 pst->set_n_tasks(1); 4872 assert(pst->valid(), "Error"); 4873 } 4874 4875 // From space 4876 { 4877 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4878 assert(!pst->valid(), "Clobbering existing data?"); 4879 size_t n_tasks = _survivor_chunk_index + 1; 4880 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4881 // Sets the condition for completion of the subtask (how many threads 4882 // need to finish in order to be done). 4883 pst->set_n_threads(n_threads); 4884 pst->set_n_tasks((int)n_tasks); 4885 assert(pst->valid(), "Error"); 4886 } 4887 } 4888 4889 // Parallel version of remark 4890 void CMSCollector::do_remark_parallel() { 4891 CMSHeap* heap = CMSHeap::heap(); 4892 WorkGang* workers = heap->workers(); 4893 assert(workers != NULL, "Need parallel worker threads."); 4894 // Choose to use the number of GC workers most recently set 4895 // into "active_workers". 4896 uint n_workers = workers->active_workers(); 4897 4898 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4899 4900 StrongRootsScope srs(n_workers); 4901 4902 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4903 4904 // We won't be iterating over the cards in the card table updating 4905 // the younger_gen cards, so we shouldn't call the following else 4906 // the verification code as well as subsequent younger_refs_iterate 4907 // code would get confused. XXX 4908 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4909 4910 // The young gen rescan work will not be done as part of 4911 // process_roots (which currently doesn't know how to 4912 // parallelize such a scan), but rather will be broken up into 4913 // a set of parallel tasks (via the sampling that the [abortable] 4914 // preclean phase did of eden, plus the [two] tasks of 4915 // scanning the [two] survivor spaces. Further fine-grain 4916 // parallelization of the scanning of the survivor spaces 4917 // themselves, and of precleaning of the young gen itself 4918 // is deferred to the future. 4919 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4920 4921 // The dirty card rescan work is broken up into a "sequence" 4922 // of parallel tasks (per constituent space) that are dynamically 4923 // claimed by the parallel threads. 4924 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4925 4926 // It turns out that even when we're using 1 thread, doing the work in a 4927 // separate thread causes wide variance in run times. We can't help this 4928 // in the multi-threaded case, but we special-case n=1 here to get 4929 // repeatable measurements of the 1-thread overhead of the parallel code. 4930 if (n_workers > 1) { 4931 // Make refs discovery MT-safe, if it isn't already: it may not 4932 // necessarily be so, since it's possible that we are doing 4933 // ST marking. 4934 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4935 workers->run_task(&tsk); 4936 } else { 4937 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4938 tsk.work(0); 4939 } 4940 4941 // restore, single-threaded for now, any preserved marks 4942 // as a result of work_q overflow 4943 restore_preserved_marks_if_any(); 4944 } 4945 4946 // Non-parallel version of remark 4947 void CMSCollector::do_remark_non_parallel() { 4948 ResourceMark rm; 4949 HandleMark hm; 4950 CMSHeap* heap = CMSHeap::heap(); 4951 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4952 4953 MarkRefsIntoAndScanClosure 4954 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4955 &_markStack, this, 4956 false /* should_yield */, false /* not precleaning */); 4957 MarkFromDirtyCardsClosure 4958 markFromDirtyCardsClosure(this, _span, 4959 NULL, // space is set further below 4960 &_markBitMap, &_markStack, &mrias_cl); 4961 { 4962 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4963 // Iterate over the dirty cards, setting the corresponding bits in the 4964 // mod union table. 4965 { 4966 ModUnionClosure modUnionClosure(&_modUnionTable); 4967 _ct->dirty_card_iterate(_cmsGen->used_region(), 4968 &modUnionClosure); 4969 } 4970 // Having transferred these marks into the modUnionTable, we just need 4971 // to rescan the marked objects on the dirty cards in the modUnionTable. 4972 // The initial marking may have been done during an asynchronous 4973 // collection so there may be dirty bits in the mod-union table. 4974 const int alignment = CardTable::card_size * BitsPerWord; 4975 { 4976 // ... First handle dirty cards in CMS gen 4977 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4978 MemRegion ur = _cmsGen->used_region(); 4979 HeapWord* lb = ur.start(); 4980 HeapWord* ub = align_up(ur.end(), alignment); 4981 MemRegion cms_span(lb, ub); 4982 _modUnionTable.dirty_range_iterate_clear(cms_span, 4983 &markFromDirtyCardsClosure); 4984 verify_work_stacks_empty(); 4985 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4986 } 4987 } 4988 if (VerifyDuringGC && 4989 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4990 HandleMark hm; // Discard invalid handles created during verification 4991 Universe::verify(); 4992 } 4993 { 4994 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4995 4996 verify_work_stacks_empty(); 4997 4998 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4999 StrongRootsScope srs(1); 5000 5001 heap->cms_process_roots(&srs, 5002 true, // young gen as roots 5003 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5004 should_unload_classes(), 5005 &mrias_cl, 5006 NULL); // The dirty klasses will be handled below 5007 5008 assert(should_unload_classes() 5009 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5010 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5011 } 5012 5013 { 5014 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 5015 5016 verify_work_stacks_empty(); 5017 5018 // Scan all class loader data objects that might have been introduced 5019 // during concurrent marking. 5020 ResourceMark rm; 5021 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5022 for (int i = 0; i < array->length(); i++) { 5023 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 5024 } 5025 5026 // We don't need to keep track of new CLDs anymore. 5027 ClassLoaderDataGraph::remember_new_clds(false); 5028 5029 verify_work_stacks_empty(); 5030 } 5031 5032 // We might have added oops to ClassLoaderData::_handles during the 5033 // concurrent marking phase. These oops do not point to newly allocated objects 5034 // that are guaranteed to be kept alive. Hence, 5035 // we do have to revisit the _handles block during the remark phase. 5036 { 5037 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 5038 5039 verify_work_stacks_empty(); 5040 5041 RemarkCLDClosure remark_closure(&mrias_cl); 5042 ClassLoaderDataGraph::cld_do(&remark_closure); 5043 5044 verify_work_stacks_empty(); 5045 } 5046 5047 verify_work_stacks_empty(); 5048 // Restore evacuated mark words, if any, used for overflow list links 5049 restore_preserved_marks_if_any(); 5050 5051 verify_overflow_empty(); 5052 } 5053 5054 //////////////////////////////////////////////////////// 5055 // Parallel Reference Processing Task Proxy Class 5056 //////////////////////////////////////////////////////// 5057 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5058 OopTaskQueueSet* _queues; 5059 TaskTerminator _terminator; 5060 public: 5061 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5062 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5063 ParallelTaskTerminator* terminator() { return _terminator.terminator(); } 5064 OopTaskQueueSet* queues() { return _queues; } 5065 }; 5066 5067 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5068 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5069 CMSCollector* _collector; 5070 CMSBitMap* _mark_bit_map; 5071 const MemRegion _span; 5072 ProcessTask& _task; 5073 5074 public: 5075 CMSRefProcTaskProxy(ProcessTask& task, 5076 CMSCollector* collector, 5077 const MemRegion& span, 5078 CMSBitMap* mark_bit_map, 5079 AbstractWorkGang* workers, 5080 OopTaskQueueSet* task_queues): 5081 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5082 task_queues, 5083 workers->active_workers()), 5084 _collector(collector), 5085 _mark_bit_map(mark_bit_map), 5086 _span(span), 5087 _task(task) 5088 { 5089 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5090 "Inconsistency in _span"); 5091 } 5092 5093 OopTaskQueueSet* task_queues() { return queues(); } 5094 5095 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5096 5097 void do_work_steal(int i, 5098 CMSParDrainMarkingStackClosure* drain, 5099 CMSParKeepAliveClosure* keep_alive); 5100 5101 virtual void work(uint worker_id); 5102 }; 5103 5104 void CMSRefProcTaskProxy::work(uint worker_id) { 5105 ResourceMark rm; 5106 HandleMark hm; 5107 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5108 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5109 _mark_bit_map, 5110 work_queue(worker_id)); 5111 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5112 _mark_bit_map, 5113 work_queue(worker_id)); 5114 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5115 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5116 if (_task.marks_oops_alive()) { 5117 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5118 } 5119 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5120 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5121 } 5122 5123 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5124 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5125 _span(span), 5126 _work_queue(work_queue), 5127 _bit_map(bit_map), 5128 _mark_and_push(collector, span, bit_map, work_queue), 5129 _low_water_mark(MIN2((work_queue->max_elems()/4), 5130 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5131 { } 5132 5133 // . see if we can share work_queues with ParNew? XXX 5134 void CMSRefProcTaskProxy::do_work_steal(int i, 5135 CMSParDrainMarkingStackClosure* drain, 5136 CMSParKeepAliveClosure* keep_alive) { 5137 OopTaskQueue* work_q = work_queue(i); 5138 NOT_PRODUCT(int num_steals = 0;) 5139 oop obj_to_scan; 5140 5141 while (true) { 5142 // Completely finish any left over work from (an) earlier round(s) 5143 drain->trim_queue(0); 5144 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5145 (size_t)ParGCDesiredObjsFromOverflowList); 5146 // Now check if there's any work in the overflow list 5147 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5148 // only affects the number of attempts made to get work from the 5149 // overflow list and does not affect the number of workers. Just 5150 // pass ParallelGCThreads so this behavior is unchanged. 5151 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5152 work_q, 5153 ParallelGCThreads)) { 5154 // Found something in global overflow list; 5155 // not yet ready to go stealing work from others. 5156 // We'd like to assert(work_q->size() != 0, ...) 5157 // because we just took work from the overflow list, 5158 // but of course we can't, since all of that might have 5159 // been already stolen from us. 5160 continue; 5161 } 5162 // Verify that we have no work before we resort to stealing 5163 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5164 // Try to steal from other queues that have work 5165 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5166 NOT_PRODUCT(num_steals++;) 5167 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5168 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5169 // Do scanning work 5170 obj_to_scan->oop_iterate(keep_alive); 5171 // Loop around, finish this work, and try to steal some more 5172 } else if (terminator()->offer_termination()) { 5173 break; // nirvana from the infinite cycle 5174 } 5175 } 5176 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5177 } 5178 5179 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5180 CMSHeap* heap = CMSHeap::heap(); 5181 WorkGang* workers = heap->workers(); 5182 assert(workers != NULL, "Need parallel worker threads."); 5183 assert(workers->active_workers() == ergo_workers, 5184 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5185 ergo_workers, workers->active_workers()); 5186 CMSRefProcTaskProxy rp_task(task, &_collector, 5187 _collector.ref_processor_span(), 5188 _collector.markBitMap(), 5189 workers, _collector.task_queues()); 5190 workers->run_task(&rp_task, workers->active_workers()); 5191 } 5192 5193 void CMSCollector::refProcessingWork() { 5194 ResourceMark rm; 5195 HandleMark hm; 5196 5197 ReferenceProcessor* rp = ref_processor(); 5198 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5199 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5200 // Process weak references. 5201 rp->setup_policy(false); 5202 verify_work_stacks_empty(); 5203 5204 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5205 { 5206 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5207 5208 // Setup keep_alive and complete closures. 5209 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5210 &_markStack, false /* !preclean */); 5211 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5212 _span, &_markBitMap, &_markStack, 5213 &cmsKeepAliveClosure, false /* !preclean */); 5214 5215 ReferenceProcessorStats stats; 5216 if (rp->processing_is_mt()) { 5217 // Set the degree of MT here. If the discovery is done MT, there 5218 // may have been a different number of threads doing the discovery 5219 // and a different number of discovered lists may have Ref objects. 5220 // That is OK as long as the Reference lists are balanced (see 5221 // balance_all_queues() and balance_queues()). 5222 CMSHeap* heap = CMSHeap::heap(); 5223 uint active_workers = ParallelGCThreads; 5224 WorkGang* workers = heap->workers(); 5225 if (workers != NULL) { 5226 active_workers = workers->active_workers(); 5227 // The expectation is that active_workers will have already 5228 // been set to a reasonable value. If it has not been set, 5229 // investigate. 5230 assert(active_workers > 0, "Should have been set during scavenge"); 5231 } 5232 rp->set_active_mt_degree(active_workers); 5233 CMSRefProcTaskExecutor task_executor(*this); 5234 stats = rp->process_discovered_references(&_is_alive_closure, 5235 &cmsKeepAliveClosure, 5236 &cmsDrainMarkingStackClosure, 5237 &task_executor, 5238 &pt); 5239 } else { 5240 stats = rp->process_discovered_references(&_is_alive_closure, 5241 &cmsKeepAliveClosure, 5242 &cmsDrainMarkingStackClosure, 5243 NULL, 5244 &pt); 5245 } 5246 _gc_tracer_cm->report_gc_reference_stats(stats); 5247 pt.print_all_references(); 5248 } 5249 5250 // This is the point where the entire marking should have completed. 5251 verify_work_stacks_empty(); 5252 5253 { 5254 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5255 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5256 } 5257 5258 if (should_unload_classes()) { 5259 { 5260 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5261 5262 // Unload classes and purge the SystemDictionary. 5263 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5264 5265 // Unload nmethods. 5266 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5267 5268 // Prune dead klasses from subklass/sibling/implementor lists. 5269 Klass::clean_weak_klass_links(purged_class); 5270 5271 #if INCLUDE_JVMCI 5272 // Clean JVMCI metadata handles. 5273 JVMCI::do_unloading(&_is_alive_closure, purged_class); 5274 #endif 5275 } 5276 } 5277 5278 // Restore any preserved marks as a result of mark stack or 5279 // work queue overflow 5280 restore_preserved_marks_if_any(); // done single-threaded for now 5281 5282 rp->set_enqueuing_is_done(true); 5283 rp->verify_no_references_recorded(); 5284 } 5285 5286 #ifndef PRODUCT 5287 void CMSCollector::check_correct_thread_executing() { 5288 Thread* t = Thread::current(); 5289 // Only the VM thread or the CMS thread should be here. 5290 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5291 "Unexpected thread type"); 5292 // If this is the vm thread, the foreground process 5293 // should not be waiting. Note that _foregroundGCIsActive is 5294 // true while the foreground collector is waiting. 