1 /* 2 * Copyright (c) 2001, 2017, 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/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "logging/logStream.hpp" 59 #include "memory/allocation.hpp" 60 #include "memory/iterator.inline.hpp" 61 #include "memory/padded.hpp" 62 #include "memory/resourceArea.hpp" 63 #include "oops/oop.inline.hpp" 64 #include "prims/jvmtiExport.hpp" 65 #include "runtime/atomic.hpp" 66 #include "runtime/globals_extension.hpp" 67 #include "runtime/handles.inline.hpp" 68 #include "runtime/java.hpp" 69 #include "runtime/orderAccess.inline.hpp" 70 #include "runtime/timer.hpp" 71 #include "runtime/vmThread.hpp" 72 #include "services/memoryService.hpp" 73 #include "services/runtimeService.hpp" 74 #include "utilities/align.hpp" 75 #include "utilities/stack.inline.hpp" 76 77 // statics 78 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 79 bool CMSCollector::_full_gc_requested = false; 80 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 81 82 ////////////////////////////////////////////////////////////////// 83 // In support of CMS/VM thread synchronization 84 ////////////////////////////////////////////////////////////////// 85 // We split use of the CGC_lock into 2 "levels". 86 // The low-level locking is of the usual CGC_lock monitor. We introduce 87 // a higher level "token" (hereafter "CMS token") built on top of the 88 // low level monitor (hereafter "CGC lock"). 89 // The token-passing protocol gives priority to the VM thread. The 90 // CMS-lock doesn't provide any fairness guarantees, but clients 91 // should ensure that it is only held for very short, bounded 92 // durations. 93 // 94 // When either of the CMS thread or the VM thread is involved in 95 // collection operations during which it does not want the other 96 // thread to interfere, it obtains the CMS token. 97 // 98 // If either thread tries to get the token while the other has 99 // it, that thread waits. However, if the VM thread and CMS thread 100 // both want the token, then the VM thread gets priority while the 101 // CMS thread waits. This ensures, for instance, that the "concurrent" 102 // phases of the CMS thread's work do not block out the VM thread 103 // for long periods of time as the CMS thread continues to hog 104 // the token. (See bug 4616232). 105 // 106 // The baton-passing functions are, however, controlled by the 107 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 108 // and here the low-level CMS lock, not the high level token, 109 // ensures mutual exclusion. 110 // 111 // Two important conditions that we have to satisfy: 112 // 1. if a thread does a low-level wait on the CMS lock, then it 113 // relinquishes the CMS token if it were holding that token 114 // when it acquired the low-level CMS lock. 115 // 2. any low-level notifications on the low-level lock 116 // should only be sent when a thread has relinquished the token. 117 // 118 // In the absence of either property, we'd have potential deadlock. 119 // 120 // We protect each of the CMS (concurrent and sequential) phases 121 // with the CMS _token_, not the CMS _lock_. 122 // 123 // The only code protected by CMS lock is the token acquisition code 124 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 125 // baton-passing code. 126 // 127 // Unfortunately, i couldn't come up with a good abstraction to factor and 128 // hide the naked CGC_lock manipulation in the baton-passing code 129 // further below. That's something we should try to do. Also, the proof 130 // of correctness of this 2-level locking scheme is far from obvious, 131 // and potentially quite slippery. We have an uneasy suspicion, for instance, 132 // that there may be a theoretical possibility of delay/starvation in the 133 // low-level lock/wait/notify scheme used for the baton-passing because of 134 // potential interference with the priority scheme embodied in the 135 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 136 // invocation further below and marked with "XXX 20011219YSR". 137 // Indeed, as we note elsewhere, this may become yet more slippery 138 // in the presence of multiple CMS and/or multiple VM threads. XXX 139 140 class CMSTokenSync: public StackObj { 141 private: 142 bool _is_cms_thread; 143 public: 144 CMSTokenSync(bool is_cms_thread): 145 _is_cms_thread(is_cms_thread) { 146 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 147 "Incorrect argument to constructor"); 148 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 149 } 150 151 ~CMSTokenSync() { 152 assert(_is_cms_thread ? 153 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 154 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 155 "Incorrect state"); 156 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 157 } 158 }; 159 160 // Convenience class that does a CMSTokenSync, and then acquires 161 // upto three locks. 162 class CMSTokenSyncWithLocks: public CMSTokenSync { 163 private: 164 // Note: locks are acquired in textual declaration order 165 // and released in the opposite order 166 MutexLockerEx _locker1, _locker2, _locker3; 167 public: 168 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 169 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 170 CMSTokenSync(is_cms_thread), 171 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 172 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 173 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 174 { } 175 }; 176 177 178 ////////////////////////////////////////////////////////////////// 179 // Concurrent Mark-Sweep Generation ///////////////////////////// 180 ////////////////////////////////////////////////////////////////// 181 182 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 183 184 // This struct contains per-thread things necessary to support parallel 185 // young-gen collection. 186 class CMSParGCThreadState: public CHeapObj<mtGC> { 187 public: 188 CompactibleFreeListSpaceLAB lab; 189 PromotionInfo promo; 190 191 // Constructor. 192 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 193 promo.setSpace(cfls); 194 } 195 }; 196 197 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 198 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 199 CardGeneration(rs, initial_byte_size, ct), 200 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 201 _did_compact(false) 202 { 203 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 204 HeapWord* end = (HeapWord*) _virtual_space.high(); 205 206 _direct_allocated_words = 0; 207 NOT_PRODUCT( 208 _numObjectsPromoted = 0; 209 _numWordsPromoted = 0; 210 _numObjectsAllocated = 0; 211 _numWordsAllocated = 0; 212 ) 213 214 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 215 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 216 _cmsSpace->_old_gen = this; 217 218 _gc_stats = new CMSGCStats(); 219 220 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 221 // offsets match. The ability to tell free chunks from objects 222 // depends on this property. 223 debug_only( 224 FreeChunk* junk = NULL; 225 assert(UseCompressedClassPointers || 226 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 227 "Offset of FreeChunk::_prev within FreeChunk must match" 228 " that of OopDesc::_klass within OopDesc"); 229 ) 230 231 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 232 for (uint i = 0; i < ParallelGCThreads; i++) { 233 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 234 } 235 236 _incremental_collection_failed = false; 237 // The "dilatation_factor" is the expansion that can occur on 238 // account of the fact that the minimum object size in the CMS 239 // generation may be larger than that in, say, a contiguous young 240 // generation. 241 // Ideally, in the calculation below, we'd compute the dilatation 242 // factor as: MinChunkSize/(promoting_gen's min object size) 243 // Since we do not have such a general query interface for the 244 // promoting generation, we'll instead just use the minimum 245 // object size (which today is a header's worth of space); 246 // note that all arithmetic is in units of HeapWords. 247 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 248 assert(_dilatation_factor >= 1.0, "from previous assert"); 249 } 250 251 252 // The field "_initiating_occupancy" represents the occupancy percentage 253 // at which we trigger a new collection cycle. Unless explicitly specified 254 // via CMSInitiatingOccupancyFraction (argument "io" below), it 255 // is calculated by: 256 // 257 // Let "f" be MinHeapFreeRatio in 258 // 259 // _initiating_occupancy = 100-f + 260 // f * (CMSTriggerRatio/100) 261 // where CMSTriggerRatio is the argument "tr" below. 262 // 263 // That is, if we assume the heap is at its desired maximum occupancy at the 264 // end of a collection, we let CMSTriggerRatio of the (purported) free 265 // space be allocated before initiating a new collection cycle. 266 // 267 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 268 assert(io <= 100 && tr <= 100, "Check the arguments"); 269 if (io >= 0) { 270 _initiating_occupancy = (double)io / 100.0; 271 } else { 272 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 273 (double)(tr * MinHeapFreeRatio) / 100.0) 274 / 100.0; 275 } 276 } 277 278 void ConcurrentMarkSweepGeneration::ref_processor_init() { 279 assert(collector() != NULL, "no collector"); 280 collector()->ref_processor_init(); 281 } 282 283 void CMSCollector::ref_processor_init() { 284 if (_ref_processor == NULL) { 285 // Allocate and initialize a reference processor 286 _ref_processor = 287 new ReferenceProcessor(_span, // span 288 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 289 ParallelGCThreads, // mt processing degree 290 _cmsGen->refs_discovery_is_mt(), // mt discovery 291 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 292 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 293 &_is_alive_closure); // closure for liveness info 294 // Initialize the _ref_processor field of CMSGen 295 _cmsGen->set_ref_processor(_ref_processor); 296 297 } 298 } 299 300 AdaptiveSizePolicy* CMSCollector::size_policy() { 301 GenCollectedHeap* gch = GenCollectedHeap::heap(); 302 return gch->gen_policy()->size_policy(); 303 } 304 305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 306 307 const char* gen_name = "old"; 308 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next young collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 375 cms_free, expected_promotion); 376 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 377 // Add 1 in case the consumption rate goes to zero. 378 return cms_free_dbl / (cms_consumption_rate() + 1.0); 379 } 380 return 0.0; 381 } 382 383 // Compare the duration of the cms collection to the 384 // time remaining before the cms generation is empty. 385 // Note that the time from the start of the cms collection 386 // to the start of the cms sweep (less than the total 387 // duration of the cms collection) can be used. This 388 // has been tried and some applications experienced 389 // promotion failures early in execution. This was 390 // possibly because the averages were not accurate 391 // enough at the beginning. 392 double CMSStats::time_until_cms_start() const { 393 // We add "gc0_period" to the "work" calculation 394 // below because this query is done (mostly) at the 395 // end of a scavenge, so we need to conservatively 396 // account for that much possible delay 397 // in the query so as to avoid concurrent mode failures 398 // due to starting the collection just a wee bit too 399 // late. 400 double work = cms_duration() + gc0_period(); 401 double deadline = time_until_cms_gen_full(); 402 // If a concurrent mode failure occurred recently, we want to be 403 // more conservative and halve our expected time_until_cms_gen_full() 404 if (work > deadline) { 405 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 406 cms_duration(), gc0_period(), time_until_cms_gen_full()); 407 return 0.0; 408 } 409 return work - deadline; 410 } 411 412 #ifndef PRODUCT 413 void CMSStats::print_on(outputStream *st) const { 414 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 415 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 416 gc0_duration(), gc0_period(), gc0_promoted()); 417 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 418 cms_duration(), cms_period(), cms_allocated()); 419 st->print(",cms_since_beg=%g,cms_since_end=%g", 420 cms_time_since_begin(), cms_time_since_end()); 421 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 422 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 423 424 if (valid()) { 425 st->print(",promo_rate=%g,cms_alloc_rate=%g", 426 promotion_rate(), cms_allocation_rate()); 427 st->print(",cms_consumption_rate=%g,time_until_full=%g", 428 cms_consumption_rate(), time_until_cms_gen_full()); 429 } 430 st->cr(); 431 } 432 #endif // #ifndef PRODUCT 433 434 CMSCollector::CollectorState CMSCollector::_collectorState = 435 CMSCollector::Idling; 436 bool CMSCollector::_foregroundGCIsActive = false; 437 bool CMSCollector::_foregroundGCShouldWait = false; 438 439 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 440 CardTableRS* ct, 441 ConcurrentMarkSweepPolicy* cp): 442 _cmsGen(cmsGen), 443 _ct(ct), 444 _ref_processor(NULL), // will be set later 445 _conc_workers(NULL), // may be set later 446 _abort_preclean(false), 447 _start_sampling(false), 448 _between_prologue_and_epilogue(false), 449 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 450 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 451 -1 /* lock-free */, "No_lock" /* dummy */), 452 _modUnionClosurePar(&_modUnionTable), 453 // Adjust my span to cover old (cms) gen 454 _span(cmsGen->reserved()), 455 // Construct the is_alive_closure with _span & markBitMap 456 _is_alive_closure(_span, &_markBitMap), 457 _restart_addr(NULL), 458 _overflow_list(NULL), 459 _stats(cmsGen), 460 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 461 //verify that this lock should be acquired with safepoint check. 462 Monitor::_safepoint_check_sometimes)), 463 _eden_chunk_array(NULL), // may be set in ctor body 464 _eden_chunk_capacity(0), // -- ditto -- 465 _eden_chunk_index(0), // -- ditto -- 466 _survivor_plab_array(NULL), // -- ditto -- 467 _survivor_chunk_array(NULL), // -- ditto -- 468 _survivor_chunk_capacity(0), // -- ditto -- 469 _survivor_chunk_index(0), // -- ditto -- 470 _ser_pmc_preclean_ovflw(0), 471 _ser_kac_preclean_ovflw(0), 472 _ser_pmc_remark_ovflw(0), 473 _par_pmc_remark_ovflw(0), 474 _ser_kac_ovflw(0), 475 _par_kac_ovflw(0), 476 #ifndef PRODUCT 477 _num_par_pushes(0), 478 #endif 479 _collection_count_start(0), 480 _verifying(false), 481 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 482 _completed_initialization(false), 483 _collector_policy(cp), 484 _should_unload_classes(CMSClassUnloadingEnabled), 485 _concurrent_cycles_since_last_unload(0), 486 _roots_scanning_options(GenCollectedHeap::SO_None), 487 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 488 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 489 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 490 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 491 _cms_start_registered(false) 492 { 493 // Now expand the span and allocate the collection support structures 494 // (MUT, marking bit map etc.) to cover both generations subject to 495 // collection. 496 497 // For use by dirty card to oop closures. 498 _cmsGen->cmsSpace()->set_collector(this); 499 500 // Allocate MUT and marking bit map 501 { 502 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 503 if (!_markBitMap.allocate(_span)) { 504 log_warning(gc)("Failed to allocate CMS Bit Map"); 505 return; 506 } 507 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 508 } 509 { 510 _modUnionTable.allocate(_span); 511 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 512 } 513 514 if (!_markStack.allocate(MarkStackSize)) { 515 log_warning(gc)("Failed to allocate CMS Marking Stack"); 516 return; 517 } 518 519 // Support for multi-threaded concurrent phases 520 if (CMSConcurrentMTEnabled) { 521 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 522 // just for now 523 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 524 } 525 if (ConcGCThreads > 1) { 526 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 527 ConcGCThreads, true); 528 if (_conc_workers == NULL) { 529 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 530 CMSConcurrentMTEnabled = false; 531 } else { 532 _conc_workers->initialize_workers(); 533 } 534 } else { 535 CMSConcurrentMTEnabled = false; 536 } 537 } 538 if (!CMSConcurrentMTEnabled) { 539 ConcGCThreads = 0; 540 } else { 541 // Turn off CMSCleanOnEnter optimization temporarily for 542 // the MT case where it's not fixed yet; see 6178663. 543 CMSCleanOnEnter = false; 544 } 545 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 546 "Inconsistency"); 547 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 548 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 549 550 // Parallel task queues; these are shared for the 551 // concurrent and stop-world phases of CMS, but 552 // are not shared with parallel scavenge (ParNew). 553 { 554 uint i; 555 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 556 557 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 558 || ParallelRefProcEnabled) 559 && num_queues > 0) { 560 _task_queues = new OopTaskQueueSet(num_queues); 561 if (_task_queues == NULL) { 562 log_warning(gc)("task_queues allocation failure."); 563 return; 564 } 565 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 566 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 567 for (i = 0; i < num_queues; i++) { 568 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 569 if (q == NULL) { 570 log_warning(gc)("work_queue allocation failure."); 571 return; 572 } 573 _task_queues->register_queue(i, q); 574 } 575 for (i = 0; i < num_queues; i++) { 576 _task_queues->queue(i)->initialize(); 577 _hash_seed[i] = 17; // copied from ParNew 578 } 579 } 580 } 581 582 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 583 584 // Clip CMSBootstrapOccupancy between 0 and 100. 585 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 586 587 // Now tell CMS generations the identity of their collector 588 ConcurrentMarkSweepGeneration::set_collector(this); 589 590 // Create & start a CMS thread for this CMS collector 591 _cmsThread = ConcurrentMarkSweepThread::start(this); 592 assert(cmsThread() != NULL, "CMS Thread should have been created"); 593 assert(cmsThread()->collector() == this, 594 "CMS Thread should refer to this gen"); 595 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 596 597 // Support for parallelizing young gen rescan 598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 599 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 600 _young_gen = (ParNewGeneration*)gch->young_gen(); 601 if (gch->supports_inline_contig_alloc()) { 602 _top_addr = gch->top_addr(); 603 _end_addr = gch->end_addr(); 604 assert(_young_gen != NULL, "no _young_gen"); 605 _eden_chunk_index = 0; 606 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 607 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 608 } 609 610 // Support for parallelizing survivor space rescan 611 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 612 const size_t max_plab_samples = 613 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 614 615 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 616 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 617 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 618 _survivor_chunk_capacity = max_plab_samples; 619 for (uint i = 0; i < ParallelGCThreads; i++) { 620 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 621 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 622 assert(cur->end() == 0, "Should be 0"); 623 assert(cur->array() == vec, "Should be vec"); 624 assert(cur->capacity() == max_plab_samples, "Error"); 625 } 626 } 627 628 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 629 _gc_counters = new CollectorCounters("CMS", 1); 630 _completed_initialization = true; 631 _inter_sweep_timer.start(); // start of time 632 } 633 634 const char* ConcurrentMarkSweepGeneration::name() const { 635 return "concurrent mark-sweep generation"; 636 } 637 void ConcurrentMarkSweepGeneration::update_counters() { 638 if (UsePerfData) { 639 _space_counters->update_all(); 640 _gen_counters->update_all(); 641 } 642 } 643 644 // this is an optimized version of update_counters(). it takes the 645 // used value as a parameter rather than computing it. 646 // 647 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 648 if (UsePerfData) { 649 _space_counters->update_used(used); 650 _space_counters->update_capacity(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 void ConcurrentMarkSweepGeneration::print() const { 656 Generation::print(); 657 cmsSpace()->print(); 658 } 659 660 #ifndef PRODUCT 661 void ConcurrentMarkSweepGeneration::print_statistics() { 662 cmsSpace()->printFLCensus(0); 663 } 664 #endif 665 666 size_t 667 ConcurrentMarkSweepGeneration::contiguous_available() const { 668 // dld proposes an improvement in precision here. If the committed 669 // part of the space ends in a free block we should add that to 670 // uncommitted size in the calculation below. Will make this 671 // change later, staying with the approximation below for the 672 // time being. -- ysr. 673 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 674 } 675 676 size_t 677 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 678 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 679 } 680 681 size_t ConcurrentMarkSweepGeneration::max_available() const { 682 return free() + _virtual_space.uncommitted_size(); 683 } 684 685 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 686 size_t available = max_available(); 687 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 688 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 689 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 690 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 691 return res; 692 } 693 694 // At a promotion failure dump information on block layout in heap 695 // (cms old generation). 696 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 697 Log(gc, promotion) log; 698 if (log.is_trace()) { 699 LogStream ls(log.trace()); 700 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 701 } 702 } 703 704 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 705 // Clear the promotion information. These pointers can be adjusted 706 // along with all the other pointers into the heap but 707 // compaction is expected to be a rare event with 708 // a heap using cms so don't do it without seeing the need. 709 for (uint i = 0; i < ParallelGCThreads; i++) { 710 _par_gc_thread_states[i]->promo.reset(); 711 } 712 } 713 714 void ConcurrentMarkSweepGeneration::compute_new_size() { 715 assert_locked_or_safepoint(Heap_lock); 716 717 // If incremental collection failed, we just want to expand 718 // to the limit. 719 if (incremental_collection_failed()) { 720 clear_incremental_collection_failed(); 721 grow_to_reserved(); 722 return; 723 } 724 725 // The heap has been compacted but not reset yet. 726 // Any metric such as free() or used() will be incorrect. 727 728 CardGeneration::compute_new_size(); 729 730 // Reset again after a possible resizing 731 if (did_compact()) { 732 cmsSpace()->reset_after_compaction(); 733 } 734 } 735 736 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 737 assert_locked_or_safepoint(Heap_lock); 738 739 // If incremental collection failed, we just want to expand 740 // to the limit. 741 if (incremental_collection_failed()) { 742 clear_incremental_collection_failed(); 743 grow_to_reserved(); 744 return; 745 } 746 747 double free_percentage = ((double) free()) / capacity(); 748 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 749 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 750 751 // compute expansion delta needed for reaching desired free percentage 752 if (free_percentage < desired_free_percentage) { 753 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 754 assert(desired_capacity >= capacity(), "invalid expansion size"); 755 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 756 Log(gc) log; 757 if (log.is_trace()) { 758 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 759 log.trace("From compute_new_size: "); 760 log.trace(" Free fraction %f", free_percentage); 761 log.trace(" Desired free fraction %f", desired_free_percentage); 762 log.trace(" Maximum free fraction %f", maximum_free_percentage); 763 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 764 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 765 GenCollectedHeap* gch = GenCollectedHeap::heap(); 766 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 767 size_t young_size = gch->young_gen()->capacity(); 768 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 769 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 770 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 771 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 772 } 773 // safe if expansion fails 774 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 775 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 776 } else { 777 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 778 assert(desired_capacity <= capacity(), "invalid expansion size"); 779 size_t shrink_bytes = capacity() - desired_capacity; 780 // Don't shrink unless the delta is greater than the minimum shrink we want 781 if (shrink_bytes >= MinHeapDeltaBytes) { 782 shrink_free_list_by(shrink_bytes); 783 } 784 } 785 } 786 787 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 788 return cmsSpace()->freelistLock(); 789 } 790 791 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 792 CMSSynchronousYieldRequest yr; 793 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 794 return have_lock_and_allocate(size, tlab); 795 } 796 797 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 798 bool tlab /* ignored */) { 799 assert_lock_strong(freelistLock()); 800 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 801 HeapWord* res = cmsSpace()->allocate(adjustedSize); 802 // Allocate the object live (grey) if the background collector has 803 // started marking. This is necessary because the marker may 804 // have passed this address and consequently this object will 805 // not otherwise be greyed and would be incorrectly swept up. 806 // Note that if this object contains references, the writing 807 // of those references will dirty the card containing this object 808 // allowing the object to be blackened (and its references scanned) 809 // either during a preclean phase or at the final checkpoint. 810 if (res != NULL) { 811 // We may block here with an uninitialized object with 812 // its mark-bit or P-bits not yet set. Such objects need 813 // to be safely navigable by block_start(). 814 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 815 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 816 collector()->direct_allocated(res, adjustedSize); 817 _direct_allocated_words += adjustedSize; 818 // allocation counters 819 NOT_PRODUCT( 820 _numObjectsAllocated++; 821 _numWordsAllocated += (int)adjustedSize; 822 ) 823 } 824 return res; 825 } 826 827 // In the case of direct allocation by mutators in a generation that 828 // is being concurrently collected, the object must be allocated 829 // live (grey) if the background collector has started marking. 830 // This is necessary because the marker may 831 // have passed this address and consequently this object will 832 // not otherwise be greyed and would be incorrectly swept up. 833 // Note that if this object contains references, the writing 834 // of those references will dirty the card containing this object 835 // allowing the object to be blackened (and its references scanned) 836 // either during a preclean phase or at the final checkpoint. 837 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 838 assert(_markBitMap.covers(start, size), "Out of bounds"); 839 if (_collectorState >= Marking) { 840 MutexLockerEx y(_markBitMap.lock(), 841 Mutex::_no_safepoint_check_flag); 842 // [see comments preceding SweepClosure::do_blk() below for details] 843 // 844 // Can the P-bits be deleted now? JJJ 845 // 846 // 1. need to mark the object as live so it isn't collected 847 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 848 // 3. need to mark the end of the object so marking, precleaning or sweeping 849 // can skip over uninitialized or unparsable objects. An allocated 850 // object is considered uninitialized for our purposes as long as 851 // its klass word is NULL. All old gen objects are parsable 852 // as soon as they are initialized.) 853 _markBitMap.mark(start); // object is live 854 _markBitMap.mark(start + 1); // object is potentially uninitialized? 855 _markBitMap.mark(start + size - 1); 856 // mark end of object 857 } 858 // check that oop looks uninitialized 859 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 860 } 861 862 void CMSCollector::promoted(bool par, HeapWord* start, 863 bool is_obj_array, size_t obj_size) { 864 assert(_markBitMap.covers(start), "Out of bounds"); 865 // See comment in direct_allocated() about when objects should 866 // be allocated live. 867 if (_collectorState >= Marking) { 868 // we already hold the marking bit map lock, taken in 869 // the prologue 870 if (par) { 871 _markBitMap.par_mark(start); 872 } else { 873 _markBitMap.mark(start); 874 } 875 // We don't need to mark the object as uninitialized (as 876 // in direct_allocated above) because this is being done with the 877 // world stopped and the object will be initialized by the 878 // time the marking, precleaning or sweeping get to look at it. 879 // But see the code for copying objects into the CMS generation, 880 // where we need to ensure that concurrent readers of the 881 // block offset table are able to safely navigate a block that 882 // is in flux from being free to being allocated (and in 883 // transition while being copied into) and subsequently 884 // becoming a bona-fide object when the copy/promotion is complete. 885 assert(SafepointSynchronize::is_at_safepoint(), 886 "expect promotion only at safepoints"); 887 888 if (_collectorState < Sweeping) { 889 // Mark the appropriate cards in the modUnionTable, so that 890 // this object gets scanned before the sweep. If this is 891 // not done, CMS generation references in the object might 892 // not get marked. 893 // For the case of arrays, which are otherwise precisely 894 // marked, we need to dirty the entire array, not just its head. 895 if (is_obj_array) { 896 // The [par_]mark_range() method expects mr.end() below to 897 // be aligned to the granularity of a bit's representation 898 // in the heap. In the case of the MUT below, that's a 899 // card size. 900 MemRegion mr(start, 901 align_up(start + obj_size, 902 CardTableModRefBS::card_size /* bytes */)); 903 if (par) { 904 _modUnionTable.par_mark_range(mr); 905 } else { 906 _modUnionTable.mark_range(mr); 907 } 908 } else { // not an obj array; we can just mark the head 909 if (par) { 910 _modUnionTable.par_mark(start); 911 } else { 912 _modUnionTable.mark(start); 913 } 914 } 915 } 916 } 917 } 918 919 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 920 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 921 // allocate, copy and if necessary update promoinfo -- 922 // delegate to underlying space. 923 assert_lock_strong(freelistLock()); 924 925 #ifndef PRODUCT 926 if (GenCollectedHeap::heap()->promotion_should_fail()) { 927 return NULL; 928 } 929 #endif // #ifndef PRODUCT 930 931 oop res = _cmsSpace->promote(obj, obj_size); 932 if (res == NULL) { 933 // expand and retry 934 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 935 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 936 // Since this is the old generation, we don't try to promote 937 // into a more senior generation. 938 res = _cmsSpace->promote(obj, obj_size); 939 } 940 if (res != NULL) { 941 // See comment in allocate() about when objects should 942 // be allocated live. 943 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 944 collector()->promoted(false, // Not parallel 945 (HeapWord*)res, obj->is_objArray(), obj_size); 946 // promotion counters 947 NOT_PRODUCT( 948 _numObjectsPromoted++; 949 _numWordsPromoted += 950 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 951 ) 952 } 953 return res; 954 } 955 956 957 // IMPORTANT: Notes on object size recognition in CMS. 958 // --------------------------------------------------- 959 // A block of storage in the CMS generation is always in 960 // one of three states. A free block (FREE), an allocated 961 // object (OBJECT) whose size() method reports the correct size, 962 // and an intermediate state (TRANSIENT) in which its size cannot 963 // be accurately determined. 964 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 965 // ----------------------------------------------------- 966 // FREE: klass_word & 1 == 1; mark_word holds block size 967 // 968 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 969 // obj->size() computes correct size 970 // 971 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 972 // 973 // STATE IDENTIFICATION: (64 bit+COOPS) 974 // ------------------------------------ 975 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 976 // 977 // OBJECT: klass_word installed; klass_word != 0; 978 // obj->size() computes correct size 979 // 980 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 981 // 982 // 983 // STATE TRANSITION DIAGRAM 984 // 985 // mut / parnew mut / parnew 986 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 987 // ^ | 988 // |------------------------ DEAD <------------------------------------| 989 // sweep mut 990 // 991 // While a block is in TRANSIENT state its size cannot be determined 992 // so readers will either need to come back later or stall until 993 // the size can be determined. Note that for the case of direct 994 // allocation, P-bits, when available, may be used to determine the 995 // size of an object that may not yet have been initialized. 996 997 // Things to support parallel young-gen collection. 998 oop 999 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1000 oop old, markOop m, 1001 size_t word_sz) { 1002 #ifndef PRODUCT 1003 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1004 return NULL; 1005 } 1006 #endif // #ifndef PRODUCT 1007 1008 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1009 PromotionInfo* promoInfo = &ps->promo; 1010 // if we are tracking promotions, then first ensure space for 1011 // promotion (including spooling space for saving header if necessary). 1012 // then allocate and copy, then track promoted info if needed. 1013 // When tracking (see PromotionInfo::track()), the mark word may 1014 // be displaced and in this case restoration of the mark word 1015 // occurs in the (oop_since_save_marks_)iterate phase. 1016 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1017 // Out of space for allocating spooling buffers; 1018 // try expanding and allocating spooling buffers. 1019 if (!expand_and_ensure_spooling_space(promoInfo)) { 1020 return NULL; 1021 } 1022 } 1023 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1024 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1025 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1026 if (obj_ptr == NULL) { 1027 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1028 if (obj_ptr == NULL) { 1029 return NULL; 1030 } 1031 } 1032 oop obj = oop(obj_ptr); 1033 OrderAccess::storestore(); 1034 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1036 // IMPORTANT: See note on object initialization for CMS above. 1037 // Otherwise, copy the object. Here we must be careful to insert the 1038 // klass pointer last, since this marks the block as an allocated object. 1039 // Except with compressed oops it's the mark word. 1040 HeapWord* old_ptr = (HeapWord*)old; 1041 // Restore the mark word copied above. 1042 obj->set_mark(m); 1043 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1044 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1045 OrderAccess::storestore(); 1046 1047 if (UseCompressedClassPointers) { 1048 // Copy gap missed by (aligned) header size calculation below 1049 obj->set_klass_gap(old->klass_gap()); 1050 } 1051 if (word_sz > (size_t)oopDesc::header_size()) { 1052 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1053 obj_ptr + oopDesc::header_size(), 1054 word_sz - oopDesc::header_size()); 1055 } 1056 1057 // Now we can track the promoted object, if necessary. We take care 1058 // to delay the transition from uninitialized to full object 1059 // (i.e., insertion of klass pointer) until after, so that it 1060 // atomically becomes a promoted object. 1061 if (promoInfo->tracking()) { 1062 promoInfo->track((PromotedObject*)obj, old->klass()); 1063 } 1064 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1065 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1066 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1067 1068 // Finally, install the klass pointer (this should be volatile). 1069 OrderAccess::storestore(); 1070 obj->set_klass(old->klass()); 1071 // We should now be able to calculate the right size for this object 1072 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1073 1074 collector()->promoted(true, // parallel 1075 obj_ptr, old->is_objArray(), word_sz); 1076 1077 NOT_PRODUCT( 1078 Atomic::inc(&_numObjectsPromoted); 1079 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1080 ) 1081 1082 return obj; 1083 } 1084 1085 void 1086 ConcurrentMarkSweepGeneration:: 1087 par_promote_alloc_done(int thread_num) { 1088 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1089 ps->lab.retire(thread_num); 1090 } 1091 1092 void 1093 ConcurrentMarkSweepGeneration:: 1094 par_oop_since_save_marks_iterate_done(int thread_num) { 1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1096 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1097 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1098 1099 // Because card-scanning has been completed, subsequent phases 1100 // (e.g., reference processing) will not need to recognize which 1101 // objects have been promoted during this GC. So, we can now disable 1102 // promotion tracking. 1103 ps->promo.stopTrackingPromotions(); 1104 } 1105 1106 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1107 size_t size, 1108 bool tlab) 1109 { 1110 // We allow a STW collection only if a full 1111 // collection was requested. 1112 return full || should_allocate(size, tlab); // FIX ME !!! 