1 /* 2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderDataGraph.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/cms/cmsGCStats.hpp" 30 #include "gc/cms/cmsHeap.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/cmsVMOperations.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/promotionInfo.inline.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/gcLocker.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.inline.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/owstTaskTerminator.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 56 #include "gc/shared/space.inline.hpp" 57 #include "gc/shared/strongRootsScope.hpp" 58 #include "gc/shared/taskqueue.inline.hpp" 59 #include "gc/shared/weakProcessor.hpp" 60 #include "gc/shared/workerPolicy.hpp" 61 #include "logging/log.hpp" 62 #include "logging/logStream.hpp" 63 #include "memory/allocation.hpp" 64 #include "memory/binaryTreeDictionary.inline.hpp" 65 #include "memory/iterator.inline.hpp" 66 #include "memory/padded.hpp" 67 #include "memory/resourceArea.hpp" 68 #include "memory/universe.hpp" 69 #include "oops/access.inline.hpp" 70 #include "oops/oop.inline.hpp" 71 #include "prims/jvmtiExport.hpp" 72 #include "runtime/atomic.hpp" 73 #include "runtime/flags/flagSetting.hpp" 74 #include "runtime/globals_extension.hpp" 75 #include "runtime/handles.inline.hpp" 76 #include "runtime/java.hpp" 77 #include "runtime/orderAccess.hpp" 78 #include "runtime/timer.hpp" 79 #include "runtime/vmThread.hpp" 80 #include "services/memoryService.hpp" 81 #include "services/runtimeService.hpp" 82 #include "utilities/align.hpp" 83 #include "utilities/stack.inline.hpp" 84 #if INCLUDE_JVMCI 85 #include "jvmci/jvmci.hpp" 86 #endif 87 88 // statics 89 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 90 bool CMSCollector::_full_gc_requested = false; 91 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 92 93 ////////////////////////////////////////////////////////////////// 94 // In support of CMS/VM thread synchronization 95 ////////////////////////////////////////////////////////////////// 96 // We split use of the CGC_lock into 2 "levels". 97 // The low-level locking is of the usual CGC_lock monitor. We introduce 98 // a higher level "token" (hereafter "CMS token") built on top of the 99 // low level monitor (hereafter "CGC lock"). 100 // The token-passing protocol gives priority to the VM thread. The 101 // CMS-lock doesn't provide any fairness guarantees, but clients 102 // should ensure that it is only held for very short, bounded 103 // durations. 104 // 105 // When either of the CMS thread or the VM thread is involved in 106 // collection operations during which it does not want the other 107 // thread to interfere, it obtains the CMS token. 108 // 109 // If either thread tries to get the token while the other has 110 // it, that thread waits. However, if the VM thread and CMS thread 111 // both want the token, then the VM thread gets priority while the 112 // CMS thread waits. This ensures, for instance, that the "concurrent" 113 // phases of the CMS thread's work do not block out the VM thread 114 // for long periods of time as the CMS thread continues to hog 115 // the token. (See bug 4616232). 116 // 117 // The baton-passing functions are, however, controlled by the 118 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 119 // and here the low-level CMS lock, not the high level token, 120 // ensures mutual exclusion. 121 // 122 // Two important conditions that we have to satisfy: 123 // 1. if a thread does a low-level wait on the CMS lock, then it 124 // relinquishes the CMS token if it were holding that token 125 // when it acquired the low-level CMS lock. 126 // 2. any low-level notifications on the low-level lock 127 // should only be sent when a thread has relinquished the token. 128 // 129 // In the absence of either property, we'd have potential deadlock. 130 // 131 // We protect each of the CMS (concurrent and sequential) phases 132 // with the CMS _token_, not the CMS _lock_. 133 // 134 // The only code protected by CMS lock is the token acquisition code 135 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 136 // baton-passing code. 137 // 138 // Unfortunately, i couldn't come up with a good abstraction to factor and 139 // hide the naked CGC_lock manipulation in the baton-passing code 140 // further below. That's something we should try to do. Also, the proof 141 // of correctness of this 2-level locking scheme is far from obvious, 142 // and potentially quite slippery. We have an uneasy suspicion, for instance, 143 // that there may be a theoretical possibility of delay/starvation in the 144 // low-level lock/wait/notify scheme used for the baton-passing because of 145 // potential interference with the priority scheme embodied in the 146 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 147 // invocation further below and marked with "XXX 20011219YSR". 148 // Indeed, as we note elsewhere, this may become yet more slippery 149 // in the presence of multiple CMS and/or multiple VM threads. XXX 150 151 class CMSTokenSync: public StackObj { 152 private: 153 bool _is_cms_thread; 154 public: 155 CMSTokenSync(bool is_cms_thread): 156 _is_cms_thread(is_cms_thread) { 157 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 158 "Incorrect argument to constructor"); 159 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 160 } 161 162 ~CMSTokenSync() { 163 assert(_is_cms_thread ? 164 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 165 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 166 "Incorrect state"); 167 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 168 } 169 }; 170 171 // Convenience class that does a CMSTokenSync, and then acquires 172 // upto three locks. 173 class CMSTokenSyncWithLocks: public CMSTokenSync { 174 private: 175 // Note: locks are acquired in textual declaration order 176 // and released in the opposite order 177 MutexLocker _locker1, _locker2, _locker3; 178 public: 179 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 180 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 181 CMSTokenSync(is_cms_thread), 182 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 183 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 184 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 185 { } 186 }; 187 188 189 ////////////////////////////////////////////////////////////////// 190 // Concurrent Mark-Sweep Generation ///////////////////////////// 191 ////////////////////////////////////////////////////////////////// 192 193 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 194 195 // This struct contains per-thread things necessary to support parallel 196 // young-gen collection. 197 class CMSParGCThreadState: public CHeapObj<mtGC> { 198 public: 199 CompactibleFreeListSpaceLAB lab; 200 PromotionInfo promo; 201 202 // Constructor. 203 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 204 promo.setSpace(cfls); 205 } 206 }; 207 208 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 209 ReservedSpace rs, 210 size_t initial_byte_size, 211 size_t min_byte_size, 212 size_t max_byte_size, 213 CardTableRS* ct) : 214 CardGeneration(rs, initial_byte_size, ct), 215 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 216 _did_compact(false) 217 { 218 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 219 HeapWord* end = (HeapWord*) _virtual_space.high(); 220 221 _direct_allocated_words = 0; 222 NOT_PRODUCT( 223 _numObjectsPromoted = 0; 224 _numWordsPromoted = 0; 225 _numObjectsAllocated = 0; 226 _numWordsAllocated = 0; 227 ) 228 229 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 230 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 231 _cmsSpace->_old_gen = this; 232 233 _gc_stats = new CMSGCStats(); 234 235 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 236 // offsets match. The ability to tell free chunks from objects 237 // depends on this property. 238 debug_only( 239 FreeChunk* junk = NULL; 240 assert(UseCompressedClassPointers || 241 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 242 "Offset of FreeChunk::_prev within FreeChunk must match" 243 " that of OopDesc::_klass within OopDesc"); 244 ) 245 246 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 247 for (uint i = 0; i < ParallelGCThreads; i++) { 248 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 249 } 250 251 _incremental_collection_failed = false; 252 // The "dilatation_factor" is the expansion that can occur on 253 // account of the fact that the minimum object size in the CMS 254 // generation may be larger than that in, say, a contiguous young 255 // generation. 256 // Ideally, in the calculation below, we'd compute the dilatation 257 // factor as: MinChunkSize/(promoting_gen's min object size) 258 // Since we do not have such a general query interface for the 259 // promoting generation, we'll instead just use the minimum 260 // object size (which today is a header's worth of space); 261 // note that all arithmetic is in units of HeapWords. 262 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 263 assert(_dilatation_factor >= 1.0, "from previous assert"); 264 265 initialize_performance_counters(min_byte_size, max_byte_size); 266 } 267 268 269 // The field "_initiating_occupancy" represents the occupancy percentage 270 // at which we trigger a new collection cycle. Unless explicitly specified 271 // via CMSInitiatingOccupancyFraction (argument "io" below), it 272 // is calculated by: 273 // 274 // Let "f" be MinHeapFreeRatio in 275 // 276 // _initiating_occupancy = 100-f + 277 // f * (CMSTriggerRatio/100) 278 // where CMSTriggerRatio is the argument "tr" below. 279 // 280 // That is, if we assume the heap is at its desired maximum occupancy at the 281 // end of a collection, we let CMSTriggerRatio of the (purported) free 282 // space be allocated before initiating a new collection cycle. 283 // 284 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 285 assert(io <= 100 && tr <= 100, "Check the arguments"); 286 if (io >= 0) { 287 _initiating_occupancy = (double)io / 100.0; 288 } else { 289 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 290 (double)(tr * MinHeapFreeRatio) / 100.0) 291 / 100.0; 292 } 293 } 294 295 void ConcurrentMarkSweepGeneration::ref_processor_init() { 296 assert(collector() != NULL, "no collector"); 297 collector()->ref_processor_init(); 298 } 299 300 void CMSCollector::ref_processor_init() { 301 if (_ref_processor == NULL) { 302 // Allocate and initialize a reference processor 303 _ref_processor = 304 new ReferenceProcessor(&_span_based_discoverer, 305 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 306 ParallelGCThreads, // mt processing degree 307 _cmsGen->refs_discovery_is_mt(), // mt discovery 308 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 309 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 310 &_is_alive_closure, // closure for liveness info 311 false); // disable adjusting number of processing threads 312 // Initialize the _ref_processor field of CMSGen 313 _cmsGen->set_ref_processor(_ref_processor); 314 315 } 316 } 317 318 AdaptiveSizePolicy* CMSCollector::size_policy() { 319 return CMSHeap::heap()->size_policy(); 320 } 321 322 void ConcurrentMarkSweepGeneration::initialize_performance_counters(size_t min_old_size, 323 size_t max_old_size) { 324 325 const char* gen_name = "old"; 326 // Generation Counters - generation 1, 1 subspace 327 _gen_counters = new GenerationCounters(gen_name, 1, 1, 328 min_old_size, max_old_size, &_virtual_space); 329 330 _space_counters = new GSpaceCounters(gen_name, 0, 331 _virtual_space.reserved_size(), 332 this, _gen_counters); 333 } 334 335 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 336 _cms_gen(cms_gen) 337 { 338 assert(alpha <= 100, "bad value"); 339 _saved_alpha = alpha; 340 341 // Initialize the alphas to the bootstrap value of 100. 342 _gc0_alpha = _cms_alpha = 100; 343 344 _cms_begin_time.update(); 345 _cms_end_time.update(); 346 347 _gc0_duration = 0.0; 348 _gc0_period = 0.0; 349 _gc0_promoted = 0; 350 351 _cms_duration = 0.0; 352 _cms_period = 0.0; 353 _cms_allocated = 0; 354 355 _cms_used_at_gc0_begin = 0; 356 _cms_used_at_gc0_end = 0; 357 _allow_duty_cycle_reduction = false; 358 _valid_bits = 0; 359 } 360 361 double CMSStats::cms_free_adjustment_factor(size_t free) const { 362 // TBD: CR 6909490 363 return 1.0; 364 } 365 366 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 367 } 368 369 // If promotion failure handling is on use 370 // the padded average size of the promotion for each 371 // young generation collection. 372 double CMSStats::time_until_cms_gen_full() const { 373 size_t cms_free = _cms_gen->cmsSpace()->free(); 374 CMSHeap* heap = CMSHeap::heap(); 375 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 376 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 377 if (cms_free > expected_promotion) { 378 // Start a cms collection if there isn't enough space to promote 379 // for the next young collection. Use the padded average as 380 // a safety factor. 381 cms_free -= expected_promotion; 382 383 // Adjust by the safety factor. 384 double cms_free_dbl = (double)cms_free; 385 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 386 // Apply a further correction factor which tries to adjust 387 // for recent occurance of concurrent mode failures. 388 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 389 cms_free_dbl = cms_free_dbl * cms_adjustment; 390 391 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 392 cms_free, expected_promotion); 393 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 394 // Add 1 in case the consumption rate goes to zero. 395 return cms_free_dbl / (cms_consumption_rate() + 1.0); 396 } 397 return 0.0; 398 } 399 400 // Compare the duration of the cms collection to the 401 // time remaining before the cms generation is empty. 402 // Note that the time from the start of the cms collection 403 // to the start of the cms sweep (less than the total 404 // duration of the cms collection) can be used. This 405 // has been tried and some applications experienced 406 // promotion failures early in execution. This was 407 // possibly because the averages were not accurate 408 // enough at the beginning. 409 double CMSStats::time_until_cms_start() const { 410 // We add "gc0_period" to the "work" calculation 411 // below because this query is done (mostly) at the 412 // end of a scavenge, so we need to conservatively 413 // account for that much possible delay 414 // in the query so as to avoid concurrent mode failures 415 // due to starting the collection just a wee bit too 416 // late. 417 double work = cms_duration() + gc0_period(); 418 double deadline = time_until_cms_gen_full(); 419 // If a concurrent mode failure occurred recently, we want to be 420 // more conservative and halve our expected time_until_cms_gen_full() 421 if (work > deadline) { 422 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 423 cms_duration(), gc0_period(), time_until_cms_gen_full()); 424 return 0.0; 425 } 426 return work - deadline; 427 } 428 429 #ifndef PRODUCT 430 void CMSStats::print_on(outputStream *st) const { 431 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 432 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 433 gc0_duration(), gc0_period(), gc0_promoted()); 434 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 435 cms_duration(), cms_period(), cms_allocated()); 436 st->print(",cms_since_beg=%g,cms_since_end=%g", 437 cms_time_since_begin(), cms_time_since_end()); 438 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 439 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 440 441 if (valid()) { 442 st->print(",promo_rate=%g,cms_alloc_rate=%g", 443 promotion_rate(), cms_allocation_rate()); 444 st->print(",cms_consumption_rate=%g,time_until_full=%g", 445 cms_consumption_rate(), time_until_cms_gen_full()); 446 } 447 st->cr(); 448 } 449 #endif // #ifndef PRODUCT 450 451 CMSCollector::CollectorState CMSCollector::_collectorState = 452 CMSCollector::Idling; 453 bool CMSCollector::_foregroundGCIsActive = false; 454 bool CMSCollector::_foregroundGCShouldWait = false; 455 456 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 457 CardTableRS* ct): 458 _overflow_list(NULL), 459 _conc_workers(NULL), // may be set later 460 _completed_initialization(false), 461 _collection_count_start(0), 462 _should_unload_classes(CMSClassUnloadingEnabled), 463 _concurrent_cycles_since_last_unload(0), 464 _roots_scanning_options(GenCollectedHeap::SO_None), 465 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 466 _verifying(false), 467 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 468 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 469 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 470 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 471 _cms_start_registered(false), 472 _cmsGen(cmsGen), 473 // Adjust span to cover old (cms) gen 474 _span(cmsGen->reserved()), 475 _ct(ct), 476 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 477 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 478 -1 /* lock-free */, "No_lock" /* dummy */), 479 _restart_addr(NULL), 480 _ser_pmc_preclean_ovflw(0), 481 _ser_pmc_remark_ovflw(0), 482 _par_pmc_remark_ovflw(0), 483 _ser_kac_preclean_ovflw(0), 484 _ser_kac_ovflw(0), 485 _par_kac_ovflw(0), 486 #ifndef PRODUCT 487 _num_par_pushes(0), 488 #endif 489 _span_based_discoverer(_span), 490 _ref_processor(NULL), // will be set later 491 // Construct the is_alive_closure with _span & markBitMap 492 _is_alive_closure(_span, &_markBitMap), 493 _modUnionClosurePar(&_modUnionTable), 494 _between_prologue_and_epilogue(false), 495 _abort_preclean(false), 496 _start_sampling(false), 497 _stats(cmsGen), 498 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 499 //verify that this lock should be acquired with safepoint check. 500 Monitor::_safepoint_check_never)), 501 _eden_chunk_array(NULL), // may be set in ctor body 502 _eden_chunk_index(0), // -- ditto -- 503 _eden_chunk_capacity(0), // -- ditto -- 504 _survivor_chunk_array(NULL), // -- ditto -- 505 _survivor_chunk_index(0), // -- ditto -- 506 _survivor_chunk_capacity(0), // -- ditto -- 507 _survivor_plab_array(NULL) // -- ditto -- 508 { 509 // Now expand the span and allocate the collection support structures 510 // (MUT, marking bit map etc.) to cover both generations subject to 511 // collection. 512 513 // For use by dirty card to oop closures. 514 _cmsGen->cmsSpace()->set_collector(this); 515 516 // Allocate MUT and marking bit map 517 { 518 MutexLocker x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 519 if (!_markBitMap.allocate(_span)) { 520 log_warning(gc)("Failed to allocate CMS Bit Map"); 521 return; 522 } 523 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 524 } 525 { 526 _modUnionTable.allocate(_span); 527 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 528 } 529 530 if (!_markStack.allocate(MarkStackSize)) { 531 log_warning(gc)("Failed to allocate CMS Marking Stack"); 532 return; 533 } 534 535 // Support for multi-threaded concurrent phases 536 if (CMSConcurrentMTEnabled) { 537 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 538 // just for now 539 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 540 } 541 if (ConcGCThreads > 1) { 542 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 543 ConcGCThreads, true); 544 if (_conc_workers == NULL) { 545 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 546 CMSConcurrentMTEnabled = false; 547 } else { 548 _conc_workers->initialize_workers(); 549 } 550 } else { 551 CMSConcurrentMTEnabled = false; 552 } 553 } 554 if (!CMSConcurrentMTEnabled) { 555 ConcGCThreads = 0; 556 } else { 557 // Turn off CMSCleanOnEnter optimization temporarily for 558 // the MT case where it's not fixed yet; see 6178663. 559 CMSCleanOnEnter = false; 560 } 561 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 562 "Inconsistency"); 563 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 564 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 565 566 // Parallel task queues; these are shared for the 567 // concurrent and stop-world phases of CMS, but 568 // are not shared with parallel scavenge (ParNew). 569 { 570 uint i; 571 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 572 573 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 574 || ParallelRefProcEnabled) 575 && num_queues > 0) { 576 _task_queues = new OopTaskQueueSet(num_queues); 577 if (_task_queues == NULL) { 578 log_warning(gc)("task_queues allocation failure."); 579 return; 580 } 581 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 582 for (i = 0; i < num_queues; i++) { 583 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 584 if (q == NULL) { 585 log_warning(gc)("work_queue allocation failure."); 586 return; 587 } 588 _task_queues->register_queue(i, q); 589 } 590 for (i = 0; i < num_queues; i++) { 591 _task_queues->queue(i)->initialize(); 592 } 593 } 594 } 595 596 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 597 598 // Clip CMSBootstrapOccupancy between 0 and 100. 599 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 600 601 // Now tell CMS generations the identity of their collector 602 ConcurrentMarkSweepGeneration::set_collector(this); 603 604 // Create & start a CMS thread for this CMS collector 605 _cmsThread = ConcurrentMarkSweepThread::start(this); 606 assert(cmsThread() != NULL, "CMS Thread should have been created"); 607 assert(cmsThread()->collector() == this, 608 "CMS Thread should refer to this gen"); 609 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 610 611 // Support for parallelizing young gen rescan 612 CMSHeap* heap = CMSHeap::heap(); 613 _young_gen = heap->young_gen(); 614 if (heap->supports_inline_contig_alloc()) { 615 _top_addr = heap->top_addr(); 616 _end_addr = heap->end_addr(); 617 assert(_young_gen != NULL, "no _young_gen"); 618 _eden_chunk_index = 0; 619 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 620 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 621 } 622 623 // Support for parallelizing survivor space rescan 624 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 625 const size_t max_plab_samples = 626 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 627 628 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 629 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 631 _survivor_chunk_capacity = max_plab_samples; 632 for (uint i = 0; i < ParallelGCThreads; i++) { 633 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 634 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 635 assert(cur->end() == 0, "Should be 0"); 636 assert(cur->array() == vec, "Should be vec"); 637 assert(cur->capacity() == max_plab_samples, "Error"); 638 } 639 } 640 641 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 642 _gc_counters = new CollectorCounters("CMS full collection pauses", 1); 643 _cgc_counters = new CollectorCounters("CMS concurrent cycle pauses", 2); 644 _completed_initialization = true; 645 _inter_sweep_timer.start(); // start of time 646 } 647 648 const char* ConcurrentMarkSweepGeneration::name() const { 649 return "concurrent mark-sweep generation"; 650 } 651 void ConcurrentMarkSweepGeneration::update_counters() { 652 if (UsePerfData) { 653 _space_counters->update_all(); 654 _gen_counters->update_all(); 655 } 656 } 657 658 // this is an optimized version of update_counters(). it takes the 659 // used value as a parameter rather than computing it. 660 // 661 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 662 if (UsePerfData) { 663 _space_counters->update_used(used); 664 _space_counters->update_capacity(); 665 _gen_counters->update_all(); 666 } 667 } 668 669 void ConcurrentMarkSweepGeneration::print() const { 670 Generation::print(); 671 cmsSpace()->print(); 672 } 673 674 #ifndef PRODUCT 675 void ConcurrentMarkSweepGeneration::print_statistics() { 676 cmsSpace()->printFLCensus(0); 677 } 678 #endif 679 680 size_t 681 ConcurrentMarkSweepGeneration::contiguous_available() const { 682 // dld proposes an improvement in precision here. If the committed 683 // part of the space ends in a free block we should add that to 684 // uncommitted size in the calculation below. Will make this 685 // change later, staying with the approximation below for the 686 // time being. -- ysr. 687 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 688 } 689 690 size_t 691 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 692 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 693 } 694 695 size_t ConcurrentMarkSweepGeneration::used_stable() const { 696 return cmsSpace()->used_stable(); 697 } 698 699 size_t ConcurrentMarkSweepGeneration::max_available() const { 700 return free() + _virtual_space.uncommitted_size(); 701 } 702 703 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 704 size_t available = max_available(); 705 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 706 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 707 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 708 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 709 return res; 710 } 711 712 // At a promotion failure dump information on block layout in heap 713 // (cms old generation). 714 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 715 Log(gc, promotion) log; 716 if (log.is_trace()) { 717 LogStream ls(log.trace()); 718 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 719 } 720 } 721 722 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 723 // Clear the promotion information. These pointers can be adjusted 724 // along with all the other pointers into the heap but 725 // compaction is expected to be a rare event with 726 // a heap using cms so don't do it without seeing the need. 727 for (uint i = 0; i < ParallelGCThreads; i++) { 728 _par_gc_thread_states[i]->promo.reset(); 729 } 730 } 731 732 void ConcurrentMarkSweepGeneration::compute_new_size() { 733 assert_locked_or_safepoint(Heap_lock); 734 735 // If incremental collection failed, we just want to expand 736 // to the limit. 737 if (incremental_collection_failed()) { 738 clear_incremental_collection_failed(); 739 grow_to_reserved(); 740 return; 741 } 742 743 // The heap has been compacted but not reset yet. 744 // Any metric such as free() or used() will be incorrect. 745 746 CardGeneration::compute_new_size(); 747 748 // Reset again after a possible resizing 749 if (did_compact()) { 750 cmsSpace()->reset_after_compaction(); 751 } 752 } 753 754 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 755 assert_locked_or_safepoint(Heap_lock); 756 757 // If incremental collection failed, we just want to expand 758 // to the limit. 759 if (incremental_collection_failed()) { 760 clear_incremental_collection_failed(); 761 grow_to_reserved(); 762 return; 763 } 764 765 double free_percentage = ((double) free()) / capacity(); 766 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 767 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 768 769 // compute expansion delta needed for reaching desired free percentage 770 if (free_percentage < desired_free_percentage) { 771 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 772 assert(desired_capacity >= capacity(), "invalid expansion size"); 773 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 774 Log(gc) log; 775 if (log.is_trace()) { 776 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 777 log.trace("From compute_new_size: "); 778 log.trace(" Free fraction %f", free_percentage); 779 log.trace(" Desired free fraction %f", desired_free_percentage); 780 log.trace(" Maximum free fraction %f", maximum_free_percentage); 781 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 782 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 783 CMSHeap* heap = CMSHeap::heap(); 784 size_t young_size = heap->young_gen()->capacity(); 785 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 786 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 787 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 788 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 789 } 790 // safe if expansion fails 791 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 792 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 793 } else { 794 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 795 assert(desired_capacity <= capacity(), "invalid expansion size"); 796 size_t shrink_bytes = capacity() - desired_capacity; 797 // Don't shrink unless the delta is greater than the minimum shrink we want 798 if (shrink_bytes >= MinHeapDeltaBytes) { 799 shrink_free_list_by(shrink_bytes); 800 } 801 } 802 } 803 804 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 805 return cmsSpace()->freelistLock(); 806 } 807 808 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 809 CMSSynchronousYieldRequest yr; 810 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 811 return have_lock_and_allocate(size, tlab); 812 } 813 814 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 815 bool tlab /* ignored */) { 816 assert_lock_strong(freelistLock()); 817 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 818 HeapWord* res = cmsSpace()->allocate(adjustedSize); 819 // Allocate the object live (grey) if the background collector has 820 // started marking. This is necessary because the marker may 821 // have passed this address and consequently this object will 822 // not otherwise be greyed and would be incorrectly swept up. 823 // Note that if this object contains references, the writing 824 // of those references will dirty the card containing this object 825 // allowing the object to be blackened (and its references scanned) 826 // either during a preclean phase or at the final checkpoint. 827 if (res != NULL) { 828 // We may block here with an uninitialized object with 829 // its mark-bit or P-bits not yet set. Such objects need 830 // to be safely navigable by block_start(). 831 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 832 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 833 collector()->direct_allocated(res, adjustedSize); 834 _direct_allocated_words += adjustedSize; 835 // allocation counters 836 NOT_PRODUCT( 837 _numObjectsAllocated++; 838 _numWordsAllocated += (int)adjustedSize; 839 ) 840 } 841 return res; 842 } 843 844 // In the case of direct allocation by mutators in a generation that 845 // is being concurrently collected, the object must be allocated 846 // live (grey) if the background collector has started marking. 847 // This is necessary because the marker may 848 // have passed this address and consequently this object will 849 // not otherwise be greyed and would be incorrectly swept up. 850 // Note that if this object contains references, the writing 851 // of those references will dirty the card containing this object 852 // allowing the object to be blackened (and its references scanned) 853 // either during a preclean phase or at the final checkpoint. 854 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 855 assert(_markBitMap.covers(start, size), "Out of bounds"); 856 if (_collectorState >= Marking) { 857 MutexLocker y(_markBitMap.lock(), 858 Mutex::_no_safepoint_check_flag); 859 // [see comments preceding SweepClosure::do_blk() below for details] 860 // 861 // Can the P-bits be deleted now? JJJ 862 // 863 // 1. need to mark the object as live so it isn't collected 864 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 865 // 3. need to mark the end of the object so marking, precleaning or sweeping 866 // can skip over uninitialized or unparsable objects. An allocated 867 // object is considered uninitialized for our purposes as long as 868 // its klass word is NULL. All old gen objects are parsable 869 // as soon as they are initialized.) 870 _markBitMap.mark(start); // object is live 871 _markBitMap.mark(start + 1); // object is potentially uninitialized? 872 _markBitMap.mark(start + size - 1); 873 // mark end of object 874 } 875 // check that oop looks uninitialized 876 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 877 } 878 879 void CMSCollector::promoted(bool par, HeapWord* start, 880 bool is_obj_array, size_t obj_size) { 881 assert(_markBitMap.covers(start), "Out of bounds"); 882 // See comment in direct_allocated() about when objects should 883 // be allocated live. 884 if (_collectorState >= Marking) { 885 // we already hold the marking bit map lock, taken in 886 // the prologue 887 if (par) { 888 _markBitMap.par_mark(start); 889 } else { 890 _markBitMap.mark(start); 891 } 892 // We don't need to mark the object as uninitialized (as 893 // in direct_allocated above) because this is being done with the 894 // world stopped and the object will be initialized by the 895 // time the marking, precleaning or sweeping get to look at it. 896 // But see the code for copying objects into the CMS generation, 897 // where we need to ensure that concurrent readers of the 898 // block offset table are able to safely navigate a block that 899 // is in flux from being free to being allocated (and in 900 // transition while being copied into) and subsequently 901 // becoming a bona-fide object when the copy/promotion is complete. 902 assert(SafepointSynchronize::is_at_safepoint(), 903 "expect promotion only at safepoints"); 904 905 if (_collectorState < Sweeping) { 906 // Mark the appropriate cards in the modUnionTable, so that 907 // this object gets scanned before the sweep. If this is 908 // not done, CMS generation references in the object might 909 // not get marked. 910 // For the case of arrays, which are otherwise precisely 911 // marked, we need to dirty the entire array, not just its head. 912 if (is_obj_array) { 913 // The [par_]mark_range() method expects mr.end() below to 914 // be aligned to the granularity of a bit's representation 915 // in the heap. In the case of the MUT below, that's a 916 // card size. 917 MemRegion mr(start, 918 align_up(start + obj_size, 919 CardTable::card_size /* bytes */)); 920 if (par) { 921 _modUnionTable.par_mark_range(mr); 922 } else { 923 _modUnionTable.mark_range(mr); 924 } 925 } else { // not an obj array; we can just mark the head 926 if (par) { 927 _modUnionTable.par_mark(start); 928 } else { 929 _modUnionTable.mark(start); 930 } 931 } 932 } 933 } 934 } 935 936 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 937 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 938 // allocate, copy and if necessary update promoinfo -- 939 // delegate to underlying space. 940 assert_lock_strong(freelistLock()); 941 942 #ifndef PRODUCT 943 if (CMSHeap::heap()->promotion_should_fail()) { 944 return NULL; 945 } 946 #endif // #ifndef PRODUCT 947 948 oop res = _cmsSpace->promote(obj, obj_size); 949 if (res == NULL) { 950 // expand and retry 951 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 952 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 953 // Since this is the old generation, we don't try to promote 954 // into a more senior generation. 955 res = _cmsSpace->promote(obj, obj_size); 956 } 957 if (res != NULL) { 958 // See comment in allocate() about when objects should 959 // be allocated live. 960 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 961 collector()->promoted(false, // Not parallel 962 (HeapWord*)res, obj->is_objArray(), obj_size); 963 // promotion counters 964 NOT_PRODUCT( 965 _numObjectsPromoted++; 966 _numWordsPromoted += 967 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 968 ) 969 } 970 return res; 971 } 972 973 974 // IMPORTANT: Notes on object size recognition in CMS. 975 // --------------------------------------------------- 976 // A block of storage in the CMS generation is always in 977 // one of three states. A free block (FREE), an allocated 978 // object (OBJECT) whose size() method reports the correct size, 979 // and an intermediate state (TRANSIENT) in which its size cannot 980 // be accurately determined. 981 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 982 // ----------------------------------------------------- 983 // FREE: klass_word & 1 == 1; mark_word holds block size 984 // 985 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 986 // obj->size() computes correct size 987 // 988 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 989 // 990 // STATE IDENTIFICATION: (64 bit+COOPS) 991 // ------------------------------------ 992 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 993 // 994 // OBJECT: klass_word installed; klass_word != 0; 995 // obj->size() computes correct size 996 // 997 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 998 // 999 // 1000 // STATE TRANSITION DIAGRAM 1001 // 1002 // mut / parnew mut / parnew 1003 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1004 // ^ | 1005 // |------------------------ DEAD <------------------------------------| 1006 // sweep mut 1007 // 1008 // While a block is in TRANSIENT state its size cannot be determined 1009 // so readers will either need to come back later or stall until 1010 // the size can be determined. Note that for the case of direct 1011 // allocation, P-bits, when available, may be used to determine the 1012 // size of an object that may not yet have been initialized. 1013 1014 // Things to support parallel young-gen collection. 1015 oop 1016 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1017 oop old, markWord m, 1018 size_t word_sz) { 1019 #ifndef PRODUCT 1020 if (CMSHeap::heap()->promotion_should_fail()) { 1021 return NULL; 1022 } 1023 #endif // #ifndef PRODUCT 1024 1025 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1026 PromotionInfo* promoInfo = &ps->promo; 1027 // if we are tracking promotions, then first ensure space for 1028 // promotion (including spooling space for saving header if necessary). 1029 // then allocate and copy, then track promoted info if needed. 1030 // When tracking (see PromotionInfo::track()), the mark word may 1031 // be displaced and in this case restoration of the mark word 1032 // occurs in the (oop_since_save_marks_)iterate phase. 1033 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1034 // Out of space for allocating spooling buffers; 1035 // try expanding and allocating spooling buffers. 1036 if (!expand_and_ensure_spooling_space(promoInfo)) { 1037 return NULL; 1038 } 1039 } 1040 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1041 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1042 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1043 if (obj_ptr == NULL) { 1044 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1045 if (obj_ptr == NULL) { 1046 return NULL; 1047 } 1048 } 1049 oop obj = oop(obj_ptr); 1050 OrderAccess::storestore(); 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 // IMPORTANT: See note on object initialization for CMS above. 1054 // Otherwise, copy the object. Here we must be careful to insert the 1055 // klass pointer last, since this marks the block as an allocated object. 1056 // Except with compressed oops it's the mark word. 1057 HeapWord* old_ptr = (HeapWord*)old; 1058 // Restore the mark word copied above. 