1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsGCStats.hpp" 33 #include "gc/cms/cmsHeap.hpp" 34 #include "gc/cms/cmsOopClosures.inline.hpp" 35 #include "gc/cms/compactibleFreeListSpace.hpp" 36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 37 #include "gc/cms/concurrentMarkSweepThread.hpp" 38 #include "gc/cms/parNewGeneration.hpp" 39 #include "gc/cms/promotionInfo.inline.hpp" 40 #include "gc/cms/vmCMSOperations.hpp" 41 #include "gc/serial/genMarkSweep.hpp" 42 #include "gc/serial/tenuredGeneration.hpp" 43 #include "gc/shared/adaptiveSizePolicy.hpp" 44 #include "gc/shared/cardGeneration.inline.hpp" 45 #include "gc/shared/cardTableRS.hpp" 46 #include "gc/shared/collectedHeap.inline.hpp" 47 #include "gc/shared/collectorCounters.hpp" 48 #include "gc/shared/collectorPolicy.hpp" 49 #include "gc/shared/gcLocker.hpp" 50 #include "gc/shared/gcPolicyCounters.hpp" 51 #include "gc/shared/gcTimer.hpp" 52 #include "gc/shared/gcTrace.hpp" 53 #include "gc/shared/gcTraceTime.inline.hpp" 54 #include "gc/shared/genCollectedHeap.hpp" 55 #include "gc/shared/genOopClosures.inline.hpp" 56 #include "gc/shared/isGCActiveMark.hpp" 57 #include "gc/shared/referencePolicy.hpp" 58 #include "gc/shared/space.inline.hpp" 59 #include "gc/shared/strongRootsScope.hpp" 60 #include "gc/shared/taskqueue.inline.hpp" 61 #include "gc/shared/weakProcessor.hpp" 62 #include "logging/log.hpp" 63 #include "logging/logStream.hpp" 64 #include "memory/allocation.hpp" 65 #include "memory/binaryTreeDictionary.inline.hpp" 66 #include "memory/iterator.inline.hpp" 67 #include "memory/padded.hpp" 68 #include "memory/resourceArea.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.inline.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 85 // statics 86 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 87 bool CMSCollector::_full_gc_requested = false; 88 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 89 90 ////////////////////////////////////////////////////////////////// 91 // In support of CMS/VM thread synchronization 92 ////////////////////////////////////////////////////////////////// 93 // We split use of the CGC_lock into 2 "levels". 94 // The low-level locking is of the usual CGC_lock monitor. We introduce 95 // a higher level "token" (hereafter "CMS token") built on top of the 96 // low level monitor (hereafter "CGC lock"). 97 // The token-passing protocol gives priority to the VM thread. The 98 // CMS-lock doesn't provide any fairness guarantees, but clients 99 // should ensure that it is only held for very short, bounded 100 // durations. 101 // 102 // When either of the CMS thread or the VM thread is involved in 103 // collection operations during which it does not want the other 104 // thread to interfere, it obtains the CMS token. 105 // 106 // If either thread tries to get the token while the other has 107 // it, that thread waits. However, if the VM thread and CMS thread 108 // both want the token, then the VM thread gets priority while the 109 // CMS thread waits. This ensures, for instance, that the "concurrent" 110 // phases of the CMS thread's work do not block out the VM thread 111 // for long periods of time as the CMS thread continues to hog 112 // the token. (See bug 4616232). 113 // 114 // The baton-passing functions are, however, controlled by the 115 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 116 // and here the low-level CMS lock, not the high level token, 117 // ensures mutual exclusion. 118 // 119 // Two important conditions that we have to satisfy: 120 // 1. if a thread does a low-level wait on the CMS lock, then it 121 // relinquishes the CMS token if it were holding that token 122 // when it acquired the low-level CMS lock. 123 // 2. any low-level notifications on the low-level lock 124 // should only be sent when a thread has relinquished the token. 125 // 126 // In the absence of either property, we'd have potential deadlock. 127 // 128 // We protect each of the CMS (concurrent and sequential) phases 129 // with the CMS _token_, not the CMS _lock_. 130 // 131 // The only code protected by CMS lock is the token acquisition code 132 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 133 // baton-passing code. 134 // 135 // Unfortunately, i couldn't come up with a good abstraction to factor and 136 // hide the naked CGC_lock manipulation in the baton-passing code 137 // further below. That's something we should try to do. Also, the proof 138 // of correctness of this 2-level locking scheme is far from obvious, 139 // and potentially quite slippery. We have an uneasy suspicion, for instance, 140 // that there may be a theoretical possibility of delay/starvation in the 141 // low-level lock/wait/notify scheme used for the baton-passing because of 142 // potential interference with the priority scheme embodied in the 143 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 144 // invocation further below and marked with "XXX 20011219YSR". 145 // Indeed, as we note elsewhere, this may become yet more slippery 146 // in the presence of multiple CMS and/or multiple VM threads. XXX 147 148 class CMSTokenSync: public StackObj { 149 private: 150 bool _is_cms_thread; 151 public: 152 CMSTokenSync(bool is_cms_thread): 153 _is_cms_thread(is_cms_thread) { 154 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 155 "Incorrect argument to constructor"); 156 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 157 } 158 159 ~CMSTokenSync() { 160 assert(_is_cms_thread ? 161 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 162 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 163 "Incorrect state"); 164 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 165 } 166 }; 167 168 // Convenience class that does a CMSTokenSync, and then acquires 169 // upto three locks. 170 class CMSTokenSyncWithLocks: public CMSTokenSync { 171 private: 172 // Note: locks are acquired in textual declaration order 173 // and released in the opposite order 174 MutexLockerEx _locker1, _locker2, _locker3; 175 public: 176 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 177 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 178 CMSTokenSync(is_cms_thread), 179 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 180 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 181 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 182 { } 183 }; 184 185 186 ////////////////////////////////////////////////////////////////// 187 // Concurrent Mark-Sweep Generation ///////////////////////////// 188 ////////////////////////////////////////////////////////////////// 189 190 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 191 192 // This struct contains per-thread things necessary to support parallel 193 // young-gen collection. 194 class CMSParGCThreadState: public CHeapObj<mtGC> { 195 public: 196 CompactibleFreeListSpaceLAB lab; 197 PromotionInfo promo; 198 199 // Constructor. 200 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 201 promo.setSpace(cfls); 202 } 203 }; 204 205 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 206 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 207 CardGeneration(rs, initial_byte_size, ct), 208 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 209 _did_compact(false) 210 { 211 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 212 HeapWord* end = (HeapWord*) _virtual_space.high(); 213 214 _direct_allocated_words = 0; 215 NOT_PRODUCT( 216 _numObjectsPromoted = 0; 217 _numWordsPromoted = 0; 218 _numObjectsAllocated = 0; 219 _numWordsAllocated = 0; 220 ) 221 222 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 223 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 224 _cmsSpace->_old_gen = this; 225 226 _gc_stats = new CMSGCStats(); 227 228 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 229 // offsets match. The ability to tell free chunks from objects 230 // depends on this property. 231 debug_only( 232 FreeChunk* junk = NULL; 233 assert(UseCompressedClassPointers || 234 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 235 "Offset of FreeChunk::_prev within FreeChunk must match" 236 " that of OopDesc::_klass within OopDesc"); 237 ) 238 239 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 240 for (uint i = 0; i < ParallelGCThreads; i++) { 241 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 242 } 243 244 _incremental_collection_failed = false; 245 // The "dilatation_factor" is the expansion that can occur on 246 // account of the fact that the minimum object size in the CMS 247 // generation may be larger than that in, say, a contiguous young 248 // generation. 249 // Ideally, in the calculation below, we'd compute the dilatation 250 // factor as: MinChunkSize/(promoting_gen's min object size) 251 // Since we do not have such a general query interface for the 252 // promoting generation, we'll instead just use the minimum 253 // object size (which today is a header's worth of space); 254 // note that all arithmetic is in units of HeapWords. 255 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 256 assert(_dilatation_factor >= 1.0, "from previous assert"); 257 } 258 259 260 // The field "_initiating_occupancy" represents the occupancy percentage 261 // at which we trigger a new collection cycle. Unless explicitly specified 262 // via CMSInitiatingOccupancyFraction (argument "io" below), it 263 // is calculated by: 264 // 265 // Let "f" be MinHeapFreeRatio in 266 // 267 // _initiating_occupancy = 100-f + 268 // f * (CMSTriggerRatio/100) 269 // where CMSTriggerRatio is the argument "tr" below. 270 // 271 // That is, if we assume the heap is at its desired maximum occupancy at the 272 // end of a collection, we let CMSTriggerRatio of the (purported) free 273 // space be allocated before initiating a new collection cycle. 274 // 275 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 276 assert(io <= 100 && tr <= 100, "Check the arguments"); 277 if (io >= 0) { 278 _initiating_occupancy = (double)io / 100.0; 279 } else { 280 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 281 (double)(tr * MinHeapFreeRatio) / 100.0) 282 / 100.0; 283 } 284 } 285 286 void ConcurrentMarkSweepGeneration::ref_processor_init() { 287 assert(collector() != NULL, "no collector"); 288 collector()->ref_processor_init(); 289 } 290 291 void CMSCollector::ref_processor_init() { 292 if (_ref_processor == NULL) { 293 // Allocate and initialize a reference processor 294 _ref_processor = 295 new ReferenceProcessor(&_span_based_discoverer, 296 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 297 ParallelGCThreads, // mt processing degree 298 _cmsGen->refs_discovery_is_mt(), // mt discovery 299 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 300 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 301 &_is_alive_closure); // closure for liveness info 302 // Initialize the _ref_processor field of CMSGen 303 _cmsGen->set_ref_processor(_ref_processor); 304 305 } 306 } 307 308 AdaptiveSizePolicy* CMSCollector::size_policy() { 309 return CMSHeap::heap()->size_policy(); 310 } 311 312 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 313 314 const char* gen_name = "old"; 315 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 316 // Generation Counters - generation 1, 1 subspace 317 _gen_counters = new GenerationCounters(gen_name, 1, 1, 318 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 319 320 _space_counters = new GSpaceCounters(gen_name, 0, 321 _virtual_space.reserved_size(), 322 this, _gen_counters); 323 } 324 325 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 326 _cms_gen(cms_gen) 327 { 328 assert(alpha <= 100, "bad value"); 329 _saved_alpha = alpha; 330 331 // Initialize the alphas to the bootstrap value of 100. 332 _gc0_alpha = _cms_alpha = 100; 333 334 _cms_begin_time.update(); 335 _cms_end_time.update(); 336 337 _gc0_duration = 0.0; 338 _gc0_period = 0.0; 339 _gc0_promoted = 0; 340 341 _cms_duration = 0.0; 342 _cms_period = 0.0; 343 _cms_allocated = 0; 344 345 _cms_used_at_gc0_begin = 0; 346 _cms_used_at_gc0_end = 0; 347 _allow_duty_cycle_reduction = false; 348 _valid_bits = 0; 349 } 350 351 double CMSStats::cms_free_adjustment_factor(size_t free) const { 352 // TBD: CR 6909490 353 return 1.0; 354 } 355 356 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 357 } 358 359 // If promotion failure handling is on use 360 // the padded average size of the promotion for each 361 // young generation collection. 362 double CMSStats::time_until_cms_gen_full() const { 363 size_t cms_free = _cms_gen->cmsSpace()->free(); 364 CMSHeap* heap = CMSHeap::heap(); 365 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 366 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 367 if (cms_free > expected_promotion) { 368 // Start a cms collection if there isn't enough space to promote 369 // for the next young collection. Use the padded average as 370 // a safety factor. 371 cms_free -= expected_promotion; 372 373 // Adjust by the safety factor. 374 double cms_free_dbl = (double)cms_free; 375 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 376 // Apply a further correction factor which tries to adjust 377 // for recent occurance of concurrent mode failures. 378 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 379 cms_free_dbl = cms_free_dbl * cms_adjustment; 380 381 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 382 cms_free, expected_promotion); 383 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 384 // Add 1 in case the consumption rate goes to zero. 385 return cms_free_dbl / (cms_consumption_rate() + 1.0); 386 } 387 return 0.0; 388 } 389 390 // Compare the duration of the cms collection to the 391 // time remaining before the cms generation is empty. 392 // Note that the time from the start of the cms collection 393 // to the start of the cms sweep (less than the total 394 // duration of the cms collection) can be used. This 395 // has been tried and some applications experienced 396 // promotion failures early in execution. This was 397 // possibly because the averages were not accurate 398 // enough at the beginning. 399 double CMSStats::time_until_cms_start() const { 400 // We add "gc0_period" to the "work" calculation 401 // below because this query is done (mostly) at the 402 // end of a scavenge, so we need to conservatively 403 // account for that much possible delay 404 // in the query so as to avoid concurrent mode failures 405 // due to starting the collection just a wee bit too 406 // late. 407 double work = cms_duration() + gc0_period(); 408 double deadline = time_until_cms_gen_full(); 409 // If a concurrent mode failure occurred recently, we want to be 410 // more conservative and halve our expected time_until_cms_gen_full() 411 if (work > deadline) { 412 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 413 cms_duration(), gc0_period(), time_until_cms_gen_full()); 414 return 0.0; 415 } 416 return work - deadline; 417 } 418 419 #ifndef PRODUCT 420 void CMSStats::print_on(outputStream *st) const { 421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 423 gc0_duration(), gc0_period(), gc0_promoted()); 424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 425 cms_duration(), cms_period(), cms_allocated()); 426 st->print(",cms_since_beg=%g,cms_since_end=%g", 427 cms_time_since_begin(), cms_time_since_end()); 428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 430 431 if (valid()) { 432 st->print(",promo_rate=%g,cms_alloc_rate=%g", 433 promotion_rate(), cms_allocation_rate()); 434 st->print(",cms_consumption_rate=%g,time_until_full=%g", 435 cms_consumption_rate(), time_until_cms_gen_full()); 436 } 437 st->cr(); 438 } 439 #endif // #ifndef PRODUCT 440 441 CMSCollector::CollectorState CMSCollector::_collectorState = 442 CMSCollector::Idling; 443 bool CMSCollector::_foregroundGCIsActive = false; 444 bool CMSCollector::_foregroundGCShouldWait = false; 445 446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 447 CardTableRS* ct, 448 ConcurrentMarkSweepPolicy* cp): 449 _cmsGen(cmsGen), 450 // Adjust span to cover old (cms) gen 451 _span(cmsGen->reserved()), 452 _ct(ct), 453 _span_based_discoverer(_span), 454 _ref_processor(NULL), // will be set later 455 _conc_workers(NULL), // may be set later 456 _abort_preclean(false), 457 _start_sampling(false), 458 _between_prologue_and_epilogue(false), 459 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 460 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 461 -1 /* lock-free */, "No_lock" /* dummy */), 462 _modUnionClosurePar(&_modUnionTable), 463 // Construct the is_alive_closure with _span & markBitMap 464 _is_alive_closure(_span, &_markBitMap), 465 _restart_addr(NULL), 466 _overflow_list(NULL), 467 _stats(cmsGen), 468 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 469 //verify that this lock should be acquired with safepoint check. 470 Monitor::_safepoint_check_sometimes)), 471 _eden_chunk_array(NULL), // may be set in ctor body 472 _eden_chunk_capacity(0), // -- ditto -- 473 _eden_chunk_index(0), // -- ditto -- 474 _survivor_plab_array(NULL), // -- ditto -- 475 _survivor_chunk_array(NULL), // -- ditto -- 476 _survivor_chunk_capacity(0), // -- ditto -- 477 _survivor_chunk_index(0), // -- ditto -- 478 _ser_pmc_preclean_ovflw(0), 479 _ser_kac_preclean_ovflw(0), 480 _ser_pmc_remark_ovflw(0), 481 _par_pmc_remark_ovflw(0), 482 _ser_kac_ovflw(0), 483 _par_kac_ovflw(0), 484 #ifndef PRODUCT 485 _num_par_pushes(0), 486 #endif 487 _collection_count_start(0), 488 _verifying(false), 489 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 490 _completed_initialization(false), 491 _collector_policy(cp), 492 _should_unload_classes(CMSClassUnloadingEnabled), 493 _concurrent_cycles_since_last_unload(0), 494 _roots_scanning_options(GenCollectedHeap::SO_None), 495 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 496 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 497 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 498 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 499 _cms_start_registered(false) 500 { 501 // Now expand the span and allocate the collection support structures 502 // (MUT, marking bit map etc.) to cover both generations subject to 503 // collection. 504 505 // For use by dirty card to oop closures. 506 _cmsGen->cmsSpace()->set_collector(this); 507 508 // Allocate MUT and marking bit map 509 { 510 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 511 if (!_markBitMap.allocate(_span)) { 512 log_warning(gc)("Failed to allocate CMS Bit Map"); 513 return; 514 } 515 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 516 } 517 { 518 _modUnionTable.allocate(_span); 519 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 520 } 521 522 if (!_markStack.allocate(MarkStackSize)) { 523 log_warning(gc)("Failed to allocate CMS Marking Stack"); 524 return; 525 } 526 527 // Support for multi-threaded concurrent phases 528 if (CMSConcurrentMTEnabled) { 529 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 530 // just for now 531 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 532 } 533 if (ConcGCThreads > 1) { 534 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 535 ConcGCThreads, true); 536 if (_conc_workers == NULL) { 537 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 538 CMSConcurrentMTEnabled = false; 539 } else { 540 _conc_workers->initialize_workers(); 541 } 542 } else { 543 CMSConcurrentMTEnabled = false; 544 } 545 } 546 if (!CMSConcurrentMTEnabled) { 547 ConcGCThreads = 0; 548 } else { 549 // Turn off CMSCleanOnEnter optimization temporarily for 550 // the MT case where it's not fixed yet; see 6178663. 551 CMSCleanOnEnter = false; 552 } 553 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 554 "Inconsistency"); 555 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 556 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 557 558 // Parallel task queues; these are shared for the 559 // concurrent and stop-world phases of CMS, but 560 // are not shared with parallel scavenge (ParNew). 561 { 562 uint i; 563 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 564 565 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 566 || ParallelRefProcEnabled) 567 && num_queues > 0) { 568 _task_queues = new OopTaskQueueSet(num_queues); 569 if (_task_queues == NULL) { 570 log_warning(gc)("task_queues allocation failure."); 571 return; 572 } 573 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 574 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 575 for (i = 0; i < num_queues; i++) { 576 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 577 if (q == NULL) { 578 log_warning(gc)("work_queue allocation failure."); 579 return; 580 } 581 _task_queues->register_queue(i, q); 582 } 583 for (i = 0; i < num_queues; i++) { 584 _task_queues->queue(i)->initialize(); 585 _hash_seed[i] = 17; // copied from ParNew 586 } 587 } 588 } 589 590 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 591 592 // Clip CMSBootstrapOccupancy between 0 and 100. 593 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 594 595 // Now tell CMS generations the identity of their collector 596 ConcurrentMarkSweepGeneration::set_collector(this); 597 598 // Create & start a CMS thread for this CMS collector 599 _cmsThread = ConcurrentMarkSweepThread::start(this); 600 assert(cmsThread() != NULL, "CMS Thread should have been created"); 601 assert(cmsThread()->collector() == this, 602 "CMS Thread should refer to this gen"); 603 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 604 605 // Support for parallelizing young gen rescan 606 CMSHeap* heap = CMSHeap::heap(); 607 _young_gen = heap->young_gen(); 608 if (heap->supports_inline_contig_alloc()) { 609 _top_addr = heap->top_addr(); 610 _end_addr = heap->end_addr(); 611 assert(_young_gen != NULL, "no _young_gen"); 612 _eden_chunk_index = 0; 613 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 614 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 615 } 616 617 // Support for parallelizing survivor space rescan 618 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 619 const size_t max_plab_samples = 620 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 621 622 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 623 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 624 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 625 _survivor_chunk_capacity = max_plab_samples; 626 for (uint i = 0; i < ParallelGCThreads; i++) { 627 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 628 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 629 assert(cur->end() == 0, "Should be 0"); 630 assert(cur->array() == vec, "Should be vec"); 631 assert(cur->capacity() == max_plab_samples, "Error"); 632 } 633 } 634 635 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 636 _gc_counters = new CollectorCounters("CMS", 1); 637 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 638 _completed_initialization = true; 639 _inter_sweep_timer.start(); // start of time 640 } 641 642 const char* ConcurrentMarkSweepGeneration::name() const { 643 return "concurrent mark-sweep generation"; 644 } 645 void ConcurrentMarkSweepGeneration::update_counters() { 646 if (UsePerfData) { 647 _space_counters->update_all(); 648 _gen_counters->update_all(); 649 } 650 } 651 652 // this is an optimized version of update_counters(). it takes the 653 // used value as a parameter rather than computing it. 654 // 655 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 656 if (UsePerfData) { 657 _space_counters->update_used(used); 658 _space_counters->update_capacity(); 659 _gen_counters->update_all(); 660 } 661 } 662 663 void ConcurrentMarkSweepGeneration::print() const { 664 Generation::print(); 665 cmsSpace()->print(); 666 } 667 668 #ifndef PRODUCT 669 void ConcurrentMarkSweepGeneration::print_statistics() { 670 cmsSpace()->printFLCensus(0); 671 } 672 #endif 673 674 size_t 675 ConcurrentMarkSweepGeneration::contiguous_available() const { 676 // dld proposes an improvement in precision here. If the committed 677 // part of the space ends in a free block we should add that to 678 // uncommitted size in the calculation below. Will make this 679 // change later, staying with the approximation below for the 680 // time being. -- ysr. 681 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 682 } 683 684 size_t 685 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 686 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 687 } 688 689 size_t ConcurrentMarkSweepGeneration::max_available() const { 690 return free() + _virtual_space.uncommitted_size(); 691 } 692 693 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 694 size_t available = max_available(); 695 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 696 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 697 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 698 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 699 return res; 700 } 701 702 // At a promotion failure dump information on block layout in heap 703 // (cms old generation). 704 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 705 Log(gc, promotion) log; 706 if (log.is_trace()) { 707 LogStream ls(log.trace()); 708 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 709 } 710 } 711 712 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 713 // Clear the promotion information. These pointers can be adjusted 714 // along with all the other pointers into the heap but 715 // compaction is expected to be a rare event with 716 // a heap using cms so don't do it without seeing the need. 717 for (uint i = 0; i < ParallelGCThreads; i++) { 718 _par_gc_thread_states[i]->promo.reset(); 719 } 720 } 721 722 void ConcurrentMarkSweepGeneration::compute_new_size() { 723 assert_locked_or_safepoint(Heap_lock); 724 725 // If incremental collection failed, we just want to expand 726 // to the limit. 727 if (incremental_collection_failed()) { 728 clear_incremental_collection_failed(); 729 grow_to_reserved(); 730 return; 731 } 732 733 // The heap has been compacted but not reset yet. 734 // Any metric such as free() or used() will be incorrect. 735 736 CardGeneration::compute_new_size(); 737 738 // Reset again after a possible resizing 739 if (did_compact()) { 740 cmsSpace()->reset_after_compaction(); 741 } 742 } 743 744 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 745 assert_locked_or_safepoint(Heap_lock); 746 747 // If incremental collection failed, we just want to expand 748 // to the limit. 749 if (incremental_collection_failed()) { 750 clear_incremental_collection_failed(); 751 grow_to_reserved(); 752 return; 753 } 754 755 double free_percentage = ((double) free()) / capacity(); 756 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 757 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 758 759 // compute expansion delta needed for reaching desired free percentage 760 if (free_percentage < desired_free_percentage) { 761 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 762 assert(desired_capacity >= capacity(), "invalid expansion size"); 763 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 764 Log(gc) log; 765 if (log.is_trace()) { 766 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 767 log.trace("From compute_new_size: "); 768 log.trace(" Free fraction %f", free_percentage); 769 log.trace(" Desired free fraction %f", desired_free_percentage); 770 log.trace(" Maximum free fraction %f", maximum_free_percentage); 771 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 772 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 773 CMSHeap* heap = CMSHeap::heap(); 774 size_t young_size = heap->young_gen()->capacity(); 775 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 776 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 777 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 778 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 779 } 780 // safe if expansion fails 781 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 782 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 783 } else { 784 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 785 assert(desired_capacity <= capacity(), "invalid expansion size"); 786 size_t shrink_bytes = capacity() - desired_capacity; 787 // Don't shrink unless the delta is greater than the minimum shrink we want 788 if (shrink_bytes >= MinHeapDeltaBytes) { 789 shrink_free_list_by(shrink_bytes); 790 } 791 } 792 } 793 794 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 795 return cmsSpace()->freelistLock(); 796 } 797 798 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 799 CMSSynchronousYieldRequest yr; 800 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 801 return have_lock_and_allocate(size, tlab); 802 } 803 804 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 805 bool tlab /* ignored */) { 806 assert_lock_strong(freelistLock()); 807 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 808 HeapWord* res = cmsSpace()->allocate(adjustedSize); 809 // Allocate the object live (grey) if the background collector has 810 // started marking. This is necessary because the marker may 811 // have passed this address and consequently this object will 812 // not otherwise be greyed and would be incorrectly swept up. 813 // Note that if this object contains references, the writing 814 // of those references will dirty the card containing this object 815 // allowing the object to be blackened (and its references scanned) 816 // either during a preclean phase or at the final checkpoint. 817 if (res != NULL) { 818 // We may block here with an uninitialized object with 819 // its mark-bit or P-bits not yet set. Such objects need 820 // to be safely navigable by block_start(). 821 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 822 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 823 collector()->direct_allocated(res, adjustedSize); 824 _direct_allocated_words += adjustedSize; 825 // allocation counters 826 NOT_PRODUCT( 827 _numObjectsAllocated++; 828 _numWordsAllocated += (int)adjustedSize; 829 ) 830 } 831 return res; 832 } 833 834 // In the case of direct allocation by mutators in a generation that 835 // is being concurrently collected, the object must be allocated 836 // live (grey) if the background collector has started marking. 837 // This is necessary because the marker may 838 // have passed this address and consequently this object will 839 // not otherwise be greyed and would be incorrectly swept up. 840 // Note that if this object contains references, the writing 841 // of those references will dirty the card containing this object 842 // allowing the object to be blackened (and its references scanned) 843 // either during a preclean phase or at the final checkpoint. 844 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 845 assert(_markBitMap.covers(start, size), "Out of bounds"); 846 if (_collectorState >= Marking) { 847 MutexLockerEx y(_markBitMap.lock(), 848 Mutex::_no_safepoint_check_flag); 849 // [see comments preceding SweepClosure::do_blk() below for details] 850 // 851 // Can the P-bits be deleted now? JJJ 852 // 853 // 1. need to mark the object as live so it isn't collected 854 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 855 // 3. need to mark the end of the object so marking, precleaning or sweeping 856 // can skip over uninitialized or unparsable objects. An allocated 857 // object is considered uninitialized for our purposes as long as 858 // its klass word is NULL. All old gen objects are parsable 859 // as soon as they are initialized.) 860 _markBitMap.mark(start); // object is live 861 _markBitMap.mark(start + 1); // object is potentially uninitialized? 862 _markBitMap.mark(start + size - 1); 863 // mark end of object 864 } 865 // check that oop looks uninitialized 866 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 867 } 868 869 void CMSCollector::promoted(bool par, HeapWord* start, 870 bool is_obj_array, size_t obj_size) { 871 assert(_markBitMap.covers(start), "Out of bounds"); 872 // See comment in direct_allocated() about when objects should 873 // be allocated live. 874 if (_collectorState >= Marking) { 875 // we already hold the marking bit map lock, taken in 876 // the prologue 877 if (par) { 878 _markBitMap.par_mark(start); 879 } else { 880 _markBitMap.mark(start); 881 } 882 // We don't need to mark the object as uninitialized (as 883 // in direct_allocated above) because this is being done with the 884 // world stopped and the object will be initialized by the 885 // time the marking, precleaning or sweeping get to look at it. 886 // But see the code for copying objects into the CMS generation, 887 // where we need to ensure that concurrent readers of the 888 // block offset table are able to safely navigate a block that 889 // is in flux from being free to being allocated (and in 890 // transition while being copied into) and subsequently 891 // becoming a bona-fide object when the copy/promotion is complete. 892 assert(SafepointSynchronize::is_at_safepoint(), 893 "expect promotion only at safepoints"); 894 895 if (_collectorState < Sweeping) { 896 // Mark the appropriate cards in the modUnionTable, so that 897 // this object gets scanned before the sweep. If this is 898 // not done, CMS generation references in the object might 899 // not get marked. 900 // For the case of arrays, which are otherwise precisely 901 // marked, we need to dirty the entire array, not just its head. 902 if (is_obj_array) { 903 // The [par_]mark_range() method expects mr.end() below to 904 // be aligned to the granularity of a bit's representation 905 // in the heap. In the case of the MUT below, that's a 906 // card size. 907 MemRegion mr(start, 908 align_up(start + obj_size, 909 CardTable::card_size /* bytes */)); 910 if (par) { 911 _modUnionTable.par_mark_range(mr); 912 } else { 913 _modUnionTable.mark_range(mr); 914 } 915 } else { // not an obj array; we can just mark the head 916 if (par) { 917 _modUnionTable.par_mark(start); 918 } else { 919 _modUnionTable.mark(start); 920 } 921 } 922 } 923 } 924 } 925 926 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 927 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 928 // allocate, copy and if necessary update promoinfo -- 929 // delegate to underlying space. 930 assert_lock_strong(freelistLock()); 931 932 #ifndef PRODUCT 933 if (CMSHeap::heap()->promotion_should_fail()) { 934 return NULL; 935 } 936 #endif // #ifndef PRODUCT 937 938 oop res = _cmsSpace->promote(obj, obj_size); 939 if (res == NULL) { 940 // expand and retry 941 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 942 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 943 // Since this is the old generation, we don't try to promote 944 // into a more senior generation. 945 res = _cmsSpace->promote(obj, obj_size); 946 } 947 if (res != NULL) { 948 // See comment in allocate() about when objects should 949 // be allocated live. 950 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 951 collector()->promoted(false, // Not parallel 952 (HeapWord*)res, obj->is_objArray(), obj_size); 953 // promotion counters 954 NOT_PRODUCT( 955 _numObjectsPromoted++; 956 _numWordsPromoted += 957 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 958 ) 959 } 960 return res; 961 } 962 963 964 // IMPORTANT: Notes on object size recognition in CMS. 965 // --------------------------------------------------- 966 // A block of storage in the CMS generation is always in 967 // one of three states. A free block (FREE), an allocated 968 // object (OBJECT) whose size() method reports the correct size, 969 // and an intermediate state (TRANSIENT) in which its size cannot 970 // be accurately determined. 971 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 972 // ----------------------------------------------------- 973 // FREE: klass_word & 1 == 1; mark_word holds block size 974 // 975 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 976 // obj->size() computes correct size 977 // 978 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 979 // 980 // STATE IDENTIFICATION: (64 bit+COOPS) 981 // ------------------------------------ 982 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 983 // 984 // OBJECT: klass_word installed; klass_word != 0; 985 // obj->size() computes correct size 986 // 987 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 988 // 989 // 990 // STATE TRANSITION DIAGRAM 991 // 992 // mut / parnew mut / parnew 993 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 994 // ^ | 995 // |------------------------ DEAD <------------------------------------| 996 // sweep mut 997 // 998 // While a block is in TRANSIENT state its size cannot be determined 999 // so readers will either need to come back later or stall until 1000 // the size can be determined. Note that for the case of direct 1001 // allocation, P-bits, when available, may be used to determine the 1002 // size of an object that may not yet have been initialized. 1003 1004 // Things to support parallel young-gen collection. 1005 oop 1006 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1007 oop old, markOop m, 1008 size_t word_sz) { 1009 #ifndef PRODUCT 1010 if (CMSHeap::heap()->promotion_should_fail()) { 1011 return NULL; 1012 } 1013 #endif // #ifndef PRODUCT 1014 1015 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1016 PromotionInfo* promoInfo = &ps->promo; 1017 // if we are tracking promotions, then first ensure space for 1018 // promotion (including spooling space for saving header if necessary). 1019 // then allocate and copy, then track promoted info if needed. 1020 // When tracking (see PromotionInfo::track()), the mark word may 1021 // be displaced and in this case restoration of the mark word 1022 // occurs in the (oop_since_save_marks_)iterate phase. 1023 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1024 // Out of space for allocating spooling buffers; 1025 // try expanding and allocating spooling buffers. 1026 if (!expand_and_ensure_spooling_space(promoInfo)) { 1027 return NULL; 1028 } 1029 } 1030 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1031 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1032 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1033 if (obj_ptr == NULL) { 1034 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1035 if (obj_ptr == NULL) { 1036 return NULL; 1037 } 1038 } 1039 oop obj = oop(obj_ptr); 1040 OrderAccess::storestore(); 1041 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1042 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1043 // IMPORTANT: See note on object initialization for CMS above. 1044 // Otherwise, copy the object. Here we must be careful to insert the 1045 // klass pointer last, since this marks the block as an allocated object. 1046 // Except with compressed oops it's the mark word. 1047 HeapWord* old_ptr = (HeapWord*)old; 1048 // Restore the mark word copied above. 