5295 if (_foregroundGCShouldWait) { 5296 // We cannot be the VM thread 5297 assert(t->is_ConcurrentGC_thread(), 5298 "Should be CMS thread"); 5299 } else { 5300 // We can be the CMS thread only if we are in a stop-world 5301 // phase of CMS collection. 5302 if (t->is_ConcurrentGC_thread()) { 5303 assert(_collectorState == InitialMarking || 5304 _collectorState == FinalMarking, 5305 "Should be a stop-world phase"); 5306 // The CMS thread should be holding the CMS_token. 5307 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5308 "Potential interference with concurrently " 5309 "executing VM thread"); 5310 } 5311 } 5312 } 5313 #endif 5314 5315 void CMSCollector::sweep() { 5316 assert(_collectorState == Sweeping, "just checking"); 5317 check_correct_thread_executing(); 5318 verify_work_stacks_empty(); 5319 verify_overflow_empty(); 5320 increment_sweep_count(); 5321 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5322 5323 _inter_sweep_timer.stop(); 5324 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5325 5326 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5327 _intra_sweep_timer.reset(); 5328 _intra_sweep_timer.start(); 5329 { 5330 GCTraceCPUTime tcpu; 5331 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5332 // First sweep the old gen 5333 { 5334 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5335 bitMapLock()); 5336 sweepWork(_cmsGen); 5337 } 5338 5339 // Update Universe::_heap_*_at_gc figures. 5340 // We need all the free list locks to make the abstract state 5341 // transition from Sweeping to Resetting. See detailed note 5342 // further below. 5343 { 5344 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5345 // Update heap occupancy information which is used as 5346 // input to soft ref clearing policy at the next gc. 5347 Universe::update_heap_info_at_gc(); 5348 _collectorState = Resizing; 5349 } 5350 } 5351 verify_work_stacks_empty(); 5352 verify_overflow_empty(); 5353 5354 if (should_unload_classes()) { 5355 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5356 // requires that the virtual spaces are stable and not deleted. 5357 ClassLoaderDataGraph::set_should_purge(true); 5358 } 5359 5360 _intra_sweep_timer.stop(); 5361 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5362 5363 _inter_sweep_timer.reset(); 5364 _inter_sweep_timer.start(); 5365 5366 // We need to use a monotonically non-decreasing time in ms 5367 // or we will see time-warp warnings and os::javaTimeMillis() 5368 // does not guarantee monotonicity. 5369 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5370 update_time_of_last_gc(now); 5371 5372 // NOTE on abstract state transitions: 5373 // Mutators allocate-live and/or mark the mod-union table dirty 5374 // based on the state of the collection. The former is done in 5375 // the interval [Marking, Sweeping] and the latter in the interval 5376 // [Marking, Sweeping). Thus the transitions into the Marking state 5377 // and out of the Sweeping state must be synchronously visible 5378 // globally to the mutators. 5379 // The transition into the Marking state happens with the world 5380 // stopped so the mutators will globally see it. Sweeping is 5381 // done asynchronously by the background collector so the transition 5382 // from the Sweeping state to the Resizing state must be done 5383 // under the freelistLock (as is the check for whether to 5384 // allocate-live and whether to dirty the mod-union table). 5385 assert(_collectorState == Resizing, "Change of collector state to" 5386 " Resizing must be done under the freelistLocks (plural)"); 5387 5388 // Now that sweeping has been completed, we clear 5389 // the incremental_collection_failed flag, 5390 // thus inviting a younger gen collection to promote into 5391 // this generation. If such a promotion may still fail, 5392 // the flag will be set again when a young collection is 5393 // attempted. 5394 CMSHeap* heap = CMSHeap::heap(); 5395 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5396 heap->update_full_collections_completed(_collection_count_start); 5397 } 5398 5399 // FIX ME!!! Looks like this belongs in CFLSpace, with 5400 // CMSGen merely delegating to it. 5401 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5402 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5403 HeapWord* minAddr = _cmsSpace->bottom(); 5404 HeapWord* largestAddr = 5405 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5406 if (largestAddr == NULL) { 5407 // The dictionary appears to be empty. In this case 5408 // try to coalesce at the end of the heap. 5409 largestAddr = _cmsSpace->end(); 5410 } 5411 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5412 size_t nearLargestOffset = 5413 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5414 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5415 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5416 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5417 } 5418 5419 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5420 return addr >= _cmsSpace->nearLargestChunk(); 5421 } 5422 5423 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5424 return _cmsSpace->find_chunk_at_end(); 5425 } 5426 5427 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5428 bool full) { 5429 // If the young generation has been collected, gather any statistics 5430 // that are of interest at this point. 5431 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5432 if (!full && current_is_young) { 5433 // Gather statistics on the young generation collection. 5434 collector()->stats().record_gc0_end(used()); 5435 } 5436 } 5437 5438 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5439 // We iterate over the space(s) underlying this generation, 5440 // checking the mark bit map to see if the bits corresponding 5441 // to specific blocks are marked or not. Blocks that are 5442 // marked are live and are not swept up. All remaining blocks 5443 // are swept up, with coalescing on-the-fly as we sweep up 5444 // contiguous free and/or garbage blocks: 5445 // We need to ensure that the sweeper synchronizes with allocators 5446 // and stop-the-world collectors. In particular, the following 5447 // locks are used: 5448 // . CMS token: if this is held, a stop the world collection cannot occur 5449 // . freelistLock: if this is held no allocation can occur from this 5450 // generation by another thread 5451 // . bitMapLock: if this is held, no other thread can access or update 5452 // 5453 5454 // Note that we need to hold the freelistLock if we use 5455 // block iterate below; else the iterator might go awry if 5456 // a mutator (or promotion) causes block contents to change 5457 // (for instance if the allocator divvies up a block). 5458 // If we hold the free list lock, for all practical purposes 5459 // young generation GC's can't occur (they'll usually need to 5460 // promote), so we might as well prevent all young generation 5461 // GC's while we do a sweeping step. For the same reason, we might 5462 // as well take the bit map lock for the entire duration 5463 5464 // check that we hold the requisite locks 5465 assert(have_cms_token(), "Should hold cms token"); 5466 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5467 assert_lock_strong(old_gen->freelistLock()); 5468 assert_lock_strong(bitMapLock()); 5469 5470 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5471 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5472 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5473 _inter_sweep_estimate.padded_average(), 5474 _intra_sweep_estimate.padded_average()); 5475 old_gen->setNearLargestChunk(); 5476 5477 { 5478 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5479 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5480 // We need to free-up/coalesce garbage/blocks from a 5481 // co-terminal free run. This is done in the SweepClosure 5482 // destructor; so, do not remove this scope, else the 5483 // end-of-sweep-census below will be off by a little bit. 5484 } 5485 old_gen->cmsSpace()->sweep_completed(); 5486 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5487 if (should_unload_classes()) { // unloaded classes this cycle, 5488 _concurrent_cycles_since_last_unload = 0; // ... reset count 5489 } else { // did not unload classes, 5490 _concurrent_cycles_since_last_unload++; // ... increment count 5491 } 5492 } 5493 5494 // Reset CMS data structures (for now just the marking bit map) 5495 // preparatory for the next cycle. 5496 void CMSCollector::reset_concurrent() { 5497 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5498 5499 // If the state is not "Resetting", the foreground thread 5500 // has done a collection and the resetting. 5501 if (_collectorState != Resetting) { 5502 assert(_collectorState == Idling, "The state should only change" 5503 " because the foreground collector has finished the collection"); 5504 return; 5505 } 5506 5507 { 5508 // Clear the mark bitmap (no grey objects to start with) 5509 // for the next cycle. 5510 GCTraceCPUTime tcpu; 5511 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5512 5513 HeapWord* curAddr = _markBitMap.startWord(); 5514 while (curAddr < _markBitMap.endWord()) { 5515 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5516 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5517 _markBitMap.clear_large_range(chunk); 5518 if (ConcurrentMarkSweepThread::should_yield() && 5519 !foregroundGCIsActive() && 5520 CMSYield) { 5521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5522 "CMS thread should hold CMS token"); 5523 assert_lock_strong(bitMapLock()); 5524 bitMapLock()->unlock(); 5525 ConcurrentMarkSweepThread::desynchronize(true); 5526 stopTimer(); 5527 incrementYields(); 5528 5529 // See the comment in coordinator_yield() 5530 for (unsigned i = 0; i < CMSYieldSleepCount && 5531 ConcurrentMarkSweepThread::should_yield() && 5532 !CMSCollector::foregroundGCIsActive(); ++i) { 5533 os::sleep(Thread::current(), 1, false); 5534 } 5535 5536 ConcurrentMarkSweepThread::synchronize(true); 5537 bitMapLock()->lock_without_safepoint_check(); 5538 startTimer(); 5539 } 5540 curAddr = chunk.end(); 5541 } 5542 // A successful mostly concurrent collection has been done. 5543 // Because only the full (i.e., concurrent mode failure) collections 5544 // are being measured for gc overhead limits, clean the "near" flag 5545 // and count. 5546 size_policy()->reset_gc_overhead_limit_count(); 5547 _collectorState = Idling; 5548 } 5549 5550 register_gc_end(); 5551 } 5552 5553 // Same as above but for STW paths 5554 void CMSCollector::reset_stw() { 5555 // already have the lock 5556 assert(_collectorState == Resetting, "just checking"); 5557 assert_lock_strong(bitMapLock()); 5558 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5559 _markBitMap.clear_all(); 5560 _collectorState = Idling; 5561 register_gc_end(); 5562 } 5563 5564 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5565 GCTraceCPUTime tcpu; 5566 TraceCollectorStats tcs_cgc(cgc_counters()); 5567 5568 switch (op) { 5569 case CMS_op_checkpointRootsInitial: { 5570 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5571 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5572 checkpointRootsInitial(); 5573 break; 5574 } 5575 case CMS_op_checkpointRootsFinal: { 5576 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5577 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5578 checkpointRootsFinal(); 5579 break; 5580 } 5581 default: 5582 fatal("No such CMS_op"); 5583 } 5584 } 5585 5586 #ifndef PRODUCT 5587 size_t const CMSCollector::skip_header_HeapWords() { 5588 return FreeChunk::header_size(); 5589 } 5590 5591 // Try and collect here conditions that should hold when 5592 // CMS thread is exiting. The idea is that the foreground GC 5593 // thread should not be blocked if it wants to terminate 5594 // the CMS thread and yet continue to run the VM for a while 5595 // after that. 5596 void CMSCollector::verify_ok_to_terminate() const { 5597 assert(Thread::current()->is_ConcurrentGC_thread(), 5598 "should be called by CMS thread"); 5599 assert(!_foregroundGCShouldWait, "should be false"); 5600 // We could check here that all the various low-level locks 5601 // are not held by the CMS thread, but that is overkill; see 5602 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5603 // is checked. 5604 } 5605 #endif 5606 5607 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5608 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5609 "missing Printezis mark?"); 5610 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5611 size_t size = pointer_delta(nextOneAddr + 1, addr); 5612 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5613 "alignment problem"); 5614 assert(size >= 3, "Necessary for Printezis marks to work"); 5615 return size; 5616 } 5617 5618 // A variant of the above (block_size_using_printezis_bits()) except 5619 // that we return 0 if the P-bits are not yet set. 5620 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5621 if (_markBitMap.isMarked(addr + 1)) { 5622 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5623 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5624 size_t size = pointer_delta(nextOneAddr + 1, addr); 5625 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5626 "alignment problem"); 5627 assert(size >= 3, "Necessary for Printezis marks to work"); 5628 return size; 5629 } 5630 return 0; 5631 } 5632 5633 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5634 size_t sz = 0; 5635 oop p = (oop)addr; 5636 if (p->klass_or_null_acquire() != NULL) { 5637 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5638 } else { 5639 sz = block_size_using_printezis_bits(addr); 5640 } 5641 assert(sz > 0, "size must be nonzero"); 5642 HeapWord* next_block = addr + sz; 5643 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5644 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5645 align_down((uintptr_t)next_card, CardTable::card_size), 5646 "must be different cards"); 5647 return next_card; 5648 } 5649 5650 5651 // CMS Bit Map Wrapper ///////////////////////////////////////// 5652 5653 // Construct a CMS bit map infrastructure, but don't create the 5654 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5655 // further below. 5656 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5657 _shifter(shifter), 5658 _bm(), 5659 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5660 Monitor::_safepoint_check_sometimes) : NULL) 5661 { 5662 _bmStartWord = 0; 5663 _bmWordSize = 0; 5664 } 5665 5666 bool CMSBitMap::allocate(MemRegion mr) { 5667 _bmStartWord = mr.start(); 5668 _bmWordSize = mr.word_size(); 5669 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5670 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5671 if (!brs.is_reserved()) { 5672 log_warning(gc)("CMS bit map allocation failure"); 5673 return false; 5674 } 5675 // For now we'll just commit all of the bit map up front. 5676 // Later on we'll try to be more parsimonious with swap. 5677 if (!