1113 // This and promotion failure handling are connected at the 1114 // hip and should be fixed by untying them. 1115 } 1116 1117 bool CMSCollector::shouldConcurrentCollect() { 1118 LogTarget(Trace, gc) log; 1119 1120 if (_full_gc_requested) { 1121 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1122 return true; 1123 } 1124 1125 FreelistLocker x(this); 1126 // ------------------------------------------------------------------ 1127 // Print out lots of information which affects the initiation of 1128 // a collection. 1129 if (log.is_enabled() && stats().valid()) { 1130 log.print("CMSCollector shouldConcurrentCollect: "); 1131 1132 LogStream out(log); 1133 stats().print_on(&out); 1134 1135 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1136 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1137 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1138 log.print("promotion_rate=%g", stats().promotion_rate()); 1139 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1140 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1141 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1142 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1143 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1144 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1145 } 1146 // ------------------------------------------------------------------ 1147 1148 // If the estimated time to complete a cms collection (cms_duration()) 1149 // is less than the estimated time remaining until the cms generation 1150 // is full, start a collection. 1151 if (!UseCMSInitiatingOccupancyOnly) { 1152 if (stats().valid()) { 1153 if (stats().time_until_cms_start() == 0.0) { 1154 return true; 1155 } 1156 } else { 1157 // We want to conservatively collect somewhat early in order 1158 // to try and "bootstrap" our CMS/promotion statistics; 1159 // this branch will not fire after the first successful CMS 1160 // collection because the stats should then be valid. 1161 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1162 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1163 _cmsGen->occupancy(), _bootstrap_occupancy); 1164 return true; 1165 } 1166 } 1167 } 1168 1169 // Otherwise, we start a collection cycle if 1170 // old gen want a collection cycle started. Each may use 1171 // an appropriate criterion for making this decision. 1172 // XXX We need to make sure that the gen expansion 1173 // criterion dovetails well with this. XXX NEED TO FIX THIS 1174 if (_cmsGen->should_concurrent_collect()) { 1175 log.print("CMS old gen initiated"); 1176 return true; 1177 } 1178 1179 // We start a collection if we believe an incremental collection may fail; 1180 // this is not likely to be productive in practice because it's probably too 1181 // late anyway. 1182 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1183 assert(gch->collector_policy()->is_generation_policy(), 1184 "You may want to check the correctness of the following"); 1185 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1186 log.print("CMSCollector: collect because incremental collection will fail "); 1187 return true; 1188 } 1189 1190 if (MetaspaceGC::should_concurrent_collect()) { 1191 log.print("CMSCollector: collect for metadata allocation "); 1192 return true; 1193 } 1194 1195 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1196 if (CMSTriggerInterval >= 0) { 1197 if (CMSTriggerInterval == 0) { 1198 // Trigger always 1199 return true; 1200 } 1201 1202 // Check the CMS time since begin (we do not check the stats validity 1203 // as we want to be able to trigger the first CMS cycle as well) 1204 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1205 if (stats().valid()) { 1206 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1207 stats().cms_time_since_begin()); 1208 } else { 1209 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1210 } 1211 return true; 1212 } 1213 } 1214 1215 return false; 1216 } 1217 1218 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1219 1220 // Clear _expansion_cause fields of constituent generations 1221 void CMSCollector::clear_expansion_cause() { 1222 _cmsGen->clear_expansion_cause(); 1223 } 1224 1225 // We should be conservative in starting a collection cycle. To 1226 // start too eagerly runs the risk of collecting too often in the 1227 // extreme. To collect too rarely falls back on full collections, 1228 // which works, even if not optimum in terms of concurrent work. 1229 // As a work around for too eagerly collecting, use the flag 1230 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1231 // giving the user an easily understandable way of controlling the 1232 // collections. 1233 // We want to start a new collection cycle if any of the following 1234 // conditions hold: 1235 // . our current occupancy exceeds the configured initiating occupancy 1236 // for this generation, or 1237 // . we recently needed to expand this space and have not, since that 1238 // expansion, done a collection of this generation, or 1239 // . the underlying space believes that it may be a good idea to initiate 1240 // a concurrent collection (this may be based on criteria such as the 1241 // following: the space uses linear allocation and linear allocation is 1242 // going to fail, or there is believed to be excessive fragmentation in 1243 // the generation, etc... or ... 1244 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1245 // the case of the old generation; see CR 6543076): 1246 // we may be approaching a point at which allocation requests may fail because 1247 // we will be out of sufficient free space given allocation rate estimates.] 1248 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1249 1250 assert_lock_strong(freelistLock()); 1251 if (occupancy() > initiating_occupancy()) { 1252 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1253 short_name(), occupancy(), initiating_occupancy()); 1254 return true; 1255 } 1256 if (UseCMSInitiatingOccupancyOnly) { 1257 return false; 1258 } 1259 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1260 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1261 return true; 1262 } 1263 return false; 1264 } 1265 1266 void ConcurrentMarkSweepGeneration::collect(bool full, 1267 bool clear_all_soft_refs, 1268 size_t size, 1269 bool tlab) 1270 { 1271 collector()->collect(full, clear_all_soft_refs, size, tlab); 1272 } 1273 1274 void CMSCollector::collect(bool full, 1275 bool clear_all_soft_refs, 1276 size_t size, 1277 bool tlab) 1278 { 1279 // The following "if" branch is present for defensive reasons. 1280 // In the current uses of this interface, it can be replaced with: 1281 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1282 // But I am not placing that assert here to allow future 1283 // generality in invoking this interface. 1284 if (GCLocker::is_active()) { 1285 // A consistency test for GCLocker 1286 assert(GCLocker::needs_gc(), "Should have been set already"); 1287 // Skip this foreground collection, instead 1288 // expanding the heap if necessary. 1289 // Need the free list locks for the call to free() in compute_new_size() 1290 compute_new_size(); 1291 return; 1292 } 1293 acquire_control_and_collect(full, clear_all_soft_refs); 1294 } 1295 1296 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1297 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1298 unsigned int gc_count = gch->total_full_collections(); 1299 if (gc_count == full_gc_count) { 1300 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1301 _full_gc_requested = true; 1302 _full_gc_cause = cause; 1303 CGC_lock->notify(); // nudge CMS thread 1304 } else { 1305 assert(gc_count > full_gc_count, "Error: causal loop"); 1306 } 1307 } 1308 1309 bool CMSCollector::is_external_interruption() { 1310 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1311 return GCCause::is_user_requested_gc(cause) || 1312 GCCause::is_serviceability_requested_gc(cause); 1313 } 1314 1315 void CMSCollector::report_concurrent_mode_interruption() { 1316 if (is_external_interruption()) { 1317 log_debug(gc)("Concurrent mode interrupted"); 1318 } else { 1319 log_debug(gc)("Concurrent mode failure"); 1320 _gc_tracer_cm->report_concurrent_mode_failure(); 1321 } 1322 } 1323 1324 1325 // The foreground and background collectors need to coordinate in order 1326 // to make sure that they do not mutually interfere with CMS collections. 1327 // When a background collection is active, 1328 // the foreground collector may need to take over (preempt) and 1329 // synchronously complete an ongoing collection. Depending on the 1330 // frequency of the background collections and the heap usage 1331 // of the application, this preemption can be seldom or frequent. 1332 // There are only certain 1333 // points in the background collection that the "collection-baton" 1334 // can be passed to the foreground collector. 1335 // 1336 // The foreground collector will wait for the baton before 1337 // starting any part of the collection. The foreground collector 1338 // will only wait at one location. 1339 // 1340 // The background collector will yield the baton before starting a new 1341 // phase of the collection (e.g., before initial marking, marking from roots, 1342 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1343 // of the loop which switches the phases. The background collector does some 1344 // of the phases (initial mark, final re-mark) with the world stopped. 1345 // Because of locking involved in stopping the world, 1346 // the foreground collector should not block waiting for the background 1347 // collector when it is doing a stop-the-world phase. The background 1348 // collector will yield the baton at an additional point just before 1349 // it enters a stop-the-world phase. Once the world is stopped, the 1350 // background collector checks the phase of the collection. If the 1351 // phase has not changed, it proceeds with the collection. If the 1352 // phase has changed, it skips that phase of the collection. See 1353 // the comments on the use of the Heap_lock in collect_in_background(). 1354 // 1355 // Variable used in baton passing. 1356 // _foregroundGCIsActive - Set to true by the foreground collector when 1357 // it wants the baton. The foreground clears it when it has finished 1358 // the collection. 1359 // _foregroundGCShouldWait - Set to true by the background collector 1360 // when it is running. The foreground collector waits while 1361 // _foregroundGCShouldWait is true. 1362 // CGC_lock - monitor used to protect access to the above variables 1363 // and to notify the foreground and background collectors. 1364 // _collectorState - current state of the CMS collection. 1365 // 1366 // The foreground collector 1367 // acquires the CGC_lock 1368 // sets _foregroundGCIsActive 1369 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1370 // various locks acquired in preparation for the collection 1371 // are released so as not to block the background collector 1372 // that is in the midst of a collection 1373 // proceeds with the collection 1374 // clears _foregroundGCIsActive 1375 // returns 1376 // 1377 // The background collector in a loop iterating on the phases of the 1378 // collection 1379 // acquires the CGC_lock 1380 // sets _foregroundGCShouldWait 1381 // if _foregroundGCIsActive is set 1382 // clears _foregroundGCShouldWait, notifies _CGC_lock 1383 // waits on _CGC_lock for _foregroundGCIsActive to become false 1384 // and exits the loop. 1385 // otherwise 1386 // proceed with that phase of the collection 1387 // if the phase is a stop-the-world phase, 1388 // yield the baton once more just before enqueueing 1389 // the stop-world CMS operation (executed by the VM thread). 1390 // returns after all phases of the collection are done 1391 // 1392 1393 void CMSCollector::acquire_control_and_collect(bool full, 1394 bool clear_all_soft_refs) { 1395 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1396 assert(!Thread::current()->is_ConcurrentGC_thread(), 1397 "shouldn't try to acquire control from self!"); 1398 1399 // Start the protocol for acquiring control of the 1400 // collection from the background collector (aka CMS thread). 1401 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1402 "VM thread should have CMS token"); 1403 // Remember the possibly interrupted state of an ongoing 1404 // concurrent collection 1405 CollectorState first_state = _collectorState; 1406 1407 // Signal to a possibly ongoing concurrent collection that 1408 // we want to do a foreground collection. 1409 _foregroundGCIsActive = true; 1410 1411 // release locks and wait for a notify from the background collector 1412 // releasing the locks in only necessary for phases which 1413 // do yields to improve the granularity of the collection. 1414 assert_lock_strong(bitMapLock()); 1415 // We need to lock the Free list lock for the space that we are 1416 // currently collecting. 1417 assert(haveFreelistLocks(), "Must be holding free list locks"); 1418 bitMapLock()->unlock(); 1419 releaseFreelistLocks(); 1420 { 1421 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1422 if (_foregroundGCShouldWait) { 1423 // We are going to be waiting for action for the CMS thread; 1424 // it had better not be gone (for instance at shutdown)! 1425 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1426 "CMS thread must be running"); 1427 // Wait here until the background collector gives us the go-ahead 1428 ConcurrentMarkSweepThread::clear_CMS_flag( 1429 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1430 // Get a possibly blocked CMS thread going: 1431 // Note that we set _foregroundGCIsActive true above, 1432 // without protection of the CGC_lock. 1433 CGC_lock->notify(); 1434 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1435 "Possible deadlock"); 1436 while (_foregroundGCShouldWait) { 1437 // wait for notification 1438 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1439 // Possibility of delay/starvation here, since CMS token does 1440 // not know to give priority to VM thread? Actually, i think 1441 // there wouldn't be any delay/starvation, but the proof of 1442 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1443 } 1444 ConcurrentMarkSweepThread::set_CMS_flag( 1445 ConcurrentMarkSweepThread::CMS_vm_has_token); 1446 } 1447 } 1448 // The CMS_token is already held. Get back the other locks. 1449 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1450 "VM thread should have CMS token"); 1451 getFreelistLocks(); 1452 bitMapLock()->lock_without_safepoint_check(); 1453 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1454 p2i(Thread::current()), first_state); 1455 log_debug(gc, state)(" gets control with state %d", _collectorState); 1456 1457 // Inform cms gen if this was due to partial collection failing. 1458 // The CMS gen may use this fact to determine its expansion policy. 1459 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1460 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1461 assert(!_cmsGen->incremental_collection_failed(), 1462 "Should have been noticed, reacted to and cleared"); 1463 _cmsGen->set_incremental_collection_failed(); 1464 } 1465 1466 if (first_state > Idling) { 1467 report_concurrent_mode_interruption(); 1468 } 1469 1470 set_did_compact(true); 1471 1472 // If the collection is being acquired from the background 1473 // collector, there may be references on the discovered 1474 // references lists. Abandon those references, since some 1475 // of them may have become unreachable after concurrent 1476 // discovery; the STW compacting collector will redo discovery 1477 // more precisely, without being subject to floating garbage. 1478 // Leaving otherwise unreachable references in the discovered 1479 // lists would require special handling. 1480 ref_processor()->disable_discovery(); 1481 ref_processor()->abandon_partial_discovery(); 1482 ref_processor()->verify_no_references_recorded(); 1483 1484 if (first_state > Idling) { 1485 save_heap_summary(); 1486 } 1487 1488 do_compaction_work(clear_all_soft_refs); 1489 1490 // Has the GC time limit been exceeded? 1491 size_t max_eden_size = _young_gen->max_eden_size(); 1492 GCCause::Cause gc_cause = gch->gc_cause(); 1493 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1494 _young_gen->eden()->used(), 1495 _cmsGen->max_capacity(), 1496 max_eden_size, 1497 full, 1498 gc_cause, 1499 gch->collector_policy()); 1500 1501 // Reset the expansion cause, now that we just completed 1502 // a collection cycle. 1503 clear_expansion_cause(); 1504 _foregroundGCIsActive = false; 1505 return; 1506 } 1507 1508 // Resize the tenured generation 1509 // after obtaining the free list locks for the 1510 // two generations. 1511 void CMSCollector::compute_new_size() { 1512 assert_locked_or_safepoint(Heap_lock); 1513 FreelistLocker z(this); 1514 MetaspaceGC::compute_new_size(); 1515 _cmsGen->compute_new_size_free_list(); 1516 } 1517 1518 // A work method used by the foreground collector to do 1519 // a mark-sweep-compact. 1520 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1521 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1522 1523 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1524 gc_timer->register_gc_start(); 1525 1526 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1527 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1528 1529 gch->pre_full_gc_dump(gc_timer); 1530 1531 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1532 1533 // Temporarily widen the span of the weak reference processing to 1534 // the entire heap. 1535 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1536 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1537 // Temporarily, clear the "is_alive_non_header" field of the 1538 // reference processor. 1539 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1540 // Temporarily make reference _processing_ single threaded (non-MT). 1541 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1542 // Temporarily make refs discovery atomic 1543 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1544 // Temporarily make reference _discovery_ single threaded (non-MT) 1545 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1546 1547 ref_processor()->set_enqueuing_is_done(false); 1548 ref_processor()->enable_discovery(); 1549 ref_processor()->setup_policy(clear_all_soft_refs); 1550 // If an asynchronous collection finishes, the _modUnionTable is 1551 // all clear. If we are assuming the collection from an asynchronous 1552 // collection, clear the _modUnionTable. 1553 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1554 "_modUnionTable should be clear if the baton was not passed"); 1555 _modUnionTable.clear_all(); 1556 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1557 "mod union for klasses should be clear if the baton was passed"); 1558 _ct->cld_rem_set()->clear_mod_union(); 1559 1560 1561 // We must adjust the allocation statistics being maintained 1562 // in the free list space. We do so by reading and clearing 1563 // the sweep timer and updating the block flux rate estimates below. 1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1565 if (_inter_sweep_timer.is_active()) { 1566 _inter_sweep_timer.stop(); 1567 // Note that we do not use this sample to update the _inter_sweep_estimate. 1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1569 _inter_sweep_estimate.padded_average(), 1570 _intra_sweep_estimate.padded_average()); 1571 } 1572 1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1574 #ifdef ASSERT 1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1576 size_t free_size = cms_space->free(); 1577 assert(free_size == 1578 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1579 * HeapWordSize, 1580 "All the free space should be compacted into one chunk at top"); 1581 assert(cms_space->dictionary()->total_chunk_size( 1582 debug_only(cms_space->freelistLock())) == 0 || 1583 cms_space->totalSizeInIndexedFreeLists() == 0, 1584 "All the free space should be in a single chunk"); 1585 size_t num = cms_space->totalCount(); 1586 assert((free_size == 0 && num == 0) || 1587 (free_size > 0 && (num == 1 || num == 2)), 1588 "There should be at most 2 free chunks after compaction"); 1589 #endif // ASSERT 1590 _collectorState = Resetting; 1591 assert(_restart_addr == NULL, 1592 "Should have been NULL'd before baton was passed"); 1593 reset_stw(); 1594 _cmsGen->reset_after_compaction(); 1595 _concurrent_cycles_since_last_unload = 0; 1596 1597 // Clear any data recorded in the PLAB chunk arrays. 1598 if (_survivor_plab_array != NULL) { 1599 reset_survivor_plab_arrays(); 1600 } 1601 1602 // Adjust the per-size allocation stats for the next epoch. 1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1604 // Restart the "inter sweep timer" for the next epoch. 1605 _inter_sweep_timer.reset(); 1606 _inter_sweep_timer.start(); 1607 1608 // No longer a need to do a concurrent collection for Metaspace. 1609 MetaspaceGC::set_should_concurrent_collect(false); 1610 1611 gch->post_full_gc_dump(gc_timer); 1612 1613 gc_timer->register_gc_end(); 1614 1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1616 1617 // For a mark-sweep-compact, compute_new_size() will be called 1618 // in the heap's do_collection() method. 1619 } 1620 1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1622 Log(gc, heap) log; 1623 if (!log.is_trace()) { 1624 return; 1625 } 1626 1627 ContiguousSpace* eden_space = _young_gen->eden(); 1628 ContiguousSpace* from_space = _young_gen->from(); 1629 ContiguousSpace* to_space = _young_gen->to(); 1630 // Eden 1631 if (_eden_chunk_array != NULL) { 1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1633 p2i(eden_space->bottom()), p2i(eden_space->top()), 1634 p2i(eden_space->end()), eden_space->capacity()); 1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1636 _eden_chunk_index, _eden_chunk_capacity); 1637 for (size_t i = 0; i < _eden_chunk_index; i++) { 1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1639 } 1640 } 1641 // Survivor 1642 if (_survivor_chunk_array != NULL) { 1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1644 p2i(from_space->bottom()), p2i(from_space->top()), 1645 p2i(from_space->end()), from_space->capacity()); 1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1647 _survivor_chunk_index, _survivor_chunk_capacity); 1648 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1650 } 1651 } 1652 } 1653 1654 void CMSCollector::getFreelistLocks() const { 1655 // Get locks for all free lists in all generations that this 1656 // collector is responsible for 1657 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1658 } 1659 1660 void CMSCollector::releaseFreelistLocks() const { 1661 // Release locks for all free lists in all generations that this 1662 // collector is responsible for 1663 _cmsGen->freelistLock()->unlock(); 1664 } 1665 1666 bool CMSCollector::haveFreelistLocks() const { 1667 // Check locks for all free lists in all generations that this 1668 // collector is responsible for 1669 assert_lock_strong(_cmsGen->freelistLock()); 1670 PRODUCT_ONLY(ShouldNotReachHere()); 1671 return true; 1672 } 1673 1674 // A utility class that is used by the CMS collector to 1675 // temporarily "release" the foreground collector from its 1676 // usual obligation to wait for the background collector to 1677 // complete an ongoing phase before proceeding. 1678 class ReleaseForegroundGC: public StackObj { 1679 private: 1680 CMSCollector* _c; 1681 public: 1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1685 // allow a potentially blocked foreground collector to proceed 1686 _c->_foregroundGCShouldWait = false; 1687 if (_c->_foregroundGCIsActive) { 1688 CGC_lock->notify(); 1689 } 1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1691 "Possible deadlock"); 1692 } 1693 1694 ~ReleaseForegroundGC() { 1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1697 _c->_foregroundGCShouldWait = true; 1698 } 1699 }; 1700 1701 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1702 assert(Thread::current()->is_ConcurrentGC_thread(), 1703 "A CMS asynchronous collection is only allowed on a CMS thread."); 1704 1705 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1706 { 1707 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1708 MutexLockerEx hl(Heap_lock, safepoint_check); 1709 FreelistLocker fll(this); 1710 MutexLockerEx x(CGC_lock, safepoint_check); 1711 if (_foregroundGCIsActive) { 1712 // The foreground collector is. Skip this 1713 // background collection. 1714 assert(!_foregroundGCShouldWait, "Should be clear"); 1715 return; 1716 } else { 1717 assert(_collectorState == Idling, "Should be idling before start."); 1718 _collectorState = InitialMarking; 1719 register_gc_start(cause); 1720 // Reset the expansion cause, now that we are about to begin 1721 // a new cycle. 1722 clear_expansion_cause(); 1723 1724 // Clear the MetaspaceGC flag since a concurrent collection 1725 // is starting but also clear it after the collection. 1726 MetaspaceGC::set_should_concurrent_collect(false); 1727 } 1728 // Decide if we want to enable class unloading as part of the 1729 // ensuing concurrent GC cycle. 1730 update_should_unload_classes(); 1731 _full_gc_requested = false; // acks all outstanding full gc requests 1732 _full_gc_cause = GCCause::_no_gc; 1733 // Signal that we are about to start a collection 1734 gch->increment_total_full_collections(); // ... starting a collection cycle 1735 _collection_count_start = gch->total_full_collections(); 1736 } 1737 1738 size_t prev_used = _cmsGen->used(); 1739 1740 // The change of the collection state is normally done at this level; 1741 // the exceptions are phases that are executed while the world is 1742 // stopped. For those phases the change of state is done while the 1743 // world is stopped. For baton passing purposes this allows the 1744 // background collector to finish the phase and change state atomically. 1745 // The foreground collector cannot wait on a phase that is done 1746 // while the world is stopped because the foreground collector already 1747 // has the world stopped and would deadlock. 1748 while (_collectorState != Idling) { 1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1750 p2i(Thread::current()), _collectorState); 1751 // The foreground collector 1752 // holds the Heap_lock throughout its collection. 1753 // holds the CMS token (but not the lock) 1754 // except while it is waiting for the background collector to yield. 1755 // 1756 // The foreground collector should be blocked (not for long) 1757 // if the background collector is about to start a phase 1758 // executed with world stopped. If the background 1759 // collector has already started such a phase, the 1760 // foreground collector is blocked waiting for the 1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1762 // are executed in the VM thread. 1763 // 1764 // The locking order is 1765 // PendingListLock (PLL) -- if applicable (FinalMarking) 1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1767 // CMS token (claimed in 1768 // stop_world_and_do() --> 1769 // safepoint_synchronize() --> 1770 // CMSThread::synchronize()) 1771 1772 { 1773 // Check if the FG collector wants us to yield. 1774 CMSTokenSync x(true); // is cms thread 1775 if (waitForForegroundGC()) { 1776 // We yielded to a foreground GC, nothing more to be 1777 // done this round. 1778 assert(_foregroundGCShouldWait == false, "We set it to false in " 1779 "waitForForegroundGC()"); 1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1781 p2i(Thread::current()), _collectorState); 1782 return; 1783 } else { 1784 // The background collector can run but check to see if the 1785 // foreground collector has done a collection while the 1786 // background collector was waiting to get the CGC_lock 1787 // above. If yes, break so that _foregroundGCShouldWait 1788 // is cleared before returning. 1789 if (_collectorState == Idling) { 1790 break; 1791 } 1792 } 1793 } 1794 1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1796 "should be waiting"); 1797 1798 switch (_collectorState) { 1799 case InitialMarking: 1800 { 1801 ReleaseForegroundGC x(this); 1802 stats().record_cms_begin(); 1803 VM_CMS_Initial_Mark initial_mark_op(this); 1804 VMThread::execute(&initial_mark_op); 1805 } 1806 // The collector state may be any legal state at this point 1807 // since the background collector may have yielded to the 1808 // foreground collector. 1809 break; 1810 case Marking: 1811 // initial marking in checkpointRootsInitialWork has been completed 1812 if (markFromRoots()) { // we were successful 1813 assert(_collectorState == Precleaning, "Collector state should " 1814 "have changed"); 1815 } else { 1816 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1817 } 1818 break; 1819 case Precleaning: 1820 // marking from roots in markFromRoots has been completed 1821 preclean(); 1822 assert(_collectorState == AbortablePreclean || 1823 _collectorState == FinalMarking, 1824 "Collector state should have changed"); 1825 break; 1826 case AbortablePreclean: 1827 abortable_preclean(); 1828 assert(_collectorState == FinalMarking, "Collector state should " 1829 "have changed"); 1830 break; 1831 case FinalMarking: 1832 { 1833 ReleaseForegroundGC x(this); 1834 1835 VM_CMS_Final_Remark final_remark_op(this); 1836 VMThread::execute(&final_remark_op); 1837 } 1838 assert(_foregroundGCShouldWait, "block post-condition"); 1839 break; 1840 case Sweeping: 1841 // final marking in checkpointRootsFinal has been completed 1842 sweep(); 1843 assert(_collectorState == Resizing, "Collector state change " 1844 "to Resizing must be done under the free_list_lock"); 1845 1846 case Resizing: { 1847 // Sweeping has been completed... 1848 // At this point the background collection has completed. 1849 // Don't move the call to compute_new_size() down 1850 // into code that might be executed if the background 1851 // collection was preempted. 1852 { 1853 ReleaseForegroundGC x(this); // unblock FG collection 1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1855 CMSTokenSync z(true); // not strictly needed. 1856 if (_collectorState == Resizing) { 1857 compute_new_size(); 1858 save_heap_summary(); 1859 _collectorState = Resetting; 1860 } else { 1861 assert(_collectorState == Idling, "The state should only change" 1862 " because the foreground collector has finished the collection"); 1863 } 1864 } 1865 break; 1866 } 1867 case Resetting: 1868 // CMS heap resizing has been completed 1869 reset_concurrent(); 1870 assert(_collectorState == Idling, "Collector state should " 1871 "have changed"); 1872 1873 MetaspaceGC::set_should_concurrent_collect(false); 1874 1875 stats().record_cms_end(); 1876 // Don't move the concurrent_phases_end() and compute_new_size() 1877 // calls to here because a preempted background collection 1878 // has it's state set to "Resetting". 1879 break; 1880 case Idling: 1881 default: 1882 ShouldNotReachHere(); 1883 break; 1884 } 1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1886 p2i(Thread::current()), _collectorState); 1887 assert(_foregroundGCShouldWait, "block post-condition"); 1888 } 1889 1890 // Should this be in gc_epilogue? 1891 collector_policy()->counters()->update_counters(); 1892 1893 { 1894 // Clear _foregroundGCShouldWait and, in the event that the 1895 // foreground collector is waiting, notify it, before 1896 // returning. 1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1898 _foregroundGCShouldWait = false; 1899 if (_foregroundGCIsActive) { 1900 CGC_lock->notify(); 1901 } 1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1903 "Possible deadlock"); 1904 } 1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1906 p2i(Thread::current()), _collectorState); 1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1909 } 1910 1911 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1912 _cms_start_registered = true; 1913 _gc_timer_cm->register_gc_start(); 1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1915 } 1916 1917 void CMSCollector::register_gc_end() { 1918 if (_cms_start_registered) { 1919 report_heap_summary(GCWhen::AfterGC); 1920 1921 _gc_timer_cm->register_gc_end(); 1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1923 _cms_start_registered = false; 1924 } 1925 } 1926 1927 void CMSCollector::save_heap_summary() { 1928 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1929 _last_heap_summary = gch->create_heap_summary(); 1930 _last_metaspace_summary = gch->create_metaspace_summary(); 1931 } 1932 1933 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1936 } 1937 1938 bool CMSCollector::waitForForegroundGC() { 1939 bool res = false; 1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1941 "CMS thread should have CMS token"); 1942 // Block the foreground collector until the 1943 // background collectors decides whether to 1944 // yield. 1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1946 _foregroundGCShouldWait = true; 1947 if (_foregroundGCIsActive) { 1948 // The background collector yields to the 1949 // foreground collector and returns a value 1950 // indicating that it has yielded. The foreground 1951 // collector can proceed. 1952 res = true; 1953 _foregroundGCShouldWait = false; 1954 ConcurrentMarkSweepThread::clear_CMS_flag( 1955 ConcurrentMarkSweepThread::CMS_cms_has_token); 1956 ConcurrentMarkSweepThread::set_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1958 // Get a possibly blocked foreground thread going 1959 CGC_lock->notify(); 1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1961 p2i(Thread::current()), _collectorState); 1962 while (_foregroundGCIsActive) { 1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1964 } 1965 ConcurrentMarkSweepThread::set_CMS_flag( 1966 ConcurrentMarkSweepThread::CMS_cms_has_token); 1967 ConcurrentMarkSweepThread::clear_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1969 } 1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 return res; 1973 } 1974 1975 // Because of the need to lock the free lists and other structures in 1976 // the collector, common to all the generations that the collector is 1977 // collecting, we need the gc_prologues of individual CMS generations 1978 // delegate to their collector. It may have been simpler had the 1979 // current infrastructure allowed one to call a prologue on a 1980 // collector. In the absence of that we have the generation's 1981 // prologue delegate to the collector, which delegates back 1982 // some "local" work to a worker method in the individual generations 1983 // that it's responsible for collecting, while itself doing any 1984 // work common to all generations it's responsible for. A similar 1985 // comment applies to the gc_epilogue()'s. 1986 // The role of the variable _between_prologue_and_epilogue is to 1987 // enforce the invocation protocol. 1988 void CMSCollector::gc_prologue(bool full) { 1989 // Call gc_prologue_work() for the CMSGen 1990 // we are responsible for. 1991 1992 // The following locking discipline assumes that we are only called 1993 // when the world is stopped. 1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1995 1996 // The CMSCollector prologue must call the gc_prologues for the 1997 // "generations" that it's responsible 1998 // for. 1999 2000 assert( Thread::current()->is_VM_thread() 2001 || ( CMSScavengeBeforeRemark 2002 && Thread::current()->is_ConcurrentGC_thread()), 2003 "Incorrect thread type for prologue execution"); 2004 2005 if (_between_prologue_and_epilogue) { 2006 // We have already been invoked; this is a gc_prologue delegation 2007 // from yet another CMS generation that we are responsible for, just 2008 // ignore it since all relevant work has already been done. 2009 return; 2010 } 2011 2012 // set a bit saying prologue has been called; cleared in epilogue 2013 _between_prologue_and_epilogue = true; 2014 // Claim locks for common data structures, then call gc_prologue_work() 2015 // for each CMSGen. 2016 2017 getFreelistLocks(); // gets free list locks on constituent spaces 2018 bitMapLock()->lock_without_safepoint_check(); 2019 2020 // Should call gc_prologue_work() for all cms gens we are responsible for 2021 bool duringMarking = _collectorState >= Marking 2022 && _collectorState < Sweeping; 2023 2024 // The young collections clear the modified oops state, which tells if 2025 // there are any modified oops in the class. The remark phase also needs 2026 // that information. Tell the young collection to save the union of all 2027 // modified klasses. 2028 if (duringMarking) { 2029 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2030 } 2031 2032 bool registerClosure = duringMarking; 2033 2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2035 2036 if (!full) { 2037 stats().record_gc0_begin(); 2038 } 2039 } 2040 2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2042 2043 _capacity_at_prologue = capacity(); 2044 _used_at_prologue = used(); 2045 2046 // We enable promotion tracking so that card-scanning can recognize 2047 // which objects have been promoted during this GC and skip them. 2048 for (uint i = 0; i < ParallelGCThreads; i++) { 2049 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2050 } 2051 2052 // Delegate to CMScollector which knows how to coordinate between 2053 // this and any other CMS generations that it is responsible for 2054 // collecting. 2055 collector()->gc_prologue(full); 2056 } 2057 2058 // This is a "private" interface for use by this generation's CMSCollector. 2059 // Not to be called directly by any other entity (for instance, 2060 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2062 bool registerClosure, ModUnionClosure* modUnionClosure) { 2063 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2065 "Should be NULL"); 2066 if (registerClosure) { 2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2068 } 2069 cmsSpace()->gc_prologue(); 2070 // Clear stat counters 2071 NOT_PRODUCT( 2072 assert(_numObjectsPromoted == 0, "check"); 2073 assert(_numWordsPromoted == 0, "check"); 2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2076 _numObjectsAllocated = 0; 2077 _numWordsAllocated = 0; 2078 ) 2079 } 2080 2081 void CMSCollector::gc_epilogue(bool full) { 2082 // The following locking discipline assumes that we are only called 2083 // when the world is stopped. 2084 assert(SafepointSynchronize::is_at_safepoint(), 2085 "world is stopped assumption"); 2086 2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2088 // if linear allocation blocks need to be appropriately marked to allow the 2089 // the blocks to be parsable. We also check here whether we need to nudge the 2090 // CMS collector thread to start a new cycle (if it's not already active). 2091 assert( Thread::current()->is_VM_thread() 2092 || ( CMSScavengeBeforeRemark 2093 && Thread::current()->is_ConcurrentGC_thread()), 2094 "Incorrect thread type for epilogue execution"); 2095 2096 if (!