1049 obj->set_mark_raw(m); 1050 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1051 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1052 OrderAccess::storestore(); 1053 1054 if (UseCompressedClassPointers) { 1055 // Copy gap missed by (aligned) header size calculation below 1056 obj->set_klass_gap(old->klass_gap()); 1057 } 1058 if (word_sz > (size_t)oopDesc::header_size()) { 1059 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1060 obj_ptr + oopDesc::header_size(), 1061 word_sz - oopDesc::header_size()); 1062 } 1063 1064 // Now we can track the promoted object, if necessary. We take care 1065 // to delay the transition from uninitialized to full object 1066 // (i.e., insertion of klass pointer) until after, so that it 1067 // atomically becomes a promoted object. 1068 if (promoInfo->tracking()) { 1069 promoInfo->track((PromotedObject*)obj, old->klass()); 1070 } 1071 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1072 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1073 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1074 1075 // Finally, install the klass pointer (this should be volatile). 1076 OrderAccess::storestore(); 1077 obj->set_klass(old->klass()); 1078 // We should now be able to calculate the right size for this object 1079 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1080 1081 collector()->promoted(true, // parallel 1082 obj_ptr, old->is_objArray(), word_sz); 1083 1084 NOT_PRODUCT( 1085 Atomic::inc(&_numObjectsPromoted); 1086 Atomic::add(alloc_sz, &_numWordsPromoted); 1087 ) 1088 1089 return obj; 1090 } 1091 1092 void 1093 ConcurrentMarkSweepGeneration:: 1094 par_promote_alloc_done(int thread_num) { 1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1096 ps->lab.retire(thread_num); 1097 } 1098 1099 void 1100 ConcurrentMarkSweepGeneration:: 1101 par_oop_since_save_marks_iterate_done(int thread_num) { 1102 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1103 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1104 ps->promo.promoted_oops_iterate(dummy_cl); 1105 1106 // Because card-scanning has been completed, subsequent phases 1107 // (e.g., reference processing) will not need to recognize which 1108 // objects have been promoted during this GC. So, we can now disable 1109 // promotion tracking. 1110 ps->promo.stopTrackingPromotions(); 1111 } 1112 1113 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1114 size_t size, 1115 bool tlab) 1116 { 1117 // We allow a STW collection only if a full 1118 // collection was requested. 1119 return full || should_allocate(size, tlab); // FIX ME !!! 1120 // This and promotion failure handling are connected at the 1121 // hip and should be fixed by untying them. 1122 } 1123 1124 bool CMSCollector::shouldConcurrentCollect() { 1125 LogTarget(Trace, gc) log; 1126 1127 if (_full_gc_requested) { 1128 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1129 return true; 1130 } 1131 1132 FreelistLocker x(this); 1133 // ------------------------------------------------------------------ 1134 // Print out lots of information which affects the initiation of 1135 // a collection. 1136 if (log.is_enabled() && stats().valid()) { 1137 log.print("CMSCollector shouldConcurrentCollect: "); 1138 1139 LogStream out(log); 1140 stats().print_on(&out); 1141 1142 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1143 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1144 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1145 log.print("promotion_rate=%g", stats().promotion_rate()); 1146 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1147 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1148 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1149 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1150 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1151 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1152 } 1153 // ------------------------------------------------------------------ 1154 1155 // If the estimated time to complete a cms collection (cms_duration()) 1156 // is less than the estimated time remaining until the cms generation 1157 // is full, start a collection. 1158 if (!UseCMSInitiatingOccupancyOnly) { 1159 if (stats().valid()) { 1160 if (stats().time_until_cms_start() == 0.0) { 1161 return true; 1162 } 1163 } else { 1164 // We want to conservatively collect somewhat early in order 1165 // to try and "bootstrap" our CMS/promotion statistics; 1166 // this branch will not fire after the first successful CMS 1167 // collection because the stats should then be valid. 1168 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1169 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1170 _cmsGen->occupancy(), _bootstrap_occupancy); 1171 return true; 1172 } 1173 } 1174 } 1175 1176 // Otherwise, we start a collection cycle if 1177 // old gen want a collection cycle started. Each may use 1178 // an appropriate criterion for making this decision. 1179 // XXX We need to make sure that the gen expansion 1180 // criterion dovetails well with this. XXX NEED TO FIX THIS 1181 if (_cmsGen->should_concurrent_collect()) { 1182 log.print("CMS old gen initiated"); 1183 return true; 1184 } 1185 1186 // We start a collection if we believe an incremental collection may fail; 1187 // this is not likely to be productive in practice because it's probably too 1188 // late anyway. 1189 CMSHeap* heap = CMSHeap::heap(); 1190 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1191 log.print("CMSCollector: collect because incremental collection will fail "); 1192 return true; 1193 } 1194 1195 if (MetaspaceGC::should_concurrent_collect()) { 1196 log.print("CMSCollector: collect for metadata allocation "); 1197 return true; 1198 } 1199 1200 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1201 if (CMSTriggerInterval >= 0) { 1202 if (CMSTriggerInterval == 0) { 1203 // Trigger always 1204 return true; 1205 } 1206 1207 // Check the CMS time since begin (we do not check the stats validity 1208 // as we want to be able to trigger the first CMS cycle as well) 1209 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1210 if (stats().valid()) { 1211 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1212 stats().cms_time_since_begin()); 1213 } else { 1214 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1215 } 1216 return true; 1217 } 1218 } 1219 1220 return false; 1221 } 1222 1223 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1224 1225 // Clear _expansion_cause fields of constituent generations 1226 void CMSCollector::clear_expansion_cause() { 1227 _cmsGen->clear_expansion_cause(); 1228 } 1229 1230 // We should be conservative in starting a collection cycle. To 1231 // start too eagerly runs the risk of collecting too often in the 1232 // extreme. To collect too rarely falls back on full collections, 1233 // which works, even if not optimum in terms of concurrent work. 1234 // As a work around for too eagerly collecting, use the flag 1235 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1236 // giving the user an easily understandable way of controlling the 1237 // collections. 1238 // We want to start a new collection cycle if any of the following 1239 // conditions hold: 1240 // . our current occupancy exceeds the configured initiating occupancy 1241 // for this generation, or 1242 // . we recently needed to expand this space and have not, since that 1243 // expansion, done a collection of this generation, or 1244 // . the underlying space believes that it may be a good idea to initiate 1245 // a concurrent collection (this may be based on criteria such as the 1246 // following: the space uses linear allocation and linear allocation is 1247 // going to fail, or there is believed to be excessive fragmentation in 1248 // the generation, etc... or ... 1249 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1250 // the case of the old generation; see CR 6543076): 1251 // we may be approaching a point at which allocation requests may fail because 1252 // we will be out of sufficient free space given allocation rate estimates.] 1253 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1254 1255 assert_lock_strong(freelistLock()); 1256 if (occupancy() > initiating_occupancy()) { 1257 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1258 short_name(), occupancy(), initiating_occupancy()); 1259 return true; 1260 } 1261 if (UseCMSInitiatingOccupancyOnly) { 1262 return false; 1263 } 1264 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1265 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1266 return true; 1267 } 1268 return false; 1269 } 1270 1271 void ConcurrentMarkSweepGeneration::collect(bool full, 1272 bool clear_all_soft_refs, 1273 size_t size, 1274 bool tlab) 1275 { 1276 collector()->collect(full, clear_all_soft_refs, size, tlab); 1277 } 1278 1279 void CMSCollector::collect(bool full, 1280 bool clear_all_soft_refs, 1281 size_t size, 1282 bool tlab) 1283 { 1284 // The following "if" branch is present for defensive reasons. 1285 // In the current uses of this interface, it can be replaced with: 1286 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1287 // But I am not placing that assert here to allow future 1288 // generality in invoking this interface. 1289 if (GCLocker::is_active()) { 1290 // A consistency test for GCLocker 1291 assert(GCLocker::needs_gc(), "Should have been set already"); 1292 // Skip this foreground collection, instead 1293 // expanding the heap if necessary. 1294 // Need the free list locks for the call to free() in compute_new_size() 1295 compute_new_size(); 1296 return; 1297 } 1298 acquire_control_and_collect(full, clear_all_soft_refs); 1299 } 1300 1301 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1302 CMSHeap* heap = CMSHeap::heap(); 1303 unsigned int gc_count = heap->total_full_collections(); 1304 if (gc_count == full_gc_count) { 1305 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1306 _full_gc_requested = true; 1307 _full_gc_cause = cause; 1308 CGC_lock->notify(); // nudge CMS thread 1309 } else { 1310 assert(gc_count > full_gc_count, "Error: causal loop"); 1311 } 1312 } 1313 1314 bool CMSCollector::is_external_interruption() { 1315 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1316 return GCCause::is_user_requested_gc(cause) || 1317 GCCause::is_serviceability_requested_gc(cause); 1318 } 1319 1320 void CMSCollector::report_concurrent_mode_interruption() { 1321 if (is_external_interruption()) { 1322 log_debug(gc)("Concurrent mode interrupted"); 1323 } else { 1324 log_debug(gc)("Concurrent mode failure"); 1325 _gc_tracer_cm->report_concurrent_mode_failure(); 1326 } 1327 } 1328 1329 1330 // The foreground and background collectors need to coordinate in order 1331 // to make sure that they do not mutually interfere with CMS collections. 1332 // When a background collection is active, 1333 // the foreground collector may need to take over (preempt) and 1334 // synchronously complete an ongoing collection. Depending on the 1335 // frequency of the background collections and the heap usage 1336 // of the application, this preemption can be seldom or frequent. 1337 // There are only certain 1338 // points in the background collection that the "collection-baton" 1339 // can be passed to the foreground collector. 1340 // 1341 // The foreground collector will wait for the baton before 1342 // starting any part of the collection. The foreground collector 1343 // will only wait at one location. 1344 // 1345 // The background collector will yield the baton before starting a new 1346 // phase of the collection (e.g., before initial marking, marking from roots, 1347 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1348 // of the loop which switches the phases. The background collector does some 1349 // of the phases (initial mark, final re-mark) with the world stopped. 1350 // Because of locking involved in stopping the world, 1351 // the foreground collector should not block waiting for the background 1352 // collector when it is doing a stop-the-world phase. The background 1353 // collector will yield the baton at an additional point just before 1354 // it enters a stop-the-world phase. Once the world is stopped, the 1355 // background collector checks the phase of the collection. If the 1356 // phase has not changed, it proceeds with the collection. If the 1357 // phase has changed, it skips that phase of the collection. See 1358 // the comments on the use of the Heap_lock in collect_in_background(). 1359 // 1360 // Variable used in baton passing. 1361 // _foregroundGCIsActive - Set to true by the foreground collector when 1362 // it wants the baton. The foreground clears it when it has finished 1363 // the collection. 1364 // _foregroundGCShouldWait - Set to true by the background collector 1365 // when it is running. The foreground collector waits while 1366 // _foregroundGCShouldWait is true. 1367 // CGC_lock - monitor used to protect access to the above variables 1368 // and to notify the foreground and background collectors. 1369 // _collectorState - current state of the CMS collection. 1370 // 1371 // The foreground collector 1372 // acquires the CGC_lock 1373 // sets _foregroundGCIsActive 1374 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1375 // various locks acquired in preparation for the collection 1376 // are released so as not to block the background collector 1377 // that is in the midst of a collection 1378 // proceeds with the collection 1379 // clears _foregroundGCIsActive 1380 // returns 1381 // 1382 // The background collector in a loop iterating on the phases of the 1383 // collection 1384 // acquires the CGC_lock 1385 // sets _foregroundGCShouldWait 1386 // if _foregroundGCIsActive is set 1387 // clears _foregroundGCShouldWait, notifies _CGC_lock 1388 // waits on _CGC_lock for _foregroundGCIsActive to become false 1389 // and exits the loop. 1390 // otherwise 1391 // proceed with that phase of the collection 1392 // if the phase is a stop-the-world phase, 1393 // yield the baton once more just before enqueueing 1394 // the stop-world CMS operation (executed by the VM thread). 1395 // returns after all phases of the collection are done 1396 // 1397 1398 void CMSCollector::acquire_control_and_collect(bool full, 1399 bool clear_all_soft_refs) { 1400 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1401 assert(!Thread::current()->is_ConcurrentGC_thread(), 1402 "shouldn't try to acquire control from self!"); 1403 1404 // Start the protocol for acquiring control of the 1405 // collection from the background collector (aka CMS thread). 1406 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1407 "VM thread should have CMS token"); 1408 // Remember the possibly interrupted state of an ongoing 1409 // concurrent collection 1410 CollectorState first_state = _collectorState; 1411 1412 // Signal to a possibly ongoing concurrent collection that 1413 // we want to do a foreground collection. 1414 _foregroundGCIsActive = true; 1415 1416 // release locks and wait for a notify from the background collector 1417 // releasing the locks in only necessary for phases which 1418 // do yields to improve the granularity of the collection. 1419 assert_lock_strong(bitMapLock()); 1420 // We need to lock the Free list lock for the space that we are 1421 // currently collecting. 1422 assert(haveFreelistLocks(), "Must be holding free list locks"); 1423 bitMapLock()->unlock(); 1424 releaseFreelistLocks(); 1425 { 1426 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1427 if (_foregroundGCShouldWait) { 1428 // We are going to be waiting for action for the CMS thread; 1429 // it had better not be gone (for instance at shutdown)! 1430 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1431 "CMS thread must be running"); 1432 // Wait here until the background collector gives us the go-ahead 1433 ConcurrentMarkSweepThread::clear_CMS_flag( 1434 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1435 // Get a possibly blocked CMS thread going: 1436 // Note that we set _foregroundGCIsActive true above, 1437 // without protection of the CGC_lock. 1438 CGC_lock->notify(); 1439 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1440 "Possible deadlock"); 1441 while (_foregroundGCShouldWait) { 1442 // wait for notification 1443 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1444 // Possibility of delay/starvation here, since CMS token does 1445 // not know to give priority to VM thread? Actually, i think 1446 // there wouldn't be any delay/starvation, but the proof of 1447 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1448 } 1449 ConcurrentMarkSweepThread::set_CMS_flag( 1450 ConcurrentMarkSweepThread::CMS_vm_has_token); 1451 } 1452 } 1453 // The CMS_token is already held. Get back the other locks. 1454 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1455 "VM thread should have CMS token"); 1456 getFreelistLocks(); 1457 bitMapLock()->lock_without_safepoint_check(); 1458 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1459 p2i(Thread::current()), first_state); 1460 log_debug(gc, state)(" gets control with state %d", _collectorState); 1461 1462 // Inform cms gen if this was due to partial collection failing. 1463 // The CMS gen may use this fact to determine its expansion policy. 1464 CMSHeap* heap = CMSHeap::heap(); 1465 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1466 assert(!_cmsGen->incremental_collection_failed(), 1467 "Should have been noticed, reacted to and cleared"); 1468 _cmsGen->set_incremental_collection_failed(); 1469 } 1470 1471 if (first_state > Idling) { 1472 report_concurrent_mode_interruption(); 1473 } 1474 1475 set_did_compact(true); 1476 1477 // If the collection is being acquired from the background 1478 // collector, there may be references on the discovered 1479 // references lists. Abandon those references, since some 1480 // of them may have become unreachable after concurrent 1481 // discovery; the STW compacting collector will redo discovery 1482 // more precisely, without being subject to floating garbage. 1483 // Leaving otherwise unreachable references in the discovered 1484 // lists would require special handling. 1485 ref_processor()->disable_discovery(); 1486 ref_processor()->abandon_partial_discovery(); 1487 ref_processor()->verify_no_references_recorded(); 1488 1489 if (first_state > Idling) { 1490 save_heap_summary(); 1491 } 1492 1493 do_compaction_work(clear_all_soft_refs); 1494 1495 // Has the GC time limit been exceeded? 1496 size_t max_eden_size = _young_gen->max_eden_size(); 1497 GCCause::Cause gc_cause = heap->gc_cause(); 1498 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1499 _young_gen->eden()->used(), 1500 _cmsGen->max_capacity(), 1501 max_eden_size, 1502 full, 1503 gc_cause, 1504 heap->soft_ref_policy()); 1505 1506 // Reset the expansion cause, now that we just completed 1507 // a collection cycle. 1508 clear_expansion_cause(); 1509 _foregroundGCIsActive = false; 1510 return; 1511 } 1512 1513 // Resize the tenured generation 1514 // after obtaining the free list locks for the 1515 // two generations. 1516 void CMSCollector::compute_new_size() { 1517 assert_locked_or_safepoint(Heap_lock); 1518 FreelistLocker z(this); 1519 MetaspaceGC::compute_new_size(); 1520 _cmsGen->compute_new_size_free_list(); 1521 } 1522 1523 // A work method used by the foreground collector to do 1524 // a mark-sweep-compact. 1525 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1526 CMSHeap* heap = CMSHeap::heap(); 1527 1528 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1529 gc_timer->register_gc_start(); 1530 1531 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1532 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1533 1534 heap->pre_full_gc_dump(gc_timer); 1535 1536 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1537 1538 // Temporarily widen the span of the weak reference processing to 1539 // the entire heap. 1540 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1541 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1542 // Temporarily, clear the "is_alive_non_header" field of the 1543 // reference processor. 1544 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1545 // Temporarily make reference _processing_ single threaded (non-MT). 1546 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1547 // Temporarily make refs discovery atomic 1548 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1549 // Temporarily make reference _discovery_ single threaded (non-MT) 1550 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1551 1552 ref_processor()->set_enqueuing_is_done(false); 1553 ref_processor()->enable_discovery(); 1554 ref_processor()->setup_policy(clear_all_soft_refs); 1555 // If an asynchronous collection finishes, the _modUnionTable is 1556 // all clear. If we are assuming the collection from an asynchronous 1557 // collection, clear the _modUnionTable. 1558 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1559 "_modUnionTable should be clear if the baton was not passed"); 1560 _modUnionTable.clear_all(); 1561 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1562 "mod union for klasses should be clear if the baton was passed"); 1563 _ct->cld_rem_set()->clear_mod_union(); 1564 1565 1566 // We must adjust the allocation statistics being maintained 1567 // in the free list space. We do so by reading and clearing 1568 // the sweep timer and updating the block flux rate estimates below. 1569 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1570 if (_inter_sweep_timer.is_active()) { 1571 _inter_sweep_timer.stop(); 1572 // Note that we do not use this sample to update the _inter_sweep_estimate. 1573 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1574 _inter_sweep_estimate.padded_average(), 1575 _intra_sweep_estimate.padded_average()); 1576 } 1577 1578 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1579 #ifdef ASSERT 1580 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1581 size_t free_size = cms_space->free(); 1582 assert(free_size == 1583 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1584 * HeapWordSize, 1585 "All the free space should be compacted into one chunk at top"); 1586 assert(cms_space->dictionary()->total_chunk_size( 1587 debug_only(cms_space->freelistLock())) == 0 || 1588 cms_space->totalSizeInIndexedFreeLists() == 0, 1589 "All the free space should be in a single chunk"); 1590 size_t num = cms_space->totalCount(); 1591 assert((free_size == 0 && num == 0) || 1592 (free_size > 0 && (num == 1 || num == 2)), 1593 "There should be at most 2 free chunks after compaction"); 1594 #endif // ASSERT 1595 _collectorState = Resetting; 1596 assert(_restart_addr == NULL, 1597 "Should have been NULL'd before baton was passed"); 1598 reset_stw(); 1599 _cmsGen->reset_after_compaction(); 1600 _concurrent_cycles_since_last_unload = 0; 1601 1602 // Clear any data recorded in the PLAB chunk arrays. 1603 if (_survivor_plab_array != NULL) { 1604 reset_survivor_plab_arrays(); 1605 } 1606 1607 // Adjust the per-size allocation stats for the next epoch. 1608 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1609 // Restart the "inter sweep timer" for the next epoch. 1610 _inter_sweep_timer.reset(); 1611 _inter_sweep_timer.start(); 1612 1613 // No longer a need to do a concurrent collection for Metaspace. 1614 MetaspaceGC::set_should_concurrent_collect(false); 1615 1616 heap->post_full_gc_dump(gc_timer); 1617 1618 gc_timer->register_gc_end(); 1619 1620 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1621 1622 // For a mark-sweep-compact, compute_new_size() will be called 1623 // in the heap's do_collection() method. 1624 } 1625 1626 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1627 Log(gc, heap) log; 1628 if (!log.is_trace()) { 1629 return; 1630 } 1631 1632 ContiguousSpace* eden_space = _young_gen->eden(); 1633 ContiguousSpace* from_space = _young_gen->from(); 1634 ContiguousSpace* to_space = _young_gen->to(); 1635 // Eden 1636 if (_eden_chunk_array != NULL) { 1637 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1638 p2i(eden_space->bottom()), p2i(eden_space->top()), 1639 p2i(eden_space->end()), eden_space->capacity()); 1640 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1641 _eden_chunk_index, _eden_chunk_capacity); 1642 for (size_t i = 0; i < _eden_chunk_index; i++) { 1643 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1644 } 1645 } 1646 // Survivor 1647 if (_survivor_chunk_array != NULL) { 1648 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1649 p2i(from_space->bottom()), p2i(from_space->top()), 1650 p2i(from_space->end()), from_space->capacity()); 1651 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1652 _survivor_chunk_index, _survivor_chunk_capacity); 1653 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1654 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1655 } 1656 } 1657 } 1658 1659 void CMSCollector::getFreelistLocks() const { 1660 // Get locks for all free lists in all generations that this 1661 // collector is responsible for 1662 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1663 } 1664 1665 void CMSCollector::releaseFreelistLocks() const { 1666 // Release locks for all free lists in all generations that this 1667 // collector is responsible for 1668 _cmsGen->freelistLock()->unlock(); 1669 } 1670 1671 bool CMSCollector::haveFreelistLocks() const { 1672 // Check locks for all free lists in all generations that this 1673 // collector is responsible for 1674 assert_lock_strong(_cmsGen->freelistLock()); 1675 PRODUCT_ONLY(ShouldNotReachHere()); 1676 return true; 1677 } 1678 1679 // A utility class that is used by the CMS collector to 1680 // temporarily "release" the foreground collector from its 1681 // usual obligation to wait for the background collector to 1682 // complete an ongoing phase before proceeding. 1683 class ReleaseForegroundGC: public StackObj { 1684 private: 1685 CMSCollector* _c; 1686 public: 1687 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1688 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1689 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1690 // allow a potentially blocked foreground collector to proceed 1691 _c->_foregroundGCShouldWait = false; 1692 if (_c->_foregroundGCIsActive) { 1693 CGC_lock->notify(); 1694 } 1695 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1696 "Possible deadlock"); 1697 } 1698 1699 ~ReleaseForegroundGC() { 1700 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1701 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1702 _c->_foregroundGCShouldWait = true; 1703 } 1704 }; 1705 1706 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1707 assert(Thread::current()->is_ConcurrentGC_thread(), 1708 "A CMS asynchronous collection is only allowed on a CMS thread."); 1709 1710 CMSHeap* heap = CMSHeap::heap(); 1711 { 1712 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1713 MutexLockerEx hl(Heap_lock, safepoint_check); 1714 FreelistLocker fll(this); 1715 MutexLockerEx x(CGC_lock, safepoint_check); 1716 if (_foregroundGCIsActive) { 1717 // The foreground collector is. Skip this 1718 // background collection. 1719 assert(!_foregroundGCShouldWait, "Should be clear"); 1720 return; 1721 } else { 1722 assert(_collectorState == Idling, "Should be idling before start."); 1723 _collectorState = InitialMarking; 1724 register_gc_start(cause); 1725 // Reset the expansion cause, now that we are about to begin 1726 // a new cycle. 1727 clear_expansion_cause(); 1728 1729 // Clear the MetaspaceGC flag since a concurrent collection 1730 // is starting but also clear it after the collection. 1731 MetaspaceGC::set_should_concurrent_collect(false); 1732 } 1733 // Decide if we want to enable class unloading as part of the 1734 // ensuing concurrent GC cycle. 1735 update_should_unload_classes(); 1736 _full_gc_requested = false; // acks all outstanding full gc requests 1737 _full_gc_cause = GCCause::_no_gc; 1738 // Signal that we are about to start a collection 1739 heap->increment_total_full_collections(); // ... starting a collection cycle 1740 _collection_count_start = heap->total_full_collections(); 1741 } 1742 1743 size_t prev_used = _cmsGen->used(); 1744 1745 // The change of the collection state is normally done at this level; 1746 // the exceptions are phases that are executed while the world is 1747 // stopped. For those phases the change of state is done while the 1748 // world is stopped. For baton passing purposes this allows the 1749 // background collector to finish the phase and change state atomically. 1750 // The foreground collector cannot wait on a phase that is done 1751 // while the world is stopped because the foreground collector already 1752 // has the world stopped and would deadlock. 1753 while (_collectorState != Idling) { 1754 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1755 p2i(Thread::current()), _collectorState); 1756 // The foreground collector 1757 // holds the Heap_lock throughout its collection. 1758 // holds the CMS token (but not the lock) 1759 // except while it is waiting for the background collector to yield. 1760 // 1761 // The foreground collector should be blocked (not for long) 1762 // if the background collector is about to start a phase 1763 // executed with world stopped. If the background 1764 // collector has already started such a phase, the 1765 // foreground collector is blocked waiting for the 1766 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1767 // are executed in the VM thread. 1768 // 1769 // The locking order is 1770 // PendingListLock (PLL) -- if applicable (FinalMarking) 1771 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1772 // CMS token (claimed in 1773 // stop_world_and_do() --> 1774 // safepoint_synchronize() --> 1775 // CMSThread::synchronize()) 1776 1777 { 1778 // Check if the FG collector wants us to yield. 1779 CMSTokenSync x(true); // is cms thread 1780 if (waitForForegroundGC()) { 1781 // We yielded to a foreground GC, nothing more to be 1782 // done this round. 1783 assert(_foregroundGCShouldWait == false, "We set it to false in " 1784 "waitForForegroundGC()"); 1785 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1786 p2i(Thread::current()), _collectorState); 1787 return; 1788 } else { 1789 // The background collector can run but check to see if the 1790 // foreground collector has done a collection while the 1791 // background collector was waiting to get the CGC_lock 1792 // above. If yes, break so that _foregroundGCShouldWait 1793 // is cleared before returning. 1794 if (_collectorState == Idling) { 1795 break; 1796 } 1797 } 1798 } 1799 1800 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1801 "should be waiting"); 1802 1803 switch (_collectorState) { 1804 case InitialMarking: 1805 { 1806 ReleaseForegroundGC x(this); 1807 stats().record_cms_begin(); 1808 VM_CMS_Initial_Mark initial_mark_op(this); 1809 VMThread::execute(&initial_mark_op); 1810 } 1811 // The collector state may be any legal state at this point 1812 // since the background collector may have yielded to the 1813 // foreground collector. 1814 break; 1815 case Marking: 1816 // initial marking in checkpointRootsInitialWork has been completed 1817 if (markFromRoots()) { // we were successful 1818 assert(_collectorState == Precleaning, "Collector state should " 1819 "have changed"); 1820 } else { 1821 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1822 } 1823 break; 1824 case Precleaning: 1825 // marking from roots in markFromRoots has been completed 1826 preclean(); 1827 assert(_collectorState == AbortablePreclean || 1828 _collectorState == FinalMarking, 1829 "Collector state should have changed"); 1830 break; 1831 case AbortablePreclean: 1832 abortable_preclean(); 1833 assert(_collectorState == FinalMarking, "Collector state should " 1834 "have changed"); 1835 break; 1836 case FinalMarking: 1837 { 1838 ReleaseForegroundGC x(this); 1839 1840 VM_CMS_Final_Remark final_remark_op(this); 1841 VMThread::execute(&final_remark_op); 1842 } 1843 assert(_foregroundGCShouldWait, "block post-condition"); 1844 break; 1845 case Sweeping: 1846 // final marking in checkpointRootsFinal has been completed 1847 sweep(); 1848 assert(_collectorState == Resizing, "Collector state change " 1849 "to Resizing must be done under the free_list_lock"); 1850 1851 case Resizing: { 1852 // Sweeping has been completed... 1853 // At this point the background collection has completed. 1854 // Don't move the call to compute_new_size() down 1855 // into code that might be executed if the background 1856 // collection was preempted. 1857 { 1858 ReleaseForegroundGC x(this); // unblock FG collection 1859 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1860 CMSTokenSync z(true); // not strictly needed. 1861 if (_collectorState == Resizing) { 1862 compute_new_size(); 1863 save_heap_summary(); 1864 _collectorState = Resetting; 1865 } else { 1866 assert(_collectorState == Idling, "The state should only change" 1867 " because the foreground collector has finished the collection"); 1868 } 1869 } 1870 break; 1871 } 1872 case Resetting: 1873 // CMS heap resizing has been completed 1874 reset_concurrent(); 1875 assert(_collectorState == Idling, "Collector state should " 1876 "have changed"); 1877 1878 MetaspaceGC::set_should_concurrent_collect(false); 1879 1880 stats().record_cms_end(); 1881 // Don't move the concurrent_phases_end() and compute_new_size() 1882 // calls to here because a preempted background collection 1883 // has it's state set to "Resetting". 1884 break; 1885 case Idling: 1886 default: 1887 ShouldNotReachHere(); 1888 break; 1889 } 1890 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1891 p2i(Thread::current()), _collectorState); 1892 assert(_foregroundGCShouldWait, "block post-condition"); 1893 } 1894 1895 // Should this be in gc_epilogue? 1896 heap->counters()->update_counters(); 1897 1898 { 1899 // Clear _foregroundGCShouldWait and, in the event that the 1900 // foreground collector is waiting, notify it, before 1901 // returning. 1902 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1903 _foregroundGCShouldWait = false; 1904 if (_foregroundGCIsActive) { 1905 CGC_lock->notify(); 1906 } 1907 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1908 "Possible deadlock"); 1909 } 1910 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1911 p2i(Thread::current()), _collectorState); 1912 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1913 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1914 } 1915 1916 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1917 _cms_start_registered = true; 1918 _gc_timer_cm->register_gc_start(); 1919 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1920 } 1921 1922 void CMSCollector::register_gc_end() { 1923 if (_cms_start_registered) { 1924 report_heap_summary(GCWhen::AfterGC); 1925 1926 _gc_timer_cm->register_gc_end(); 1927 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1928 _cms_start_registered = false; 1929 } 1930 } 1931 1932 void CMSCollector::save_heap_summary() { 1933 CMSHeap* heap = CMSHeap::heap(); 1934 _last_heap_summary = heap->create_heap_summary(); 1935 _last_metaspace_summary = heap->create_metaspace_summary(); 1936 } 1937 1938 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1939 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1940 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1941 } 1942 1943 bool CMSCollector::waitForForegroundGC() { 1944 bool res = false; 1945 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1946 "CMS thread should have CMS token"); 1947 // Block the foreground collector until the 1948 // background collectors decides whether to 1949 // yield. 1950 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1951 _foregroundGCShouldWait = true; 1952 if (_foregroundGCIsActive) { 1953 // The background collector yields to the 1954 // foreground collector and returns a value 1955 // indicating that it has yielded. The foreground 1956 // collector can proceed. 1957 res = true; 1958 _foregroundGCShouldWait = false; 1959 ConcurrentMarkSweepThread::clear_CMS_flag( 1960 ConcurrentMarkSweepThread::CMS_cms_has_token); 1961 ConcurrentMarkSweepThread::set_CMS_flag( 1962 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1963 // Get a possibly blocked foreground thread going 1964 CGC_lock->notify(); 1965 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1966 p2i(Thread::current()), _collectorState); 1967 while (_foregroundGCIsActive) { 1968 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1969 } 1970 ConcurrentMarkSweepThread::set_CMS_flag( 1971 ConcurrentMarkSweepThread::CMS_cms_has_token); 1972 ConcurrentMarkSweepThread::clear_CMS_flag( 1973 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1974 } 1975 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1976 p2i(Thread::current()), _collectorState); 1977 return res; 1978 } 1979 1980 // Because of the need to lock the free lists and other structures in 1981 // the collector, common to all the generations that the collector is 1982 // collecting, we need the gc_prologues of individual CMS generations 1983 // delegate to their collector. It may have been simpler had the 1984 // current infrastructure allowed one to call a prologue on a 1985 // collector. In the absence of that we have the generation's 1986 // prologue delegate to the collector, which delegates back 1987 // some "local" work to a worker method in the individual generations 1988 // that it's responsible for collecting, while itself doing any 1989 // work common to all generations it's responsible for. A similar 1990 // comment applies to the gc_epilogue()'s. 1991 // The role of the variable _between_prologue_and_epilogue is to 1992 // enforce the invocation protocol. 1993 void CMSCollector::gc_prologue(bool full) { 1994 // Call gc_prologue_work() for the CMSGen 1995 // we are responsible for. 1996 1997 // The following locking discipline assumes that we are only called 1998 // when the world is stopped. 1999 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2000 2001 // The CMSCollector prologue must call the gc_prologues for the 2002 // "generations" that it's responsible 2003 // for. 2004 2005 assert( Thread::current()->is_VM_thread() 2006 || ( CMSScavengeBeforeRemark 2007 && Thread::current()->is_ConcurrentGC_thread()), 2008 "Incorrect thread type for prologue execution"); 2009 2010 if (_between_prologue_and_epilogue) { 2011 // We have already been invoked; this is a gc_prologue delegation 2012 // from yet another CMS generation that we are responsible for, just 2013 // ignore it since all relevant work has already been done. 2014 return; 2015 } 2016 2017 // set a bit saying prologue has been called; cleared in epilogue 2018 _between_prologue_and_epilogue = true; 2019 // Claim locks for common data structures, then call gc_prologue_work() 2020 // for each CMSGen. 2021 2022 getFreelistLocks(); // gets free list locks on constituent spaces 2023 bitMapLock()->lock_without_safepoint_check(); 2024 2025 // Should call gc_prologue_work() for all cms gens we are responsible for 2026 bool duringMarking = _collectorState >= Marking 2027 && _collectorState < Sweeping; 2028 2029 // The young collections clear the modified oops state, which tells if 2030 // there are any modified oops in the class. The remark phase also needs 2031 // that information. Tell the young collection to save the union of all 2032 // modified klasses. 