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/cmsHeap.hpp" 33 #include "gc/cms/cmsOopClosures.inline.hpp" 34 #include "gc/cms/compactibleFreeListSpace.hpp" 35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc/cms/concurrentMarkSweepThread.hpp" 37 #include "gc/cms/parNewGeneration.hpp" 38 #include "gc/cms/vmCMSOperations.hpp" 39 #include "gc/serial/genMarkSweep.hpp" 40 #include "gc/serial/tenuredGeneration.hpp" 41 #include "gc/shared/adaptiveSizePolicy.hpp" 42 #include "gc/shared/cardGeneration.inline.hpp" 43 #include "gc/shared/cardTableRS.hpp" 44 #include "gc/shared/collectedHeap.inline.hpp" 45 #include "gc/shared/collectorCounters.hpp" 46 #include "gc/shared/collectorPolicy.hpp" 47 #include "gc/shared/gcLocker.inline.hpp" 48 #include "gc/shared/gcPolicyCounters.hpp" 49 #include "gc/shared/gcTimer.hpp" 50 #include "gc/shared/gcTrace.hpp" 51 #include "gc/shared/gcTraceTime.inline.hpp" 52 #include "gc/shared/genCollectedHeap.hpp" 53 #include "gc/shared/genOopClosures.inline.hpp" 54 #include "gc/shared/isGCActiveMark.hpp" 55 #include "gc/shared/referencePolicy.hpp" 56 #include "gc/shared/strongRootsScope.hpp" 57 #include "gc/shared/taskqueue.inline.hpp" 58 #include "gc/shared/weakProcessor.hpp" 59 #include "logging/log.hpp" 60 #include "logging/logStream.hpp" 61 #include "memory/allocation.hpp" 62 #include "memory/iterator.inline.hpp" 63 #include "memory/padded.hpp" 64 #include "memory/resourceArea.hpp" 65 #include "oops/oop.inline.hpp" 66 #include "prims/jvmtiExport.hpp" 67 #include "runtime/atomic.hpp" 68 #include "runtime/globals_extension.hpp" 69 #include "runtime/handles.inline.hpp" 70 #include "runtime/java.hpp" 71 #include "runtime/orderAccess.inline.hpp" 72 #include "runtime/timer.hpp" 73 #include "runtime/vmThread.hpp" 74 #include "services/memoryService.hpp" 75 #include "services/runtimeService.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/stack.inline.hpp" 78 79 // statics 80 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 81 bool CMSCollector::_full_gc_requested = false; 82 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 83 84 ////////////////////////////////////////////////////////////////// 85 // In support of CMS/VM thread synchronization 86 ////////////////////////////////////////////////////////////////// 87 // We split use of the CGC_lock into 2 "levels". 88 // The low-level locking is of the usual CGC_lock monitor. We introduce 89 // a higher level "token" (hereafter "CMS token") built on top of the 90 // low level monitor (hereafter "CGC lock"). 91 // The token-passing protocol gives priority to the VM thread. The 92 // CMS-lock doesn't provide any fairness guarantees, but clients 93 // should ensure that it is only held for very short, bounded 94 // durations. 95 // 96 // When either of the CMS thread or the VM thread is involved in 97 // collection operations during which it does not want the other 98 // thread to interfere, it obtains the CMS token. 99 // 100 // If either thread tries to get the token while the other has 101 // it, that thread waits. However, if the VM thread and CMS thread 102 // both want the token, then the VM thread gets priority while the 103 // CMS thread waits. This ensures, for instance, that the "concurrent" 104 // phases of the CMS thread's work do not block out the VM thread 105 // for long periods of time as the CMS thread continues to hog 106 // the token. (See bug 4616232). 107 // 108 // The baton-passing functions are, however, controlled by the 109 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 110 // and here the low-level CMS lock, not the high level token, 111 // ensures mutual exclusion. 112 // 113 // Two important conditions that we have to satisfy: 114 // 1. if a thread does a low-level wait on the CMS lock, then it 115 // relinquishes the CMS token if it were holding that token 116 // when it acquired the low-level CMS lock. 117 // 2. any low-level notifications on the low-level lock 118 // should only be sent when a thread has relinquished the token. 119 // 120 // In the absence of either property, we'd have potential deadlock. 121 // 122 // We protect each of the CMS (concurrent and sequential) phases 123 // with the CMS _token_, not the CMS _lock_. 124 // 125 // The only code protected by CMS lock is the token acquisition code 126 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 127 // baton-passing code. 128 // 129 // Unfortunately, i couldn't come up with a good abstraction to factor and 130 // hide the naked CGC_lock manipulation in the baton-passing code 131 // further below. That's something we should try to do. Also, the proof 132 // of correctness of this 2-level locking scheme is far from obvious, 133 // and potentially quite slippery. We have an uneasy suspicion, for instance, 134 // that there may be a theoretical possibility of delay/starvation in the 135 // low-level lock/wait/notify scheme used for the baton-passing because of 136 // potential interference with the priority scheme embodied in the 137 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 138 // invocation further below and marked with "XXX 20011219YSR". 139 // Indeed, as we note elsewhere, this may become yet more slippery 140 // in the presence of multiple CMS and/or multiple VM threads. XXX 141 142 class CMSTokenSync: public StackObj { 143 private: 144 bool _is_cms_thread; 145 public: 146 CMSTokenSync(bool is_cms_thread): 147 _is_cms_thread(is_cms_thread) { 148 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 149 "Incorrect argument to constructor"); 150 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 151 } 152 153 ~CMSTokenSync() { 154 assert(_is_cms_thread ? 155 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 156 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 157 "Incorrect state"); 158 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 159 } 160 }; 161 162 // Convenience class that does a CMSTokenSync, and then acquires 163 // upto three locks. 164 class CMSTokenSyncWithLocks: public CMSTokenSync { 165 private: 166 // Note: locks are acquired in textual declaration order 167 // and released in the opposite order 168 MutexLockerEx _locker1, _locker2, _locker3; 169 public: 170 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 171 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 172 CMSTokenSync(is_cms_thread), 173 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 174 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 175 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 176 { } 177 }; 178 179 180 ////////////////////////////////////////////////////////////////// 181 // Concurrent Mark-Sweep Generation ///////////////////////////// 182 ////////////////////////////////////////////////////////////////// 183 184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 185 186 // This struct contains per-thread things necessary to support parallel 187 // young-gen collection. 188 class CMSParGCThreadState: public CHeapObj<mtGC> { 189 public: 190 CompactibleFreeListSpaceLAB lab; 191 PromotionInfo promo; 192 193 // Constructor. 194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 195 promo.setSpace(cfls); 196 } 197 }; 198 199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 200 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 201 CardGeneration(rs, initial_byte_size, ct), 202 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 203 _did_compact(false) 204 { 205 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 206 HeapWord* end = (HeapWord*) _virtual_space.high(); 207 208 _direct_allocated_words = 0; 209 NOT_PRODUCT( 210 _numObjectsPromoted = 0; 211 _numWordsPromoted = 0; 212 _numObjectsAllocated = 0; 213 _numWordsAllocated = 0; 214 ) 215 216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 217 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 218 _cmsSpace->_old_gen = this; 219 220 _gc_stats = new CMSGCStats(); 221 222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 223 // offsets match. The ability to tell free chunks from objects 224 // depends on this property. 225 debug_only( 226 FreeChunk* junk = NULL; 227 assert(UseCompressedClassPointers || 228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 229 "Offset of FreeChunk::_prev within FreeChunk must match" 230 " that of OopDesc::_klass within OopDesc"); 231 ) 232 233 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 234 for (uint i = 0; i < ParallelGCThreads; i++) { 235 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 236 } 237 238 _incremental_collection_failed = false; 239 // The "dilatation_factor" is the expansion that can occur on 240 // account of the fact that the minimum object size in the CMS 241 // generation may be larger than that in, say, a contiguous young 242 // generation. 243 // Ideally, in the calculation below, we'd compute the dilatation 244 // factor as: MinChunkSize/(promoting_gen's min object size) 245 // Since we do not have such a general query interface for the 246 // promoting generation, we'll instead just use the minimum 247 // object size (which today is a header's worth of space); 248 // note that all arithmetic is in units of HeapWords. 249 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 250 assert(_dilatation_factor >= 1.0, "from previous assert"); 251 } 252 253 254 // The field "_initiating_occupancy" represents the occupancy percentage 255 // at which we trigger a new collection cycle. Unless explicitly specified 256 // via CMSInitiatingOccupancyFraction (argument "io" below), it 257 // is calculated by: 258 // 259 // Let "f" be MinHeapFreeRatio in 260 // 261 // _initiating_occupancy = 100-f + 262 // f * (CMSTriggerRatio/100) 263 // where CMSTriggerRatio is the argument "tr" below. 264 // 265 // That is, if we assume the heap is at its desired maximum occupancy at the 266 // end of a collection, we let CMSTriggerRatio of the (purported) free 267 // space be allocated before initiating a new collection cycle. 268 // 269 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 270 assert(io <= 100 && tr <= 100, "Check the arguments"); 271 if (io >= 0) { 272 _initiating_occupancy = (double)io / 100.0; 273 } else { 274 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 275 (double)(tr * MinHeapFreeRatio) / 100.0) 276 / 100.0; 277 } 278 } 279 280 void ConcurrentMarkSweepGeneration::ref_processor_init() { 281 assert(collector() != NULL, "no collector"); 282 collector()->ref_processor_init(); 283 } 284 285 void CMSCollector::ref_processor_init() { 286 if (_ref_processor == NULL) { 287 // Allocate and initialize a reference processor 288 _ref_processor = 289 new ReferenceProcessor(_span, // span 290 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 291 ParallelGCThreads, // mt processing degree 292 _cmsGen->refs_discovery_is_mt(), // mt discovery 293 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 294 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 295 &_is_alive_closure); // closure for liveness info 296 // Initialize the _ref_processor field of CMSGen 297 _cmsGen->set_ref_processor(_ref_processor); 298 299 } 300 } 301 302 AdaptiveSizePolicy* CMSCollector::size_policy() { 303 CMSHeap* heap = CMSHeap::heap(); 304 return heap->gen_policy()->size_policy(); 305 } 306 307 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 308 309 const char* gen_name = "old"; 310 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 311 // Generation Counters - generation 1, 1 subspace 312 _gen_counters = new GenerationCounters(gen_name, 1, 1, 313 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 314 315 _space_counters = new GSpaceCounters(gen_name, 0, 316 _virtual_space.reserved_size(), 317 this, _gen_counters); 318 } 319 320 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 321 _cms_gen(cms_gen) 322 { 323 assert(alpha <= 100, "bad value"); 324 _saved_alpha = alpha; 325 326 // Initialize the alphas to the bootstrap value of 100. 327 _gc0_alpha = _cms_alpha = 100; 328 329 _cms_begin_time.update(); 330 _cms_end_time.update(); 331 332 _gc0_duration = 0.0; 333 _gc0_period = 0.0; 334 _gc0_promoted = 0; 335 336 _cms_duration = 0.0; 337 _cms_period = 0.0; 338 _cms_allocated = 0; 339 340 _cms_used_at_gc0_begin = 0; 341 _cms_used_at_gc0_end = 0; 342 _allow_duty_cycle_reduction = false; 343 _valid_bits = 0; 344 } 345 346 double CMSStats::cms_free_adjustment_factor(size_t free) const { 347 // TBD: CR 6909490 348 return 1.0; 349 } 350 351 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 352 } 353 354 // If promotion failure handling is on use 355 // the padded average size of the promotion for each 356 // young generation collection. 357 double CMSStats::time_until_cms_gen_full() const { 358 size_t cms_free = _cms_gen->cmsSpace()->free(); 359 CMSHeap* heap = CMSHeap::heap(); 360 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 361 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 362 if (cms_free > expected_promotion) { 363 // Start a cms collection if there isn't enough space to promote 364 // for the next young collection. Use the padded average as 365 // a safety factor. 366 cms_free -= expected_promotion; 367 368 // Adjust by the safety factor. 369 double cms_free_dbl = (double)cms_free; 370 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 371 // Apply a further correction factor which tries to adjust 372 // for recent occurance of concurrent mode failures. 373 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 374 cms_free_dbl = cms_free_dbl * cms_adjustment; 375 376 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 377 cms_free, expected_promotion); 378 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 379 // Add 1 in case the consumption rate goes to zero. 380 return cms_free_dbl / (cms_consumption_rate() + 1.0); 381 } 382 return 0.0; 383 } 384 385 // Compare the duration of the cms collection to the 386 // time remaining before the cms generation is empty. 387 // Note that the time from the start of the cms collection 388 // to the start of the cms sweep (less than the total 389 // duration of the cms collection) can be used. This 390 // has been tried and some applications experienced 391 // promotion failures early in execution. This was 392 // possibly because the averages were not accurate 393 // enough at the beginning. 394 double CMSStats::time_until_cms_start() const { 395 // We add "gc0_period" to the "work" calculation 396 // below because this query is done (mostly) at the 397 // end of a scavenge, so we need to conservatively 398 // account for that much possible delay 399 // in the query so as to avoid concurrent mode failures 400 // due to starting the collection just a wee bit too 401 // late. 402 double work = cms_duration() + gc0_period(); 403 double deadline = time_until_cms_gen_full(); 404 // If a concurrent mode failure occurred recently, we want to be 405 // more conservative and halve our expected time_until_cms_gen_full() 406 if (work > deadline) { 407 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 408 cms_duration(), gc0_period(), time_until_cms_gen_full()); 409 return 0.0; 410 } 411 return work - deadline; 412 } 413 414 #ifndef PRODUCT 415 void CMSStats::print_on(outputStream *st) const { 416 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 417 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 418 gc0_duration(), gc0_period(), gc0_promoted()); 419 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 420 cms_duration(), cms_period(), cms_allocated()); 421 st->print(",cms_since_beg=%g,cms_since_end=%g", 422 cms_time_since_begin(), cms_time_since_end()); 423 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 424 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 425 426 if (valid()) { 427 st->print(",promo_rate=%g,cms_alloc_rate=%g", 428 promotion_rate(), cms_allocation_rate()); 429 st->print(",cms_consumption_rate=%g,time_until_full=%g", 430 cms_consumption_rate(), time_until_cms_gen_full()); 431 } 432 st->cr(); 433 } 434 #endif // #ifndef PRODUCT 435 436 CMSCollector::CollectorState CMSCollector::_collectorState = 437 CMSCollector::Idling; 438 bool CMSCollector::_foregroundGCIsActive = false; 439 bool CMSCollector::_foregroundGCShouldWait = false; 440 441 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 442 CardTableRS* ct, 443 ConcurrentMarkSweepPolicy* cp): 444 _cmsGen(cmsGen), 445 _ct(ct), 446 _ref_processor(NULL), // will be set later 447 _conc_workers(NULL), // may be set later 448 _abort_preclean(false), 449 _start_sampling(false), 450 _between_prologue_and_epilogue(false), 451 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 452 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 453 -1 /* lock-free */, "No_lock" /* dummy */), 454 _modUnionClosurePar(&_modUnionTable), 455 // Adjust my span to cover old (cms) gen 456 _span(cmsGen->reserved()), 457 // Construct the is_alive_closure with _span & markBitMap 458 _is_alive_closure(_span, &_markBitMap), 459 _restart_addr(NULL), 460 _overflow_list(NULL), 461 _stats(cmsGen), 462 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 463 //verify that this lock should be acquired with safepoint check. 464 Monitor::_safepoint_check_sometimes)), 465 _eden_chunk_array(NULL), // may be set in ctor body 466 _eden_chunk_capacity(0), // -- ditto -- 467 _eden_chunk_index(0), // -- ditto -- 468 _survivor_plab_array(NULL), // -- ditto -- 469 _survivor_chunk_array(NULL), // -- ditto -- 470 _survivor_chunk_capacity(0), // -- ditto -- 471 _survivor_chunk_index(0), // -- ditto -- 472 _ser_pmc_preclean_ovflw(0), 473 _ser_kac_preclean_ovflw(0), 474 _ser_pmc_remark_ovflw(0), 475 _par_pmc_remark_ovflw(0), 476 _ser_kac_ovflw(0), 477 _par_kac_ovflw(0), 478 #ifndef PRODUCT 479 _num_par_pushes(0), 480 #endif 481 _collection_count_start(0), 482 _verifying(false), 483 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 484 _completed_initialization(false), 485 _collector_policy(cp), 486 _should_unload_classes(CMSClassUnloadingEnabled), 487 _concurrent_cycles_since_last_unload(0), 488 _roots_scanning_options(GenCollectedHeap::SO_None), 489 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 490 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 491 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 492 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 493 _cms_start_registered(false) 494 { 495 // Now expand the span and allocate the collection support structures 496 // (MUT, marking bit map etc.) to cover both generations subject to 497 // collection. 498 499 // For use by dirty card to oop closures. 500 _cmsGen->cmsSpace()->set_collector(this); 501 502 // Allocate MUT and marking bit map 503 { 504 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 505 if (!_markBitMap.allocate(_span)) { 506 log_warning(gc)("Failed to allocate CMS Bit Map"); 507 return; 508 } 509 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 510 } 511 { 512 _modUnionTable.allocate(_span); 513 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 514 } 515 516 if (!_markStack.allocate(MarkStackSize)) { 517 log_warning(gc)("Failed to allocate CMS Marking Stack"); 518 return; 519 } 520 521 // Support for multi-threaded concurrent phases 522 if (CMSConcurrentMTEnabled) { 523 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 524 // just for now 525 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 526 } 527 if (ConcGCThreads > 1) { 528 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 529 ConcGCThreads, true); 530 if (_conc_workers == NULL) { 531 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 532 CMSConcurrentMTEnabled = false; 533 } else { 534 _conc_workers->initialize_workers(); 535 } 536 } else { 537 CMSConcurrentMTEnabled = false; 538 } 539 } 540 if (!CMSConcurrentMTEnabled) { 541 ConcGCThreads = 0; 542 } else { 543 // Turn off CMSCleanOnEnter optimization temporarily for 544 // the MT case where it's not fixed yet; see 6178663. 545 CMSCleanOnEnter = false; 546 } 547 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 548 "Inconsistency"); 549 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 550 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 551 552 // Parallel task queues; these are shared for the 553 // concurrent and stop-world phases of CMS, but 554 // are not shared with parallel scavenge (ParNew). 555 { 556 uint i; 557 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 558 559 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 560 || ParallelRefProcEnabled) 561 && num_queues > 0) { 562 _task_queues = new OopTaskQueueSet(num_queues); 563 if (_task_queues == NULL) { 564 log_warning(gc)("task_queues allocation failure."); 565 return; 566 } 567 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 568 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 569 for (i = 0; i < num_queues; i++) { 570 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 571 if (q == NULL) { 572 log_warning(gc)("work_queue allocation failure."); 573 return; 574 } 575 _task_queues->register_queue(i, q); 576 } 577 for (i = 0; i < num_queues; i++) { 578 _task_queues->queue(i)->initialize(); 579 _hash_seed[i] = 17; // copied from ParNew 580 } 581 } 582 } 583 584 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 585 586 // Clip CMSBootstrapOccupancy between 0 and 100. 587 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 588 589 // Now tell CMS generations the identity of their collector 590 ConcurrentMarkSweepGeneration::set_collector(this); 591 592 // Create & start a CMS thread for this CMS collector 593 _cmsThread = ConcurrentMarkSweepThread::start(this); 594 assert(cmsThread() != NULL, "CMS Thread should have been created"); 595 assert(cmsThread()->collector() == this, 596 "CMS Thread should refer to this gen"); 597 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 598 599 // Support for parallelizing young gen rescan 600 CMSHeap* heap = CMSHeap::heap(); 601 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 602 _young_gen = (ParNewGeneration*)heap->young_gen(); 603 if (heap->supports_inline_contig_alloc()) { 604 _top_addr = heap->top_addr(); 605 _end_addr = heap->end_addr(); 606 assert(_young_gen != NULL, "no _young_gen"); 607 _eden_chunk_index = 0; 608 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 609 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 610 } 611 612 // Support for parallelizing survivor space rescan 613 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 614 const size_t max_plab_samples = 615 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 616 617 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 618 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 619 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 620 _survivor_chunk_capacity = max_plab_samples; 621 for (uint i = 0; i < ParallelGCThreads; i++) { 622 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 623 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 624 assert(cur->end() == 0, "Should be 0"); 625 assert(cur->array() == vec, "Should be vec"); 626 assert(cur->capacity() == max_plab_samples, "Error"); 627 } 628 } 629 630 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 631 _gc_counters = new CollectorCounters("CMS", 1); 632 _completed_initialization = true; 633 _inter_sweep_timer.start(); // start of time 634 } 635 636 const char* ConcurrentMarkSweepGeneration::name() const { 637 return "concurrent mark-sweep generation"; 638 } 639 void ConcurrentMarkSweepGeneration::update_counters() { 640 if (UsePerfData) { 641 _space_counters->update_all(); 642 _gen_counters->update_all(); 643 } 644 } 645 646 // this is an optimized version of update_counters(). it takes the 647 // used value as a parameter rather than computing it. 648 // 649 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 650 if (UsePerfData) { 651 _space_counters->update_used(used); 652 _space_counters->update_capacity(); 653 _gen_counters->update_all(); 654 } 655 } 656 657 void ConcurrentMarkSweepGeneration::print() const { 658 Generation::print(); 659 cmsSpace()->print(); 660 } 661 662 #ifndef PRODUCT 663 void ConcurrentMarkSweepGeneration::print_statistics() { 664 cmsSpace()->printFLCensus(0); 665 } 666 #endif 667 668 size_t 669 ConcurrentMarkSweepGeneration::contiguous_available() const { 670 // dld proposes an improvement in precision here. If the committed 671 // part of the space ends in a free block we should add that to 672 // uncommitted size in the calculation below. Will make this 673 // change later, staying with the approximation below for the 674 // time being. -- ysr. 675 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 676 } 677 678 size_t 679 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 680 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 681 } 682 683 size_t ConcurrentMarkSweepGeneration::max_available() const { 684 return free() + _virtual_space.uncommitted_size(); 685 } 686 687 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 688 size_t available = max_available(); 689 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 690 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 691 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 692 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 693 return res; 694 } 695 696 // At a promotion failure dump information on block layout in heap 697 // (cms old generation). 698 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 699 Log(gc, promotion) log; 700 if (log.is_trace()) { 701 LogStream ls(log.trace()); 702 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 703 } 704 } 705 706 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 707 // Clear the promotion information. These pointers can be adjusted 708 // along with all the other pointers into the heap but 709 // compaction is expected to be a rare event with 710 // a heap using cms so don't do it without seeing the need. 711 for (uint i = 0; i < ParallelGCThreads; i++) { 712 _par_gc_thread_states[i]->promo.reset(); 713 } 714 } 715 716 void ConcurrentMarkSweepGeneration::compute_new_size() { 717 assert_locked_or_safepoint(Heap_lock); 718 719 // If incremental collection failed, we just want to expand 720 // to the limit. 721 if (incremental_collection_failed()) { 722 clear_incremental_collection_failed(); 723 grow_to_reserved(); 724 return; 725 } 726 727 // The heap has been compacted but not reset yet. 728 // Any metric such as free() or used() will be incorrect. 