_between_prologue_and_epilogue) { 2097 // We have already been invoked; this is a gc_epilogue delegation 2098 // from yet another CMS generation that we are responsible for, just 2099 // ignore it since all relevant work has already been done. 2100 return; 2101 } 2102 assert(haveFreelistLocks(), "must have freelist locks"); 2103 assert_lock_strong(bitMapLock()); 2104 2105 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2106 2107 _cmsGen->gc_epilogue_work(full); 2108 2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2110 // in case sampling was not already enabled, enable it 2111 _start_sampling = true; 2112 } 2113 // reset _eden_chunk_array so sampling starts afresh 2114 _eden_chunk_index = 0; 2115 2116 size_t cms_used = _cmsGen->cmsSpace()->used(); 2117 2118 // update performance counters - this uses a special version of 2119 // update_counters() that allows the utilization to be passed as a 2120 // parameter, avoiding multiple calls to used(). 2121 // 2122 _cmsGen->update_counters(cms_used); 2123 2124 bitMapLock()->unlock(); 2125 releaseFreelistLocks(); 2126 2127 if (!CleanChunkPoolAsync) { 2128 Chunk::clean_chunk_pool(); 2129 } 2130 2131 set_did_compact(false); 2132 _between_prologue_and_epilogue = false; // ready for next cycle 2133 } 2134 2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2136 collector()->gc_epilogue(full); 2137 2138 // When using ParNew, promotion tracking should have already been 2139 // disabled. However, the prologue (which enables promotion 2140 // tracking) and epilogue are called irrespective of the type of 2141 // GC. So they will also be called before and after Full GCs, during 2142 // which promotion tracking will not be explicitly disabled. So, 2143 // it's safer to also disable it here too (to be symmetric with 2144 // enabling it in the prologue). 2145 for (uint i = 0; i < ParallelGCThreads; i++) { 2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2147 } 2148 } 2149 2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2151 assert(!incremental_collection_failed(), "Should have been cleared"); 2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2153 cmsSpace()->gc_epilogue(); 2154 // Print stat counters 2155 NOT_PRODUCT( 2156 assert(_numObjectsAllocated == 0, "check"); 2157 assert(_numWordsAllocated == 0, "check"); 2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2160 _numObjectsPromoted = 0; 2161 _numWordsPromoted = 0; 2162 ) 2163 2164 // Call down the chain in contiguous_available needs the freelistLock 2165 // so print this out before releasing the freeListLock. 2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2167 } 2168 2169 #ifndef PRODUCT 2170 bool CMSCollector::have_cms_token() { 2171 Thread* thr = Thread::current(); 2172 if (thr->is_VM_thread()) { 2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2174 } else if (thr->is_ConcurrentGC_thread()) { 2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2176 } else if (thr->is_GC_task_thread()) { 2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2178 ParGCRareEvent_lock->owned_by_self(); 2179 } 2180 return false; 2181 } 2182 2183 // Check reachability of the given heap address in CMS generation, 2184 // treating all other generations as roots. 2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2186 // We could "guarantee" below, rather than assert, but I'll 2187 // leave these as "asserts" so that an adventurous debugger 2188 // could try this in the product build provided some subset of 2189 // the conditions were met, provided they were interested in the 2190 // results and knew that the computation below wouldn't interfere 2191 // with other concurrent computations mutating the structures 2192 // being read or written. 2193 assert(SafepointSynchronize::is_at_safepoint(), 2194 "Else mutations in object graph will make answer suspect"); 2195 assert(have_cms_token(), "Should hold cms token"); 2196 assert(haveFreelistLocks(), "must hold free list locks"); 2197 assert_lock_strong(bitMapLock()); 2198 2199 // Clear the marking bit map array before starting, but, just 2200 // for kicks, first report if the given address is already marked 2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2202 _markBitMap.isMarked(addr) ? "" : " not"); 2203 2204 if (verify_after_remark()) { 2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2206 bool result = verification_mark_bm()->isMarked(addr); 2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2208 result ? "IS" : "is NOT"); 2209 return result; 2210 } else { 2211 tty->print_cr("Could not compute result"); 2212 return false; 2213 } 2214 } 2215 #endif 2216 2217 void 2218 CMSCollector::print_on_error(outputStream* st) { 2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2220 if (collector != NULL) { 2221 CMSBitMap* bitmap = &collector->_markBitMap; 2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2223 bitmap->print_on_error(st, " Bits: "); 2224 2225 st->cr(); 2226 2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2229 mut_bitmap->print_on_error(st, " Bits: "); 2230 } 2231 } 2232 2233 //////////////////////////////////////////////////////// 2234 // CMS Verification Support 2235 //////////////////////////////////////////////////////// 2236 // Following the remark phase, the following invariant 2237 // should hold -- each object in the CMS heap which is 2238 // marked in markBitMap() should be marked in the verification_mark_bm(). 2239 2240 class VerifyMarkedClosure: public BitMapClosure { 2241 CMSBitMap* _marks; 2242 bool _failed; 2243 2244 public: 2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2246 2247 bool do_bit(size_t offset) { 2248 HeapWord* addr = _marks->offsetToHeapWord(offset); 2249 if (!_marks->isMarked(addr)) { 2250 Log(gc, verify) log; 2251 ResourceMark rm; 2252 LogStream ls(log.error()); 2253 oop(addr)->print_on(&ls); 2254 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2255 _failed = true; 2256 } 2257 return true; 2258 } 2259 2260 bool failed() { return _failed; } 2261 }; 2262 2263 bool CMSCollector::verify_after_remark() { 2264 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2265 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2266 static bool init = false; 2267 2268 assert(SafepointSynchronize::is_at_safepoint(), 2269 "Else mutations in object graph will make answer suspect"); 2270 assert(have_cms_token(), 2271 "Else there may be mutual interference in use of " 2272 " verification data structures"); 2273 assert(_collectorState > Marking && _collectorState <= Sweeping, 2274 "Else marking info checked here may be obsolete"); 2275 assert(haveFreelistLocks(), "must hold free list locks"); 2276 assert_lock_strong(bitMapLock()); 2277 2278 2279 // Allocate marking bit map if not already allocated 2280 if (!init) { // first time 2281 if (!verification_mark_bm()->allocate(_span)) { 2282 return false; 2283 } 2284 init = true; 2285 } 2286 2287 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2288 2289 // Turn off refs discovery -- so we will be tracing through refs. 2290 // This is as intended, because by this time 2291 // GC must already have cleared any refs that need to be cleared, 2292 // and traced those that need to be marked; moreover, 2293 // the marking done here is not going to interfere in any 2294 // way with the marking information used by GC. 2295 NoRefDiscovery no_discovery(ref_processor()); 2296 2297 #if defined(COMPILER2) || INCLUDE_JVMCI 2298 DerivedPointerTableDeactivate dpt_deact; 2299 #endif 2300 2301 // Clear any marks from a previous round 2302 verification_mark_bm()->clear_all(); 2303 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2304 verify_work_stacks_empty(); 2305 2306 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2307 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2308 // Update the saved marks which may affect the root scans. 2309 gch->save_marks(); 2310 2311 if (CMSRemarkVerifyVariant == 1) { 2312 // In this first variant of verification, we complete 2313 // all marking, then check if the new marks-vector is 2314 // a subset of the CMS marks-vector. 2315 verify_after_remark_work_1(); 2316 } else { 2317 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2318 // In this second variant of verification, we flag an error 2319 // (i.e. an object reachable in the new marks-vector not reachable 2320 // in the CMS marks-vector) immediately, also indicating the 2321 // identify of an object (A) that references the unmarked object (B) -- 2322 // presumably, a mutation to A failed to be picked up by preclean/remark? 2323 verify_after_remark_work_2(); 2324 } 2325 2326 return true; 2327 } 2328 2329 void CMSCollector::verify_after_remark_work_1() { 2330 ResourceMark rm; 2331 HandleMark hm; 2332 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2333 2334 // Get a clear set of claim bits for the roots processing to work with. 2335 ClassLoaderDataGraph::clear_claimed_marks(); 2336 2337 // Mark from roots one level into CMS 2338 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2339 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2340 2341 { 2342 StrongRootsScope srs(1); 2343 2344 gch->cms_process_roots(&srs, 2345 true, // young gen as roots 2346 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2347 should_unload_classes(), 2348 ¬Older, 2349 NULL); 2350 } 2351 2352 // Now mark from the roots 2353 MarkFromRootsClosure markFromRootsClosure(this, _span, 2354 verification_mark_bm(), verification_mark_stack(), 2355 false /* don't yield */, true /* verifying */); 2356 assert(_restart_addr == NULL, "Expected pre-condition"); 2357 verification_mark_bm()->iterate(&markFromRootsClosure); 2358 while (_restart_addr != NULL) { 2359 // Deal with stack overflow: by restarting at the indicated 2360 // address. 2361 HeapWord* ra = _restart_addr; 2362 markFromRootsClosure.reset(ra); 2363 _restart_addr = NULL; 2364 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2365 } 2366 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2367 verify_work_stacks_empty(); 2368 2369 // Marking completed -- now verify that each bit marked in 2370 // verification_mark_bm() is also marked in markBitMap(); flag all 2371 // errors by printing corresponding objects. 2372 VerifyMarkedClosure vcl(markBitMap()); 2373 verification_mark_bm()->iterate(&vcl); 2374 if (vcl.failed()) { 2375 Log(gc, verify) log; 2376 log.error("Failed marking verification after remark"); 2377 ResourceMark rm; 2378 LogStream ls(log.error()); 2379 gch->print_on(&ls); 2380 fatal("CMS: failed marking verification after remark"); 2381 } 2382 } 2383 2384 class VerifyCLDOopsCLDClosure : public CLDClosure { 2385 class VerifyCLDOopsClosure : public OopClosure { 2386 CMSBitMap* _bitmap; 2387 public: 2388 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2389 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2390 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2391 } _oop_closure; 2392 public: 2393 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2394 void do_cld(ClassLoaderData* cld) { 2395 cld->oops_do(&_oop_closure, false, false); 2396 } 2397 }; 2398 2399 void CMSCollector::verify_after_remark_work_2() { 2400 ResourceMark rm; 2401 HandleMark hm; 2402 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2403 2404 // Get a clear set of claim bits for the roots processing to work with. 2405 ClassLoaderDataGraph::clear_claimed_marks(); 2406 2407 // Mark from roots one level into CMS 2408 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2409 markBitMap()); 2410 CLDToOopClosure cld_closure(¬Older, true); 2411 2412 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2413 2414 { 2415 StrongRootsScope srs(1); 2416 2417 gch->cms_process_roots(&srs, 2418 true, // young gen as roots 2419 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2420 should_unload_classes(), 2421 ¬Older, 2422 &cld_closure); 2423 } 2424 2425 // Now mark from the roots 2426 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2427 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2428 assert(_restart_addr == NULL, "Expected pre-condition"); 2429 verification_mark_bm()->iterate(&markFromRootsClosure); 2430 while (_restart_addr != NULL) { 2431 // Deal with stack overflow: by restarting at the indicated 2432 // address. 2433 HeapWord* ra = _restart_addr; 2434 markFromRootsClosure.reset(ra); 2435 _restart_addr = NULL; 2436 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2437 } 2438 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2439 verify_work_stacks_empty(); 2440 2441 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2442 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2443 2444 // Marking completed -- now verify that each bit marked in 2445 // verification_mark_bm() is also marked in markBitMap(); flag all 2446 // errors by printing corresponding objects. 2447 VerifyMarkedClosure vcl(markBitMap()); 2448 verification_mark_bm()->iterate(&vcl); 2449 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2450 } 2451 2452 void ConcurrentMarkSweepGeneration::save_marks() { 2453 // delegate to CMS space 2454 cmsSpace()->save_marks(); 2455 } 2456 2457 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2458 return cmsSpace()->no_allocs_since_save_marks(); 2459 } 2460 2461 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2462 \ 2463 void ConcurrentMarkSweepGeneration:: \ 2464 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2465 cl->set_generation(this); \ 2466 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2467 cl->reset_generation(); \ 2468 save_marks(); \ 2469 } 2470 2471 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2472 2473 void 2474 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2475 if (freelistLock()->owned_by_self()) { 2476 Generation::oop_iterate(cl); 2477 } else { 2478 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2479 Generation::oop_iterate(cl); 2480 } 2481 } 2482 2483 void 2484 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2485 if (freelistLock()->owned_by_self()) { 2486 Generation::object_iterate(cl); 2487 } else { 2488 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2489 Generation::object_iterate(cl); 2490 } 2491 } 2492 2493 void 2494 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2495 if (freelistLock()->owned_by_self()) { 2496 Generation::safe_object_iterate(cl); 2497 } else { 2498 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2499 Generation::safe_object_iterate(cl); 2500 } 2501 } 2502 2503 void 2504 ConcurrentMarkSweepGeneration::post_compact() { 2505 } 2506 2507 void 2508 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2509 // Fix the linear allocation blocks to look like free blocks. 2510 2511 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2512 // are not called when the heap is verified during universe initialization and 2513 // at vm shutdown. 2514 if (freelistLock()->owned_by_self()) { 2515 cmsSpace()->prepare_for_verify(); 2516 } else { 2517 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2518 cmsSpace()->prepare_for_verify(); 2519 } 2520 } 2521 2522 void 2523 ConcurrentMarkSweepGeneration::verify() { 2524 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2525 // are not called when the heap is verified during universe initialization and 2526 // at vm shutdown. 2527 if (freelistLock()->owned_by_self()) { 2528 cmsSpace()->verify(); 2529 } else { 2530 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2531 cmsSpace()->verify(); 2532 } 2533 } 2534 2535 void CMSCollector::verify() { 2536 _cmsGen->verify(); 2537 } 2538 2539 #ifndef PRODUCT 2540 bool CMSCollector::overflow_list_is_empty() const { 2541 assert(_num_par_pushes >= 0, "Inconsistency"); 2542 if (_overflow_list == NULL) { 2543 assert(_num_par_pushes == 0, "Inconsistency"); 2544 } 2545 return _overflow_list == NULL; 2546 } 2547 2548 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2549 // merely consolidate assertion checks that appear to occur together frequently. 2550 void CMSCollector::verify_work_stacks_empty() const { 2551 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2552 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2553 } 2554 2555 void CMSCollector::verify_overflow_empty() const { 2556 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2557 assert(no_preserved_marks(), "No preserved marks"); 2558 } 2559 #endif // PRODUCT 2560 2561 // Decide if we want to enable class unloading as part of the 2562 // ensuing concurrent GC cycle. We will collect and 2563 // unload classes if it's the case that: 2564 // (a) class unloading is enabled at the command line, and 2565 // (b) old gen is getting really full 2566 // NOTE: Provided there is no change in the state of the heap between 2567 // calls to this method, it should have idempotent results. Moreover, 2568 // its results should be monotonically increasing (i.e. going from 0 to 1, 2569 // but not 1 to 0) between successive calls between which the heap was 2570 // not collected. For the implementation below, it must thus rely on 2571 // the property that concurrent_cycles_since_last_unload() 2572 // will not decrease unless a collection cycle happened and that 2573 // _cmsGen->is_too_full() are 2574 // themselves also monotonic in that sense. See check_monotonicity() 2575 // below. 2576 void CMSCollector::update_should_unload_classes() { 2577 _should_unload_classes = false; 2578 if (CMSClassUnloadingEnabled) { 2579 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2580 CMSClassUnloadingMaxInterval) 2581 || _cmsGen->is_too_full(); 2582 } 2583 } 2584 2585 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2586 bool res = should_concurrent_collect(); 2587 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2588 return res; 2589 } 2590 2591 void CMSCollector::setup_cms_unloading_and_verification_state() { 2592 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2593 || VerifyBeforeExit; 2594 const int rso = GenCollectedHeap::SO_AllCodeCache; 2595 2596 // We set the proper root for this CMS cycle here. 2597 if (should_unload_classes()) { // Should unload classes this cycle 2598 remove_root_scanning_option(rso); // Shrink the root set appropriately 2599 set_verifying(should_verify); // Set verification state for this cycle 2600 return; // Nothing else needs to be done at this time 2601 } 2602 2603 // Not unloading classes this cycle 2604 assert(!should_unload_classes(), "Inconsistency!"); 2605 2606 // If we are not unloading classes then add SO_AllCodeCache to root 2607 // scanning options. 2608 add_root_scanning_option(rso); 2609 2610 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2611 set_verifying(true); 2612 } else if (verifying() && !should_verify) { 2613 // We were verifying, but some verification flags got disabled. 2614 set_verifying(false); 2615 // Exclude symbols, strings and code cache elements from root scanning to 2616 // reduce IM and RM pauses. 2617 remove_root_scanning_option(rso); 2618 } 2619 } 2620 2621 2622 #ifndef PRODUCT 2623 HeapWord* CMSCollector::block_start(const void* p) const { 2624 const HeapWord* addr = (HeapWord*)p; 2625 if (_span.contains(p)) { 2626 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2627 return _cmsGen->cmsSpace()->block_start(p); 2628 } 2629 } 2630 return NULL; 2631 } 2632 #endif 2633 2634 HeapWord* 2635 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2636 bool tlab, 2637 bool parallel) { 2638 CMSSynchronousYieldRequest yr; 2639 assert(!tlab, "Can't deal with TLAB allocation"); 2640 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2641 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2642 if (GCExpandToAllocateDelayMillis > 0) { 2643 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2644 } 2645 return have_lock_and_allocate(word_size, tlab); 2646 } 2647 2648 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2649 size_t bytes, 2650 size_t expand_bytes, 2651 CMSExpansionCause::Cause cause) 2652 { 2653 2654 bool success = expand(bytes, expand_bytes); 2655 2656 // remember why we expanded; this information is used 2657 // by shouldConcurrentCollect() when making decisions on whether to start 2658 // a new CMS cycle. 2659 if (success) { 2660 set_expansion_cause(cause); 2661 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2662 } 2663 } 2664 2665 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2666 HeapWord* res = NULL; 2667 MutexLocker x(ParGCRareEvent_lock); 2668 while (true) { 2669 // Expansion by some other thread might make alloc OK now: 2670 res = ps->lab.alloc(word_sz); 2671 if (res != NULL) return res; 2672 // If there's not enough expansion space available, give up. 2673 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2674 return NULL; 2675 } 2676 // Otherwise, we try expansion. 2677 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2678 // Now go around the loop and try alloc again; 2679 // A competing par_promote might beat us to the expansion space, 2680 // so we may go around the loop again if promotion fails again. 2681 if (GCExpandToAllocateDelayMillis > 0) { 2682 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2683 } 2684 } 2685 } 2686 2687 2688 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2689 PromotionInfo* promo) { 2690 MutexLocker x(ParGCRareEvent_lock); 2691 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2692 while (true) { 2693 // Expansion by some other thread might make alloc OK now: 2694 if (promo->ensure_spooling_space()) { 2695 assert(promo->has_spooling_space(), 2696 "Post-condition of successful ensure_spooling_space()"); 2697 return true; 2698 } 2699 // If there's not enough expansion space available, give up. 2700 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2701 return false; 2702 } 2703 // Otherwise, we try expansion. 2704 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2705 // Now go around the loop and try alloc again; 2706 // A competing allocation might beat us to the expansion space, 2707 // so we may go around the loop again if allocation fails again. 2708 if (GCExpandToAllocateDelayMillis > 0) { 2709 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2710 } 2711 } 2712 } 2713 2714 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2715 // Only shrink if a compaction was done so that all the free space 2716 // in the generation is in a contiguous block at the end. 2717 if (did_compact()) { 2718 CardGeneration::shrink(bytes); 2719 } 2720 } 2721 2722 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2723 assert_locked_or_safepoint(Heap_lock); 2724 } 2725 2726 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2727 assert_locked_or_safepoint(Heap_lock); 2728 assert_lock_strong(freelistLock()); 2729 log_trace(gc)("Shrinking of CMS not yet implemented"); 2730 return; 2731 } 2732 2733 2734 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2735 // phases. 2736 class CMSPhaseAccounting: public StackObj { 2737 public: 2738 CMSPhaseAccounting(CMSCollector *collector, 2739 const char *title); 2740 ~CMSPhaseAccounting(); 2741 2742 private: 2743 CMSCollector *_collector; 2744 const char *_title; 2745 GCTraceConcTime(Info, gc) _trace_time; 2746 2747 public: 2748 // Not MT-safe; so do not pass around these StackObj's 2749 // where they may be accessed by other threads. 2750 double wallclock_millis() { 2751 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2752 } 2753 }; 2754 2755 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2756 const char *title) : 2757 _collector(collector), _title(title), _trace_time(title) { 2758 2759 _collector->resetYields(); 2760 _collector->resetTimer(); 2761 _collector->startTimer(); 2762 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2763 } 2764 2765 CMSPhaseAccounting::~CMSPhaseAccounting() { 2766 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2767 _collector->stopTimer(); 2768 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2769 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2770 } 2771 2772 // CMS work 2773 2774 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2775 class CMSParMarkTask : public AbstractGangTask { 2776 protected: 2777 CMSCollector* _collector; 2778 uint _n_workers; 2779 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2780 AbstractGangTask(name), 2781 _collector(collector), 2782 _n_workers(n_workers) {} 2783 // Work method in support of parallel rescan ... of young gen spaces 2784 void do_young_space_rescan(OopsInGenClosure* cl, 2785 ContiguousSpace* space, 2786 HeapWord** chunk_array, size_t chunk_top); 2787 void work_on_young_gen_roots(OopsInGenClosure* cl); 2788 }; 2789 2790 // Parallel initial mark task 2791 class CMSParInitialMarkTask: public CMSParMarkTask { 2792 StrongRootsScope* _strong_roots_scope; 2793 public: 2794 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2795 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2796 _strong_roots_scope(strong_roots_scope) {} 2797 void work(uint worker_id); 2798 }; 2799 2800 // Checkpoint the roots into this generation from outside 2801 // this generation. [Note this initial checkpoint need only 2802 // be approximate -- we'll do a catch up phase subsequently.] 2803 void CMSCollector::checkpointRootsInitial() { 2804 assert(_collectorState == InitialMarking, "Wrong collector state"); 2805 check_correct_thread_executing(); 2806 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2807 2808 save_heap_summary(); 2809 report_heap_summary(GCWhen::BeforeGC); 2810 2811 ReferenceProcessor* rp = ref_processor(); 2812 assert(_restart_addr == NULL, "Control point invariant"); 2813 { 2814 // acquire locks for subsequent manipulations 2815 MutexLockerEx x(bitMapLock(), 2816 Mutex::_no_safepoint_check_flag); 2817 checkpointRootsInitialWork(); 2818 // enable ("weak") refs discovery 2819 rp->enable_discovery(); 2820 _collectorState = Marking; 2821 } 2822 } 2823 2824 void CMSCollector::checkpointRootsInitialWork() { 2825 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2826 assert(_collectorState == InitialMarking, "just checking"); 2827 2828 // Already have locks. 2829 assert_lock_strong(bitMapLock()); 2830 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2831 2832 // Setup the verification and class unloading state for this 2833 // CMS collection cycle. 2834 setup_cms_unloading_and_verification_state(); 2835 2836 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2837 2838 // Reset all the PLAB chunk arrays if necessary. 2839 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2840 reset_survivor_plab_arrays(); 2841 } 2842 2843 ResourceMark rm; 2844 HandleMark hm; 2845 2846 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2847 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2848 2849 verify_work_stacks_empty(); 2850 verify_overflow_empty(); 2851 2852 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2853 // Update the saved marks which may affect the root scans. 2854 gch->save_marks(); 2855 2856 // weak reference processing has not started yet. 2857 ref_processor()->set_enqueuing_is_done(false); 2858 2859 // Need to remember all newly created CLDs, 2860 // so that we can guarantee that the remark finds them. 2861 ClassLoaderDataGraph::remember_new_clds(true); 2862 2863 // Whenever a CLD is found, it will be claimed before proceeding to mark 2864 // the klasses. The claimed marks need to be cleared before marking starts. 2865 ClassLoaderDataGraph::clear_claimed_marks(); 2866 2867 print_eden_and_survivor_chunk_arrays(); 2868 2869 { 2870 #if defined(COMPILER2) || INCLUDE_JVMCI 2871 DerivedPointerTableDeactivate dpt_deact; 2872 #endif 2873 if (CMSParallelInitialMarkEnabled) { 2874 // The parallel version. 2875 WorkGang* workers = gch->workers(); 2876 assert(workers != NULL, "Need parallel worker threads."); 2877 uint n_workers = workers->active_workers(); 2878 2879 StrongRootsScope srs(n_workers); 2880 2881 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2882 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2883 // If the total workers is greater than 1, then multiple workers 2884 // may be used at some time and the initialization has been set 2885 // such that the single threaded path cannot be used. 2886 if (workers->total_workers() > 1) { 2887 workers->run_task(&tsk); 2888 } else { 2889 tsk.work(0); 2890 } 2891 } else { 2892 // The serial version. 2893 CLDToOopClosure cld_closure(¬Older, true); 2894 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2895 2896 StrongRootsScope srs(1); 2897 2898 gch->cms_process_roots(&srs, 2899 true, // young gen as roots 2900 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2901 should_unload_classes(), 2902 ¬Older, 2903 &cld_closure); 2904 } 2905 } 2906 2907 // Clear mod-union table; it will be dirtied in the prologue of 2908 // CMS generation per each young generation collection. 2909 2910 assert(_modUnionTable.isAllClear(), 2911 "Was cleared in most recent final checkpoint phase" 2912 " or no bits are set in the gc_prologue before the start of the next " 2913 "subsequent marking phase."); 2914 2915 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2916 2917 // Save the end of the used_region of the constituent generations 2918 // to be used to limit the extent of sweep in each generation. 2919 save_sweep_limits(); 2920 verify_overflow_empty(); 2921 } 2922 2923 bool CMSCollector::markFromRoots() { 2924 // we might be tempted to assert that: 2925 // assert(!SafepointSynchronize::is_at_safepoint(), 2926 // "inconsistent argument?"); 2927 // However that wouldn't be right, because it's possible that 2928 // a safepoint is indeed in progress as a young generation 2929 // stop-the-world GC happens even as we mark in this generation. 2930 assert(_collectorState == Marking, "inconsistent state?"); 2931 check_correct_thread_executing(); 2932 verify_overflow_empty(); 2933 2934 // Weak ref discovery note: We may be discovering weak 2935 // refs in this generation concurrent (but interleaved) with 2936 // weak ref discovery by the young generation collector. 2937 2938 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2939 GCTraceCPUTime tcpu; 2940 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2941 bool res = markFromRootsWork(); 2942 if (res) { 2943 _collectorState = Precleaning; 2944 } else { // We failed and a foreground collection wants to take over 2945 assert(_foregroundGCIsActive, "internal state inconsistency"); 2946 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2947 log_debug(gc)("bailing out to foreground collection"); 2948 } 2949 verify_overflow_empty(); 2950 return res; 2951 } 2952 2953 bool CMSCollector::markFromRootsWork() { 2954 // iterate over marked bits in bit map, doing a full scan and mark 2955 // from these roots using the following algorithm: 2956 // . if oop is to the right of the current scan pointer, 2957 // mark corresponding bit (we'll process it later) 2958 // . else (oop is to left of current scan pointer) 2959 // push oop on marking stack 2960 // . drain the marking stack 2961 2962 // Note that when we do a marking step we need to hold the 2963 // bit map lock -- recall that direct allocation (by mutators) 2964 // and promotion (by the young generation collector) is also 2965 // marking the bit map. [the so-called allocate live policy.] 2966 // Because the implementation of bit map marking is not 2967 // robust wrt simultaneous marking of bits in the same word, 2968 // we need to make sure that there is no such interference 2969 // between concurrent such updates. 2970 2971 // already have locks 2972 assert_lock_strong(bitMapLock()); 2973 2974 verify_work_stacks_empty(); 2975 verify_overflow_empty(); 2976 bool result = false; 2977 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2978 result = do_marking_mt(); 2979 } else { 2980 result = do_marking_st(); 2981 } 2982 return result; 2983 } 2984 2985 // Forward decl 2986 class CMSConcMarkingTask; 2987 2988 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2989 CMSCollector* _collector; 2990 CMSConcMarkingTask* _task; 2991 public: 2992 virtual void yield(); 2993 2994 // "n_threads" is the number of threads to be terminated. 2995 // "queue_set" is a set of work queues of other threads. 2996 // "collector" is the CMS collector associated with this task terminator. 2997 // "yield" indicates whether we need the gang as a whole to yield. 2998 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2999 ParallelTaskTerminator(n_threads, queue_set), 3000 _collector(collector) { } 3001 3002 void set_task(CMSConcMarkingTask* task) { 3003 _task = task; 3004 } 3005 }; 3006 3007 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3008 CMSConcMarkingTask* _task; 3009 public: 3010 bool should_exit_termination(); 3011 void set_task(CMSConcMarkingTask* task) { 3012 _task = task; 3013 } 3014 }; 3015 3016 // MT Concurrent Marking Task 3017 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3018 CMSCollector* _collector; 3019 uint _n_workers; // requested/desired # workers 3020 bool _result; 3021 CompactibleFreeListSpace* _cms_space; 3022 char _pad_front[64]; // padding to ... 3023 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3024 char _pad_back[64]; 3025 HeapWord* _restart_addr; 3026 3027 // Exposed here for yielding support 3028 Mutex* const _bit_map_lock; 3029 3030 // The per thread work queues, available here for stealing 3031 OopTaskQueueSet* _task_queues; 3032 3033 // Termination (and yielding) support 3034 CMSConcMarkingTerminator _term; 3035 CMSConcMarkingTerminatorTerminator _term_term; 3036 3037 public: 3038 CMSConcMarkingTask(CMSCollector* collector, 3039 CompactibleFreeListSpace* cms_space, 3040 YieldingFlexibleWorkGang* workers, 3041 OopTaskQueueSet* task_queues): 3042 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3043 _collector(collector), 3044 _cms_space(cms_space), 3045 _n_workers(0), _result(true), 3046 _task_queues(task_queues), 3047 _term(_n_workers, task_queues, _collector), 3048 _bit_map_lock(collector->bitMapLock()) 3049 { 3050 _requested_size = _n_workers; 3051 _term.set_task(this); 3052 _term_term.set_task(this); 3053 _restart_addr = _global_finger = _cms_space->bottom(); 3054 } 3055 3056 3057 OopTaskQueueSet* task_queues() { return _task_queues; } 3058 3059 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3060 3061 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3062 3063 CMSConcMarkingTerminator* terminator() { return &_term; } 3064 3065 virtual void set_for_termination(uint active_workers) { 3066 terminator()->reset_for_reuse(active_workers); 3067 } 3068 3069 void work(uint worker_id); 3070 bool should_yield() { 3071 return ConcurrentMarkSweepThread::should_yield() 3072 && !_collector->foregroundGCIsActive(); 3073 } 3074 3075 virtual void coordinator_yield(); // stuff done by coordinator 3076 bool result() { return _result; } 3077 3078 void reset(HeapWord* ra) { 3079 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3080 _restart_addr = _global_finger = ra; 3081 _term.reset_for_reuse(); 3082 } 3083 3084 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3085 OopTaskQueue* work_q); 3086 3087 private: 3088 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3089 void do_work_steal(int i); 3090 void bump_global_finger(HeapWord* f); 3091 }; 3092 3093 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3094 assert(_task != NULL, "Error"); 3095 return _task->yielding(); 3096 // Note that we do not need the disjunct || _task->should_yield() above 3097 // because we want terminating threads to yield only if the task 3098 // is already in the midst of yielding, which happens only after at least one 3099 // thread has yielded. 3100 } 3101 3102 void CMSConcMarkingTerminator::yield() { 3103 if (_task->should_yield()) { 3104 _task->yield(); 3105 } else { 3106 ParallelTaskTerminator::yield(); 3107 } 3108 } 3109 3110 //////////////////////////////////////////////////////////////// 3111 // Concurrent Marking Algorithm Sketch 3112 //////////////////////////////////////////////////////////////// 3113 // Until all tasks exhausted (both spaces): 3114 // -- claim next available chunk 3115 // -- bump global finger via CAS 3116 // -- find first object that starts in this chunk 3117 // and start scanning bitmap from that position 3118 // -- scan marked objects for oops 3119 // -- CAS-mark target, and if successful: 3120 // . if target oop is above global finger (volatile read) 3121 // nothing to do 3122 // . if target oop is in chunk and above local finger 3123 // then nothing to do 3124 // . else push on work-queue 3125 // -- Deal with possible overflow issues: 3126 // . local work-queue overflow causes stuff to be pushed on 3127 // global (common) overflow queue 3128 // . always first empty local work queue 3129 // . then get a batch of oops from global work queue if any 3130 // . then do work stealing 3131 // -- When all tasks claimed (both spaces) 3132 // and local work queue empty, 3133 // then in a loop do: 3134 // . check global overflow stack; steal a batch of oops and trace 3135 // . try to steal from other threads oif GOS is empty 3136 // . if neither is available, offer termination 3137 // -- Terminate and return result 3138 // 3139 void CMSConcMarkingTask::work(uint worker_id) { 3140 elapsedTimer _timer; 3141 ResourceMark rm; 3142 HandleMark hm; 3143 3144 DEBUG_ONLY(_collector->verify_overflow_empty();) 3145 3146 // Before we begin work, our work queue should be empty 3147 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3148 // Scan the bitmap covering _cms_space, tracing through grey objects. 3149 _timer.start(); 3150 do_scan_and_mark(worker_id, _cms_space); 3151 _timer.stop(); 3152 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3153 3154 // ... do work stealing 3155 _timer.reset(); 3156 _timer.start(); 3157 do_work_steal(worker_id); 3158 _timer.stop(); 3159 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3160 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3161 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3162 // Note that under the current task protocol, the 3163 // following assertion is true even of the spaces 3164 // expanded since the completion of the concurrent 3165 // marking. XXX This will likely change under a strict 3166 // ABORT semantics. 3167 // After perm removal the comparison was changed to 3168 // greater than or equal to from strictly greater than. 3169 // Before perm removal the highest address sweep would 3170 // have been at the end of perm gen but now is at the 3171 // end of the tenured gen. 3172 assert(_global_finger >= _cms_space->end(), 3173 "All tasks have been completed"); 3174 DEBUG_ONLY(_collector->verify_overflow_empty();) 3175 } 3176 3177 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3178 HeapWord* read = _global_finger; 3179 HeapWord* cur = read; 3180 while (f > read) { 3181 cur = read; 3182 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3183 if (cur == read) { 3184 // our cas succeeded 3185 assert(_global_finger >= f, "protocol consistency"); 3186 break; 3187 } 3188 } 3189 } 3190 3191 // This is really inefficient, and should be redone by 3192 // using (not yet available) block-read and -write interfaces to the 3193 // stack and the work_queue. XXX FIX ME !!! 