2033 if (duringMarking) { 2034 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2035 } 2036 2037 bool registerClosure = duringMarking; 2038 2039 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2040 2041 if (!full) { 2042 stats().record_gc0_begin(); 2043 } 2044 } 2045 2046 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2047 2048 _capacity_at_prologue = capacity(); 2049 _used_at_prologue = used(); 2050 2051 // We enable promotion tracking so that card-scanning can recognize 2052 // which objects have been promoted during this GC and skip them. 2053 for (uint i = 0; i < ParallelGCThreads; i++) { 2054 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2055 } 2056 2057 // Delegate to CMScollector which knows how to coordinate between 2058 // this and any other CMS generations that it is responsible for 2059 // collecting. 2060 collector()->gc_prologue(full); 2061 } 2062 2063 // This is a "private" interface for use by this generation's CMSCollector. 2064 // Not to be called directly by any other entity (for instance, 2065 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2066 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2067 bool registerClosure, ModUnionClosure* modUnionClosure) { 2068 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2069 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2070 "Should be NULL"); 2071 if (registerClosure) { 2072 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2073 } 2074 cmsSpace()->gc_prologue(); 2075 // Clear stat counters 2076 NOT_PRODUCT( 2077 assert(_numObjectsPromoted == 0, "check"); 2078 assert(_numWordsPromoted == 0, "check"); 2079 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2080 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2081 _numObjectsAllocated = 0; 2082 _numWordsAllocated = 0; 2083 ) 2084 } 2085 2086 void CMSCollector::gc_epilogue(bool full) { 2087 // The following locking discipline assumes that we are only called 2088 // when the world is stopped. 2089 assert(SafepointSynchronize::is_at_safepoint(), 2090 "world is stopped assumption"); 2091 2092 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2093 // if linear allocation blocks need to be appropriately marked to allow the 2094 // the blocks to be parsable. We also check here whether we need to nudge the 2095 // CMS collector thread to start a new cycle (if it's not already active). 2096 assert( Thread::current()->is_VM_thread() 2097 || ( CMSScavengeBeforeRemark 2098 && Thread::current()->is_ConcurrentGC_thread()), 2099 "Incorrect thread type for epilogue execution"); 2100 2101 if (!_between_prologue_and_epilogue) { 2102 // We have already been invoked; this is a gc_epilogue delegation 2103 // from yet another CMS generation that we are responsible for, just 2104 // ignore it since all relevant work has already been done. 2105 return; 2106 } 2107 assert(haveFreelistLocks(), "must have freelist locks"); 2108 assert_lock_strong(bitMapLock()); 2109 2110 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2111 2112 _cmsGen->gc_epilogue_work(full); 2113 2114 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2115 // in case sampling was not already enabled, enable it 2116 _start_sampling = true; 2117 } 2118 // reset _eden_chunk_array so sampling starts afresh 2119 _eden_chunk_index = 0; 2120 2121 size_t cms_used = _cmsGen->cmsSpace()->used(); 2122 2123 // update performance counters - this uses a special version of 2124 // update_counters() that allows the utilization to be passed as a 2125 // parameter, avoiding multiple calls to used(). 2126 // 2127 _cmsGen->update_counters(cms_used); 2128 2129 bitMapLock()->unlock(); 2130 releaseFreelistLocks(); 2131 2132 if (!CleanChunkPoolAsync) { 2133 Chunk::clean_chunk_pool(); 2134 } 2135 2136 set_did_compact(false); 2137 _between_prologue_and_epilogue = false; // ready for next cycle 2138 } 2139 2140 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2141 collector()->gc_epilogue(full); 2142 2143 // When using ParNew, promotion tracking should have already been 2144 // disabled. However, the prologue (which enables promotion 2145 // tracking) and epilogue are called irrespective of the type of 2146 // GC. So they will also be called before and after Full GCs, during 2147 // which promotion tracking will not be explicitly disabled. So, 2148 // it's safer to also disable it here too (to be symmetric with 2149 // enabling it in the prologue). 2150 for (uint i = 0; i < ParallelGCThreads; i++) { 2151 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2152 } 2153 } 2154 2155 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2156 assert(!incremental_collection_failed(), "Should have been cleared"); 2157 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2158 cmsSpace()->gc_epilogue(); 2159 // Print stat counters 2160 NOT_PRODUCT( 2161 assert(_numObjectsAllocated == 0, "check"); 2162 assert(_numWordsAllocated == 0, "check"); 2163 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2164 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2165 _numObjectsPromoted = 0; 2166 _numWordsPromoted = 0; 2167 ) 2168 2169 // Call down the chain in contiguous_available needs the freelistLock 2170 // so print this out before releasing the freeListLock. 2171 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2172 } 2173 2174 #ifndef PRODUCT 2175 bool CMSCollector::have_cms_token() { 2176 Thread* thr = Thread::current(); 2177 if (thr->is_VM_thread()) { 2178 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2179 } else if (thr->is_ConcurrentGC_thread()) { 2180 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2181 } else if (thr->is_GC_task_thread()) { 2182 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2183 ParGCRareEvent_lock->owned_by_self(); 2184 } 2185 return false; 2186 } 2187 2188 // Check reachability of the given heap address in CMS generation, 2189 // treating all other generations as roots. 2190 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2191 // We could "guarantee" below, rather than assert, but I'll 2192 // leave these as "asserts" so that an adventurous debugger 2193 // could try this in the product build provided some subset of 2194 // the conditions were met, provided they were interested in the 2195 // results and knew that the computation below wouldn't interfere 2196 // with other concurrent computations mutating the structures 2197 // being read or written. 2198 assert(SafepointSynchronize::is_at_safepoint(), 2199 "Else mutations in object graph will make answer suspect"); 2200 assert(have_cms_token(), "Should hold cms token"); 2201 assert(haveFreelistLocks(), "must hold free list locks"); 2202 assert_lock_strong(bitMapLock()); 2203 2204 // Clear the marking bit map array before starting, but, just 2205 // for kicks, first report if the given address is already marked 2206 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2207 _markBitMap.isMarked(addr) ? "" : " not"); 2208 2209 if (verify_after_remark()) { 2210 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2211 bool result = verification_mark_bm()->isMarked(addr); 2212 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2213 result ? "IS" : "is NOT"); 2214 return result; 2215 } else { 2216 tty->print_cr("Could not compute result"); 2217 return false; 2218 } 2219 } 2220 #endif 2221 2222 void 2223 CMSCollector::print_on_error(outputStream* st) { 2224 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2225 if (collector != NULL) { 2226 CMSBitMap* bitmap = &collector->_markBitMap; 2227 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2228 bitmap->print_on_error(st, " Bits: "); 2229 2230 st->cr(); 2231 2232 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2233 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2234 mut_bitmap->print_on_error(st, " Bits: "); 2235 } 2236 } 2237 2238 //////////////////////////////////////////////////////// 2239 // CMS Verification Support 2240 //////////////////////////////////////////////////////// 2241 // Following the remark phase, the following invariant 2242 // should hold -- each object in the CMS heap which is 2243 // marked in markBitMap() should be marked in the verification_mark_bm(). 2244 2245 class VerifyMarkedClosure: public BitMapClosure { 2246 CMSBitMap* _marks; 2247 bool _failed; 2248 2249 public: 2250 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2251 2252 bool do_bit(size_t offset) { 2253 HeapWord* addr = _marks->offsetToHeapWord(offset); 2254 if (!_marks->isMarked(addr)) { 2255 Log(gc, verify) log; 2256 ResourceMark rm; 2257 LogStream ls(log.error()); 2258 oop(addr)->print_on(&ls); 2259 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2260 _failed = true; 2261 } 2262 return true; 2263 } 2264 2265 bool failed() { return _failed; } 2266 }; 2267 2268 bool CMSCollector::verify_after_remark() { 2269 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2270 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2271 static bool init = false; 2272 2273 assert(SafepointSynchronize::is_at_safepoint(), 2274 "Else mutations in object graph will make answer suspect"); 2275 assert(have_cms_token(), 2276 "Else there may be mutual interference in use of " 2277 " verification data structures"); 2278 assert(_collectorState > Marking && _collectorState <= Sweeping, 2279 "Else marking info checked here may be obsolete"); 2280 assert(haveFreelistLocks(), "must hold free list locks"); 2281 assert_lock_strong(bitMapLock()); 2282 2283 2284 // Allocate marking bit map if not already allocated 2285 if (!init) { // first time 2286 if (!verification_mark_bm()->allocate(_span)) { 2287 return false; 2288 } 2289 init = true; 2290 } 2291 2292 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2293 2294 // Turn off refs discovery -- so we will be tracing through refs. 2295 // This is as intended, because by this time 2296 // GC must already have cleared any refs that need to be cleared, 2297 // and traced those that need to be marked; moreover, 2298 // the marking done here is not going to interfere in any 2299 // way with the marking information used by GC. 2300 NoRefDiscovery no_discovery(ref_processor()); 2301 2302 #if COMPILER2_OR_JVMCI 2303 DerivedPointerTableDeactivate dpt_deact; 2304 #endif 2305 2306 // Clear any marks from a previous round 2307 verification_mark_bm()->clear_all(); 2308 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2309 verify_work_stacks_empty(); 2310 2311 CMSHeap* heap = CMSHeap::heap(); 2312 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2313 // Update the saved marks which may affect the root scans. 2314 heap->save_marks(); 2315 2316 if (CMSRemarkVerifyVariant == 1) { 2317 // In this first variant of verification, we complete 2318 // all marking, then check if the new marks-vector is 2319 // a subset of the CMS marks-vector. 2320 verify_after_remark_work_1(); 2321 } else { 2322 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2323 // In this second variant of verification, we flag an error 2324 // (i.e. an object reachable in the new marks-vector not reachable 2325 // in the CMS marks-vector) immediately, also indicating the 2326 // identify of an object (A) that references the unmarked object (B) -- 2327 // presumably, a mutation to A failed to be picked up by preclean/remark? 2328 verify_after_remark_work_2(); 2329 } 2330 2331 return true; 2332 } 2333 2334 void CMSCollector::verify_after_remark_work_1() { 2335 ResourceMark rm; 2336 HandleMark hm; 2337 CMSHeap* heap = CMSHeap::heap(); 2338 2339 // Get a clear set of claim bits for the roots processing to work with. 2340 ClassLoaderDataGraph::clear_claimed_marks(); 2341 2342 // Mark from roots one level into CMS 2343 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2344 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2345 2346 { 2347 StrongRootsScope srs(1); 2348 2349 heap->cms_process_roots(&srs, 2350 true, // young gen as roots 2351 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2352 should_unload_classes(), 2353 ¬Older, 2354 NULL); 2355 } 2356 2357 // Now mark from the roots 2358 MarkFromRootsClosure markFromRootsClosure(this, _span, 2359 verification_mark_bm(), verification_mark_stack(), 2360 false /* don't yield */, true /* verifying */); 2361 assert(_restart_addr == NULL, "Expected pre-condition"); 2362 verification_mark_bm()->iterate(&markFromRootsClosure); 2363 while (_restart_addr != NULL) { 2364 // Deal with stack overflow: by restarting at the indicated 2365 // address. 2366 HeapWord* ra = _restart_addr; 2367 markFromRootsClosure.reset(ra); 2368 _restart_addr = NULL; 2369 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2370 } 2371 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2372 verify_work_stacks_empty(); 2373 2374 // Marking completed -- now verify that each bit marked in 2375 // verification_mark_bm() is also marked in markBitMap(); flag all 2376 // errors by printing corresponding objects. 2377 VerifyMarkedClosure vcl(markBitMap()); 2378 verification_mark_bm()->iterate(&vcl); 2379 if (vcl.failed()) { 2380 Log(gc, verify) log; 2381 log.error("Failed marking verification after remark"); 2382 ResourceMark rm; 2383 LogStream ls(log.error()); 2384 heap->print_on(&ls); 2385 fatal("CMS: failed marking verification after remark"); 2386 } 2387 } 2388 2389 class VerifyCLDOopsCLDClosure : public CLDClosure { 2390 class VerifyCLDOopsClosure : public OopClosure { 2391 CMSBitMap* _bitmap; 2392 public: 2393 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2394 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2395 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2396 } _oop_closure; 2397 public: 2398 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2399 void do_cld(ClassLoaderData* cld) { 2400 cld->oops_do(&_oop_closure, false, false); 2401 } 2402 }; 2403 2404 void CMSCollector::verify_after_remark_work_2() { 2405 ResourceMark rm; 2406 HandleMark hm; 2407 CMSHeap* heap = CMSHeap::heap(); 2408 2409 // Get a clear set of claim bits for the roots processing to work with. 2410 ClassLoaderDataGraph::clear_claimed_marks(); 2411 2412 // Mark from roots one level into CMS 2413 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2414 markBitMap()); 2415 CLDToOopClosure cld_closure(¬Older, true); 2416 2417 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2418 2419 { 2420 StrongRootsScope srs(1); 2421 2422 heap->cms_process_roots(&srs, 2423 true, // young gen as roots 2424 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2425 should_unload_classes(), 2426 ¬Older, 2427 &cld_closure); 2428 } 2429 2430 // Now mark from the roots 2431 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2432 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2433 assert(_restart_addr == NULL, "Expected pre-condition"); 2434 verification_mark_bm()->iterate(&markFromRootsClosure); 2435 while (_restart_addr != NULL) { 2436 // Deal with stack overflow: by restarting at the indicated 2437 // address. 2438 HeapWord* ra = _restart_addr; 2439 markFromRootsClosure.reset(ra); 2440 _restart_addr = NULL; 2441 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2442 } 2443 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2444 verify_work_stacks_empty(); 2445 2446 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2447 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2448 2449 // Marking completed -- now verify that each bit marked in 2450 // verification_mark_bm() is also marked in markBitMap(); flag all 2451 // errors by printing corresponding objects. 2452 VerifyMarkedClosure vcl(markBitMap()); 2453 verification_mark_bm()->iterate(&vcl); 2454 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2455 } 2456 2457 void ConcurrentMarkSweepGeneration::save_marks() { 2458 // delegate to CMS space 2459 cmsSpace()->save_marks(); 2460 } 2461 2462 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2463 return cmsSpace()->no_allocs_since_save_marks(); 2464 } 2465 2466 void 2467 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2468 if (freelistLock()->owned_by_self()) { 2469 Generation::oop_iterate(cl); 2470 } else { 2471 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2472 Generation::oop_iterate(cl); 2473 } 2474 } 2475 2476 void 2477 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2478 if (freelistLock()->owned_by_self()) { 2479 Generation::object_iterate(cl); 2480 } else { 2481 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2482 Generation::object_iterate(cl); 2483 } 2484 } 2485 2486 void 2487 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2488 if (freelistLock()->owned_by_self()) { 2489 Generation::safe_object_iterate(cl); 2490 } else { 2491 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2492 Generation::safe_object_iterate(cl); 2493 } 2494 } 2495 2496 void 2497 ConcurrentMarkSweepGeneration::post_compact() { 2498 } 2499 2500 void 2501 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2502 // Fix the linear allocation blocks to look like free blocks. 2503 2504 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2505 // are not called when the heap is verified during universe initialization and 2506 // at vm shutdown. 2507 if (freelistLock()->owned_by_self()) { 2508 cmsSpace()->prepare_for_verify(); 2509 } else { 2510 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2511 cmsSpace()->prepare_for_verify(); 2512 } 2513 } 2514 2515 void 2516 ConcurrentMarkSweepGeneration::verify() { 2517 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2518 // are not called when the heap is verified during universe initialization and 2519 // at vm shutdown. 2520 if (freelistLock()->owned_by_self()) { 2521 cmsSpace()->verify(); 2522 } else { 2523 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2524 cmsSpace()->verify(); 2525 } 2526 } 2527 2528 void CMSCollector::verify() { 2529 _cmsGen->verify(); 2530 } 2531 2532 #ifndef PRODUCT 2533 bool CMSCollector::overflow_list_is_empty() const { 2534 assert(_num_par_pushes >= 0, "Inconsistency"); 2535 if (_overflow_list == NULL) { 2536 assert(_num_par_pushes == 0, "Inconsistency"); 2537 } 2538 return _overflow_list == NULL; 2539 } 2540 2541 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2542 // merely consolidate assertion checks that appear to occur together frequently. 2543 void CMSCollector::verify_work_stacks_empty() const { 2544 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2545 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2546 } 2547 2548 void CMSCollector::verify_overflow_empty() const { 2549 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2550 assert(no_preserved_marks(), "No preserved marks"); 2551 } 2552 #endif // PRODUCT 2553 2554 // Decide if we want to enable class unloading as part of the 2555 // ensuing concurrent GC cycle. We will collect and 2556 // unload classes if it's the case that: 2557 // (a) class unloading is enabled at the command line, and 2558 // (b) old gen is getting really full 2559 // NOTE: Provided there is no change in the state of the heap between 2560 // calls to this method, it should have idempotent results. Moreover, 2561 // its results should be monotonically increasing (i.e. going from 0 to 1, 2562 // but not 1 to 0) between successive calls between which the heap was 2563 // not collected. For the implementation below, it must thus rely on 2564 // the property that concurrent_cycles_since_last_unload() 2565 // will not decrease unless a collection cycle happened and that 2566 // _cmsGen->is_too_full() are 2567 // themselves also monotonic in that sense. See check_monotonicity() 2568 // below. 2569 void CMSCollector::update_should_unload_classes() { 2570 _should_unload_classes = false; 2571 if (CMSClassUnloadingEnabled) { 2572 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2573 CMSClassUnloadingMaxInterval) 2574 || _cmsGen->is_too_full(); 2575 } 2576 } 2577 2578 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2579 bool res = should_concurrent_collect(); 2580 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2581 return res; 2582 } 2583 2584 void CMSCollector::setup_cms_unloading_and_verification_state() { 2585 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2586 || VerifyBeforeExit; 2587 const int rso = GenCollectedHeap::SO_AllCodeCache; 2588 2589 // We set the proper root for this CMS cycle here. 2590 if (should_unload_classes()) { // Should unload classes this cycle 2591 remove_root_scanning_option(rso); // Shrink the root set appropriately 2592 set_verifying(should_verify); // Set verification state for this cycle 2593 return; // Nothing else needs to be done at this time 2594 } 2595 2596 // Not unloading classes this cycle 2597 assert(!should_unload_classes(), "Inconsistency!"); 2598 2599 // If we are not unloading classes then add SO_AllCodeCache to root 2600 // scanning options. 2601 add_root_scanning_option(rso); 2602 2603 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2604 set_verifying(true); 2605 } else if (verifying() && !should_verify) { 2606 // We were verifying, but some verification flags got disabled. 2607 set_verifying(false); 2608 // Exclude symbols, strings and code cache elements from root scanning to 2609 // reduce IM and RM pauses. 2610 remove_root_scanning_option(rso); 2611 } 2612 } 2613 2614 2615 #ifndef PRODUCT 2616 HeapWord* CMSCollector::block_start(const void* p) const { 2617 const HeapWord* addr = (HeapWord*)p; 2618 if (_span.contains(p)) { 2619 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2620 return _cmsGen->cmsSpace()->block_start(p); 2621 } 2622 } 2623 return NULL; 2624 } 2625 #endif 2626 2627 HeapWord* 2628 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2629 bool tlab, 2630 bool parallel) { 2631 CMSSynchronousYieldRequest yr; 2632 assert(!tlab, "Can't deal with TLAB allocation"); 2633 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2634 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2635 if (GCExpandToAllocateDelayMillis > 0) { 2636 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2637 } 2638 return have_lock_and_allocate(word_size, tlab); 2639 } 2640 2641 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2642 size_t bytes, 2643 size_t expand_bytes, 2644 CMSExpansionCause::Cause cause) 2645 { 2646 2647 bool success = expand(bytes, expand_bytes); 2648 2649 // remember why we expanded; this information is used 2650 // by shouldConcurrentCollect() when making decisions on whether to start 2651 // a new CMS cycle. 2652 if (success) { 2653 set_expansion_cause(cause); 2654 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2655 } 2656 } 2657 2658 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2659 HeapWord* res = NULL; 2660 MutexLocker x(ParGCRareEvent_lock); 2661 while (true) { 2662 // Expansion by some other thread might make alloc OK now: 2663 res = ps->lab.alloc(word_sz); 2664 if (res != NULL) return res; 2665 // If there's not enough expansion space available, give up. 2666 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2667 return NULL; 2668 } 2669 // Otherwise, we try expansion. 2670 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2671 // Now go around the loop and try alloc again; 2672 // A competing par_promote might beat us to the expansion space, 2673 // so we may go around the loop again if promotion fails again. 2674 if (GCExpandToAllocateDelayMillis > 0) { 2675 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2676 } 2677 } 2678 } 2679 2680 2681 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2682 PromotionInfo* promo) { 2683 MutexLocker x(ParGCRareEvent_lock); 2684 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2685 while (true) { 2686 // Expansion by some other thread might make alloc OK now: 2687 if (promo->ensure_spooling_space()) { 2688 assert(promo->has_spooling_space(), 2689 "Post-condition of successful ensure_spooling_space()"); 2690 return true; 2691 } 2692 // If there's not enough expansion space available, give up. 2693 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2694 return false; 2695 } 2696 // Otherwise, we try expansion. 2697 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2698 // Now go around the loop and try alloc again; 2699 // A competing allocation might beat us to the expansion space, 2700 // so we may go around the loop again if allocation fails again. 2701 if (GCExpandToAllocateDelayMillis > 0) { 2702 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2703 } 2704 } 2705 } 2706 2707 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2708 // Only shrink if a compaction was done so that all the free space 2709 // in the generation is in a contiguous block at the end. 2710 if (did_compact()) { 2711 CardGeneration::shrink(bytes); 2712 } 2713 } 2714 2715 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2716 assert_locked_or_safepoint(Heap_lock); 2717 } 2718 2719 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2720 assert_locked_or_safepoint(Heap_lock); 2721 assert_lock_strong(freelistLock()); 2722 log_trace(gc)("Shrinking of CMS not yet implemented"); 2723 return; 2724 } 2725 2726 2727 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2728 // phases. 2729 class CMSPhaseAccounting: public StackObj { 2730 public: 2731 CMSPhaseAccounting(CMSCollector *collector, 2732 const char *title); 2733 ~CMSPhaseAccounting(); 2734 2735 private: 2736 CMSCollector *_collector; 2737 const char *_title; 2738 GCTraceConcTime(Info, gc) _trace_time; 2739 2740 public: 2741 // Not MT-safe; so do not pass around these StackObj's 2742 // where they may be accessed by other threads. 2743 double wallclock_millis() { 2744 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2745 } 2746 }; 2747 2748 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2749 const char *title) : 2750 _collector(collector), _title(title), _trace_time(title) { 2751 2752 _collector->resetYields(); 2753 _collector->resetTimer(); 2754 _collector->startTimer(); 2755 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2756 } 2757 2758 CMSPhaseAccounting::~CMSPhaseAccounting() { 2759 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2760 _collector->stopTimer(); 2761 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2762 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2763 } 2764 2765 // CMS work 2766 2767 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2768 class CMSParMarkTask : public AbstractGangTask { 2769 protected: 2770 CMSCollector* _collector; 2771 uint _n_workers; 2772 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2773 AbstractGangTask(name), 2774 _collector(collector), 2775 _n_workers(n_workers) {} 2776 // Work method in support of parallel rescan ... of young gen spaces 2777 void do_young_space_rescan(OopsInGenClosure* cl, 2778 ContiguousSpace* space, 2779 HeapWord** chunk_array, size_t chunk_top); 2780 void work_on_young_gen_roots(OopsInGenClosure* cl); 2781 }; 2782 2783 // Parallel initial mark task 2784 class CMSParInitialMarkTask: public CMSParMarkTask { 2785 StrongRootsScope* _strong_roots_scope; 2786 public: 2787 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2788 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2789 _strong_roots_scope(strong_roots_scope) {} 2790 void work(uint worker_id); 2791 }; 2792 2793 // Checkpoint the roots into this generation from outside 2794 // this generation. [Note this initial checkpoint need only 2795 // be approximate -- we'll do a catch up phase subsequently.] 2796 void CMSCollector::checkpointRootsInitial() { 2797 assert(_collectorState == InitialMarking, "Wrong collector state"); 2798 check_correct_thread_executing(); 2799 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2800 2801 save_heap_summary(); 2802 report_heap_summary(GCWhen::BeforeGC); 2803 2804 ReferenceProcessor* rp = ref_processor(); 2805 assert(_restart_addr == NULL, "Control point invariant"); 2806 { 2807 // acquire locks for subsequent manipulations 2808 MutexLockerEx x(bitMapLock(), 2809 Mutex::_no_safepoint_check_flag); 2810 checkpointRootsInitialWork(); 2811 // enable ("weak") refs discovery 2812 rp->enable_discovery(); 2813 _collectorState = Marking; 2814 } 2815 } 2816 2817 void CMSCollector::checkpointRootsInitialWork() { 2818 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2819 assert(_collectorState == InitialMarking, "just checking"); 2820 2821 // Already have locks. 2822 assert_lock_strong(bitMapLock()); 2823 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2824 2825 // Setup the verification and class unloading state for this 2826 // CMS collection cycle. 2827 setup_cms_unloading_and_verification_state(); 2828 2829 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2830 2831 // Reset all the PLAB chunk arrays if necessary. 2832 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2833 reset_survivor_plab_arrays(); 2834 } 2835 2836 ResourceMark rm; 2837 HandleMark hm; 2838 2839 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2840 CMSHeap* heap = CMSHeap::heap(); 2841 2842 verify_work_stacks_empty(); 2843 verify_overflow_empty(); 2844 2845 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2846 // Update the saved marks which may affect the root scans. 2847 heap->save_marks(); 2848 2849 // weak reference processing has not started yet. 2850 ref_processor()->set_enqueuing_is_done(false); 2851 2852 // Need to remember all newly created CLDs, 2853 // so that we can guarantee that the remark finds them. 2854 ClassLoaderDataGraph::remember_new_clds(true); 2855 2856 // Whenever a CLD is found, it will be claimed before proceeding to mark 2857 // the klasses. The claimed marks need to be cleared before marking starts. 2858 ClassLoaderDataGraph::clear_claimed_marks(); 2859 2860 print_eden_and_survivor_chunk_arrays(); 2861 2862 { 2863 #if COMPILER2_OR_JVMCI 2864 DerivedPointerTableDeactivate dpt_deact; 2865 #endif 2866 if (CMSParallelInitialMarkEnabled) { 2867 // The parallel version. 2868 WorkGang* workers = heap->workers(); 2869 assert(workers != NULL, "Need parallel worker threads."); 2870 uint n_workers = workers->active_workers(); 2871 2872 StrongRootsScope srs(n_workers); 2873 2874 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2875 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2876 // If the total workers is greater than 1, then multiple workers 2877 // may be used at some time and the initialization has been set 2878 // such that the single threaded path cannot be used. 2879 if (workers->total_workers() > 1) { 2880 workers->run_task(&tsk); 2881 } else { 2882 tsk.work(0); 2883 } 2884 } else { 2885 // The serial version. 2886 CLDToOopClosure cld_closure(¬Older, true); 2887 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2888 2889 StrongRootsScope srs(1); 2890 2891 heap->cms_process_roots(&srs, 2892 true, // young gen as roots 2893 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2894 should_unload_classes(), 2895 ¬Older, 2896 &cld_closure); 2897 } 2898 } 2899 2900 // Clear mod-union table; it will be dirtied in the prologue of 2901 // CMS generation per each young generation collection. 2902 2903 assert(_modUnionTable.isAllClear(), 2904 "Was cleared in most recent final checkpoint phase" 2905 " or no bits are set in the gc_prologue before the start of the next " 2906 "subsequent marking phase."); 2907 2908 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2909 2910 // Save the end of the used_region of the constituent generations 2911 // to be used to limit the extent of sweep in each generation. 2912 save_sweep_limits(); 2913 verify_overflow_empty(); 2914 } 2915 2916 bool CMSCollector::markFromRoots() { 2917 // we might be tempted to assert that: 2918 // assert(!SafepointSynchronize::is_at_safepoint(), 2919 // "inconsistent argument?"); 2920 // However that wouldn't be right, because it's possible that 2921 // a safepoint is indeed in progress as a young generation 2922 // stop-the-world GC happens even as we mark in this generation. 2923 assert(_collectorState == Marking, "inconsistent state?"); 2924 check_correct_thread_executing(); 2925 verify_overflow_empty(); 2926 2927 // Weak ref discovery note: We may be discovering weak 2928 // refs in this generation concurrent (but interleaved) with 2929 // weak ref discovery by the young generation collector. 2930 2931 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2932 GCTraceCPUTime tcpu; 2933 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2934 bool res = markFromRootsWork(); 2935 if (res) { 2936 _collectorState = Precleaning; 2937 } else { // We failed and a foreground collection wants to take over 2938 assert(_foregroundGCIsActive, "internal state inconsistency"); 2939 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2940 log_debug(gc)("bailing out to foreground collection"); 2941 } 2942 verify_overflow_empty(); 2943 return res; 2944 } 2945 2946 bool CMSCollector::markFromRootsWork() { 2947 // iterate over marked bits in bit map, doing a full scan and mark 2948 // from these roots using the following algorithm: 2949 // . if oop is to the right of the current scan pointer, 2950 // mark corresponding bit (we'll process it later) 2951 // . else (oop is to left of current scan pointer) 2952 // push oop on marking stack 2953 // . drain the marking stack 2954 2955 // Note that when we do a marking step we need to hold the 2956 // bit map lock -- recall that direct allocation (by mutators) 2957 // and promotion (by the young generation collector) is also 2958 // marking the bit map. [the so-called allocate live policy.] 2959 // Because the implementation of bit map marking is not 2960 // robust wrt simultaneous marking of bits in the same word, 2961 // we need to make sure that there is no such interference 2962 // between concurrent such updates. 2963 2964 // already have locks 2965 assert_lock_strong(bitMapLock()); 2966 2967 verify_work_stacks_empty(); 2968 verify_overflow_empty(); 2969 bool result = false; 2970 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2971 result = do_marking_mt(); 2972 } else { 2973 result = do_marking_st(); 2974 } 2975 return result; 2976 } 2977 2978 // Forward decl 2979 class CMSConcMarkingTask; 2980 2981 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2982 CMSCollector* _collector; 2983 CMSConcMarkingTask* _task; 2984 public: 2985 virtual void yield(); 2986 2987 // "n_threads" is the number of threads to be terminated. 2988 // "queue_set" is a set of work queues of other threads. 2989 // "collector" is the CMS collector associated with this task terminator. 2990 // "yield" indicates whether we need the gang as a whole to yield. 2991 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2992 ParallelTaskTerminator(n_threads, queue_set), 2993 _collector(collector) { } 2994 2995 void set_task(CMSConcMarkingTask* task) { 2996 _task = task; 2997 } 2998 }; 2999 3000 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3001 CMSConcMarkingTask* _task; 3002 public: 3003 bool should_exit_termination(); 3004 void set_task(CMSConcMarkingTask* task) { 3005 _task = task; 3006 } 3007 }; 3008 3009 // MT Concurrent Marking Task 3010 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3011 CMSCollector* _collector; 3012 uint _n_workers; // requested/desired # workers 3013 bool _result; 3014 CompactibleFreeListSpace* _cms_space; 3015 char _pad_front[64]; // padding to ... 3016 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3017 char _pad_back[64]; 3018 HeapWord* _restart_addr; 3019 3020 // Exposed here for yielding support 3021 Mutex* const _bit_map_lock; 3022 3023 // The per thread work queues, available here for stealing 3024 OopTaskQueueSet* _task_queues; 3025 3026 // Termination (and yielding) support 3027 CMSConcMarkingTerminator _term; 3028 CMSConcMarkingTerminatorTerminator _term_term; 3029 3030 public: 3031 CMSConcMarkingTask(CMSCollector* collector, 3032 CompactibleFreeListSpace* cms_space, 3033 YieldingFlexibleWorkGang* workers, 3034 OopTaskQueueSet* task_queues): 3035 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3036 _collector(collector), 3037 _cms_space(cms_space), 3038 _n_workers(0), _result(true), 3039 _task_queues(task_queues), 3040 _term(_n_workers, task_queues, _collector), 3041 _bit_map_lock(collector->bitMapLock()) 3042 { 3043 _requested_size = _n_workers; 3044 _term.set_task(this); 3045 _term_term.set_task(this); 3046 _restart_addr = _global_finger = _cms_space->bottom(); 3047 } 3048 3049 3050 OopTaskQueueSet* task_queues() { return _task_queues; } 3051 3052 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3053 3054 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3055 3056 CMSConcMarkingTerminator* terminator() { return &_term; } 3057 3058 virtual void set_for_termination(uint active_workers) { 3059 terminator()->reset_for_reuse(active_workers); 3060 } 3061 3062 void work(uint worker_id); 3063 bool should_yield() { 3064 return ConcurrentMarkSweepThread::should_yield() 3065 && !_collector->foregroundGCIsActive(); 3066 } 3067 3068 virtual void coordinator_yield(); // stuff done by coordinator 3069 bool result() { return _result; } 3070 3071 void reset(HeapWord* ra) { 3072 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3073 _restart_addr = _global_finger = ra; 3074 _term.reset_for_reuse(); 3075 } 3076 3077 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3078 OopTaskQueue* work_q); 3079 3080 private: 3081 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3082 void do_work_steal(int i); 3083 void bump_global_finger(HeapWord* f); 3084 }; 3085 3086 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3087 assert(_task != NULL, "Error"); 3088 return _task->yielding(); 3089 // Note that we do not need the disjunct || _task->should_yield() above 3090 // because we want terminating threads to yield only if the task 3091 // is already in the midst of yielding, which happens only after at least one 3092 // thread has yielded. 3093 } 3094 3095 void CMSConcMarkingTerminator::yield() { 3096 if (_task->should_yield()) { 3097 _task->yield(); 3098 } else { 3099 ParallelTaskTerminator::yield(); 3100 } 3101 } 3102 3103 //////////////////////////////////////////////////////////////// 3104 // Concurrent Marking Algorithm Sketch 3105 //////////////////////////////////////////////////////////////// 3106 // Until all tasks exhausted (both spaces): 3107 // -- claim next available chunk 3108 // -- bump global finger via CAS 3109 // -- find first object that starts in this chunk 3110 // and start scanning bitmap from that position 3111 // -- scan marked objects for oops 3112 // -- CAS-mark target, and if successful: 3113 // . if target oop is above global finger (volatile read) 3114 // nothing to do 3115 // . if target oop is in chunk and above local finger 3116 // then nothing to do 3117 // . else push on work-queue 3118 // -- Deal with possible overflow issues: 3119 // . local work-queue overflow causes stuff to be pushed on 3120 // global (common) overflow queue 3121 // . always first empty local work queue 3122 // . then get a batch of oops from global work queue if any 3123 // . then do work stealing 3124 // -- When all tasks claimed (both spaces) 3125 // and local work queue empty, 3126 // then in a loop do: 3127 // . check global overflow stack; steal a batch of oops and trace 3128 // . try to steal from other threads oif GOS is empty 3129 // . if neither is available, offer termination 3130 // -- Terminate and return result 3131 // 3132 void CMSConcMarkingTask::work(uint worker_id) { 3133 elapsedTimer _timer; 3134 ResourceMark rm; 3135 HandleMark hm; 3136 3137 DEBUG_ONLY(_collector->verify_overflow_empty();) 3138 3139 // Before we begin work, our work queue should be empty 3140 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3141 // Scan the bitmap covering _cms_space, tracing through grey objects. 