729 730 CardGeneration::compute_new_size(); 731 732 // Reset again after a possible resizing 733 if (did_compact()) { 734 cmsSpace()->reset_after_compaction(); 735 } 736 } 737 738 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 739 assert_locked_or_safepoint(Heap_lock); 740 741 // If incremental collection failed, we just want to expand 742 // to the limit. 743 if (incremental_collection_failed()) { 744 clear_incremental_collection_failed(); 745 grow_to_reserved(); 746 return; 747 } 748 749 double free_percentage = ((double) free()) / capacity(); 750 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 751 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 752 753 // compute expansion delta needed for reaching desired free percentage 754 if (free_percentage < desired_free_percentage) { 755 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 756 assert(desired_capacity >= capacity(), "invalid expansion size"); 757 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 758 Log(gc) log; 759 if (log.is_trace()) { 760 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 761 log.trace("From compute_new_size: "); 762 log.trace(" Free fraction %f", free_percentage); 763 log.trace(" Desired free fraction %f", desired_free_percentage); 764 log.trace(" Maximum free fraction %f", maximum_free_percentage); 765 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 766 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 767 CMSHeap* heap = CMSHeap::heap(); 768 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation"); 769 size_t young_size = heap->young_gen()->capacity(); 770 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 771 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 772 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 773 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 774 } 775 // safe if expansion fails 776 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 777 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 778 } else { 779 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 780 assert(desired_capacity <= capacity(), "invalid expansion size"); 781 size_t shrink_bytes = capacity() - desired_capacity; 782 // Don't shrink unless the delta is greater than the minimum shrink we want 783 if (shrink_bytes >= MinHeapDeltaBytes) { 784 shrink_free_list_by(shrink_bytes); 785 } 786 } 787 } 788 789 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 790 return cmsSpace()->freelistLock(); 791 } 792 793 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 794 CMSSynchronousYieldRequest yr; 795 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 796 return have_lock_and_allocate(size, tlab); 797 } 798 799 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 800 bool tlab /* ignored */) { 801 assert_lock_strong(freelistLock()); 802 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 803 HeapWord* res = cmsSpace()->allocate(adjustedSize); 804 // Allocate the object live (grey) if the background collector has 805 // started marking. This is necessary because the marker may 806 // have passed this address and consequently this object will 807 // not otherwise be greyed and would be incorrectly swept up. 808 // Note that if this object contains references, the writing 809 // of those references will dirty the card containing this object 810 // allowing the object to be blackened (and its references scanned) 811 // either during a preclean phase or at the final checkpoint. 812 if (res != NULL) { 813 // We may block here with an uninitialized object with 814 // its mark-bit or P-bits not yet set. Such objects need 815 // to be safely navigable by block_start(). 816 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 817 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 818 collector()->direct_allocated(res, adjustedSize); 819 _direct_allocated_words += adjustedSize; 820 // allocation counters 821 NOT_PRODUCT( 822 _numObjectsAllocated++; 823 _numWordsAllocated += (int)adjustedSize; 824 ) 825 } 826 return res; 827 } 828 829 // In the case of direct allocation by mutators in a generation that 830 // is being concurrently collected, the object must be allocated 831 // live (grey) if the background collector has started marking. 832 // This is necessary because the marker may 833 // have passed this address and consequently this object will 834 // not otherwise be greyed and would be incorrectly swept up. 835 // Note that if this object contains references, the writing 836 // of those references will dirty the card containing this object 837 // allowing the object to be blackened (and its references scanned) 838 // either during a preclean phase or at the final checkpoint. 839 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 840 assert(_markBitMap.covers(start, size), "Out of bounds"); 841 if (_collectorState >= Marking) { 842 MutexLockerEx y(_markBitMap.lock(), 843 Mutex::_no_safepoint_check_flag); 844 // [see comments preceding SweepClosure::do_blk() below for details] 845 // 846 // Can the P-bits be deleted now? JJJ 847 // 848 // 1. need to mark the object as live so it isn't collected 849 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 850 // 3. need to mark the end of the object so marking, precleaning or sweeping 851 // can skip over uninitialized or unparsable objects. An allocated 852 // object is considered uninitialized for our purposes as long as 853 // its klass word is NULL. All old gen objects are parsable 854 // as soon as they are initialized.) 855 _markBitMap.mark(start); // object is live 856 _markBitMap.mark(start + 1); // object is potentially uninitialized? 857 _markBitMap.mark(start + size - 1); 858 // mark end of object 859 } 860 // check that oop looks uninitialized 861 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 862 } 863 864 void CMSCollector::promoted(bool par, HeapWord* start, 865 bool is_obj_array, size_t obj_size) { 866 assert(_markBitMap.covers(start), "Out of bounds"); 867 // See comment in direct_allocated() about when objects should 868 // be allocated live. 869 if (_collectorState >= Marking) { 870 // we already hold the marking bit map lock, taken in 871 // the prologue 872 if (par) { 873 _markBitMap.par_mark(start); 874 } else { 875 _markBitMap.mark(start); 876 } 877 // We don't need to mark the object as uninitialized (as 878 // in direct_allocated above) because this is being done with the 879 // world stopped and the object will be initialized by the 880 // time the marking, precleaning or sweeping get to look at it. 881 // But see the code for copying objects into the CMS generation, 882 // where we need to ensure that concurrent readers of the 883 // block offset table are able to safely navigate a block that 884 // is in flux from being free to being allocated (and in 885 // transition while being copied into) and subsequently 886 // becoming a bona-fide object when the copy/promotion is complete. 887 assert(SafepointSynchronize::is_at_safepoint(), 888 "expect promotion only at safepoints"); 889 890 if (_collectorState < Sweeping) { 891 // Mark the appropriate cards in the modUnionTable, so that 892 // this object gets scanned before the sweep. If this is 893 // not done, CMS generation references in the object might 894 // not get marked. 895 // For the case of arrays, which are otherwise precisely 896 // marked, we need to dirty the entire array, not just its head. 897 if (is_obj_array) { 898 // The [par_]mark_range() method expects mr.end() below to 899 // be aligned to the granularity of a bit's representation 900 // in the heap. In the case of the MUT below, that's a 901 // card size. 902 MemRegion mr(start, 903 align_up(start + obj_size, 904 CardTableModRefBS::card_size /* bytes */)); 905 if (par) { 906 _modUnionTable.par_mark_range(mr); 907 } else { 908 _modUnionTable.mark_range(mr); 909 } 910 } else { // not an obj array; we can just mark the head 911 if (par) { 912 _modUnionTable.par_mark(start); 913 } else { 914 _modUnionTable.mark(start); 915 } 916 } 917 } 918 } 919 } 920 921 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 922 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 923 // allocate, copy and if necessary update promoinfo -- 924 // delegate to underlying space. 925 assert_lock_strong(freelistLock()); 926 927 #ifndef PRODUCT 928 if (CMSHeap::heap()->promotion_should_fail()) { 929 return NULL; 930 } 931 #endif // #ifndef PRODUCT 932 933 oop res = _cmsSpace->promote(obj, obj_size); 934 if (res == NULL) { 935 // expand and retry 936 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 937 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 938 // Since this is the old generation, we don't try to promote 939 // into a more senior generation. 940 res = _cmsSpace->promote(obj, obj_size); 941 } 942 if (res != NULL) { 943 // See comment in allocate() about when objects should 944 // be allocated live. 945 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 946 collector()->promoted(false, // Not parallel 947 (HeapWord*)res, obj->is_objArray(), obj_size); 948 // promotion counters 949 NOT_PRODUCT( 950 _numObjectsPromoted++; 951 _numWordsPromoted += 952 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 953 ) 954 } 955 return res; 956 } 957 958 959 // IMPORTANT: Notes on object size recognition in CMS. 960 // --------------------------------------------------- 961 // A block of storage in the CMS generation is always in 962 // one of three states. A free block (FREE), an allocated 963 // object (OBJECT) whose size() method reports the correct size, 964 // and an intermediate state (TRANSIENT) in which its size cannot 965 // be accurately determined. 966 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 967 // ----------------------------------------------------- 968 // FREE: klass_word & 1 == 1; mark_word holds block size 969 // 970 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 971 // obj->size() computes correct size 972 // 973 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 974 // 975 // STATE IDENTIFICATION: (64 bit+COOPS) 976 // ------------------------------------ 977 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 978 // 979 // OBJECT: klass_word installed; klass_word != 0; 980 // obj->size() computes correct size 981 // 982 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 983 // 984 // 985 // STATE TRANSITION DIAGRAM 986 // 987 // mut / parnew mut / parnew 988 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 989 // ^ | 990 // |------------------------ DEAD <------------------------------------| 991 // sweep mut 992 // 993 // While a block is in TRANSIENT state its size cannot be determined 994 // so readers will either need to come back later or stall until 995 // the size can be determined. Note that for the case of direct 996 // allocation, P-bits, when available, may be used to determine the 997 // size of an object that may not yet have been initialized. 998 999 // Things to support parallel young-gen collection. 1000 oop 1001 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1002 oop old, markOop m, 1003 size_t word_sz) { 1004 #ifndef PRODUCT 1005 if (CMSHeap::heap()->promotion_should_fail()) { 1006 return NULL; 1007 } 1008 #endif // #ifndef PRODUCT 1009 1010 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1011 PromotionInfo* promoInfo = &ps->promo; 1012 // if we are tracking promotions, then first ensure space for 1013 // promotion (including spooling space for saving header if necessary). 1014 // then allocate and copy, then track promoted info if needed. 1015 // When tracking (see PromotionInfo::track()), the mark word may 1016 // be displaced and in this case restoration of the mark word 1017 // occurs in the (oop_since_save_marks_)iterate phase. 1018 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1019 // Out of space for allocating spooling buffers; 1020 // try expanding and allocating spooling buffers. 1021 if (!expand_and_ensure_spooling_space(promoInfo)) { 1022 return NULL; 1023 } 1024 } 1025 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1026 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1027 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1028 if (obj_ptr == NULL) { 1029 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1030 if (obj_ptr == NULL) { 1031 return NULL; 1032 } 1033 } 1034 oop obj = oop(obj_ptr); 1035 OrderAccess::storestore(); 1036 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1037 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1038 // IMPORTANT: See note on object initialization for CMS above. 1039 // Otherwise, copy the object. Here we must be careful to insert the 1040 // klass pointer last, since this marks the block as an allocated object. 1041 // Except with compressed oops it's the mark word. 1042 HeapWord* old_ptr = (HeapWord*)old; 1043 // Restore the mark word copied above. 1044 obj->set_mark(m); 1045 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1046 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1047 OrderAccess::storestore(); 1048 1049 if (UseCompressedClassPointers) { 1050 // Copy gap missed by (aligned) header size calculation below 1051 obj->set_klass_gap(old->klass_gap()); 1052 } 1053 if (word_sz > (size_t)oopDesc::header_size()) { 1054 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1055 obj_ptr + oopDesc::header_size(), 1056 word_sz - oopDesc::header_size()); 1057 } 1058 1059 // Now we can track the promoted object, if necessary. We take care 1060 // to delay the transition from uninitialized to full object 1061 // (i.e., insertion of klass pointer) until after, so that it 1062 // atomically becomes a promoted object. 1063 if (promoInfo->tracking()) { 1064 promoInfo->track((PromotedObject*)obj, old->klass()); 1065 } 1066 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1067 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1068 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1069 1070 // Finally, install the klass pointer (this should be volatile). 1071 OrderAccess::storestore(); 1072 obj->set_klass(old->klass()); 1073 // We should now be able to calculate the right size for this object 1074 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1075 1076 collector()->promoted(true, // parallel 1077 obj_ptr, old->is_objArray(), word_sz); 1078 1079 NOT_PRODUCT( 1080 Atomic::inc(&_numObjectsPromoted); 1081 Atomic::add(alloc_sz, &_numWordsPromoted); 1082 ) 1083 1084 return obj; 1085 } 1086 1087 void 1088 ConcurrentMarkSweepGeneration:: 1089 par_promote_alloc_done(int thread_num) { 1090 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1091 ps->lab.retire(thread_num); 1092 } 1093 1094 void 1095 ConcurrentMarkSweepGeneration:: 1096 par_oop_since_save_marks_iterate_done(int thread_num) { 1097 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1098 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1099 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1100 1101 // Because card-scanning has been completed, subsequent phases 1102 // (e.g., reference processing) will not need to recognize which 1103 // objects have been promoted during this GC. So, we can now disable 1104 // promotion tracking. 1105 ps->promo.stopTrackingPromotions(); 1106 } 1107 1108 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1109 size_t size, 1110 bool tlab) 1111 { 1112 // We allow a STW collection only if a full 1113 // collection was requested. 1114 return full || should_allocate(size, tlab); // FIX ME !!! 1115 // This and promotion failure handling are connected at the 1116 // hip and should be fixed by untying them. 1117 } 1118 1119 bool CMSCollector::shouldConcurrentCollect() { 1120 LogTarget(Trace, gc) log; 1121 1122 if (_full_gc_requested) { 1123 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1124 return true; 1125 } 1126 1127 FreelistLocker x(this); 1128 // ------------------------------------------------------------------ 1129 // Print out lots of information which affects the initiation of 1130 // a collection. 1131 if (log.is_enabled() && stats().valid()) { 1132 log.print("CMSCollector shouldConcurrentCollect: "); 1133 1134 LogStream out(log); 1135 stats().print_on(&out); 1136 1137 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1138 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1139 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1140 log.print("promotion_rate=%g", stats().promotion_rate()); 1141 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1142 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1143 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1144 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1145 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1146 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1147 } 1148 // ------------------------------------------------------------------ 1149 1150 // If the estimated time to complete a cms collection (cms_duration()) 1151 // is less than the estimated time remaining until the cms generation 1152 // is full, start a collection. 1153 if (!UseCMSInitiatingOccupancyOnly) { 1154 if (stats().valid()) { 1155 if (stats().time_until_cms_start() == 0.0) { 1156 return true; 1157 } 1158 } else { 1159 // We want to conservatively collect somewhat early in order 1160 // to try and "bootstrap" our CMS/promotion statistics; 1161 // this branch will not fire after the first successful CMS 1162 // collection because the stats should then be valid. 1163 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1164 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1165 _cmsGen->occupancy(), _bootstrap_occupancy); 1166 return true; 1167 } 1168 } 1169 } 1170 1171 // Otherwise, we start a collection cycle if 1172 // old gen want a collection cycle started. Each may use 1173 // an appropriate criterion for making this decision. 1174 // XXX We need to make sure that the gen expansion 1175 // criterion dovetails well with this. XXX NEED TO FIX THIS 1176 if (_cmsGen->should_concurrent_collect()) { 1177 log.print("CMS old gen initiated"); 1178 return true; 1179 } 1180 1181 // We start a collection if we believe an incremental collection may fail; 1182 // this is not likely to be productive in practice because it's probably too 1183 // late anyway. 1184 CMSHeap* heap = CMSHeap::heap(); 1185 assert(heap->collector_policy()->is_generation_policy(), 1186 "You may want to check the correctness of the following"); 1187 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1188 log.print("CMSCollector: collect because incremental collection will fail "); 1189 return true; 1190 } 1191 1192 if (MetaspaceGC::should_concurrent_collect()) { 1193 log.print("CMSCollector: collect for metadata allocation "); 1194 return true; 1195 } 1196 1197 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1198 if (CMSTriggerInterval >= 0) { 1199 if (CMSTriggerInterval == 0) { 1200 // Trigger always 1201 return true; 1202 } 1203 1204 // Check the CMS time since begin (we do not check the stats validity 1205 // as we want to be able to trigger the first CMS cycle as well) 1206 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1207 if (stats().valid()) { 1208 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1209 stats().cms_time_since_begin()); 1210 } else { 1211 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1212 } 1213 return true; 1214 } 1215 } 1216 1217 return false; 1218 } 1219 1220 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1221 1222 // Clear _expansion_cause fields of constituent generations 1223 void CMSCollector::clear_expansion_cause() { 1224 _cmsGen->clear_expansion_cause(); 1225 } 1226 1227 // We should be conservative in starting a collection cycle. To 1228 // start too eagerly runs the risk of collecting too often in the 1229 // extreme. To collect too rarely falls back on full collections, 1230 // which works, even if not optimum in terms of concurrent work. 1231 // As a work around for too eagerly collecting, use the flag 1232 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1233 // giving the user an easily understandable way of controlling the 1234 // collections. 1235 // We want to start a new collection cycle if any of the following 1236 // conditions hold: 1237 // . our current occupancy exceeds the configured initiating occupancy 1238 // for this generation, or 1239 // . we recently needed to expand this space and have not, since that 1240 // expansion, done a collection of this generation, or 1241 // . the underlying space believes that it may be a good idea to initiate 1242 // a concurrent collection (this may be based on criteria such as the 1243 // following: the space uses linear allocation and linear allocation is 1244 // going to fail, or there is believed to be excessive fragmentation in 1245 // the generation, etc... or ... 1246 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1247 // the case of the old generation; see CR 6543076): 1248 // we may be approaching a point at which allocation requests may fail because 1249 // we will be out of sufficient free space given allocation rate estimates.] 1250 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1251 1252 assert_lock_strong(freelistLock()); 1253 if (occupancy() > initiating_occupancy()) { 1254 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1255 short_name(), occupancy(), initiating_occupancy()); 1256 return true; 1257 } 1258 if (UseCMSInitiatingOccupancyOnly) { 1259 return false; 1260 } 1261 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1262 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1263 return true; 1264 } 1265 return false; 1266 } 1267 1268 void ConcurrentMarkSweepGeneration::collect(bool full, 1269 bool clear_all_soft_refs, 1270 size_t size, 1271 bool tlab) 1272 { 1273 collector()->collect(full, clear_all_soft_refs, size, tlab); 1274 } 1275 1276 void CMSCollector::collect(bool full, 1277 bool clear_all_soft_refs, 1278 size_t size, 1279 bool tlab) 1280 { 1281 // The following "if" branch is present for defensive reasons. 1282 // In the current uses of this interface, it can be replaced with: 1283 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1284 // But I am not placing that assert here to allow future 1285 // generality in invoking this interface. 1286 if (GCLocker::is_active()) { 1287 // A consistency test for GCLocker 1288 assert(GCLocker::needs_gc(), "Should have been set already"); 1289 // Skip this foreground collection, instead 1290 // expanding the heap if necessary. 1291 // Need the free list locks for the call to free() in compute_new_size() 1292 compute_new_size(); 1293 return; 1294 } 1295 acquire_control_and_collect(full, clear_all_soft_refs); 1296 } 1297 1298 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1299 CMSHeap* heap = CMSHeap::heap(); 1300 unsigned int gc_count = heap->total_full_collections(); 1301 if (gc_count == full_gc_count) { 1302 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1303 _full_gc_requested = true; 1304 _full_gc_cause = cause; 1305 CGC_lock->notify(); // nudge CMS thread 1306 } else { 1307 assert(gc_count > full_gc_count, "Error: causal loop"); 1308 } 1309 } 1310 1311 bool CMSCollector::is_external_interruption() { 1312 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1313 return GCCause::is_user_requested_gc(cause) || 1314 GCCause::is_serviceability_requested_gc(cause); 1315 } 1316 1317 void CMSCollector::report_concurrent_mode_interruption() { 1318 if (is_external_interruption()) { 1319 log_debug(gc)("Concurrent mode interrupted"); 1320 } else { 1321 log_debug(gc)("Concurrent mode failure"); 1322 _gc_tracer_cm->report_concurrent_mode_failure(); 1323 } 1324 } 1325 1326 1327 // The foreground and background collectors need to coordinate in order 1328 // to make sure that they do not mutually interfere with CMS collections. 1329 // When a background collection is active, 1330 // the foreground collector may need to take over (preempt) and 1331 // synchronously complete an ongoing collection. Depending on the 1332 // frequency of the background collections and the heap usage 1333 // of the application, this preemption can be seldom or frequent. 1334 // There are only certain 1335 // points in the background collection that the "collection-baton" 1336 // can be passed to the foreground collector. 1337 // 1338 // The foreground collector will wait for the baton before 1339 // starting any part of the collection. The foreground collector 1340 // will only wait at one location. 1341 // 1342 // The background collector will yield the baton before starting a new 1343 // phase of the collection (e.g., before initial marking, marking from roots, 1344 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1345 // of the loop which switches the phases. The background collector does some 1346 // of the phases (initial mark, final re-mark) with the world stopped. 1347 // Because of locking involved in stopping the world, 1348 // the foreground collector should not block waiting for the background 1349 // collector when it is doing a stop-the-world phase. The background 1350 // collector will yield the baton at an additional point just before 1351 // it enters a stop-the-world phase. Once the world is stopped, the 1352 // background collector checks the phase of the collection. If the 1353 // phase has not changed, it proceeds with the collection. If the 1354 // phase has changed, it skips that phase of the collection. See 1355 // the comments on the use of the Heap_lock in collect_in_background(). 1356 // 1357 // Variable used in baton passing. 1358 // _foregroundGCIsActive - Set to true by the foreground collector when 1359 // it wants the baton. The foreground clears it when it has finished 1360 // the collection. 1361 // _foregroundGCShouldWait - Set to true by the background collector 1362 // when it is running. The foreground collector waits while 1363 // _foregroundGCShouldWait is true. 1364 // CGC_lock - monitor used to protect access to the above variables 1365 // and to notify the foreground and background collectors. 1366 // _collectorState - current state of the CMS collection. 