1059 obj->set_mark_raw(m); 1060 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1061 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1062 OrderAccess::storestore(); 1063 1064 if (UseCompressedClassPointers) { 1065 // Copy gap missed by (aligned) header size calculation below 1066 obj->set_klass_gap(old->klass_gap()); 1067 } 1068 if (word_sz > (size_t)oopDesc::header_size()) { 1069 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1070 obj_ptr + oopDesc::header_size(), 1071 word_sz - oopDesc::header_size()); 1072 } 1073 1074 // Now we can track the promoted object, if necessary. We take care 1075 // to delay the transition from uninitialized to full object 1076 // (i.e., insertion of klass pointer) until after, so that it 1077 // atomically becomes a promoted object. 1078 if (promoInfo->tracking()) { 1079 promoInfo->track((PromotedObject*)obj, old->klass()); 1080 } 1081 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1082 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1083 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1084 1085 // Finally, install the klass pointer (this should be volatile). 1086 OrderAccess::storestore(); 1087 obj->set_klass(old->klass()); 1088 // We should now be able to calculate the right size for this object 1089 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1090 1091 collector()->promoted(true, // parallel 1092 obj_ptr, old->is_objArray(), word_sz); 1093 1094 NOT_PRODUCT( 1095 Atomic::inc(&_numObjectsPromoted); 1096 Atomic::add(alloc_sz, &_numWordsPromoted); 1097 ) 1098 1099 return obj; 1100 } 1101 1102 void 1103 ConcurrentMarkSweepGeneration:: 1104 par_promote_alloc_done(int thread_num) { 1105 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1106 ps->lab.retire(thread_num); 1107 } 1108 1109 void 1110 ConcurrentMarkSweepGeneration:: 1111 par_oop_since_save_marks_iterate_done(int thread_num) { 1112 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1113 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1114 ps->promo.promoted_oops_iterate(dummy_cl); 1115 1116 // Because card-scanning has been completed, subsequent phases 1117 // (e.g., reference processing) will not need to recognize which 1118 // objects have been promoted during this GC. So, we can now disable 1119 // promotion tracking. 1120 ps->promo.stopTrackingPromotions(); 1121 } 1122 1123 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1124 size_t size, 1125 bool tlab) 1126 { 1127 // We allow a STW collection only if a full 1128 // collection was requested. 1129 return full || should_allocate(size, tlab); // FIX ME !!! 1130 // This and promotion failure handling are connected at the 1131 // hip and should be fixed by untying them. 1132 } 1133 1134 bool CMSCollector::shouldConcurrentCollect() { 1135 LogTarget(Trace, gc) log; 1136 1137 if (_full_gc_requested) { 1138 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1139 return true; 1140 } 1141 1142 FreelistLocker x(this); 1143 // ------------------------------------------------------------------ 1144 // Print out lots of information which affects the initiation of 1145 // a collection. 1146 if (log.is_enabled() && stats().valid()) { 1147 log.print("CMSCollector shouldConcurrentCollect: "); 1148 1149 LogStream out(log); 1150 stats().print_on(&out); 1151 1152 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1153 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1154 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1155 log.print("promotion_rate=%g", stats().promotion_rate()); 1156 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1157 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1158 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1159 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1160 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1161 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1162 } 1163 // ------------------------------------------------------------------ 1164 1165 // If the estimated time to complete a cms collection (cms_duration()) 1166 // is less than the estimated time remaining until the cms generation 1167 // is full, start a collection. 1168 if (!UseCMSInitiatingOccupancyOnly) { 1169 if (stats().valid()) { 1170 if (stats().time_until_cms_start() == 0.0) { 1171 return true; 1172 } 1173 } else { 1174 // We want to conservatively collect somewhat early in order 1175 // to try and "bootstrap" our CMS/promotion statistics; 1176 // this branch will not fire after the first successful CMS 1177 // collection because the stats should then be valid. 1178 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1179 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1180 _cmsGen->occupancy(), _bootstrap_occupancy); 1181 return true; 1182 } 1183 } 1184 } 1185 1186 // Otherwise, we start a collection cycle if 1187 // old gen want a collection cycle started. Each may use 1188 // an appropriate criterion for making this decision. 1189 // XXX We need to make sure that the gen expansion 1190 // criterion dovetails well with this. XXX NEED TO FIX THIS 1191 if (_cmsGen->should_concurrent_collect()) { 1192 log.print("CMS old gen initiated"); 1193 return true; 1194 } 1195 1196 // We start a collection if we believe an incremental collection may fail; 1197 // this is not likely to be productive in practice because it's probably too 1198 // late anyway. 1199 CMSHeap* heap = CMSHeap::heap(); 1200 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1201 log.print("CMSCollector: collect because incremental collection will fail "); 1202 return true; 1203 } 1204 1205 if (MetaspaceGC::should_concurrent_collect()) { 1206 log.print("CMSCollector: collect for metadata allocation "); 1207 return true; 1208 } 1209 1210 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1211 if (CMSTriggerInterval >= 0) { 1212 if (CMSTriggerInterval == 0) { 1213 // Trigger always 1214 return true; 1215 } 1216 1217 // Check the CMS time since begin (we do not check the stats validity 1218 // as we want to be able to trigger the first CMS cycle as well) 1219 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1220 if (stats().valid()) { 1221 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1222 stats().cms_time_since_begin()); 1223 } else { 1224 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1225 } 1226 return true; 1227 } 1228 } 1229 1230 return false; 1231 } 1232 1233 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1234 1235 // Clear _expansion_cause fields of constituent generations 1236 void CMSCollector::clear_expansion_cause() { 1237 _cmsGen->clear_expansion_cause(); 1238 } 1239 1240 // We should be conservative in starting a collection cycle. To 1241 // start too eagerly runs the risk of collecting too often in the 1242 // extreme. To collect too rarely falls back on full collections, 1243 // which works, even if not optimum in terms of concurrent work. 1244 // As a work around for too eagerly collecting, use the flag 1245 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1246 // giving the user an easily understandable way of controlling the 1247 // collections. 1248 // We want to start a new collection cycle if any of the following 1249 // conditions hold: 1250 // . our current occupancy exceeds the configured initiating occupancy 1251 // for this generation, or 1252 // . we recently needed to expand this space and have not, since that 1253 // expansion, done a collection of this generation, or 1254 // . the underlying space believes that it may be a good idea to initiate 1255 // a concurrent collection (this may be based on criteria such as the 1256 // following: the space uses linear allocation and linear allocation is 1257 // going to fail, or there is believed to be excessive fragmentation in 1258 // the generation, etc... or ... 1259 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1260 // the case of the old generation; see CR 6543076): 1261 // we may be approaching a point at which allocation requests may fail because 1262 // we will be out of sufficient free space given allocation rate estimates.] 1263 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1264 1265 assert_lock_strong(freelistLock()); 1266 if (occupancy() > initiating_occupancy()) { 1267 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1268 short_name(), occupancy(), initiating_occupancy()); 1269 return true; 1270 } 1271 if (UseCMSInitiatingOccupancyOnly) { 1272 return false; 1273 } 1274 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1275 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1276 return true; 1277 } 1278 return false; 1279 } 1280 1281 void ConcurrentMarkSweepGeneration::collect(bool full, 1282 bool clear_all_soft_refs, 1283 size_t size, 1284 bool tlab) 1285 { 1286 collector()->collect(full, clear_all_soft_refs, size, tlab); 1287 } 1288 1289 void CMSCollector::collect(bool full, 1290 bool clear_all_soft_refs, 1291 size_t size, 1292 bool tlab) 1293 { 1294 // The following "if" branch is present for defensive reasons. 1295 // In the current uses of this interface, it can be replaced with: 1296 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1297 // But I am not placing that assert here to allow future 1298 // generality in invoking this interface. 1299 if (GCLocker::is_active()) { 1300 // A consistency test for GCLocker 1301 assert(GCLocker::needs_gc(), "Should have been set already"); 1302 // Skip this foreground collection, instead 1303 // expanding the heap if necessary. 1304 // Need the free list locks for the call to free() in compute_new_size() 1305 compute_new_size(); 1306 return; 1307 } 1308 acquire_control_and_collect(full, clear_all_soft_refs); 1309 } 1310 1311 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1312 CMSHeap* heap = CMSHeap::heap(); 1313 unsigned int gc_count = heap->total_full_collections(); 1314 if (gc_count == full_gc_count) { 1315 MutexLocker y(CGC_lock, Mutex::_no_safepoint_check_flag); 1316 _full_gc_requested = true; 1317 _full_gc_cause = cause; 1318 CGC_lock->notify(); // nudge CMS thread 1319 } else { 1320 assert(gc_count > full_gc_count, "Error: causal loop"); 1321 } 1322 } 1323 1324 bool CMSCollector::is_external_interruption() { 1325 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1326 return GCCause::is_user_requested_gc(cause) || 1327 GCCause::is_serviceability_requested_gc(cause); 1328 } 1329 1330 void CMSCollector::report_concurrent_mode_interruption() { 1331 if (is_external_interruption()) { 1332 log_debug(gc)("Concurrent mode interrupted"); 1333 } else { 1334 log_debug(gc)("Concurrent mode failure"); 1335 _gc_tracer_cm->report_concurrent_mode_failure(); 1336 } 1337 } 1338 1339 1340 // The foreground and background collectors need to coordinate in order 1341 // to make sure that they do not mutually interfere with CMS collections. 1342 // When a background collection is active, 1343 // the foreground collector may need to take over (preempt) and 1344 // synchronously complete an ongoing collection. Depending on the 1345 // frequency of the background collections and the heap usage 1346 // of the application, this preemption can be seldom or frequent. 1347 // There are only certain 1348 // points in the background collection that the "collection-baton" 1349 // can be passed to the foreground collector. 1350 // 1351 // The foreground collector will wait for the baton before 1352 // starting any part of the collection. The foreground collector 1353 // will only wait at one location. 1354 // 1355 // The background collector will yield the baton before starting a new 1356 // phase of the collection (e.g., before initial marking, marking from roots, 1357 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1358 // of the loop which switches the phases. The background collector does some 1359 // of the phases (initial mark, final re-mark) with the world stopped. 1360 // Because of locking involved in stopping the world, 1361 // the foreground collector should not block waiting for the background 1362 // collector when it is doing a stop-the-world phase. The background 1363 // collector will yield the baton at an additional point just before 1364 // it enters a stop-the-world phase. Once the world is stopped, the 1365 // background collector checks the phase of the collection. If the 1366 // phase has not changed, it proceeds with the collection. If the 1367 // phase has changed, it skips that phase of the collection. See 1368 // the comments on the use of the Heap_lock in collect_in_background(). 1369 // 1370 // Variable used in baton passing. 1371 // _foregroundGCIsActive - Set to true by the foreground collector when 1372 // it wants the baton. The foreground clears it when it has finished 1373 // the collection. 1374 // _foregroundGCShouldWait - Set to true by the background collector 1375 // when it is running. The foreground collector waits while 1376 // _foregroundGCShouldWait is true. 1377 // CGC_lock - monitor used to protect access to the above variables 1378 // and to notify the foreground and background collectors. 1379 // _collectorState - current state of the CMS collection. 1380 // 1381 // The foreground collector 1382 // acquires the CGC_lock 1383 // sets _foregroundGCIsActive 1384 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1385 // various locks acquired in preparation for the collection 1386 // are released so as not to block the background collector 1387 // that is in the midst of a collection 1388 // proceeds with the collection 1389 // clears _foregroundGCIsActive 1390 // returns 1391 // 1392 // The background collector in a loop iterating on the phases of the 1393 // collection 1394 // acquires the CGC_lock 1395 // sets _foregroundGCShouldWait 1396 // if _foregroundGCIsActive is set 1397 // clears _foregroundGCShouldWait, notifies _CGC_lock 1398 // waits on _CGC_lock for _foregroundGCIsActive to become false 1399 // and exits the loop. 1400 // otherwise 1401 // proceed with that phase of the collection 1402 // if the phase is a stop-the-world phase, 1403 // yield the baton once more just before enqueueing 1404 // the stop-world CMS operation (executed by the VM thread). 1405 // returns after all phases of the collection are done 1406 // 1407 1408 void CMSCollector::acquire_control_and_collect(bool full, 1409 bool clear_all_soft_refs) { 1410 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1411 assert(!Thread::current()->is_ConcurrentGC_thread(), 1412 "shouldn't try to acquire control from self!"); 1413 1414 // Start the protocol for acquiring control of the 1415 // collection from the background collector (aka CMS thread). 1416 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1417 "VM thread should have CMS token"); 1418 // Remember the possibly interrupted state of an ongoing 1419 // concurrent collection 1420 CollectorState first_state = _collectorState; 1421 1422 // Signal to a possibly ongoing concurrent collection that 1423 // we want to do a foreground collection. 1424 _foregroundGCIsActive = true; 1425 1426 // release locks and wait for a notify from the background collector 1427 // releasing the locks in only necessary for phases which 1428 // do yields to improve the granularity of the collection. 1429 assert_lock_strong(bitMapLock()); 1430 // We need to lock the Free list lock for the space that we are 1431 // currently collecting. 1432 assert(haveFreelistLocks(), "Must be holding free list locks"); 1433 bitMapLock()->unlock(); 1434 releaseFreelistLocks(); 1435 { 1436 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1437 if (_foregroundGCShouldWait) { 1438 // We are going to be waiting for action for the CMS thread; 1439 // it had better not be gone (for instance at shutdown)! 1440 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1441 "CMS thread must be running"); 1442 // Wait here until the background collector gives us the go-ahead 1443 ConcurrentMarkSweepThread::clear_CMS_flag( 1444 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1445 // Get a possibly blocked CMS thread going: 1446 // Note that we set _foregroundGCIsActive true above, 1447 // without protection of the CGC_lock. 1448 CGC_lock->notify(); 1449 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1450 "Possible deadlock"); 1451 while (_foregroundGCShouldWait) { 1452 // wait for notification 1453 CGC_lock->wait_without_safepoint_check(); 1454 // Possibility of delay/starvation here, since CMS token does 1455 // not know to give priority to VM thread? Actually, i think 1456 // there wouldn't be any delay/starvation, but the proof of 1457 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1458 } 1459 ConcurrentMarkSweepThread::set_CMS_flag( 1460 ConcurrentMarkSweepThread::CMS_vm_has_token); 1461 } 1462 } 1463 // The CMS_token is already held. Get back the other locks. 1464 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1465 "VM thread should have CMS token"); 1466 getFreelistLocks(); 1467 bitMapLock()->lock_without_safepoint_check(); 1468 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1469 p2i(Thread::current()), first_state); 1470 log_debug(gc, state)(" gets control with state %d", _collectorState); 1471 1472 // Inform cms gen if this was due to partial collection failing. 1473 // The CMS gen may use this fact to determine its expansion policy. 1474 CMSHeap* heap = CMSHeap::heap(); 1475 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1476 assert(!_cmsGen->incremental_collection_failed(), 1477 "Should have been noticed, reacted to and cleared"); 1478 _cmsGen->set_incremental_collection_failed(); 1479 } 1480 1481 if (first_state > Idling) { 1482 report_concurrent_mode_interruption(); 1483 } 1484 1485 set_did_compact(true); 1486 1487 // If the collection is being acquired from the background 1488 // collector, there may be references on the discovered 1489 // references lists. Abandon those references, since some 1490 // of them may have become unreachable after concurrent 1491 // discovery; the STW compacting collector will redo discovery 1492 // more precisely, without being subject to floating garbage. 1493 // Leaving otherwise unreachable references in the discovered 1494 // lists would require special handling. 1495 ref_processor()->disable_discovery(); 1496 ref_processor()->abandon_partial_discovery(); 1497 ref_processor()->verify_no_references_recorded(); 1498 1499 if (first_state > Idling) { 1500 save_heap_summary(); 1501 } 1502 1503 do_compaction_work(clear_all_soft_refs); 1504 1505 // Has the GC time limit been exceeded? 1506 size_t max_eden_size = _young_gen->max_eden_size(); 1507 GCCause::Cause gc_cause = heap->gc_cause(); 1508 size_policy()->check_gc_overhead_limit(_young_gen->eden()->used(), 1509 _cmsGen->max_capacity(), 1510 max_eden_size, 1511 full, 1512 gc_cause, 1513 heap->soft_ref_policy()); 1514 1515 // Reset the expansion cause, now that we just completed 1516 // a collection cycle. 1517 clear_expansion_cause(); 1518 _foregroundGCIsActive = false; 1519 return; 1520 } 1521 1522 // Resize the tenured generation 1523 // after obtaining the free list locks for the 1524 // two generations. 1525 void CMSCollector::compute_new_size() { 1526 assert_locked_or_safepoint(Heap_lock); 1527 FreelistLocker z(this); 1528 MetaspaceGC::compute_new_size(); 1529 _cmsGen->compute_new_size_free_list(); 1530 // recalculate CMS used space after CMS collection 1531 _cmsGen->cmsSpace()->recalculate_used_stable(); 1532 } 1533 1534 // A work method used by the foreground collector to do 1535 // a mark-sweep-compact. 1536 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1537 CMSHeap* heap = CMSHeap::heap(); 1538 1539 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1540 gc_timer->register_gc_start(); 1541 1542 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1543 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1544 1545 heap->pre_full_gc_dump(gc_timer); 1546 1547 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1548 1549 // Temporarily widen the span of the weak reference processing to 1550 // the entire heap. 1551 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1552 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1553 // Temporarily, clear the "is_alive_non_header" field of the 1554 // reference processor. 1555 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1556 // Temporarily make reference _processing_ single threaded (non-MT). 1557 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1558 // Temporarily make refs discovery atomic 1559 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1560 // Temporarily make reference _discovery_ single threaded (non-MT) 1561 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1562 1563 ref_processor()->set_enqueuing_is_done(false); 1564 ref_processor()->enable_discovery(); 1565 ref_processor()->setup_policy(clear_all_soft_refs); 1566 // If an asynchronous collection finishes, the _modUnionTable is 1567 // all clear. If we are assuming the collection from an asynchronous 1568 // collection, clear the _modUnionTable. 1569 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1570 "_modUnionTable should be clear if the baton was not passed"); 1571 _modUnionTable.clear_all(); 1572 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1573 "mod union for klasses should be clear if the baton was passed"); 1574 _ct->cld_rem_set()->clear_mod_union(); 1575 1576 1577 // We must adjust the allocation statistics being maintained 1578 // in the free list space. We do so by reading and clearing 1579 // the sweep timer and updating the block flux rate estimates below. 1580 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1581 if (_inter_sweep_timer.is_active()) { 1582 _inter_sweep_timer.stop(); 1583 // Note that we do not use this sample to update the _inter_sweep_estimate. 1584 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1585 _inter_sweep_estimate.padded_average(), 1586 _intra_sweep_estimate.padded_average()); 1587 } 1588 1589 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1590 #ifdef ASSERT 1591 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1592 size_t free_size = cms_space->free(); 1593 assert(free_size == 1594 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1595 * HeapWordSize, 1596 "All the free space should be compacted into one chunk at top"); 1597 assert(cms_space->dictionary()->total_chunk_size( 1598 debug_only(cms_space->freelistLock())) == 0 || 1599 cms_space->totalSizeInIndexedFreeLists() == 0, 1600 "All the free space should be in a single chunk"); 1601 size_t num = cms_space->totalCount(); 1602 assert((free_size == 0 && num == 0) || 1603 (free_size > 0 && (num == 1 || num == 2)), 1604 "There should be at most 2 free chunks after compaction"); 1605 #endif // ASSERT 1606 _collectorState = Resetting; 1607 assert(_restart_addr == NULL, 1608 "Should have been NULL'd before baton was passed"); 1609 reset_stw(); 1610 _cmsGen->reset_after_compaction(); 1611 _concurrent_cycles_since_last_unload = 0; 1612 1613 // Clear any data recorded in the PLAB chunk arrays. 1614 if (_survivor_plab_array != NULL) { 1615 reset_survivor_plab_arrays(); 1616 } 1617 1618 // Adjust the per-size allocation stats for the next epoch. 1619 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1620 // Restart the "inter sweep timer" for the next epoch. 1621 _inter_sweep_timer.reset(); 1622 _inter_sweep_timer.start(); 1623 1624 // No longer a need to do a concurrent collection for Metaspace. 1625 MetaspaceGC::set_should_concurrent_collect(false); 1626 1627 heap->post_full_gc_dump(gc_timer); 1628 1629 gc_timer->register_gc_end(); 1630 1631 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1632 1633 // For a mark-sweep-compact, compute_new_size() will be called 1634 // in the heap's do_collection() method. 1635 } 1636 1637 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1638 Log(gc, heap) log; 1639 if (!log.is_trace()) { 1640 return; 1641 } 1642 1643 ContiguousSpace* eden_space = _young_gen->eden(); 1644 ContiguousSpace* from_space = _young_gen->from(); 1645 ContiguousSpace* to_space = _young_gen->to(); 1646 // Eden 1647 if (_eden_chunk_array != NULL) { 1648 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1649 p2i(eden_space->bottom()), p2i(eden_space->top()), 1650 p2i(eden_space->end()), eden_space->capacity()); 1651 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1652 _eden_chunk_index, _eden_chunk_capacity); 1653 for (size_t i = 0; i < _eden_chunk_index; i++) { 1654 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1655 } 1656 } 1657 // Survivor 1658 if (_survivor_chunk_array != NULL) { 1659 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1660 p2i(from_space->bottom()), p2i(from_space->top()), 1661 p2i(from_space->end()), from_space->capacity()); 1662 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1663 _survivor_chunk_index, _survivor_chunk_capacity); 1664 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1665 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1666 } 1667 } 1668 } 1669 1670 void CMSCollector::getFreelistLocks() const { 1671 // Get locks for all free lists in all generations that this 1672 // collector is responsible for 1673 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1674 } 1675 1676 void CMSCollector::releaseFreelistLocks() const { 1677 // Release locks for all free lists in all generations that this 1678 // collector is responsible for 1679 _cmsGen->freelistLock()->unlock(); 1680 } 1681 1682 bool CMSCollector::haveFreelistLocks() const { 1683 // Check locks for all free lists in all generations that this 1684 // collector is responsible for 1685 assert_lock_strong(_cmsGen->freelistLock()); 1686 PRODUCT_ONLY(ShouldNotReachHere()); 1687 return true; 1688 } 1689 1690 // A utility class that is used by the CMS collector to 1691 // temporarily "release" the foreground collector from its 1692 // usual obligation to wait for the background collector to 1693 // complete an ongoing phase before proceeding. 1694 class ReleaseForegroundGC: public StackObj { 1695 private: 1696 CMSCollector* _c; 1697 public: 1698 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1699 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1700 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1701 // allow a potentially blocked foreground collector to proceed 1702 _c->_foregroundGCShouldWait = false; 1703 if (_c->_foregroundGCIsActive) { 1704 CGC_lock->notify(); 1705 } 1706 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1707 "Possible deadlock"); 1708 } 1709 1710 ~ReleaseForegroundGC() { 1711 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1712 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1713 _c->_foregroundGCShouldWait = true; 1714 } 1715 }; 1716 1717 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1718 assert(Thread::current()->is_ConcurrentGC_thread(), 1719 "A CMS asynchronous collection is only allowed on a CMS thread."); 1720 1721 CMSHeap* heap = CMSHeap::heap(); 1722 { 1723 MutexLocker hl(Heap_lock, Mutex::_no_safepoint_check_flag); 1724 FreelistLocker fll(this); 1725 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1726 if (_foregroundGCIsActive) { 1727 // The foreground collector is. Skip this 1728 // background collection. 1729 assert(!_foregroundGCShouldWait, "Should be clear"); 1730 return; 1731 } else { 1732 assert(_collectorState == Idling, "Should be idling before start."); 1733 _collectorState = InitialMarking; 1734 register_gc_start(cause); 1735 // Reset the expansion cause, now that we are about to begin 1736 // a new cycle. 1737 clear_expansion_cause(); 1738 1739 // Clear the MetaspaceGC flag since a concurrent collection 1740 // is starting but also clear it after the collection. 1741 MetaspaceGC::set_should_concurrent_collect(false); 1742 } 1743 // Decide if we want to enable class unloading as part of the 1744 // ensuing concurrent GC cycle. 1745 update_should_unload_classes(); 1746 _full_gc_requested = false; // acks all outstanding full gc requests 1747 _full_gc_cause = GCCause::_no_gc; 1748 // Signal that we are about to start a collection 1749 heap->increment_total_full_collections(); // ... starting a collection cycle 1750 _collection_count_start = heap->total_full_collections(); 1751 } 1752 1753 size_t prev_used = _cmsGen->used(); 1754 1755 // The change of the collection state is normally done at this level; 1756 // the exceptions are phases that are executed while the world is 1757 // stopped. For those phases the change of state is done while the 1758 // world is stopped. For baton passing purposes this allows the 1759 // background collector to finish the phase and change state atomically. 1760 // The foreground collector cannot wait on a phase that is done 1761 // while the world is stopped because the foreground collector already 1762 // has the world stopped and would deadlock. 1763 while (_collectorState != Idling) { 1764 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1765 p2i(Thread::current()), _collectorState); 1766 // The foreground collector 1767 // holds the Heap_lock throughout its collection. 1768 // holds the CMS token (but not the lock) 1769 // except while it is waiting for the background collector to yield. 1770 // 1771 // The foreground collector should be blocked (not for long) 1772 // if the background collector is about to start a phase 1773 // executed with world stopped. If the background 1774 // collector has already started such a phase, the 1775 // foreground collector is blocked waiting for the 1776 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1777 // are executed in the VM thread. 1778 // 1779 // The locking order is 1780 // PendingListLock (PLL) -- if applicable (FinalMarking) 1781 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1782 // CMS token (claimed in 1783 // stop_world_and_do() --> 1784 // safepoint_synchronize() --> 1785 // CMSThread::synchronize()) 1786 1787 { 1788 // Check if the FG collector wants us to yield. 1789 CMSTokenSync x(true); // is cms thread 1790 if (waitForForegroundGC()) { 1791 // We yielded to a foreground GC, nothing more to be 1792 // done this round. 1793 assert(_foregroundGCShouldWait == false, "We set it to false in " 1794 "waitForForegroundGC()"); 1795 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1796 p2i(Thread::current()), _collectorState); 1797 return; 1798 } else { 1799 // The background collector can run but check to see if the 1800 // foreground collector has done a collection while the 1801 // background collector was waiting to get the CGC_lock 1802 // above. If yes, break so that _foregroundGCShouldWait 1803 // is cleared before returning. 1804 if (_collectorState == Idling) { 1805 break; 1806 } 1807 } 1808 } 1809 1810 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1811 "should be waiting"); 1812 1813 switch (_collectorState) { 1814 case InitialMarking: 1815 { 1816 ReleaseForegroundGC x(this); 1817 stats().record_cms_begin(); 1818 VM_CMS_Initial_Mark initial_mark_op(this); 1819 VMThread::execute(&initial_mark_op); 1820 } 1821 // The collector state may be any legal state at this point 1822 // since the background collector may have yielded to the 1823 // foreground collector. 1824 break; 1825 case Marking: 1826 // initial marking in checkpointRootsInitialWork has been completed 1827 if (markFromRoots()) { // we were successful 1828 assert(_collectorState == Precleaning, "Collector state should " 1829 "have changed"); 1830 } else { 1831 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1832 } 1833 break; 1834 case Precleaning: 1835 // marking from roots in markFromRoots has been completed 1836 preclean(); 1837 assert(_collectorState == AbortablePreclean || 1838 _collectorState == FinalMarking, 1839 "Collector state should have changed"); 1840 break; 1841 case AbortablePreclean: 1842 abortable_preclean(); 1843 assert(_collectorState == FinalMarking, "Collector state should " 1844 "have changed"); 1845 break; 1846 case FinalMarking: 1847 { 1848 ReleaseForegroundGC x(this); 1849 1850 VM_CMS_Final_Remark final_remark_op(this); 1851 VMThread::execute(&final_remark_op); 1852 } 1853 assert(_foregroundGCShouldWait, "block post-condition"); 1854 break; 1855 case Sweeping: 1856 // final marking in checkpointRootsFinal has been completed 1857 sweep(); 1858 assert(_collectorState == Resizing, "Collector state change " 1859 "to Resizing must be done under the free_list_lock"); 1860 1861 case Resizing: { 1862 // Sweeping has been completed... 1863 // At this point the background collection has completed. 1864 // Don't move the call to compute_new_size() down 1865 // into code that might be executed if the background 1866 // collection was preempted. 1867 { 1868 ReleaseForegroundGC x(this); // unblock FG collection 1869 MutexLocker y(Heap_lock, Mutex::_no_safepoint_check_flag); 1870 CMSTokenSync z(true); // not strictly needed. 1871 if (_collectorState == Resizing) { 1872 compute_new_size(); 1873 save_heap_summary(); 1874 _collectorState = Resetting; 1875 } else { 1876 assert(_collectorState == Idling, "The state should only change" 1877 " because the foreground collector has finished the collection"); 1878 } 1879 } 1880 break; 1881 } 1882 case Resetting: 1883 // CMS heap resizing has been completed 1884 reset_concurrent(); 1885 assert(_collectorState == Idling, "Collector state should " 1886 "have changed"); 1887 1888 MetaspaceGC::set_should_concurrent_collect(false); 1889 1890 stats().record_cms_end(); 1891 // Don't move the concurrent_phases_end() and compute_new_size() 1892 // calls to here because a preempted background collection 1893 // has it's state set to "Resetting". 1894 break; 1895 case Idling: 1896 default: 1897 ShouldNotReachHere(); 1898 break; 1899 } 1900 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1901 p2i(Thread::current()), _collectorState); 1902 assert(_foregroundGCShouldWait, "block post-condition"); 1903 } 1904 1905 // Should this be in gc_epilogue? 1906 heap->counters()->update_counters(); 1907 1908 { 1909 // Clear _foregroundGCShouldWait and, in the event that the 1910 // foreground collector is waiting, notify it, before 1911 // returning. 1912 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1913 _foregroundGCShouldWait = false; 1914 if (_foregroundGCIsActive) { 1915 CGC_lock->notify(); 1916 } 1917 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1918 "Possible deadlock"); 1919 } 1920 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1921 p2i(Thread::current()), _collectorState); 1922 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1923 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1924 } 1925 1926 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1927 _cms_start_registered = true; 1928 _gc_timer_cm->register_gc_start(); 1929 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1930 } 1931 1932 void CMSCollector::register_gc_end() { 1933 if (_cms_start_registered) { 1934 report_heap_summary(GCWhen::AfterGC); 1935 1936 _gc_timer_cm->register_gc_end(); 1937 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1938 _cms_start_registered = false; 1939 } 1940 } 1941 1942 void CMSCollector::save_heap_summary() { 1943 CMSHeap* heap = CMSHeap::heap(); 1944 _last_heap_summary = heap->create_heap_summary(); 1945 _last_metaspace_summary = heap->create_metaspace_summary(); 1946 } 1947 1948 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1949 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1950 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1951 } 1952 1953 bool CMSCollector::waitForForegroundGC() { 1954 bool res = false; 1955 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1956 "CMS thread should have CMS token"); 1957 // Block the foreground collector until the 1958 // background collectors decides whether to 1959 // yield. 1960 MutexLocker x(CGC_lock, Mutex::_no_safepoint_check_flag); 1961 _foregroundGCShouldWait = true; 1962 if (_foregroundGCIsActive) { 1963 // The background collector yields to the 1964 // foreground collector and returns a value 1965 // indicating that it has yielded. The foreground 1966 // collector can proceed. 1967 res = true; 1968 _foregroundGCShouldWait = false; 1969 ConcurrentMarkSweepThread::clear_CMS_flag( 1970 ConcurrentMarkSweepThread::CMS_cms_has_token); 1971 ConcurrentMarkSweepThread::set_CMS_flag( 1972 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1973 // Get a possibly blocked foreground thread going 1974 CGC_lock->notify(); 1975 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1976 p2i(Thread::current()), _collectorState); 1977 while (_foregroundGCIsActive) { 1978 CGC_lock->wait_without_safepoint_check(); 1979 } 1980 ConcurrentMarkSweepThread::set_CMS_flag( 1981 ConcurrentMarkSweepThread::CMS_cms_has_token); 1982 ConcurrentMarkSweepThread::clear_CMS_flag( 1983 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1984 } 1985 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1986 p2i(Thread::current()), _collectorState); 1987 return res; 1988 } 1989 1990 // Because of the need to lock the free lists and other structures in 1991 // the collector, common to all the generations that the collector is 1992 // collecting, we need the gc_prologues of individual CMS generations 1993 // delegate to their collector. It may have been simpler had the 1994 // current infrastructure allowed one to call a prologue on a 1995 // collector. In the absence of that we have the generation's 1996 // prologue delegate to the collector, which delegates back 1997 // some "local" work to a worker method in the individual generations 1998 // that it's responsible for collecting, while itself doing any 1999 // work common to all generations it's responsible for. A similar 2000 // comment applies to the gc_epilogue()'s. 2001 // The role of the variable _between_prologue_and_epilogue is to 2002 // enforce the invocation protocol. 2003 void CMSCollector::gc_prologue(bool full) { 2004 // Call gc_prologue_work() for the CMSGen 2005 // we are responsible for. 2006 2007 // The following locking discipline assumes that we are only called 2008 // when the world is stopped. 2009 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2010 2011 // The CMSCollector prologue must call the gc_prologues for the 2012 // "generations" that it's responsible 2013 // for. 2014 2015 assert( Thread::current()->is_VM_thread() 2016 || ( CMSScavengeBeforeRemark 2017 && Thread::current()->is_ConcurrentGC_thread()), 2018 "Incorrect thread type for prologue execution"); 2019 2020 if (_between_prologue_and_epilogue) { 2021 // We have already been invoked; this is a gc_prologue delegation 2022 // from yet another CMS generation that we are responsible for, just 2023 // ignore it since all relevant work has already been done. 2024 return; 2025 } 2026 2027 // set a bit saying prologue has been called; cleared in epilogue 2028 _between_prologue_and_epilogue = true; 2029 // Claim locks for common data structures, then call gc_prologue_work() 2030 // for each CMSGen. 2031 2032 getFreelistLocks(); // gets free list locks on constituent spaces 2033 bitMapLock()->lock_without_safepoint_check(); 2034 2035 // Should call gc_prologue_work() for all cms gens we are responsible for 2036 bool duringMarking = _collectorState >= Marking 2037 && _collectorState < Sweeping; 2038 2039 // The young collections clear the modified oops state, which tells if 2040 // there are any modified oops in the class. The remark phase also needs 2041 // that information. Tell the young collection to save the union of all 2042 // modified klasses. 