_virtual_space.initialize(brs, brs.size())) { 5678 log_warning(gc)("CMS bit map backing store failure"); 5679 return false; 5680 } 5681 assert(_virtual_space.committed_size() == brs.size(), 5682 "didn't reserve backing store for all of CMS bit map?"); 5683 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5684 _bmWordSize, "inconsistency in bit map sizing"); 5685 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5686 5687 // bm.clear(); // can we rely on getting zero'd memory? verify below 5688 assert(isAllClear(), 5689 "Expected zero'd memory from ReservedSpace constructor"); 5690 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5691 "consistency check"); 5692 return true; 5693 } 5694 5695 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5696 HeapWord *next_addr, *end_addr, *last_addr; 5697 assert_locked(); 5698 assert(covers(mr), "out-of-range error"); 5699 // XXX assert that start and end are appropriately aligned 5700 for (next_addr = mr.start(), end_addr = mr.end(); 5701 next_addr < end_addr; next_addr = last_addr) { 5702 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5703 last_addr = dirty_region.end(); 5704 if (!dirty_region.is_empty()) { 5705 cl->do_MemRegion(dirty_region); 5706 } else { 5707 assert(last_addr == end_addr, "program logic"); 5708 return; 5709 } 5710 } 5711 } 5712 5713 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5714 _bm.print_on_error(st, prefix); 5715 } 5716 5717 #ifndef PRODUCT 5718 void CMSBitMap::assert_locked() const { 5719 CMSLockVerifier::assert_locked(lock()); 5720 } 5721 5722 bool CMSBitMap::covers(MemRegion mr) const { 5723 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5724 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5725 "size inconsistency"); 5726 return (mr.start() >= _bmStartWord) && 5727 (mr.end() <= endWord()); 5728 } 5729 5730 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5731 return (start >= _bmStartWord && (start + size) <= endWord()); 5732 } 5733 5734 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5735 // verify that there are no 1 bits in the interval [left, right) 5736 FalseBitMapClosure falseBitMapClosure; 5737 iterate(&falseBitMapClosure, left, right); 5738 } 5739 5740 void CMSBitMap::region_invariant(MemRegion mr) 5741 { 5742 assert_locked(); 5743 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5744 assert(!mr.is_empty(), "unexpected empty region"); 5745 assert(covers(mr), "mr should be covered by bit map"); 5746 // convert address range into offset range 5747 size_t start_ofs = heapWordToOffset(mr.start()); 5748 // Make sure that end() is appropriately aligned 5749 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5750 "Misaligned mr.end()"); 5751 size_t end_ofs = heapWordToOffset(mr.end()); 5752 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5753 } 5754 5755 #endif 5756 5757 bool CMSMarkStack::allocate(size_t size) { 5758 // allocate a stack of the requisite depth 5759 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5760 size * sizeof(oop))); 5761 if (!rs.is_reserved()) { 5762 log_warning(gc)("CMSMarkStack allocation failure"); 5763 return false; 5764 } 5765 if (!_virtual_space.initialize(rs, rs.size())) { 5766 log_warning(gc)("CMSMarkStack backing store failure"); 5767 return false; 5768 } 5769 assert(_virtual_space.committed_size() == rs.size(), 5770 "didn't reserve backing store for all of CMS stack?"); 5771 _base = (oop*)(_virtual_space.low()); 5772 _index = 0; 5773 _capacity = size; 5774 NOT_PRODUCT(_max_depth = 0); 5775 return true; 5776 } 5777 5778 // XXX FIX ME !!! In the MT case we come in here holding a 5779 // leaf lock. For printing we need to take a further lock 5780 // which has lower rank. We need to recalibrate the two 5781 // lock-ranks involved in order to be able to print the 5782 // messages below. (Or defer the printing to the caller. 5783 // For now we take the expedient path of just disabling the 5784 // messages for the problematic case.) 5785 void CMSMarkStack::expand() { 5786 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5787 if (_capacity == MarkStackSizeMax) { 5788 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5789 // We print a warning message only once per CMS cycle. 5790 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5791 } 5792 return; 5793 } 5794 // Double capacity if possible 5795 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5796 // Do not give up existing stack until we have managed to 5797 // get the double capacity that we desired. 5798 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5799 new_capacity * sizeof(oop))); 5800 if (rs.is_reserved()) { 5801 // Release the backing store associated with old stack 5802 _virtual_space.release(); 5803 // Reinitialize virtual space for new stack 5804 if (!_virtual_space.initialize(rs, rs.size())) { 5805 fatal("Not enough swap for expanded marking stack"); 5806 } 5807 _base = (oop*)(_virtual_space.low()); 5808 _index = 0; 5809 _capacity = new_capacity; 5810 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5811 // Failed to double capacity, continue; 5812 // we print a detail message only once per CMS cycle. 5813 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5814 _capacity / K, new_capacity / K); 5815 } 5816 } 5817 5818 5819 // Closures 5820 // XXX: there seems to be a lot of code duplication here; 5821 // should refactor and consolidate common code. 5822 5823 // This closure is used to mark refs into the CMS generation in 5824 // the CMS bit map. Called at the first checkpoint. This closure 5825 // assumes that we do not need to re-mark dirty cards; if the CMS 5826 // generation on which this is used is not an oldest 5827 // generation then this will lose younger_gen cards! 5828 5829 MarkRefsIntoClosure::MarkRefsIntoClosure( 5830 MemRegion span, CMSBitMap* bitMap): 5831 _span(span), 5832 _bitMap(bitMap) 5833 { 5834 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5835 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5836 } 5837 5838 void MarkRefsIntoClosure::do_oop(oop obj) { 5839 // if p points into _span, then mark corresponding bit in _markBitMap 5840 assert(oopDesc::is_oop(obj), "expected an oop"); 5841 HeapWord* addr = (HeapWord*)obj; 5842 if (_span.contains(addr)) { 5843 // this should be made more efficient 5844 _bitMap->mark(addr); 5845 } 5846 } 5847 5848 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5849 MemRegion span, CMSBitMap* bitMap): 5850 _span(span), 5851 _bitMap(bitMap) 5852 { 5853 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5854 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5855 } 5856 5857 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5858 // if p points into _span, then mark corresponding bit in _markBitMap 5859 assert(oopDesc::is_oop(obj), "expected an oop"); 5860 HeapWord* addr = (HeapWord*)obj; 5861 if (_span.contains(addr)) { 5862 // this should be made more efficient 5863 _bitMap->par_mark(addr); 5864 } 5865 } 5866 5867 // A variant of the above, used for CMS marking verification. 5868 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5869 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5870 _span(span), 5871 _verification_bm(verification_bm), 5872 _cms_bm(cms_bm) 5873 { 5874 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5875 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5876 } 5877 5878 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5879 // if p points into _span, then mark corresponding bit in _markBitMap 5880 assert(oopDesc::is_oop(obj), "expected an oop"); 5881 HeapWord* addr = (HeapWord*)obj; 5882 if (_span.contains(addr)) { 5883 _verification_bm->mark(addr); 5884 if (!_cms_bm->isMarked(addr)) { 5885 Log(gc, verify) log; 5886 ResourceMark rm; 5887 LogStream ls(log.error()); 5888 oop(addr)->print_on(&ls); 5889 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5890 fatal("... aborting"); 5891 } 5892 } 5893 } 5894 5895 ////////////////////////////////////////////////// 5896 // MarkRefsIntoAndScanClosure 5897 ////////////////////////////////////////////////// 5898 5899 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5900 ReferenceDiscoverer* rd, 5901 CMSBitMap* bit_map, 5902 CMSBitMap* mod_union_table, 5903 CMSMarkStack* mark_stack, 5904 CMSCollector* collector, 5905 bool should_yield, 5906 bool concurrent_precleaning): 5907 _span(span), 5908 _bit_map(bit_map), 5909 _mark_stack(mark_stack), 5910 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5911 mark_stack, concurrent_precleaning), 5912 _collector(collector), 5913 _freelistLock(NULL), 5914 _yield(should_yield), 5915 _concurrent_precleaning(concurrent_precleaning) 5916 { 5917 // FIXME: Should initialize in base class constructor. 5918 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5919 set_ref_discoverer_internal(rd); 5920 } 5921 5922 // This closure is used to mark refs into the CMS generation at the 5923 // second (final) checkpoint, and to scan and transitively follow 5924 // the unmarked oops. It is also used during the concurrent precleaning 5925 // phase while scanning objects on dirty cards in the CMS generation. 5926 // The marks are made in the marking bit map and the marking stack is 5927 // used for keeping the (newly) grey objects during the scan. 5928 // The parallel version (Par_...) appears further below. 5929 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5930 if (obj != NULL) { 5931 assert(oopDesc::is_oop(obj), "expected an oop"); 5932 HeapWord* addr = (HeapWord*)obj; 5933 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5934 assert(_collector->overflow_list_is_empty(), 5935 "overflow list should be empty"); 5936 if (_span.contains(addr) && 5937 !_bit_map->isMarked(addr)) { 5938 // mark bit map (object is now grey) 5939 _bit_map->mark(addr); 5940 // push on marking stack (stack should be empty), and drain the 5941 // stack by applying this closure to the oops in the oops popped 5942 // from the stack (i.e. blacken the grey objects) 5943 bool res = _mark_stack->push(obj); 5944 assert(res, "Should have space to push on empty stack"); 5945 do { 5946 oop new_oop = _mark_stack->pop(); 5947 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5948 assert(_bit_map->isMarked((HeapWord*)new_oop), 5949 "only grey objects on this stack"); 5950 // iterate over the oops in this oop, marking and pushing 5951 // the ones in CMS heap (i.e. in _span). 5952 new_oop->oop_iterate(&_pushAndMarkClosure); 5953 // check if it's time to yield 5954 do_yield_check(); 5955 } while (!_mark_stack->isEmpty() || 5956 (!_concurrent_precleaning && take_from_overflow_list())); 5957 // if marking stack is empty, and we are not doing this 5958 // during precleaning, then check the overflow list 5959 } 5960 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5961 assert(_collector->overflow_list_is_empty(), 5962 "overflow list was drained above"); 5963 5964 assert(_collector->no_preserved_marks(), 5965 "All preserved marks should have been restored above"); 5966 } 5967 } 5968 5969 void MarkRefsIntoAndScanClosure::do_yield_work() { 5970 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5971 "CMS thread should hold CMS token"); 5972 assert_lock_strong(_freelistLock); 5973 assert_lock_strong(_bit_map->lock()); 5974 // relinquish the free_list_lock and bitMaplock() 5975 _bit_map->lock()->unlock(); 5976 _freelistLock->unlock(); 5977 ConcurrentMarkSweepThread::desynchronize(true); 5978 _collector->stopTimer(); 5979 _collector->incrementYields(); 5980 5981 // See the comment in coordinator_yield() 5982 for (unsigned i = 0; 5983 i < CMSYieldSleepCount && 5984 ConcurrentMarkSweepThread::should_yield() && 5985 !CMSCollector::foregroundGCIsActive(); 5986 ++i) { 5987 os::sleep(Thread::current(), 1, false); 5988 } 5989 5990 ConcurrentMarkSweepThread::synchronize(true); 5991 _freelistLock->lock_without_safepoint_check(); 5992 _bit_map->lock()->lock_without_safepoint_check(); 5993 _collector->startTimer(); 5994 } 5995 5996 /////////////////////////////////////////////////////////// 5997 // ParMarkRefsIntoAndScanClosure: a parallel version of 5998 // MarkRefsIntoAndScanClosure 5999 /////////////////////////////////////////////////////////// 6000 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6001 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 6002 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6003 _span(span), 6004 _bit_map(bit_map), 6005 _work_queue(work_queue), 6006 _low_water_mark(MIN2((work_queue->max_elems()/4), 6007 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6008 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 6009 { 6010 // FIXME: Should initialize in base class constructor. 6011 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 6012 set_ref_discoverer_internal(rd); 6013 } 6014 6015 // This closure is used to mark refs into the CMS generation at the 6016 // second (final) checkpoint, and to scan and transitively follow 6017 // the unmarked oops. The marks are made in the marking bit map and 6018 // the work_queue is used for keeping the (newly) grey objects during 6019 // the scan phase whence they are also available for stealing by parallel 6020 // threads. Since the marking bit map is shared, updates are 6021 // synchronized (via CAS). 6022 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6023 if (obj != NULL) { 6024 // Ignore mark word because this could be an already marked oop 6025 // that may be chained at the end of the overflow list. 6026 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6027 HeapWord* addr = (HeapWord*)obj; 6028 if (_span.contains(addr) && 6029 !_bit_map->isMarked(addr)) { 6030 // mark bit map (object will become grey): 6031 // It is possible for several threads to be 6032 // trying to "claim" this object concurrently; 6033 // the unique thread that succeeds in marking the 6034 // object first will do the subsequent push on 6035 // to the work queue (or overflow list). 6036 if (_bit_map->par_mark(addr)) { 6037 // push on work_queue (which may not be empty), and trim the 6038 // queue to an appropriate length by applying this closure to 6039 // the oops in the oops popped from the stack (i.e. blacken the 6040 // grey objects) 6041 bool res = _work_queue->push(obj); 6042 assert(res, "Low water mark should be less than capacity?"); 6043 trim_queue(_low_water_mark); 6044 } // Else, another thread claimed the object 6045 } 6046 } 6047 } 6048 6049 // This closure is used to rescan the marked objects on the dirty cards 6050 // in the mod union table and the card table proper. 6051 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6052 oop p, MemRegion mr) { 6053 6054 size_t size = 0; 6055 HeapWord* addr = (HeapWord*)p; 6056 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6057 assert(_span.contains(addr), "we are scanning the CMS generation"); 6058 // check if it's time to yield 6059 if (do_yield_check()) { 6060 // We yielded for some foreground stop-world work, 6061 // and we have been asked to abort this ongoing preclean cycle. 6062 return 0; 6063 } 6064 if (_bitMap->isMarked(addr)) { 6065 // it's marked; is it potentially uninitialized? 6066 if (p->klass_or_null_acquire() != NULL) { 6067 // an initialized object; ignore mark word in verification below 6068 // since we are running concurrent with mutators 6069 assert(oopDesc::is_oop(p, true), "should be an oop"); 6070 if (p->is_objArray()) { 6071 // objArrays are precisely marked; restrict scanning 6072 // to dirty cards only. 6073 size = CompactibleFreeListSpace::adjustObjectSize( 6074 p->oop_iterate_size(_scanningClosure, mr)); 6075 } else { 6076 // A non-array may have been imprecisely marked; we need 6077 // to scan object in its entirety. 6078 size = CompactibleFreeListSpace::adjustObjectSize( 6079 p->oop_iterate_size(_scanningClosure)); 6080 } 6081 #ifdef ASSERT 6082 size_t direct_size = 6083 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6084 assert(size == direct_size, "Inconsistency in size"); 6085 assert(size >= 3, "Necessary for Printezis marks to work"); 6086 HeapWord* start_pbit = addr + 1; 6087 HeapWord* end_pbit = addr + size - 1; 6088 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6089 "inconsistent Printezis mark"); 6090 // Verify inner mark bits (between Printezis bits) are clear, 6091 // but don't repeat if there are multiple dirty regions for 6092 // the same object, to avoid potential O(N^2) performance. 6093 if (addr != _last_scanned_object) { 6094 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6095 _last_scanned_object = addr; 6096 } 6097 #endif // ASSERT 6098 } else { 6099 // An uninitialized object. 6100 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6101 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6102 size = pointer_delta(nextOneAddr + 1, addr); 6103 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6104 "alignment problem"); 6105 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6106 // will dirty the card when the klass pointer is installed in the 6107 // object (signaling the completion of initialization). 6108 } 6109 } else { 6110 // Either a not yet marked object or an uninitialized object 6111 if (p->klass_or_null_acquire() == NULL) { 6112 // An uninitialized object, skip to the next card, since 6113 // we may not be able to read its P-bits yet. 6114 assert(size == 0, "Initial value"); 6115 } else { 6116 // An object not (yet) reached by marking: we merely need to 6117 // compute its size so as to go look at the next block. 