1367 // 1368 // The foreground collector 1369 // acquires the CGC_lock 1370 // sets _foregroundGCIsActive 1371 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1372 // various locks acquired in preparation for the collection 1373 // are released so as not to block the background collector 1374 // that is in the midst of a collection 1375 // proceeds with the collection 1376 // clears _foregroundGCIsActive 1377 // returns 1378 // 1379 // The background collector in a loop iterating on the phases of the 1380 // collection 1381 // acquires the CGC_lock 1382 // sets _foregroundGCShouldWait 1383 // if _foregroundGCIsActive is set 1384 // clears _foregroundGCShouldWait, notifies _CGC_lock 1385 // waits on _CGC_lock for _foregroundGCIsActive to become false 1386 // and exits the loop. 1387 // otherwise 1388 // proceed with that phase of the collection 1389 // if the phase is a stop-the-world phase, 1390 // yield the baton once more just before enqueueing 1391 // the stop-world CMS operation (executed by the VM thread). 1392 // returns after all phases of the collection are done 1393 // 1394 1395 void CMSCollector::acquire_control_and_collect(bool full, 1396 bool clear_all_soft_refs) { 1397 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1398 assert(!Thread::current()->is_ConcurrentGC_thread(), 1399 "shouldn't try to acquire control from self!"); 1400 1401 // Start the protocol for acquiring control of the 1402 // collection from the background collector (aka CMS thread). 1403 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1404 "VM thread should have CMS token"); 1405 // Remember the possibly interrupted state of an ongoing 1406 // concurrent collection 1407 CollectorState first_state = _collectorState; 1408 1409 // Signal to a possibly ongoing concurrent collection that 1410 // we want to do a foreground collection. 1411 _foregroundGCIsActive = true; 1412 1413 // release locks and wait for a notify from the background collector 1414 // releasing the locks in only necessary for phases which 1415 // do yields to improve the granularity of the collection. 1416 assert_lock_strong(bitMapLock()); 1417 // We need to lock the Free list lock for the space that we are 1418 // currently collecting. 1419 assert(haveFreelistLocks(), "Must be holding free list locks"); 1420 bitMapLock()->unlock(); 1421 releaseFreelistLocks(); 1422 { 1423 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1424 if (_foregroundGCShouldWait) { 1425 // We are going to be waiting for action for the CMS thread; 1426 // it had better not be gone (for instance at shutdown)! 1427 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1428 "CMS thread must be running"); 1429 // Wait here until the background collector gives us the go-ahead 1430 ConcurrentMarkSweepThread::clear_CMS_flag( 1431 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1432 // Get a possibly blocked CMS thread going: 1433 // Note that we set _foregroundGCIsActive true above, 1434 // without protection of the CGC_lock. 1435 CGC_lock->notify(); 1436 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1437 "Possible deadlock"); 1438 while (_foregroundGCShouldWait) { 1439 // wait for notification 1440 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1441 // Possibility of delay/starvation here, since CMS token does 1442 // not know to give priority to VM thread? Actually, i think 1443 // there wouldn't be any delay/starvation, but the proof of 1444 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1445 } 1446 ConcurrentMarkSweepThread::set_CMS_flag( 1447 ConcurrentMarkSweepThread::CMS_vm_has_token); 1448 } 1449 } 1450 // The CMS_token is already held. Get back the other locks. 1451 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1452 "VM thread should have CMS token"); 1453 getFreelistLocks(); 1454 bitMapLock()->lock_without_safepoint_check(); 1455 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1456 p2i(Thread::current()), first_state); 1457 log_debug(gc, state)(" gets control with state %d", _collectorState); 1458 1459 // Inform cms gen if this was due to partial collection failing. 1460 // The CMS gen may use this fact to determine its expansion policy. 1461 CMSHeap* heap = CMSHeap::heap(); 1462 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1463 assert(!_cmsGen->incremental_collection_failed(), 1464 "Should have been noticed, reacted to and cleared"); 1465 _cmsGen->set_incremental_collection_failed(); 1466 } 1467 1468 if (first_state > Idling) { 1469 report_concurrent_mode_interruption(); 1470 } 1471 1472 set_did_compact(true); 1473 1474 // If the collection is being acquired from the background 1475 // collector, there may be references on the discovered 1476 // references lists. Abandon those references, since some 1477 // of them may have become unreachable after concurrent 1478 // discovery; the STW compacting collector will redo discovery 1479 // more precisely, without being subject to floating garbage. 1480 // Leaving otherwise unreachable references in the discovered 1481 // lists would require special handling. 1482 ref_processor()->disable_discovery(); 1483 ref_processor()->abandon_partial_discovery(); 1484 ref_processor()->verify_no_references_recorded(); 1485 1486 if (first_state > Idling) { 1487 save_heap_summary(); 1488 } 1489 1490 do_compaction_work(clear_all_soft_refs); 1491 1492 // Has the GC time limit been exceeded? 1493 size_t max_eden_size = _young_gen->max_eden_size(); 1494 GCCause::Cause gc_cause = heap->gc_cause(); 1495 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1496 _young_gen->eden()->used(), 1497 _cmsGen->max_capacity(), 1498 max_eden_size, 1499 full, 1500 gc_cause, 1501 heap->collector_policy()); 1502 1503 // Reset the expansion cause, now that we just completed 1504 // a collection cycle. 1505 clear_expansion_cause(); 1506 _foregroundGCIsActive = false; 1507 return; 1508 } 1509 1510 // Resize the tenured generation 1511 // after obtaining the free list locks for the 1512 // two generations. 1513 void CMSCollector::compute_new_size() { 1514 assert_locked_or_safepoint(Heap_lock); 1515 FreelistLocker z(this); 1516 MetaspaceGC::compute_new_size(); 1517 _cmsGen->compute_new_size_free_list(); 1518 } 1519 1520 // A work method used by the foreground collector to do 1521 // a mark-sweep-compact. 1522 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1523 CMSHeap* heap = CMSHeap::heap(); 1524 1525 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1526 gc_timer->register_gc_start(); 1527 1528 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1529 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1530 1531 heap->pre_full_gc_dump(gc_timer); 1532 1533 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1534 1535 // Temporarily widen the span of the weak reference processing to 1536 // the entire heap. 1537 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1538 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1539 // Temporarily, clear the "is_alive_non_header" field of the 1540 // reference processor. 1541 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1542 // Temporarily make reference _processing_ single threaded (non-MT). 1543 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1544 // Temporarily make refs discovery atomic 1545 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1546 // Temporarily make reference _discovery_ single threaded (non-MT) 1547 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1548 1549 ref_processor()->set_enqueuing_is_done(false); 1550 ref_processor()->enable_discovery(); 1551 ref_processor()->setup_policy(clear_all_soft_refs); 1552 // If an asynchronous collection finishes, the _modUnionTable is 1553 // all clear. If we are assuming the collection from an asynchronous 1554 // collection, clear the _modUnionTable. 1555 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1556 "_modUnionTable should be clear if the baton was not passed"); 1557 _modUnionTable.clear_all(); 1558 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1559 "mod union for klasses should be clear if the baton was passed"); 1560 _ct->cld_rem_set()->clear_mod_union(); 1561 1562 1563 // We must adjust the allocation statistics being maintained 1564 // in the free list space. We do so by reading and clearing 1565 // the sweep timer and updating the block flux rate estimates below. 1566 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1567 if (_inter_sweep_timer.is_active()) { 1568 _inter_sweep_timer.stop(); 1569 // Note that we do not use this sample to update the _inter_sweep_estimate. 1570 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1571 _inter_sweep_estimate.padded_average(), 1572 _intra_sweep_estimate.padded_average()); 1573 } 1574 1575 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1576 #ifdef ASSERT 1577 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1578 size_t free_size = cms_space->free(); 1579 assert(free_size == 1580 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1581 * HeapWordSize, 1582 "All the free space should be compacted into one chunk at top"); 1583 assert(cms_space->dictionary()->total_chunk_size( 1584 debug_only(cms_space->freelistLock())) == 0 || 1585 cms_space->totalSizeInIndexedFreeLists() == 0, 1586 "All the free space should be in a single chunk"); 1587 size_t num = cms_space->totalCount(); 1588 assert((free_size == 0 && num == 0) || 1589 (free_size > 0 && (num == 1 || num == 2)), 1590 "There should be at most 2 free chunks after compaction"); 1591 #endif // ASSERT 1592 _collectorState = Resetting; 1593 assert(_restart_addr == NULL, 1594 "Should have been NULL'd before baton was passed"); 1595 reset_stw(); 1596 _cmsGen->reset_after_compaction(); 1597 _concurrent_cycles_since_last_unload = 0; 1598 1599 // Clear any data recorded in the PLAB chunk arrays. 1600 if (_survivor_plab_array != NULL) { 1601 reset_survivor_plab_arrays(); 1602 } 1603 1604 // Adjust the per-size allocation stats for the next epoch. 1605 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1606 // Restart the "inter sweep timer" for the next epoch. 1607 _inter_sweep_timer.reset(); 1608 _inter_sweep_timer.start(); 1609 1610 // No longer a need to do a concurrent collection for Metaspace. 1611 MetaspaceGC::set_should_concurrent_collect(false); 1612 1613 heap->post_full_gc_dump(gc_timer); 1614 1615 gc_timer->register_gc_end(); 1616 1617 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1618 1619 // For a mark-sweep-compact, compute_new_size() will be called 1620 // in the heap's do_collection() method. 1621 } 1622 1623 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1624 Log(gc, heap) log; 1625 if (!log.is_trace()) { 1626 return; 1627 } 1628 1629 ContiguousSpace* eden_space = _young_gen->eden(); 1630 ContiguousSpace* from_space = _young_gen->from(); 1631 ContiguousSpace* to_space = _young_gen->to(); 1632 // Eden 1633 if (_eden_chunk_array != NULL) { 1634 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1635 p2i(eden_space->bottom()), p2i(eden_space->top()), 1636 p2i(eden_space->end()), eden_space->capacity()); 1637 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1638 _eden_chunk_index, _eden_chunk_capacity); 1639 for (size_t i = 0; i < _eden_chunk_index; i++) { 1640 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1641 } 1642 } 1643 // Survivor 1644 if (_survivor_chunk_array != NULL) { 1645 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1646 p2i(from_space->bottom()), p2i(from_space->top()), 1647 p2i(from_space->end()), from_space->capacity()); 1648 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1649 _survivor_chunk_index, _survivor_chunk_capacity); 1650 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1651 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1652 } 1653 } 1654 } 1655 1656 void CMSCollector::getFreelistLocks() const { 1657 // Get locks for all free lists in all generations that this 1658 // collector is responsible for 1659 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1660 } 1661 1662 void CMSCollector::releaseFreelistLocks() const { 1663 // Release locks for all free lists in all generations that this 1664 // collector is responsible for 1665 _cmsGen->freelistLock()->unlock(); 1666 } 1667 1668 bool CMSCollector::haveFreelistLocks() const { 1669 // Check locks for all free lists in all generations that this 1670 // collector is responsible for 1671 assert_lock_strong(_cmsGen->freelistLock()); 1672 PRODUCT_ONLY(ShouldNotReachHere()); 1673 return true; 1674 } 1675 1676 // A utility class that is used by the CMS collector to 1677 // temporarily "release" the foreground collector from its 1678 // usual obligation to wait for the background collector to 1679 // complete an ongoing phase before proceeding. 1680 class ReleaseForegroundGC: public StackObj { 1681 private: 1682 CMSCollector* _c; 1683 public: 1684 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1685 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1686 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1687 // allow a potentially blocked foreground collector to proceed 1688 _c->_foregroundGCShouldWait = false; 1689 if (_c->_foregroundGCIsActive) { 1690 CGC_lock->notify(); 1691 } 1692 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1693 "Possible deadlock"); 1694 } 1695 1696 ~ReleaseForegroundGC() { 1697 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1698 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1699 _c->_foregroundGCShouldWait = true; 1700 } 1701 }; 1702 1703 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1704 assert(Thread::current()->is_ConcurrentGC_thread(), 1705 "A CMS asynchronous collection is only allowed on a CMS thread."); 1706 1707 CMSHeap* heap = CMSHeap::heap(); 1708 { 1709 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1710 MutexLockerEx hl(Heap_lock, safepoint_check); 1711 FreelistLocker fll(this); 1712 MutexLockerEx x(CGC_lock, safepoint_check); 1713 if (_foregroundGCIsActive) { 1714 // The foreground collector is. Skip this 1715 // background collection. 1716 assert(!_foregroundGCShouldWait, "Should be clear"); 1717 return; 1718 } else { 1719 assert(_collectorState == Idling, "Should be idling before start."); 1720 _collectorState = InitialMarking; 1721 register_gc_start(cause); 1722 // Reset the expansion cause, now that we are about to begin 1723 // a new cycle. 1724 clear_expansion_cause(); 1725 1726 // Clear the MetaspaceGC flag since a concurrent collection 1727 // is starting but also clear it after the collection. 1728 MetaspaceGC::set_should_concurrent_collect(false); 1729 } 1730 // Decide if we want to enable class unloading as part of the 1731 // ensuing concurrent GC cycle. 1732 update_should_unload_classes(); 1733 _full_gc_requested = false; // acks all outstanding full gc requests 1734 _full_gc_cause = GCCause::_no_gc; 1735 // Signal that we are about to start a collection 1736 heap->increment_total_full_collections(); // ... starting a collection cycle 1737 _collection_count_start = heap->total_full_collections(); 1738 } 1739 1740 size_t prev_used = _cmsGen->used(); 1741 1742 // The change of the collection state is normally done at this level; 1743 // the exceptions are phases that are executed while the world is 1744 // stopped. For those phases the change of state is done while the 1745 // world is stopped. For baton passing purposes this allows the 1746 // background collector to finish the phase and change state atomically. 1747 // The foreground collector cannot wait on a phase that is done 1748 // while the world is stopped because the foreground collector already 1749 // has the world stopped and would deadlock. 1750 while (_collectorState != Idling) { 1751 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1752 p2i(Thread::current()), _collectorState); 1753 // The foreground collector 1754 // holds the Heap_lock throughout its collection. 1755 // holds the CMS token (but not the lock) 1756 // except while it is waiting for the background collector to yield. 1757 // 1758 // The foreground collector should be blocked (not for long) 1759 // if the background collector is about to start a phase 1760 // executed with world stopped. If the background 1761 // collector has already started such a phase, the 1762 // foreground collector is blocked waiting for the 1763 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1764 // are executed in the VM thread. 1765 // 1766 // The locking order is 1767 // PendingListLock (PLL) -- if applicable (FinalMarking) 1768 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1769 // CMS token (claimed in 1770 // stop_world_and_do() --> 1771 // safepoint_synchronize() --> 1772 // CMSThread::synchronize()) 1773 1774 { 1775 // Check if the FG collector wants us to yield. 1776 CMSTokenSync x(true); // is cms thread 1777 if (waitForForegroundGC()) { 1778 // We yielded to a foreground GC, nothing more to be 1779 // done this round. 1780 assert(_foregroundGCShouldWait == false, "We set it to false in " 1781 "waitForForegroundGC()"); 1782 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1783 p2i(Thread::current()), _collectorState); 1784 return; 1785 } else { 1786 // The background collector can run but check to see if the 1787 // foreground collector has done a collection while the 1788 // background collector was waiting to get the CGC_lock 1789 // above. If yes, break so that _foregroundGCShouldWait 1790 // is cleared before returning. 1791 if (_collectorState == Idling) { 1792 break; 1793 } 1794 } 1795 } 1796 1797 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1798 "should be waiting"); 1799 1800 switch (_collectorState) { 1801 case InitialMarking: 1802 { 1803 ReleaseForegroundGC x(this); 1804 stats().record_cms_begin(); 1805 VM_CMS_Initial_Mark initial_mark_op(this); 1806 VMThread::execute(&initial_mark_op); 1807 } 1808 // The collector state may be any legal state at this point 1809 // since the background collector may have yielded to the 1810 // foreground collector. 1811 break; 1812 case Marking: 1813 // initial marking in checkpointRootsInitialWork has been completed 1814 if (markFromRoots()) { // we were successful 1815 assert(_collectorState == Precleaning, "Collector state should " 1816 "have changed"); 1817 } else { 1818 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1819 } 1820 break; 1821 case Precleaning: 1822 // marking from roots in markFromRoots has been completed 1823 preclean(); 1824 assert(_collectorState == AbortablePreclean || 1825 _collectorState == FinalMarking, 1826 "Collector state should have changed"); 1827 break; 1828 case AbortablePreclean: 1829 abortable_preclean(); 1830 assert(_collectorState == FinalMarking, "Collector state should " 1831 "have changed"); 1832 break; 1833 case FinalMarking: 1834 { 1835 ReleaseForegroundGC x(this); 1836 1837 VM_CMS_Final_Remark final_remark_op(this); 1838 VMThread::execute(&final_remark_op); 1839 } 1840 assert(_foregroundGCShouldWait, "block post-condition"); 1841 break; 1842 case Sweeping: 1843 // final marking in checkpointRootsFinal has been completed 1844 sweep(); 1845 assert(_collectorState == Resizing, "Collector state change " 1846 "to Resizing must be done under the free_list_lock"); 1847 1848 case Resizing: { 1849 // Sweeping has been completed... 1850 // At this point the background collection has completed. 1851 // Don't move the call to compute_new_size() down 1852 // into code that might be executed if the background 1853 // collection was preempted. 1854 { 1855 ReleaseForegroundGC x(this); // unblock FG collection 1856 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1857 CMSTokenSync z(true); // not strictly needed. 1858 if (_collectorState == Resizing) { 1859 compute_new_size(); 1860 save_heap_summary(); 1861 _collectorState = Resetting; 1862 } else { 1863 assert(_collectorState == Idling, "The state should only change" 1864 " because the foreground collector has finished the collection"); 1865 } 1866 } 1867 break; 1868 } 1869 case Resetting: 1870 // CMS heap resizing has been completed 1871 reset_concurrent(); 1872 assert(_collectorState == Idling, "Collector state should " 1873 "have changed"); 1874 1875 MetaspaceGC::set_should_concurrent_collect(false); 1876 1877 stats().record_cms_end(); 1878 // Don't move the concurrent_phases_end() and compute_new_size() 1879 // calls to here because a preempted background collection 1880 // has it's state set to "Resetting". 1881 break; 1882 case Idling: 1883 default: 1884 ShouldNotReachHere(); 1885 break; 1886 } 1887 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1888 p2i(Thread::current()), _collectorState); 1889 assert(_foregroundGCShouldWait, "block post-condition"); 1890 } 1891 1892 // Should this be in gc_epilogue? 1893 collector_policy()->counters()->update_counters(); 1894 1895 { 1896 // Clear _foregroundGCShouldWait and, in the event that the 1897 // foreground collector is waiting, notify it, before 1898 // returning. 1899 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1900 _foregroundGCShouldWait = false; 1901 if (_foregroundGCIsActive) { 1902 CGC_lock->notify(); 1903 } 1904 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1905 "Possible deadlock"); 1906 } 1907 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1908 p2i(Thread::current()), _collectorState); 1909 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1910 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1911 } 1912 1913 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1914 _cms_start_registered = true; 1915 _gc_timer_cm->register_gc_start(); 1916 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1917 } 1918 1919 void CMSCollector::register_gc_end() { 1920 if (_cms_start_registered) { 1921 report_heap_summary(GCWhen::AfterGC); 1922 1923 _gc_timer_cm->register_gc_end(); 1924 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1925 _cms_start_registered = false; 1926 } 1927 } 1928 1929 void CMSCollector::save_heap_summary() { 1930 CMSHeap* heap = CMSHeap::heap(); 1931 _last_heap_summary = heap->create_heap_summary(); 1932 _last_metaspace_summary = heap->create_metaspace_summary(); 1933 } 1934 1935 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1936 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1937 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1938 } 1939 1940 bool CMSCollector::waitForForegroundGC() { 1941 bool res = false; 1942 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1943 "CMS thread should have CMS token"); 1944 // Block the foreground collector until the 1945 // background collectors decides whether to 1946 // yield. 1947 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1948 _foregroundGCShouldWait = true; 1949 if (_foregroundGCIsActive) { 1950 // The background collector yields to the 1951 // foreground collector and returns a value 1952 // indicating that it has yielded. The foreground 1953 // collector can proceed. 1954 res = true; 1955 _foregroundGCShouldWait = false; 1956 ConcurrentMarkSweepThread::clear_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_has_token); 1958 ConcurrentMarkSweepThread::set_CMS_flag( 1959 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1960 // Get a possibly blocked foreground thread going 1961 CGC_lock->notify(); 1962 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1963 p2i(Thread::current()), _collectorState); 1964 while (_foregroundGCIsActive) { 1965 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1966 } 1967 ConcurrentMarkSweepThread::set_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_has_token); 1969 ConcurrentMarkSweepThread::clear_CMS_flag( 1970 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1971 } 1972 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1973 p2i(Thread::current()), _collectorState); 1974 return res; 1975 } 1976 1977 // Because of the need to lock the free lists and other structures in 1978 // the collector, common to all the generations that the collector is 1979 // collecting, we need the gc_prologues of individual CMS generations 1980 // delegate to their collector. It may have been simpler had the 1981 // current infrastructure allowed one to call a prologue on a 1982 // collector. In the absence of that we have the generation's 1983 // prologue delegate to the collector, which delegates back 1984 // some "local" work to a worker method in the individual generations 1985 // that it's responsible for collecting, while itself doing any 1986 // work common to all generations it's responsible for. A similar 1987 // comment applies to the gc_epilogue()'s. 1988 // The role of the variable _between_prologue_and_epilogue is to 1989 // enforce the invocation protocol. 1990 void CMSCollector::gc_prologue(bool full) { 1991 // Call gc_prologue_work() for the CMSGen 1992 // we are responsible for. 1993 1994 // The following locking discipline assumes that we are only called 1995 // when the world is stopped. 1996 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1997 1998 // The CMSCollector prologue must call the gc_prologues for the 1999 // "generations" that it's responsible 2000 // for. 2001 2002 assert( Thread::current()->is_VM_thread() 2003 || ( CMSScavengeBeforeRemark 2004 && Thread::current()->is_ConcurrentGC_thread()), 2005 "Incorrect thread type for prologue execution"); 2006 2007 if (_between_prologue_and_epilogue) { 2008 // We have already been invoked; this is a gc_prologue delegation 2009 // from yet another CMS generation that we are responsible for, just 2010 // ignore it since all relevant work has already been done. 2011 return; 2012 } 2013 2014 // set a bit saying prologue has been called; cleared in epilogue 2015 _between_prologue_and_epilogue = true; 2016 // Claim locks for common data structures, then call gc_prologue_work() 2017 // for each CMSGen. 2018 2019 getFreelistLocks(); // gets free list locks on constituent spaces 2020 bitMapLock()->lock_without_safepoint_check(); 2021 2022 // Should call gc_prologue_work() for all cms gens we are responsible for 2023 bool duringMarking = _collectorState >= Marking 2024 && _collectorState < Sweeping; 2025 2026 // The young collections clear the modified oops state, which tells if 2027 // there are any modified oops in the class. The remark phase also needs 2028 // that information. Tell the young collection to save the union of all 2029 // modified klasses. 2030 if (duringMarking) { 2031 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2032 } 2033 2034 bool registerClosure = duringMarking; 2035 2036 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2037 2038 if (!full) { 2039 stats().record_gc0_begin(); 2040 } 2041 } 2042 2043 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2044 2045 _capacity_at_prologue = capacity(); 2046 _used_at_prologue = used(); 2047 2048 // We enable promotion tracking so that card-scanning can recognize 2049 // which objects have been promoted during this GC and skip them. 2050 for (uint i = 0; i < ParallelGCThreads; i++) { 2051 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2052 } 2053 2054 // Delegate to CMScollector which knows how to coordinate between 2055 // this and any other CMS generations that it is responsible for 2056 // collecting. 