1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/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/compactibleFreeListSpace.hpp" 36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 37 #include "gc/cms/concurrentMarkSweepThread.hpp" 38 #include "gc/cms/parNewGeneration.hpp" 39 #include "gc/cms/promotionInfo.inline.hpp" 40 #include "gc/cms/vmCMSOperations.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/referencePolicy.hpp" 59 #include "gc/shared/space.inline.hpp" 60 #include "gc/shared/strongRootsScope.hpp" 61 #include "gc/shared/taskqueue.inline.hpp" 62 #include "gc/shared/weakProcessor.hpp" 63 #include "logging/log.hpp" 64 #include "logging/logStream.hpp" 65 #include "memory/allocation.hpp" 66 #include "memory/binaryTreeDictionary.inline.hpp" 67 #include "memory/iterator.inline.hpp" 68 #include "memory/padded.hpp" 69 #include "memory/resourceArea.hpp" 70 #include "oops/access.inline.hpp" 71 #include "oops/oop.inline.hpp" 72 #include "prims/jvmtiExport.hpp" 73 #include "runtime/atomic.hpp" 74 #include "runtime/flags/flagSetting.hpp" 75 #include "runtime/globals_extension.hpp" 76 #include "runtime/handles.inline.hpp" 77 #include "runtime/java.hpp" 78 #include "runtime/orderAccess.inline.hpp" 79 #include "runtime/timer.hpp" 80 #include "runtime/vmThread.hpp" 81 #include "services/memoryService.hpp" 82 #include "services/runtimeService.hpp" 83 #include "utilities/align.hpp" 84 #include "utilities/stack.inline.hpp" 85 86 // statics 87 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 88 bool CMSCollector::_full_gc_requested = false; 89 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 90 91 ////////////////////////////////////////////////////////////////// 92 // In support of CMS/VM thread synchronization 93 ////////////////////////////////////////////////////////////////// 94 // We split use of the CGC_lock into 2 "levels". 95 // The low-level locking is of the usual CGC_lock monitor. We introduce 96 // a higher level "token" (hereafter "CMS token") built on top of the 97 // low level monitor (hereafter "CGC lock"). 98 // The token-passing protocol gives priority to the VM thread. The 99 // CMS-lock doesn't provide any fairness guarantees, but clients 100 // should ensure that it is only held for very short, bounded 101 // durations. 102 // 103 // When either of the CMS thread or the VM thread is involved in 104 // collection operations during which it does not want the other 105 // thread to interfere, it obtains the CMS token. 106 // 107 // If either thread tries to get the token while the other has 108 // it, that thread waits. However, if the VM thread and CMS thread 109 // both want the token, then the VM thread gets priority while the 110 // CMS thread waits. This ensures, for instance, that the "concurrent" 111 // phases of the CMS thread's work do not block out the VM thread 112 // for long periods of time as the CMS thread continues to hog 113 // the token. (See bug 4616232). 114 // 115 // The baton-passing functions are, however, controlled by the 116 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 117 // and here the low-level CMS lock, not the high level token, 118 // ensures mutual exclusion. 119 // 120 // Two important conditions that we have to satisfy: 121 // 1. if a thread does a low-level wait on the CMS lock, then it 122 // relinquishes the CMS token if it were holding that token 123 // when it acquired the low-level CMS lock. 124 // 2. any low-level notifications on the low-level lock 125 // should only be sent when a thread has relinquished the token. 126 // 127 // In the absence of either property, we'd have potential deadlock. 128 // 129 // We protect each of the CMS (concurrent and sequential) phases 130 // with the CMS _token_, not the CMS _lock_. 131 // 132 // The only code protected by CMS lock is the token acquisition code 133 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 134 // baton-passing code. 135 // 136 // Unfortunately, i couldn't come up with a good abstraction to factor and 137 // hide the naked CGC_lock manipulation in the baton-passing code 138 // further below. That's something we should try to do. Also, the proof 139 // of correctness of this 2-level locking scheme is far from obvious, 140 // and potentially quite slippery. We have an uneasy suspicion, for instance, 141 // that there may be a theoretical possibility of delay/starvation in the 142 // low-level lock/wait/notify scheme used for the baton-passing because of 143 // potential interference with the priority scheme embodied in the 144 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 145 // invocation further below and marked with "XXX 20011219YSR". 146 // Indeed, as we note elsewhere, this may become yet more slippery 147 // in the presence of multiple CMS and/or multiple VM threads. XXX 148 149 class CMSTokenSync: public StackObj { 150 private: 151 bool _is_cms_thread; 152 public: 153 CMSTokenSync(bool is_cms_thread): 154 _is_cms_thread(is_cms_thread) { 155 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 156 "Incorrect argument to constructor"); 157 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 158 } 159 160 ~CMSTokenSync() { 161 assert(_is_cms_thread ? 162 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 163 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 164 "Incorrect state"); 165 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 166 } 167 }; 168 169 // Convenience class that does a CMSTokenSync, and then acquires 170 // upto three locks. 171 class CMSTokenSyncWithLocks: public CMSTokenSync { 172 private: 173 // Note: locks are acquired in textual declaration order 174 // and released in the opposite order 175 MutexLockerEx _locker1, _locker2, _locker3; 176 public: 177 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 178 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 179 CMSTokenSync(is_cms_thread), 180 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 181 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 182 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 183 { } 184 }; 185 186 187 ////////////////////////////////////////////////////////////////// 188 // Concurrent Mark-Sweep Generation ///////////////////////////// 189 ////////////////////////////////////////////////////////////////// 190 191 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 192 193 // This struct contains per-thread things necessary to support parallel 194 // young-gen collection. 195 class CMSParGCThreadState: public CHeapObj<mtGC> { 196 public: 197 CompactibleFreeListSpaceLAB lab; 198 PromotionInfo promo; 199 200 // Constructor. 201 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 202 promo.setSpace(cfls); 203 } 204 }; 205 206 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 207 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 208 CardGeneration(rs, initial_byte_size, ct), 209 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 210 _did_compact(false) 211 { 212 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 213 HeapWord* end = (HeapWord*) _virtual_space.high(); 214 215 _direct_allocated_words = 0; 216 NOT_PRODUCT( 217 _numObjectsPromoted = 0; 218 _numWordsPromoted = 0; 219 _numObjectsAllocated = 0; 220 _numWordsAllocated = 0; 221 ) 222 223 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 224 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 225 _cmsSpace->_old_gen = this; 226 227 _gc_stats = new CMSGCStats(); 228 229 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 230 // offsets match. The ability to tell free chunks from objects 231 // depends on this property. 232 debug_only( 233 FreeChunk* junk = NULL; 234 assert(UseCompressedClassPointers || 235 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 236 "Offset of FreeChunk::_prev within FreeChunk must match" 237 " that of OopDesc::_klass within OopDesc"); 238 ) 239 240 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 241 for (uint i = 0; i < ParallelGCThreads; i++) { 242 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 243 } 244 245 _incremental_collection_failed = false; 246 // The "dilatation_factor" is the expansion that can occur on 247 // account of the fact that the minimum object size in the CMS 248 // generation may be larger than that in, say, a contiguous young 249 // generation. 250 // Ideally, in the calculation below, we'd compute the dilatation 251 // factor as: MinChunkSize/(promoting_gen's min object size) 252 // Since we do not have such a general query interface for the 253 // promoting generation, we'll instead just use the minimum 254 // object size (which today is a header's worth of space); 255 // note that all arithmetic is in units of HeapWords. 256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 257 assert(_dilatation_factor >= 1.0, "from previous assert"); 258 } 259 260 261 // The field "_initiating_occupancy" represents the occupancy percentage 262 // at which we trigger a new collection cycle. Unless explicitly specified 263 // via CMSInitiatingOccupancyFraction (argument "io" below), it 264 // is calculated by: 265 // 266 // Let "f" be MinHeapFreeRatio in 267 // 268 // _initiating_occupancy = 100-f + 269 // f * (CMSTriggerRatio/100) 270 // where CMSTriggerRatio is the argument "tr" below. 271 // 272 // That is, if we assume the heap is at its desired maximum occupancy at the 273 // end of a collection, we let CMSTriggerRatio of the (purported) free 274 // space be allocated before initiating a new collection cycle. 275 // 276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 277 assert(io <= 100 && tr <= 100, "Check the arguments"); 278 if (io >= 0) { 279 _initiating_occupancy = (double)io / 100.0; 280 } else { 281 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 282 (double)(tr * MinHeapFreeRatio) / 100.0) 283 / 100.0; 284 } 285 } 286 287 void ConcurrentMarkSweepGeneration::ref_processor_init() { 288 assert(collector() != NULL, "no collector"); 289 collector()->ref_processor_init(); 290 } 291 292 void CMSCollector::ref_processor_init() { 293 if (_ref_processor == NULL) { 294 // Allocate and initialize a reference processor 295 _ref_processor = 296 new ReferenceProcessor(&_span_based_discoverer, 297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 298 ParallelGCThreads, // mt processing degree 299 _cmsGen->refs_discovery_is_mt(), // mt discovery 300 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 302 &_is_alive_closure); // closure for liveness info 303 // Initialize the _ref_processor field of CMSGen 304 _cmsGen->set_ref_processor(_ref_processor); 305 306 } 307 } 308 309 AdaptiveSizePolicy* CMSCollector::size_policy() { 310 return CMSHeap::heap()->size_policy(); 311 } 312 313 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 314 315 const char* gen_name = "old"; 316 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 317 // Generation Counters - generation 1, 1 subspace 318 _gen_counters = new GenerationCounters(gen_name, 1, 1, 319 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 320 321 _space_counters = new GSpaceCounters(gen_name, 0, 322 _virtual_space.reserved_size(), 323 this, _gen_counters); 324 } 325 326 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 327 _cms_gen(cms_gen) 328 { 329 assert(alpha <= 100, "bad value"); 330 _saved_alpha = alpha; 331 332 // Initialize the alphas to the bootstrap value of 100. 333 _gc0_alpha = _cms_alpha = 100; 334 335 _cms_begin_time.update(); 336 _cms_end_time.update(); 337 338 _gc0_duration = 0.0; 339 _gc0_period = 0.0; 340 _gc0_promoted = 0; 341 342 _cms_duration = 0.0; 343 _cms_period = 0.0; 344 _cms_allocated = 0; 345 346 _cms_used_at_gc0_begin = 0; 347 _cms_used_at_gc0_end = 0; 348 _allow_duty_cycle_reduction = false; 349 _valid_bits = 0; 350 } 351 352 double CMSStats::cms_free_adjustment_factor(size_t free) const { 353 // TBD: CR 6909490 354 return 1.0; 355 } 356 357 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 358 } 359 360 // If promotion failure handling is on use 361 // the padded average size of the promotion for each 362 // young generation collection. 363 double CMSStats::time_until_cms_gen_full() const { 364 size_t cms_free = _cms_gen->cmsSpace()->free(); 365 CMSHeap* heap = CMSHeap::heap(); 366 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 367 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 368 if (cms_free > expected_promotion) { 369 // Start a cms collection if there isn't enough space to promote 370 // for the next young collection. Use the padded average as 371 // a safety factor. 372 cms_free -= expected_promotion; 373 374 // Adjust by the safety factor. 375 double cms_free_dbl = (double)cms_free; 376 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 377 // Apply a further correction factor which tries to adjust 378 // for recent occurance of concurrent mode failures. 379 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 380 cms_free_dbl = cms_free_dbl * cms_adjustment; 381 382 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 383 cms_free, expected_promotion); 384 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 385 // Add 1 in case the consumption rate goes to zero. 386 return cms_free_dbl / (cms_consumption_rate() + 1.0); 387 } 388 return 0.0; 389 } 390 391 // Compare the duration of the cms collection to the 392 // time remaining before the cms generation is empty. 393 // Note that the time from the start of the cms collection 394 // to the start of the cms sweep (less than the total 395 // duration of the cms collection) can be used. This 396 // has been tried and some applications experienced 397 // promotion failures early in execution. This was 398 // possibly because the averages were not accurate 399 // enough at the beginning. 400 double CMSStats::time_until_cms_start() const { 401 // We add "gc0_period" to the "work" calculation 402 // below because this query is done (mostly) at the 403 // end of a scavenge, so we need to conservatively 404 // account for that much possible delay 405 // in the query so as to avoid concurrent mode failures 406 // due to starting the collection just a wee bit too 407 // late. 408 double work = cms_duration() + gc0_period(); 409 double deadline = time_until_cms_gen_full(); 410 // If a concurrent mode failure occurred recently, we want to be 411 // more conservative and halve our expected time_until_cms_gen_full() 412 if (work > deadline) { 413 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 414 cms_duration(), gc0_period(), time_until_cms_gen_full()); 415 return 0.0; 416 } 417 return work - deadline; 418 } 419 420 #ifndef PRODUCT 421 void CMSStats::print_on(outputStream *st) const { 422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 424 gc0_duration(), gc0_period(), gc0_promoted()); 425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 426 cms_duration(), cms_period(), cms_allocated()); 427 st->print(",cms_since_beg=%g,cms_since_end=%g", 428 cms_time_since_begin(), cms_time_since_end()); 429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 431 432 if (valid()) { 433 st->print(",promo_rate=%g,cms_alloc_rate=%g", 434 promotion_rate(), cms_allocation_rate()); 435 st->print(",cms_consumption_rate=%g,time_until_full=%g", 436 cms_consumption_rate(), time_until_cms_gen_full()); 437 } 438 st->cr(); 439 } 440 #endif // #ifndef PRODUCT 441 442 CMSCollector::CollectorState CMSCollector::_collectorState = 443 CMSCollector::Idling; 444 bool CMSCollector::_foregroundGCIsActive = false; 445 bool CMSCollector::_foregroundGCShouldWait = false; 446 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 448 CardTableRS* ct, 449 ConcurrentMarkSweepPolicy* cp): 450 _cmsGen(cmsGen), 451 // Adjust span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 _ct(ct), 454 _span_based_discoverer(_span), 455 _ref_processor(NULL), // will be set later 456 _conc_workers(NULL), // may be set later 457 _abort_preclean(false), 458 _start_sampling(false), 459 _between_prologue_and_epilogue(false), 460 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 461 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 462 -1 /* lock-free */, "No_lock" /* dummy */), 463 _modUnionClosurePar(&_modUnionTable), 464 // Construct the is_alive_closure with _span & markBitMap 465 _is_alive_closure(_span, &_markBitMap), 466 _restart_addr(NULL), 467 _overflow_list(NULL), 468 _stats(cmsGen), 469 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 470 //verify that this lock should be acquired with safepoint check. 471 Monitor::_safepoint_check_sometimes)), 472 _eden_chunk_array(NULL), // may be set in ctor body 473 _eden_chunk_capacity(0), // -- ditto -- 474 _eden_chunk_index(0), // -- ditto -- 475 _survivor_plab_array(NULL), // -- ditto -- 476 _survivor_chunk_array(NULL), // -- ditto -- 477 _survivor_chunk_capacity(0), // -- ditto -- 478 _survivor_chunk_index(0), // -- ditto -- 479 _ser_pmc_preclean_ovflw(0), 480 _ser_kac_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_ovflw(0), 484 _par_kac_ovflw(0), 485 #ifndef PRODUCT 486 _num_par_pushes(0), 487 #endif 488 _collection_count_start(0), 489 _verifying(false), 490 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 491 _completed_initialization(false), 492 _collector_policy(cp), 493 _should_unload_classes(CMSClassUnloadingEnabled), 494 _concurrent_cycles_since_last_unload(0), 495 _roots_scanning_options(GenCollectedHeap::SO_None), 496 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 498 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 499 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 500 _cms_start_registered(false) 501 { 502 // Now expand the span and allocate the collection support structures 503 // (MUT, marking bit map etc.) to cover both generations subject to 504 // collection. 505 506 // For use by dirty card to oop closures. 507 _cmsGen->cmsSpace()->set_collector(this); 508 509 // Allocate MUT and marking bit map 510 { 511 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 512 if (!_markBitMap.allocate(_span)) { 513 log_warning(gc)("Failed to allocate CMS Bit Map"); 514 return; 515 } 516 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 517 } 518 { 519 _modUnionTable.allocate(_span); 520 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 521 } 522 523 if (!_markStack.allocate(MarkStackSize)) { 524 log_warning(gc)("Failed to allocate CMS Marking Stack"); 525 return; 526 } 527 528 // Support for multi-threaded concurrent phases 529 if (CMSConcurrentMTEnabled) { 530 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 531 // just for now 532 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 533 } 534 if (ConcGCThreads > 1) { 535 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 536 ConcGCThreads, true); 537 if (_conc_workers == NULL) { 538 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 539 CMSConcurrentMTEnabled = false; 540 } else { 541 _conc_workers->initialize_workers(); 542 } 543 } else { 544 CMSConcurrentMTEnabled = false; 545 } 546 } 547 if (!CMSConcurrentMTEnabled) { 548 ConcGCThreads = 0; 549 } else { 550 // Turn off CMSCleanOnEnter optimization temporarily for 551 // the MT case where it's not fixed yet; see 6178663. 552 CMSCleanOnEnter = false; 553 } 554 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 555 "Inconsistency"); 556 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 557 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 558 559 // Parallel task queues; these are shared for the 560 // concurrent and stop-world phases of CMS, but 561 // are not shared with parallel scavenge (ParNew). 562 { 563 uint i; 564 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 565 566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 567 || ParallelRefProcEnabled) 568 && num_queues > 0) { 569 _task_queues = new OopTaskQueueSet(num_queues); 570 if (_task_queues == NULL) { 571 log_warning(gc)("task_queues allocation failure."); 572 return; 573 } 574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 576 for (i = 0; i < num_queues; i++) { 577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 578 if (q == NULL) { 579 log_warning(gc)("work_queue allocation failure."); 580 return; 581 } 582 _task_queues->register_queue(i, q); 583 } 584 for (i = 0; i < num_queues; i++) { 585 _task_queues->queue(i)->initialize(); 586 _hash_seed[i] = 17; // copied from ParNew 587 } 588 } 589 } 590 591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 592 593 // Clip CMSBootstrapOccupancy between 0 and 100. 594 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 595 596 // Now tell CMS generations the identity of their collector 597 ConcurrentMarkSweepGeneration::set_collector(this); 598 599 // Create & start a CMS thread for this CMS collector 600 _cmsThread = ConcurrentMarkSweepThread::start(this); 601 assert(cmsThread() != NULL, "CMS Thread should have been created"); 602 assert(cmsThread()->collector() == this, 603 "CMS Thread should refer to this gen"); 604 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 605 606 // Support for parallelizing young gen rescan 607 CMSHeap* heap = CMSHeap::heap(); 608 _young_gen = heap->young_gen(); 609 if (heap->supports_inline_contig_alloc()) { 610 _top_addr = heap->top_addr(); 611 _end_addr = heap->end_addr(); 612 assert(_young_gen != NULL, "no _young_gen"); 613 _eden_chunk_index = 0; 614 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 615 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 616 } 617 618 // Support for parallelizing survivor space rescan 619 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 620 const size_t max_plab_samples = 621 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 622 623 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 624 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 625 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 626 _survivor_chunk_capacity = max_plab_samples; 627 for (uint i = 0; i < ParallelGCThreads; i++) { 628 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 629 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 630 assert(cur->end() == 0, "Should be 0"); 631 assert(cur->array() == vec, "Should be vec"); 632 assert(cur->capacity() == max_plab_samples, "Error"); 633 } 634 } 635 636 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 637 _gc_counters = new CollectorCounters("CMS", 1); 638 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 639 _completed_initialization = true; 640 _inter_sweep_timer.start(); // start of time 641 } 642 643 const char* ConcurrentMarkSweepGeneration::name() const { 644 return "concurrent mark-sweep generation"; 645 } 646 void ConcurrentMarkSweepGeneration::update_counters() { 647 if (UsePerfData) { 648 _space_counters->update_all(); 649 _gen_counters->update_all(); 650 } 651 } 652 653 // this is an optimized version of update_counters(). it takes the 654 // used value as a parameter rather than computing it. 655 // 656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 657 if (UsePerfData) { 658 _space_counters->update_used(used); 659 _space_counters->update_capacity(); 660 _gen_counters->update_all(); 661 } 662 } 663 664 void ConcurrentMarkSweepGeneration::print() const { 665 Generation::print(); 666 cmsSpace()->print(); 667 } 668 669 #ifndef PRODUCT 670 void ConcurrentMarkSweepGeneration::print_statistics() { 671 cmsSpace()->printFLCensus(0); 672 } 673 #endif 674 675 size_t 676 ConcurrentMarkSweepGeneration::contiguous_available() const { 677 // dld proposes an improvement in precision here. If the committed 678 // part of the space ends in a free block we should add that to 679 // uncommitted size in the calculation below. Will make this 680 // change later, staying with the approximation below for the 681 // time being. -- ysr. 682 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 683 } 684 685 size_t 686 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 687 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 688 } 689 690 size_t ConcurrentMarkSweepGeneration::max_available() const { 691 return free() + _virtual_space.uncommitted_size(); 692 } 693 694 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 695 size_t available = max_available(); 696 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 697 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 698 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 699 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 700 return res; 701 } 702 703 // At a promotion failure dump information on block layout in heap 704 // (cms old generation). 705 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 706 Log(gc, promotion) log; 707 if (log.is_trace()) { 708 LogStream ls(log.trace()); 709 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 710 } 711 } 712 713 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 714 // Clear the promotion information. These pointers can be adjusted 715 // along with all the other pointers into the heap but 716 // compaction is expected to be a rare event with 717 // a heap using cms so don't do it without seeing the need. 718 for (uint i = 0; i < ParallelGCThreads; i++) { 719 _par_gc_thread_states[i]->promo.reset(); 720 } 721 } 722 723 void ConcurrentMarkSweepGeneration::compute_new_size() { 724 assert_locked_or_safepoint(Heap_lock); 725 726 // If incremental collection failed, we just want to expand 727 // to the limit. 728 if (incremental_collection_failed()) { 729 clear_incremental_collection_failed(); 730 grow_to_reserved(); 731 return; 732 } 733 734 // The heap has been compacted but not reset yet. 735 // Any metric such as free() or used() will be incorrect. 736 737 CardGeneration::compute_new_size(); 738 739 // Reset again after a possible resizing 740 if (did_compact()) { 741 cmsSpace()->reset_after_compaction(); 742 } 743 } 744 745 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 746 assert_locked_or_safepoint(Heap_lock); 747 748 // If incremental collection failed, we just want to expand 749 // to the limit. 750 if (incremental_collection_failed()) { 751 clear_incremental_collection_failed(); 752 grow_to_reserved(); 753 return; 754 } 755 756 double free_percentage = ((double) free()) / capacity(); 757 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 758 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 759 760 // compute expansion delta needed for reaching desired free percentage 761 if (free_percentage < desired_free_percentage) { 762 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 763 assert(desired_capacity >= capacity(), "invalid expansion size"); 764 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 765 Log(gc) log; 766 if (log.is_trace()) { 767 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 768 log.trace("From compute_new_size: "); 769 log.trace(" Free fraction %f", free_percentage); 770 log.trace(" Desired free fraction %f", desired_free_percentage); 771 log.trace(" Maximum free fraction %f", maximum_free_percentage); 772 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 773 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 774 CMSHeap* heap = CMSHeap::heap(); 775 size_t young_size = heap->young_gen()->capacity(); 776 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 777 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 778 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 779 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 780 } 781 // safe if expansion fails 782 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 783 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 784 } else { 785 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 786 assert(desired_capacity <= capacity(), "invalid expansion size"); 787 size_t shrink_bytes = capacity() - desired_capacity; 788 // Don't shrink unless the delta is greater than the minimum shrink we want 789 if (shrink_bytes >= MinHeapDeltaBytes) { 790 shrink_free_list_by(shrink_bytes); 791 } 792 } 793 } 794 795 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 796 return cmsSpace()->freelistLock(); 797 } 798 799 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 800 CMSSynchronousYieldRequest yr; 801 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 802 return have_lock_and_allocate(size, tlab); 803 } 804 805 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 806 bool tlab /* ignored */) { 807 assert_lock_strong(freelistLock()); 808 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 809 HeapWord* res = cmsSpace()->allocate(adjustedSize); 810 // Allocate the object live (grey) if the background collector has 811 // started marking. This is necessary because the marker may 812 // have passed this address and consequently this object will 813 // not otherwise be greyed and would be incorrectly swept up. 814 // Note that if this object contains references, the writing 815 // of those references will dirty the card containing this object 816 // allowing the object to be blackened (and its references scanned) 817 // either during a preclean phase or at the final checkpoint. 818 if (res != NULL) { 819 // We may block here with an uninitialized object with 820 // its mark-bit or P-bits not yet set. Such objects need 821 // to be safely navigable by block_start(). 822 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 823 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 824 collector()->direct_allocated(res, adjustedSize); 825 _direct_allocated_words += adjustedSize; 826 // allocation counters 827 NOT_PRODUCT( 828 _numObjectsAllocated++; 829 _numWordsAllocated += (int)adjustedSize; 830 ) 831 } 832 return res; 833 } 834 835 // In the case of direct allocation by mutators in a generation that 836 // is being concurrently collected, the object must be allocated 837 // live (grey) if the background collector has started marking. 838 // This is necessary because the marker may 839 // have passed this address and consequently this object will 840 // not otherwise be greyed and would be incorrectly swept up. 841 // Note that if this object contains references, the writing 842 // of those references will dirty the card containing this object 843 // allowing the object to be blackened (and its references scanned) 844 // either during a preclean phase or at the final checkpoint. 845 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 846 assert(_markBitMap.covers(start, size), "Out of bounds"); 847 if (_collectorState >= Marking) { 848 MutexLockerEx y(_markBitMap.lock(), 849 Mutex::_no_safepoint_check_flag); 850 // [see comments preceding SweepClosure::do_blk() below for details] 851 // 852 // Can the P-bits be deleted now? JJJ 853 // 854 // 1. need to mark the object as live so it isn't collected 855 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 856 // 3. need to mark the end of the object so marking, precleaning or sweeping 857 // can skip over uninitialized or unparsable objects. An allocated 858 // object is considered uninitialized for our purposes as long as 859 // its klass word is NULL. All old gen objects are parsable 860 // as soon as they are initialized.) 861 _markBitMap.mark(start); // object is live 862 _markBitMap.mark(start + 1); // object is potentially uninitialized? 863 _markBitMap.mark(start + size - 1); 864 // mark end of object 865 } 866 // check that oop looks uninitialized 867 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 868 } 869 870 void CMSCollector::promoted(bool par, HeapWord* start, 871 bool is_obj_array, size_t obj_size) { 872 assert(_markBitMap.covers(start), "Out of bounds"); 873 // See comment in direct_allocated() about when objects should 874 // be allocated live. 875 if (_collectorState >= Marking) { 876 // we already hold the marking bit map lock, taken in 877 // the prologue 878 if (par) { 879 _markBitMap.par_mark(start); 880 } else { 881 _markBitMap.mark(start); 882 } 883 // We don't need to mark the object as uninitialized (as 884 // in direct_allocated above) because this is being done with the 885 // world stopped and the object will be initialized by the 886 // time the marking, precleaning or sweeping get to look at it. 887 // But see the code for copying objects into the CMS generation, 888 // where we need to ensure that concurrent readers of the 889 // block offset table are able to safely navigate a block that 890 // is in flux from being free to being allocated (and in 891 // transition while being copied into) and subsequently 892 // becoming a bona-fide object when the copy/promotion is complete. 893 assert(SafepointSynchronize::is_at_safepoint(), 894 "expect promotion only at safepoints"); 895 896 if (_collectorState < Sweeping) { 897 // Mark the appropriate cards in the modUnionTable, so that 898 // this object gets scanned before the sweep. If this is 899 // not done, CMS generation references in the object might 900 // not get marked. 901 // For the case of arrays, which are otherwise precisely 902 // marked, we need to dirty the entire array, not just its head. 903 if (is_obj_array) { 904 // The [par_]mark_range() method expects mr.end() below to 905 // be aligned to the granularity of a bit's representation 906 // in the heap. In the case of the MUT below, that's a 907 // card size. 908 MemRegion mr(start, 909 align_up(start + obj_size, 910 CardTable::card_size /* bytes */)); 911 if (par) { 912 _modUnionTable.par_mark_range(mr); 913 } else { 914 _modUnionTable.mark_range(mr); 915 } 916 } else { // not an obj array; we can just mark the head 917 if (par) { 918 _modUnionTable.par_mark(start); 919 } else { 920 _modUnionTable.mark(start); 921 } 922 } 923 } 924 } 925 } 926 927 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 928 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 929 // allocate, copy and if necessary update promoinfo -- 930 // delegate to underlying space. 931 assert_lock_strong(freelistLock()); 932 933 #ifndef PRODUCT 934 if (CMSHeap::heap()->promotion_should_fail()) { 935 return NULL; 936 } 937 #endif // #ifndef PRODUCT 938 939 oop res = _cmsSpace->promote(obj, obj_size); 940 if (res == NULL) { 941 // expand and retry 942 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 943 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 944 // Since this is the old generation, we don't try to promote 945 // into a more senior generation. 946 res = _cmsSpace->promote(obj, obj_size); 947 } 948 if (res != NULL) { 949 // See comment in allocate() about when objects should 950 // be allocated live. 951 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 952 collector()->promoted(false, // Not parallel 953 (HeapWord*)res, obj->is_objArray(), obj_size); 954 // promotion counters 955 NOT_PRODUCT( 956 _numObjectsPromoted++; 957 _numWordsPromoted += 958 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 959 ) 960 } 961 return res; 962 } 963 964 965 // IMPORTANT: Notes on object size recognition in CMS. 966 // --------------------------------------------------- 967 // A block of storage in the CMS generation is always in 968 // one of three states. A free block (FREE), an allocated 969 // object (OBJECT) whose size() method reports the correct size, 970 // and an intermediate state (TRANSIENT) in which its size cannot 971 // be accurately determined. 972 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 973 // ----------------------------------------------------- 974 // FREE: klass_word & 1 == 1; mark_word holds block size 975 // 976 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 977 // obj->size() computes correct size 978 // 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 980 // 981 // STATE IDENTIFICATION: (64 bit+COOPS) 982 // ------------------------------------ 983 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 984 // 985 // OBJECT: klass_word installed; klass_word != 0; 986 // obj->size() computes correct size 987 // 988 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 989 // 990 // 991 // STATE TRANSITION DIAGRAM 992 // 993 // mut / parnew mut / parnew 994 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 995 // ^ | 996 // |------------------------ DEAD <------------------------------------| 997 // sweep mut 998 // 999 // While a block is in TRANSIENT state its size cannot be determined 1000 // so readers will either need to come back later or stall until 1001 // the size can be determined. Note that for the case of direct 1002 // allocation, P-bits, when available, may be used to determine the 1003 // size of an object that may not yet have been initialized. 1004 1005 // Things to support parallel young-gen collection. 1006 oop 1007 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1008 oop old, markOop m, 1009 size_t word_sz) { 1010 #ifndef PRODUCT 1011 if (CMSHeap::heap()->promotion_should_fail()) { 1012 return NULL; 1013 } 1014 #endif // #ifndef PRODUCT 1015 1016 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1017 PromotionInfo* promoInfo = &ps->promo; 1018 // if we are tracking promotions, then first ensure space for 1019 // promotion (including spooling space for saving header if necessary). 1020 // then allocate and copy, then track promoted info if needed. 1021 // When tracking (see PromotionInfo::track()), the mark word may 1022 // be displaced and in this case restoration of the mark word 1023 // occurs in the (oop_since_save_marks_)iterate phase. 1024 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1025 // Out of space for allocating spooling buffers; 1026 // try expanding and allocating spooling buffers. 1027 if (!expand_and_ensure_spooling_space(promoInfo)) { 1028 return NULL; 1029 } 1030 } 1031 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1032 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1033 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1034 if (obj_ptr == NULL) { 1035 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1036 if (obj_ptr == NULL) { 1037 return NULL; 1038 } 1039 } 1040 oop obj = oop(obj_ptr); 1041 OrderAccess::storestore(); 1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1044 // IMPORTANT: See note on object initialization for CMS above. 1045 // Otherwise, copy the object. Here we must be careful to insert the 1046 // klass pointer last, since this marks the block as an allocated object. 1047 // Except with compressed oops it's the mark word. 1048 HeapWord* old_ptr = (HeapWord*)old; 1049 // Restore the mark word copied above. 1050 obj->set_mark_raw(m); 1051 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1052 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1053 OrderAccess::storestore(); 1054 1055 if (UseCompressedClassPointers) { 1056 // Copy gap missed by (aligned) header size calculation below 1057 obj->set_klass_gap(old->klass_gap()); 1058 } 1059 if (word_sz > (size_t)oopDesc::header_size()) { 1060 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1061 obj_ptr + oopDesc::header_size(), 1062 word_sz - oopDesc::header_size()); 1063 } 1064 1065 // Now we can track the promoted object, if necessary. We take care 1066 // to delay the transition from uninitialized to full object 1067 // (i.e., insertion of klass pointer) until after, so that it 1068 // atomically becomes a promoted object. 