6118 assert(oopDesc::is_oop(p, true), "should be an oop"); 6119 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6120 } 6121 } 6122 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6123 return size; 6124 } 6125 6126 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6127 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6128 "CMS thread should hold CMS token"); 6129 assert_lock_strong(_freelistLock); 6130 assert_lock_strong(_bitMap->lock()); 6131 // relinquish the free_list_lock and bitMaplock() 6132 _bitMap->lock()->unlock(); 6133 _freelistLock->unlock(); 6134 ConcurrentMarkSweepThread::desynchronize(true); 6135 _collector->stopTimer(); 6136 _collector->incrementYields(); 6137 6138 // See the comment in coordinator_yield() 6139 for (unsigned i = 0; i < CMSYieldSleepCount && 6140 ConcurrentMarkSweepThread::should_yield() && 6141 !CMSCollector::foregroundGCIsActive(); ++i) { 6142 os::sleep(Thread::current(), 1, false); 6143 } 6144 6145 ConcurrentMarkSweepThread::synchronize(true); 6146 _freelistLock->lock_without_safepoint_check(); 6147 _bitMap->lock()->lock_without_safepoint_check(); 6148 _collector->startTimer(); 6149 } 6150 6151 6152 ////////////////////////////////////////////////////////////////// 6153 // SurvivorSpacePrecleanClosure 6154 ////////////////////////////////////////////////////////////////// 6155 // This (single-threaded) closure is used to preclean the oops in 6156 // the survivor spaces. 6157 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6158 6159 HeapWord* addr = (HeapWord*)p; 6160 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6161 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6162 assert(p->klass_or_null() != NULL, "object should be initialized"); 6163 // an initialized object; ignore mark word in verification below 6164 // since we are running concurrent with mutators 6165 assert(oopDesc::is_oop(p, true), "should be an oop"); 6166 // Note that we do not yield while we iterate over 6167 // the interior oops of p, pushing the relevant ones 6168 // on our marking stack. 6169 size_t size = p->oop_iterate_size(_scanning_closure); 6170 do_yield_check(); 6171 // Observe that below, we do not abandon the preclean 6172 // phase as soon as we should; rather we empty the 6173 // marking stack before returning. This is to satisfy 6174 // some existing assertions. In general, it may be a 6175 // good idea to abort immediately and complete the marking 6176 // from the grey objects at a later time. 6177 while (!_mark_stack->isEmpty()) { 6178 oop new_oop = _mark_stack->pop(); 6179 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6180 assert(_bit_map->isMarked((HeapWord*)new_oop), 6181 "only grey objects on this stack"); 6182 // iterate over the oops in this oop, marking and pushing 6183 // the ones in CMS heap (i.e. in _span). 6184 new_oop->oop_iterate(_scanning_closure); 6185 // check if it's time to yield 6186 do_yield_check(); 6187 } 6188 unsigned int after_count = 6189 CMSHeap::heap()->total_collections(); 6190 bool abort = (_before_count != after_count) || 6191 _collector->should_abort_preclean(); 6192 return abort ? 0 : size; 6193 } 6194 6195 void SurvivorSpacePrecleanClosure::do_yield_work() { 6196 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6197 "CMS thread should hold CMS token"); 6198 assert_lock_strong(_bit_map->lock()); 6199 // Relinquish the bit map lock 6200 _bit_map->lock()->unlock(); 6201 ConcurrentMarkSweepThread::desynchronize(true); 6202 _collector->stopTimer(); 6203 _collector->incrementYields(); 6204 6205 // See the comment in coordinator_yield() 6206 for (unsigned i = 0; i < CMSYieldSleepCount && 6207 ConcurrentMarkSweepThread::should_yield() && 6208 !CMSCollector::foregroundGCIsActive(); ++i) { 6209 os::sleep(Thread::current(), 1, false); 6210 } 6211 6212 ConcurrentMarkSweepThread::synchronize(true); 6213 _bit_map->lock()->lock_without_safepoint_check(); 6214 _collector->startTimer(); 6215 } 6216 6217 // This closure is used to rescan the marked objects on the dirty cards 6218 // in the mod union table and the card table proper. In the parallel 6219 // case, although the bitMap is shared, we do a single read so the 6220 // isMarked() query is "safe". 6221 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6222 // Ignore mark word because we are running concurrent with mutators 6223 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6224 HeapWord* addr = (HeapWord*)p; 6225 assert(_span.contains(addr), "we are scanning the CMS generation"); 6226 bool is_obj_array = false; 6227 #ifdef ASSERT 6228 if (!_parallel) { 6229 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6230 assert(_collector->overflow_list_is_empty(), 6231 "overflow list should be empty"); 6232 6233 } 6234 #endif // ASSERT 6235 if (_bit_map->isMarked(addr)) { 6236 // Obj arrays are precisely marked, non-arrays are not; 6237 // so we scan objArrays precisely and non-arrays in their 6238 // entirety. 6239 if (p->is_objArray()) { 6240 is_obj_array = true; 6241 if (_parallel) { 6242 p->oop_iterate(_par_scan_closure, mr); 6243 } else { 6244 p->oop_iterate(_scan_closure, mr); 6245 } 6246 } else { 6247 if (_parallel) { 6248 p->oop_iterate(_par_scan_closure); 6249 } else { 6250 p->oop_iterate(_scan_closure); 6251 } 6252 } 6253 } 6254 #ifdef ASSERT 6255 if (!_parallel) { 6256 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6257 assert(_collector->overflow_list_is_empty(), 6258 "overflow list should be empty"); 6259 6260 } 6261 #endif // ASSERT 6262 return is_obj_array; 6263 } 6264 6265 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6266 MemRegion span, 6267 CMSBitMap* bitMap, CMSMarkStack* markStack, 6268 bool should_yield, bool verifying): 6269 _collector(collector), 6270 _span(span), 6271 _bitMap(bitMap), 6272 _mut(&collector->_modUnionTable), 6273 _markStack(markStack), 6274 _yield(should_yield), 6275 _skipBits(0) 6276 { 6277 assert(_markStack->isEmpty(), "stack should be empty"); 6278 _finger = _bitMap->startWord(); 6279 _threshold = _finger; 6280 assert(_collector->_restart_addr == NULL, "Sanity check"); 6281 assert(_span.contains(_finger), "Out of bounds _finger?"); 6282 DEBUG_ONLY(_verifying = verifying;) 6283 } 6284 6285 void MarkFromRootsClosure::reset(HeapWord* addr) { 6286 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6287 assert(_span.contains(addr), "Out of bounds _finger?"); 6288 _finger = addr; 6289 _threshold = align_up(_finger, CardTable::card_size); 6290 } 6291 6292 // Should revisit to see if this should be restructured for 6293 // greater efficiency. 6294 bool MarkFromRootsClosure::do_bit(size_t offset) { 6295 if (_skipBits > 0) { 6296 _skipBits--; 6297 return true; 6298 } 6299 // convert offset into a HeapWord* 6300 HeapWord* addr = _bitMap->startWord() + offset; 6301 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6302 "address out of range"); 6303 assert(_bitMap->isMarked(addr), "tautology"); 6304 if (_bitMap->isMarked(addr+1)) { 6305 // this is an allocated but not yet initialized object 6306 assert(_skipBits == 0, "tautology"); 6307 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6308 oop p = oop(addr); 6309 if (p->klass_or_null_acquire() == NULL) { 6310 DEBUG_ONLY(if (!_verifying) {) 6311 // We re-dirty the cards on which this object lies and increase 6312 // the _threshold so that we'll come back to scan this object 6313 // during the preclean or remark phase. (CMSCleanOnEnter) 6314 if (CMSCleanOnEnter) { 6315 size_t sz = _collector->block_size_using_printezis_bits(addr); 6316 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6317 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6318 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6319 // Bump _threshold to end_card_addr; note that 6320 // _threshold cannot possibly exceed end_card_addr, anyhow. 6321 // This prevents future clearing of the card as the scan proceeds 6322 // to the right. 6323 assert(_threshold <= end_card_addr, 6324 "Because we are just scanning into this object"); 6325 if (_threshold < end_card_addr) { 6326 _threshold = end_card_addr; 6327 } 6328 if (p->klass_or_null_acquire() != NULL) { 6329 // Redirty the range of cards... 6330 _mut->mark_range(redirty_range); 6331 } // ...else the setting of klass will dirty the card anyway. 6332 } 6333 DEBUG_ONLY(}) 6334 return true; 6335 } 6336 } 6337 scanOopsInOop(addr); 6338 return true; 6339 } 6340 6341 // We take a break if we've been at this for a while, 6342 // so as to avoid monopolizing the locks involved. 6343 void MarkFromRootsClosure::do_yield_work() { 6344 // First give up the locks, then yield, then re-lock 6345 // We should probably use a constructor/destructor idiom to 6346 // do this unlock/lock or modify the MutexUnlocker class to 6347 // serve our purpose. XXX 6348 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6349 "CMS thread should hold CMS token"); 6350 assert_lock_strong(_bitMap->lock()); 6351 _bitMap->lock()->unlock(); 6352 ConcurrentMarkSweepThread::desynchronize(true); 6353 _collector->stopTimer(); 6354 _collector->incrementYields(); 6355 6356 // See the comment in coordinator_yield() 6357 for (unsigned i = 0; i < CMSYieldSleepCount && 6358 ConcurrentMarkSweepThread::should_yield() && 6359 !CMSCollector::foregroundGCIsActive(); ++i) { 6360 os::sleep(Thread::current(), 1, false); 6361 } 6362 6363 ConcurrentMarkSweepThread::synchronize(true); 6364 _bitMap->lock()->lock_without_safepoint_check(); 6365 _collector->startTimer(); 6366 } 6367 6368 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6369 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6370 assert(_markStack->isEmpty(), 6371 "should drain stack to limit stack usage"); 6372 // convert ptr to an oop preparatory to scanning 6373 oop obj = oop(ptr); 6374 // Ignore mark word in verification below, since we 6375 // may be running concurrent with mutators. 6376 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6377 assert(_finger <= ptr, "_finger runneth ahead"); 6378 // advance the finger to right end of this object 6379 _finger = ptr + obj->size(); 6380 assert(_finger > ptr, "we just incremented it above"); 6381 // On large heaps, it may take us some time to get through 6382 // the marking phase. During 6383 // this time it's possible that a lot of mutations have 6384 // accumulated in the card table and the mod union table -- 6385 // these mutation records are redundant until we have 6386 // actually traced into the corresponding card. 6387 // Here, we check whether advancing the finger would make 6388 // us cross into a new card, and if so clear corresponding 6389 // cards in the MUT (preclean them in the card-table in the 6390 // future). 6391 6392 DEBUG_ONLY(if (!_verifying) {) 6393 // The clean-on-enter optimization is disabled by default, 6394 // until we fix 6178663. 6395 if (CMSCleanOnEnter && (_finger > _threshold)) { 6396 // [_threshold, _finger) represents the interval 6397 // of cards to be cleared in MUT (or precleaned in card table). 6398 // The set of cards to be cleared is all those that overlap 6399 // with the interval [_threshold, _finger); note that 6400 // _threshold is always kept card-aligned but _finger isn't 6401 // always card-aligned. 6402 HeapWord* old_threshold = _threshold; 6403 assert(is_aligned(old_threshold, CardTable::card_size), 6404 "_threshold should always be card-aligned"); 6405 _threshold = align_up(_finger, CardTable::card_size); 6406 MemRegion mr(old_threshold, _threshold); 6407 assert(!mr.is_empty(), "Control point invariant"); 6408 assert(_span.contains(mr), "Should clear within span"); 6409 _mut->clear_range(mr); 6410 } 6411 DEBUG_ONLY(}) 6412 // Note: the finger doesn't advance while we drain 6413 // the stack below. 6414 PushOrMarkClosure pushOrMarkClosure(_collector, 6415 _span, _bitMap, _markStack, 6416 _finger, this); 6417 bool res = _markStack->push(obj); 6418 assert(res, "Empty non-zero size stack should have space for single push"); 6419 while (!_markStack->isEmpty()) { 6420 oop new_oop = _markStack->pop(); 6421 // Skip verifying header mark word below because we are 6422 // running concurrent with mutators. 6423 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6424 // now scan this oop's oops 6425 new_oop->oop_iterate(&pushOrMarkClosure); 6426 do_yield_check(); 6427 } 6428 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6429 } 6430 6431 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6432 CMSCollector* collector, MemRegion span, 6433 CMSBitMap* bit_map, 6434 OopTaskQueue* work_queue, 6435 CMSMarkStack* overflow_stack): 6436 _collector(collector), 6437 _whole_span(collector->_span), 6438 _span(span), 6439 _bit_map(bit_map), 6440 _mut(&collector->_modUnionTable), 6441 _work_queue(work_queue), 6442 _overflow_stack(overflow_stack), 6443 _skip_bits(0), 6444 _task(task) 6445 { 6446 assert(_work_queue->size() == 0, "work_queue should be empty"); 6447 _finger = span.start(); 6448 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6449 assert(_span.contains(_finger), "Out of bounds _finger?"); 6450 } 6451 6452 // Should revisit to see if this should be restructured for 6453 // greater efficiency. 6454 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6455 if (_skip_bits > 0) { 6456 _skip_bits--; 6457 return true; 6458 } 6459 // convert offset into a HeapWord* 6460 HeapWord* addr = _bit_map->startWord() + offset; 6461 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6462 "address out of range"); 6463 assert(_bit_map->isMarked(addr), "tautology"); 6464 if (_bit_map->isMarked(addr+1)) { 6465 // this is an allocated object that might not yet be initialized 6466 assert(_skip_bits == 0, "tautology"); 6467 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6468 oop p = oop(addr); 6469 if (p->klass_or_null_acquire() == NULL) { 6470 // in the case of Clean-on-Enter optimization, redirty card 6471 // and avoid clearing card by increasing the threshold. 6472 return true; 6473 } 6474 } 6475 scan_oops_in_oop(addr); 6476 return true; 6477 } 6478 6479 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6480 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6481 // Should we assert that our work queue is empty or 6482 // below some drain limit? 6483 assert(_work_queue->size() == 0, 6484 "should drain stack to limit stack usage"); 6485 // convert ptr to an oop preparatory to scanning 6486 oop obj = oop(ptr); 6487 // Ignore mark word in verification below, since we 6488 // may be running concurrent with mutators. 6489 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6490 assert(_finger <= ptr, "_finger runneth ahead"); 6491 // advance the finger to right end of this object 6492 _finger = ptr + obj->size(); 6493 assert(_finger > ptr, "we just incremented it above"); 6494 // On large heaps, it may take us some time to get through 6495 // the marking phase. During 6496 // this time it's possible that a lot of mutations have 6497 // accumulated in the card table and the mod union table -- 6498 // these mutation records are redundant until we have 6499 // actually traced into the corresponding card. 6500 // Here, we check whether advancing the finger would make 6501 // us cross into a new card, and if so clear corresponding 6502 // cards in the MUT (preclean them in the card-table in the 6503 // future). 6504 6505 // The clean-on-enter optimization is disabled by default, 6506 // until we fix 6178663. 6507 if (CMSCleanOnEnter && (_finger > _threshold)) { 6508 // [_threshold, _finger) represents the interval 6509 // of cards to be cleared in MUT (or precleaned in card table). 6510 // The set of cards to be cleared is all those that overlap 6511 // with the interval [_threshold, _finger); note that 6512 // _threshold is always kept card-aligned but _finger isn't 6513 // always card-aligned. 