2057 collector()->gc_prologue(full); 2058 } 2059 2060 // This is a "private" interface for use by this generation's CMSCollector. 2061 // Not to be called directly by any other entity (for instance, 2062 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2063 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2064 bool registerClosure, ModUnionClosure* modUnionClosure) { 2065 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2066 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2067 "Should be NULL"); 2068 if (registerClosure) { 2069 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2070 } 2071 cmsSpace()->gc_prologue(); 2072 // Clear stat counters 2073 NOT_PRODUCT( 2074 assert(_numObjectsPromoted == 0, "check"); 2075 assert(_numWordsPromoted == 0, "check"); 2076 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2077 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2078 _numObjectsAllocated = 0; 2079 _numWordsAllocated = 0; 2080 ) 2081 } 2082 2083 void CMSCollector::gc_epilogue(bool full) { 2084 // The following locking discipline assumes that we are only called 2085 // when the world is stopped. 2086 assert(SafepointSynchronize::is_at_safepoint(), 2087 "world is stopped assumption"); 2088 2089 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2090 // if linear allocation blocks need to be appropriately marked to allow the 2091 // the blocks to be parsable. We also check here whether we need to nudge the 2092 // CMS collector thread to start a new cycle (if it's not already active). 2093 assert( Thread::current()->is_VM_thread() 2094 || ( CMSScavengeBeforeRemark 2095 && Thread::current()->is_ConcurrentGC_thread()), 2096 "Incorrect thread type for epilogue execution"); 2097 2098 if (!_between_prologue_and_epilogue) { 2099 // We have already been invoked; this is a gc_epilogue delegation 2100 // from yet another CMS generation that we are responsible for, just 2101 // ignore it since all relevant work has already been done. 2102 return; 2103 } 2104 assert(haveFreelistLocks(), "must have freelist locks"); 2105 assert_lock_strong(bitMapLock()); 2106 2107 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2108 2109 _cmsGen->gc_epilogue_work(full); 2110 2111 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2112 // in case sampling was not already enabled, enable it 2113 _start_sampling = true; 2114 } 2115 // reset _eden_chunk_array so sampling starts afresh 2116 _eden_chunk_index = 0; 2117 2118 size_t cms_used = _cmsGen->cmsSpace()->used(); 2119 2120 // update performance counters - this uses a special version of 2121 // update_counters() that allows the utilization to be passed as a 2122 // parameter, avoiding multiple calls to used(). 2123 // 2124 _cmsGen->update_counters(cms_used); 2125 2126 bitMapLock()->unlock(); 2127 releaseFreelistLocks(); 2128 2129 if (!CleanChunkPoolAsync) { 2130 Chunk::clean_chunk_pool(); 2131 } 2132 2133 set_did_compact(false); 2134 _between_prologue_and_epilogue = false; // ready for next cycle 2135 } 2136 2137 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2138 collector()->gc_epilogue(full); 2139 2140 // When using ParNew, promotion tracking should have already been 2141 // disabled. However, the prologue (which enables promotion 2142 // tracking) and epilogue are called irrespective of the type of 2143 // GC. So they will also be called before and after Full GCs, during 2144 // which promotion tracking will not be explicitly disabled. So, 2145 // it's safer to also disable it here too (to be symmetric with 2146 // enabling it in the prologue). 2147 for (uint i = 0; i < ParallelGCThreads; i++) { 2148 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2149 } 2150 } 2151 2152 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2153 assert(!incremental_collection_failed(), "Should have been cleared"); 2154 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2155 cmsSpace()->gc_epilogue(); 2156 // Print stat counters 2157 NOT_PRODUCT( 2158 assert(_numObjectsAllocated == 0, "check"); 2159 assert(_numWordsAllocated == 0, "check"); 2160 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2161 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2162 _numObjectsPromoted = 0; 2163 _numWordsPromoted = 0; 2164 ) 2165 2166 // Call down the chain in contiguous_available needs the freelistLock 2167 // so print this out before releasing the freeListLock. 2168 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2169 } 2170 2171 #ifndef PRODUCT 2172 bool CMSCollector::have_cms_token() { 2173 Thread* thr = Thread::current(); 2174 if (thr->is_VM_thread()) { 2175 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2176 } else if (thr->is_ConcurrentGC_thread()) { 2177 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2178 } else if (thr->is_GC_task_thread()) { 2179 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2180 ParGCRareEvent_lock->owned_by_self(); 2181 } 2182 return false; 2183 } 2184 2185 // Check reachability of the given heap address in CMS generation, 2186 // treating all other generations as roots. 2187 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2188 // We could "guarantee" below, rather than assert, but I'll 2189 // leave these as "asserts" so that an adventurous debugger 2190 // could try this in the product build provided some subset of 2191 // the conditions were met, provided they were interested in the 2192 // results and knew that the computation below wouldn't interfere 2193 // with other concurrent computations mutating the structures 2194 // being read or written. 2195 assert(SafepointSynchronize::is_at_safepoint(), 2196 "Else mutations in object graph will make answer suspect"); 2197 assert(have_cms_token(), "Should hold cms token"); 2198 assert(haveFreelistLocks(), "must hold free list locks"); 2199 assert_lock_strong(bitMapLock()); 2200 2201 // Clear the marking bit map array before starting, but, just 2202 // for kicks, first report if the given address is already marked 2203 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2204 _markBitMap.isMarked(addr) ? "" : " not"); 2205 2206 if (verify_after_remark()) { 2207 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2208 bool result = verification_mark_bm()->isMarked(addr); 2209 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2210 result ? "IS" : "is NOT"); 2211 return result; 2212 } else { 2213 tty->print_cr("Could not compute result"); 2214 return false; 2215 } 2216 } 2217 #endif 2218 2219 void 2220 CMSCollector::print_on_error(outputStream* st) { 2221 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2222 if (collector != NULL) { 2223 CMSBitMap* bitmap = &collector->_markBitMap; 2224 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2225 bitmap->print_on_error(st, " Bits: "); 2226 2227 st->cr(); 2228 2229 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2230 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2231 mut_bitmap->print_on_error(st, " Bits: "); 2232 } 2233 } 2234 2235 //////////////////////////////////////////////////////// 2236 // CMS Verification Support 2237 //////////////////////////////////////////////////////// 2238 // Following the remark phase, the following invariant 2239 // should hold -- each object in the CMS heap which is 2240 // marked in markBitMap() should be marked in the verification_mark_bm(). 2241 2242 class VerifyMarkedClosure: public BitMapClosure { 2243 CMSBitMap* _marks; 2244 bool _failed; 2245 2246 public: 2247 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2248 2249 bool do_bit(size_t offset) { 2250 HeapWord* addr = _marks->offsetToHeapWord(offset); 2251 if (!_marks->isMarked(addr)) { 2252 Log(gc, verify) log; 2253 ResourceMark rm; 2254 LogStream ls(log.error()); 2255 oop(addr)->print_on(&ls); 2256 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2257 _failed = true; 2258 } 2259 return true; 2260 } 2261 2262 bool failed() { return _failed; } 2263 }; 2264 2265 bool CMSCollector::verify_after_remark() { 2266 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2267 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2268 static bool init = false; 2269 2270 assert(SafepointSynchronize::is_at_safepoint(), 2271 "Else mutations in object graph will make answer suspect"); 2272 assert(have_cms_token(), 2273 "Else there may be mutual interference in use of " 2274 " verification data structures"); 2275 assert(_collectorState > Marking && _collectorState <= Sweeping, 2276 "Else marking info checked here may be obsolete"); 2277 assert(haveFreelistLocks(), "must hold free list locks"); 2278 assert_lock_strong(bitMapLock()); 2279 2280 2281 // Allocate marking bit map if not already allocated 2282 if (!init) { // first time 2283 if (!verification_mark_bm()->allocate(_span)) { 2284 return false; 2285 } 2286 init = true; 2287 } 2288 2289 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2290 2291 // Turn off refs discovery -- so we will be tracing through refs. 2292 // This is as intended, because by this time 2293 // GC must already have cleared any refs that need to be cleared, 2294 // and traced those that need to be marked; moreover, 2295 // the marking done here is not going to interfere in any 2296 // way with the marking information used by GC. 2297 NoRefDiscovery no_discovery(ref_processor()); 2298 2299 #if COMPILER2_OR_JVMCI 2300 DerivedPointerTableDeactivate dpt_deact; 2301 #endif 2302 2303 // Clear any marks from a previous round 2304 verification_mark_bm()->clear_all(); 2305 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2306 verify_work_stacks_empty(); 2307 2308 CMSHeap* heap = CMSHeap::heap(); 2309 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2310 // Update the saved marks which may affect the root scans. 2311 heap->save_marks(); 2312 2313 if (CMSRemarkVerifyVariant == 1) { 2314 // In this first variant of verification, we complete 2315 // all marking, then check if the new marks-vector is 2316 // a subset of the CMS marks-vector. 2317 verify_after_remark_work_1(); 2318 } else { 2319 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2320 // In this second variant of verification, we flag an error 2321 // (i.e. an object reachable in the new marks-vector not reachable 2322 // in the CMS marks-vector) immediately, also indicating the 2323 // identify of an object (A) that references the unmarked object (B) -- 2324 // presumably, a mutation to A failed to be picked up by preclean/remark? 2325 verify_after_remark_work_2(); 2326 } 2327 2328 return true; 2329 } 2330 2331 void CMSCollector::verify_after_remark_work_1() { 2332 ResourceMark rm; 2333 HandleMark hm; 2334 CMSHeap* heap = CMSHeap::heap(); 2335 2336 // Get a clear set of claim bits for the roots processing to work with. 2337 ClassLoaderDataGraph::clear_claimed_marks(); 2338 2339 // Mark from roots one level into CMS 2340 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2341 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2342 2343 { 2344 StrongRootsScope srs(1); 2345 2346 heap->cms_process_roots(&srs, 2347 true, // young gen as roots 2348 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2349 should_unload_classes(), 2350 ¬Older, 2351 NULL); 2352 } 2353 2354 // Now mark from the roots 2355 MarkFromRootsClosure markFromRootsClosure(this, _span, 2356 verification_mark_bm(), verification_mark_stack(), 2357 false /* don't yield */, true /* verifying */); 2358 assert(_restart_addr == NULL, "Expected pre-condition"); 2359 verification_mark_bm()->iterate(&markFromRootsClosure); 2360 while (_restart_addr != NULL) { 2361 // Deal with stack overflow: by restarting at the indicated 2362 // address. 2363 HeapWord* ra = _restart_addr; 2364 markFromRootsClosure.reset(ra); 2365 _restart_addr = NULL; 2366 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2367 } 2368 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2369 verify_work_stacks_empty(); 2370 2371 // Marking completed -- now verify that each bit marked in 2372 // verification_mark_bm() is also marked in markBitMap(); flag all 2373 // errors by printing corresponding objects. 2374 VerifyMarkedClosure vcl(markBitMap()); 2375 verification_mark_bm()->iterate(&vcl); 2376 if (vcl.failed()) { 2377 Log(gc, verify) log; 2378 log.error("Failed marking verification after remark"); 2379 ResourceMark rm; 2380 LogStream ls(log.error()); 2381 heap->print_on(&ls); 2382 fatal("CMS: failed marking verification after remark"); 2383 } 2384 } 2385 2386 class VerifyCLDOopsCLDClosure : public CLDClosure { 2387 class VerifyCLDOopsClosure : public OopClosure { 2388 CMSBitMap* _bitmap; 2389 public: 2390 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2391 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2392 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2393 } _oop_closure; 2394 public: 2395 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2396 void do_cld(ClassLoaderData* cld) { 2397 cld->oops_do(&_oop_closure, false, false); 2398 } 2399 }; 2400 2401 void CMSCollector::verify_after_remark_work_2() { 2402 ResourceMark rm; 2403 HandleMark hm; 2404 CMSHeap* heap = CMSHeap::heap(); 2405 2406 // Get a clear set of claim bits for the roots processing to work with. 2407 ClassLoaderDataGraph::clear_claimed_marks(); 2408 2409 // Mark from roots one level into CMS 2410 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2411 markBitMap()); 2412 CLDToOopClosure cld_closure(¬Older, true); 2413 2414 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2415 2416 { 2417 StrongRootsScope srs(1); 2418 2419 heap->cms_process_roots(&srs, 2420 true, // young gen as roots 2421 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2422 should_unload_classes(), 2423 ¬Older, 2424 &cld_closure); 2425 } 2426 2427 // Now mark from the roots 2428 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2429 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2430 assert(_restart_addr == NULL, "Expected pre-condition"); 2431 verification_mark_bm()->iterate(&markFromRootsClosure); 2432 while (_restart_addr != NULL) { 2433 // Deal with stack overflow: by restarting at the indicated 2434 // address. 2435 HeapWord* ra = _restart_addr; 2436 markFromRootsClosure.reset(ra); 2437 _restart_addr = NULL; 2438 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2439 } 2440 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2441 verify_work_stacks_empty(); 2442 2443 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2444 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2445 2446 // Marking completed -- now verify that each bit marked in 2447 // verification_mark_bm() is also marked in markBitMap(); flag all 2448 // errors by printing corresponding objects. 2449 VerifyMarkedClosure vcl(markBitMap()); 2450 verification_mark_bm()->iterate(&vcl); 2451 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2452 } 2453 2454 void ConcurrentMarkSweepGeneration::save_marks() { 2455 // delegate to CMS space 2456 cmsSpace()->save_marks(); 2457 } 2458 2459 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2460 return cmsSpace()->no_allocs_since_save_marks(); 2461 } 2462 2463 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2464 \ 2465 void ConcurrentMarkSweepGeneration:: \ 2466 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2467 cl->set_generation(this); \ 2468 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2469 cl->reset_generation(); \ 2470 save_marks(); \ 2471 } 2472 2473 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2474 2475 void 2476 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2477 if (freelistLock()->owned_by_self()) { 2478 Generation::oop_iterate(cl); 2479 } else { 2480 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2481 Generation::oop_iterate(cl); 2482 } 2483 } 2484 2485 void 2486 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2487 if (freelistLock()->owned_by_self()) { 2488 Generation::object_iterate(cl); 2489 } else { 2490 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2491 Generation::object_iterate(cl); 2492 } 2493 } 2494 2495 void 2496 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2497 if (freelistLock()->owned_by_self()) { 2498 Generation::safe_object_iterate(cl); 2499 } else { 2500 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2501 Generation::safe_object_iterate(cl); 2502 } 2503 } 2504 2505 void 2506 ConcurrentMarkSweepGeneration::post_compact() { 2507 } 2508 2509 void 2510 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2511 // Fix the linear allocation blocks to look like free blocks. 2512 2513 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2514 // are not called when the heap is verified during universe initialization and 2515 // at vm shutdown. 2516 if (freelistLock()->owned_by_self()) { 2517 cmsSpace()->prepare_for_verify(); 2518 } else { 2519 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2520 cmsSpace()->prepare_for_verify(); 2521 } 2522 } 2523 2524 void 2525 ConcurrentMarkSweepGeneration::verify() { 2526 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2527 // are not called when the heap is verified during universe initialization and 2528 // at vm shutdown. 2529 if (freelistLock()->owned_by_self()) { 2530 cmsSpace()->verify(); 2531 } else { 2532 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2533 cmsSpace()->verify(); 2534 } 2535 } 2536 2537 void CMSCollector::verify() { 2538 _cmsGen->verify(); 2539 } 2540 2541 #ifndef PRODUCT 2542 bool CMSCollector::overflow_list_is_empty() const { 2543 assert(_num_par_pushes >= 0, "Inconsistency"); 2544 if (_overflow_list == NULL) { 2545 assert(_num_par_pushes == 0, "Inconsistency"); 2546 } 2547 return _overflow_list == NULL; 2548 } 2549 2550 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2551 // merely consolidate assertion checks that appear to occur together frequently. 2552 void CMSCollector::verify_work_stacks_empty() const { 2553 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2554 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2555 } 2556 2557 void CMSCollector::verify_overflow_empty() const { 2558 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2559 assert(no_preserved_marks(), "No preserved marks"); 2560 } 2561 #endif // PRODUCT 2562 2563 // Decide if we want to enable class unloading as part of the 2564 // ensuing concurrent GC cycle. We will collect and 2565 // unload classes if it's the case that: 2566 // (a) class unloading is enabled at the command line, and 2567 // (b) old gen is getting really full 2568 // NOTE: Provided there is no change in the state of the heap between 2569 // calls to this method, it should have idempotent results. Moreover, 2570 // its results should be monotonically increasing (i.e. going from 0 to 1, 2571 // but not 1 to 0) between successive calls between which the heap was 2572 // not collected. For the implementation below, it must thus rely on 2573 // the property that concurrent_cycles_since_last_unload() 2574 // will not decrease unless a collection cycle happened and that 2575 // _cmsGen->is_too_full() are 2576 // themselves also monotonic in that sense. See check_monotonicity() 2577 // below. 2578 void CMSCollector::update_should_unload_classes() { 2579 _should_unload_classes = false; 2580 if (CMSClassUnloadingEnabled) { 2581 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2582 CMSClassUnloadingMaxInterval) 2583 || _cmsGen->is_too_full(); 2584 } 2585 } 2586 2587 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2588 bool res = should_concurrent_collect(); 2589 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2590 return res; 2591 } 2592 2593 void CMSCollector::setup_cms_unloading_and_verification_state() { 2594 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2595 || VerifyBeforeExit; 2596 const int rso = GenCollectedHeap::SO_AllCodeCache; 2597 2598 // We set the proper root for this CMS cycle here. 2599 if (should_unload_classes()) { // Should unload classes this cycle 2600 remove_root_scanning_option(rso); // Shrink the root set appropriately 2601 set_verifying(should_verify); // Set verification state for this cycle 2602 return; // Nothing else needs to be done at this time 2603 } 2604 2605 // Not unloading classes this cycle 2606 assert(!should_unload_classes(), "Inconsistency!"); 2607 2608 // If we are not unloading classes then add SO_AllCodeCache to root 2609 // scanning options. 2610 add_root_scanning_option(rso); 2611 2612 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2613 set_verifying(true); 2614 } else if (verifying() && !should_verify) { 2615 // We were verifying, but some verification flags got disabled. 2616 set_verifying(false); 2617 // Exclude symbols, strings and code cache elements from root scanning to 2618 // reduce IM and RM pauses. 2619 remove_root_scanning_option(rso); 2620 } 2621 } 2622 2623 2624 #ifndef PRODUCT 2625 HeapWord* CMSCollector::block_start(const void* p) const { 2626 const HeapWord* addr = (HeapWord*)p; 2627 if (_span.contains(p)) { 2628 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2629 return _cmsGen->cmsSpace()->block_start(p); 2630 } 2631 } 2632 return NULL; 2633 } 2634 #endif 2635 2636 HeapWord* 2637 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2638 bool tlab, 2639 bool parallel) { 2640 CMSSynchronousYieldRequest yr; 2641 assert(!tlab, "Can't deal with TLAB allocation"); 2642 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2643 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2644 if (GCExpandToAllocateDelayMillis > 0) { 2645 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2646 } 2647 return have_lock_and_allocate(word_size, tlab); 2648 } 2649 2650 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2651 size_t bytes, 2652 size_t expand_bytes, 2653 CMSExpansionCause::Cause cause) 2654 { 2655 2656 bool success = expand(bytes, expand_bytes); 2657 2658 // remember why we expanded; this information is used 2659 // by shouldConcurrentCollect() when making decisions on whether to start 2660 // a new CMS cycle. 2661 if (success) { 2662 set_expansion_cause(cause); 2663 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2664 } 2665 } 2666 2667 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2668 HeapWord* res = NULL; 2669 MutexLocker x(ParGCRareEvent_lock); 2670 while (true) { 2671 // Expansion by some other thread might make alloc OK now: 2672 res = ps->lab.alloc(word_sz); 2673 if (res != NULL) return res; 2674 // If there's not enough expansion space available, give up. 2675 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2676 return NULL; 2677 } 2678 // Otherwise, we try expansion. 2679 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2680 // Now go around the loop and try alloc again; 2681 // A competing par_promote might beat us to the expansion space, 2682 // so we may go around the loop again if promotion fails again. 2683 if (GCExpandToAllocateDelayMillis > 0) { 2684 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2685 } 2686 } 2687 } 2688 2689 2690 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2691 PromotionInfo* promo) { 2692 MutexLocker x(ParGCRareEvent_lock); 2693 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2694 while (true) { 2695 // Expansion by some other thread might make alloc OK now: 2696 if (promo->ensure_spooling_space()) { 2697 assert(promo->has_spooling_space(), 2698 "Post-condition of successful ensure_spooling_space()"); 2699 return true; 2700 } 2701 // If there's not enough expansion space available, give up. 2702 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2703 return false; 2704 } 2705 // Otherwise, we try expansion. 2706 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2707 // Now go around the loop and try alloc again; 2708 // A competing allocation might beat us to the expansion space, 2709 // so we may go around the loop again if allocation fails again. 2710 if (GCExpandToAllocateDelayMillis > 0) { 2711 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2712 } 2713 } 2714 } 2715 2716 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2717 // Only shrink if a compaction was done so that all the free space 2718 // in the generation is in a contiguous block at the end. 2719 if (did_compact()) { 2720 CardGeneration::shrink(bytes); 2721 } 2722 } 2723 2724 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2725 assert_locked_or_safepoint(Heap_lock); 2726 } 2727 2728 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2729 assert_locked_or_safepoint(Heap_lock); 2730 assert_lock_strong(freelistLock()); 2731 log_trace(gc)("Shrinking of CMS not yet implemented"); 2732 return; 2733 } 2734 2735 2736 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2737 // phases. 2738 class CMSPhaseAccounting: public StackObj { 2739 public: 2740 CMSPhaseAccounting(CMSCollector *collector, 2741 const char *title); 2742 ~CMSPhaseAccounting(); 2743 2744 private: 2745 CMSCollector *_collector; 2746 const char *_title; 2747 GCTraceConcTime(Info, gc) _trace_time; 2748 2749 public: 2750 // Not MT-safe; so do not pass around these StackObj's 2751 // where they may be accessed by other threads. 2752 double wallclock_millis() { 2753 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2754 } 2755 }; 2756 2757 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2758 const char *title) : 2759 _collector(collector), _title(title), _trace_time(title) { 2760 2761 _collector->resetYields(); 2762 _collector->resetTimer(); 2763 _collector->startTimer(); 2764 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2765 } 2766 2767 CMSPhaseAccounting::~CMSPhaseAccounting() { 2768 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2769 _collector->stopTimer(); 2770 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2771 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2772 } 2773 2774 // CMS work 2775 2776 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2777 class CMSParMarkTask : public AbstractGangTask { 2778 protected: 2779 CMSCollector* _collector; 2780 uint _n_workers; 2781 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2782 AbstractGangTask(name), 2783 _collector(collector), 2784 _n_workers(n_workers) {} 2785 // Work method in support of parallel rescan ... of young gen spaces 2786 void do_young_space_rescan(OopsInGenClosure* cl, 2787 ContiguousSpace* space, 2788 HeapWord** chunk_array, size_t chunk_top); 2789 void work_on_young_gen_roots(OopsInGenClosure* cl); 2790 }; 2791 2792 // Parallel initial mark task 2793 class CMSParInitialMarkTask: public CMSParMarkTask { 2794 StrongRootsScope* _strong_roots_scope; 2795 public: 2796 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2797 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2798 _strong_roots_scope(strong_roots_scope) {} 2799 void work(uint worker_id); 2800 }; 2801 2802 // Checkpoint the roots into this generation from outside 2803 // this generation. [Note this initial checkpoint need only 2804 // be approximate -- we'll do a catch up phase subsequently.] 