2043 if (duringMarking) { 2044 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2045 } 2046 2047 bool registerClosure = duringMarking; 2048 2049 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2050 2051 if (!full) { 2052 stats().record_gc0_begin(); 2053 } 2054 } 2055 2056 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2057 2058 _capacity_at_prologue = capacity(); 2059 _used_at_prologue = used(); 2060 _cmsSpace->recalculate_used_stable(); 2061 2062 // We enable promotion tracking so that card-scanning can recognize 2063 // which objects have been promoted during this GC and skip them. 2064 for (uint i = 0; i < ParallelGCThreads; i++) { 2065 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2066 } 2067 2068 // Delegate to CMScollector which knows how to coordinate between 2069 // this and any other CMS generations that it is responsible for 2070 // collecting. 2071 collector()->gc_prologue(full); 2072 } 2073 2074 // This is a "private" interface for use by this generation's CMSCollector. 2075 // Not to be called directly by any other entity (for instance, 2076 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2077 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2078 bool registerClosure, ModUnionClosure* modUnionClosure) { 2079 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2080 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2081 "Should be NULL"); 2082 if (registerClosure) { 2083 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2084 } 2085 cmsSpace()->gc_prologue(); 2086 // Clear stat counters 2087 NOT_PRODUCT( 2088 assert(_numObjectsPromoted == 0, "check"); 2089 assert(_numWordsPromoted == 0, "check"); 2090 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2091 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2092 _numObjectsAllocated = 0; 2093 _numWordsAllocated = 0; 2094 ) 2095 } 2096 2097 void CMSCollector::gc_epilogue(bool full) { 2098 // The following locking discipline assumes that we are only called 2099 // when the world is stopped. 2100 assert(SafepointSynchronize::is_at_safepoint(), 2101 "world is stopped assumption"); 2102 2103 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2104 // if linear allocation blocks need to be appropriately marked to allow the 2105 // the blocks to be parsable. We also check here whether we need to nudge the 2106 // CMS collector thread to start a new cycle (if it's not already active). 2107 assert( Thread::current()->is_VM_thread() 2108 || ( CMSScavengeBeforeRemark 2109 && Thread::current()->is_ConcurrentGC_thread()), 2110 "Incorrect thread type for epilogue execution"); 2111 2112 if (!_between_prologue_and_epilogue) { 2113 // We have already been invoked; this is a gc_epilogue delegation 2114 // from yet another CMS generation that we are responsible for, just 2115 // ignore it since all relevant work has already been done. 2116 return; 2117 } 2118 assert(haveFreelistLocks(), "must have freelist locks"); 2119 assert_lock_strong(bitMapLock()); 2120 2121 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2122 2123 _cmsGen->gc_epilogue_work(full); 2124 2125 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2126 // in case sampling was not already enabled, enable it 2127 _start_sampling = true; 2128 } 2129 // reset _eden_chunk_array so sampling starts afresh 2130 _eden_chunk_index = 0; 2131 2132 size_t cms_used = _cmsGen->cmsSpace()->used(); 2133 _cmsGen->cmsSpace()->recalculate_used_stable(); 2134 2135 // update performance counters - this uses a special version of 2136 // update_counters() that allows the utilization to be passed as a 2137 // parameter, avoiding multiple calls to used(). 2138 // 2139 _cmsGen->update_counters(cms_used); 2140 2141 bitMapLock()->unlock(); 2142 releaseFreelistLocks(); 2143 2144 if (!CleanChunkPoolAsync) { 2145 Chunk::clean_chunk_pool(); 2146 } 2147 2148 set_did_compact(false); 2149 _between_prologue_and_epilogue = false; // ready for next cycle 2150 } 2151 2152 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2153 collector()->gc_epilogue(full); 2154 2155 // When using ParNew, promotion tracking should have already been 2156 // disabled. However, the prologue (which enables promotion 2157 // tracking) and epilogue are called irrespective of the type of 2158 // GC. So they will also be called before and after Full GCs, during 2159 // which promotion tracking will not be explicitly disabled. So, 2160 // it's safer to also disable it here too (to be symmetric with 2161 // enabling it in the prologue). 2162 for (uint i = 0; i < ParallelGCThreads; i++) { 2163 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2164 } 2165 } 2166 2167 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2168 assert(!incremental_collection_failed(), "Should have been cleared"); 2169 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2170 cmsSpace()->gc_epilogue(); 2171 // Print stat counters 2172 NOT_PRODUCT( 2173 assert(_numObjectsAllocated == 0, "check"); 2174 assert(_numWordsAllocated == 0, "check"); 2175 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2176 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2177 _numObjectsPromoted = 0; 2178 _numWordsPromoted = 0; 2179 ) 2180 2181 // Call down the chain in contiguous_available needs the freelistLock 2182 // so print this out before releasing the freeListLock. 2183 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2184 } 2185 2186 #ifndef PRODUCT 2187 bool CMSCollector::have_cms_token() { 2188 Thread* thr = Thread::current(); 2189 if (thr->is_VM_thread()) { 2190 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2191 } else if (thr->is_ConcurrentGC_thread()) { 2192 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2193 } else if (thr->is_GC_task_thread()) { 2194 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2195 ParGCRareEvent_lock->owned_by_self(); 2196 } 2197 return false; 2198 } 2199 2200 // Check reachability of the given heap address in CMS generation, 2201 // treating all other generations as roots. 2202 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2203 // We could "guarantee" below, rather than assert, but I'll 2204 // leave these as "asserts" so that an adventurous debugger 2205 // could try this in the product build provided some subset of 2206 // the conditions were met, provided they were interested in the 2207 // results and knew that the computation below wouldn't interfere 2208 // with other concurrent computations mutating the structures 2209 // being read or written. 2210 assert(SafepointSynchronize::is_at_safepoint(), 2211 "Else mutations in object graph will make answer suspect"); 2212 assert(have_cms_token(), "Should hold cms token"); 2213 assert(haveFreelistLocks(), "must hold free list locks"); 2214 assert_lock_strong(bitMapLock()); 2215 2216 // Clear the marking bit map array before starting, but, just 2217 // for kicks, first report if the given address is already marked 2218 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2219 _markBitMap.isMarked(addr) ? "" : " not"); 2220 2221 if (verify_after_remark()) { 2222 MutexLocker x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2223 bool result = verification_mark_bm()->isMarked(addr); 2224 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2225 result ? "IS" : "is NOT"); 2226 return result; 2227 } else { 2228 tty->print_cr("Could not compute result"); 2229 return false; 2230 } 2231 } 2232 #endif 2233 2234 void 2235 CMSCollector::print_on_error(outputStream* st) { 2236 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2237 if (collector != NULL) { 2238 CMSBitMap* bitmap = &collector->_markBitMap; 2239 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2240 bitmap->print_on_error(st, " Bits: "); 2241 2242 st->cr(); 2243 2244 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2245 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2246 mut_bitmap->print_on_error(st, " Bits: "); 2247 } 2248 } 2249 2250 //////////////////////////////////////////////////////// 2251 // CMS Verification Support 2252 //////////////////////////////////////////////////////// 2253 // Following the remark phase, the following invariant 2254 // should hold -- each object in the CMS heap which is 2255 // marked in markBitMap() should be marked in the verification_mark_bm(). 2256 2257 class VerifyMarkedClosure: public BitMapClosure { 2258 CMSBitMap* _marks; 2259 bool _failed; 2260 2261 public: 2262 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2263 2264 bool do_bit(size_t offset) { 2265 HeapWord* addr = _marks->offsetToHeapWord(offset); 2266 if (!_marks->isMarked(addr)) { 2267 Log(gc, verify) log; 2268 ResourceMark rm; 2269 LogStream ls(log.error()); 2270 oop(addr)->print_on(&ls); 2271 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2272 _failed = true; 2273 } 2274 return true; 2275 } 2276 2277 bool failed() { return _failed; } 2278 }; 2279 2280 bool CMSCollector::verify_after_remark() { 2281 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2282 MutexLocker ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2283 static bool init = false; 2284 2285 assert(SafepointSynchronize::is_at_safepoint(), 2286 "Else mutations in object graph will make answer suspect"); 2287 assert(have_cms_token(), 2288 "Else there may be mutual interference in use of " 2289 " verification data structures"); 2290 assert(_collectorState > Marking && _collectorState <= Sweeping, 2291 "Else marking info checked here may be obsolete"); 2292 assert(haveFreelistLocks(), "must hold free list locks"); 2293 assert_lock_strong(bitMapLock()); 2294 2295 2296 // Allocate marking bit map if not already allocated 2297 if (!init) { // first time 2298 if (!verification_mark_bm()->allocate(_span)) { 2299 return false; 2300 } 2301 init = true; 2302 } 2303 2304 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2305 2306 // Turn off refs discovery -- so we will be tracing through refs. 2307 // This is as intended, because by this time 2308 // GC must already have cleared any refs that need to be cleared, 2309 // and traced those that need to be marked; moreover, 2310 // the marking done here is not going to interfere in any 2311 // way with the marking information used by GC. 2312 NoRefDiscovery no_discovery(ref_processor()); 2313 2314 #if COMPILER2_OR_JVMCI 2315 DerivedPointerTableDeactivate dpt_deact; 2316 #endif 2317 2318 // Clear any marks from a previous round 2319 verification_mark_bm()->clear_all(); 2320 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2321 verify_work_stacks_empty(); 2322 2323 CMSHeap* heap = CMSHeap::heap(); 2324 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2325 // Update the saved marks which may affect the root scans. 2326 heap->save_marks(); 2327 2328 if (CMSRemarkVerifyVariant == 1) { 2329 // In this first variant of verification, we complete 2330 // all marking, then check if the new marks-vector is 2331 // a subset of the CMS marks-vector. 2332 verify_after_remark_work_1(); 2333 } else { 2334 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2335 // In this second variant of verification, we flag an error 2336 // (i.e. an object reachable in the new marks-vector not reachable 2337 // in the CMS marks-vector) immediately, also indicating the 2338 // identify of an object (A) that references the unmarked object (B) -- 2339 // presumably, a mutation to A failed to be picked up by preclean/remark? 2340 verify_after_remark_work_2(); 2341 } 2342 2343 return true; 2344 } 2345 2346 void CMSCollector::verify_after_remark_work_1() { 2347 ResourceMark rm; 2348 HandleMark hm; 2349 CMSHeap* heap = CMSHeap::heap(); 2350 2351 // Get a clear set of claim bits for the roots processing to work with. 2352 ClassLoaderDataGraph::clear_claimed_marks(); 2353 2354 // Mark from roots one level into CMS 2355 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2356 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2357 2358 { 2359 StrongRootsScope srs(1); 2360 2361 heap->cms_process_roots(&srs, 2362 true, // young gen as roots 2363 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2364 should_unload_classes(), 2365 ¬Older, 2366 NULL); 2367 } 2368 2369 // Now mark from the roots 2370 MarkFromRootsClosure markFromRootsClosure(this, _span, 2371 verification_mark_bm(), verification_mark_stack(), 2372 false /* don't yield */, true /* verifying */); 2373 assert(_restart_addr == NULL, "Expected pre-condition"); 2374 verification_mark_bm()->iterate(&markFromRootsClosure); 2375 while (_restart_addr != NULL) { 2376 // Deal with stack overflow: by restarting at the indicated 2377 // address. 2378 HeapWord* ra = _restart_addr; 2379 markFromRootsClosure.reset(ra); 2380 _restart_addr = NULL; 2381 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2382 } 2383 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2384 verify_work_stacks_empty(); 2385 2386 // Marking completed -- now verify that each bit marked in 2387 // verification_mark_bm() is also marked in markBitMap(); flag all 2388 // errors by printing corresponding objects. 2389 VerifyMarkedClosure vcl(markBitMap()); 2390 verification_mark_bm()->iterate(&vcl); 2391 if (vcl.failed()) { 2392 Log(gc, verify) log; 2393 log.error("Failed marking verification after remark"); 2394 ResourceMark rm; 2395 LogStream ls(log.error()); 2396 heap->print_on(&ls); 2397 fatal("CMS: failed marking verification after remark"); 2398 } 2399 } 2400 2401 class VerifyCLDOopsCLDClosure : public CLDClosure { 2402 class VerifyCLDOopsClosure : public OopClosure { 2403 CMSBitMap* _bitmap; 2404 public: 2405 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2406 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2407 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2408 } _oop_closure; 2409 public: 2410 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2411 void do_cld(ClassLoaderData* cld) { 2412 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2413 } 2414 }; 2415 2416 void CMSCollector::verify_after_remark_work_2() { 2417 ResourceMark rm; 2418 HandleMark hm; 2419 CMSHeap* heap = CMSHeap::heap(); 2420 2421 // Get a clear set of claim bits for the roots processing to work with. 2422 ClassLoaderDataGraph::clear_claimed_marks(); 2423 2424 // Mark from roots one level into CMS 2425 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2426 markBitMap()); 2427 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2428 2429 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2430 2431 { 2432 StrongRootsScope srs(1); 2433 2434 heap->cms_process_roots(&srs, 2435 true, // young gen as roots 2436 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2437 should_unload_classes(), 2438 ¬Older, 2439 &cld_closure); 2440 } 2441 2442 // Now mark from the roots 2443 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2444 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2445 assert(_restart_addr == NULL, "Expected pre-condition"); 2446 verification_mark_bm()->iterate(&markFromRootsClosure); 2447 while (_restart_addr != NULL) { 2448 // Deal with stack overflow: by restarting at the indicated 2449 // address. 2450 HeapWord* ra = _restart_addr; 2451 markFromRootsClosure.reset(ra); 2452 _restart_addr = NULL; 2453 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2454 } 2455 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2456 verify_work_stacks_empty(); 2457 2458 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2459 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2460 2461 // Marking completed -- now verify that each bit marked in 2462 // verification_mark_bm() is also marked in markBitMap(); flag all 2463 // errors by printing corresponding objects. 2464 VerifyMarkedClosure vcl(markBitMap()); 2465 verification_mark_bm()->iterate(&vcl); 2466 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2467 } 2468 2469 void ConcurrentMarkSweepGeneration::save_marks() { 2470 // delegate to CMS space 2471 cmsSpace()->save_marks(); 2472 } 2473 2474 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2475 return cmsSpace()->no_allocs_since_save_marks(); 2476 } 2477 2478 void 2479 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2480 if (freelistLock()->owned_by_self()) { 2481 Generation::oop_iterate(cl); 2482 } else { 2483 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2484 Generation::oop_iterate(cl); 2485 } 2486 } 2487 2488 void 2489 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2490 if (freelistLock()->owned_by_self()) { 2491 Generation::object_iterate(cl); 2492 } else { 2493 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2494 Generation::object_iterate(cl); 2495 } 2496 } 2497 2498 void 2499 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2500 if (freelistLock()->owned_by_self()) { 2501 Generation::safe_object_iterate(cl); 2502 } else { 2503 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2504 Generation::safe_object_iterate(cl); 2505 } 2506 } 2507 2508 void 2509 ConcurrentMarkSweepGeneration::post_compact() { 2510 } 2511 2512 void 2513 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2514 // Fix the linear allocation blocks to look like free blocks. 2515 2516 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2517 // are not called when the heap is verified during universe initialization and 2518 // at vm shutdown. 2519 if (freelistLock()->owned_by_self()) { 2520 cmsSpace()->prepare_for_verify(); 2521 } else { 2522 MutexLocker fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2523 cmsSpace()->prepare_for_verify(); 2524 } 2525 } 2526 2527 void 2528 ConcurrentMarkSweepGeneration::verify() { 2529 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2530 // are not called when the heap is verified during universe initialization and 2531 // at vm shutdown. 2532 if (freelistLock()->owned_by_self()) { 2533 cmsSpace()->verify(); 2534 } else { 2535 MutexLocker fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2536 cmsSpace()->verify(); 2537 } 2538 } 2539 2540 void CMSCollector::verify() { 2541 _cmsGen->verify(); 2542 } 2543 2544 #ifndef PRODUCT 2545 bool CMSCollector::overflow_list_is_empty() const { 2546 assert(_num_par_pushes >= 0, "Inconsistency"); 2547 if (_overflow_list == NULL) { 2548 assert(_num_par_pushes == 0, "Inconsistency"); 2549 } 2550 return _overflow_list == NULL; 2551 } 2552 2553 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2554 // merely consolidate assertion checks that appear to occur together frequently. 2555 void CMSCollector::verify_work_stacks_empty() const { 2556 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2557 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2558 } 2559 2560 void CMSCollector::verify_overflow_empty() const { 2561 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2562 assert(no_preserved_marks(), "No preserved marks"); 2563 } 2564 #endif // PRODUCT 2565 2566 // Decide if we want to enable class unloading as part of the 2567 // ensuing concurrent GC cycle. We will collect and 2568 // unload classes if it's the case that: 2569 // (a) class unloading is enabled at the command line, and 2570 // (b) old gen is getting really full 2571 // NOTE: Provided there is no change in the state of the heap between 2572 // calls to this method, it should have idempotent results. Moreover, 2573 // its results should be monotonically increasing (i.e. going from 0 to 1, 2574 // but not 1 to 0) between successive calls between which the heap was 2575 // not collected. For the implementation below, it must thus rely on 2576 // the property that concurrent_cycles_since_last_unload() 2577 // will not decrease unless a collection cycle happened and that 2578 // _cmsGen->is_too_full() are 2579 // themselves also monotonic in that sense. See check_monotonicity() 2580 // below. 2581 void CMSCollector::update_should_unload_classes() { 2582 _should_unload_classes = false; 2583 if (CMSClassUnloadingEnabled) { 2584 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2585 CMSClassUnloadingMaxInterval) 2586 || _cmsGen->is_too_full(); 2587 } 2588 } 2589 2590 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2591 bool res = should_concurrent_collect(); 2592 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2593 return res; 2594 } 2595 2596 void CMSCollector::setup_cms_unloading_and_verification_state() { 2597 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2598 || VerifyBeforeExit; 2599 const int rso = GenCollectedHeap::SO_AllCodeCache; 2600 2601 // We set the proper root for this CMS cycle here. 2602 if (should_unload_classes()) { // Should unload classes this cycle 2603 remove_root_scanning_option(rso); // Shrink the root set appropriately 2604 set_verifying(should_verify); // Set verification state for this cycle 2605 return; // Nothing else needs to be done at this time 2606 } 2607 2608 // Not unloading classes this cycle 2609 assert(!should_unload_classes(), "Inconsistency!"); 2610 2611 // If we are not unloading classes then add SO_AllCodeCache to root 2612 // scanning options. 2613 add_root_scanning_option(rso); 2614 2615 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2616 set_verifying(true); 2617 } else if (verifying() && !should_verify) { 2618 // We were verifying, but some verification flags got disabled. 2619 set_verifying(false); 2620 // Exclude symbols, strings and code cache elements from root scanning to 2621 // reduce IM and RM pauses. 2622 remove_root_scanning_option(rso); 2623 } 2624 } 2625 2626 2627 #ifndef PRODUCT 2628 HeapWord* CMSCollector::block_start(const void* p) const { 2629 const HeapWord* addr = (HeapWord*)p; 2630 if (_span.contains(p)) { 2631 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2632 return _cmsGen->cmsSpace()->block_start(p); 2633 } 2634 } 2635 return NULL; 2636 } 2637 #endif 2638 2639 HeapWord* 2640 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2641 bool tlab, 2642 bool parallel) { 2643 CMSSynchronousYieldRequest yr; 2644 assert(!tlab, "Can't deal with TLAB allocation"); 2645 MutexLocker x(freelistLock(), Mutex::_no_safepoint_check_flag); 2646 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2647 if (GCExpandToAllocateDelayMillis > 0) { 2648 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2649 } 2650 return have_lock_and_allocate(word_size, tlab); 2651 } 2652 2653 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2654 size_t bytes, 2655 size_t expand_bytes, 2656 CMSExpansionCause::Cause cause) 2657 { 2658 2659 bool success = expand(bytes, expand_bytes); 2660 2661 // remember why we expanded; this information is used 2662 // by shouldConcurrentCollect() when making decisions on whether to start 2663 // a new CMS cycle. 2664 if (success) { 2665 set_expansion_cause(cause); 2666 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2667 } 2668 } 2669 2670 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2671 HeapWord* res = NULL; 2672 MutexLocker x(ParGCRareEvent_lock); 2673 while (true) { 2674 // Expansion by some other thread might make alloc OK now: 2675 res = ps->lab.alloc(word_sz); 2676 if (res != NULL) return res; 2677 // If there's not enough expansion space available, give up. 2678 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2679 return NULL; 2680 } 2681 // Otherwise, we try expansion. 2682 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2683 // Now go around the loop and try alloc again; 2684 // A competing par_promote might beat us to the expansion space, 2685 // so we may go around the loop again if promotion fails again. 2686 if (GCExpandToAllocateDelayMillis > 0) { 2687 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2688 } 2689 } 2690 } 2691 2692 2693 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2694 PromotionInfo* promo) { 2695 MutexLocker x(ParGCRareEvent_lock); 2696 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2697 while (true) { 2698 // Expansion by some other thread might make alloc OK now: 2699 if (promo->ensure_spooling_space()) { 2700 assert(promo->has_spooling_space(), 2701 "Post-condition of successful ensure_spooling_space()"); 2702 return true; 2703 } 2704 // If there's not enough expansion space available, give up. 2705 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2706 return false; 2707 } 2708 // Otherwise, we try expansion. 2709 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2710 // Now go around the loop and try alloc again; 2711 // A competing allocation might beat us to the expansion space, 2712 // so we may go around the loop again if allocation fails again. 2713 if (GCExpandToAllocateDelayMillis > 0) { 2714 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2715 } 2716 } 2717 } 2718 2719 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2720 // Only shrink if a compaction was done so that all the free space 2721 // in the generation is in a contiguous block at the end. 2722 if (did_compact()) { 2723 CardGeneration::shrink(bytes); 2724 } 2725 } 2726 2727 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2728 assert_locked_or_safepoint(Heap_lock); 2729 } 2730 2731 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2732 assert_locked_or_safepoint(Heap_lock); 2733 assert_lock_strong(freelistLock()); 2734 log_trace(gc)("Shrinking of CMS not yet implemented"); 2735 return; 2736 } 2737 2738 2739 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2740 // phases. 2741 class CMSPhaseAccounting: public StackObj { 2742 public: 2743 CMSPhaseAccounting(CMSCollector *collector, 2744 const char *title); 2745 ~CMSPhaseAccounting(); 2746 2747 private: 2748 CMSCollector *_collector; 2749 const char *_title; 2750 GCTraceConcTime(Info, gc) _trace_time; 2751 2752 public: 2753 // Not MT-safe; so do not pass around these StackObj's 2754 // where they may be accessed by other threads. 2755 double wallclock_millis() { 2756 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2757 } 2758 }; 2759 2760 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2761 const char *title) : 2762 _collector(collector), _title(title), _trace_time(title) { 2763 2764 _collector->resetYields(); 2765 _collector->resetTimer(); 2766 _collector->startTimer(); 2767 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2768 } 2769 2770 CMSPhaseAccounting::~CMSPhaseAccounting() { 2771 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2772 _collector->stopTimer(); 2773 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_millis(_collector->timerTicks())); 2774 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2775 } 2776 2777 // CMS work 2778 2779 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2780 class CMSParMarkTask : public AbstractGangTask { 2781 protected: 2782 CMSCollector* _collector; 2783 uint _n_workers; 2784 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2785 AbstractGangTask(name), 2786 _collector(collector), 2787 _n_workers(n_workers) {} 2788 // Work method in support of parallel rescan ... of young gen spaces 2789 void do_young_space_rescan(OopsInGenClosure* cl, 2790 ContiguousSpace* space, 2791 HeapWord** chunk_array, size_t chunk_top); 2792 void work_on_young_gen_roots(OopsInGenClosure* cl); 2793 }; 2794 2795 // Parallel initial mark task 2796 class CMSParInitialMarkTask: public CMSParMarkTask { 2797 StrongRootsScope* _strong_roots_scope; 2798 public: 2799 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2800 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2801 _strong_roots_scope(strong_roots_scope) {} 2802 void work(uint worker_id); 2803 }; 2804 2805 // Checkpoint the roots into this generation from outside 2806 // this generation. [Note this initial checkpoint need only 2807 // be approximate -- we'll do a catch up phase subsequently.] 2808 void CMSCollector::checkpointRootsInitial() { 2809 assert(_collectorState == InitialMarking, "Wrong collector state"); 2810 check_correct_thread_executing(); 2811 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2812 2813 save_heap_summary(); 2814 report_heap_summary(GCWhen::BeforeGC); 2815 2816 ReferenceProcessor* rp = ref_processor(); 2817 assert(_restart_addr == NULL, "Control point invariant"); 2818 { 2819 // acquire locks for subsequent manipulations 2820 MutexLocker x(bitMapLock(), 2821 Mutex::_no_safepoint_check_flag); 2822 checkpointRootsInitialWork(); 2823 // enable ("weak") refs discovery 2824 rp->enable_discovery(); 2825 _collectorState = Marking; 2826 } 2827 2828 _cmsGen->cmsSpace()->recalculate_used_stable(); 2829 } 2830 2831 void CMSCollector::checkpointRootsInitialWork() { 2832 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2833 assert(_collectorState == InitialMarking, "just checking"); 2834 2835 // Already have locks. 2836 assert_lock_strong(bitMapLock()); 2837 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2838 2839 // Setup the verification and class unloading state for this 2840 // CMS collection cycle. 2841 setup_cms_unloading_and_verification_state(); 2842 2843 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2844 2845 // Reset all the PLAB chunk arrays if necessary. 2846 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2847 reset_survivor_plab_arrays(); 2848 } 2849 2850 ResourceMark rm; 2851 HandleMark hm; 2852 2853 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2854 CMSHeap* heap = CMSHeap::heap(); 2855 2856 verify_work_stacks_empty(); 2857 verify_overflow_empty(); 2858 2859 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2860 // Update the saved marks which may affect the root scans. 2861 heap->save_marks(); 2862 2863 // weak reference processing has not started yet. 2864 ref_processor()->set_enqueuing_is_done(false); 2865 2866 // Need to remember all newly created CLDs, 2867 // so that we can guarantee that the remark finds them. 2868 ClassLoaderDataGraph::remember_new_clds(true); 2869 2870 // Whenever a CLD is found, it will be claimed before proceeding to mark 2871 // the klasses. The claimed marks need to be cleared before marking starts. 2872 ClassLoaderDataGraph::clear_claimed_marks(); 2873 2874 print_eden_and_survivor_chunk_arrays(); 2875 2876 { 2877 #if COMPILER2_OR_JVMCI 2878 DerivedPointerTableDeactivate dpt_deact; 2879 #endif 2880 if (CMSParallelInitialMarkEnabled) { 2881 // The parallel version. 2882 WorkGang* workers = heap->workers(); 2883 assert(workers != NULL, "Need parallel worker threads."); 2884 uint n_workers = workers->active_workers(); 2885 2886 StrongRootsScope srs(n_workers); 2887 2888 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2889 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2890 // If the total workers is greater than 1, then multiple workers 2891 // may be used at some time and the initialization has been set 2892 // such that the single threaded path cannot be used. 2893 if (workers->total_workers() > 1) { 2894 workers->run_task(&tsk); 2895 } else { 2896 tsk.work(0); 2897 } 2898 } else { 2899 // The serial version. 2900 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2901 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2902 2903 StrongRootsScope srs(1); 2904 2905 heap->cms_process_roots(&srs, 2906 true, // young gen as roots 2907 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2908 should_unload_classes(), 2909 ¬Older, 2910 &cld_closure); 2911 } 2912 } 2913 2914 // Clear mod-union table; it will be dirtied in the prologue of 2915 // CMS generation per each young generation collection. 2916 2917 assert(_modUnionTable.isAllClear(), 2918 "Was cleared in most recent final checkpoint phase" 2919 " or no bits are set in the gc_prologue before the start of the next " 2920 "subsequent marking phase."); 2921 2922 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2923 2924 // Save the end of the used_region of the constituent generations 2925 // to be used to limit the extent of sweep in each generation. 2926 save_sweep_limits(); 2927 verify_overflow_empty(); 2928 } 2929 2930 bool CMSCollector::markFromRoots() { 2931 // we might be tempted to assert that: 2932 // assert(!SafepointSynchronize::is_at_safepoint(), 2933 // "inconsistent argument?"); 2934 // However that wouldn't be right, because it's possible that 2935 // a safepoint is indeed in progress as a young generation 2936 // stop-the-world GC happens even as we mark in this generation. 2937 assert(_collectorState == Marking, "inconsistent state?"); 2938 check_correct_thread_executing(); 2939 verify_overflow_empty(); 2940 2941 // Weak ref discovery note: We may be discovering weak 2942 // refs in this generation concurrent (but interleaved) with 2943 // weak ref discovery by the young generation collector. 2944 2945 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2946 GCTraceCPUTime tcpu; 2947 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2948 bool res = markFromRootsWork(); 2949 if (res) { 2950 _collectorState = Precleaning; 2951 } else { // We failed and a foreground collection wants to take over 2952 assert(_foregroundGCIsActive, "internal state inconsistency"); 2953 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2954 log_debug(gc)("bailing out to foreground collection"); 2955 } 2956 verify_overflow_empty(); 2957 return res; 2958 } 2959 2960 bool CMSCollector::markFromRootsWork() { 2961 // iterate over marked bits in bit map, doing a full scan and mark 2962 // from these roots using the following algorithm: 2963 // . if oop is to the right of the current scan pointer, 2964 // mark corresponding bit (we'll process it later) 2965 // . else (oop is to left of current scan pointer) 2966 // push oop on marking stack 2967 // . drain the marking stack 2968 2969 // Note that when we do a marking step we need to hold the 2970 // bit map lock -- recall that direct allocation (by mutators) 2971 // and promotion (by the young generation collector) is also 2972 // marking the bit map. [the so-called allocate live policy.] 2973 // Because the implementation of bit map marking is not 2974 // robust wrt simultaneous marking of bits in the same word, 2975 // we need to make sure that there is no such interference 2976 // between concurrent such updates. 2977 2978 // already have locks 2979 assert_lock_strong(bitMapLock()); 2980 2981 verify_work_stacks_empty(); 2982 verify_overflow_empty(); 2983 bool result = false; 2984 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2985 result = do_marking_mt(); 2986 } else { 2987 result = do_marking_st(); 2988 } 2989 return result; 2990 } 2991 2992 // Forward decl 2993 class CMSConcMarkingTask; 2994 2995 class CMSConcMarkingParallelTerminator: public ParallelTaskTerminator { 2996 CMSCollector* _collector; 2997 CMSConcMarkingTask* _task; 2998 public: 2999 virtual void yield(); 3000 3001 // "n_threads" is the number of threads to be terminated. 3002 // "queue_set" is a set of work queues of other threads. 3003 // "collector" is the CMS collector associated with this task terminator. 3004 // "yield" indicates whether we need the gang as a whole to yield. 3005 CMSConcMarkingParallelTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3006 ParallelTaskTerminator(n_threads, queue_set), 3007 _collector(collector) { } 3008 3009 void set_task(CMSConcMarkingTask* task) { 3010 _task = task; 3011 } 3012 }; 3013 3014 class CMSConcMarkingOWSTTerminator: public OWSTTaskTerminator { 3015 CMSCollector* _collector; 3016 CMSConcMarkingTask* _task; 3017 public: 3018 virtual void yield(); 3019 3020 // "n_threads" is the number of threads to be terminated. 3021 // "queue_set" is a set of work queues of other threads. 3022 // "collector" is the CMS collector associated with this task terminator. 3023 // "yield" indicates whether we need the gang as a whole to yield. 3024 CMSConcMarkingOWSTTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3025 OWSTTaskTerminator(n_threads, queue_set), 3026 _collector(collector) { } 3027 3028 void set_task(CMSConcMarkingTask* task) { 3029 _task = task; 3030 } 3031 }; 3032 3033 class CMSConcMarkingTaskTerminator { 3034 private: 3035 ParallelTaskTerminator* _term; 3036 public: 3037 CMSConcMarkingTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) { 3038 if (UseOWSTTaskTerminator) { 3039 _term = new CMSConcMarkingOWSTTerminator(n_threads, queue_set, collector); 3040 } else { 3041 _term = new CMSConcMarkingParallelTerminator(n_threads, queue_set, collector); 3042 } 3043 } 3044 ~CMSConcMarkingTaskTerminator() { 3045 assert(_term != NULL, "Must not be NULL"); 3046 delete _term; 3047 } 3048 3049 void set_task(CMSConcMarkingTask* task); 3050 ParallelTaskTerminator* terminator() const { return _term; } 3051 }; 3052 3053 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3054 CMSConcMarkingTask* _task; 3055 public: 3056 bool should_exit_termination(); 3057 void set_task(CMSConcMarkingTask* task) { 3058 _task = task; 3059 } 3060 }; 3061 3062 // MT Concurrent Marking Task 3063 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3064 CMSCollector* _collector; 3065 uint _n_workers; // requested/desired # workers 3066 bool _result; 3067 CompactibleFreeListSpace* _cms_space; 3068 char _pad_front[64]; // padding to ... 3069 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3070 char _pad_back[64]; 3071 HeapWord* _restart_addr; 3072 3073 // Exposed here for yielding support 3074 Mutex* const _bit_map_lock; 3075 3076 // The per thread work queues, available here for stealing 3077 OopTaskQueueSet* _task_queues; 3078 3079 // Termination (and yielding) support 3080 CMSConcMarkingTaskTerminator _term; 3081 CMSConcMarkingTerminatorTerminator _term_term; 3082 3083 public: 3084 CMSConcMarkingTask(CMSCollector* collector, 3085 CompactibleFreeListSpace* cms_space, 3086 YieldingFlexibleWorkGang* workers, 3087 OopTaskQueueSet* task_queues): 3088 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3089 _collector(collector), 3090 _n_workers(0), 3091 _result(true), 3092 _cms_space(cms_space), 3093 _bit_map_lock(collector->bitMapLock()), 3094 _task_queues(task_queues), 3095 _term(_n_workers, task_queues, _collector) 3096 { 3097 _requested_size = _n_workers; 3098 _term.set_task(this); 3099 _term_term.set_task(this); 3100 _restart_addr = _global_finger = _cms_space->bottom(); 3101 } 3102 3103 3104 OopTaskQueueSet* task_queues() { return _task_queues; } 3105 3106 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3107 3108 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3109 3110 ParallelTaskTerminator* terminator() { return _term.terminator(); } 3111 3112 virtual void set_for_termination(uint active_workers) { 3113 terminator()->reset_for_reuse(active_workers); 3114 } 3115 3116 void work(uint worker_id); 3117 bool should_yield() { 3118 return ConcurrentMarkSweepThread::should_yield() 3119 && !_collector->foregroundGCIsActive(); 3120 } 3121 3122 virtual void coordinator_yield(); // stuff done by coordinator 3123 bool result() { return _result; } 3124 3125 void reset(HeapWord* ra) { 3126 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3127 _restart_addr = _global_finger = ra; 3128 _term.terminator()->reset_for_reuse(); 3129 } 3130 3131 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3132 OopTaskQueue* work_q); 3133 3134 private: 3135 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3136 void do_work_steal(int i); 3137 void bump_global_finger(HeapWord* f); 3138 }; 3139 3140 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3141 assert(_task != NULL, "Error"); 3142 return _task->yielding(); 3143 // Note that we do not need the disjunct || _task->should_yield() above 3144 // because we want terminating threads to yield only if the task 3145 // is already in the midst of yielding, which happens only after at least one 3146 // thread has yielded. 