6514 HeapWord* old_threshold = _threshold; 6515 assert(is_aligned(old_threshold, CardTable::card_size), 6516 "_threshold should always be card-aligned"); 6517 _threshold = align_up(_finger, CardTable::card_size); 6518 MemRegion mr(old_threshold, _threshold); 6519 assert(!mr.is_empty(), "Control point invariant"); 6520 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6521 _mut->clear_range(mr); 6522 } 6523 6524 // Note: the local finger doesn't advance while we drain 6525 // the stack below, but the global finger sure can and will. 6526 HeapWord* volatile* gfa = _task->global_finger_addr(); 6527 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6528 _span, _bit_map, 6529 _work_queue, 6530 _overflow_stack, 6531 _finger, 6532 gfa, this); 6533 bool res = _work_queue->push(obj); // overflow could occur here 6534 assert(res, "Will hold once we use workqueues"); 6535 while (true) { 6536 oop new_oop; 6537 if (!_work_queue->pop_local(new_oop)) { 6538 // We emptied our work_queue; check if there's stuff that can 6539 // be gotten from the overflow stack. 6540 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6541 _overflow_stack, _work_queue)) { 6542 do_yield_check(); 6543 continue; 6544 } else { // done 6545 break; 6546 } 6547 } 6548 // Skip verifying header mark word below because we are 6549 // running concurrent with mutators. 6550 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6551 // now scan this oop's oops 6552 new_oop->oop_iterate(&pushOrMarkClosure); 6553 do_yield_check(); 6554 } 6555 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6556 } 6557 6558 // Yield in response to a request from VM Thread or 6559 // from mutators. 6560 void ParMarkFromRootsClosure::do_yield_work() { 6561 assert(_task != NULL, "sanity"); 6562 _task->yield(); 6563 } 6564 6565 // A variant of the above used for verifying CMS marking work. 6566 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6567 MemRegion span, 6568 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6569 CMSMarkStack* mark_stack): 6570 _collector(collector), 6571 _span(span), 6572 _verification_bm(verification_bm), 6573 _cms_bm(cms_bm), 6574 _mark_stack(mark_stack), 6575 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6576 mark_stack) 6577 { 6578 assert(_mark_stack->isEmpty(), "stack should be empty"); 6579 _finger = _verification_bm->startWord(); 6580 assert(_collector->_restart_addr == NULL, "Sanity check"); 6581 assert(_span.contains(_finger), "Out of bounds _finger?"); 6582 } 6583 6584 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6585 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6586 assert(_span.contains(addr), "Out of bounds _finger?"); 6587 _finger = addr; 6588 } 6589 6590 // Should revisit to see if this should be restructured for 6591 // greater efficiency. 6592 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6593 // convert offset into a HeapWord* 6594 HeapWord* addr = _verification_bm->startWord() + offset; 6595 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6596 "address out of range"); 6597 assert(_verification_bm->isMarked(addr), "tautology"); 6598 assert(_cms_bm->isMarked(addr), "tautology"); 6599 6600 assert(_mark_stack->isEmpty(), 6601 "should drain stack to limit stack usage"); 6602 // convert addr to an oop preparatory to scanning 6603 oop obj = oop(addr); 6604 assert(oopDesc::is_oop(obj), "should be an oop"); 6605 assert(_finger <= addr, "_finger runneth ahead"); 6606 // advance the finger to right end of this object 6607 _finger = addr + obj->size(); 6608 assert(_finger > addr, "we just incremented it above"); 6609 // Note: the finger doesn't advance while we drain 6610 // the stack below. 6611 bool res = _mark_stack->push(obj); 6612 assert(res, "Empty non-zero size stack should have space for single push"); 6613 while (!_mark_stack->isEmpty()) { 6614 oop new_oop = _mark_stack->pop(); 6615 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6616 // now scan this oop's oops 6617 new_oop->oop_iterate(&_pam_verify_closure); 6618 } 6619 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6620 return true; 6621 } 6622 6623 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6624 CMSCollector* collector, MemRegion span, 6625 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6626 CMSMarkStack* mark_stack): 6627 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6628 _collector(collector), 6629 _span(span), 6630 _verification_bm(verification_bm), 6631 _cms_bm(cms_bm), 6632 _mark_stack(mark_stack) 6633 { } 6634 6635 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6636 oop obj = RawAccess<>::oop_load(p); 6637 do_oop(obj); 6638 } 6639 6640 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6641 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6642 6643 // Upon stack overflow, we discard (part of) the stack, 6644 // remembering the least address amongst those discarded 6645 // in CMSCollector's _restart_address. 6646 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6647 // Remember the least grey address discarded 6648 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6649 _collector->lower_restart_addr(ra); 6650 _mark_stack->reset(); // discard stack contents 6651 _mark_stack->expand(); // expand the stack if possible 6652 } 6653 6654 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6655 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6656 HeapWord* addr = (HeapWord*)obj; 6657 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6658 // Oop lies in _span and isn't yet grey or black 6659 _verification_bm->mark(addr); // now grey 6660 if (!_cms_bm->isMarked(addr)) { 6661 Log(gc, verify) log; 6662 ResourceMark rm; 6663 LogStream ls(log.error()); 6664 oop(addr)->print_on(&ls); 6665 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6666 fatal("... aborting"); 6667 } 6668 6669 if (!_mark_stack->push(obj)) { // stack overflow 6670 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6671 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6672 handle_stack_overflow(addr); 6673 } 6674 // anything including and to the right of _finger 6675 // will be scanned as we iterate over the remainder of the 6676 // bit map 6677 } 6678 } 6679 6680 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6681 MemRegion span, 6682 CMSBitMap* bitMap, CMSMarkStack* markStack, 6683 HeapWord* finger, MarkFromRootsClosure* parent) : 6684 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6685 _collector(collector), 6686 _span(span), 6687 _bitMap(bitMap), 6688 _markStack(markStack), 6689 _finger(finger), 6690 _parent(parent) 6691 { } 6692 6693 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6694 MemRegion span, 6695 CMSBitMap* bit_map, 6696 OopTaskQueue* work_queue, 6697 CMSMarkStack* overflow_stack, 6698 HeapWord* finger, 6699 HeapWord* volatile* global_finger_addr, 6700 ParMarkFromRootsClosure* parent) : 6701 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6702 _collector(collector), 6703 _whole_span(collector->_span), 6704 _span(span), 6705 _bit_map(bit_map), 6706 _work_queue(work_queue), 6707 _overflow_stack(overflow_stack), 6708 _finger(finger), 6709 _global_finger_addr(global_finger_addr), 6710 _parent(parent) 6711 { } 6712 6713 // Assumes thread-safe access by callers, who are 6714 // responsible for mutual exclusion. 6715 void CMSCollector::lower_restart_addr(HeapWord* low) { 6716 assert(_span.contains(low), "Out of bounds addr"); 6717 if (_restart_addr == NULL) { 6718 _restart_addr = low; 6719 } else { 6720 _restart_addr = MIN2(_restart_addr, low); 6721 } 6722 } 6723 6724 // Upon stack overflow, we discard (part of) the stack, 6725 // remembering the least address amongst those discarded 6726 // in CMSCollector's _restart_address. 6727 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6728 // Remember the least grey address discarded 6729 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6730 _collector->lower_restart_addr(ra); 6731 _markStack->reset(); // discard stack contents 6732 _markStack->expand(); // expand the stack if possible 6733 } 6734 6735 // Upon stack overflow, we discard (part of) the stack, 6736 // remembering the least address amongst those discarded 6737 // in CMSCollector's _restart_address. 6738 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6739 // We need to do this under a mutex to prevent other 6740 // workers from interfering with the work done below. 6741 MutexLockerEx ml(_overflow_stack->par_lock(), 6742 Mutex::_no_safepoint_check_flag); 6743 // Remember the least grey address discarded 6744 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6745 _collector->lower_restart_addr(ra); 6746 _overflow_stack->reset(); // discard stack contents 6747 _overflow_stack->expand(); // expand the stack if possible 6748 } 6749 6750 void PushOrMarkClosure::do_oop(oop obj) { 6751 // Ignore mark word because we are running concurrent with mutators. 6752 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6753 HeapWord* addr = (HeapWord*)obj; 6754 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6755 // Oop lies in _span and isn't yet grey or black 6756 _bitMap->mark(addr); // now grey 6757 if (addr < _finger) { 6758 // the bit map iteration has already either passed, or 6759 // sampled, this bit in the bit map; we'll need to 6760 // use the marking stack to scan this oop's oops. 6761 bool simulate_overflow = false; 6762 NOT_PRODUCT( 6763 if (CMSMarkStackOverflowALot && 6764 _collector->simulate_overflow()) { 6765 // simulate a stack overflow 6766 simulate_overflow = true; 6767 } 6768 ) 6769 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6770 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6771 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6772 handle_stack_overflow(addr); 6773 } 6774 } 6775 // anything including and to the right of _finger 6776 // will be scanned as we iterate over the remainder of the 6777 // bit map 6778 do_yield_check(); 6779 } 6780 } 6781 6782 void ParPushOrMarkClosure::do_oop(oop obj) { 6783 // Ignore mark word because we are running concurrent with mutators. 6784 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6785 HeapWord* addr = (HeapWord*)obj; 6786 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6787 // Oop lies in _span and isn't yet grey or black 6788 // We read the global_finger (volatile read) strictly after marking oop 6789 bool res = _bit_map->par_mark(addr); // now grey 6790 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6791 // Should we push this marked oop on our stack? 6792 // -- if someone else marked it, nothing to do 6793 // -- if target oop is above global finger nothing to do 6794 // -- if target oop is in chunk and above local finger 6795 // then nothing to do 6796 // -- else push on work queue 6797 if ( !res // someone else marked it, they will deal with it 6798 || (addr >= *gfa) // will be scanned in a later task 6799 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6800 return; 6801 } 6802 // the bit map iteration has already either passed, or 6803 // sampled, this bit in the bit map; we'll need to 6804 // use the marking stack to scan this oop's oops. 6805 bool simulate_overflow = false; 6806 NOT_PRODUCT( 6807 if (CMSMarkStackOverflowALot && 6808 _collector->simulate_overflow()) { 6809 // simulate a stack overflow 6810 simulate_overflow = true; 6811 } 6812 ) 6813 if (simulate_overflow || 6814 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6815 // stack overflow 6816 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6817 // We cannot assert that the overflow stack is full because 6818 // it may have been emptied since. 6819 assert(simulate_overflow || 6820 _work_queue->size() == _work_queue->max_elems(), 6821 "Else push should have succeeded"); 6822 handle_stack_overflow(addr); 6823 } 6824 do_yield_check(); 6825 } 6826 } 6827 6828 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6829 MemRegion span, 6830 ReferenceDiscoverer* rd, 6831 CMSBitMap* bit_map, 6832 CMSBitMap* mod_union_table, 6833 CMSMarkStack* mark_stack, 6834 bool concurrent_precleaning): 6835 MetadataVisitingOopIterateClosure(rd), 6836 _collector(collector), 6837 _span(span), 6838 _bit_map(bit_map), 6839 _mod_union_table(mod_union_table), 6840 _mark_stack(mark_stack), 6841 _concurrent_precleaning(concurrent_precleaning) 6842 { 6843 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6844 } 6845 6846 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6847 // the non-parallel version (the parallel version appears further below.) 6848 void PushAndMarkClosure::do_oop(oop obj) { 6849 // Ignore mark word verification. If during concurrent precleaning, 6850 // the object monitor may be locked. If during the checkpoint 6851 // phases, the object may already have been reached by a different 6852 // path and may be at the end of the global overflow list (so 6853 // the mark word may be NULL). 6854 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6855 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6856 HeapWord* addr = (HeapWord*)obj; 6857 // Check if oop points into the CMS generation 6858 // and is not marked 6859 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6860 // a white object ... 6861 _bit_map->mark(addr); // ... now grey 6862 // push on the marking stack (grey set) 6863 bool simulate_overflow = false; 6864 NOT_PRODUCT( 6865 if (CMSMarkStackOverflowALot && 6866 _collector->simulate_overflow()) { 6867 // simulate a stack overflow 6868 simulate_overflow = true; 6869 } 6870 ) 6871 if (simulate_overflow || !_mark_stack->push(obj)) { 6872 if (_concurrent_precleaning) { 6873 // During precleaning we can just dirty the appropriate card(s) 6874 // in the mod union table, thus ensuring that the object remains 6875 // in the grey set and continue. In the case of object arrays 6876 // we need to dirty all of the cards that the object spans, 6877 // since the rescan of object arrays will be limited to the 6878 // dirty cards. 6879 // Note that no one can be interfering with us in this action 6880 // of dirtying the mod union table, so no locking or atomics 6881 // are required. 6882 if (obj->is_objArray()) { 6883 size_t sz = obj->size(); 6884 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6885 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6886 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6887 _mod_union_table->mark_range(redirty_range); 6888 } else { 6889 _mod_union_table->mark(addr); 6890 } 6891 _collector->_ser_pmc_preclean_ovflw++; 6892 } else { 6893 // During the remark phase, we need to remember this oop 6894 // in the overflow list. 6895 _collector->push_on_overflow_list(obj); 6896 _collector->_ser_pmc_remark_ovflw++; 6897 } 6898 } 6899 } 6900 } 6901 6902 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6903 MemRegion span, 6904 ReferenceDiscoverer* rd, 6905 CMSBitMap* bit_map, 6906 OopTaskQueue* work_queue): 6907 MetadataVisitingOopIterateClosure(rd), 6908 _collector(collector), 6909 _span(span), 6910 _bit_map(bit_map), 6911 _work_queue(work_queue) 6912 { 6913 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6914 } 6915 6916 // Grey object rescan during second checkpoint phase -- 6917 // the parallel version. 6918 void ParPushAndMarkClosure::do_oop(oop obj) { 6919 // In the assert below, we ignore the mark word because 6920 // this oop may point to an already visited object that is 6921 // on the overflow stack (in which case the mark word has 6922 // been hijacked for chaining into the overflow stack -- 6923 // if this is the last object in the overflow stack then 6924 // its mark word will be NULL). Because this object may 6925 // have been subsequently popped off the global overflow 6926 // stack, and the mark word possibly restored to the prototypical 6927 // value, by the time we get to examined this failing assert in 6928 // the debugger, is_oop_or_null(false) may subsequently start 6929 // to hold. 6930 assert(oopDesc::is_oop_or_null(obj, true), 6931 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6932 HeapWord* addr = (HeapWord*)obj; 6933 // Check if oop points into the CMS generation 6934 // and is not marked 6935 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6936 // a white object ... 