2805 void CMSCollector::checkpointRootsInitial() { 2806 assert(_collectorState == InitialMarking, "Wrong collector state"); 2807 check_correct_thread_executing(); 2808 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2809 2810 save_heap_summary(); 2811 report_heap_summary(GCWhen::BeforeGC); 2812 2813 ReferenceProcessor* rp = ref_processor(); 2814 assert(_restart_addr == NULL, "Control point invariant"); 2815 { 2816 // acquire locks for subsequent manipulations 2817 MutexLockerEx x(bitMapLock(), 2818 Mutex::_no_safepoint_check_flag); 2819 checkpointRootsInitialWork(); 2820 // enable ("weak") refs discovery 2821 rp->enable_discovery(); 2822 _collectorState = Marking; 2823 } 2824 } 2825 2826 void CMSCollector::checkpointRootsInitialWork() { 2827 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2828 assert(_collectorState == InitialMarking, "just checking"); 2829 2830 // Already have locks. 2831 assert_lock_strong(bitMapLock()); 2832 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2833 2834 // Setup the verification and class unloading state for this 2835 // CMS collection cycle. 2836 setup_cms_unloading_and_verification_state(); 2837 2838 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2839 2840 // Reset all the PLAB chunk arrays if necessary. 2841 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2842 reset_survivor_plab_arrays(); 2843 } 2844 2845 ResourceMark rm; 2846 HandleMark hm; 2847 2848 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2849 CMSHeap* heap = CMSHeap::heap(); 2850 2851 verify_work_stacks_empty(); 2852 verify_overflow_empty(); 2853 2854 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2855 // Update the saved marks which may affect the root scans. 2856 heap->save_marks(); 2857 2858 // weak reference processing has not started yet. 2859 ref_processor()->set_enqueuing_is_done(false); 2860 2861 // Need to remember all newly created CLDs, 2862 // so that we can guarantee that the remark finds them. 2863 ClassLoaderDataGraph::remember_new_clds(true); 2864 2865 // Whenever a CLD is found, it will be claimed before proceeding to mark 2866 // the klasses. The claimed marks need to be cleared before marking starts. 2867 ClassLoaderDataGraph::clear_claimed_marks(); 2868 2869 print_eden_and_survivor_chunk_arrays(); 2870 2871 { 2872 #if COMPILER2_OR_JVMCI 2873 DerivedPointerTableDeactivate dpt_deact; 2874 #endif 2875 if (CMSParallelInitialMarkEnabled) { 2876 // The parallel version. 2877 WorkGang* workers = heap->workers(); 2878 assert(workers != NULL, "Need parallel worker threads."); 2879 uint n_workers = workers->active_workers(); 2880 2881 StrongRootsScope srs(n_workers); 2882 2883 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2884 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2885 // If the total workers is greater than 1, then multiple workers 2886 // may be used at some time and the initialization has been set 2887 // such that the single threaded path cannot be used. 2888 if (workers->total_workers() > 1) { 2889 workers->run_task(&tsk); 2890 } else { 2891 tsk.work(0); 2892 } 2893 } else { 2894 // The serial version. 2895 CLDToOopClosure cld_closure(¬Older, true); 2896 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2897 2898 StrongRootsScope srs(1); 2899 2900 heap->cms_process_roots(&srs, 2901 true, // young gen as roots 2902 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2903 should_unload_classes(), 2904 ¬Older, 2905 &cld_closure); 2906 } 2907 } 2908 2909 // Clear mod-union table; it will be dirtied in the prologue of 2910 // CMS generation per each young generation collection. 2911 2912 assert(_modUnionTable.isAllClear(), 2913 "Was cleared in most recent final checkpoint phase" 2914 " or no bits are set in the gc_prologue before the start of the next " 2915 "subsequent marking phase."); 2916 2917 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2918 2919 // Save the end of the used_region of the constituent generations 2920 // to be used to limit the extent of sweep in each generation. 2921 save_sweep_limits(); 2922 verify_overflow_empty(); 2923 } 2924 2925 bool CMSCollector::markFromRoots() { 2926 // we might be tempted to assert that: 2927 // assert(!SafepointSynchronize::is_at_safepoint(), 2928 // "inconsistent argument?"); 2929 // However that wouldn't be right, because it's possible that 2930 // a safepoint is indeed in progress as a young generation 2931 // stop-the-world GC happens even as we mark in this generation. 2932 assert(_collectorState == Marking, "inconsistent state?"); 2933 check_correct_thread_executing(); 2934 verify_overflow_empty(); 2935 2936 // Weak ref discovery note: We may be discovering weak 2937 // refs in this generation concurrent (but interleaved) with 2938 // weak ref discovery by the young generation collector. 2939 2940 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2941 GCTraceCPUTime tcpu; 2942 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2943 bool res = markFromRootsWork(); 2944 if (res) { 2945 _collectorState = Precleaning; 2946 } else { // We failed and a foreground collection wants to take over 2947 assert(_foregroundGCIsActive, "internal state inconsistency"); 2948 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2949 log_debug(gc)("bailing out to foreground collection"); 2950 } 2951 verify_overflow_empty(); 2952 return res; 2953 } 2954 2955 bool CMSCollector::markFromRootsWork() { 2956 // iterate over marked bits in bit map, doing a full scan and mark 2957 // from these roots using the following algorithm: 2958 // . if oop is to the right of the current scan pointer, 2959 // mark corresponding bit (we'll process it later) 2960 // . else (oop is to left of current scan pointer) 2961 // push oop on marking stack 2962 // . drain the marking stack 2963 2964 // Note that when we do a marking step we need to hold the 2965 // bit map lock -- recall that direct allocation (by mutators) 2966 // and promotion (by the young generation collector) is also 2967 // marking the bit map. [the so-called allocate live policy.] 2968 // Because the implementation of bit map marking is not 2969 // robust wrt simultaneous marking of bits in the same word, 2970 // we need to make sure that there is no such interference 2971 // between concurrent such updates. 2972 2973 // already have locks 2974 assert_lock_strong(bitMapLock()); 2975 2976 verify_work_stacks_empty(); 2977 verify_overflow_empty(); 2978 bool result = false; 2979 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2980 result = do_marking_mt(); 2981 } else { 2982 result = do_marking_st(); 2983 } 2984 return result; 2985 } 2986 2987 // Forward decl 2988 class CMSConcMarkingTask; 2989 2990 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2991 CMSCollector* _collector; 2992 CMSConcMarkingTask* _task; 2993 public: 2994 virtual void yield(); 2995 2996 // "n_threads" is the number of threads to be terminated. 2997 // "queue_set" is a set of work queues of other threads. 2998 // "collector" is the CMS collector associated with this task terminator. 2999 // "yield" indicates whether we need the gang as a whole to yield. 3000 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3001 ParallelTaskTerminator(n_threads, queue_set), 3002 _collector(collector) { } 3003 3004 void set_task(CMSConcMarkingTask* task) { 3005 _task = task; 3006 } 3007 }; 3008 3009 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3010 CMSConcMarkingTask* _task; 3011 public: 3012 bool should_exit_termination(); 3013 void set_task(CMSConcMarkingTask* task) { 3014 _task = task; 3015 } 3016 }; 3017 3018 // MT Concurrent Marking Task 3019 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3020 CMSCollector* _collector; 3021 uint _n_workers; // requested/desired # workers 3022 bool _result; 3023 CompactibleFreeListSpace* _cms_space; 3024 char _pad_front[64]; // padding to ... 3025 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3026 char _pad_back[64]; 3027 HeapWord* _restart_addr; 3028 3029 // Exposed here for yielding support 3030 Mutex* const _bit_map_lock; 3031 3032 // The per thread work queues, available here for stealing 3033 OopTaskQueueSet* _task_queues; 3034 3035 // Termination (and yielding) support 3036 CMSConcMarkingTerminator _term; 3037 CMSConcMarkingTerminatorTerminator _term_term; 3038 3039 public: 3040 CMSConcMarkingTask(CMSCollector* collector, 3041 CompactibleFreeListSpace* cms_space, 3042 YieldingFlexibleWorkGang* workers, 3043 OopTaskQueueSet* task_queues): 3044 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3045 _collector(collector), 3046 _cms_space(cms_space), 3047 _n_workers(0), _result(true), 3048 _task_queues(task_queues), 3049 _term(_n_workers, task_queues, _collector), 3050 _bit_map_lock(collector->bitMapLock()) 3051 { 3052 _requested_size = _n_workers; 3053 _term.set_task(this); 3054 _term_term.set_task(this); 3055 _restart_addr = _global_finger = _cms_space->bottom(); 3056 } 3057 3058 3059 OopTaskQueueSet* task_queues() { return _task_queues; } 3060 3061 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3062 3063 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3064 3065 CMSConcMarkingTerminator* terminator() { return &_term; } 3066 3067 virtual void set_for_termination(uint active_workers) { 3068 terminator()->reset_for_reuse(active_workers); 3069 } 3070 3071 void work(uint worker_id); 3072 bool should_yield() { 3073 return ConcurrentMarkSweepThread::should_yield() 3074 && !_collector->foregroundGCIsActive(); 3075 } 3076 3077 virtual void coordinator_yield(); // stuff done by coordinator 3078 bool result() { return _result; } 3079 3080 void reset(HeapWord* ra) { 3081 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3082 _restart_addr = _global_finger = ra; 3083 _term.reset_for_reuse(); 3084 } 3085 3086 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3087 OopTaskQueue* work_q); 3088 3089 private: 3090 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3091 void do_work_steal(int i); 3092 void bump_global_finger(HeapWord* f); 3093 }; 3094 3095 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3096 assert(_task != NULL, "Error"); 3097 return _task->yielding(); 3098 // Note that we do not need the disjunct || _task->should_yield() above 3099 // because we want terminating threads to yield only if the task 3100 // is already in the midst of yielding, which happens only after at least one 3101 // thread has yielded. 3102 } 3103 3104 void CMSConcMarkingTerminator::yield() { 3105 if (_task->should_yield()) { 3106 _task->yield(); 3107 } else { 3108 ParallelTaskTerminator::yield(); 3109 } 3110 } 3111 3112 //////////////////////////////////////////////////////////////// 3113 // Concurrent Marking Algorithm Sketch 3114 //////////////////////////////////////////////////////////////// 3115 // Until all tasks exhausted (both spaces): 3116 // -- claim next available chunk 3117 // -- bump global finger via CAS 3118 // -- find first object that starts in this chunk 3119 // and start scanning bitmap from that position 3120 // -- scan marked objects for oops 3121 // -- CAS-mark target, and if successful: 3122 // . if target oop is above global finger (volatile read) 3123 // nothing to do 3124 // . if target oop is in chunk and above local finger 3125 // then nothing to do 3126 // . else push on work-queue 3127 // -- Deal with possible overflow issues: 3128 // . local work-queue overflow causes stuff to be pushed on 3129 // global (common) overflow queue 3130 // . always first empty local work queue 3131 // . then get a batch of oops from global work queue if any 3132 // . then do work stealing 3133 // -- When all tasks claimed (both spaces) 3134 // and local work queue empty, 3135 // then in a loop do: 3136 // . check global overflow stack; steal a batch of oops and trace 3137 // . try to steal from other threads oif GOS is empty 3138 // . if neither is available, offer termination 3139 // -- Terminate and return result 3140 // 3141 void CMSConcMarkingTask::work(uint worker_id) { 3142 elapsedTimer _timer; 3143 ResourceMark rm; 3144 HandleMark hm; 3145 3146 DEBUG_ONLY(_collector->verify_overflow_empty();) 3147 3148 // Before we begin work, our work queue should be empty 3149 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3150 // Scan the bitmap covering _cms_space, tracing through grey objects. 3151 _timer.start(); 3152 do_scan_and_mark(worker_id, _cms_space); 3153 _timer.stop(); 3154 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3155 3156 // ... do work stealing 3157 _timer.reset(); 3158 _timer.start(); 3159 do_work_steal(worker_id); 3160 _timer.stop(); 3161 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3162 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3163 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3164 // Note that under the current task protocol, the 3165 // following assertion is true even of the spaces 3166 // expanded since the completion of the concurrent 3167 // marking. XXX This will likely change under a strict 3168 // ABORT semantics. 3169 // After perm removal the comparison was changed to 3170 // greater than or equal to from strictly greater than. 3171 // Before perm removal the highest address sweep would 3172 // have been at the end of perm gen but now is at the 3173 // end of the tenured gen. 3174 assert(_global_finger >= _cms_space->end(), 3175 "All tasks have been completed"); 3176 DEBUG_ONLY(_collector->verify_overflow_empty();) 3177 } 3178 3179 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3180 HeapWord* read = _global_finger; 3181 HeapWord* cur = read; 3182 while (f > read) { 3183 cur = read; 3184 read = Atomic::cmpxchg(f, &_global_finger, cur); 3185 if (cur == read) { 3186 // our cas succeeded 3187 assert(_global_finger >= f, "protocol consistency"); 3188 break; 3189 } 3190 } 3191 } 3192 3193 // This is really inefficient, and should be redone by 3194 // using (not yet available) block-read and -write interfaces to the 3195 // stack and the work_queue. XXX FIX ME !!! 3196 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3197 OopTaskQueue* work_q) { 3198 // Fast lock-free check 3199 if (ovflw_stk->length() == 0) { 3200 return false; 3201 } 3202 assert(work_q->size() == 0, "Shouldn't steal"); 3203 MutexLockerEx ml(ovflw_stk->par_lock(), 3204 Mutex::_no_safepoint_check_flag); 3205 // Grab up to 1/4 the size of the work queue 3206 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3207 (size_t)ParGCDesiredObjsFromOverflowList); 3208 num = MIN2(num, ovflw_stk->length()); 3209 for (int i = (int) num; i > 0; i--) { 3210 oop cur = ovflw_stk->pop(); 3211 assert(cur != NULL, "Counted wrong?"); 3212 work_q->push(cur); 3213 } 3214 return num > 0; 3215 } 3216 3217 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3218 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3219 int n_tasks = pst->n_tasks(); 3220 // We allow that there may be no tasks to do here because 3221 // we are restarting after a stack overflow. 3222 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3223 uint nth_task = 0; 3224 3225 HeapWord* aligned_start = sp->bottom(); 3226 if (sp->used_region().contains(_restart_addr)) { 3227 // Align down to a card boundary for the start of 0th task 3228 // for this space. 3229 aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size); 3230 } 3231 3232 size_t chunk_size = sp->marking_task_size(); 3233 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3234 // Having claimed the nth task in this space, 3235 // compute the chunk that it corresponds to: 3236 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3237 aligned_start + (nth_task+1)*chunk_size); 3238 // Try and bump the global finger via a CAS; 3239 // note that we need to do the global finger bump 3240 // _before_ taking the intersection below, because 3241 // the task corresponding to that region will be 3242 // deemed done even if the used_region() expands 3243 // because of allocation -- as it almost certainly will 3244 // during start-up while the threads yield in the 3245 // closure below. 3246 HeapWord* finger = span.end(); 3247 bump_global_finger(finger); // atomically 3248 // There are null tasks here corresponding to chunks 3249 // beyond the "top" address of the space. 3250 span = span.intersection(sp->used_region()); 3251 if (!span.is_empty()) { // Non-null task 3252 HeapWord* prev_obj; 3253 assert(!span.contains(_restart_addr) || nth_task == 0, 3254 "Inconsistency"); 3255 if (nth_task == 0) { 3256 // For the 0th task, we'll not need to compute a block_start. 3257 if (span.contains(_restart_addr)) { 3258 // In the case of a restart because of stack overflow, 3259 // we might additionally skip a chunk prefix. 3260 prev_obj = _restart_addr; 3261 } else { 3262 prev_obj = span.start(); 3263 } 3264 } else { 3265 // We want to skip the first object because 3266 // the protocol is to scan any object in its entirety 3267 // that _starts_ in this span; a fortiori, any 3268 // object starting in an earlier span is scanned 3269 // as part of an earlier claimed task. 3270 // Below we use the "careful" version of block_start 3271 // so we do not try to navigate uninitialized objects. 3272 prev_obj = sp->block_start_careful(span.start()); 3273 // Below we use a variant of block_size that uses the 3274 // Printezis bits to avoid waiting for allocated 3275 // objects to become initialized/parsable. 3276 while (prev_obj < span.start()) { 3277 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3278 if (sz > 0) { 3279 prev_obj += sz; 3280 } else { 3281 // In this case we may end up doing a bit of redundant 3282 // scanning, but that appears unavoidable, short of 3283 // locking the free list locks; see bug 6324141. 3284 break; 3285 } 3286 } 3287 } 3288 if (prev_obj < span.end()) { 3289 MemRegion my_span = MemRegion(prev_obj, span.end()); 3290 // Do the marking work within a non-empty span -- 3291 // the last argument to the constructor indicates whether the 3292 // iteration should be incremental with periodic yields. 3293 ParMarkFromRootsClosure cl(this, _collector, my_span, 3294 &_collector->_markBitMap, 3295 work_queue(i), 3296 &_collector->_markStack); 3297 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3298 } // else nothing to do for this task 3299 } // else nothing to do for this task 3300 } 3301 // We'd be tempted to assert here that since there are no 3302 // more tasks left to claim in this space, the global_finger 3303 // must exceed space->top() and a fortiori space->end(). However, 3304 // that would not quite be correct because the bumping of 3305 // global_finger occurs strictly after the claiming of a task, 3306 // so by the time we reach here the global finger may not yet 3307 // have been bumped up by the thread that claimed the last 3308 // task. 3309 pst->all_tasks_completed(); 3310 } 3311 3312 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3313 private: 3314 CMSCollector* _collector; 3315 CMSConcMarkingTask* _task; 3316 MemRegion _span; 3317 CMSBitMap* _bit_map; 3318 CMSMarkStack* _overflow_stack; 3319 OopTaskQueue* _work_queue; 3320 protected: 3321 DO_OOP_WORK_DEFN 3322 public: 3323 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3324 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3325 MetadataAwareOopClosure(collector->ref_processor()), 3326 _collector(collector), 3327 _task(task), 3328 _span(collector->_span), 3329 _work_queue(work_queue), 3330 _bit_map(bit_map), 3331 _overflow_stack(overflow_stack) 3332 { } 3333 virtual void do_oop(oop* p); 3334 virtual void do_oop(narrowOop* p); 3335 3336 void trim_queue(size_t max); 3337 void handle_stack_overflow(HeapWord* lost); 3338 void do_yield_check() { 3339 if (_task->should_yield()) { 3340 _task->yield(); 3341 } 3342 } 3343 }; 3344 3345 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3346 3347 // Grey object scanning during work stealing phase -- 3348 // the salient assumption here is that any references 3349 // that are in these stolen objects being scanned must 3350 // already have been initialized (else they would not have 3351 // been published), so we do not need to check for 3352 // uninitialized objects before pushing here. 3353 void ParConcMarkingClosure::do_oop(oop obj) { 3354 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3355 HeapWord* addr = (HeapWord*)obj; 3356 // Check if oop points into the CMS generation 3357 // and is not marked 3358 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3359 // a white object ... 3360 // If we manage to "claim" the object, by being the 3361 // first thread to mark it, then we push it on our 3362 // marking stack 3363 if (_bit_map->par_mark(addr)) { // ... now grey 3364 // push on work queue (grey set) 3365 bool simulate_overflow = false; 3366 NOT_PRODUCT( 3367 if (CMSMarkStackOverflowALot && 3368 _collector->simulate_overflow()) { 3369 // simulate a stack overflow 3370 simulate_overflow = true; 3371 } 3372 ) 3373 if (simulate_overflow || 3374 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3375 // stack overflow 3376 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3377 // We cannot assert that the overflow stack is full because 3378 // it may have been emptied since. 3379 assert(simulate_overflow || 3380 _work_queue->size() == _work_queue->max_elems(), 3381 "Else push should have succeeded"); 3382 handle_stack_overflow(addr); 3383 } 3384 } // Else, some other thread got there first 3385 do_yield_check(); 3386 } 3387 } 3388 3389 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3390 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3391 3392 void ParConcMarkingClosure::trim_queue(size_t max) { 3393 while (_work_queue->size() > max) { 3394 oop new_oop; 3395 if (_work_queue->pop_local(new_oop)) { 3396 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3397 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3398 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3399 new_oop->oop_iterate(this); // do_oop() above 3400 do_yield_check(); 3401 } 3402 } 3403 } 3404 3405 // Upon stack overflow, we discard (part of) the stack, 3406 // remembering the least address amongst those discarded 3407 // in CMSCollector's _restart_address. 3408 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3409 // We need to do this under a mutex to prevent other 3410 // workers from interfering with the work done below. 3411 MutexLockerEx ml(_overflow_stack->par_lock(), 3412 Mutex::_no_safepoint_check_flag); 3413 // Remember the least grey address discarded 3414 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3415 _collector->lower_restart_addr(ra); 3416 _overflow_stack->reset(); // discard stack contents 3417 _overflow_stack->expand(); // expand the stack if possible 3418 } 3419 3420 3421 void CMSConcMarkingTask::do_work_steal(int i) { 3422 OopTaskQueue* work_q = work_queue(i); 3423 oop obj_to_scan; 3424 CMSBitMap* bm = &(_collector->_markBitMap); 3425 CMSMarkStack* ovflw = &(_collector->_markStack); 3426 int* seed = _collector->hash_seed(i); 3427 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3428 while (true) { 3429 cl.trim_queue(0); 3430 assert(work_q->size() == 0, "Should have been emptied above"); 3431 if (get_work_from_overflow_stack(ovflw, work_q)) { 3432 // Can't assert below because the work obtained from the 3433 // overflow stack may already have been stolen from us. 3434 // assert(work_q->size() > 0, "Work from overflow stack"); 3435 continue; 3436 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3437 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3438 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3439 obj_to_scan->oop_iterate(&cl); 3440 } else if (terminator()->offer_termination(&_term_term)) { 3441 assert(work_q->size() == 0, "Impossible!"); 3442 break; 3443 } else if (yielding() || should_yield()) { 3444 yield(); 3445 } 3446 } 3447 } 3448 3449 // This is run by the CMS (coordinator) thread. 3450 void CMSConcMarkingTask::coordinator_yield() { 3451 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3452 "CMS thread should hold CMS token"); 3453 // First give up the locks, then yield, then re-lock 3454 // We should probably use a constructor/destructor idiom to 3455 // do this unlock/lock or modify the MutexUnlocker class to 3456 // serve our purpose. XXX 3457 assert_lock_strong(_bit_map_lock); 3458 _bit_map_lock->unlock(); 3459 ConcurrentMarkSweepThread::desynchronize(true); 3460 _collector->stopTimer(); 3461 _collector->incrementYields(); 3462 3463 // It is possible for whichever thread initiated the yield request 3464 // not to get a chance to wake up and take the bitmap lock between 3465 // this thread releasing it and reacquiring it. So, while the 3466 // should_yield() flag is on, let's sleep for a bit to give the 3467 // other thread a chance to wake up. The limit imposed on the number 3468 // of iterations is defensive, to avoid any unforseen circumstances 3469 // putting us into an infinite loop. Since it's always been this 3470 // (coordinator_yield()) method that was observed to cause the 3471 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3472 // which is by default non-zero. For the other seven methods that 3473 // also perform the yield operation, as are using a different 3474 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3475 // can enable the sleeping for those methods too, if necessary. 3476 // See 6442774. 3477 // 3478 // We really need to reconsider the synchronization between the GC 3479 // thread and the yield-requesting threads in the future and we 3480 // should really use wait/notify, which is the recommended 3481 // way of doing this type of interaction. Additionally, we should 3482 // consolidate the eight methods that do the yield operation and they 3483 // are almost identical into one for better maintainability and 3484 // readability. See 6445193. 