3142 _timer.start(); 3143 do_scan_and_mark(worker_id, _cms_space); 3144 _timer.stop(); 3145 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3146 3147 // ... do work stealing 3148 _timer.reset(); 3149 _timer.start(); 3150 do_work_steal(worker_id); 3151 _timer.stop(); 3152 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3153 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3154 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3155 // Note that under the current task protocol, the 3156 // following assertion is true even of the spaces 3157 // expanded since the completion of the concurrent 3158 // marking. XXX This will likely change under a strict 3159 // ABORT semantics. 3160 // After perm removal the comparison was changed to 3161 // greater than or equal to from strictly greater than. 3162 // Before perm removal the highest address sweep would 3163 // have been at the end of perm gen but now is at the 3164 // end of the tenured gen. 3165 assert(_global_finger >= _cms_space->end(), 3166 "All tasks have been completed"); 3167 DEBUG_ONLY(_collector->verify_overflow_empty();) 3168 } 3169 3170 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3171 HeapWord* read = _global_finger; 3172 HeapWord* cur = read; 3173 while (f > read) { 3174 cur = read; 3175 read = Atomic::cmpxchg(f, &_global_finger, cur); 3176 if (cur == read) { 3177 // our cas succeeded 3178 assert(_global_finger >= f, "protocol consistency"); 3179 break; 3180 } 3181 } 3182 } 3183 3184 // This is really inefficient, and should be redone by 3185 // using (not yet available) block-read and -write interfaces to the 3186 // stack and the work_queue. XXX FIX ME !!! 3187 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3188 OopTaskQueue* work_q) { 3189 // Fast lock-free check 3190 if (ovflw_stk->length() == 0) { 3191 return false; 3192 } 3193 assert(work_q->size() == 0, "Shouldn't steal"); 3194 MutexLockerEx ml(ovflw_stk->par_lock(), 3195 Mutex::_no_safepoint_check_flag); 3196 // Grab up to 1/4 the size of the work queue 3197 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3198 (size_t)ParGCDesiredObjsFromOverflowList); 3199 num = MIN2(num, ovflw_stk->length()); 3200 for (int i = (int) num; i > 0; i--) { 3201 oop cur = ovflw_stk->pop(); 3202 assert(cur != NULL, "Counted wrong?"); 3203 work_q->push(cur); 3204 } 3205 return num > 0; 3206 } 3207 3208 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3209 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3210 int n_tasks = pst->n_tasks(); 3211 // We allow that there may be no tasks to do here because 3212 // we are restarting after a stack overflow. 3213 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3214 uint nth_task = 0; 3215 3216 HeapWord* aligned_start = sp->bottom(); 3217 if (sp->used_region().contains(_restart_addr)) { 3218 // Align down to a card boundary for the start of 0th task 3219 // for this space. 3220 aligned_start = align_down(_restart_addr, CardTable::card_size); 3221 } 3222 3223 size_t chunk_size = sp->marking_task_size(); 3224 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3225 // Having claimed the nth task in this space, 3226 // compute the chunk that it corresponds to: 3227 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3228 aligned_start + (nth_task+1)*chunk_size); 3229 // Try and bump the global finger via a CAS; 3230 // note that we need to do the global finger bump 3231 // _before_ taking the intersection below, because 3232 // the task corresponding to that region will be 3233 // deemed done even if the used_region() expands 3234 // because of allocation -- as it almost certainly will 3235 // during start-up while the threads yield in the 3236 // closure below. 3237 HeapWord* finger = span.end(); 3238 bump_global_finger(finger); // atomically 3239 // There are null tasks here corresponding to chunks 3240 // beyond the "top" address of the space. 3241 span = span.intersection(sp->used_region()); 3242 if (!span.is_empty()) { // Non-null task 3243 HeapWord* prev_obj; 3244 assert(!span.contains(_restart_addr) || nth_task == 0, 3245 "Inconsistency"); 3246 if (nth_task == 0) { 3247 // For the 0th task, we'll not need to compute a block_start. 3248 if (span.contains(_restart_addr)) { 3249 // In the case of a restart because of stack overflow, 3250 // we might additionally skip a chunk prefix. 3251 prev_obj = _restart_addr; 3252 } else { 3253 prev_obj = span.start(); 3254 } 3255 } else { 3256 // We want to skip the first object because 3257 // the protocol is to scan any object in its entirety 3258 // that _starts_ in this span; a fortiori, any 3259 // object starting in an earlier span is scanned 3260 // as part of an earlier claimed task. 3261 // Below we use the "careful" version of block_start 3262 // so we do not try to navigate uninitialized objects. 3263 prev_obj = sp->block_start_careful(span.start()); 3264 // Below we use a variant of block_size that uses the 3265 // Printezis bits to avoid waiting for allocated 3266 // objects to become initialized/parsable. 3267 while (prev_obj < span.start()) { 3268 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3269 if (sz > 0) { 3270 prev_obj += sz; 3271 } else { 3272 // In this case we may end up doing a bit of redundant 3273 // scanning, but that appears unavoidable, short of 3274 // locking the free list locks; see bug 6324141. 3275 break; 3276 } 3277 } 3278 } 3279 if (prev_obj < span.end()) { 3280 MemRegion my_span = MemRegion(prev_obj, span.end()); 3281 // Do the marking work within a non-empty span -- 3282 // the last argument to the constructor indicates whether the 3283 // iteration should be incremental with periodic yields. 3284 ParMarkFromRootsClosure cl(this, _collector, my_span, 3285 &_collector->_markBitMap, 3286 work_queue(i), 3287 &_collector->_markStack); 3288 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3289 } // else nothing to do for this task 3290 } // else nothing to do for this task 3291 } 3292 // We'd be tempted to assert here that since there are no 3293 // more tasks left to claim in this space, the global_finger 3294 // must exceed space->top() and a fortiori space->end(). However, 3295 // that would not quite be correct because the bumping of 3296 // global_finger occurs strictly after the claiming of a task, 3297 // so by the time we reach here the global finger may not yet 3298 // have been bumped up by the thread that claimed the last 3299 // task. 3300 pst->all_tasks_completed(); 3301 } 3302 3303 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3304 private: 3305 CMSCollector* _collector; 3306 CMSConcMarkingTask* _task; 3307 MemRegion _span; 3308 CMSBitMap* _bit_map; 3309 CMSMarkStack* _overflow_stack; 3310 OopTaskQueue* _work_queue; 3311 protected: 3312 DO_OOP_WORK_DEFN 3313 public: 3314 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3315 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3316 MetadataAwareOopClosure(collector->ref_processor()), 3317 _collector(collector), 3318 _task(task), 3319 _span(collector->_span), 3320 _work_queue(work_queue), 3321 _bit_map(bit_map), 3322 _overflow_stack(overflow_stack) 3323 { } 3324 virtual void do_oop(oop* p); 3325 virtual void do_oop(narrowOop* p); 3326 3327 void trim_queue(size_t max); 3328 void handle_stack_overflow(HeapWord* lost); 3329 void do_yield_check() { 3330 if (_task->should_yield()) { 3331 _task->yield(); 3332 } 3333 } 3334 }; 3335 3336 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3337 3338 // Grey object scanning during work stealing phase -- 3339 // the salient assumption here is that any references 3340 // that are in these stolen objects being scanned must 3341 // already have been initialized (else they would not have 3342 // been published), so we do not need to check for 3343 // uninitialized objects before pushing here. 3344 void ParConcMarkingClosure::do_oop(oop obj) { 3345 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3346 HeapWord* addr = (HeapWord*)obj; 3347 // Check if oop points into the CMS generation 3348 // and is not marked 3349 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3350 // a white object ... 3351 // If we manage to "claim" the object, by being the 3352 // first thread to mark it, then we push it on our 3353 // marking stack 3354 if (_bit_map->par_mark(addr)) { // ... now grey 3355 // push on work queue (grey set) 3356 bool simulate_overflow = false; 3357 NOT_PRODUCT( 3358 if (CMSMarkStackOverflowALot && 3359 _collector->simulate_overflow()) { 3360 // simulate a stack overflow 3361 simulate_overflow = true; 3362 } 3363 ) 3364 if (simulate_overflow || 3365 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3366 // stack overflow 3367 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3368 // We cannot assert that the overflow stack is full because 3369 // it may have been emptied since. 3370 assert(simulate_overflow || 3371 _work_queue->size() == _work_queue->max_elems(), 3372 "Else push should have succeeded"); 3373 handle_stack_overflow(addr); 3374 } 3375 } // Else, some other thread got there first 3376 do_yield_check(); 3377 } 3378 } 3379 3380 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3381 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3382 3383 void ParConcMarkingClosure::trim_queue(size_t max) { 3384 while (_work_queue->size() > max) { 3385 oop new_oop; 3386 if (_work_queue->pop_local(new_oop)) { 3387 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3388 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3389 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3390 new_oop->oop_iterate(this); // do_oop() above 3391 do_yield_check(); 3392 } 3393 } 3394 } 3395 3396 // Upon stack overflow, we discard (part of) the stack, 3397 // remembering the least address amongst those discarded 3398 // in CMSCollector's _restart_address. 3399 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3400 // We need to do this under a mutex to prevent other 3401 // workers from interfering with the work done below. 3402 MutexLockerEx ml(_overflow_stack->par_lock(), 3403 Mutex::_no_safepoint_check_flag); 3404 // Remember the least grey address discarded 3405 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3406 _collector->lower_restart_addr(ra); 3407 _overflow_stack->reset(); // discard stack contents 3408 _overflow_stack->expand(); // expand the stack if possible 3409 } 3410 3411 3412 void CMSConcMarkingTask::do_work_steal(int i) { 3413 OopTaskQueue* work_q = work_queue(i); 3414 oop obj_to_scan; 3415 CMSBitMap* bm = &(_collector->_markBitMap); 3416 CMSMarkStack* ovflw = &(_collector->_markStack); 3417 int* seed = _collector->hash_seed(i); 3418 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3419 while (true) { 3420 cl.trim_queue(0); 3421 assert(work_q->size() == 0, "Should have been emptied above"); 3422 if (get_work_from_overflow_stack(ovflw, work_q)) { 3423 // Can't assert below because the work obtained from the 3424 // overflow stack may already have been stolen from us. 3425 // assert(work_q->size() > 0, "Work from overflow stack"); 3426 continue; 3427 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3428 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3429 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3430 obj_to_scan->oop_iterate(&cl); 3431 } else if (terminator()->offer_termination(&_term_term)) { 3432 assert(work_q->size() == 0, "Impossible!"); 3433 break; 3434 } else if (yielding() || should_yield()) { 3435 yield(); 3436 } 3437 } 3438 } 3439 3440 // This is run by the CMS (coordinator) thread. 3441 void CMSConcMarkingTask::coordinator_yield() { 3442 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3443 "CMS thread should hold CMS token"); 3444 // First give up the locks, then yield, then re-lock 3445 // We should probably use a constructor/destructor idiom to 3446 // do this unlock/lock or modify the MutexUnlocker class to 3447 // serve our purpose. XXX 3448 assert_lock_strong(_bit_map_lock); 3449 _bit_map_lock->unlock(); 3450 ConcurrentMarkSweepThread::desynchronize(true); 3451 _collector->stopTimer(); 3452 _collector->incrementYields(); 3453 3454 // It is possible for whichever thread initiated the yield request 3455 // not to get a chance to wake up and take the bitmap lock between 3456 // this thread releasing it and reacquiring it. So, while the 3457 // should_yield() flag is on, let's sleep for a bit to give the 3458 // other thread a chance to wake up. The limit imposed on the number 3459 // of iterations is defensive, to avoid any unforseen circumstances 3460 // putting us into an infinite loop. Since it's always been this 3461 // (coordinator_yield()) method that was observed to cause the 3462 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3463 // which is by default non-zero. For the other seven methods that 3464 // also perform the yield operation, as are using a different 3465 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3466 // can enable the sleeping for those methods too, if necessary. 3467 // See 6442774. 3468 // 3469 // We really need to reconsider the synchronization between the GC 3470 // thread and the yield-requesting threads in the future and we 3471 // should really use wait/notify, which is the recommended 3472 // way of doing this type of interaction. Additionally, we should 3473 // consolidate the eight methods that do the yield operation and they 3474 // are almost identical into one for better maintainability and 3475 // readability. See 6445193. 3476 // 3477 // Tony 2006.06.29 3478 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3479 ConcurrentMarkSweepThread::should_yield() && 3480 !CMSCollector::foregroundGCIsActive(); ++i) { 3481 os::sleep(Thread::current(), 1, false); 3482 } 3483 3484 ConcurrentMarkSweepThread::synchronize(true); 3485 _bit_map_lock->lock_without_safepoint_check(); 3486 _collector->startTimer(); 3487 } 3488 3489 bool CMSCollector::do_marking_mt() { 3490 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3491 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3492 conc_workers()->active_workers(), 3493 Threads::number_of_non_daemon_threads()); 3494 num_workers = conc_workers()->update_active_workers(num_workers); 3495 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3496 3497 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3498 3499 CMSConcMarkingTask tsk(this, 3500 cms_space, 3501 conc_workers(), 3502 task_queues()); 3503 3504 // Since the actual number of workers we get may be different 3505 // from the number we requested above, do we need to do anything different 3506 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3507 // class?? XXX 3508 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3509 3510 // Refs discovery is already non-atomic. 3511 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3512 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3513 conc_workers()->start_task(&tsk); 3514 while (tsk.yielded()) { 3515 tsk.coordinator_yield(); 3516 conc_workers()->continue_task(&tsk); 3517 } 3518 // If the task was aborted, _restart_addr will be non-NULL 3519 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3520 while (_restart_addr != NULL) { 3521 // XXX For now we do not make use of ABORTED state and have not 3522 // yet implemented the right abort semantics (even in the original 3523 // single-threaded CMS case). That needs some more investigation 3524 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3525 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3526 // If _restart_addr is non-NULL, a marking stack overflow 3527 // occurred; we need to do a fresh marking iteration from the 3528 // indicated restart address. 3529 if (_foregroundGCIsActive) { 3530 // We may be running into repeated stack overflows, having 3531 // reached the limit of the stack size, while making very 3532 // slow forward progress. It may be best to bail out and 3533 // let the foreground collector do its job. 3534 // Clear _restart_addr, so that foreground GC 3535 // works from scratch. This avoids the headache of 3536 // a "rescan" which would otherwise be needed because 3537 // of the dirty mod union table & card table. 3538 _restart_addr = NULL; 3539 return false; 3540 } 3541 // Adjust the task to restart from _restart_addr 3542 tsk.reset(_restart_addr); 3543 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3544 _restart_addr); 3545 _restart_addr = NULL; 3546 // Get the workers going again 3547 conc_workers()->start_task(&tsk); 3548 while (tsk.yielded()) { 3549 tsk.coordinator_yield(); 3550 conc_workers()->continue_task(&tsk); 3551 } 3552 } 3553 assert(tsk.completed(), "Inconsistency"); 3554 assert(tsk.result() == true, "Inconsistency"); 3555 return true; 3556 } 3557 3558 bool CMSCollector::do_marking_st() { 3559 ResourceMark rm; 3560 HandleMark hm; 3561 3562 // Temporarily make refs discovery single threaded (non-MT) 3563 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3564 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3565 &_markStack, CMSYield); 3566 // the last argument to iterate indicates whether the iteration 3567 // should be incremental with periodic yields. 3568 _markBitMap.iterate(&markFromRootsClosure); 3569 // If _restart_addr is non-NULL, a marking stack overflow 3570 // occurred; we need to do a fresh iteration from the 3571 // indicated restart address. 3572 while (_restart_addr != NULL) { 3573 if (_foregroundGCIsActive) { 3574 // We may be running into repeated stack overflows, having 3575 // reached the limit of the stack size, while making very 3576 // slow forward progress. It may be best to bail out and 3577 // let the foreground collector do its job. 3578 // Clear _restart_addr, so that foreground GC 3579 // works from scratch. This avoids the headache of 3580 // a "rescan" which would otherwise be needed because 3581 // of the dirty mod union table & card table. 3582 _restart_addr = NULL; 3583 return false; // indicating failure to complete marking 3584 } 3585 // Deal with stack overflow: 3586 // we restart marking from _restart_addr 3587 HeapWord* ra = _restart_addr; 3588 markFromRootsClosure.reset(ra); 3589 _restart_addr = NULL; 3590 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3591 } 3592 return true; 3593 } 3594 3595 void CMSCollector::preclean() { 3596 check_correct_thread_executing(); 3597 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3598 verify_work_stacks_empty(); 3599 verify_overflow_empty(); 3600 _abort_preclean = false; 3601 if (CMSPrecleaningEnabled) { 3602 if (!CMSEdenChunksRecordAlways) { 3603 _eden_chunk_index = 0; 3604 } 3605 size_t used = get_eden_used(); 3606 size_t capacity = get_eden_capacity(); 3607 // Don't start sampling unless we will get sufficiently 3608 // many samples. 3609 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3610 * CMSScheduleRemarkEdenPenetration)) { 3611 _start_sampling = true; 3612 } else { 3613 _start_sampling = false; 3614 } 3615 GCTraceCPUTime tcpu; 3616 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3617 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3618 } 3619 CMSTokenSync x(true); // is cms thread 3620 if (CMSPrecleaningEnabled) { 3621 sample_eden(); 3622 _collectorState = AbortablePreclean; 3623 } else { 3624 _collectorState = FinalMarking; 3625 } 3626 verify_work_stacks_empty(); 3627 verify_overflow_empty(); 3628 } 3629 3630 // Try and schedule the remark such that young gen 3631 // occupancy is CMSScheduleRemarkEdenPenetration %. 3632 void CMSCollector::abortable_preclean() { 3633 check_correct_thread_executing(); 3634 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3635 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3636 3637 // If Eden's current occupancy is below this threshold, 3638 // immediately schedule the remark; else preclean 3639 // past the next scavenge in an effort to 3640 // schedule the pause as described above. By choosing 3641 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3642 // we will never do an actual abortable preclean cycle. 3643 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3644 GCTraceCPUTime tcpu; 3645 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3646 // We need more smarts in the abortable preclean 3647 // loop below to deal with cases where allocation 3648 // in young gen is very very slow, and our precleaning 3649 // is running a losing race against a horde of 3650 // mutators intent on flooding us with CMS updates 3651 // (dirty cards). 3652 // One, admittedly dumb, strategy is to give up 3653 // after a certain number of abortable precleaning loops 3654 // or after a certain maximum time. We want to make 3655 // this smarter in the next iteration. 3656 // XXX FIX ME!!! YSR 3657 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3658 while (!(should_abort_preclean() || 3659 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3660 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3661 cumworkdone += workdone; 3662 loops++; 3663 // Voluntarily terminate abortable preclean phase if we have 3664 // been at it for too long. 3665 if ((CMSMaxAbortablePrecleanLoops != 0) && 3666 loops >= CMSMaxAbortablePrecleanLoops) { 3667 log_debug(gc)(" CMS: abort preclean due to loops "); 3668 break; 3669 } 3670 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3671 log_debug(gc)(" CMS: abort preclean due to time "); 3672 break; 3673 } 3674 // If we are doing little work each iteration, we should 3675 // take a short break. 3676 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3677 // Sleep for some time, waiting for work to accumulate 3678 stopTimer(); 3679 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3680 startTimer(); 3681 waited++; 3682 } 3683 } 3684 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3685 loops, waited, cumworkdone); 3686 } 3687 CMSTokenSync x(true); // is cms thread 3688 if (_collectorState != Idling) { 3689 assert(_collectorState == AbortablePreclean, 3690 "Spontaneous state transition?"); 3691 _collectorState = FinalMarking; 3692 } // Else, a foreground collection completed this CMS cycle. 3693 return; 3694 } 3695 3696 // Respond to an Eden sampling opportunity 3697 void CMSCollector::sample_eden() { 3698 // Make sure a young gc cannot sneak in between our 3699 // reading and recording of a sample. 3700 assert(Thread::current()->is_ConcurrentGC_thread(), 3701 "Only the cms thread may collect Eden samples"); 3702 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3703 "Should collect samples while holding CMS token"); 3704 if (!_start_sampling) { 3705 return; 3706 } 3707 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3708 // is populated by the young generation. 3709 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3710 if (_eden_chunk_index < _eden_chunk_capacity) { 3711 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3712 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3713 "Unexpected state of Eden"); 3714 // We'd like to check that what we just sampled is an oop-start address; 3715 // however, we cannot do that here since the object may not yet have been 3716 // initialized. So we'll instead do the check when we _use_ this sample 3717 // later. 3718 if (_eden_chunk_index == 0 || 3719 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3720 _eden_chunk_array[_eden_chunk_index-1]) 3721 >= CMSSamplingGrain)) { 3722 _eden_chunk_index++; // commit sample 3723 } 3724 } 3725 } 3726 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3727 size_t used = get_eden_used(); 3728 size_t capacity = get_eden_capacity(); 3729 assert(used <= capacity, "Unexpected state of Eden"); 3730 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3731 _abort_preclean = true; 3732 } 3733 } 3734 } 3735 3736 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3737 assert(_collectorState == Precleaning || 3738 _collectorState == AbortablePreclean, "incorrect state"); 3739 ResourceMark rm; 3740 HandleMark hm; 3741 3742 // Precleaning is currently not MT but the reference processor 3743 // may be set for MT. Disable it temporarily here. 3744 ReferenceProcessor* rp = ref_processor(); 3745 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3746 3747 // Do one pass of scrubbing the discovered reference lists 3748 // to remove any reference objects with strongly-reachable 3749 // referents. 3750 if (clean_refs) { 3751 CMSPrecleanRefsYieldClosure yield_cl(this); 3752 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3753 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3754 &_markStack, true /* preclean */); 3755 CMSDrainMarkingStackClosure complete_trace(this, 3756 _span, &_markBitMap, &_markStack, 3757 &keep_alive, true /* preclean */); 3758 3759 // We don't want this step to interfere with a young 3760 // collection because we don't want to take CPU 3761 // or memory bandwidth away from the young GC threads 3762 // (which may be as many as there are CPUs). 3763 // Note that we don't need to protect ourselves from 3764 // interference with mutators because they can't 3765 // manipulate the discovered reference lists nor affect 3766 // the computed reachability of the referents, the 3767 // only properties manipulated by the precleaning 3768 // of these reference lists. 3769 stopTimer(); 3770 CMSTokenSyncWithLocks x(true /* is cms thread */, 3771 bitMapLock()); 3772 startTimer(); 3773 sample_eden(); 3774 3775 // The following will yield to allow foreground 3776 // collection to proceed promptly. XXX YSR: 3777 // The code in this method may need further 3778 // tweaking for better performance and some restructuring 3779 // for cleaner interfaces. 3780 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3781 rp->preclean_discovered_references( 3782 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3783 gc_timer); 3784 } 3785 3786 if (clean_survivor) { // preclean the active survivor space(s) 3787 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3788 &_markBitMap, &_modUnionTable, 3789 &_markStack, true /* precleaning phase */); 3790 stopTimer(); 3791 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3792 bitMapLock()); 3793 startTimer(); 3794 unsigned int before_count = 3795 CMSHeap::heap()->total_collections(); 3796 SurvivorSpacePrecleanClosure 3797 sss_cl(this, _span, &_markBitMap, &_markStack, 3798 &pam_cl, before_count, CMSYield); 3799 _young_gen->from()->object_iterate_careful(&sss_cl); 3800 _young_gen->to()->object_iterate_careful(&sss_cl); 3801 } 3802 MarkRefsIntoAndScanClosure 3803 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3804 &_markStack, this, CMSYield, 3805 true /* precleaning phase */); 3806 // CAUTION: The following closure has persistent state that may need to 3807 // be reset upon a decrease in the sequence of addresses it 3808 // processes. 3809 ScanMarkedObjectsAgainCarefullyClosure 3810 smoac_cl(this, _span, 3811 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3812 3813 // Preclean dirty cards in ModUnionTable and CardTable using 3814 // appropriate convergence criterion; 3815 // repeat CMSPrecleanIter times unless we find that 3816 // we are losing. 3817 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3818 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3819 "Bad convergence multiplier"); 3820 assert(CMSPrecleanThreshold >= 100, 3821 "Unreasonably low CMSPrecleanThreshold"); 3822 3823 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3824 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3825 numIter < CMSPrecleanIter; 3826 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3827 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3828 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3829 // Either there are very few dirty cards, so re-mark 3830 // pause will be small anyway, or our pre-cleaning isn't 3831 // that much faster than the rate at which cards are being 3832 // dirtied, so we might as well stop and re-mark since 3833 // precleaning won't improve our re-mark time by much. 3834 if (curNumCards <= CMSPrecleanThreshold || 3835 (numIter > 0 && 3836 (curNumCards * CMSPrecleanDenominator > 3837 lastNumCards * CMSPrecleanNumerator))) { 3838 numIter++; 3839 cumNumCards += curNumCards; 3840 break; 3841 } 3842 } 3843 3844 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3845 3846 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3847 cumNumCards += curNumCards; 3848 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3849 curNumCards, cumNumCards, numIter); 3850 return cumNumCards; // as a measure of useful work done 3851 } 3852 3853 // PRECLEANING NOTES: 3854 // Precleaning involves: 3855 // . reading the bits of the modUnionTable and clearing the set bits. 3856 // . For the cards corresponding to the set bits, we scan the 3857 // objects on those cards. This means we need the free_list_lock 3858 // so that we can safely iterate over the CMS space when scanning 3859 // for oops. 3860 // . When we scan the objects, we'll be both reading and setting 3861 // marks in the marking bit map, so we'll need the marking bit map. 3862 // . For protecting _collector_state transitions, we take the CGC_lock. 3863 // Note that any races in the reading of of card table entries by the 3864 // CMS thread on the one hand and the clearing of those entries by the 3865 // VM thread or the setting of those entries by the mutator threads on the 3866 // other are quite benign. However, for efficiency it makes sense to keep 3867 // the VM thread from racing with the CMS thread while the latter is 3868 // dirty card info to the modUnionTable. We therefore also use the 3869 // CGC_lock to protect the reading of the card table and the mod union 3870 // table by the CM thread. 3871 // . We run concurrently with mutator updates, so scanning 3872 // needs to be done carefully -- we should not try to scan 3873 // potentially uninitialized objects. 3874 // 3875 // Locking strategy: While holding the CGC_lock, we scan over and 3876 // reset a maximal dirty range of the mod union / card tables, then lock 3877 // the free_list_lock and bitmap lock to do a full marking, then 3878 // release these locks; and repeat the cycle. This allows for a 3879 // certain amount of fairness in the sharing of these locks between 3880 // the CMS collector on the one hand, and the VM thread and the 3881 // mutators on the other. 3882 3883 // NOTE: preclean_mod_union_table() and preclean_card_table() 3884 // further below are largely identical; if you need to modify 3885 // one of these methods, please check the other method too. 3886 3887 size_t CMSCollector::preclean_mod_union_table( 3888 ConcurrentMarkSweepGeneration* old_gen, 3889 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3890 verify_work_stacks_empty(); 3891 verify_overflow_empty(); 3892 3893 // strategy: starting with the first card, accumulate contiguous 3894 // ranges of dirty cards; clear these cards, then scan the region 3895 // covered by these cards. 3896 3897 // Since all of the MUT is committed ahead, we can just use 3898 // that, in case the generations expand while we are precleaning. 3899 // It might also be fine to just use the committed part of the 3900 // generation, but we might potentially miss cards when the 3901 // generation is rapidly expanding while we are in the midst 3902 // of precleaning. 3903 HeapWord* startAddr = old_gen->reserved().start(); 3904 HeapWord* endAddr = old_gen->reserved().end(); 3905 3906 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3907 3908 size_t numDirtyCards, cumNumDirtyCards; 3909 HeapWord *nextAddr, *lastAddr; 3910 for (cumNumDirtyCards = numDirtyCards = 0, 3911 nextAddr = lastAddr = startAddr; 3912 nextAddr < endAddr; 3913 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3914 3915 ResourceMark rm; 3916 HandleMark hm; 3917 3918 MemRegion dirtyRegion; 3919 { 3920 stopTimer(); 3921 // Potential yield point 3922 CMSTokenSync ts(true); 3923 startTimer(); 3924 sample_eden(); 3925 // Get dirty region starting at nextOffset (inclusive), 3926 // simultaneously clearing it. 3927 dirtyRegion = 3928 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3929 assert(dirtyRegion.start() >= nextAddr, 3930 "returned region inconsistent?"); 3931 } 3932 // Remember where the next search should begin. 3933 // The returned region (if non-empty) is a right open interval, 3934 // so lastOffset is obtained from the right end of that 3935 // interval. 3936 lastAddr = dirtyRegion.end(); 3937 // Should do something more transparent and less hacky XXX 3938 numDirtyCards = 3939 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3940 3941 // We'll scan the cards in the dirty region (with periodic 3942 // yields for foreground GC as needed). 3943 if (!dirtyRegion.is_empty()) { 3944 assert(numDirtyCards > 0, "consistency check"); 3945 HeapWord* stop_point = NULL; 3946 stopTimer(); 3947 // Potential yield point 3948 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3949 bitMapLock()); 3950 startTimer(); 3951 { 3952 verify_work_stacks_empty(); 3953 verify_overflow_empty(); 3954 sample_eden(); 3955 stop_point = 3956 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3957 } 3958 if (stop_point != NULL) { 3959 // The careful iteration stopped early either because it found an 3960 // uninitialized object, or because we were in the midst of an 3961 // "abortable preclean", which should now be aborted. Redirty 3962 // the bits corresponding to the partially-scanned or unscanned 3963 // cards. We'll either restart at the next block boundary or 3964 // abort the preclean. 3965 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3966 "Should only be AbortablePreclean."); 3967 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3968 if (should_abort_preclean()) { 3969 break; // out of preclean loop 3970 } else { 3971 // Compute the next address at which preclean should pick up; 3972 // might need bitMapLock in order to read P-bits. 3973 lastAddr = next_card_start_after_block(stop_point); 3974 } 3975 } 3976 } else { 3977 assert(lastAddr == endAddr, "consistency check"); 3978 assert(numDirtyCards == 0, "consistency check"); 3979 break; 3980 } 3981 } 3982 verify_work_stacks_empty(); 3983 verify_overflow_empty(); 3984 return cumNumDirtyCards; 3985 } 3986 3987 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3988 // below are largely identical; if you need to modify 3989 // one of these methods, please check the other method too. 3990 3991 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3992 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3993 // strategy: it's similar to precleamModUnionTable above, in that 3994 // we accumulate contiguous ranges of dirty cards, mark these cards 3995 // precleaned, then scan the region covered by these cards. 3996 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3997 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3998 3999 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4000 4001 size_t numDirtyCards, cumNumDirtyCards; 4002 HeapWord *lastAddr, *nextAddr; 4003 4004 for (cumNumDirtyCards = numDirtyCards = 0, 4005 nextAddr = lastAddr = startAddr; 4006 nextAddr < endAddr; 4007 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4008 4009 ResourceMark rm; 4010 HandleMark hm; 4011 4012 MemRegion dirtyRegion; 4013 { 4014 // See comments in "Precleaning notes" above on why we 4015 // do this locking. XXX Could the locking overheads be 4016 // too high when dirty cards are sparse? [I don't think so.] 4017 stopTimer(); 4018 CMSTokenSync x(true); // is cms thread 4019 startTimer(); 4020 sample_eden(); 4021 // Get and clear dirty region from card table 4022 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4023 true, 4024 CardTable::precleaned_card_val()); 4025 4026 assert(dirtyRegion.start() >= nextAddr, 4027 "returned region inconsistent?"); 4028 } 4029 lastAddr = dirtyRegion.end(); 4030 numDirtyCards = 4031 dirtyRegion.word_size()/CardTable::card_size_in_words; 4032 4033 if (!dirtyRegion.is_empty()) { 4034 stopTimer(); 4035 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4036 startTimer(); 4037 sample_eden(); 4038 verify_work_stacks_empty(); 4039 verify_overflow_empty(); 4040 HeapWord* stop_point = 4041 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4042 if (stop_point != NULL) { 4043 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4044 "Should only be AbortablePreclean."); 4045 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4046 if (should_abort_preclean()) { 4047 break; // out of preclean loop 4048 } else { 4049 // Compute the next address at which preclean should pick up. 4050 lastAddr = next_card_start_after_block(stop_point); 4051 } 4052 } 4053 } else { 4054 break; 4055 } 4056 } 4057 verify_work_stacks_empty(); 4058 verify_overflow_empty(); 4059 return cumNumDirtyCards; 4060 } 4061 4062 class PrecleanCLDClosure : public CLDClosure { 4063 MetadataAwareOopsInGenClosure* _cm_closure; 4064 public: 4065 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4066 void do_cld(ClassLoaderData* cld) { 4067 if (cld->has_accumulated_modified_oops()) { 4068 cld->clear_accumulated_modified_oops(); 4069 4070 _cm_closure->do_cld(cld); 4071 } 4072 } 4073 }; 4074 4075 // The freelist lock is needed to prevent asserts, is it really needed? 