6937 // If we manage to "claim" the object, by being the 6938 // first thread to mark it, then we push it on our 6939 // marking stack 6940 if (_bit_map->par_mark(addr)) { // ... now grey 6941 // push on work queue (grey set) 6942 bool simulate_overflow = false; 6943 NOT_PRODUCT( 6944 if (CMSMarkStackOverflowALot && 6945 _collector->par_simulate_overflow()) { 6946 // simulate a stack overflow 6947 simulate_overflow = true; 6948 } 6949 ) 6950 if (simulate_overflow || !_work_queue->push(obj)) { 6951 _collector->par_push_on_overflow_list(obj); 6952 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6953 } 6954 } // Else, some other thread got there first 6955 } 6956 } 6957 6958 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6959 Mutex* bml = _collector->bitMapLock(); 6960 assert_lock_strong(bml); 6961 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6962 "CMS thread should hold CMS token"); 6963 6964 bml->unlock(); 6965 ConcurrentMarkSweepThread::desynchronize(true); 6966 6967 _collector->stopTimer(); 6968 _collector->incrementYields(); 6969 6970 // See the comment in coordinator_yield() 6971 for (unsigned i = 0; i < CMSYieldSleepCount && 6972 ConcurrentMarkSweepThread::should_yield() && 6973 !CMSCollector::foregroundGCIsActive(); ++i) { 6974 os::sleep(Thread::current(), 1, false); 6975 } 6976 6977 ConcurrentMarkSweepThread::synchronize(true); 6978 bml->lock(); 6979 6980 _collector->startTimer(); 6981 } 6982 6983 bool CMSPrecleanRefsYieldClosure::should_return() { 6984 if (ConcurrentMarkSweepThread::should_yield()) { 6985 do_yield_work(); 6986 } 6987 return _collector->foregroundGCIsActive(); 6988 } 6989 6990 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6991 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6992 "mr should be aligned to start at a card boundary"); 6993 // We'd like to assert: 6994 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6995 // "mr should be a range of cards"); 6996 // However, that would be too strong in one case -- the last 6997 // partition ends at _unallocated_block which, in general, can be 6998 // an arbitrary boundary, not necessarily card aligned. 6999 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7000 _space->object_iterate_mem(mr, &_scan_cl); 7001 } 7002 7003 SweepClosure::SweepClosure(CMSCollector* collector, 7004 ConcurrentMarkSweepGeneration* g, 7005 CMSBitMap* bitMap, bool should_yield) : 7006 _collector(collector), 7007 _g(g), 7008 _sp(g->cmsSpace()), 7009 _limit(_sp->sweep_limit()), 7010 _freelistLock(_sp->freelistLock()), 7011 _bitMap(bitMap), 7012 _inFreeRange(false), // No free range at beginning of sweep 7013 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7014 _lastFreeRangeCoalesced(false), 7015 _yield(should_yield), 7016 _freeFinger(g->used_region().start()) 7017 { 7018 NOT_PRODUCT( 7019 _numObjectsFreed = 0; 7020 _numWordsFreed = 0; 7021 _numObjectsLive = 0; 7022 _numWordsLive = 0; 7023 _numObjectsAlreadyFree = 0; 7024 _numWordsAlreadyFree = 0; 7025 _last_fc = NULL; 7026 7027 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7028 _sp->dictionary()->initialize_dict_returned_bytes(); 7029 ) 7030 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7031 "sweep _limit out of bounds"); 7032 log_develop_trace(gc, sweep)("===================="); 7033 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7034 } 7035 7036 void SweepClosure::print_on(outputStream* st) const { 7037 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7038 p2i(_sp->bottom()), p2i(_sp->end())); 7039 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7040 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7041 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7042 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7043 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7044 } 7045 7046 #ifndef PRODUCT 7047 // Assertion checking only: no useful work in product mode -- 7048 // however, if any of the flags below become product flags, 7049 // you may need to review this code to see if it needs to be 7050 // enabled in product mode. 7051 SweepClosure::~SweepClosure() { 7052 assert_lock_strong(_freelistLock); 7053 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7054 "sweep _limit out of bounds"); 7055 if (inFreeRange()) { 7056 Log(gc, sweep) log; 7057 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7058 ResourceMark rm; 7059 LogStream ls(log.error()); 7060 print_on(&ls); 7061 ShouldNotReachHere(); 7062 } 7063 7064 if (log_is_enabled(Debug, gc, sweep)) { 7065 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7066 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7067 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7068 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7069 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7070 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7071 } 7072 7073 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7074 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7075 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7076 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7077 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7078 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7079 } 7080 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7081 log_develop_trace(gc, sweep)("================"); 7082 } 7083 #endif // PRODUCT 7084 7085 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7086 bool freeRangeInFreeLists) { 7087 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7088 p2i(freeFinger), freeRangeInFreeLists); 7089 assert(!inFreeRange(), "Trampling existing free range"); 7090 set_inFreeRange(true); 7091 set_lastFreeRangeCoalesced(false); 7092 7093 set_freeFinger(freeFinger); 7094 set_freeRangeInFreeLists(freeRangeInFreeLists); 7095 if (CMSTestInFreeList) { 7096 if (freeRangeInFreeLists) { 7097 FreeChunk* fc = (FreeChunk*) freeFinger; 7098 assert(fc->is_free(), "A chunk on the free list should be free."); 7099 assert(fc->size() > 0, "Free range should have a size"); 7100 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7101 } 7102 } 7103 } 7104 7105 // Note that the sweeper runs concurrently with mutators. Thus, 7106 // it is possible for direct allocation in this generation to happen 7107 // in the middle of the sweep. Note that the sweeper also coalesces 7108 // contiguous free blocks. Thus, unless the sweeper and the allocator 7109 // synchronize appropriately freshly allocated blocks may get swept up. 7110 // This is accomplished by the sweeper locking the free lists while 7111 // it is sweeping. Thus blocks that are determined to be free are 7112 // indeed free. There is however one additional complication: 7113 // blocks that have been allocated since the final checkpoint and 7114 // mark, will not have been marked and so would be treated as 7115 // unreachable and swept up. To prevent this, the allocator marks 7116 // the bit map when allocating during the sweep phase. This leads, 7117 // however, to a further complication -- objects may have been allocated 7118 // but not yet initialized -- in the sense that the header isn't yet 7119 // installed. The sweeper can not then determine the size of the block 7120 // in order to skip over it. To deal with this case, we use a technique 7121 // (due to Printezis) to encode such uninitialized block sizes in the 7122 // bit map. Since the bit map uses a bit per every HeapWord, but the 7123 // CMS generation has a minimum object size of 3 HeapWords, it follows 7124 // that "normal marks" won't be adjacent in the bit map (there will 7125 // always be at least two 0 bits between successive 1 bits). We make use 7126 // of these "unused" bits to represent uninitialized blocks -- the bit 7127 // corresponding to the start of the uninitialized object and the next 7128 // bit are both set. Finally, a 1 bit marks the end of the object that 7129 // started with the two consecutive 1 bits to indicate its potentially 7130 // uninitialized state. 7131 7132 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7133 FreeChunk* fc = (FreeChunk*)addr; 7134 size_t res; 7135 7136 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7137 // than "addr == _limit" because although _limit was a block boundary when 7138 // we started the sweep, it may no longer be one because heap expansion 7139 // may have caused us to coalesce the block ending at the address _limit 7140 // with a newly expanded chunk (this happens when _limit was set to the 7141 // previous _end of the space), so we may have stepped past _limit: 7142 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7143 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7144 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7145 "sweep _limit out of bounds"); 7146 assert(addr < _sp->end(), "addr out of bounds"); 7147 // Flush any free range we might be holding as a single 7148 // coalesced chunk to the appropriate free list. 7149 if (inFreeRange()) { 7150 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7151 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7152 flush_cur_free_chunk(freeFinger(), 7153 pointer_delta(addr, freeFinger())); 7154 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7155 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7156 lastFreeRangeCoalesced() ? 1 : 0); 7157 } 7158 7159 // help the iterator loop finish 7160 return pointer_delta(_sp->end(), addr); 7161 } 7162 7163 assert(addr < _limit, "sweep invariant"); 7164 // check if we should yield 7165 do_yield_check(addr); 7166 if (fc->is_free()) { 7167 // Chunk that is already free 7168 res = fc->size(); 7169 do_already_free_chunk(fc); 7170 debug_only(_sp->verifyFreeLists()); 7171 // If we flush the chunk at hand in lookahead_and_flush() 7172 // and it's coalesced with a preceding chunk, then the 7173 // process of "mangling" the payload of the coalesced block 7174 // will cause erasure of the size information from the 7175 // (erstwhile) header of all the coalesced blocks but the 7176 // first, so the first disjunct in the assert will not hold 7177 // in that specific case (in which case the second disjunct 7178 // will hold). 7179 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7180 "Otherwise the size info doesn't change at this step"); 7181 NOT_PRODUCT( 7182 _numObjectsAlreadyFree++; 7183 _numWordsAlreadyFree += res; 7184 ) 7185 NOT_PRODUCT(_last_fc = fc;) 7186 } else if (!_bitMap->isMarked(addr)) { 7187 // Chunk is fresh garbage 7188 res = do_garbage_chunk(fc); 7189 debug_only(_sp->verifyFreeLists()); 7190 NOT_PRODUCT( 7191 _numObjectsFreed++; 7192 _numWordsFreed += res; 7193 ) 7194 } else { 7195 // Chunk that is alive. 7196 res = do_live_chunk(fc); 7197 debug_only(_sp->verifyFreeLists()); 7198 NOT_PRODUCT( 7199 _numObjectsLive++; 7200 _numWordsLive += res; 7201 ) 7202 } 7203 return res; 7204 } 7205 7206 // For the smart allocation, record following 7207 // split deaths - a free chunk is removed from its free list because 7208 // it is being split into two or more chunks. 7209 // split birth - a free chunk is being added to its free list because 7210 // a larger free chunk has been split and resulted in this free chunk. 7211 // coal death - a free chunk is being removed from its free list because 7212 // it is being coalesced into a large free chunk. 7213 // coal birth - a free chunk is being added to its free list because 7214 // it was created when two or more free chunks where coalesced into 7215 // this free chunk. 7216 // 7217 // These statistics are used to determine the desired number of free 7218 // chunks of a given size. The desired number is chosen to be relative 7219 // to the end of a CMS sweep. The desired number at the end of a sweep 7220 // is the 7221 // count-at-end-of-previous-sweep (an amount that was enough) 7222 // - count-at-beginning-of-current-sweep (the excess) 7223 // + split-births (gains in this size during interval) 7224 // - split-deaths (demands on this size during interval) 7225 // where the interval is from the end of one sweep to the end of the 7226 // next. 7227 // 7228 // When sweeping the sweeper maintains an accumulated chunk which is 7229 // the chunk that is made up of chunks that have been coalesced. That 7230 // will be termed the left-hand chunk. A new chunk of garbage that 7231 // is being considered for coalescing will be referred to as the 7232 // right-hand chunk. 7233 // 7234 // When making a decision on whether to coalesce a right-hand chunk with 7235 // the current left-hand chunk, the current count vs. the desired count 7236 // of the left-hand chunk is considered. Also if the right-hand chunk 7237 // is near the large chunk at the end of the heap (see 7238 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7239 // left-hand chunk is coalesced. 7240 // 7241 // When making a decision about whether to split a chunk, the desired count 7242 // vs. the current count of the candidate to be split is also considered. 7243 // If the candidate is underpopulated (currently fewer chunks than desired) 7244 // a chunk of an overpopulated (currently more chunks than desired) size may 7245 // be chosen. The "hint" associated with a free list, if non-null, points 7246 // to a free list which may be overpopulated. 7247 // 7248 7249 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7250 const size_t size = fc->size(); 7251 // Chunks that cannot be coalesced are not in the 7252 // free lists. 7253 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7254 assert(_sp->verify_chunk_in_free_list(fc), 7255 "free chunk should be in free lists"); 7256 } 7257 // a chunk that is already free, should not have been 7258 // marked in the bit map 7259 HeapWord* const addr = (HeapWord*) fc; 7260 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7261 // Verify that the bit map has no bits marked between 7262 // addr and purported end of this block. 7263 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7264 7265 // Some chunks cannot be coalesced under any circumstances. 7266 // See the definition of cantCoalesce(). 7267 if (!fc->cantCoalesce()) { 7268 // This chunk can potentially be coalesced. 7269 // All the work is done in 7270 do_post_free_or_garbage_chunk(fc, size); 7271 // Note that if the chunk is not coalescable (the else arm 7272 // below), we unconditionally flush, without needing to do 7273 // a "lookahead," as we do below. 7274 if (inFreeRange()) lookahead_and_flush(fc, size); 7275 } else { 7276 // Code path common to both original and adaptive free lists. 7277 7278 // cant coalesce with previous block; this should be treated 7279 // as the end of a free run if any 7280 if (inFreeRange()) { 7281 // we kicked some butt; time to pick up the garbage 7282 assert(freeFinger() < addr, "freeFinger points too high"); 7283 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7284 } 7285 // else, nothing to do, just continue 7286 } 7287 } 7288 7289 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7290 // This is a chunk of garbage. It is not in any free list. 7291 // Add it to a free list or let it possibly be coalesced into 7292 // a larger chunk. 7293 HeapWord* const addr = (HeapWord*) fc; 7294 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7295 7296 // Verify that the bit map has no bits marked between 7297 // addr and purported end of just dead object. 7298 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7299 do_post_free_or_garbage_chunk(fc, size); 7300 7301 assert(_limit >= addr + size, 7302 "A freshly garbage chunk can't possibly straddle over _limit"); 7303 if (inFreeRange()) lookahead_and_flush(fc, size); 7304 return size; 7305 } 7306 7307 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7308 HeapWord* addr = (HeapWord*) fc; 7309 // The sweeper has just found a live object. Return any accumulated 7310 // left hand chunk to the free lists. 