3485 // 3486 // Tony 2006.06.29 3487 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3488 ConcurrentMarkSweepThread::should_yield() && 3489 !CMSCollector::foregroundGCIsActive(); ++i) { 3490 os::sleep(Thread::current(), 1, false); 3491 } 3492 3493 ConcurrentMarkSweepThread::synchronize(true); 3494 _bit_map_lock->lock_without_safepoint_check(); 3495 _collector->startTimer(); 3496 } 3497 3498 bool CMSCollector::do_marking_mt() { 3499 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3500 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3501 conc_workers()->active_workers(), 3502 Threads::number_of_non_daemon_threads()); 3503 num_workers = conc_workers()->update_active_workers(num_workers); 3504 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3505 3506 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3507 3508 CMSConcMarkingTask tsk(this, 3509 cms_space, 3510 conc_workers(), 3511 task_queues()); 3512 3513 // Since the actual number of workers we get may be different 3514 // from the number we requested above, do we need to do anything different 3515 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3516 // class?? XXX 3517 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3518 3519 // Refs discovery is already non-atomic. 3520 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3521 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3522 conc_workers()->start_task(&tsk); 3523 while (tsk.yielded()) { 3524 tsk.coordinator_yield(); 3525 conc_workers()->continue_task(&tsk); 3526 } 3527 // If the task was aborted, _restart_addr will be non-NULL 3528 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3529 while (_restart_addr != NULL) { 3530 // XXX For now we do not make use of ABORTED state and have not 3531 // yet implemented the right abort semantics (even in the original 3532 // single-threaded CMS case). That needs some more investigation 3533 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3534 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3535 // If _restart_addr is non-NULL, a marking stack overflow 3536 // occurred; we need to do a fresh marking iteration from the 3537 // indicated restart address. 3538 if (_foregroundGCIsActive) { 3539 // We may be running into repeated stack overflows, having 3540 // reached the limit of the stack size, while making very 3541 // slow forward progress. It may be best to bail out and 3542 // let the foreground collector do its job. 3543 // Clear _restart_addr, so that foreground GC 3544 // works from scratch. This avoids the headache of 3545 // a "rescan" which would otherwise be needed because 3546 // of the dirty mod union table & card table. 3547 _restart_addr = NULL; 3548 return false; 3549 } 3550 // Adjust the task to restart from _restart_addr 3551 tsk.reset(_restart_addr); 3552 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3553 _restart_addr); 3554 _restart_addr = NULL; 3555 // Get the workers going again 3556 conc_workers()->start_task(&tsk); 3557 while (tsk.yielded()) { 3558 tsk.coordinator_yield(); 3559 conc_workers()->continue_task(&tsk); 3560 } 3561 } 3562 assert(tsk.completed(), "Inconsistency"); 3563 assert(tsk.result() == true, "Inconsistency"); 3564 return true; 3565 } 3566 3567 bool CMSCollector::do_marking_st() { 3568 ResourceMark rm; 3569 HandleMark hm; 3570 3571 // Temporarily make refs discovery single threaded (non-MT) 3572 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3573 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3574 &_markStack, CMSYield); 3575 // the last argument to iterate indicates whether the iteration 3576 // should be incremental with periodic yields. 3577 _markBitMap.iterate(&markFromRootsClosure); 3578 // If _restart_addr is non-NULL, a marking stack overflow 3579 // occurred; we need to do a fresh iteration from the 3580 // indicated restart address. 3581 while (_restart_addr != NULL) { 3582 if (_foregroundGCIsActive) { 3583 // We may be running into repeated stack overflows, having 3584 // reached the limit of the stack size, while making very 3585 // slow forward progress. It may be best to bail out and 3586 // let the foreground collector do its job. 3587 // Clear _restart_addr, so that foreground GC 3588 // works from scratch. This avoids the headache of 3589 // a "rescan" which would otherwise be needed because 3590 // of the dirty mod union table & card table. 3591 _restart_addr = NULL; 3592 return false; // indicating failure to complete marking 3593 } 3594 // Deal with stack overflow: 3595 // we restart marking from _restart_addr 3596 HeapWord* ra = _restart_addr; 3597 markFromRootsClosure.reset(ra); 3598 _restart_addr = NULL; 3599 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3600 } 3601 return true; 3602 } 3603 3604 void CMSCollector::preclean() { 3605 check_correct_thread_executing(); 3606 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3607 verify_work_stacks_empty(); 3608 verify_overflow_empty(); 3609 _abort_preclean = false; 3610 if (CMSPrecleaningEnabled) { 3611 if (!CMSEdenChunksRecordAlways) { 3612 _eden_chunk_index = 0; 3613 } 3614 size_t used = get_eden_used(); 3615 size_t capacity = get_eden_capacity(); 3616 // Don't start sampling unless we will get sufficiently 3617 // many samples. 3618 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3619 * CMSScheduleRemarkEdenPenetration)) { 3620 _start_sampling = true; 3621 } else { 3622 _start_sampling = false; 3623 } 3624 GCTraceCPUTime tcpu; 3625 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3626 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3627 } 3628 CMSTokenSync x(true); // is cms thread 3629 if (CMSPrecleaningEnabled) { 3630 sample_eden(); 3631 _collectorState = AbortablePreclean; 3632 } else { 3633 _collectorState = FinalMarking; 3634 } 3635 verify_work_stacks_empty(); 3636 verify_overflow_empty(); 3637 } 3638 3639 // Try and schedule the remark such that young gen 3640 // occupancy is CMSScheduleRemarkEdenPenetration %. 3641 void CMSCollector::abortable_preclean() { 3642 check_correct_thread_executing(); 3643 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3644 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3645 3646 // If Eden's current occupancy is below this threshold, 3647 // immediately schedule the remark; else preclean 3648 // past the next scavenge in an effort to 3649 // schedule the pause as described above. By choosing 3650 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3651 // we will never do an actual abortable preclean cycle. 3652 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3653 GCTraceCPUTime tcpu; 3654 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3655 // We need more smarts in the abortable preclean 3656 // loop below to deal with cases where allocation 3657 // in young gen is very very slow, and our precleaning 3658 // is running a losing race against a horde of 3659 // mutators intent on flooding us with CMS updates 3660 // (dirty cards). 3661 // One, admittedly dumb, strategy is to give up 3662 // after a certain number of abortable precleaning loops 3663 // or after a certain maximum time. We want to make 3664 // this smarter in the next iteration. 3665 // XXX FIX ME!!! YSR 3666 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3667 while (!(should_abort_preclean() || 3668 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3669 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3670 cumworkdone += workdone; 3671 loops++; 3672 // Voluntarily terminate abortable preclean phase if we have 3673 // been at it for too long. 3674 if ((CMSMaxAbortablePrecleanLoops != 0) && 3675 loops >= CMSMaxAbortablePrecleanLoops) { 3676 log_debug(gc)(" CMS: abort preclean due to loops "); 3677 break; 3678 } 3679 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3680 log_debug(gc)(" CMS: abort preclean due to time "); 3681 break; 3682 } 3683 // If we are doing little work each iteration, we should 3684 // take a short break. 3685 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3686 // Sleep for some time, waiting for work to accumulate 3687 stopTimer(); 3688 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3689 startTimer(); 3690 waited++; 3691 } 3692 } 3693 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3694 loops, waited, cumworkdone); 3695 } 3696 CMSTokenSync x(true); // is cms thread 3697 if (_collectorState != Idling) { 3698 assert(_collectorState == AbortablePreclean, 3699 "Spontaneous state transition?"); 3700 _collectorState = FinalMarking; 3701 } // Else, a foreground collection completed this CMS cycle. 3702 return; 3703 } 3704 3705 // Respond to an Eden sampling opportunity 3706 void CMSCollector::sample_eden() { 3707 // Make sure a young gc cannot sneak in between our 3708 // reading and recording of a sample. 3709 assert(Thread::current()->is_ConcurrentGC_thread(), 3710 "Only the cms thread may collect Eden samples"); 3711 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3712 "Should collect samples while holding CMS token"); 3713 if (!_start_sampling) { 3714 return; 3715 } 3716 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3717 // is populated by the young generation. 3718 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3719 if (_eden_chunk_index < _eden_chunk_capacity) { 3720 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3721 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3722 "Unexpected state of Eden"); 3723 // We'd like to check that what we just sampled is an oop-start address; 3724 // however, we cannot do that here since the object may not yet have been 3725 // initialized. So we'll instead do the check when we _use_ this sample 3726 // later. 3727 if (_eden_chunk_index == 0 || 3728 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3729 _eden_chunk_array[_eden_chunk_index-1]) 3730 >= CMSSamplingGrain)) { 3731 _eden_chunk_index++; // commit sample 3732 } 3733 } 3734 } 3735 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3736 size_t used = get_eden_used(); 3737 size_t capacity = get_eden_capacity(); 3738 assert(used <= capacity, "Unexpected state of Eden"); 3739 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3740 _abort_preclean = true; 3741 } 3742 } 3743 } 3744 3745 3746 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3747 assert(_collectorState == Precleaning || 3748 _collectorState == AbortablePreclean, "incorrect state"); 3749 ResourceMark rm; 3750 HandleMark hm; 3751 3752 // Precleaning is currently not MT but the reference processor 3753 // may be set for MT. Disable it temporarily here. 3754 ReferenceProcessor* rp = ref_processor(); 3755 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3756 3757 // Do one pass of scrubbing the discovered reference lists 3758 // to remove any reference objects with strongly-reachable 3759 // referents. 3760 if (clean_refs) { 3761 CMSPrecleanRefsYieldClosure yield_cl(this); 3762 assert(rp->span().equals(_span), "Spans should be equal"); 3763 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3764 &_markStack, true /* preclean */); 3765 CMSDrainMarkingStackClosure complete_trace(this, 3766 _span, &_markBitMap, &_markStack, 3767 &keep_alive, true /* preclean */); 3768 3769 // We don't want this step to interfere with a young 3770 // collection because we don't want to take CPU 3771 // or memory bandwidth away from the young GC threads 3772 // (which may be as many as there are CPUs). 3773 // Note that we don't need to protect ourselves from 3774 // interference with mutators because they can't 3775 // manipulate the discovered reference lists nor affect 3776 // the computed reachability of the referents, the 3777 // only properties manipulated by the precleaning 3778 // of these reference lists. 3779 stopTimer(); 3780 CMSTokenSyncWithLocks x(true /* is cms thread */, 3781 bitMapLock()); 3782 startTimer(); 3783 sample_eden(); 3784 3785 // The following will yield to allow foreground 3786 // collection to proceed promptly. XXX YSR: 3787 // The code in this method may need further 3788 // tweaking for better performance and some restructuring 3789 // for cleaner interfaces. 3790 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3791 rp->preclean_discovered_references( 3792 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3793 gc_timer); 3794 } 3795 3796 if (clean_survivor) { // preclean the active survivor space(s) 3797 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3798 &_markBitMap, &_modUnionTable, 3799 &_markStack, true /* precleaning phase */); 3800 stopTimer(); 3801 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3802 bitMapLock()); 3803 startTimer(); 3804 unsigned int before_count = 3805 CMSHeap::heap()->total_collections(); 3806 SurvivorSpacePrecleanClosure 3807 sss_cl(this, _span, &_markBitMap, &_markStack, 3808 &pam_cl, before_count, CMSYield); 3809 _young_gen->from()->object_iterate_careful(&sss_cl); 3810 _young_gen->to()->object_iterate_careful(&sss_cl); 3811 } 3812 MarkRefsIntoAndScanClosure 3813 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3814 &_markStack, this, CMSYield, 3815 true /* precleaning phase */); 3816 // CAUTION: The following closure has persistent state that may need to 3817 // be reset upon a decrease in the sequence of addresses it 3818 // processes. 3819 ScanMarkedObjectsAgainCarefullyClosure 3820 smoac_cl(this, _span, 3821 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3822 3823 // Preclean dirty cards in ModUnionTable and CardTable using 3824 // appropriate convergence criterion; 3825 // repeat CMSPrecleanIter times unless we find that 3826 // we are losing. 3827 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3828 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3829 "Bad convergence multiplier"); 3830 assert(CMSPrecleanThreshold >= 100, 3831 "Unreasonably low CMSPrecleanThreshold"); 3832 3833 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3834 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3835 numIter < CMSPrecleanIter; 3836 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3837 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3838 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3839 // Either there are very few dirty cards, so re-mark 3840 // pause will be small anyway, or our pre-cleaning isn't 3841 // that much faster than the rate at which cards are being 3842 // dirtied, so we might as well stop and re-mark since 3843 // precleaning won't improve our re-mark time by much. 3844 if (curNumCards <= CMSPrecleanThreshold || 3845 (numIter > 0 && 3846 (curNumCards * CMSPrecleanDenominator > 3847 lastNumCards * CMSPrecleanNumerator))) { 3848 numIter++; 3849 cumNumCards += curNumCards; 3850 break; 3851 } 3852 } 3853 3854 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3855 3856 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3857 cumNumCards += curNumCards; 3858 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3859 curNumCards, cumNumCards, numIter); 3860 return cumNumCards; // as a measure of useful work done 3861 } 3862 3863 // PRECLEANING NOTES: 3864 // Precleaning involves: 3865 // . reading the bits of the modUnionTable and clearing the set bits. 3866 // . For the cards corresponding to the set bits, we scan the 3867 // objects on those cards. This means we need the free_list_lock 3868 // so that we can safely iterate over the CMS space when scanning 3869 // for oops. 3870 // . When we scan the objects, we'll be both reading and setting 3871 // marks in the marking bit map, so we'll need the marking bit map. 3872 // . For protecting _collector_state transitions, we take the CGC_lock. 3873 // Note that any races in the reading of of card table entries by the 3874 // CMS thread on the one hand and the clearing of those entries by the 3875 // VM thread or the setting of those entries by the mutator threads on the 3876 // other are quite benign. However, for efficiency it makes sense to keep 3877 // the VM thread from racing with the CMS thread while the latter is 3878 // dirty card info to the modUnionTable. We therefore also use the 3879 // CGC_lock to protect the reading of the card table and the mod union 3880 // table by the CM thread. 3881 // . We run concurrently with mutator updates, so scanning 3882 // needs to be done carefully -- we should not try to scan 3883 // potentially uninitialized objects. 3884 // 3885 // Locking strategy: While holding the CGC_lock, we scan over and 3886 // reset a maximal dirty range of the mod union / card tables, then lock 3887 // the free_list_lock and bitmap lock to do a full marking, then 3888 // release these locks; and repeat the cycle. This allows for a 3889 // certain amount of fairness in the sharing of these locks between 3890 // the CMS collector on the one hand, and the VM thread and the 3891 // mutators on the other. 3892 3893 // NOTE: preclean_mod_union_table() and preclean_card_table() 3894 // further below are largely identical; if you need to modify 3895 // one of these methods, please check the other method too. 3896 3897 size_t CMSCollector::preclean_mod_union_table( 3898 ConcurrentMarkSweepGeneration* old_gen, 3899 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3900 verify_work_stacks_empty(); 3901 verify_overflow_empty(); 3902 3903 // strategy: starting with the first card, accumulate contiguous 3904 // ranges of dirty cards; clear these cards, then scan the region 3905 // covered by these cards. 3906 3907 // Since all of the MUT is committed ahead, we can just use 3908 // that, in case the generations expand while we are precleaning. 3909 // It might also be fine to just use the committed part of the 3910 // generation, but we might potentially miss cards when the 3911 // generation is rapidly expanding while we are in the midst 3912 // of precleaning. 3913 HeapWord* startAddr = old_gen->reserved().start(); 3914 HeapWord* endAddr = old_gen->reserved().end(); 3915 3916 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3917 3918 size_t numDirtyCards, cumNumDirtyCards; 3919 HeapWord *nextAddr, *lastAddr; 3920 for (cumNumDirtyCards = numDirtyCards = 0, 3921 nextAddr = lastAddr = startAddr; 3922 nextAddr < endAddr; 3923 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3924 3925 ResourceMark rm; 3926 HandleMark hm; 3927 3928 MemRegion dirtyRegion; 3929 { 3930 stopTimer(); 3931 // Potential yield point 3932 CMSTokenSync ts(true); 3933 startTimer(); 3934 sample_eden(); 3935 // Get dirty region starting at nextOffset (inclusive), 3936 // simultaneously clearing it. 3937 dirtyRegion = 3938 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3939 assert(dirtyRegion.start() >= nextAddr, 3940 "returned region inconsistent?"); 3941 } 3942 // Remember where the next search should begin. 3943 // The returned region (if non-empty) is a right open interval, 3944 // so lastOffset is obtained from the right end of that 3945 // interval. 3946 lastAddr = dirtyRegion.end(); 3947 // Should do something more transparent and less hacky XXX 3948 numDirtyCards = 3949 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3950 3951 // We'll scan the cards in the dirty region (with periodic 3952 // yields for foreground GC as needed). 3953 if (!dirtyRegion.is_empty()) { 3954 assert(numDirtyCards > 0, "consistency check"); 3955 HeapWord* stop_point = NULL; 3956 stopTimer(); 3957 // Potential yield point 3958 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3959 bitMapLock()); 3960 startTimer(); 3961 { 3962 verify_work_stacks_empty(); 3963 verify_overflow_empty(); 3964 sample_eden(); 3965 stop_point = 3966 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3967 } 3968 if (stop_point != NULL) { 3969 // The careful iteration stopped early either because it found an 3970 // uninitialized object, or because we were in the midst of an 3971 // "abortable preclean", which should now be aborted. Redirty 3972 // the bits corresponding to the partially-scanned or unscanned 3973 // cards. We'll either restart at the next block boundary or 3974 // abort the preclean. 3975 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3976 "Should only be AbortablePreclean."); 3977 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3978 if (should_abort_preclean()) { 3979 break; // out of preclean loop 3980 } else { 3981 // Compute the next address at which preclean should pick up; 3982 // might need bitMapLock in order to read P-bits. 3983 lastAddr = next_card_start_after_block(stop_point); 3984 } 3985 } 3986 } else { 3987 assert(lastAddr == endAddr, "consistency check"); 3988 assert(numDirtyCards == 0, "consistency check"); 3989 break; 3990 } 3991 } 3992 verify_work_stacks_empty(); 3993 verify_overflow_empty(); 3994 return cumNumDirtyCards; 3995 } 3996 3997 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3998 // below are largely identical; if you need to modify 3999 // one of these methods, please check the other method too. 4000 4001 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4002 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4003 // strategy: it's similar to precleamModUnionTable above, in that 4004 // we accumulate contiguous ranges of dirty cards, mark these cards 4005 // precleaned, then scan the region covered by these cards. 4006 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4007 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4008 4009 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4010 4011 size_t numDirtyCards, cumNumDirtyCards; 4012 HeapWord *lastAddr, *nextAddr; 4013 4014 for (cumNumDirtyCards = numDirtyCards = 0, 4015 nextAddr = lastAddr = startAddr; 4016 nextAddr < endAddr; 4017 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4018 4019 ResourceMark rm; 4020 HandleMark hm; 4021 4022 MemRegion dirtyRegion; 4023 { 4024 // See comments in "Precleaning notes" above on why we 4025 // do this locking. XXX Could the locking overheads be 4026 // too high when dirty cards are sparse? [I don't think so.] 4027 stopTimer(); 4028 CMSTokenSync x(true); // is cms thread 4029 startTimer(); 4030 sample_eden(); 4031 // Get and clear dirty region from card table 4032 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4033 MemRegion(nextAddr, endAddr), 4034 true, 4035 CardTableModRefBS::precleaned_card_val()); 4036 4037 assert(dirtyRegion.start() >= nextAddr, 4038 "returned region inconsistent?"); 4039 } 4040 lastAddr = dirtyRegion.end(); 4041 numDirtyCards = 4042 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4043 4044 if (!dirtyRegion.is_empty()) { 4045 stopTimer(); 4046 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4047 startTimer(); 4048 sample_eden(); 4049 verify_work_stacks_empty(); 4050 verify_overflow_empty(); 4051 HeapWord* stop_point = 4052 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4053 if (stop_point != NULL) { 4054 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4055 "Should only be AbortablePreclean."); 4056 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4057 if (should_abort_preclean()) { 4058 break; // out of preclean loop 4059 } else { 4060 // Compute the next address at which preclean should pick up. 4061 lastAddr = next_card_start_after_block(stop_point); 4062 } 4063 } 4064 } else { 4065 break; 4066 } 4067 } 4068 verify_work_stacks_empty(); 4069 verify_overflow_empty(); 4070 return cumNumDirtyCards; 4071 } 4072 4073 class PrecleanCLDClosure : public CLDClosure { 4074 MetadataAwareOopsInGenClosure* _cm_closure; 4075 public: 4076 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4077 void do_cld(ClassLoaderData* cld) { 4078 if (cld->has_accumulated_modified_oops()) { 4079 cld->clear_accumulated_modified_oops(); 4080 4081 _cm_closure->do_cld(cld); 4082 } 4083 } 4084 }; 4085 4086 // The freelist lock is needed to prevent asserts, is it really needed? 4087 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4088 4089 cl->set_freelistLock(freelistLock); 4090 4091 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4092 4093 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4094 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4095 PrecleanCLDClosure preclean_closure(cl); 4096 ClassLoaderDataGraph::cld_do(&preclean_closure); 4097 4098 verify_work_stacks_empty(); 4099 verify_overflow_empty(); 4100 } 4101 4102 void CMSCollector::checkpointRootsFinal() { 4103 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4104 check_correct_thread_executing(); 4105 // world is stopped at this checkpoint 4106 assert(SafepointSynchronize::is_at_safepoint(), 4107 "world should be stopped"); 4108 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4109 4110 verify_work_stacks_empty(); 4111 verify_overflow_empty(); 4112 4113 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4114 _young_gen->used() / K, _young_gen->capacity() / K); 4115 { 4116 if (CMSScavengeBeforeRemark) { 4117 CMSHeap* heap = CMSHeap::heap(); 4118 // Temporarily set flag to false, GCH->do_collection will 4119 // expect it to be false and set to true 4120 FlagSetting fl(heap->_is_gc_active, false); 4121 4122 heap->do_collection(true, // full (i.e. force, see below) 4123 false, // !clear_all_soft_refs 4124 0, // size 4125 false, // is_tlab 4126 GenCollectedHeap::YoungGen // type 4127 ); 4128 } 4129 FreelistLocker x(this); 4130 MutexLockerEx y(bitMapLock(), 4131 Mutex::_no_safepoint_check_flag); 4132 checkpointRootsFinalWork(); 4133 } 4134 verify_work_stacks_empty(); 4135 verify_overflow_empty(); 4136 } 4137 4138 void CMSCollector::checkpointRootsFinalWork() { 4139 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4140 4141 assert(haveFreelistLocks(), "must have free list locks"); 4142 assert_lock_strong(bitMapLock()); 4143 4144 ResourceMark rm; 4145 HandleMark hm; 4146 4147 CMSHeap* heap = CMSHeap::heap(); 4148 4149 if (should_unload_classes()) { 4150 CodeCache::gc_prologue(); 4151 } 4152 assert(haveFreelistLocks(), "must have free list locks"); 4153 assert_lock_strong(bitMapLock()); 4154 4155 // We might assume that we need not fill TLAB's when 4156 // CMSScavengeBeforeRemark is set, because we may have just done 4157 // a scavenge which would have filled all TLAB's -- and besides 4158 // Eden would be empty. This however may not always be the case -- 4159 // for instance although we asked for a scavenge, it may not have 4160 // happened because of a JNI critical section. We probably need 4161 // a policy for deciding whether we can in that case wait until 4162 // the critical section releases and then do the remark following 4163 // the scavenge, and skip it here. In the absence of that policy, 4164 // or of an indication of whether the scavenge did indeed occur, 4165 // we cannot rely on TLAB's having been filled and must do 4166 // so here just in case a scavenge did not happen. 