1069 if (promoInfo->tracking()) { 1070 promoInfo->track((PromotedObject*)obj, old->klass()); 1071 } 1072 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1073 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1074 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1075 1076 // Finally, install the klass pointer (this should be volatile). 1077 OrderAccess::storestore(); 1078 obj->set_klass(old->klass()); 1079 // We should now be able to calculate the right size for this object 1080 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1081 1082 collector()->promoted(true, // parallel 1083 obj_ptr, old->is_objArray(), word_sz); 1084 1085 NOT_PRODUCT( 1086 Atomic::inc(&_numObjectsPromoted); 1087 Atomic::add(alloc_sz, &_numWordsPromoted); 1088 ) 1089 1090 return obj; 1091 } 1092 1093 void 1094 ConcurrentMarkSweepGeneration:: 1095 par_promote_alloc_done(int thread_num) { 1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1097 ps->lab.retire(thread_num); 1098 } 1099 1100 void 1101 ConcurrentMarkSweepGeneration:: 1102 par_oop_since_save_marks_iterate_done(int thread_num) { 1103 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1104 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1105 ps->promo.promoted_oops_iterate(dummy_cl); 1106 1107 // Because card-scanning has been completed, subsequent phases 1108 // (e.g., reference processing) will not need to recognize which 1109 // objects have been promoted during this GC. So, we can now disable 1110 // promotion tracking. 1111 ps->promo.stopTrackingPromotions(); 1112 } 1113 1114 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1115 size_t size, 1116 bool tlab) 1117 { 1118 // We allow a STW collection only if a full 1119 // collection was requested. 1120 return full || should_allocate(size, tlab); // FIX ME !!! 1121 // This and promotion failure handling are connected at the 1122 // hip and should be fixed by untying them. 1123 } 1124 1125 bool CMSCollector::shouldConcurrentCollect() { 1126 LogTarget(Trace, gc) log; 1127 1128 if (_full_gc_requested) { 1129 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1130 return true; 1131 } 1132 1133 FreelistLocker x(this); 1134 // ------------------------------------------------------------------ 1135 // Print out lots of information which affects the initiation of 1136 // a collection. 1137 if (log.is_enabled() && stats().valid()) { 1138 log.print("CMSCollector shouldConcurrentCollect: "); 1139 1140 LogStream out(log); 1141 stats().print_on(&out); 1142 1143 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1144 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1145 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1146 log.print("promotion_rate=%g", stats().promotion_rate()); 1147 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1148 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1149 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1150 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1151 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1152 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1153 } 1154 // ------------------------------------------------------------------ 1155 1156 // If the estimated time to complete a cms collection (cms_duration()) 1157 // is less than the estimated time remaining until the cms generation 1158 // is full, start a collection. 1159 if (!UseCMSInitiatingOccupancyOnly) { 1160 if (stats().valid()) { 1161 if (stats().time_until_cms_start() == 0.0) { 1162 return true; 1163 } 1164 } else { 1165 // We want to conservatively collect somewhat early in order 1166 // to try and "bootstrap" our CMS/promotion statistics; 1167 // this branch will not fire after the first successful CMS 1168 // collection because the stats should then be valid. 1169 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1170 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1171 _cmsGen->occupancy(), _bootstrap_occupancy); 1172 return true; 1173 } 1174 } 1175 } 1176 1177 // Otherwise, we start a collection cycle if 1178 // old gen want a collection cycle started. Each may use 1179 // an appropriate criterion for making this decision. 1180 // XXX We need to make sure that the gen expansion 1181 // criterion dovetails well with this. XXX NEED TO FIX THIS 1182 if (_cmsGen->should_concurrent_collect()) { 1183 log.print("CMS old gen initiated"); 1184 return true; 1185 } 1186 1187 // We start a collection if we believe an incremental collection may fail; 1188 // this is not likely to be productive in practice because it's probably too 1189 // late anyway. 1190 CMSHeap* heap = CMSHeap::heap(); 1191 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1192 log.print("CMSCollector: collect because incremental collection will fail "); 1193 return true; 1194 } 1195 1196 if (MetaspaceGC::should_concurrent_collect()) { 1197 log.print("CMSCollector: collect for metadata allocation "); 1198 return true; 1199 } 1200 1201 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1202 if (CMSTriggerInterval >= 0) { 1203 if (CMSTriggerInterval == 0) { 1204 // Trigger always 1205 return true; 1206 } 1207 1208 // Check the CMS time since begin (we do not check the stats validity 1209 // as we want to be able to trigger the first CMS cycle as well) 1210 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1211 if (stats().valid()) { 1212 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1213 stats().cms_time_since_begin()); 1214 } else { 1215 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1216 } 1217 return true; 1218 } 1219 } 1220 1221 return false; 1222 } 1223 1224 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1225 1226 // Clear _expansion_cause fields of constituent generations 1227 void CMSCollector::clear_expansion_cause() { 1228 _cmsGen->clear_expansion_cause(); 1229 } 1230 1231 // We should be conservative in starting a collection cycle. To 1232 // start too eagerly runs the risk of collecting too often in the 1233 // extreme. To collect too rarely falls back on full collections, 1234 // which works, even if not optimum in terms of concurrent work. 1235 // As a work around for too eagerly collecting, use the flag 1236 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1237 // giving the user an easily understandable way of controlling the 1238 // collections. 1239 // We want to start a new collection cycle if any of the following 1240 // conditions hold: 1241 // . our current occupancy exceeds the configured initiating occupancy 1242 // for this generation, or 1243 // . we recently needed to expand this space and have not, since that 1244 // expansion, done a collection of this generation, or 1245 // . the underlying space believes that it may be a good idea to initiate 1246 // a concurrent collection (this may be based on criteria such as the 1247 // following: the space uses linear allocation and linear allocation is 1248 // going to fail, or there is believed to be excessive fragmentation in 1249 // the generation, etc... or ... 1250 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1251 // the case of the old generation; see CR 6543076): 1252 // we may be approaching a point at which allocation requests may fail because 1253 // we will be out of sufficient free space given allocation rate estimates.] 1254 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1255 1256 assert_lock_strong(freelistLock()); 1257 if (occupancy() > initiating_occupancy()) { 1258 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1259 short_name(), occupancy(), initiating_occupancy()); 1260 return true; 1261 } 1262 if (UseCMSInitiatingOccupancyOnly) { 1263 return false; 1264 } 1265 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1266 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1267 return true; 1268 } 1269 return false; 1270 } 1271 1272 void ConcurrentMarkSweepGeneration::collect(bool full, 1273 bool clear_all_soft_refs, 1274 size_t size, 1275 bool tlab) 1276 { 1277 collector()->collect(full, clear_all_soft_refs, size, tlab); 1278 } 1279 1280 void CMSCollector::collect(bool full, 1281 bool clear_all_soft_refs, 1282 size_t size, 1283 bool tlab) 1284 { 1285 // The following "if" branch is present for defensive reasons. 1286 // In the current uses of this interface, it can be replaced with: 1287 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1288 // But I am not placing that assert here to allow future 1289 // generality in invoking this interface. 1290 if (GCLocker::is_active()) { 1291 // A consistency test for GCLocker 1292 assert(GCLocker::needs_gc(), "Should have been set already"); 1293 // Skip this foreground collection, instead 1294 // expanding the heap if necessary. 1295 // Need the free list locks for the call to free() in compute_new_size() 1296 compute_new_size(); 1297 return; 1298 } 1299 acquire_control_and_collect(full, clear_all_soft_refs); 1300 } 1301 1302 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1303 CMSHeap* heap = CMSHeap::heap(); 1304 unsigned int gc_count = heap->total_full_collections(); 1305 if (gc_count == full_gc_count) { 1306 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1307 _full_gc_requested = true; 1308 _full_gc_cause = cause; 1309 CGC_lock->notify(); // nudge CMS thread 1310 } else { 1311 assert(gc_count > full_gc_count, "Error: causal loop"); 1312 } 1313 } 1314 1315 bool CMSCollector::is_external_interruption() { 1316 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1317 return GCCause::is_user_requested_gc(cause) || 1318 GCCause::is_serviceability_requested_gc(cause); 1319 } 1320 1321 void CMSCollector::report_concurrent_mode_interruption() { 1322 if (is_external_interruption()) { 1323 log_debug(gc)("Concurrent mode interrupted"); 1324 } else { 1325 log_debug(gc)("Concurrent mode failure"); 1326 _gc_tracer_cm->report_concurrent_mode_failure(); 1327 } 1328 } 1329 1330 1331 // The foreground and background collectors need to coordinate in order 1332 // to make sure that they do not mutually interfere with CMS collections. 1333 // When a background collection is active, 1334 // the foreground collector may need to take over (preempt) and 1335 // synchronously complete an ongoing collection. Depending on the 1336 // frequency of the background collections and the heap usage 1337 // of the application, this preemption can be seldom or frequent. 1338 // There are only certain 1339 // points in the background collection that the "collection-baton" 1340 // can be passed to the foreground collector. 1341 // 1342 // The foreground collector will wait for the baton before 1343 // starting any part of the collection. The foreground collector 1344 // will only wait at one location. 1345 // 1346 // The background collector will yield the baton before starting a new 1347 // phase of the collection (e.g., before initial marking, marking from roots, 1348 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1349 // of the loop which switches the phases. The background collector does some 1350 // of the phases (initial mark, final re-mark) with the world stopped. 1351 // Because of locking involved in stopping the world, 1352 // the foreground collector should not block waiting for the background 1353 // collector when it is doing a stop-the-world phase. The background 1354 // collector will yield the baton at an additional point just before 1355 // it enters a stop-the-world phase. Once the world is stopped, the 1356 // background collector checks the phase of the collection. If the 1357 // phase has not changed, it proceeds with the collection. If the 1358 // phase has changed, it skips that phase of the collection. See 1359 // the comments on the use of the Heap_lock in collect_in_background(). 1360 // 1361 // Variable used in baton passing. 1362 // _foregroundGCIsActive - Set to true by the foreground collector when 1363 // it wants the baton. The foreground clears it when it has finished 1364 // the collection. 1365 // _foregroundGCShouldWait - Set to true by the background collector 1366 // when it is running. The foreground collector waits while 1367 // _foregroundGCShouldWait is true. 1368 // CGC_lock - monitor used to protect access to the above variables 1369 // and to notify the foreground and background collectors. 1370 // _collectorState - current state of the CMS collection. 1371 // 1372 // The foreground collector 1373 // acquires the CGC_lock 1374 // sets _foregroundGCIsActive 1375 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1376 // various locks acquired in preparation for the collection 1377 // are released so as not to block the background collector 1378 // that is in the midst of a collection 1379 // proceeds with the collection 1380 // clears _foregroundGCIsActive 1381 // returns 1382 // 1383 // The background collector in a loop iterating on the phases of the 1384 // collection 1385 // acquires the CGC_lock 1386 // sets _foregroundGCShouldWait 1387 // if _foregroundGCIsActive is set 1388 // clears _foregroundGCShouldWait, notifies _CGC_lock 1389 // waits on _CGC_lock for _foregroundGCIsActive to become false 1390 // and exits the loop. 1391 // otherwise 1392 // proceed with that phase of the collection 1393 // if the phase is a stop-the-world phase, 1394 // yield the baton once more just before enqueueing 1395 // the stop-world CMS operation (executed by the VM thread). 1396 // returns after all phases of the collection are done 1397 // 1398 1399 void CMSCollector::acquire_control_and_collect(bool full, 1400 bool clear_all_soft_refs) { 1401 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1402 assert(!Thread::current()->is_ConcurrentGC_thread(), 1403 "shouldn't try to acquire control from self!"); 1404 1405 // Start the protocol for acquiring control of the 1406 // collection from the background collector (aka CMS thread). 1407 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1408 "VM thread should have CMS token"); 1409 // Remember the possibly interrupted state of an ongoing 1410 // concurrent collection 1411 CollectorState first_state = _collectorState; 1412 1413 // Signal to a possibly ongoing concurrent collection that 1414 // we want to do a foreground collection. 1415 _foregroundGCIsActive = true; 1416 1417 // release locks and wait for a notify from the background collector 1418 // releasing the locks in only necessary for phases which 1419 // do yields to improve the granularity of the collection. 1420 assert_lock_strong(bitMapLock()); 1421 // We need to lock the Free list lock for the space that we are 1422 // currently collecting. 1423 assert(haveFreelistLocks(), "Must be holding free list locks"); 1424 bitMapLock()->unlock(); 1425 releaseFreelistLocks(); 1426 { 1427 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1428 if (_foregroundGCShouldWait) { 1429 // We are going to be waiting for action for the CMS thread; 1430 // it had better not be gone (for instance at shutdown)! 1431 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1432 "CMS thread must be running"); 1433 // Wait here until the background collector gives us the go-ahead 1434 ConcurrentMarkSweepThread::clear_CMS_flag( 1435 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1436 // Get a possibly blocked CMS thread going: 1437 // Note that we set _foregroundGCIsActive true above, 1438 // without protection of the CGC_lock. 1439 CGC_lock->notify(); 1440 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1441 "Possible deadlock"); 1442 while (_foregroundGCShouldWait) { 1443 // wait for notification 1444 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1445 // Possibility of delay/starvation here, since CMS token does 1446 // not know to give priority to VM thread? Actually, i think 1447 // there wouldn't be any delay/starvation, but the proof of 1448 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1449 } 1450 ConcurrentMarkSweepThread::set_CMS_flag( 1451 ConcurrentMarkSweepThread::CMS_vm_has_token); 1452 } 1453 } 1454 // The CMS_token is already held. Get back the other locks. 1455 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1456 "VM thread should have CMS token"); 1457 getFreelistLocks(); 1458 bitMapLock()->lock_without_safepoint_check(); 1459 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1460 p2i(Thread::current()), first_state); 1461 log_debug(gc, state)(" gets control with state %d", _collectorState); 1462 1463 // Inform cms gen if this was due to partial collection failing. 1464 // The CMS gen may use this fact to determine its expansion policy. 1465 CMSHeap* heap = CMSHeap::heap(); 1466 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1467 assert(!_cmsGen->incremental_collection_failed(), 1468 "Should have been noticed, reacted to and cleared"); 1469 _cmsGen->set_incremental_collection_failed(); 1470 } 1471 1472 if (first_state > Idling) { 1473 report_concurrent_mode_interruption(); 1474 } 1475 1476 set_did_compact(true); 1477 1478 // If the collection is being acquired from the background 1479 // collector, there may be references on the discovered 1480 // references lists. Abandon those references, since some 1481 // of them may have become unreachable after concurrent 1482 // discovery; the STW compacting collector will redo discovery 1483 // more precisely, without being subject to floating garbage. 1484 // Leaving otherwise unreachable references in the discovered 1485 // lists would require special handling. 1486 ref_processor()->disable_discovery(); 1487 ref_processor()->abandon_partial_discovery(); 1488 ref_processor()->verify_no_references_recorded(); 1489 1490 if (first_state > Idling) { 1491 save_heap_summary(); 1492 } 1493 1494 do_compaction_work(clear_all_soft_refs); 1495 1496 // Has the GC time limit been exceeded? 1497 size_t max_eden_size = _young_gen->max_eden_size(); 1498 GCCause::Cause gc_cause = heap->gc_cause(); 1499 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1500 _young_gen->eden()->used(), 1501 _cmsGen->max_capacity(), 1502 max_eden_size, 1503 full, 1504 gc_cause, 1505 heap->soft_ref_policy()); 1506 1507 // Reset the expansion cause, now that we just completed 1508 // a collection cycle. 1509 clear_expansion_cause(); 1510 _foregroundGCIsActive = false; 1511 return; 1512 } 1513 1514 // Resize the tenured generation 1515 // after obtaining the free list locks for the 1516 // two generations. 1517 void CMSCollector::compute_new_size() { 1518 assert_locked_or_safepoint(Heap_lock); 1519 FreelistLocker z(this); 1520 MetaspaceGC::compute_new_size(); 1521 _cmsGen->compute_new_size_free_list(); 1522 } 1523 1524 // A work method used by the foreground collector to do 1525 // a mark-sweep-compact. 1526 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1527 CMSHeap* heap = CMSHeap::heap(); 1528 1529 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1530 gc_timer->register_gc_start(); 1531 1532 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1533 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1534 1535 heap->pre_full_gc_dump(gc_timer); 1536 1537 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1538 1539 // Temporarily widen the span of the weak reference processing to 1540 // the entire heap. 1541 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1542 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1543 // Temporarily, clear the "is_alive_non_header" field of the 1544 // reference processor. 1545 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1546 // Temporarily make reference _processing_ single threaded (non-MT). 1547 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1548 // Temporarily make refs discovery atomic 1549 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1550 // Temporarily make reference _discovery_ single threaded (non-MT) 1551 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1552 1553 ref_processor()->set_enqueuing_is_done(false); 1554 ref_processor()->enable_discovery(); 1555 ref_processor()->setup_policy(clear_all_soft_refs); 1556 // If an asynchronous collection finishes, the _modUnionTable is 1557 // all clear. If we are assuming the collection from an asynchronous 1558 // collection, clear the _modUnionTable. 1559 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1560 "_modUnionTable should be clear if the baton was not passed"); 1561 _modUnionTable.clear_all(); 1562 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1563 "mod union for klasses should be clear if the baton was passed"); 1564 _ct->cld_rem_set()->clear_mod_union(); 1565 1566 1567 // We must adjust the allocation statistics being maintained 1568 // in the free list space. We do so by reading and clearing 1569 // the sweep timer and updating the block flux rate estimates below. 1570 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1571 if (_inter_sweep_timer.is_active()) { 1572 _inter_sweep_timer.stop(); 1573 // Note that we do not use this sample to update the _inter_sweep_estimate. 1574 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1575 _inter_sweep_estimate.padded_average(), 1576 _intra_sweep_estimate.padded_average()); 1577 } 1578 1579 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1580 #ifdef ASSERT 1581 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1582 size_t free_size = cms_space->free(); 1583 assert(free_size == 1584 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1585 * HeapWordSize, 1586 "All the free space should be compacted into one chunk at top"); 1587 assert(cms_space->dictionary()->total_chunk_size( 1588 debug_only(cms_space->freelistLock())) == 0 || 1589 cms_space->totalSizeInIndexedFreeLists() == 0, 1590 "All the free space should be in a single chunk"); 1591 size_t num = cms_space->totalCount(); 1592 assert((free_size == 0 && num == 0) || 1593 (free_size > 0 && (num == 1 || num == 2)), 1594 "There should be at most 2 free chunks after compaction"); 1595 #endif // ASSERT 1596 _collectorState = Resetting; 1597 assert(_restart_addr == NULL, 1598 "Should have been NULL'd before baton was passed"); 1599 reset_stw(); 1600 _cmsGen->reset_after_compaction(); 1601 _concurrent_cycles_since_last_unload = 0; 1602 1603 // Clear any data recorded in the PLAB chunk arrays. 1604 if (_survivor_plab_array != NULL) { 1605 reset_survivor_plab_arrays(); 1606 } 1607 1608 // Adjust the per-size allocation stats for the next epoch. 1609 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1610 // Restart the "inter sweep timer" for the next epoch. 1611 _inter_sweep_timer.reset(); 1612 _inter_sweep_timer.start(); 1613 1614 // No longer a need to do a concurrent collection for Metaspace. 1615 MetaspaceGC::set_should_concurrent_collect(false); 1616 1617 heap->post_full_gc_dump(gc_timer); 1618 1619 gc_timer->register_gc_end(); 1620 1621 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1622 1623 // For a mark-sweep-compact, compute_new_size() will be called 1624 // in the heap's do_collection() method. 1625 } 1626 1627 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1628 Log(gc, heap) log; 1629 if (!log.is_trace()) { 1630 return; 1631 } 1632 1633 ContiguousSpace* eden_space = _young_gen->eden(); 1634 ContiguousSpace* from_space = _young_gen->from(); 1635 ContiguousSpace* to_space = _young_gen->to(); 1636 // Eden 1637 if (_eden_chunk_array != NULL) { 1638 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1639 p2i(eden_space->bottom()), p2i(eden_space->top()), 1640 p2i(eden_space->end()), eden_space->capacity()); 1641 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1642 _eden_chunk_index, _eden_chunk_capacity); 1643 for (size_t i = 0; i < _eden_chunk_index; i++) { 1644 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1645 } 1646 } 1647 // Survivor 1648 if (_survivor_chunk_array != NULL) { 1649 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1650 p2i(from_space->bottom()), p2i(from_space->top()), 1651 p2i(from_space->end()), from_space->capacity()); 1652 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1653 _survivor_chunk_index, _survivor_chunk_capacity); 1654 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1655 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1656 } 1657 } 1658 } 1659 1660 void CMSCollector::getFreelistLocks() const { 1661 // Get locks for all free lists in all generations that this 1662 // collector is responsible for 1663 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1664 } 1665 1666 void CMSCollector::releaseFreelistLocks() const { 1667 // Release locks for all free lists in all generations that this 1668 // collector is responsible for 1669 _cmsGen->freelistLock()->unlock(); 1670 } 1671 1672 bool CMSCollector::haveFreelistLocks() const { 1673 // Check locks for all free lists in all generations that this 1674 // collector is responsible for 1675 assert_lock_strong(_cmsGen->freelistLock()); 1676 PRODUCT_ONLY(ShouldNotReachHere()); 1677 return true; 1678 } 1679 1680 // A utility class that is used by the CMS collector to 1681 // temporarily "release" the foreground collector from its 1682 // usual obligation to wait for the background collector to 1683 // complete an ongoing phase before proceeding. 1684 class ReleaseForegroundGC: public StackObj { 1685 private: 1686 CMSCollector* _c; 1687 public: 1688 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1689 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1690 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1691 // allow a potentially blocked foreground collector to proceed 1692 _c->_foregroundGCShouldWait = false; 1693 if (_c->_foregroundGCIsActive) { 1694 CGC_lock->notify(); 1695 } 1696 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1697 "Possible deadlock"); 1698 } 1699 1700 ~ReleaseForegroundGC() { 1701 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1702 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1703 _c->_foregroundGCShouldWait = true; 1704 } 1705 }; 1706 1707 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1708 assert(Thread::current()->is_ConcurrentGC_thread(), 1709 "A CMS asynchronous collection is only allowed on a CMS thread."); 1710 1711 CMSHeap* heap = CMSHeap::heap(); 1712 { 1713 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1714 MutexLockerEx hl(Heap_lock, safepoint_check); 1715 FreelistLocker fll(this); 1716 MutexLockerEx x(CGC_lock, safepoint_check); 1717 if (_foregroundGCIsActive) { 1718 // The foreground collector is. Skip this 1719 // background collection. 1720 assert(!_foregroundGCShouldWait, "Should be clear"); 1721 return; 1722 } else { 1723 assert(_collectorState == Idling, "Should be idling before start."); 1724 _collectorState = InitialMarking; 1725 register_gc_start(cause); 1726 // Reset the expansion cause, now that we are about to begin 1727 // a new cycle. 1728 clear_expansion_cause(); 1729 1730 // Clear the MetaspaceGC flag since a concurrent collection 1731 // is starting but also clear it after the collection. 1732 MetaspaceGC::set_should_concurrent_collect(false); 1733 } 1734 // Decide if we want to enable class unloading as part of the 1735 // ensuing concurrent GC cycle. 1736 update_should_unload_classes(); 1737 _full_gc_requested = false; // acks all outstanding full gc requests 1738 _full_gc_cause = GCCause::_no_gc; 1739 // Signal that we are about to start a collection 1740 heap->increment_total_full_collections(); // ... starting a collection cycle 1741 _collection_count_start = heap->total_full_collections(); 1742 } 1743 1744 size_t prev_used = _cmsGen->used(); 1745 1746 // The change of the collection state is normally done at this level; 1747 // the exceptions are phases that are executed while the world is 1748 // stopped. For those phases the change of state is done while the 1749 // world is stopped. For baton passing purposes this allows the 1750 // background collector to finish the phase and change state atomically. 1751 // The foreground collector cannot wait on a phase that is done 1752 // while the world is stopped because the foreground collector already 1753 // has the world stopped and would deadlock. 1754 while (_collectorState != Idling) { 1755 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1756 p2i(Thread::current()), _collectorState); 1757 // The foreground collector 1758 // holds the Heap_lock throughout its collection. 1759 // holds the CMS token (but not the lock) 1760 // except while it is waiting for the background collector to yield. 1761 // 1762 // The foreground collector should be blocked (not for long) 1763 // if the background collector is about to start a phase 1764 // executed with world stopped. If the background 1765 // collector has already started such a phase, the 1766 // foreground collector is blocked waiting for the 1767 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1768 // are executed in the VM thread. 1769 // 1770 // The locking order is 1771 // PendingListLock (PLL) -- if applicable (FinalMarking) 1772 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1773 // CMS token (claimed in 1774 // stop_world_and_do() --> 1775 // safepoint_synchronize() --> 1776 // CMSThread::synchronize()) 1777 1778 { 1779 // Check if the FG collector wants us to yield. 1780 CMSTokenSync x(true); // is cms thread 1781 if (waitForForegroundGC()) { 1782 // We yielded to a foreground GC, nothing more to be 1783 // done this round. 1784 assert(_foregroundGCShouldWait == false, "We set it to false in " 1785 "waitForForegroundGC()"); 1786 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1787 p2i(Thread::current()), _collectorState); 1788 return; 1789 } else { 1790 // The background collector can run but check to see if the 1791 // foreground collector has done a collection while the 1792 // background collector was waiting to get the CGC_lock 1793 // above. If yes, break so that _foregroundGCShouldWait 1794 // is cleared before returning. 1795 if (_collectorState == Idling) { 1796 break; 1797 } 1798 } 1799 } 1800 1801 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1802 "should be waiting"); 1803 1804 switch (_collectorState) { 1805 case InitialMarking: 1806 { 1807 ReleaseForegroundGC x(this); 1808 stats().record_cms_begin(); 1809 VM_CMS_Initial_Mark initial_mark_op(this); 1810 VMThread::execute(&initial_mark_op); 1811 } 1812 // The collector state may be any legal state at this point 1813 // since the background collector may have yielded to the 1814 // foreground collector. 1815 break; 1816 case Marking: 1817 // initial marking in checkpointRootsInitialWork has been completed 1818 if (markFromRoots()) { // we were successful 1819 assert(_collectorState == Precleaning, "Collector state should " 1820 "have changed"); 1821 } else { 1822 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1823 } 1824 break; 1825 case Precleaning: 1826 // marking from roots in markFromRoots has been completed 1827 preclean(); 1828 assert(_collectorState == AbortablePreclean || 1829 _collectorState == FinalMarking, 1830 "Collector state should have changed"); 1831 break; 1832 case AbortablePreclean: 1833 abortable_preclean(); 1834 assert(_collectorState == FinalMarking, "Collector state should " 1835 "have changed"); 1836 break; 1837 case FinalMarking: 1838 { 1839 ReleaseForegroundGC x(this); 1840 1841 VM_CMS_Final_Remark final_remark_op(this); 1842 VMThread::execute(&final_remark_op); 1843 } 1844 assert(_foregroundGCShouldWait, "block post-condition"); 1845 break; 1846 case Sweeping: 1847 // final marking in checkpointRootsFinal has been completed 1848 sweep(); 1849 assert(_collectorState == Resizing, "Collector state change " 1850 "to Resizing must be done under the free_list_lock"); 1851 1852 case Resizing: { 1853 // Sweeping has been completed... 1854 // At this point the background collection has completed. 1855 // Don't move the call to compute_new_size() down 1856 // into code that might be executed if the background 1857 // collection was preempted. 1858 { 1859 ReleaseForegroundGC x(this); // unblock FG collection 1860 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1861 CMSTokenSync z(true); // not strictly needed. 1862 if (_collectorState == Resizing) { 1863 compute_new_size(); 1864 save_heap_summary(); 1865 _collectorState = Resetting; 1866 } else { 1867 assert(_collectorState == Idling, "The state should only change" 1868 " because the foreground collector has finished the collection"); 1869 } 1870 } 1871 break; 1872 } 1873 case Resetting: 1874 // CMS heap resizing has been completed 1875 reset_concurrent(); 1876 assert(_collectorState == Idling, "Collector state should " 1877 "have changed"); 1878 1879 MetaspaceGC::set_should_concurrent_collect(false); 1880 1881 stats().record_cms_end(); 1882 // Don't move the concurrent_phases_end() and compute_new_size() 1883 // calls to here because a preempted background collection 1884 // has it's state set to "Resetting". 1885 break; 1886 case Idling: 1887 default: 1888 ShouldNotReachHere(); 1889 break; 1890 } 1891 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1892 p2i(Thread::current()), _collectorState); 1893 assert(_foregroundGCShouldWait, "block post-condition"); 1894 } 1895 1896 // Should this be in gc_epilogue? 1897 heap->counters()->update_counters(); 1898 1899 { 1900 // Clear _foregroundGCShouldWait and, in the event that the 1901 // foreground collector is waiting, notify it, before 1902 // returning. 1903 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1904 _foregroundGCShouldWait = false; 1905 if (_foregroundGCIsActive) { 1906 CGC_lock->notify(); 1907 } 1908 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1909 "Possible deadlock"); 1910 } 1911 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1912 p2i(Thread::current()), _collectorState); 1913 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1914 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1915 } 1916 1917 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1918 _cms_start_registered = true; 1919 _gc_timer_cm->register_gc_start(); 1920 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1921 } 1922 1923 void CMSCollector::register_gc_end() { 1924 if (_cms_start_registered) { 1925 report_heap_summary(GCWhen::AfterGC); 1926 1927 _gc_timer_cm->register_gc_end(); 1928 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1929 _cms_start_registered = false; 1930 } 1931 } 1932 1933 void CMSCollector::save_heap_summary() { 1934 CMSHeap* heap = CMSHeap::heap(); 1935 _last_heap_summary = heap->create_heap_summary(); 1936 _last_metaspace_summary = heap->create_metaspace_summary(); 1937 } 1938 1939 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1940 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1941 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1942 } 1943 1944 bool CMSCollector::waitForForegroundGC() { 1945 bool res = false; 1946 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1947 "CMS thread should have CMS token"); 1948 // Block the foreground collector until the 1949 // background collectors decides whether to 1950 // yield. 1951 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1952 _foregroundGCShouldWait = true; 1953 if (_foregroundGCIsActive) { 1954 // The background collector yields to the 1955 // foreground collector and returns a value 1956 // indicating that it has yielded. The foreground 1957 // collector can proceed. 1958 res = true; 1959 _foregroundGCShouldWait = false; 1960 ConcurrentMarkSweepThread::clear_CMS_flag( 1961 ConcurrentMarkSweepThread::CMS_cms_has_token); 1962 ConcurrentMarkSweepThread::set_CMS_flag( 1963 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1964 // Get a possibly blocked foreground thread going 1965 CGC_lock->notify(); 1966 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1967 p2i(Thread::current()), _collectorState); 1968 while (_foregroundGCIsActive) { 1969 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1970 } 1971 ConcurrentMarkSweepThread::set_CMS_flag( 1972 ConcurrentMarkSweepThread::CMS_cms_has_token); 1973 ConcurrentMarkSweepThread::clear_CMS_flag( 1974 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1975 } 1976 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1977 p2i(Thread::current()), _collectorState); 1978 return res; 1979 } 1980 1981 // Because of the need to lock the free lists and other structures in 1982 // the collector, common to all the generations that the collector is 1983 // collecting, we need the gc_prologues of individual CMS generations 1984 // delegate to their collector. It may have been simpler had the 1985 // current infrastructure allowed one to call a prologue on a 1986 // collector. In the absence of that we have the generation's 1987 // prologue delegate to the collector, which delegates back 1988 // some "local" work to a worker method in the individual generations 1989 // that it's responsible for collecting, while itself doing any 1990 // work common to all generations it's responsible for. A similar 1991 // comment applies to the gc_epilogue()'s. 1992 // The role of the variable _between_prologue_and_epilogue is to 1993 // enforce the invocation protocol. 1994 void CMSCollector::gc_prologue(bool full) { 1995 // Call gc_prologue_work() for the CMSGen 1996 // we are responsible for. 1997 1998 // The following locking discipline assumes that we are only called 1999 // when the world is stopped. 2000 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2001 2002 // The CMSCollector prologue must call the gc_prologues for the 2003 // "generations" that it's responsible 2004 // for. 2005 2006 assert( Thread::current()->is_VM_thread() 2007 || ( CMSScavengeBeforeRemark 2008 && Thread::current()->is_ConcurrentGC_thread()), 2009 "Incorrect thread type for prologue execution"); 2010 2011 if (_between_prologue_and_epilogue) { 2012 // We have already been invoked; this is a gc_prologue delegation 2013 // from yet another CMS generation that we are responsible for, just 2014 // ignore it since all relevant work has already been done. 2015 return; 2016 } 2017 2018 // set a bit saying prologue has been called; cleared in epilogue 2019 _between_prologue_and_epilogue = true; 2020 // Claim locks for common data structures, then call gc_prologue_work() 2021 // for each CMSGen. 2022 2023 getFreelistLocks(); // gets free list locks on constituent spaces 2024 bitMapLock()->lock_without_safepoint_check(); 2025 2026 // Should call gc_prologue_work() for all cms gens we are responsible for 2027 bool duringMarking = _collectorState >= Marking 2028 && _collectorState < Sweeping; 2029 2030 // The young collections clear the modified oops state, which tells if 2031 // there are any modified oops in the class. The remark phase also needs 2032 // that information. Tell the young collection to save the union of all 2033 // modified klasses. 2034 if (duringMarking) { 2035 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2036 } 2037 2038 bool registerClosure = duringMarking; 2039 2040 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2041 2042 if (!full) { 2043 stats().record_gc0_begin(); 2044 } 2045 } 2046 2047 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2048 2049 _capacity_at_prologue = capacity(); 2050 _used_at_prologue = used(); 2051 2052 // We enable promotion tracking so that card-scanning can recognize 2053 // which objects have been promoted during this GC and skip them. 2054 for (uint i = 0; i < ParallelGCThreads; i++) { 2055 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2056 } 2057 2058 // Delegate to CMScollector which knows how to coordinate between 2059 // this and any other CMS generations that it is responsible for 2060 // collecting. 