4076 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4077 4078 cl->set_freelistLock(freelistLock); 4079 4080 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4081 4082 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4083 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4084 PrecleanCLDClosure preclean_closure(cl); 4085 ClassLoaderDataGraph::cld_do(&preclean_closure); 4086 4087 verify_work_stacks_empty(); 4088 verify_overflow_empty(); 4089 } 4090 4091 void CMSCollector::checkpointRootsFinal() { 4092 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4093 check_correct_thread_executing(); 4094 // world is stopped at this checkpoint 4095 assert(SafepointSynchronize::is_at_safepoint(), 4096 "world should be stopped"); 4097 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4098 4099 verify_work_stacks_empty(); 4100 verify_overflow_empty(); 4101 4102 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4103 _young_gen->used() / K, _young_gen->capacity() / K); 4104 { 4105 if (CMSScavengeBeforeRemark) { 4106 CMSHeap* heap = CMSHeap::heap(); 4107 // Temporarily set flag to false, GCH->do_collection will 4108 // expect it to be false and set to true 4109 FlagSetting fl(heap->_is_gc_active, false); 4110 4111 heap->do_collection(true, // full (i.e. force, see below) 4112 false, // !clear_all_soft_refs 4113 0, // size 4114 false, // is_tlab 4115 GenCollectedHeap::YoungGen // type 4116 ); 4117 } 4118 FreelistLocker x(this); 4119 MutexLockerEx y(bitMapLock(), 4120 Mutex::_no_safepoint_check_flag); 4121 checkpointRootsFinalWork(); 4122 } 4123 verify_work_stacks_empty(); 4124 verify_overflow_empty(); 4125 } 4126 4127 void CMSCollector::checkpointRootsFinalWork() { 4128 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4129 4130 assert(haveFreelistLocks(), "must have free list locks"); 4131 assert_lock_strong(bitMapLock()); 4132 4133 ResourceMark rm; 4134 HandleMark hm; 4135 4136 CMSHeap* heap = CMSHeap::heap(); 4137 4138 if (should_unload_classes()) { 4139 CodeCache::gc_prologue(); 4140 } 4141 assert(haveFreelistLocks(), "must have free list locks"); 4142 assert_lock_strong(bitMapLock()); 4143 4144 // We might assume that we need not fill TLAB's when 4145 // CMSScavengeBeforeRemark is set, because we may have just done 4146 // a scavenge which would have filled all TLAB's -- and besides 4147 // Eden would be empty. This however may not always be the case -- 4148 // for instance although we asked for a scavenge, it may not have 4149 // happened because of a JNI critical section. We probably need 4150 // a policy for deciding whether we can in that case wait until 4151 // the critical section releases and then do the remark following 4152 // the scavenge, and skip it here. In the absence of that policy, 4153 // or of an indication of whether the scavenge did indeed occur, 4154 // we cannot rely on TLAB's having been filled and must do 4155 // so here just in case a scavenge did not happen. 4156 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4157 // Update the saved marks which may affect the root scans. 4158 heap->save_marks(); 4159 4160 print_eden_and_survivor_chunk_arrays(); 4161 4162 { 4163 #if COMPILER2_OR_JVMCI 4164 DerivedPointerTableDeactivate dpt_deact; 4165 #endif 4166 4167 // Note on the role of the mod union table: 4168 // Since the marker in "markFromRoots" marks concurrently with 4169 // mutators, it is possible for some reachable objects not to have been 4170 // scanned. For instance, an only reference to an object A was 4171 // placed in object B after the marker scanned B. Unless B is rescanned, 4172 // A would be collected. Such updates to references in marked objects 4173 // are detected via the mod union table which is the set of all cards 4174 // dirtied since the first checkpoint in this GC cycle and prior to 4175 // the most recent young generation GC, minus those cleaned up by the 4176 // concurrent precleaning. 4177 if (CMSParallelRemarkEnabled) { 4178 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4179 do_remark_parallel(); 4180 } else { 4181 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4182 do_remark_non_parallel(); 4183 } 4184 } 4185 verify_work_stacks_empty(); 4186 verify_overflow_empty(); 4187 4188 { 4189 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4190 refProcessingWork(); 4191 } 4192 verify_work_stacks_empty(); 4193 verify_overflow_empty(); 4194 4195 if (should_unload_classes()) { 4196 CodeCache::gc_epilogue(); 4197 } 4198 JvmtiExport::gc_epilogue(); 4199 4200 // If we encountered any (marking stack / work queue) overflow 4201 // events during the current CMS cycle, take appropriate 4202 // remedial measures, where possible, so as to try and avoid 4203 // recurrence of that condition. 4204 assert(_markStack.isEmpty(), "No grey objects"); 4205 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4206 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4207 if (ser_ovflw > 0) { 4208 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4209 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4210 _markStack.expand(); 4211 _ser_pmc_remark_ovflw = 0; 4212 _ser_pmc_preclean_ovflw = 0; 4213 _ser_kac_preclean_ovflw = 0; 4214 _ser_kac_ovflw = 0; 4215 } 4216 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4217 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4218 _par_pmc_remark_ovflw, _par_kac_ovflw); 4219 _par_pmc_remark_ovflw = 0; 4220 _par_kac_ovflw = 0; 4221 } 4222 if (_markStack._hit_limit > 0) { 4223 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4224 _markStack._hit_limit); 4225 } 4226 if (_markStack._failed_double > 0) { 4227 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4228 _markStack._failed_double, _markStack.capacity()); 4229 } 4230 _markStack._hit_limit = 0; 4231 _markStack._failed_double = 0; 4232 4233 if ((VerifyAfterGC || VerifyDuringGC) && 4234 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4235 verify_after_remark(); 4236 } 4237 4238 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4239 4240 // Change under the freelistLocks. 4241 _collectorState = Sweeping; 4242 // Call isAllClear() under bitMapLock 4243 assert(_modUnionTable.isAllClear(), 4244 "Should be clear by end of the final marking"); 4245 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4246 "Should be clear by end of the final marking"); 4247 } 4248 4249 void CMSParInitialMarkTask::work(uint worker_id) { 4250 elapsedTimer _timer; 4251 ResourceMark rm; 4252 HandleMark hm; 4253 4254 // ---------- scan from roots -------------- 4255 _timer.start(); 4256 CMSHeap* heap = CMSHeap::heap(); 4257 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4258 4259 // ---------- young gen roots -------------- 4260 { 4261 work_on_young_gen_roots(&par_mri_cl); 4262 _timer.stop(); 4263 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4264 } 4265 4266 // ---------- remaining roots -------------- 4267 _timer.reset(); 4268 _timer.start(); 4269 4270 CLDToOopClosure cld_closure(&par_mri_cl, true); 4271 4272 heap->cms_process_roots(_strong_roots_scope, 4273 false, // yg was scanned above 4274 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4275 _collector->should_unload_classes(), 4276 &par_mri_cl, 4277 &cld_closure); 4278 assert(_collector->should_unload_classes() 4279 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4280 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4281 _timer.stop(); 4282 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4283 } 4284 4285 // Parallel remark task 4286 class CMSParRemarkTask: public CMSParMarkTask { 4287 CompactibleFreeListSpace* _cms_space; 4288 4289 // The per-thread work queues, available here for stealing. 4290 OopTaskQueueSet* _task_queues; 4291 ParallelTaskTerminator _term; 4292 StrongRootsScope* _strong_roots_scope; 4293 4294 public: 4295 // A value of 0 passed to n_workers will cause the number of 4296 // workers to be taken from the active workers in the work gang. 4297 CMSParRemarkTask(CMSCollector* collector, 4298 CompactibleFreeListSpace* cms_space, 4299 uint n_workers, WorkGang* workers, 4300 OopTaskQueueSet* task_queues, 4301 StrongRootsScope* strong_roots_scope): 4302 CMSParMarkTask("Rescan roots and grey objects in parallel", 4303 collector, n_workers), 4304 _cms_space(cms_space), 4305 _task_queues(task_queues), 4306 _term(n_workers, task_queues), 4307 _strong_roots_scope(strong_roots_scope) { } 4308 4309 OopTaskQueueSet* task_queues() { return _task_queues; } 4310 4311 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4312 4313 ParallelTaskTerminator* terminator() { return &_term; } 4314 uint n_workers() { return _n_workers; } 4315 4316 void work(uint worker_id); 4317 4318 private: 4319 // ... of dirty cards in old space 4320 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4321 ParMarkRefsIntoAndScanClosure* cl); 4322 4323 // ... work stealing for the above 4324 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4325 }; 4326 4327 class RemarkCLDClosure : public CLDClosure { 4328 CLDToOopClosure _cm_closure; 4329 public: 4330 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4331 void do_cld(ClassLoaderData* cld) { 4332 // Check if we have modified any oops in the CLD during the concurrent marking. 4333 if (cld->has_accumulated_modified_oops()) { 4334 cld->clear_accumulated_modified_oops(); 4335 4336 // We could have transfered the current modified marks to the accumulated marks, 4337 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4338 } else if (cld->has_modified_oops()) { 4339 // Don't clear anything, this info is needed by the next young collection. 4340 } else { 4341 // No modified oops in the ClassLoaderData. 4342 return; 4343 } 4344 4345 // The klass has modified fields, need to scan the klass. 4346 _cm_closure.do_cld(cld); 4347 } 4348 }; 4349 4350 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4351 ParNewGeneration* young_gen = _collector->_young_gen; 4352 ContiguousSpace* eden_space = young_gen->eden(); 4353 ContiguousSpace* from_space = young_gen->from(); 4354 ContiguousSpace* to_space = young_gen->to(); 4355 4356 HeapWord** eca = _collector->_eden_chunk_array; 4357 size_t ect = _collector->_eden_chunk_index; 4358 HeapWord** sca = _collector->_survivor_chunk_array; 4359 size_t sct = _collector->_survivor_chunk_index; 4360 4361 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4362 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4363 4364 do_young_space_rescan(cl, to_space, NULL, 0); 4365 do_young_space_rescan(cl, from_space, sca, sct); 4366 do_young_space_rescan(cl, eden_space, eca, ect); 4367 } 4368 4369 // work_queue(i) is passed to the closure 4370 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4371 // also is passed to do_dirty_card_rescan_tasks() and to 4372 // do_work_steal() to select the i-th task_queue. 4373 4374 void CMSParRemarkTask::work(uint worker_id) { 4375 elapsedTimer _timer; 4376 ResourceMark rm; 4377 HandleMark hm; 4378 4379 // ---------- rescan from roots -------------- 4380 _timer.start(); 4381 CMSHeap* heap = CMSHeap::heap(); 4382 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4383 _collector->_span, _collector->ref_processor(), 4384 &(_collector->_markBitMap), 4385 work_queue(worker_id)); 4386 4387 // Rescan young gen roots first since these are likely 4388 // coarsely partitioned and may, on that account, constitute 4389 // the critical path; thus, it's best to start off that 4390 // work first. 4391 // ---------- young gen roots -------------- 4392 { 4393 work_on_young_gen_roots(&par_mrias_cl); 4394 _timer.stop(); 4395 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4396 } 4397 4398 // ---------- remaining roots -------------- 4399 _timer.reset(); 4400 _timer.start(); 4401 heap->cms_process_roots(_strong_roots_scope, 4402 false, // yg was scanned above 4403 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4404 _collector->should_unload_classes(), 4405 &par_mrias_cl, 4406 NULL); // The dirty klasses will be handled below 4407 4408 assert(_collector->should_unload_classes() 4409 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4410 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4411 _timer.stop(); 4412 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4413 4414 // ---------- unhandled CLD scanning ---------- 4415 if (worker_id == 0) { // Single threaded at the moment. 4416 _timer.reset(); 4417 _timer.start(); 4418 4419 // Scan all new class loader data objects and new dependencies that were 4420 // introduced during concurrent marking. 4421 ResourceMark rm; 4422 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4423 for (int i = 0; i < array->length(); i++) { 4424 par_mrias_cl.do_cld_nv(array->at(i)); 4425 } 4426 4427 // We don't need to keep track of new CLDs anymore. 4428 ClassLoaderDataGraph::remember_new_clds(false); 4429 4430 _timer.stop(); 4431 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4432 } 4433 4434 // We might have added oops to ClassLoaderData::_handles during the 4435 // concurrent marking phase. These oops do not always point to newly allocated objects 4436 // that are guaranteed to be kept alive. Hence, 4437 // we do have to revisit the _handles block during the remark phase. 4438 4439 // ---------- dirty CLD scanning ---------- 4440 if (worker_id == 0) { // Single threaded at the moment. 4441 _timer.reset(); 4442 _timer.start(); 4443 4444 // Scan all classes that was dirtied during the concurrent marking phase. 4445 RemarkCLDClosure remark_closure(&par_mrias_cl); 4446 ClassLoaderDataGraph::cld_do(&remark_closure); 4447 4448 _timer.stop(); 4449 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4450 } 4451 4452 4453 // ---------- rescan dirty cards ------------ 4454 _timer.reset(); 4455 _timer.start(); 4456 4457 // Do the rescan tasks for each of the two spaces 4458 // (cms_space) in turn. 4459 // "worker_id" is passed to select the task_queue for "worker_id" 4460 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4461 _timer.stop(); 4462 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4463 4464 // ---------- steal work from other threads ... 4465 // ---------- ... and drain overflow list. 4466 _timer.reset(); 4467 _timer.start(); 4468 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4469 _timer.stop(); 4470 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4471 } 4472 4473 void 4474 CMSParMarkTask::do_young_space_rescan( 4475 OopsInGenClosure* cl, ContiguousSpace* space, 4476 HeapWord** chunk_array, size_t chunk_top) { 4477 // Until all tasks completed: 4478 // . claim an unclaimed task 4479 // . compute region boundaries corresponding to task claimed 4480 // using chunk_array 4481 // . par_oop_iterate(cl) over that region 4482 4483 ResourceMark rm; 4484 HandleMark hm; 4485 4486 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4487 4488 uint nth_task = 0; 4489 uint n_tasks = pst->n_tasks(); 4490 4491 if (n_tasks > 0) { 4492 assert(pst->valid(), "Uninitialized use?"); 4493 HeapWord *start, *end; 4494 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4495 // We claimed task # nth_task; compute its boundaries. 4496 if (chunk_top == 0) { // no samples were taken 4497 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4498 start = space->bottom(); 4499 end = space->top(); 4500 } else if (nth_task == 0) { 4501 start = space->bottom(); 4502 end = chunk_array[nth_task]; 4503 } else if (nth_task < (uint)chunk_top) { 4504 assert(nth_task >= 1, "Control point invariant"); 4505 start = chunk_array[nth_task - 1]; 4506 end = chunk_array[nth_task]; 4507 } else { 4508 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4509 start = chunk_array[chunk_top - 1]; 4510 end = space->top(); 4511 } 4512 MemRegion mr(start, end); 4513 // Verify that mr is in space 4514 assert(mr.is_empty() || space->used_region().contains(mr), 4515 "Should be in space"); 4516 // Verify that "start" is an object boundary 4517 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4518 "Should be an oop"); 4519 space->par_oop_iterate(mr, cl); 4520 } 4521 pst->all_tasks_completed(); 4522 } 4523 } 4524 4525 void 4526 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4527 CompactibleFreeListSpace* sp, int i, 4528 ParMarkRefsIntoAndScanClosure* cl) { 4529 // Until all tasks completed: 4530 // . claim an unclaimed task 4531 // . compute region boundaries corresponding to task claimed 4532 // . transfer dirty bits ct->mut for that region 4533 // . apply rescanclosure to dirty mut bits for that region 4534 4535 ResourceMark rm; 4536 HandleMark hm; 4537 4538 OopTaskQueue* work_q = work_queue(i); 4539 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4540 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4541 // CAUTION: This closure has state that persists across calls to 4542 // the work method dirty_range_iterate_clear() in that it has 4543 // embedded in it a (subtype of) UpwardsObjectClosure. The 4544 // use of that state in the embedded UpwardsObjectClosure instance 4545 // assumes that the cards are always iterated (even if in parallel 4546 // by several threads) in monotonically increasing order per each 4547 // thread. This is true of the implementation below which picks 4548 // card ranges (chunks) in monotonically increasing order globally 4549 // and, a-fortiori, in monotonically increasing order per thread 4550 // (the latter order being a subsequence of the former). 4551 // If the work code below is ever reorganized into a more chaotic 4552 // work-partitioning form than the current "sequential tasks" 4553 // paradigm, the use of that persistent state will have to be 4554 // revisited and modified appropriately. See also related 4555 // bug 4756801 work on which should examine this code to make 4556 // sure that the changes there do not run counter to the 4557 // assumptions made here and necessary for correctness and 4558 // efficiency. Note also that this code might yield inefficient 4559 // behavior in the case of very large objects that span one or 4560 // more work chunks. Such objects would potentially be scanned 4561 // several times redundantly. Work on 4756801 should try and 4562 // address that performance anomaly if at all possible. XXX 4563 MemRegion full_span = _collector->_span; 4564 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4565 MarkFromDirtyCardsClosure 4566 greyRescanClosure(_collector, full_span, // entire span of interest 4567 sp, bm, work_q, cl); 4568 4569 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4570 assert(pst->valid(), "Uninitialized use?"); 4571 uint nth_task = 0; 4572 const int alignment = CardTable::card_size * BitsPerWord; 4573 MemRegion span = sp->used_region(); 4574 HeapWord* start_addr = span.start(); 4575 HeapWord* end_addr = align_up(span.end(), alignment); 4576 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4577 assert(is_aligned(start_addr, alignment), "Check alignment"); 4578 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4579 4580 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4581 // Having claimed the nth_task, compute corresponding mem-region, 4582 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4583 // The alignment restriction ensures that we do not need any 4584 // synchronization with other gang-workers while setting or 4585 // clearing bits in thus chunk of the MUT. 4586 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4587 start_addr + (nth_task+1)*chunk_size); 4588 // The last chunk's end might be way beyond end of the 4589 // used region. In that case pull back appropriately. 4590 if (this_span.end() > end_addr) { 4591 this_span.set_end(end_addr); 4592 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4593 } 4594 // Iterate over the dirty cards covering this chunk, marking them 4595 // precleaned, and setting the corresponding bits in the mod union 4596 // table. Since we have been careful to partition at Card and MUT-word 4597 // boundaries no synchronization is needed between parallel threads. 4598 _collector->_ct->dirty_card_iterate(this_span, 4599 &modUnionClosure); 4600 4601 // Having transferred these marks into the modUnionTable, 4602 // rescan the marked objects on the dirty cards in the modUnionTable. 4603 // Even if this is at a synchronous collection, the initial marking 4604 // may have been done during an asynchronous collection so there 4605 // may be dirty bits in the mod-union table. 4606 _collector->_modUnionTable.dirty_range_iterate_clear( 4607 this_span, &greyRescanClosure); 4608 _collector->_modUnionTable.verifyNoOneBitsInRange( 4609 this_span.start(), 4610 this_span.end()); 4611 } 4612 pst->all_tasks_completed(); // declare that i am done 4613 } 4614 4615 // . see if we can share work_queues with ParNew? XXX 4616 void 4617 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4618 int* seed) { 4619 OopTaskQueue* work_q = work_queue(i); 4620 NOT_PRODUCT(int num_steals = 0;) 4621 oop obj_to_scan; 4622 CMSBitMap* bm = &(_collector->_markBitMap); 4623 4624 while (true) { 4625 // Completely finish any left over work from (an) earlier round(s) 4626 cl->trim_queue(0); 4627 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4628 (size_t)ParGCDesiredObjsFromOverflowList); 4629 // Now check if there's any work in the overflow list 4630 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4631 // only affects the number of attempts made to get work from the 4632 // overflow list and does not affect the number of workers. Just 4633 // pass ParallelGCThreads so this behavior is unchanged. 4634 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4635 work_q, 4636 ParallelGCThreads)) { 4637 // found something in global overflow list; 4638 // not yet ready to go stealing work from others. 4639 // We'd like to assert(work_q->size() != 0, ...) 4640 // because we just took work from the overflow list, 4641 // but of course we can't since all of that could have 4642 // been already stolen from us. 4643 // "He giveth and He taketh away." 4644 continue; 4645 } 4646 // Verify that we have no work before we resort to stealing 4647 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4648 // Try to steal from other queues that have work 4649 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4650 NOT_PRODUCT(num_steals++;) 4651 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4652 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4653 // Do scanning work 4654 obj_to_scan->oop_iterate(cl); 4655 // Loop around, finish this work, and try to steal some more 4656 } else if (terminator()->offer_termination()) { 4657 break; // nirvana from the infinite cycle 4658 } 4659 } 4660 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4661 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4662 "Else our work is not yet done"); 4663 } 4664 4665 // Record object boundaries in _eden_chunk_array by sampling the eden 4666 // top in the slow-path eden object allocation code path and record 4667 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4668 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4669 // sampling in sample_eden() that activates during the part of the 4670 // preclean phase. 4671 void CMSCollector::sample_eden_chunk() { 4672 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4673 if (_eden_chunk_lock->try_lock()) { 4674 // Record a sample. This is the critical section. The contents 4675 // of the _eden_chunk_array have to be non-decreasing in the 4676 // address order. 4677 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4678 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4679 "Unexpected state of Eden"); 4680 if (_eden_chunk_index == 0 || 4681 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4682 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4683 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4684 _eden_chunk_index++; // commit sample 4685 } 4686 _eden_chunk_lock->unlock(); 4687 } 4688 } 4689 } 4690 4691 // Return a thread-local PLAB recording array, as appropriate. 4692 void* CMSCollector::get_data_recorder(int thr_num) { 4693 if (_survivor_plab_array != NULL && 4694 (CMSPLABRecordAlways || 4695 (_collectorState > Marking && _collectorState < FinalMarking))) { 4696 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4697 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4698 ca->reset(); // clear it so that fresh data is recorded 4699 return (void*) ca; 4700 } else { 4701 return NULL; 4702 } 4703 } 4704 4705 // Reset all the thread-local PLAB recording arrays 4706 void CMSCollector::reset_survivor_plab_arrays() { 4707 for (uint i = 0; i < ParallelGCThreads; i++) { 4708 _survivor_plab_array[i].reset(); 4709 } 4710 } 4711 4712 // Merge the per-thread plab arrays into the global survivor chunk 4713 // array which will provide the partitioning of the survivor space 4714 // for CMS initial scan and rescan. 4715 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4716 int no_of_gc_threads) { 4717 assert(_survivor_plab_array != NULL, "Error"); 4718 assert(_survivor_chunk_array != NULL, "Error"); 4719 assert(_collectorState == FinalMarking || 4720 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4721 for (int j = 0; j < no_of_gc_threads; j++) { 4722 _cursor[j] = 0; 4723 } 4724 HeapWord* top = surv->top(); 4725 size_t i; 4726 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4727 HeapWord* min_val = top; // Higher than any PLAB address 4728 uint min_tid = 0; // position of min_val this round 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4731 if (_cursor[j] == cur_sca->end()) { 4732 continue; 4733 } 4734 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4735 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4736 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4737 if (cur_val < min_val) { 4738 min_tid = j; 4739 min_val = cur_val; 4740 } else { 4741 assert(cur_val < top, "All recorded addresses should be less"); 4742 } 4743 } 4744 // At this point min_val and min_tid are respectively 4745 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4746 // and the thread (j) that witnesses that address. 4747 // We record this address in the _survivor_chunk_array[i] 4748 // and increment _cursor[min_tid] prior to the next round i. 4749 if (min_val == top) { 4750 break; 4751 } 4752 _survivor_chunk_array[i] = min_val; 4753 _cursor[min_tid]++; 4754 } 4755 // We are all done; record the size of the _survivor_chunk_array 4756 _survivor_chunk_index = i; // exclusive: [0, i) 4757 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4758 // Verify that we used up all the recorded entries 4759 #ifdef ASSERT 4760 size_t total = 0; 4761 for (int j = 0; j < no_of_gc_threads; j++) { 4762 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4763 total += _cursor[j]; 4764 } 4765 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4766 // Check that the merged array is in sorted order 4767 if (total > 0) { 4768 for (size_t i = 0; i < total - 1; i++) { 4769 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4770 i, p2i(_survivor_chunk_array[i])); 4771 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4772 "Not sorted"); 4773 } 4774 } 4775 #endif // ASSERT 4776 } 4777 4778 // Set up the space's par_seq_tasks structure for work claiming 4779 // for parallel initial scan and rescan of young gen. 4780 // See ParRescanTask where this is currently used. 4781 void 4782 CMSCollector:: 4783 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4784 assert(n_threads > 0, "Unexpected n_threads argument"); 4785 4786 // Eden space 4787 if (!_young_gen->eden()->is_empty()) { 4788 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4789 assert(!pst->valid(), "Clobbering existing data?"); 4790 // Each valid entry in [0, _eden_chunk_index) represents a task. 4791 size_t n_tasks = _eden_chunk_index + 1; 4792 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4793 // Sets the condition for completion of the subtask (how many threads 4794 // need to finish in order to be done). 4795 pst->set_n_threads(n_threads); 4796 pst->set_n_tasks((int)n_tasks); 4797 } 4798 4799 // Merge the survivor plab arrays into _survivor_chunk_array 4800 if (_survivor_plab_array != NULL) { 4801 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4802 } else { 4803 assert(_survivor_chunk_index == 0, "Error"); 4804 } 4805 4806 // To space 4807 { 4808 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4809 assert(!pst->valid(), "Clobbering existing data?"); 4810 // Sets the condition for completion of the subtask (how many threads 4811 // need to finish in order to be done). 4812 pst->set_n_threads(n_threads); 4813 pst->set_n_tasks(1); 4814 assert(pst->valid(), "Error"); 4815 } 4816 4817 // From space 4818 { 4819 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4820 assert(!pst->valid(), "Clobbering existing data?"); 4821 size_t n_tasks = _survivor_chunk_index + 1; 4822 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4823 // Sets the condition for completion of the subtask (how many threads 4824 // need to finish in order to be done). 4825 pst->set_n_threads(n_threads); 4826 pst->set_n_tasks((int)n_tasks); 4827 assert(pst->valid(), "Error"); 4828 } 4829 } 4830 4831 // Parallel version of remark 4832 void CMSCollector::do_remark_parallel() { 4833 CMSHeap* heap = CMSHeap::heap(); 4834 WorkGang* workers = heap->workers(); 4835 assert(workers != NULL, "Need parallel worker threads."); 4836 // Choose to use the number of GC workers most recently set 4837 // into "active_workers". 4838 uint n_workers = workers->active_workers(); 4839 4840 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4841 4842 StrongRootsScope srs(n_workers); 4843 4844 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4845 4846 // We won't be iterating over the cards in the card table updating 4847 // the younger_gen cards, so we shouldn't call the following else 4848 // the verification code as well as subsequent younger_refs_iterate 4849 // code would get confused. XXX 4850 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4851 4852 // The young gen rescan work will not be done as part of 4853 // process_roots (which currently doesn't know how to 4854 // parallelize such a scan), but rather will be broken up into 4855 // a set of parallel tasks (via the sampling that the [abortable] 4856 // preclean phase did of eden, plus the [two] tasks of 4857 // scanning the [two] survivor spaces. Further fine-grain 4858 // parallelization of the scanning of the survivor spaces 4859 // themselves, and of precleaning of the young gen itself 4860 // is deferred to the future. 4861 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4862 4863 // The dirty card rescan work is broken up into a "sequence" 4864 // of parallel tasks (per constituent space) that are dynamically 4865 // claimed by the parallel threads. 4866 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4867 4868 // It turns out that even when we're using 1 thread, doing the work in a 4869 // separate thread causes wide variance in run times. We can't help this 4870 // in the multi-threaded case, but we special-case n=1 here to get 4871 // repeatable measurements of the 1-thread overhead of the parallel code. 4872 if (n_workers > 1) { 4873 // Make refs discovery MT-safe, if it isn't already: it may not 4874 // necessarily be so, since it's possible that we are doing 4875 // ST marking. 4876 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4877 workers->run_task(&tsk); 4878 } else { 4879 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4880 tsk.work(0); 4881 } 4882 4883 // restore, single-threaded for now, any preserved marks 4884 // as a result of work_q overflow 4885 restore_preserved_marks_if_any(); 4886 } 4887 4888 // Non-parallel version of remark 4889 void CMSCollector::do_remark_non_parallel() { 4890 ResourceMark rm; 4891 HandleMark hm; 4892 CMSHeap* heap = CMSHeap::heap(); 4893 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4894 4895 MarkRefsIntoAndScanClosure 4896 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4897 &_markStack, this, 4898 false /* should_yield */, false /* not precleaning */); 4899 MarkFromDirtyCardsClosure 4900 markFromDirtyCardsClosure(this, _span, 4901 NULL, // space is set further below 4902 &_markBitMap, &_markStack, &mrias_cl); 4903 { 4904 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4905 // Iterate over the dirty cards, setting the corresponding bits in the 4906 // mod union table. 4907 { 4908 ModUnionClosure modUnionClosure(&_modUnionTable); 4909 _ct->dirty_card_iterate(_cmsGen->used_region(), 4910 &modUnionClosure); 4911 } 4912 // Having transferred these marks into the modUnionTable, we just need 4913 // to rescan the marked objects on the dirty cards in the modUnionTable. 4914 // The initial marking may have been done during an asynchronous 4915 // collection so there may be dirty bits in the mod-union table. 4916 const int alignment = CardTable::card_size * BitsPerWord; 4917 { 4918 // ... First handle dirty cards in CMS gen 4919 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4920 MemRegion ur = _cmsGen->used_region(); 4921 HeapWord* lb = ur.start(); 4922 HeapWord* ub = align_up(ur.end(), alignment); 4923 MemRegion cms_span(lb, ub); 4924 _modUnionTable.dirty_range_iterate_clear(cms_span, 4925 &markFromDirtyCardsClosure); 4926 verify_work_stacks_empty(); 4927 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4928 } 4929 } 4930 if (VerifyDuringGC && 4931 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4932 HandleMark hm; // Discard invalid handles created during verification 4933 Universe::verify(); 4934 } 4935 { 4936 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4937 4938 verify_work_stacks_empty(); 4939 4940 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4941 StrongRootsScope srs(1); 4942 4943 heap->cms_process_roots(&srs, 4944 true, // young gen as roots 4945 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4946 should_unload_classes(), 4947 &mrias_cl, 4948 NULL); // The dirty klasses will be handled below 4949 4950 assert(should_unload_classes() 4951 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4952 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4953 } 4954 4955 { 4956 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4957 4958 verify_work_stacks_empty(); 4959 4960 // Scan all class loader data objects that might have been introduced 4961 // during concurrent marking. 4962 ResourceMark rm; 4963 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4964 for (int i = 0; i < array->length(); i++) { 4965 mrias_cl.do_cld_nv(array->at(i)); 4966 } 4967 4968 // We don't need to keep track of new CLDs anymore. 4969 ClassLoaderDataGraph::remember_new_clds(false); 4970 4971 verify_work_stacks_empty(); 4972 } 4973 4974 // We might have added oops to ClassLoaderData::_handles during the 4975 // concurrent marking phase. These oops do not point to newly allocated objects 4976 // that are guaranteed to be kept alive. Hence, 4977 // we do have to revisit the _handles block during the remark phase. 4978 { 4979 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4980 4981 verify_work_stacks_empty(); 4982 4983 RemarkCLDClosure remark_closure(&mrias_cl); 4984 ClassLoaderDataGraph::cld_do(&remark_closure); 4985 4986 verify_work_stacks_empty(); 4987 } 4988 4989 verify_work_stacks_empty(); 4990 // Restore evacuated mark words, if any, used for overflow list links 4991 restore_preserved_marks_if_any(); 4992 4993 verify_overflow_empty(); 4994 } 4995 4996 //////////////////////////////////////////////////////// 4997 // Parallel Reference Processing Task Proxy Class 4998 //////////////////////////////////////////////////////// 4999 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5000 OopTaskQueueSet* _queues; 5001 ParallelTaskTerminator _terminator; 5002 public: 5003 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5004 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5005 ParallelTaskTerminator* terminator() { return &_terminator; } 5006 OopTaskQueueSet* queues() { return _queues; } 5007 }; 5008 5009 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5010 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5011 CMSCollector* _collector; 5012 CMSBitMap* _mark_bit_map; 5013 const MemRegion _span; 5014 ProcessTask& _task; 5015 5016 public: 5017 CMSRefProcTaskProxy(ProcessTask& task, 5018 CMSCollector* collector, 5019 const MemRegion& span, 5020 CMSBitMap* mark_bit_map, 5021 AbstractWorkGang* workers, 5022 OopTaskQueueSet* task_queues): 5023 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5024 task_queues, 5025 workers->active_workers()), 5026 _task(task), 5027 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5028 { 5029 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5030 "Inconsistency in _span"); 5031 } 5032 5033 OopTaskQueueSet* task_queues() { return queues(); } 5034 5035 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5036 5037 void do_work_steal(int i, 5038 CMSParDrainMarkingStackClosure* drain, 5039 CMSParKeepAliveClosure* keep_alive, 5040 int* seed); 5041 5042 virtual void work(uint worker_id); 5043 }; 5044 5045 void CMSRefProcTaskProxy::work(uint worker_id) { 5046 ResourceMark rm; 5047 HandleMark hm; 5048 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5049 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5050 _mark_bit_map, 5051 work_queue(worker_id)); 5052 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5053 _mark_bit_map, 5054 work_queue(worker_id)); 5055 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5056 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5057 if (_task.marks_oops_alive()) { 5058 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5059 _collector->hash_seed(worker_id)); 5060 } 5061 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5062 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5063 } 5064 5065 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5066 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5067 EnqueueTask& _task; 5068 5069 public: 5070 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5071 : AbstractGangTask("Enqueue reference objects in parallel"), 5072 _task(task) 5073 { } 5074 5075 virtual void work(uint worker_id) 5076 { 5077 _task.work(worker_id); 5078 } 5079 }; 5080 5081 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5082 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5083 _span(span), 5084 _bit_map(bit_map), 5085 _work_queue(work_queue), 5086 _mark_and_push(collector, span, bit_map, work_queue), 5087 _low_water_mark(MIN2((work_queue->max_elems()/4), 5088 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5089 { } 5090 5091 // . see if we can share work_queues with ParNew? XXX 5092 void CMSRefProcTaskProxy::do_work_steal(int i, 5093 CMSParDrainMarkingStackClosure* drain, 5094 CMSParKeepAliveClosure* keep_alive, 5095 int* seed) { 5096 OopTaskQueue* work_q = work_queue(i); 5097 NOT_PRODUCT(int num_steals = 0;) 5098 oop obj_to_scan; 5099 5100 while (true) { 5101 // Completely finish any left over work from (an) earlier round(s) 5102 drain->trim_queue(0); 5103 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5104 (size_t)ParGCDesiredObjsFromOverflowList); 5105 // Now check if there's any work in the overflow list 5106 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5107 // only affects the number of attempts made to get work from the 5108 // overflow list and does not affect the number of workers. Just 5109 // pass ParallelGCThreads so this behavior is unchanged. 5110 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5111 work_q, 5112 ParallelGCThreads)) { 5113 // Found something in global overflow list; 5114 // not yet ready to go stealing work from others. 5115 // We'd like to assert(work_q->size() != 0, ...) 