4167 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4168 // Update the saved marks which may affect the root scans. 4169 heap->save_marks(); 4170 4171 print_eden_and_survivor_chunk_arrays(); 4172 4173 { 4174 #if COMPILER2_OR_JVMCI 4175 DerivedPointerTableDeactivate dpt_deact; 4176 #endif 4177 4178 // Note on the role of the mod union table: 4179 // Since the marker in "markFromRoots" marks concurrently with 4180 // mutators, it is possible for some reachable objects not to have been 4181 // scanned. For instance, an only reference to an object A was 4182 // placed in object B after the marker scanned B. Unless B is rescanned, 4183 // A would be collected. Such updates to references in marked objects 4184 // are detected via the mod union table which is the set of all cards 4185 // dirtied since the first checkpoint in this GC cycle and prior to 4186 // the most recent young generation GC, minus those cleaned up by the 4187 // concurrent precleaning. 4188 if (CMSParallelRemarkEnabled) { 4189 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4190 do_remark_parallel(); 4191 } else { 4192 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4193 do_remark_non_parallel(); 4194 } 4195 } 4196 verify_work_stacks_empty(); 4197 verify_overflow_empty(); 4198 4199 { 4200 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4201 refProcessingWork(); 4202 } 4203 verify_work_stacks_empty(); 4204 verify_overflow_empty(); 4205 4206 if (should_unload_classes()) { 4207 CodeCache::gc_epilogue(); 4208 } 4209 JvmtiExport::gc_epilogue(); 4210 4211 // If we encountered any (marking stack / work queue) overflow 4212 // events during the current CMS cycle, take appropriate 4213 // remedial measures, where possible, so as to try and avoid 4214 // recurrence of that condition. 4215 assert(_markStack.isEmpty(), "No grey objects"); 4216 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4217 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4218 if (ser_ovflw > 0) { 4219 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4220 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4221 _markStack.expand(); 4222 _ser_pmc_remark_ovflw = 0; 4223 _ser_pmc_preclean_ovflw = 0; 4224 _ser_kac_preclean_ovflw = 0; 4225 _ser_kac_ovflw = 0; 4226 } 4227 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4228 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4229 _par_pmc_remark_ovflw, _par_kac_ovflw); 4230 _par_pmc_remark_ovflw = 0; 4231 _par_kac_ovflw = 0; 4232 } 4233 if (_markStack._hit_limit > 0) { 4234 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4235 _markStack._hit_limit); 4236 } 4237 if (_markStack._failed_double > 0) { 4238 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4239 _markStack._failed_double, _markStack.capacity()); 4240 } 4241 _markStack._hit_limit = 0; 4242 _markStack._failed_double = 0; 4243 4244 if ((VerifyAfterGC || VerifyDuringGC) && 4245 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4246 verify_after_remark(); 4247 } 4248 4249 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4250 4251 // Change under the freelistLocks. 4252 _collectorState = Sweeping; 4253 // Call isAllClear() under bitMapLock 4254 assert(_modUnionTable.isAllClear(), 4255 "Should be clear by end of the final marking"); 4256 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4257 "Should be clear by end of the final marking"); 4258 } 4259 4260 void CMSParInitialMarkTask::work(uint worker_id) { 4261 elapsedTimer _timer; 4262 ResourceMark rm; 4263 HandleMark hm; 4264 4265 // ---------- scan from roots -------------- 4266 _timer.start(); 4267 CMSHeap* heap = CMSHeap::heap(); 4268 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4269 4270 // ---------- young gen roots -------------- 4271 { 4272 work_on_young_gen_roots(&par_mri_cl); 4273 _timer.stop(); 4274 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4275 } 4276 4277 // ---------- remaining roots -------------- 4278 _timer.reset(); 4279 _timer.start(); 4280 4281 CLDToOopClosure cld_closure(&par_mri_cl, true); 4282 4283 heap->cms_process_roots(_strong_roots_scope, 4284 false, // yg was scanned above 4285 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4286 _collector->should_unload_classes(), 4287 &par_mri_cl, 4288 &cld_closure); 4289 assert(_collector->should_unload_classes() 4290 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4291 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4292 _timer.stop(); 4293 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4294 } 4295 4296 // Parallel remark task 4297 class CMSParRemarkTask: public CMSParMarkTask { 4298 CompactibleFreeListSpace* _cms_space; 4299 4300 // The per-thread work queues, available here for stealing. 4301 OopTaskQueueSet* _task_queues; 4302 ParallelTaskTerminator _term; 4303 StrongRootsScope* _strong_roots_scope; 4304 4305 public: 4306 // A value of 0 passed to n_workers will cause the number of 4307 // workers to be taken from the active workers in the work gang. 4308 CMSParRemarkTask(CMSCollector* collector, 4309 CompactibleFreeListSpace* cms_space, 4310 uint n_workers, WorkGang* workers, 4311 OopTaskQueueSet* task_queues, 4312 StrongRootsScope* strong_roots_scope): 4313 CMSParMarkTask("Rescan roots and grey objects in parallel", 4314 collector, n_workers), 4315 _cms_space(cms_space), 4316 _task_queues(task_queues), 4317 _term(n_workers, task_queues), 4318 _strong_roots_scope(strong_roots_scope) { } 4319 4320 OopTaskQueueSet* task_queues() { return _task_queues; } 4321 4322 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4323 4324 ParallelTaskTerminator* terminator() { return &_term; } 4325 uint n_workers() { return _n_workers; } 4326 4327 void work(uint worker_id); 4328 4329 private: 4330 // ... of dirty cards in old space 4331 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4332 ParMarkRefsIntoAndScanClosure* cl); 4333 4334 // ... work stealing for the above 4335 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4336 }; 4337 4338 class RemarkCLDClosure : public CLDClosure { 4339 CLDToOopClosure _cm_closure; 4340 public: 4341 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4342 void do_cld(ClassLoaderData* cld) { 4343 // Check if we have modified any oops in the CLD during the concurrent marking. 4344 if (cld->has_accumulated_modified_oops()) { 4345 cld->clear_accumulated_modified_oops(); 4346 4347 // We could have transfered the current modified marks to the accumulated marks, 4348 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4349 } else if (cld->has_modified_oops()) { 4350 // Don't clear anything, this info is needed by the next young collection. 4351 } else { 4352 // No modified oops in the ClassLoaderData. 4353 return; 4354 } 4355 4356 // The klass has modified fields, need to scan the klass. 4357 _cm_closure.do_cld(cld); 4358 } 4359 }; 4360 4361 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4362 ParNewGeneration* young_gen = _collector->_young_gen; 4363 ContiguousSpace* eden_space = young_gen->eden(); 4364 ContiguousSpace* from_space = young_gen->from(); 4365 ContiguousSpace* to_space = young_gen->to(); 4366 4367 HeapWord** eca = _collector->_eden_chunk_array; 4368 size_t ect = _collector->_eden_chunk_index; 4369 HeapWord** sca = _collector->_survivor_chunk_array; 4370 size_t sct = _collector->_survivor_chunk_index; 4371 4372 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4373 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4374 4375 do_young_space_rescan(cl, to_space, NULL, 0); 4376 do_young_space_rescan(cl, from_space, sca, sct); 4377 do_young_space_rescan(cl, eden_space, eca, ect); 4378 } 4379 4380 // work_queue(i) is passed to the closure 4381 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4382 // also is passed to do_dirty_card_rescan_tasks() and to 4383 // do_work_steal() to select the i-th task_queue. 4384 4385 void CMSParRemarkTask::work(uint worker_id) { 4386 elapsedTimer _timer; 4387 ResourceMark rm; 4388 HandleMark hm; 4389 4390 // ---------- rescan from roots -------------- 4391 _timer.start(); 4392 CMSHeap* heap = CMSHeap::heap(); 4393 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4394 _collector->_span, _collector->ref_processor(), 4395 &(_collector->_markBitMap), 4396 work_queue(worker_id)); 4397 4398 // Rescan young gen roots first since these are likely 4399 // coarsely partitioned and may, on that account, constitute 4400 // the critical path; thus, it's best to start off that 4401 // work first. 4402 // ---------- young gen roots -------------- 4403 { 4404 work_on_young_gen_roots(&par_mrias_cl); 4405 _timer.stop(); 4406 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4407 } 4408 4409 // ---------- remaining roots -------------- 4410 _timer.reset(); 4411 _timer.start(); 4412 heap->cms_process_roots(_strong_roots_scope, 4413 false, // yg was scanned above 4414 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4415 _collector->should_unload_classes(), 4416 &par_mrias_cl, 4417 NULL); // The dirty klasses will be handled below 4418 4419 assert(_collector->should_unload_classes() 4420 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4421 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4422 _timer.stop(); 4423 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4424 4425 // ---------- unhandled CLD scanning ---------- 4426 if (worker_id == 0) { // Single threaded at the moment. 4427 _timer.reset(); 4428 _timer.start(); 4429 4430 // Scan all new class loader data objects and new dependencies that were 4431 // introduced during concurrent marking. 4432 ResourceMark rm; 4433 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4434 for (int i = 0; i < array->length(); i++) { 4435 par_mrias_cl.do_cld_nv(array->at(i)); 4436 } 4437 4438 // We don't need to keep track of new CLDs anymore. 4439 ClassLoaderDataGraph::remember_new_clds(false); 4440 4441 _timer.stop(); 4442 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4443 } 4444 4445 // We might have added oops to ClassLoaderData::_handles during the 4446 // concurrent marking phase. These oops do not always point to newly allocated objects 4447 // that are guaranteed to be kept alive. Hence, 4448 // we do have to revisit the _handles block during the remark phase. 4449 4450 // ---------- dirty CLD scanning ---------- 4451 if (worker_id == 0) { // Single threaded at the moment. 4452 _timer.reset(); 4453 _timer.start(); 4454 4455 // Scan all classes that was dirtied during the concurrent marking phase. 4456 RemarkCLDClosure remark_closure(&par_mrias_cl); 4457 ClassLoaderDataGraph::cld_do(&remark_closure); 4458 4459 _timer.stop(); 4460 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4461 } 4462 4463 4464 // ---------- rescan dirty cards ------------ 4465 _timer.reset(); 4466 _timer.start(); 4467 4468 // Do the rescan tasks for each of the two spaces 4469 // (cms_space) in turn. 4470 // "worker_id" is passed to select the task_queue for "worker_id" 4471 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4472 _timer.stop(); 4473 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4474 4475 // ---------- steal work from other threads ... 4476 // ---------- ... and drain overflow list. 4477 _timer.reset(); 4478 _timer.start(); 4479 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4480 _timer.stop(); 4481 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4482 } 4483 4484 void 4485 CMSParMarkTask::do_young_space_rescan( 4486 OopsInGenClosure* cl, ContiguousSpace* space, 4487 HeapWord** chunk_array, size_t chunk_top) { 4488 // Until all tasks completed: 4489 // . claim an unclaimed task 4490 // . compute region boundaries corresponding to task claimed 4491 // using chunk_array 4492 // . par_oop_iterate(cl) over that region 4493 4494 ResourceMark rm; 4495 HandleMark hm; 4496 4497 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4498 4499 uint nth_task = 0; 4500 uint n_tasks = pst->n_tasks(); 4501 4502 if (n_tasks > 0) { 4503 assert(pst->valid(), "Uninitialized use?"); 4504 HeapWord *start, *end; 4505 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4506 // We claimed task # nth_task; compute its boundaries. 4507 if (chunk_top == 0) { // no samples were taken 4508 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4509 start = space->bottom(); 4510 end = space->top(); 4511 } else if (nth_task == 0) { 4512 start = space->bottom(); 4513 end = chunk_array[nth_task]; 4514 } else if (nth_task < (uint)chunk_top) { 4515 assert(nth_task >= 1, "Control point invariant"); 4516 start = chunk_array[nth_task - 1]; 4517 end = chunk_array[nth_task]; 4518 } else { 4519 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4520 start = chunk_array[chunk_top - 1]; 4521 end = space->top(); 4522 } 4523 MemRegion mr(start, end); 4524 // Verify that mr is in space 4525 assert(mr.is_empty() || space->used_region().contains(mr), 4526 "Should be in space"); 4527 // Verify that "start" is an object boundary 4528 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4529 "Should be an oop"); 4530 space->par_oop_iterate(mr, cl); 4531 } 4532 pst->all_tasks_completed(); 4533 } 4534 } 4535 4536 void 4537 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4538 CompactibleFreeListSpace* sp, int i, 4539 ParMarkRefsIntoAndScanClosure* cl) { 4540 // Until all tasks completed: 4541 // . claim an unclaimed task 4542 // . compute region boundaries corresponding to task claimed 4543 // . transfer dirty bits ct->mut for that region 4544 // . apply rescanclosure to dirty mut bits for that region 4545 4546 ResourceMark rm; 4547 HandleMark hm; 4548 4549 OopTaskQueue* work_q = work_queue(i); 4550 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4551 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4552 // CAUTION: This closure has state that persists across calls to 4553 // the work method dirty_range_iterate_clear() in that it has 4554 // embedded in it a (subtype of) UpwardsObjectClosure. The 4555 // use of that state in the embedded UpwardsObjectClosure instance 4556 // assumes that the cards are always iterated (even if in parallel 4557 // by several threads) in monotonically increasing order per each 4558 // thread. This is true of the implementation below which picks 4559 // card ranges (chunks) in monotonically increasing order globally 4560 // and, a-fortiori, in monotonically increasing order per thread 4561 // (the latter order being a subsequence of the former). 4562 // If the work code below is ever reorganized into a more chaotic 4563 // work-partitioning form than the current "sequential tasks" 4564 // paradigm, the use of that persistent state will have to be 4565 // revisited and modified appropriately. See also related 4566 // bug 4756801 work on which should examine this code to make 4567 // sure that the changes there do not run counter to the 4568 // assumptions made here and necessary for correctness and 4569 // efficiency. Note also that this code might yield inefficient 4570 // behavior in the case of very large objects that span one or 4571 // more work chunks. Such objects would potentially be scanned 4572 // several times redundantly. Work on 4756801 should try and 4573 // address that performance anomaly if at all possible. XXX 4574 MemRegion full_span = _collector->_span; 4575 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4576 MarkFromDirtyCardsClosure 4577 greyRescanClosure(_collector, full_span, // entire span of interest 4578 sp, bm, work_q, cl); 4579 4580 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4581 assert(pst->valid(), "Uninitialized use?"); 4582 uint nth_task = 0; 4583 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4584 MemRegion span = sp->used_region(); 4585 HeapWord* start_addr = span.start(); 4586 HeapWord* end_addr = align_up(span.end(), alignment); 4587 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4588 assert(is_aligned(start_addr, alignment), "Check alignment"); 4589 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4590 4591 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4592 // Having claimed the nth_task, compute corresponding mem-region, 4593 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4594 // The alignment restriction ensures that we do not need any 4595 // synchronization with other gang-workers while setting or 4596 // clearing bits in thus chunk of the MUT. 4597 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4598 start_addr + (nth_task+1)*chunk_size); 4599 // The last chunk's end might be way beyond end of the 4600 // used region. In that case pull back appropriately. 4601 if (this_span.end() > end_addr) { 4602 this_span.set_end(end_addr); 4603 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4604 } 4605 // Iterate over the dirty cards covering this chunk, marking them 4606 // precleaned, and setting the corresponding bits in the mod union 4607 // table. Since we have been careful to partition at Card and MUT-word 4608 // boundaries no synchronization is needed between parallel threads. 4609 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4610 &modUnionClosure); 4611 4612 // Having transferred these marks into the modUnionTable, 4613 // rescan the marked objects on the dirty cards in the modUnionTable. 4614 // Even if this is at a synchronous collection, the initial marking 4615 // may have been done during an asynchronous collection so there 4616 // may be dirty bits in the mod-union table. 4617 _collector->_modUnionTable.dirty_range_iterate_clear( 4618 this_span, &greyRescanClosure); 4619 _collector->_modUnionTable.verifyNoOneBitsInRange( 4620 this_span.start(), 4621 this_span.end()); 4622 } 4623 pst->all_tasks_completed(); // declare that i am done 4624 } 4625 4626 // . see if we can share work_queues with ParNew? XXX 4627 void 4628 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4629 int* seed) { 4630 OopTaskQueue* work_q = work_queue(i); 4631 NOT_PRODUCT(int num_steals = 0;) 4632 oop obj_to_scan; 4633 CMSBitMap* bm = &(_collector->_markBitMap); 4634 4635 while (true) { 4636 // Completely finish any left over work from (an) earlier round(s) 4637 cl->trim_queue(0); 4638 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4639 (size_t)ParGCDesiredObjsFromOverflowList); 4640 // Now check if there's any work in the overflow list 4641 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4642 // only affects the number of attempts made to get work from the 4643 // overflow list and does not affect the number of workers. Just 4644 // pass ParallelGCThreads so this behavior is unchanged. 4645 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4646 work_q, 4647 ParallelGCThreads)) { 4648 // found something in global overflow list; 4649 // not yet ready to go stealing work from others. 4650 // We'd like to assert(work_q->size() != 0, ...) 4651 // because we just took work from the overflow list, 4652 // but of course we can't since all of that could have 4653 // been already stolen from us. 4654 // "He giveth and He taketh away." 4655 continue; 4656 } 4657 // Verify that we have no work before we resort to stealing 4658 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4659 // Try to steal from other queues that have work 4660 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4661 NOT_PRODUCT(num_steals++;) 4662 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4663 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4664 // Do scanning work 4665 obj_to_scan->oop_iterate(cl); 4666 // Loop around, finish this work, and try to steal some more 4667 } else if (terminator()->offer_termination()) { 4668 break; // nirvana from the infinite cycle 4669 } 4670 } 4671 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4672 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4673 "Else our work is not yet done"); 4674 } 4675 4676 // Record object boundaries in _eden_chunk_array by sampling the eden 4677 // top in the slow-path eden object allocation code path and record 4678 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4679 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4680 // sampling in sample_eden() that activates during the part of the 4681 // preclean phase. 4682 void CMSCollector::sample_eden_chunk() { 4683 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4684 if (_eden_chunk_lock->try_lock()) { 4685 // Record a sample. This is the critical section. The contents 4686 // of the _eden_chunk_array have to be non-decreasing in the 4687 // address order. 4688 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4689 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4690 "Unexpected state of Eden"); 4691 if (_eden_chunk_index == 0 || 4692 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4693 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4694 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4695 _eden_chunk_index++; // commit sample 4696 } 4697 _eden_chunk_lock->unlock(); 4698 } 4699 } 4700 } 4701 4702 // Return a thread-local PLAB recording array, as appropriate. 4703 void* CMSCollector::get_data_recorder(int thr_num) { 4704 if (_survivor_plab_array != NULL && 4705 (CMSPLABRecordAlways || 4706 (_collectorState > Marking && _collectorState < FinalMarking))) { 4707 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4708 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4709 ca->reset(); // clear it so that fresh data is recorded 4710 return (void*) ca; 4711 } else { 4712 return NULL; 4713 } 4714 } 4715 4716 // Reset all the thread-local PLAB recording arrays 4717 void CMSCollector::reset_survivor_plab_arrays() { 4718 for (uint i = 0; i < ParallelGCThreads; i++) { 4719 _survivor_plab_array[i].reset(); 4720 } 4721 } 4722 4723 // Merge the per-thread plab arrays into the global survivor chunk 4724 // array which will provide the partitioning of the survivor space 4725 // for CMS initial scan and rescan. 4726 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4727 int no_of_gc_threads) { 4728 assert(_survivor_plab_array != NULL, "Error"); 4729 assert(_survivor_chunk_array != NULL, "Error"); 4730 assert(_collectorState == FinalMarking || 4731 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4732 for (int j = 0; j < no_of_gc_threads; j++) { 4733 _cursor[j] = 0; 4734 } 4735 HeapWord* top = surv->top(); 4736 size_t i; 4737 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4738 HeapWord* min_val = top; // Higher than any PLAB address 4739 uint min_tid = 0; // position of min_val this round 4740 for (int j = 0; j < no_of_gc_threads; j++) { 4741 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4742 if (_cursor[j] == cur_sca->end()) { 4743 continue; 4744 } 4745 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4746 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4747 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4748 if (cur_val < min_val) { 4749 min_tid = j; 4750 min_val = cur_val; 4751 } else { 4752 assert(cur_val < top, "All recorded addresses should be less"); 4753 } 4754 } 4755 // At this point min_val and min_tid are respectively 4756 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4757 // and the thread (j) that witnesses that address. 4758 // We record this address in the _survivor_chunk_array[i] 4759 // and increment _cursor[min_tid] prior to the next round i. 4760 if (min_val == top) { 4761 break; 4762 } 4763 _survivor_chunk_array[i] = min_val; 4764 _cursor[min_tid]++; 4765 } 4766 // We are all done; record the size of the _survivor_chunk_array 4767 _survivor_chunk_index = i; // exclusive: [0, i) 4768 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4769 // Verify that we used up all the recorded entries 4770 #ifdef ASSERT 4771 size_t total = 0; 4772 for (int j = 0; j < no_of_gc_threads; j++) { 4773 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4774 total += _cursor[j]; 4775 } 4776 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4777 // Check that the merged array is in sorted order 4778 if (total > 0) { 4779 for (size_t i = 0; i < total - 1; i++) { 4780 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4781 i, p2i(_survivor_chunk_array[i])); 4782 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4783 "Not sorted"); 4784 } 4785 } 4786 #endif // ASSERT 4787 } 4788 4789 // Set up the space's par_seq_tasks structure for work claiming 4790 // for parallel initial scan and rescan of young gen. 4791 // See ParRescanTask where this is currently used. 4792 void 4793 CMSCollector:: 4794 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4795 assert(n_threads > 0, "Unexpected n_threads argument"); 4796 4797 // Eden space 4798 if (!_young_gen->eden()->is_empty()) { 4799 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4800 assert(!pst->valid(), "Clobbering existing data?"); 4801 // Each valid entry in [0, _eden_chunk_index) represents a task. 4802 size_t n_tasks = _eden_chunk_index + 1; 4803 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4804 // Sets the condition for completion of the subtask (how many threads 4805 // need to finish in order to be done). 4806 pst->set_n_threads(n_threads); 4807 pst->set_n_tasks((int)n_tasks); 4808 } 4809 4810 // Merge the survivor plab arrays into _survivor_chunk_array 4811 if (_survivor_plab_array != NULL) { 4812 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4813 } else { 4814 assert(_survivor_chunk_index == 0, "Error"); 4815 } 4816 4817 // To space 4818 { 4819 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4820 assert(!pst->valid(), "Clobbering existing data?"); 4821 // Sets the condition for completion of the subtask (how many threads 4822 // need to finish in order to be done). 4823 pst->set_n_threads(n_threads); 4824 pst->set_n_tasks(1); 4825 assert(pst->valid(), "Error"); 4826 } 4827 4828 // From space 4829 { 4830 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4831 assert(!pst->valid(), "Clobbering existing data?"); 4832 size_t n_tasks = _survivor_chunk_index + 1; 4833 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4834 // Sets the condition for completion of the subtask (how many threads 4835 // need to finish in order to be done). 4836 pst->set_n_threads(n_threads); 4837 pst->set_n_tasks((int)n_tasks); 4838 assert(pst->valid(), "Error"); 4839 } 4840 } 4841 4842 // Parallel version of remark 4843 void CMSCollector::do_remark_parallel() { 4844 CMSHeap* heap = CMSHeap::heap(); 4845 WorkGang* workers = heap->workers(); 4846 assert(workers != NULL, "Need parallel worker threads."); 4847 // Choose to use the number of GC workers most recently set 4848 // into "active_workers". 