2061 collector()->gc_prologue(full); 2062 } 2063 2064 // This is a "private" interface for use by this generation's CMSCollector. 2065 // Not to be called directly by any other entity (for instance, 2066 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2067 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2068 bool registerClosure, ModUnionClosure* modUnionClosure) { 2069 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2070 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2071 "Should be NULL"); 2072 if (registerClosure) { 2073 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2074 } 2075 cmsSpace()->gc_prologue(); 2076 // Clear stat counters 2077 NOT_PRODUCT( 2078 assert(_numObjectsPromoted == 0, "check"); 2079 assert(_numWordsPromoted == 0, "check"); 2080 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2081 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2082 _numObjectsAllocated = 0; 2083 _numWordsAllocated = 0; 2084 ) 2085 } 2086 2087 void CMSCollector::gc_epilogue(bool full) { 2088 // The following locking discipline assumes that we are only called 2089 // when the world is stopped. 2090 assert(SafepointSynchronize::is_at_safepoint(), 2091 "world is stopped assumption"); 2092 2093 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2094 // if linear allocation blocks need to be appropriately marked to allow the 2095 // the blocks to be parsable. We also check here whether we need to nudge the 2096 // CMS collector thread to start a new cycle (if it's not already active). 2097 assert( Thread::current()->is_VM_thread() 2098 || ( CMSScavengeBeforeRemark 2099 && Thread::current()->is_ConcurrentGC_thread()), 2100 "Incorrect thread type for epilogue execution"); 2101 2102 if (!_between_prologue_and_epilogue) { 2103 // We have already been invoked; this is a gc_epilogue delegation 2104 // from yet another CMS generation that we are responsible for, just 2105 // ignore it since all relevant work has already been done. 2106 return; 2107 } 2108 assert(haveFreelistLocks(), "must have freelist locks"); 2109 assert_lock_strong(bitMapLock()); 2110 2111 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2112 2113 _cmsGen->gc_epilogue_work(full); 2114 2115 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2116 // in case sampling was not already enabled, enable it 2117 _start_sampling = true; 2118 } 2119 // reset _eden_chunk_array so sampling starts afresh 2120 _eden_chunk_index = 0; 2121 2122 size_t cms_used = _cmsGen->cmsSpace()->used(); 2123 2124 // update performance counters - this uses a special version of 2125 // update_counters() that allows the utilization to be passed as a 2126 // parameter, avoiding multiple calls to used(). 2127 // 2128 _cmsGen->update_counters(cms_used); 2129 2130 bitMapLock()->unlock(); 2131 releaseFreelistLocks(); 2132 2133 if (!CleanChunkPoolAsync) { 2134 Chunk::clean_chunk_pool(); 2135 } 2136 2137 set_did_compact(false); 2138 _between_prologue_and_epilogue = false; // ready for next cycle 2139 } 2140 2141 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2142 collector()->gc_epilogue(full); 2143 2144 // When using ParNew, promotion tracking should have already been 2145 // disabled. However, the prologue (which enables promotion 2146 // tracking) and epilogue are called irrespective of the type of 2147 // GC. So they will also be called before and after Full GCs, during 2148 // which promotion tracking will not be explicitly disabled. So, 2149 // it's safer to also disable it here too (to be symmetric with 2150 // enabling it in the prologue). 2151 for (uint i = 0; i < ParallelGCThreads; i++) { 2152 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2153 } 2154 } 2155 2156 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2157 assert(!incremental_collection_failed(), "Should have been cleared"); 2158 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2159 cmsSpace()->gc_epilogue(); 2160 // Print stat counters 2161 NOT_PRODUCT( 2162 assert(_numObjectsAllocated == 0, "check"); 2163 assert(_numWordsAllocated == 0, "check"); 2164 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2165 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2166 _numObjectsPromoted = 0; 2167 _numWordsPromoted = 0; 2168 ) 2169 2170 // Call down the chain in contiguous_available needs the freelistLock 2171 // so print this out before releasing the freeListLock. 2172 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2173 } 2174 2175 #ifndef PRODUCT 2176 bool CMSCollector::have_cms_token() { 2177 Thread* thr = Thread::current(); 2178 if (thr->is_VM_thread()) { 2179 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2180 } else if (thr->is_ConcurrentGC_thread()) { 2181 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2182 } else if (thr->is_GC_task_thread()) { 2183 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2184 ParGCRareEvent_lock->owned_by_self(); 2185 } 2186 return false; 2187 } 2188 2189 // Check reachability of the given heap address in CMS generation, 2190 // treating all other generations as roots. 2191 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2192 // We could "guarantee" below, rather than assert, but I'll 2193 // leave these as "asserts" so that an adventurous debugger 2194 // could try this in the product build provided some subset of 2195 // the conditions were met, provided they were interested in the 2196 // results and knew that the computation below wouldn't interfere 2197 // with other concurrent computations mutating the structures 2198 // being read or written. 2199 assert(SafepointSynchronize::is_at_safepoint(), 2200 "Else mutations in object graph will make answer suspect"); 2201 assert(have_cms_token(), "Should hold cms token"); 2202 assert(haveFreelistLocks(), "must hold free list locks"); 2203 assert_lock_strong(bitMapLock()); 2204 2205 // Clear the marking bit map array before starting, but, just 2206 // for kicks, first report if the given address is already marked 2207 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2208 _markBitMap.isMarked(addr) ? "" : " not"); 2209 2210 if (verify_after_remark()) { 2211 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2212 bool result = verification_mark_bm()->isMarked(addr); 2213 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2214 result ? "IS" : "is NOT"); 2215 return result; 2216 } else { 2217 tty->print_cr("Could not compute result"); 2218 return false; 2219 } 2220 } 2221 #endif 2222 2223 void 2224 CMSCollector::print_on_error(outputStream* st) { 2225 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2226 if (collector != NULL) { 2227 CMSBitMap* bitmap = &collector->_markBitMap; 2228 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2229 bitmap->print_on_error(st, " Bits: "); 2230 2231 st->cr(); 2232 2233 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2234 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2235 mut_bitmap->print_on_error(st, " Bits: "); 2236 } 2237 } 2238 2239 //////////////////////////////////////////////////////// 2240 // CMS Verification Support 2241 //////////////////////////////////////////////////////// 2242 // Following the remark phase, the following invariant 2243 // should hold -- each object in the CMS heap which is 2244 // marked in markBitMap() should be marked in the verification_mark_bm(). 2245 2246 class VerifyMarkedClosure: public BitMapClosure { 2247 CMSBitMap* _marks; 2248 bool _failed; 2249 2250 public: 2251 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2252 2253 bool do_bit(size_t offset) { 2254 HeapWord* addr = _marks->offsetToHeapWord(offset); 2255 if (!_marks->isMarked(addr)) { 2256 Log(gc, verify) log; 2257 ResourceMark rm; 2258 LogStream ls(log.error()); 2259 oop(addr)->print_on(&ls); 2260 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2261 _failed = true; 2262 } 2263 return true; 2264 } 2265 2266 bool failed() { return _failed; } 2267 }; 2268 2269 bool CMSCollector::verify_after_remark() { 2270 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2271 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2272 static bool init = false; 2273 2274 assert(SafepointSynchronize::is_at_safepoint(), 2275 "Else mutations in object graph will make answer suspect"); 2276 assert(have_cms_token(), 2277 "Else there may be mutual interference in use of " 2278 " verification data structures"); 2279 assert(_collectorState > Marking && _collectorState <= Sweeping, 2280 "Else marking info checked here may be obsolete"); 2281 assert(haveFreelistLocks(), "must hold free list locks"); 2282 assert_lock_strong(bitMapLock()); 2283 2284 2285 // Allocate marking bit map if not already allocated 2286 if (!init) { // first time 2287 if (!verification_mark_bm()->allocate(_span)) { 2288 return false; 2289 } 2290 init = true; 2291 } 2292 2293 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2294 2295 // Turn off refs discovery -- so we will be tracing through refs. 2296 // This is as intended, because by this time 2297 // GC must already have cleared any refs that need to be cleared, 2298 // and traced those that need to be marked; moreover, 2299 // the marking done here is not going to interfere in any 2300 // way with the marking information used by GC. 2301 NoRefDiscovery no_discovery(ref_processor()); 2302 2303 #if COMPILER2_OR_JVMCI 2304 DerivedPointerTableDeactivate dpt_deact; 2305 #endif 2306 2307 // Clear any marks from a previous round 2308 verification_mark_bm()->clear_all(); 2309 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2310 verify_work_stacks_empty(); 2311 2312 CMSHeap* heap = CMSHeap::heap(); 2313 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2314 // Update the saved marks which may affect the root scans. 2315 heap->save_marks(); 2316 2317 if (CMSRemarkVerifyVariant == 1) { 2318 // In this first variant of verification, we complete 2319 // all marking, then check if the new marks-vector is 2320 // a subset of the CMS marks-vector. 2321 verify_after_remark_work_1(); 2322 } else { 2323 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2324 // In this second variant of verification, we flag an error 2325 // (i.e. an object reachable in the new marks-vector not reachable 2326 // in the CMS marks-vector) immediately, also indicating the 2327 // identify of an object (A) that references the unmarked object (B) -- 2328 // presumably, a mutation to A failed to be picked up by preclean/remark? 2329 verify_after_remark_work_2(); 2330 } 2331 2332 return true; 2333 } 2334 2335 void CMSCollector::verify_after_remark_work_1() { 2336 ResourceMark rm; 2337 HandleMark hm; 2338 CMSHeap* heap = CMSHeap::heap(); 2339 2340 // Get a clear set of claim bits for the roots processing to work with. 2341 ClassLoaderDataGraph::clear_claimed_marks(); 2342 2343 // Mark from roots one level into CMS 2344 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2345 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2346 2347 { 2348 StrongRootsScope srs(1); 2349 2350 heap->cms_process_roots(&srs, 2351 true, // young gen as roots 2352 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2353 should_unload_classes(), 2354 ¬Older, 2355 NULL); 2356 } 2357 2358 // Now mark from the roots 2359 MarkFromRootsClosure markFromRootsClosure(this, _span, 2360 verification_mark_bm(), verification_mark_stack(), 2361 false /* don't yield */, true /* verifying */); 2362 assert(_restart_addr == NULL, "Expected pre-condition"); 2363 verification_mark_bm()->iterate(&markFromRootsClosure); 2364 while (_restart_addr != NULL) { 2365 // Deal with stack overflow: by restarting at the indicated 2366 // address. 2367 HeapWord* ra = _restart_addr; 2368 markFromRootsClosure.reset(ra); 2369 _restart_addr = NULL; 2370 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2371 } 2372 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2373 verify_work_stacks_empty(); 2374 2375 // Marking completed -- now verify that each bit marked in 2376 // verification_mark_bm() is also marked in markBitMap(); flag all 2377 // errors by printing corresponding objects. 2378 VerifyMarkedClosure vcl(markBitMap()); 2379 verification_mark_bm()->iterate(&vcl); 2380 if (vcl.failed()) { 2381 Log(gc, verify) log; 2382 log.error("Failed marking verification after remark"); 2383 ResourceMark rm; 2384 LogStream ls(log.error()); 2385 heap->print_on(&ls); 2386 fatal("CMS: failed marking verification after remark"); 2387 } 2388 } 2389 2390 class VerifyCLDOopsCLDClosure : public CLDClosure { 2391 class VerifyCLDOopsClosure : public OopClosure { 2392 CMSBitMap* _bitmap; 2393 public: 2394 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2395 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2396 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2397 } _oop_closure; 2398 public: 2399 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2400 void do_cld(ClassLoaderData* cld) { 2401 cld->oops_do(&_oop_closure, false, false); 2402 } 2403 }; 2404 2405 void CMSCollector::verify_after_remark_work_2() { 2406 ResourceMark rm; 2407 HandleMark hm; 2408 CMSHeap* heap = CMSHeap::heap(); 2409 2410 // Get a clear set of claim bits for the roots processing to work with. 2411 ClassLoaderDataGraph::clear_claimed_marks(); 2412 2413 // Mark from roots one level into CMS 2414 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2415 markBitMap()); 2416 CLDToOopClosure cld_closure(¬Older, true); 2417 2418 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2419 2420 { 2421 StrongRootsScope srs(1); 2422 2423 heap->cms_process_roots(&srs, 2424 true, // young gen as roots 2425 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2426 should_unload_classes(), 2427 ¬Older, 2428 &cld_closure); 2429 } 2430 2431 // Now mark from the roots 2432 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2433 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2434 assert(_restart_addr == NULL, "Expected pre-condition"); 2435 verification_mark_bm()->iterate(&markFromRootsClosure); 2436 while (_restart_addr != NULL) { 2437 // Deal with stack overflow: by restarting at the indicated 2438 // address. 2439 HeapWord* ra = _restart_addr; 2440 markFromRootsClosure.reset(ra); 2441 _restart_addr = NULL; 2442 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2443 } 2444 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2445 verify_work_stacks_empty(); 2446 2447 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2448 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2449 2450 // Marking completed -- now verify that each bit marked in 2451 // verification_mark_bm() is also marked in markBitMap(); flag all 2452 // errors by printing corresponding objects. 2453 VerifyMarkedClosure vcl(markBitMap()); 2454 verification_mark_bm()->iterate(&vcl); 2455 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2456 } 2457 2458 void ConcurrentMarkSweepGeneration::save_marks() { 2459 // delegate to CMS space 2460 cmsSpace()->save_marks(); 2461 } 2462 2463 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2464 return cmsSpace()->no_allocs_since_save_marks(); 2465 } 2466 2467 void 2468 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2469 if (freelistLock()->owned_by_self()) { 2470 Generation::oop_iterate(cl); 2471 } else { 2472 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2473 Generation::oop_iterate(cl); 2474 } 2475 } 2476 2477 void 2478 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2479 if (freelistLock()->owned_by_self()) { 2480 Generation::object_iterate(cl); 2481 } else { 2482 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2483 Generation::object_iterate(cl); 2484 } 2485 } 2486 2487 void 2488 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2489 if (freelistLock()->owned_by_self()) { 2490 Generation::safe_object_iterate(cl); 2491 } else { 2492 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2493 Generation::safe_object_iterate(cl); 2494 } 2495 } 2496 2497 void 2498 ConcurrentMarkSweepGeneration::post_compact() { 2499 } 2500 2501 void 2502 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2503 // Fix the linear allocation blocks to look like free blocks. 2504 2505 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2506 // are not called when the heap is verified during universe initialization and 2507 // at vm shutdown. 2508 if (freelistLock()->owned_by_self()) { 2509 cmsSpace()->prepare_for_verify(); 2510 } else { 2511 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2512 cmsSpace()->prepare_for_verify(); 2513 } 2514 } 2515 2516 void 2517 ConcurrentMarkSweepGeneration::verify() { 2518 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2519 // are not called when the heap is verified during universe initialization and 2520 // at vm shutdown. 2521 if (freelistLock()->owned_by_self()) { 2522 cmsSpace()->verify(); 2523 } else { 2524 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2525 cmsSpace()->verify(); 2526 } 2527 } 2528 2529 void CMSCollector::verify() { 2530 _cmsGen->verify(); 2531 } 2532 2533 #ifndef PRODUCT 2534 bool CMSCollector::overflow_list_is_empty() const { 2535 assert(_num_par_pushes >= 0, "Inconsistency"); 2536 if (_overflow_list == NULL) { 2537 assert(_num_par_pushes == 0, "Inconsistency"); 2538 } 2539 return _overflow_list == NULL; 2540 } 2541 2542 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2543 // merely consolidate assertion checks that appear to occur together frequently. 2544 void CMSCollector::verify_work_stacks_empty() const { 2545 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2546 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2547 } 2548 2549 void CMSCollector::verify_overflow_empty() const { 2550 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2551 assert(no_preserved_marks(), "No preserved marks"); 2552 } 2553 #endif // PRODUCT 2554 2555 // Decide if we want to enable class unloading as part of the 2556 // ensuing concurrent GC cycle. We will collect and 2557 // unload classes if it's the case that: 2558 // (a) class unloading is enabled at the command line, and 2559 // (b) old gen is getting really full 2560 // NOTE: Provided there is no change in the state of the heap between 2561 // calls to this method, it should have idempotent results. Moreover, 2562 // its results should be monotonically increasing (i.e. going from 0 to 1, 2563 // but not 1 to 0) between successive calls between which the heap was 2564 // not collected. For the implementation below, it must thus rely on 2565 // the property that concurrent_cycles_since_last_unload() 2566 // will not decrease unless a collection cycle happened and that 2567 // _cmsGen->is_too_full() are 2568 // themselves also monotonic in that sense. See check_monotonicity() 2569 // below. 2570 void CMSCollector::update_should_unload_classes() { 2571 _should_unload_classes = false; 2572 if (CMSClassUnloadingEnabled) { 2573 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2574 CMSClassUnloadingMaxInterval) 2575 || _cmsGen->is_too_full(); 2576 } 2577 } 2578 2579 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2580 bool res = should_concurrent_collect(); 2581 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2582 return res; 2583 } 2584 2585 void CMSCollector::setup_cms_unloading_and_verification_state() { 2586 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2587 || VerifyBeforeExit; 2588 const int rso = GenCollectedHeap::SO_AllCodeCache; 2589 2590 // We set the proper root for this CMS cycle here. 2591 if (should_unload_classes()) { // Should unload classes this cycle 2592 remove_root_scanning_option(rso); // Shrink the root set appropriately 2593 set_verifying(should_verify); // Set verification state for this cycle 2594 return; // Nothing else needs to be done at this time 2595 } 2596 2597 // Not unloading classes this cycle 2598 assert(!should_unload_classes(), "Inconsistency!"); 2599 2600 // If we are not unloading classes then add SO_AllCodeCache to root 2601 // scanning options. 2602 add_root_scanning_option(rso); 2603 2604 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2605 set_verifying(true); 2606 } else if (verifying() && !should_verify) { 2607 // We were verifying, but some verification flags got disabled. 2608 set_verifying(false); 2609 // Exclude symbols, strings and code cache elements from root scanning to 2610 // reduce IM and RM pauses. 2611 remove_root_scanning_option(rso); 2612 } 2613 } 2614 2615 2616 #ifndef PRODUCT 2617 HeapWord* CMSCollector::block_start(const void* p) const { 2618 const HeapWord* addr = (HeapWord*)p; 2619 if (_span.contains(p)) { 2620 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2621 return _cmsGen->cmsSpace()->block_start(p); 2622 } 2623 } 2624 return NULL; 2625 } 2626 #endif 2627 2628 HeapWord* 2629 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2630 bool tlab, 2631 bool parallel) { 2632 CMSSynchronousYieldRequest yr; 2633 assert(!tlab, "Can't deal with TLAB allocation"); 2634 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2635 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2636 if (GCExpandToAllocateDelayMillis > 0) { 2637 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2638 } 2639 return have_lock_and_allocate(word_size, tlab); 2640 } 2641 2642 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2643 size_t bytes, 2644 size_t expand_bytes, 2645 CMSExpansionCause::Cause cause) 2646 { 2647 2648 bool success = expand(bytes, expand_bytes); 2649 2650 // remember why we expanded; this information is used 2651 // by shouldConcurrentCollect() when making decisions on whether to start 2652 // a new CMS cycle. 2653 if (success) { 2654 set_expansion_cause(cause); 2655 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2656 } 2657 } 2658 2659 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2660 HeapWord* res = NULL; 2661 MutexLocker x(ParGCRareEvent_lock); 2662 while (true) { 2663 // Expansion by some other thread might make alloc OK now: 2664 res = ps->lab.alloc(word_sz); 2665 if (res != NULL) return res; 2666 // If there's not enough expansion space available, give up. 2667 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2668 return NULL; 2669 } 2670 // Otherwise, we try expansion. 2671 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2672 // Now go around the loop and try alloc again; 2673 // A competing par_promote might beat us to the expansion space, 2674 // so we may go around the loop again if promotion fails again. 2675 if (GCExpandToAllocateDelayMillis > 0) { 2676 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2677 } 2678 } 2679 } 2680 2681 2682 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2683 PromotionInfo* promo) { 2684 MutexLocker x(ParGCRareEvent_lock); 2685 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2686 while (true) { 2687 // Expansion by some other thread might make alloc OK now: 2688 if (promo->ensure_spooling_space()) { 2689 assert(promo->has_spooling_space(), 2690 "Post-condition of successful ensure_spooling_space()"); 2691 return true; 2692 } 2693 // If there's not enough expansion space available, give up. 2694 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2695 return false; 2696 } 2697 // Otherwise, we try expansion. 2698 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2699 // Now go around the loop and try alloc again; 2700 // A competing allocation might beat us to the expansion space, 2701 // so we may go around the loop again if allocation fails again. 2702 if (GCExpandToAllocateDelayMillis > 0) { 2703 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2704 } 2705 } 2706 } 2707 2708 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2709 // Only shrink if a compaction was done so that all the free space 2710 // in the generation is in a contiguous block at the end. 2711 if (did_compact()) { 2712 CardGeneration::shrink(bytes); 2713 } 2714 } 2715 2716 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2717 assert_locked_or_safepoint(Heap_lock); 2718 } 2719 2720 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2721 assert_locked_or_safepoint(Heap_lock); 2722 assert_lock_strong(freelistLock()); 2723 log_trace(gc)("Shrinking of CMS not yet implemented"); 2724 return; 2725 } 2726 2727 2728 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2729 // phases. 2730 class CMSPhaseAccounting: public StackObj { 2731 public: 2732 CMSPhaseAccounting(CMSCollector *collector, 2733 const char *title); 2734 ~CMSPhaseAccounting(); 2735 2736 private: 2737 CMSCollector *_collector; 2738 const char *_title; 2739 GCTraceConcTime(Info, gc) _trace_time; 2740 2741 public: 2742 // Not MT-safe; so do not pass around these StackObj's 2743 // where they may be accessed by other threads. 2744 double wallclock_millis() { 2745 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2746 } 2747 }; 2748 2749 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2750 const char *title) : 2751 _collector(collector), _title(title), _trace_time(title) { 2752 2753 _collector->resetYields(); 2754 _collector->resetTimer(); 2755 _collector->startTimer(); 2756 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2757 } 2758 2759 CMSPhaseAccounting::~CMSPhaseAccounting() { 2760 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2761 _collector->stopTimer(); 2762 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2763 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2764 } 2765 2766 // CMS work 2767 2768 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2769 class CMSParMarkTask : public AbstractGangTask { 2770 protected: 2771 CMSCollector* _collector; 2772 uint _n_workers; 2773 OopStorage::ParState<false, false> _par_state_string; 2774 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2775 AbstractGangTask(name), 2776 _collector(collector), 2777 _n_workers(n_workers), 2778 _par_state_string(StringTable::weak_storage()) {} 2779 // Work method in support of parallel rescan ... of young gen spaces 2780 void do_young_space_rescan(OopsInGenClosure* cl, 2781 ContiguousSpace* space, 2782 HeapWord** chunk_array, size_t chunk_top); 2783 void work_on_young_gen_roots(OopsInGenClosure* cl); 2784 }; 2785 2786 // Parallel initial mark task 2787 class CMSParInitialMarkTask: public CMSParMarkTask { 2788 StrongRootsScope* _strong_roots_scope; 2789 public: 2790 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2791 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2792 _strong_roots_scope(strong_roots_scope) {} 2793 void work(uint worker_id); 2794 }; 2795 2796 // Checkpoint the roots into this generation from outside 2797 // this generation. [Note this initial checkpoint need only 2798 // be approximate -- we'll do a catch up phase subsequently.] 2799 void CMSCollector::checkpointRootsInitial() { 2800 assert(_collectorState == InitialMarking, "Wrong collector state"); 2801 check_correct_thread_executing(); 2802 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2803 2804 save_heap_summary(); 2805 report_heap_summary(GCWhen::BeforeGC); 2806 2807 ReferenceProcessor* rp = ref_processor(); 2808 assert(_restart_addr == NULL, "Control point invariant"); 2809 { 2810 // acquire locks for subsequent manipulations 2811 MutexLockerEx x(bitMapLock(), 2812 Mutex::_no_safepoint_check_flag); 2813 checkpointRootsInitialWork(); 2814 // enable ("weak") refs discovery 2815 rp->enable_discovery(); 2816 _collectorState = Marking; 2817 } 2818 } 2819 2820 void CMSCollector::checkpointRootsInitialWork() { 2821 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2822 assert(_collectorState == InitialMarking, "just checking"); 2823 2824 // Already have locks. 2825 assert_lock_strong(bitMapLock()); 2826 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2827 2828 // Setup the verification and class unloading state for this 2829 // CMS collection cycle. 2830 setup_cms_unloading_and_verification_state(); 2831 2832 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2833 2834 // Reset all the PLAB chunk arrays if necessary. 2835 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2836 reset_survivor_plab_arrays(); 2837 } 2838 2839 ResourceMark rm; 2840 HandleMark hm; 2841 2842 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2843 CMSHeap* heap = CMSHeap::heap(); 2844 2845 verify_work_stacks_empty(); 2846 verify_overflow_empty(); 2847 2848 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2849 // Update the saved marks which may affect the root scans. 2850 heap->save_marks(); 2851 2852 // weak reference processing has not started yet. 2853 ref_processor()->set_enqueuing_is_done(false); 2854 2855 // Need to remember all newly created CLDs, 2856 // so that we can guarantee that the remark finds them. 2857 ClassLoaderDataGraph::remember_new_clds(true); 2858 2859 // Whenever a CLD is found, it will be claimed before proceeding to mark 2860 // the klasses. The claimed marks need to be cleared before marking starts. 2861 ClassLoaderDataGraph::clear_claimed_marks(); 2862 2863 print_eden_and_survivor_chunk_arrays(); 2864 2865 { 2866 #if COMPILER2_OR_JVMCI 2867 DerivedPointerTableDeactivate dpt_deact; 2868 #endif 2869 if (CMSParallelInitialMarkEnabled) { 2870 // The parallel version. 2871 WorkGang* workers = heap->workers(); 2872 assert(workers != NULL, "Need parallel worker threads."); 2873 uint n_workers = workers->active_workers(); 2874 2875 StrongRootsScope srs(n_workers); 2876 2877 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2878 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2879 // If the total workers is greater than 1, then multiple workers 2880 // may be used at some time and the initialization has been set 2881 // such that the single threaded path cannot be used. 2882 if (workers->total_workers() > 1) { 2883 workers->run_task(&tsk); 2884 } else { 2885 tsk.work(0); 2886 } 2887 } else { 2888 // The serial version. 2889 CLDToOopClosure cld_closure(¬Older, true); 2890 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2891 2892 StrongRootsScope srs(1); 2893 2894 heap->cms_process_roots(&srs, 2895 true, // young gen as roots 2896 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2897 should_unload_classes(), 2898 ¬Older, 2899 &cld_closure); 2900 } 2901 } 2902 2903 // Clear mod-union table; it will be dirtied in the prologue of 2904 // CMS generation per each young generation collection. 2905 2906 assert(_modUnionTable.isAllClear(), 2907 "Was cleared in most recent final checkpoint phase" 2908 " or no bits are set in the gc_prologue before the start of the next " 2909 "subsequent marking phase."); 2910 2911 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2912 2913 // Save the end of the used_region of the constituent generations 2914 // to be used to limit the extent of sweep in each generation. 2915 save_sweep_limits(); 2916 verify_overflow_empty(); 2917 } 2918 2919 bool CMSCollector::markFromRoots() { 2920 // we might be tempted to assert that: 2921 // assert(!SafepointSynchronize::is_at_safepoint(), 2922 // "inconsistent argument?"); 2923 // However that wouldn't be right, because it's possible that 2924 // a safepoint is indeed in progress as a young generation 2925 // stop-the-world GC happens even as we mark in this generation. 2926 assert(_collectorState == Marking, "inconsistent state?"); 2927 check_correct_thread_executing(); 2928 verify_overflow_empty(); 2929 2930 // Weak ref discovery note: We may be discovering weak 2931 // refs in this generation concurrent (but interleaved) with 2932 // weak ref discovery by the young generation collector. 2933 2934 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2935 GCTraceCPUTime tcpu; 2936 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2937 bool res = markFromRootsWork(); 2938 if (res) { 2939 _collectorState = Precleaning; 2940 } else { // We failed and a foreground collection wants to take over 2941 assert(_foregroundGCIsActive, "internal state inconsistency"); 2942 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2943 log_debug(gc)("bailing out to foreground collection"); 2944 } 2945 verify_overflow_empty(); 2946 return res; 2947 } 2948 2949 bool CMSCollector::markFromRootsWork() { 2950 // iterate over marked bits in bit map, doing a full scan and mark 2951 // from these roots using the following algorithm: 2952 // . if oop is to the right of the current scan pointer, 2953 // mark corresponding bit (we'll process it later) 2954 // . else (oop is to left of current scan pointer) 2955 // push oop on marking stack 2956 // . drain the marking stack 2957 2958 // Note that when we do a marking step we need to hold the 2959 // bit map lock -- recall that direct allocation (by mutators) 2960 // and promotion (by the young generation collector) is also 2961 // marking the bit map. [the so-called allocate live policy.] 2962 // Because the implementation of bit map marking is not 2963 // robust wrt simultaneous marking of bits in the same word, 2964 // we need to make sure that there is no such interference 2965 // between concurrent such updates. 2966 2967 // already have locks 2968 assert_lock_strong(bitMapLock()); 2969 2970 verify_work_stacks_empty(); 2971 verify_overflow_empty(); 2972 bool result = false; 2973 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2974 result = do_marking_mt(); 2975 } else { 2976 result = do_marking_st(); 2977 } 2978 return result; 2979 } 2980 2981 // Forward decl 2982 class CMSConcMarkingTask; 2983 2984 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2985 CMSCollector* _collector; 2986 CMSConcMarkingTask* _task; 2987 public: 2988 virtual void yield(); 2989 2990 // "n_threads" is the number of threads to be terminated. 2991 // "queue_set" is a set of work queues of other threads. 2992 // "collector" is the CMS collector associated with this task terminator. 2993 // "yield" indicates whether we need the gang as a whole to yield. 2994 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2995 ParallelTaskTerminator(n_threads, queue_set), 2996 _collector(collector) { } 2997 2998 void set_task(CMSConcMarkingTask* task) { 2999 _task = task; 3000 } 3001 }; 3002 3003 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3004 CMSConcMarkingTask* _task; 3005 public: 3006 bool should_exit_termination(); 3007 void set_task(CMSConcMarkingTask* task) { 3008 _task = task; 3009 } 3010 }; 3011 3012 // MT Concurrent Marking Task 3013 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3014 CMSCollector* _collector; 3015 uint _n_workers; // requested/desired # workers 3016 bool _result; 3017 CompactibleFreeListSpace* _cms_space; 3018 char _pad_front[64]; // padding to ... 3019 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3020 char _pad_back[64]; 3021 HeapWord* _restart_addr; 3022 3023 // Exposed here for yielding support 3024 Mutex* const _bit_map_lock; 3025 3026 // The per thread work queues, available here for stealing 3027 OopTaskQueueSet* _task_queues; 3028 3029 // Termination (and yielding) support 3030 CMSConcMarkingTerminator _term; 3031 CMSConcMarkingTerminatorTerminator _term_term; 3032 3033 public: 3034 CMSConcMarkingTask(CMSCollector* collector, 3035 CompactibleFreeListSpace* cms_space, 3036 YieldingFlexibleWorkGang* workers, 3037 OopTaskQueueSet* task_queues): 3038 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3039 _collector(collector), 3040 _cms_space(cms_space), 3041 _n_workers(0), _result(true), 3042 _task_queues(task_queues), 3043 _term(_n_workers, task_queues, _collector), 3044 _bit_map_lock(collector->bitMapLock()) 3045 { 3046 _requested_size = _n_workers; 3047 _term.set_task(this); 3048 _term_term.set_task(this); 3049 _restart_addr = _global_finger = _cms_space->bottom(); 3050 } 3051 3052 3053 OopTaskQueueSet* task_queues() { return _task_queues; } 3054 3055 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3056 3057 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3058 3059 CMSConcMarkingTerminator* terminator() { return &_term; } 3060 3061 virtual void set_for_termination(uint active_workers) { 3062 terminator()->reset_for_reuse(active_workers); 3063 } 3064 3065 void work(uint worker_id); 3066 bool should_yield() { 3067 return ConcurrentMarkSweepThread::should_yield() 3068 && !_collector->foregroundGCIsActive(); 3069 } 3070 3071 virtual void coordinator_yield(); // stuff done by coordinator 3072 bool result() { return _result; } 3073 3074 void reset(HeapWord* ra) { 3075 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3076 _restart_addr = _global_finger = ra; 3077 _term.reset_for_reuse(); 3078 } 3079 3080 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3081 OopTaskQueue* work_q); 3082 3083 private: 3084 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3085 void do_work_steal(int i); 3086 void bump_global_finger(HeapWord* f); 3087 }; 3088 3089 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3090 assert(_task != NULL, "Error"); 3091 return _task->yielding(); 3092 // Note that we do not need the disjunct || _task->should_yield() above 3093 // because we want terminating threads to yield only if the task 3094 // is already in the midst of yielding, which happens only after at least one 3095 // thread has yielded. 3096 } 3097 3098 void CMSConcMarkingTerminator::yield() { 3099 if (_task->should_yield()) { 3100 _task->yield(); 3101 } else { 3102 ParallelTaskTerminator::yield(); 3103 } 3104 } 3105 3106 //////////////////////////////////////////////////////////////// 3107 // Concurrent Marking Algorithm Sketch 3108 //////////////////////////////////////////////////////////////// 3109 // Until all tasks exhausted (both spaces): 3110 // -- claim next available chunk 3111 // -- bump global finger via CAS 3112 // -- find first object that starts in this chunk 3113 // and start scanning bitmap from that position 3114 // -- scan marked objects for oops 3115 // -- CAS-mark target, and if successful: 3116 // . if target oop is above global finger (volatile read) 3117 // nothing to do 3118 // . if target oop is in chunk and above local finger 3119 // then nothing to do 3120 // . else push on work-queue 3121 // -- Deal with possible overflow issues: 3122 // . local work-queue overflow causes stuff to be pushed on 3123 // global (common) overflow queue 3124 // . always first empty local work queue 3125 // . then get a batch of oops from global work queue if any 3126 // . then do work stealing 3127 // -- When all tasks claimed (both spaces) 3128 // and local work queue empty, 3129 // then in a loop do: 3130 // . check global overflow stack; steal a batch of oops and trace 3131 // . try to steal from other threads oif GOS is empty 3132 // . if neither is available, offer termination 3133 // -- Terminate and return result 3134 // 3135 void CMSConcMarkingTask::work(uint worker_id) { 3136 elapsedTimer _timer; 3137 ResourceMark rm; 3138 HandleMark hm; 3139 3140 DEBUG_ONLY(_collector->verify_overflow_empty();) 3141 3142 // Before we begin work, our work queue should be empty 3143 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3144 // Scan the bitmap covering _cms_space, tracing through grey objects. 