7311 if (inFreeRange()) { 7312 assert(freeFinger() < addr, "freeFinger points too high"); 7313 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7314 } 7315 7316 // This object is live: we'd normally expect this to be 7317 // an oop, and like to assert the following: 7318 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7319 // However, as we commented above, this may be an object whose 7320 // header hasn't yet been initialized. 7321 size_t size; 7322 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7323 if (_bitMap->isMarked(addr + 1)) { 7324 // Determine the size from the bit map, rather than trying to 7325 // compute it from the object header. 7326 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7327 size = pointer_delta(nextOneAddr + 1, addr); 7328 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7329 "alignment problem"); 7330 7331 #ifdef ASSERT 7332 if (oop(addr)->klass_or_null_acquire() != NULL) { 7333 // Ignore mark word because we are running concurrent with mutators 7334 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7335 assert(size == 7336 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7337 "P-mark and computed size do not agree"); 7338 } 7339 #endif 7340 7341 } else { 7342 // This should be an initialized object that's alive. 7343 assert(oop(addr)->klass_or_null_acquire() != NULL, 7344 "Should be an initialized object"); 7345 // Ignore mark word because we are running concurrent with mutators 7346 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7347 // Verify that the bit map has no bits marked between 7348 // addr and purported end of this block. 7349 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7350 assert(size >= 3, "Necessary for Printezis marks to work"); 7351 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7352 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7353 } 7354 return size; 7355 } 7356 7357 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7358 size_t chunkSize) { 7359 // do_post_free_or_garbage_chunk() should only be called in the case 7360 // of the adaptive free list allocator. 7361 const bool fcInFreeLists = fc->is_free(); 7362 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7363 if (CMSTestInFreeList && fcInFreeLists) { 7364 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7365 } 7366 7367 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7368 7369 HeapWord* const fc_addr = (HeapWord*) fc; 7370 7371 bool coalesce = false; 7372 const size_t left = pointer_delta(fc_addr, freeFinger()); 7373 const size_t right = chunkSize; 7374 switch (FLSCoalescePolicy) { 7375 // numeric value forms a coalition aggressiveness metric 7376 case 0: { // never coalesce 7377 coalesce = false; 7378 break; 7379 } 7380 case 1: { // coalesce if left & right chunks on overpopulated lists 7381 coalesce = _sp->coalOverPopulated(left) && 7382 _sp->coalOverPopulated(right); 7383 break; 7384 } 7385 case 2: { // coalesce if left chunk on overpopulated list (default) 7386 coalesce = _sp->coalOverPopulated(left); 7387 break; 7388 } 7389 case 3: { // coalesce if left OR right chunk on overpopulated list 7390 coalesce = _sp->coalOverPopulated(left) || 7391 _sp->coalOverPopulated(right); 7392 break; 7393 } 7394 case 4: { // always coalesce 7395 coalesce = true; 7396 break; 7397 } 7398 default: 7399 ShouldNotReachHere(); 7400 } 7401 7402 // Should the current free range be coalesced? 7403 // If the chunk is in a free range and either we decided to coalesce above 7404 // or the chunk is near the large block at the end of the heap 7405 // (isNearLargestChunk() returns true), then coalesce this chunk. 7406 const bool doCoalesce = inFreeRange() 7407 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7408 if (doCoalesce) { 7409 // Coalesce the current free range on the left with the new 7410 // chunk on the right. If either is on a free list, 7411 // it must be removed from the list and stashed in the closure. 7412 if (freeRangeInFreeLists()) { 7413 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7414 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7415 "Size of free range is inconsistent with chunk size."); 7416 if (CMSTestInFreeList) { 7417 assert(_sp->verify_chunk_in_free_list(ffc), 7418 "Chunk is not in free lists"); 7419 } 7420 _sp->coalDeath(ffc->size()); 7421 _sp->removeFreeChunkFromFreeLists(ffc); 7422 set_freeRangeInFreeLists(false); 7423 } 7424 if (fcInFreeLists) { 7425 _sp->coalDeath(chunkSize); 7426 assert(fc->size() == chunkSize, 7427 "The chunk has the wrong size or is not in the free lists"); 7428 _sp->removeFreeChunkFromFreeLists(fc); 7429 } 7430 set_lastFreeRangeCoalesced(true); 7431 print_free_block_coalesced(fc); 7432 } else { // not in a free range and/or should not coalesce 7433 // Return the current free range and start a new one. 7434 if (inFreeRange()) { 7435 // In a free range but cannot coalesce with the right hand chunk. 7436 // Put the current free range into the free lists. 7437 flush_cur_free_chunk(freeFinger(), 7438 pointer_delta(fc_addr, freeFinger())); 7439 } 7440 // Set up for new free range. Pass along whether the right hand 7441 // chunk is in the free lists. 7442 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7443 } 7444 } 7445 7446 // Lookahead flush: 7447 // If we are tracking a free range, and this is the last chunk that 7448 // we'll look at because its end crosses past _limit, we'll preemptively 7449 // flush it along with any free range we may be holding on to. Note that 7450 // this can be the case only for an already free or freshly garbage 7451 // chunk. If this block is an object, it can never straddle 7452 // over _limit. The "straddling" occurs when _limit is set at 7453 // the previous end of the space when this cycle started, and 7454 // a subsequent heap expansion caused the previously co-terminal 7455 // free block to be coalesced with the newly expanded portion, 7456 // thus rendering _limit a non-block-boundary making it dangerous 7457 // for the sweeper to step over and examine. 7458 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7459 assert(inFreeRange(), "Should only be called if currently in a free range."); 7460 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7461 assert(_sp->used_region().contains(eob - 1), 7462 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7463 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7464 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7465 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7466 if (eob >= _limit) { 7467 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7468 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7469 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7470 "[" PTR_FORMAT "," PTR_FORMAT ")", 7471 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7472 // Return the storage we are tracking back into the free lists. 7473 log_develop_trace(gc, sweep)("Flushing ... "); 7474 assert(freeFinger() < eob, "Error"); 7475 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7476 } 7477 } 7478 7479 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7480 assert(inFreeRange(), "Should only be called if currently in a free range."); 7481 assert(size > 0, 7482 "A zero sized chunk cannot be added to the free lists."); 7483 if (!freeRangeInFreeLists()) { 7484 if (CMSTestInFreeList) { 7485 FreeChunk* fc = (FreeChunk*) chunk; 7486 fc->set_size(size); 7487 assert(!_sp->verify_chunk_in_free_list(fc), 7488 "chunk should not be in free lists yet"); 7489 } 7490 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7491 // A new free range is going to be starting. The current 7492 // free range has not been added to the free lists yet or 7493 // was removed so add it back. 7494 // If the current free range was coalesced, then the death 7495 // of the free range was recorded. Record a birth now. 7496 if (lastFreeRangeCoalesced()) { 7497 _sp->coalBirth(size); 7498 } 7499 _sp->addChunkAndRepairOffsetTable(chunk, size, 7500 lastFreeRangeCoalesced()); 7501 } else { 7502 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7503 } 7504 set_inFreeRange(false); 7505 set_freeRangeInFreeLists(false); 7506 } 7507 7508 // We take a break if we've been at this for a while, 7509 // so as to avoid monopolizing the locks involved. 7510 void SweepClosure::do_yield_work(HeapWord* addr) { 7511 // Return current free chunk being used for coalescing (if any) 7512 // to the appropriate freelist. After yielding, the next 7513 // free block encountered will start a coalescing range of 7514 // free blocks. If the next free block is adjacent to the 7515 // chunk just flushed, they will need to wait for the next 7516 // sweep to be coalesced. 7517 if (inFreeRange()) { 7518 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7519 } 7520 7521 // First give up the locks, then yield, then re-lock. 7522 // We should probably use a constructor/destructor idiom to 7523 // do this unlock/lock or modify the MutexUnlocker class to 7524 // serve our purpose. XXX 7525 assert_lock_strong(_bitMap->lock()); 7526 assert_lock_strong(_freelistLock); 7527 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7528 "CMS thread should hold CMS token"); 7529 _bitMap->lock()->unlock(); 7530 _freelistLock->unlock(); 7531 ConcurrentMarkSweepThread::desynchronize(true); 7532 _collector->stopTimer(); 7533 _collector->incrementYields(); 7534 7535 // See the comment in coordinator_yield() 7536 for (unsigned i = 0; i < CMSYieldSleepCount && 7537 ConcurrentMarkSweepThread::should_yield() && 7538 !CMSCollector::foregroundGCIsActive(); ++i) { 7539 os::sleep(Thread::current(), 1, false); 7540 } 7541 7542 ConcurrentMarkSweepThread::synchronize(true); 7543 _freelistLock->lock(); 7544 _bitMap->lock()->lock_without_safepoint_check(); 7545 _collector->startTimer(); 7546 } 7547 7548 #ifndef PRODUCT 7549 // This is actually very useful in a product build if it can 7550 // be called from the debugger. Compile it into the product 7551 // as needed. 7552 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7553 return debug_cms_space->verify_chunk_in_free_list(fc); 7554 } 7555 #endif 7556 7557 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7558 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7559 p2i(fc), fc->size()); 7560 } 7561 7562 // CMSIsAliveClosure 7563 bool CMSIsAliveClosure::do_object_b(oop obj) { 7564 HeapWord* addr = (HeapWord*)obj; 7565 return addr != NULL && 7566 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7567 } 7568 7569 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7570 MemRegion span, 7571 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7572 bool cpc): 7573 _collector(collector), 7574 _span(span), 7575 _mark_stack(mark_stack), 7576 _bit_map(bit_map), 7577 _concurrent_precleaning(cpc) { 7578 assert(!_span.is_empty(), "Empty span could spell trouble"); 7579 } 7580 7581 7582 // CMSKeepAliveClosure: the serial version 7583 void CMSKeepAliveClosure::do_oop(oop obj) { 7584 HeapWord* addr = (HeapWord*)obj; 7585 if (_span.contains(addr) && 7586 !_bit_map->isMarked(addr)) { 7587 _bit_map->mark(addr); 7588 bool simulate_overflow = false; 7589 NOT_PRODUCT( 7590 if (CMSMarkStackOverflowALot && 7591 _collector->simulate_overflow()) { 7592 // simulate a stack overflow 7593 simulate_overflow = true; 7594 } 7595 ) 7596 if (simulate_overflow || !_mark_stack->push(obj)) { 7597 if (_concurrent_precleaning) { 7598 // We dirty the overflown object and let the remark 7599 // phase deal with it. 7600 assert(_collector->overflow_list_is_empty(), "Error"); 7601 // In the case of object arrays, we need to dirty all of 7602 // the cards that the object spans. No locking or atomics 7603 // are needed since no one else can be mutating the mod union 7604 // table. 7605 if (obj->is_objArray()) { 7606 size_t sz = obj->size(); 7607 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7608 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7609 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7610 _collector->_modUnionTable.mark_range(redirty_range); 7611 } else { 7612 _collector->_modUnionTable.mark(addr); 7613 } 7614 _collector->_ser_kac_preclean_ovflw++; 7615 } else { 7616 _collector->push_on_overflow_list(obj); 7617 _collector->_ser_kac_ovflw++; 7618 } 7619 } 7620 } 7621 } 7622 7623 // CMSParKeepAliveClosure: a parallel version of the above. 7624 // The work queues are private to each closure (thread), 7625 // but (may be) available for stealing by other threads. 7626 void CMSParKeepAliveClosure::do_oop(oop obj) { 7627 HeapWord* addr = (HeapWord*)obj; 7628 if (_span.contains(addr) && 7629 !_bit_map->isMarked(addr)) { 7630 // In general, during recursive tracing, several threads 7631 // may be concurrently getting here; the first one to 7632 // "tag" it, claims it. 7633 if (_bit_map->par_mark(addr)) { 7634 bool res = _work_queue->push(obj); 7635 assert(res, "Low water mark should be much less than capacity"); 7636 // Do a recursive trim in the hope that this will keep 7637 // stack usage lower, but leave some oops for potential stealers 7638 trim_queue(_low_water_mark); 7639 } // Else, another thread got there first 7640 } 7641 } 7642 7643 void CMSParKeepAliveClosure::trim_queue(uint max) { 7644 while (_work_queue->size() > max) { 7645 oop new_oop; 7646 if (_work_queue->pop_local(new_oop)) { 7647 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7648 assert(_bit_map->isMarked((HeapWord*)new_oop), 7649 "no white objects on this stack!"); 7650 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7651 // iterate over the oops in this oop, marking and pushing 7652 // the ones in CMS heap (i.e. in _span). 7653 new_oop->oop_iterate(&_mark_and_push); 7654 } 7655 } 7656 } 7657 7658 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7659 CMSCollector* collector, 7660 MemRegion span, CMSBitMap* bit_map, 7661 OopTaskQueue* work_queue): 7662 _collector(collector), 7663 _span(span), 7664 _work_queue(work_queue), 7665 _bit_map(bit_map) { } 7666 7667 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7668 HeapWord* addr = (HeapWord*)obj; 7669 if (_span.contains(addr) && 7670 !_bit_map->isMarked(addr)) { 7671 if (_bit_map->par_mark(addr)) { 7672 bool simulate_overflow = false; 7673 NOT_PRODUCT( 7674 if (CMSMarkStackOverflowALot && 7675 _collector->par_simulate_overflow()) { 7676 // simulate a stack overflow 7677 simulate_overflow = true; 7678 } 7679 ) 7680 if (simulate_overflow || !_work_queue->push(obj)) { 7681 _collector->par_push_on_overflow_list(obj); 7682 _collector->_par_kac_ovflw++; 7683 } 7684 } // Else another thread got there already 7685 } 7686 } 7687 7688 ////////////////////////////////////////////////////////////////// 7689 // CMSExpansionCause ///////////////////////////// 7690 ////////////////////////////////////////////////////////////////// 7691 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7692 switch (cause) { 7693 case _no_expansion: 7694 return "No expansion"; 7695 case _satisfy_free_ratio: 7696 return "Free ratio"; 7697 case _satisfy_promotion: 7698 return "Satisfy promotion"; 7699 case _satisfy_allocation: 7700 return "allocation"; 7701 case _allocate_par_lab: 7702 return "Par LAB"; 7703 case _allocate_par_spooling_space: 7704 return "Par Spooling Space"; 7705 case _adaptive_size_policy: 7706 return "Ergonomics"; 7707 default: 7708 return "unknown"; 7709 } 7710 } 7711 7712 void CMSDrainMarkingStackClosure::do_void() { 7713 // the max number to take from overflow list at a time 7714 const size_t num = _mark_stack->capacity()/4; 7715 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7716 "Overflow list should be NULL during concurrent phases"); 7717 while (!_mark_stack->isEmpty() || 7718 // if stack is empty, check the overflow list 7719 _collector->take_from_overflow_list(num, _mark_stack)) { 7720 oop obj = _mark_stack->pop(); 7721 HeapWord* addr = (HeapWord*)obj; 7722 assert(_span.contains(addr), "Should be within span"); 7723 assert(_bit_map->isMarked(addr), "Should be marked"); 7724 assert(oopDesc::is_oop(obj), "Should be an oop"); 7725 obj->oop_iterate(_keep_alive); 7726 } 7727 } 7728 7729 void CMSParDrainMarkingStackClosure::do_void() { 7730 // drain queue 7731 trim_queue(0); 7732 } 7733 7734 // Trim our work_queue so its length is below max at return 7735 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7736 while (_work_queue->size() > max) { 7737 oop new_oop; 7738 if (_work_queue->pop_local(new_oop)) { 7739 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7740 assert(_bit_map->isMarked((HeapWord*)new_oop), 7741 "no white objects on this stack!"); 7742 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7743 // iterate over the oops in this oop, marking and pushing 7744 // the ones in CMS heap (i.e. in _span). 