3194 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3195 OopTaskQueue* work_q) { 3196 // Fast lock-free check 3197 if (ovflw_stk->length() == 0) { 3198 return false; 3199 } 3200 assert(work_q->size() == 0, "Shouldn't steal"); 3201 MutexLockerEx ml(ovflw_stk->par_lock(), 3202 Mutex::_no_safepoint_check_flag); 3203 // Grab up to 1/4 the size of the work queue 3204 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3205 (size_t)ParGCDesiredObjsFromOverflowList); 3206 num = MIN2(num, ovflw_stk->length()); 3207 for (int i = (int) num; i > 0; i--) { 3208 oop cur = ovflw_stk->pop(); 3209 assert(cur != NULL, "Counted wrong?"); 3210 work_q->push(cur); 3211 } 3212 return num > 0; 3213 } 3214 3215 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3216 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3217 int n_tasks = pst->n_tasks(); 3218 // We allow that there may be no tasks to do here because 3219 // we are restarting after a stack overflow. 3220 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3221 uint nth_task = 0; 3222 3223 HeapWord* aligned_start = sp->bottom(); 3224 if (sp->used_region().contains(_restart_addr)) { 3225 // Align down to a card boundary for the start of 0th task 3226 // for this space. 3227 aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size); 3228 } 3229 3230 size_t chunk_size = sp->marking_task_size(); 3231 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3232 // Having claimed the nth task in this space, 3233 // compute the chunk that it corresponds to: 3234 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3235 aligned_start + (nth_task+1)*chunk_size); 3236 // Try and bump the global finger via a CAS; 3237 // note that we need to do the global finger bump 3238 // _before_ taking the intersection below, because 3239 // the task corresponding to that region will be 3240 // deemed done even if the used_region() expands 3241 // because of allocation -- as it almost certainly will 3242 // during start-up while the threads yield in the 3243 // closure below. 3244 HeapWord* finger = span.end(); 3245 bump_global_finger(finger); // atomically 3246 // There are null tasks here corresponding to chunks 3247 // beyond the "top" address of the space. 3248 span = span.intersection(sp->used_region()); 3249 if (!span.is_empty()) { // Non-null task 3250 HeapWord* prev_obj; 3251 assert(!span.contains(_restart_addr) || nth_task == 0, 3252 "Inconsistency"); 3253 if (nth_task == 0) { 3254 // For the 0th task, we'll not need to compute a block_start. 3255 if (span.contains(_restart_addr)) { 3256 // In the case of a restart because of stack overflow, 3257 // we might additionally skip a chunk prefix. 3258 prev_obj = _restart_addr; 3259 } else { 3260 prev_obj = span.start(); 3261 } 3262 } else { 3263 // We want to skip the first object because 3264 // the protocol is to scan any object in its entirety 3265 // that _starts_ in this span; a fortiori, any 3266 // object starting in an earlier span is scanned 3267 // as part of an earlier claimed task. 3268 // Below we use the "careful" version of block_start 3269 // so we do not try to navigate uninitialized objects. 3270 prev_obj = sp->block_start_careful(span.start()); 3271 // Below we use a variant of block_size that uses the 3272 // Printezis bits to avoid waiting for allocated 3273 // objects to become initialized/parsable. 3274 while (prev_obj < span.start()) { 3275 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3276 if (sz > 0) { 3277 prev_obj += sz; 3278 } else { 3279 // In this case we may end up doing a bit of redundant 3280 // scanning, but that appears unavoidable, short of 3281 // locking the free list locks; see bug 6324141. 3282 break; 3283 } 3284 } 3285 } 3286 if (prev_obj < span.end()) { 3287 MemRegion my_span = MemRegion(prev_obj, span.end()); 3288 // Do the marking work within a non-empty span -- 3289 // the last argument to the constructor indicates whether the 3290 // iteration should be incremental with periodic yields. 3291 ParMarkFromRootsClosure cl(this, _collector, my_span, 3292 &_collector->_markBitMap, 3293 work_queue(i), 3294 &_collector->_markStack); 3295 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3296 } // else nothing to do for this task 3297 } // else nothing to do for this task 3298 } 3299 // We'd be tempted to assert here that since there are no 3300 // more tasks left to claim in this space, the global_finger 3301 // must exceed space->top() and a fortiori space->end(). However, 3302 // that would not quite be correct because the bumping of 3303 // global_finger occurs strictly after the claiming of a task, 3304 // so by the time we reach here the global finger may not yet 3305 // have been bumped up by the thread that claimed the last 3306 // task. 3307 pst->all_tasks_completed(); 3308 } 3309 3310 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3311 private: 3312 CMSCollector* _collector; 3313 CMSConcMarkingTask* _task; 3314 MemRegion _span; 3315 CMSBitMap* _bit_map; 3316 CMSMarkStack* _overflow_stack; 3317 OopTaskQueue* _work_queue; 3318 protected: 3319 DO_OOP_WORK_DEFN 3320 public: 3321 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3322 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3323 MetadataAwareOopClosure(collector->ref_processor()), 3324 _collector(collector), 3325 _task(task), 3326 _span(collector->_span), 3327 _work_queue(work_queue), 3328 _bit_map(bit_map), 3329 _overflow_stack(overflow_stack) 3330 { } 3331 virtual void do_oop(oop* p); 3332 virtual void do_oop(narrowOop* p); 3333 3334 void trim_queue(size_t max); 3335 void handle_stack_overflow(HeapWord* lost); 3336 void do_yield_check() { 3337 if (_task->should_yield()) { 3338 _task->yield(); 3339 } 3340 } 3341 }; 3342 3343 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3344 3345 // Grey object scanning during work stealing phase -- 3346 // the salient assumption here is that any references 3347 // that are in these stolen objects being scanned must 3348 // already have been initialized (else they would not have 3349 // been published), so we do not need to check for 3350 // uninitialized objects before pushing here. 3351 void ParConcMarkingClosure::do_oop(oop obj) { 3352 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3353 HeapWord* addr = (HeapWord*)obj; 3354 // Check if oop points into the CMS generation 3355 // and is not marked 3356 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3357 // a white object ... 3358 // If we manage to "claim" the object, by being the 3359 // first thread to mark it, then we push it on our 3360 // marking stack 3361 if (_bit_map->par_mark(addr)) { // ... now grey 3362 // push on work queue (grey set) 3363 bool simulate_overflow = false; 3364 NOT_PRODUCT( 3365 if (CMSMarkStackOverflowALot && 3366 _collector->simulate_overflow()) { 3367 // simulate a stack overflow 3368 simulate_overflow = true; 3369 } 3370 ) 3371 if (simulate_overflow || 3372 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3373 // stack overflow 3374 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3375 // We cannot assert that the overflow stack is full because 3376 // it may have been emptied since. 3377 assert(simulate_overflow || 3378 _work_queue->size() == _work_queue->max_elems(), 3379 "Else push should have succeeded"); 3380 handle_stack_overflow(addr); 3381 } 3382 } // Else, some other thread got there first 3383 do_yield_check(); 3384 } 3385 } 3386 3387 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3388 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3389 3390 void ParConcMarkingClosure::trim_queue(size_t max) { 3391 while (_work_queue->size() > max) { 3392 oop new_oop; 3393 if (_work_queue->pop_local(new_oop)) { 3394 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3395 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3396 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3397 new_oop->oop_iterate(this); // do_oop() above 3398 do_yield_check(); 3399 } 3400 } 3401 } 3402 3403 // Upon stack overflow, we discard (part of) the stack, 3404 // remembering the least address amongst those discarded 3405 // in CMSCollector's _restart_address. 3406 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3407 // We need to do this under a mutex to prevent other 3408 // workers from interfering with the work done below. 3409 MutexLockerEx ml(_overflow_stack->par_lock(), 3410 Mutex::_no_safepoint_check_flag); 3411 // Remember the least grey address discarded 3412 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3413 _collector->lower_restart_addr(ra); 3414 _overflow_stack->reset(); // discard stack contents 3415 _overflow_stack->expand(); // expand the stack if possible 3416 } 3417 3418 3419 void CMSConcMarkingTask::do_work_steal(int i) { 3420 OopTaskQueue* work_q = work_queue(i); 3421 oop obj_to_scan; 3422 CMSBitMap* bm = &(_collector->_markBitMap); 3423 CMSMarkStack* ovflw = &(_collector->_markStack); 3424 int* seed = _collector->hash_seed(i); 3425 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3426 while (true) { 3427 cl.trim_queue(0); 3428 assert(work_q->size() == 0, "Should have been emptied above"); 3429 if (get_work_from_overflow_stack(ovflw, work_q)) { 3430 // Can't assert below because the work obtained from the 3431 // overflow stack may already have been stolen from us. 3432 // assert(work_q->size() > 0, "Work from overflow stack"); 3433 continue; 3434 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3435 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3436 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3437 obj_to_scan->oop_iterate(&cl); 3438 } else if (terminator()->offer_termination(&_term_term)) { 3439 assert(work_q->size() == 0, "Impossible!"); 3440 break; 3441 } else if (yielding() || should_yield()) { 3442 yield(); 3443 } 3444 } 3445 } 3446 3447 // This is run by the CMS (coordinator) thread. 3448 void CMSConcMarkingTask::coordinator_yield() { 3449 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3450 "CMS thread should hold CMS token"); 3451 // First give up the locks, then yield, then re-lock 3452 // We should probably use a constructor/destructor idiom to 3453 // do this unlock/lock or modify the MutexUnlocker class to 3454 // serve our purpose. XXX 3455 assert_lock_strong(_bit_map_lock); 3456 _bit_map_lock->unlock(); 3457 ConcurrentMarkSweepThread::desynchronize(true); 3458 _collector->stopTimer(); 3459 _collector->incrementYields(); 3460 3461 // It is possible for whichever thread initiated the yield request 3462 // not to get a chance to wake up and take the bitmap lock between 3463 // this thread releasing it and reacquiring it. So, while the 3464 // should_yield() flag is on, let's sleep for a bit to give the 3465 // other thread a chance to wake up. The limit imposed on the number 3466 // of iterations is defensive, to avoid any unforseen circumstances 3467 // putting us into an infinite loop. Since it's always been this 3468 // (coordinator_yield()) method that was observed to cause the 3469 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3470 // which is by default non-zero. For the other seven methods that 3471 // also perform the yield operation, as are using a different 3472 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3473 // can enable the sleeping for those methods too, if necessary. 3474 // See 6442774. 3475 // 3476 // We really need to reconsider the synchronization between the GC 3477 // thread and the yield-requesting threads in the future and we 3478 // should really use wait/notify, which is the recommended 3479 // way of doing this type of interaction. Additionally, we should 3480 // consolidate the eight methods that do the yield operation and they 3481 // are almost identical into one for better maintainability and 3482 // readability. See 6445193. 3483 // 3484 // Tony 2006.06.29 3485 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3486 ConcurrentMarkSweepThread::should_yield() && 3487 !CMSCollector::foregroundGCIsActive(); ++i) { 3488 os::sleep(Thread::current(), 1, false); 3489 } 3490 3491 ConcurrentMarkSweepThread::synchronize(true); 3492 _bit_map_lock->lock_without_safepoint_check(); 3493 _collector->startTimer(); 3494 } 3495 3496 bool CMSCollector::do_marking_mt() { 3497 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3498 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3499 conc_workers()->active_workers(), 3500 Threads::number_of_non_daemon_threads()); 3501 num_workers = conc_workers()->update_active_workers(num_workers); 3502 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3503 3504 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3505 3506 CMSConcMarkingTask tsk(this, 3507 cms_space, 3508 conc_workers(), 3509 task_queues()); 3510 3511 // Since the actual number of workers we get may be different 3512 // from the number we requested above, do we need to do anything different 3513 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3514 // class?? XXX 3515 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3516 3517 // Refs discovery is already non-atomic. 3518 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3519 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3520 conc_workers()->start_task(&tsk); 3521 while (tsk.yielded()) { 3522 tsk.coordinator_yield(); 3523 conc_workers()->continue_task(&tsk); 3524 } 3525 // If the task was aborted, _restart_addr will be non-NULL 3526 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3527 while (_restart_addr != NULL) { 3528 // XXX For now we do not make use of ABORTED state and have not 3529 // yet implemented the right abort semantics (even in the original 3530 // single-threaded CMS case). That needs some more investigation 3531 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3532 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3533 // If _restart_addr is non-NULL, a marking stack overflow 3534 // occurred; we need to do a fresh marking iteration from the 3535 // indicated restart address. 3536 if (_foregroundGCIsActive) { 3537 // We may be running into repeated stack overflows, having 3538 // reached the limit of the stack size, while making very 3539 // slow forward progress. It may be best to bail out and 3540 // let the foreground collector do its job. 3541 // Clear _restart_addr, so that foreground GC 3542 // works from scratch. This avoids the headache of 3543 // a "rescan" which would otherwise be needed because 3544 // of the dirty mod union table & card table. 3545 _restart_addr = NULL; 3546 return false; 3547 } 3548 // Adjust the task to restart from _restart_addr 3549 tsk.reset(_restart_addr); 3550 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3551 _restart_addr); 3552 _restart_addr = NULL; 3553 // Get the workers going again 3554 conc_workers()->start_task(&tsk); 3555 while (tsk.yielded()) { 3556 tsk.coordinator_yield(); 3557 conc_workers()->continue_task(&tsk); 3558 } 3559 } 3560 assert(tsk.completed(), "Inconsistency"); 3561 assert(tsk.result() == true, "Inconsistency"); 3562 return true; 3563 } 3564 3565 bool CMSCollector::do_marking_st() { 3566 ResourceMark rm; 3567 HandleMark hm; 3568 3569 // Temporarily make refs discovery single threaded (non-MT) 3570 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3571 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3572 &_markStack, CMSYield); 3573 // the last argument to iterate indicates whether the iteration 3574 // should be incremental with periodic yields. 3575 _markBitMap.iterate(&markFromRootsClosure); 3576 // If _restart_addr is non-NULL, a marking stack overflow 3577 // occurred; we need to do a fresh iteration from the 3578 // indicated restart address. 3579 while (_restart_addr != NULL) { 3580 if (_foregroundGCIsActive) { 3581 // We may be running into repeated stack overflows, having 3582 // reached the limit of the stack size, while making very 3583 // slow forward progress. It may be best to bail out and 3584 // let the foreground collector do its job. 3585 // Clear _restart_addr, so that foreground GC 3586 // works from scratch. This avoids the headache of 3587 // a "rescan" which would otherwise be needed because 3588 // of the dirty mod union table & card table. 3589 _restart_addr = NULL; 3590 return false; // indicating failure to complete marking 3591 } 3592 // Deal with stack overflow: 3593 // we restart marking from _restart_addr 3594 HeapWord* ra = _restart_addr; 3595 markFromRootsClosure.reset(ra); 3596 _restart_addr = NULL; 3597 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3598 } 3599 return true; 3600 } 3601 3602 void CMSCollector::preclean() { 3603 check_correct_thread_executing(); 3604 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3605 verify_work_stacks_empty(); 3606 verify_overflow_empty(); 3607 _abort_preclean = false; 3608 if (CMSPrecleaningEnabled) { 3609 if (!CMSEdenChunksRecordAlways) { 3610 _eden_chunk_index = 0; 3611 } 3612 size_t used = get_eden_used(); 3613 size_t capacity = get_eden_capacity(); 3614 // Don't start sampling unless we will get sufficiently 3615 // many samples. 3616 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3617 * CMSScheduleRemarkEdenPenetration)) { 3618 _start_sampling = true; 3619 } else { 3620 _start_sampling = false; 3621 } 3622 GCTraceCPUTime tcpu; 3623 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3624 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3625 } 3626 CMSTokenSync x(true); // is cms thread 3627 if (CMSPrecleaningEnabled) { 3628 sample_eden(); 3629 _collectorState = AbortablePreclean; 3630 } else { 3631 _collectorState = FinalMarking; 3632 } 3633 verify_work_stacks_empty(); 3634 verify_overflow_empty(); 3635 } 3636 3637 // Try and schedule the remark such that young gen 3638 // occupancy is CMSScheduleRemarkEdenPenetration %. 3639 void CMSCollector::abortable_preclean() { 3640 check_correct_thread_executing(); 3641 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3642 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3643 3644 // If Eden's current occupancy is below this threshold, 3645 // immediately schedule the remark; else preclean 3646 // past the next scavenge in an effort to 3647 // schedule the pause as described above. By choosing 3648 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3649 // we will never do an actual abortable preclean cycle. 3650 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3651 GCTraceCPUTime tcpu; 3652 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3653 // We need more smarts in the abortable preclean 3654 // loop below to deal with cases where allocation 3655 // in young gen is very very slow, and our precleaning 3656 // is running a losing race against a horde of 3657 // mutators intent on flooding us with CMS updates 3658 // (dirty cards). 3659 // One, admittedly dumb, strategy is to give up 3660 // after a certain number of abortable precleaning loops 3661 // or after a certain maximum time. We want to make 3662 // this smarter in the next iteration. 3663 // XXX FIX ME!!! YSR 3664 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3665 while (!(should_abort_preclean() || 3666 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3667 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3668 cumworkdone += workdone; 3669 loops++; 3670 // Voluntarily terminate abortable preclean phase if we have 3671 // been at it for too long. 3672 if ((CMSMaxAbortablePrecleanLoops != 0) && 3673 loops >= CMSMaxAbortablePrecleanLoops) { 3674 log_debug(gc)(" CMS: abort preclean due to loops "); 3675 break; 3676 } 3677 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3678 log_debug(gc)(" CMS: abort preclean due to time "); 3679 break; 3680 } 3681 // If we are doing little work each iteration, we should 3682 // take a short break. 3683 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3684 // Sleep for some time, waiting for work to accumulate 3685 stopTimer(); 3686 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3687 startTimer(); 3688 waited++; 3689 } 3690 } 3691 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3692 loops, waited, cumworkdone); 3693 } 3694 CMSTokenSync x(true); // is cms thread 3695 if (_collectorState != Idling) { 3696 assert(_collectorState == AbortablePreclean, 3697 "Spontaneous state transition?"); 3698 _collectorState = FinalMarking; 3699 } // Else, a foreground collection completed this CMS cycle. 3700 return; 3701 } 3702 3703 // Respond to an Eden sampling opportunity 3704 void CMSCollector::sample_eden() { 3705 // Make sure a young gc cannot sneak in between our 3706 // reading and recording of a sample. 3707 assert(Thread::current()->is_ConcurrentGC_thread(), 3708 "Only the cms thread may collect Eden samples"); 3709 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3710 "Should collect samples while holding CMS token"); 3711 if (!_start_sampling) { 3712 return; 3713 } 3714 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3715 // is populated by the young generation. 3716 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3717 if (_eden_chunk_index < _eden_chunk_capacity) { 3718 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3719 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3720 "Unexpected state of Eden"); 3721 // We'd like to check that what we just sampled is an oop-start address; 3722 // however, we cannot do that here since the object may not yet have been 3723 // initialized. So we'll instead do the check when we _use_ this sample 3724 // later. 3725 if (_eden_chunk_index == 0 || 3726 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3727 _eden_chunk_array[_eden_chunk_index-1]) 3728 >= CMSSamplingGrain)) { 3729 _eden_chunk_index++; // commit sample 3730 } 3731 } 3732 } 3733 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3734 size_t used = get_eden_used(); 3735 size_t capacity = get_eden_capacity(); 3736 assert(used <= capacity, "Unexpected state of Eden"); 3737 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3738 _abort_preclean = true; 3739 } 3740 } 3741 } 3742 3743 3744 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3745 assert(_collectorState == Precleaning || 3746 _collectorState == AbortablePreclean, "incorrect state"); 3747 ResourceMark rm; 3748 HandleMark hm; 3749 3750 // Precleaning is currently not MT but the reference processor 3751 // may be set for MT. Disable it temporarily here. 3752 ReferenceProcessor* rp = ref_processor(); 3753 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3754 3755 // Do one pass of scrubbing the discovered reference lists 3756 // to remove any reference objects with strongly-reachable 3757 // referents. 3758 if (clean_refs) { 3759 CMSPrecleanRefsYieldClosure yield_cl(this); 3760 assert(rp->span().equals(_span), "Spans should be equal"); 3761 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3762 &_markStack, true /* preclean */); 3763 CMSDrainMarkingStackClosure complete_trace(this, 3764 _span, &_markBitMap, &_markStack, 3765 &keep_alive, true /* preclean */); 3766 3767 // We don't want this step to interfere with a young 3768 // collection because we don't want to take CPU 3769 // or memory bandwidth away from the young GC threads 3770 // (which may be as many as there are CPUs). 3771 // Note that we don't need to protect ourselves from 3772 // interference with mutators because they can't 3773 // manipulate the discovered reference lists nor affect 3774 // the computed reachability of the referents, the 3775 // only properties manipulated by the precleaning 3776 // of these reference lists. 3777 stopTimer(); 3778 CMSTokenSyncWithLocks x(true /* is cms thread */, 3779 bitMapLock()); 3780 startTimer(); 3781 sample_eden(); 3782 3783 // The following will yield to allow foreground 3784 // collection to proceed promptly. XXX YSR: 3785 // The code in this method may need further 3786 // tweaking for better performance and some restructuring 3787 // for cleaner interfaces. 3788 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3789 rp->preclean_discovered_references( 3790 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3791 gc_timer); 3792 } 3793 3794 if (clean_survivor) { // preclean the active survivor space(s) 3795 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3796 &_markBitMap, &_modUnionTable, 3797 &_markStack, true /* precleaning phase */); 3798 stopTimer(); 3799 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3800 bitMapLock()); 3801 startTimer(); 3802 unsigned int before_count = 3803 GenCollectedHeap::heap()->total_collections(); 3804 SurvivorSpacePrecleanClosure 3805 sss_cl(this, _span, &_markBitMap, &_markStack, 3806 &pam_cl, before_count, CMSYield); 3807 _young_gen->from()->object_iterate_careful(&sss_cl); 3808 _young_gen->to()->object_iterate_careful(&sss_cl); 3809 } 3810 MarkRefsIntoAndScanClosure 3811 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3812 &_markStack, this, CMSYield, 3813 true /* precleaning phase */); 3814 // CAUTION: The following closure has persistent state that may need to 3815 // be reset upon a decrease in the sequence of addresses it 3816 // processes. 3817 ScanMarkedObjectsAgainCarefullyClosure 3818 smoac_cl(this, _span, 3819 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3820 3821 // Preclean dirty cards in ModUnionTable and CardTable using 3822 // appropriate convergence criterion; 3823 // repeat CMSPrecleanIter times unless we find that 3824 // we are losing. 3825 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3826 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3827 "Bad convergence multiplier"); 3828 assert(CMSPrecleanThreshold >= 100, 3829 "Unreasonably low CMSPrecleanThreshold"); 3830 3831 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3832 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3833 numIter < CMSPrecleanIter; 3834 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3835 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3836 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3837 // Either there are very few dirty cards, so re-mark 3838 // pause will be small anyway, or our pre-cleaning isn't 3839 // that much faster than the rate at which cards are being 3840 // dirtied, so we might as well stop and re-mark since 3841 // precleaning won't improve our re-mark time by much. 3842 if (curNumCards <= CMSPrecleanThreshold || 3843 (numIter > 0 && 3844 (curNumCards * CMSPrecleanDenominator > 3845 lastNumCards * CMSPrecleanNumerator))) { 3846 numIter++; 3847 cumNumCards += curNumCards; 3848 break; 3849 } 3850 } 3851 3852 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3853 3854 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3855 cumNumCards += curNumCards; 3856 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3857 curNumCards, cumNumCards, numIter); 3858 return cumNumCards; // as a measure of useful work done 3859 } 3860 3861 // PRECLEANING NOTES: 3862 // Precleaning involves: 3863 // . reading the bits of the modUnionTable and clearing the set bits. 3864 // . For the cards corresponding to the set bits, we scan the 3865 // objects on those cards. This means we need the free_list_lock 3866 // so that we can safely iterate over the CMS space when scanning 3867 // for oops. 3868 // . When we scan the objects, we'll be both reading and setting 3869 // marks in the marking bit map, so we'll need the marking bit map. 3870 // . For protecting _collector_state transitions, we take the CGC_lock. 3871 // Note that any races in the reading of of card table entries by the 3872 // CMS thread on the one hand and the clearing of those entries by the 3873 // VM thread or the setting of those entries by the mutator threads on the 3874 // other are quite benign. However, for efficiency it makes sense to keep 3875 // the VM thread from racing with the CMS thread while the latter is 3876 // dirty card info to the modUnionTable. We therefore also use the 3877 // CGC_lock to protect the reading of the card table and the mod union 3878 // table by the CM thread. 3879 // . We run concurrently with mutator updates, so scanning 3880 // needs to be done carefully -- we should not try to scan 3881 // potentially uninitialized objects. 3882 // 3883 // Locking strategy: While holding the CGC_lock, we scan over and 3884 // reset a maximal dirty range of the mod union / card tables, then lock 3885 // the free_list_lock and bitmap lock to do a full marking, then 3886 // release these locks; and repeat the cycle. This allows for a 3887 // certain amount of fairness in the sharing of these locks between 3888 // the CMS collector on the one hand, and the VM thread and the 3889 // mutators on the other. 3890 3891 // NOTE: preclean_mod_union_table() and preclean_card_table() 3892 // further below are largely identical; if you need to modify 3893 // one of these methods, please check the other method too. 3894 3895 size_t CMSCollector::preclean_mod_union_table( 3896 ConcurrentMarkSweepGeneration* old_gen, 3897 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3898 verify_work_stacks_empty(); 3899 verify_overflow_empty(); 3900 3901 // strategy: starting with the first card, accumulate contiguous 3902 // ranges of dirty cards; clear these cards, then scan the region 3903 // covered by these cards. 3904 3905 // Since all of the MUT is committed ahead, we can just use 3906 // that, in case the generations expand while we are precleaning. 3907 // It might also be fine to just use the committed part of the 3908 // generation, but we might potentially miss cards when the 3909 // generation is rapidly expanding while we are in the midst 3910 // of precleaning. 3911 HeapWord* startAddr = old_gen->reserved().start(); 3912 HeapWord* endAddr = old_gen->reserved().end(); 3913 3914 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3915 3916 size_t numDirtyCards, cumNumDirtyCards; 3917 HeapWord *nextAddr, *lastAddr; 3918 for (cumNumDirtyCards = numDirtyCards = 0, 3919 nextAddr = lastAddr = startAddr; 3920 nextAddr < endAddr; 3921 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3922 3923 ResourceMark rm; 3924 HandleMark hm; 3925 3926 MemRegion dirtyRegion; 3927 { 3928 stopTimer(); 3929 // Potential yield point 3930 CMSTokenSync ts(true); 3931 startTimer(); 3932 sample_eden(); 3933 // Get dirty region starting at nextOffset (inclusive), 3934 // simultaneously clearing it. 3935 dirtyRegion = 3936 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3937 assert(dirtyRegion.start() >= nextAddr, 3938 "returned region inconsistent?"); 3939 } 3940 // Remember where the next search should begin. 3941 // The returned region (if non-empty) is a right open interval, 3942 // so lastOffset is obtained from the right end of that 3943 // interval. 3944 lastAddr = dirtyRegion.end(); 3945 // Should do something more transparent and less hacky XXX 3946 numDirtyCards = 3947 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3948 3949 // We'll scan the cards in the dirty region (with periodic 3950 // yields for foreground GC as needed). 3951 if (!dirtyRegion.is_empty()) { 3952 assert(numDirtyCards > 0, "consistency check"); 3953 HeapWord* stop_point = NULL; 3954 stopTimer(); 3955 // Potential yield point 3956 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3957 bitMapLock()); 3958 startTimer(); 3959 { 3960 verify_work_stacks_empty(); 3961 verify_overflow_empty(); 3962 sample_eden(); 3963 stop_point = 3964 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3965 } 3966 if (stop_point != NULL) { 3967 // The careful iteration stopped early either because it found an 3968 // uninitialized object, or because we were in the midst of an 3969 // "abortable preclean", which should now be aborted. Redirty 3970 // the bits corresponding to the partially-scanned or unscanned 3971 // cards. We'll either restart at the next block boundary or 3972 // abort the preclean. 3973 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3974 "Should only be AbortablePreclean."); 3975 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3976 if (should_abort_preclean()) { 3977 break; // out of preclean loop 3978 } else { 3979 // Compute the next address at which preclean should pick up; 3980 // might need bitMapLock in order to read P-bits. 3981 lastAddr = next_card_start_after_block(stop_point); 3982 } 3983 } 3984 } else { 3985 assert(lastAddr == endAddr, "consistency check"); 3986 assert(numDirtyCards == 0, "consistency check"); 3987 break; 3988 } 3989 } 3990 verify_work_stacks_empty(); 3991 verify_overflow_empty(); 3992 return cumNumDirtyCards; 3993 } 3994 3995 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3996 // below are largely identical; if you need to modify 3997 // one of these methods, please check the other method too. 3998 3999 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4000 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4001 // strategy: it's similar to precleamModUnionTable above, in that 4002 // we accumulate contiguous ranges of dirty cards, mark these cards 4003 // precleaned, then scan the region covered by these cards. 4004 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4005 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4006 4007 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4008 4009 size_t numDirtyCards, cumNumDirtyCards; 4010 HeapWord *lastAddr, *nextAddr; 4011 4012 for (cumNumDirtyCards = numDirtyCards = 0, 4013 nextAddr = lastAddr = startAddr; 4014 nextAddr < endAddr; 4015 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4016 4017 ResourceMark rm; 4018 HandleMark hm; 4019 4020 MemRegion dirtyRegion; 4021 { 4022 // See comments in "Precleaning notes" above on why we 4023 // do this locking. XXX Could the locking overheads be 4024 // too high when dirty cards are sparse? [I don't think so.] 4025 stopTimer(); 4026 CMSTokenSync x(true); // is cms thread 4027 startTimer(); 4028 sample_eden(); 4029 // Get and clear dirty region from card table 4030 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4031 MemRegion(nextAddr, endAddr), 4032 true, 4033 CardTableModRefBS::precleaned_card_val()); 4034 4035 assert(dirtyRegion.start() >= nextAddr, 4036 "returned region inconsistent?"); 4037 } 4038 lastAddr = dirtyRegion.end(); 4039 numDirtyCards = 4040 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4041 4042 if (!dirtyRegion.is_empty()) { 4043 stopTimer(); 4044 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4045 startTimer(); 4046 sample_eden(); 4047 verify_work_stacks_empty(); 4048 verify_overflow_empty(); 4049 HeapWord* stop_point = 4050 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4051 if (stop_point != NULL) { 4052 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4053 "Should only be AbortablePreclean."); 4054 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4055 if (should_abort_preclean()) { 4056 break; // out of preclean loop 4057 } else { 4058 // Compute the next address at which preclean should pick up. 4059 lastAddr = next_card_start_after_block(stop_point); 4060 } 4061 } 4062 } else { 4063 break; 4064 } 4065 } 4066 verify_work_stacks_empty(); 4067 verify_overflow_empty(); 4068 return cumNumDirtyCards; 4069 } 4070 4071 class PrecleanCLDClosure : public CLDClosure { 4072 MetadataAwareOopsInGenClosure* _cm_closure; 4073 public: 4074 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4075 void do_cld(ClassLoaderData* cld) { 4076 if (cld->has_accumulated_modified_oops()) { 4077 cld->clear_accumulated_modified_oops(); 4078 4079 _cm_closure->do_cld(cld); 4080 } 4081 } 4082 }; 4083 4084 // The freelist lock is needed to prevent asserts, is it really needed? 4085 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4086 4087 cl->set_freelistLock(freelistLock); 4088 4089 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4090 4091 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4092 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4093 PrecleanCLDClosure preclean_closure(cl); 4094 ClassLoaderDataGraph::cld_do(&preclean_closure); 4095 4096 verify_work_stacks_empty(); 4097 verify_overflow_empty(); 4098 } 4099 4100 void CMSCollector::checkpointRootsFinal() { 4101 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4102 check_correct_thread_executing(); 4103 // world is stopped at this checkpoint 4104 assert(SafepointSynchronize::is_at_safepoint(), 4105 "world should be stopped"); 4106 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4107 4108 verify_work_stacks_empty(); 4109 verify_overflow_empty(); 4110 4111 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4112 _young_gen->used() / K, _young_gen->capacity() / K); 4113 { 4114 if (CMSScavengeBeforeRemark) { 4115 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4116 // Temporarily set flag to false, GCH->do_collection will 4117 // expect it to be false and set to true 4118 FlagSetting fl(gch->_is_gc_active, false); 4119 4120 gch->do_collection(true, // full (i.e. force, see below) 4121 false, // !clear_all_soft_refs 4122 0, // size 4123 false, // is_tlab 4124 GenCollectedHeap::YoungGen // type 4125 ); 4126 } 4127 FreelistLocker x(this); 4128 MutexLockerEx y(bitMapLock(), 4129 Mutex::_no_safepoint_check_flag); 4130 checkpointRootsFinalWork(); 4131 } 4132 verify_work_stacks_empty(); 4133 verify_overflow_empty(); 4134 } 4135 4136 void CMSCollector::checkpointRootsFinalWork() { 4137 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4138 4139 assert(haveFreelistLocks(), "must have free list locks"); 4140 assert_lock_strong(bitMapLock()); 4141 4142 ResourceMark rm; 4143 HandleMark hm; 4144 4145 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4146 4147 if (should_unload_classes()) { 4148 CodeCache::gc_prologue(); 4149 } 4150 assert(haveFreelistLocks(), "must have free list locks"); 4151 assert_lock_strong(bitMapLock()); 4152 4153 // We might assume that we need not fill TLAB's when 4154 // CMSScavengeBeforeRemark is set, because we may have just done 4155 // a scavenge which would have filled all TLAB's -- and besides 4156 // Eden would be empty. This however may not always be the case -- 4157 // for instance although we asked for a scavenge, it may not have 4158 // happened because of a JNI critical section. We probably need 4159 // a policy for deciding whether we can in that case wait until 4160 // the critical section releases and then do the remark following 4161 // the scavenge, and skip it here. In the absence of that policy, 4162 // or of an indication of whether the scavenge did indeed occur, 4163 // we cannot rely on TLAB's having been filled and must do 4164 // so here just in case a scavenge did not happen. 4165 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4166 // Update the saved marks which may affect the root scans. 