3145 _timer.start(); 3146 do_scan_and_mark(worker_id, _cms_space); 3147 _timer.stop(); 3148 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3149 3150 // ... do work stealing 3151 _timer.reset(); 3152 _timer.start(); 3153 do_work_steal(worker_id); 3154 _timer.stop(); 3155 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3156 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3157 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3158 // Note that under the current task protocol, the 3159 // following assertion is true even of the spaces 3160 // expanded since the completion of the concurrent 3161 // marking. XXX This will likely change under a strict 3162 // ABORT semantics. 3163 // After perm removal the comparison was changed to 3164 // greater than or equal to from strictly greater than. 3165 // Before perm removal the highest address sweep would 3166 // have been at the end of perm gen but now is at the 3167 // end of the tenured gen. 3168 assert(_global_finger >= _cms_space->end(), 3169 "All tasks have been completed"); 3170 DEBUG_ONLY(_collector->verify_overflow_empty();) 3171 } 3172 3173 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3174 HeapWord* read = _global_finger; 3175 HeapWord* cur = read; 3176 while (f > read) { 3177 cur = read; 3178 read = Atomic::cmpxchg(f, &_global_finger, cur); 3179 if (cur == read) { 3180 // our cas succeeded 3181 assert(_global_finger >= f, "protocol consistency"); 3182 break; 3183 } 3184 } 3185 } 3186 3187 // This is really inefficient, and should be redone by 3188 // using (not yet available) block-read and -write interfaces to the 3189 // stack and the work_queue. XXX FIX ME !!! 3190 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3191 OopTaskQueue* work_q) { 3192 // Fast lock-free check 3193 if (ovflw_stk->length() == 0) { 3194 return false; 3195 } 3196 assert(work_q->size() == 0, "Shouldn't steal"); 3197 MutexLockerEx ml(ovflw_stk->par_lock(), 3198 Mutex::_no_safepoint_check_flag); 3199 // Grab up to 1/4 the size of the work queue 3200 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3201 (size_t)ParGCDesiredObjsFromOverflowList); 3202 num = MIN2(num, ovflw_stk->length()); 3203 for (int i = (int) num; i > 0; i--) { 3204 oop cur = ovflw_stk->pop(); 3205 assert(cur != NULL, "Counted wrong?"); 3206 work_q->push(cur); 3207 } 3208 return num > 0; 3209 } 3210 3211 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3212 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3213 int n_tasks = pst->n_tasks(); 3214 // We allow that there may be no tasks to do here because 3215 // we are restarting after a stack overflow. 3216 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3217 uint nth_task = 0; 3218 3219 HeapWord* aligned_start = sp->bottom(); 3220 if (sp->used_region().contains(_restart_addr)) { 3221 // Align down to a card boundary for the start of 0th task 3222 // for this space. 3223 aligned_start = align_down(_restart_addr, CardTable::card_size); 3224 } 3225 3226 size_t chunk_size = sp->marking_task_size(); 3227 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3228 // Having claimed the nth task in this space, 3229 // compute the chunk that it corresponds to: 3230 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3231 aligned_start + (nth_task+1)*chunk_size); 3232 // Try and bump the global finger via a CAS; 3233 // note that we need to do the global finger bump 3234 // _before_ taking the intersection below, because 3235 // the task corresponding to that region will be 3236 // deemed done even if the used_region() expands 3237 // because of allocation -- as it almost certainly will 3238 // during start-up while the threads yield in the 3239 // closure below. 3240 HeapWord* finger = span.end(); 3241 bump_global_finger(finger); // atomically 3242 // There are null tasks here corresponding to chunks 3243 // beyond the "top" address of the space. 3244 span = span.intersection(sp->used_region()); 3245 if (!span.is_empty()) { // Non-null task 3246 HeapWord* prev_obj; 3247 assert(!span.contains(_restart_addr) || nth_task == 0, 3248 "Inconsistency"); 3249 if (nth_task == 0) { 3250 // For the 0th task, we'll not need to compute a block_start. 3251 if (span.contains(_restart_addr)) { 3252 // In the case of a restart because of stack overflow, 3253 // we might additionally skip a chunk prefix. 3254 prev_obj = _restart_addr; 3255 } else { 3256 prev_obj = span.start(); 3257 } 3258 } else { 3259 // We want to skip the first object because 3260 // the protocol is to scan any object in its entirety 3261 // that _starts_ in this span; a fortiori, any 3262 // object starting in an earlier span is scanned 3263 // as part of an earlier claimed task. 3264 // Below we use the "careful" version of block_start 3265 // so we do not try to navigate uninitialized objects. 3266 prev_obj = sp->block_start_careful(span.start()); 3267 // Below we use a variant of block_size that uses the 3268 // Printezis bits to avoid waiting for allocated 3269 // objects to become initialized/parsable. 3270 while (prev_obj < span.start()) { 3271 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3272 if (sz > 0) { 3273 prev_obj += sz; 3274 } else { 3275 // In this case we may end up doing a bit of redundant 3276 // scanning, but that appears unavoidable, short of 3277 // locking the free list locks; see bug 6324141. 3278 break; 3279 } 3280 } 3281 } 3282 if (prev_obj < span.end()) { 3283 MemRegion my_span = MemRegion(prev_obj, span.end()); 3284 // Do the marking work within a non-empty span -- 3285 // the last argument to the constructor indicates whether the 3286 // iteration should be incremental with periodic yields. 3287 ParMarkFromRootsClosure cl(this, _collector, my_span, 3288 &_collector->_markBitMap, 3289 work_queue(i), 3290 &_collector->_markStack); 3291 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3292 } // else nothing to do for this task 3293 } // else nothing to do for this task 3294 } 3295 // We'd be tempted to assert here that since there are no 3296 // more tasks left to claim in this space, the global_finger 3297 // must exceed space->top() and a fortiori space->end(). However, 3298 // that would not quite be correct because the bumping of 3299 // global_finger occurs strictly after the claiming of a task, 3300 // so by the time we reach here the global finger may not yet 3301 // have been bumped up by the thread that claimed the last 3302 // task. 3303 pst->all_tasks_completed(); 3304 } 3305 3306 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3307 private: 3308 CMSCollector* _collector; 3309 CMSConcMarkingTask* _task; 3310 MemRegion _span; 3311 CMSBitMap* _bit_map; 3312 CMSMarkStack* _overflow_stack; 3313 OopTaskQueue* _work_queue; 3314 protected: 3315 DO_OOP_WORK_DEFN 3316 public: 3317 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3318 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3319 MetadataAwareOopClosure(collector->ref_processor()), 3320 _collector(collector), 3321 _task(task), 3322 _span(collector->_span), 3323 _work_queue(work_queue), 3324 _bit_map(bit_map), 3325 _overflow_stack(overflow_stack) 3326 { } 3327 virtual void do_oop(oop* p); 3328 virtual void do_oop(narrowOop* p); 3329 3330 void trim_queue(size_t max); 3331 void handle_stack_overflow(HeapWord* lost); 3332 void do_yield_check() { 3333 if (_task->should_yield()) { 3334 _task->yield(); 3335 } 3336 } 3337 }; 3338 3339 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3340 3341 // Grey object scanning during work stealing phase -- 3342 // the salient assumption here is that any references 3343 // that are in these stolen objects being scanned must 3344 // already have been initialized (else they would not have 3345 // been published), so we do not need to check for 3346 // uninitialized objects before pushing here. 3347 void ParConcMarkingClosure::do_oop(oop obj) { 3348 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3349 HeapWord* addr = (HeapWord*)obj; 3350 // Check if oop points into the CMS generation 3351 // and is not marked 3352 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3353 // a white object ... 3354 // If we manage to "claim" the object, by being the 3355 // first thread to mark it, then we push it on our 3356 // marking stack 3357 if (_bit_map->par_mark(addr)) { // ... now grey 3358 // push on work queue (grey set) 3359 bool simulate_overflow = false; 3360 NOT_PRODUCT( 3361 if (CMSMarkStackOverflowALot && 3362 _collector->simulate_overflow()) { 3363 // simulate a stack overflow 3364 simulate_overflow = true; 3365 } 3366 ) 3367 if (simulate_overflow || 3368 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3369 // stack overflow 3370 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3371 // We cannot assert that the overflow stack is full because 3372 // it may have been emptied since. 3373 assert(simulate_overflow || 3374 _work_queue->size() == _work_queue->max_elems(), 3375 "Else push should have succeeded"); 3376 handle_stack_overflow(addr); 3377 } 3378 } // Else, some other thread got there first 3379 do_yield_check(); 3380 } 3381 } 3382 3383 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3384 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3385 3386 void ParConcMarkingClosure::trim_queue(size_t max) { 3387 while (_work_queue->size() > max) { 3388 oop new_oop; 3389 if (_work_queue->pop_local(new_oop)) { 3390 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3391 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3392 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3393 new_oop->oop_iterate(this); // do_oop() above 3394 do_yield_check(); 3395 } 3396 } 3397 } 3398 3399 // Upon stack overflow, we discard (part of) the stack, 3400 // remembering the least address amongst those discarded 3401 // in CMSCollector's _restart_address. 3402 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3403 // We need to do this under a mutex to prevent other 3404 // workers from interfering with the work done below. 3405 MutexLockerEx ml(_overflow_stack->par_lock(), 3406 Mutex::_no_safepoint_check_flag); 3407 // Remember the least grey address discarded 3408 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3409 _collector->lower_restart_addr(ra); 3410 _overflow_stack->reset(); // discard stack contents 3411 _overflow_stack->expand(); // expand the stack if possible 3412 } 3413 3414 3415 void CMSConcMarkingTask::do_work_steal(int i) { 3416 OopTaskQueue* work_q = work_queue(i); 3417 oop obj_to_scan; 3418 CMSBitMap* bm = &(_collector->_markBitMap); 3419 CMSMarkStack* ovflw = &(_collector->_markStack); 3420 int* seed = _collector->hash_seed(i); 3421 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3422 while (true) { 3423 cl.trim_queue(0); 3424 assert(work_q->size() == 0, "Should have been emptied above"); 3425 if (get_work_from_overflow_stack(ovflw, work_q)) { 3426 // Can't assert below because the work obtained from the 3427 // overflow stack may already have been stolen from us. 3428 // assert(work_q->size() > 0, "Work from overflow stack"); 3429 continue; 3430 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3431 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3432 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3433 obj_to_scan->oop_iterate(&cl); 3434 } else if (terminator()->offer_termination(&_term_term)) { 3435 assert(work_q->size() == 0, "Impossible!"); 3436 break; 3437 } else if (yielding() || should_yield()) { 3438 yield(); 3439 } 3440 } 3441 } 3442 3443 // This is run by the CMS (coordinator) thread. 3444 void CMSConcMarkingTask::coordinator_yield() { 3445 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3446 "CMS thread should hold CMS token"); 3447 // First give up the locks, then yield, then re-lock 3448 // We should probably use a constructor/destructor idiom to 3449 // do this unlock/lock or modify the MutexUnlocker class to 3450 // serve our purpose. XXX 3451 assert_lock_strong(_bit_map_lock); 3452 _bit_map_lock->unlock(); 3453 ConcurrentMarkSweepThread::desynchronize(true); 3454 _collector->stopTimer(); 3455 _collector->incrementYields(); 3456 3457 // It is possible for whichever thread initiated the yield request 3458 // not to get a chance to wake up and take the bitmap lock between 3459 // this thread releasing it and reacquiring it. So, while the 3460 // should_yield() flag is on, let's sleep for a bit to give the 3461 // other thread a chance to wake up. The limit imposed on the number 3462 // of iterations is defensive, to avoid any unforseen circumstances 3463 // putting us into an infinite loop. Since it's always been this 3464 // (coordinator_yield()) method that was observed to cause the 3465 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3466 // which is by default non-zero. For the other seven methods that 3467 // also perform the yield operation, as are using a different 3468 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3469 // can enable the sleeping for those methods too, if necessary. 3470 // See 6442774. 3471 // 3472 // We really need to reconsider the synchronization between the GC 3473 // thread and the yield-requesting threads in the future and we 3474 // should really use wait/notify, which is the recommended 3475 // way of doing this type of interaction. Additionally, we should 3476 // consolidate the eight methods that do the yield operation and they 3477 // are almost identical into one for better maintainability and 3478 // readability. See 6445193. 3479 // 3480 // Tony 2006.06.29 3481 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3482 ConcurrentMarkSweepThread::should_yield() && 3483 !CMSCollector::foregroundGCIsActive(); ++i) { 3484 os::sleep(Thread::current(), 1, false); 3485 } 3486 3487 ConcurrentMarkSweepThread::synchronize(true); 3488 _bit_map_lock->lock_without_safepoint_check(); 3489 _collector->startTimer(); 3490 } 3491 3492 bool CMSCollector::do_marking_mt() { 3493 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3494 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3495 conc_workers()->active_workers(), 3496 Threads::number_of_non_daemon_threads()); 3497 num_workers = conc_workers()->update_active_workers(num_workers); 3498 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3499 3500 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3501 3502 CMSConcMarkingTask tsk(this, 3503 cms_space, 3504 conc_workers(), 3505 task_queues()); 3506 3507 // Since the actual number of workers we get may be different 3508 // from the number we requested above, do we need to do anything different 3509 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3510 // class?? XXX 3511 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3512 3513 // Refs discovery is already non-atomic. 3514 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3515 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3516 conc_workers()->start_task(&tsk); 3517 while (tsk.yielded()) { 3518 tsk.coordinator_yield(); 3519 conc_workers()->continue_task(&tsk); 3520 } 3521 // If the task was aborted, _restart_addr will be non-NULL 3522 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3523 while (_restart_addr != NULL) { 3524 // XXX For now we do not make use of ABORTED state and have not 3525 // yet implemented the right abort semantics (even in the original 3526 // single-threaded CMS case). That needs some more investigation 3527 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3528 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3529 // If _restart_addr is non-NULL, a marking stack overflow 3530 // occurred; we need to do a fresh marking iteration from the 3531 // indicated restart address. 3532 if (_foregroundGCIsActive) { 3533 // We may be running into repeated stack overflows, having 3534 // reached the limit of the stack size, while making very 3535 // slow forward progress. It may be best to bail out and 3536 // let the foreground collector do its job. 3537 // Clear _restart_addr, so that foreground GC 3538 // works from scratch. This avoids the headache of 3539 // a "rescan" which would otherwise be needed because 3540 // of the dirty mod union table & card table. 3541 _restart_addr = NULL; 3542 return false; 3543 } 3544 // Adjust the task to restart from _restart_addr 3545 tsk.reset(_restart_addr); 3546 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3547 _restart_addr); 3548 _restart_addr = NULL; 3549 // Get the workers going again 3550 conc_workers()->start_task(&tsk); 3551 while (tsk.yielded()) { 3552 tsk.coordinator_yield(); 3553 conc_workers()->continue_task(&tsk); 3554 } 3555 } 3556 assert(tsk.completed(), "Inconsistency"); 3557 assert(tsk.result() == true, "Inconsistency"); 3558 return true; 3559 } 3560 3561 bool CMSCollector::do_marking_st() { 3562 ResourceMark rm; 3563 HandleMark hm; 3564 3565 // Temporarily make refs discovery single threaded (non-MT) 3566 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3567 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3568 &_markStack, CMSYield); 3569 // the last argument to iterate indicates whether the iteration 3570 // should be incremental with periodic yields. 3571 _markBitMap.iterate(&markFromRootsClosure); 3572 // If _restart_addr is non-NULL, a marking stack overflow 3573 // occurred; we need to do a fresh iteration from the 3574 // indicated restart address. 3575 while (_restart_addr != NULL) { 3576 if (_foregroundGCIsActive) { 3577 // We may be running into repeated stack overflows, having 3578 // reached the limit of the stack size, while making very 3579 // slow forward progress. It may be best to bail out and 3580 // let the foreground collector do its job. 3581 // Clear _restart_addr, so that foreground GC 3582 // works from scratch. This avoids the headache of 3583 // a "rescan" which would otherwise be needed because 3584 // of the dirty mod union table & card table. 3585 _restart_addr = NULL; 3586 return false; // indicating failure to complete marking 3587 } 3588 // Deal with stack overflow: 3589 // we restart marking from _restart_addr 3590 HeapWord* ra = _restart_addr; 3591 markFromRootsClosure.reset(ra); 3592 _restart_addr = NULL; 3593 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3594 } 3595 return true; 3596 } 3597 3598 void CMSCollector::preclean() { 3599 check_correct_thread_executing(); 3600 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3601 verify_work_stacks_empty(); 3602 verify_overflow_empty(); 3603 _abort_preclean = false; 3604 if (CMSPrecleaningEnabled) { 3605 if (!CMSEdenChunksRecordAlways) { 3606 _eden_chunk_index = 0; 3607 } 3608 size_t used = get_eden_used(); 3609 size_t capacity = get_eden_capacity(); 3610 // Don't start sampling unless we will get sufficiently 3611 // many samples. 3612 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3613 * CMSScheduleRemarkEdenPenetration)) { 3614 _start_sampling = true; 3615 } else { 3616 _start_sampling = false; 3617 } 3618 GCTraceCPUTime tcpu; 3619 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3620 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3621 } 3622 CMSTokenSync x(true); // is cms thread 3623 if (CMSPrecleaningEnabled) { 3624 sample_eden(); 3625 _collectorState = AbortablePreclean; 3626 } else { 3627 _collectorState = FinalMarking; 3628 } 3629 verify_work_stacks_empty(); 3630 verify_overflow_empty(); 3631 } 3632 3633 // Try and schedule the remark such that young gen 3634 // occupancy is CMSScheduleRemarkEdenPenetration %. 3635 void CMSCollector::abortable_preclean() { 3636 check_correct_thread_executing(); 3637 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3638 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3639 3640 // If Eden's current occupancy is below this threshold, 3641 // immediately schedule the remark; else preclean 3642 // past the next scavenge in an effort to 3643 // schedule the pause as described above. By choosing 3644 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3645 // we will never do an actual abortable preclean cycle. 3646 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3647 GCTraceCPUTime tcpu; 3648 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3649 // We need more smarts in the abortable preclean 3650 // loop below to deal with cases where allocation 3651 // in young gen is very very slow, and our precleaning 3652 // is running a losing race against a horde of 3653 // mutators intent on flooding us with CMS updates 3654 // (dirty cards). 3655 // One, admittedly dumb, strategy is to give up 3656 // after a certain number of abortable precleaning loops 3657 // or after a certain maximum time. We want to make 3658 // this smarter in the next iteration. 3659 // XXX FIX ME!!! YSR 3660 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3661 while (!(should_abort_preclean() || 3662 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3663 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3664 cumworkdone += workdone; 3665 loops++; 3666 // Voluntarily terminate abortable preclean phase if we have 3667 // been at it for too long. 3668 if ((CMSMaxAbortablePrecleanLoops != 0) && 3669 loops >= CMSMaxAbortablePrecleanLoops) { 3670 log_debug(gc)(" CMS: abort preclean due to loops "); 3671 break; 3672 } 3673 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3674 log_debug(gc)(" CMS: abort preclean due to time "); 3675 break; 3676 } 3677 // If we are doing little work each iteration, we should 3678 // take a short break. 3679 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3680 // Sleep for some time, waiting for work to accumulate 3681 stopTimer(); 3682 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3683 startTimer(); 3684 waited++; 3685 } 3686 } 3687 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3688 loops, waited, cumworkdone); 3689 } 3690 CMSTokenSync x(true); // is cms thread 3691 if (_collectorState != Idling) { 3692 assert(_collectorState == AbortablePreclean, 3693 "Spontaneous state transition?"); 3694 _collectorState = FinalMarking; 3695 } // Else, a foreground collection completed this CMS cycle. 3696 return; 3697 } 3698 3699 // Respond to an Eden sampling opportunity 3700 void CMSCollector::sample_eden() { 3701 // Make sure a young gc cannot sneak in between our 3702 // reading and recording of a sample. 3703 assert(Thread::current()->is_ConcurrentGC_thread(), 3704 "Only the cms thread may collect Eden samples"); 3705 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3706 "Should collect samples while holding CMS token"); 3707 if (!_start_sampling) { 3708 return; 3709 } 3710 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3711 // is populated by the young generation. 3712 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3713 if (_eden_chunk_index < _eden_chunk_capacity) { 3714 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3715 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3716 "Unexpected state of Eden"); 3717 // We'd like to check that what we just sampled is an oop-start address; 3718 // however, we cannot do that here since the object may not yet have been 3719 // initialized. So we'll instead do the check when we _use_ this sample 3720 // later. 3721 if (_eden_chunk_index == 0 || 3722 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3723 _eden_chunk_array[_eden_chunk_index-1]) 3724 >= CMSSamplingGrain)) { 3725 _eden_chunk_index++; // commit sample 3726 } 3727 } 3728 } 3729 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3730 size_t used = get_eden_used(); 3731 size_t capacity = get_eden_capacity(); 3732 assert(used <= capacity, "Unexpected state of Eden"); 3733 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3734 _abort_preclean = true; 3735 } 3736 } 3737 } 3738 3739 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3740 assert(_collectorState == Precleaning || 3741 _collectorState == AbortablePreclean, "incorrect state"); 3742 ResourceMark rm; 3743 HandleMark hm; 3744 3745 // Precleaning is currently not MT but the reference processor 3746 // may be set for MT. Disable it temporarily here. 3747 ReferenceProcessor* rp = ref_processor(); 3748 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3749 3750 // Do one pass of scrubbing the discovered reference lists 3751 // to remove any reference objects with strongly-reachable 3752 // referents. 3753 if (clean_refs) { 3754 CMSPrecleanRefsYieldClosure yield_cl(this); 3755 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3756 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3757 &_markStack, true /* preclean */); 3758 CMSDrainMarkingStackClosure complete_trace(this, 3759 _span, &_markBitMap, &_markStack, 3760 &keep_alive, true /* preclean */); 3761 3762 // We don't want this step to interfere with a young 3763 // collection because we don't want to take CPU 3764 // or memory bandwidth away from the young GC threads 3765 // (which may be as many as there are CPUs). 3766 // Note that we don't need to protect ourselves from 3767 // interference with mutators because they can't 3768 // manipulate the discovered reference lists nor affect 3769 // the computed reachability of the referents, the 3770 // only properties manipulated by the precleaning 3771 // of these reference lists. 3772 stopTimer(); 3773 CMSTokenSyncWithLocks x(true /* is cms thread */, 3774 bitMapLock()); 3775 startTimer(); 3776 sample_eden(); 3777 3778 // The following will yield to allow foreground 3779 // collection to proceed promptly. XXX YSR: 3780 // The code in this method may need further 3781 // tweaking for better performance and some restructuring 3782 // for cleaner interfaces. 3783 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3784 rp->preclean_discovered_references( 3785 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3786 gc_timer); 3787 } 3788 3789 if (clean_survivor) { // preclean the active survivor space(s) 3790 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3791 &_markBitMap, &_modUnionTable, 3792 &_markStack, true /* precleaning phase */); 3793 stopTimer(); 3794 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3795 bitMapLock()); 3796 startTimer(); 3797 unsigned int before_count = 3798 CMSHeap::heap()->total_collections(); 3799 SurvivorSpacePrecleanClosure 3800 sss_cl(this, _span, &_markBitMap, &_markStack, 3801 &pam_cl, before_count, CMSYield); 3802 _young_gen->from()->object_iterate_careful(&sss_cl); 3803 _young_gen->to()->object_iterate_careful(&sss_cl); 3804 } 3805 MarkRefsIntoAndScanClosure 3806 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3807 &_markStack, this, CMSYield, 3808 true /* precleaning phase */); 3809 // CAUTION: The following closure has persistent state that may need to 3810 // be reset upon a decrease in the sequence of addresses it 3811 // processes. 3812 ScanMarkedObjectsAgainCarefullyClosure 3813 smoac_cl(this, _span, 3814 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3815 3816 // Preclean dirty cards in ModUnionTable and CardTable using 3817 // appropriate convergence criterion; 3818 // repeat CMSPrecleanIter times unless we find that 3819 // we are losing. 3820 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3821 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3822 "Bad convergence multiplier"); 3823 assert(CMSPrecleanThreshold >= 100, 3824 "Unreasonably low CMSPrecleanThreshold"); 3825 3826 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3827 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3828 numIter < CMSPrecleanIter; 3829 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3830 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3831 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3832 // Either there are very few dirty cards, so re-mark 3833 // pause will be small anyway, or our pre-cleaning isn't 3834 // that much faster than the rate at which cards are being 3835 // dirtied, so we might as well stop and re-mark since 3836 // precleaning won't improve our re-mark time by much. 3837 if (curNumCards <= CMSPrecleanThreshold || 3838 (numIter > 0 && 3839 (curNumCards * CMSPrecleanDenominator > 3840 lastNumCards * CMSPrecleanNumerator))) { 3841 numIter++; 3842 cumNumCards += curNumCards; 3843 break; 3844 } 3845 } 3846 3847 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3848 3849 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3850 cumNumCards += curNumCards; 3851 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3852 curNumCards, cumNumCards, numIter); 3853 return cumNumCards; // as a measure of useful work done 3854 } 3855 3856 // PRECLEANING NOTES: 3857 // Precleaning involves: 3858 // . reading the bits of the modUnionTable and clearing the set bits. 3859 // . For the cards corresponding to the set bits, we scan the 3860 // objects on those cards. This means we need the free_list_lock 3861 // so that we can safely iterate over the CMS space when scanning 3862 // for oops. 3863 // . When we scan the objects, we'll be both reading and setting 3864 // marks in the marking bit map, so we'll need the marking bit map. 3865 // . For protecting _collector_state transitions, we take the CGC_lock. 3866 // Note that any races in the reading of of card table entries by the 3867 // CMS thread on the one hand and the clearing of those entries by the 3868 // VM thread or the setting of those entries by the mutator threads on the 3869 // other are quite benign. However, for efficiency it makes sense to keep 3870 // the VM thread from racing with the CMS thread while the latter is 3871 // dirty card info to the modUnionTable. We therefore also use the 3872 // CGC_lock to protect the reading of the card table and the mod union 3873 // table by the CM thread. 3874 // . We run concurrently with mutator updates, so scanning 3875 // needs to be done carefully -- we should not try to scan 3876 // potentially uninitialized objects. 3877 // 3878 // Locking strategy: While holding the CGC_lock, we scan over and 3879 // reset a maximal dirty range of the mod union / card tables, then lock 3880 // the free_list_lock and bitmap lock to do a full marking, then 3881 // release these locks; and repeat the cycle. This allows for a 3882 // certain amount of fairness in the sharing of these locks between 3883 // the CMS collector on the one hand, and the VM thread and the 3884 // mutators on the other. 3885 3886 // NOTE: preclean_mod_union_table() and preclean_card_table() 3887 // further below are largely identical; if you need to modify 3888 // one of these methods, please check the other method too. 3889 3890 size_t CMSCollector::preclean_mod_union_table( 3891 ConcurrentMarkSweepGeneration* old_gen, 3892 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3893 verify_work_stacks_empty(); 3894 verify_overflow_empty(); 3895 3896 // strategy: starting with the first card, accumulate contiguous 3897 // ranges of dirty cards; clear these cards, then scan the region 3898 // covered by these cards. 3899 3900 // Since all of the MUT is committed ahead, we can just use 3901 // that, in case the generations expand while we are precleaning. 3902 // It might also be fine to just use the committed part of the 3903 // generation, but we might potentially miss cards when the 3904 // generation is rapidly expanding while we are in the midst 3905 // of precleaning. 3906 HeapWord* startAddr = old_gen->reserved().start(); 3907 HeapWord* endAddr = old_gen->reserved().end(); 3908 3909 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3910 3911 size_t numDirtyCards, cumNumDirtyCards; 3912 HeapWord *nextAddr, *lastAddr; 3913 for (cumNumDirtyCards = numDirtyCards = 0, 3914 nextAddr = lastAddr = startAddr; 3915 nextAddr < endAddr; 3916 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3917 3918 ResourceMark rm; 3919 HandleMark hm; 3920 3921 MemRegion dirtyRegion; 3922 { 3923 stopTimer(); 3924 // Potential yield point 3925 CMSTokenSync ts(true); 3926 startTimer(); 3927 sample_eden(); 3928 // Get dirty region starting at nextOffset (inclusive), 3929 // simultaneously clearing it. 3930 dirtyRegion = 3931 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3932 assert(dirtyRegion.start() >= nextAddr, 3933 "returned region inconsistent?"); 3934 } 3935 // Remember where the next search should begin. 3936 // The returned region (if non-empty) is a right open interval, 3937 // so lastOffset is obtained from the right end of that 3938 // interval. 3939 lastAddr = dirtyRegion.end(); 3940 // Should do something more transparent and less hacky XXX 3941 numDirtyCards = 3942 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3943 3944 // We'll scan the cards in the dirty region (with periodic 3945 // yields for foreground GC as needed). 3946 if (!dirtyRegion.is_empty()) { 3947 assert(numDirtyCards > 0, "consistency check"); 3948 HeapWord* stop_point = NULL; 3949 stopTimer(); 3950 // Potential yield point 3951 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3952 bitMapLock()); 3953 startTimer(); 3954 { 3955 verify_work_stacks_empty(); 3956 verify_overflow_empty(); 3957 sample_eden(); 3958 stop_point = 3959 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3960 } 3961 if (stop_point != NULL) { 3962 // The careful iteration stopped early either because it found an 3963 // uninitialized object, or because we were in the midst of an 3964 // "abortable preclean", which should now be aborted. Redirty 3965 // the bits corresponding to the partially-scanned or unscanned 3966 // cards. We'll either restart at the next block boundary or 3967 // abort the preclean. 3968 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3969 "Should only be AbortablePreclean."); 3970 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3971 if (should_abort_preclean()) { 3972 break; // out of preclean loop 3973 } else { 3974 // Compute the next address at which preclean should pick up; 3975 // might need bitMapLock in order to read P-bits. 3976 lastAddr = next_card_start_after_block(stop_point); 3977 } 3978 } 3979 } else { 3980 assert(lastAddr == endAddr, "consistency check"); 3981 assert(numDirtyCards == 0, "consistency check"); 3982 break; 3983 } 3984 } 3985 verify_work_stacks_empty(); 3986 verify_overflow_empty(); 3987 return cumNumDirtyCards; 3988 } 3989 3990 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3991 // below are largely identical; if you need to modify 3992 // one of these methods, please check the other method too. 3993 3994 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3995 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3996 // strategy: it's similar to precleamModUnionTable above, in that 3997 // we accumulate contiguous ranges of dirty cards, mark these cards 3998 // precleaned, then scan the region covered by these cards. 3999 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4000 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4001 4002 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4003 4004 size_t numDirtyCards, cumNumDirtyCards; 4005 HeapWord *lastAddr, *nextAddr; 4006 4007 for (cumNumDirtyCards = numDirtyCards = 0, 4008 nextAddr = lastAddr = startAddr; 4009 nextAddr < endAddr; 4010 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4011 4012 ResourceMark rm; 4013 HandleMark hm; 4014 4015 MemRegion dirtyRegion; 4016 { 4017 // See comments in "Precleaning notes" above on why we 4018 // do this locking. XXX Could the locking overheads be 4019 // too high when dirty cards are sparse? [I don't think so.] 4020 stopTimer(); 4021 CMSTokenSync x(true); // is cms thread 4022 startTimer(); 4023 sample_eden(); 4024 // Get and clear dirty region from card table 4025 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4026 true, 4027 CardTable::precleaned_card_val()); 4028 4029 assert(dirtyRegion.start() >= nextAddr, 4030 "returned region inconsistent?"); 4031 } 4032 lastAddr = dirtyRegion.end(); 4033 numDirtyCards = 4034 dirtyRegion.word_size()/CardTable::card_size_in_words; 4035 4036 if (!dirtyRegion.is_empty()) { 4037 stopTimer(); 4038 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4039 startTimer(); 4040 sample_eden(); 4041 verify_work_stacks_empty(); 4042 verify_overflow_empty(); 4043 HeapWord* stop_point = 4044 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4045 if (stop_point != NULL) { 4046 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4047 "Should only be AbortablePreclean."); 4048 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4049 if (should_abort_preclean()) { 4050 break; // out of preclean loop 4051 } else { 4052 // Compute the next address at which preclean should pick up. 4053 lastAddr = next_card_start_after_block(stop_point); 4054 } 4055 } 4056 } else { 4057 break; 4058 } 4059 } 4060 verify_work_stacks_empty(); 4061 verify_overflow_empty(); 4062 return cumNumDirtyCards; 4063 } 4064 4065 class PrecleanCLDClosure : public CLDClosure { 4066 MetadataAwareOopsInGenClosure* _cm_closure; 4067 public: 4068 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4069 void do_cld(ClassLoaderData* cld) { 4070 if (cld->has_accumulated_modified_oops()) { 4071 cld->clear_accumulated_modified_oops(); 4072 4073 _cm_closure->do_cld(cld); 4074 } 4075 } 4076 }; 4077 4078 // The freelist lock is needed to prevent asserts, is it really needed? 4079 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4080 4081 cl->set_freelistLock(freelistLock); 4082 4083 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4084 4085 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4086 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4087 PrecleanCLDClosure preclean_closure(cl); 4088 ClassLoaderDataGraph::cld_do(&preclean_closure); 4089 4090 verify_work_stacks_empty(); 4091 verify_overflow_empty(); 4092 } 4093 4094 void CMSCollector::checkpointRootsFinal() { 4095 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4096 check_correct_thread_executing(); 4097 // world is stopped at this checkpoint 4098 assert(SafepointSynchronize::is_at_safepoint(), 4099 "world should be stopped"); 4100 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4101 4102 verify_work_stacks_empty(); 4103 verify_overflow_empty(); 4104 4105 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4106 _young_gen->used() / K, _young_gen->capacity() / K); 4107 { 4108 if (CMSScavengeBeforeRemark) { 4109 CMSHeap* heap = CMSHeap::heap(); 4110 // Temporarily set flag to false, GCH->do_collection will 4111 // expect it to be false and set to true 4112 FlagSetting fl(heap->_is_gc_active, false); 4113 4114 heap->do_collection(true, // full (i.e. force, see below) 4115 false, // !clear_all_soft_refs 4116 0, // size 4117 false, // is_tlab 4118 GenCollectedHeap::YoungGen // type 4119 ); 4120 } 4121 FreelistLocker x(this); 4122 MutexLockerEx y(bitMapLock(), 4123 Mutex::_no_safepoint_check_flag); 4124 checkpointRootsFinalWork(); 4125 } 4126 verify_work_stacks_empty(); 4127 verify_overflow_empty(); 4128 } 4129 4130 void CMSCollector::checkpointRootsFinalWork() { 4131 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4132 4133 assert(haveFreelistLocks(), "must have free list locks"); 4134 assert_lock_strong(bitMapLock()); 4135 4136 ResourceMark rm; 4137 HandleMark hm; 4138 4139 CMSHeap* heap = CMSHeap::heap(); 4140 4141 if (should_unload_classes()) { 4142 CodeCache::gc_prologue(); 4143 } 4144 assert(haveFreelistLocks(), "must have free list locks"); 4145 assert_lock_strong(bitMapLock()); 4146 4147 // We might assume that we need not fill TLAB's when 4148 // CMSScavengeBeforeRemark is set, because we may have just done 4149 // a scavenge which would have filled all TLAB's -- and besides 4150 // Eden would be empty. This however may not always be the case -- 4151 // for instance although we asked for a scavenge, it may not have 4152 // happened because of a JNI critical section. We probably need 4153 // a policy for deciding whether we can in that case wait until 4154 // the critical section releases and then do the remark following 4155 // the scavenge, and skip it here. In the absence of that policy, 4156 // or of an indication of whether the scavenge did indeed occur, 4157 // we cannot rely on TLAB's having been filled and must do 4158 // so here just in case a scavenge did not happen. 