5116 // because we just took work from the overflow list, 5117 // but of course we can't, since all of that might have 5118 // been already stolen from us. 5119 continue; 5120 } 5121 // Verify that we have no work before we resort to stealing 5122 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5123 // Try to steal from other queues that have work 5124 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5125 NOT_PRODUCT(num_steals++;) 5126 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5127 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5128 // Do scanning work 5129 obj_to_scan->oop_iterate(keep_alive); 5130 // Loop around, finish this work, and try to steal some more 5131 } else if (terminator()->offer_termination()) { 5132 break; // nirvana from the infinite cycle 5133 } 5134 } 5135 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5136 } 5137 5138 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5139 { 5140 CMSHeap* heap = CMSHeap::heap(); 5141 WorkGang* workers = heap->workers(); 5142 assert(workers != NULL, "Need parallel worker threads."); 5143 CMSRefProcTaskProxy rp_task(task, &_collector, 5144 _collector.ref_processor_span(), 5145 _collector.markBitMap(), 5146 workers, _collector.task_queues()); 5147 workers->run_task(&rp_task); 5148 } 5149 5150 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5151 { 5152 5153 CMSHeap* heap = CMSHeap::heap(); 5154 WorkGang* workers = heap->workers(); 5155 assert(workers != NULL, "Need parallel worker threads."); 5156 CMSRefEnqueueTaskProxy enq_task(task); 5157 workers->run_task(&enq_task); 5158 } 5159 5160 void CMSCollector::refProcessingWork() { 5161 ResourceMark rm; 5162 HandleMark hm; 5163 5164 ReferenceProcessor* rp = ref_processor(); 5165 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5166 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5167 // Process weak references. 5168 rp->setup_policy(false); 5169 verify_work_stacks_empty(); 5170 5171 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_queues()); 5172 { 5173 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5174 5175 // Setup keep_alive and complete closures. 5176 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5177 &_markStack, false /* !preclean */); 5178 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5179 _span, &_markBitMap, &_markStack, 5180 &cmsKeepAliveClosure, false /* !preclean */); 5181 5182 ReferenceProcessorStats stats; 5183 if (rp->processing_is_mt()) { 5184 // Set the degree of MT here. If the discovery is done MT, there 5185 // may have been a different number of threads doing the discovery 5186 // and a different number of discovered lists may have Ref objects. 5187 // That is OK as long as the Reference lists are balanced (see 5188 // balance_all_queues() and balance_queues()). 5189 CMSHeap* heap = CMSHeap::heap(); 5190 uint active_workers = ParallelGCThreads; 5191 WorkGang* workers = heap->workers(); 5192 if (workers != NULL) { 5193 active_workers = workers->active_workers(); 5194 // The expectation is that active_workers will have already 5195 // been set to a reasonable value. If it has not been set, 5196 // investigate. 5197 assert(active_workers > 0, "Should have been set during scavenge"); 5198 } 5199 rp->set_active_mt_degree(active_workers); 5200 CMSRefProcTaskExecutor task_executor(*this); 5201 stats = rp->process_discovered_references(&_is_alive_closure, 5202 &cmsKeepAliveClosure, 5203 &cmsDrainMarkingStackClosure, 5204 &task_executor, 5205 &pt); 5206 } else { 5207 stats = rp->process_discovered_references(&_is_alive_closure, 5208 &cmsKeepAliveClosure, 5209 &cmsDrainMarkingStackClosure, 5210 NULL, 5211 &pt); 5212 } 5213 _gc_tracer_cm->report_gc_reference_stats(stats); 5214 pt.print_all_references(); 5215 } 5216 5217 // This is the point where the entire marking should have completed. 5218 verify_work_stacks_empty(); 5219 5220 { 5221 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5222 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5223 } 5224 5225 if (should_unload_classes()) { 5226 { 5227 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5228 5229 // Unload classes and purge the SystemDictionary. 5230 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5231 5232 // Unload nmethods. 5233 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5234 5235 // Prune dead klasses from subklass/sibling/implementor lists. 5236 Klass::clean_weak_klass_links(purged_class); 5237 } 5238 5239 { 5240 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5241 // Clean up unreferenced symbols in symbol table. 5242 SymbolTable::unlink(); 5243 } 5244 5245 { 5246 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5247 // Delete entries for dead interned strings. 5248 StringTable::unlink(&_is_alive_closure); 5249 } 5250 } 5251 5252 // Restore any preserved marks as a result of mark stack or 5253 // work queue overflow 5254 restore_preserved_marks_if_any(); // done single-threaded for now 5255 5256 rp->set_enqueuing_is_done(true); 5257 if (rp->processing_is_mt()) { 5258 rp->balance_all_queues(); 5259 CMSRefProcTaskExecutor task_executor(*this); 5260 rp->enqueue_discovered_references(&task_executor, &pt); 5261 } else { 5262 rp->enqueue_discovered_references(NULL, &pt); 5263 } 5264 rp->verify_no_references_recorded(); 5265 pt.print_enqueue_phase(); 5266 assert(!rp->discovery_enabled(), "should have been disabled"); 5267 } 5268 5269 #ifndef PRODUCT 5270 void CMSCollector::check_correct_thread_executing() { 5271 Thread* t = Thread::current(); 5272 // Only the VM thread or the CMS thread should be here. 5273 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5274 "Unexpected thread type"); 5275 // If this is the vm thread, the foreground process 5276 // should not be waiting. Note that _foregroundGCIsActive is 5277 // true while the foreground collector is waiting. 5278 if (_foregroundGCShouldWait) { 5279 // We cannot be the VM thread 5280 assert(t->is_ConcurrentGC_thread(), 5281 "Should be CMS thread"); 5282 } else { 5283 // We can be the CMS thread only if we are in a stop-world 5284 // phase of CMS collection. 5285 if (t->is_ConcurrentGC_thread()) { 5286 assert(_collectorState == InitialMarking || 5287 _collectorState == FinalMarking, 5288 "Should be a stop-world phase"); 5289 // The CMS thread should be holding the CMS_token. 5290 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5291 "Potential interference with concurrently " 5292 "executing VM thread"); 5293 } 5294 } 5295 } 5296 #endif 5297 5298 void CMSCollector::sweep() { 5299 assert(_collectorState == Sweeping, "just checking"); 5300 check_correct_thread_executing(); 5301 verify_work_stacks_empty(); 5302 verify_overflow_empty(); 5303 increment_sweep_count(); 5304 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5305 5306 _inter_sweep_timer.stop(); 5307 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5308 5309 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5310 _intra_sweep_timer.reset(); 5311 _intra_sweep_timer.start(); 5312 { 5313 GCTraceCPUTime tcpu; 5314 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5315 // First sweep the old gen 5316 { 5317 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5318 bitMapLock()); 5319 sweepWork(_cmsGen); 5320 } 5321 5322 // Update Universe::_heap_*_at_gc figures. 5323 // We need all the free list locks to make the abstract state 5324 // transition from Sweeping to Resetting. See detailed note 5325 // further below. 5326 { 5327 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5328 // Update heap occupancy information which is used as 5329 // input to soft ref clearing policy at the next gc. 5330 Universe::update_heap_info_at_gc(); 5331 _collectorState = Resizing; 5332 } 5333 } 5334 verify_work_stacks_empty(); 5335 verify_overflow_empty(); 5336 5337 if (should_unload_classes()) { 5338 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5339 // requires that the virtual spaces are stable and not deleted. 5340 ClassLoaderDataGraph::set_should_purge(true); 5341 } 5342 5343 _intra_sweep_timer.stop(); 5344 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5345 5346 _inter_sweep_timer.reset(); 5347 _inter_sweep_timer.start(); 5348 5349 // We need to use a monotonically non-decreasing time in ms 5350 // or we will see time-warp warnings and os::javaTimeMillis() 5351 // does not guarantee monotonicity. 5352 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5353 update_time_of_last_gc(now); 5354 5355 // NOTE on abstract state transitions: 5356 // Mutators allocate-live and/or mark the mod-union table dirty 5357 // based on the state of the collection. The former is done in 5358 // the interval [Marking, Sweeping] and the latter in the interval 5359 // [Marking, Sweeping). Thus the transitions into the Marking state 5360 // and out of the Sweeping state must be synchronously visible 5361 // globally to the mutators. 5362 // The transition into the Marking state happens with the world 5363 // stopped so the mutators will globally see it. Sweeping is 5364 // done asynchronously by the background collector so the transition 5365 // from the Sweeping state to the Resizing state must be done 5366 // under the freelistLock (as is the check for whether to 5367 // allocate-live and whether to dirty the mod-union table). 5368 assert(_collectorState == Resizing, "Change of collector state to" 5369 " Resizing must be done under the freelistLocks (plural)"); 5370 5371 // Now that sweeping has been completed, we clear 5372 // the incremental_collection_failed flag, 5373 // thus inviting a younger gen collection to promote into 5374 // this generation. If such a promotion may still fail, 5375 // the flag will be set again when a young collection is 5376 // attempted. 5377 CMSHeap* heap = CMSHeap::heap(); 5378 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5379 heap->update_full_collections_completed(_collection_count_start); 5380 } 5381 5382 // FIX ME!!! Looks like this belongs in CFLSpace, with 5383 // CMSGen merely delegating to it. 5384 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5385 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5386 HeapWord* minAddr = _cmsSpace->bottom(); 5387 HeapWord* largestAddr = 5388 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5389 if (largestAddr == NULL) { 5390 // The dictionary appears to be empty. In this case 5391 // try to coalesce at the end of the heap. 5392 largestAddr = _cmsSpace->end(); 5393 } 5394 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5395 size_t nearLargestOffset = 5396 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5397 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5398 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5399 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5400 } 5401 5402 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5403 return addr >= _cmsSpace->nearLargestChunk(); 5404 } 5405 5406 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5407 return _cmsSpace->find_chunk_at_end(); 5408 } 5409 5410 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5411 bool full) { 5412 // If the young generation has been collected, gather any statistics 5413 // that are of interest at this point. 5414 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5415 if (!full && current_is_young) { 5416 // Gather statistics on the young generation collection. 5417 collector()->stats().record_gc0_end(used()); 5418 } 5419 } 5420 5421 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5422 // We iterate over the space(s) underlying this generation, 5423 // checking the mark bit map to see if the bits corresponding 5424 // to specific blocks are marked or not. Blocks that are 5425 // marked are live and are not swept up. All remaining blocks 5426 // are swept up, with coalescing on-the-fly as we sweep up 5427 // contiguous free and/or garbage blocks: 5428 // We need to ensure that the sweeper synchronizes with allocators 5429 // and stop-the-world collectors. In particular, the following 5430 // locks are used: 5431 // . CMS token: if this is held, a stop the world collection cannot occur 5432 // . freelistLock: if this is held no allocation can occur from this 5433 // generation by another thread 5434 // . bitMapLock: if this is held, no other thread can access or update 5435 // 5436 5437 // Note that we need to hold the freelistLock if we use 5438 // block iterate below; else the iterator might go awry if 5439 // a mutator (or promotion) causes block contents to change 5440 // (for instance if the allocator divvies up a block). 5441 // If we hold the free list lock, for all practical purposes 5442 // young generation GC's can't occur (they'll usually need to 5443 // promote), so we might as well prevent all young generation 5444 // GC's while we do a sweeping step. For the same reason, we might 5445 // as well take the bit map lock for the entire duration 5446 5447 // check that we hold the requisite locks 5448 assert(have_cms_token(), "Should hold cms token"); 5449 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5450 assert_lock_strong(old_gen->freelistLock()); 5451 assert_lock_strong(bitMapLock()); 5452 5453 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5454 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5455 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5456 _inter_sweep_estimate.padded_average(), 5457 _intra_sweep_estimate.padded_average()); 5458 old_gen->setNearLargestChunk(); 5459 5460 { 5461 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5462 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5463 // We need to free-up/coalesce garbage/blocks from a 5464 // co-terminal free run. This is done in the SweepClosure 5465 // destructor; so, do not remove this scope, else the 5466 // end-of-sweep-census below will be off by a little bit. 5467 } 5468 old_gen->cmsSpace()->sweep_completed(); 5469 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5470 if (should_unload_classes()) { // unloaded classes this cycle, 5471 _concurrent_cycles_since_last_unload = 0; // ... reset count 5472 } else { // did not unload classes, 5473 _concurrent_cycles_since_last_unload++; // ... increment count 5474 } 5475 } 5476 5477 // Reset CMS data structures (for now just the marking bit map) 5478 // preparatory for the next cycle. 5479 void CMSCollector::reset_concurrent() { 5480 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5481 5482 // If the state is not "Resetting", the foreground thread 5483 // has done a collection and the resetting. 5484 if (_collectorState != Resetting) { 5485 assert(_collectorState == Idling, "The state should only change" 5486 " because the foreground collector has finished the collection"); 5487 return; 5488 } 5489 5490 { 5491 // Clear the mark bitmap (no grey objects to start with) 5492 // for the next cycle. 5493 GCTraceCPUTime tcpu; 5494 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5495 5496 HeapWord* curAddr = _markBitMap.startWord(); 5497 while (curAddr < _markBitMap.endWord()) { 5498 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5499 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5500 _markBitMap.clear_large_range(chunk); 5501 if (ConcurrentMarkSweepThread::should_yield() && 5502 !foregroundGCIsActive() && 5503 CMSYield) { 5504 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5505 "CMS thread should hold CMS token"); 5506 assert_lock_strong(bitMapLock()); 5507 bitMapLock()->unlock(); 5508 ConcurrentMarkSweepThread::desynchronize(true); 5509 stopTimer(); 5510 incrementYields(); 5511 5512 // See the comment in coordinator_yield() 5513 for (unsigned i = 0; i < CMSYieldSleepCount && 5514 ConcurrentMarkSweepThread::should_yield() && 5515 !CMSCollector::foregroundGCIsActive(); ++i) { 5516 os::sleep(Thread::current(), 1, false); 5517 } 5518 5519 ConcurrentMarkSweepThread::synchronize(true); 5520 bitMapLock()->lock_without_safepoint_check(); 5521 startTimer(); 5522 } 5523 curAddr = chunk.end(); 5524 } 5525 // A successful mostly concurrent collection has been done. 5526 // Because only the full (i.e., concurrent mode failure) collections 5527 // are being measured for gc overhead limits, clean the "near" flag 5528 // and count. 5529 size_policy()->reset_gc_overhead_limit_count(); 5530 _collectorState = Idling; 5531 } 5532 5533 register_gc_end(); 5534 } 5535 5536 // Same as above but for STW paths 5537 void CMSCollector::reset_stw() { 5538 // already have the lock 5539 assert(_collectorState == Resetting, "just checking"); 5540 assert_lock_strong(bitMapLock()); 5541 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5542 _markBitMap.clear_all(); 5543 _collectorState = Idling; 5544 register_gc_end(); 5545 } 5546 5547 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5548 GCTraceCPUTime tcpu; 5549 TraceCollectorStats tcs_cgc(cgc_counters()); 5550 5551 switch (op) { 5552 case CMS_op_checkpointRootsInitial: { 5553 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5554 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5555 checkpointRootsInitial(); 5556 break; 5557 } 5558 case CMS_op_checkpointRootsFinal: { 5559 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5560 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5561 checkpointRootsFinal(); 5562 break; 5563 } 5564 default: 5565 fatal("No such CMS_op"); 5566 } 5567 } 5568 5569 #ifndef PRODUCT 5570 size_t const CMSCollector::skip_header_HeapWords() { 5571 return FreeChunk::header_size(); 5572 } 5573 5574 // Try and collect here conditions that should hold when 5575 // CMS thread is exiting. The idea is that the foreground GC 5576 // thread should not be blocked if it wants to terminate 5577 // the CMS thread and yet continue to run the VM for a while 5578 // after that. 5579 void CMSCollector::verify_ok_to_terminate() const { 5580 assert(Thread::current()->is_ConcurrentGC_thread(), 5581 "should be called by CMS thread"); 5582 assert(!_foregroundGCShouldWait, "should be false"); 5583 // We could check here that all the various low-level locks 5584 // are not held by the CMS thread, but that is overkill; see 5585 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5586 // is checked. 5587 } 5588 #endif 5589 5590 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5591 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5592 "missing Printezis mark?"); 5593 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5594 size_t size = pointer_delta(nextOneAddr + 1, addr); 5595 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5596 "alignment problem"); 5597 assert(size >= 3, "Necessary for Printezis marks to work"); 5598 return size; 5599 } 5600 5601 // A variant of the above (block_size_using_printezis_bits()) except 5602 // that we return 0 if the P-bits are not yet set. 5603 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5604 if (_markBitMap.isMarked(addr + 1)) { 5605 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5606 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5607 size_t size = pointer_delta(nextOneAddr + 1, addr); 5608 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5609 "alignment problem"); 5610 assert(size >= 3, "Necessary for Printezis marks to work"); 5611 return size; 5612 } 5613 return 0; 5614 } 5615 5616 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5617 size_t sz = 0; 5618 oop p = (oop)addr; 5619 if (p->klass_or_null_acquire() != NULL) { 5620 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5621 } else { 5622 sz = block_size_using_printezis_bits(addr); 5623 } 5624 assert(sz > 0, "size must be nonzero"); 5625 HeapWord* next_block = addr + sz; 5626 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5627 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5628 align_down((uintptr_t)next_card, CardTable::card_size), 5629 "must be different cards"); 5630 return next_card; 5631 } 5632 5633 5634 // CMS Bit Map Wrapper ///////////////////////////////////////// 5635 5636 // Construct a CMS bit map infrastructure, but don't create the 5637 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5638 // further below. 5639 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5640 _bm(), 5641 _shifter(shifter), 5642 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5643 Monitor::_safepoint_check_sometimes) : NULL) 5644 { 5645 _bmStartWord = 0; 5646 _bmWordSize = 0; 5647 } 5648 5649 bool CMSBitMap::allocate(MemRegion mr) { 5650 _bmStartWord = mr.start(); 5651 _bmWordSize = mr.word_size(); 5652 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5653 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5654 if (!brs.is_reserved()) { 5655 log_warning(gc)("CMS bit map allocation failure"); 5656 return false; 5657 } 5658 // For now we'll just commit all of the bit map up front. 5659 // Later on we'll try to be more parsimonious with swap. 5660 if (!_virtual_space.initialize(brs, brs.size())) { 5661 log_warning(gc)("CMS bit map backing store failure"); 5662 return false; 5663 } 5664 assert(_virtual_space.committed_size() == brs.size(), 5665 "didn't reserve backing store for all of CMS bit map?"); 5666 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5667 _bmWordSize, "inconsistency in bit map sizing"); 5668 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5669 5670 // bm.clear(); // can we rely on getting zero'd memory? verify below 5671 assert(isAllClear(), 5672 "Expected zero'd memory from ReservedSpace constructor"); 5673 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5674 "consistency check"); 5675 return true; 5676 } 5677 5678 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5679 HeapWord *next_addr, *end_addr, *last_addr; 5680 assert_locked(); 5681 assert(covers(mr), "out-of-range error"); 5682 // XXX assert that start and end are appropriately aligned 5683 for (next_addr = mr.start(), end_addr = mr.end(); 5684 next_addr < end_addr; next_addr = last_addr) { 5685 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5686 last_addr = dirty_region.end(); 5687 if (!dirty_region.is_empty()) { 5688 cl->do_MemRegion(dirty_region); 5689 } else { 5690 assert(last_addr == end_addr, "program logic"); 5691 return; 5692 } 5693 } 5694 } 5695 5696 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5697 _bm.print_on_error(st, prefix); 5698 } 5699 5700 #ifndef PRODUCT 5701 void CMSBitMap::assert_locked() const { 5702 CMSLockVerifier::assert_locked(lock()); 5703 } 5704 5705 bool CMSBitMap::covers(MemRegion mr) const { 5706 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5707 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5708 "size inconsistency"); 5709 return (mr.start() >= _bmStartWord) && 5710 (mr.end() <= endWord()); 5711 } 5712 5713 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5714 return (start >= _bmStartWord && (start + size) <= endWord()); 5715 } 5716 5717 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5718 // verify that there are no 1 bits in the interval [left, right) 5719 FalseBitMapClosure falseBitMapClosure; 5720 iterate(&falseBitMapClosure, left, right); 5721 } 5722 5723 void CMSBitMap::region_invariant(MemRegion mr) 5724 { 5725 assert_locked(); 5726 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5727 assert(!mr.is_empty(), "unexpected empty region"); 5728 assert(covers(mr), "mr should be covered by bit map"); 5729 // convert address range into offset range 5730 size_t start_ofs = heapWordToOffset(mr.start()); 5731 // Make sure that end() is appropriately aligned 5732 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5733 "Misaligned mr.end()"); 5734 size_t end_ofs = heapWordToOffset(mr.end()); 5735 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5736 } 5737 5738 #endif 5739 5740 bool CMSMarkStack::allocate(size_t size) { 5741 // allocate a stack of the requisite depth 5742 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5743 size * sizeof(oop))); 5744 if (!rs.is_reserved()) { 5745 log_warning(gc)("CMSMarkStack allocation failure"); 5746 return false; 5747 } 5748 if (!_virtual_space.initialize(rs, rs.size())) { 5749 log_warning(gc)("CMSMarkStack backing store failure"); 5750 return false; 5751 } 5752 assert(_virtual_space.committed_size() == rs.size(), 5753 "didn't reserve backing store for all of CMS stack?"); 5754 _base = (oop*)(_virtual_space.low()); 5755 _index = 0; 5756 _capacity = size; 5757 NOT_PRODUCT(_max_depth = 0); 5758 return true; 5759 } 5760 5761 // XXX FIX ME !!! In the MT case we come in here holding a 5762 // leaf lock. For printing we need to take a further lock 5763 // which has lower rank. We need to recalibrate the two 5764 // lock-ranks involved in order to be able to print the 5765 // messages below. (Or defer the printing to the caller. 5766 // For now we take the expedient path of just disabling the 5767 // messages for the problematic case.) 5768 void CMSMarkStack::expand() { 5769 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5770 if (_capacity == MarkStackSizeMax) { 5771 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5772 // We print a warning message only once per CMS cycle. 5773 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5774 } 5775 return; 5776 } 5777 // Double capacity if possible 5778 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5779 // Do not give up existing stack until we have managed to 5780 // get the double capacity that we desired. 5781 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5782 new_capacity * sizeof(oop))); 5783 if (rs.is_reserved()) { 5784 // Release the backing store associated with old stack 5785 _virtual_space.release(); 5786 // Reinitialize virtual space for new stack 5787 if (!_virtual_space.initialize(rs, rs.size())) { 5788 fatal("Not enough swap for expanded marking stack"); 5789 } 5790 _base = (oop*)(_virtual_space.low()); 5791 _index = 0; 5792 _capacity = new_capacity; 5793 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5794 // Failed to double capacity, continue; 5795 // we print a detail message only once per CMS cycle. 5796 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5797 _capacity / K, new_capacity / K); 5798 } 5799 } 5800 5801 5802 // Closures 5803 // XXX: there seems to be a lot of code duplication here; 5804 // should refactor and consolidate common code. 5805 5806 // This closure is used to mark refs into the CMS generation in 5807 // the CMS bit map. Called at the first checkpoint. This closure 5808 // assumes that we do not need to re-mark dirty cards; if the CMS 5809 // generation on which this is used is not an oldest 5810 // generation then this will lose younger_gen cards! 5811 5812 MarkRefsIntoClosure::MarkRefsIntoClosure( 5813 MemRegion span, CMSBitMap* bitMap): 5814 _span(span), 5815 _bitMap(bitMap) 5816 { 5817 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5818 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5819 } 5820 5821 void MarkRefsIntoClosure::do_oop(oop obj) { 5822 // if p points into _span, then mark corresponding bit in _markBitMap 5823 assert(oopDesc::is_oop(obj), "expected an oop"); 5824 HeapWord* addr = (HeapWord*)obj; 5825 if (_span.contains(addr)) { 5826 // this should be made more efficient 5827 _bitMap->mark(addr); 5828 } 5829 } 5830 5831 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5832 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5833 5834 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5835 MemRegion span, CMSBitMap* bitMap): 5836 _span(span), 5837 _bitMap(bitMap) 5838 { 5839 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5840 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5841 } 5842 5843 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5844 // if p points into _span, then mark corresponding bit in _markBitMap 5845 assert(oopDesc::is_oop(obj), "expected an oop"); 5846 HeapWord* addr = (HeapWord*)obj; 5847 if (_span.contains(addr)) { 5848 // this should be made more efficient 5849 _bitMap->par_mark(addr); 5850 } 5851 } 5852 5853 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5854 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5855 5856 // A variant of the above, used for CMS marking verification. 5857 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5858 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5859 _span(span), 5860 _verification_bm(verification_bm), 5861 _cms_bm(cms_bm) 5862 { 5863 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5864 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5865 } 5866 5867 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5868 // if p points into _span, then mark corresponding bit in _markBitMap 5869 assert(oopDesc::is_oop(obj), "expected an oop"); 5870 HeapWord* addr = (HeapWord*)obj; 5871 if (_span.contains(addr)) { 5872 _verification_bm->mark(addr); 5873 if (!_cms_bm->isMarked(addr)) { 5874 Log(gc, verify) log; 5875 ResourceMark rm; 5876 LogStream ls(log.error()); 5877 oop(addr)->print_on(&ls); 5878 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5879 fatal("... aborting"); 5880 } 5881 } 5882 } 5883 5884 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5885 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5886 5887 ////////////////////////////////////////////////// 5888 // MarkRefsIntoAndScanClosure 5889 ////////////////////////////////////////////////// 5890 5891 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5892 ReferenceDiscoverer* rd, 5893 CMSBitMap* bit_map, 5894 CMSBitMap* mod_union_table, 5895 CMSMarkStack* mark_stack, 5896 CMSCollector* collector, 5897 bool should_yield, 5898 bool concurrent_precleaning): 5899 _collector(collector), 5900 _span(span), 5901 _bit_map(bit_map), 5902 _mark_stack(mark_stack), 5903 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5904 mark_stack, concurrent_precleaning), 5905 _yield(should_yield), 5906 _concurrent_precleaning(concurrent_precleaning), 5907 _freelistLock(NULL) 5908 { 5909 // FIXME: Should initialize in base class constructor. 5910 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5911 set_ref_discoverer_internal(rd); 5912 } 5913 5914 // This closure is used to mark refs into the CMS generation at the 5915 // second (final) checkpoint, and to scan and transitively follow 5916 // the unmarked oops. It is also used during the concurrent precleaning 5917 // phase while scanning objects on dirty cards in the CMS generation. 5918 // The marks are made in the marking bit map and the marking stack is 5919 // used for keeping the (newly) grey objects during the scan. 5920 // The parallel version (Par_...) appears further below. 5921 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5922 if (obj != NULL) { 5923 assert(oopDesc::is_oop(obj), "expected an oop"); 5924 HeapWord* addr = (HeapWord*)obj; 5925 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5926 assert(_collector->overflow_list_is_empty(), 5927 "overflow list should be empty"); 5928 if (_span.contains(addr) && 5929 !_bit_map->isMarked(addr)) { 5930 // mark bit map (object is now grey) 5931 _bit_map->mark(addr); 5932 // push on marking stack (stack should be empty), and drain the 5933 // stack by applying this closure to the oops in the oops popped 5934 // from the stack (i.e. blacken the grey objects) 5935 bool res = _mark_stack->push(obj); 5936 assert(res, "Should have space to push on empty stack"); 5937 do { 5938 oop new_oop = _mark_stack->pop(); 5939 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5940 assert(_bit_map->isMarked((HeapWord*)new_oop), 5941 "only grey objects on this stack"); 5942 // iterate over the oops in this oop, marking and pushing 5943 // the ones in CMS heap (i.e. in _span). 5944 new_oop->oop_iterate(&_pushAndMarkClosure); 5945 // check if it's time to yield 5946 do_yield_check(); 5947 } while (!_mark_stack->isEmpty() || 5948 (!_concurrent_precleaning && take_from_overflow_list())); 5949 // if marking stack is empty, and we are not doing this 5950 // during precleaning, then check the overflow list 5951 } 5952 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5953 assert(_collector->overflow_list_is_empty(), 5954 "overflow list was drained above"); 5955 5956 assert(_collector->no_preserved_marks(), 5957 "All preserved marks should have been restored above"); 5958 } 5959 } 5960 5961 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5962 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5963 5964 void MarkRefsIntoAndScanClosure::do_yield_work() { 5965 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5966 "CMS thread should hold CMS token"); 5967 assert_lock_strong(_freelistLock); 5968 assert_lock_strong(_bit_map->lock()); 5969 // relinquish the free_list_lock and bitMaplock() 5970 _bit_map->lock()->unlock(); 5971 _freelistLock->unlock(); 5972 ConcurrentMarkSweepThread::desynchronize(true); 5973 _collector->stopTimer(); 5974 _collector->incrementYields(); 5975 5976 // See the comment in coordinator_yield() 5977 for (unsigned i = 0; 5978 i < CMSYieldSleepCount && 5979 ConcurrentMarkSweepThread::should_yield() && 5980 !CMSCollector::foregroundGCIsActive(); 5981 ++i) { 5982 os::sleep(Thread::current(), 1, false); 5983 } 5984 5985 ConcurrentMarkSweepThread::synchronize(true); 5986 _freelistLock->lock_without_safepoint_check(); 5987 _bit_map->lock()->lock_without_safepoint_check(); 5988 _collector->startTimer(); 5989 } 5990 5991 /////////////////////////////////////////////////////////// 5992 // ParMarkRefsIntoAndScanClosure: a parallel version of 5993 // MarkRefsIntoAndScanClosure 5994 /////////////////////////////////////////////////////////// 5995 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5996 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5997 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5998 _span(span), 5999 _bit_map(bit_map), 6000 _work_queue(work_queue), 6001 _low_water_mark(MIN2((work_queue->max_elems()/4), 6002 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6003 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 6004 { 6005 // FIXME: Should initialize in base class constructor. 6006 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 6007 set_ref_discoverer_internal(rd); 6008 } 6009 6010 // This closure is used to mark refs into the CMS generation at the 6011 // second (final) checkpoint, and to scan and transitively follow 6012 // the unmarked oops. The marks are made in the marking bit map and 6013 // the work_queue is used for keeping the (newly) grey objects during 6014 // the scan phase whence they are also available for stealing by parallel 6015 // threads. Since the marking bit map is shared, updates are 6016 // synchronized (via CAS). 6017 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6018 if (obj != NULL) { 6019 // Ignore mark word because this could be an already marked oop 6020 // that may be chained at the end of the overflow list. 6021 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6022 HeapWord* addr = (HeapWord*)obj; 6023 if (_span.contains(addr) && 6024 !_bit_map->isMarked(addr)) { 6025 // mark bit map (object will become grey): 6026 // It is possible for several threads to be 6027 // trying to "claim" this object concurrently; 6028 // the unique thread that succeeds in marking the 6029 // object first will do the subsequent push on 6030 // to the work queue (or overflow list). 6031 if (_bit_map->par_mark(addr)) { 6032 // push on work_queue (which may not be empty), and trim the 6033 // queue to an appropriate length by applying this closure to 6034 // the oops in the oops popped from the stack (i.e. blacken the 6035 // grey objects) 6036 bool res = _work_queue->push(obj); 6037 assert(res, "Low water mark should be less than capacity?"); 6038 trim_queue(_low_water_mark); 6039 } // Else, another thread claimed the object 6040 } 6041 } 6042 } 6043 6044 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6045 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6046 6047 // This closure is used to rescan the marked objects on the dirty cards 6048 // in the mod union table and the card table proper. 6049 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6050 oop p, MemRegion mr) { 6051 6052 size_t size = 0; 6053 HeapWord* addr = (HeapWord*)p; 6054 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6055 assert(_span.contains(addr), "we are scanning the CMS generation"); 6056 // check if it's time to yield 6057 if (do_yield_check()) { 6058 // We yielded for some foreground stop-world work, 6059 // and we have been asked to abort this ongoing preclean cycle. 6060 return 0; 6061 } 6062 if (_bitMap->isMarked(addr)) { 6063 // it's marked; is it potentially uninitialized? 6064 if (p->klass_or_null_acquire() != NULL) { 6065 // an initialized object; ignore mark word in verification below 6066 // since we are running concurrent with mutators 6067 assert(oopDesc::is_oop(p, true), "should be an oop"); 6068 if (p->is_objArray()) { 6069 // objArrays are precisely marked; restrict scanning 6070 // to dirty cards only. 6071 size = CompactibleFreeListSpace::adjustObjectSize( 6072 p->oop_iterate_size(_scanningClosure, mr)); 6073 } else { 6074 // A non-array may have been imprecisely marked; we need 6075 // to scan object in its entirety. 6076 size = CompactibleFreeListSpace::adjustObjectSize( 6077 p->oop_iterate_size(_scanningClosure)); 6078 } 6079 #ifdef ASSERT 6080 size_t direct_size = 6081 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6082 assert(size == direct_size, "Inconsistency in size"); 6083 assert(size >= 3, "Necessary for Printezis marks to work"); 6084 HeapWord* start_pbit = addr + 1; 6085 HeapWord* end_pbit = addr + size - 1; 6086 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6087 "inconsistent Printezis mark"); 6088 // Verify inner mark bits (between Printezis bits) are clear, 6089 // but don't repeat if there are multiple dirty regions for 6090 // the same object, to avoid potential O(N^2) performance. 6091 if (addr != _last_scanned_object) { 6092 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6093 _last_scanned_object = addr; 6094 } 6095 #endif // ASSERT 6096 } else { 6097 // An uninitialized object. 6098 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6099 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6100 size = pointer_delta(nextOneAddr + 1, addr); 6101 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6102 "alignment problem"); 6103 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6104 // will dirty the card when the klass pointer is installed in the 6105 // object (signaling the completion of initialization). 6106 } 6107 } else { 6108 // Either a not yet marked object or an uninitialized object 6109 if (p->klass_or_null_acquire() == NULL) { 6110 // An uninitialized object, skip to the next card, since 6111 // we may not be able to read its P-bits yet. 6112 assert(size == 0, "Initial value"); 6113 } else { 6114 // An object not (yet) reached by marking: we merely need to 6115 // compute its size so as to go look at the next block. 6116 assert(oopDesc::is_oop(p, true), "should be an oop"); 6117 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6118 } 6119 } 6120 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6121 return size; 6122 } 6123 6124 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6125 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6126 "CMS thread should hold CMS token"); 6127 assert_lock_strong(_freelistLock); 6128 assert_lock_strong(_bitMap->lock()); 6129 // relinquish the free_list_lock and bitMaplock() 6130 _bitMap->lock()->unlock(); 6131 _freelistLock->unlock(); 6132 ConcurrentMarkSweepThread::desynchronize(true); 6133 _collector->stopTimer(); 6134 _collector->incrementYields(); 6135 6136 // See the comment in coordinator_yield() 6137 for (unsigned i = 0; i < CMSYieldSleepCount && 6138 ConcurrentMarkSweepThread::should_yield() && 6139 !CMSCollector::foregroundGCIsActive(); ++i) { 6140 os::sleep(Thread::current(), 1, false); 6141 } 6142 6143 ConcurrentMarkSweepThread::synchronize(true); 6144 _freelistLock->lock_without_safepoint_check(); 6145 _bitMap->lock()->lock_without_safepoint_check(); 6146 _collector->startTimer(); 6147 } 6148 6149 6150 ////////////////////////////////////////////////////////////////// 6151 // SurvivorSpacePrecleanClosure 6152 ////////////////////////////////////////////////////////////////// 6153 // This (single-threaded) closure is used to preclean the oops in 6154 // the survivor spaces. 