4167 gch->save_marks(); 4168 4169 print_eden_and_survivor_chunk_arrays(); 4170 4171 { 4172 #if defined(COMPILER2) || INCLUDE_JVMCI 4173 DerivedPointerTableDeactivate dpt_deact; 4174 #endif 4175 4176 // Note on the role of the mod union table: 4177 // Since the marker in "markFromRoots" marks concurrently with 4178 // mutators, it is possible for some reachable objects not to have been 4179 // scanned. For instance, an only reference to an object A was 4180 // placed in object B after the marker scanned B. Unless B is rescanned, 4181 // A would be collected. Such updates to references in marked objects 4182 // are detected via the mod union table which is the set of all cards 4183 // dirtied since the first checkpoint in this GC cycle and prior to 4184 // the most recent young generation GC, minus those cleaned up by the 4185 // concurrent precleaning. 4186 if (CMSParallelRemarkEnabled) { 4187 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4188 do_remark_parallel(); 4189 } else { 4190 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4191 do_remark_non_parallel(); 4192 } 4193 } 4194 verify_work_stacks_empty(); 4195 verify_overflow_empty(); 4196 4197 { 4198 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4199 refProcessingWork(); 4200 } 4201 verify_work_stacks_empty(); 4202 verify_overflow_empty(); 4203 4204 if (should_unload_classes()) { 4205 CodeCache::gc_epilogue(); 4206 } 4207 JvmtiExport::gc_epilogue(); 4208 4209 // If we encountered any (marking stack / work queue) overflow 4210 // events during the current CMS cycle, take appropriate 4211 // remedial measures, where possible, so as to try and avoid 4212 // recurrence of that condition. 4213 assert(_markStack.isEmpty(), "No grey objects"); 4214 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4215 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4216 if (ser_ovflw > 0) { 4217 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4218 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4219 _markStack.expand(); 4220 _ser_pmc_remark_ovflw = 0; 4221 _ser_pmc_preclean_ovflw = 0; 4222 _ser_kac_preclean_ovflw = 0; 4223 _ser_kac_ovflw = 0; 4224 } 4225 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4226 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4227 _par_pmc_remark_ovflw, _par_kac_ovflw); 4228 _par_pmc_remark_ovflw = 0; 4229 _par_kac_ovflw = 0; 4230 } 4231 if (_markStack._hit_limit > 0) { 4232 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4233 _markStack._hit_limit); 4234 } 4235 if (_markStack._failed_double > 0) { 4236 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4237 _markStack._failed_double, _markStack.capacity()); 4238 } 4239 _markStack._hit_limit = 0; 4240 _markStack._failed_double = 0; 4241 4242 if ((VerifyAfterGC || VerifyDuringGC) && 4243 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4244 verify_after_remark(); 4245 } 4246 4247 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4248 4249 // Change under the freelistLocks. 4250 _collectorState = Sweeping; 4251 // Call isAllClear() under bitMapLock 4252 assert(_modUnionTable.isAllClear(), 4253 "Should be clear by end of the final marking"); 4254 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4255 "Should be clear by end of the final marking"); 4256 } 4257 4258 void CMSParInitialMarkTask::work(uint worker_id) { 4259 elapsedTimer _timer; 4260 ResourceMark rm; 4261 HandleMark hm; 4262 4263 // ---------- scan from roots -------------- 4264 _timer.start(); 4265 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4266 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4267 4268 // ---------- young gen roots -------------- 4269 { 4270 work_on_young_gen_roots(&par_mri_cl); 4271 _timer.stop(); 4272 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4273 } 4274 4275 // ---------- remaining roots -------------- 4276 _timer.reset(); 4277 _timer.start(); 4278 4279 CLDToOopClosure cld_closure(&par_mri_cl, true); 4280 4281 gch->cms_process_roots(_strong_roots_scope, 4282 false, // yg was scanned above 4283 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4284 _collector->should_unload_classes(), 4285 &par_mri_cl, 4286 &cld_closure); 4287 assert(_collector->should_unload_classes() 4288 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4289 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4290 _timer.stop(); 4291 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4292 } 4293 4294 // Parallel remark task 4295 class CMSParRemarkTask: public CMSParMarkTask { 4296 CompactibleFreeListSpace* _cms_space; 4297 4298 // The per-thread work queues, available here for stealing. 4299 OopTaskQueueSet* _task_queues; 4300 ParallelTaskTerminator _term; 4301 StrongRootsScope* _strong_roots_scope; 4302 4303 public: 4304 // A value of 0 passed to n_workers will cause the number of 4305 // workers to be taken from the active workers in the work gang. 4306 CMSParRemarkTask(CMSCollector* collector, 4307 CompactibleFreeListSpace* cms_space, 4308 uint n_workers, WorkGang* workers, 4309 OopTaskQueueSet* task_queues, 4310 StrongRootsScope* strong_roots_scope): 4311 CMSParMarkTask("Rescan roots and grey objects in parallel", 4312 collector, n_workers), 4313 _cms_space(cms_space), 4314 _task_queues(task_queues), 4315 _term(n_workers, task_queues), 4316 _strong_roots_scope(strong_roots_scope) { } 4317 4318 OopTaskQueueSet* task_queues() { return _task_queues; } 4319 4320 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4321 4322 ParallelTaskTerminator* terminator() { return &_term; } 4323 uint n_workers() { return _n_workers; } 4324 4325 void work(uint worker_id); 4326 4327 private: 4328 // ... of dirty cards in old space 4329 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4330 ParMarkRefsIntoAndScanClosure* cl); 4331 4332 // ... work stealing for the above 4333 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4334 }; 4335 4336 class RemarkCLDClosure : public CLDClosure { 4337 CLDToOopClosure _cm_closure; 4338 public: 4339 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4340 void do_cld(ClassLoaderData* cld) { 4341 // Check if we have modified any oops in the CLD during the concurrent marking. 4342 if (cld->has_accumulated_modified_oops()) { 4343 cld->clear_accumulated_modified_oops(); 4344 4345 // We could have transfered the current modified marks to the accumulated marks, 4346 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4347 } else if (cld->has_modified_oops()) { 4348 // Don't clear anything, this info is needed by the next young collection. 4349 } else { 4350 // No modified oops in the ClassLoaderData. 4351 return; 4352 } 4353 4354 // The klass has modified fields, need to scan the klass. 4355 _cm_closure.do_cld(cld); 4356 } 4357 }; 4358 4359 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4360 ParNewGeneration* young_gen = _collector->_young_gen; 4361 ContiguousSpace* eden_space = young_gen->eden(); 4362 ContiguousSpace* from_space = young_gen->from(); 4363 ContiguousSpace* to_space = young_gen->to(); 4364 4365 HeapWord** eca = _collector->_eden_chunk_array; 4366 size_t ect = _collector->_eden_chunk_index; 4367 HeapWord** sca = _collector->_survivor_chunk_array; 4368 size_t sct = _collector->_survivor_chunk_index; 4369 4370 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4371 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4372 4373 do_young_space_rescan(cl, to_space, NULL, 0); 4374 do_young_space_rescan(cl, from_space, sca, sct); 4375 do_young_space_rescan(cl, eden_space, eca, ect); 4376 } 4377 4378 // work_queue(i) is passed to the closure 4379 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4380 // also is passed to do_dirty_card_rescan_tasks() and to 4381 // do_work_steal() to select the i-th task_queue. 4382 4383 void CMSParRemarkTask::work(uint worker_id) { 4384 elapsedTimer _timer; 4385 ResourceMark rm; 4386 HandleMark hm; 4387 4388 // ---------- rescan from roots -------------- 4389 _timer.start(); 4390 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4391 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4392 _collector->_span, _collector->ref_processor(), 4393 &(_collector->_markBitMap), 4394 work_queue(worker_id)); 4395 4396 // Rescan young gen roots first since these are likely 4397 // coarsely partitioned and may, on that account, constitute 4398 // the critical path; thus, it's best to start off that 4399 // work first. 4400 // ---------- young gen roots -------------- 4401 { 4402 work_on_young_gen_roots(&par_mrias_cl); 4403 _timer.stop(); 4404 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4405 } 4406 4407 // ---------- remaining roots -------------- 4408 _timer.reset(); 4409 _timer.start(); 4410 gch->cms_process_roots(_strong_roots_scope, 4411 false, // yg was scanned above 4412 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4413 _collector->should_unload_classes(), 4414 &par_mrias_cl, 4415 NULL); // The dirty klasses will be handled below 4416 4417 assert(_collector->should_unload_classes() 4418 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4419 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4420 _timer.stop(); 4421 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4422 4423 // ---------- unhandled CLD scanning ---------- 4424 if (worker_id == 0) { // Single threaded at the moment. 4425 _timer.reset(); 4426 _timer.start(); 4427 4428 // Scan all new class loader data objects and new dependencies that were 4429 // introduced during concurrent marking. 4430 ResourceMark rm; 4431 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4432 for (int i = 0; i < array->length(); i++) { 4433 par_mrias_cl.do_cld_nv(array->at(i)); 4434 } 4435 4436 // We don't need to keep track of new CLDs anymore. 4437 ClassLoaderDataGraph::remember_new_clds(false); 4438 4439 _timer.stop(); 4440 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4441 } 4442 4443 // We might have added oops to ClassLoaderData::_handles during the 4444 // concurrent marking phase. These oops do not always point to newly allocated objects 4445 // that are guaranteed to be kept alive. Hence, 4446 // we do have to revisit the _handles block during the remark phase. 4447 4448 // ---------- dirty CLD scanning ---------- 4449 if (worker_id == 0) { // Single threaded at the moment. 4450 _timer.reset(); 4451 _timer.start(); 4452 4453 // Scan all classes that was dirtied during the concurrent marking phase. 4454 RemarkCLDClosure remark_closure(&par_mrias_cl); 4455 ClassLoaderDataGraph::cld_do(&remark_closure); 4456 4457 _timer.stop(); 4458 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4459 } 4460 4461 4462 // ---------- rescan dirty cards ------------ 4463 _timer.reset(); 4464 _timer.start(); 4465 4466 // Do the rescan tasks for each of the two spaces 4467 // (cms_space) in turn. 4468 // "worker_id" is passed to select the task_queue for "worker_id" 4469 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4470 _timer.stop(); 4471 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4472 4473 // ---------- steal work from other threads ... 4474 // ---------- ... and drain overflow list. 4475 _timer.reset(); 4476 _timer.start(); 4477 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4478 _timer.stop(); 4479 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4480 } 4481 4482 void 4483 CMSParMarkTask::do_young_space_rescan( 4484 OopsInGenClosure* cl, ContiguousSpace* space, 4485 HeapWord** chunk_array, size_t chunk_top) { 4486 // Until all tasks completed: 4487 // . claim an unclaimed task 4488 // . compute region boundaries corresponding to task claimed 4489 // using chunk_array 4490 // . par_oop_iterate(cl) over that region 4491 4492 ResourceMark rm; 4493 HandleMark hm; 4494 4495 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4496 4497 uint nth_task = 0; 4498 uint n_tasks = pst->n_tasks(); 4499 4500 if (n_tasks > 0) { 4501 assert(pst->valid(), "Uninitialized use?"); 4502 HeapWord *start, *end; 4503 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4504 // We claimed task # nth_task; compute its boundaries. 4505 if (chunk_top == 0) { // no samples were taken 4506 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4507 start = space->bottom(); 4508 end = space->top(); 4509 } else if (nth_task == 0) { 4510 start = space->bottom(); 4511 end = chunk_array[nth_task]; 4512 } else if (nth_task < (uint)chunk_top) { 4513 assert(nth_task >= 1, "Control point invariant"); 4514 start = chunk_array[nth_task - 1]; 4515 end = chunk_array[nth_task]; 4516 } else { 4517 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4518 start = chunk_array[chunk_top - 1]; 4519 end = space->top(); 4520 } 4521 MemRegion mr(start, end); 4522 // Verify that mr is in space 4523 assert(mr.is_empty() || space->used_region().contains(mr), 4524 "Should be in space"); 4525 // Verify that "start" is an object boundary 4526 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4527 "Should be an oop"); 4528 space->par_oop_iterate(mr, cl); 4529 } 4530 pst->all_tasks_completed(); 4531 } 4532 } 4533 4534 void 4535 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4536 CompactibleFreeListSpace* sp, int i, 4537 ParMarkRefsIntoAndScanClosure* cl) { 4538 // Until all tasks completed: 4539 // . claim an unclaimed task 4540 // . compute region boundaries corresponding to task claimed 4541 // . transfer dirty bits ct->mut for that region 4542 // . apply rescanclosure to dirty mut bits for that region 4543 4544 ResourceMark rm; 4545 HandleMark hm; 4546 4547 OopTaskQueue* work_q = work_queue(i); 4548 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4549 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4550 // CAUTION: This closure has state that persists across calls to 4551 // the work method dirty_range_iterate_clear() in that it has 4552 // embedded in it a (subtype of) UpwardsObjectClosure. The 4553 // use of that state in the embedded UpwardsObjectClosure instance 4554 // assumes that the cards are always iterated (even if in parallel 4555 // by several threads) in monotonically increasing order per each 4556 // thread. This is true of the implementation below which picks 4557 // card ranges (chunks) in monotonically increasing order globally 4558 // and, a-fortiori, in monotonically increasing order per thread 4559 // (the latter order being a subsequence of the former). 4560 // If the work code below is ever reorganized into a more chaotic 4561 // work-partitioning form than the current "sequential tasks" 4562 // paradigm, the use of that persistent state will have to be 4563 // revisited and modified appropriately. See also related 4564 // bug 4756801 work on which should examine this code to make 4565 // sure that the changes there do not run counter to the 4566 // assumptions made here and necessary for correctness and 4567 // efficiency. Note also that this code might yield inefficient 4568 // behavior in the case of very large objects that span one or 4569 // more work chunks. Such objects would potentially be scanned 4570 // several times redundantly. Work on 4756801 should try and 4571 // address that performance anomaly if at all possible. XXX 4572 MemRegion full_span = _collector->_span; 4573 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4574 MarkFromDirtyCardsClosure 4575 greyRescanClosure(_collector, full_span, // entire span of interest 4576 sp, bm, work_q, cl); 4577 4578 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4579 assert(pst->valid(), "Uninitialized use?"); 4580 uint nth_task = 0; 4581 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4582 MemRegion span = sp->used_region(); 4583 HeapWord* start_addr = span.start(); 4584 HeapWord* end_addr = align_up(span.end(), alignment); 4585 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4586 assert(is_aligned(start_addr, alignment), "Check alignment"); 4587 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4588 4589 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4590 // Having claimed the nth_task, compute corresponding mem-region, 4591 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4592 // The alignment restriction ensures that we do not need any 4593 // synchronization with other gang-workers while setting or 4594 // clearing bits in thus chunk of the MUT. 4595 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4596 start_addr + (nth_task+1)*chunk_size); 4597 // The last chunk's end might be way beyond end of the 4598 // used region. In that case pull back appropriately. 4599 if (this_span.end() > end_addr) { 4600 this_span.set_end(end_addr); 4601 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4602 } 4603 // Iterate over the dirty cards covering this chunk, marking them 4604 // precleaned, and setting the corresponding bits in the mod union 4605 // table. Since we have been careful to partition at Card and MUT-word 4606 // boundaries no synchronization is needed between parallel threads. 4607 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4608 &modUnionClosure); 4609 4610 // Having transferred these marks into the modUnionTable, 4611 // rescan the marked objects on the dirty cards in the modUnionTable. 4612 // Even if this is at a synchronous collection, the initial marking 4613 // may have been done during an asynchronous collection so there 4614 // may be dirty bits in the mod-union table. 4615 _collector->_modUnionTable.dirty_range_iterate_clear( 4616 this_span, &greyRescanClosure); 4617 _collector->_modUnionTable.verifyNoOneBitsInRange( 4618 this_span.start(), 4619 this_span.end()); 4620 } 4621 pst->all_tasks_completed(); // declare that i am done 4622 } 4623 4624 // . see if we can share work_queues with ParNew? XXX 4625 void 4626 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4627 int* seed) { 4628 OopTaskQueue* work_q = work_queue(i); 4629 NOT_PRODUCT(int num_steals = 0;) 4630 oop obj_to_scan; 4631 CMSBitMap* bm = &(_collector->_markBitMap); 4632 4633 while (true) { 4634 // Completely finish any left over work from (an) earlier round(s) 4635 cl->trim_queue(0); 4636 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4637 (size_t)ParGCDesiredObjsFromOverflowList); 4638 // Now check if there's any work in the overflow list 4639 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4640 // only affects the number of attempts made to get work from the 4641 // overflow list and does not affect the number of workers. Just 4642 // pass ParallelGCThreads so this behavior is unchanged. 4643 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4644 work_q, 4645 ParallelGCThreads)) { 4646 // found something in global overflow list; 4647 // not yet ready to go stealing work from others. 4648 // We'd like to assert(work_q->size() != 0, ...) 4649 // because we just took work from the overflow list, 4650 // but of course we can't since all of that could have 4651 // been already stolen from us. 4652 // "He giveth and He taketh away." 4653 continue; 4654 } 4655 // Verify that we have no work before we resort to stealing 4656 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4657 // Try to steal from other queues that have work 4658 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4659 NOT_PRODUCT(num_steals++;) 4660 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4661 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4662 // Do scanning work 4663 obj_to_scan->oop_iterate(cl); 4664 // Loop around, finish this work, and try to steal some more 4665 } else if (terminator()->offer_termination()) { 4666 break; // nirvana from the infinite cycle 4667 } 4668 } 4669 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4670 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4671 "Else our work is not yet done"); 4672 } 4673 4674 // Record object boundaries in _eden_chunk_array by sampling the eden 4675 // top in the slow-path eden object allocation code path and record 4676 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4677 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4678 // sampling in sample_eden() that activates during the part of the 4679 // preclean phase. 4680 void CMSCollector::sample_eden_chunk() { 4681 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4682 if (_eden_chunk_lock->try_lock()) { 4683 // Record a sample. This is the critical section. The contents 4684 // of the _eden_chunk_array have to be non-decreasing in the 4685 // address order. 4686 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4687 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4688 "Unexpected state of Eden"); 4689 if (_eden_chunk_index == 0 || 4690 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4691 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4692 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4693 _eden_chunk_index++; // commit sample 4694 } 4695 _eden_chunk_lock->unlock(); 4696 } 4697 } 4698 } 4699 4700 // Return a thread-local PLAB recording array, as appropriate. 4701 void* CMSCollector::get_data_recorder(int thr_num) { 4702 if (_survivor_plab_array != NULL && 4703 (CMSPLABRecordAlways || 4704 (_collectorState > Marking && _collectorState < FinalMarking))) { 4705 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4706 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4707 ca->reset(); // clear it so that fresh data is recorded 4708 return (void*) ca; 4709 } else { 4710 return NULL; 4711 } 4712 } 4713 4714 // Reset all the thread-local PLAB recording arrays 4715 void CMSCollector::reset_survivor_plab_arrays() { 4716 for (uint i = 0; i < ParallelGCThreads; i++) { 4717 _survivor_plab_array[i].reset(); 4718 } 4719 } 4720 4721 // Merge the per-thread plab arrays into the global survivor chunk 4722 // array which will provide the partitioning of the survivor space 4723 // for CMS initial scan and rescan. 4724 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4725 int no_of_gc_threads) { 4726 assert(_survivor_plab_array != NULL, "Error"); 4727 assert(_survivor_chunk_array != NULL, "Error"); 4728 assert(_collectorState == FinalMarking || 4729 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4730 for (int j = 0; j < no_of_gc_threads; j++) { 4731 _cursor[j] = 0; 4732 } 4733 HeapWord* top = surv->top(); 4734 size_t i; 4735 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4736 HeapWord* min_val = top; // Higher than any PLAB address 4737 uint min_tid = 0; // position of min_val this round 4738 for (int j = 0; j < no_of_gc_threads; j++) { 4739 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4740 if (_cursor[j] == cur_sca->end()) { 4741 continue; 4742 } 4743 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4744 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4745 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4746 if (cur_val < min_val) { 4747 min_tid = j; 4748 min_val = cur_val; 4749 } else { 4750 assert(cur_val < top, "All recorded addresses should be less"); 4751 } 4752 } 4753 // At this point min_val and min_tid are respectively 4754 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4755 // and the thread (j) that witnesses that address. 4756 // We record this address in the _survivor_chunk_array[i] 4757 // and increment _cursor[min_tid] prior to the next round i. 4758 if (min_val == top) { 4759 break; 4760 } 4761 _survivor_chunk_array[i] = min_val; 4762 _cursor[min_tid]++; 4763 } 4764 // We are all done; record the size of the _survivor_chunk_array 4765 _survivor_chunk_index = i; // exclusive: [0, i) 4766 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4767 // Verify that we used up all the recorded entries 4768 #ifdef ASSERT 4769 size_t total = 0; 4770 for (int j = 0; j < no_of_gc_threads; j++) { 4771 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4772 total += _cursor[j]; 4773 } 4774 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4775 // Check that the merged array is in sorted order 4776 if (total > 0) { 4777 for (size_t i = 0; i < total - 1; i++) { 4778 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4779 i, p2i(_survivor_chunk_array[i])); 4780 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4781 "Not sorted"); 4782 } 4783 } 4784 #endif // ASSERT 4785 } 4786 4787 // Set up the space's par_seq_tasks structure for work claiming 4788 // for parallel initial scan and rescan of young gen. 4789 // See ParRescanTask where this is currently used. 4790 void 4791 CMSCollector:: 4792 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4793 assert(n_threads > 0, "Unexpected n_threads argument"); 4794 4795 // Eden space 4796 if (!_young_gen->eden()->is_empty()) { 4797 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4798 assert(!pst->valid(), "Clobbering existing data?"); 4799 // Each valid entry in [0, _eden_chunk_index) represents a task. 4800 size_t n_tasks = _eden_chunk_index + 1; 4801 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4802 // Sets the condition for completion of the subtask (how many threads 4803 // need to finish in order to be done). 4804 pst->set_n_threads(n_threads); 4805 pst->set_n_tasks((int)n_tasks); 4806 } 4807 4808 // Merge the survivor plab arrays into _survivor_chunk_array 4809 if (_survivor_plab_array != NULL) { 4810 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4811 } else { 4812 assert(_survivor_chunk_index == 0, "Error"); 4813 } 4814 4815 // To space 4816 { 4817 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4818 assert(!pst->valid(), "Clobbering existing data?"); 4819 // Sets the condition for completion of the subtask (how many threads 4820 // need to finish in order to be done). 4821 pst->set_n_threads(n_threads); 4822 pst->set_n_tasks(1); 4823 assert(pst->valid(), "Error"); 4824 } 4825 4826 // From space 4827 { 4828 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4829 assert(!pst->valid(), "Clobbering existing data?"); 4830 size_t n_tasks = _survivor_chunk_index + 1; 4831 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4832 // Sets the condition for completion of the subtask (how many threads 4833 // need to finish in order to be done). 4834 pst->set_n_threads(n_threads); 4835 pst->set_n_tasks((int)n_tasks); 4836 assert(pst->valid(), "Error"); 4837 } 4838 } 4839 4840 // Parallel version of remark 4841 void CMSCollector::do_remark_parallel() { 4842 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4843 WorkGang* workers = gch->workers(); 4844 assert(workers != NULL, "Need parallel worker threads."); 4845 // Choose to use the number of GC workers most recently set 4846 // into "active_workers". 4847 uint n_workers = workers->active_workers(); 4848 4849 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4850 4851 StrongRootsScope srs(n_workers); 4852 4853 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4854 4855 // We won't be iterating over the cards in the card table updating 4856 // the younger_gen cards, so we shouldn't call the following else 4857 // the verification code as well as subsequent younger_refs_iterate 4858 // code would get confused. XXX 4859 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4860 4861 // The young gen rescan work will not be done as part of 4862 // process_roots (which currently doesn't know how to 4863 // parallelize such a scan), but rather will be broken up into 4864 // a set of parallel tasks (via the sampling that the [abortable] 4865 // preclean phase did of eden, plus the [two] tasks of 4866 // scanning the [two] survivor spaces. Further fine-grain 4867 // parallelization of the scanning of the survivor spaces 4868 // themselves, and of precleaning of the young gen itself 4869 // is deferred to the future. 4870 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4871 4872 // The dirty card rescan work is broken up into a "sequence" 4873 // of parallel tasks (per constituent space) that are dynamically 4874 // claimed by the parallel threads. 4875 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4876 4877 // It turns out that even when we're using 1 thread, doing the work in a 4878 // separate thread causes wide variance in run times. We can't help this 4879 // in the multi-threaded case, but we special-case n=1 here to get 4880 // repeatable measurements of the 1-thread overhead of the parallel code. 4881 if (n_workers > 1) { 4882 // Make refs discovery MT-safe, if it isn't already: it may not 4883 // necessarily be so, since it's possible that we are doing 4884 // ST marking. 4885 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4886 workers->run_task(&tsk); 4887 } else { 4888 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4889 tsk.work(0); 4890 } 4891 4892 // restore, single-threaded for now, any preserved marks 4893 // as a result of work_q overflow 4894 restore_preserved_marks_if_any(); 4895 } 4896 4897 // Non-parallel version of remark 4898 void CMSCollector::do_remark_non_parallel() { 4899 ResourceMark rm; 4900 HandleMark hm; 4901 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4902 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4903 4904 MarkRefsIntoAndScanClosure 4905 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4906 &_markStack, this, 4907 false /* should_yield */, false /* not precleaning */); 4908 MarkFromDirtyCardsClosure 4909 markFromDirtyCardsClosure(this, _span, 4910 NULL, // space is set further below 4911 &_markBitMap, &_markStack, &mrias_cl); 4912 { 4913 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4914 // Iterate over the dirty cards, setting the corresponding bits in the 4915 // mod union table. 4916 { 4917 ModUnionClosure modUnionClosure(&_modUnionTable); 4918 _ct->ct_bs()->dirty_card_iterate( 4919 _cmsGen->used_region(), 4920 &modUnionClosure); 4921 } 4922 // Having transferred these marks into the modUnionTable, we just need 4923 // to rescan the marked objects on the dirty cards in the modUnionTable. 4924 // The initial marking may have been done during an asynchronous 4925 // collection so there may be dirty bits in the mod-union table. 4926 const int alignment = 4927 CardTableModRefBS::card_size * BitsPerWord; 4928 { 4929 // ... First handle dirty cards in CMS gen 4930 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4931 MemRegion ur = _cmsGen->used_region(); 4932 HeapWord* lb = ur.start(); 4933 HeapWord* ub = align_up(ur.end(), alignment); 4934 MemRegion cms_span(lb, ub); 4935 _modUnionTable.dirty_range_iterate_clear(cms_span, 4936 &markFromDirtyCardsClosure); 4937 verify_work_stacks_empty(); 4938 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4939 } 4940 } 4941 if (VerifyDuringGC && 4942 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4943 HandleMark hm; // Discard invalid handles created during verification 4944 Universe::verify(); 4945 } 4946 { 4947 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4948 4949 verify_work_stacks_empty(); 4950 4951 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4952 StrongRootsScope srs(1); 4953 4954 gch->cms_process_roots(&srs, 4955 true, // young gen as roots 4956 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4957 should_unload_classes(), 4958 &mrias_cl, 4959 NULL); // The dirty klasses will be handled below 4960 4961 assert(should_unload_classes() 4962 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4963 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4964 } 4965 4966 { 4967 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4968 4969 verify_work_stacks_empty(); 4970 4971 // Scan all class loader data objects that might have been introduced 4972 // during concurrent marking. 4973 ResourceMark rm; 4974 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4975 for (int i = 0; i < array->length(); i++) { 4976 mrias_cl.do_cld_nv(array->at(i)); 4977 } 4978 4979 // We don't need to keep track of new CLDs anymore. 4980 ClassLoaderDataGraph::remember_new_clds(false); 4981 4982 verify_work_stacks_empty(); 4983 } 4984 4985 // We might have added oops to ClassLoaderData::_handles during the 4986 // concurrent marking phase. These oops do not point to newly allocated objects 4987 // that are guaranteed to be kept alive. Hence, 4988 // we do have to revisit the _handles block during the remark phase. 4989 { 4990 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4991 4992 verify_work_stacks_empty(); 4993 4994 RemarkCLDClosure remark_closure(&mrias_cl); 4995 ClassLoaderDataGraph::cld_do(&remark_closure); 4996 4997 verify_work_stacks_empty(); 4998 } 4999 5000 verify_work_stacks_empty(); 5001 // Restore evacuated mark words, if any, used for overflow list links 5002 restore_preserved_marks_if_any(); 5003 5004 verify_overflow_empty(); 5005 } 5006 5007 //////////////////////////////////////////////////////// 5008 // Parallel Reference Processing Task Proxy Class 5009 //////////////////////////////////////////////////////// 5010 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5011 OopTaskQueueSet* _queues; 5012 ParallelTaskTerminator _terminator; 5013 public: 5014 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5015 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5016 ParallelTaskTerminator* terminator() { return &_terminator; } 5017 OopTaskQueueSet* queues() { return _queues; } 5018 }; 5019 5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5021 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5022 CMSCollector* _collector; 5023 CMSBitMap* _mark_bit_map; 5024 const MemRegion _span; 5025 ProcessTask& _task; 5026 5027 public: 5028 CMSRefProcTaskProxy(ProcessTask& task, 5029 CMSCollector* collector, 5030 const MemRegion& span, 5031 CMSBitMap* mark_bit_map, 5032 AbstractWorkGang* workers, 5033 OopTaskQueueSet* task_queues): 5034 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5035 task_queues, 5036 workers->active_workers()), 5037 _task(task), 5038 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5039 { 5040 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5041 "Inconsistency in _span"); 5042 } 5043 5044 OopTaskQueueSet* task_queues() { return queues(); } 5045 5046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5047 5048 void do_work_steal(int i, 5049 CMSParDrainMarkingStackClosure* drain, 5050 CMSParKeepAliveClosure* keep_alive, 5051 int* seed); 5052 5053 virtual void work(uint worker_id); 5054 }; 5055 5056 void CMSRefProcTaskProxy::work(uint worker_id) { 5057 ResourceMark rm; 5058 HandleMark hm; 5059 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5060 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5061 _mark_bit_map, 5062 work_queue(worker_id)); 5063 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5064 _mark_bit_map, 5065 work_queue(worker_id)); 5066 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5067 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5068 if (_task.marks_oops_alive()) { 5069 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5070 _collector->hash_seed(worker_id)); 5071 } 5072 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5073 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5074 } 5075 5076 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5077 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5078 EnqueueTask& _task; 5079 5080 public: 5081 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5082 : AbstractGangTask("Enqueue reference objects in parallel"), 5083 _task(task) 5084 { } 5085 5086 virtual void work(uint worker_id) 5087 { 5088 _task.work(worker_id); 5089 } 5090 }; 5091 5092 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5093 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5094 _span(span), 5095 _bit_map(bit_map), 5096 _work_queue(work_queue), 5097 _mark_and_push(collector, span, bit_map, work_queue), 5098 _low_water_mark(MIN2((work_queue->max_elems()/4), 5099 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5100 { } 5101 5102 // . see if we can share work_queues with ParNew? XXX 5103 void CMSRefProcTaskProxy::do_work_steal(int i, 5104 CMSParDrainMarkingStackClosure* drain, 5105 CMSParKeepAliveClosure* keep_alive, 5106 int* seed) { 5107 OopTaskQueue* work_q = work_queue(i); 5108 NOT_PRODUCT(int num_steals = 0;) 5109 oop obj_to_scan; 5110 5111 while (true) { 5112 // Completely finish any left over work from (an) earlier round(s) 5113 drain->trim_queue(0); 5114 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5115 (size_t)ParGCDesiredObjsFromOverflowList); 5116 // Now check if there's any work in the overflow list 5117 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5118 // only affects the number of attempts made to get work from the 5119 // overflow list and does not affect the number of workers. Just 5120 // pass ParallelGCThreads so this behavior is unchanged. 5121 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5122 work_q, 5123 ParallelGCThreads)) { 5124 // Found something in global overflow list; 5125 // not yet ready to go stealing work from others. 5126 // We'd like to assert(work_q->size() != 0, ...) 5127 // because we just took work from the overflow list, 5128 // but of course we can't, since all of that might have 5129 // been already stolen from us. 5130 continue; 5131 } 5132 // Verify that we have no work before we resort to stealing 5133 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5134 // Try to steal from other queues that have work 5135 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5136 NOT_PRODUCT(num_steals++;) 5137 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5138 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5139 // Do scanning work 5140 obj_to_scan->oop_iterate(keep_alive); 5141 // Loop around, finish this work, and try to steal some more 5142 } else if (terminator()->offer_termination()) { 5143 break; // nirvana from the infinite cycle 5144 } 5145 } 5146 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5147 } 5148 5149 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5150 { 5151 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5152 WorkGang* workers = gch->workers(); 5153 assert(workers != NULL, "Need parallel worker threads."); 5154 CMSRefProcTaskProxy rp_task(task, &_collector, 5155 _collector.ref_processor()->span(), 5156 _collector.markBitMap(), 5157 workers, _collector.task_queues()); 5158 workers->run_task(&rp_task); 5159 } 5160 5161 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5162 { 5163 5164 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5165 WorkGang* workers = gch->workers(); 5166 assert(workers != NULL, "Need parallel worker threads."); 5167 CMSRefEnqueueTaskProxy enq_task(task); 5168 workers->run_task(&enq_task); 5169 } 5170 5171 void CMSCollector::refProcessingWork() { 5172 ResourceMark rm; 5173 HandleMark hm; 5174 5175 ReferenceProcessor* rp = ref_processor(); 5176 assert(rp->span().equals(_span), "Spans should be equal"); 5177 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5178 // Process weak references. 5179 rp->setup_policy(false); 5180 verify_work_stacks_empty(); 5181 5182 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5183 &_markStack, false /* !preclean */); 5184 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5185 _span, &_markBitMap, &_markStack, 5186 &cmsKeepAliveClosure, false /* !preclean */); 5187 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q()); 5188 { 5189 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5190 5191 ReferenceProcessorStats stats; 5192 if (rp->processing_is_mt()) { 5193 // Set the degree of MT here. If the discovery is done MT, there 5194 // may have been a different number of threads doing the discovery 5195 // and a different number of discovered lists may have Ref objects. 