4159 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4160 // Update the saved marks which may affect the root scans. 4161 heap->save_marks(); 4162 4163 print_eden_and_survivor_chunk_arrays(); 4164 4165 { 4166 #if COMPILER2_OR_JVMCI 4167 DerivedPointerTableDeactivate dpt_deact; 4168 #endif 4169 4170 // Note on the role of the mod union table: 4171 // Since the marker in "markFromRoots" marks concurrently with 4172 // mutators, it is possible for some reachable objects not to have been 4173 // scanned. For instance, an only reference to an object A was 4174 // placed in object B after the marker scanned B. Unless B is rescanned, 4175 // A would be collected. Such updates to references in marked objects 4176 // are detected via the mod union table which is the set of all cards 4177 // dirtied since the first checkpoint in this GC cycle and prior to 4178 // the most recent young generation GC, minus those cleaned up by the 4179 // concurrent precleaning. 4180 if (CMSParallelRemarkEnabled) { 4181 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4182 do_remark_parallel(); 4183 } else { 4184 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4185 do_remark_non_parallel(); 4186 } 4187 } 4188 verify_work_stacks_empty(); 4189 verify_overflow_empty(); 4190 4191 { 4192 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4193 refProcessingWork(); 4194 } 4195 verify_work_stacks_empty(); 4196 verify_overflow_empty(); 4197 4198 if (should_unload_classes()) { 4199 CodeCache::gc_epilogue(); 4200 } 4201 JvmtiExport::gc_epilogue(); 4202 4203 // If we encountered any (marking stack / work queue) overflow 4204 // events during the current CMS cycle, take appropriate 4205 // remedial measures, where possible, so as to try and avoid 4206 // recurrence of that condition. 4207 assert(_markStack.isEmpty(), "No grey objects"); 4208 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4209 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4210 if (ser_ovflw > 0) { 4211 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4212 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4213 _markStack.expand(); 4214 _ser_pmc_remark_ovflw = 0; 4215 _ser_pmc_preclean_ovflw = 0; 4216 _ser_kac_preclean_ovflw = 0; 4217 _ser_kac_ovflw = 0; 4218 } 4219 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4220 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4221 _par_pmc_remark_ovflw, _par_kac_ovflw); 4222 _par_pmc_remark_ovflw = 0; 4223 _par_kac_ovflw = 0; 4224 } 4225 if (_markStack._hit_limit > 0) { 4226 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4227 _markStack._hit_limit); 4228 } 4229 if (_markStack._failed_double > 0) { 4230 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4231 _markStack._failed_double, _markStack.capacity()); 4232 } 4233 _markStack._hit_limit = 0; 4234 _markStack._failed_double = 0; 4235 4236 if ((VerifyAfterGC || VerifyDuringGC) && 4237 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4238 verify_after_remark(); 4239 } 4240 4241 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4242 4243 // Change under the freelistLocks. 4244 _collectorState = Sweeping; 4245 // Call isAllClear() under bitMapLock 4246 assert(_modUnionTable.isAllClear(), 4247 "Should be clear by end of the final marking"); 4248 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4249 "Should be clear by end of the final marking"); 4250 } 4251 4252 void CMSParInitialMarkTask::work(uint worker_id) { 4253 elapsedTimer _timer; 4254 ResourceMark rm; 4255 HandleMark hm; 4256 4257 // ---------- scan from roots -------------- 4258 _timer.start(); 4259 CMSHeap* heap = CMSHeap::heap(); 4260 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4261 4262 // ---------- young gen roots -------------- 4263 { 4264 work_on_young_gen_roots(&par_mri_cl); 4265 _timer.stop(); 4266 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4267 } 4268 4269 // ---------- remaining roots -------------- 4270 _timer.reset(); 4271 _timer.start(); 4272 4273 CLDToOopClosure cld_closure(&par_mri_cl, true); 4274 4275 heap->cms_process_roots(_strong_roots_scope, 4276 false, // yg was scanned above 4277 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4278 _collector->should_unload_classes(), 4279 &par_mri_cl, 4280 &cld_closure, 4281 &_par_state_string); 4282 4283 assert(_collector->should_unload_classes() 4284 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4285 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4286 _timer.stop(); 4287 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4288 } 4289 4290 // Parallel remark task 4291 class CMSParRemarkTask: public CMSParMarkTask { 4292 CompactibleFreeListSpace* _cms_space; 4293 4294 // The per-thread work queues, available here for stealing. 4295 OopTaskQueueSet* _task_queues; 4296 ParallelTaskTerminator _term; 4297 StrongRootsScope* _strong_roots_scope; 4298 4299 public: 4300 // A value of 0 passed to n_workers will cause the number of 4301 // workers to be taken from the active workers in the work gang. 4302 CMSParRemarkTask(CMSCollector* collector, 4303 CompactibleFreeListSpace* cms_space, 4304 uint n_workers, WorkGang* workers, 4305 OopTaskQueueSet* task_queues, 4306 StrongRootsScope* strong_roots_scope): 4307 CMSParMarkTask("Rescan roots and grey objects in parallel", 4308 collector, n_workers), 4309 _cms_space(cms_space), 4310 _task_queues(task_queues), 4311 _term(n_workers, task_queues), 4312 _strong_roots_scope(strong_roots_scope) { } 4313 4314 OopTaskQueueSet* task_queues() { return _task_queues; } 4315 4316 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4317 4318 ParallelTaskTerminator* terminator() { return &_term; } 4319 uint n_workers() { return _n_workers; } 4320 4321 void work(uint worker_id); 4322 4323 private: 4324 // ... of dirty cards in old space 4325 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4326 ParMarkRefsIntoAndScanClosure* cl); 4327 4328 // ... work stealing for the above 4329 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4330 }; 4331 4332 class RemarkCLDClosure : public CLDClosure { 4333 CLDToOopClosure _cm_closure; 4334 public: 4335 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4336 void do_cld(ClassLoaderData* cld) { 4337 // Check if we have modified any oops in the CLD during the concurrent marking. 4338 if (cld->has_accumulated_modified_oops()) { 4339 cld->clear_accumulated_modified_oops(); 4340 4341 // We could have transfered the current modified marks to the accumulated marks, 4342 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4343 } else if (cld->has_modified_oops()) { 4344 // Don't clear anything, this info is needed by the next young collection. 4345 } else { 4346 // No modified oops in the ClassLoaderData. 4347 return; 4348 } 4349 4350 // The klass has modified fields, need to scan the klass. 4351 _cm_closure.do_cld(cld); 4352 } 4353 }; 4354 4355 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4356 ParNewGeneration* young_gen = _collector->_young_gen; 4357 ContiguousSpace* eden_space = young_gen->eden(); 4358 ContiguousSpace* from_space = young_gen->from(); 4359 ContiguousSpace* to_space = young_gen->to(); 4360 4361 HeapWord** eca = _collector->_eden_chunk_array; 4362 size_t ect = _collector->_eden_chunk_index; 4363 HeapWord** sca = _collector->_survivor_chunk_array; 4364 size_t sct = _collector->_survivor_chunk_index; 4365 4366 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4367 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4368 4369 do_young_space_rescan(cl, to_space, NULL, 0); 4370 do_young_space_rescan(cl, from_space, sca, sct); 4371 do_young_space_rescan(cl, eden_space, eca, ect); 4372 } 4373 4374 // work_queue(i) is passed to the closure 4375 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4376 // also is passed to do_dirty_card_rescan_tasks() and to 4377 // do_work_steal() to select the i-th task_queue. 4378 4379 void CMSParRemarkTask::work(uint worker_id) { 4380 elapsedTimer _timer; 4381 ResourceMark rm; 4382 HandleMark hm; 4383 4384 // ---------- rescan from roots -------------- 4385 _timer.start(); 4386 CMSHeap* heap = CMSHeap::heap(); 4387 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4388 _collector->_span, _collector->ref_processor(), 4389 &(_collector->_markBitMap), 4390 work_queue(worker_id)); 4391 4392 // Rescan young gen roots first since these are likely 4393 // coarsely partitioned and may, on that account, constitute 4394 // the critical path; thus, it's best to start off that 4395 // work first. 4396 // ---------- young gen roots -------------- 4397 { 4398 work_on_young_gen_roots(&par_mrias_cl); 4399 _timer.stop(); 4400 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4401 } 4402 4403 // ---------- remaining roots -------------- 4404 _timer.reset(); 4405 _timer.start(); 4406 heap->cms_process_roots(_strong_roots_scope, 4407 false, // yg was scanned above 4408 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4409 _collector->should_unload_classes(), 4410 &par_mrias_cl, 4411 NULL, // The dirty klasses will be handled below 4412 &_par_state_string); 4413 4414 assert(_collector->should_unload_classes() 4415 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4416 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4417 _timer.stop(); 4418 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4419 4420 // ---------- unhandled CLD scanning ---------- 4421 if (worker_id == 0) { // Single threaded at the moment. 4422 _timer.reset(); 4423 _timer.start(); 4424 4425 // Scan all new class loader data objects and new dependencies that were 4426 // introduced during concurrent marking. 4427 ResourceMark rm; 4428 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4429 for (int i = 0; i < array->length(); i++) { 4430 par_mrias_cl.do_cld_nv(array->at(i)); 4431 } 4432 4433 // We don't need to keep track of new CLDs anymore. 4434 ClassLoaderDataGraph::remember_new_clds(false); 4435 4436 _timer.stop(); 4437 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4438 } 4439 4440 // We might have added oops to ClassLoaderData::_handles during the 4441 // concurrent marking phase. These oops do not always point to newly allocated objects 4442 // that are guaranteed to be kept alive. Hence, 4443 // we do have to revisit the _handles block during the remark phase. 4444 4445 // ---------- dirty CLD scanning ---------- 4446 if (worker_id == 0) { // Single threaded at the moment. 4447 _timer.reset(); 4448 _timer.start(); 4449 4450 // Scan all classes that was dirtied during the concurrent marking phase. 4451 RemarkCLDClosure remark_closure(&par_mrias_cl); 4452 ClassLoaderDataGraph::cld_do(&remark_closure); 4453 4454 _timer.stop(); 4455 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4456 } 4457 4458 4459 // ---------- rescan dirty cards ------------ 4460 _timer.reset(); 4461 _timer.start(); 4462 4463 // Do the rescan tasks for each of the two spaces 4464 // (cms_space) in turn. 4465 // "worker_id" is passed to select the task_queue for "worker_id" 4466 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4467 _timer.stop(); 4468 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4469 4470 // ---------- steal work from other threads ... 4471 // ---------- ... and drain overflow list. 4472 _timer.reset(); 4473 _timer.start(); 4474 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4475 _timer.stop(); 4476 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4477 } 4478 4479 void 4480 CMSParMarkTask::do_young_space_rescan( 4481 OopsInGenClosure* cl, ContiguousSpace* space, 4482 HeapWord** chunk_array, size_t chunk_top) { 4483 // Until all tasks completed: 4484 // . claim an unclaimed task 4485 // . compute region boundaries corresponding to task claimed 4486 // using chunk_array 4487 // . par_oop_iterate(cl) over that region 4488 4489 ResourceMark rm; 4490 HandleMark hm; 4491 4492 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4493 4494 uint nth_task = 0; 4495 uint n_tasks = pst->n_tasks(); 4496 4497 if (n_tasks > 0) { 4498 assert(pst->valid(), "Uninitialized use?"); 4499 HeapWord *start, *end; 4500 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4501 // We claimed task # nth_task; compute its boundaries. 4502 if (chunk_top == 0) { // no samples were taken 4503 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4504 start = space->bottom(); 4505 end = space->top(); 4506 } else if (nth_task == 0) { 4507 start = space->bottom(); 4508 end = chunk_array[nth_task]; 4509 } else if (nth_task < (uint)chunk_top) { 4510 assert(nth_task >= 1, "Control point invariant"); 4511 start = chunk_array[nth_task - 1]; 4512 end = chunk_array[nth_task]; 4513 } else { 4514 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4515 start = chunk_array[chunk_top - 1]; 4516 end = space->top(); 4517 } 4518 MemRegion mr(start, end); 4519 // Verify that mr is in space 4520 assert(mr.is_empty() || space->used_region().contains(mr), 4521 "Should be in space"); 4522 // Verify that "start" is an object boundary 4523 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4524 "Should be an oop"); 4525 space->par_oop_iterate(mr, cl); 4526 } 4527 pst->all_tasks_completed(); 4528 } 4529 } 4530 4531 void 4532 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4533 CompactibleFreeListSpace* sp, int i, 4534 ParMarkRefsIntoAndScanClosure* cl) { 4535 // Until all tasks completed: 4536 // . claim an unclaimed task 4537 // . compute region boundaries corresponding to task claimed 4538 // . transfer dirty bits ct->mut for that region 4539 // . apply rescanclosure to dirty mut bits for that region 4540 4541 ResourceMark rm; 4542 HandleMark hm; 4543 4544 OopTaskQueue* work_q = work_queue(i); 4545 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4546 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4547 // CAUTION: This closure has state that persists across calls to 4548 // the work method dirty_range_iterate_clear() in that it has 4549 // embedded in it a (subtype of) UpwardsObjectClosure. The 4550 // use of that state in the embedded UpwardsObjectClosure instance 4551 // assumes that the cards are always iterated (even if in parallel 4552 // by several threads) in monotonically increasing order per each 4553 // thread. This is true of the implementation below which picks 4554 // card ranges (chunks) in monotonically increasing order globally 4555 // and, a-fortiori, in monotonically increasing order per thread 4556 // (the latter order being a subsequence of the former). 4557 // If the work code below is ever reorganized into a more chaotic 4558 // work-partitioning form than the current "sequential tasks" 4559 // paradigm, the use of that persistent state will have to be 4560 // revisited and modified appropriately. See also related 4561 // bug 4756801 work on which should examine this code to make 4562 // sure that the changes there do not run counter to the 4563 // assumptions made here and necessary for correctness and 4564 // efficiency. Note also that this code might yield inefficient 4565 // behavior in the case of very large objects that span one or 4566 // more work chunks. Such objects would potentially be scanned 4567 // several times redundantly. Work on 4756801 should try and 4568 // address that performance anomaly if at all possible. XXX 4569 MemRegion full_span = _collector->_span; 4570 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4571 MarkFromDirtyCardsClosure 4572 greyRescanClosure(_collector, full_span, // entire span of interest 4573 sp, bm, work_q, cl); 4574 4575 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4576 assert(pst->valid(), "Uninitialized use?"); 4577 uint nth_task = 0; 4578 const int alignment = CardTable::card_size * BitsPerWord; 4579 MemRegion span = sp->used_region(); 4580 HeapWord* start_addr = span.start(); 4581 HeapWord* end_addr = align_up(span.end(), alignment); 4582 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4583 assert(is_aligned(start_addr, alignment), "Check alignment"); 4584 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4585 4586 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4587 // Having claimed the nth_task, compute corresponding mem-region, 4588 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4589 // The alignment restriction ensures that we do not need any 4590 // synchronization with other gang-workers while setting or 4591 // clearing bits in thus chunk of the MUT. 4592 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4593 start_addr + (nth_task+1)*chunk_size); 4594 // The last chunk's end might be way beyond end of the 4595 // used region. In that case pull back appropriately. 4596 if (this_span.end() > end_addr) { 4597 this_span.set_end(end_addr); 4598 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4599 } 4600 // Iterate over the dirty cards covering this chunk, marking them 4601 // precleaned, and setting the corresponding bits in the mod union 4602 // table. Since we have been careful to partition at Card and MUT-word 4603 // boundaries no synchronization is needed between parallel threads. 4604 _collector->_ct->dirty_card_iterate(this_span, 4605 &modUnionClosure); 4606 4607 // Having transferred these marks into the modUnionTable, 4608 // rescan the marked objects on the dirty cards in the modUnionTable. 4609 // Even if this is at a synchronous collection, the initial marking 4610 // may have been done during an asynchronous collection so there 4611 // may be dirty bits in the mod-union table. 4612 _collector->_modUnionTable.dirty_range_iterate_clear( 4613 this_span, &greyRescanClosure); 4614 _collector->_modUnionTable.verifyNoOneBitsInRange( 4615 this_span.start(), 4616 this_span.end()); 4617 } 4618 pst->all_tasks_completed(); // declare that i am done 4619 } 4620 4621 // . see if we can share work_queues with ParNew? XXX 4622 void 4623 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4624 int* seed) { 4625 OopTaskQueue* work_q = work_queue(i); 4626 NOT_PRODUCT(int num_steals = 0;) 4627 oop obj_to_scan; 4628 CMSBitMap* bm = &(_collector->_markBitMap); 4629 4630 while (true) { 4631 // Completely finish any left over work from (an) earlier round(s) 4632 cl->trim_queue(0); 4633 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4634 (size_t)ParGCDesiredObjsFromOverflowList); 4635 // Now check if there's any work in the overflow list 4636 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4637 // only affects the number of attempts made to get work from the 4638 // overflow list and does not affect the number of workers. Just 4639 // pass ParallelGCThreads so this behavior is unchanged. 4640 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4641 work_q, 4642 ParallelGCThreads)) { 4643 // found something in global overflow list; 4644 // not yet ready to go stealing work from others. 4645 // We'd like to assert(work_q->size() != 0, ...) 4646 // because we just took work from the overflow list, 4647 // but of course we can't since all of that could have 4648 // been already stolen from us. 4649 // "He giveth and He taketh away." 4650 continue; 4651 } 4652 // Verify that we have no work before we resort to stealing 4653 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4654 // Try to steal from other queues that have work 4655 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4656 NOT_PRODUCT(num_steals++;) 4657 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4658 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4659 // Do scanning work 4660 obj_to_scan->oop_iterate(cl); 4661 // Loop around, finish this work, and try to steal some more 4662 } else if (terminator()->offer_termination()) { 4663 break; // nirvana from the infinite cycle 4664 } 4665 } 4666 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4667 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4668 "Else our work is not yet done"); 4669 } 4670 4671 // Record object boundaries in _eden_chunk_array by sampling the eden 4672 // top in the slow-path eden object allocation code path and record 4673 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4674 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4675 // sampling in sample_eden() that activates during the part of the 4676 // preclean phase. 4677 void CMSCollector::sample_eden_chunk() { 4678 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4679 if (_eden_chunk_lock->try_lock()) { 4680 // Record a sample. This is the critical section. The contents 4681 // of the _eden_chunk_array have to be non-decreasing in the 4682 // address order. 4683 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4684 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4685 "Unexpected state of Eden"); 4686 if (_eden_chunk_index == 0 || 4687 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4688 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4689 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4690 _eden_chunk_index++; // commit sample 4691 } 4692 _eden_chunk_lock->unlock(); 4693 } 4694 } 4695 } 4696 4697 // Return a thread-local PLAB recording array, as appropriate. 4698 void* CMSCollector::get_data_recorder(int thr_num) { 4699 if (_survivor_plab_array != NULL && 4700 (CMSPLABRecordAlways || 4701 (_collectorState > Marking && _collectorState < FinalMarking))) { 4702 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4703 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4704 ca->reset(); // clear it so that fresh data is recorded 4705 return (void*) ca; 4706 } else { 4707 return NULL; 4708 } 4709 } 4710 4711 // Reset all the thread-local PLAB recording arrays 4712 void CMSCollector::reset_survivor_plab_arrays() { 4713 for (uint i = 0; i < ParallelGCThreads; i++) { 4714 _survivor_plab_array[i].reset(); 4715 } 4716 } 4717 4718 // Merge the per-thread plab arrays into the global survivor chunk 4719 // array which will provide the partitioning of the survivor space 4720 // for CMS initial scan and rescan. 4721 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4722 int no_of_gc_threads) { 4723 assert(_survivor_plab_array != NULL, "Error"); 4724 assert(_survivor_chunk_array != NULL, "Error"); 4725 assert(_collectorState == FinalMarking || 4726 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4727 for (int j = 0; j < no_of_gc_threads; j++) { 4728 _cursor[j] = 0; 4729 } 4730 HeapWord* top = surv->top(); 4731 size_t i; 4732 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4733 HeapWord* min_val = top; // Higher than any PLAB address 4734 uint min_tid = 0; // position of min_val this round 4735 for (int j = 0; j < no_of_gc_threads; j++) { 4736 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4737 if (_cursor[j] == cur_sca->end()) { 4738 continue; 4739 } 4740 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4741 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4742 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4743 if (cur_val < min_val) { 4744 min_tid = j; 4745 min_val = cur_val; 4746 } else { 4747 assert(cur_val < top, "All recorded addresses should be less"); 4748 } 4749 } 4750 // At this point min_val and min_tid are respectively 4751 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4752 // and the thread (j) that witnesses that address. 4753 // We record this address in the _survivor_chunk_array[i] 4754 // and increment _cursor[min_tid] prior to the next round i. 4755 if (min_val == top) { 4756 break; 4757 } 4758 _survivor_chunk_array[i] = min_val; 4759 _cursor[min_tid]++; 4760 } 4761 // We are all done; record the size of the _survivor_chunk_array 4762 _survivor_chunk_index = i; // exclusive: [0, i) 4763 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4764 // Verify that we used up all the recorded entries 4765 #ifdef ASSERT 4766 size_t total = 0; 4767 for (int j = 0; j < no_of_gc_threads; j++) { 4768 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4769 total += _cursor[j]; 4770 } 4771 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4772 // Check that the merged array is in sorted order 4773 if (total > 0) { 4774 for (size_t i = 0; i < total - 1; i++) { 4775 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4776 i, p2i(_survivor_chunk_array[i])); 4777 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4778 "Not sorted"); 4779 } 4780 } 4781 #endif // ASSERT 4782 } 4783 4784 // Set up the space's par_seq_tasks structure for work claiming 4785 // for parallel initial scan and rescan of young gen. 4786 // See ParRescanTask where this is currently used. 4787 void 4788 CMSCollector:: 4789 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4790 assert(n_threads > 0, "Unexpected n_threads argument"); 4791 4792 // Eden space 4793 if (!_young_gen->eden()->is_empty()) { 4794 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4795 assert(!pst->valid(), "Clobbering existing data?"); 4796 // Each valid entry in [0, _eden_chunk_index) represents a task. 4797 size_t n_tasks = _eden_chunk_index + 1; 4798 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4799 // Sets the condition for completion of the subtask (how many threads 4800 // need to finish in order to be done). 4801 pst->set_n_threads(n_threads); 4802 pst->set_n_tasks((int)n_tasks); 4803 } 4804 4805 // Merge the survivor plab arrays into _survivor_chunk_array 4806 if (_survivor_plab_array != NULL) { 4807 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4808 } else { 4809 assert(_survivor_chunk_index == 0, "Error"); 4810 } 4811 4812 // To space 4813 { 4814 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4815 assert(!pst->valid(), "Clobbering existing data?"); 4816 // Sets the condition for completion of the subtask (how many threads 4817 // need to finish in order to be done). 4818 pst->set_n_threads(n_threads); 4819 pst->set_n_tasks(1); 4820 assert(pst->valid(), "Error"); 4821 } 4822 4823 // From space 4824 { 4825 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4826 assert(!pst->valid(), "Clobbering existing data?"); 4827 size_t n_tasks = _survivor_chunk_index + 1; 4828 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4829 // Sets the condition for completion of the subtask (how many threads 4830 // need to finish in order to be done). 4831 pst->set_n_threads(n_threads); 4832 pst->set_n_tasks((int)n_tasks); 4833 assert(pst->valid(), "Error"); 4834 } 4835 } 4836 4837 // Parallel version of remark 4838 void CMSCollector::do_remark_parallel() { 4839 CMSHeap* heap = CMSHeap::heap(); 4840 WorkGang* workers = heap->workers(); 4841 assert(workers != NULL, "Need parallel worker threads."); 4842 // Choose to use the number of GC workers most recently set 4843 // into "active_workers". 4844 uint n_workers = workers->active_workers(); 4845 4846 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4847 4848 StrongRootsScope srs(n_workers); 4849 4850 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4851 4852 // We won't be iterating over the cards in the card table updating 4853 // the younger_gen cards, so we shouldn't call the following else 4854 // the verification code as well as subsequent younger_refs_iterate 4855 // code would get confused. XXX 4856 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4857 4858 // The young gen rescan work will not be done as part of 4859 // process_roots (which currently doesn't know how to 4860 // parallelize such a scan), but rather will be broken up into 4861 // a set of parallel tasks (via the sampling that the [abortable] 4862 // preclean phase did of eden, plus the [two] tasks of 4863 // scanning the [two] survivor spaces. Further fine-grain 4864 // parallelization of the scanning of the survivor spaces 4865 // themselves, and of precleaning of the young gen itself 4866 // is deferred to the future. 4867 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4868 4869 // The dirty card rescan work is broken up into a "sequence" 4870 // of parallel tasks (per constituent space) that are dynamically 4871 // claimed by the parallel threads. 4872 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4873 4874 // It turns out that even when we're using 1 thread, doing the work in a 4875 // separate thread causes wide variance in run times. We can't help this 4876 // in the multi-threaded case, but we special-case n=1 here to get 4877 // repeatable measurements of the 1-thread overhead of the parallel code. 4878 if (n_workers > 1) { 4879 // Make refs discovery MT-safe, if it isn't already: it may not 4880 // necessarily be so, since it's possible that we are doing 4881 // ST marking. 4882 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4883 workers->run_task(&tsk); 4884 } else { 4885 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4886 tsk.work(0); 4887 } 4888 4889 // restore, single-threaded for now, any preserved marks 4890 // as a result of work_q overflow 4891 restore_preserved_marks_if_any(); 4892 } 4893 4894 // Non-parallel version of remark 4895 void CMSCollector::do_remark_non_parallel() { 4896 ResourceMark rm; 4897 HandleMark hm; 4898 CMSHeap* heap = CMSHeap::heap(); 4899 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4900 4901 MarkRefsIntoAndScanClosure 4902 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4903 &_markStack, this, 4904 false /* should_yield */, false /* not precleaning */); 4905 MarkFromDirtyCardsClosure 4906 markFromDirtyCardsClosure(this, _span, 4907 NULL, // space is set further below 4908 &_markBitMap, &_markStack, &mrias_cl); 4909 { 4910 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4911 // Iterate over the dirty cards, setting the corresponding bits in the 4912 // mod union table. 4913 { 4914 ModUnionClosure modUnionClosure(&_modUnionTable); 4915 _ct->dirty_card_iterate(_cmsGen->used_region(), 4916 &modUnionClosure); 4917 } 4918 // Having transferred these marks into the modUnionTable, we just need 4919 // to rescan the marked objects on the dirty cards in the modUnionTable. 4920 // The initial marking may have been done during an asynchronous 4921 // collection so there may be dirty bits in the mod-union table. 4922 const int alignment = CardTable::card_size * BitsPerWord; 4923 { 4924 // ... First handle dirty cards in CMS gen 4925 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4926 MemRegion ur = _cmsGen->used_region(); 4927 HeapWord* lb = ur.start(); 4928 HeapWord* ub = align_up(ur.end(), alignment); 4929 MemRegion cms_span(lb, ub); 4930 _modUnionTable.dirty_range_iterate_clear(cms_span, 4931 &markFromDirtyCardsClosure); 4932 verify_work_stacks_empty(); 4933 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4934 } 4935 } 4936 if (VerifyDuringGC && 4937 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4938 HandleMark hm; // Discard invalid handles created during verification 4939 Universe::verify(); 4940 } 4941 { 4942 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4943 4944 verify_work_stacks_empty(); 4945 4946 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4947 StrongRootsScope srs(1); 4948 4949 heap->cms_process_roots(&srs, 4950 true, // young gen as roots 4951 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4952 should_unload_classes(), 4953 &mrias_cl, 4954 NULL); // The dirty klasses will be handled below 4955 4956 assert(should_unload_classes() 4957 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4958 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4959 } 4960 4961 { 4962 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4963 4964 verify_work_stacks_empty(); 4965 4966 // Scan all class loader data objects that might have been introduced 4967 // during concurrent marking. 4968 ResourceMark rm; 4969 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4970 for (int i = 0; i < array->length(); i++) { 4971 mrias_cl.do_cld_nv(array->at(i)); 4972 } 4973 4974 // We don't need to keep track of new CLDs anymore. 4975 ClassLoaderDataGraph::remember_new_clds(false); 4976 4977 verify_work_stacks_empty(); 4978 } 4979 4980 // We might have added oops to ClassLoaderData::_handles during the 4981 // concurrent marking phase. These oops do not point to newly allocated objects 4982 // that are guaranteed to be kept alive. Hence, 4983 // we do have to revisit the _handles block during the remark phase. 4984 { 4985 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4986 4987 verify_work_stacks_empty(); 4988 4989 RemarkCLDClosure remark_closure(&mrias_cl); 4990 ClassLoaderDataGraph::cld_do(&remark_closure); 4991 4992 verify_work_stacks_empty(); 4993 } 4994 4995 verify_work_stacks_empty(); 4996 // Restore evacuated mark words, if any, used for overflow list links 4997 restore_preserved_marks_if_any(); 4998 4999 verify_overflow_empty(); 5000 } 5001 5002 //////////////////////////////////////////////////////// 5003 // Parallel Reference Processing Task Proxy Class 5004 //////////////////////////////////////////////////////// 5005 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5006 OopTaskQueueSet* _queues; 5007 ParallelTaskTerminator _terminator; 5008 public: 5009 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5010 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5011 ParallelTaskTerminator* terminator() { return &_terminator; } 5012 OopTaskQueueSet* queues() { return _queues; } 5013 }; 5014 5015 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5016 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5017 CMSCollector* _collector; 5018 CMSBitMap* _mark_bit_map; 5019 const MemRegion _span; 5020 ProcessTask& _task; 5021 5022 public: 5023 CMSRefProcTaskProxy(ProcessTask& task, 5024 CMSCollector* collector, 5025 const MemRegion& span, 5026 CMSBitMap* mark_bit_map, 5027 AbstractWorkGang* workers, 5028 OopTaskQueueSet* task_queues): 5029 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5030 task_queues, 5031 workers->active_workers()), 5032 _task(task), 5033 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5034 { 5035 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5036 "Inconsistency in _span"); 5037 } 5038 5039 OopTaskQueueSet* task_queues() { return queues(); } 5040 5041 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5042 5043 void do_work_steal(int i, 5044 CMSParDrainMarkingStackClosure* drain, 5045 CMSParKeepAliveClosure* keep_alive, 5046 int* seed); 5047 5048 virtual void work(uint worker_id); 5049 }; 5050 5051 void CMSRefProcTaskProxy::work(uint worker_id) { 5052 ResourceMark rm; 5053 HandleMark hm; 5054 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5055 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5056 _mark_bit_map, 5057 work_queue(worker_id)); 5058 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5059 _mark_bit_map, 5060 work_queue(worker_id)); 5061 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5062 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5063 if (_task.marks_oops_alive()) { 5064 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5065 _collector->hash_seed(worker_id)); 5066 } 5067 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5068 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5069 } 5070 5071 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5072 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5073 _span(span), 5074 _bit_map(bit_map), 5075 _work_queue(work_queue), 5076 _mark_and_push(collector, span, bit_map, work_queue), 5077 _low_water_mark(MIN2((work_queue->max_elems()/4), 5078 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5079 { } 5080 5081 // . see if we can share work_queues with ParNew? XXX 5082 void CMSRefProcTaskProxy::do_work_steal(int i, 5083 CMSParDrainMarkingStackClosure* drain, 5084 CMSParKeepAliveClosure* keep_alive, 5085 int* seed) { 5086 OopTaskQueue* work_q = work_queue(i); 5087 NOT_PRODUCT(int num_steals = 0;) 5088 oop obj_to_scan; 5089 5090 while (true) { 5091 // Completely finish any left over work from (an) earlier round(s) 5092 drain->trim_queue(0); 5093 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5094 (size_t)ParGCDesiredObjsFromOverflowList); 5095 // Now check if there's any work in the overflow list 5096 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5097 // only affects the number of attempts made to get work from the 5098 // overflow list and does not affect the number of workers. Just 5099 // pass ParallelGCThreads so this behavior is unchanged. 5100 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5101 work_q, 5102 ParallelGCThreads)) { 5103 // Found something in global overflow list; 5104 // not yet ready to go stealing work from others. 5105 // We'd like to assert(work_q->size() != 0, ...) 5106 // because we just took work from the overflow list, 5107 // but of course we can't, since all of that might have 5108 // been already stolen from us. 5109 continue; 5110 } 5111 // Verify that we have no work before we resort to stealing 5112 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5113 // Try to steal from other queues that have work 5114 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5115 NOT_PRODUCT(num_steals++;) 5116 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5117 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5118 // Do scanning work 5119 obj_to_scan->oop_iterate(keep_alive); 5120 // Loop around, finish this work, and try to steal some more 5121 } else if (terminator()->offer_termination()) { 5122 break; // nirvana from the infinite cycle 5123 } 5124 } 5125 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5126 } 5127 5128 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5129 { 5130 CMSHeap* heap = CMSHeap::heap(); 5131 WorkGang* workers = heap->workers(); 5132 assert(workers != NULL, "Need parallel worker threads."); 5133 CMSRefProcTaskProxy rp_task(task, &_collector, 5134 _collector.ref_processor_span(), 5135 _collector.markBitMap(), 5136 workers, _collector.task_queues()); 5137 workers->run_task(&rp_task); 5138 } 5139 5140 void CMSCollector::refProcessingWork() { 5141 ResourceMark rm; 5142 HandleMark hm; 5143 5144 ReferenceProcessor* rp = ref_processor(); 5145 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5146 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5147 // Process weak references. 