3147 } 3148 3149 void CMSConcMarkingParallelTerminator::yield() { 3150 if (_task->should_yield()) { 3151 _task->yield(); 3152 } else { 3153 ParallelTaskTerminator::yield(); 3154 } 3155 } 3156 3157 void CMSConcMarkingOWSTTerminator::yield() { 3158 if (_task->should_yield()) { 3159 _task->yield(); 3160 } else { 3161 OWSTTaskTerminator::yield(); 3162 } 3163 } 3164 3165 void CMSConcMarkingTaskTerminator::set_task(CMSConcMarkingTask* task) { 3166 if (UseOWSTTaskTerminator) { 3167 ((CMSConcMarkingOWSTTerminator*)_term)->set_task(task); 3168 } else { 3169 ((CMSConcMarkingParallelTerminator*)_term)->set_task(task); 3170 } 3171 } 3172 3173 //////////////////////////////////////////////////////////////// 3174 // Concurrent Marking Algorithm Sketch 3175 //////////////////////////////////////////////////////////////// 3176 // Until all tasks exhausted (both spaces): 3177 // -- claim next available chunk 3178 // -- bump global finger via CAS 3179 // -- find first object that starts in this chunk 3180 // and start scanning bitmap from that position 3181 // -- scan marked objects for oops 3182 // -- CAS-mark target, and if successful: 3183 // . if target oop is above global finger (volatile read) 3184 // nothing to do 3185 // . if target oop is in chunk and above local finger 3186 // then nothing to do 3187 // . else push on work-queue 3188 // -- Deal with possible overflow issues: 3189 // . local work-queue overflow causes stuff to be pushed on 3190 // global (common) overflow queue 3191 // . always first empty local work queue 3192 // . then get a batch of oops from global work queue if any 3193 // . then do work stealing 3194 // -- When all tasks claimed (both spaces) 3195 // and local work queue empty, 3196 // then in a loop do: 3197 // . check global overflow stack; steal a batch of oops and trace 3198 // . try to steal from other threads oif GOS is empty 3199 // . if neither is available, offer termination 3200 // -- Terminate and return result 3201 // 3202 void CMSConcMarkingTask::work(uint worker_id) { 3203 elapsedTimer _timer; 3204 ResourceMark rm; 3205 HandleMark hm; 3206 3207 DEBUG_ONLY(_collector->verify_overflow_empty();) 3208 3209 // Before we begin work, our work queue should be empty 3210 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3211 // Scan the bitmap covering _cms_space, tracing through grey objects. 3212 _timer.start(); 3213 do_scan_and_mark(worker_id, _cms_space); 3214 _timer.stop(); 3215 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3216 3217 // ... do work stealing 3218 _timer.reset(); 3219 _timer.start(); 3220 do_work_steal(worker_id); 3221 _timer.stop(); 3222 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3223 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3224 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3225 // Note that under the current task protocol, the 3226 // following assertion is true even of the spaces 3227 // expanded since the completion of the concurrent 3228 // marking. XXX This will likely change under a strict 3229 // ABORT semantics. 3230 // After perm removal the comparison was changed to 3231 // greater than or equal to from strictly greater than. 3232 // Before perm removal the highest address sweep would 3233 // have been at the end of perm gen but now is at the 3234 // end of the tenured gen. 3235 assert(_global_finger >= _cms_space->end(), 3236 "All tasks have been completed"); 3237 DEBUG_ONLY(_collector->verify_overflow_empty();) 3238 } 3239 3240 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3241 HeapWord* read = _global_finger; 3242 HeapWord* cur = read; 3243 while (f > read) { 3244 cur = read; 3245 read = Atomic::cmpxchg(f, &_global_finger, cur); 3246 if (cur == read) { 3247 // our cas succeeded 3248 assert(_global_finger >= f, "protocol consistency"); 3249 break; 3250 } 3251 } 3252 } 3253 3254 // This is really inefficient, and should be redone by 3255 // using (not yet available) block-read and -write interfaces to the 3256 // stack and the work_queue. XXX FIX ME !!! 3257 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3258 OopTaskQueue* work_q) { 3259 // Fast lock-free check 3260 if (ovflw_stk->length() == 0) { 3261 return false; 3262 } 3263 assert(work_q->size() == 0, "Shouldn't steal"); 3264 MutexLocker ml(ovflw_stk->par_lock(), 3265 Mutex::_no_safepoint_check_flag); 3266 // Grab up to 1/4 the size of the work queue 3267 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3268 (size_t)ParGCDesiredObjsFromOverflowList); 3269 num = MIN2(num, ovflw_stk->length()); 3270 for (int i = (int) num; i > 0; i--) { 3271 oop cur = ovflw_stk->pop(); 3272 assert(cur != NULL, "Counted wrong?"); 3273 work_q->push(cur); 3274 } 3275 return num > 0; 3276 } 3277 3278 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3279 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3280 int n_tasks = pst->n_tasks(); 3281 // We allow that there may be no tasks to do here because 3282 // we are restarting after a stack overflow. 3283 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3284 uint nth_task = 0; 3285 3286 HeapWord* aligned_start = sp->bottom(); 3287 if (sp->used_region().contains(_restart_addr)) { 3288 // Align down to a card boundary for the start of 0th task 3289 // for this space. 3290 aligned_start = align_down(_restart_addr, CardTable::card_size); 3291 } 3292 3293 size_t chunk_size = sp->marking_task_size(); 3294 while (pst->try_claim_task(/* reference */ nth_task)) { 3295 // Having claimed the nth task in this space, 3296 // compute the chunk that it corresponds to: 3297 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3298 aligned_start + (nth_task+1)*chunk_size); 3299 // Try and bump the global finger via a CAS; 3300 // note that we need to do the global finger bump 3301 // _before_ taking the intersection below, because 3302 // the task corresponding to that region will be 3303 // deemed done even if the used_region() expands 3304 // because of allocation -- as it almost certainly will 3305 // during start-up while the threads yield in the 3306 // closure below. 3307 HeapWord* finger = span.end(); 3308 bump_global_finger(finger); // atomically 3309 // There are null tasks here corresponding to chunks 3310 // beyond the "top" address of the space. 3311 span = span.intersection(sp->used_region()); 3312 if (!span.is_empty()) { // Non-null task 3313 HeapWord* prev_obj; 3314 assert(!span.contains(_restart_addr) || nth_task == 0, 3315 "Inconsistency"); 3316 if (nth_task == 0) { 3317 // For the 0th task, we'll not need to compute a block_start. 3318 if (span.contains(_restart_addr)) { 3319 // In the case of a restart because of stack overflow, 3320 // we might additionally skip a chunk prefix. 3321 prev_obj = _restart_addr; 3322 } else { 3323 prev_obj = span.start(); 3324 } 3325 } else { 3326 // We want to skip the first object because 3327 // the protocol is to scan any object in its entirety 3328 // that _starts_ in this span; a fortiori, any 3329 // object starting in an earlier span is scanned 3330 // as part of an earlier claimed task. 3331 // Below we use the "careful" version of block_start 3332 // so we do not try to navigate uninitialized objects. 3333 prev_obj = sp->block_start_careful(span.start()); 3334 // Below we use a variant of block_size that uses the 3335 // Printezis bits to avoid waiting for allocated 3336 // objects to become initialized/parsable. 3337 while (prev_obj < span.start()) { 3338 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3339 if (sz > 0) { 3340 prev_obj += sz; 3341 } else { 3342 // In this case we may end up doing a bit of redundant 3343 // scanning, but that appears unavoidable, short of 3344 // locking the free list locks; see bug 6324141. 3345 break; 3346 } 3347 } 3348 } 3349 if (prev_obj < span.end()) { 3350 MemRegion my_span = MemRegion(prev_obj, span.end()); 3351 // Do the marking work within a non-empty span -- 3352 // the last argument to the constructor indicates whether the 3353 // iteration should be incremental with periodic yields. 3354 ParMarkFromRootsClosure cl(this, _collector, my_span, 3355 &_collector->_markBitMap, 3356 work_queue(i), 3357 &_collector->_markStack); 3358 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3359 } // else nothing to do for this task 3360 } // else nothing to do for this task 3361 } 3362 // We'd be tempted to assert here that since there are no 3363 // more tasks left to claim in this space, the global_finger 3364 // must exceed space->top() and a fortiori space->end(). However, 3365 // that would not quite be correct because the bumping of 3366 // global_finger occurs strictly after the claiming of a task, 3367 // so by the time we reach here the global finger may not yet 3368 // have been bumped up by the thread that claimed the last 3369 // task. 3370 pst->all_tasks_completed(); 3371 } 3372 3373 class ParConcMarkingClosure: public MetadataVisitingOopIterateClosure { 3374 private: 3375 CMSCollector* _collector; 3376 CMSConcMarkingTask* _task; 3377 MemRegion _span; 3378 CMSBitMap* _bit_map; 3379 CMSMarkStack* _overflow_stack; 3380 OopTaskQueue* _work_queue; 3381 protected: 3382 DO_OOP_WORK_DEFN 3383 public: 3384 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3385 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3386 MetadataVisitingOopIterateClosure(collector->ref_processor()), 3387 _collector(collector), 3388 _task(task), 3389 _span(collector->_span), 3390 _bit_map(bit_map), 3391 _overflow_stack(overflow_stack), 3392 _work_queue(work_queue) 3393 { } 3394 virtual void do_oop(oop* p); 3395 virtual void do_oop(narrowOop* p); 3396 3397 void trim_queue(size_t max); 3398 void handle_stack_overflow(HeapWord* lost); 3399 void do_yield_check() { 3400 if (_task->should_yield()) { 3401 _task->yield(); 3402 } 3403 } 3404 }; 3405 3406 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3407 3408 // Grey object scanning during work stealing phase -- 3409 // the salient assumption here is that any references 3410 // that are in these stolen objects being scanned must 3411 // already have been initialized (else they would not have 3412 // been published), so we do not need to check for 3413 // uninitialized objects before pushing here. 3414 void ParConcMarkingClosure::do_oop(oop obj) { 3415 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3416 HeapWord* addr = (HeapWord*)obj; 3417 // Check if oop points into the CMS generation 3418 // and is not marked 3419 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3420 // a white object ... 3421 // If we manage to "claim" the object, by being the 3422 // first thread to mark it, then we push it on our 3423 // marking stack 3424 if (_bit_map->par_mark(addr)) { // ... now grey 3425 // push on work queue (grey set) 3426 bool simulate_overflow = false; 3427 NOT_PRODUCT( 3428 if (CMSMarkStackOverflowALot && 3429 _collector->simulate_overflow()) { 3430 // simulate a stack overflow 3431 simulate_overflow = true; 3432 } 3433 ) 3434 if (simulate_overflow || 3435 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3436 // stack overflow 3437 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3438 // We cannot assert that the overflow stack is full because 3439 // it may have been emptied since. 3440 assert(simulate_overflow || 3441 _work_queue->size() == _work_queue->max_elems(), 3442 "Else push should have succeeded"); 3443 handle_stack_overflow(addr); 3444 } 3445 } // Else, some other thread got there first 3446 do_yield_check(); 3447 } 3448 } 3449 3450 void ParConcMarkingClosure::trim_queue(size_t max) { 3451 while (_work_queue->size() > max) { 3452 oop new_oop; 3453 if (_work_queue->pop_local(new_oop)) { 3454 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3455 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3456 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3457 new_oop->oop_iterate(this); // do_oop() above 3458 do_yield_check(); 3459 } 3460 } 3461 } 3462 3463 // Upon stack overflow, we discard (part of) the stack, 3464 // remembering the least address amongst those discarded 3465 // in CMSCollector's _restart_address. 3466 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3467 // We need to do this under a mutex to prevent other 3468 // workers from interfering with the work done below. 3469 MutexLocker ml(_overflow_stack->par_lock(), 3470 Mutex::_no_safepoint_check_flag); 3471 // Remember the least grey address discarded 3472 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3473 _collector->lower_restart_addr(ra); 3474 _overflow_stack->reset(); // discard stack contents 3475 _overflow_stack->expand(); // expand the stack if possible 3476 } 3477 3478 3479 void CMSConcMarkingTask::do_work_steal(int i) { 3480 OopTaskQueue* work_q = work_queue(i); 3481 oop obj_to_scan; 3482 CMSBitMap* bm = &(_collector->_markBitMap); 3483 CMSMarkStack* ovflw = &(_collector->_markStack); 3484 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3485 while (true) { 3486 cl.trim_queue(0); 3487 assert(work_q->size() == 0, "Should have been emptied above"); 3488 if (get_work_from_overflow_stack(ovflw, work_q)) { 3489 // Can't assert below because the work obtained from the 3490 // overflow stack may already have been stolen from us. 3491 // assert(work_q->size() > 0, "Work from overflow stack"); 3492 continue; 3493 } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 3494 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3495 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3496 obj_to_scan->oop_iterate(&cl); 3497 } else if (terminator()->offer_termination(&_term_term)) { 3498 assert(work_q->size() == 0, "Impossible!"); 3499 break; 3500 } else if (yielding() || should_yield()) { 3501 yield(); 3502 } 3503 } 3504 } 3505 3506 // This is run by the CMS (coordinator) thread. 3507 void CMSConcMarkingTask::coordinator_yield() { 3508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3509 "CMS thread should hold CMS token"); 3510 // First give up the locks, then yield, then re-lock 3511 // We should probably use a constructor/destructor idiom to 3512 // do this unlock/lock or modify the MutexUnlocker class to 3513 // serve our purpose. XXX 3514 assert_lock_strong(_bit_map_lock); 3515 _bit_map_lock->unlock(); 3516 ConcurrentMarkSweepThread::desynchronize(true); 3517 _collector->stopTimer(); 3518 _collector->incrementYields(); 3519 3520 // It is possible for whichever thread initiated the yield request 3521 // not to get a chance to wake up and take the bitmap lock between 3522 // this thread releasing it and reacquiring it. So, while the 3523 // should_yield() flag is on, let's sleep for a bit to give the 3524 // other thread a chance to wake up. The limit imposed on the number 3525 // of iterations is defensive, to avoid any unforseen circumstances 3526 // putting us into an infinite loop. Since it's always been this 3527 // (coordinator_yield()) method that was observed to cause the 3528 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3529 // which is by default non-zero. For the other seven methods that 3530 // also perform the yield operation, as are using a different 3531 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3532 // can enable the sleeping for those methods too, if necessary. 3533 // See 6442774. 3534 // 3535 // We really need to reconsider the synchronization between the GC 3536 // thread and the yield-requesting threads in the future and we 3537 // should really use wait/notify, which is the recommended 3538 // way of doing this type of interaction. Additionally, we should 3539 // consolidate the eight methods that do the yield operation and they 3540 // are almost identical into one for better maintainability and 3541 // readability. See 6445193. 3542 // 3543 // Tony 2006.06.29 3544 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3545 ConcurrentMarkSweepThread::should_yield() && 3546 !CMSCollector::foregroundGCIsActive(); ++i) { 3547 os::sleep(Thread::current(), 1, false); 3548 } 3549 3550 ConcurrentMarkSweepThread::synchronize(true); 3551 _bit_map_lock->lock_without_safepoint_check(); 3552 _collector->startTimer(); 3553 } 3554 3555 bool CMSCollector::do_marking_mt() { 3556 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3557 uint num_workers = WorkerPolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3558 conc_workers()->active_workers(), 3559 Threads::number_of_non_daemon_threads()); 3560 num_workers = conc_workers()->update_active_workers(num_workers); 3561 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3562 3563 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3564 3565 CMSConcMarkingTask tsk(this, 3566 cms_space, 3567 conc_workers(), 3568 task_queues()); 3569 3570 // Since the actual number of workers we get may be different 3571 // from the number we requested above, do we need to do anything different 3572 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3573 // class?? XXX 3574 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3575 3576 // Refs discovery is already non-atomic. 3577 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3578 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3579 conc_workers()->start_task(&tsk); 3580 while (tsk.yielded()) { 3581 tsk.coordinator_yield(); 3582 conc_workers()->continue_task(&tsk); 3583 } 3584 // If the task was aborted, _restart_addr will be non-NULL 3585 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3586 while (_restart_addr != NULL) { 3587 // XXX For now we do not make use of ABORTED state and have not 3588 // yet implemented the right abort semantics (even in the original 3589 // single-threaded CMS case). That needs some more investigation 3590 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3591 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3592 // If _restart_addr is non-NULL, a marking stack overflow 3593 // occurred; we need to do a fresh marking iteration from the 3594 // indicated restart address. 3595 if (_foregroundGCIsActive) { 3596 // We may be running into repeated stack overflows, having 3597 // reached the limit of the stack size, while making very 3598 // slow forward progress. It may be best to bail out and 3599 // let the foreground collector do its job. 3600 // Clear _restart_addr, so that foreground GC 3601 // works from scratch. This avoids the headache of 3602 // a "rescan" which would otherwise be needed because 3603 // of the dirty mod union table & card table. 3604 _restart_addr = NULL; 3605 return false; 3606 } 3607 // Adjust the task to restart from _restart_addr 3608 tsk.reset(_restart_addr); 3609 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3610 _restart_addr); 3611 _restart_addr = NULL; 3612 // Get the workers going again 3613 conc_workers()->start_task(&tsk); 3614 while (tsk.yielded()) { 3615 tsk.coordinator_yield(); 3616 conc_workers()->continue_task(&tsk); 3617 } 3618 } 3619 assert(tsk.completed(), "Inconsistency"); 3620 assert(tsk.result() == true, "Inconsistency"); 3621 return true; 3622 } 3623 3624 bool CMSCollector::do_marking_st() { 3625 ResourceMark rm; 3626 HandleMark hm; 3627 3628 // Temporarily make refs discovery single threaded (non-MT) 3629 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3630 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3631 &_markStack, CMSYield); 3632 // the last argument to iterate indicates whether the iteration 3633 // should be incremental with periodic yields. 3634 _markBitMap.iterate(&markFromRootsClosure); 3635 // If _restart_addr is non-NULL, a marking stack overflow 3636 // occurred; we need to do a fresh iteration from the 3637 // indicated restart address. 3638 while (_restart_addr != NULL) { 3639 if (_foregroundGCIsActive) { 3640 // We may be running into repeated stack overflows, having 3641 // reached the limit of the stack size, while making very 3642 // slow forward progress. It may be best to bail out and 3643 // let the foreground collector do its job. 3644 // Clear _restart_addr, so that foreground GC 3645 // works from scratch. This avoids the headache of 3646 // a "rescan" which would otherwise be needed because 3647 // of the dirty mod union table & card table. 3648 _restart_addr = NULL; 3649 return false; // indicating failure to complete marking 3650 } 3651 // Deal with stack overflow: 3652 // we restart marking from _restart_addr 3653 HeapWord* ra = _restart_addr; 3654 markFromRootsClosure.reset(ra); 3655 _restart_addr = NULL; 3656 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3657 } 3658 return true; 3659 } 3660 3661 void CMSCollector::preclean() { 3662 check_correct_thread_executing(); 3663 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3664 verify_work_stacks_empty(); 3665 verify_overflow_empty(); 3666 _abort_preclean = false; 3667 if (CMSPrecleaningEnabled) { 3668 if (!CMSEdenChunksRecordAlways) { 3669 _eden_chunk_index = 0; 3670 } 3671 size_t used = get_eden_used(); 3672 size_t capacity = get_eden_capacity(); 3673 // Don't start sampling unless we will get sufficiently 3674 // many samples. 3675 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3676 * CMSScheduleRemarkEdenPenetration)) { 3677 _start_sampling = true; 3678 } else { 3679 _start_sampling = false; 3680 } 3681 GCTraceCPUTime tcpu; 3682 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3683 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3684 } 3685 CMSTokenSync x(true); // is cms thread 3686 if (CMSPrecleaningEnabled) { 3687 sample_eden(); 3688 _collectorState = AbortablePreclean; 3689 } else { 3690 _collectorState = FinalMarking; 3691 } 3692 verify_work_stacks_empty(); 3693 verify_overflow_empty(); 3694 } 3695 3696 // Try and schedule the remark such that young gen 3697 // occupancy is CMSScheduleRemarkEdenPenetration %. 3698 void CMSCollector::abortable_preclean() { 3699 check_correct_thread_executing(); 3700 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3701 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3702 3703 // If Eden's current occupancy is below this threshold, 3704 // immediately schedule the remark; else preclean 3705 // past the next scavenge in an effort to 3706 // schedule the pause as described above. By choosing 3707 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3708 // we will never do an actual abortable preclean cycle. 3709 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3710 GCTraceCPUTime tcpu; 3711 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3712 // We need more smarts in the abortable preclean 3713 // loop below to deal with cases where allocation 3714 // in young gen is very very slow, and our precleaning 3715 // is running a losing race against a horde of 3716 // mutators intent on flooding us with CMS updates 3717 // (dirty cards). 3718 // One, admittedly dumb, strategy is to give up 3719 // after a certain number of abortable precleaning loops 3720 // or after a certain maximum time. We want to make 3721 // this smarter in the next iteration. 3722 // XXX FIX ME!!! YSR 3723 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3724 while (!(should_abort_preclean() || 3725 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3726 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3727 cumworkdone += workdone; 3728 loops++; 3729 // Voluntarily terminate abortable preclean phase if we have 3730 // been at it for too long. 3731 if ((CMSMaxAbortablePrecleanLoops != 0) && 3732 loops >= CMSMaxAbortablePrecleanLoops) { 3733 log_debug(gc)(" CMS: abort preclean due to loops "); 3734 break; 3735 } 3736 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3737 log_debug(gc)(" CMS: abort preclean due to time "); 3738 break; 3739 } 3740 // If we are doing little work each iteration, we should 3741 // take a short break. 3742 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3743 // Sleep for some time, waiting for work to accumulate 3744 stopTimer(); 3745 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3746 startTimer(); 3747 waited++; 3748 } 3749 } 3750 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3751 loops, waited, cumworkdone); 3752 } 3753 CMSTokenSync x(true); // is cms thread 3754 if (_collectorState != Idling) { 3755 assert(_collectorState == AbortablePreclean, 3756 "Spontaneous state transition?"); 3757 _collectorState = FinalMarking; 3758 } // Else, a foreground collection completed this CMS cycle. 3759 return; 3760 } 3761 3762 // Respond to an Eden sampling opportunity 3763 void CMSCollector::sample_eden() { 3764 // Make sure a young gc cannot sneak in between our 3765 // reading and recording of a sample. 3766 assert(Thread::current()->is_ConcurrentGC_thread(), 3767 "Only the cms thread may collect Eden samples"); 3768 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3769 "Should collect samples while holding CMS token"); 3770 if (!_start_sampling) { 3771 return; 3772 } 3773 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3774 // is populated by the young generation. 3775 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3776 if (_eden_chunk_index < _eden_chunk_capacity) { 3777 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3778 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3779 "Unexpected state of Eden"); 3780 // We'd like to check that what we just sampled is an oop-start address; 3781 // however, we cannot do that here since the object may not yet have been 3782 // initialized. So we'll instead do the check when we _use_ this sample 3783 // later. 3784 if (_eden_chunk_index == 0 || 3785 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3786 _eden_chunk_array[_eden_chunk_index-1]) 3787 >= CMSSamplingGrain)) { 3788 _eden_chunk_index++; // commit sample 3789 } 3790 } 3791 } 3792 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3793 size_t used = get_eden_used(); 3794 size_t capacity = get_eden_capacity(); 3795 assert(used <= capacity, "Unexpected state of Eden"); 3796 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3797 _abort_preclean = true; 3798 } 3799 } 3800 } 3801 3802 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3803 assert(_collectorState == Precleaning || 3804 _collectorState == AbortablePreclean, "incorrect state"); 3805 ResourceMark rm; 3806 HandleMark hm; 3807 3808 // Precleaning is currently not MT but the reference processor 3809 // may be set for MT. Disable it temporarily here. 3810 ReferenceProcessor* rp = ref_processor(); 3811 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3812 3813 // Do one pass of scrubbing the discovered reference lists 3814 // to remove any reference objects with strongly-reachable 3815 // referents. 3816 if (clean_refs) { 3817 CMSPrecleanRefsYieldClosure yield_cl(this); 3818 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3819 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3820 &_markStack, true /* preclean */); 3821 CMSDrainMarkingStackClosure complete_trace(this, 3822 _span, &_markBitMap, &_markStack, 3823 &keep_alive, true /* preclean */); 3824 3825 // We don't want this step to interfere with a young 3826 // collection because we don't want to take CPU 3827 // or memory bandwidth away from the young GC threads 3828 // (which may be as many as there are CPUs). 3829 // Note that we don't need to protect ourselves from 3830 // interference with mutators because they can't 3831 // manipulate the discovered reference lists nor affect 3832 // the computed reachability of the referents, the 3833 // only properties manipulated by the precleaning 3834 // of these reference lists. 3835 stopTimer(); 3836 CMSTokenSyncWithLocks x(true /* is cms thread */, 3837 bitMapLock()); 3838 startTimer(); 3839 sample_eden(); 3840 3841 // The following will yield to allow foreground 3842 // collection to proceed promptly. XXX YSR: 3843 // The code in this method may need further 3844 // tweaking for better performance and some restructuring 3845 // for cleaner interfaces. 3846 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3847 rp->preclean_discovered_references( 3848 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3849 gc_timer); 3850 } 3851 3852 if (clean_survivor) { // preclean the active survivor space(s) 3853 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3854 &_markBitMap, &_modUnionTable, 3855 &_markStack, true /* precleaning phase */); 3856 stopTimer(); 3857 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3858 bitMapLock()); 3859 startTimer(); 3860 unsigned int before_count = 3861 CMSHeap::heap()->total_collections(); 3862 SurvivorSpacePrecleanClosure 3863 sss_cl(this, _span, &_markBitMap, &_markStack, 3864 &pam_cl, before_count, CMSYield); 3865 _young_gen->from()->object_iterate_careful(&sss_cl); 3866 _young_gen->to()->object_iterate_careful(&sss_cl); 3867 } 3868 MarkRefsIntoAndScanClosure 3869 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3870 &_markStack, this, CMSYield, 3871 true /* precleaning phase */); 3872 // CAUTION: The following closure has persistent state that may need to 3873 // be reset upon a decrease in the sequence of addresses it 3874 // processes. 3875 ScanMarkedObjectsAgainCarefullyClosure 3876 smoac_cl(this, _span, 3877 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3878 3879 // Preclean dirty cards in ModUnionTable and CardTable using 3880 // appropriate convergence criterion; 3881 // repeat CMSPrecleanIter times unless we find that 3882 // we are losing. 3883 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3884 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3885 "Bad convergence multiplier"); 3886 assert(CMSPrecleanThreshold >= 100, 3887 "Unreasonably low CMSPrecleanThreshold"); 3888 3889 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3890 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3891 numIter < CMSPrecleanIter; 3892 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3893 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3894 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3895 // Either there are very few dirty cards, so re-mark 3896 // pause will be small anyway, or our pre-cleaning isn't 3897 // that much faster than the rate at which cards are being 3898 // dirtied, so we might as well stop and re-mark since 3899 // precleaning won't improve our re-mark time by much. 3900 if (curNumCards <= CMSPrecleanThreshold || 3901 (numIter > 0 && 3902 (curNumCards * CMSPrecleanDenominator > 3903 lastNumCards * CMSPrecleanNumerator))) { 3904 numIter++; 3905 cumNumCards += curNumCards; 3906 break; 3907 } 3908 } 3909 3910 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3911 3912 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3913 cumNumCards += curNumCards; 3914 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3915 curNumCards, cumNumCards, numIter); 3916 return cumNumCards; // as a measure of useful work done 3917 } 3918 3919 // PRECLEANING NOTES: 3920 // Precleaning involves: 3921 // . reading the bits of the modUnionTable and clearing the set bits. 3922 // . For the cards corresponding to the set bits, we scan the 3923 // objects on those cards. This means we need the free_list_lock 3924 // so that we can safely iterate over the CMS space when scanning 3925 // for oops. 3926 // . When we scan the objects, we'll be both reading and setting 3927 // marks in the marking bit map, so we'll need the marking bit map. 3928 // . For protecting _collector_state transitions, we take the CGC_lock. 3929 // Note that any races in the reading of of card table entries by the 3930 // CMS thread on the one hand and the clearing of those entries by the 3931 // VM thread or the setting of those entries by the mutator threads on the 3932 // other are quite benign. However, for efficiency it makes sense to keep 3933 // the VM thread from racing with the CMS thread while the latter is 3934 // dirty card info to the modUnionTable. We therefore also use the 3935 // CGC_lock to protect the reading of the card table and the mod union 3936 // table by the CM thread. 3937 // . We run concurrently with mutator updates, so scanning 3938 // needs to be done carefully -- we should not try to scan 3939 // potentially uninitialized objects. 3940 // 3941 // Locking strategy: While holding the CGC_lock, we scan over and 3942 // reset a maximal dirty range of the mod union / card tables, then lock 3943 // the free_list_lock and bitmap lock to do a full marking, then 3944 // release these locks; and repeat the cycle. This allows for a 3945 // certain amount of fairness in the sharing of these locks between 3946 // the CMS collector on the one hand, and the VM thread and the 3947 // mutators on the other. 3948 3949 // NOTE: preclean_mod_union_table() and preclean_card_table() 3950 // further below are largely identical; if you need to modify 3951 // one of these methods, please check the other method too. 3952 3953 size_t CMSCollector::preclean_mod_union_table( 3954 ConcurrentMarkSweepGeneration* old_gen, 3955 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3956 verify_work_stacks_empty(); 3957 verify_overflow_empty(); 3958 3959 // strategy: starting with the first card, accumulate contiguous 3960 // ranges of dirty cards; clear these cards, then scan the region 3961 // covered by these cards. 3962 3963 // Since all of the MUT is committed ahead, we can just use 3964 // that, in case the generations expand while we are precleaning. 3965 // It might also be fine to just use the committed part of the 3966 // generation, but we might potentially miss cards when the 3967 // generation is rapidly expanding while we are in the midst 3968 // of precleaning. 3969 HeapWord* startAddr = old_gen->reserved().start(); 3970 HeapWord* endAddr = old_gen->reserved().end(); 3971 3972 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3973 3974 size_t numDirtyCards, cumNumDirtyCards; 3975 HeapWord *nextAddr, *lastAddr; 3976 for (cumNumDirtyCards = numDirtyCards = 0, 3977 nextAddr = lastAddr = startAddr; 3978 nextAddr < endAddr; 3979 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3980 3981 ResourceMark rm; 3982 HandleMark hm; 3983 3984 MemRegion dirtyRegion; 3985 { 3986 stopTimer(); 3987 // Potential yield point 3988 CMSTokenSync ts(true); 3989 startTimer(); 3990 sample_eden(); 3991 // Get dirty region starting at nextOffset (inclusive), 3992 // simultaneously clearing it. 3993 dirtyRegion = 3994 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3995 assert(dirtyRegion.start() >= nextAddr, 3996 "returned region inconsistent?"); 3997 } 3998 // Remember where the next search should begin. 3999 // The returned region (if non-empty) is a right open interval, 4000 // so lastOffset is obtained from the right end of that 4001 // interval. 4002 lastAddr = dirtyRegion.end(); 4003 // Should do something more transparent and less hacky XXX 4004 numDirtyCards = 4005 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4006 4007 // We'll scan the cards in the dirty region (with periodic 4008 // yields for foreground GC as needed). 4009 if (!dirtyRegion.is_empty()) { 4010 assert(numDirtyCards > 0, "consistency check"); 4011 HeapWord* stop_point = NULL; 4012 stopTimer(); 4013 // Potential yield point 4014 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4015 bitMapLock()); 4016 startTimer(); 4017 { 4018 verify_work_stacks_empty(); 4019 verify_overflow_empty(); 4020 sample_eden(); 4021 stop_point = 4022 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4023 } 4024 if (stop_point != NULL) { 4025 // The careful iteration stopped early either because it found an 4026 // uninitialized object, or because we were in the midst of an 4027 // "abortable preclean", which should now be aborted. Redirty 4028 // the bits corresponding to the partially-scanned or unscanned 4029 // cards. We'll either restart at the next block boundary or 4030 // abort the preclean. 4031 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4032 "Should only be AbortablePreclean."); 4033 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4034 if (should_abort_preclean()) { 4035 break; // out of preclean loop 4036 } else { 4037 // Compute the next address at which preclean should pick up; 4038 // might need bitMapLock in order to read P-bits. 4039 lastAddr = next_card_start_after_block(stop_point); 4040 } 4041 } 4042 } else { 4043 assert(lastAddr == endAddr, "consistency check"); 4044 assert(numDirtyCards == 0, "consistency check"); 4045 break; 4046 } 4047 } 4048 verify_work_stacks_empty(); 4049 verify_overflow_empty(); 4050 return cumNumDirtyCards; 4051 } 4052 4053 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4054 // below are largely identical; if you need to modify 4055 // one of these methods, please check the other method too. 4056 4057 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4058 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4059 // strategy: it's similar to precleamModUnionTable above, in that 4060 // we accumulate contiguous ranges of dirty cards, mark these cards 4061 // precleaned, then scan the region covered by these cards. 4062 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4063 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4064 4065 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4066 4067 size_t numDirtyCards, cumNumDirtyCards; 4068 HeapWord *lastAddr, *nextAddr; 4069 4070 for (cumNumDirtyCards = numDirtyCards = 0, 4071 nextAddr = lastAddr = startAddr; 4072 nextAddr < endAddr; 4073 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4074 4075 ResourceMark rm; 4076 HandleMark hm; 4077 4078 MemRegion dirtyRegion; 4079 { 4080 // See comments in "Precleaning notes" above on why we 4081 // do this locking. XXX Could the locking overheads be 4082 // too high when dirty cards are sparse? [I don't think so.] 4083 stopTimer(); 4084 CMSTokenSync x(true); // is cms thread 4085 startTimer(); 4086 sample_eden(); 4087 // Get and clear dirty region from card table 4088 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4089 true, 4090 CardTable::precleaned_card_val()); 4091 4092 assert(dirtyRegion.start() >= nextAddr, 4093 "returned region inconsistent?"); 4094 } 4095 lastAddr = dirtyRegion.end(); 4096 numDirtyCards = 4097 dirtyRegion.word_size()/CardTable::card_size_in_words; 4098 4099 if (!dirtyRegion.is_empty()) { 4100 stopTimer(); 4101 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4102 startTimer(); 4103 sample_eden(); 4104 verify_work_stacks_empty(); 4105 verify_overflow_empty(); 4106 HeapWord* stop_point = 4107 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4108 if (stop_point != NULL) { 4109 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4110 "Should only be AbortablePreclean."); 4111 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4112 if (should_abort_preclean()) { 4113 break; // out of preclean loop 4114 } else { 4115 // Compute the next address at which preclean should pick up. 4116 lastAddr = next_card_start_after_block(stop_point); 4117 } 4118 } 4119 } else { 4120 break; 4121 } 4122 } 4123 verify_work_stacks_empty(); 4124 verify_overflow_empty(); 4125 return cumNumDirtyCards; 4126 } 4127 4128 class PrecleanCLDClosure : public CLDClosure { 4129 MetadataVisitingOopsInGenClosure* _cm_closure; 4130 public: 4131 PrecleanCLDClosure(MetadataVisitingOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4132 void do_cld(ClassLoaderData* cld) { 4133 if (cld->has_accumulated_modified_oops()) { 4134 cld->clear_accumulated_modified_oops(); 4135 4136 _cm_closure->do_cld(cld); 4137 } 4138 } 4139 }; 4140 4141 // The freelist lock is needed to prevent asserts, is it really needed? 