7745 new_oop->oop_iterate(&_mark_and_push); 7746 } 7747 } 7748 } 7749 7750 //////////////////////////////////////////////////////////////////// 7751 // Support for Marking Stack Overflow list handling and related code 7752 //////////////////////////////////////////////////////////////////// 7753 // Much of the following code is similar in shape and spirit to the 7754 // code used in ParNewGC. We should try and share that code 7755 // as much as possible in the future. 7756 7757 #ifndef PRODUCT 7758 // Debugging support for CMSStackOverflowALot 7759 7760 // It's OK to call this multi-threaded; the worst thing 7761 // that can happen is that we'll get a bunch of closely 7762 // spaced simulated overflows, but that's OK, in fact 7763 // probably good as it would exercise the overflow code 7764 // under contention. 7765 bool CMSCollector::simulate_overflow() { 7766 if (_overflow_counter-- <= 0) { // just being defensive 7767 _overflow_counter = CMSMarkStackOverflowInterval; 7768 return true; 7769 } else { 7770 return false; 7771 } 7772 } 7773 7774 bool CMSCollector::par_simulate_overflow() { 7775 return simulate_overflow(); 7776 } 7777 #endif 7778 7779 // Single-threaded 7780 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7781 assert(stack->isEmpty(), "Expected precondition"); 7782 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7783 size_t i = num; 7784 oop cur = _overflow_list; 7785 const markOop proto = markOopDesc::prototype(); 7786 NOT_PRODUCT(ssize_t n = 0;) 7787 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7788 next = oop(cur->mark_raw()); 7789 cur->set_mark_raw(proto); // until proven otherwise 7790 assert(oopDesc::is_oop(cur), "Should be an oop"); 7791 bool res = stack->push(cur); 7792 assert(res, "Bit off more than can chew?"); 7793 NOT_PRODUCT(n++;) 7794 } 7795 _overflow_list = cur; 7796 #ifndef PRODUCT 7797 assert(_num_par_pushes >= n, "Too many pops?"); 7798 _num_par_pushes -=n; 7799 #endif 7800 return !stack->isEmpty(); 7801 } 7802 7803 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7804 // (MT-safe) Get a prefix of at most "num" from the list. 7805 // The overflow list is chained through the mark word of 7806 // each object in the list. We fetch the entire list, 7807 // break off a prefix of the right size and return the 7808 // remainder. If other threads try to take objects from 7809 // the overflow list at that time, they will wait for 7810 // some time to see if data becomes available. If (and 7811 // only if) another thread places one or more object(s) 7812 // on the global list before we have returned the suffix 7813 // to the global list, we will walk down our local list 7814 // to find its end and append the global list to 7815 // our suffix before returning it. This suffix walk can 7816 // prove to be expensive (quadratic in the amount of traffic) 7817 // when there are many objects in the overflow list and 7818 // there is much producer-consumer contention on the list. 7819 // *NOTE*: The overflow list manipulation code here and 7820 // in ParNewGeneration:: are very similar in shape, 7821 // except that in the ParNew case we use the old (from/eden) 7822 // copy of the object to thread the list via its klass word. 7823 // Because of the common code, if you make any changes in 7824 // the code below, please check the ParNew version to see if 7825 // similar changes might be needed. 7826 // CR 6797058 has been filed to consolidate the common code. 7827 bool CMSCollector::par_take_from_overflow_list(size_t num, 7828 OopTaskQueue* work_q, 7829 int no_of_gc_threads) { 7830 assert(work_q->size() == 0, "First empty local work queue"); 7831 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7832 if (_overflow_list == NULL) { 7833 return false; 7834 } 7835 // Grab the entire list; we'll put back a suffix 7836 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7837 Thread* tid = Thread::current(); 7838 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7839 // set to ParallelGCThreads. 7840 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7841 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7842 // If the list is busy, we spin for a short while, 7843 // sleeping between attempts to get the list. 7844 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7845 os::sleep(tid, sleep_time_millis, false); 7846 if (_overflow_list == NULL) { 7847 // Nothing left to take 7848 return false; 7849 } else if (_overflow_list != BUSY) { 7850 // Try and grab the prefix 7851 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7852 } 7853 } 7854 // If the list was found to be empty, or we spun long 7855 // enough, we give up and return empty-handed. If we leave 7856 // the list in the BUSY state below, it must be the case that 7857 // some other thread holds the overflow list and will set it 7858 // to a non-BUSY state in the future. 7859 if (prefix == NULL || prefix == BUSY) { 7860 // Nothing to take or waited long enough 7861 if (prefix == NULL) { 7862 // Write back the NULL in case we overwrote it with BUSY above 7863 // and it is still the same value. 7864 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7865 } 7866 return false; 7867 } 7868 assert(prefix != NULL && prefix != BUSY, "Error"); 7869 size_t i = num; 7870 oop cur = prefix; 7871 // Walk down the first "num" objects, unless we reach the end. 7872 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7873 if (cur->mark_raw() == NULL) { 7874 // We have "num" or fewer elements in the list, so there 7875 // is nothing to return to the global list. 7876 // Write back the NULL in lieu of the BUSY we wrote 7877 // above, if it is still the same value. 7878 if (_overflow_list == BUSY) { 7879 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7880 } 7881 } else { 7882 // Chop off the suffix and return it to the global list. 7883 assert(cur->mark_raw() != BUSY, "Error"); 7884 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7885 cur->set_mark_raw(NULL); // break off suffix 7886 // It's possible that the list is still in the empty(busy) state 7887 // we left it in a short while ago; in that case we may be 7888 // able to place back the suffix without incurring the cost 7889 // of a walk down the list. 7890 oop observed_overflow_list = _overflow_list; 7891 oop cur_overflow_list = observed_overflow_list; 7892 bool attached = false; 7893 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7894 observed_overflow_list = 7895 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7896 if (cur_overflow_list == observed_overflow_list) { 7897 attached = true; 7898 break; 7899 } else cur_overflow_list = observed_overflow_list; 7900 } 7901 if (!attached) { 7902 // Too bad, someone else sneaked in (at least) an element; we'll need 7903 // to do a splice. Find tail of suffix so we can prepend suffix to global 7904 // list. 7905 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7906 oop suffix_tail = cur; 7907 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7908 "Tautology"); 7909 observed_overflow_list = _overflow_list; 7910 do { 7911 cur_overflow_list = observed_overflow_list; 7912 if (cur_overflow_list != BUSY) { 7913 // Do the splice ... 7914 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7915 } else { // cur_overflow_list == BUSY 7916 suffix_tail->set_mark_raw(NULL); 7917 } 7918 // ... and try to place spliced list back on overflow_list ... 7919 observed_overflow_list = 7920 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7921 } while (cur_overflow_list != observed_overflow_list); 7922 // ... until we have succeeded in doing so. 7923 } 7924 } 7925 7926 // Push the prefix elements on work_q 7927 assert(prefix != NULL, "control point invariant"); 7928 const markOop proto = markOopDesc::prototype(); 7929 oop next; 7930 NOT_PRODUCT(ssize_t n = 0;) 7931 for (cur = prefix; cur != NULL; cur = next) { 7932 next = oop(cur->mark_raw()); 7933 cur->set_mark_raw(proto); // until proven otherwise 7934 assert(oopDesc::is_oop(cur), "Should be an oop"); 7935 bool res = work_q->push(cur); 7936 assert(res, "Bit off more than we can chew?"); 7937 NOT_PRODUCT(n++;) 7938 } 7939 #ifndef PRODUCT 7940 assert(_num_par_pushes >= n, "Too many pops?"); 7941 Atomic::sub(n, &_num_par_pushes); 7942 #endif 7943 return true; 7944 } 7945 7946 // Single-threaded 7947 void CMSCollector::push_on_overflow_list(oop p) { 7948 NOT_PRODUCT(_num_par_pushes++;) 7949 assert(oopDesc::is_oop(p), "Not an oop"); 7950 preserve_mark_if_necessary(p); 7951 p->set_mark_raw((markOop)_overflow_list); 7952 _overflow_list = p; 7953 } 7954 7955 // Multi-threaded; use CAS to prepend to overflow list 7956 void CMSCollector::par_push_on_overflow_list(oop p) { 7957 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7958 assert(oopDesc::is_oop(p), "Not an oop"); 7959 par_preserve_mark_if_necessary(p); 7960 oop observed_overflow_list = _overflow_list; 7961 oop cur_overflow_list; 7962 do { 7963 cur_overflow_list = observed_overflow_list; 7964 if (cur_overflow_list != BUSY) { 7965 p->set_mark_raw(markOop(cur_overflow_list)); 7966 } else { 7967 p->set_mark_raw(NULL); 7968 } 7969 observed_overflow_list = 7970 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7971 } while (cur_overflow_list != observed_overflow_list); 7972 } 7973 #undef BUSY 7974 7975 // Single threaded 7976 // General Note on GrowableArray: pushes may silently fail 7977 // because we are (temporarily) out of C-heap for expanding 7978 // the stack. The problem is quite ubiquitous and affects 7979 // a lot of code in the JVM. The prudent thing for GrowableArray 7980 // to do (for now) is to exit with an error. However, that may 7981 // be too draconian in some cases because the caller may be 7982 // able to recover without much harm. For such cases, we 7983 // should probably introduce a "soft_push" method which returns 7984 // an indication of success or failure with the assumption that 7985 // the caller may be able to recover from a failure; code in 7986 // the VM can then be changed, incrementally, to deal with such 7987 // failures where possible, thus, incrementally hardening the VM 7988 // in such low resource situations. 7989 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7990 _preserved_oop_stack.push(p); 7991 _preserved_mark_stack.push(m); 7992 assert(m == p->mark_raw(), "Mark word changed"); 7993 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7994 "bijection"); 7995 } 7996 7997 // Single threaded 7998 void CMSCollector::preserve_mark_if_necessary(oop p) { 7999 markOop m = p->mark_raw(); 8000 if (m->must_be_preserved(p)) { 8001 preserve_mark_work(p, m); 8002 } 8003 } 8004 8005 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8006 markOop m = p->mark_raw(); 8007 if (m->must_be_preserved(p)) { 8008 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8009 // Even though we read the mark word without holding 8010 // the lock, we are assured that it will not change 8011 // because we "own" this oop, so no other thread can 8012 // be trying to push it on the overflow list; see 8013 // the assertion in preserve_mark_work() that checks 8014 // that m == p->mark_raw(). 8015 preserve_mark_work(p, m); 8016 } 8017 } 8018 8019 // We should be able to do this multi-threaded, 8020 // a chunk of stack being a task (this is 8021 // correct because each oop only ever appears 8022 // once in the overflow list. However, it's 8023 // not very easy to completely overlap this with 8024 // other operations, so will generally not be done 8025 // until all work's been completed. Because we 8026 // expect the preserved oop stack (set) to be small, 8027 // it's probably fine to do this single-threaded. 8028 // We can explore cleverer concurrent/overlapped/parallel 8029 // processing of preserved marks if we feel the 8030 // need for this in the future. Stack overflow should 8031 // be so rare in practice and, when it happens, its 8032 // effect on performance so great that this will 8033 // likely just be in the noise anyway. 8034 void CMSCollector::restore_preserved_marks_if_any() { 8035 assert(SafepointSynchronize::is_at_safepoint(), 8036 "world should be stopped"); 8037 assert(Thread::current()->is_ConcurrentGC_thread() || 8038 Thread::current()->is_VM_thread(), 8039 "should be single-threaded"); 8040 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8041 "bijection"); 8042 8043 while (!_preserved_oop_stack.is_empty()) { 8044 oop p = _preserved_oop_stack.pop(); 8045 assert(oopDesc::is_oop(p), "Should be an oop"); 8046 assert(_span.contains(p), "oop should be in _span"); 8047 assert(p->mark_raw() == markOopDesc::prototype(), 8048 "Set when taken from overflow list"); 8049 markOop m = _preserved_mark_stack.pop(); 8050 p->set_mark_raw(m); 8051 } 8052 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8053 "stacks were cleared above"); 8054 } 8055 8056 #ifndef PRODUCT 8057 bool CMSCollector::no_preserved_marks() const { 8058 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8059 } 8060 #endif 8061 8062 // Transfer some number of overflown objects to usual marking 8063 // stack. Return true if some objects were transferred. 8064 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8065 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8066 (size_t)ParGCDesiredObjsFromOverflowList); 8067 8068 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8069 assert(_collector->overflow_list_is_empty() || res, 8070 "If list is not empty, we should have taken something"); 8071 assert(!res || !_mark_stack->isEmpty(), 8072 "If we took something, it should now be on our stack"); 8073 return res; 8074 } 8075 8076 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8077 size_t res = _sp->block_size_no_stall(addr, _collector); 8078 if (_sp->block_is_obj(addr)) { 8079 if (_live_bit_map->isMarked(addr)) { 8080 // It can't have been dead in a previous cycle 8081 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8082 } else { 8083 _dead_bit_map->mark(addr); // mark the dead object 8084 } 8085 } 8086 // Could be 0, if the block size could not be computed without stalling. 8087 return res; 8088 } 8089 8090 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8091 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8092 switch (phase) { 8093 case CMSCollector::InitialMarking: 8094 initialize(manager /* GC manager */ , 8095 cause /* cause of the GC */, 8096 true /* allMemoryPoolsAffected */, 8097 true /* recordGCBeginTime */, 8098 true /* recordPreGCUsage */, 8099 false /* recordPeakUsage */, 8100 false /* recordPostGCusage */, 8101 true /* recordAccumulatedGCTime */, 8102 false /* recordGCEndTime */, 8103 false /* countCollection */ ); 8104 break; 8105 8106 case CMSCollector::FinalMarking: 8107 initialize(manager /* GC manager */ , 8108 cause /* cause of the GC */, 8109 true /* allMemoryPoolsAffected */, 8110 false /* recordGCBeginTime */, 8111 false /* recordPreGCUsage */, 8112 false /* recordPeakUsage */, 8113 false /* recordPostGCusage */, 8114 true /* recordAccumulatedGCTime */, 8115 false /* recordGCEndTime */, 8116 false /* countCollection */ ); 8117 break; 8118 8119 case CMSCollector::Sweeping: 8120 initialize(manager /* GC manager */ , 8121 cause /* cause of the GC */, 8122 true /* allMemoryPoolsAffected */, 8123 false /* recordGCBeginTime */, 8124 false /* recordPreGCUsage */, 8125 true /* recordPeakUsage */, 8126 true /* recordPostGCusage */, 8127 false /* recordAccumulatedGCTime */, 8128 true /* recordGCEndTime */, 8129 true /* countCollection */ ); 8130 break; 8131 8132 default: 8133 ShouldNotReachHere(); 8134 } 8135 }