6155 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6156 6157 HeapWord* addr = (HeapWord*)p; 6158 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6159 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6160 assert(p->klass_or_null() != NULL, "object should be initialized"); 6161 // an initialized object; ignore mark word in verification below 6162 // since we are running concurrent with mutators 6163 assert(oopDesc::is_oop(p, true), "should be an oop"); 6164 // Note that we do not yield while we iterate over 6165 // the interior oops of p, pushing the relevant ones 6166 // on our marking stack. 6167 size_t size = p->oop_iterate_size(_scanning_closure); 6168 do_yield_check(); 6169 // Observe that below, we do not abandon the preclean 6170 // phase as soon as we should; rather we empty the 6171 // marking stack before returning. This is to satisfy 6172 // some existing assertions. In general, it may be a 6173 // good idea to abort immediately and complete the marking 6174 // from the grey objects at a later time. 6175 while (!_mark_stack->isEmpty()) { 6176 oop new_oop = _mark_stack->pop(); 6177 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6178 assert(_bit_map->isMarked((HeapWord*)new_oop), 6179 "only grey objects on this stack"); 6180 // iterate over the oops in this oop, marking and pushing 6181 // the ones in CMS heap (i.e. in _span). 6182 new_oop->oop_iterate(_scanning_closure); 6183 // check if it's time to yield 6184 do_yield_check(); 6185 } 6186 unsigned int after_count = 6187 CMSHeap::heap()->total_collections(); 6188 bool abort = (_before_count != after_count) || 6189 _collector->should_abort_preclean(); 6190 return abort ? 0 : size; 6191 } 6192 6193 void SurvivorSpacePrecleanClosure::do_yield_work() { 6194 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6195 "CMS thread should hold CMS token"); 6196 assert_lock_strong(_bit_map->lock()); 6197 // Relinquish the bit map lock 6198 _bit_map->lock()->unlock(); 6199 ConcurrentMarkSweepThread::desynchronize(true); 6200 _collector->stopTimer(); 6201 _collector->incrementYields(); 6202 6203 // See the comment in coordinator_yield() 6204 for (unsigned i = 0; i < CMSYieldSleepCount && 6205 ConcurrentMarkSweepThread::should_yield() && 6206 !CMSCollector::foregroundGCIsActive(); ++i) { 6207 os::sleep(Thread::current(), 1, false); 6208 } 6209 6210 ConcurrentMarkSweepThread::synchronize(true); 6211 _bit_map->lock()->lock_without_safepoint_check(); 6212 _collector->startTimer(); 6213 } 6214 6215 // This closure is used to rescan the marked objects on the dirty cards 6216 // in the mod union table and the card table proper. In the parallel 6217 // case, although the bitMap is shared, we do a single read so the 6218 // isMarked() query is "safe". 6219 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6220 // Ignore mark word because we are running concurrent with mutators 6221 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6222 HeapWord* addr = (HeapWord*)p; 6223 assert(_span.contains(addr), "we are scanning the CMS generation"); 6224 bool is_obj_array = false; 6225 #ifdef ASSERT 6226 if (!_parallel) { 6227 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6228 assert(_collector->overflow_list_is_empty(), 6229 "overflow list should be empty"); 6230 6231 } 6232 #endif // ASSERT 6233 if (_bit_map->isMarked(addr)) { 6234 // Obj arrays are precisely marked, non-arrays are not; 6235 // so we scan objArrays precisely and non-arrays in their 6236 // entirety. 6237 if (p->is_objArray()) { 6238 is_obj_array = true; 6239 if (_parallel) { 6240 p->oop_iterate(_par_scan_closure, mr); 6241 } else { 6242 p->oop_iterate(_scan_closure, mr); 6243 } 6244 } else { 6245 if (_parallel) { 6246 p->oop_iterate(_par_scan_closure); 6247 } else { 6248 p->oop_iterate(_scan_closure); 6249 } 6250 } 6251 } 6252 #ifdef ASSERT 6253 if (!_parallel) { 6254 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6255 assert(_collector->overflow_list_is_empty(), 6256 "overflow list should be empty"); 6257 6258 } 6259 #endif // ASSERT 6260 return is_obj_array; 6261 } 6262 6263 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6264 MemRegion span, 6265 CMSBitMap* bitMap, CMSMarkStack* markStack, 6266 bool should_yield, bool verifying): 6267 _collector(collector), 6268 _span(span), 6269 _bitMap(bitMap), 6270 _mut(&collector->_modUnionTable), 6271 _markStack(markStack), 6272 _yield(should_yield), 6273 _skipBits(0) 6274 { 6275 assert(_markStack->isEmpty(), "stack should be empty"); 6276 _finger = _bitMap->startWord(); 6277 _threshold = _finger; 6278 assert(_collector->_restart_addr == NULL, "Sanity check"); 6279 assert(_span.contains(_finger), "Out of bounds _finger?"); 6280 DEBUG_ONLY(_verifying = verifying;) 6281 } 6282 6283 void MarkFromRootsClosure::reset(HeapWord* addr) { 6284 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6285 assert(_span.contains(addr), "Out of bounds _finger?"); 6286 _finger = addr; 6287 _threshold = align_up(_finger, CardTable::card_size); 6288 } 6289 6290 // Should revisit to see if this should be restructured for 6291 // greater efficiency. 6292 bool MarkFromRootsClosure::do_bit(size_t offset) { 6293 if (_skipBits > 0) { 6294 _skipBits--; 6295 return true; 6296 } 6297 // convert offset into a HeapWord* 6298 HeapWord* addr = _bitMap->startWord() + offset; 6299 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6300 "address out of range"); 6301 assert(_bitMap->isMarked(addr), "tautology"); 6302 if (_bitMap->isMarked(addr+1)) { 6303 // this is an allocated but not yet initialized object 6304 assert(_skipBits == 0, "tautology"); 6305 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6306 oop p = oop(addr); 6307 if (p->klass_or_null_acquire() == NULL) { 6308 DEBUG_ONLY(if (!_verifying) {) 6309 // We re-dirty the cards on which this object lies and increase 6310 // the _threshold so that we'll come back to scan this object 6311 // during the preclean or remark phase. (CMSCleanOnEnter) 6312 if (CMSCleanOnEnter) { 6313 size_t sz = _collector->block_size_using_printezis_bits(addr); 6314 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6315 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6316 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6317 // Bump _threshold to end_card_addr; note that 6318 // _threshold cannot possibly exceed end_card_addr, anyhow. 6319 // This prevents future clearing of the card as the scan proceeds 6320 // to the right. 6321 assert(_threshold <= end_card_addr, 6322 "Because we are just scanning into this object"); 6323 if (_threshold < end_card_addr) { 6324 _threshold = end_card_addr; 6325 } 6326 if (p->klass_or_null_acquire() != NULL) { 6327 // Redirty the range of cards... 6328 _mut->mark_range(redirty_range); 6329 } // ...else the setting of klass will dirty the card anyway. 6330 } 6331 DEBUG_ONLY(}) 6332 return true; 6333 } 6334 } 6335 scanOopsInOop(addr); 6336 return true; 6337 } 6338 6339 // We take a break if we've been at this for a while, 6340 // so as to avoid monopolizing the locks involved. 6341 void MarkFromRootsClosure::do_yield_work() { 6342 // First give up the locks, then yield, then re-lock 6343 // We should probably use a constructor/destructor idiom to 6344 // do this unlock/lock or modify the MutexUnlocker class to 6345 // serve our purpose. XXX 6346 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6347 "CMS thread should hold CMS token"); 6348 assert_lock_strong(_bitMap->lock()); 6349 _bitMap->lock()->unlock(); 6350 ConcurrentMarkSweepThread::desynchronize(true); 6351 _collector->stopTimer(); 6352 _collector->incrementYields(); 6353 6354 // See the comment in coordinator_yield() 6355 for (unsigned i = 0; i < CMSYieldSleepCount && 6356 ConcurrentMarkSweepThread::should_yield() && 6357 !CMSCollector::foregroundGCIsActive(); ++i) { 6358 os::sleep(Thread::current(), 1, false); 6359 } 6360 6361 ConcurrentMarkSweepThread::synchronize(true); 6362 _bitMap->lock()->lock_without_safepoint_check(); 6363 _collector->startTimer(); 6364 } 6365 6366 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6367 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6368 assert(_markStack->isEmpty(), 6369 "should drain stack to limit stack usage"); 6370 // convert ptr to an oop preparatory to scanning 6371 oop obj = oop(ptr); 6372 // Ignore mark word in verification below, since we 6373 // may be running concurrent with mutators. 6374 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6375 assert(_finger <= ptr, "_finger runneth ahead"); 6376 // advance the finger to right end of this object 6377 _finger = ptr + obj->size(); 6378 assert(_finger > ptr, "we just incremented it above"); 6379 // On large heaps, it may take us some time to get through 6380 // the marking phase. During 6381 // this time it's possible that a lot of mutations have 6382 // accumulated in the card table and the mod union table -- 6383 // these mutation records are redundant until we have 6384 // actually traced into the corresponding card. 6385 // Here, we check whether advancing the finger would make 6386 // us cross into a new card, and if so clear corresponding 6387 // cards in the MUT (preclean them in the card-table in the 6388 // future). 6389 6390 DEBUG_ONLY(if (!_verifying) {) 6391 // The clean-on-enter optimization is disabled by default, 6392 // until we fix 6178663. 6393 if (CMSCleanOnEnter && (_finger > _threshold)) { 6394 // [_threshold, _finger) represents the interval 6395 // of cards to be cleared in MUT (or precleaned in card table). 6396 // The set of cards to be cleared is all those that overlap 6397 // with the interval [_threshold, _finger); note that 6398 // _threshold is always kept card-aligned but _finger isn't 6399 // always card-aligned. 6400 HeapWord* old_threshold = _threshold; 6401 assert(is_aligned(old_threshold, CardTable::card_size), 6402 "_threshold should always be card-aligned"); 6403 _threshold = align_up(_finger, CardTable::card_size); 6404 MemRegion mr(old_threshold, _threshold); 6405 assert(!mr.is_empty(), "Control point invariant"); 6406 assert(_span.contains(mr), "Should clear within span"); 6407 _mut->clear_range(mr); 6408 } 6409 DEBUG_ONLY(}) 6410 // Note: the finger doesn't advance while we drain 6411 // the stack below. 6412 PushOrMarkClosure pushOrMarkClosure(_collector, 6413 _span, _bitMap, _markStack, 6414 _finger, this); 6415 bool res = _markStack->push(obj); 6416 assert(res, "Empty non-zero size stack should have space for single push"); 6417 while (!_markStack->isEmpty()) { 6418 oop new_oop = _markStack->pop(); 6419 // Skip verifying header mark word below because we are 6420 // running concurrent with mutators. 6421 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6422 // now scan this oop's oops 6423 new_oop->oop_iterate(&pushOrMarkClosure); 6424 do_yield_check(); 6425 } 6426 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6427 } 6428 6429 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6430 CMSCollector* collector, MemRegion span, 6431 CMSBitMap* bit_map, 6432 OopTaskQueue* work_queue, 6433 CMSMarkStack* overflow_stack): 6434 _collector(collector), 6435 _whole_span(collector->_span), 6436 _span(span), 6437 _bit_map(bit_map), 6438 _mut(&collector->_modUnionTable), 6439 _work_queue(work_queue), 6440 _overflow_stack(overflow_stack), 6441 _skip_bits(0), 6442 _task(task) 6443 { 6444 assert(_work_queue->size() == 0, "work_queue should be empty"); 6445 _finger = span.start(); 6446 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6447 assert(_span.contains(_finger), "Out of bounds _finger?"); 6448 } 6449 6450 // Should revisit to see if this should be restructured for 6451 // greater efficiency. 6452 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6453 if (_skip_bits > 0) { 6454 _skip_bits--; 6455 return true; 6456 } 6457 // convert offset into a HeapWord* 6458 HeapWord* addr = _bit_map->startWord() + offset; 6459 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6460 "address out of range"); 6461 assert(_bit_map->isMarked(addr), "tautology"); 6462 if (_bit_map->isMarked(addr+1)) { 6463 // this is an allocated object that might not yet be initialized 6464 assert(_skip_bits == 0, "tautology"); 6465 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6466 oop p = oop(addr); 6467 if (p->klass_or_null_acquire() == NULL) { 6468 // in the case of Clean-on-Enter optimization, redirty card 6469 // and avoid clearing card by increasing the threshold. 6470 return true; 6471 } 6472 } 6473 scan_oops_in_oop(addr); 6474 return true; 6475 } 6476 6477 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6478 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6479 // Should we assert that our work queue is empty or 6480 // below some drain limit? 6481 assert(_work_queue->size() == 0, 6482 "should drain stack to limit stack usage"); 6483 // convert ptr to an oop preparatory to scanning 6484 oop obj = oop(ptr); 6485 // Ignore mark word in verification below, since we 6486 // may be running concurrent with mutators. 6487 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6488 assert(_finger <= ptr, "_finger runneth ahead"); 6489 // advance the finger to right end of this object 6490 _finger = ptr + obj->size(); 6491 assert(_finger > ptr, "we just incremented it above"); 6492 // On large heaps, it may take us some time to get through 6493 // the marking phase. During 6494 // this time it's possible that a lot of mutations have 6495 // accumulated in the card table and the mod union table -- 6496 // these mutation records are redundant until we have 6497 // actually traced into the corresponding card. 6498 // Here, we check whether advancing the finger would make 6499 // us cross into a new card, and if so clear corresponding 6500 // cards in the MUT (preclean them in the card-table in the 6501 // future). 6502 6503 // The clean-on-enter optimization is disabled by default, 6504 // until we fix 6178663. 6505 if (CMSCleanOnEnter && (_finger > _threshold)) { 6506 // [_threshold, _finger) represents the interval 6507 // of cards to be cleared in MUT (or precleaned in card table). 6508 // The set of cards to be cleared is all those that overlap 6509 // with the interval [_threshold, _finger); note that 6510 // _threshold is always kept card-aligned but _finger isn't 6511 // always card-aligned. 6512 HeapWord* old_threshold = _threshold; 6513 assert(is_aligned(old_threshold, CardTable::card_size), 6514 "_threshold should always be card-aligned"); 6515 _threshold = align_up(_finger, CardTable::card_size); 6516 MemRegion mr(old_threshold, _threshold); 6517 assert(!mr.is_empty(), "Control point invariant"); 6518 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6519 _mut->clear_range(mr); 6520 } 6521 6522 // Note: the local finger doesn't advance while we drain 6523 // the stack below, but the global finger sure can and will. 6524 HeapWord* volatile* gfa = _task->global_finger_addr(); 6525 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6526 _span, _bit_map, 6527 _work_queue, 6528 _overflow_stack, 6529 _finger, 6530 gfa, this); 6531 bool res = _work_queue->push(obj); // overflow could occur here 6532 assert(res, "Will hold once we use workqueues"); 6533 while (true) { 6534 oop new_oop; 6535 if (!_work_queue->pop_local(new_oop)) { 6536 // We emptied our work_queue; check if there's stuff that can 6537 // be gotten from the overflow stack. 6538 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6539 _overflow_stack, _work_queue)) { 6540 do_yield_check(); 6541 continue; 6542 } else { // done 6543 break; 6544 } 6545 } 6546 // Skip verifying header mark word below because we are 6547 // running concurrent with mutators. 6548 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6549 // now scan this oop's oops 6550 new_oop->oop_iterate(&pushOrMarkClosure); 6551 do_yield_check(); 6552 } 6553 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6554 } 6555 6556 // Yield in response to a request from VM Thread or 6557 // from mutators. 6558 void ParMarkFromRootsClosure::do_yield_work() { 6559 assert(_task != NULL, "sanity"); 6560 _task->yield(); 6561 } 6562 6563 // A variant of the above used for verifying CMS marking work. 6564 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6565 MemRegion span, 6566 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6567 CMSMarkStack* mark_stack): 6568 _collector(collector), 6569 _span(span), 6570 _verification_bm(verification_bm), 6571 _cms_bm(cms_bm), 6572 _mark_stack(mark_stack), 6573 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6574 mark_stack) 6575 { 6576 assert(_mark_stack->isEmpty(), "stack should be empty"); 6577 _finger = _verification_bm->startWord(); 6578 assert(_collector->_restart_addr == NULL, "Sanity check"); 6579 assert(_span.contains(_finger), "Out of bounds _finger?"); 6580 } 6581 6582 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6583 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6584 assert(_span.contains(addr), "Out of bounds _finger?"); 6585 _finger = addr; 6586 } 6587 6588 // Should revisit to see if this should be restructured for 6589 // greater efficiency. 6590 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6591 // convert offset into a HeapWord* 6592 HeapWord* addr = _verification_bm->startWord() + offset; 6593 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6594 "address out of range"); 6595 assert(_verification_bm->isMarked(addr), "tautology"); 6596 assert(_cms_bm->isMarked(addr), "tautology"); 6597 6598 assert(_mark_stack->isEmpty(), 6599 "should drain stack to limit stack usage"); 6600 // convert addr to an oop preparatory to scanning 6601 oop obj = oop(addr); 6602 assert(oopDesc::is_oop(obj), "should be an oop"); 6603 assert(_finger <= addr, "_finger runneth ahead"); 6604 // advance the finger to right end of this object 6605 _finger = addr + obj->size(); 6606 assert(_finger > addr, "we just incremented it above"); 6607 // Note: the finger doesn't advance while we drain 6608 // the stack below. 6609 bool res = _mark_stack->push(obj); 6610 assert(res, "Empty non-zero size stack should have space for single push"); 6611 while (!_mark_stack->isEmpty()) { 6612 oop new_oop = _mark_stack->pop(); 6613 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6614 // now scan this oop's oops 6615 new_oop->oop_iterate(&_pam_verify_closure); 6616 } 6617 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6618 return true; 6619 } 6620 6621 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6622 CMSCollector* collector, MemRegion span, 6623 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6624 CMSMarkStack* mark_stack): 6625 MetadataAwareOopClosure(collector->ref_processor()), 6626 _collector(collector), 6627 _span(span), 6628 _verification_bm(verification_bm), 6629 _cms_bm(cms_bm), 6630 _mark_stack(mark_stack) 6631 { } 6632 6633 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6634 oop obj = RawAccess<>::oop_load(p); 6635 do_oop(obj); 6636 } 6637 6638 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6639 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6640 6641 // Upon stack overflow, we discard (part of) the stack, 6642 // remembering the least address amongst those discarded 6643 // in CMSCollector's _restart_address. 6644 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6645 // Remember the least grey address discarded 6646 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6647 _collector->lower_restart_addr(ra); 6648 _mark_stack->reset(); // discard stack contents 6649 _mark_stack->expand(); // expand the stack if possible 6650 } 6651 6652 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6653 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6654 HeapWord* addr = (HeapWord*)obj; 6655 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6656 // Oop lies in _span and isn't yet grey or black 6657 _verification_bm->mark(addr); // now grey 6658 if (!_cms_bm->isMarked(addr)) { 6659 Log(gc, verify) log; 6660 ResourceMark rm; 6661 LogStream ls(log.error()); 6662 oop(addr)->print_on(&ls); 6663 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6664 fatal("... aborting"); 6665 } 6666 6667 if (!_mark_stack->push(obj)) { // stack overflow 6668 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6669 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6670 handle_stack_overflow(addr); 6671 } 6672 // anything including and to the right of _finger 6673 // will be scanned as we iterate over the remainder of the 6674 // bit map 6675 } 6676 } 6677 6678 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6679 MemRegion span, 6680 CMSBitMap* bitMap, CMSMarkStack* markStack, 6681 HeapWord* finger, MarkFromRootsClosure* parent) : 6682 MetadataAwareOopClosure(collector->ref_processor()), 6683 _collector(collector), 6684 _span(span), 6685 _bitMap(bitMap), 6686 _markStack(markStack), 6687 _finger(finger), 6688 _parent(parent) 6689 { } 6690 6691 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6692 MemRegion span, 6693 CMSBitMap* bit_map, 6694 OopTaskQueue* work_queue, 6695 CMSMarkStack* overflow_stack, 6696 HeapWord* finger, 6697 HeapWord* volatile* global_finger_addr, 6698 ParMarkFromRootsClosure* parent) : 6699 MetadataAwareOopClosure(collector->ref_processor()), 6700 _collector(collector), 6701 _whole_span(collector->_span), 6702 _span(span), 6703 _bit_map(bit_map), 6704 _work_queue(work_queue), 6705 _overflow_stack(overflow_stack), 6706 _finger(finger), 6707 _global_finger_addr(global_finger_addr), 6708 _parent(parent) 6709 { } 6710 6711 // Assumes thread-safe access by callers, who are 6712 // responsible for mutual exclusion. 6713 void CMSCollector::lower_restart_addr(HeapWord* low) { 6714 assert(_span.contains(low), "Out of bounds addr"); 6715 if (_restart_addr == NULL) { 6716 _restart_addr = low; 6717 } else { 6718 _restart_addr = MIN2(_restart_addr, low); 6719 } 6720 } 6721 6722 // Upon stack overflow, we discard (part of) the stack, 6723 // remembering the least address amongst those discarded 6724 // in CMSCollector's _restart_address. 6725 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6726 // Remember the least grey address discarded 6727 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6728 _collector->lower_restart_addr(ra); 6729 _markStack->reset(); // discard stack contents 6730 _markStack->expand(); // expand the stack if possible 6731 } 6732 6733 // Upon stack overflow, we discard (part of) the stack, 6734 // remembering the least address amongst those discarded 6735 // in CMSCollector's _restart_address. 6736 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6737 // We need to do this under a mutex to prevent other 6738 // workers from interfering with the work done below. 6739 MutexLockerEx ml(_overflow_stack->par_lock(), 6740 Mutex::_no_safepoint_check_flag); 6741 // Remember the least grey address discarded 6742 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6743 _collector->lower_restart_addr(ra); 6744 _overflow_stack->reset(); // discard stack contents 6745 _overflow_stack->expand(); // expand the stack if possible 6746 } 6747 6748 void PushOrMarkClosure::do_oop(oop obj) { 6749 // Ignore mark word because we are running concurrent with mutators. 6750 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6751 HeapWord* addr = (HeapWord*)obj; 6752 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6753 // Oop lies in _span and isn't yet grey or black 6754 _bitMap->mark(addr); // now grey 6755 if (addr < _finger) { 6756 // the bit map iteration has already either passed, or 6757 // sampled, this bit in the bit map; we'll need to 6758 // use the marking stack to scan this oop's oops. 6759 bool simulate_overflow = false; 6760 NOT_PRODUCT( 6761 if (CMSMarkStackOverflowALot && 6762 _collector->simulate_overflow()) { 6763 // simulate a stack overflow 6764 simulate_overflow = true; 6765 } 6766 ) 6767 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6768 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6769 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6770 handle_stack_overflow(addr); 6771 } 6772 } 6773 // anything including and to the right of _finger 6774 // will be scanned as we iterate over the remainder of the 6775 // bit map 6776 do_yield_check(); 6777 } 6778 } 6779 6780 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6781 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6782 6783 void ParPushOrMarkClosure::do_oop(oop obj) { 6784 // Ignore mark word because we are running concurrent with mutators. 6785 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6786 HeapWord* addr = (HeapWord*)obj; 6787 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6788 // Oop lies in _span and isn't yet grey or black 6789 // We read the global_finger (volatile read) strictly after marking oop 6790 bool res = _bit_map->par_mark(addr); // now grey 6791 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6792 // Should we push this marked oop on our stack? 6793 // -- if someone else marked it, nothing to do 6794 // -- if target oop is above global finger nothing to do 6795 // -- if target oop is in chunk and above local finger 6796 // then nothing to do 6797 // -- else push on work queue 6798 if ( !res // someone else marked it, they will deal with it 6799 || (addr >= *gfa) // will be scanned in a later task 6800 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6801 return; 6802 } 6803 // the bit map iteration has already either passed, or 6804 // sampled, this bit in the bit map; we'll need to 6805 // use the marking stack to scan this oop's oops. 6806 bool simulate_overflow = false; 6807 NOT_PRODUCT( 6808 if (CMSMarkStackOverflowALot && 6809 _collector->simulate_overflow()) { 6810 // simulate a stack overflow 6811 simulate_overflow = true; 6812 } 6813 ) 6814 if (simulate_overflow || 6815 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6816 // stack overflow 6817 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6818 // We cannot assert that the overflow stack is full because 6819 // it may have been emptied since. 6820 assert(simulate_overflow || 6821 _work_queue->size() == _work_queue->max_elems(), 6822 "Else push should have succeeded"); 6823 handle_stack_overflow(addr); 6824 } 6825 do_yield_check(); 6826 } 6827 } 6828 6829 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6830 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6831 6832 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6833 MemRegion span, 6834 ReferenceDiscoverer* rd, 6835 CMSBitMap* bit_map, 6836 CMSBitMap* mod_union_table, 6837 CMSMarkStack* mark_stack, 6838 bool concurrent_precleaning): 6839 MetadataAwareOopClosure(rd), 6840 _collector(collector), 6841 _span(span), 6842 _bit_map(bit_map), 6843 _mod_union_table(mod_union_table), 6844 _mark_stack(mark_stack), 6845 _concurrent_precleaning(concurrent_precleaning) 6846 { 6847 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6848 } 6849 6850 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6851 // the non-parallel version (the parallel version appears further below.) 6852 void PushAndMarkClosure::do_oop(oop obj) { 6853 // Ignore mark word verification. If during concurrent precleaning, 6854 // the object monitor may be locked. If during the checkpoint 6855 // phases, the object may already have been reached by a different 6856 // path and may be at the end of the global overflow list (so 6857 // the mark word may be NULL). 6858 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6859 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6860 HeapWord* addr = (HeapWord*)obj; 6861 // Check if oop points into the CMS generation 6862 // and is not marked 6863 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6864 // a white object ... 6865 _bit_map->mark(addr); // ... now grey 6866 // push on the marking stack (grey set) 6867 bool simulate_overflow = false; 6868 NOT_PRODUCT( 6869 if (CMSMarkStackOverflowALot && 6870 _collector->simulate_overflow()) { 6871 // simulate a stack overflow 6872 simulate_overflow = true; 6873 } 6874 ) 6875 if (simulate_overflow || !_mark_stack->push(obj)) { 6876 if (_concurrent_precleaning) { 6877 // During precleaning we can just dirty the appropriate card(s) 6878 // in the mod union table, thus ensuring that the object remains 6879 // in the grey set and continue. In the case of object arrays 6880 // we need to dirty all of the cards that the object spans, 6881 // since the rescan of object arrays will be limited to the 6882 // dirty cards. 6883 // Note that no one can be interfering with us in this action 6884 // of dirtying the mod union table, so no locking or atomics 6885 // are required. 6886 if (obj->is_objArray()) { 6887 size_t sz = obj->size(); 6888 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6889 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6890 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6891 _mod_union_table->mark_range(redirty_range); 6892 } else { 6893 _mod_union_table->mark(addr); 6894 } 6895 _collector->_ser_pmc_preclean_ovflw++; 6896 } else { 6897 // During the remark phase, we need to remember this oop 6898 // in the overflow list. 6899 _collector->push_on_overflow_list(obj); 6900 _collector->_ser_pmc_remark_ovflw++; 6901 } 6902 } 6903 } 6904 } 6905 6906 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6907 MemRegion span, 6908 ReferenceDiscoverer* rd, 6909 CMSBitMap* bit_map, 6910 OopTaskQueue* work_queue): 6911 MetadataAwareOopClosure(rd), 6912 _collector(collector), 6913 _span(span), 6914 _bit_map(bit_map), 6915 _work_queue(work_queue) 6916 { 6917 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6918 } 6919 6920 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6921 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6922 6923 // Grey object rescan during second checkpoint phase -- 6924 // the parallel version. 6925 void ParPushAndMarkClosure::do_oop(oop obj) { 6926 // In the assert below, we ignore the mark word because 6927 // this oop may point to an already visited object that is 6928 // on the overflow stack (in which case the mark word has 6929 // been hijacked for chaining into the overflow stack -- 6930 // if this is the last object in the overflow stack then 6931 // its mark word will be NULL). Because this object may 6932 // have been subsequently popped off the global overflow 6933 // stack, and the mark word possibly restored to the prototypical 6934 // value, by the time we get to examined this failing assert in 6935 // the debugger, is_oop_or_null(false) may subsequently start 6936 // to hold. 6937 assert(oopDesc::is_oop_or_null(obj, true), 6938 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6939 HeapWord* addr = (HeapWord*)obj; 6940 // Check if oop points into the CMS generation 6941 // and is not marked 6942 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6943 // a white object ... 6944 // If we manage to "claim" the object, by being the 6945 // first thread to mark it, then we push it on our 6946 // marking stack 6947 if (_bit_map->par_mark(addr)) { // ... now grey 6948 // push on work queue (grey set) 6949 bool simulate_overflow = false; 6950 NOT_PRODUCT( 6951 if (CMSMarkStackOverflowALot && 6952 _collector->par_simulate_overflow()) { 6953 // simulate a stack overflow 6954 simulate_overflow = true; 6955 } 6956 ) 6957 if (simulate_overflow || !_work_queue->push(obj)) { 6958 _collector->par_push_on_overflow_list(obj); 6959 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6960 } 6961 } // Else, some other thread got there first 6962 } 6963 } 6964 6965 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6966 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6967 6968 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6969 Mutex* bml = _collector->bitMapLock(); 6970 assert_lock_strong(bml); 6971 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6972 "CMS thread should hold CMS token"); 6973 6974 bml->unlock(); 6975 ConcurrentMarkSweepThread::desynchronize(true); 6976 6977 _collector->stopTimer(); 6978 _collector->incrementYields(); 6979 6980 // See the comment in coordinator_yield() 6981 for (unsigned i = 0; i < CMSYieldSleepCount && 6982 ConcurrentMarkSweepThread::should_yield() && 6983 !CMSCollector::foregroundGCIsActive(); ++i) { 6984 os::sleep(Thread::current(), 1, false); 6985 } 6986 6987 ConcurrentMarkSweepThread::synchronize(true); 6988 bml->lock(); 6989 6990 _collector->startTimer(); 6991 } 6992 6993 bool CMSPrecleanRefsYieldClosure::should_return() { 6994 if (ConcurrentMarkSweepThread::should_yield()) { 6995 do_yield_work(); 6996 } 6997 return _collector->foregroundGCIsActive(); 6998 } 6999 7000 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7001 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7002 "mr should be aligned to start at a card boundary"); 7003 // We'd like to assert: 7004 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7005 // "mr should be a range of cards"); 7006 // However, that would be too strong in one case -- the last 7007 // partition ends at _unallocated_block which, in general, can be 7008 // an arbitrary boundary, not necessarily card aligned. 7009 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7010 _space->object_iterate_mem(mr, &_scan_cl); 7011 } 7012 7013 SweepClosure::SweepClosure(CMSCollector* collector, 7014 ConcurrentMarkSweepGeneration* g, 7015 CMSBitMap* bitMap, bool should_yield) : 7016 _collector(collector), 7017 _g(g), 7018 _sp(g->cmsSpace()), 7019 _limit(_sp->sweep_limit()), 7020 _freelistLock(_sp->freelistLock()), 7021 _bitMap(bitMap), 7022 _yield(should_yield), 7023 _inFreeRange(false), // No free range at beginning of sweep 7024 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7025 _lastFreeRangeCoalesced(false), 7026 _freeFinger(g->used_region().start()) 7027 { 7028 NOT_PRODUCT( 7029 _numObjectsFreed = 0; 7030 _numWordsFreed = 0; 7031 _numObjectsLive = 0; 7032 _numWordsLive = 0; 7033 _numObjectsAlreadyFree = 0; 7034 _numWordsAlreadyFree = 0; 7035 _last_fc = NULL; 7036 7037 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7038 _sp->dictionary()->initialize_dict_returned_bytes(); 7039 ) 7040 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7041 "sweep _limit out of bounds"); 7042 log_develop_trace(gc, sweep)("===================="); 7043 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7044 } 7045 7046 void SweepClosure::print_on(outputStream* st) const { 7047 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7048 p2i(_sp->bottom()), p2i(_sp->end())); 7049 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7050 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7051 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7052 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7053 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7054 } 7055 7056 #ifndef PRODUCT 7057 // Assertion checking only: no useful work in product mode -- 7058 // however, if any of the flags below become product flags, 7059 // you may need to review this code to see if it needs to be 7060 // enabled in product mode. 7061 SweepClosure::~SweepClosure() { 7062 assert_lock_strong(_freelistLock); 7063 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7064 "sweep _limit out of bounds"); 7065 if (inFreeRange()) { 7066 Log(gc, sweep) log; 7067 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7068 ResourceMark rm; 7069 LogStream ls(log.error()); 7070 print_on(&ls); 7071 ShouldNotReachHere(); 7072 } 7073 7074 if (log_is_enabled(Debug, gc, sweep)) { 7075 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7076 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7077 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7078 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7079 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7080 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7081 } 7082 7083 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7084 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7085 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7086 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7087 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7088 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7089 } 7090 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7091 log_develop_trace(gc, sweep)("================"); 7092 } 7093 #endif // PRODUCT 7094 7095 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7096 bool freeRangeInFreeLists) { 7097 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7098 p2i(freeFinger), freeRangeInFreeLists); 7099 assert(!inFreeRange(), "Trampling existing free range"); 7100 set_inFreeRange(true); 7101 set_lastFreeRangeCoalesced(false); 7102 7103 set_freeFinger(freeFinger); 7104 set_freeRangeInFreeLists(freeRangeInFreeLists); 7105 if (CMSTestInFreeList) { 7106 if (freeRangeInFreeLists) { 7107 FreeChunk* fc = (FreeChunk*) freeFinger; 7108 assert(fc->is_free(), "A chunk on the free list should be free."); 7109 assert(fc->size() > 0, "Free range should have a size"); 7110 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7111 } 7112 } 7113 } 7114 7115 // Note that the sweeper runs concurrently with mutators. Thus, 7116 // it is possible for direct allocation in this generation to happen 7117 // in the middle of the sweep. Note that the sweeper also coalesces 7118 // contiguous free blocks. Thus, unless the sweeper and the allocator 7119 // synchronize appropriately freshly allocated blocks may get swept up. 7120 // This is accomplished by the sweeper locking the free lists while 7121 // it is sweeping. Thus blocks that are determined to be free are 7122 // indeed free. There is however one additional complication: 7123 // blocks that have been allocated since the final checkpoint and 7124 // mark, will not have been marked and so would be treated as 7125 // unreachable and swept up. To prevent this, the allocator marks 7126 // the bit map when allocating during the sweep phase. This leads, 7127 // however, to a further complication -- objects may have been allocated 7128 // but not yet initialized -- in the sense that the header isn't yet 7129 // installed. The sweeper can not then determine the size of the block 7130 // in order to skip over it. To deal with this case, we use a technique 7131 // (due to Printezis) to encode such uninitialized block sizes in the 7132 // bit map. Since the bit map uses a bit per every HeapWord, but the 7133 // CMS generation has a minimum object size of 3 HeapWords, it follows 7134 // that "normal marks" won't be adjacent in the bit map (there will 7135 // always be at least two 0 bits between successive 1 bits). We make use 7136 // of these "unused" bits to represent uninitialized blocks -- the bit 7137 // corresponding to the start of the uninitialized object and the next 7138 // bit are both set. Finally, a 1 bit marks the end of the object that 7139 // started with the two consecutive 1 bits to indicate its potentially 7140 // uninitialized state. 