5148 rp->setup_policy(false); 5149 verify_work_stacks_empty(); 5150 5151 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_queues()); 5152 { 5153 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5154 5155 // Setup keep_alive and complete closures. 5156 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5157 &_markStack, false /* !preclean */); 5158 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5159 _span, &_markBitMap, &_markStack, 5160 &cmsKeepAliveClosure, false /* !preclean */); 5161 5162 ReferenceProcessorStats stats; 5163 if (rp->processing_is_mt()) { 5164 // Set the degree of MT here. If the discovery is done MT, there 5165 // may have been a different number of threads doing the discovery 5166 // and a different number of discovered lists may have Ref objects. 5167 // That is OK as long as the Reference lists are balanced (see 5168 // balance_all_queues() and balance_queues()). 5169 CMSHeap* heap = CMSHeap::heap(); 5170 uint active_workers = ParallelGCThreads; 5171 WorkGang* workers = heap->workers(); 5172 if (workers != NULL) { 5173 active_workers = workers->active_workers(); 5174 // The expectation is that active_workers will have already 5175 // been set to a reasonable value. If it has not been set, 5176 // investigate. 5177 assert(active_workers > 0, "Should have been set during scavenge"); 5178 } 5179 rp->set_active_mt_degree(active_workers); 5180 CMSRefProcTaskExecutor task_executor(*this); 5181 stats = rp->process_discovered_references(&_is_alive_closure, 5182 &cmsKeepAliveClosure, 5183 &cmsDrainMarkingStackClosure, 5184 &task_executor, 5185 &pt); 5186 } else { 5187 stats = rp->process_discovered_references(&_is_alive_closure, 5188 &cmsKeepAliveClosure, 5189 &cmsDrainMarkingStackClosure, 5190 NULL, 5191 &pt); 5192 } 5193 _gc_tracer_cm->report_gc_reference_stats(stats); 5194 pt.print_all_references(); 5195 } 5196 5197 // This is the point where the entire marking should have completed. 5198 verify_work_stacks_empty(); 5199 5200 { 5201 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5202 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5203 } 5204 5205 if (should_unload_classes()) { 5206 { 5207 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5208 5209 // Unload classes and purge the SystemDictionary. 5210 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5211 5212 // Unload nmethods. 5213 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5214 5215 // Prune dead klasses from subklass/sibling/implementor lists. 5216 Klass::clean_weak_klass_links(purged_class); 5217 } 5218 5219 { 5220 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5221 // Clean up unreferenced symbols in symbol table. 5222 SymbolTable::unlink(); 5223 } 5224 5225 { 5226 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5227 // Delete entries for dead interned strings. 5228 StringTable::unlink(&_is_alive_closure); 5229 } 5230 } 5231 5232 // Restore any preserved marks as a result of mark stack or 5233 // work queue overflow 5234 restore_preserved_marks_if_any(); // done single-threaded for now 5235 5236 rp->set_enqueuing_is_done(true); 5237 rp->verify_no_references_recorded(); 5238 } 5239 5240 #ifndef PRODUCT 5241 void CMSCollector::check_correct_thread_executing() { 5242 Thread* t = Thread::current(); 5243 // Only the VM thread or the CMS thread should be here. 5244 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5245 "Unexpected thread type"); 5246 // If this is the vm thread, the foreground process 5247 // should not be waiting. Note that _foregroundGCIsActive is 5248 // true while the foreground collector is waiting. 5249 if (_foregroundGCShouldWait) { 5250 // We cannot be the VM thread 5251 assert(t->is_ConcurrentGC_thread(), 5252 "Should be CMS thread"); 5253 } else { 5254 // We can be the CMS thread only if we are in a stop-world 5255 // phase of CMS collection. 5256 if (t->is_ConcurrentGC_thread()) { 5257 assert(_collectorState == InitialMarking || 5258 _collectorState == FinalMarking, 5259 "Should be a stop-world phase"); 5260 // The CMS thread should be holding the CMS_token. 5261 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5262 "Potential interference with concurrently " 5263 "executing VM thread"); 5264 } 5265 } 5266 } 5267 #endif 5268 5269 void CMSCollector::sweep() { 5270 assert(_collectorState == Sweeping, "just checking"); 5271 check_correct_thread_executing(); 5272 verify_work_stacks_empty(); 5273 verify_overflow_empty(); 5274 increment_sweep_count(); 5275 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5276 5277 _inter_sweep_timer.stop(); 5278 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5279 5280 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5281 _intra_sweep_timer.reset(); 5282 _intra_sweep_timer.start(); 5283 { 5284 GCTraceCPUTime tcpu; 5285 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5286 // First sweep the old gen 5287 { 5288 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5289 bitMapLock()); 5290 sweepWork(_cmsGen); 5291 } 5292 5293 // Update Universe::_heap_*_at_gc figures. 5294 // We need all the free list locks to make the abstract state 5295 // transition from Sweeping to Resetting. See detailed note 5296 // further below. 5297 { 5298 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5299 // Update heap occupancy information which is used as 5300 // input to soft ref clearing policy at the next gc. 5301 Universe::update_heap_info_at_gc(); 5302 _collectorState = Resizing; 5303 } 5304 } 5305 verify_work_stacks_empty(); 5306 verify_overflow_empty(); 5307 5308 if (should_unload_classes()) { 5309 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5310 // requires that the virtual spaces are stable and not deleted. 5311 ClassLoaderDataGraph::set_should_purge(true); 5312 } 5313 5314 _intra_sweep_timer.stop(); 5315 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5316 5317 _inter_sweep_timer.reset(); 5318 _inter_sweep_timer.start(); 5319 5320 // We need to use a monotonically non-decreasing time in ms 5321 // or we will see time-warp warnings and os::javaTimeMillis() 5322 // does not guarantee monotonicity. 5323 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5324 update_time_of_last_gc(now); 5325 5326 // NOTE on abstract state transitions: 5327 // Mutators allocate-live and/or mark the mod-union table dirty 5328 // based on the state of the collection. The former is done in 5329 // the interval [Marking, Sweeping] and the latter in the interval 5330 // [Marking, Sweeping). Thus the transitions into the Marking state 5331 // and out of the Sweeping state must be synchronously visible 5332 // globally to the mutators. 5333 // The transition into the Marking state happens with the world 5334 // stopped so the mutators will globally see it. Sweeping is 5335 // done asynchronously by the background collector so the transition 5336 // from the Sweeping state to the Resizing state must be done 5337 // under the freelistLock (as is the check for whether to 5338 // allocate-live and whether to dirty the mod-union table). 5339 assert(_collectorState == Resizing, "Change of collector state to" 5340 " Resizing must be done under the freelistLocks (plural)"); 5341 5342 // Now that sweeping has been completed, we clear 5343 // the incremental_collection_failed flag, 5344 // thus inviting a younger gen collection to promote into 5345 // this generation. If such a promotion may still fail, 5346 // the flag will be set again when a young collection is 5347 // attempted. 5348 CMSHeap* heap = CMSHeap::heap(); 5349 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5350 heap->update_full_collections_completed(_collection_count_start); 5351 } 5352 5353 // FIX ME!!! Looks like this belongs in CFLSpace, with 5354 // CMSGen merely delegating to it. 5355 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5356 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5357 HeapWord* minAddr = _cmsSpace->bottom(); 5358 HeapWord* largestAddr = 5359 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5360 if (largestAddr == NULL) { 5361 // The dictionary appears to be empty. In this case 5362 // try to coalesce at the end of the heap. 5363 largestAddr = _cmsSpace->end(); 5364 } 5365 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5366 size_t nearLargestOffset = 5367 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5368 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5369 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5370 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5371 } 5372 5373 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5374 return addr >= _cmsSpace->nearLargestChunk(); 5375 } 5376 5377 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5378 return _cmsSpace->find_chunk_at_end(); 5379 } 5380 5381 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5382 bool full) { 5383 // If the young generation has been collected, gather any statistics 5384 // that are of interest at this point. 5385 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5386 if (!full && current_is_young) { 5387 // Gather statistics on the young generation collection. 5388 collector()->stats().record_gc0_end(used()); 5389 } 5390 } 5391 5392 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5393 // We iterate over the space(s) underlying this generation, 5394 // checking the mark bit map to see if the bits corresponding 5395 // to specific blocks are marked or not. Blocks that are 5396 // marked are live and are not swept up. All remaining blocks 5397 // are swept up, with coalescing on-the-fly as we sweep up 5398 // contiguous free and/or garbage blocks: 5399 // We need to ensure that the sweeper synchronizes with allocators 5400 // and stop-the-world collectors. In particular, the following 5401 // locks are used: 5402 // . CMS token: if this is held, a stop the world collection cannot occur 5403 // . freelistLock: if this is held no allocation can occur from this 5404 // generation by another thread 5405 // . bitMapLock: if this is held, no other thread can access or update 5406 // 5407 5408 // Note that we need to hold the freelistLock if we use 5409 // block iterate below; else the iterator might go awry if 5410 // a mutator (or promotion) causes block contents to change 5411 // (for instance if the allocator divvies up a block). 5412 // If we hold the free list lock, for all practical purposes 5413 // young generation GC's can't occur (they'll usually need to 5414 // promote), so we might as well prevent all young generation 5415 // GC's while we do a sweeping step. For the same reason, we might 5416 // as well take the bit map lock for the entire duration 5417 5418 // check that we hold the requisite locks 5419 assert(have_cms_token(), "Should hold cms token"); 5420 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5421 assert_lock_strong(old_gen->freelistLock()); 5422 assert_lock_strong(bitMapLock()); 5423 5424 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5425 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5426 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5427 _inter_sweep_estimate.padded_average(), 5428 _intra_sweep_estimate.padded_average()); 5429 old_gen->setNearLargestChunk(); 5430 5431 { 5432 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5433 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5434 // We need to free-up/coalesce garbage/blocks from a 5435 // co-terminal free run. This is done in the SweepClosure 5436 // destructor; so, do not remove this scope, else the 5437 // end-of-sweep-census below will be off by a little bit. 5438 } 5439 old_gen->cmsSpace()->sweep_completed(); 5440 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5441 if (should_unload_classes()) { // unloaded classes this cycle, 5442 _concurrent_cycles_since_last_unload = 0; // ... reset count 5443 } else { // did not unload classes, 5444 _concurrent_cycles_since_last_unload++; // ... increment count 5445 } 5446 } 5447 5448 // Reset CMS data structures (for now just the marking bit map) 5449 // preparatory for the next cycle. 5450 void CMSCollector::reset_concurrent() { 5451 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5452 5453 // If the state is not "Resetting", the foreground thread 5454 // has done a collection and the resetting. 5455 if (_collectorState != Resetting) { 5456 assert(_collectorState == Idling, "The state should only change" 5457 " because the foreground collector has finished the collection"); 5458 return; 5459 } 5460 5461 { 5462 // Clear the mark bitmap (no grey objects to start with) 5463 // for the next cycle. 5464 GCTraceCPUTime tcpu; 5465 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5466 5467 HeapWord* curAddr = _markBitMap.startWord(); 5468 while (curAddr < _markBitMap.endWord()) { 5469 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5470 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5471 _markBitMap.clear_large_range(chunk); 5472 if (ConcurrentMarkSweepThread::should_yield() && 5473 !foregroundGCIsActive() && 5474 CMSYield) { 5475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5476 "CMS thread should hold CMS token"); 5477 assert_lock_strong(bitMapLock()); 5478 bitMapLock()->unlock(); 5479 ConcurrentMarkSweepThread::desynchronize(true); 5480 stopTimer(); 5481 incrementYields(); 5482 5483 // See the comment in coordinator_yield() 5484 for (unsigned i = 0; i < CMSYieldSleepCount && 5485 ConcurrentMarkSweepThread::should_yield() && 5486 !CMSCollector::foregroundGCIsActive(); ++i) { 5487 os::sleep(Thread::current(), 1, false); 5488 } 5489 5490 ConcurrentMarkSweepThread::synchronize(true); 5491 bitMapLock()->lock_without_safepoint_check(); 5492 startTimer(); 5493 } 5494 curAddr = chunk.end(); 5495 } 5496 // A successful mostly concurrent collection has been done. 5497 // Because only the full (i.e., concurrent mode failure) collections 5498 // are being measured for gc overhead limits, clean the "near" flag 5499 // and count. 5500 size_policy()->reset_gc_overhead_limit_count(); 5501 _collectorState = Idling; 5502 } 5503 5504 register_gc_end(); 5505 } 5506 5507 // Same as above but for STW paths 5508 void CMSCollector::reset_stw() { 5509 // already have the lock 5510 assert(_collectorState == Resetting, "just checking"); 5511 assert_lock_strong(bitMapLock()); 5512 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5513 _markBitMap.clear_all(); 5514 _collectorState = Idling; 5515 register_gc_end(); 5516 } 5517 5518 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5519 GCTraceCPUTime tcpu; 5520 TraceCollectorStats tcs_cgc(cgc_counters()); 5521 5522 switch (op) { 5523 case CMS_op_checkpointRootsInitial: { 5524 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5525 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5526 checkpointRootsInitial(); 5527 break; 5528 } 5529 case CMS_op_checkpointRootsFinal: { 5530 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5531 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5532 checkpointRootsFinal(); 5533 break; 5534 } 5535 default: 5536 fatal("No such CMS_op"); 5537 } 5538 } 5539 5540 #ifndef PRODUCT 5541 size_t const CMSCollector::skip_header_HeapWords() { 5542 return FreeChunk::header_size(); 5543 } 5544 5545 // Try and collect here conditions that should hold when 5546 // CMS thread is exiting. The idea is that the foreground GC 5547 // thread should not be blocked if it wants to terminate 5548 // the CMS thread and yet continue to run the VM for a while 5549 // after that. 5550 void CMSCollector::verify_ok_to_terminate() const { 5551 assert(Thread::current()->is_ConcurrentGC_thread(), 5552 "should be called by CMS thread"); 5553 assert(!_foregroundGCShouldWait, "should be false"); 5554 // We could check here that all the various low-level locks 5555 // are not held by the CMS thread, but that is overkill; see 5556 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5557 // is checked. 5558 } 5559 #endif 5560 5561 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5562 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5563 "missing Printezis mark?"); 5564 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5565 size_t size = pointer_delta(nextOneAddr + 1, addr); 5566 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5567 "alignment problem"); 5568 assert(size >= 3, "Necessary for Printezis marks to work"); 5569 return size; 5570 } 5571 5572 // A variant of the above (block_size_using_printezis_bits()) except 5573 // that we return 0 if the P-bits are not yet set. 5574 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5575 if (_markBitMap.isMarked(addr + 1)) { 5576 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5577 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5578 size_t size = pointer_delta(nextOneAddr + 1, addr); 5579 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5580 "alignment problem"); 5581 assert(size >= 3, "Necessary for Printezis marks to work"); 5582 return size; 5583 } 5584 return 0; 5585 } 5586 5587 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5588 size_t sz = 0; 5589 oop p = (oop)addr; 5590 if (p->klass_or_null_acquire() != NULL) { 5591 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5592 } else { 5593 sz = block_size_using_printezis_bits(addr); 5594 } 5595 assert(sz > 0, "size must be nonzero"); 5596 HeapWord* next_block = addr + sz; 5597 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5598 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5599 align_down((uintptr_t)next_card, CardTable::card_size), 5600 "must be different cards"); 5601 return next_card; 5602 } 5603 5604 5605 // CMS Bit Map Wrapper ///////////////////////////////////////// 5606 5607 // Construct a CMS bit map infrastructure, but don't create the 5608 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5609 // further below. 5610 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5611 _bm(), 5612 _shifter(shifter), 5613 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5614 Monitor::_safepoint_check_sometimes) : NULL) 5615 { 5616 _bmStartWord = 0; 5617 _bmWordSize = 0; 5618 } 5619 5620 bool CMSBitMap::allocate(MemRegion mr) { 5621 _bmStartWord = mr.start(); 5622 _bmWordSize = mr.word_size(); 5623 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5624 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5625 if (!brs.is_reserved()) { 5626 log_warning(gc)("CMS bit map allocation failure"); 5627 return false; 5628 } 5629 // For now we'll just commit all of the bit map up front. 5630 // Later on we'll try to be more parsimonious with swap. 5631 if (!_virtual_space.initialize(brs, brs.size())) { 5632 log_warning(gc)("CMS bit map backing store failure"); 5633 return false; 5634 } 5635 assert(_virtual_space.committed_size() == brs.size(), 5636 "didn't reserve backing store for all of CMS bit map?"); 5637 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5638 _bmWordSize, "inconsistency in bit map sizing"); 5639 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5640 5641 // bm.clear(); // can we rely on getting zero'd memory? verify below 5642 assert(isAllClear(), 5643 "Expected zero'd memory from ReservedSpace constructor"); 5644 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5645 "consistency check"); 5646 return true; 5647 } 5648 5649 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5650 HeapWord *next_addr, *end_addr, *last_addr; 5651 assert_locked(); 5652 assert(covers(mr), "out-of-range error"); 5653 // XXX assert that start and end are appropriately aligned 5654 for (next_addr = mr.start(), end_addr = mr.end(); 5655 next_addr < end_addr; next_addr = last_addr) { 5656 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5657 last_addr = dirty_region.end(); 5658 if (!dirty_region.is_empty()) { 5659 cl->do_MemRegion(dirty_region); 5660 } else { 5661 assert(last_addr == end_addr, "program logic"); 5662 return; 5663 } 5664 } 5665 } 5666 5667 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5668 _bm.print_on_error(st, prefix); 5669 } 5670 5671 #ifndef PRODUCT 5672 void CMSBitMap::assert_locked() const { 5673 CMSLockVerifier::assert_locked(lock()); 5674 } 5675 5676 bool CMSBitMap::covers(MemRegion mr) const { 5677 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5678 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5679 "size inconsistency"); 5680 return (mr.start() >= _bmStartWord) && 5681 (mr.end() <= endWord()); 5682 } 5683 5684 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5685 return (start >= _bmStartWord && (start + size) <= endWord()); 5686 } 5687 5688 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5689 // verify that there are no 1 bits in the interval [left, right) 5690 FalseBitMapClosure falseBitMapClosure; 5691 iterate(&falseBitMapClosure, left, right); 5692 } 5693 5694 void CMSBitMap::region_invariant(MemRegion mr) 5695 { 5696 assert_locked(); 5697 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5698 assert(!mr.is_empty(), "unexpected empty region"); 5699 assert(covers(mr), "mr should be covered by bit map"); 5700 // convert address range into offset range 5701 size_t start_ofs = heapWordToOffset(mr.start()); 5702 // Make sure that end() is appropriately aligned 5703 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5704 "Misaligned mr.end()"); 5705 size_t end_ofs = heapWordToOffset(mr.end()); 5706 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5707 } 5708 5709 #endif 5710 5711 bool CMSMarkStack::allocate(size_t size) { 5712 // allocate a stack of the requisite depth 5713 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5714 size * sizeof(oop))); 5715 if (!rs.is_reserved()) { 5716 log_warning(gc)("CMSMarkStack allocation failure"); 5717 return false; 5718 } 5719 if (!_virtual_space.initialize(rs, rs.size())) { 5720 log_warning(gc)("CMSMarkStack backing store failure"); 5721 return false; 5722 } 5723 assert(_virtual_space.committed_size() == rs.size(), 5724 "didn't reserve backing store for all of CMS stack?"); 5725 _base = (oop*)(_virtual_space.low()); 5726 _index = 0; 5727 _capacity = size; 5728 NOT_PRODUCT(_max_depth = 0); 5729 return true; 5730 } 5731 5732 // XXX FIX ME !!! In the MT case we come in here holding a 5733 // leaf lock. For printing we need to take a further lock 5734 // which has lower rank. We need to recalibrate the two 5735 // lock-ranks involved in order to be able to print the 5736 // messages below. (Or defer the printing to the caller. 5737 // For now we take the expedient path of just disabling the 5738 // messages for the problematic case.) 5739 void CMSMarkStack::expand() { 5740 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5741 if (_capacity == MarkStackSizeMax) { 5742 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5743 // We print a warning message only once per CMS cycle. 5744 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5745 } 5746 return; 5747 } 5748 // Double capacity if possible 5749 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5750 // Do not give up existing stack until we have managed to 5751 // get the double capacity that we desired. 5752 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5753 new_capacity * sizeof(oop))); 5754 if (rs.is_reserved()) { 5755 // Release the backing store associated with old stack 5756 _virtual_space.release(); 5757 // Reinitialize virtual space for new stack 5758 if (!_virtual_space.initialize(rs, rs.size())) { 5759 fatal("Not enough swap for expanded marking stack"); 5760 } 5761 _base = (oop*)(_virtual_space.low()); 5762 _index = 0; 5763 _capacity = new_capacity; 5764 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5765 // Failed to double capacity, continue; 5766 // we print a detail message only once per CMS cycle. 5767 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5768 _capacity / K, new_capacity / K); 5769 } 5770 } 5771 5772 5773 // Closures 5774 // XXX: there seems to be a lot of code duplication here; 5775 // should refactor and consolidate common code. 5776 5777 // This closure is used to mark refs into the CMS generation in 5778 // the CMS bit map. Called at the first checkpoint. This closure 5779 // assumes that we do not need to re-mark dirty cards; if the CMS 5780 // generation on which this is used is not an oldest 5781 // generation then this will lose younger_gen cards! 5782 5783 MarkRefsIntoClosure::MarkRefsIntoClosure( 5784 MemRegion span, CMSBitMap* bitMap): 5785 _span(span), 5786 _bitMap(bitMap) 5787 { 5788 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5789 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5790 } 5791 5792 void MarkRefsIntoClosure::do_oop(oop obj) { 5793 // if p points into _span, then mark corresponding bit in _markBitMap 5794 assert(oopDesc::is_oop(obj), "expected an oop"); 5795 HeapWord* addr = (HeapWord*)obj; 5796 if (_span.contains(addr)) { 5797 // this should be made more efficient 5798 _bitMap->mark(addr); 5799 } 5800 } 5801 5802 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5803 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5804 5805 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5806 MemRegion span, CMSBitMap* bitMap): 5807 _span(span), 5808 _bitMap(bitMap) 5809 { 5810 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5811 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5812 } 5813 5814 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5815 // if p points into _span, then mark corresponding bit in _markBitMap 5816 assert(oopDesc::is_oop(obj), "expected an oop"); 5817 HeapWord* addr = (HeapWord*)obj; 5818 if (_span.contains(addr)) { 5819 // this should be made more efficient 5820 _bitMap->par_mark(addr); 5821 } 5822 } 5823 5824 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5825 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5826 5827 // A variant of the above, used for CMS marking verification. 5828 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5829 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5830 _span(span), 5831 _verification_bm(verification_bm), 5832 _cms_bm(cms_bm) 5833 { 5834 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5835 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5836 } 5837 5838 void MarkRefsIntoVerifyClosure::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 _verification_bm->mark(addr); 5844 if (!_cms_bm->isMarked(addr)) { 5845 Log(gc, verify) log; 5846 ResourceMark rm; 5847 LogStream ls(log.error()); 5848 oop(addr)->print_on(&ls); 5849 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5850 fatal("... aborting"); 5851 } 5852 } 5853 } 5854 5855 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5856 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5857 5858 ////////////////////////////////////////////////// 5859 // MarkRefsIntoAndScanClosure 5860 ////////////////////////////////////////////////// 5861 5862 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5863 ReferenceDiscoverer* rd, 5864 CMSBitMap* bit_map, 5865 CMSBitMap* mod_union_table, 5866 CMSMarkStack* mark_stack, 5867 CMSCollector* collector, 5868 bool should_yield, 5869 bool concurrent_precleaning): 5870 _collector(collector), 5871 _span(span), 5872 _bit_map(bit_map), 5873 _mark_stack(mark_stack), 5874 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5875 mark_stack, concurrent_precleaning), 5876 _yield(should_yield), 5877 _concurrent_precleaning(concurrent_precleaning), 5878 _freelistLock(NULL) 5879 { 5880 // FIXME: Should initialize in base class constructor. 5881 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5882 set_ref_discoverer_internal(rd); 5883 } 5884 5885 // This closure is used to mark refs into the CMS generation at the 5886 // second (final) checkpoint, and to scan and transitively follow 5887 // the unmarked oops. It is also used during the concurrent precleaning 5888 // phase while scanning objects on dirty cards in the CMS generation. 5889 // The marks are made in the marking bit map and the marking stack is 5890 // used for keeping the (newly) grey objects during the scan. 5891 // The parallel version (Par_...) appears further below. 5892 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5893 if (obj != NULL) { 5894 assert(oopDesc::is_oop(obj), "expected an oop"); 5895 HeapWord* addr = (HeapWord*)obj; 5896 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5897 assert(_collector->overflow_list_is_empty(), 5898 "overflow list should be empty"); 5899 if (_span.contains(addr) && 5900 !_bit_map->isMarked(addr)) { 5901 // mark bit map (object is now grey) 5902 _bit_map->mark(addr); 5903 // push on marking stack (stack should be empty), and drain the 5904 // stack by applying this closure to the oops in the oops popped 5905 // from the stack (i.e. blacken the grey objects) 5906 bool res = _mark_stack->push(obj); 5907 assert(res, "Should have space to push on empty stack"); 5908 do { 5909 oop new_oop = _mark_stack->pop(); 5910 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5911 assert(_bit_map->isMarked((HeapWord*)new_oop), 5912 "only grey objects on this stack"); 5913 // iterate over the oops in this oop, marking and pushing 5914 // the ones in CMS heap (i.e. in _span). 5915 new_oop->oop_iterate(&_pushAndMarkClosure); 5916 // check if it's time to yield 5917 do_yield_check(); 5918 } while (!_mark_stack->isEmpty() || 5919 (!_concurrent_precleaning && take_from_overflow_list())); 5920 // if marking stack is empty, and we are not doing this 5921 // during precleaning, then check the overflow list 5922 } 5923 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5924 assert(_collector->overflow_list_is_empty(), 5925 "overflow list was drained above"); 5926 5927 assert(_collector->no_preserved_marks(), 5928 "All preserved marks should have been restored above"); 5929 } 5930 } 5931 5932 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5933 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5934 5935 void MarkRefsIntoAndScanClosure::do_yield_work() { 5936 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5937 "CMS thread should hold CMS token"); 5938 assert_lock_strong(_freelistLock); 5939 assert_lock_strong(_bit_map->lock()); 5940 // relinquish the free_list_lock and bitMaplock() 5941 _bit_map->lock()->unlock(); 5942 _freelistLock->unlock(); 5943 ConcurrentMarkSweepThread::desynchronize(true); 5944 _collector->stopTimer(); 5945 _collector->incrementYields(); 5946 5947 // See the comment in coordinator_yield() 5948 for (unsigned i = 0; 5949 i < CMSYieldSleepCount && 5950 ConcurrentMarkSweepThread::should_yield() && 5951 !CMSCollector::foregroundGCIsActive(); 5952 ++i) { 5953 os::sleep(Thread::current(), 1, false); 5954 } 5955 5956 ConcurrentMarkSweepThread::synchronize(true); 5957 _freelistLock->lock_without_safepoint_check(); 5958 _bit_map->lock()->lock_without_safepoint_check(); 5959 _collector->startTimer(); 5960 } 5961 5962 /////////////////////////////////////////////////////////// 5963 // ParMarkRefsIntoAndScanClosure: a parallel version of 5964 // MarkRefsIntoAndScanClosure 5965 /////////////////////////////////////////////////////////// 5966 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5967 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5968 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5969 _span(span), 5970 _bit_map(bit_map), 5971 _work_queue(work_queue), 5972 _low_water_mark(MIN2((work_queue->max_elems()/4), 5973 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5974 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 5975 { 5976 // FIXME: Should initialize in base class constructor. 5977 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5978 set_ref_discoverer_internal(rd); 5979 } 5980 5981 // This closure is used to mark refs into the CMS generation at the 5982 // second (final) checkpoint, and to scan and transitively follow 5983 // the unmarked oops. The marks are made in the marking bit map and 5984 // the work_queue is used for keeping the (newly) grey objects during 5985 // the scan phase whence they are also available for stealing by parallel 5986 // threads. Since the marking bit map is shared, updates are 5987 // synchronized (via CAS). 5988 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 5989 if (obj != NULL) { 5990 // Ignore mark word because this could be an already marked oop 5991 // that may be chained at the end of the overflow list. 5992 assert(oopDesc::is_oop(obj, true), "expected an oop"); 5993 HeapWord* addr = (HeapWord*)obj; 5994 if (_span.contains(addr) && 5995 !_bit_map->isMarked(addr)) { 5996 // mark bit map (object will become grey): 5997 // It is possible for several threads to be 5998 // trying to "claim" this object concurrently; 5999 // the unique thread that succeeds in marking the 6000 // object first will do the subsequent push on 6001 // to the work queue (or overflow list). 6002 if (_bit_map->par_mark(addr)) { 6003 // push on work_queue (which may not be empty), and trim the 6004 // queue to an appropriate length by applying this closure to 6005 // the oops in the oops popped from the stack (i.e. blacken the 6006 // grey objects) 6007 bool res = _work_queue->push(obj); 6008 assert(res, "Low water mark should be less than capacity?"); 6009 trim_queue(_low_water_mark); 6010 } // Else, another thread claimed the object 6011 } 6012 } 6013 } 6014 6015 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6016 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6017 6018 // This closure is used to rescan the marked objects on the dirty cards 6019 // in the mod union table and the card table proper. 6020 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6021 oop p, MemRegion mr) { 6022 6023 size_t size = 0; 6024 HeapWord* addr = (HeapWord*)p; 6025 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6026 assert(_span.contains(addr), "we are scanning the CMS generation"); 6027 // check if it's time to yield 6028 if (do_yield_check()) { 6029 // We yielded for some foreground stop-world work, 6030 // and we have been asked to abort this ongoing preclean cycle. 6031 return 0; 6032 } 6033 if (_bitMap->isMarked(addr)) { 6034 // it's marked; is it potentially uninitialized? 6035 if (p->klass_or_null_acquire() != NULL) { 6036 // an initialized object; ignore mark word in verification below 6037 // since we are running concurrent with mutators 6038 assert(oopDesc::is_oop(p, true), "should be an oop"); 6039 if (p->is_objArray()) { 6040 // objArrays are precisely marked; restrict scanning 6041 // to dirty cards only. 6042 size = CompactibleFreeListSpace::adjustObjectSize( 6043 p->oop_iterate_size(_scanningClosure, mr)); 6044 } else { 6045 // A non-array may have been imprecisely marked; we need 6046 // to scan object in its entirety. 6047 size = CompactibleFreeListSpace::adjustObjectSize( 6048 p->oop_iterate_size(_scanningClosure)); 6049 } 6050 #ifdef ASSERT 6051 size_t direct_size = 6052 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6053 assert(size == direct_size, "Inconsistency in size"); 6054 assert(size >= 3, "Necessary for Printezis marks to work"); 6055 HeapWord* start_pbit = addr + 1; 6056 HeapWord* end_pbit = addr + size - 1; 6057 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6058 "inconsistent Printezis mark"); 6059 // Verify inner mark bits (between Printezis bits) are clear, 6060 // but don't repeat if there are multiple dirty regions for 6061 // the same object, to avoid potential O(N^2) performance. 6062 if (addr != _last_scanned_object) { 6063 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6064 _last_scanned_object = addr; 6065 } 6066 #endif // ASSERT 6067 } else { 6068 // An uninitialized object. 6069 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6070 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6071 size = pointer_delta(nextOneAddr + 1, addr); 6072 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6073 "alignment problem"); 6074 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6075 // will dirty the card when the klass pointer is installed in the 6076 // object (signaling the completion of initialization). 6077 } 6078 } else { 6079 // Either a not yet marked object or an uninitialized object 6080 if (p->klass_or_null_acquire() == NULL) { 6081 // An uninitialized object, skip to the next card, since 6082 // we may not be able to read its P-bits yet. 6083 assert(size == 0, "Initial value"); 6084 } else { 6085 // An object not (yet) reached by marking: we merely need to 6086 // compute its size so as to go look at the next block. 6087 assert(oopDesc::is_oop(p, true), "should be an oop"); 6088 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6089 } 6090 } 6091 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6092 return size; 6093 } 6094 6095 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6096 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6097 "CMS thread should hold CMS token"); 6098 assert_lock_strong(_freelistLock); 6099 assert_lock_strong(_bitMap->lock()); 6100 // relinquish the free_list_lock and bitMaplock() 6101 _bitMap->lock()->unlock(); 6102 _freelistLock->unlock(); 6103 ConcurrentMarkSweepThread::desynchronize(true); 6104 _collector->stopTimer(); 6105 _collector->incrementYields(); 6106 6107 // See the comment in coordinator_yield() 6108 for (unsigned i = 0; i < CMSYieldSleepCount && 6109 ConcurrentMarkSweepThread::should_yield() && 6110 !CMSCollector::foregroundGCIsActive(); ++i) { 6111 os::sleep(Thread::current(), 1, false); 6112 } 6113 6114 ConcurrentMarkSweepThread::synchronize(true); 6115 _freelistLock->lock_without_safepoint_check(); 6116 _bitMap->lock()->lock_without_safepoint_check(); 6117 _collector->startTimer(); 6118 } 6119 6120 6121 ////////////////////////////////////////////////////////////////// 6122 // SurvivorSpacePrecleanClosure 6123 ////////////////////////////////////////////////////////////////// 6124 // This (single-threaded) closure is used to preclean the oops in 6125 // the survivor spaces. 6126 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6127 6128 HeapWord* addr = (HeapWord*)p; 6129 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6130 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6131 assert(p->klass_or_null() != NULL, "object should be initialized"); 6132 // an initialized object; ignore mark word in verification below 6133 // since we are running concurrent with mutators 6134 assert(oopDesc::is_oop(p, true), "should be an oop"); 6135 // Note that we do not yield while we iterate over 6136 // the interior oops of p, pushing the relevant ones 6137 // on our marking stack. 6138 size_t size = p->oop_iterate_size(_scanning_closure); 6139 do_yield_check(); 6140 // Observe that below, we do not abandon the preclean 6141 // phase as soon as we should; rather we empty the 6142 // marking stack before returning. This is to satisfy 6143 // some existing assertions. In general, it may be a 6144 // good idea to abort immediately and complete the marking 6145 // from the grey objects at a later time. 6146 while (!