5196 // That is OK as long as the Reference lists are balanced (see 5197 // balance_all_queues() and balance_queues()). 5198 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5199 uint active_workers = ParallelGCThreads; 5200 WorkGang* workers = gch->workers(); 5201 if (workers != NULL) { 5202 active_workers = workers->active_workers(); 5203 // The expectation is that active_workers will have already 5204 // been set to a reasonable value. If it has not been set, 5205 // investigate. 5206 assert(active_workers > 0, "Should have been set during scavenge"); 5207 } 5208 rp->set_active_mt_degree(active_workers); 5209 CMSRefProcTaskExecutor task_executor(*this); 5210 stats = rp->process_discovered_references(&_is_alive_closure, 5211 &cmsKeepAliveClosure, 5212 &cmsDrainMarkingStackClosure, 5213 &task_executor, 5214 &pt); 5215 } else { 5216 stats = rp->process_discovered_references(&_is_alive_closure, 5217 &cmsKeepAliveClosure, 5218 &cmsDrainMarkingStackClosure, 5219 NULL, 5220 &pt); 5221 } 5222 _gc_tracer_cm->report_gc_reference_stats(stats); 5223 pt.print_all_references(); 5224 } 5225 5226 // This is the point where the entire marking should have completed. 5227 verify_work_stacks_empty(); 5228 5229 if (should_unload_classes()) { 5230 { 5231 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5232 5233 // Unload classes and purge the SystemDictionary. 5234 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5235 5236 // Unload nmethods. 5237 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5238 5239 // Prune dead klasses from subklass/sibling/implementor lists. 5240 Klass::clean_weak_klass_links(&_is_alive_closure); 5241 } 5242 5243 { 5244 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5245 // Clean up unreferenced symbols in symbol table. 5246 SymbolTable::unlink(); 5247 } 5248 5249 { 5250 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5251 // Delete entries for dead interned strings. 5252 StringTable::unlink(&_is_alive_closure); 5253 } 5254 } 5255 5256 // Restore any preserved marks as a result of mark stack or 5257 // work queue overflow 5258 restore_preserved_marks_if_any(); // done single-threaded for now 5259 5260 rp->set_enqueuing_is_done(true); 5261 if (rp->processing_is_mt()) { 5262 rp->balance_all_queues(); 5263 CMSRefProcTaskExecutor task_executor(*this); 5264 rp->enqueue_discovered_references(&task_executor, &pt); 5265 } else { 5266 rp->enqueue_discovered_references(NULL, &pt); 5267 } 5268 rp->verify_no_references_recorded(); 5269 pt.print_enqueue_phase(); 5270 assert(!rp->discovery_enabled(), "should have been disabled"); 5271 } 5272 5273 #ifndef PRODUCT 5274 void CMSCollector::check_correct_thread_executing() { 5275 Thread* t = Thread::current(); 5276 // Only the VM thread or the CMS thread should be here. 5277 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5278 "Unexpected thread type"); 5279 // If this is the vm thread, the foreground process 5280 // should not be waiting. Note that _foregroundGCIsActive is 5281 // true while the foreground collector is waiting. 5282 if (_foregroundGCShouldWait) { 5283 // We cannot be the VM thread 5284 assert(t->is_ConcurrentGC_thread(), 5285 "Should be CMS thread"); 5286 } else { 5287 // We can be the CMS thread only if we are in a stop-world 5288 // phase of CMS collection. 5289 if (t->is_ConcurrentGC_thread()) { 5290 assert(_collectorState == InitialMarking || 5291 _collectorState == FinalMarking, 5292 "Should be a stop-world phase"); 5293 // The CMS thread should be holding the CMS_token. 5294 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5295 "Potential interference with concurrently " 5296 "executing VM thread"); 5297 } 5298 } 5299 } 5300 #endif 5301 5302 void CMSCollector::sweep() { 5303 assert(_collectorState == Sweeping, "just checking"); 5304 check_correct_thread_executing(); 5305 verify_work_stacks_empty(); 5306 verify_overflow_empty(); 5307 increment_sweep_count(); 5308 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5309 5310 _inter_sweep_timer.stop(); 5311 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5312 5313 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5314 _intra_sweep_timer.reset(); 5315 _intra_sweep_timer.start(); 5316 { 5317 GCTraceCPUTime tcpu; 5318 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5319 // First sweep the old gen 5320 { 5321 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5322 bitMapLock()); 5323 sweepWork(_cmsGen); 5324 } 5325 5326 // Update Universe::_heap_*_at_gc figures. 5327 // We need all the free list locks to make the abstract state 5328 // transition from Sweeping to Resetting. See detailed note 5329 // further below. 5330 { 5331 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5332 // Update heap occupancy information which is used as 5333 // input to soft ref clearing policy at the next gc. 5334 Universe::update_heap_info_at_gc(); 5335 _collectorState = Resizing; 5336 } 5337 } 5338 verify_work_stacks_empty(); 5339 verify_overflow_empty(); 5340 5341 if (should_unload_classes()) { 5342 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5343 // requires that the virtual spaces are stable and not deleted. 5344 ClassLoaderDataGraph::set_should_purge(true); 5345 } 5346 5347 _intra_sweep_timer.stop(); 5348 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5349 5350 _inter_sweep_timer.reset(); 5351 _inter_sweep_timer.start(); 5352 5353 // We need to use a monotonically non-decreasing time in ms 5354 // or we will see time-warp warnings and os::javaTimeMillis() 5355 // does not guarantee monotonicity. 5356 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5357 update_time_of_last_gc(now); 5358 5359 // NOTE on abstract state transitions: 5360 // Mutators allocate-live and/or mark the mod-union table dirty 5361 // based on the state of the collection. The former is done in 5362 // the interval [Marking, Sweeping] and the latter in the interval 5363 // [Marking, Sweeping). Thus the transitions into the Marking state 5364 // and out of the Sweeping state must be synchronously visible 5365 // globally to the mutators. 5366 // The transition into the Marking state happens with the world 5367 // stopped so the mutators will globally see it. Sweeping is 5368 // done asynchronously by the background collector so the transition 5369 // from the Sweeping state to the Resizing state must be done 5370 // under the freelistLock (as is the check for whether to 5371 // allocate-live and whether to dirty the mod-union table). 5372 assert(_collectorState == Resizing, "Change of collector state to" 5373 " Resizing must be done under the freelistLocks (plural)"); 5374 5375 // Now that sweeping has been completed, we clear 5376 // the incremental_collection_failed flag, 5377 // thus inviting a younger gen collection to promote into 5378 // this generation. If such a promotion may still fail, 5379 // the flag will be set again when a young collection is 5380 // attempted. 5381 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5382 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5383 gch->update_full_collections_completed(_collection_count_start); 5384 } 5385 5386 // FIX ME!!! Looks like this belongs in CFLSpace, with 5387 // CMSGen merely delegating to it. 5388 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5389 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5390 HeapWord* minAddr = _cmsSpace->bottom(); 5391 HeapWord* largestAddr = 5392 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5393 if (largestAddr == NULL) { 5394 // The dictionary appears to be empty. In this case 5395 // try to coalesce at the end of the heap. 5396 largestAddr = _cmsSpace->end(); 5397 } 5398 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5399 size_t nearLargestOffset = 5400 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5401 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5402 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5403 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5404 } 5405 5406 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5407 return addr >= _cmsSpace->nearLargestChunk(); 5408 } 5409 5410 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5411 return _cmsSpace->find_chunk_at_end(); 5412 } 5413 5414 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5415 bool full) { 5416 // If the young generation has been collected, gather any statistics 5417 // that are of interest at this point. 5418 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5419 if (!full && current_is_young) { 5420 // Gather statistics on the young generation collection. 5421 collector()->stats().record_gc0_end(used()); 5422 } 5423 } 5424 5425 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5426 // We iterate over the space(s) underlying this generation, 5427 // checking the mark bit map to see if the bits corresponding 5428 // to specific blocks are marked or not. Blocks that are 5429 // marked are live and are not swept up. All remaining blocks 5430 // are swept up, with coalescing on-the-fly as we sweep up 5431 // contiguous free and/or garbage blocks: 5432 // We need to ensure that the sweeper synchronizes with allocators 5433 // and stop-the-world collectors. In particular, the following 5434 // locks are used: 5435 // . CMS token: if this is held, a stop the world collection cannot occur 5436 // . freelistLock: if this is held no allocation can occur from this 5437 // generation by another thread 5438 // . bitMapLock: if this is held, no other thread can access or update 5439 // 5440 5441 // Note that we need to hold the freelistLock if we use 5442 // block iterate below; else the iterator might go awry if 5443 // a mutator (or promotion) causes block contents to change 5444 // (for instance if the allocator divvies up a block). 5445 // If we hold the free list lock, for all practical purposes 5446 // young generation GC's can't occur (they'll usually need to 5447 // promote), so we might as well prevent all young generation 5448 // GC's while we do a sweeping step. For the same reason, we might 5449 // as well take the bit map lock for the entire duration 5450 5451 // check that we hold the requisite locks 5452 assert(have_cms_token(), "Should hold cms token"); 5453 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5454 assert_lock_strong(old_gen->freelistLock()); 5455 assert_lock_strong(bitMapLock()); 5456 5457 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5458 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5459 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5460 _inter_sweep_estimate.padded_average(), 5461 _intra_sweep_estimate.padded_average()); 5462 old_gen->setNearLargestChunk(); 5463 5464 { 5465 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5466 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5467 // We need to free-up/coalesce garbage/blocks from a 5468 // co-terminal free run. This is done in the SweepClosure 5469 // destructor; so, do not remove this scope, else the 5470 // end-of-sweep-census below will be off by a little bit. 5471 } 5472 old_gen->cmsSpace()->sweep_completed(); 5473 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5474 if (should_unload_classes()) { // unloaded classes this cycle, 5475 _concurrent_cycles_since_last_unload = 0; // ... reset count 5476 } else { // did not unload classes, 5477 _concurrent_cycles_since_last_unload++; // ... increment count 5478 } 5479 } 5480 5481 // Reset CMS data structures (for now just the marking bit map) 5482 // preparatory for the next cycle. 5483 void CMSCollector::reset_concurrent() { 5484 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5485 5486 // If the state is not "Resetting", the foreground thread 5487 // has done a collection and the resetting. 5488 if (_collectorState != Resetting) { 5489 assert(_collectorState == Idling, "The state should only change" 5490 " because the foreground collector has finished the collection"); 5491 return; 5492 } 5493 5494 { 5495 // Clear the mark bitmap (no grey objects to start with) 5496 // for the next cycle. 5497 GCTraceCPUTime tcpu; 5498 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5499 5500 HeapWord* curAddr = _markBitMap.startWord(); 5501 while (curAddr < _markBitMap.endWord()) { 5502 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5503 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5504 _markBitMap.clear_large_range(chunk); 5505 if (ConcurrentMarkSweepThread::should_yield() && 5506 !foregroundGCIsActive() && 5507 CMSYield) { 5508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5509 "CMS thread should hold CMS token"); 5510 assert_lock_strong(bitMapLock()); 5511 bitMapLock()->unlock(); 5512 ConcurrentMarkSweepThread::desynchronize(true); 5513 stopTimer(); 5514 incrementYields(); 5515 5516 // See the comment in coordinator_yield() 5517 for (unsigned i = 0; i < CMSYieldSleepCount && 5518 ConcurrentMarkSweepThread::should_yield() && 5519 !CMSCollector::foregroundGCIsActive(); ++i) { 5520 os::sleep(Thread::current(), 1, false); 5521 } 5522 5523 ConcurrentMarkSweepThread::synchronize(true); 5524 bitMapLock()->lock_without_safepoint_check(); 5525 startTimer(); 5526 } 5527 curAddr = chunk.end(); 5528 } 5529 // A successful mostly concurrent collection has been done. 5530 // Because only the full (i.e., concurrent mode failure) collections 5531 // are being measured for gc overhead limits, clean the "near" flag 5532 // and count. 5533 size_policy()->reset_gc_overhead_limit_count(); 5534 _collectorState = Idling; 5535 } 5536 5537 register_gc_end(); 5538 } 5539 5540 // Same as above but for STW paths 5541 void CMSCollector::reset_stw() { 5542 // already have the lock 5543 assert(_collectorState == Resetting, "just checking"); 5544 assert_lock_strong(bitMapLock()); 5545 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5546 _markBitMap.clear_all(); 5547 _collectorState = Idling; 5548 register_gc_end(); 5549 } 5550 5551 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5552 GCTraceCPUTime tcpu; 5553 TraceCollectorStats tcs(counters()); 5554 5555 switch (op) { 5556 case CMS_op_checkpointRootsInitial: { 5557 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5558 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5559 checkpointRootsInitial(); 5560 break; 5561 } 5562 case CMS_op_checkpointRootsFinal: { 5563 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5564 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5565 checkpointRootsFinal(); 5566 break; 5567 } 5568 default: 5569 fatal("No such CMS_op"); 5570 } 5571 } 5572 5573 #ifndef PRODUCT 5574 size_t const CMSCollector::skip_header_HeapWords() { 5575 return FreeChunk::header_size(); 5576 } 5577 5578 // Try and collect here conditions that should hold when 5579 // CMS thread is exiting. The idea is that the foreground GC 5580 // thread should not be blocked if it wants to terminate 5581 // the CMS thread and yet continue to run the VM for a while 5582 // after that. 5583 void CMSCollector::verify_ok_to_terminate() const { 5584 assert(Thread::current()->is_ConcurrentGC_thread(), 5585 "should be called by CMS thread"); 5586 assert(!_foregroundGCShouldWait, "should be false"); 5587 // We could check here that all the various low-level locks 5588 // are not held by the CMS thread, but that is overkill; see 5589 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5590 // is checked. 5591 } 5592 #endif 5593 5594 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5595 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5596 "missing Printezis mark?"); 5597 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5598 size_t size = pointer_delta(nextOneAddr + 1, addr); 5599 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5600 "alignment problem"); 5601 assert(size >= 3, "Necessary for Printezis marks to work"); 5602 return size; 5603 } 5604 5605 // A variant of the above (block_size_using_printezis_bits()) except 5606 // that we return 0 if the P-bits are not yet set. 5607 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5608 if (_markBitMap.isMarked(addr + 1)) { 5609 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5610 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5611 size_t size = pointer_delta(nextOneAddr + 1, addr); 5612 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5613 "alignment problem"); 5614 assert(size >= 3, "Necessary for Printezis marks to work"); 5615 return size; 5616 } 5617 return 0; 5618 } 5619 5620 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5621 size_t sz = 0; 5622 oop p = (oop)addr; 5623 if (p->klass_or_null_acquire() != NULL) { 5624 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5625 } else { 5626 sz = block_size_using_printezis_bits(addr); 5627 } 5628 assert(sz > 0, "size must be nonzero"); 5629 HeapWord* next_block = addr + sz; 5630 HeapWord* next_card = align_up(next_block, CardTableModRefBS::card_size); 5631 assert(align_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5632 align_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5633 "must be different cards"); 5634 return next_card; 5635 } 5636 5637 5638 // CMS Bit Map Wrapper ///////////////////////////////////////// 5639 5640 // Construct a CMS bit map infrastructure, but don't create the 5641 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5642 // further below. 5643 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5644 _bm(), 5645 _shifter(shifter), 5646 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5647 Monitor::_safepoint_check_sometimes) : NULL) 5648 { 5649 _bmStartWord = 0; 5650 _bmWordSize = 0; 5651 } 5652 5653 bool CMSBitMap::allocate(MemRegion mr) { 5654 _bmStartWord = mr.start(); 5655 _bmWordSize = mr.word_size(); 5656 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5657 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5658 if (!brs.is_reserved()) { 5659 log_warning(gc)("CMS bit map allocation failure"); 5660 return false; 5661 } 5662 // For now we'll just commit all of the bit map up front. 5663 // Later on we'll try to be more parsimonious with swap. 5664 if (!_virtual_space.initialize(brs, brs.size())) { 5665 log_warning(gc)("CMS bit map backing store failure"); 5666 return false; 5667 } 5668 assert(_virtual_space.committed_size() == brs.size(), 5669 "didn't reserve backing store for all of CMS bit map?"); 5670 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5671 _bmWordSize, "inconsistency in bit map sizing"); 5672 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5673 5674 // bm.clear(); // can we rely on getting zero'd memory? verify below 5675 assert(isAllClear(), 5676 "Expected zero'd memory from ReservedSpace constructor"); 5677 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5678 "consistency check"); 5679 return true; 5680 } 5681 5682 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5683 HeapWord *next_addr, *end_addr, *last_addr; 5684 assert_locked(); 5685 assert(covers(mr), "out-of-range error"); 5686 // XXX assert that start and end are appropriately aligned 5687 for (next_addr = mr.start(), end_addr = mr.end(); 5688 next_addr < end_addr; next_addr = last_addr) { 5689 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5690 last_addr = dirty_region.end(); 5691 if (!dirty_region.is_empty()) { 5692 cl->do_MemRegion(dirty_region); 5693 } else { 5694 assert(last_addr == end_addr, "program logic"); 5695 return; 5696 } 5697 } 5698 } 5699 5700 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5701 _bm.print_on_error(st, prefix); 5702 } 5703 5704 #ifndef PRODUCT 5705 void CMSBitMap::assert_locked() const { 5706 CMSLockVerifier::assert_locked(lock()); 5707 } 5708 5709 bool CMSBitMap::covers(MemRegion mr) const { 5710 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5711 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5712 "size inconsistency"); 5713 return (mr.start() >= _bmStartWord) && 5714 (mr.end() <= endWord()); 5715 } 5716 5717 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5718 return (start >= _bmStartWord && (start + size) <= endWord()); 5719 } 5720 5721 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5722 // verify that there are no 1 bits in the interval [left, right) 5723 FalseBitMapClosure falseBitMapClosure; 5724 iterate(&falseBitMapClosure, left, right); 5725 } 5726 5727 void CMSBitMap::region_invariant(MemRegion mr) 5728 { 5729 assert_locked(); 5730 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5731 assert(!mr.is_empty(), "unexpected empty region"); 5732 assert(covers(mr), "mr should be covered by bit map"); 5733 // convert address range into offset range 5734 size_t start_ofs = heapWordToOffset(mr.start()); 5735 // Make sure that end() is appropriately aligned 5736 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5737 "Misaligned mr.end()"); 5738 size_t end_ofs = heapWordToOffset(mr.end()); 5739 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5740 } 5741 5742 #endif 5743 5744 bool CMSMarkStack::allocate(size_t size) { 5745 // allocate a stack of the requisite depth 5746 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5747 size * sizeof(oop))); 5748 if (!rs.is_reserved()) { 5749 log_warning(gc)("CMSMarkStack allocation failure"); 5750 return false; 5751 } 5752 if (!_virtual_space.initialize(rs, rs.size())) { 5753 log_warning(gc)("CMSMarkStack backing store failure"); 5754 return false; 5755 } 5756 assert(_virtual_space.committed_size() == rs.size(), 5757 "didn't reserve backing store for all of CMS stack?"); 5758 _base = (oop*)(_virtual_space.low()); 5759 _index = 0; 5760 _capacity = size; 5761 NOT_PRODUCT(_max_depth = 0); 5762 return true; 5763 } 5764 5765 // XXX FIX ME !!! In the MT case we come in here holding a 5766 // leaf lock. For printing we need to take a further lock 5767 // which has lower rank. We need to recalibrate the two 5768 // lock-ranks involved in order to be able to print the 5769 // messages below. (Or defer the printing to the caller. 5770 // For now we take the expedient path of just disabling the 5771 // messages for the problematic case.) 5772 void CMSMarkStack::expand() { 5773 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5774 if (_capacity == MarkStackSizeMax) { 5775 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5776 // We print a warning message only once per CMS cycle. 5777 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5778 } 5779 return; 5780 } 5781 // Double capacity if possible 5782 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5783 // Do not give up existing stack until we have managed to 5784 // get the double capacity that we desired. 5785 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5786 new_capacity * sizeof(oop))); 5787 if (rs.is_reserved()) { 5788 // Release the backing store associated with old stack 5789 _virtual_space.release(); 5790 // Reinitialize virtual space for new stack 5791 if (!_virtual_space.initialize(rs, rs.size())) { 5792 fatal("Not enough swap for expanded marking stack"); 5793 } 5794 _base = (oop*)(_virtual_space.low()); 5795 _index = 0; 5796 _capacity = new_capacity; 5797 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5798 // Failed to double capacity, continue; 5799 // we print a detail message only once per CMS cycle. 5800 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5801 _capacity / K, new_capacity / K); 5802 } 5803 } 5804 5805 5806 // Closures 5807 // XXX: there seems to be a lot of code duplication here; 5808 // should refactor and consolidate common code. 5809 5810 // This closure is used to mark refs into the CMS generation in 5811 // the CMS bit map. Called at the first checkpoint. This closure 5812 // assumes that we do not need to re-mark dirty cards; if the CMS 5813 // generation on which this is used is not an oldest 5814 // generation then this will lose younger_gen cards! 5815 5816 MarkRefsIntoClosure::MarkRefsIntoClosure( 5817 MemRegion span, CMSBitMap* bitMap): 5818 _span(span), 5819 _bitMap(bitMap) 5820 { 5821 assert(ref_processor() == NULL, "deliberately left NULL"); 5822 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5823 } 5824 5825 void MarkRefsIntoClosure::do_oop(oop obj) { 5826 // if p points into _span, then mark corresponding bit in _markBitMap 5827 assert(oopDesc::is_oop(obj), "expected an oop"); 5828 HeapWord* addr = (HeapWord*)obj; 5829 if (_span.contains(addr)) { 5830 // this should be made more efficient 5831 _bitMap->mark(addr); 5832 } 5833 } 5834 5835 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5836 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5837 5838 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5839 MemRegion span, CMSBitMap* bitMap): 5840 _span(span), 5841 _bitMap(bitMap) 5842 { 5843 assert(ref_processor() == NULL, "deliberately left NULL"); 5844 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5845 } 5846 5847 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5848 // if p points into _span, then mark corresponding bit in _markBitMap 5849 assert(oopDesc::is_oop(obj), "expected an oop"); 5850 HeapWord* addr = (HeapWord*)obj; 5851 if (_span.contains(addr)) { 5852 // this should be made more efficient 5853 _bitMap->par_mark(addr); 5854 } 5855 } 5856 5857 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5858 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5859 5860 // A variant of the above, used for CMS marking verification. 5861 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5862 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5863 _span(span), 5864 _verification_bm(verification_bm), 5865 _cms_bm(cms_bm) 5866 { 5867 assert(ref_processor() == NULL, "deliberately left NULL"); 5868 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5869 } 5870 5871 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5872 // if p points into _span, then mark corresponding bit in _markBitMap 5873 assert(oopDesc::is_oop(obj), "expected an oop"); 5874 HeapWord* addr = (HeapWord*)obj; 5875 if (_span.contains(addr)) { 5876 _verification_bm->mark(addr); 5877 if (!_cms_bm->isMarked(addr)) { 5878 Log(gc, verify) log; 5879 ResourceMark rm; 5880 LogStream ls(log.error()); 5881 oop(addr)->print_on(&ls); 5882 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5883 fatal("... aborting"); 5884 } 5885 } 5886 } 5887 5888 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5889 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5890 5891 ////////////////////////////////////////////////// 5892 // MarkRefsIntoAndScanClosure 5893 ////////////////////////////////////////////////// 5894 5895 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5896 ReferenceProcessor* rp, 5897 CMSBitMap* bit_map, 5898 CMSBitMap* mod_union_table, 5899 CMSMarkStack* mark_stack, 5900 CMSCollector* collector, 5901 bool should_yield, 5902 bool concurrent_precleaning): 5903 _collector(collector), 5904 _span(span), 5905 _bit_map(bit_map), 5906 _mark_stack(mark_stack), 5907 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5908 mark_stack, concurrent_precleaning), 5909 _yield(should_yield), 5910 _concurrent_precleaning(concurrent_precleaning), 5911 _freelistLock(NULL) 5912 { 5913 // FIXME: Should initialize in base class constructor. 5914 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5915 set_ref_processor_internal(rp); 5916 } 5917 5918 // This closure is used to mark refs into the CMS generation at the 5919 // second (final) checkpoint, and to scan and transitively follow 5920 // the unmarked oops. It is also used during the concurrent precleaning 5921 // phase while scanning objects on dirty cards in the CMS generation. 5922 // The marks are made in the marking bit map and the marking stack is 5923 // used for keeping the (newly) grey objects during the scan. 5924 // The parallel version (Par_...) appears further below. 5925 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5926 if (obj != NULL) { 5927 assert(oopDesc::is_oop(obj), "expected an oop"); 5928 HeapWord* addr = (HeapWord*)obj; 5929 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5930 assert(_collector->overflow_list_is_empty(), 5931 "overflow list should be empty"); 5932 if (_span.contains(addr) && 5933 !_bit_map->isMarked(addr)) { 5934 // mark bit map (object is now grey) 5935 _bit_map->mark(addr); 5936 // push on marking stack (stack should be empty), and drain the 5937 // stack by applying this closure to the oops in the oops popped 5938 // from the stack (i.e. blacken the grey objects) 5939 bool res = _mark_stack->push(obj); 5940 assert(res, "Should have space to push on empty stack"); 5941 do { 5942 oop new_oop = _mark_stack->pop(); 5943 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5944 assert(_bit_map->isMarked((HeapWord*)new_oop), 5945 "only grey objects on this stack"); 5946 // iterate over the oops in this oop, marking and pushing 5947 // the ones in CMS heap (i.e. in _span). 5948 new_oop->oop_iterate(&_pushAndMarkClosure); 5949 // check if it's time to yield 5950 do_yield_check(); 5951 } while (!_mark_stack->isEmpty() || 5952 (!_concurrent_precleaning && take_from_overflow_list())); 5953 // if marking stack is empty, and we are not doing this 5954 // during precleaning, then check the overflow list 5955 } 5956 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5957 assert(_collector->overflow_list_is_empty(), 5958 "overflow list was drained above"); 5959 5960 assert(_collector->no_preserved_marks(), 5961 "All preserved marks should have been restored above"); 5962 } 5963 } 5964 5965 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5966 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5967 5968 void MarkRefsIntoAndScanClosure::do_yield_work() { 5969 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5970 "CMS thread should hold CMS token"); 5971 assert_lock_strong(_freelistLock); 5972 assert_lock_strong(_bit_map->lock()); 5973 // relinquish the free_list_lock and bitMaplock() 5974 _bit_map->lock()->unlock(); 5975 _freelistLock->unlock(); 5976 ConcurrentMarkSweepThread::desynchronize(true); 5977 _collector->stopTimer(); 5978 _collector->incrementYields(); 5979 5980 // See the comment in coordinator_yield() 5981 for (unsigned i = 0; 5982 i < CMSYieldSleepCount && 5983 ConcurrentMarkSweepThread::should_yield() && 5984 !CMSCollector::foregroundGCIsActive(); 5985 ++i) { 5986 os::sleep(Thread::current(), 1, false); 5987 } 5988 5989 ConcurrentMarkSweepThread::synchronize(true); 5990 _freelistLock->lock_without_safepoint_check(); 5991 _bit_map->lock()->lock_without_safepoint_check(); 5992 _collector->startTimer(); 5993 } 5994 5995 /////////////////////////////////////////////////////////// 5996 // ParMarkRefsIntoAndScanClosure: a parallel version of 5997 // MarkRefsIntoAndScanClosure 5998 /////////////////////////////////////////////////////////// 5999 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6000 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6001 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6002 _span(span), 6003 _bit_map(bit_map), 6004 _work_queue(work_queue), 6005 _low_water_mark(MIN2((work_queue->max_elems()/4), 6006 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6007 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6008 { 6009 // FIXME: Should initialize in base class constructor. 6010 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6011 set_ref_processor_internal(rp); 6012 } 6013 6014 // This closure is used to mark refs into the CMS generation at the 6015 // second (final) checkpoint, and to scan and transitively follow 6016 // the unmarked oops. The marks are made in the marking bit map and 6017 // the work_queue is used for keeping the (newly) grey objects during 6018 // the scan phase whence they are also available for stealing by parallel 6019 // threads. Since the marking bit map is shared, updates are 6020 // synchronized (via CAS). 6021 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6022 if (obj != NULL) { 6023 // Ignore mark word because this could be an already marked oop 6024 // that may be chained at the end of the overflow list. 6025 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6026 HeapWord* addr = (HeapWord*)obj; 6027 if (_span.contains(addr) && 6028 !_bit_map->isMarked(addr)) { 6029 // mark bit map (object will become grey): 6030 // It is possible for several threads to be 6031 // trying to "claim" this object concurrently; 6032 // the unique thread that succeeds in marking the 6033 // object first will do the subsequent push on 6034 // to the work queue (or overflow list). 6035 if (_bit_map->par_mark(addr)) { 6036 // push on work_queue (which may not be empty), and trim the 6037 // queue to an appropriate length by applying this closure to 6038 // the oops in the oops popped from the stack (i.e. blacken the 6039 // grey objects) 6040 bool res = _work_queue->push(obj); 6041 assert(res, "Low water mark should be less than capacity?"); 6042 trim_queue(_low_water_mark); 6043 } // Else, another thread claimed the object 6044 } 6045 } 6046 } 6047 6048 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6049 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6050 6051 // This closure is used to rescan the marked objects on the dirty cards 6052 // in the mod union table and the card table proper. 6053 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6054 oop p, MemRegion mr) { 6055 6056 size_t size = 0; 6057 HeapWord* addr = (HeapWord*)p; 6058 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6059 assert(_span.contains(addr), "we are scanning the CMS generation"); 6060 // check if it's time to yield 6061 if (do_yield_check()) { 6062 // We yielded for some foreground stop-world work, 6063 // and we have been asked to abort this ongoing preclean cycle. 6064 return 0; 6065 } 6066 if (_bitMap->isMarked(addr)) { 6067 // it's marked; is it potentially uninitialized? 6068 if (p->klass_or_null_acquire() != NULL) { 6069 // an initialized object; ignore mark word in verification below 6070 // since we are running concurrent with mutators 6071 assert(oopDesc::is_oop(p, true), "should be an oop"); 6072 if (p->is_objArray()) { 6073 // objArrays are precisely marked; restrict scanning 6074 // to dirty cards only. 6075 size = CompactibleFreeListSpace::adjustObjectSize( 6076 p->oop_iterate_size(_scanningClosure, mr)); 6077 } else { 6078 // A non-array may have been imprecisely marked; we need 6079 // to scan object in its entirety. 6080 size = CompactibleFreeListSpace::adjustObjectSize( 6081 p->oop_iterate_size(_scanningClosure)); 6082 } 6083 #ifdef ASSERT 6084 size_t direct_size = 6085 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6086 assert(size == direct_size, "Inconsistency in size"); 6087 assert(size >= 3, "Necessary for Printezis marks to work"); 6088 HeapWord* start_pbit = addr + 1; 6089 HeapWord* end_pbit = addr + size - 1; 6090 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6091 "inconsistent Printezis mark"); 6092 // Verify inner mark bits (between Printezis bits) are clear, 6093 // but don't repeat if there are multiple dirty regions for 6094 // the same object, to avoid potential O(N^2) performance. 6095 if (addr != _last_scanned_object) { 6096 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6097 _last_scanned_object = addr; 6098 } 6099 #endif // ASSERT 6100 } else { 6101 // An uninitialized object. 6102 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6103 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6104 size = pointer_delta(nextOneAddr + 1, addr); 6105 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6106 "alignment problem"); 6107 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6108 // will dirty the card when the klass pointer is installed in the 6109 // object (signaling the completion of initialization). 6110 } 6111 } else { 6112 // Either a not yet marked object or an uninitialized object 6113 if (p->klass_or_null_acquire() == NULL) { 6114 // An uninitialized object, skip to the next card, since 6115 // we may not be able to read its P-bits yet. 6116 assert(size == 0, "Initial value"); 6117 } else { 6118 // An object not (yet) reached by marking: we merely need to 6119 // compute its size so as to go look at the next block. 6120 assert(oopDesc::is_oop(p, true), "should be an oop"); 6121 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6122 } 6123 } 6124 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6125 return size; 6126 } 6127 6128 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6129 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6130 "CMS thread should hold CMS token"); 6131 assert_lock_strong(_freelistLock); 6132 assert_lock_strong(_bitMap->lock()); 6133 // relinquish the free_list_lock and bitMaplock() 6134 _bitMap->lock()->unlock(); 6135 _freelistLock->unlock(); 6136 ConcurrentMarkSweepThread::desynchronize(true); 6137 _collector->stopTimer(); 6138 _collector->incrementYields(); 6139 6140 // See the comment in coordinator_yield() 6141 for (unsigned i = 0; i < CMSYieldSleepCount && 6142 ConcurrentMarkSweepThread::should_yield() && 6143 !CMSCollector::foregroundGCIsActive(); ++i) { 6144 os::sleep(Thread::current(), 1, false); 6145 } 6146 6147 ConcurrentMarkSweepThread::synchronize(true); 6148 _freelistLock->lock_without_safepoint_check(); 6149 _bitMap->lock()->lock_without_safepoint_check(); 6150 _collector->startTimer(); 6151 } 6152 6153 6154 ////////////////////////////////////////////////////////////////// 6155 // SurvivorSpacePrecleanClosure 6156 ////////////////////////////////////////////////////////////////// 6157 // This (single-threaded) closure is used to preclean the oops in 6158 // the survivor spaces. 6159 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6160 6161 HeapWord* addr = (HeapWord*)p; 6162 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6163 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6164 assert(p->klass_or_null() != NULL, "object should be initialized"); 6165 // an initialized object; ignore mark word in verification below 6166 // since we are running concurrent with mutators 6167 assert(oopDesc::is_oop(p, true), "should be an oop"); 6168 // Note that we do not yield while we iterate over 6169 // the interior oops of p, pushing the relevant ones 6170 // on our marking stack. 6171 size_t size = p->oop_iterate_size(_scanning_closure); 6172 do_yield_check(); 6173 // Observe that below, we do not abandon the preclean 6174 // phase as soon as we should; rather we empty the 6175 // marking stack before returning. This is to satisfy 6176 // some existing assertions. In general, it may be a 6177 // good idea to abort immediately and complete the marking 6178 // from the grey objects at a later time. 6179 while (!_mark_stack->isEmpty()) { 6180 oop new_oop = _mark_stack->pop(); 6181 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6182 assert(_bit_map->isMarked((HeapWord*)new_oop), 6183 "only grey objects on this stack"); 6184 // iterate over the oops in this oop, marking and pushing 6185 // the ones in CMS heap (i.e. in _span). 