7141 7142 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7143 FreeChunk* fc = (FreeChunk*)addr; 7144 size_t res; 7145 7146 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7147 // than "addr == _limit" because although _limit was a block boundary when 7148 // we started the sweep, it may no longer be one because heap expansion 7149 // may have caused us to coalesce the block ending at the address _limit 7150 // with a newly expanded chunk (this happens when _limit was set to the 7151 // previous _end of the space), so we may have stepped past _limit: 7152 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7153 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7154 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7155 "sweep _limit out of bounds"); 7156 assert(addr < _sp->end(), "addr out of bounds"); 7157 // Flush any free range we might be holding as a single 7158 // coalesced chunk to the appropriate free list. 7159 if (inFreeRange()) { 7160 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7161 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7162 flush_cur_free_chunk(freeFinger(), 7163 pointer_delta(addr, freeFinger())); 7164 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7165 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7166 lastFreeRangeCoalesced() ? 1 : 0); 7167 } 7168 7169 // help the iterator loop finish 7170 return pointer_delta(_sp->end(), addr); 7171 } 7172 7173 assert(addr < _limit, "sweep invariant"); 7174 // check if we should yield 7175 do_yield_check(addr); 7176 if (fc->is_free()) { 7177 // Chunk that is already free 7178 res = fc->size(); 7179 do_already_free_chunk(fc); 7180 debug_only(_sp->verifyFreeLists()); 7181 // If we flush the chunk at hand in lookahead_and_flush() 7182 // and it's coalesced with a preceding chunk, then the 7183 // process of "mangling" the payload of the coalesced block 7184 // will cause erasure of the size information from the 7185 // (erstwhile) header of all the coalesced blocks but the 7186 // first, so the first disjunct in the assert will not hold 7187 // in that specific case (in which case the second disjunct 7188 // will hold). 7189 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7190 "Otherwise the size info doesn't change at this step"); 7191 NOT_PRODUCT( 7192 _numObjectsAlreadyFree++; 7193 _numWordsAlreadyFree += res; 7194 ) 7195 NOT_PRODUCT(_last_fc = fc;) 7196 } else if (!_bitMap->isMarked(addr)) { 7197 // Chunk is fresh garbage 7198 res = do_garbage_chunk(fc); 7199 debug_only(_sp->verifyFreeLists()); 7200 NOT_PRODUCT( 7201 _numObjectsFreed++; 7202 _numWordsFreed += res; 7203 ) 7204 } else { 7205 // Chunk that is alive. 7206 res = do_live_chunk(fc); 7207 debug_only(_sp->verifyFreeLists()); 7208 NOT_PRODUCT( 7209 _numObjectsLive++; 7210 _numWordsLive += res; 7211 ) 7212 } 7213 return res; 7214 } 7215 7216 // For the smart allocation, record following 7217 // split deaths - a free chunk is removed from its free list because 7218 // it is being split into two or more chunks. 7219 // split birth - a free chunk is being added to its free list because 7220 // a larger free chunk has been split and resulted in this free chunk. 7221 // coal death - a free chunk is being removed from its free list because 7222 // it is being coalesced into a large free chunk. 7223 // coal birth - a free chunk is being added to its free list because 7224 // it was created when two or more free chunks where coalesced into 7225 // this free chunk. 7226 // 7227 // These statistics are used to determine the desired number of free 7228 // chunks of a given size. The desired number is chosen to be relative 7229 // to the end of a CMS sweep. The desired number at the end of a sweep 7230 // is the 7231 // count-at-end-of-previous-sweep (an amount that was enough) 7232 // - count-at-beginning-of-current-sweep (the excess) 7233 // + split-births (gains in this size during interval) 7234 // - split-deaths (demands on this size during interval) 7235 // where the interval is from the end of one sweep to the end of the 7236 // next. 7237 // 7238 // When sweeping the sweeper maintains an accumulated chunk which is 7239 // the chunk that is made up of chunks that have been coalesced. That 7240 // will be termed the left-hand chunk. A new chunk of garbage that 7241 // is being considered for coalescing will be referred to as the 7242 // right-hand chunk. 7243 // 7244 // When making a decision on whether to coalesce a right-hand chunk with 7245 // the current left-hand chunk, the current count vs. the desired count 7246 // of the left-hand chunk is considered. Also if the right-hand chunk 7247 // is near the large chunk at the end of the heap (see 7248 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7249 // left-hand chunk is coalesced. 7250 // 7251 // When making a decision about whether to split a chunk, the desired count 7252 // vs. the current count of the candidate to be split is also considered. 7253 // If the candidate is underpopulated (currently fewer chunks than desired) 7254 // a chunk of an overpopulated (currently more chunks than desired) size may 7255 // be chosen. The "hint" associated with a free list, if non-null, points 7256 // to a free list which may be overpopulated. 7257 // 7258 7259 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7260 const size_t size = fc->size(); 7261 // Chunks that cannot be coalesced are not in the 7262 // free lists. 7263 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7264 assert(_sp->verify_chunk_in_free_list(fc), 7265 "free chunk should be in free lists"); 7266 } 7267 // a chunk that is already free, should not have been 7268 // marked in the bit map 7269 HeapWord* const addr = (HeapWord*) fc; 7270 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7271 // Verify that the bit map has no bits marked between 7272 // addr and purported end of this block. 7273 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7274 7275 // Some chunks cannot be coalesced under any circumstances. 7276 // See the definition of cantCoalesce(). 7277 if (!fc->cantCoalesce()) { 7278 // This chunk can potentially be coalesced. 7279 // All the work is done in 7280 do_post_free_or_garbage_chunk(fc, size); 7281 // Note that if the chunk is not coalescable (the else arm 7282 // below), we unconditionally flush, without needing to do 7283 // a "lookahead," as we do below. 7284 if (inFreeRange()) lookahead_and_flush(fc, size); 7285 } else { 7286 // Code path common to both original and adaptive free lists. 7287 7288 // cant coalesce with previous block; this should be treated 7289 // as the end of a free run if any 7290 if (inFreeRange()) { 7291 // we kicked some butt; time to pick up the garbage 7292 assert(freeFinger() < addr, "freeFinger points too high"); 7293 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7294 } 7295 // else, nothing to do, just continue 7296 } 7297 } 7298 7299 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7300 // This is a chunk of garbage. It is not in any free list. 7301 // Add it to a free list or let it possibly be coalesced into 7302 // a larger chunk. 7303 HeapWord* const addr = (HeapWord*) fc; 7304 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7305 7306 // Verify that the bit map has no bits marked between 7307 // addr and purported end of just dead object. 7308 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7309 do_post_free_or_garbage_chunk(fc, size); 7310 7311 assert(_limit >= addr + size, 7312 "A freshly garbage chunk can't possibly straddle over _limit"); 7313 if (inFreeRange()) lookahead_and_flush(fc, size); 7314 return size; 7315 } 7316 7317 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7318 HeapWord* addr = (HeapWord*) fc; 7319 // The sweeper has just found a live object. Return any accumulated 7320 // left hand chunk to the free lists. 7321 if (inFreeRange()) { 7322 assert(freeFinger() < addr, "freeFinger points too high"); 7323 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7324 } 7325 7326 // This object is live: we'd normally expect this to be 7327 // an oop, and like to assert the following: 7328 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7329 // However, as we commented above, this may be an object whose 7330 // header hasn't yet been initialized. 7331 size_t size; 7332 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7333 if (_bitMap->isMarked(addr + 1)) { 7334 // Determine the size from the bit map, rather than trying to 7335 // compute it from the object header. 7336 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7337 size = pointer_delta(nextOneAddr + 1, addr); 7338 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7339 "alignment problem"); 7340 7341 #ifdef ASSERT 7342 if (oop(addr)->klass_or_null_acquire() != NULL) { 7343 // Ignore mark word because we are running concurrent with mutators 7344 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7345 assert(size == 7346 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7347 "P-mark and computed size do not agree"); 7348 } 7349 #endif 7350 7351 } else { 7352 // This should be an initialized object that's alive. 7353 assert(oop(addr)->klass_or_null_acquire() != NULL, 7354 "Should be an initialized object"); 7355 // Ignore mark word because we are running concurrent with mutators 7356 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7357 // Verify that the bit map has no bits marked between 7358 // addr and purported end of this block. 7359 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7360 assert(size >= 3, "Necessary for Printezis marks to work"); 7361 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7362 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7363 } 7364 return size; 7365 } 7366 7367 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7368 size_t chunkSize) { 7369 // do_post_free_or_garbage_chunk() should only be called in the case 7370 // of the adaptive free list allocator. 7371 const bool fcInFreeLists = fc->is_free(); 7372 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7373 if (CMSTestInFreeList && fcInFreeLists) { 7374 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7375 } 7376 7377 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7378 7379 HeapWord* const fc_addr = (HeapWord*) fc; 7380 7381 bool coalesce = false; 7382 const size_t left = pointer_delta(fc_addr, freeFinger()); 7383 const size_t right = chunkSize; 7384 switch (FLSCoalescePolicy) { 7385 // numeric value forms a coalition aggressiveness metric 7386 case 0: { // never coalesce 7387 coalesce = false; 7388 break; 7389 } 7390 case 1: { // coalesce if left & right chunks on overpopulated lists 7391 coalesce = _sp->coalOverPopulated(left) && 7392 _sp->coalOverPopulated(right); 7393 break; 7394 } 7395 case 2: { // coalesce if left chunk on overpopulated list (default) 7396 coalesce = _sp->coalOverPopulated(left); 7397 break; 7398 } 7399 case 3: { // coalesce if left OR right chunk on overpopulated list 7400 coalesce = _sp->coalOverPopulated(left) || 7401 _sp->coalOverPopulated(right); 7402 break; 7403 } 7404 case 4: { // always coalesce 7405 coalesce = true; 7406 break; 7407 } 7408 default: 7409 ShouldNotReachHere(); 7410 } 7411 7412 // Should the current free range be coalesced? 7413 // If the chunk is in a free range and either we decided to coalesce above 7414 // or the chunk is near the large block at the end of the heap 7415 // (isNearLargestChunk() returns true), then coalesce this chunk. 7416 const bool doCoalesce = inFreeRange() 7417 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7418 if (doCoalesce) { 7419 // Coalesce the current free range on the left with the new 7420 // chunk on the right. If either is on a free list, 7421 // it must be removed from the list and stashed in the closure. 7422 if (freeRangeInFreeLists()) { 7423 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7424 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7425 "Size of free range is inconsistent with chunk size."); 7426 if (CMSTestInFreeList) { 7427 assert(_sp->verify_chunk_in_free_list(ffc), 7428 "Chunk is not in free lists"); 7429 } 7430 _sp->coalDeath(ffc->size()); 7431 _sp->removeFreeChunkFromFreeLists(ffc); 7432 set_freeRangeInFreeLists(false); 7433 } 7434 if (fcInFreeLists) { 7435 _sp->coalDeath(chunkSize); 7436 assert(fc->size() == chunkSize, 7437 "The chunk has the wrong size or is not in the free lists"); 7438 _sp->removeFreeChunkFromFreeLists(fc); 7439 } 7440 set_lastFreeRangeCoalesced(true); 7441 print_free_block_coalesced(fc); 7442 } else { // not in a free range and/or should not coalesce 7443 // Return the current free range and start a new one. 7444 if (inFreeRange()) { 7445 // In a free range but cannot coalesce with the right hand chunk. 7446 // Put the current free range into the free lists. 7447 flush_cur_free_chunk(freeFinger(), 7448 pointer_delta(fc_addr, freeFinger())); 7449 } 7450 // Set up for new free range. Pass along whether the right hand 7451 // chunk is in the free lists. 7452 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7453 } 7454 } 7455 7456 // Lookahead flush: 7457 // If we are tracking a free range, and this is the last chunk that 7458 // we'll look at because its end crosses past _limit, we'll preemptively 7459 // flush it along with any free range we may be holding on to. Note that 7460 // this can be the case only for an already free or freshly garbage 7461 // chunk. If this block is an object, it can never straddle 7462 // over _limit. The "straddling" occurs when _limit is set at 7463 // the previous end of the space when this cycle started, and 7464 // a subsequent heap expansion caused the previously co-terminal 7465 // free block to be coalesced with the newly expanded portion, 7466 // thus rendering _limit a non-block-boundary making it dangerous 7467 // for the sweeper to step over and examine. 7468 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7469 assert(inFreeRange(), "Should only be called if currently in a free range."); 7470 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7471 assert(_sp->used_region().contains(eob - 1), 7472 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7473 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7474 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7475 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7476 if (eob >= _limit) { 7477 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7478 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7479 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7480 "[" PTR_FORMAT "," PTR_FORMAT ")", 7481 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7482 // Return the storage we are tracking back into the free lists. 7483 log_develop_trace(gc, sweep)("Flushing ... "); 7484 assert(freeFinger() < eob, "Error"); 7485 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7486 } 7487 } 7488 7489 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7490 assert(inFreeRange(), "Should only be called if currently in a free range."); 7491 assert(size > 0, 7492 "A zero sized chunk cannot be added to the free lists."); 7493 if (!freeRangeInFreeLists()) { 7494 if (CMSTestInFreeList) { 7495 FreeChunk* fc = (FreeChunk*) chunk; 7496 fc->set_size(size); 7497 assert(!_sp->verify_chunk_in_free_list(fc), 7498 "chunk should not be in free lists yet"); 7499 } 7500 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7501 // A new free range is going to be starting. The current 7502 // free range has not been added to the free lists yet or 7503 // was removed so add it back. 7504 // If the current free range was coalesced, then the death 7505 // of the free range was recorded. Record a birth now. 7506 if (lastFreeRangeCoalesced()) { 7507 _sp->coalBirth(size); 7508 } 7509 _sp->addChunkAndRepairOffsetTable(chunk, size, 7510 lastFreeRangeCoalesced()); 7511 } else { 7512 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7513 } 7514 set_inFreeRange(false); 7515 set_freeRangeInFreeLists(false); 7516 } 7517 7518 // We take a break if we've been at this for a while, 7519 // so as to avoid monopolizing the locks involved. 7520 void SweepClosure::do_yield_work(HeapWord* addr) { 7521 // Return current free chunk being used for coalescing (if any) 7522 // to the appropriate freelist. After yielding, the next 7523 // free block encountered will start a coalescing range of 7524 // free blocks. If the next free block is adjacent to the 7525 // chunk just flushed, they will need to wait for the next 7526 // sweep to be coalesced. 7527 if (inFreeRange()) { 7528 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7529 } 7530 7531 // First give up the locks, then yield, then re-lock. 7532 // We should probably use a constructor/destructor idiom to 7533 // do this unlock/lock or modify the MutexUnlocker class to 7534 // serve our purpose. XXX 7535 assert_lock_strong(_bitMap->lock()); 7536 assert_lock_strong(_freelistLock); 7537 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7538 "CMS thread should hold CMS token"); 7539 _bitMap->lock()->unlock(); 7540 _freelistLock->unlock(); 7541 ConcurrentMarkSweepThread::desynchronize(true); 7542 _collector->stopTimer(); 7543 _collector->incrementYields(); 7544 7545 // See the comment in coordinator_yield() 7546 for (unsigned i = 0; i < CMSYieldSleepCount && 7547 ConcurrentMarkSweepThread::should_yield() && 7548 !CMSCollector::foregroundGCIsActive(); ++i) { 7549 os::sleep(Thread::current(), 1, false); 7550 } 7551 7552 ConcurrentMarkSweepThread::synchronize(true); 7553 _freelistLock->lock(); 7554 _bitMap->lock()->lock_without_safepoint_check(); 7555 _collector->startTimer(); 7556 } 7557 7558 #ifndef PRODUCT 7559 // This is actually very useful in a product build if it can 7560 // be called from the debugger. Compile it into the product 7561 // as needed. 7562 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7563 return debug_cms_space->verify_chunk_in_free_list(fc); 7564 } 7565 #endif 7566 7567 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7568 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7569 p2i(fc), fc->size()); 7570 } 7571 7572 // CMSIsAliveClosure 7573 bool CMSIsAliveClosure::do_object_b(oop obj) { 7574 HeapWord* addr = (HeapWord*)obj; 7575 return addr != NULL && 7576 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7577 } 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 _bit_map(bit_map), 7587 _mark_stack(mark_stack), 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 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7635 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7636 7637 // CMSParKeepAliveClosure: a parallel version of the above. 7638 // The work queues are private to each closure (thread), 7639 // but (may be) available for stealing by other threads. 7640 void CMSParKeepAliveClosure::do_oop(oop obj) { 7641 HeapWord* addr = (HeapWord*)obj; 7642 if (_span.contains(addr) && 7643 !_bit_map->isMarked(addr)) { 7644 // In general, during recursive tracing, several threads 7645 // may be concurrently getting here; the first one to 7646 // "tag" it, claims it. 7647 if (_bit_map->par_mark(addr)) { 7648 bool res = _work_queue->push(obj); 7649 assert(res, "Low water mark should be much less than capacity"); 7650 // Do a recursive trim in the hope that this will keep 7651 // stack usage lower, but leave some oops for potential stealers 7652 trim_queue(_low_water_mark); 7653 } // Else, another thread got there first 7654 } 7655 } 7656 7657 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7658 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7659 7660 void CMSParKeepAliveClosure::trim_queue(uint max) { 7661 while (_work_queue->size() > max) { 7662 oop new_oop; 7663 if (_work_queue->pop_local(new_oop)) { 7664 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7665 assert(_bit_map->isMarked((HeapWord*)new_oop), 7666 "no white objects on this stack!"); 7667 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7668 // iterate over the oops in this oop, marking and pushing 7669 // the ones in CMS heap (i.e. in _span). 7670 new_oop->oop_iterate(&_mark_and_push); 7671 } 7672 } 7673 } 7674 7675 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7676 CMSCollector* collector, 7677 MemRegion span, CMSBitMap* bit_map, 7678 OopTaskQueue* work_queue): 7679 _collector(collector), 7680 _span(span), 7681 _bit_map(bit_map), 7682 _work_queue(work_queue) { } 7683 7684 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7685 HeapWord* addr = (HeapWord*)obj; 7686 if (_span.contains(addr) && 7687 !_bit_map->isMarked(addr)) { 7688 if (_bit_map->par_mark(addr)) { 7689 bool simulate_overflow = false; 7690 NOT_PRODUCT( 7691 if (CMSMarkStackOverflowALot && 7692 _collector->par_simulate_overflow()) { 7693 // simulate a stack overflow 7694 simulate_overflow = true; 7695 } 7696 ) 7697 if (simulate_overflow || !_work_queue->push(obj)) { 7698 _collector->par_push_on_overflow_list(obj); 7699 _collector->_par_kac_ovflw++; 7700 } 7701 } // Else another thread got there already 7702 } 7703 } 7704 7705 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7706 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7707 7708 ////////////////////////////////////////////////////////////////// 7709 // CMSExpansionCause ///////////////////////////// 7710 ////////////////////////////////////////////////////////////////// 7711 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7712 switch (cause) { 7713 case _no_expansion: 7714 return "No expansion"; 7715 case _satisfy_free_ratio: 7716 return "Free ratio"; 7717 case _satisfy_promotion: 7718 return "Satisfy promotion"; 7719 case _satisfy_allocation: 7720 return "allocation"; 7721 case _allocate_par_lab: 7722 return "Par LAB"; 7723 case _allocate_par_spooling_space: 7724 return "Par Spooling Space"; 7725 case _adaptive_size_policy: 7726 return "Ergonomics"; 7727 default: 7728 return "unknown"; 7729 } 7730 } 7731 7732 void CMSDrainMarkingStackClosure::do_void() { 7733 // the max number to take from overflow list at a time 7734 const size_t num = _mark_stack->capacity()/4; 7735 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7736 "Overflow list should be NULL during concurrent phases"); 7737 while (!_mark_stack->isEmpty() || 7738 // if stack is empty, check the overflow list 7739 _collector->take_from_overflow_list(num, _mark_stack)) { 7740 oop obj = _mark_stack->pop(); 7741 HeapWord* addr = (HeapWord*)obj; 7742 assert(_span.contains(addr), "Should be within span"); 7743 assert(_bit_map->isMarked(addr), "Should be marked"); 7744 assert(oopDesc::is_oop(obj), "Should be an oop"); 7745 obj->oop_iterate(_keep_alive); 7746 } 7747 } 7748 7749 void CMSParDrainMarkingStackClosure::do_void() { 7750 // drain queue 7751 trim_queue(0); 7752 } 7753 7754 // Trim our work_queue so its length is below max at return 7755 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7756 while (_work_queue->size() > max) { 7757 oop new_oop; 7758 if (_work_queue->pop_local(new_oop)) { 7759 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7760 assert(_bit_map->isMarked((HeapWord*)new_oop), 7761 "no white objects on this stack!"); 7762 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7763 // iterate over the oops in this oop, marking and pushing 7764 // the ones in CMS heap (i.e. in _span). 7765 new_oop->oop_iterate(&_mark_and_push); 7766 } 7767 } 7768 } 7769 7770 //////////////////////////////////////////////////////////////////// 7771 // Support for Marking Stack Overflow list handling and related code 7772 //////////////////////////////////////////////////////////////////// 7773 // Much of the following code is similar in shape and spirit to the 7774 // code used in ParNewGC. We should try and share that code 7775 // as much as possible in the future. 7776 7777 #ifndef PRODUCT 7778 // Debugging support for CMSStackOverflowALot 7779 7780 // It's OK to call this multi-threaded; the worst thing 7781 // that can happen is that we'll get a bunch of closely 7782 // spaced simulated overflows, but that's OK, in fact 7783 // probably good as it would exercise the overflow code 7784 // under contention. 7785 bool CMSCollector::simulate_overflow() { 7786 if (_overflow_counter-- <= 0) { // just being defensive 7787 _overflow_counter = CMSMarkStackOverflowInterval; 7788 return true; 7789 } else { 7790 return false; 7791 } 7792 } 7793 7794 bool CMSCollector::par_simulate_overflow() { 7795 return simulate_overflow(); 7796 } 7797 #endif 7798 7799 // Single-threaded 7800 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7801 assert(stack->isEmpty(), "Expected precondition"); 7802 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7803 size_t i = num; 7804 oop cur = _overflow_list; 7805 const markOop proto = markOopDesc::prototype(); 7806 NOT_PRODUCT(ssize_t n = 0;) 7807 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7808 next = oop(cur->mark_raw()); 7809 cur->set_mark_raw(proto); // until proven otherwise 7810 assert(oopDesc::is_oop(cur), "Should be an oop"); 7811 bool res = stack->push(cur); 7812 assert(res, "Bit off more than can chew?"); 7813 NOT_PRODUCT(n++;) 7814 } 7815 _overflow_list = cur; 7816 #ifndef PRODUCT 7817 assert(_num_par_pushes >= n, "Too many pops?"); 7818 _num_par_pushes -=n; 7819 #endif 7820 return !stack->isEmpty(); 7821 } 7822 7823 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7824 // (MT-safe) Get a prefix of at most "num" from the list. 7825 // The overflow list is chained through the mark word of 7826 // each object in the list. We fetch the entire list, 7827 // break off a prefix of the right size and return the 7828 // remainder. If other threads try to take objects from 7829 // the overflow list at that time, they will wait for 7830 // some time to see if data becomes available. If (and 7831 // only if) another thread places one or more object(s) 7832 // on the global list before we have returned the suffix 7833 // to the global list, we will walk down our local list 7834 // to find its end and append the global list to 7835 // our suffix before returning it. This suffix walk can 7836 // prove to be expensive (quadratic in the amount of traffic) 7837 // when there are many objects in the overflow list and 7838 // there is much producer-consumer contention on the list. 7839 // *NOTE*: The overflow list manipulation code here and 7840 // in ParNewGeneration:: are very similar in shape, 7841 // except that in the ParNew case we use the old (from/eden) 7842 // copy of the object to thread the list via its klass word. 7843 // Because of the common code, if you make any changes in 7844 // the code below, please check the ParNew version to see if 7845 // similar changes might be needed. 7846 // CR 6797058 has been filed to consolidate the common code. 7847 bool CMSCollector::par_take_from_overflow_list(size_t num, 7848 OopTaskQueue* work_q, 7849 int no_of_gc_threads) { 7850 assert(work_q->size() == 0, "First empty local work queue"); 7851 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7852 if (_overflow_list == NULL) { 7853 return false; 7854 } 7855 // Grab the entire list; we'll put back a suffix 7856 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7857 Thread* tid = Thread::current(); 7858 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7859 // set to ParallelGCThreads. 7860 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7861 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7862 // If the list is busy, we spin for a short while, 7863 // sleeping between attempts to get the list. 7864 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7865 os::sleep(tid, sleep_time_millis, false); 7866 if (_overflow_list == NULL) { 7867 // Nothing left to take 7868 return false; 7869 } else if (_overflow_list != BUSY) { 7870 // Try and grab the prefix 7871 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7872 } 7873 } 7874 // If the list was found to be empty, or we spun long 7875 // enough, we give up and return empty-handed. If we leave 7876 // the list in the BUSY state below, it must be the case that 7877 // some other thread holds the overflow list and will set it 7878 // to a non-BUSY state in the future. 7879 if (prefix == NULL || prefix == BUSY) { 7880 // Nothing to take or waited long enough 7881 if (prefix == NULL) { 7882 // Write back the NULL in case we overwrote it with BUSY above 7883 // and it is still the same value. 7884 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7885 } 7886 return false; 7887 } 7888 assert(prefix != NULL && prefix != BUSY, "Error"); 7889 size_t i = num; 7890 oop cur = prefix; 7891 // Walk down the first "num" objects, unless we reach the end. 7892 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7893 if (cur->mark_raw() == NULL) { 7894 // We have "num" or fewer elements in the list, so there 7895 // is nothing to return to the global list. 7896 // Write back the NULL in lieu of the BUSY we wrote 7897 // above, if it is still the same value. 7898 if (_overflow_list == BUSY) { 7899 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7900 } 7901 } else { 7902 // Chop off the suffix and return it to the global list. 7903 assert(cur->mark_raw() != BUSY, "Error"); 7904 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7905 cur->set_mark_raw(NULL); // break off suffix 7906 // It's possible that the list is still in the empty(busy) state 7907 // we left it in a short while ago; in that case we may be 7908 // able to place back the suffix without incurring the cost 7909 // of a walk down the list. 7910 oop observed_overflow_list = _overflow_list; 7911 oop cur_overflow_list = observed_overflow_list; 7912 bool attached = false; 7913 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7914 observed_overflow_list = 7915 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7916 if (cur_overflow_list == observed_overflow_list) { 7917 attached = true; 7918 break; 7919 } else cur_overflow_list = observed_overflow_list; 7920 } 7921 if (!attached) { 7922 // Too bad, someone else sneaked in (at least) an element; we'll need 7923 // to do a splice. Find tail of suffix so we can prepend suffix to global 7924 // list. 7925 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7926 oop suffix_tail = cur; 7927 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7928 "Tautology"); 7929 observed_overflow_list = _overflow_list; 7930 do { 7931 cur_overflow_list = observed_overflow_list; 7932 if (cur_overflow_list != BUSY) { 7933 // Do the splice ... 7934 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7935 } else { // cur_overflow_list == BUSY 7936 suffix_tail->set_mark_raw(NULL); 7937 } 7938 // ... and try to place spliced list back on overflow_list ... 7939 observed_overflow_list = 7940 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7941 } while (cur_overflow_list != observed_overflow_list); 7942 // ... until we have succeeded in doing so. 7943 } 7944 } 7945 7946 // Push the prefix elements on work_q 7947 assert(prefix != NULL, "control point invariant"); 7948 const markOop proto = markOopDesc::prototype(); 7949 oop next; 7950 NOT_PRODUCT(ssize_t n = 0;) 7951 for (cur = prefix; cur != NULL; cur = next) { 7952 next = oop(cur->mark_raw()); 7953 cur->set_mark_raw(proto); // until proven otherwise 7954 assert(oopDesc::is_oop(cur), "Should be an oop"); 7955 bool res = work_q->push(cur); 7956 assert(res, "Bit off more than we can chew?"); 7957 NOT_PRODUCT(n++;) 7958 } 7959 #ifndef PRODUCT 7960 assert(_num_par_pushes >= n, "Too many pops?"); 7961 Atomic::sub(n, &_num_par_pushes); 7962 #endif 7963 return true; 7964 } 7965 7966 // Single-threaded 7967 void CMSCollector::push_on_overflow_list(oop p) { 7968 NOT_PRODUCT(_num_par_pushes++;) 7969 assert(oopDesc::is_oop(p), "Not an oop"); 7970 preserve_mark_if_necessary(p); 7971 p->set_mark_raw((markOop)_overflow_list); 7972 _overflow_list = p; 7973 } 7974 7975 // Multi-threaded; use CAS to prepend to overflow list 7976 void CMSCollector::par_push_on_overflow_list(oop p) { 7977 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7978 assert(oopDesc::is_oop(p), "Not an oop"); 7979 par_preserve_mark_if_necessary(p); 7980 oop observed_overflow_list = _overflow_list; 7981 oop cur_overflow_list; 7982 do { 7983 cur_overflow_list = observed_overflow_list; 7984 if (cur_overflow_list != BUSY) { 7985 p->set_mark_raw(markOop(cur_overflow_list)); 7986 } else { 7987 p->set_mark_raw(NULL); 7988 } 7989 observed_overflow_list = 7990 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7991 } while (cur_overflow_list != observed_overflow_list); 7992 } 7993 #undef BUSY 7994 7995 // Single threaded 7996 // General Note on GrowableArray: pushes may silently fail 7997 // because we are (temporarily) out of C-heap for expanding 7998 // the stack. The problem is quite ubiquitous and affects 7999 // a lot of code in the JVM. The prudent thing for GrowableArray 8000 // to do (for now) is to exit with an error. However, that may 8001 // be too draconian in some cases because the caller may be 8002 // able to recover without much harm. For such cases, we 8003 // should probably introduce a "soft_push" method which returns 8004 // an indication of success or failure with the assumption that 8005 // the caller may be able to recover from a failure; code in 8006 // the VM can then be changed, incrementally, to deal with such 8007 // failures where possible, thus, incrementally hardening the VM 8008 // in such low resource situations. 8009 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8010 _preserved_oop_stack.push(p); 8011 _preserved_mark_stack.push(m); 8012 assert(m == p->mark_raw(), "Mark word changed"); 8013 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8014 "bijection"); 8015 } 8016 8017 // Single threaded 8018 void CMSCollector::preserve_mark_if_necessary(oop p) { 8019 markOop m = p->mark_raw(); 8020 if (m->must_be_preserved(p)) { 8021 preserve_mark_work(p, m); 8022 } 8023 } 8024 8025 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8026 markOop m = p->mark_raw(); 8027 if (m->must_be_preserved(p)) { 8028 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8029 // Even though we read the mark word without holding 8030 // the lock, we are assured that it will not change 8031 // because we "own" this oop, so no other thread can 8032 // be trying to push it on the overflow list; see 8033 // the assertion in preserve_mark_work() that checks 8034 // that m == p->mark_raw(). 8035 preserve_mark_work(p, m); 8036 } 8037 } 8038 8039 // We should be able to do this multi-threaded, 8040 // a chunk of stack being a task (this is 8041 // correct because each oop only ever appears 8042 // once in the overflow list. However, it's 8043 // not very easy to completely overlap this with 8044 // other operations, so will generally not be done 8045 // until all work's been completed. Because we 8046 // expect the preserved oop stack (set) to be small, 8047 // it's probably fine to do this single-threaded. 8048 // We can explore cleverer concurrent/overlapped/parallel 8049 // processing of preserved marks if we feel the 8050 // need for this in the future. Stack overflow should 8051 // be so rare in practice and, when it happens, its 8052 // effect on performance so great that this will 8053 // likely just be in the noise anyway. 8054 void CMSCollector::restore_preserved_marks_if_any() { 8055 assert(SafepointSynchronize::is_at_safepoint(), 8056 "world should be stopped"); 8057 assert(Thread::current()->is_ConcurrentGC_thread() || 8058 Thread::current()->is_VM_thread(), 8059 "should be single-threaded"); 8060 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8061 "bijection"); 8062 8063 while (!_preserved_oop_stack.is_empty()) { 8064 oop p = _preserved_oop_stack.pop(); 8065 assert(oopDesc::is_oop(p), "Should be an oop"); 8066 assert(_span.contains(p), "oop should be in _span"); 8067 assert(p->mark_raw() == markOopDesc::prototype(), 8068 "Set when taken from overflow list"); 8069 markOop m = _preserved_mark_stack.pop(); 8070 p->set_mark_raw(m); 8071 } 8072 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8073 "stacks were cleared above"); 8074 } 8075 8076 #ifndef PRODUCT 8077 bool CMSCollector::no_preserved_marks() const { 8078 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8079 } 8080 #endif 8081 8082 // Transfer some number of overflown objects to usual marking 8083 // stack. Return true if some objects were transferred. 8084 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8085 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8086 (size_t)ParGCDesiredObjsFromOverflowList); 8087 8088 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8089 assert(_collector->overflow_list_is_empty() || res, 8090 "If list is not empty, we should have taken something"); 8091 assert(!res || !_mark_stack->isEmpty(), 8092 "If we took something, it should now be on our stack"); 8093 return res; 8094 } 8095 8096 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8097 size_t res = _sp->block_size_no_stall(addr, _collector); 8098 if (_sp->block_is_obj(addr)) { 8099 if (_live_bit_map->isMarked(addr)) { 8100 // It can't have been dead in a previous cycle 8101 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8102 } else { 8103 _dead_bit_map->mark(addr); // mark the dead object 8104 } 8105 } 8106 // Could be 0, if the block size could not be computed without stalling. 8107 return res; 8108 } 8109 8110 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8111 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8112 switch (phase) { 8113 case CMSCollector::InitialMarking: 8114 initialize(manager /* GC manager */ , 8115 cause /* cause of the GC */, 8116 true /* recordGCBeginTime */, 8117 true /* recordPreGCUsage */, 8118 false /* recordPeakUsage */, 8119 false /* recordPostGCusage */, 8120 true /* recordAccumulatedGCTime */, 8121 false /* recordGCEndTime */, 8122 false /* countCollection */ ); 8123 break; 8124 8125 case CMSCollector::FinalMarking: 8126 initialize(manager /* GC manager */ , 8127 cause /* cause of the GC */, 8128 false /* recordGCBeginTime */, 8129 false /* recordPreGCUsage */, 8130 false /* recordPeakUsage */, 8131 false /* recordPostGCusage */, 8132 true /* recordAccumulatedGCTime */, 8133 false /* recordGCEndTime */, 8134 false /* countCollection */ ); 8135 break; 8136 8137 case CMSCollector::Sweeping: 8138 initialize(manager /* GC manager */ , 8139 cause /* cause of the GC */, 8140 false /* recordGCBeginTime */, 8141 false /* recordPreGCUsage */, 8142 true /* recordPeakUsage */, 8143 true /* recordPostGCusage */, 8144 false /* recordAccumulatedGCTime */, 8145 true /* recordGCEndTime */, 8146 true /* countCollection */ ); 8147 break; 8148 8149 default: 8150 ShouldNotReachHere(); 8151 } 8152 }