_mark_stack->isEmpty()) { 6147 oop new_oop = _mark_stack->pop(); 6148 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6149 assert(_bit_map->isMarked((HeapWord*)new_oop), 6150 "only grey objects on this stack"); 6151 // iterate over the oops in this oop, marking and pushing 6152 // the ones in CMS heap (i.e. in _span). 6153 new_oop->oop_iterate(_scanning_closure); 6154 // check if it's time to yield 6155 do_yield_check(); 6156 } 6157 unsigned int after_count = 6158 CMSHeap::heap()->total_collections(); 6159 bool abort = (_before_count != after_count) || 6160 _collector->should_abort_preclean(); 6161 return abort ? 0 : size; 6162 } 6163 6164 void SurvivorSpacePrecleanClosure::do_yield_work() { 6165 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6166 "CMS thread should hold CMS token"); 6167 assert_lock_strong(_bit_map->lock()); 6168 // Relinquish the bit map lock 6169 _bit_map->lock()->unlock(); 6170 ConcurrentMarkSweepThread::desynchronize(true); 6171 _collector->stopTimer(); 6172 _collector->incrementYields(); 6173 6174 // See the comment in coordinator_yield() 6175 for (unsigned i = 0; i < CMSYieldSleepCount && 6176 ConcurrentMarkSweepThread::should_yield() && 6177 !CMSCollector::foregroundGCIsActive(); ++i) { 6178 os::sleep(Thread::current(), 1, false); 6179 } 6180 6181 ConcurrentMarkSweepThread::synchronize(true); 6182 _bit_map->lock()->lock_without_safepoint_check(); 6183 _collector->startTimer(); 6184 } 6185 6186 // This closure is used to rescan the marked objects on the dirty cards 6187 // in the mod union table and the card table proper. In the parallel 6188 // case, although the bitMap is shared, we do a single read so the 6189 // isMarked() query is "safe". 6190 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6191 // Ignore mark word because we are running concurrent with mutators 6192 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6193 HeapWord* addr = (HeapWord*)p; 6194 assert(_span.contains(addr), "we are scanning the CMS generation"); 6195 bool is_obj_array = false; 6196 #ifdef ASSERT 6197 if (!_parallel) { 6198 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6199 assert(_collector->overflow_list_is_empty(), 6200 "overflow list should be empty"); 6201 6202 } 6203 #endif // ASSERT 6204 if (_bit_map->isMarked(addr)) { 6205 // Obj arrays are precisely marked, non-arrays are not; 6206 // so we scan objArrays precisely and non-arrays in their 6207 // entirety. 6208 if (p->is_objArray()) { 6209 is_obj_array = true; 6210 if (_parallel) { 6211 p->oop_iterate(_par_scan_closure, mr); 6212 } else { 6213 p->oop_iterate(_scan_closure, mr); 6214 } 6215 } else { 6216 if (_parallel) { 6217 p->oop_iterate(_par_scan_closure); 6218 } else { 6219 p->oop_iterate(_scan_closure); 6220 } 6221 } 6222 } 6223 #ifdef ASSERT 6224 if (!_parallel) { 6225 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6226 assert(_collector->overflow_list_is_empty(), 6227 "overflow list should be empty"); 6228 6229 } 6230 #endif // ASSERT 6231 return is_obj_array; 6232 } 6233 6234 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6235 MemRegion span, 6236 CMSBitMap* bitMap, CMSMarkStack* markStack, 6237 bool should_yield, bool verifying): 6238 _collector(collector), 6239 _span(span), 6240 _bitMap(bitMap), 6241 _mut(&collector->_modUnionTable), 6242 _markStack(markStack), 6243 _yield(should_yield), 6244 _skipBits(0) 6245 { 6246 assert(_markStack->isEmpty(), "stack should be empty"); 6247 _finger = _bitMap->startWord(); 6248 _threshold = _finger; 6249 assert(_collector->_restart_addr == NULL, "Sanity check"); 6250 assert(_span.contains(_finger), "Out of bounds _finger?"); 6251 DEBUG_ONLY(_verifying = verifying;) 6252 } 6253 6254 void MarkFromRootsClosure::reset(HeapWord* addr) { 6255 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6256 assert(_span.contains(addr), "Out of bounds _finger?"); 6257 _finger = addr; 6258 _threshold = align_up(_finger, CardTable::card_size); 6259 } 6260 6261 // Should revisit to see if this should be restructured for 6262 // greater efficiency. 6263 bool MarkFromRootsClosure::do_bit(size_t offset) { 6264 if (_skipBits > 0) { 6265 _skipBits--; 6266 return true; 6267 } 6268 // convert offset into a HeapWord* 6269 HeapWord* addr = _bitMap->startWord() + offset; 6270 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6271 "address out of range"); 6272 assert(_bitMap->isMarked(addr), "tautology"); 6273 if (_bitMap->isMarked(addr+1)) { 6274 // this is an allocated but not yet initialized object 6275 assert(_skipBits == 0, "tautology"); 6276 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6277 oop p = oop(addr); 6278 if (p->klass_or_null_acquire() == NULL) { 6279 DEBUG_ONLY(if (!_verifying) {) 6280 // We re-dirty the cards on which this object lies and increase 6281 // the _threshold so that we'll come back to scan this object 6282 // during the preclean or remark phase. (CMSCleanOnEnter) 6283 if (CMSCleanOnEnter) { 6284 size_t sz = _collector->block_size_using_printezis_bits(addr); 6285 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6286 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6287 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6288 // Bump _threshold to end_card_addr; note that 6289 // _threshold cannot possibly exceed end_card_addr, anyhow. 6290 // This prevents future clearing of the card as the scan proceeds 6291 // to the right. 6292 assert(_threshold <= end_card_addr, 6293 "Because we are just scanning into this object"); 6294 if (_threshold < end_card_addr) { 6295 _threshold = end_card_addr; 6296 } 6297 if (p->klass_or_null_acquire() != NULL) { 6298 // Redirty the range of cards... 6299 _mut->mark_range(redirty_range); 6300 } // ...else the setting of klass will dirty the card anyway. 6301 } 6302 DEBUG_ONLY(}) 6303 return true; 6304 } 6305 } 6306 scanOopsInOop(addr); 6307 return true; 6308 } 6309 6310 // We take a break if we've been at this for a while, 6311 // so as to avoid monopolizing the locks involved. 6312 void MarkFromRootsClosure::do_yield_work() { 6313 // First give up the locks, then yield, then re-lock 6314 // We should probably use a constructor/destructor idiom to 6315 // do this unlock/lock or modify the MutexUnlocker class to 6316 // serve our purpose. XXX 6317 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6318 "CMS thread should hold CMS token"); 6319 assert_lock_strong(_bitMap->lock()); 6320 _bitMap->lock()->unlock(); 6321 ConcurrentMarkSweepThread::desynchronize(true); 6322 _collector->stopTimer(); 6323 _collector->incrementYields(); 6324 6325 // See the comment in coordinator_yield() 6326 for (unsigned i = 0; i < CMSYieldSleepCount && 6327 ConcurrentMarkSweepThread::should_yield() && 6328 !CMSCollector::foregroundGCIsActive(); ++i) { 6329 os::sleep(Thread::current(), 1, false); 6330 } 6331 6332 ConcurrentMarkSweepThread::synchronize(true); 6333 _bitMap->lock()->lock_without_safepoint_check(); 6334 _collector->startTimer(); 6335 } 6336 6337 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6338 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6339 assert(_markStack->isEmpty(), 6340 "should drain stack to limit stack usage"); 6341 // convert ptr to an oop preparatory to scanning 6342 oop obj = oop(ptr); 6343 // Ignore mark word in verification below, since we 6344 // may be running concurrent with mutators. 6345 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6346 assert(_finger <= ptr, "_finger runneth ahead"); 6347 // advance the finger to right end of this object 6348 _finger = ptr + obj->size(); 6349 assert(_finger > ptr, "we just incremented it above"); 6350 // On large heaps, it may take us some time to get through 6351 // the marking phase. During 6352 // this time it's possible that a lot of mutations have 6353 // accumulated in the card table and the mod union table -- 6354 // these mutation records are redundant until we have 6355 // actually traced into the corresponding card. 6356 // Here, we check whether advancing the finger would make 6357 // us cross into a new card, and if so clear corresponding 6358 // cards in the MUT (preclean them in the card-table in the 6359 // future). 6360 6361 DEBUG_ONLY(if (!_verifying) {) 6362 // The clean-on-enter optimization is disabled by default, 6363 // until we fix 6178663. 6364 if (CMSCleanOnEnter && (_finger > _threshold)) { 6365 // [_threshold, _finger) represents the interval 6366 // of cards to be cleared in MUT (or precleaned in card table). 6367 // The set of cards to be cleared is all those that overlap 6368 // with the interval [_threshold, _finger); note that 6369 // _threshold is always kept card-aligned but _finger isn't 6370 // always card-aligned. 6371 HeapWord* old_threshold = _threshold; 6372 assert(is_aligned(old_threshold, CardTable::card_size), 6373 "_threshold should always be card-aligned"); 6374 _threshold = align_up(_finger, CardTable::card_size); 6375 MemRegion mr(old_threshold, _threshold); 6376 assert(!mr.is_empty(), "Control point invariant"); 6377 assert(_span.contains(mr), "Should clear within span"); 6378 _mut->clear_range(mr); 6379 } 6380 DEBUG_ONLY(}) 6381 // Note: the finger doesn't advance while we drain 6382 // the stack below. 6383 PushOrMarkClosure pushOrMarkClosure(_collector, 6384 _span, _bitMap, _markStack, 6385 _finger, this); 6386 bool res = _markStack->push(obj); 6387 assert(res, "Empty non-zero size stack should have space for single push"); 6388 while (!_markStack->isEmpty()) { 6389 oop new_oop = _markStack->pop(); 6390 // Skip verifying header mark word below because we are 6391 // running concurrent with mutators. 6392 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6393 // now scan this oop's oops 6394 new_oop->oop_iterate(&pushOrMarkClosure); 6395 do_yield_check(); 6396 } 6397 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6398 } 6399 6400 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6401 CMSCollector* collector, MemRegion span, 6402 CMSBitMap* bit_map, 6403 OopTaskQueue* work_queue, 6404 CMSMarkStack* overflow_stack): 6405 _collector(collector), 6406 _whole_span(collector->_span), 6407 _span(span), 6408 _bit_map(bit_map), 6409 _mut(&collector->_modUnionTable), 6410 _work_queue(work_queue), 6411 _overflow_stack(overflow_stack), 6412 _skip_bits(0), 6413 _task(task) 6414 { 6415 assert(_work_queue->size() == 0, "work_queue should be empty"); 6416 _finger = span.start(); 6417 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6418 assert(_span.contains(_finger), "Out of bounds _finger?"); 6419 } 6420 6421 // Should revisit to see if this should be restructured for 6422 // greater efficiency. 6423 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6424 if (_skip_bits > 0) { 6425 _skip_bits--; 6426 return true; 6427 } 6428 // convert offset into a HeapWord* 6429 HeapWord* addr = _bit_map->startWord() + offset; 6430 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6431 "address out of range"); 6432 assert(_bit_map->isMarked(addr), "tautology"); 6433 if (_bit_map->isMarked(addr+1)) { 6434 // this is an allocated object that might not yet be initialized 6435 assert(_skip_bits == 0, "tautology"); 6436 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6437 oop p = oop(addr); 6438 if (p->klass_or_null_acquire() == NULL) { 6439 // in the case of Clean-on-Enter optimization, redirty card 6440 // and avoid clearing card by increasing the threshold. 6441 return true; 6442 } 6443 } 6444 scan_oops_in_oop(addr); 6445 return true; 6446 } 6447 6448 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6449 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6450 // Should we assert that our work queue is empty or 6451 // below some drain limit? 6452 assert(_work_queue->size() == 0, 6453 "should drain stack to limit stack usage"); 6454 // convert ptr to an oop preparatory to scanning 6455 oop obj = oop(ptr); 6456 // Ignore mark word in verification below, since we 6457 // may be running concurrent with mutators. 6458 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6459 assert(_finger <= ptr, "_finger runneth ahead"); 6460 // advance the finger to right end of this object 6461 _finger = ptr + obj->size(); 6462 assert(_finger > ptr, "we just incremented it above"); 6463 // On large heaps, it may take us some time to get through 6464 // the marking phase. During 6465 // this time it's possible that a lot of mutations have 6466 // accumulated in the card table and the mod union table -- 6467 // these mutation records are redundant until we have 6468 // actually traced into the corresponding card. 6469 // Here, we check whether advancing the finger would make 6470 // us cross into a new card, and if so clear corresponding 6471 // cards in the MUT (preclean them in the card-table in the 6472 // future). 6473 6474 // The clean-on-enter optimization is disabled by default, 6475 // until we fix 6178663. 6476 if (CMSCleanOnEnter && (_finger > _threshold)) { 6477 // [_threshold, _finger) represents the interval 6478 // of cards to be cleared in MUT (or precleaned in card table). 6479 // The set of cards to be cleared is all those that overlap 6480 // with the interval [_threshold, _finger); note that 6481 // _threshold is always kept card-aligned but _finger isn't 6482 // always card-aligned. 6483 HeapWord* old_threshold = _threshold; 6484 assert(is_aligned(old_threshold, CardTable::card_size), 6485 "_threshold should always be card-aligned"); 6486 _threshold = align_up(_finger, CardTable::card_size); 6487 MemRegion mr(old_threshold, _threshold); 6488 assert(!mr.is_empty(), "Control point invariant"); 6489 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6490 _mut->clear_range(mr); 6491 } 6492 6493 // Note: the local finger doesn't advance while we drain 6494 // the stack below, but the global finger sure can and will. 6495 HeapWord* volatile* gfa = _task->global_finger_addr(); 6496 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6497 _span, _bit_map, 6498 _work_queue, 6499 _overflow_stack, 6500 _finger, 6501 gfa, this); 6502 bool res = _work_queue->push(obj); // overflow could occur here 6503 assert(res, "Will hold once we use workqueues"); 6504 while (true) { 6505 oop new_oop; 6506 if (!_work_queue->pop_local(new_oop)) { 6507 // We emptied our work_queue; check if there's stuff that can 6508 // be gotten from the overflow stack. 6509 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6510 _overflow_stack, _work_queue)) { 6511 do_yield_check(); 6512 continue; 6513 } else { // done 6514 break; 6515 } 6516 } 6517 // Skip verifying header mark word below because we are 6518 // running concurrent with mutators. 6519 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6520 // now scan this oop's oops 6521 new_oop->oop_iterate(&pushOrMarkClosure); 6522 do_yield_check(); 6523 } 6524 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6525 } 6526 6527 // Yield in response to a request from VM Thread or 6528 // from mutators. 6529 void ParMarkFromRootsClosure::do_yield_work() { 6530 assert(_task != NULL, "sanity"); 6531 _task->yield(); 6532 } 6533 6534 // A variant of the above used for verifying CMS marking work. 6535 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6536 MemRegion span, 6537 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6538 CMSMarkStack* mark_stack): 6539 _collector(collector), 6540 _span(span), 6541 _verification_bm(verification_bm), 6542 _cms_bm(cms_bm), 6543 _mark_stack(mark_stack), 6544 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6545 mark_stack) 6546 { 6547 assert(_mark_stack->isEmpty(), "stack should be empty"); 6548 _finger = _verification_bm->startWord(); 6549 assert(_collector->_restart_addr == NULL, "Sanity check"); 6550 assert(_span.contains(_finger), "Out of bounds _finger?"); 6551 } 6552 6553 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6554 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6555 assert(_span.contains(addr), "Out of bounds _finger?"); 6556 _finger = addr; 6557 } 6558 6559 // Should revisit to see if this should be restructured for 6560 // greater efficiency. 6561 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6562 // convert offset into a HeapWord* 6563 HeapWord* addr = _verification_bm->startWord() + offset; 6564 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6565 "address out of range"); 6566 assert(_verification_bm->isMarked(addr), "tautology"); 6567 assert(_cms_bm->isMarked(addr), "tautology"); 6568 6569 assert(_mark_stack->isEmpty(), 6570 "should drain stack to limit stack usage"); 6571 // convert addr to an oop preparatory to scanning 6572 oop obj = oop(addr); 6573 assert(oopDesc::is_oop(obj), "should be an oop"); 6574 assert(_finger <= addr, "_finger runneth ahead"); 6575 // advance the finger to right end of this object 6576 _finger = addr + obj->size(); 6577 assert(_finger > addr, "we just incremented it above"); 6578 // Note: the finger doesn't advance while we drain 6579 // the stack below. 6580 bool res = _mark_stack->push(obj); 6581 assert(res, "Empty non-zero size stack should have space for single push"); 6582 while (!_mark_stack->isEmpty()) { 6583 oop new_oop = _mark_stack->pop(); 6584 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6585 // now scan this oop's oops 6586 new_oop->oop_iterate(&_pam_verify_closure); 6587 } 6588 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6589 return true; 6590 } 6591 6592 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6593 CMSCollector* collector, MemRegion span, 6594 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6595 CMSMarkStack* mark_stack): 6596 MetadataAwareOopClosure(collector->ref_processor()), 6597 _collector(collector), 6598 _span(span), 6599 _verification_bm(verification_bm), 6600 _cms_bm(cms_bm), 6601 _mark_stack(mark_stack) 6602 { } 6603 6604 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6605 oop obj = RawAccess<>::oop_load(p); 6606 do_oop(obj); 6607 } 6608 6609 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6610 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6611 6612 // Upon stack overflow, we discard (part of) the stack, 6613 // remembering the least address amongst those discarded 6614 // in CMSCollector's _restart_address. 6615 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6616 // Remember the least grey address discarded 6617 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6618 _collector->lower_restart_addr(ra); 6619 _mark_stack->reset(); // discard stack contents 6620 _mark_stack->expand(); // expand the stack if possible 6621 } 6622 6623 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6624 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6625 HeapWord* addr = (HeapWord*)obj; 6626 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6627 // Oop lies in _span and isn't yet grey or black 6628 _verification_bm->mark(addr); // now grey 6629 if (!_cms_bm->isMarked(addr)) { 6630 Log(gc, verify) log; 6631 ResourceMark rm; 6632 LogStream ls(log.error()); 6633 oop(addr)->print_on(&ls); 6634 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6635 fatal("... aborting"); 6636 } 6637 6638 if (!_mark_stack->push(obj)) { // stack overflow 6639 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6640 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6641 handle_stack_overflow(addr); 6642 } 6643 // anything including and to the right of _finger 6644 // will be scanned as we iterate over the remainder of the 6645 // bit map 6646 } 6647 } 6648 6649 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6650 MemRegion span, 6651 CMSBitMap* bitMap, CMSMarkStack* markStack, 6652 HeapWord* finger, MarkFromRootsClosure* parent) : 6653 MetadataAwareOopClosure(collector->ref_processor()), 6654 _collector(collector), 6655 _span(span), 6656 _bitMap(bitMap), 6657 _markStack(markStack), 6658 _finger(finger), 6659 _parent(parent) 6660 { } 6661 6662 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6663 MemRegion span, 6664 CMSBitMap* bit_map, 6665 OopTaskQueue* work_queue, 6666 CMSMarkStack* overflow_stack, 6667 HeapWord* finger, 6668 HeapWord* volatile* global_finger_addr, 6669 ParMarkFromRootsClosure* parent) : 6670 MetadataAwareOopClosure(collector->ref_processor()), 6671 _collector(collector), 6672 _whole_span(collector->_span), 6673 _span(span), 6674 _bit_map(bit_map), 6675 _work_queue(work_queue), 6676 _overflow_stack(overflow_stack), 6677 _finger(finger), 6678 _global_finger_addr(global_finger_addr), 6679 _parent(parent) 6680 { } 6681 6682 // Assumes thread-safe access by callers, who are 6683 // responsible for mutual exclusion. 6684 void CMSCollector::lower_restart_addr(HeapWord* low) { 6685 assert(_span.contains(low), "Out of bounds addr"); 6686 if (_restart_addr == NULL) { 6687 _restart_addr = low; 6688 } else { 6689 _restart_addr = MIN2(_restart_addr, low); 6690 } 6691 } 6692 6693 // Upon stack overflow, we discard (part of) the stack, 6694 // remembering the least address amongst those discarded 6695 // in CMSCollector's _restart_address. 6696 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6697 // Remember the least grey address discarded 6698 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6699 _collector->lower_restart_addr(ra); 6700 _markStack->reset(); // discard stack contents 6701 _markStack->expand(); // expand the stack if possible 6702 } 6703 6704 // Upon stack overflow, we discard (part of) the stack, 6705 // remembering the least address amongst those discarded 6706 // in CMSCollector's _restart_address. 6707 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6708 // We need to do this under a mutex to prevent other 6709 // workers from interfering with the work done below. 6710 MutexLockerEx ml(_overflow_stack->par_lock(), 6711 Mutex::_no_safepoint_check_flag); 6712 // Remember the least grey address discarded 6713 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6714 _collector->lower_restart_addr(ra); 6715 _overflow_stack->reset(); // discard stack contents 6716 _overflow_stack->expand(); // expand the stack if possible 6717 } 6718 6719 void PushOrMarkClosure::do_oop(oop obj) { 6720 // Ignore mark word because we are running concurrent with mutators. 6721 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6722 HeapWord* addr = (HeapWord*)obj; 6723 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6724 // Oop lies in _span and isn't yet grey or black 6725 _bitMap->mark(addr); // now grey 6726 if (addr < _finger) { 6727 // the bit map iteration has already either passed, or 6728 // sampled, this bit in the bit map; we'll need to 6729 // use the marking stack to scan this oop's oops. 6730 bool simulate_overflow = false; 6731 NOT_PRODUCT( 6732 if (CMSMarkStackOverflowALot && 6733 _collector->simulate_overflow()) { 6734 // simulate a stack overflow 6735 simulate_overflow = true; 6736 } 6737 ) 6738 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6739 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6740 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6741 handle_stack_overflow(addr); 6742 } 6743 } 6744 // anything including and to the right of _finger 6745 // will be scanned as we iterate over the remainder of the 6746 // bit map 6747 do_yield_check(); 6748 } 6749 } 6750 6751 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6752 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6753 6754 void ParPushOrMarkClosure::do_oop(oop obj) { 6755 // Ignore mark word because we are running concurrent with mutators. 6756 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6757 HeapWord* addr = (HeapWord*)obj; 6758 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6759 // Oop lies in _span and isn't yet grey or black 6760 // We read the global_finger (volatile read) strictly after marking oop 6761 bool res = _bit_map->par_mark(addr); // now grey 6762 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6763 // Should we push this marked oop on our stack? 6764 // -- if someone else marked it, nothing to do 6765 // -- if target oop is above global finger nothing to do 6766 // -- if target oop is in chunk and above local finger 6767 // then nothing to do 6768 // -- else push on work queue 6769 if ( !res // someone else marked it, they will deal with it 6770 || (addr >= *gfa) // will be scanned in a later task 6771 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6772 return; 6773 } 6774 // the bit map iteration has already either passed, or 6775 // sampled, this bit in the bit map; we'll need to 6776 // use the marking stack to scan this oop's oops. 6777 bool simulate_overflow = false; 6778 NOT_PRODUCT( 6779 if (CMSMarkStackOverflowALot && 6780 _collector->simulate_overflow()) { 6781 // simulate a stack overflow 6782 simulate_overflow = true; 6783 } 6784 ) 6785 if (simulate_overflow || 6786 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6787 // stack overflow 6788 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6789 // We cannot assert that the overflow stack is full because 6790 // it may have been emptied since. 6791 assert(simulate_overflow || 6792 _work_queue->size() == _work_queue->max_elems(), 6793 "Else push should have succeeded"); 6794 handle_stack_overflow(addr); 6795 } 6796 do_yield_check(); 6797 } 6798 } 6799 6800 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6801 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6802 6803 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6804 MemRegion span, 6805 ReferenceDiscoverer* rd, 6806 CMSBitMap* bit_map, 6807 CMSBitMap* mod_union_table, 6808 CMSMarkStack* mark_stack, 6809 bool concurrent_precleaning): 6810 MetadataAwareOopClosure(rd), 6811 _collector(collector), 6812 _span(span), 6813 _bit_map(bit_map), 6814 _mod_union_table(mod_union_table), 6815 _mark_stack(mark_stack), 6816 _concurrent_precleaning(concurrent_precleaning) 6817 { 6818 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6819 } 6820 6821 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6822 // the non-parallel version (the parallel version appears further below.) 6823 void PushAndMarkClosure::do_oop(oop obj) { 6824 // Ignore mark word verification. If during concurrent precleaning, 6825 // the object monitor may be locked. If during the checkpoint 6826 // phases, the object may already have been reached by a different 6827 // path and may be at the end of the global overflow list (so 6828 // the mark word may be NULL). 6829 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6830 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6831 HeapWord* addr = (HeapWord*)obj; 6832 // Check if oop points into the CMS generation 6833 // and is not marked 6834 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6835 // a white object ... 6836 _bit_map->mark(addr); // ... now grey 6837 // push on the marking stack (grey set) 6838 bool simulate_overflow = false; 6839 NOT_PRODUCT( 6840 if (CMSMarkStackOverflowALot && 6841 _collector->simulate_overflow()) { 6842 // simulate a stack overflow 6843 simulate_overflow = true; 6844 } 6845 ) 6846 if (simulate_overflow || !_mark_stack->push(obj)) { 6847 if (_concurrent_precleaning) { 6848 // During precleaning we can just dirty the appropriate card(s) 6849 // in the mod union table, thus ensuring that the object remains 6850 // in the grey set and continue. In the case of object arrays 6851 // we need to dirty all of the cards that the object spans, 6852 // since the rescan of object arrays will be limited to the 6853 // dirty cards. 6854 // Note that no one can be interfering with us in this action 6855 // of dirtying the mod union table, so no locking or atomics 6856 // are required. 6857 if (obj->is_objArray()) { 6858 size_t sz = obj->size(); 6859 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6860 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6861 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6862 _mod_union_table->mark_range(redirty_range); 6863 } else { 6864 _mod_union_table->mark(addr); 6865 } 6866 _collector->_ser_pmc_preclean_ovflw++; 6867 } else { 6868 // During the remark phase, we need to remember this oop 6869 // in the overflow list. 6870 _collector->push_on_overflow_list(obj); 6871 _collector->_ser_pmc_remark_ovflw++; 6872 } 6873 } 6874 } 6875 } 6876 6877 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6878 MemRegion span, 6879 ReferenceDiscoverer* rd, 6880 CMSBitMap* bit_map, 6881 OopTaskQueue* work_queue): 6882 MetadataAwareOopClosure(rd), 6883 _collector(collector), 6884 _span(span), 6885 _bit_map(bit_map), 6886 _work_queue(work_queue) 6887 { 6888 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6889 } 6890 6891 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6892 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6893 6894 // Grey object rescan during second checkpoint phase -- 6895 // the parallel version. 6896 void ParPushAndMarkClosure::do_oop(oop obj) { 6897 // In the assert below, we ignore the mark word because 6898 // this oop may point to an already visited object that is 6899 // on the overflow stack (in which case the mark word has 6900 // been hijacked for chaining into the overflow stack -- 6901 // if this is the last object in the overflow stack then 6902 // its mark word will be NULL). Because this object may 6903 // have been subsequently popped off the global overflow 6904 // stack, and the mark word possibly restored to the prototypical 6905 // value, by the time we get to examined this failing assert in 6906 // the debugger, is_oop_or_null(false) may subsequently start 6907 // to hold. 6908 assert(oopDesc::is_oop_or_null(obj, true), 6909 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6910 HeapWord* addr = (HeapWord*)obj; 6911 // Check if oop points into the CMS generation 6912 // and is not marked 6913 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6914 // a white object ... 6915 // If we manage to "claim" the object, by being the 6916 // first thread to mark it, then we push it on our 6917 // marking stack 6918 if (_bit_map->par_mark(addr)) { // ... now grey 6919 // push on work queue (grey set) 6920 bool simulate_overflow = false; 6921 NOT_PRODUCT( 6922 if (CMSMarkStackOverflowALot && 6923 _collector->par_simulate_overflow()) { 6924 // simulate a stack overflow 6925 simulate_overflow = true; 6926 } 6927 ) 6928 if (simulate_overflow || !_work_queue->push(obj)) { 6929 _collector->par_push_on_overflow_list(obj); 6930 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6931 } 6932 } // Else, some other thread got there first 6933 } 6934 } 6935 6936 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6937 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6938 6939 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6940 Mutex* bml = _collector->bitMapLock(); 6941 assert_lock_strong(bml); 6942 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6943 "CMS thread should hold CMS token"); 6944 6945 bml->unlock(); 6946 ConcurrentMarkSweepThread::desynchronize(true); 6947 6948 _collector->stopTimer(); 6949 _collector->incrementYields(); 6950 6951 // See the comment in coordinator_yield() 6952 for (unsigned i = 0; i < CMSYieldSleepCount && 6953 ConcurrentMarkSweepThread::should_yield() && 6954 !CMSCollector::foregroundGCIsActive(); ++i) { 6955 os::sleep(Thread::current(), 1, false); 6956 } 6957 6958 ConcurrentMarkSweepThread::synchronize(true); 6959 bml->lock(); 6960 6961 _collector->startTimer(); 6962 } 6963 6964 bool CMSPrecleanRefsYieldClosure::should_return() { 6965 if (ConcurrentMarkSweepThread::should_yield()) { 6966 do_yield_work(); 6967 } 6968 return _collector->foregroundGCIsActive(); 6969 } 6970 6971 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6972 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6973 "mr should be aligned to start at a card boundary"); 6974 // We'd like to assert: 6975 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6976 // "mr should be a range of cards"); 6977 // However, that would be too strong in one case -- the last 6978 // partition ends at _unallocated_block which, in general, can be 6979 // an arbitrary boundary, not necessarily card aligned. 6980 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 6981 _space->object_iterate_mem(mr, &_scan_cl); 6982 } 6983 6984 SweepClosure::SweepClosure(CMSCollector* collector, 6985 ConcurrentMarkSweepGeneration* g, 6986 CMSBitMap* bitMap, bool should_yield) : 6987 _collector(collector), 6988 _g(g), 6989 _sp(g->cmsSpace()), 6990 _limit(_sp->sweep_limit()), 6991 _freelistLock(_sp->freelistLock()), 6992 _bitMap(bitMap), 6993 _yield(should_yield), 6994 _inFreeRange(false), // No free range at beginning of sweep 6995 _freeRangeInFreeLists(false), // No free range at beginning of sweep 6996 _lastFreeRangeCoalesced(false), 6997 _freeFinger(g->used_region().start()) 6998 { 6999 NOT_PRODUCT( 7000 _numObjectsFreed = 0; 7001 _numWordsFreed = 0; 7002 _numObjectsLive = 0; 7003 _numWordsLive = 0; 7004 _numObjectsAlreadyFree = 0; 7005 _numWordsAlreadyFree = 0; 7006 _last_fc = NULL; 7007 7008 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7009 _sp->dictionary()->initialize_dict_returned_bytes(); 7010 ) 7011 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7012 "sweep _limit out of bounds"); 7013 log_develop_trace(gc, sweep)("===================="); 7014 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7015 } 7016 7017 void SweepClosure::print_on(outputStream* st) const { 7018 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7019 p2i(_sp->bottom()), p2i(_sp->end())); 7020 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7021 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7022 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7023 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7024 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7025 } 7026 7027 #ifndef PRODUCT 7028 // Assertion checking only: no useful work in product mode -- 7029 // however, if any of the flags below become product flags, 7030 // you may need to review this code to see if it needs to be 7031 // enabled in product mode. 7032 SweepClosure::~SweepClosure() { 7033 assert_lock_strong(_freelistLock); 7034 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7035 "sweep _limit out of bounds"); 7036 if (inFreeRange()) { 7037 Log(gc, sweep) log; 7038 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7039 ResourceMark rm; 7040 LogStream ls(log.error()); 7041 print_on(&ls); 7042 ShouldNotReachHere(); 7043 } 7044 7045 if (log_is_enabled(Debug, gc, sweep)) { 7046 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7047 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7048 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7049 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7050 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7051 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7052 } 7053 7054 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7055 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7056 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7057 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7058 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7059 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7060 } 7061 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7062 log_develop_trace(gc, sweep)("================"); 7063 } 7064 #endif // PRODUCT 7065 7066 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7067 bool freeRangeInFreeLists) { 7068 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7069 p2i(freeFinger), freeRangeInFreeLists); 7070 assert(!inFreeRange(), "Trampling existing free range"); 7071 set_inFreeRange(true); 7072 set_lastFreeRangeCoalesced(false); 7073 7074 set_freeFinger(freeFinger); 7075 set_freeRangeInFreeLists(freeRangeInFreeLists); 7076 if (CMSTestInFreeList) { 7077 if (freeRangeInFreeLists) { 7078 FreeChunk* fc = (FreeChunk*) freeFinger; 7079 assert(fc->is_free(), "A chunk on the free list should be free."); 7080 assert(fc->size() > 0, "Free range should have a size"); 7081 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7082 } 7083 } 7084 } 7085 7086 // Note that the sweeper runs concurrently with mutators. Thus, 7087 // it is possible for direct allocation in this generation to happen 7088 // in the middle of the sweep. Note that the sweeper also coalesces 7089 // contiguous free blocks. Thus, unless the sweeper and the allocator 7090 // synchronize appropriately freshly allocated blocks may get swept up. 7091 // This is accomplished by the sweeper locking the free lists while 7092 // it is sweeping. Thus blocks that are determined to be free are 7093 // indeed free. There is however one additional complication: 7094 // blocks that have been allocated since the final checkpoint and 7095 // mark, will not have been marked and so would be treated as 7096 // unreachable and swept up. To prevent this, the allocator marks 7097 // the bit map when allocating during the sweep phase. This leads, 7098 // however, to a further complication -- objects may have been allocated 7099 // but not yet initialized -- in the sense that the header isn't yet 7100 // installed. The sweeper can not then determine the size of the block 7101 // in order to skip over it. To deal with this case, we use a technique 7102 // (due to Printezis) to encode such uninitialized block sizes in the 7103 // bit map. Since the bit map uses a bit per every HeapWord, but the 7104 // CMS generation has a minimum object size of 3 HeapWords, it follows 7105 // that "normal marks" won't be adjacent in the bit map (there will 7106 // always be at least two 0 bits between successive 1 bits). We make use 7107 // of these "unused" bits to represent uninitialized blocks -- the bit 7108 // corresponding to the start of the uninitialized object and the next 7109 // bit are both set. Finally, a 1 bit marks the end of the object that 7110 // started with the two consecutive 1 bits to indicate its potentially 7111 // uninitialized state. 7112 7113 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7114 FreeChunk* fc = (FreeChunk*)addr; 7115 size_t res; 7116 7117 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7118 // than "addr == _limit" because although _limit was a block boundary when 7119 // we started the sweep, it may no longer be one because heap expansion 7120 // may have caused us to coalesce the block ending at the address _limit 7121 // with a newly expanded chunk (this happens when _limit was set to the 7122 // previous _end of the space), so we may have stepped past _limit: 7123 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7124 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7125 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7126 "sweep _limit out of bounds"); 7127 assert(addr < _sp->end(), "addr out of bounds"); 7128 // Flush any free range we might be holding as a single 7129 // coalesced chunk to the appropriate free list. 7130 if (inFreeRange()) { 7131 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7132 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7133 flush_cur_free_chunk(freeFinger(), 7134 pointer_delta(addr, freeFinger())); 7135 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7136 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7137 lastFreeRangeCoalesced() ? 