4142 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4143 // Needed to walk CLDG 4144 MutexLocker ml(ClassLoaderDataGraph_lock); 4145 4146 cl->set_freelistLock(freelistLock); 4147 4148 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4149 4150 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4151 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4152 PrecleanCLDClosure preclean_closure(cl); 4153 ClassLoaderDataGraph::cld_do(&preclean_closure); 4154 4155 verify_work_stacks_empty(); 4156 verify_overflow_empty(); 4157 } 4158 4159 void CMSCollector::checkpointRootsFinal() { 4160 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4161 check_correct_thread_executing(); 4162 // world is stopped at this checkpoint 4163 assert(SafepointSynchronize::is_at_safepoint(), 4164 "world should be stopped"); 4165 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4166 4167 verify_work_stacks_empty(); 4168 verify_overflow_empty(); 4169 4170 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4171 _young_gen->used() / K, _young_gen->capacity() / K); 4172 { 4173 if (CMSScavengeBeforeRemark) { 4174 CMSHeap* heap = CMSHeap::heap(); 4175 // Temporarily set flag to false, GCH->do_collection will 4176 // expect it to be false and set to true 4177 FlagSetting fl(heap->_is_gc_active, false); 4178 4179 heap->do_collection(true, // full (i.e. force, see below) 4180 false, // !clear_all_soft_refs 4181 0, // size 4182 false, // is_tlab 4183 GenCollectedHeap::YoungGen // type 4184 ); 4185 } 4186 FreelistLocker x(this); 4187 MutexLocker y(bitMapLock(), 4188 Mutex::_no_safepoint_check_flag); 4189 checkpointRootsFinalWork(); 4190 _cmsGen->cmsSpace()->recalculate_used_stable(); 4191 } 4192 verify_work_stacks_empty(); 4193 verify_overflow_empty(); 4194 } 4195 4196 void CMSCollector::checkpointRootsFinalWork() { 4197 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4198 4199 assert(haveFreelistLocks(), "must have free list locks"); 4200 assert_lock_strong(bitMapLock()); 4201 4202 ResourceMark rm; 4203 HandleMark hm; 4204 4205 CMSHeap* heap = CMSHeap::heap(); 4206 4207 assert(haveFreelistLocks(), "must have free list locks"); 4208 assert_lock_strong(bitMapLock()); 4209 4210 // We might assume that we need not fill TLAB's when 4211 // CMSScavengeBeforeRemark is set, because we may have just done 4212 // a scavenge which would have filled all TLAB's -- and besides 4213 // Eden would be empty. This however may not always be the case -- 4214 // for instance although we asked for a scavenge, it may not have 4215 // happened because of a JNI critical section. We probably need 4216 // a policy for deciding whether we can in that case wait until 4217 // the critical section releases and then do the remark following 4218 // the scavenge, and skip it here. In the absence of that policy, 4219 // or of an indication of whether the scavenge did indeed occur, 4220 // we cannot rely on TLAB's having been filled and must do 4221 // so here just in case a scavenge did not happen. 4222 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4223 // Update the saved marks which may affect the root scans. 4224 heap->save_marks(); 4225 4226 print_eden_and_survivor_chunk_arrays(); 4227 4228 { 4229 #if COMPILER2_OR_JVMCI 4230 DerivedPointerTableDeactivate dpt_deact; 4231 #endif 4232 4233 // Note on the role of the mod union table: 4234 // Since the marker in "markFromRoots" marks concurrently with 4235 // mutators, it is possible for some reachable objects not to have been 4236 // scanned. For instance, an only reference to an object A was 4237 // placed in object B after the marker scanned B. Unless B is rescanned, 4238 // A would be collected. Such updates to references in marked objects 4239 // are detected via the mod union table which is the set of all cards 4240 // dirtied since the first checkpoint in this GC cycle and prior to 4241 // the most recent young generation GC, minus those cleaned up by the 4242 // concurrent precleaning. 4243 if (CMSParallelRemarkEnabled) { 4244 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4245 do_remark_parallel(); 4246 } else { 4247 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4248 do_remark_non_parallel(); 4249 } 4250 } 4251 verify_work_stacks_empty(); 4252 verify_overflow_empty(); 4253 4254 { 4255 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4256 refProcessingWork(); 4257 } 4258 verify_work_stacks_empty(); 4259 verify_overflow_empty(); 4260 4261 if (should_unload_classes()) { 4262 heap->prune_scavengable_nmethods(); 4263 } 4264 4265 // If we encountered any (marking stack / work queue) overflow 4266 // events during the current CMS cycle, take appropriate 4267 // remedial measures, where possible, so as to try and avoid 4268 // recurrence of that condition. 4269 assert(_markStack.isEmpty(), "No grey objects"); 4270 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4271 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4272 if (ser_ovflw > 0) { 4273 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4274 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4275 _markStack.expand(); 4276 _ser_pmc_remark_ovflw = 0; 4277 _ser_pmc_preclean_ovflw = 0; 4278 _ser_kac_preclean_ovflw = 0; 4279 _ser_kac_ovflw = 0; 4280 } 4281 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4282 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4283 _par_pmc_remark_ovflw, _par_kac_ovflw); 4284 _par_pmc_remark_ovflw = 0; 4285 _par_kac_ovflw = 0; 4286 } 4287 if (_markStack._hit_limit > 0) { 4288 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4289 _markStack._hit_limit); 4290 } 4291 if (_markStack._failed_double > 0) { 4292 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4293 _markStack._failed_double, _markStack.capacity()); 4294 } 4295 _markStack._hit_limit = 0; 4296 _markStack._failed_double = 0; 4297 4298 if ((VerifyAfterGC || VerifyDuringGC) && 4299 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4300 verify_after_remark(); 4301 } 4302 4303 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4304 4305 // Change under the freelistLocks. 4306 _collectorState = Sweeping; 4307 // Call isAllClear() under bitMapLock 4308 assert(_modUnionTable.isAllClear(), 4309 "Should be clear by end of the final marking"); 4310 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4311 "Should be clear by end of the final marking"); 4312 } 4313 4314 void CMSParInitialMarkTask::work(uint worker_id) { 4315 elapsedTimer _timer; 4316 ResourceMark rm; 4317 HandleMark hm; 4318 4319 // ---------- scan from roots -------------- 4320 _timer.start(); 4321 CMSHeap* heap = CMSHeap::heap(); 4322 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4323 4324 // ---------- young gen roots -------------- 4325 { 4326 work_on_young_gen_roots(&par_mri_cl); 4327 _timer.stop(); 4328 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4329 } 4330 4331 // ---------- remaining roots -------------- 4332 _timer.reset(); 4333 _timer.start(); 4334 4335 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4336 4337 heap->cms_process_roots(_strong_roots_scope, 4338 false, // yg was scanned above 4339 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4340 _collector->should_unload_classes(), 4341 &par_mri_cl, 4342 &cld_closure); 4343 4344 assert(_collector->should_unload_classes() 4345 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4346 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4347 _timer.stop(); 4348 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4349 } 4350 4351 // Parallel remark task 4352 class CMSParRemarkTask: public CMSParMarkTask { 4353 CompactibleFreeListSpace* _cms_space; 4354 4355 // The per-thread work queues, available here for stealing. 4356 OopTaskQueueSet* _task_queues; 4357 TaskTerminator _term; 4358 StrongRootsScope* _strong_roots_scope; 4359 4360 public: 4361 // A value of 0 passed to n_workers will cause the number of 4362 // workers to be taken from the active workers in the work gang. 4363 CMSParRemarkTask(CMSCollector* collector, 4364 CompactibleFreeListSpace* cms_space, 4365 uint n_workers, WorkGang* workers, 4366 OopTaskQueueSet* task_queues, 4367 StrongRootsScope* strong_roots_scope): 4368 CMSParMarkTask("Rescan roots and grey objects in parallel", 4369 collector, n_workers), 4370 _cms_space(cms_space), 4371 _task_queues(task_queues), 4372 _term(n_workers, task_queues), 4373 _strong_roots_scope(strong_roots_scope) { } 4374 4375 OopTaskQueueSet* task_queues() { return _task_queues; } 4376 4377 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4378 4379 ParallelTaskTerminator* terminator() { return _term.terminator(); } 4380 uint n_workers() { return _n_workers; } 4381 4382 void work(uint worker_id); 4383 4384 private: 4385 // ... of dirty cards in old space 4386 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4387 ParMarkRefsIntoAndScanClosure* cl); 4388 4389 // ... work stealing for the above 4390 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4391 }; 4392 4393 class RemarkCLDClosure : public CLDClosure { 4394 CLDToOopClosure _cm_closure; 4395 public: 4396 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4397 void do_cld(ClassLoaderData* cld) { 4398 // Check if we have modified any oops in the CLD during the concurrent marking. 4399 if (cld->has_accumulated_modified_oops()) { 4400 cld->clear_accumulated_modified_oops(); 4401 4402 // We could have transfered the current modified marks to the accumulated marks, 4403 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4404 } else if (cld->has_modified_oops()) { 4405 // Don't clear anything, this info is needed by the next young collection. 4406 } else { 4407 // No modified oops in the ClassLoaderData. 4408 return; 4409 } 4410 4411 // The klass has modified fields, need to scan the klass. 4412 _cm_closure.do_cld(cld); 4413 } 4414 }; 4415 4416 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4417 ParNewGeneration* young_gen = _collector->_young_gen; 4418 ContiguousSpace* eden_space = young_gen->eden(); 4419 ContiguousSpace* from_space = young_gen->from(); 4420 ContiguousSpace* to_space = young_gen->to(); 4421 4422 HeapWord** eca = _collector->_eden_chunk_array; 4423 size_t ect = _collector->_eden_chunk_index; 4424 HeapWord** sca = _collector->_survivor_chunk_array; 4425 size_t sct = _collector->_survivor_chunk_index; 4426 4427 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4428 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4429 4430 do_young_space_rescan(cl, to_space, NULL, 0); 4431 do_young_space_rescan(cl, from_space, sca, sct); 4432 do_young_space_rescan(cl, eden_space, eca, ect); 4433 } 4434 4435 // work_queue(i) is passed to the closure 4436 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4437 // also is passed to do_dirty_card_rescan_tasks() and to 4438 // do_work_steal() to select the i-th task_queue. 4439 4440 void CMSParRemarkTask::work(uint worker_id) { 4441 elapsedTimer _timer; 4442 ResourceMark rm; 4443 HandleMark hm; 4444 4445 // ---------- rescan from roots -------------- 4446 _timer.start(); 4447 CMSHeap* heap = CMSHeap::heap(); 4448 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4449 _collector->_span, _collector->ref_processor(), 4450 &(_collector->_markBitMap), 4451 work_queue(worker_id)); 4452 4453 // Rescan young gen roots first since these are likely 4454 // coarsely partitioned and may, on that account, constitute 4455 // the critical path; thus, it's best to start off that 4456 // work first. 4457 // ---------- young gen roots -------------- 4458 { 4459 work_on_young_gen_roots(&par_mrias_cl); 4460 _timer.stop(); 4461 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4462 } 4463 4464 // ---------- remaining roots -------------- 4465 _timer.reset(); 4466 _timer.start(); 4467 heap->cms_process_roots(_strong_roots_scope, 4468 false, // yg was scanned above 4469 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4470 _collector->should_unload_classes(), 4471 &par_mrias_cl, 4472 NULL); // The dirty klasses will be handled below 4473 4474 assert(_collector->should_unload_classes() 4475 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4476 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4477 _timer.stop(); 4478 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4479 4480 // ---------- unhandled CLD scanning ---------- 4481 if (worker_id == 0) { // Single threaded at the moment. 4482 _timer.reset(); 4483 _timer.start(); 4484 4485 // Scan all new class loader data objects and new dependencies that were 4486 // introduced during concurrent marking. 4487 ResourceMark rm; 4488 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4489 for (int i = 0; i < array->length(); i++) { 4490 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4491 } 4492 4493 // We don't need to keep track of new CLDs anymore. 4494 ClassLoaderDataGraph::remember_new_clds(false); 4495 4496 _timer.stop(); 4497 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4498 } 4499 4500 // We might have added oops to ClassLoaderData::_handles during the 4501 // concurrent marking phase. These oops do not always point to newly allocated objects 4502 // that are guaranteed to be kept alive. Hence, 4503 // we do have to revisit the _handles block during the remark phase. 4504 4505 // ---------- dirty CLD scanning ---------- 4506 if (worker_id == 0) { // Single threaded at the moment. 4507 _timer.reset(); 4508 _timer.start(); 4509 4510 // Scan all classes that was dirtied during the concurrent marking phase. 4511 RemarkCLDClosure remark_closure(&par_mrias_cl); 4512 ClassLoaderDataGraph::cld_do(&remark_closure); 4513 4514 _timer.stop(); 4515 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4516 } 4517 4518 4519 // ---------- rescan dirty cards ------------ 4520 _timer.reset(); 4521 _timer.start(); 4522 4523 // Do the rescan tasks for each of the two spaces 4524 // (cms_space) in turn. 4525 // "worker_id" is passed to select the task_queue for "worker_id" 4526 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4527 _timer.stop(); 4528 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4529 4530 // ---------- steal work from other threads ... 4531 // ---------- ... and drain overflow list. 4532 _timer.reset(); 4533 _timer.start(); 4534 do_work_steal(worker_id, &par_mrias_cl); 4535 _timer.stop(); 4536 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4537 } 4538 4539 void 4540 CMSParMarkTask::do_young_space_rescan( 4541 OopsInGenClosure* cl, ContiguousSpace* space, 4542 HeapWord** chunk_array, size_t chunk_top) { 4543 // Until all tasks completed: 4544 // . claim an unclaimed task 4545 // . compute region boundaries corresponding to task claimed 4546 // using chunk_array 4547 // . par_oop_iterate(cl) over that region 4548 4549 ResourceMark rm; 4550 HandleMark hm; 4551 4552 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4553 4554 uint nth_task = 0; 4555 uint n_tasks = pst->n_tasks(); 4556 4557 if (n_tasks > 0) { 4558 assert(pst->valid(), "Uninitialized use?"); 4559 HeapWord *start, *end; 4560 while (pst->try_claim_task(/* reference */ nth_task)) { 4561 // We claimed task # nth_task; compute its boundaries. 4562 if (chunk_top == 0) { // no samples were taken 4563 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4564 start = space->bottom(); 4565 end = space->top(); 4566 } else if (nth_task == 0) { 4567 start = space->bottom(); 4568 end = chunk_array[nth_task]; 4569 } else if (nth_task < (uint)chunk_top) { 4570 assert(nth_task >= 1, "Control point invariant"); 4571 start = chunk_array[nth_task - 1]; 4572 end = chunk_array[nth_task]; 4573 } else { 4574 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4575 start = chunk_array[chunk_top - 1]; 4576 end = space->top(); 4577 } 4578 MemRegion mr(start, end); 4579 // Verify that mr is in space 4580 assert(mr.is_empty() || space->used_region().contains(mr), 4581 "Should be in space"); 4582 // Verify that "start" is an object boundary 4583 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4584 "Should be an oop"); 4585 space->par_oop_iterate(mr, cl); 4586 } 4587 pst->all_tasks_completed(); 4588 } 4589 } 4590 4591 void 4592 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4593 CompactibleFreeListSpace* sp, int i, 4594 ParMarkRefsIntoAndScanClosure* cl) { 4595 // Until all tasks completed: 4596 // . claim an unclaimed task 4597 // . compute region boundaries corresponding to task claimed 4598 // . transfer dirty bits ct->mut for that region 4599 // . apply rescanclosure to dirty mut bits for that region 4600 4601 ResourceMark rm; 4602 HandleMark hm; 4603 4604 OopTaskQueue* work_q = work_queue(i); 4605 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4606 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4607 // CAUTION: This closure has state that persists across calls to 4608 // the work method dirty_range_iterate_clear() in that it has 4609 // embedded in it a (subtype of) UpwardsObjectClosure. The 4610 // use of that state in the embedded UpwardsObjectClosure instance 4611 // assumes that the cards are always iterated (even if in parallel 4612 // by several threads) in monotonically increasing order per each 4613 // thread. This is true of the implementation below which picks 4614 // card ranges (chunks) in monotonically increasing order globally 4615 // and, a-fortiori, in monotonically increasing order per thread 4616 // (the latter order being a subsequence of the former). 4617 // If the work code below is ever reorganized into a more chaotic 4618 // work-partitioning form than the current "sequential tasks" 4619 // paradigm, the use of that persistent state will have to be 4620 // revisited and modified appropriately. See also related 4621 // bug 4756801 work on which should examine this code to make 4622 // sure that the changes there do not run counter to the 4623 // assumptions made here and necessary for correctness and 4624 // efficiency. Note also that this code might yield inefficient 4625 // behavior in the case of very large objects that span one or 4626 // more work chunks. Such objects would potentially be scanned 4627 // several times redundantly. Work on 4756801 should try and 4628 // address that performance anomaly if at all possible. XXX 4629 MemRegion full_span = _collector->_span; 4630 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4631 MarkFromDirtyCardsClosure 4632 greyRescanClosure(_collector, full_span, // entire span of interest 4633 sp, bm, work_q, cl); 4634 4635 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4636 assert(pst->valid(), "Uninitialized use?"); 4637 uint nth_task = 0; 4638 const int alignment = CardTable::card_size * BitsPerWord; 4639 MemRegion span = sp->used_region(); 4640 HeapWord* start_addr = span.start(); 4641 HeapWord* end_addr = align_up(span.end(), alignment); 4642 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4643 assert(is_aligned(start_addr, alignment), "Check alignment"); 4644 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4645 4646 while (pst->try_claim_task(/* reference */ nth_task)) { 4647 // Having claimed the nth_task, compute corresponding mem-region, 4648 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4649 // The alignment restriction ensures that we do not need any 4650 // synchronization with other gang-workers while setting or 4651 // clearing bits in thus chunk of the MUT. 4652 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4653 start_addr + (nth_task+1)*chunk_size); 4654 // The last chunk's end might be way beyond end of the 4655 // used region. In that case pull back appropriately. 4656 if (this_span.end() > end_addr) { 4657 this_span.set_end(end_addr); 4658 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4659 } 4660 // Iterate over the dirty cards covering this chunk, marking them 4661 // precleaned, and setting the corresponding bits in the mod union 4662 // table. Since we have been careful to partition at Card and MUT-word 4663 // boundaries no synchronization is needed between parallel threads. 4664 _collector->_ct->dirty_card_iterate(this_span, 4665 &modUnionClosure); 4666 4667 // Having transferred these marks into the modUnionTable, 4668 // rescan the marked objects on the dirty cards in the modUnionTable. 4669 // Even if this is at a synchronous collection, the initial marking 4670 // may have been done during an asynchronous collection so there 4671 // may be dirty bits in the mod-union table. 4672 _collector->_modUnionTable.dirty_range_iterate_clear( 4673 this_span, &greyRescanClosure); 4674 _collector->_modUnionTable.verifyNoOneBitsInRange( 4675 this_span.start(), 4676 this_span.end()); 4677 } 4678 pst->all_tasks_completed(); // declare that i am done 4679 } 4680 4681 // . see if we can share work_queues with ParNew? XXX 4682 void 4683 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4684 OopTaskQueue* work_q = work_queue(i); 4685 NOT_PRODUCT(int num_steals = 0;) 4686 oop obj_to_scan; 4687 CMSBitMap* bm = &(_collector->_markBitMap); 4688 4689 while (true) { 4690 // Completely finish any left over work from (an) earlier round(s) 4691 cl->trim_queue(0); 4692 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4693 (size_t)ParGCDesiredObjsFromOverflowList); 4694 // Now check if there's any work in the overflow list 4695 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4696 // only affects the number of attempts made to get work from the 4697 // overflow list and does not affect the number of workers. Just 4698 // pass ParallelGCThreads so this behavior is unchanged. 4699 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4700 work_q, 4701 ParallelGCThreads)) { 4702 // found something in global overflow list; 4703 // not yet ready to go stealing work from others. 4704 // We'd like to assert(work_q->size() != 0, ...) 4705 // because we just took work from the overflow list, 4706 // but of course we can't since all of that could have 4707 // been already stolen from us. 4708 // "He giveth and He taketh away." 4709 continue; 4710 } 4711 // Verify that we have no work before we resort to stealing 4712 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4713 // Try to steal from other queues that have work 4714 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4715 NOT_PRODUCT(num_steals++;) 4716 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4717 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4718 // Do scanning work 4719 obj_to_scan->oop_iterate(cl); 4720 // Loop around, finish this work, and try to steal some more 4721 } else if (terminator()->offer_termination()) { 4722 break; // nirvana from the infinite cycle 4723 } 4724 } 4725 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4726 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4727 "Else our work is not yet done"); 4728 } 4729 4730 // Record object boundaries in _eden_chunk_array by sampling the eden 4731 // top in the slow-path eden object allocation code path and record 4732 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4733 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4734 // sampling in sample_eden() that activates during the part of the 4735 // preclean phase. 4736 void CMSCollector::sample_eden_chunk() { 4737 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4738 if (_eden_chunk_lock->try_lock()) { 4739 // Record a sample. This is the critical section. The contents 4740 // of the _eden_chunk_array have to be non-decreasing in the 4741 // address order. 4742 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4743 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4744 "Unexpected state of Eden"); 4745 if (_eden_chunk_index == 0 || 4746 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4747 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4748 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4749 _eden_chunk_index++; // commit sample 4750 } 4751 _eden_chunk_lock->unlock(); 4752 } 4753 } 4754 } 4755 4756 // Return a thread-local PLAB recording array, as appropriate. 4757 void* CMSCollector::get_data_recorder(int thr_num) { 4758 if (_survivor_plab_array != NULL && 4759 (CMSPLABRecordAlways || 4760 (_collectorState > Marking && _collectorState < FinalMarking))) { 4761 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4762 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4763 ca->reset(); // clear it so that fresh data is recorded 4764 return (void*) ca; 4765 } else { 4766 return NULL; 4767 } 4768 } 4769 4770 // Reset all the thread-local PLAB recording arrays 4771 void CMSCollector::reset_survivor_plab_arrays() { 4772 for (uint i = 0; i < ParallelGCThreads; i++) { 4773 _survivor_plab_array[i].reset(); 4774 } 4775 } 4776 4777 // Merge the per-thread plab arrays into the global survivor chunk 4778 // array which will provide the partitioning of the survivor space 4779 // for CMS initial scan and rescan. 4780 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4781 int no_of_gc_threads) { 4782 assert(_survivor_plab_array != NULL, "Error"); 4783 assert(_survivor_chunk_array != NULL, "Error"); 4784 assert(_collectorState == FinalMarking || 4785 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4786 for (int j = 0; j < no_of_gc_threads; j++) { 4787 _cursor[j] = 0; 4788 } 4789 HeapWord* top = surv->top(); 4790 size_t i; 4791 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4792 HeapWord* min_val = top; // Higher than any PLAB address 4793 uint min_tid = 0; // position of min_val this round 4794 for (int j = 0; j < no_of_gc_threads; j++) { 4795 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4796 if (_cursor[j] == cur_sca->end()) { 4797 continue; 4798 } 4799 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4800 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4801 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4802 if (cur_val < min_val) { 4803 min_tid = j; 4804 min_val = cur_val; 4805 } else { 4806 assert(cur_val < top, "All recorded addresses should be less"); 4807 } 4808 } 4809 // At this point min_val and min_tid are respectively 4810 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4811 // and the thread (j) that witnesses that address. 4812 // We record this address in the _survivor_chunk_array[i] 4813 // and increment _cursor[min_tid] prior to the next round i. 4814 if (min_val == top) { 4815 break; 4816 } 4817 _survivor_chunk_array[i] = min_val; 4818 _cursor[min_tid]++; 4819 } 4820 // We are all done; record the size of the _survivor_chunk_array 4821 _survivor_chunk_index = i; // exclusive: [0, i) 4822 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4823 // Verify that we used up all the recorded entries 4824 #ifdef ASSERT 4825 size_t total = 0; 4826 for (int j = 0; j < no_of_gc_threads; j++) { 4827 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4828 total += _cursor[j]; 4829 } 4830 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4831 // Check that the merged array is in sorted order 4832 if (total > 0) { 4833 for (size_t i = 0; i < total - 1; i++) { 4834 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4835 i, p2i(_survivor_chunk_array[i])); 4836 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4837 "Not sorted"); 4838 } 4839 } 4840 #endif // ASSERT 4841 } 4842 4843 // Set up the space's par_seq_tasks structure for work claiming 4844 // for parallel initial scan and rescan of young gen. 4845 // See ParRescanTask where this is currently used. 4846 void 4847 CMSCollector:: 4848 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4849 assert(n_threads > 0, "Unexpected n_threads argument"); 4850 4851 // Eden space 4852 if (!_young_gen->eden()->is_empty()) { 4853 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4854 assert(!pst->valid(), "Clobbering existing data?"); 4855 // Each valid entry in [0, _eden_chunk_index) represents a task. 4856 size_t n_tasks = _eden_chunk_index + 1; 4857 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4858 // Sets the condition for completion of the subtask (how many threads 4859 // need to finish in order to be done). 4860 pst->set_n_threads(n_threads); 4861 pst->set_n_tasks((int)n_tasks); 4862 } 4863 4864 // Merge the survivor plab arrays into _survivor_chunk_array 4865 if (_survivor_plab_array != NULL) { 4866 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4867 } else { 4868 assert(_survivor_chunk_index == 0, "Error"); 4869 } 4870 4871 // To space 4872 { 4873 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4874 assert(!pst->valid(), "Clobbering existing data?"); 4875 // Sets the condition for completion of the subtask (how many threads 4876 // need to finish in order to be done). 4877 pst->set_n_threads(n_threads); 4878 pst->set_n_tasks(1); 4879 assert(pst->valid(), "Error"); 4880 } 4881 4882 // From space 4883 { 4884 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4885 assert(!pst->valid(), "Clobbering existing data?"); 4886 size_t n_tasks = _survivor_chunk_index + 1; 4887 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4888 // Sets the condition for completion of the subtask (how many threads 4889 // need to finish in order to be done). 4890 pst->set_n_threads(n_threads); 4891 pst->set_n_tasks((int)n_tasks); 4892 assert(pst->valid(), "Error"); 4893 } 4894 } 4895 4896 // Parallel version of remark 4897 void CMSCollector::do_remark_parallel() { 4898 CMSHeap* heap = CMSHeap::heap(); 4899 WorkGang* workers = heap->workers(); 4900 assert(workers != NULL, "Need parallel worker threads."); 4901 // Choose to use the number of GC workers most recently set 4902 // into "active_workers". 4903 uint n_workers = workers->active_workers(); 4904 4905 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4906 4907 StrongRootsScope srs(n_workers); 4908 4909 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4910 4911 // We won't be iterating over the cards in the card table updating 4912 // the younger_gen cards, so we shouldn't call the following else 4913 // the verification code as well as subsequent younger_refs_iterate 4914 // code would get confused. XXX 4915 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4916 4917 // The young gen rescan work will not be done as part of 4918 // process_roots (which currently doesn't know how to 4919 // parallelize such a scan), but rather will be broken up into 4920 // a set of parallel tasks (via the sampling that the [abortable] 4921 // preclean phase did of eden, plus the [two] tasks of 4922 // scanning the [two] survivor spaces. Further fine-grain 4923 // parallelization of the scanning of the survivor spaces 4924 // themselves, and of precleaning of the young gen itself 4925 // is deferred to the future. 4926 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4927 4928 // The dirty card rescan work is broken up into a "sequence" 4929 // of parallel tasks (per constituent space) that are dynamically 4930 // claimed by the parallel threads. 4931 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4932 4933 // It turns out that even when we're using 1 thread, doing the work in a 4934 // separate thread causes wide variance in run times. We can't help this 4935 // in the multi-threaded case, but we special-case n=1 here to get 4936 // repeatable measurements of the 1-thread overhead of the parallel code. 4937 if (n_workers > 1) { 4938 // Make refs discovery MT-safe, if it isn't already: it may not 4939 // necessarily be so, since it's possible that we are doing 4940 // ST marking. 4941 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4942 workers->run_task(&tsk); 4943 } else { 4944 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4945 tsk.work(0); 4946 } 4947 4948 // restore, single-threaded for now, any preserved marks 4949 // as a result of work_q overflow 4950 restore_preserved_marks_if_any(); 4951 } 4952 4953 // Non-parallel version of remark 4954 void CMSCollector::do_remark_non_parallel() { 4955 ResourceMark rm; 4956 HandleMark hm; 4957 CMSHeap* heap = CMSHeap::heap(); 4958 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4959 4960 MarkRefsIntoAndScanClosure 4961 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4962 &_markStack, this, 4963 false /* should_yield */, false /* not precleaning */); 4964 MarkFromDirtyCardsClosure 4965 markFromDirtyCardsClosure(this, _span, 4966 NULL, // space is set further below 4967 &_markBitMap, &_markStack, &mrias_cl); 4968 { 4969 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4970 // Iterate over the dirty cards, setting the corresponding bits in the 4971 // mod union table. 4972 { 4973 ModUnionClosure modUnionClosure(&_modUnionTable); 4974 _ct->dirty_card_iterate(_cmsGen->used_region(), 4975 &modUnionClosure); 4976 } 4977 // Having transferred these marks into the modUnionTable, we just need 4978 // to rescan the marked objects on the dirty cards in the modUnionTable. 4979 // The initial marking may have been done during an asynchronous 4980 // collection so there may be dirty bits in the mod-union table. 4981 const int alignment = CardTable::card_size * BitsPerWord; 4982 { 4983 // ... First handle dirty cards in CMS gen 4984 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4985 MemRegion ur = _cmsGen->used_region(); 4986 HeapWord* lb = ur.start(); 4987 HeapWord* ub = align_up(ur.end(), alignment); 4988 MemRegion cms_span(lb, ub); 4989 _modUnionTable.dirty_range_iterate_clear(cms_span, 4990 &markFromDirtyCardsClosure); 4991 verify_work_stacks_empty(); 4992 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4993 } 4994 } 4995 if (VerifyDuringGC && 4996 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4997 HandleMark hm; // Discard invalid handles created during verification 4998 Universe::verify(); 4999 } 5000 { 5001 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 5002 5003 verify_work_stacks_empty(); 5004 5005 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5006 StrongRootsScope srs(1); 5007 5008 heap->cms_process_roots(&srs, 5009 true, // young gen as roots 5010 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5011 should_unload_classes(), 5012 &mrias_cl, 5013 NULL); // The dirty klasses will be handled below 5014 5015 assert(should_unload_classes() 5016 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5017 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5018 } 5019 5020 { 5021 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 5022 5023 verify_work_stacks_empty(); 5024 5025 // Scan all class loader data objects that might have been introduced 5026 // during concurrent marking. 5027 ResourceMark rm; 5028 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5029 for (int i = 0; i < array->length(); i++) { 5030 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 5031 } 5032 5033 // We don't need to keep track of new CLDs anymore. 5034 ClassLoaderDataGraph::remember_new_clds(false); 5035 5036 verify_work_stacks_empty(); 5037 } 5038 5039 // We might have added oops to ClassLoaderData::_handles during the 5040 // concurrent marking phase. These oops do not point to newly allocated objects 5041 // that are guaranteed to be kept alive. Hence, 5042 // we do have to revisit the _handles block during the remark phase. 5043 { 5044 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 5045 5046 verify_work_stacks_empty(); 5047 5048 RemarkCLDClosure remark_closure(&mrias_cl); 5049 ClassLoaderDataGraph::cld_do(&remark_closure); 5050 5051 verify_work_stacks_empty(); 5052 } 5053 5054 verify_work_stacks_empty(); 5055 // Restore evacuated mark words, if any, used for overflow list links 5056 restore_preserved_marks_if_any(); 5057 5058 verify_overflow_empty(); 5059 } 5060 5061 //////////////////////////////////////////////////////// 5062 // Parallel Reference Processing Task Proxy Class 5063 //////////////////////////////////////////////////////// 5064 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5065 OopTaskQueueSet* _queues; 5066 TaskTerminator _terminator; 5067 public: 5068 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5069 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5070 ParallelTaskTerminator* terminator() { return _terminator.terminator(); } 5071 OopTaskQueueSet* queues() { return _queues; } 5072 }; 5073 5074 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5075 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5076 CMSCollector* _collector; 5077 CMSBitMap* _mark_bit_map; 5078 const MemRegion _span; 5079 ProcessTask& _task; 5080 5081 public: 5082 CMSRefProcTaskProxy(ProcessTask& task, 5083 CMSCollector* collector, 5084 const MemRegion& span, 5085 CMSBitMap* mark_bit_map, 5086 AbstractWorkGang* workers, 5087 OopTaskQueueSet* task_queues): 5088 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5089 task_queues, 5090 workers->active_workers()), 5091 _collector(collector), 5092 _mark_bit_map(mark_bit_map), 5093 _span(span), 5094 _task(task) 5095 { 5096 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5097 "Inconsistency in _span"); 5098 } 5099 5100 OopTaskQueueSet* task_queues() { return queues(); } 5101 5102 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5103 5104 void do_work_steal(int i, 5105 CMSParDrainMarkingStackClosure* drain, 5106 CMSParKeepAliveClosure* keep_alive); 5107 5108 virtual void work(uint worker_id); 5109 }; 5110 5111 void CMSRefProcTaskProxy::work(uint worker_id) { 5112 ResourceMark rm; 5113 HandleMark hm; 5114 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5115 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5116 _mark_bit_map, 5117 work_queue(worker_id)); 5118 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5119 _mark_bit_map, 5120 work_queue(worker_id)); 5121 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5122 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5123 if (_task.marks_oops_alive()) { 5124 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5125 } 5126 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5127 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5128 } 5129 5130 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5131 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5132 _span(span), 5133 _work_queue(work_queue), 5134 _bit_map(bit_map), 5135 _mark_and_push(collector, span, bit_map, work_queue), 5136 _low_water_mark(MIN2((work_queue->max_elems()/4), 5137 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5138 { } 5139 5140 // . see if we can share work_queues with ParNew? XXX 5141 void CMSRefProcTaskProxy::do_work_steal(int i, 5142 CMSParDrainMarkingStackClosure* drain, 5143 CMSParKeepAliveClosure* keep_alive) { 5144 OopTaskQueue* work_q = work_queue(i); 5145 NOT_PRODUCT(int num_steals = 0;) 5146 oop obj_to_scan; 5147 5148 while (true) { 5149 // Completely finish any left over work from (an) earlier round(s) 5150 drain->trim_queue(0); 5151 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5152 (size_t)ParGCDesiredObjsFromOverflowList); 5153 // Now check if there's any work in the overflow list 5154 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5155 // only affects the number of attempts made to get work from the 5156 // overflow list and does not affect the number of workers. Just 5157 // pass ParallelGCThreads so this behavior is unchanged. 5158 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5159 work_q, 5160 ParallelGCThreads)) { 5161 // Found something in global overflow list; 5162 // not yet ready to go stealing work from others. 5163 // We'd like to assert(work_q->size() != 0, ...) 