6186 new_oop->oop_iterate(_scanning_closure); 6187 // check if it's time to yield 6188 do_yield_check(); 6189 } 6190 unsigned int after_count = 6191 GenCollectedHeap::heap()->total_collections(); 6192 bool abort = (_before_count != after_count) || 6193 _collector->should_abort_preclean(); 6194 return abort ? 0 : size; 6195 } 6196 6197 void SurvivorSpacePrecleanClosure::do_yield_work() { 6198 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6199 "CMS thread should hold CMS token"); 6200 assert_lock_strong(_bit_map->lock()); 6201 // Relinquish the bit map lock 6202 _bit_map->lock()->unlock(); 6203 ConcurrentMarkSweepThread::desynchronize(true); 6204 _collector->stopTimer(); 6205 _collector->incrementYields(); 6206 6207 // See the comment in coordinator_yield() 6208 for (unsigned i = 0; i < CMSYieldSleepCount && 6209 ConcurrentMarkSweepThread::should_yield() && 6210 !CMSCollector::foregroundGCIsActive(); ++i) { 6211 os::sleep(Thread::current(), 1, false); 6212 } 6213 6214 ConcurrentMarkSweepThread::synchronize(true); 6215 _bit_map->lock()->lock_without_safepoint_check(); 6216 _collector->startTimer(); 6217 } 6218 6219 // This closure is used to rescan the marked objects on the dirty cards 6220 // in the mod union table and the card table proper. In the parallel 6221 // case, although the bitMap is shared, we do a single read so the 6222 // isMarked() query is "safe". 6223 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6224 // Ignore mark word because we are running concurrent with mutators 6225 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6226 HeapWord* addr = (HeapWord*)p; 6227 assert(_span.contains(addr), "we are scanning the CMS generation"); 6228 bool is_obj_array = false; 6229 #ifdef ASSERT 6230 if (!_parallel) { 6231 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6232 assert(_collector->overflow_list_is_empty(), 6233 "overflow list should be empty"); 6234 6235 } 6236 #endif // ASSERT 6237 if (_bit_map->isMarked(addr)) { 6238 // Obj arrays are precisely marked, non-arrays are not; 6239 // so we scan objArrays precisely and non-arrays in their 6240 // entirety. 6241 if (p->is_objArray()) { 6242 is_obj_array = true; 6243 if (_parallel) { 6244 p->oop_iterate(_par_scan_closure, mr); 6245 } else { 6246 p->oop_iterate(_scan_closure, mr); 6247 } 6248 } else { 6249 if (_parallel) { 6250 p->oop_iterate(_par_scan_closure); 6251 } else { 6252 p->oop_iterate(_scan_closure); 6253 } 6254 } 6255 } 6256 #ifdef ASSERT 6257 if (!_parallel) { 6258 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6259 assert(_collector->overflow_list_is_empty(), 6260 "overflow list should be empty"); 6261 6262 } 6263 #endif // ASSERT 6264 return is_obj_array; 6265 } 6266 6267 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6268 MemRegion span, 6269 CMSBitMap* bitMap, CMSMarkStack* markStack, 6270 bool should_yield, bool verifying): 6271 _collector(collector), 6272 _span(span), 6273 _bitMap(bitMap), 6274 _mut(&collector->_modUnionTable), 6275 _markStack(markStack), 6276 _yield(should_yield), 6277 _skipBits(0) 6278 { 6279 assert(_markStack->isEmpty(), "stack should be empty"); 6280 _finger = _bitMap->startWord(); 6281 _threshold = _finger; 6282 assert(_collector->_restart_addr == NULL, "Sanity check"); 6283 assert(_span.contains(_finger), "Out of bounds _finger?"); 6284 DEBUG_ONLY(_verifying = verifying;) 6285 } 6286 6287 void MarkFromRootsClosure::reset(HeapWord* addr) { 6288 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6289 assert(_span.contains(addr), "Out of bounds _finger?"); 6290 _finger = addr; 6291 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6292 } 6293 6294 // Should revisit to see if this should be restructured for 6295 // greater efficiency. 6296 bool MarkFromRootsClosure::do_bit(size_t offset) { 6297 if (_skipBits > 0) { 6298 _skipBits--; 6299 return true; 6300 } 6301 // convert offset into a HeapWord* 6302 HeapWord* addr = _bitMap->startWord() + offset; 6303 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6304 "address out of range"); 6305 assert(_bitMap->isMarked(addr), "tautology"); 6306 if (_bitMap->isMarked(addr+1)) { 6307 // this is an allocated but not yet initialized object 6308 assert(_skipBits == 0, "tautology"); 6309 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6310 oop p = oop(addr); 6311 if (p->klass_or_null_acquire() == NULL) { 6312 DEBUG_ONLY(if (!_verifying) {) 6313 // We re-dirty the cards on which this object lies and increase 6314 // the _threshold so that we'll come back to scan this object 6315 // during the preclean or remark phase. (CMSCleanOnEnter) 6316 if (CMSCleanOnEnter) { 6317 size_t sz = _collector->block_size_using_printezis_bits(addr); 6318 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6319 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6320 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6321 // Bump _threshold to end_card_addr; note that 6322 // _threshold cannot possibly exceed end_card_addr, anyhow. 6323 // This prevents future clearing of the card as the scan proceeds 6324 // to the right. 6325 assert(_threshold <= end_card_addr, 6326 "Because we are just scanning into this object"); 6327 if (_threshold < end_card_addr) { 6328 _threshold = end_card_addr; 6329 } 6330 if (p->klass_or_null_acquire() != NULL) { 6331 // Redirty the range of cards... 6332 _mut->mark_range(redirty_range); 6333 } // ...else the setting of klass will dirty the card anyway. 6334 } 6335 DEBUG_ONLY(}) 6336 return true; 6337 } 6338 } 6339 scanOopsInOop(addr); 6340 return true; 6341 } 6342 6343 // We take a break if we've been at this for a while, 6344 // so as to avoid monopolizing the locks involved. 6345 void MarkFromRootsClosure::do_yield_work() { 6346 // First give up the locks, then yield, then re-lock 6347 // We should probably use a constructor/destructor idiom to 6348 // do this unlock/lock or modify the MutexUnlocker class to 6349 // serve our purpose. XXX 6350 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6351 "CMS thread should hold CMS token"); 6352 assert_lock_strong(_bitMap->lock()); 6353 _bitMap->lock()->unlock(); 6354 ConcurrentMarkSweepThread::desynchronize(true); 6355 _collector->stopTimer(); 6356 _collector->incrementYields(); 6357 6358 // See the comment in coordinator_yield() 6359 for (unsigned i = 0; i < CMSYieldSleepCount && 6360 ConcurrentMarkSweepThread::should_yield() && 6361 !CMSCollector::foregroundGCIsActive(); ++i) { 6362 os::sleep(Thread::current(), 1, false); 6363 } 6364 6365 ConcurrentMarkSweepThread::synchronize(true); 6366 _bitMap->lock()->lock_without_safepoint_check(); 6367 _collector->startTimer(); 6368 } 6369 6370 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6371 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6372 assert(_markStack->isEmpty(), 6373 "should drain stack to limit stack usage"); 6374 // convert ptr to an oop preparatory to scanning 6375 oop obj = oop(ptr); 6376 // Ignore mark word in verification below, since we 6377 // may be running concurrent with mutators. 6378 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6379 assert(_finger <= ptr, "_finger runneth ahead"); 6380 // advance the finger to right end of this object 6381 _finger = ptr + obj->size(); 6382 assert(_finger > ptr, "we just incremented it above"); 6383 // On large heaps, it may take us some time to get through 6384 // the marking phase. During 6385 // this time it's possible that a lot of mutations have 6386 // accumulated in the card table and the mod union table -- 6387 // these mutation records are redundant until we have 6388 // actually traced into the corresponding card. 6389 // Here, we check whether advancing the finger would make 6390 // us cross into a new card, and if so clear corresponding 6391 // cards in the MUT (preclean them in the card-table in the 6392 // future). 6393 6394 DEBUG_ONLY(if (!_verifying) {) 6395 // The clean-on-enter optimization is disabled by default, 6396 // until we fix 6178663. 6397 if (CMSCleanOnEnter && (_finger > _threshold)) { 6398 // [_threshold, _finger) represents the interval 6399 // of cards to be cleared in MUT (or precleaned in card table). 6400 // The set of cards to be cleared is all those that overlap 6401 // with the interval [_threshold, _finger); note that 6402 // _threshold is always kept card-aligned but _finger isn't 6403 // always card-aligned. 6404 HeapWord* old_threshold = _threshold; 6405 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6406 "_threshold should always be card-aligned"); 6407 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6408 MemRegion mr(old_threshold, _threshold); 6409 assert(!mr.is_empty(), "Control point invariant"); 6410 assert(_span.contains(mr), "Should clear within span"); 6411 _mut->clear_range(mr); 6412 } 6413 DEBUG_ONLY(}) 6414 // Note: the finger doesn't advance while we drain 6415 // the stack below. 6416 PushOrMarkClosure pushOrMarkClosure(_collector, 6417 _span, _bitMap, _markStack, 6418 _finger, this); 6419 bool res = _markStack->push(obj); 6420 assert(res, "Empty non-zero size stack should have space for single push"); 6421 while (!_markStack->isEmpty()) { 6422 oop new_oop = _markStack->pop(); 6423 // Skip verifying header mark word below because we are 6424 // running concurrent with mutators. 6425 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6426 // now scan this oop's oops 6427 new_oop->oop_iterate(&pushOrMarkClosure); 6428 do_yield_check(); 6429 } 6430 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6431 } 6432 6433 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6434 CMSCollector* collector, MemRegion span, 6435 CMSBitMap* bit_map, 6436 OopTaskQueue* work_queue, 6437 CMSMarkStack* overflow_stack): 6438 _collector(collector), 6439 _whole_span(collector->_span), 6440 _span(span), 6441 _bit_map(bit_map), 6442 _mut(&collector->_modUnionTable), 6443 _work_queue(work_queue), 6444 _overflow_stack(overflow_stack), 6445 _skip_bits(0), 6446 _task(task) 6447 { 6448 assert(_work_queue->size() == 0, "work_queue should be empty"); 6449 _finger = span.start(); 6450 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6451 assert(_span.contains(_finger), "Out of bounds _finger?"); 6452 } 6453 6454 // Should revisit to see if this should be restructured for 6455 // greater efficiency. 6456 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6457 if (_skip_bits > 0) { 6458 _skip_bits--; 6459 return true; 6460 } 6461 // convert offset into a HeapWord* 6462 HeapWord* addr = _bit_map->startWord() + offset; 6463 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6464 "address out of range"); 6465 assert(_bit_map->isMarked(addr), "tautology"); 6466 if (_bit_map->isMarked(addr+1)) { 6467 // this is an allocated object that might not yet be initialized 6468 assert(_skip_bits == 0, "tautology"); 6469 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6470 oop p = oop(addr); 6471 if (p->klass_or_null_acquire() == NULL) { 6472 // in the case of Clean-on-Enter optimization, redirty card 6473 // and avoid clearing card by increasing the threshold. 6474 return true; 6475 } 6476 } 6477 scan_oops_in_oop(addr); 6478 return true; 6479 } 6480 6481 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6482 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6483 // Should we assert that our work queue is empty or 6484 // below some drain limit? 6485 assert(_work_queue->size() == 0, 6486 "should drain stack to limit stack usage"); 6487 // convert ptr to an oop preparatory to scanning 6488 oop obj = oop(ptr); 6489 // Ignore mark word in verification below, since we 6490 // may be running concurrent with mutators. 6491 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6492 assert(_finger <= ptr, "_finger runneth ahead"); 6493 // advance the finger to right end of this object 6494 _finger = ptr + obj->size(); 6495 assert(_finger > ptr, "we just incremented it above"); 6496 // On large heaps, it may take us some time to get through 6497 // the marking phase. During 6498 // this time it's possible that a lot of mutations have 6499 // accumulated in the card table and the mod union table -- 6500 // these mutation records are redundant until we have 6501 // actually traced into the corresponding card. 6502 // Here, we check whether advancing the finger would make 6503 // us cross into a new card, and if so clear corresponding 6504 // cards in the MUT (preclean them in the card-table in the 6505 // future). 6506 6507 // The clean-on-enter optimization is disabled by default, 6508 // until we fix 6178663. 6509 if (CMSCleanOnEnter && (_finger > _threshold)) { 6510 // [_threshold, _finger) represents the interval 6511 // of cards to be cleared in MUT (or precleaned in card table). 6512 // The set of cards to be cleared is all those that overlap 6513 // with the interval [_threshold, _finger); note that 6514 // _threshold is always kept card-aligned but _finger isn't 6515 // always card-aligned. 6516 HeapWord* old_threshold = _threshold; 6517 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6518 "_threshold should always be card-aligned"); 6519 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6520 MemRegion mr(old_threshold, _threshold); 6521 assert(!mr.is_empty(), "Control point invariant"); 6522 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6523 _mut->clear_range(mr); 6524 } 6525 6526 // Note: the local finger doesn't advance while we drain 6527 // the stack below, but the global finger sure can and will. 6528 HeapWord* volatile* gfa = _task->global_finger_addr(); 6529 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6530 _span, _bit_map, 6531 _work_queue, 6532 _overflow_stack, 6533 _finger, 6534 gfa, this); 6535 bool res = _work_queue->push(obj); // overflow could occur here 6536 assert(res, "Will hold once we use workqueues"); 6537 while (true) { 6538 oop new_oop; 6539 if (!_work_queue->pop_local(new_oop)) { 6540 // We emptied our work_queue; check if there's stuff that can 6541 // be gotten from the overflow stack. 6542 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6543 _overflow_stack, _work_queue)) { 6544 do_yield_check(); 6545 continue; 6546 } else { // done 6547 break; 6548 } 6549 } 6550 // Skip verifying header mark word below because we are 6551 // running concurrent with mutators. 6552 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6553 // now scan this oop's oops 6554 new_oop->oop_iterate(&pushOrMarkClosure); 6555 do_yield_check(); 6556 } 6557 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6558 } 6559 6560 // Yield in response to a request from VM Thread or 6561 // from mutators. 6562 void ParMarkFromRootsClosure::do_yield_work() { 6563 assert(_task != NULL, "sanity"); 6564 _task->yield(); 6565 } 6566 6567 // A variant of the above used for verifying CMS marking work. 6568 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6569 MemRegion span, 6570 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6571 CMSMarkStack* mark_stack): 6572 _collector(collector), 6573 _span(span), 6574 _verification_bm(verification_bm), 6575 _cms_bm(cms_bm), 6576 _mark_stack(mark_stack), 6577 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6578 mark_stack) 6579 { 6580 assert(_mark_stack->isEmpty(), "stack should be empty"); 6581 _finger = _verification_bm->startWord(); 6582 assert(_collector->_restart_addr == NULL, "Sanity check"); 6583 assert(_span.contains(_finger), "Out of bounds _finger?"); 6584 } 6585 6586 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6587 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6588 assert(_span.contains(addr), "Out of bounds _finger?"); 6589 _finger = addr; 6590 } 6591 6592 // Should revisit to see if this should be restructured for 6593 // greater efficiency. 6594 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6595 // convert offset into a HeapWord* 6596 HeapWord* addr = _verification_bm->startWord() + offset; 6597 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6598 "address out of range"); 6599 assert(_verification_bm->isMarked(addr), "tautology"); 6600 assert(_cms_bm->isMarked(addr), "tautology"); 6601 6602 assert(_mark_stack->isEmpty(), 6603 "should drain stack to limit stack usage"); 6604 // convert addr to an oop preparatory to scanning 6605 oop obj = oop(addr); 6606 assert(oopDesc::is_oop(obj), "should be an oop"); 6607 assert(_finger <= addr, "_finger runneth ahead"); 6608 // advance the finger to right end of this object 6609 _finger = addr + obj->size(); 6610 assert(_finger > addr, "we just incremented it above"); 6611 // Note: the finger doesn't advance while we drain 6612 // the stack below. 6613 bool res = _mark_stack->push(obj); 6614 assert(res, "Empty non-zero size stack should have space for single push"); 6615 while (!_mark_stack->isEmpty()) { 6616 oop new_oop = _mark_stack->pop(); 6617 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6618 // now scan this oop's oops 6619 new_oop->oop_iterate(&_pam_verify_closure); 6620 } 6621 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6622 return true; 6623 } 6624 6625 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6626 CMSCollector* collector, MemRegion span, 6627 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6628 CMSMarkStack* mark_stack): 6629 MetadataAwareOopClosure(collector->ref_processor()), 6630 _collector(collector), 6631 _span(span), 6632 _verification_bm(verification_bm), 6633 _cms_bm(cms_bm), 6634 _mark_stack(mark_stack) 6635 { } 6636 6637 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6638 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6639 6640 // Upon stack overflow, we discard (part of) the stack, 6641 // remembering the least address amongst those discarded 6642 // in CMSCollector's _restart_address. 6643 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6644 // Remember the least grey address discarded 6645 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6646 _collector->lower_restart_addr(ra); 6647 _mark_stack->reset(); // discard stack contents 6648 _mark_stack->expand(); // expand the stack if possible 6649 } 6650 6651 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6652 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6653 HeapWord* addr = (HeapWord*)obj; 6654 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6655 // Oop lies in _span and isn't yet grey or black 6656 _verification_bm->mark(addr); // now grey 6657 if (!_cms_bm->isMarked(addr)) { 6658 Log(gc, verify) log; 6659 ResourceMark rm; 6660 LogStream ls(log.error()); 6661 oop(addr)->print_on(&ls); 6662 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6663 fatal("... aborting"); 6664 } 6665 6666 if (!_mark_stack->push(obj)) { // stack overflow 6667 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6668 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6669 handle_stack_overflow(addr); 6670 } 6671 // anything including and to the right of _finger 6672 // will be scanned as we iterate over the remainder of the 6673 // bit map 6674 } 6675 } 6676 6677 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6678 MemRegion span, 6679 CMSBitMap* bitMap, CMSMarkStack* markStack, 6680 HeapWord* finger, MarkFromRootsClosure* parent) : 6681 MetadataAwareOopClosure(collector->ref_processor()), 6682 _collector(collector), 6683 _span(span), 6684 _bitMap(bitMap), 6685 _markStack(markStack), 6686 _finger(finger), 6687 _parent(parent) 6688 { } 6689 6690 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6691 MemRegion span, 6692 CMSBitMap* bit_map, 6693 OopTaskQueue* work_queue, 6694 CMSMarkStack* overflow_stack, 6695 HeapWord* finger, 6696 HeapWord* volatile* global_finger_addr, 6697 ParMarkFromRootsClosure* parent) : 6698 MetadataAwareOopClosure(collector->ref_processor()), 6699 _collector(collector), 6700 _whole_span(collector->_span), 6701 _span(span), 6702 _bit_map(bit_map), 6703 _work_queue(work_queue), 6704 _overflow_stack(overflow_stack), 6705 _finger(finger), 6706 _global_finger_addr(global_finger_addr), 6707 _parent(parent) 6708 { } 6709 6710 // Assumes thread-safe access by callers, who are 6711 // responsible for mutual exclusion. 6712 void CMSCollector::lower_restart_addr(HeapWord* low) { 6713 assert(_span.contains(low), "Out of bounds addr"); 6714 if (_restart_addr == NULL) { 6715 _restart_addr = low; 6716 } else { 6717 _restart_addr = MIN2(_restart_addr, low); 6718 } 6719 } 6720 6721 // Upon stack overflow, we discard (part of) the stack, 6722 // remembering the least address amongst those discarded 6723 // in CMSCollector's _restart_address. 6724 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6725 // Remember the least grey address discarded 6726 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6727 _collector->lower_restart_addr(ra); 6728 _markStack->reset(); // discard stack contents 6729 _markStack->expand(); // expand the stack if possible 6730 } 6731 6732 // Upon stack overflow, we discard (part of) the stack, 6733 // remembering the least address amongst those discarded 6734 // in CMSCollector's _restart_address. 6735 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6736 // We need to do this under a mutex to prevent other 6737 // workers from interfering with the work done below. 6738 MutexLockerEx ml(_overflow_stack->par_lock(), 6739 Mutex::_no_safepoint_check_flag); 6740 // Remember the least grey address discarded 6741 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6742 _collector->lower_restart_addr(ra); 6743 _overflow_stack->reset(); // discard stack contents 6744 _overflow_stack->expand(); // expand the stack if possible 6745 } 6746 6747 void PushOrMarkClosure::do_oop(oop obj) { 6748 // Ignore mark word because we are running concurrent with mutators. 6749 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6750 HeapWord* addr = (HeapWord*)obj; 6751 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6752 // Oop lies in _span and isn't yet grey or black 6753 _bitMap->mark(addr); // now grey 6754 if (addr < _finger) { 6755 // the bit map iteration has already either passed, or 6756 // sampled, this bit in the bit map; we'll need to 6757 // use the marking stack to scan this oop's oops. 6758 bool simulate_overflow = false; 6759 NOT_PRODUCT( 6760 if (CMSMarkStackOverflowALot && 6761 _collector->simulate_overflow()) { 6762 // simulate a stack overflow 6763 simulate_overflow = true; 6764 } 6765 ) 6766 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6767 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6768 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6769 handle_stack_overflow(addr); 6770 } 6771 } 6772 // anything including and to the right of _finger 6773 // will be scanned as we iterate over the remainder of the 6774 // bit map 6775 do_yield_check(); 6776 } 6777 } 6778 6779 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6780 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6781 6782 void ParPushOrMarkClosure::do_oop(oop obj) { 6783 // Ignore mark word because we are running concurrent with mutators. 6784 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6785 HeapWord* addr = (HeapWord*)obj; 6786 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6787 // Oop lies in _span and isn't yet grey or black 6788 // We read the global_finger (volatile read) strictly after marking oop 6789 bool res = _bit_map->par_mark(addr); // now grey 6790 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6791 // Should we push this marked oop on our stack? 6792 // -- if someone else marked it, nothing to do 6793 // -- if target oop is above global finger nothing to do 6794 // -- if target oop is in chunk and above local finger 6795 // then nothing to do 6796 // -- else push on work queue 6797 if ( !res // someone else marked it, they will deal with it 6798 || (addr >= *gfa) // will be scanned in a later task 6799 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6800 return; 6801 } 6802 // the bit map iteration has already either passed, or 6803 // sampled, this bit in the bit map; we'll need to 6804 // use the marking stack to scan this oop's oops. 6805 bool simulate_overflow = false; 6806 NOT_PRODUCT( 6807 if (CMSMarkStackOverflowALot && 6808 _collector->simulate_overflow()) { 6809 // simulate a stack overflow 6810 simulate_overflow = true; 6811 } 6812 ) 6813 if (simulate_overflow || 6814 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6815 // stack overflow 6816 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6817 // We cannot assert that the overflow stack is full because 6818 // it may have been emptied since. 6819 assert(simulate_overflow || 6820 _work_queue->size() == _work_queue->max_elems(), 6821 "Else push should have succeeded"); 6822 handle_stack_overflow(addr); 6823 } 6824 do_yield_check(); 6825 } 6826 } 6827 6828 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6829 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6830 6831 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6832 MemRegion span, 6833 ReferenceProcessor* rp, 6834 CMSBitMap* bit_map, 6835 CMSBitMap* mod_union_table, 6836 CMSMarkStack* mark_stack, 6837 bool concurrent_precleaning): 6838 MetadataAwareOopClosure(rp), 6839 _collector(collector), 6840 _span(span), 6841 _bit_map(bit_map), 6842 _mod_union_table(mod_union_table), 6843 _mark_stack(mark_stack), 6844 _concurrent_precleaning(concurrent_precleaning) 6845 { 6846 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6847 } 6848 6849 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6850 // the non-parallel version (the parallel version appears further below.) 6851 void PushAndMarkClosure::do_oop(oop obj) { 6852 // Ignore mark word verification. If during concurrent precleaning, 6853 // the object monitor may be locked. If during the checkpoint 6854 // phases, the object may already have been reached by a different 6855 // path and may be at the end of the global overflow list (so 6856 // the mark word may be NULL). 6857 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6858 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6859 HeapWord* addr = (HeapWord*)obj; 6860 // Check if oop points into the CMS generation 6861 // and is not marked 6862 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6863 // a white object ... 6864 _bit_map->mark(addr); // ... now grey 6865 // push on the marking stack (grey set) 6866 bool simulate_overflow = false; 6867 NOT_PRODUCT( 6868 if (CMSMarkStackOverflowALot && 6869 _collector->simulate_overflow()) { 6870 // simulate a stack overflow 6871 simulate_overflow = true; 6872 } 6873 ) 6874 if (simulate_overflow || !_mark_stack->push(obj)) { 6875 if (_concurrent_precleaning) { 6876 // During precleaning we can just dirty the appropriate card(s) 6877 // in the mod union table, thus ensuring that the object remains 6878 // in the grey set and continue. In the case of object arrays 6879 // we need to dirty all of the cards that the object spans, 6880 // since the rescan of object arrays will be limited to the 6881 // dirty cards. 6882 // Note that no one can be interfering with us in this action 6883 // of dirtying the mod union table, so no locking or atomics 6884 // are required. 6885 if (obj->is_objArray()) { 6886 size_t sz = obj->size(); 6887 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6888 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6889 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6890 _mod_union_table->mark_range(redirty_range); 6891 } else { 6892 _mod_union_table->mark(addr); 6893 } 6894 _collector->_ser_pmc_preclean_ovflw++; 6895 } else { 6896 // During the remark phase, we need to remember this oop 6897 // in the overflow list. 6898 _collector->push_on_overflow_list(obj); 6899 _collector->_ser_pmc_remark_ovflw++; 6900 } 6901 } 6902 } 6903 } 6904 6905 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6906 MemRegion span, 6907 ReferenceProcessor* rp, 6908 CMSBitMap* bit_map, 6909 OopTaskQueue* work_queue): 6910 MetadataAwareOopClosure(rp), 6911 _collector(collector), 6912 _span(span), 6913 _bit_map(bit_map), 6914 _work_queue(work_queue) 6915 { 6916 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6917 } 6918 6919 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6920 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6921 6922 // Grey object rescan during second checkpoint phase -- 6923 // the parallel version. 6924 void ParPushAndMarkClosure::do_oop(oop obj) { 6925 // In the assert below, we ignore the mark word because 6926 // this oop may point to an already visited object that is 6927 // on the overflow stack (in which case the mark word has 6928 // been hijacked for chaining into the overflow stack -- 6929 // if this is the last object in the overflow stack then 6930 // its mark word will be NULL). Because this object may 6931 // have been subsequently popped off the global overflow 6932 // stack, and the mark word possibly restored to the prototypical 6933 // value, by the time we get to examined this failing assert in 6934 // the debugger, is_oop_or_null(false) may subsequently start 6935 // to hold. 6936 assert(oopDesc::is_oop_or_null(obj, true), 6937 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6938 HeapWord* addr = (HeapWord*)obj; 6939 // Check if oop points into the CMS generation 6940 // and is not marked 6941 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6942 // a white object ... 6943 // If we manage to "claim" the object, by being the 6944 // first thread to mark it, then we push it on our 6945 // marking stack 6946 if (_bit_map->par_mark(addr)) { // ... now grey 6947 // push on work queue (grey set) 6948 bool simulate_overflow = false; 6949 NOT_PRODUCT( 6950 if (CMSMarkStackOverflowALot && 6951 _collector->par_simulate_overflow()) { 6952 // simulate a stack overflow 6953 simulate_overflow = true; 6954 } 6955 ) 6956 if (simulate_overflow || !_work_queue->push(obj)) { 6957 _collector->par_push_on_overflow_list(obj); 6958 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6959 } 6960 } // Else, some other thread got there first 6961 } 6962 } 6963 6964 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6965 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6966 6967 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6968 Mutex* bml = _collector->bitMapLock(); 6969 assert_lock_strong(bml); 6970 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6971 "CMS thread should hold CMS token"); 6972 6973 bml->unlock(); 6974 ConcurrentMarkSweepThread::desynchronize(true); 6975 6976 _collector->stopTimer(); 6977 _collector->incrementYields(); 6978 6979 // See the comment in coordinator_yield() 6980 for (unsigned i = 0; i < CMSYieldSleepCount && 6981 ConcurrentMarkSweepThread::should_yield() && 6982 !CMSCollector::foregroundGCIsActive(); ++i) { 6983 os::sleep(Thread::current(), 1, false); 6984 } 6985 6986 ConcurrentMarkSweepThread::synchronize(true); 6987 bml->lock(); 6988 6989 _collector->startTimer(); 6990 } 6991 6992 bool CMSPrecleanRefsYieldClosure::should_return() { 6993 if (ConcurrentMarkSweepThread::should_yield()) { 6994 do_yield_work(); 6995 } 6996 return _collector->foregroundGCIsActive(); 6997 } 6998 6999 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7000 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7001 "mr should be aligned to start at a card boundary"); 7002 // We'd like to assert: 7003 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7004 // "mr should be a range of cards"); 7005 // However, that would be too strong in one case -- the last 7006 // partition ends at _unallocated_block which, in general, can be 7007 // an arbitrary boundary, not necessarily card aligned. 7008 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7009 _space->object_iterate_mem(mr, &_scan_cl); 7010 } 7011 7012 SweepClosure::SweepClosure(CMSCollector* collector, 7013 ConcurrentMarkSweepGeneration* g, 7014 CMSBitMap* bitMap, bool should_yield) : 7015 _collector(collector), 7016 _g(g), 7017 _sp(g->cmsSpace()), 7018 _limit(_sp->sweep_limit()), 7019 _freelistLock(_sp->freelistLock()), 7020 _bitMap(bitMap), 7021 _yield(should_yield), 7022 _inFreeRange(false), // No free range at beginning of sweep 7023 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7024 _lastFreeRangeCoalesced(false), 7025 _freeFinger(g->used_region().start()) 7026 { 7027 NOT_PRODUCT( 7028 _numObjectsFreed = 0; 7029 _numWordsFreed = 0; 7030 _numObjectsLive = 0; 7031 _numWordsLive = 0; 7032 _numObjectsAlreadyFree = 0; 7033 _numWordsAlreadyFree = 0; 7034 _last_fc = NULL; 7035 7036 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7037 _sp->dictionary()->initialize_dict_returned_bytes(); 7038 ) 7039 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7040 "sweep _limit out of bounds"); 7041 log_develop_trace(gc, sweep)("===================="); 7042 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7043 } 7044 7045 void SweepClosure::print_on(outputStream* st) const { 7046 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7047 p2i(_sp->bottom()), p2i(_sp->end())); 7048 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7049 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7050 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7051 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7052 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7053 } 7054 7055 #ifndef PRODUCT 7056 // Assertion checking only: no useful work in product mode -- 7057 // however, if any of the flags below become product flags, 7058 // you may need to review this code to see if it needs to be 7059 // enabled in product mode. 7060 SweepClosure::~SweepClosure() { 7061 assert_lock_strong(_freelistLock); 7062 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7063 "sweep _limit out of bounds"); 7064 if (inFreeRange()) { 7065 Log(gc, sweep) log; 7066 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7067 ResourceMark rm; 7068 LogStream ls(log.error()); 7069 print_on(&ls); 7070 ShouldNotReachHere(); 7071 } 7072 7073 if (log_is_enabled(Debug, gc, sweep)) { 7074 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7075 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7076 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7077 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7078 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7079 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7080 } 7081 7082 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7083 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7084 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7085 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7086 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7087 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7088 } 7089 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7090 log_develop_trace(gc, sweep)("================"); 7091 } 7092 #endif // PRODUCT 7093 7094 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7095 bool freeRangeInFreeLists) { 7096 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7097 p2i(freeFinger), freeRangeInFreeLists); 7098 assert(!inFreeRange(), "Trampling existing free range"); 7099 set_inFreeRange(true); 7100 set_lastFreeRangeCoalesced(false); 7101 7102 set_freeFinger(freeFinger); 7103 set_freeRangeInFreeLists(freeRangeInFreeLists); 7104 if (CMSTestInFreeList) { 7105 if (freeRangeInFreeLists) { 7106 FreeChunk* fc = (FreeChunk*) freeFinger; 7107 assert(fc->is_free(), "A chunk on the free list should be free."); 7108 assert(fc->size() > 0, "Free range should have a size"); 7109 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7110 } 7111 } 7112 } 7113 7114 // Note that the sweeper runs concurrently with mutators. Thus, 7115 // it is possible for direct allocation in this generation to happen 7116 // in the middle of the sweep. Note that the sweeper also coalesces 7117 // contiguous free blocks. Thus, unless the sweeper and the allocator 7118 // synchronize appropriately freshly allocated blocks may get swept up. 7119 // This is accomplished by the sweeper locking the free lists while 7120 // it is sweeping. Thus blocks that are determined to be free are 7121 // indeed free. There is however one additional complication: 7122 // blocks that have been allocated since the final checkpoint and 7123 // mark, will not have been marked and so would be treated as 7124 // unreachable and swept up. To prevent this, the allocator marks 7125 // the bit map when allocating during the sweep phase. This leads, 7126 // however, to a further complication -- objects may have been allocated 7127 // but not yet initialized -- in the sense that the header isn't yet 7128 // installed. The sweeper can not then determine the size of the block 7129 // in order to skip over it. To deal with this case, we use a technique 7130 // (due to Printezis) to encode such uninitialized block sizes in the 7131 // bit map. Since the bit map uses a bit per every HeapWord, but the 7132 // CMS generation has a minimum object size of 3 HeapWords, it follows 7133 // that "normal marks" won't be adjacent in the bit map (there will 7134 // always be at least two 0 bits between successive 1 bits). We make use 7135 // of these "unused" bits to represent uninitialized blocks -- the bit 7136 // corresponding to the start of the uninitialized object and the next 7137 // bit are both set. Finally, a 1 bit marks the end of the object that 7138 // started with the two consecutive 1 bits to indicate its potentially 7139 // uninitialized state. 7140 7141 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7142 FreeChunk* fc = (FreeChunk*)addr; 7143 size_t res; 7144 7145 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7146 // than "addr == _limit" because although _limit was a block boundary when 7147 // we started the sweep, it may no longer be one because heap expansion 7148 // may have caused us to coalesce the block ending at the address _limit 7149 // with a newly expanded chunk (this happens when _limit was set to the 7150 // previous _end of the space), so we may have stepped past _limit: 7151 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7152 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7153 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7154 "sweep _limit out of bounds"); 7155 assert(addr < _sp->end(), "addr out of bounds"); 7156 // Flush any free range we might be holding as a single 7157 // coalesced chunk to the appropriate free list. 7158 if (inFreeRange()) { 7159 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7160 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7161 flush_cur_free_chunk(freeFinger(), 7162 pointer_delta(addr, freeFinger())); 7163 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7164 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7165 lastFreeRangeCoalesced() ? 