1 : 0); 7138 } 7139 7140 // help the iterator loop finish 7141 return pointer_delta(_sp->end(), addr); 7142 } 7143 7144 assert(addr < _limit, "sweep invariant"); 7145 // check if we should yield 7146 do_yield_check(addr); 7147 if (fc->is_free()) { 7148 // Chunk that is already free 7149 res = fc->size(); 7150 do_already_free_chunk(fc); 7151 debug_only(_sp->verifyFreeLists()); 7152 // If we flush the chunk at hand in lookahead_and_flush() 7153 // and it's coalesced with a preceding chunk, then the 7154 // process of "mangling" the payload of the coalesced block 7155 // will cause erasure of the size information from the 7156 // (erstwhile) header of all the coalesced blocks but the 7157 // first, so the first disjunct in the assert will not hold 7158 // in that specific case (in which case the second disjunct 7159 // will hold). 7160 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7161 "Otherwise the size info doesn't change at this step"); 7162 NOT_PRODUCT( 7163 _numObjectsAlreadyFree++; 7164 _numWordsAlreadyFree += res; 7165 ) 7166 NOT_PRODUCT(_last_fc = fc;) 7167 } else if (!_bitMap->isMarked(addr)) { 7168 // Chunk is fresh garbage 7169 res = do_garbage_chunk(fc); 7170 debug_only(_sp->verifyFreeLists()); 7171 NOT_PRODUCT( 7172 _numObjectsFreed++; 7173 _numWordsFreed += res; 7174 ) 7175 } else { 7176 // Chunk that is alive. 7177 res = do_live_chunk(fc); 7178 debug_only(_sp->verifyFreeLists()); 7179 NOT_PRODUCT( 7180 _numObjectsLive++; 7181 _numWordsLive += res; 7182 ) 7183 } 7184 return res; 7185 } 7186 7187 // For the smart allocation, record following 7188 // split deaths - a free chunk is removed from its free list because 7189 // it is being split into two or more chunks. 7190 // split birth - a free chunk is being added to its free list because 7191 // a larger free chunk has been split and resulted in this free chunk. 7192 // coal death - a free chunk is being removed from its free list because 7193 // it is being coalesced into a large free chunk. 7194 // coal birth - a free chunk is being added to its free list because 7195 // it was created when two or more free chunks where coalesced into 7196 // this free chunk. 7197 // 7198 // These statistics are used to determine the desired number of free 7199 // chunks of a given size. The desired number is chosen to be relative 7200 // to the end of a CMS sweep. The desired number at the end of a sweep 7201 // is the 7202 // count-at-end-of-previous-sweep (an amount that was enough) 7203 // - count-at-beginning-of-current-sweep (the excess) 7204 // + split-births (gains in this size during interval) 7205 // - split-deaths (demands on this size during interval) 7206 // where the interval is from the end of one sweep to the end of the 7207 // next. 7208 // 7209 // When sweeping the sweeper maintains an accumulated chunk which is 7210 // the chunk that is made up of chunks that have been coalesced. That 7211 // will be termed the left-hand chunk. A new chunk of garbage that 7212 // is being considered for coalescing will be referred to as the 7213 // right-hand chunk. 7214 // 7215 // When making a decision on whether to coalesce a right-hand chunk with 7216 // the current left-hand chunk, the current count vs. the desired count 7217 // of the left-hand chunk is considered. Also if the right-hand chunk 7218 // is near the large chunk at the end of the heap (see 7219 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7220 // left-hand chunk is coalesced. 7221 // 7222 // When making a decision about whether to split a chunk, the desired count 7223 // vs. the current count of the candidate to be split is also considered. 7224 // If the candidate is underpopulated (currently fewer chunks than desired) 7225 // a chunk of an overpopulated (currently more chunks than desired) size may 7226 // be chosen. The "hint" associated with a free list, if non-null, points 7227 // to a free list which may be overpopulated. 7228 // 7229 7230 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7231 const size_t size = fc->size(); 7232 // Chunks that cannot be coalesced are not in the 7233 // free lists. 7234 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7235 assert(_sp->verify_chunk_in_free_list(fc), 7236 "free chunk should be in free lists"); 7237 } 7238 // a chunk that is already free, should not have been 7239 // marked in the bit map 7240 HeapWord* const addr = (HeapWord*) fc; 7241 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7242 // Verify that the bit map has no bits marked between 7243 // addr and purported end of this block. 7244 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7245 7246 // Some chunks cannot be coalesced under any circumstances. 7247 // See the definition of cantCoalesce(). 7248 if (!fc->cantCoalesce()) { 7249 // This chunk can potentially be coalesced. 7250 // All the work is done in 7251 do_post_free_or_garbage_chunk(fc, size); 7252 // Note that if the chunk is not coalescable (the else arm 7253 // below), we unconditionally flush, without needing to do 7254 // a "lookahead," as we do below. 7255 if (inFreeRange()) lookahead_and_flush(fc, size); 7256 } else { 7257 // Code path common to both original and adaptive free lists. 7258 7259 // cant coalesce with previous block; this should be treated 7260 // as the end of a free run if any 7261 if (inFreeRange()) { 7262 // we kicked some butt; time to pick up the garbage 7263 assert(freeFinger() < addr, "freeFinger points too high"); 7264 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7265 } 7266 // else, nothing to do, just continue 7267 } 7268 } 7269 7270 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7271 // This is a chunk of garbage. It is not in any free list. 7272 // Add it to a free list or let it possibly be coalesced into 7273 // a larger chunk. 7274 HeapWord* const addr = (HeapWord*) fc; 7275 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7276 7277 // Verify that the bit map has no bits marked between 7278 // addr and purported end of just dead object. 7279 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7280 do_post_free_or_garbage_chunk(fc, size); 7281 7282 assert(_limit >= addr + size, 7283 "A freshly garbage chunk can't possibly straddle over _limit"); 7284 if (inFreeRange()) lookahead_and_flush(fc, size); 7285 return size; 7286 } 7287 7288 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7289 HeapWord* addr = (HeapWord*) fc; 7290 // The sweeper has just found a live object. Return any accumulated 7291 // left hand chunk to the free lists. 7292 if (inFreeRange()) { 7293 assert(freeFinger() < addr, "freeFinger points too high"); 7294 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7295 } 7296 7297 // This object is live: we'd normally expect this to be 7298 // an oop, and like to assert the following: 7299 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7300 // However, as we commented above, this may be an object whose 7301 // header hasn't yet been initialized. 7302 size_t size; 7303 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7304 if (_bitMap->isMarked(addr + 1)) { 7305 // Determine the size from the bit map, rather than trying to 7306 // compute it from the object header. 7307 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7308 size = pointer_delta(nextOneAddr + 1, addr); 7309 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7310 "alignment problem"); 7311 7312 #ifdef ASSERT 7313 if (oop(addr)->klass_or_null_acquire() != NULL) { 7314 // Ignore mark word because we are running concurrent with mutators 7315 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7316 assert(size == 7317 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7318 "P-mark and computed size do not agree"); 7319 } 7320 #endif 7321 7322 } else { 7323 // This should be an initialized object that's alive. 7324 assert(oop(addr)->klass_or_null_acquire() != NULL, 7325 "Should be an initialized object"); 7326 // Ignore mark word because we are running concurrent with mutators 7327 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7328 // Verify that the bit map has no bits marked between 7329 // addr and purported end of this block. 7330 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7331 assert(size >= 3, "Necessary for Printezis marks to work"); 7332 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7333 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7334 } 7335 return size; 7336 } 7337 7338 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7339 size_t chunkSize) { 7340 // do_post_free_or_garbage_chunk() should only be called in the case 7341 // of the adaptive free list allocator. 7342 const bool fcInFreeLists = fc->is_free(); 7343 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7344 if (CMSTestInFreeList && fcInFreeLists) { 7345 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7346 } 7347 7348 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7349 7350 HeapWord* const fc_addr = (HeapWord*) fc; 7351 7352 bool coalesce = false; 7353 const size_t left = pointer_delta(fc_addr, freeFinger()); 7354 const size_t right = chunkSize; 7355 switch (FLSCoalescePolicy) { 7356 // numeric value forms a coalition aggressiveness metric 7357 case 0: { // never coalesce 7358 coalesce = false; 7359 break; 7360 } 7361 case 1: { // coalesce if left & right chunks on overpopulated lists 7362 coalesce = _sp->coalOverPopulated(left) && 7363 _sp->coalOverPopulated(right); 7364 break; 7365 } 7366 case 2: { // coalesce if left chunk on overpopulated list (default) 7367 coalesce = _sp->coalOverPopulated(left); 7368 break; 7369 } 7370 case 3: { // coalesce if left OR right chunk on overpopulated list 7371 coalesce = _sp->coalOverPopulated(left) || 7372 _sp->coalOverPopulated(right); 7373 break; 7374 } 7375 case 4: { // always coalesce 7376 coalesce = true; 7377 break; 7378 } 7379 default: 7380 ShouldNotReachHere(); 7381 } 7382 7383 // Should the current free range be coalesced? 7384 // If the chunk is in a free range and either we decided to coalesce above 7385 // or the chunk is near the large block at the end of the heap 7386 // (isNearLargestChunk() returns true), then coalesce this chunk. 7387 const bool doCoalesce = inFreeRange() 7388 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7389 if (doCoalesce) { 7390 // Coalesce the current free range on the left with the new 7391 // chunk on the right. If either is on a free list, 7392 // it must be removed from the list and stashed in the closure. 7393 if (freeRangeInFreeLists()) { 7394 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7395 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7396 "Size of free range is inconsistent with chunk size."); 7397 if (CMSTestInFreeList) { 7398 assert(_sp->verify_chunk_in_free_list(ffc), 7399 "Chunk is not in free lists"); 7400 } 7401 _sp->coalDeath(ffc->size()); 7402 _sp->removeFreeChunkFromFreeLists(ffc); 7403 set_freeRangeInFreeLists(false); 7404 } 7405 if (fcInFreeLists) { 7406 _sp->coalDeath(chunkSize); 7407 assert(fc->size() == chunkSize, 7408 "The chunk has the wrong size or is not in the free lists"); 7409 _sp->removeFreeChunkFromFreeLists(fc); 7410 } 7411 set_lastFreeRangeCoalesced(true); 7412 print_free_block_coalesced(fc); 7413 } else { // not in a free range and/or should not coalesce 7414 // Return the current free range and start a new one. 7415 if (inFreeRange()) { 7416 // In a free range but cannot coalesce with the right hand chunk. 7417 // Put the current free range into the free lists. 7418 flush_cur_free_chunk(freeFinger(), 7419 pointer_delta(fc_addr, freeFinger())); 7420 } 7421 // Set up for new free range. Pass along whether the right hand 7422 // chunk is in the free lists. 7423 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7424 } 7425 } 7426 7427 // Lookahead flush: 7428 // If we are tracking a free range, and this is the last chunk that 7429 // we'll look at because its end crosses past _limit, we'll preemptively 7430 // flush it along with any free range we may be holding on to. Note that 7431 // this can be the case only for an already free or freshly garbage 7432 // chunk. If this block is an object, it can never straddle 7433 // over _limit. The "straddling" occurs when _limit is set at 7434 // the previous end of the space when this cycle started, and 7435 // a subsequent heap expansion caused the previously co-terminal 7436 // free block to be coalesced with the newly expanded portion, 7437 // thus rendering _limit a non-block-boundary making it dangerous 7438 // for the sweeper to step over and examine. 7439 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7440 assert(inFreeRange(), "Should only be called if currently in a free range."); 7441 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7442 assert(_sp->used_region().contains(eob - 1), 7443 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7444 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7445 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7446 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7447 if (eob >= _limit) { 7448 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7449 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7450 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7451 "[" PTR_FORMAT "," PTR_FORMAT ")", 7452 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7453 // Return the storage we are tracking back into the free lists. 7454 log_develop_trace(gc, sweep)("Flushing ... "); 7455 assert(freeFinger() < eob, "Error"); 7456 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7457 } 7458 } 7459 7460 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7461 assert(inFreeRange(), "Should only be called if currently in a free range."); 7462 assert(size > 0, 7463 "A zero sized chunk cannot be added to the free lists."); 7464 if (!freeRangeInFreeLists()) { 7465 if (CMSTestInFreeList) { 7466 FreeChunk* fc = (FreeChunk*) chunk; 7467 fc->set_size(size); 7468 assert(!_sp->verify_chunk_in_free_list(fc), 7469 "chunk should not be in free lists yet"); 7470 } 7471 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7472 // A new free range is going to be starting. The current 7473 // free range has not been added to the free lists yet or 7474 // was removed so add it back. 7475 // If the current free range was coalesced, then the death 7476 // of the free range was recorded. Record a birth now. 7477 if (lastFreeRangeCoalesced()) { 7478 _sp->coalBirth(size); 7479 } 7480 _sp->addChunkAndRepairOffsetTable(chunk, size, 7481 lastFreeRangeCoalesced()); 7482 } else { 7483 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7484 } 7485 set_inFreeRange(false); 7486 set_freeRangeInFreeLists(false); 7487 } 7488 7489 // We take a break if we've been at this for a while, 7490 // so as to avoid monopolizing the locks involved. 7491 void SweepClosure::do_yield_work(HeapWord* addr) { 7492 // Return current free chunk being used for coalescing (if any) 7493 // to the appropriate freelist. After yielding, the next 7494 // free block encountered will start a coalescing range of 7495 // free blocks. If the next free block is adjacent to the 7496 // chunk just flushed, they will need to wait for the next 7497 // sweep to be coalesced. 7498 if (inFreeRange()) { 7499 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7500 } 7501 7502 // First give up the locks, then yield, then re-lock. 7503 // We should probably use a constructor/destructor idiom to 7504 // do this unlock/lock or modify the MutexUnlocker class to 7505 // serve our purpose. XXX 7506 assert_lock_strong(_bitMap->lock()); 7507 assert_lock_strong(_freelistLock); 7508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7509 "CMS thread should hold CMS token"); 7510 _bitMap->lock()->unlock(); 7511 _freelistLock->unlock(); 7512 ConcurrentMarkSweepThread::desynchronize(true); 7513 _collector->stopTimer(); 7514 _collector->incrementYields(); 7515 7516 // See the comment in coordinator_yield() 7517 for (unsigned i = 0; i < CMSYieldSleepCount && 7518 ConcurrentMarkSweepThread::should_yield() && 7519 !CMSCollector::foregroundGCIsActive(); ++i) { 7520 os::sleep(Thread::current(), 1, false); 7521 } 7522 7523 ConcurrentMarkSweepThread::synchronize(true); 7524 _freelistLock->lock(); 7525 _bitMap->lock()->lock_without_safepoint_check(); 7526 _collector->startTimer(); 7527 } 7528 7529 #ifndef PRODUCT 7530 // This is actually very useful in a product build if it can 7531 // be called from the debugger. Compile it into the product 7532 // as needed. 7533 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7534 return debug_cms_space->verify_chunk_in_free_list(fc); 7535 } 7536 #endif 7537 7538 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7539 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7540 p2i(fc), fc->size()); 7541 } 7542 7543 // CMSIsAliveClosure 7544 bool CMSIsAliveClosure::do_object_b(oop obj) { 7545 HeapWord* addr = (HeapWord*)obj; 7546 return addr != NULL && 7547 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7548 } 7549 7550 7551 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7552 MemRegion span, 7553 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7554 bool cpc): 7555 _collector(collector), 7556 _span(span), 7557 _bit_map(bit_map), 7558 _mark_stack(mark_stack), 7559 _concurrent_precleaning(cpc) { 7560 assert(!_span.is_empty(), "Empty span could spell trouble"); 7561 } 7562 7563 7564 // CMSKeepAliveClosure: the serial version 7565 void CMSKeepAliveClosure::do_oop(oop obj) { 7566 HeapWord* addr = (HeapWord*)obj; 7567 if (_span.contains(addr) && 7568 !_bit_map->isMarked(addr)) { 7569 _bit_map->mark(addr); 7570 bool simulate_overflow = false; 7571 NOT_PRODUCT( 7572 if (CMSMarkStackOverflowALot && 7573 _collector->simulate_overflow()) { 7574 // simulate a stack overflow 7575 simulate_overflow = true; 7576 } 7577 ) 7578 if (simulate_overflow || !_mark_stack->push(obj)) { 7579 if (_concurrent_precleaning) { 7580 // We dirty the overflown object and let the remark 7581 // phase deal with it. 7582 assert(_collector->overflow_list_is_empty(), "Error"); 7583 // In the case of object arrays, we need to dirty all of 7584 // the cards that the object spans. No locking or atomics 7585 // are needed since no one else can be mutating the mod union 7586 // table. 7587 if (obj->is_objArray()) { 7588 size_t sz = obj->size(); 7589 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7590 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7591 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7592 _collector->_modUnionTable.mark_range(redirty_range); 7593 } else { 7594 _collector->_modUnionTable.mark(addr); 7595 } 7596 _collector->_ser_kac_preclean_ovflw++; 7597 } else { 7598 _collector->push_on_overflow_list(obj); 7599 _collector->_ser_kac_ovflw++; 7600 } 7601 } 7602 } 7603 } 7604 7605 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7606 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7607 7608 // CMSParKeepAliveClosure: a parallel version of the above. 7609 // The work queues are private to each closure (thread), 7610 // but (may be) available for stealing by other threads. 7611 void CMSParKeepAliveClosure::do_oop(oop obj) { 7612 HeapWord* addr = (HeapWord*)obj; 7613 if (_span.contains(addr) && 7614 !_bit_map->isMarked(addr)) { 7615 // In general, during recursive tracing, several threads 7616 // may be concurrently getting here; the first one to 7617 // "tag" it, claims it. 7618 if (_bit_map->par_mark(addr)) { 7619 bool res = _work_queue->push(obj); 7620 assert(res, "Low water mark should be much less than capacity"); 7621 // Do a recursive trim in the hope that this will keep 7622 // stack usage lower, but leave some oops for potential stealers 7623 trim_queue(_low_water_mark); 7624 } // Else, another thread got there first 7625 } 7626 } 7627 7628 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7629 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7630 7631 void CMSParKeepAliveClosure::trim_queue(uint max) { 7632 while (_work_queue->size() > max) { 7633 oop new_oop; 7634 if (_work_queue->pop_local(new_oop)) { 7635 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7636 assert(_bit_map->isMarked((HeapWord*)new_oop), 7637 "no white objects on this stack!"); 7638 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7639 // iterate over the oops in this oop, marking and pushing 7640 // the ones in CMS heap (i.e. in _span). 7641 new_oop->oop_iterate(&_mark_and_push); 7642 } 7643 } 7644 } 7645 7646 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7647 CMSCollector* collector, 7648 MemRegion span, CMSBitMap* bit_map, 7649 OopTaskQueue* work_queue): 7650 _collector(collector), 7651 _span(span), 7652 _bit_map(bit_map), 7653 _work_queue(work_queue) { } 7654 7655 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7656 HeapWord* addr = (HeapWord*)obj; 7657 if (_span.contains(addr) && 7658 !_bit_map->isMarked(addr)) { 7659 if (_bit_map->par_mark(addr)) { 7660 bool simulate_overflow = false; 7661 NOT_PRODUCT( 7662 if (CMSMarkStackOverflowALot && 7663 _collector->par_simulate_overflow()) { 7664 // simulate a stack overflow 7665 simulate_overflow = true; 7666 } 7667 ) 7668 if (simulate_overflow || !_work_queue->push(obj)) { 7669 _collector->par_push_on_overflow_list(obj); 7670 _collector->_par_kac_ovflw++; 7671 } 7672 } // Else another thread got there already 7673 } 7674 } 7675 7676 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7677 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7678 7679 ////////////////////////////////////////////////////////////////// 7680 // CMSExpansionCause ///////////////////////////// 7681 ////////////////////////////////////////////////////////////////// 7682 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7683 switch (cause) { 7684 case _no_expansion: 7685 return "No expansion"; 7686 case _satisfy_free_ratio: 7687 return "Free ratio"; 7688 case _satisfy_promotion: 7689 return "Satisfy promotion"; 7690 case _satisfy_allocation: 7691 return "allocation"; 7692 case _allocate_par_lab: 7693 return "Par LAB"; 7694 case _allocate_par_spooling_space: 7695 return "Par Spooling Space"; 7696 case _adaptive_size_policy: 7697 return "Ergonomics"; 7698 default: 7699 return "unknown"; 7700 } 7701 } 7702 7703 void CMSDrainMarkingStackClosure::do_void() { 7704 // the max number to take from overflow list at a time 7705 const size_t num = _mark_stack->capacity()/4; 7706 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7707 "Overflow list should be NULL during concurrent phases"); 7708 while (!_mark_stack->isEmpty() || 7709 // if stack is empty, check the overflow list 7710 _collector->take_from_overflow_list(num, _mark_stack)) { 7711 oop obj = _mark_stack->pop(); 7712 HeapWord* addr = (HeapWord*)obj; 7713 assert(_span.contains(addr), "Should be within span"); 7714 assert(_bit_map->isMarked(addr), "Should be marked"); 7715 assert(oopDesc::is_oop(obj), "Should be an oop"); 7716 obj->oop_iterate(_keep_alive); 7717 } 7718 } 7719 7720 void CMSParDrainMarkingStackClosure::do_void() { 7721 // drain queue 7722 trim_queue(0); 7723 } 7724 7725 // Trim our work_queue so its length is below max at return 7726 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7727 while (_work_queue->size() > max) { 7728 oop new_oop; 7729 if (_work_queue->pop_local(new_oop)) { 7730 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7731 assert(_bit_map->isMarked((HeapWord*)new_oop), 7732 "no white objects on this stack!"); 7733 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7734 // iterate over the oops in this oop, marking and pushing 7735 // the ones in CMS heap (i.e. in _span). 7736 new_oop->oop_iterate(&_mark_and_push); 7737 } 7738 } 7739 } 7740 7741 //////////////////////////////////////////////////////////////////// 7742 // Support for Marking Stack Overflow list handling and related code 7743 //////////////////////////////////////////////////////////////////// 7744 // Much of the following code is similar in shape and spirit to the 7745 // code used in ParNewGC. We should try and share that code 7746 // as much as possible in the future. 7747 7748 #ifndef PRODUCT 7749 // Debugging support for CMSStackOverflowALot 7750 7751 // It's OK to call this multi-threaded; the worst thing 7752 // that can happen is that we'll get a bunch of closely 7753 // spaced simulated overflows, but that's OK, in fact 7754 // probably good as it would exercise the overflow code 7755 // under contention. 7756 bool CMSCollector::simulate_overflow() { 7757 if (_overflow_counter-- <= 0) { // just being defensive 7758 _overflow_counter = CMSMarkStackOverflowInterval; 7759 return true; 7760 } else { 7761 return false; 7762 } 7763 } 7764 7765 bool CMSCollector::par_simulate_overflow() { 7766 return simulate_overflow(); 7767 } 7768 #endif 7769 7770 // Single-threaded 7771 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7772 assert(stack->isEmpty(), "Expected precondition"); 7773 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7774 size_t i = num; 7775 oop cur = _overflow_list; 7776 const markOop proto = markOopDesc::prototype(); 7777 NOT_PRODUCT(ssize_t n = 0;) 7778 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7779 next = oop(cur->mark_raw()); 7780 cur->set_mark_raw(proto); // until proven otherwise 7781 assert(oopDesc::is_oop(cur), "Should be an oop"); 7782 bool res = stack->push(cur); 7783 assert(res, "Bit off more than can chew?"); 7784 NOT_PRODUCT(n++;) 7785 } 7786 _overflow_list = cur; 7787 #ifndef PRODUCT 7788 assert(_num_par_pushes >= n, "Too many pops?"); 7789 _num_par_pushes -=n; 7790 #endif 7791 return !stack->isEmpty(); 7792 } 7793 7794 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7795 // (MT-safe) Get a prefix of at most "num" from the list. 7796 // The overflow list is chained through the mark word of 7797 // each object in the list. We fetch the entire list, 7798 // break off a prefix of the right size and return the 7799 // remainder. If other threads try to take objects from 7800 // the overflow list at that time, they will wait for 7801 // some time to see if data becomes available. If (and 7802 // only if) another thread places one or more object(s) 7803 // on the global list before we have returned the suffix 7804 // to the global list, we will walk down our local list 7805 // to find its end and append the global list to 7806 // our suffix before returning it. This suffix walk can 7807 // prove to be expensive (quadratic in the amount of traffic) 7808 // when there are many objects in the overflow list and 7809 // there is much producer-consumer contention on the list. 7810 // *NOTE*: The overflow list manipulation code here and 7811 // in ParNewGeneration:: are very similar in shape, 7812 // except that in the ParNew case we use the old (from/eden) 7813 // copy of the object to thread the list via its klass word. 7814 // Because of the common code, if you make any changes in 7815 // the code below, please check the ParNew version to see if 7816 // similar changes might be needed. 7817 // CR 6797058 has been filed to consolidate the common code. 7818 bool CMSCollector::par_take_from_overflow_list(size_t num, 7819 OopTaskQueue* work_q, 7820 int no_of_gc_threads) { 7821 assert(work_q->size() == 0, "First empty local work queue"); 7822 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7823 if (_overflow_list == NULL) { 7824 return false; 7825 } 7826 // Grab the entire list; we'll put back a suffix 7827 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7828 Thread* tid = Thread::current(); 7829 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7830 // set to ParallelGCThreads. 7831 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7832 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7833 // If the list is busy, we spin for a short while, 7834 // sleeping between attempts to get the list. 7835 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7836 os::sleep(tid, sleep_time_millis, false); 7837 if (_overflow_list == NULL) { 7838 // Nothing left to take 7839 return false; 7840 } else if (_overflow_list != BUSY) { 7841 // Try and grab the prefix 7842 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7843 } 7844 } 7845 // If the list was found to be empty, or we spun long 7846 // enough, we give up and return empty-handed. If we leave 7847 // the list in the BUSY state below, it must be the case that 7848 // some other thread holds the overflow list and will set it 7849 // to a non-BUSY state in the future. 7850 if (prefix == NULL || prefix == BUSY) { 7851 // Nothing to take or waited long enough 7852 if (prefix == NULL) { 7853 // Write back the NULL in case we overwrote it with BUSY above 7854 // and it is still the same value. 7855 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7856 } 7857 return false; 7858 } 7859 assert(prefix != NULL && prefix != BUSY, "Error"); 7860 size_t i = num; 7861 oop cur = prefix; 7862 // Walk down the first "num" objects, unless we reach the end. 7863 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7864 if (cur->mark_raw() == NULL) { 7865 // We have "num" or fewer elements in the list, so there 7866 // is nothing to return to the global list. 7867 // Write back the NULL in lieu of the BUSY we wrote 7868 // above, if it is still the same value. 7869 if (_overflow_list == BUSY) { 7870 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7871 } 7872 } else { 7873 // Chop off the suffix and return it to the global list. 7874 assert(cur->mark_raw() != BUSY, "Error"); 7875 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7876 cur->set_mark_raw(NULL); // break off suffix 7877 // It's possible that the list is still in the empty(busy) state 7878 // we left it in a short while ago; in that case we may be 7879 // able to place back the suffix without incurring the cost 7880 // of a walk down the list. 7881 oop observed_overflow_list = _overflow_list; 7882 oop cur_overflow_list = observed_overflow_list; 7883 bool attached = false; 7884 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7885 observed_overflow_list = 7886 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7887 if (cur_overflow_list == observed_overflow_list) { 7888 attached = true; 7889 break; 7890 } else cur_overflow_list = observed_overflow_list; 7891 } 7892 if (!attached) { 7893 // Too bad, someone else sneaked in (at least) an element; we'll need 7894 // to do a splice. Find tail of suffix so we can prepend suffix to global 7895 // list. 7896 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7897 oop suffix_tail = cur; 7898 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7899 "Tautology"); 7900 observed_overflow_list = _overflow_list; 7901 do { 7902 cur_overflow_list = observed_overflow_list; 7903 if (cur_overflow_list != BUSY) { 7904 // Do the splice ... 7905 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7906 } else { // cur_overflow_list == BUSY 7907 suffix_tail->set_mark_raw(NULL); 7908 } 7909 // ... and try to place spliced list back on overflow_list ... 7910 observed_overflow_list = 7911 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7912 } while (cur_overflow_list != observed_overflow_list); 7913 // ... until we have succeeded in doing so. 7914 } 7915 } 7916 7917 // Push the prefix elements on work_q 7918 assert(prefix != NULL, "control point invariant"); 7919 const markOop proto = markOopDesc::prototype(); 7920 oop next; 7921 NOT_PRODUCT(ssize_t n = 0;) 7922 for (cur = prefix; cur != NULL; cur = next) { 7923 next = oop(cur->mark_raw()); 7924 cur->set_mark_raw(proto); // until proven otherwise 7925 assert(oopDesc::is_oop(cur), "Should be an oop"); 7926 bool res = work_q->push(cur); 7927 assert(res, "Bit off more than we can chew?"); 7928 NOT_PRODUCT(n++;) 7929 } 7930 #ifndef PRODUCT 7931 assert(_num_par_pushes >= n, "Too many pops?"); 7932 Atomic::sub(n, &_num_par_pushes); 7933 #endif 7934 return true; 7935 } 7936 7937 // Single-threaded 7938 void CMSCollector::push_on_overflow_list(oop p) { 7939 NOT_PRODUCT(_num_par_pushes++;) 7940 assert(oopDesc::is_oop(p), "Not an oop"); 7941 preserve_mark_if_necessary(p); 7942 p->set_mark_raw((markOop)_overflow_list); 7943 _overflow_list = p; 7944 } 7945 7946 // Multi-threaded; use CAS to prepend to overflow list 7947 void CMSCollector::par_push_on_overflow_list(oop p) { 7948 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7949 assert(oopDesc::is_oop(p), "Not an oop"); 7950 par_preserve_mark_if_necessary(p); 7951 oop observed_overflow_list = _overflow_list; 7952 oop cur_overflow_list; 7953 do { 7954 cur_overflow_list = observed_overflow_list; 7955 if (cur_overflow_list != BUSY) { 7956 p->set_mark_raw(markOop(cur_overflow_list)); 7957 } else { 7958 p->set_mark_raw(NULL); 7959 } 7960 observed_overflow_list = 7961 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7962 } while (cur_overflow_list != observed_overflow_list); 7963 } 7964 #undef BUSY 7965 7966 // Single threaded 7967 // General Note on GrowableArray: pushes may silently fail 7968 // because we are (temporarily) out of C-heap for expanding 7969 // the stack. The problem is quite ubiquitous and affects 7970 // a lot of code in the JVM. The prudent thing for GrowableArray 7971 // to do (for now) is to exit with an error. However, that may 7972 // be too draconian in some cases because the caller may be 7973 // able to recover without much harm. For such cases, we 7974 // should probably introduce a "soft_push" method which returns 7975 // an indication of success or failure with the assumption that 7976 // the caller may be able to recover from a failure; code in 7977 // the VM can then be changed, incrementally, to deal with such 7978 // failures where possible, thus, incrementally hardening the VM 7979 // in such low resource situations. 7980 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7981 _preserved_oop_stack.push(p); 7982 _preserved_mark_stack.push(m); 7983 assert(m == p->mark_raw(), "Mark word changed"); 7984 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7985 "bijection"); 7986 } 7987 7988 // Single threaded 7989 void CMSCollector::preserve_mark_if_necessary(oop p) { 7990 markOop m = p->mark_raw(); 7991 if (m->must_be_preserved(p)) { 7992 preserve_mark_work(p, m); 7993 } 7994 } 7995 7996 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 7997 markOop m = p->mark_raw(); 7998 if (m->must_be_preserved(p)) { 7999 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8000 // Even though we read the mark word without holding 8001 // the lock, we are assured that it will not change 8002 // because we "own" this oop, so no other thread can 8003 // be trying to push it on the overflow list; see 8004 // the assertion in preserve_mark_work() that checks 8005 // that m == p->mark_raw(). 8006 preserve_mark_work(p, m); 8007 } 8008 } 8009 8010 // We should be able to do this multi-threaded, 8011 // a chunk of stack being a task (this is 8012 // correct because each oop only ever appears 8013 // once in the overflow list. However, it's 8014 // not very easy to completely overlap this with 8015 // other operations, so will generally not be done 8016 // until all work's been completed. Because we 8017 // expect the preserved oop stack (set) to be small, 8018 // it's probably fine to do this single-threaded. 8019 // We can explore cleverer concurrent/overlapped/parallel 8020 // processing of preserved marks if we feel the 8021 // need for this in the future. Stack overflow should 8022 // be so rare in practice and, when it happens, its 8023 // effect on performance so great that this will 8024 // likely just be in the noise anyway. 8025 void CMSCollector::restore_preserved_marks_if_any() { 8026 assert(SafepointSynchronize::is_at_safepoint(), 8027 "world should be stopped"); 8028 assert(Thread::current()->is_ConcurrentGC_thread() || 8029 Thread::current()->is_VM_thread(), 8030 "should be single-threaded"); 8031 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8032 "bijection"); 8033 8034 while (!_preserved_oop_stack.is_empty()) { 8035 oop p = _preserved_oop_stack.pop(); 8036 assert(oopDesc::is_oop(p), "Should be an oop"); 8037 assert(_span.contains(p), "oop should be in _span"); 8038 assert(p->mark_raw() == markOopDesc::prototype(), 8039 "Set when taken from overflow list"); 8040 markOop m = _preserved_mark_stack.pop(); 8041 p->set_mark_raw(m); 8042 } 8043 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8044 "stacks were cleared above"); 8045 } 8046 8047 #ifndef PRODUCT 8048 bool CMSCollector::no_preserved_marks() const { 8049 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8050 } 8051 #endif 8052 8053 // Transfer some number of overflown objects to usual marking 8054 // stack. Return true if some objects were transferred. 8055 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8056 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8057 (size_t)ParGCDesiredObjsFromOverflowList); 8058 8059 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8060 assert(_collector->overflow_list_is_empty() || res, 8061 "If list is not empty, we should have taken something"); 8062 assert(!res || !_mark_stack->isEmpty(), 8063 "If we took something, it should now be on our stack"); 8064 return res; 8065 } 8066 8067 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8068 size_t res = _sp->block_size_no_stall(addr, _collector); 8069 if (_sp->block_is_obj(addr)) { 8070 if (_live_bit_map->isMarked(addr)) { 8071 // It can't have been dead in a previous cycle 8072 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8073 } else { 8074 _dead_bit_map->mark(addr); // mark the dead object 8075 } 8076 } 8077 // Could be 0, if the block size could not be computed without stalling. 8078 return res; 8079 } 8080 8081 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8082 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8083 switch (phase) { 8084 case CMSCollector::InitialMarking: 8085 initialize(manager /* GC manager */ , 8086 cause /* cause of the GC */, 8087 true /* recordGCBeginTime */, 8088 true /* recordPreGCUsage */, 8089 false /* recordPeakUsage */, 8090 false /* recordPostGCusage */, 8091 true /* recordAccumulatedGCTime */, 8092 false /* recordGCEndTime */, 8093 false /* countCollection */ ); 8094 break; 8095 8096 case CMSCollector::FinalMarking: 8097 initialize(manager /* GC manager */ , 8098 cause /* cause of the GC */, 8099 false /* recordGCBeginTime */, 8100 false /* recordPreGCUsage */, 8101 false /* recordPeakUsage */, 8102 false /* recordPostGCusage */, 8103 true /* recordAccumulatedGCTime */, 8104 false /* recordGCEndTime */, 8105 false /* countCollection */ ); 8106 break; 8107 8108 case CMSCollector::Sweeping: 8109 initialize(manager /* GC manager */ , 8110 cause /* cause of the GC */, 8111 false /* recordGCBeginTime */, 8112 false /* recordPreGCUsage */, 8113 true /* recordPeakUsage */, 8114 true /* recordPostGCusage */, 8115 false /* recordAccumulatedGCTime */, 8116 true /* recordGCEndTime */, 8117 true /* countCollection */ ); 8118 break; 8119 8120 default: 8121 ShouldNotReachHere(); 8122 } 8123 }