5164 // because we just took work from the overflow list, 5165 // but of course we can't, since all of that might have 5166 // been already stolen from us. 5167 continue; 5168 } 5169 // Verify that we have no work before we resort to stealing 5170 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5171 // Try to steal from other queues that have work 5172 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5173 NOT_PRODUCT(num_steals++;) 5174 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5175 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5176 // Do scanning work 5177 obj_to_scan->oop_iterate(keep_alive); 5178 // Loop around, finish this work, and try to steal some more 5179 } else if (terminator()->offer_termination()) { 5180 break; // nirvana from the infinite cycle 5181 } 5182 } 5183 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5184 } 5185 5186 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5187 CMSHeap* heap = CMSHeap::heap(); 5188 WorkGang* workers = heap->workers(); 5189 assert(workers != NULL, "Need parallel worker threads."); 5190 assert(workers->active_workers() == ergo_workers, 5191 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5192 ergo_workers, workers->active_workers()); 5193 CMSRefProcTaskProxy rp_task(task, &_collector, 5194 _collector.ref_processor_span(), 5195 _collector.markBitMap(), 5196 workers, _collector.task_queues()); 5197 workers->run_task(&rp_task, workers->active_workers()); 5198 } 5199 5200 void CMSCollector::refProcessingWork() { 5201 ResourceMark rm; 5202 HandleMark hm; 5203 5204 ReferenceProcessor* rp = ref_processor(); 5205 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5206 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5207 // Process weak references. 5208 rp->setup_policy(false); 5209 verify_work_stacks_empty(); 5210 5211 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5212 { 5213 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5214 5215 // Setup keep_alive and complete closures. 5216 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5217 &_markStack, false /* !preclean */); 5218 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5219 _span, &_markBitMap, &_markStack, 5220 &cmsKeepAliveClosure, false /* !preclean */); 5221 5222 ReferenceProcessorStats stats; 5223 if (rp->processing_is_mt()) { 5224 // Set the degree of MT here. If the discovery is done MT, there 5225 // may have been a different number of threads doing the discovery 5226 // and a different number of discovered lists may have Ref objects. 5227 // That is OK as long as the Reference lists are balanced (see 5228 // balance_all_queues() and balance_queues()). 5229 CMSHeap* heap = CMSHeap::heap(); 5230 uint active_workers = ParallelGCThreads; 5231 WorkGang* workers = heap->workers(); 5232 if (workers != NULL) { 5233 active_workers = workers->active_workers(); 5234 // The expectation is that active_workers will have already 5235 // been set to a reasonable value. If it has not been set, 5236 // investigate. 5237 assert(active_workers > 0, "Should have been set during scavenge"); 5238 } 5239 rp->set_active_mt_degree(active_workers); 5240 CMSRefProcTaskExecutor task_executor(*this); 5241 stats = rp->process_discovered_references(&_is_alive_closure, 5242 &cmsKeepAliveClosure, 5243 &cmsDrainMarkingStackClosure, 5244 &task_executor, 5245 &pt); 5246 } else { 5247 stats = rp->process_discovered_references(&_is_alive_closure, 5248 &cmsKeepAliveClosure, 5249 &cmsDrainMarkingStackClosure, 5250 NULL, 5251 &pt); 5252 } 5253 _gc_tracer_cm->report_gc_reference_stats(stats); 5254 pt.print_all_references(); 5255 } 5256 5257 // This is the point where the entire marking should have completed. 5258 verify_work_stacks_empty(); 5259 5260 { 5261 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5262 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5263 } 5264 5265 if (should_unload_classes()) { 5266 { 5267 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5268 5269 // Unload classes and purge the SystemDictionary. 5270 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5271 5272 // Unload nmethods. 5273 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5274 5275 // Prune dead klasses from subklass/sibling/implementor lists. 5276 Klass::clean_weak_klass_links(purged_class); 5277 5278 // Clean JVMCI metadata handles. 5279 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 5280 } 5281 } 5282 5283 // Restore any preserved marks as a result of mark stack or 5284 // work queue overflow 5285 restore_preserved_marks_if_any(); // done single-threaded for now 5286 5287 rp->set_enqueuing_is_done(true); 5288 rp->verify_no_references_recorded(); 5289 } 5290 5291 #ifndef PRODUCT 5292 void CMSCollector::check_correct_thread_executing() { 5293 Thread* t = Thread::current(); 5294 // Only the VM thread or the CMS thread should be here. 5295 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5296 "Unexpected thread type"); 5297 // If this is the vm thread, the foreground process 5298 // should not be waiting. Note that _foregroundGCIsActive is 5299 // true while the foreground collector is waiting. 5300 if (_foregroundGCShouldWait) { 5301 // We cannot be the VM thread 5302 assert(t->is_ConcurrentGC_thread(), 5303 "Should be CMS thread"); 5304 } else { 5305 // We can be the CMS thread only if we are in a stop-world 5306 // phase of CMS collection. 5307 if (t->is_ConcurrentGC_thread()) { 5308 assert(_collectorState == InitialMarking || 5309 _collectorState == FinalMarking, 5310 "Should be a stop-world phase"); 5311 // The CMS thread should be holding the CMS_token. 5312 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5313 "Potential interference with concurrently " 5314 "executing VM thread"); 5315 } 5316 } 5317 } 5318 #endif 5319 5320 void CMSCollector::sweep() { 5321 assert(_collectorState == Sweeping, "just checking"); 5322 check_correct_thread_executing(); 5323 verify_work_stacks_empty(); 5324 verify_overflow_empty(); 5325 increment_sweep_count(); 5326 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5327 5328 _inter_sweep_timer.stop(); 5329 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5330 5331 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5332 _intra_sweep_timer.reset(); 5333 _intra_sweep_timer.start(); 5334 { 5335 GCTraceCPUTime tcpu; 5336 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5337 // First sweep the old gen 5338 { 5339 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5340 bitMapLock()); 5341 sweepWork(_cmsGen); 5342 } 5343 5344 // Update Universe::_heap_*_at_gc figures. 5345 // We need all the free list locks to make the abstract state 5346 // transition from Sweeping to Resetting. See detailed note 5347 // further below. 5348 { 5349 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5350 5351 // Update heap occupancy information which is used as 5352 // input to soft ref clearing policy at the next gc. 5353 Universe::update_heap_info_at_gc(); 5354 5355 // recalculate CMS used space after CMS collection 5356 _cmsGen->cmsSpace()->recalculate_used_stable(); 5357 5358 _collectorState = Resizing; 5359 } 5360 } 5361 verify_work_stacks_empty(); 5362 verify_overflow_empty(); 5363 5364 if (should_unload_classes()) { 5365 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5366 // requires that the virtual spaces are stable and not deleted. 5367 ClassLoaderDataGraph::set_should_purge(true); 5368 } 5369 5370 _intra_sweep_timer.stop(); 5371 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5372 5373 _inter_sweep_timer.reset(); 5374 _inter_sweep_timer.start(); 5375 5376 // We need to use a monotonically non-decreasing time in ms 5377 // or we will see time-warp warnings and os::javaTimeMillis() 5378 // does not guarantee monotonicity. 5379 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5380 update_time_of_last_gc(now); 5381 5382 // NOTE on abstract state transitions: 5383 // Mutators allocate-live and/or mark the mod-union table dirty 5384 // based on the state of the collection. The former is done in 5385 // the interval [Marking, Sweeping] and the latter in the interval 5386 // [Marking, Sweeping). Thus the transitions into the Marking state 5387 // and out of the Sweeping state must be synchronously visible 5388 // globally to the mutators. 5389 // The transition into the Marking state happens with the world 5390 // stopped so the mutators will globally see it. Sweeping is 5391 // done asynchronously by the background collector so the transition 5392 // from the Sweeping state to the Resizing state must be done 5393 // under the freelistLock (as is the check for whether to 5394 // allocate-live and whether to dirty the mod-union table). 5395 assert(_collectorState == Resizing, "Change of collector state to" 5396 " Resizing must be done under the freelistLocks (plural)"); 5397 5398 // Now that sweeping has been completed, we clear 5399 // the incremental_collection_failed flag, 5400 // thus inviting a younger gen collection to promote into 5401 // this generation. If such a promotion may still fail, 5402 // the flag will be set again when a young collection is 5403 // attempted. 5404 CMSHeap* heap = CMSHeap::heap(); 5405 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5406 heap->update_full_collections_completed(_collection_count_start); 5407 } 5408 5409 // FIX ME!!! Looks like this belongs in CFLSpace, with 5410 // CMSGen merely delegating to it. 5411 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5412 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5413 HeapWord* minAddr = _cmsSpace->bottom(); 5414 HeapWord* largestAddr = 5415 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5416 if (largestAddr == NULL) { 5417 // The dictionary appears to be empty. In this case 5418 // try to coalesce at the end of the heap. 5419 largestAddr = _cmsSpace->end(); 5420 } 5421 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5422 size_t nearLargestOffset = 5423 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5424 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5425 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5426 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5427 } 5428 5429 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5430 return addr >= _cmsSpace->nearLargestChunk(); 5431 } 5432 5433 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5434 return _cmsSpace->find_chunk_at_end(); 5435 } 5436 5437 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5438 bool full) { 5439 // If the young generation has been collected, gather any statistics 5440 // that are of interest at this point. 5441 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5442 if (!full && current_is_young) { 5443 // Gather statistics on the young generation collection. 5444 collector()->stats().record_gc0_end(used()); 5445 } 5446 _cmsSpace->recalculate_used_stable(); 5447 } 5448 5449 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5450 // We iterate over the space(s) underlying this generation, 5451 // checking the mark bit map to see if the bits corresponding 5452 // to specific blocks are marked or not. Blocks that are 5453 // marked are live and are not swept up. All remaining blocks 5454 // are swept up, with coalescing on-the-fly as we sweep up 5455 // contiguous free and/or garbage blocks: 5456 // We need to ensure that the sweeper synchronizes with allocators 5457 // and stop-the-world collectors. In particular, the following 5458 // locks are used: 5459 // . CMS token: if this is held, a stop the world collection cannot occur 5460 // . freelistLock: if this is held no allocation can occur from this 5461 // generation by another thread 5462 // . bitMapLock: if this is held, no other thread can access or update 5463 // 5464 5465 // Note that we need to hold the freelistLock if we use 5466 // block iterate below; else the iterator might go awry if 5467 // a mutator (or promotion) causes block contents to change 5468 // (for instance if the allocator divvies up a block). 5469 // If we hold the free list lock, for all practical purposes 5470 // young generation GC's can't occur (they'll usually need to 5471 // promote), so we might as well prevent all young generation 5472 // GC's while we do a sweeping step. For the same reason, we might 5473 // as well take the bit map lock for the entire duration 5474 5475 // check that we hold the requisite locks 5476 assert(have_cms_token(), "Should hold cms token"); 5477 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5478 assert_lock_strong(old_gen->freelistLock()); 5479 assert_lock_strong(bitMapLock()); 5480 5481 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5482 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5483 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5484 _inter_sweep_estimate.padded_average(), 5485 _intra_sweep_estimate.padded_average()); 5486 old_gen->setNearLargestChunk(); 5487 5488 { 5489 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5490 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5491 // We need to free-up/coalesce garbage/blocks from a 5492 // co-terminal free run. This is done in the SweepClosure 5493 // destructor; so, do not remove this scope, else the 5494 // end-of-sweep-census below will be off by a little bit. 5495 } 5496 old_gen->cmsSpace()->sweep_completed(); 5497 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5498 if (should_unload_classes()) { // unloaded classes this cycle, 5499 _concurrent_cycles_since_last_unload = 0; // ... reset count 5500 } else { // did not unload classes, 5501 _concurrent_cycles_since_last_unload++; // ... increment count 5502 } 5503 } 5504 5505 // Reset CMS data structures (for now just the marking bit map) 5506 // preparatory for the next cycle. 5507 void CMSCollector::reset_concurrent() { 5508 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5509 5510 // If the state is not "Resetting", the foreground thread 5511 // has done a collection and the resetting. 5512 if (_collectorState != Resetting) { 5513 assert(_collectorState == Idling, "The state should only change" 5514 " because the foreground collector has finished the collection"); 5515 return; 5516 } 5517 5518 { 5519 // Clear the mark bitmap (no grey objects to start with) 5520 // for the next cycle. 5521 GCTraceCPUTime tcpu; 5522 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5523 5524 HeapWord* curAddr = _markBitMap.startWord(); 5525 while (curAddr < _markBitMap.endWord()) { 5526 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5527 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5528 _markBitMap.clear_large_range(chunk); 5529 if (ConcurrentMarkSweepThread::should_yield() && 5530 !foregroundGCIsActive() && 5531 CMSYield) { 5532 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5533 "CMS thread should hold CMS token"); 5534 assert_lock_strong(bitMapLock()); 5535 bitMapLock()->unlock(); 5536 ConcurrentMarkSweepThread::desynchronize(true); 5537 stopTimer(); 5538 incrementYields(); 5539 5540 // See the comment in coordinator_yield() 5541 for (unsigned i = 0; i < CMSYieldSleepCount && 5542 ConcurrentMarkSweepThread::should_yield() && 5543 !CMSCollector::foregroundGCIsActive(); ++i) { 5544 os::sleep(Thread::current(), 1, false); 5545 } 5546 5547 ConcurrentMarkSweepThread::synchronize(true); 5548 bitMapLock()->lock_without_safepoint_check(); 5549 startTimer(); 5550 } 5551 curAddr = chunk.end(); 5552 } 5553 // A successful mostly concurrent collection has been done. 5554 // Because only the full (i.e., concurrent mode failure) collections 5555 // are being measured for gc overhead limits, clean the "near" flag 5556 // and count. 5557 size_policy()->reset_gc_overhead_limit_count(); 5558 _collectorState = Idling; 5559 } 5560 5561 register_gc_end(); 5562 } 5563 5564 // Same as above but for STW paths 5565 void CMSCollector::reset_stw() { 5566 // already have the lock 5567 assert(_collectorState == Resetting, "just checking"); 5568 assert_lock_strong(bitMapLock()); 5569 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5570 _markBitMap.clear_all(); 5571 _collectorState = Idling; 5572 register_gc_end(); 5573 } 5574 5575 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5576 GCTraceCPUTime tcpu; 5577 TraceCollectorStats tcs_cgc(cgc_counters()); 5578 5579 switch (op) { 5580 case CMS_op_checkpointRootsInitial: { 5581 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5582 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5583 checkpointRootsInitial(); 5584 break; 5585 } 5586 case CMS_op_checkpointRootsFinal: { 5587 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5588 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5589 checkpointRootsFinal(); 5590 break; 5591 } 5592 default: 5593 fatal("No such CMS_op"); 5594 } 5595 } 5596 5597 #ifndef PRODUCT 5598 size_t const CMSCollector::skip_header_HeapWords() { 5599 return FreeChunk::header_size(); 5600 } 5601 5602 // Try and collect here conditions that should hold when 5603 // CMS thread is exiting. The idea is that the foreground GC 5604 // thread should not be blocked if it wants to terminate 5605 // the CMS thread and yet continue to run the VM for a while 5606 // after that. 5607 void CMSCollector::verify_ok_to_terminate() const { 5608 assert(Thread::current()->is_ConcurrentGC_thread(), 5609 "should be called by CMS thread"); 5610 assert(!_foregroundGCShouldWait, "should be false"); 5611 // We could check here that all the various low-level locks 5612 // are not held by the CMS thread, but that is overkill; see 5613 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5614 // is checked. 5615 } 5616 #endif 5617 5618 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5619 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5620 "missing Printezis mark?"); 5621 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5622 size_t size = pointer_delta(nextOneAddr + 1, addr); 5623 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5624 "alignment problem"); 5625 assert(size >= 3, "Necessary for Printezis marks to work"); 5626 return size; 5627 } 5628 5629 // A variant of the above (block_size_using_printezis_bits()) except 5630 // that we return 0 if the P-bits are not yet set. 5631 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5632 if (_markBitMap.isMarked(addr + 1)) { 5633 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5634 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5635 size_t size = pointer_delta(nextOneAddr + 1, addr); 5636 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5637 "alignment problem"); 5638 assert(size >= 3, "Necessary for Printezis marks to work"); 5639 return size; 5640 } 5641 return 0; 5642 } 5643 5644 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5645 size_t sz = 0; 5646 oop p = (oop)addr; 5647 if (p->klass_or_null_acquire() != NULL) { 5648 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5649 } else { 5650 sz = block_size_using_printezis_bits(addr); 5651 } 5652 assert(sz > 0, "size must be nonzero"); 5653 HeapWord* next_block = addr + sz; 5654 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5655 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5656 align_down((uintptr_t)next_card, CardTable::card_size), 5657 "must be different cards"); 5658 return next_card; 5659 } 5660 5661 5662 // CMS Bit Map Wrapper ///////////////////////////////////////// 5663 5664 // Construct a CMS bit map infrastructure, but don't create the 5665 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5666 // further below. 5667 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5668 _shifter(shifter), 5669 _bm(), 5670 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5671 Monitor::_safepoint_check_never) : NULL) 5672 { 5673 _bmStartWord = 0; 5674 _bmWordSize = 0; 5675 } 5676 5677 bool CMSBitMap::allocate(MemRegion mr) { 5678 _bmStartWord = mr.start(); 5679 _bmWordSize = mr.word_size(); 5680 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5681 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5682 if (!brs.is_reserved()) { 5683 log_warning(gc)("CMS bit map allocation failure"); 5684 return false; 5685 } 5686 // For now we'll just commit all of the bit map up front. 5687 // Later on we'll try to be more parsimonious with swap. 5688 if (!_virtual_space.initialize(brs, brs.size())) { 5689 log_warning(gc)("CMS bit map backing store failure"); 5690 return false; 5691 } 5692 assert(_virtual_space.committed_size() == brs.size(), 5693 "didn't reserve backing store for all of CMS bit map?"); 5694 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5695 _bmWordSize, "inconsistency in bit map sizing"); 5696 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5697 5698 // bm.clear(); // can we rely on getting zero'd memory? verify below 5699 assert(isAllClear(), 5700 "Expected zero'd memory from ReservedSpace constructor"); 5701 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5702 "consistency check"); 5703 return true; 5704 } 5705 5706 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5707 HeapWord *next_addr, *end_addr, *last_addr; 5708 assert_locked(); 5709 assert(covers(mr), "out-of-range error"); 5710 // XXX assert that start and end are appropriately aligned 5711 for (next_addr = mr.start(), end_addr = mr.end(); 5712 next_addr < end_addr; next_addr = last_addr) { 5713 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5714 last_addr = dirty_region.end(); 5715 if (!dirty_region.is_empty()) { 5716 cl->do_MemRegion(dirty_region); 5717 } else { 5718 assert(last_addr == end_addr, "program logic"); 5719 return; 5720 } 5721 } 5722 } 5723 5724 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5725 _bm.print_on_error(st, prefix); 5726 } 5727 5728 #ifndef PRODUCT 5729 void CMSBitMap::assert_locked() const { 5730 CMSLockVerifier::assert_locked(lock()); 5731 } 5732 5733 bool CMSBitMap::covers(MemRegion mr) const { 5734 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5735 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5736 "size inconsistency"); 5737 return (mr.start() >= _bmStartWord) && 5738 (mr.end() <= endWord()); 5739 } 5740 5741 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5742 return (start >= _bmStartWord && (start + size) <= endWord()); 5743 } 5744 5745 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5746 // verify that there are no 1 bits in the interval [left, right) 5747 FalseBitMapClosure falseBitMapClosure; 5748 iterate(&falseBitMapClosure, left, right); 5749 } 5750 5751 void CMSBitMap::region_invariant(MemRegion mr) 5752 { 5753 assert_locked(); 5754 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5755 assert(!mr.is_empty(), "unexpected empty region"); 5756 assert(covers(mr), "mr should be covered by bit map"); 5757 // convert address range into offset range 5758 size_t start_ofs = heapWordToOffset(mr.start()); 5759 // Make sure that end() is appropriately aligned 5760 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5761 "Misaligned mr.end()"); 5762 size_t end_ofs = heapWordToOffset(mr.end()); 5763 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5764 } 5765 5766 #endif 5767 5768 bool CMSMarkStack::allocate(size_t size) { 5769 // allocate a stack of the requisite depth 5770 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5771 size * sizeof(oop))); 5772 if (!rs.is_reserved()) { 5773 log_warning(gc)("CMSMarkStack allocation failure"); 5774 return false; 5775 } 5776 if (!_virtual_space.initialize(rs, rs.size())) { 5777 log_warning(gc)("CMSMarkStack backing store failure"); 5778 return false; 5779 } 5780 assert(_virtual_space.committed_size() == rs.size(), 5781 "didn't reserve backing store for all of CMS stack?"); 5782 _base = (oop*)(_virtual_space.low()); 5783 _index = 0; 5784 _capacity = size; 5785 NOT_PRODUCT(_max_depth = 0); 5786 return true; 5787 } 5788 5789 // XXX FIX ME !!! In the MT case we come in here holding a 5790 // leaf lock. For printing we need to take a further lock 5791 // which has lower rank. We need to recalibrate the two 5792 // lock-ranks involved in order to be able to print the 5793 // messages below. (Or defer the printing to the caller. 5794 // For now we take the expedient path of just disabling the 5795 // messages for the problematic case.) 5796 void CMSMarkStack::expand() { 5797 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5798 if (_capacity == MarkStackSizeMax) { 5799 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5800 // We print a warning message only once per CMS cycle. 5801 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5802 } 5803 return; 5804 } 5805 // Double capacity if possible 5806 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5807 // Do not give up existing stack until we have managed to 5808 // get the double capacity that we desired. 5809 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5810 new_capacity * sizeof(oop))); 5811 if (rs.is_reserved()) { 5812 // Release the backing store associated with old stack 5813 _virtual_space.release(); 5814 // Reinitialize virtual space for new stack 5815 if (!_virtual_space.initialize(rs, rs.size())) { 5816 fatal("Not enough swap for expanded marking stack"); 5817 } 5818 _base = (oop*)(_virtual_space.low()); 5819 _index = 0; 5820 _capacity = new_capacity; 5821 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5822 // Failed to double capacity, continue; 5823 // we print a detail message only once per CMS cycle. 5824 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5825 _capacity / K, new_capacity / K); 5826 } 5827 } 5828 5829 5830 // Closures 5831 // XXX: there seems to be a lot of code duplication here; 5832 // should refactor and consolidate common code. 5833 5834 // This closure is used to mark refs into the CMS generation in 5835 // the CMS bit map. Called at the first checkpoint. This closure 5836 // assumes that we do not need to re-mark dirty cards; if the CMS 5837 // generation on which this is used is not an oldest 5838 // generation then this will lose younger_gen cards! 5839 5840 MarkRefsIntoClosure::MarkRefsIntoClosure( 5841 MemRegion span, CMSBitMap* bitMap): 5842 _span(span), 5843 _bitMap(bitMap) 5844 { 5845 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5846 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5847 } 5848 5849 void MarkRefsIntoClosure::do_oop(oop obj) { 5850 // if p points into _span, then mark corresponding bit in _markBitMap 5851 assert(oopDesc::is_oop(obj), "expected an oop"); 5852 HeapWord* addr = (HeapWord*)obj; 5853 if (_span.contains(addr)) { 5854 // this should be made more efficient 5855 _bitMap->mark(addr); 5856 } 5857 } 5858 5859 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5860 MemRegion span, CMSBitMap* bitMap): 5861 _span(span), 5862 _bitMap(bitMap) 5863 { 5864 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5865 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5866 } 5867 5868 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5869 // if p points into _span, then mark corresponding bit in _markBitMap 5870 assert(oopDesc::is_oop(obj), "expected an oop"); 5871 HeapWord* addr = (HeapWord*)obj; 5872 if (_span.contains(addr)) { 5873 // this should be made more efficient 5874 _bitMap->par_mark(addr); 5875 } 5876 } 5877 5878 // A variant of the above, used for CMS marking verification. 5879 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5880 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5881 _span(span), 5882 _verification_bm(verification_bm), 5883 _cms_bm(cms_bm) 5884 { 5885 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5886 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5887 } 5888 5889 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5890 // if p points into _span, then mark corresponding bit in _markBitMap 5891 assert(oopDesc::is_oop(obj), "expected an oop"); 5892 HeapWord* addr = (HeapWord*)obj; 5893 if (_span.contains(addr)) { 5894 _verification_bm->mark(addr); 5895 if (!_cms_bm->isMarked(addr)) { 5896 Log(gc, verify) log; 5897 ResourceMark rm; 5898 LogStream ls(log.error()); 5899 oop(addr)->print_on(&ls); 5900 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5901 fatal("... aborting"); 5902 } 5903 } 5904 } 5905 5906 ////////////////////////////////////////////////// 5907 // MarkRefsIntoAndScanClosure 5908 ////////////////////////////////////////////////// 5909 5910 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5911 ReferenceDiscoverer* rd, 5912 CMSBitMap* bit_map, 5913 CMSBitMap* mod_union_table, 5914 CMSMarkStack* mark_stack, 5915 CMSCollector* collector, 5916 bool should_yield, 5917 bool concurrent_precleaning): 5918 _span(span), 5919 _bit_map(bit_map), 5920 _mark_stack(mark_stack), 5921 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5922 mark_stack, concurrent_precleaning), 5923 _collector(collector), 5924 _freelistLock(NULL), 5925 _yield(should_yield), 5926 _concurrent_precleaning(concurrent_precleaning) 5927 { 5928 // FIXME: Should initialize in base class constructor. 5929 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5930 set_ref_discoverer_internal(rd); 5931 } 5932 5933 // This closure is used to mark refs into the CMS generation at the 5934 // second (final) checkpoint, and to scan and transitively follow 5935 // the unmarked oops. It is also used during the concurrent precleaning 5936 // phase while scanning objects on dirty cards in the CMS generation. 5937 // The marks are made in the marking bit map and the marking stack is 5938 // used for keeping the (newly) grey objects during the scan. 5939 // The parallel version (Par_...) appears further below. 5940 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5941 if (obj != NULL) { 5942 assert(oopDesc::is_oop(obj), "expected an oop"); 5943 HeapWord* addr = (HeapWord*)obj; 5944 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5945 assert(_collector->overflow_list_is_empty(), 5946 "overflow list should be empty"); 5947 if (_span.contains(addr) && 5948 !_bit_map->isMarked(addr)) { 5949 // mark bit map (object is now grey) 5950 _bit_map->mark(addr); 5951 // push on marking stack (stack should be empty), and drain the 5952 // stack by applying this closure to the oops in the oops popped 5953 // from the stack (i.e. blacken the grey objects) 5954 bool res = _mark_stack->push(obj); 5955 assert(res, "Should have space to push on empty stack"); 5956 do { 5957 oop new_oop = _mark_stack->pop(); 5958 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5959 assert(_bit_map->isMarked((HeapWord*)new_oop), 5960 "only grey objects on this stack"); 5961 // iterate over the oops in this oop, marking and pushing 5962 // the ones in CMS heap (i.e. in _span). 5963 new_oop->oop_iterate(&_pushAndMarkClosure); 5964 // check if it's time to yield 5965 do_yield_check(); 5966 } while (!_mark_stack->isEmpty() || 5967 (!_concurrent_precleaning && take_from_overflow_list())); 5968 // if marking stack is empty, and we are not doing this 5969 // during precleaning, then check the overflow list 5970 } 5971 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5972 assert(_collector->overflow_list_is_empty(), 5973 "overflow list was drained above"); 5974 5975 assert(_collector->no_preserved_marks(), 5976 "All preserved marks should have been restored above"); 5977 } 5978 } 5979 5980 void MarkRefsIntoAndScanClosure::do_yield_work() { 5981 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5982 "CMS thread should hold CMS token"); 5983 assert_lock_strong(_freelistLock); 5984 assert_lock_strong(_bit_map->lock()); 5985 // relinquish the free_list_lock and bitMaplock() 5986 _bit_map->lock()->unlock(); 5987 _freelistLock->unlock(); 5988 ConcurrentMarkSweepThread::desynchronize(true); 5989 _collector->stopTimer(); 5990 _collector->incrementYields(); 5991 5992 // See the comment in coordinator_yield() 5993 for (unsigned i = 0; 5994 i < CMSYieldSleepCount && 5995 ConcurrentMarkSweepThread::should_yield() && 5996 !CMSCollector::foregroundGCIsActive(); 5997 ++i) { 5998 os::sleep(Thread::current(), 1, false); 5999 } 6000 6001 ConcurrentMarkSweepThread::synchronize(true); 6002 _freelistLock->lock_without_safepoint_check(); 6003 _bit_map->lock()->lock_without_safepoint_check(); 6004 _collector->startTimer(); 6005 } 6006 6007 /////////////////////////////////////////////////////////// 6008 // ParMarkRefsIntoAndScanClosure: a parallel version of 6009 // MarkRefsIntoAndScanClosure 6010 /////////////////////////////////////////////////////////// 6011 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6012 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 6013 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6014 _span(span), 6015 _bit_map(bit_map), 6016 _work_queue(work_queue), 6017 _low_water_mark(MIN2((work_queue->max_elems()/4), 6018 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6019 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 6020 { 6021 // FIXME: Should initialize in base class constructor. 6022 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 6023 set_ref_discoverer_internal(rd); 6024 } 6025 6026 // This closure is used to mark refs into the CMS generation at the 6027 // second (final) checkpoint, and to scan and transitively follow 6028 // the unmarked oops. The marks are made in the marking bit map and 6029 // the work_queue is used for keeping the (newly) grey objects during 6030 // the scan phase whence they are also available for stealing by parallel 6031 // threads. Since the marking bit map is shared, updates are 6032 // synchronized (via CAS). 6033 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6034 if (obj != NULL) { 6035 // Ignore mark word because this could be an already marked oop 6036 // that may be chained at the end of the overflow list. 6037 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6038 HeapWord* addr = (HeapWord*)obj; 6039 if (_span.contains(addr) && 6040 !_bit_map->isMarked(addr)) { 6041 // mark bit map (object will become grey): 6042 // It is possible for several threads to be 6043 // trying to "claim" this object concurrently; 6044 // the unique thread that succeeds in marking the 6045 // object first will do the subsequent push on 6046 // to the work queue (or overflow list). 6047 if (_bit_map->par_mark(addr)) { 6048 // push on work_queue (which may not be empty), and trim the 6049 // queue to an appropriate length by applying this closure to 6050 // the oops in the oops popped from the stack (i.e. blacken the 6051 // grey objects) 6052 bool res = _work_queue->push(obj); 6053 assert(res, "Low water mark should be less than capacity?"); 6054 trim_queue(_low_water_mark); 6055 } // Else, another thread claimed the object 6056 } 6057 } 6058 } 6059 6060 // This closure is used to rescan the marked objects on the dirty cards 6061 // in the mod union table and the card table proper. 6062 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6063 oop p, MemRegion mr) { 6064 6065 size_t size = 0; 6066 HeapWord* addr = (HeapWord*)p; 6067 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6068 assert(_span.contains(addr), "we are scanning the CMS generation"); 6069 // check if it's time to yield 6070 if (do_yield_check()) { 6071 // We yielded for some foreground stop-world work, 6072 // and we have been asked to abort this ongoing preclean cycle. 6073 return 0; 6074 } 6075 if (_bitMap->isMarked(addr)) { 6076 // it's marked; is it potentially uninitialized? 6077 if (p->klass_or_null_acquire() != NULL) { 6078 // an initialized object; ignore mark word in verification below 6079 // since we are running concurrent with mutators 6080 assert(oopDesc::is_oop(p, true), "should be an oop"); 6081 if (p->is_objArray()) { 6082 // objArrays are precisely marked; restrict scanning 6083 // to dirty cards only. 6084 size = CompactibleFreeListSpace::adjustObjectSize( 6085 p->oop_iterate_size(_scanningClosure, mr)); 6086 } else { 6087 // A non-array may have been imprecisely marked; we need 6088 // to scan object in its entirety. 6089 size = CompactibleFreeListSpace::adjustObjectSize( 6090 p->oop_iterate_size(_scanningClosure)); 6091 } 6092 #ifdef ASSERT 6093 size_t direct_size = 6094 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6095 assert(size == direct_size, "Inconsistency in size"); 6096 assert(size >= 3, "Necessary for Printezis marks to work"); 6097 HeapWord* start_pbit = addr + 1; 6098 HeapWord* end_pbit = addr + size - 1; 6099 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6100 "inconsistent Printezis mark"); 6101 // Verify inner mark bits (between Printezis bits) are clear, 6102 // but don't repeat if there are multiple dirty regions for 6103 // the same object, to avoid potential O(N^2) performance. 6104 if (addr != _last_scanned_object) { 6105 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6106 _last_scanned_object = addr; 6107 } 6108 #endif // ASSERT 6109 } else { 6110 // An uninitialized object. 6111 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6112 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6113 size = pointer_delta(nextOneAddr + 1, addr); 6114 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6115 "alignment problem"); 6116 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6117 // will dirty the card when the klass pointer is installed in the 6118 // object (signaling the completion of initialization). 6119 } 6120 } else { 6121 // Either a not yet marked object or an uninitialized object 6122 if (p->klass_or_null_acquire() == NULL) { 6123 // An uninitialized object, skip to the next card, since 6124 // we may not be able to read its P-bits yet. 6125 assert(size == 0, "Initial value"); 6126 } else { 6127 // An object not (yet) reached by marking: we merely need to 6128 // compute its size so as to go look at the next block. 6129 assert(oopDesc::is_oop(p, true), "should be an oop"); 6130 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6131 } 6132 } 6133 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6134 return size; 6135 } 6136 6137 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6138 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6139 "CMS thread should hold CMS token"); 6140 assert_lock_strong(_freelistLock); 6141 assert_lock_strong(_bitMap->lock()); 6142 // relinquish the free_list_lock and bitMaplock() 6143 _bitMap->lock()->unlock(); 6144 _freelistLock->unlock(); 6145 ConcurrentMarkSweepThread::desynchronize(true); 6146 _collector->stopTimer(); 6147 _collector->incrementYields(); 6148 6149 // See the comment in coordinator_yield() 6150 for (unsigned i = 0; i < CMSYieldSleepCount && 6151 ConcurrentMarkSweepThread::should_yield() && 6152 !CMSCollector::foregroundGCIsActive(); ++i) { 6153 os::sleep(Thread::current(), 1, false); 6154 } 6155 6156 ConcurrentMarkSweepThread::synchronize(true); 6157 _freelistLock->lock_without_safepoint_check(); 6158 _bitMap->lock()->lock_without_safepoint_check(); 6159 _collector->startTimer(); 6160 } 6161 6162 6163 ////////////////////////////////////////////////////////////////// 6164 // SurvivorSpacePrecleanClosure 6165 ////////////////////////////////////////////////////////////////// 6166 // This (single-threaded) closure is used to preclean the oops in 6167 // the survivor spaces. 