1 : 0); 7166 } 7167 7168 // help the iterator loop finish 7169 return pointer_delta(_sp->end(), addr); 7170 } 7171 7172 assert(addr < _limit, "sweep invariant"); 7173 // check if we should yield 7174 do_yield_check(addr); 7175 if (fc->is_free()) { 7176 // Chunk that is already free 7177 res = fc->size(); 7178 do_already_free_chunk(fc); 7179 debug_only(_sp->verifyFreeLists()); 7180 // If we flush the chunk at hand in lookahead_and_flush() 7181 // and it's coalesced with a preceding chunk, then the 7182 // process of "mangling" the payload of the coalesced block 7183 // will cause erasure of the size information from the 7184 // (erstwhile) header of all the coalesced blocks but the 7185 // first, so the first disjunct in the assert will not hold 7186 // in that specific case (in which case the second disjunct 7187 // will hold). 7188 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7189 "Otherwise the size info doesn't change at this step"); 7190 NOT_PRODUCT( 7191 _numObjectsAlreadyFree++; 7192 _numWordsAlreadyFree += res; 7193 ) 7194 NOT_PRODUCT(_last_fc = fc;) 7195 } else if (!_bitMap->isMarked(addr)) { 7196 // Chunk is fresh garbage 7197 res = do_garbage_chunk(fc); 7198 debug_only(_sp->verifyFreeLists()); 7199 NOT_PRODUCT( 7200 _numObjectsFreed++; 7201 _numWordsFreed += res; 7202 ) 7203 } else { 7204 // Chunk that is alive. 7205 res = do_live_chunk(fc); 7206 debug_only(_sp->verifyFreeLists()); 7207 NOT_PRODUCT( 7208 _numObjectsLive++; 7209 _numWordsLive += res; 7210 ) 7211 } 7212 return res; 7213 } 7214 7215 // For the smart allocation, record following 7216 // split deaths - a free chunk is removed from its free list because 7217 // it is being split into two or more chunks. 7218 // split birth - a free chunk is being added to its free list because 7219 // a larger free chunk has been split and resulted in this free chunk. 7220 // coal death - a free chunk is being removed from its free list because 7221 // it is being coalesced into a large free chunk. 7222 // coal birth - a free chunk is being added to its free list because 7223 // it was created when two or more free chunks where coalesced into 7224 // this free chunk. 7225 // 7226 // These statistics are used to determine the desired number of free 7227 // chunks of a given size. The desired number is chosen to be relative 7228 // to the end of a CMS sweep. The desired number at the end of a sweep 7229 // is the 7230 // count-at-end-of-previous-sweep (an amount that was enough) 7231 // - count-at-beginning-of-current-sweep (the excess) 7232 // + split-births (gains in this size during interval) 7233 // - split-deaths (demands on this size during interval) 7234 // where the interval is from the end of one sweep to the end of the 7235 // next. 7236 // 7237 // When sweeping the sweeper maintains an accumulated chunk which is 7238 // the chunk that is made up of chunks that have been coalesced. That 7239 // will be termed the left-hand chunk. A new chunk of garbage that 7240 // is being considered for coalescing will be referred to as the 7241 // right-hand chunk. 7242 // 7243 // When making a decision on whether to coalesce a right-hand chunk with 7244 // the current left-hand chunk, the current count vs. the desired count 7245 // of the left-hand chunk is considered. Also if the right-hand chunk 7246 // is near the large chunk at the end of the heap (see 7247 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7248 // left-hand chunk is coalesced. 7249 // 7250 // When making a decision about whether to split a chunk, the desired count 7251 // vs. the current count of the candidate to be split is also considered. 7252 // If the candidate is underpopulated (currently fewer chunks than desired) 7253 // a chunk of an overpopulated (currently more chunks than desired) size may 7254 // be chosen. The "hint" associated with a free list, if non-null, points 7255 // to a free list which may be overpopulated. 7256 // 7257 7258 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7259 const size_t size = fc->size(); 7260 // Chunks that cannot be coalesced are not in the 7261 // free lists. 7262 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7263 assert(_sp->verify_chunk_in_free_list(fc), 7264 "free chunk should be in free lists"); 7265 } 7266 // a chunk that is already free, should not have been 7267 // marked in the bit map 7268 HeapWord* const addr = (HeapWord*) fc; 7269 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7270 // Verify that the bit map has no bits marked between 7271 // addr and purported end of this block. 7272 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7273 7274 // Some chunks cannot be coalesced under any circumstances. 7275 // See the definition of cantCoalesce(). 7276 if (!fc->cantCoalesce()) { 7277 // This chunk can potentially be coalesced. 7278 // All the work is done in 7279 do_post_free_or_garbage_chunk(fc, size); 7280 // Note that if the chunk is not coalescable (the else arm 7281 // below), we unconditionally flush, without needing to do 7282 // a "lookahead," as we do below. 7283 if (inFreeRange()) lookahead_and_flush(fc, size); 7284 } else { 7285 // Code path common to both original and adaptive free lists. 7286 7287 // cant coalesce with previous block; this should be treated 7288 // as the end of a free run if any 7289 if (inFreeRange()) { 7290 // we kicked some butt; time to pick up the garbage 7291 assert(freeFinger() < addr, "freeFinger points too high"); 7292 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7293 } 7294 // else, nothing to do, just continue 7295 } 7296 } 7297 7298 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7299 // This is a chunk of garbage. It is not in any free list. 7300 // Add it to a free list or let it possibly be coalesced into 7301 // a larger chunk. 7302 HeapWord* const addr = (HeapWord*) fc; 7303 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7304 7305 // Verify that the bit map has no bits marked between 7306 // addr and purported end of just dead object. 7307 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7308 do_post_free_or_garbage_chunk(fc, size); 7309 7310 assert(_limit >= addr + size, 7311 "A freshly garbage chunk can't possibly straddle over _limit"); 7312 if (inFreeRange()) lookahead_and_flush(fc, size); 7313 return size; 7314 } 7315 7316 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7317 HeapWord* addr = (HeapWord*) fc; 7318 // The sweeper has just found a live object. Return any accumulated 7319 // left hand chunk to the free lists. 7320 if (inFreeRange()) { 7321 assert(freeFinger() < addr, "freeFinger points too high"); 7322 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7323 } 7324 7325 // This object is live: we'd normally expect this to be 7326 // an oop, and like to assert the following: 7327 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7328 // However, as we commented above, this may be an object whose 7329 // header hasn't yet been initialized. 7330 size_t size; 7331 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7332 if (_bitMap->isMarked(addr + 1)) { 7333 // Determine the size from the bit map, rather than trying to 7334 // compute it from the object header. 7335 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7336 size = pointer_delta(nextOneAddr + 1, addr); 7337 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7338 "alignment problem"); 7339 7340 #ifdef ASSERT 7341 if (oop(addr)->klass_or_null_acquire() != NULL) { 7342 // Ignore mark word because we are running concurrent with mutators 7343 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7344 assert(size == 7345 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7346 "P-mark and computed size do not agree"); 7347 } 7348 #endif 7349 7350 } else { 7351 // This should be an initialized object that's alive. 7352 assert(oop(addr)->klass_or_null_acquire() != NULL, 7353 "Should be an initialized object"); 7354 // Ignore mark word because we are running concurrent with mutators 7355 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7356 // Verify that the bit map has no bits marked between 7357 // addr and purported end of this block. 7358 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7359 assert(size >= 3, "Necessary for Printezis marks to work"); 7360 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7361 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7362 } 7363 return size; 7364 } 7365 7366 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7367 size_t chunkSize) { 7368 // do_post_free_or_garbage_chunk() should only be called in the case 7369 // of the adaptive free list allocator. 7370 const bool fcInFreeLists = fc->is_free(); 7371 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7372 if (CMSTestInFreeList && fcInFreeLists) { 7373 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7374 } 7375 7376 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7377 7378 HeapWord* const fc_addr = (HeapWord*) fc; 7379 7380 bool coalesce = false; 7381 const size_t left = pointer_delta(fc_addr, freeFinger()); 7382 const size_t right = chunkSize; 7383 switch (FLSCoalescePolicy) { 7384 // numeric value forms a coalition aggressiveness metric 7385 case 0: { // never coalesce 7386 coalesce = false; 7387 break; 7388 } 7389 case 1: { // coalesce if left & right chunks on overpopulated lists 7390 coalesce = _sp->coalOverPopulated(left) && 7391 _sp->coalOverPopulated(right); 7392 break; 7393 } 7394 case 2: { // coalesce if left chunk on overpopulated list (default) 7395 coalesce = _sp->coalOverPopulated(left); 7396 break; 7397 } 7398 case 3: { // coalesce if left OR right chunk on overpopulated list 7399 coalesce = _sp->coalOverPopulated(left) || 7400 _sp->coalOverPopulated(right); 7401 break; 7402 } 7403 case 4: { // always coalesce 7404 coalesce = true; 7405 break; 7406 } 7407 default: 7408 ShouldNotReachHere(); 7409 } 7410 7411 // Should the current free range be coalesced? 7412 // If the chunk is in a free range and either we decided to coalesce above 7413 // or the chunk is near the large block at the end of the heap 7414 // (isNearLargestChunk() returns true), then coalesce this chunk. 7415 const bool doCoalesce = inFreeRange() 7416 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7417 if (doCoalesce) { 7418 // Coalesce the current free range on the left with the new 7419 // chunk on the right. If either is on a free list, 7420 // it must be removed from the list and stashed in the closure. 7421 if (freeRangeInFreeLists()) { 7422 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7423 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7424 "Size of free range is inconsistent with chunk size."); 7425 if (CMSTestInFreeList) { 7426 assert(_sp->verify_chunk_in_free_list(ffc), 7427 "Chunk is not in free lists"); 7428 } 7429 _sp->coalDeath(ffc->size()); 7430 _sp->removeFreeChunkFromFreeLists(ffc); 7431 set_freeRangeInFreeLists(false); 7432 } 7433 if (fcInFreeLists) { 7434 _sp->coalDeath(chunkSize); 7435 assert(fc->size() == chunkSize, 7436 "The chunk has the wrong size or is not in the free lists"); 7437 _sp->removeFreeChunkFromFreeLists(fc); 7438 } 7439 set_lastFreeRangeCoalesced(true); 7440 print_free_block_coalesced(fc); 7441 } else { // not in a free range and/or should not coalesce 7442 // Return the current free range and start a new one. 7443 if (inFreeRange()) { 7444 // In a free range but cannot coalesce with the right hand chunk. 7445 // Put the current free range into the free lists. 7446 flush_cur_free_chunk(freeFinger(), 7447 pointer_delta(fc_addr, freeFinger())); 7448 } 7449 // Set up for new free range. Pass along whether the right hand 7450 // chunk is in the free lists. 7451 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7452 } 7453 } 7454 7455 // Lookahead flush: 7456 // If we are tracking a free range, and this is the last chunk that 7457 // we'll look at because its end crosses past _limit, we'll preemptively 7458 // flush it along with any free range we may be holding on to. Note that 7459 // this can be the case only for an already free or freshly garbage 7460 // chunk. If this block is an object, it can never straddle 7461 // over _limit. The "straddling" occurs when _limit is set at 7462 // the previous end of the space when this cycle started, and 7463 // a subsequent heap expansion caused the previously co-terminal 7464 // free block to be coalesced with the newly expanded portion, 7465 // thus rendering _limit a non-block-boundary making it dangerous 7466 // for the sweeper to step over and examine. 7467 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7468 assert(inFreeRange(), "Should only be called if currently in a free range."); 7469 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7470 assert(_sp->used_region().contains(eob - 1), 7471 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7472 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7473 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7474 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7475 if (eob >= _limit) { 7476 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7477 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7478 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7479 "[" PTR_FORMAT "," PTR_FORMAT ")", 7480 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7481 // Return the storage we are tracking back into the free lists. 7482 log_develop_trace(gc, sweep)("Flushing ... "); 7483 assert(freeFinger() < eob, "Error"); 7484 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7485 } 7486 } 7487 7488 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7489 assert(inFreeRange(), "Should only be called if currently in a free range."); 7490 assert(size > 0, 7491 "A zero sized chunk cannot be added to the free lists."); 7492 if (!freeRangeInFreeLists()) { 7493 if (CMSTestInFreeList) { 7494 FreeChunk* fc = (FreeChunk*) chunk; 7495 fc->set_size(size); 7496 assert(!_sp->verify_chunk_in_free_list(fc), 7497 "chunk should not be in free lists yet"); 7498 } 7499 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7500 // A new free range is going to be starting. The current 7501 // free range has not been added to the free lists yet or 7502 // was removed so add it back. 7503 // If the current free range was coalesced, then the death 7504 // of the free range was recorded. Record a birth now. 7505 if (lastFreeRangeCoalesced()) { 7506 _sp->coalBirth(size); 7507 } 7508 _sp->addChunkAndRepairOffsetTable(chunk, size, 7509 lastFreeRangeCoalesced()); 7510 } else { 7511 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7512 } 7513 set_inFreeRange(false); 7514 set_freeRangeInFreeLists(false); 7515 } 7516 7517 // We take a break if we've been at this for a while, 7518 // so as to avoid monopolizing the locks involved. 7519 void SweepClosure::do_yield_work(HeapWord* addr) { 7520 // Return current free chunk being used for coalescing (if any) 7521 // to the appropriate freelist. After yielding, the next 7522 // free block encountered will start a coalescing range of 7523 // free blocks. If the next free block is adjacent to the 7524 // chunk just flushed, they will need to wait for the next 7525 // sweep to be coalesced. 7526 if (inFreeRange()) { 7527 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7528 } 7529 7530 // First give up the locks, then yield, then re-lock. 7531 // We should probably use a constructor/destructor idiom to 7532 // do this unlock/lock or modify the MutexUnlocker class to 7533 // serve our purpose. XXX 7534 assert_lock_strong(_bitMap->lock()); 7535 assert_lock_strong(_freelistLock); 7536 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7537 "CMS thread should hold CMS token"); 7538 _bitMap->lock()->unlock(); 7539 _freelistLock->unlock(); 7540 ConcurrentMarkSweepThread::desynchronize(true); 7541 _collector->stopTimer(); 7542 _collector->incrementYields(); 7543 7544 // See the comment in coordinator_yield() 7545 for (unsigned i = 0; i < CMSYieldSleepCount && 7546 ConcurrentMarkSweepThread::should_yield() && 7547 !CMSCollector::foregroundGCIsActive(); ++i) { 7548 os::sleep(Thread::current(), 1, false); 7549 } 7550 7551 ConcurrentMarkSweepThread::synchronize(true); 7552 _freelistLock->lock(); 7553 _bitMap->lock()->lock_without_safepoint_check(); 7554 _collector->startTimer(); 7555 } 7556 7557 #ifndef PRODUCT 7558 // This is actually very useful in a product build if it can 7559 // be called from the debugger. Compile it into the product 7560 // as needed. 7561 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7562 return debug_cms_space->verify_chunk_in_free_list(fc); 7563 } 7564 #endif 7565 7566 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7567 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7568 p2i(fc), fc->size()); 7569 } 7570 7571 // CMSIsAliveClosure 7572 bool CMSIsAliveClosure::do_object_b(oop obj) { 7573 HeapWord* addr = (HeapWord*)obj; 7574 return addr != NULL && 7575 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7576 } 7577 7578 7579 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7580 MemRegion span, 7581 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7582 bool cpc): 7583 _collector(collector), 7584 _span(span), 7585 _bit_map(bit_map), 7586 _mark_stack(mark_stack), 7587 _concurrent_precleaning(cpc) { 7588 assert(!_span.is_empty(), "Empty span could spell trouble"); 7589 } 7590 7591 7592 // CMSKeepAliveClosure: the serial version 7593 void CMSKeepAliveClosure::do_oop(oop obj) { 7594 HeapWord* addr = (HeapWord*)obj; 7595 if (_span.contains(addr) && 7596 !_bit_map->isMarked(addr)) { 7597 _bit_map->mark(addr); 7598 bool simulate_overflow = false; 7599 NOT_PRODUCT( 7600 if (CMSMarkStackOverflowALot && 7601 _collector->simulate_overflow()) { 7602 // simulate a stack overflow 7603 simulate_overflow = true; 7604 } 7605 ) 7606 if (simulate_overflow || !_mark_stack->push(obj)) { 7607 if (_concurrent_precleaning) { 7608 // We dirty the overflown object and let the remark 7609 // phase deal with it. 7610 assert(_collector->overflow_list_is_empty(), "Error"); 7611 // In the case of object arrays, we need to dirty all of 7612 // the cards that the object spans. No locking or atomics 7613 // are needed since no one else can be mutating the mod union 7614 // table. 7615 if (obj->is_objArray()) { 7616 size_t sz = obj->size(); 7617 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 7618 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7619 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7620 _collector->_modUnionTable.mark_range(redirty_range); 7621 } else { 7622 _collector->_modUnionTable.mark(addr); 7623 } 7624 _collector->_ser_kac_preclean_ovflw++; 7625 } else { 7626 _collector->push_on_overflow_list(obj); 7627 _collector->_ser_kac_ovflw++; 7628 } 7629 } 7630 } 7631 } 7632 7633 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7634 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7635 7636 // CMSParKeepAliveClosure: a parallel version of the above. 7637 // The work queues are private to each closure (thread), 7638 // but (may be) available for stealing by other threads. 7639 void CMSParKeepAliveClosure::do_oop(oop obj) { 7640 HeapWord* addr = (HeapWord*)obj; 7641 if (_span.contains(addr) && 7642 !_bit_map->isMarked(addr)) { 7643 // In general, during recursive tracing, several threads 7644 // may be concurrently getting here; the first one to 7645 // "tag" it, claims it. 7646 if (_bit_map->par_mark(addr)) { 7647 bool res = _work_queue->push(obj); 7648 assert(res, "Low water mark should be much less than capacity"); 7649 // Do a recursive trim in the hope that this will keep 7650 // stack usage lower, but leave some oops for potential stealers 7651 trim_queue(_low_water_mark); 7652 } // Else, another thread got there first 7653 } 7654 } 7655 7656 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7657 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7658 7659 void CMSParKeepAliveClosure::trim_queue(uint max) { 7660 while (_work_queue->size() > max) { 7661 oop new_oop; 7662 if (_work_queue->pop_local(new_oop)) { 7663 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7664 assert(_bit_map->isMarked((HeapWord*)new_oop), 7665 "no white objects on this stack!"); 7666 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7667 // iterate over the oops in this oop, marking and pushing 7668 // the ones in CMS heap (i.e. in _span). 7669 new_oop->oop_iterate(&_mark_and_push); 7670 } 7671 } 7672 } 7673 7674 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7675 CMSCollector* collector, 7676 MemRegion span, CMSBitMap* bit_map, 7677 OopTaskQueue* work_queue): 7678 _collector(collector), 7679 _span(span), 7680 _bit_map(bit_map), 7681 _work_queue(work_queue) { } 7682 7683 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7684 HeapWord* addr = (HeapWord*)obj; 7685 if (_span.contains(addr) && 7686 !_bit_map->isMarked(addr)) { 7687 if (_bit_map->par_mark(addr)) { 7688 bool simulate_overflow = false; 7689 NOT_PRODUCT( 7690 if (CMSMarkStackOverflowALot && 7691 _collector->par_simulate_overflow()) { 7692 // simulate a stack overflow 7693 simulate_overflow = true; 7694 } 7695 ) 7696 if (simulate_overflow || !_work_queue->push(obj)) { 7697 _collector->par_push_on_overflow_list(obj); 7698 _collector->_par_kac_ovflw++; 7699 } 7700 } // Else another thread got there already 7701 } 7702 } 7703 7704 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7705 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7706 7707 ////////////////////////////////////////////////////////////////// 7708 // CMSExpansionCause ///////////////////////////// 7709 ////////////////////////////////////////////////////////////////// 7710 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7711 switch (cause) { 7712 case _no_expansion: 7713 return "No expansion"; 7714 case _satisfy_free_ratio: 7715 return "Free ratio"; 7716 case _satisfy_promotion: 7717 return "Satisfy promotion"; 7718 case _satisfy_allocation: 7719 return "allocation"; 7720 case _allocate_par_lab: 7721 return "Par LAB"; 7722 case _allocate_par_spooling_space: 7723 return "Par Spooling Space"; 7724 case _adaptive_size_policy: 7725 return "Ergonomics"; 7726 default: 7727 return "unknown"; 7728 } 7729 } 7730 7731 void CMSDrainMarkingStackClosure::do_void() { 7732 // the max number to take from overflow list at a time 7733 const size_t num = _mark_stack->capacity()/4; 7734 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7735 "Overflow list should be NULL during concurrent phases"); 7736 while (!_mark_stack->isEmpty() || 7737 // if stack is empty, check the overflow list 7738 _collector->take_from_overflow_list(num, _mark_stack)) { 7739 oop obj = _mark_stack->pop(); 7740 HeapWord* addr = (HeapWord*)obj; 7741 assert(_span.contains(addr), "Should be within span"); 7742 assert(_bit_map->isMarked(addr), "Should be marked"); 7743 assert(oopDesc::is_oop(obj), "Should be an oop"); 7744 obj->oop_iterate(_keep_alive); 7745 } 7746 } 7747 7748 void CMSParDrainMarkingStackClosure::do_void() { 7749 // drain queue 7750 trim_queue(0); 7751 } 7752 7753 // Trim our work_queue so its length is below max at return 7754 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7755 while (_work_queue->size() > max) { 7756 oop new_oop; 7757 if (_work_queue->pop_local(new_oop)) { 7758 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7759 assert(_bit_map->isMarked((HeapWord*)new_oop), 7760 "no white objects on this stack!"); 7761 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7762 // iterate over the oops in this oop, marking and pushing 7763 // the ones in CMS heap (i.e. in _span). 7764 new_oop->oop_iterate(&_mark_and_push); 7765 } 7766 } 7767 } 7768 7769 //////////////////////////////////////////////////////////////////// 7770 // Support for Marking Stack Overflow list handling and related code 7771 //////////////////////////////////////////////////////////////////// 7772 // Much of the following code is similar in shape and spirit to the 7773 // code used in ParNewGC. We should try and share that code 7774 // as much as possible in the future. 7775 7776 #ifndef PRODUCT 7777 // Debugging support for CMSStackOverflowALot 7778 7779 // It's OK to call this multi-threaded; the worst thing 7780 // that can happen is that we'll get a bunch of closely 7781 // spaced simulated overflows, but that's OK, in fact 7782 // probably good as it would exercise the overflow code 7783 // under contention. 7784 bool CMSCollector::simulate_overflow() { 7785 if (_overflow_counter-- <= 0) { // just being defensive 7786 _overflow_counter = CMSMarkStackOverflowInterval; 7787 return true; 7788 } else { 7789 return false; 7790 } 7791 } 7792 7793 bool CMSCollector::par_simulate_overflow() { 7794 return simulate_overflow(); 7795 } 7796 #endif 7797 7798 // Single-threaded 7799 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7800 assert(stack->isEmpty(), "Expected precondition"); 7801 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7802 size_t i = num; 7803 oop cur = _overflow_list; 7804 const markOop proto = markOopDesc::prototype(); 7805 NOT_PRODUCT(ssize_t n = 0;) 7806 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7807 next = oop(cur->mark()); 7808 cur->set_mark(proto); // until proven otherwise 7809 assert(oopDesc::is_oop(cur), "Should be an oop"); 7810 bool res = stack->push(cur); 7811 assert(res, "Bit off more than can chew?"); 7812 NOT_PRODUCT(n++;) 7813 } 7814 _overflow_list = cur; 7815 #ifndef PRODUCT 7816 assert(_num_par_pushes >= n, "Too many pops?"); 7817 _num_par_pushes -=n; 7818 #endif 7819 return !stack->isEmpty(); 7820 } 7821 7822 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7823 // (MT-safe) Get a prefix of at most "num" from the list. 7824 // The overflow list is chained through the mark word of 7825 // each object in the list. We fetch the entire list, 7826 // break off a prefix of the right size and return the 7827 // remainder. If other threads try to take objects from 7828 // the overflow list at that time, they will wait for 7829 // some time to see if data becomes available. If (and 7830 // only if) another thread places one or more object(s) 7831 // on the global list before we have returned the suffix 7832 // to the global list, we will walk down our local list 7833 // to find its end and append the global list to 7834 // our suffix before returning it. This suffix walk can 7835 // prove to be expensive (quadratic in the amount of traffic) 7836 // when there are many objects in the overflow list and 7837 // there is much producer-consumer contention on the list. 7838 // *NOTE*: The overflow list manipulation code here and 7839 // in ParNewGeneration:: are very similar in shape, 7840 // except that in the ParNew case we use the old (from/eden) 7841 // copy of the object to thread the list via its klass word. 7842 // Because of the common code, if you make any changes in 7843 // the code below, please check the ParNew version to see if 7844 // similar changes might be needed. 7845 // CR 6797058 has been filed to consolidate the common code. 7846 bool CMSCollector::par_take_from_overflow_list(size_t num, 7847 OopTaskQueue* work_q, 7848 int no_of_gc_threads) { 7849 assert(work_q->size() == 0, "First empty local work queue"); 7850 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7851 if (_overflow_list == NULL) { 7852 return false; 7853 } 7854 // Grab the entire list; we'll put back a suffix 7855 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7856 Thread* tid = Thread::current(); 7857 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7858 // set to ParallelGCThreads. 7859 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7860 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7861 // If the list is busy, we spin for a short while, 7862 // sleeping between attempts to get the list. 7863 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7864 os::sleep(tid, sleep_time_millis, false); 7865 if (_overflow_list == NULL) { 7866 // Nothing left to take 7867 return false; 7868 } else if (_overflow_list != BUSY) { 7869 // Try and grab the prefix 7870 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7871 } 7872 } 7873 // If the list was found to be empty, or we spun long 7874 // enough, we give up and return empty-handed. If we leave 7875 // the list in the BUSY state below, it must be the case that 7876 // some other thread holds the overflow list and will set it 7877 // to a non-BUSY state in the future. 7878 if (prefix == NULL || prefix == BUSY) { 7879 // Nothing to take or waited long enough 7880 if (prefix == NULL) { 7881 // Write back the NULL in case we overwrote it with BUSY above 7882 // and it is still the same value. 7883 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7884 } 7885 return false; 7886 } 7887 assert(prefix != NULL && prefix != BUSY, "Error"); 7888 size_t i = num; 7889 oop cur = prefix; 7890 // Walk down the first "num" objects, unless we reach the end. 7891 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7892 if (cur->mark() == NULL) { 7893 // We have "num" or fewer elements in the list, so there 7894 // is nothing to return to the global list. 7895 // Write back the NULL in lieu of the BUSY we wrote 7896 // above, if it is still the same value. 7897 if (_overflow_list == BUSY) { 7898 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7899 } 7900 } else { 7901 // Chop off the suffix and return it to the global list. 7902 assert(cur->mark() != BUSY, "Error"); 7903 oop suffix_head = cur->mark(); // suffix will be put back on global list 7904 cur->set_mark(NULL); // break off suffix 7905 // It's possible that the list is still in the empty(busy) state 7906 // we left it in a short while ago; in that case we may be 7907 // able to place back the suffix without incurring the cost 7908 // of a walk down the list. 7909 oop observed_overflow_list = _overflow_list; 7910 oop cur_overflow_list = observed_overflow_list; 7911 bool attached = false; 7912 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7913 observed_overflow_list = 7914 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7915 if (cur_overflow_list == observed_overflow_list) { 7916 attached = true; 7917 break; 7918 } else cur_overflow_list = observed_overflow_list; 7919 } 7920 if (!attached) { 7921 // Too bad, someone else sneaked in (at least) an element; we'll need 7922 // to do a splice. Find tail of suffix so we can prepend suffix to global 7923 // list. 7924 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7925 oop suffix_tail = cur; 7926 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7927 "Tautology"); 7928 observed_overflow_list = _overflow_list; 7929 do { 7930 cur_overflow_list = observed_overflow_list; 7931 if (cur_overflow_list != BUSY) { 7932 // Do the splice ... 7933 suffix_tail->set_mark(markOop(cur_overflow_list)); 7934 } else { // cur_overflow_list == BUSY 7935 suffix_tail->set_mark(NULL); 7936 } 7937 // ... and try to place spliced list back on overflow_list ... 7938 observed_overflow_list = 7939 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7940 } while (cur_overflow_list != observed_overflow_list); 7941 // ... until we have succeeded in doing so. 7942 } 7943 } 7944 7945 // Push the prefix elements on work_q 7946 assert(prefix != NULL, "control point invariant"); 7947 const markOop proto = markOopDesc::prototype(); 7948 oop next; 7949 NOT_PRODUCT(ssize_t n = 0;) 7950 for (cur = prefix; cur != NULL; cur = next) { 7951 next = oop(cur->mark()); 7952 cur->set_mark(proto); // until proven otherwise 7953 assert(oopDesc::is_oop(cur), "Should be an oop"); 7954 bool res = work_q->push(cur); 7955 assert(res, "Bit off more than we can chew?"); 7956 NOT_PRODUCT(n++;) 7957 } 7958 #ifndef PRODUCT 7959 assert(_num_par_pushes >= n, "Too many pops?"); 7960 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7961 #endif 7962 return true; 7963 } 7964 7965 // Single-threaded 7966 void CMSCollector::push_on_overflow_list(oop p) { 7967 NOT_PRODUCT(_num_par_pushes++;) 7968 assert(oopDesc::is_oop(p), "Not an oop"); 7969 preserve_mark_if_necessary(p); 7970 p->set_mark((markOop)_overflow_list); 7971 _overflow_list = p; 7972 } 7973 7974 // Multi-threaded; use CAS to prepend to overflow list 7975 void CMSCollector::par_push_on_overflow_list(oop p) { 7976 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7977 assert(oopDesc::is_oop(p), "Not an oop"); 7978 par_preserve_mark_if_necessary(p); 7979 oop observed_overflow_list = _overflow_list; 7980 oop cur_overflow_list; 7981 do { 7982 cur_overflow_list = observed_overflow_list; 7983 if (cur_overflow_list != BUSY) { 7984 p->set_mark(markOop(cur_overflow_list)); 7985 } else { 7986 p->set_mark(NULL); 7987 } 7988 observed_overflow_list = 7989 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7990 } while (cur_overflow_list != observed_overflow_list); 7991 } 7992 #undef BUSY 7993 7994 // Single threaded 7995 // General Note on GrowableArray: pushes may silently fail 7996 // because we are (temporarily) out of C-heap for expanding 7997 // the stack. The problem is quite ubiquitous and affects 7998 // a lot of code in the JVM. The prudent thing for GrowableArray 7999 // to do (for now) is to exit with an error. However, that may 8000 // be too draconian in some cases because the caller may be 8001 // able to recover without much harm. For such cases, we 8002 // should probably introduce a "soft_push" method which returns 8003 // an indication of success or failure with the assumption that 8004 // the caller may be able to recover from a failure; code in 8005 // the VM can then be changed, incrementally, to deal with such 8006 // failures where possible, thus, incrementally hardening the VM 8007 // in such low resource situations. 8008 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8009 _preserved_oop_stack.push(p); 8010 _preserved_mark_stack.push(m); 8011 assert(m == p->mark(), "Mark word changed"); 8012 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8013 "bijection"); 8014 } 8015 8016 // Single threaded 8017 void CMSCollector::preserve_mark_if_necessary(oop p) { 8018 markOop m = p->mark(); 8019 if (m->must_be_preserved(p)) { 8020 preserve_mark_work(p, m); 8021 } 8022 } 8023 8024 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8025 markOop m = p->mark(); 8026 if (m->must_be_preserved(p)) { 8027 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8028 // Even though we read the mark word without holding 8029 // the lock, we are assured that it will not change 8030 // because we "own" this oop, so no other thread can 8031 // be trying to push it on the overflow list; see 8032 // the assertion in preserve_mark_work() that checks 8033 // that m == p->mark(). 8034 preserve_mark_work(p, m); 8035 } 8036 } 8037 8038 // We should be able to do this multi-threaded, 8039 // a chunk of stack being a task (this is 8040 // correct because each oop only ever appears 8041 // once in the overflow list. However, it's 8042 // not very easy to completely overlap this with 8043 // other operations, so will generally not be done 8044 // until all work's been completed. Because we 8045 // expect the preserved oop stack (set) to be small, 8046 // it's probably fine to do this single-threaded. 8047 // We can explore cleverer concurrent/overlapped/parallel 8048 // processing of preserved marks if we feel the 8049 // need for this in the future. Stack overflow should 8050 // be so rare in practice and, when it happens, its 8051 // effect on performance so great that this will 8052 // likely just be in the noise anyway. 8053 void CMSCollector::restore_preserved_marks_if_any() { 8054 assert(SafepointSynchronize::is_at_safepoint(), 8055 "world should be stopped"); 8056 assert(Thread::current()->is_ConcurrentGC_thread() || 8057 Thread::current()->is_VM_thread(), 8058 "should be single-threaded"); 8059 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8060 "bijection"); 8061 8062 while (!_preserved_oop_stack.is_empty()) { 8063 oop p = _preserved_oop_stack.pop(); 8064 assert(oopDesc::is_oop(p), "Should be an oop"); 8065 assert(_span.contains(p), "oop should be in _span"); 8066 assert(p->mark() == markOopDesc::prototype(), 8067 "Set when taken from overflow list"); 8068 markOop m = _preserved_mark_stack.pop(); 8069 p->set_mark(m); 8070 } 8071 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8072 "stacks were cleared above"); 8073 } 8074 8075 #ifndef PRODUCT 8076 bool CMSCollector::no_preserved_marks() const { 8077 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8078 } 8079 #endif 8080 8081 // Transfer some number of overflown objects to usual marking 8082 // stack. Return true if some objects were transferred. 8083 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8084 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8085 (size_t)ParGCDesiredObjsFromOverflowList); 8086 8087 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8088 assert(_collector->overflow_list_is_empty() || res, 8089 "If list is not empty, we should have taken something"); 8090 assert(!res || !_mark_stack->isEmpty(), 8091 "If we took something, it should now be on our stack"); 8092 return res; 8093 } 8094 8095 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8096 size_t res = _sp->block_size_no_stall(addr, _collector); 8097 if (_sp->block_is_obj(addr)) { 8098 if (_live_bit_map->isMarked(addr)) { 8099 // It can't have been dead in a previous cycle 8100 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8101 } else { 8102 _dead_bit_map->mark(addr); // mark the dead object 8103 } 8104 } 8105 // Could be 0, if the block size could not be computed without stalling. 8106 return res; 8107 } 8108 8109 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8110 8111 switch (phase) { 8112 case CMSCollector::InitialMarking: 8113 initialize(true /* fullGC */ , 8114 cause /* cause of the GC */, 8115 true /* recordGCBeginTime */, 8116 true /* recordPreGCUsage */, 8117 false /* recordPeakUsage */, 8118 false /* recordPostGCusage */, 8119 true /* recordAccumulatedGCTime */, 8120 false /* recordGCEndTime */, 8121 false /* countCollection */ ); 8122 break; 8123 8124 case CMSCollector::FinalMarking: 8125 initialize(true /* fullGC */ , 8126 cause /* cause of the GC */, 8127 false /* recordGCBeginTime */, 8128 false /* recordPreGCUsage */, 8129 false /* recordPeakUsage */, 8130 false /* recordPostGCusage */, 8131 true /* recordAccumulatedGCTime */, 8132 false /* recordGCEndTime */, 8133 false /* countCollection */ ); 8134 break; 8135 8136 case CMSCollector::Sweeping: 8137 initialize(true /* fullGC */ , 8138 cause /* cause of the GC */, 8139 false /* recordGCBeginTime */, 8140 false /* recordPreGCUsage */, 8141 true /* recordPeakUsage */, 8142 true /* recordPostGCusage */, 8143 false /* recordAccumulatedGCTime */, 8144 true /* recordGCEndTime */, 8145 true /* countCollection */ ); 8146 break; 8147 8148 default: 8149 ShouldNotReachHere(); 8150 } 8151 }