6168 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6169 6170 HeapWord* addr = (HeapWord*)p; 6171 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6172 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6173 assert(p->klass_or_null() != NULL, "object should be initialized"); 6174 // an initialized object; ignore mark word in verification below 6175 // since we are running concurrent with mutators 6176 assert(oopDesc::is_oop(p, true), "should be an oop"); 6177 // Note that we do not yield while we iterate over 6178 // the interior oops of p, pushing the relevant ones 6179 // on our marking stack. 6180 size_t size = p->oop_iterate_size(_scanning_closure); 6181 do_yield_check(); 6182 // Observe that below, we do not abandon the preclean 6183 // phase as soon as we should; rather we empty the 6184 // marking stack before returning. This is to satisfy 6185 // some existing assertions. In general, it may be a 6186 // good idea to abort immediately and complete the marking 6187 // from the grey objects at a later time. 6188 while (!_mark_stack->isEmpty()) { 6189 oop new_oop = _mark_stack->pop(); 6190 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6191 assert(_bit_map->isMarked((HeapWord*)new_oop), 6192 "only grey objects on this stack"); 6193 // iterate over the oops in this oop, marking and pushing 6194 // the ones in CMS heap (i.e. in _span). 6195 new_oop->oop_iterate(_scanning_closure); 6196 // check if it's time to yield 6197 do_yield_check(); 6198 } 6199 unsigned int after_count = 6200 CMSHeap::heap()->total_collections(); 6201 bool abort = (_before_count != after_count) || 6202 _collector->should_abort_preclean(); 6203 return abort ? 0 : size; 6204 } 6205 6206 void SurvivorSpacePrecleanClosure::do_yield_work() { 6207 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6208 "CMS thread should hold CMS token"); 6209 assert_lock_strong(_bit_map->lock()); 6210 // Relinquish the bit map lock 6211 _bit_map->lock()->unlock(); 6212 ConcurrentMarkSweepThread::desynchronize(true); 6213 _collector->stopTimer(); 6214 _collector->incrementYields(); 6215 6216 // See the comment in coordinator_yield() 6217 for (unsigned i = 0; i < CMSYieldSleepCount && 6218 ConcurrentMarkSweepThread::should_yield() && 6219 !CMSCollector::foregroundGCIsActive(); ++i) { 6220 os::sleep(Thread::current(), 1, false); 6221 } 6222 6223 ConcurrentMarkSweepThread::synchronize(true); 6224 _bit_map->lock()->lock_without_safepoint_check(); 6225 _collector->startTimer(); 6226 } 6227 6228 // This closure is used to rescan the marked objects on the dirty cards 6229 // in the mod union table and the card table proper. In the parallel 6230 // case, although the bitMap is shared, we do a single read so the 6231 // isMarked() query is "safe". 6232 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6233 // Ignore mark word because we are running concurrent with mutators 6234 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6235 HeapWord* addr = (HeapWord*)p; 6236 assert(_span.contains(addr), "we are scanning the CMS generation"); 6237 bool is_obj_array = false; 6238 #ifdef ASSERT 6239 if (!_parallel) { 6240 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6241 assert(_collector->overflow_list_is_empty(), 6242 "overflow list should be empty"); 6243 6244 } 6245 #endif // ASSERT 6246 if (_bit_map->isMarked(addr)) { 6247 // Obj arrays are precisely marked, non-arrays are not; 6248 // so we scan objArrays precisely and non-arrays in their 6249 // entirety. 6250 if (p->is_objArray()) { 6251 is_obj_array = true; 6252 if (_parallel) { 6253 p->oop_iterate(_par_scan_closure, mr); 6254 } else { 6255 p->oop_iterate(_scan_closure, mr); 6256 } 6257 } else { 6258 if (_parallel) { 6259 p->oop_iterate(_par_scan_closure); 6260 } else { 6261 p->oop_iterate(_scan_closure); 6262 } 6263 } 6264 } 6265 #ifdef ASSERT 6266 if (!_parallel) { 6267 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6268 assert(_collector->overflow_list_is_empty(), 6269 "overflow list should be empty"); 6270 6271 } 6272 #endif // ASSERT 6273 return is_obj_array; 6274 } 6275 6276 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6277 MemRegion span, 6278 CMSBitMap* bitMap, CMSMarkStack* markStack, 6279 bool should_yield, bool verifying): 6280 _collector(collector), 6281 _span(span), 6282 _bitMap(bitMap), 6283 _mut(&collector->_modUnionTable), 6284 _markStack(markStack), 6285 _yield(should_yield), 6286 _skipBits(0) 6287 { 6288 assert(_markStack->isEmpty(), "stack should be empty"); 6289 _finger = _bitMap->startWord(); 6290 _threshold = _finger; 6291 assert(_collector->_restart_addr == NULL, "Sanity check"); 6292 assert(_span.contains(_finger), "Out of bounds _finger?"); 6293 DEBUG_ONLY(_verifying = verifying;) 6294 } 6295 6296 void MarkFromRootsClosure::reset(HeapWord* addr) { 6297 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6298 assert(_span.contains(addr), "Out of bounds _finger?"); 6299 _finger = addr; 6300 _threshold = align_up(_finger, CardTable::card_size); 6301 } 6302 6303 // Should revisit to see if this should be restructured for 6304 // greater efficiency. 6305 bool MarkFromRootsClosure::do_bit(size_t offset) { 6306 if (_skipBits > 0) { 6307 _skipBits--; 6308 return true; 6309 } 6310 // convert offset into a HeapWord* 6311 HeapWord* addr = _bitMap->startWord() + offset; 6312 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6313 "address out of range"); 6314 assert(_bitMap->isMarked(addr), "tautology"); 6315 if (_bitMap->isMarked(addr+1)) { 6316 // this is an allocated but not yet initialized object 6317 assert(_skipBits == 0, "tautology"); 6318 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6319 oop p = oop(addr); 6320 if (p->klass_or_null_acquire() == NULL) { 6321 DEBUG_ONLY(if (!_verifying) {) 6322 // We re-dirty the cards on which this object lies and increase 6323 // the _threshold so that we'll come back to scan this object 6324 // during the preclean or remark phase. (CMSCleanOnEnter) 6325 if (CMSCleanOnEnter) { 6326 size_t sz = _collector->block_size_using_printezis_bits(addr); 6327 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6328 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6329 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6330 // Bump _threshold to end_card_addr; note that 6331 // _threshold cannot possibly exceed end_card_addr, anyhow. 6332 // This prevents future clearing of the card as the scan proceeds 6333 // to the right. 6334 assert(_threshold <= end_card_addr, 6335 "Because we are just scanning into this object"); 6336 if (_threshold < end_card_addr) { 6337 _threshold = end_card_addr; 6338 } 6339 if (p->klass_or_null_acquire() != NULL) { 6340 // Redirty the range of cards... 6341 _mut->mark_range(redirty_range); 6342 } // ...else the setting of klass will dirty the card anyway. 6343 } 6344 DEBUG_ONLY(}) 6345 return true; 6346 } 6347 } 6348 scanOopsInOop(addr); 6349 return true; 6350 } 6351 6352 // We take a break if we've been at this for a while, 6353 // so as to avoid monopolizing the locks involved. 6354 void MarkFromRootsClosure::do_yield_work() { 6355 // First give up the locks, then yield, then re-lock 6356 // We should probably use a constructor/destructor idiom to 6357 // do this unlock/lock or modify the MutexUnlocker class to 6358 // serve our purpose. XXX 6359 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6360 "CMS thread should hold CMS token"); 6361 assert_lock_strong(_bitMap->lock()); 6362 _bitMap->lock()->unlock(); 6363 ConcurrentMarkSweepThread::desynchronize(true); 6364 _collector->stopTimer(); 6365 _collector->incrementYields(); 6366 6367 // See the comment in coordinator_yield() 6368 for (unsigned i = 0; i < CMSYieldSleepCount && 6369 ConcurrentMarkSweepThread::should_yield() && 6370 !CMSCollector::foregroundGCIsActive(); ++i) { 6371 os::sleep(Thread::current(), 1, false); 6372 } 6373 6374 ConcurrentMarkSweepThread::synchronize(true); 6375 _bitMap->lock()->lock_without_safepoint_check(); 6376 _collector->startTimer(); 6377 } 6378 6379 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6380 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6381 assert(_markStack->isEmpty(), 6382 "should drain stack to limit stack usage"); 6383 // convert ptr to an oop preparatory to scanning 6384 oop obj = oop(ptr); 6385 // Ignore mark word in verification below, since we 6386 // may be running concurrent with mutators. 6387 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6388 assert(_finger <= ptr, "_finger runneth ahead"); 6389 // advance the finger to right end of this object 6390 _finger = ptr + obj->size(); 6391 assert(_finger > ptr, "we just incremented it above"); 6392 // On large heaps, it may take us some time to get through 6393 // the marking phase. During 6394 // this time it's possible that a lot of mutations have 6395 // accumulated in the card table and the mod union table -- 6396 // these mutation records are redundant until we have 6397 // actually traced into the corresponding card. 6398 // Here, we check whether advancing the finger would make 6399 // us cross into a new card, and if so clear corresponding 6400 // cards in the MUT (preclean them in the card-table in the 6401 // future). 6402 6403 DEBUG_ONLY(if (!_verifying) {) 6404 // The clean-on-enter optimization is disabled by default, 6405 // until we fix 6178663. 6406 if (CMSCleanOnEnter && (_finger > _threshold)) { 6407 // [_threshold, _finger) represents the interval 6408 // of cards to be cleared in MUT (or precleaned in card table). 6409 // The set of cards to be cleared is all those that overlap 6410 // with the interval [_threshold, _finger); note that 6411 // _threshold is always kept card-aligned but _finger isn't 6412 // always card-aligned. 6413 HeapWord* old_threshold = _threshold; 6414 assert(is_aligned(old_threshold, CardTable::card_size), 6415 "_threshold should always be card-aligned"); 6416 _threshold = align_up(_finger, CardTable::card_size); 6417 MemRegion mr(old_threshold, _threshold); 6418 assert(!mr.is_empty(), "Control point invariant"); 6419 assert(_span.contains(mr), "Should clear within span"); 6420 _mut->clear_range(mr); 6421 } 6422 DEBUG_ONLY(}) 6423 // Note: the finger doesn't advance while we drain 6424 // the stack below. 6425 PushOrMarkClosure pushOrMarkClosure(_collector, 6426 _span, _bitMap, _markStack, 6427 _finger, this); 6428 bool res = _markStack->push(obj); 6429 assert(res, "Empty non-zero size stack should have space for single push"); 6430 while (!_markStack->isEmpty()) { 6431 oop new_oop = _markStack->pop(); 6432 // Skip verifying header mark word below because we are 6433 // running concurrent with mutators. 6434 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6435 // now scan this oop's oops 6436 new_oop->oop_iterate(&pushOrMarkClosure); 6437 do_yield_check(); 6438 } 6439 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6440 } 6441 6442 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6443 CMSCollector* collector, MemRegion span, 6444 CMSBitMap* bit_map, 6445 OopTaskQueue* work_queue, 6446 CMSMarkStack* overflow_stack): 6447 _collector(collector), 6448 _whole_span(collector->_span), 6449 _span(span), 6450 _bit_map(bit_map), 6451 _mut(&collector->_modUnionTable), 6452 _work_queue(work_queue), 6453 _overflow_stack(overflow_stack), 6454 _skip_bits(0), 6455 _task(task) 6456 { 6457 assert(_work_queue->size() == 0, "work_queue should be empty"); 6458 _finger = span.start(); 6459 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6460 assert(_span.contains(_finger), "Out of bounds _finger?"); 6461 } 6462 6463 // Should revisit to see if this should be restructured for 6464 // greater efficiency. 6465 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6466 if (_skip_bits > 0) { 6467 _skip_bits--; 6468 return true; 6469 } 6470 // convert offset into a HeapWord* 6471 HeapWord* addr = _bit_map->startWord() + offset; 6472 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6473 "address out of range"); 6474 assert(_bit_map->isMarked(addr), "tautology"); 6475 if (_bit_map->isMarked(addr+1)) { 6476 // this is an allocated object that might not yet be initialized 6477 assert(_skip_bits == 0, "tautology"); 6478 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6479 oop p = oop(addr); 6480 if (p->klass_or_null_acquire() == NULL) { 6481 // in the case of Clean-on-Enter optimization, redirty card 6482 // and avoid clearing card by increasing the threshold. 6483 return true; 6484 } 6485 } 6486 scan_oops_in_oop(addr); 6487 return true; 6488 } 6489 6490 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6491 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6492 // Should we assert that our work queue is empty or 6493 // below some drain limit? 6494 assert(_work_queue->size() == 0, 6495 "should drain stack to limit stack usage"); 6496 // convert ptr to an oop preparatory to scanning 6497 oop obj = oop(ptr); 6498 // Ignore mark word in verification below, since we 6499 // may be running concurrent with mutators. 6500 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6501 assert(_finger <= ptr, "_finger runneth ahead"); 6502 // advance the finger to right end of this object 6503 _finger = ptr + obj->size(); 6504 assert(_finger > ptr, "we just incremented it above"); 6505 // On large heaps, it may take us some time to get through 6506 // the marking phase. During 6507 // this time it's possible that a lot of mutations have 6508 // accumulated in the card table and the mod union table -- 6509 // these mutation records are redundant until we have 6510 // actually traced into the corresponding card. 6511 // Here, we check whether advancing the finger would make 6512 // us cross into a new card, and if so clear corresponding 6513 // cards in the MUT (preclean them in the card-table in the 6514 // future). 6515 6516 // The clean-on-enter optimization is disabled by default, 6517 // until we fix 6178663. 6518 if (CMSCleanOnEnter && (_finger > _threshold)) { 6519 // [_threshold, _finger) represents the interval 6520 // of cards to be cleared in MUT (or precleaned in card table). 6521 // The set of cards to be cleared is all those that overlap 6522 // with the interval [_threshold, _finger); note that 6523 // _threshold is always kept card-aligned but _finger isn't 6524 // always card-aligned. 6525 HeapWord* old_threshold = _threshold; 6526 assert(is_aligned(old_threshold, CardTable::card_size), 6527 "_threshold should always be card-aligned"); 6528 _threshold = align_up(_finger, CardTable::card_size); 6529 MemRegion mr(old_threshold, _threshold); 6530 assert(!mr.is_empty(), "Control point invariant"); 6531 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6532 _mut->clear_range(mr); 6533 } 6534 6535 // Note: the local finger doesn't advance while we drain 6536 // the stack below, but the global finger sure can and will. 6537 HeapWord* volatile* gfa = _task->global_finger_addr(); 6538 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6539 _span, _bit_map, 6540 _work_queue, 6541 _overflow_stack, 6542 _finger, 6543 gfa, this); 6544 bool res = _work_queue->push(obj); // overflow could occur here 6545 assert(res, "Will hold once we use workqueues"); 6546 while (true) { 6547 oop new_oop; 6548 if (!_work_queue->pop_local(new_oop)) { 6549 // We emptied our work_queue; check if there's stuff that can 6550 // be gotten from the overflow stack. 6551 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6552 _overflow_stack, _work_queue)) { 6553 do_yield_check(); 6554 continue; 6555 } else { // done 6556 break; 6557 } 6558 } 6559 // Skip verifying header mark word below because we are 6560 // running concurrent with mutators. 6561 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6562 // now scan this oop's oops 6563 new_oop->oop_iterate(&pushOrMarkClosure); 6564 do_yield_check(); 6565 } 6566 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6567 } 6568 6569 // Yield in response to a request from VM Thread or 6570 // from mutators. 6571 void ParMarkFromRootsClosure::do_yield_work() { 6572 assert(_task != NULL, "sanity"); 6573 _task->yield(); 6574 } 6575 6576 // A variant of the above used for verifying CMS marking work. 6577 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6578 MemRegion span, 6579 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6580 CMSMarkStack* mark_stack): 6581 _collector(collector), 6582 _span(span), 6583 _verification_bm(verification_bm), 6584 _cms_bm(cms_bm), 6585 _mark_stack(mark_stack), 6586 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6587 mark_stack) 6588 { 6589 assert(_mark_stack->isEmpty(), "stack should be empty"); 6590 _finger = _verification_bm->startWord(); 6591 assert(_collector->_restart_addr == NULL, "Sanity check"); 6592 assert(_span.contains(_finger), "Out of bounds _finger?"); 6593 } 6594 6595 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6596 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6597 assert(_span.contains(addr), "Out of bounds _finger?"); 6598 _finger = addr; 6599 } 6600 6601 // Should revisit to see if this should be restructured for 6602 // greater efficiency. 6603 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6604 // convert offset into a HeapWord* 6605 HeapWord* addr = _verification_bm->startWord() + offset; 6606 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6607 "address out of range"); 6608 assert(_verification_bm->isMarked(addr), "tautology"); 6609 assert(_cms_bm->isMarked(addr), "tautology"); 6610 6611 assert(_mark_stack->isEmpty(), 6612 "should drain stack to limit stack usage"); 6613 // convert addr to an oop preparatory to scanning 6614 oop obj = oop(addr); 6615 assert(oopDesc::is_oop(obj), "should be an oop"); 6616 assert(_finger <= addr, "_finger runneth ahead"); 6617 // advance the finger to right end of this object 6618 _finger = addr + obj->size(); 6619 assert(_finger > addr, "we just incremented it above"); 6620 // Note: the finger doesn't advance while we drain 6621 // the stack below. 6622 bool res = _mark_stack->push(obj); 6623 assert(res, "Empty non-zero size stack should have space for single push"); 6624 while (!_mark_stack->isEmpty()) { 6625 oop new_oop = _mark_stack->pop(); 6626 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6627 // now scan this oop's oops 6628 new_oop->oop_iterate(&_pam_verify_closure); 6629 } 6630 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6631 return true; 6632 } 6633 6634 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6635 CMSCollector* collector, MemRegion span, 6636 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6637 CMSMarkStack* mark_stack): 6638 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6639 _collector(collector), 6640 _span(span), 6641 _verification_bm(verification_bm), 6642 _cms_bm(cms_bm), 6643 _mark_stack(mark_stack) 6644 { } 6645 6646 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6647 oop obj = RawAccess<>::oop_load(p); 6648 do_oop(obj); 6649 } 6650 6651 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6652 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6653 6654 // Upon stack overflow, we discard (part of) the stack, 6655 // remembering the least address amongst those discarded 6656 // in CMSCollector's _restart_address. 6657 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6658 // Remember the least grey address discarded 6659 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6660 _collector->lower_restart_addr(ra); 6661 _mark_stack->reset(); // discard stack contents 6662 _mark_stack->expand(); // expand the stack if possible 6663 } 6664 6665 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6666 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6667 HeapWord* addr = (HeapWord*)obj; 6668 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6669 // Oop lies in _span and isn't yet grey or black 6670 _verification_bm->mark(addr); // now grey 6671 if (!_cms_bm->isMarked(addr)) { 6672 Log(gc, verify) log; 6673 ResourceMark rm; 6674 LogStream ls(log.error()); 6675 oop(addr)->print_on(&ls); 6676 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6677 fatal("... aborting"); 6678 } 6679 6680 if (!_mark_stack->push(obj)) { // stack overflow 6681 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6682 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6683 handle_stack_overflow(addr); 6684 } 6685 // anything including and to the right of _finger 6686 // will be scanned as we iterate over the remainder of the 6687 // bit map 6688 } 6689 } 6690 6691 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6692 MemRegion span, 6693 CMSBitMap* bitMap, CMSMarkStack* markStack, 6694 HeapWord* finger, MarkFromRootsClosure* parent) : 6695 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6696 _collector(collector), 6697 _span(span), 6698 _bitMap(bitMap), 6699 _markStack(markStack), 6700 _finger(finger), 6701 _parent(parent) 6702 { } 6703 6704 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6705 MemRegion span, 6706 CMSBitMap* bit_map, 6707 OopTaskQueue* work_queue, 6708 CMSMarkStack* overflow_stack, 6709 HeapWord* finger, 6710 HeapWord* volatile* global_finger_addr, 6711 ParMarkFromRootsClosure* parent) : 6712 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6713 _collector(collector), 6714 _whole_span(collector->_span), 6715 _span(span), 6716 _bit_map(bit_map), 6717 _work_queue(work_queue), 6718 _overflow_stack(overflow_stack), 6719 _finger(finger), 6720 _global_finger_addr(global_finger_addr), 6721 _parent(parent) 6722 { } 6723 6724 // Assumes thread-safe access by callers, who are 6725 // responsible for mutual exclusion. 6726 void CMSCollector::lower_restart_addr(HeapWord* low) { 6727 assert(_span.contains(low), "Out of bounds addr"); 6728 if (_restart_addr == NULL) { 6729 _restart_addr = low; 6730 } else { 6731 _restart_addr = MIN2(_restart_addr, low); 6732 } 6733 } 6734 6735 // Upon stack overflow, we discard (part of) the stack, 6736 // remembering the least address amongst those discarded 6737 // in CMSCollector's _restart_address. 6738 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6739 // Remember the least grey address discarded 6740 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6741 _collector->lower_restart_addr(ra); 6742 _markStack->reset(); // discard stack contents 6743 _markStack->expand(); // expand the stack if possible 6744 } 6745 6746 // Upon stack overflow, we discard (part of) the stack, 6747 // remembering the least address amongst those discarded 6748 // in CMSCollector's _restart_address. 6749 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6750 // We need to do this under a mutex to prevent other 6751 // workers from interfering with the work done below. 6752 MutexLocker ml(_overflow_stack->par_lock(), 6753 Mutex::_no_safepoint_check_flag); 6754 // Remember the least grey address discarded 6755 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6756 _collector->lower_restart_addr(ra); 6757 _overflow_stack->reset(); // discard stack contents 6758 _overflow_stack->expand(); // expand the stack if possible 6759 } 6760 6761 void PushOrMarkClosure::do_oop(oop obj) { 6762 // Ignore mark word because we are running concurrent with mutators. 6763 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6764 HeapWord* addr = (HeapWord*)obj; 6765 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6766 // Oop lies in _span and isn't yet grey or black 6767 _bitMap->mark(addr); // now grey 6768 if (addr < _finger) { 6769 // the bit map iteration has already either passed, or 6770 // sampled, this bit in the bit map; we'll need to 6771 // use the marking stack to scan this oop's oops. 6772 bool simulate_overflow = false; 6773 NOT_PRODUCT( 6774 if (CMSMarkStackOverflowALot && 6775 _collector->simulate_overflow()) { 6776 // simulate a stack overflow 6777 simulate_overflow = true; 6778 } 6779 ) 6780 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6781 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6782 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6783 handle_stack_overflow(addr); 6784 } 6785 } 6786 // anything including and to the right of _finger 6787 // will be scanned as we iterate over the remainder of the 6788 // bit map 6789 do_yield_check(); 6790 } 6791 } 6792 6793 void ParPushOrMarkClosure::do_oop(oop obj) { 6794 // Ignore mark word because we are running concurrent with mutators. 6795 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6796 HeapWord* addr = (HeapWord*)obj; 6797 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6798 // Oop lies in _span and isn't yet grey or black 6799 // We read the global_finger (volatile read) strictly after marking oop 6800 bool res = _bit_map->par_mark(addr); // now grey 6801 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6802 // Should we push this marked oop on our stack? 6803 // -- if someone else marked it, nothing to do 6804 // -- if target oop is above global finger nothing to do 6805 // -- if target oop is in chunk and above local finger 6806 // then nothing to do 6807 // -- else push on work queue 6808 if ( !res // someone else marked it, they will deal with it 6809 || (addr >= *gfa) // will be scanned in a later task 6810 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6811 return; 6812 } 6813 // the bit map iteration has already either passed, or 6814 // sampled, this bit in the bit map; we'll need to 6815 // use the marking stack to scan this oop's oops. 6816 bool simulate_overflow = false; 6817 NOT_PRODUCT( 6818 if (CMSMarkStackOverflowALot && 6819 _collector->simulate_overflow()) { 6820 // simulate a stack overflow 6821 simulate_overflow = true; 6822 } 6823 ) 6824 if (simulate_overflow || 6825 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6826 // stack overflow 6827 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6828 // We cannot assert that the overflow stack is full because 6829 // it may have been emptied since. 6830 assert(simulate_overflow || 6831 _work_queue->size() == _work_queue->max_elems(), 6832 "Else push should have succeeded"); 6833 handle_stack_overflow(addr); 6834 } 6835 do_yield_check(); 6836 } 6837 } 6838 6839 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6840 MemRegion span, 6841 ReferenceDiscoverer* rd, 6842 CMSBitMap* bit_map, 6843 CMSBitMap* mod_union_table, 6844 CMSMarkStack* mark_stack, 6845 bool concurrent_precleaning): 6846 MetadataVisitingOopIterateClosure(rd), 6847 _collector(collector), 6848 _span(span), 6849 _bit_map(bit_map), 6850 _mod_union_table(mod_union_table), 6851 _mark_stack(mark_stack), 6852 _concurrent_precleaning(concurrent_precleaning) 6853 { 6854 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6855 } 6856 6857 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6858 // the non-parallel version (the parallel version appears further below.) 6859 void PushAndMarkClosure::do_oop(oop obj) { 6860 // Ignore mark word verification. If during concurrent precleaning, 6861 // the object monitor may be locked. If during the checkpoint 6862 // phases, the object may already have been reached by a different 6863 // path and may be at the end of the global overflow list (so 6864 // the mark word may be NULL). 6865 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6866 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6867 HeapWord* addr = (HeapWord*)obj; 6868 // Check if oop points into the CMS generation 6869 // and is not marked 6870 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6871 // a white object ... 6872 _bit_map->mark(addr); // ... now grey 6873 // push on the marking stack (grey set) 6874 bool simulate_overflow = false; 6875 NOT_PRODUCT( 6876 if (CMSMarkStackOverflowALot && 6877 _collector->simulate_overflow()) { 6878 // simulate a stack overflow 6879 simulate_overflow = true; 6880 } 6881 ) 6882 if (simulate_overflow || !_mark_stack->push(obj)) { 6883 if (_concurrent_precleaning) { 6884 // During precleaning we can just dirty the appropriate card(s) 6885 // in the mod union table, thus ensuring that the object remains 6886 // in the grey set and continue. In the case of object arrays 6887 // we need to dirty all of the cards that the object spans, 6888 // since the rescan of object arrays will be limited to the 6889 // dirty cards. 6890 // Note that no one can be interfering with us in this action 6891 // of dirtying the mod union table, so no locking or atomics 6892 // are required. 6893 if (obj->is_objArray()) { 6894 size_t sz = obj->size(); 6895 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6896 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6897 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6898 _mod_union_table->mark_range(redirty_range); 6899 } else { 6900 _mod_union_table->mark(addr); 6901 } 6902 _collector->_ser_pmc_preclean_ovflw++; 6903 } else { 6904 // During the remark phase, we need to remember this oop 6905 // in the overflow list. 6906 _collector->push_on_overflow_list(obj); 6907 _collector->_ser_pmc_remark_ovflw++; 6908 } 6909 } 6910 } 6911 } 6912 6913 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6914 MemRegion span, 6915 ReferenceDiscoverer* rd, 6916 CMSBitMap* bit_map, 6917 OopTaskQueue* work_queue): 6918 MetadataVisitingOopIterateClosure(rd), 6919 _collector(collector), 6920 _span(span), 6921 _bit_map(bit_map), 6922 _work_queue(work_queue) 6923 { 6924 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6925 } 6926 6927 // Grey object rescan during second checkpoint phase -- 6928 // the parallel version. 6929 void ParPushAndMarkClosure::do_oop(oop obj) { 6930 // In the assert below, we ignore the mark word because 6931 // this oop may point to an already visited object that is 6932 // on the overflow stack (in which case the mark word has 6933 // been hijacked for chaining into the overflow stack -- 6934 // if this is the last object in the overflow stack then 6935 // its mark word will be NULL). Because this object may 6936 // have been subsequently popped off the global overflow 6937 // stack, and the mark word possibly restored to the prototypical 6938 // value, by the time we get to examined this failing assert in 6939 // the debugger, is_oop_or_null(false) may subsequently start 6940 // to hold. 6941 assert(oopDesc::is_oop_or_null(obj, true), 6942 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6943 HeapWord* addr = (HeapWord*)obj; 6944 // Check if oop points into the CMS generation 6945 // and is not marked 6946 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6947 // a white object ... 6948 // If we manage to "claim" the object, by being the 6949 // first thread to mark it, then we push it on our 6950 // marking stack 6951 if (_bit_map->par_mark(addr)) { // ... now grey 6952 // push on work queue (grey set) 6953 bool simulate_overflow = false; 6954 NOT_PRODUCT( 6955 if (CMSMarkStackOverflowALot && 6956 _collector->par_simulate_overflow()) { 6957 // simulate a stack overflow 6958 simulate_overflow = true; 6959 } 6960 ) 6961 if (simulate_overflow || !_work_queue->push(obj)) { 6962 _collector->par_push_on_overflow_list(obj); 6963 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6964 } 6965 } // Else, some other thread got there first 6966 } 6967 } 6968 6969 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6970 Mutex* bml = _collector->bitMapLock(); 6971 assert_lock_strong(bml); 6972 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6973 "CMS thread should hold CMS token"); 6974 6975 bml->unlock(); 6976 ConcurrentMarkSweepThread::desynchronize(true); 6977 6978 _collector->stopTimer(); 6979 _collector->incrementYields(); 6980 6981 // See the comment in coordinator_yield() 6982 for (unsigned i = 0; i < CMSYieldSleepCount && 6983 ConcurrentMarkSweepThread::should_yield() && 6984 !CMSCollector::foregroundGCIsActive(); ++i) { 6985 os::sleep(Thread::current(), 1, false); 6986 } 6987 6988 ConcurrentMarkSweepThread::synchronize(true); 6989 bml->lock_without_safepoint_check(); 6990 6991 _collector->startTimer(); 6992 } 6993 6994 bool CMSPrecleanRefsYieldClosure::should_return() { 6995 if (ConcurrentMarkSweepThread::should_yield()) { 6996 do_yield_work(); 6997 } 6998 return _collector->foregroundGCIsActive(); 6999 } 7000 7001 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7002 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7003 "mr should be aligned to start at a card boundary"); 7004 // We'd like to assert: 7005 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7006 // "mr should be a range of cards"); 7007 // However, that would be too strong in one case -- the last 7008 // partition ends at _unallocated_block which, in general, can be 7009 // an arbitrary boundary, not necessarily card aligned. 7010 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7011 _space->object_iterate_mem(mr, &_scan_cl); 7012 } 7013 7014 SweepClosure::SweepClosure(CMSCollector* collector, 7015 ConcurrentMarkSweepGeneration* g, 7016 CMSBitMap* bitMap, bool should_yield) : 7017 _collector(collector), 7018 _g(g), 7019 _sp(g->cmsSpace()), 7020 _limit(_sp->sweep_limit()), 7021 _freelistLock(_sp->freelistLock()), 7022 _bitMap(bitMap), 7023 _inFreeRange(false), // No free range at beginning of sweep 7024 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7025 _lastFreeRangeCoalesced(false), 7026 _yield(should_yield), 7027 _freeFinger(g->used_region().start()) 7028 { 7029 NOT_PRODUCT( 7030 _numObjectsFreed = 0; 7031 _numWordsFreed = 0; 7032 _numObjectsLive = 0; 7033 _numWordsLive = 0; 7034 _numObjectsAlreadyFree = 0; 7035 _numWordsAlreadyFree = 0; 7036 _last_fc = NULL; 7037 7038 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7039 _sp->dictionary()->initialize_dict_returned_bytes(); 7040 ) 7041 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7042 "sweep _limit out of bounds"); 7043 log_develop_trace(gc, sweep)("===================="); 7044 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7045 } 7046 7047 void SweepClosure::print_on(outputStream* st) const { 7048 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7049 p2i(_sp->bottom()), p2i(_sp->end())); 7050 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7051 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7052 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7053 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7054 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7055 } 7056 7057 #ifndef PRODUCT 7058 // Assertion checking only: no useful work in product mode -- 7059 // however, if any of the flags below become product flags, 7060 // you may need to review this code to see if it needs to be 7061 // enabled in product mode. 7062 SweepClosure::~SweepClosure() { 7063 assert_lock_strong(_freelistLock); 7064 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7065 "sweep _limit out of bounds"); 7066 if (inFreeRange()) { 7067 Log(gc, sweep) log; 7068 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7069 ResourceMark rm; 7070 LogStream ls(log.error()); 7071 print_on(&ls); 7072 ShouldNotReachHere(); 7073 } 7074 7075 if (log_is_enabled(Debug, gc, sweep)) { 7076 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7077 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7078 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7079 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7080 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7081 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7082 } 7083 7084 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7085 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7086 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7087 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7088 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7089 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7090 } 7091 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7092 log_develop_trace(gc, sweep)("================"); 7093 } 7094 #endif // PRODUCT 7095 7096 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7097 bool freeRangeInFreeLists) { 7098 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7099 p2i(freeFinger), freeRangeInFreeLists); 7100 assert(!inFreeRange(), "Trampling existing free range"); 7101 set_inFreeRange(true); 7102 set_lastFreeRangeCoalesced(false); 7103 7104 set_freeFinger(freeFinger); 7105 set_freeRangeInFreeLists(freeRangeInFreeLists); 7106 if (CMSTestInFreeList) { 7107 if (freeRangeInFreeLists) { 7108 FreeChunk* fc = (FreeChunk*) freeFinger; 7109 assert(fc->is_free(), "A chunk on the free list should be free."); 7110 assert(fc->size() > 0, "Free range should have a size"); 7111 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7112 } 7113 } 7114 } 7115 7116 // Note that the sweeper runs concurrently with mutators. Thus, 7117 // it is possible for direct allocation in this generation to happen 7118 // in the middle of the sweep. Note that the sweeper also coalesces 7119 // contiguous free blocks. Thus, unless the sweeper and the allocator 7120 // synchronize appropriately freshly allocated blocks may get swept up. 7121 // This is accomplished by the sweeper locking the free lists while 7122 // it is sweeping. Thus blocks that are determined to be free are 7123 // indeed free. There is however one additional complication: 7124 // blocks that have been allocated since the final checkpoint and 7125 // mark, will not have been marked and so would be treated as 7126 // unreachable and swept up. To prevent this, the allocator marks 7127 // the bit map when allocating during the sweep phase. This leads, 7128 // however, to a further complication -- objects may have been allocated 7129 // but not yet initialized -- in the sense that the header isn't yet 7130 // installed. The sweeper can not then determine the size of the block 7131 // in order to skip over it. To deal with this case, we use a technique 7132 // (due to Printezis) to encode such uninitialized block sizes in the 7133 // bit map. Since the bit map uses a bit per every HeapWord, but the 7134 // CMS generation has a minimum object size of 3 HeapWords, it follows 7135 // that "normal marks" won't be adjacent in the bit map (there will 7136 // always be at least two 0 bits between successive 1 bits). We make use 7137 // of these "unused" bits to represent uninitialized blocks -- the bit 7138 // corresponding to the start of the uninitialized object and the next 7139 // bit are both set. Finally, a 1 bit marks the end of the object that 7140 // started with the two consecutive 1 bits to indicate its potentially 7141 // uninitialized state. 7142 7143 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7144 FreeChunk* fc = (FreeChunk*)addr; 7145 size_t res; 7146 7147 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7148 // than "addr == _limit" because although _limit was a block boundary when 7149 // we started the sweep, it may no longer be one because heap expansion 7150 // may have caused us to coalesce the block ending at the address _limit 7151 // with a newly expanded chunk (this happens when _limit was set to the 7152 // previous _end of the space), so we may have stepped past _limit: 7153 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7154 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7155 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7156 "sweep _limit out of bounds"); 7157 assert(addr < _sp->end(), "addr out of bounds"); 7158 // Flush any free range we might be holding as a single 7159 // coalesced chunk to the appropriate free list. 7160 if (inFreeRange()) { 7161 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7162 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7163 flush_cur_free_chunk(freeFinger(), 7164 pointer_delta(addr, freeFinger())); 7165 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7166 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7167 lastFreeRangeCoalesced() ? 1 : 0); 7168 } 7169 7170 // help the iterator loop finish 7171 return pointer_delta(_sp->end(), addr); 7172 } 7173 7174 assert(addr < _limit, "sweep invariant"); 7175 // check if we should yield 7176 do_yield_check(addr); 7177 if (fc->is_free()) { 7178 // Chunk that is already free 7179 res = fc->size(); 7180 do_already_free_chunk(fc); 7181 debug_only(_sp->verifyFreeLists()); 7182 // If we flush the chunk at hand in lookahead_and_flush() 7183 // and it's coalesced with a preceding chunk, then the 7184 // process of "mangling" the payload of the coalesced block 7185 // will cause erasure of the size information from the 7186 // (erstwhile) header of all the coalesced blocks but the 7187 // first, so the first disjunct in the assert will not hold 7188 // in that specific case (in which case the second disjunct 7189 // will hold). 7190 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7191 "Otherwise the size info doesn't change at this step"); 7192 NOT_PRODUCT( 7193 _numObjectsAlreadyFree++; 7194 _numWordsAlreadyFree += res; 7195 ) 7196 NOT_PRODUCT(_last_fc = fc;) 7197 } else if (!_bitMap->isMarked(addr)) { 7198 // Chunk is fresh garbage 7199 res = do_garbage_chunk(fc); 7200 debug_only(_sp->verifyFreeLists()); 7201 NOT_PRODUCT( 7202 _numObjectsFreed++; 7203 _numWordsFreed += res; 7204 ) 7205 } else { 7206 // Chunk that is alive. 7207 res = do_live_chunk(fc); 7208 debug_only(_sp->verifyFreeLists()); 7209 NOT_PRODUCT( 7210 _numObjectsLive++; 7211 _numWordsLive += res; 7212 ) 7213 } 7214 return res; 7215 } 7216 7217 // For the smart allocation, record following 7218 // split deaths - a free chunk is removed from its free list because 7219 // it is being split into two or more chunks. 7220 // split birth - a free chunk is being added to its free list because 7221 // a larger free chunk has been split and resulted in this free chunk. 7222 // coal death - a free chunk is being removed from its free list because 7223 // it is being coalesced into a large free chunk. 7224 // coal birth - a free chunk is being added to its free list because 7225 // it was created when two or more free chunks where coalesced into 7226 // this free chunk. 7227 // 7228 // These statistics are used to determine the desired number of free 7229 // chunks of a given size. The desired number is chosen to be relative 7230 // to the end of a CMS sweep. The desired number at the end of a sweep 7231 // is the 7232 // count-at-end-of-previous-sweep (an amount that was enough) 7233 // - count-at-beginning-of-current-sweep (the excess) 7234 // + split-births (gains in this size during interval) 7235 // - split-deaths (demands on this size during interval) 7236 // where the interval is from the end of one sweep to the end of the 7237 // next. 7238 // 7239 // When sweeping the sweeper maintains an accumulated chunk which is 7240 // the chunk that is made up of chunks that have been coalesced. That 7241 // will be termed the left-hand chunk. A new chunk of garbage that 7242 // is being considered for coalescing will be referred to as the 7243 // right-hand chunk. 7244 // 7245 // When making a decision on whether to coalesce a right-hand chunk with 7246 // the current left-hand chunk, the current count vs. the desired count 7247 // of the left-hand chunk is considered. Also if the right-hand chunk 7248 // is near the large chunk at the end of the heap (see 7249 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7250 // left-hand chunk is coalesced. 7251 // 7252 // When making a decision about whether to split a chunk, the desired count 7253 // vs. the current count of the candidate to be split is also considered. 7254 // If the candidate is underpopulated (currently fewer chunks than desired) 7255 // a chunk of an overpopulated (currently more chunks than desired) size may 7256 // be chosen. The "hint" associated with a free list, if non-null, points 7257 // to a free list which may be overpopulated. 7258 // 7259 7260 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7261 const size_t size = fc->size(); 7262 // Chunks that cannot be coalesced are not in the 7263 // free lists. 7264 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7265 assert(_sp->verify_chunk_in_free_list(fc), 7266 "free chunk should be in free lists"); 7267 } 7268 // a chunk that is already free, should not have been 7269 // marked in the bit map 7270 HeapWord* const addr = (HeapWord*) fc; 7271 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7272 // Verify that the bit map has no bits marked between 7273 // addr and purported end of this block. 7274 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7275 7276 // Some chunks cannot be coalesced under any circumstances. 7277 // See the definition of cantCoalesce(). 7278 if (!fc->cantCoalesce()) { 7279 // This chunk can potentially be coalesced. 7280 // All the work is done in 7281 do_post_free_or_garbage_chunk(fc, size); 7282 // Note that if the chunk is not coalescable (the else arm 7283 // below), we unconditionally flush, without needing to do 7284 // a "lookahead," as we do below. 7285 if (inFreeRange()) lookahead_and_flush(fc, size); 7286 } else { 7287 // Code path common to both original and adaptive free lists. 7288 7289 // cant coalesce with previous block; this should be treated 7290 // as the end of a free run if any 7291 if (inFreeRange()) { 7292 // we kicked some butt; time to pick up the garbage 7293 assert(freeFinger() < addr, "freeFinger points too high"); 7294 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7295 } 7296 // else, nothing to do, just continue 7297 } 7298 } 7299 7300 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7301 // This is a chunk of garbage. It is not in any free list. 7302 // Add it to a free list or let it possibly be coalesced into 7303 // a larger chunk. 7304 HeapWord* const addr = (HeapWord*) fc; 7305 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7306 7307 // Verify that the bit map has no bits marked between 7308 // addr and purported end of just dead object. 7309 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7310 do_post_free_or_garbage_chunk(fc, size); 7311 7312 assert(_limit >= addr + size, 7313 "A freshly garbage chunk can't possibly straddle over _limit"); 7314 if (inFreeRange()) lookahead_and_flush(fc, size); 7315 return size; 7316 } 7317 7318 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7319 HeapWord* addr = (HeapWord*) fc; 7320 // The sweeper has just found a live object. Return any accumulated 7321 // left hand chunk to the free lists. 7322 if (inFreeRange()) { 7323 assert(freeFinger() < addr, "freeFinger points too high"); 7324 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7325 } 7326 7327 // This object is live: we'd normally expect this to be 7328 // an oop, and like to assert the following: 7329 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7330 // However, as we commented above, this may be an object whose 7331 // header hasn't yet been initialized. 7332 size_t size; 7333 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7334 if (_bitMap->isMarked(addr + 1)) { 7335 // Determine the size from the bit map, rather than trying to 7336 // compute it from the object header. 7337 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7338 size = pointer_delta(nextOneAddr + 1, addr); 7339 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7340 "alignment problem"); 7341 7342 #ifdef ASSERT 7343 if (oop(addr)->klass_or_null_acquire() != NULL) { 7344 // Ignore mark word because we are running concurrent with mutators 7345 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7346 assert(size == 7347 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7348 "P-mark and computed size do not agree"); 7349 } 7350 #endif 7351 7352 } else { 7353 // This should be an initialized object that's alive. 7354 assert(oop(addr)->klass_or_null_acquire() != NULL, 7355 "Should be an initialized object"); 7356 // Ignore mark word because we are running concurrent with mutators 7357 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7358 // Verify that the bit map has no bits marked between 7359 // addr and purported end of this block. 7360 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7361 assert(size >= 3, "Necessary for Printezis marks to work"); 7362 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7363 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7364 } 7365 return size; 7366 } 7367 7368 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7369 size_t chunkSize) { 7370 // do_post_free_or_garbage_chunk() should only be called in the case 7371 // of the adaptive free list allocator. 7372 const bool fcInFreeLists = fc->is_free(); 7373 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7374 if (CMSTestInFreeList && fcInFreeLists) { 7375 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7376 } 7377 7378 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7379 7380 HeapWord* const fc_addr = (HeapWord*) fc; 7381 7382 bool coalesce = false; 7383 const size_t left = pointer_delta(fc_addr, freeFinger()); 7384 const size_t right = chunkSize; 7385 switch (FLSCoalescePolicy) { 7386 // numeric value forms a coalition aggressiveness metric 7387 case 0: { // never coalesce 7388 coalesce = false; 7389 break; 7390 } 7391 case 1: { // coalesce if left & right chunks on overpopulated lists 7392 coalesce = _sp->coalOverPopulated(left) && 7393 _sp->coalOverPopulated(right); 7394 break; 7395 } 7396 case 2: { // coalesce if left chunk on overpopulated list (default) 7397 coalesce = _sp->coalOverPopulated(left); 7398 break; 7399 } 7400 case 3: { // coalesce if left OR right chunk on overpopulated list 7401 coalesce = _sp->coalOverPopulated(left) || 7402 _sp->coalOverPopulated(right); 7403 break; 7404 } 7405 case 4: { // always coalesce 7406 coalesce = true; 7407 break; 7408 } 7409 default: 7410 ShouldNotReachHere(); 7411 } 7412 7413 // Should the current free range be coalesced? 7414 // If the chunk is in a free range and either we decided to coalesce above 7415 // or the chunk is near the large block at the end of the heap 7416 // (isNearLargestChunk() returns true), then coalesce this chunk. 7417 const bool doCoalesce = inFreeRange() 7418 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7419 if (doCoalesce) { 7420 // Coalesce the current free range on the left with the new 7421 // chunk on the right. If either is on a free list, 7422 // it must be removed from the list and stashed in the closure. 7423 if (freeRangeInFreeLists()) { 7424 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7425 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7426 "Size of free range is inconsistent with chunk size."); 7427 if (CMSTestInFreeList) { 7428 assert(_sp->verify_chunk_in_free_list(ffc), 7429 "Chunk is not in free lists"); 7430 } 7431 _sp->coalDeath(ffc->size()); 7432 _sp->removeFreeChunkFromFreeLists(ffc); 7433 set_freeRangeInFreeLists(false); 7434 } 7435 if (fcInFreeLists) { 7436 _sp->coalDeath(chunkSize); 7437 assert(fc->size() == chunkSize, 7438 "The chunk has the wrong size or is not in the free lists"); 7439 _sp->removeFreeChunkFromFreeLists(fc); 7440 } 7441 set_lastFreeRangeCoalesced(true); 7442 print_free_block_coalesced(fc); 7443 } else { // not in a free range and/or should not coalesce 7444 // Return the current free range and start a new one. 7445 if (inFreeRange()) { 7446 // In a free range but cannot coalesce with the right hand chunk. 7447 // Put the current free range into the free lists. 7448 flush_cur_free_chunk(freeFinger(), 7449 pointer_delta(fc_addr, freeFinger())); 7450 } 7451 // Set up for new free range. Pass along whether the right hand 7452 // chunk is in the free lists. 7453 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7454 } 7455 } 7456 7457 // Lookahead flush: 7458 // If we are tracking a free range, and this is the last chunk that 7459 // we'll look at because its end crosses past _limit, we'll preemptively 7460 // flush it along with any free range we may be holding on to. Note that 7461 // this can be the case only for an already free or freshly garbage 7462 // chunk. If this block is an object, it can never straddle 7463 // over _limit. The "straddling" occurs when _limit is set at 7464 // the previous end of the space when this cycle started, and 7465 // a subsequent heap expansion caused the previously co-terminal 7466 // free block to be coalesced with the newly expanded portion, 7467 // thus rendering _limit a non-block-boundary making it dangerous 7468 // for the sweeper to step over and examine. 7469 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7470 assert(inFreeRange(), "Should only be called if currently in a free range."); 7471 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7472 assert(_sp->used_region().contains(eob - 1), 7473 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7474 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7475 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7476 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7477 if (eob >= _limit) { 7478 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7479 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7480 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7481 "[" PTR_FORMAT "," PTR_FORMAT ")", 7482 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7483 // Return the storage we are tracking back into the free lists. 7484 log_develop_trace(gc, sweep)("Flushing ... "); 7485 assert(freeFinger() < eob, "Error"); 7486 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7487 } 7488 } 7489 7490 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7491 assert(inFreeRange(), "Should only be called if currently in a free range."); 7492 assert(size > 0, 7493 "A zero sized chunk cannot be added to the free lists."); 7494 if (!freeRangeInFreeLists()) { 7495 if (CMSTestInFreeList) { 7496 FreeChunk* fc = (FreeChunk*) chunk; 7497 fc->set_size(size); 7498 assert(!_sp->verify_chunk_in_free_list(fc), 7499 "chunk should not be in free lists yet"); 7500 } 7501 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7502 // A new free range is going to be starting. The current 7503 // free range has not been added to the free lists yet or 7504 // was removed so add it back. 7505 // If the current free range was coalesced, then the death 7506 // of the free range was recorded. Record a birth now. 7507 if (lastFreeRangeCoalesced()) { 7508 _sp->coalBirth(size); 7509 } 7510 _sp->addChunkAndRepairOffsetTable(chunk, size, 7511 lastFreeRangeCoalesced()); 7512 } else { 7513 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7514 } 7515 set_inFreeRange(false); 7516 set_freeRangeInFreeLists(false); 7517 } 7518 7519 // We take a break if we've been at this for a while, 7520 // so as to avoid monopolizing the locks involved. 7521 void SweepClosure::do_yield_work(HeapWord* addr) { 7522 // Return current free chunk being used for coalescing (if any) 7523 // to the appropriate freelist. After yielding, the next 7524 // free block encountered will start a coalescing range of 7525 // free blocks. If the next free block is adjacent to the 7526 // chunk just flushed, they will need to wait for the next 7527 // sweep to be coalesced. 7528 if (inFreeRange()) { 7529 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7530 } 7531 7532 // First give up the locks, then yield, then re-lock. 7533 // We should probably use a constructor/destructor idiom to 7534 // do this unlock/lock or modify the MutexUnlocker class to 7535 // serve our purpose. XXX 7536 assert_lock_strong(_bitMap->lock()); 7537 assert_lock_strong(_freelistLock); 7538 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7539 "CMS thread should hold CMS token"); 7540 _bitMap->lock()->unlock(); 7541 _freelistLock->unlock(); 7542 ConcurrentMarkSweepThread::desynchronize(true); 7543 _collector->stopTimer(); 7544 _collector->incrementYields(); 7545 7546 // See the comment in coordinator_yield() 7547 for (unsigned i = 0; i < CMSYieldSleepCount && 7548 ConcurrentMarkSweepThread::should_yield() && 7549 !CMSCollector::foregroundGCIsActive(); ++i) { 7550 os::sleep(Thread::current(), 1, false); 7551 } 7552 7553 ConcurrentMarkSweepThread::synchronize(true); 7554 _freelistLock->lock_without_safepoint_check(); 7555 _bitMap->lock()->lock_without_safepoint_check(); 7556 _collector->startTimer(); 7557 } 7558 7559 #ifndef PRODUCT 7560 // This is actually very useful in a product build if it can 7561 // be called from the debugger. Compile it into the product 7562 // as needed. 7563 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7564 return debug_cms_space->verify_chunk_in_free_list(fc); 7565 } 7566 #endif 7567 7568 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7569 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7570 p2i(fc), fc->size()); 7571 } 7572 7573 // CMSIsAliveClosure 7574 bool CMSIsAliveClosure::do_object_b(oop obj) { 7575 HeapWord* addr = (HeapWord*)obj; 7576 return addr != NULL && 7577 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7578 } 7579 7580 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7581 MemRegion span, 7582 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7583 bool cpc): 7584 _collector(collector), 7585 _span(span), 7586 _mark_stack(mark_stack), 7587 _bit_map(bit_map), 7588 _concurrent_precleaning(cpc) { 7589 assert(!_span.is_empty(), "Empty span could spell trouble"); 7590 } 7591 7592 7593 // CMSKeepAliveClosure: the serial version 7594 void CMSKeepAliveClosure::do_oop(oop obj) { 7595 HeapWord* addr = (HeapWord*)obj; 7596 if (_span.contains(addr) && 7597 !_bit_map->isMarked(addr)) { 7598 _bit_map->mark(addr); 7599 bool simulate_overflow = false; 7600 NOT_PRODUCT( 7601 if (CMSMarkStackOverflowALot && 7602 _collector->simulate_overflow()) { 7603 // simulate a stack overflow 7604 simulate_overflow = true; 7605 } 7606 ) 7607 if (simulate_overflow || !_mark_stack->push(obj)) { 7608 if (_concurrent_precleaning) { 7609 // We dirty the overflown object and let the remark 7610 // phase deal with it. 7611 assert(_collector->overflow_list_is_empty(), "Error"); 7612 // In the case of object arrays, we need to dirty all of 7613 // the cards that the object spans. No locking or atomics 7614 // are needed since no one else can be mutating the mod union 7615 // table. 7616 if (obj->is_objArray()) { 7617 size_t sz = obj->size(); 7618 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7619 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7620 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7621 _collector->_modUnionTable.mark_range(redirty_range); 7622 } else { 7623 _collector->_modUnionTable.mark(addr); 7624 } 7625 _collector->_ser_kac_preclean_ovflw++; 7626 } else { 7627 _collector->push_on_overflow_list(obj); 7628 _collector->_ser_kac_ovflw++; 7629 } 7630 } 7631 } 7632 } 7633 7634 // CMSParKeepAliveClosure: a parallel version of the above. 7635 // The work queues are private to each closure (thread), 7636 // but (may be) available for stealing by other threads. 7637 void CMSParKeepAliveClosure::do_oop(oop obj) { 7638 HeapWord* addr = (HeapWord*)obj; 7639 if (_span.contains(addr) && 7640 !_bit_map->isMarked(addr)) { 7641 // In general, during recursive tracing, several threads 7642 // may be concurrently getting here; the first one to 7643 // "tag" it, claims it. 7644 if (_bit_map->par_mark(addr)) { 7645 bool res = _work_queue->push(obj); 7646 assert(res, "Low water mark should be much less than capacity"); 7647 // Do a recursive trim in the hope that this will keep 7648 // stack usage lower, but leave some oops for potential stealers 7649 trim_queue(_low_water_mark); 7650 } // Else, another thread got there first 7651 } 7652 } 7653 7654 void CMSParKeepAliveClosure::trim_queue(uint max) { 7655 while (_work_queue->size() > max) { 7656 oop new_oop; 7657 if (_work_queue->pop_local(new_oop)) { 7658 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7659 assert(_bit_map->isMarked((HeapWord*)new_oop), 7660 "no white objects on this stack!"); 7661 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7662 // iterate over the oops in this oop, marking and pushing 7663 // the ones in CMS heap (i.e. in _span). 7664 new_oop->oop_iterate(&_mark_and_push); 7665 } 7666 } 7667 } 7668 7669 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7670 CMSCollector* collector, 7671 MemRegion span, CMSBitMap* bit_map, 7672 OopTaskQueue* work_queue): 7673 _collector(collector), 7674 _span(span), 7675 _work_queue(work_queue), 7676 _bit_map(bit_map) { } 7677 7678 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7679 HeapWord* addr = (HeapWord*)obj; 7680 if (_span.contains(addr) && 7681 !_bit_map->isMarked(addr)) { 7682 if (_bit_map->par_mark(addr)) { 7683 bool simulate_overflow = false; 7684 NOT_PRODUCT( 7685 if (CMSMarkStackOverflowALot && 7686 _collector->par_simulate_overflow()) { 7687 // simulate a stack overflow 7688 simulate_overflow = true; 7689 } 7690 ) 7691 if (simulate_overflow || !_work_queue->push(obj)) { 7692 _collector->par_push_on_overflow_list(obj); 7693 _collector->_par_kac_ovflw++; 7694 } 7695 } // Else another thread got there already 7696 } 7697 } 7698 7699 ////////////////////////////////////////////////////////////////// 7700 // CMSExpansionCause ///////////////////////////// 7701 ////////////////////////////////////////////////////////////////// 7702 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7703 switch (cause) { 7704 case _no_expansion: 7705 return "No expansion"; 7706 case _satisfy_free_ratio: 7707 return "Free ratio"; 7708 case _satisfy_promotion: 7709 return "Satisfy promotion"; 7710 case _satisfy_allocation: 7711 return "allocation"; 7712 case _allocate_par_lab: 7713 return "Par LAB"; 7714 case _allocate_par_spooling_space: 7715 return "Par Spooling Space"; 7716 case _adaptive_size_policy: 7717 return "Ergonomics"; 7718 default: 7719 return "unknown"; 7720 } 7721 } 7722 7723 void CMSDrainMarkingStackClosure::do_void() { 7724 // the max number to take from overflow list at a time 7725 const size_t num = _mark_stack->capacity()/4; 7726 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7727 "Overflow list should be NULL during concurrent phases"); 7728 while (!_mark_stack->isEmpty() || 7729 // if stack is empty, check the overflow list 7730 _collector->take_from_overflow_list(num, _mark_stack)) { 7731 oop obj = _mark_stack->pop(); 7732 HeapWord* addr = (HeapWord*)obj; 7733 assert(_span.contains(addr), "Should be within span"); 7734 assert(_bit_map->isMarked(addr), "Should be marked"); 7735 assert(oopDesc::is_oop(obj), "Should be an oop"); 7736 obj->oop_iterate(_keep_alive); 7737 } 7738 } 7739 7740 void CMSParDrainMarkingStackClosure::do_void() { 7741 // drain queue 7742 trim_queue(0); 7743 } 7744 7745 // Trim our work_queue so its length is below max at return 7746 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7747 while (_work_queue->size() > max) { 7748 oop new_oop; 7749 if (_work_queue->pop_local(new_oop)) { 7750 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7751 assert(_bit_map->isMarked((HeapWord*)new_oop), 7752 "no white objects on this stack!"); 7753 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7754 // iterate over the oops in this oop, marking and pushing 7755 // the ones in CMS heap (i.e. in _span). 7756 new_oop->oop_iterate(&_mark_and_push); 7757 } 7758 } 7759 } 7760 7761 //////////////////////////////////////////////////////////////////// 7762 // Support for Marking Stack Overflow list handling and related code 7763 //////////////////////////////////////////////////////////////////// 7764 // Much of the following code is similar in shape and spirit to the 7765 // code used in ParNewGC. We should try and share that code 7766 // as much as possible in the future. 7767 7768 #ifndef PRODUCT 7769 // Debugging support for CMSStackOverflowALot 7770 7771 // It's OK to call this multi-threaded; the worst thing 7772 // that can happen is that we'll get a bunch of closely 7773 // spaced simulated overflows, but that's OK, in fact 7774 // probably good as it would exercise the overflow code 7775 // under contention. 7776 bool CMSCollector::simulate_overflow() { 7777 if (_overflow_counter-- <= 0) { // just being defensive 7778 _overflow_counter = CMSMarkStackOverflowInterval; 7779 return true; 7780 } else { 7781 return false; 7782 } 7783 } 7784 7785 bool CMSCollector::par_simulate_overflow() { 7786 return simulate_overflow(); 7787 } 7788 #endif 7789 7790 // Single-threaded 7791 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7792 assert(stack->isEmpty(), "Expected precondition"); 7793 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7794 size_t i = num; 7795 oop cur = _overflow_list; 7796 const markWord proto = markWord::prototype(); 7797 NOT_PRODUCT(ssize_t n = 0;) 7798 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7799 next = oop(cur->mark_raw().to_pointer()); 7800 cur->set_mark_raw(proto); // until proven otherwise 7801 assert(oopDesc::is_oop(cur), "Should be an oop"); 7802 bool res = stack->push(cur); 7803 assert(res, "Bit off more than can chew?"); 7804 NOT_PRODUCT(n++;) 7805 } 7806 _overflow_list = cur; 7807 #ifndef PRODUCT 7808 assert(_num_par_pushes >= n, "Too many pops?"); 7809 _num_par_pushes -=n; 7810 #endif 7811 return !stack->isEmpty(); 7812 } 7813 7814 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7815 // (MT-safe) Get a prefix of at most "num" from the list. 7816 // The overflow list is chained through the mark word of 7817 // each object in the list. We fetch the entire list, 7818 // break off a prefix of the right size and return the 7819 // remainder. If other threads try to take objects from 7820 // the overflow list at that time, they will wait for 7821 // some time to see if data becomes available. If (and 7822 // only if) another thread places one or more object(s) 7823 // on the global list before we have returned the suffix 7824 // to the global list, we will walk down our local list 7825 // to find its end and append the global list to 7826 // our suffix before returning it. This suffix walk can 7827 // prove to be expensive (quadratic in the amount of traffic) 7828 // when there are many objects in the overflow list and 7829 // there is much producer-consumer contention on the list. 7830 // *NOTE*: The overflow list manipulation code here and 7831 // in ParNewGeneration:: are very similar in shape, 7832 // except that in the ParNew case we use the old (from/eden) 7833 // copy of the object to thread the list via its klass word. 7834 // Because of the common code, if you make any changes in 7835 // the code below, please check the ParNew version to see if 7836 // similar changes might be needed. 7837 // CR 6797058 has been filed to consolidate the common code. 7838 bool CMSCollector::par_take_from_overflow_list(size_t num, 7839 OopTaskQueue* work_q, 7840 int no_of_gc_threads) { 7841 assert(work_q->size() == 0, "First empty local work queue"); 7842 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7843 if (_overflow_list == NULL) { 7844 return false; 7845 } 7846 // Grab the entire list; we'll put back a suffix 7847 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7848 Thread* tid = Thread::current(); 7849 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7850 // set to ParallelGCThreads. 7851 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7852 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7853 // If the list is busy, we spin for a short while, 7854 // sleeping between attempts to get the list. 7855 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7856 os::sleep(tid, sleep_time_millis, false); 7857 if (_overflow_list == NULL) { 7858 // Nothing left to take 7859 return false; 7860 } else if (_overflow_list != BUSY) { 7861 // Try and grab the prefix 7862 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7863 } 7864 } 7865 // If the list was found to be empty, or we spun long 7866 // enough, we give up and return empty-handed. If we leave 7867 // the list in the BUSY state below, it must be the case that 7868 // some other thread holds the overflow list and will set it 7869 // to a non-BUSY state in the future. 7870 if (prefix == NULL || prefix == BUSY) { 7871 // Nothing to take or waited long enough 7872 if (prefix == NULL) { 7873 // Write back the NULL in case we overwrote it with BUSY above 7874 // and it is still the same value. 7875 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7876 } 7877 return false; 7878 } 7879 assert(prefix != NULL && prefix != BUSY, "Error"); 7880 size_t i = num; 7881 oop cur = prefix; 7882 // Walk down the first "num" objects, unless we reach the end. 7883 for (; i > 1 && cur->mark_raw().to_pointer() != NULL; cur = oop(cur->mark_raw().to_pointer()), i--); 7884 if (cur->mark_raw().to_pointer() == NULL) { 7885 // We have "num" or fewer elements in the list, so there 7886 // is nothing to return to the global list. 7887 // Write back the NULL in lieu of the BUSY we wrote 7888 // above, if it is still the same value. 7889 if (_overflow_list == BUSY) { 7890 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7891 } 7892 } else { 7893 // Chop off the suffix and return it to the global list. 7894 assert(cur->mark_raw().to_pointer() != (void*)BUSY, "Error"); 7895 oop suffix_head = oop(cur->mark_raw().to_pointer()); // suffix will be put back on global list 7896 cur->set_mark_raw(markWord::from_pointer(NULL)); // break off suffix 7897 // It's possible that the list is still in the empty(busy) state 7898 // we left it in a short while ago; in that case we may be 7899 // able to place back the suffix without incurring the cost 7900 // of a walk down the list. 7901 oop observed_overflow_list = _overflow_list; 7902 oop cur_overflow_list = observed_overflow_list; 7903 bool attached = false; 7904 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7905 observed_overflow_list = 7906 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7907 if (cur_overflow_list == observed_overflow_list) { 7908 attached = true; 7909 break; 7910 } else cur_overflow_list = observed_overflow_list; 7911 } 7912 if (!attached) { 7913 // Too bad, someone else sneaked in (at least) an element; we'll need 7914 // to do a splice. Find tail of suffix so we can prepend suffix to global 7915 // list. 7916 for (cur = suffix_head; cur->mark_raw().to_pointer() != NULL; cur = (oop)(cur->mark_raw().to_pointer())); 7917 oop suffix_tail = cur; 7918 assert(suffix_tail != NULL && suffix_tail->mark_raw().to_pointer() == NULL, 7919 "Tautology"); 7920 observed_overflow_list = _overflow_list; 7921 do { 7922 cur_overflow_list = observed_overflow_list; 7923 if (cur_overflow_list != BUSY) { 7924 // Do the splice ... 7925 suffix_tail->set_mark_raw(markWord::from_pointer((void*)cur_overflow_list)); 7926 } else { // cur_overflow_list == BUSY 7927 suffix_tail->set_mark_raw(markWord::from_pointer(NULL)); 7928 } 7929 // ... and try to place spliced list back on overflow_list ... 7930 observed_overflow_list = 7931 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7932 } while (cur_overflow_list != observed_overflow_list); 7933 // ... until we have succeeded in doing so. 7934 } 7935 } 7936 7937 // Push the prefix elements on work_q 7938 assert(prefix != NULL, "control point invariant"); 7939 const markWord proto = markWord::prototype(); 7940 oop next; 7941 NOT_PRODUCT(ssize_t n = 0;) 7942 for (cur = prefix; cur != NULL; cur = next) { 7943 next = oop(cur->mark_raw().to_pointer()); 7944 cur->set_mark_raw(proto); // until proven otherwise 7945 assert(oopDesc::is_oop(cur), "Should be an oop"); 7946 bool res = work_q->push(cur); 7947 assert(res, "Bit off more than we can chew?"); 7948 NOT_PRODUCT(n++;) 7949 } 7950 #ifndef PRODUCT 7951 assert(_num_par_pushes >= n, "Too many pops?"); 7952 Atomic::sub(n, &_num_par_pushes); 7953 #endif 7954 return true; 7955 } 7956 7957 // Single-threaded 7958 void CMSCollector::push_on_overflow_list(oop p) { 7959 NOT_PRODUCT(_num_par_pushes++;) 7960 assert(oopDesc::is_oop(p), "Not an oop"); 7961 preserve_mark_if_necessary(p); 7962 p->set_mark_raw(markWord::from_pointer(_overflow_list)); 7963 _overflow_list = p; 7964 } 7965 7966 // Multi-threaded; use CAS to prepend to overflow list 7967 void CMSCollector::par_push_on_overflow_list(oop p) { 7968 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7969 assert(oopDesc::is_oop(p), "Not an oop"); 7970 par_preserve_mark_if_necessary(p); 7971 oop observed_overflow_list = _overflow_list; 7972 oop cur_overflow_list; 7973 do { 7974 cur_overflow_list = observed_overflow_list; 7975 if (cur_overflow_list != BUSY) { 7976 p->set_mark_raw(markWord::from_pointer((void*)cur_overflow_list)); 7977 } else { 7978 p->set_mark_raw(markWord::from_pointer(NULL)); 7979 } 7980 observed_overflow_list = 7981 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7982 } while (cur_overflow_list != observed_overflow_list); 7983 } 7984 #undef BUSY 7985 7986 // Single threaded 7987 // General Note on GrowableArray: pushes may silently fail 7988 // because we are (temporarily) out of C-heap for expanding 7989 // the stack. The problem is quite ubiquitous and affects 7990 // a lot of code in the JVM. The prudent thing for GrowableArray 7991 // to do (for now) is to exit with an error. However, that may 7992 // be too draconian in some cases because the caller may be 7993 // able to recover without much harm. For such cases, we 7994 // should probably introduce a "soft_push" method which returns 7995 // an indication of success or failure with the assumption that 7996 // the caller may be able to recover from a failure; code in 7997 // the VM can then be changed, incrementally, to deal with such 7998 // failures where possible, thus, incrementally hardening the VM 7999 // in such low resource situations. 8000 void CMSCollector::preserve_mark_work(oop p, markWord m) { 8001 _preserved_oop_stack.push(p); 8002 _preserved_mark_stack.push(m); 8003 assert(m == p->mark_raw(), "Mark word changed"); 8004 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8005 "bijection"); 8006 } 8007 8008 // Single threaded 8009 void CMSCollector::preserve_mark_if_necessary(oop p) { 8010 markWord m = p->mark_raw(); 8011 if (p->mark_must_be_preserved(m)) { 8012 preserve_mark_work(p, m); 8013 } 8014 } 8015 8016 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8017 markWord m = p->mark_raw(); 8018 if (p->mark_must_be_preserved(m)) { 8019 MutexLocker x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8020 // Even though we read the mark word without holding 8021 // the lock, we are assured that it will not change 8022 // because we "own" this oop, so no other thread can 8023 // be trying to push it on the overflow list; see 8024 // the assertion in preserve_mark_work() that checks 8025 // that m == p->mark_raw(). 8026 preserve_mark_work(p, m); 8027 } 8028 } 8029 8030 // We should be able to do this multi-threaded, 8031 // a chunk of stack being a task (this is 8032 // correct because each oop only ever appears 8033 // once in the overflow list. However, it's 8034 // not very easy to completely overlap this with 8035 // other operations, so will generally not be done 8036 // until all work's been completed. Because we 8037 // expect the preserved oop stack (set) to be small, 8038 // it's probably fine to do this single-threaded. 8039 // We can explore cleverer concurrent/overlapped/parallel 8040 // processing of preserved marks if we feel the 8041 // need for this in the future. Stack overflow should 8042 // be so rare in practice and, when it happens, its 8043 // effect on performance so great that this will 8044 // likely just be in the noise anyway. 8045 void CMSCollector::restore_preserved_marks_if_any() { 8046 assert(SafepointSynchronize::is_at_safepoint(), 8047 "world should be stopped"); 8048 assert(Thread::current()->is_ConcurrentGC_thread() || 8049 Thread::current()->is_VM_thread(), 8050 "should be single-threaded"); 8051 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8052 "bijection"); 8053 8054 while (!_preserved_oop_stack.is_empty()) { 8055 oop p = _preserved_oop_stack.pop(); 8056 assert(oopDesc::is_oop(p), "Should be an oop"); 8057 assert(_span.contains(p), "oop should be in _span"); 8058 assert(p->mark_raw() == markWord::prototype(), 8059 "Set when taken from overflow list"); 8060 markWord m = _preserved_mark_stack.pop(); 8061 p->set_mark_raw(m); 8062 } 8063 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8064 "stacks were cleared above"); 8065 } 8066 8067 #ifndef PRODUCT 8068 bool CMSCollector::no_preserved_marks() const { 8069 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8070 } 8071 #endif 8072 8073 // Transfer some number of overflown objects to usual marking 8074 // stack. Return true if some objects were transferred. 8075 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8076 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8077 (size_t)ParGCDesiredObjsFromOverflowList); 8078 8079 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8080 assert(_collector->overflow_list_is_empty() || res, 8081 "If list is not empty, we should have taken something"); 8082 assert(!res || !_mark_stack->isEmpty(), 8083 "If we took something, it should now be on our stack"); 8084 return res; 8085 } 8086 8087 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8088 size_t res = _sp->block_size_no_stall(addr, _collector); 8089 if (_sp->block_is_obj(addr)) { 8090 if (_live_bit_map->isMarked(addr)) { 8091 // It can't have been dead in a previous cycle 8092 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8093 } else { 8094 _dead_bit_map->mark(addr); // mark the dead object 8095 } 8096 } 8097 // Could be 0, if the block size could not be computed without stalling. 8098 return res; 8099 } 8100 8101 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8102 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8103 switch (phase) { 8104 case CMSCollector::InitialMarking: 8105 initialize(manager /* GC manager */ , 8106 cause /* cause of the GC */, 8107 true /* allMemoryPoolsAffected */, 8108 true /* recordGCBeginTime */, 8109 true /* recordPreGCUsage */, 8110 false /* recordPeakUsage */, 8111 false /* recordPostGCusage */, 8112 true /* recordAccumulatedGCTime */, 8113 false /* recordGCEndTime */, 8114 false /* countCollection */ ); 8115 break; 8116 8117 case CMSCollector::FinalMarking: 8118 initialize(manager /* GC manager */ , 8119 cause /* cause of the GC */, 8120 true /* allMemoryPoolsAffected */, 8121 false /* recordGCBeginTime */, 8122 false /* recordPreGCUsage */, 8123 false /* recordPeakUsage */, 8124 false /* recordPostGCusage */, 8125 true /* recordAccumulatedGCTime */, 8126 false /* recordGCEndTime */, 8127 false /* countCollection */ ); 8128 break; 8129 8130 case CMSCollector::Sweeping: 8131 initialize(manager /* GC manager */ , 8132 cause /* cause of the GC */, 8133 true /* allMemoryPoolsAffected */, 8134 false /* recordGCBeginTime */, 8135 false /* recordPreGCUsage */, 8136 true /* recordPeakUsage */, 8137 true /* recordPostGCusage */, 8138 false /* recordAccumulatedGCTime */, 8139 true /* recordGCEndTime */, 8140 true /* countCollection */ ); 8141 break; 8142 8143 default: 8144 ShouldNotReachHere(); 8145 } 8146 }