4849 uint n_workers = workers->active_workers(); 4850 4851 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4852 4853 StrongRootsScope srs(n_workers); 4854 4855 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4856 4857 // We won't be iterating over the cards in the card table updating 4858 // the younger_gen cards, so we shouldn't call the following else 4859 // the verification code as well as subsequent younger_refs_iterate 4860 // code would get confused. XXX 4861 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4862 4863 // The young gen rescan work will not be done as part of 4864 // process_roots (which currently doesn't know how to 4865 // parallelize such a scan), but rather will be broken up into 4866 // a set of parallel tasks (via the sampling that the [abortable] 4867 // preclean phase did of eden, plus the [two] tasks of 4868 // scanning the [two] survivor spaces. Further fine-grain 4869 // parallelization of the scanning of the survivor spaces 4870 // themselves, and of precleaning of the young gen itself 4871 // is deferred to the future. 4872 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4873 4874 // The dirty card rescan work is broken up into a "sequence" 4875 // of parallel tasks (per constituent space) that are dynamically 4876 // claimed by the parallel threads. 4877 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4878 4879 // It turns out that even when we're using 1 thread, doing the work in a 4880 // separate thread causes wide variance in run times. We can't help this 4881 // in the multi-threaded case, but we special-case n=1 here to get 4882 // repeatable measurements of the 1-thread overhead of the parallel code. 4883 if (n_workers > 1) { 4884 // Make refs discovery MT-safe, if it isn't already: it may not 4885 // necessarily be so, since it's possible that we are doing 4886 // ST marking. 4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4888 workers->run_task(&tsk); 4889 } else { 4890 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4891 tsk.work(0); 4892 } 4893 4894 // restore, single-threaded for now, any preserved marks 4895 // as a result of work_q overflow 4896 restore_preserved_marks_if_any(); 4897 } 4898 4899 // Non-parallel version of remark 4900 void CMSCollector::do_remark_non_parallel() { 4901 ResourceMark rm; 4902 HandleMark hm; 4903 CMSHeap* heap = CMSHeap::heap(); 4904 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4905 4906 MarkRefsIntoAndScanClosure 4907 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4908 &_markStack, this, 4909 false /* should_yield */, false /* not precleaning */); 4910 MarkFromDirtyCardsClosure 4911 markFromDirtyCardsClosure(this, _span, 4912 NULL, // space is set further below 4913 &_markBitMap, &_markStack, &mrias_cl); 4914 { 4915 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4916 // Iterate over the dirty cards, setting the corresponding bits in the 4917 // mod union table. 4918 { 4919 ModUnionClosure modUnionClosure(&_modUnionTable); 4920 _ct->ct_bs()->dirty_card_iterate( 4921 _cmsGen->used_region(), 4922 &modUnionClosure); 4923 } 4924 // Having transferred these marks into the modUnionTable, we just need 4925 // to rescan the marked objects on the dirty cards in the modUnionTable. 4926 // The initial marking may have been done during an asynchronous 4927 // collection so there may be dirty bits in the mod-union table. 4928 const int alignment = 4929 CardTableModRefBS::card_size * BitsPerWord; 4930 { 4931 // ... First handle dirty cards in CMS gen 4932 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4933 MemRegion ur = _cmsGen->used_region(); 4934 HeapWord* lb = ur.start(); 4935 HeapWord* ub = align_up(ur.end(), alignment); 4936 MemRegion cms_span(lb, ub); 4937 _modUnionTable.dirty_range_iterate_clear(cms_span, 4938 &markFromDirtyCardsClosure); 4939 verify_work_stacks_empty(); 4940 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4941 } 4942 } 4943 if (VerifyDuringGC && 4944 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4945 HandleMark hm; // Discard invalid handles created during verification 4946 Universe::verify(); 4947 } 4948 { 4949 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4950 4951 verify_work_stacks_empty(); 4952 4953 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4954 StrongRootsScope srs(1); 4955 4956 heap->cms_process_roots(&srs, 4957 true, // young gen as roots 4958 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4959 should_unload_classes(), 4960 &mrias_cl, 4961 NULL); // The dirty klasses will be handled below 4962 4963 assert(should_unload_classes() 4964 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4965 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4966 } 4967 4968 { 4969 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4970 4971 verify_work_stacks_empty(); 4972 4973 // Scan all class loader data objects that might have been introduced 4974 // during concurrent marking. 4975 ResourceMark rm; 4976 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4977 for (int i = 0; i < array->length(); i++) { 4978 mrias_cl.do_cld_nv(array->at(i)); 4979 } 4980 4981 // We don't need to keep track of new CLDs anymore. 4982 ClassLoaderDataGraph::remember_new_clds(false); 4983 4984 verify_work_stacks_empty(); 4985 } 4986 4987 // We might have added oops to ClassLoaderData::_handles during the 4988 // concurrent marking phase. These oops do not point to newly allocated objects 4989 // that are guaranteed to be kept alive. Hence, 4990 // we do have to revisit the _handles block during the remark phase. 4991 { 4992 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4993 4994 verify_work_stacks_empty(); 4995 4996 RemarkCLDClosure remark_closure(&mrias_cl); 4997 ClassLoaderDataGraph::cld_do(&remark_closure); 4998 4999 verify_work_stacks_empty(); 5000 } 5001 5002 verify_work_stacks_empty(); 5003 // Restore evacuated mark words, if any, used for overflow list links 5004 restore_preserved_marks_if_any(); 5005 5006 verify_overflow_empty(); 5007 } 5008 5009 //////////////////////////////////////////////////////// 5010 // Parallel Reference Processing Task Proxy Class 5011 //////////////////////////////////////////////////////// 5012 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5013 OopTaskQueueSet* _queues; 5014 ParallelTaskTerminator _terminator; 5015 public: 5016 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5017 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5018 ParallelTaskTerminator* terminator() { return &_terminator; } 5019 OopTaskQueueSet* queues() { return _queues; } 5020 }; 5021 5022 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5023 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5024 CMSCollector* _collector; 5025 CMSBitMap* _mark_bit_map; 5026 const MemRegion _span; 5027 ProcessTask& _task; 5028 5029 public: 5030 CMSRefProcTaskProxy(ProcessTask& task, 5031 CMSCollector* collector, 5032 const MemRegion& span, 5033 CMSBitMap* mark_bit_map, 5034 AbstractWorkGang* workers, 5035 OopTaskQueueSet* task_queues): 5036 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5037 task_queues, 5038 workers->active_workers()), 5039 _task(task), 5040 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5041 { 5042 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5043 "Inconsistency in _span"); 5044 } 5045 5046 OopTaskQueueSet* task_queues() { return queues(); } 5047 5048 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5049 5050 void do_work_steal(int i, 5051 CMSParDrainMarkingStackClosure* drain, 5052 CMSParKeepAliveClosure* keep_alive, 5053 int* seed); 5054 5055 virtual void work(uint worker_id); 5056 }; 5057 5058 void CMSRefProcTaskProxy::work(uint worker_id) { 5059 ResourceMark rm; 5060 HandleMark hm; 5061 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5062 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5063 _mark_bit_map, 5064 work_queue(worker_id)); 5065 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5066 _mark_bit_map, 5067 work_queue(worker_id)); 5068 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5069 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5070 if (_task.marks_oops_alive()) { 5071 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5072 _collector->hash_seed(worker_id)); 5073 } 5074 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5075 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5076 } 5077 5078 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5079 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5080 EnqueueTask& _task; 5081 5082 public: 5083 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5084 : AbstractGangTask("Enqueue reference objects in parallel"), 5085 _task(task) 5086 { } 5087 5088 virtual void work(uint worker_id) 5089 { 5090 _task.work(worker_id); 5091 } 5092 }; 5093 5094 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5095 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5096 _span(span), 5097 _bit_map(bit_map), 5098 _work_queue(work_queue), 5099 _mark_and_push(collector, span, bit_map, work_queue), 5100 _low_water_mark(MIN2((work_queue->max_elems()/4), 5101 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5102 { } 5103 5104 // . see if we can share work_queues with ParNew? XXX 5105 void CMSRefProcTaskProxy::do_work_steal(int i, 5106 CMSParDrainMarkingStackClosure* drain, 5107 CMSParKeepAliveClosure* keep_alive, 5108 int* seed) { 5109 OopTaskQueue* work_q = work_queue(i); 5110 NOT_PRODUCT(int num_steals = 0;) 5111 oop obj_to_scan; 5112 5113 while (true) { 5114 // Completely finish any left over work from (an) earlier round(s) 5115 drain->trim_queue(0); 5116 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5117 (size_t)ParGCDesiredObjsFromOverflowList); 5118 // Now check if there's any work in the overflow list 5119 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5120 // only affects the number of attempts made to get work from the 5121 // overflow list and does not affect the number of workers. Just 5122 // pass ParallelGCThreads so this behavior is unchanged. 5123 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5124 work_q, 5125 ParallelGCThreads)) { 5126 // Found something in global overflow list; 5127 // not yet ready to go stealing work from others. 5128 // We'd like to assert(work_q->size() != 0, ...) 5129 // because we just took work from the overflow list, 5130 // but of course we can't, since all of that might have 5131 // been already stolen from us. 5132 continue; 5133 } 5134 // Verify that we have no work before we resort to stealing 5135 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5136 // Try to steal from other queues that have work 5137 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5138 NOT_PRODUCT(num_steals++;) 5139 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5140 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5141 // Do scanning work 5142 obj_to_scan->oop_iterate(keep_alive); 5143 // Loop around, finish this work, and try to steal some more 5144 } else if (terminator()->offer_termination()) { 5145 break; // nirvana from the infinite cycle 5146 } 5147 } 5148 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5149 } 5150 5151 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5152 { 5153 CMSHeap* heap = CMSHeap::heap(); 5154 WorkGang* workers = heap->workers(); 5155 assert(workers != NULL, "Need parallel worker threads."); 5156 CMSRefProcTaskProxy rp_task(task, &_collector, 5157 _collector.ref_processor()->span(), 5158 _collector.markBitMap(), 5159 workers, _collector.task_queues()); 5160 workers->run_task(&rp_task); 5161 } 5162 5163 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5164 { 5165 5166 CMSHeap* heap = CMSHeap::heap(); 5167 WorkGang* workers = heap->workers(); 5168 assert(workers != NULL, "Need parallel worker threads."); 5169 CMSRefEnqueueTaskProxy enq_task(task); 5170 workers->run_task(&enq_task); 5171 } 5172 5173 void CMSCollector::refProcessingWork() { 5174 ResourceMark rm; 5175 HandleMark hm; 5176 5177 ReferenceProcessor* rp = ref_processor(); 5178 assert(rp->span().equals(_span), "Spans should be equal"); 5179 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5180 // Process weak references. 5181 rp->setup_policy(false); 5182 verify_work_stacks_empty(); 5183 5184 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q()); 5185 { 5186 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5187 5188 // Setup keep_alive and complete closures. 5189 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5190 &_markStack, false /* !preclean */); 5191 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5192 _span, &_markBitMap, &_markStack, 5193 &cmsKeepAliveClosure, false /* !preclean */); 5194 5195 ReferenceProcessorStats stats; 5196 if (rp->processing_is_mt()) { 5197 // Set the degree of MT here. If the discovery is done MT, there 5198 // may have been a different number of threads doing the discovery 5199 // and a different number of discovered lists may have Ref objects. 5200 // That is OK as long as the Reference lists are balanced (see 5201 // balance_all_queues() and balance_queues()). 5202 CMSHeap* heap = CMSHeap::heap(); 5203 uint active_workers = ParallelGCThreads; 5204 WorkGang* workers = heap->workers(); 5205 if (workers != NULL) { 5206 active_workers = workers->active_workers(); 5207 // The expectation is that active_workers will have already 5208 // been set to a reasonable value. If it has not been set, 5209 // investigate. 5210 assert(active_workers > 0, "Should have been set during scavenge"); 5211 } 5212 rp->set_active_mt_degree(active_workers); 5213 CMSRefProcTaskExecutor task_executor(*this); 5214 stats = rp->process_discovered_references(&_is_alive_closure, 5215 &cmsKeepAliveClosure, 5216 &cmsDrainMarkingStackClosure, 5217 &task_executor, 5218 &pt); 5219 } else { 5220 stats = rp->process_discovered_references(&_is_alive_closure, 5221 &cmsKeepAliveClosure, 5222 &cmsDrainMarkingStackClosure, 5223 NULL, 5224 &pt); 5225 } 5226 _gc_tracer_cm->report_gc_reference_stats(stats); 5227 pt.print_all_references(); 5228 } 5229 5230 // This is the point where the entire marking should have completed. 5231 verify_work_stacks_empty(); 5232 5233 { 5234 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5235 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5236 } 5237 5238 if (should_unload_classes()) { 5239 { 5240 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5241 5242 // Unload classes and purge the SystemDictionary. 5243 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5244 5245 // Unload nmethods. 5246 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5247 5248 // Prune dead klasses from subklass/sibling/implementor lists. 5249 Klass::clean_weak_klass_links(&_is_alive_closure); 5250 } 5251 5252 { 5253 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5254 // Clean up unreferenced symbols in symbol table. 5255 SymbolTable::unlink(); 5256 } 5257 5258 { 5259 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5260 // Delete entries for dead interned strings. 5261 StringTable::unlink(&_is_alive_closure); 5262 } 5263 } 5264 5265 // Restore any preserved marks as a result of mark stack or 5266 // work queue overflow 5267 restore_preserved_marks_if_any(); // done single-threaded for now 5268 5269 rp->set_enqueuing_is_done(true); 5270 if (rp->processing_is_mt()) { 5271 rp->balance_all_queues(); 5272 CMSRefProcTaskExecutor task_executor(*this); 5273 rp->enqueue_discovered_references(&task_executor, &pt); 5274 } else { 5275 rp->enqueue_discovered_references(NULL, &pt); 5276 } 5277 rp->verify_no_references_recorded(); 5278 pt.print_enqueue_phase(); 5279 assert(!rp->discovery_enabled(), "should have been disabled"); 5280 } 5281 5282 #ifndef PRODUCT 5283 void CMSCollector::check_correct_thread_executing() { 5284 Thread* t = Thread::current(); 5285 // Only the VM thread or the CMS thread should be here. 5286 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5287 "Unexpected thread type"); 5288 // If this is the vm thread, the foreground process 5289 // should not be waiting. Note that _foregroundGCIsActive is 5290 // true while the foreground collector is waiting. 5291 if (_foregroundGCShouldWait) { 5292 // We cannot be the VM thread 5293 assert(t->is_ConcurrentGC_thread(), 5294 "Should be CMS thread"); 5295 } else { 5296 // We can be the CMS thread only if we are in a stop-world 5297 // phase of CMS collection. 5298 if (t->is_ConcurrentGC_thread()) { 5299 assert(_collectorState == InitialMarking || 5300 _collectorState == FinalMarking, 5301 "Should be a stop-world phase"); 5302 // The CMS thread should be holding the CMS_token. 5303 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5304 "Potential interference with concurrently " 5305 "executing VM thread"); 5306 } 5307 } 5308 } 5309 #endif 5310 5311 void CMSCollector::sweep() { 5312 assert(_collectorState == Sweeping, "just checking"); 5313 check_correct_thread_executing(); 5314 verify_work_stacks_empty(); 5315 verify_overflow_empty(); 5316 increment_sweep_count(); 5317 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5318 5319 _inter_sweep_timer.stop(); 5320 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5321 5322 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5323 _intra_sweep_timer.reset(); 5324 _intra_sweep_timer.start(); 5325 { 5326 GCTraceCPUTime tcpu; 5327 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5328 // First sweep the old gen 5329 { 5330 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5331 bitMapLock()); 5332 sweepWork(_cmsGen); 5333 } 5334 5335 // Update Universe::_heap_*_at_gc figures. 5336 // We need all the free list locks to make the abstract state 5337 // transition from Sweeping to Resetting. See detailed note 5338 // further below. 5339 { 5340 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5341 // Update heap occupancy information which is used as 5342 // input to soft ref clearing policy at the next gc. 5343 Universe::update_heap_info_at_gc(); 5344 _collectorState = Resizing; 5345 } 5346 } 5347 verify_work_stacks_empty(); 5348 verify_overflow_empty(); 5349 5350 if (should_unload_classes()) { 5351 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5352 // requires that the virtual spaces are stable and not deleted. 5353 ClassLoaderDataGraph::set_should_purge(true); 5354 } 5355 5356 _intra_sweep_timer.stop(); 5357 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5358 5359 _inter_sweep_timer.reset(); 5360 _inter_sweep_timer.start(); 5361 5362 // We need to use a monotonically non-decreasing time in ms 5363 // or we will see time-warp warnings and os::javaTimeMillis() 5364 // does not guarantee monotonicity. 5365 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5366 update_time_of_last_gc(now); 5367 5368 // NOTE on abstract state transitions: 5369 // Mutators allocate-live and/or mark the mod-union table dirty 5370 // based on the state of the collection. The former is done in 5371 // the interval [Marking, Sweeping] and the latter in the interval 5372 // [Marking, Sweeping). Thus the transitions into the Marking state 5373 // and out of the Sweeping state must be synchronously visible 5374 // globally to the mutators. 5375 // The transition into the Marking state happens with the world 5376 // stopped so the mutators will globally see it. Sweeping is 5377 // done asynchronously by the background collector so the transition 5378 // from the Sweeping state to the Resizing state must be done 5379 // under the freelistLock (as is the check for whether to 5380 // allocate-live and whether to dirty the mod-union table). 5381 assert(_collectorState == Resizing, "Change of collector state to" 5382 " Resizing must be done under the freelistLocks (plural)"); 5383 5384 // Now that sweeping has been completed, we clear 5385 // the incremental_collection_failed flag, 5386 // thus inviting a younger gen collection to promote into 5387 // this generation. If such a promotion may still fail, 5388 // the flag will be set again when a young collection is 5389 // attempted. 5390 CMSHeap* heap = CMSHeap::heap(); 5391 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5392 heap->update_full_collections_completed(_collection_count_start); 5393 } 5394 5395 // FIX ME!!! Looks like this belongs in CFLSpace, with 5396 // CMSGen merely delegating to it. 5397 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5398 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5399 HeapWord* minAddr = _cmsSpace->bottom(); 5400 HeapWord* largestAddr = 5401 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5402 if (largestAddr == NULL) { 5403 // The dictionary appears to be empty. In this case 5404 // try to coalesce at the end of the heap. 5405 largestAddr = _cmsSpace->end(); 5406 } 5407 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5408 size_t nearLargestOffset = 5409 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5410 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5411 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5412 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5413 } 5414 5415 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5416 return addr >= _cmsSpace->nearLargestChunk(); 5417 } 5418 5419 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5420 return _cmsSpace->find_chunk_at_end(); 5421 } 5422 5423 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5424 bool full) { 5425 // If the young generation has been collected, gather any statistics 5426 // that are of interest at this point. 5427 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5428 if (!full && current_is_young) { 5429 // Gather statistics on the young generation collection. 5430 collector()->stats().record_gc0_end(used()); 5431 } 5432 } 5433 5434 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5435 // We iterate over the space(s) underlying this generation, 5436 // checking the mark bit map to see if the bits corresponding 5437 // to specific blocks are marked or not. Blocks that are 5438 // marked are live and are not swept up. All remaining blocks 5439 // are swept up, with coalescing on-the-fly as we sweep up 5440 // contiguous free and/or garbage blocks: 5441 // We need to ensure that the sweeper synchronizes with allocators 5442 // and stop-the-world collectors. In particular, the following 5443 // locks are used: 5444 // . CMS token: if this is held, a stop the world collection cannot occur 5445 // . freelistLock: if this is held no allocation can occur from this 5446 // generation by another thread 5447 // . bitMapLock: if this is held, no other thread can access or update 5448 // 5449 5450 // Note that we need to hold the freelistLock if we use 5451 // block iterate below; else the iterator might go awry if 5452 // a mutator (or promotion) causes block contents to change 5453 // (for instance if the allocator divvies up a block). 5454 // If we hold the free list lock, for all practical purposes 5455 // young generation GC's can't occur (they'll usually need to 5456 // promote), so we might as well prevent all young generation 5457 // GC's while we do a sweeping step. For the same reason, we might 5458 // as well take the bit map lock for the entire duration 5459 5460 // check that we hold the requisite locks 5461 assert(have_cms_token(), "Should hold cms token"); 5462 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5463 assert_lock_strong(old_gen->freelistLock()); 5464 assert_lock_strong(bitMapLock()); 5465 5466 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5467 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5468 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5469 _inter_sweep_estimate.padded_average(), 5470 _intra_sweep_estimate.padded_average()); 5471 old_gen->setNearLargestChunk(); 5472 5473 { 5474 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5475 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5476 // We need to free-up/coalesce garbage/blocks from a 5477 // co-terminal free run. This is done in the SweepClosure 5478 // destructor; so, do not remove this scope, else the 5479 // end-of-sweep-census below will be off by a little bit. 5480 } 5481 old_gen->cmsSpace()->sweep_completed(); 5482 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5483 if (should_unload_classes()) { // unloaded classes this cycle, 5484 _concurrent_cycles_since_last_unload = 0; // ... reset count 5485 } else { // did not unload classes, 5486 _concurrent_cycles_since_last_unload++; // ... increment count 5487 } 5488 } 5489 5490 // Reset CMS data structures (for now just the marking bit map) 5491 // preparatory for the next cycle. 5492 void CMSCollector::reset_concurrent() { 5493 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5494 5495 // If the state is not "Resetting", the foreground thread 5496 // has done a collection and the resetting. 5497 if (_collectorState != Resetting) { 5498 assert(_collectorState == Idling, "The state should only change" 5499 " because the foreground collector has finished the collection"); 5500 return; 5501 } 5502 5503 { 5504 // Clear the mark bitmap (no grey objects to start with) 5505 // for the next cycle. 5506 GCTraceCPUTime tcpu; 5507 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5508 5509 HeapWord* curAddr = _markBitMap.startWord(); 5510 while (curAddr < _markBitMap.endWord()) { 5511 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5512 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5513 _markBitMap.clear_large_range(chunk); 5514 if (ConcurrentMarkSweepThread::should_yield() && 5515 !foregroundGCIsActive() && 5516 CMSYield) { 5517 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5518 "CMS thread should hold CMS token"); 5519 assert_lock_strong(bitMapLock()); 5520 bitMapLock()->unlock(); 5521 ConcurrentMarkSweepThread::desynchronize(true); 5522 stopTimer(); 5523 incrementYields(); 5524 5525 // See the comment in coordinator_yield() 5526 for (unsigned i = 0; i < CMSYieldSleepCount && 5527 ConcurrentMarkSweepThread::should_yield() && 5528 !CMSCollector::foregroundGCIsActive(); ++i) { 5529 os::sleep(Thread::current(), 1, false); 5530 } 5531 5532 ConcurrentMarkSweepThread::synchronize(true); 5533 bitMapLock()->lock_without_safepoint_check(); 5534 startTimer(); 5535 } 5536 curAddr = chunk.end(); 5537 } 5538 // A successful mostly concurrent collection has been done. 5539 // Because only the full (i.e., concurrent mode failure) collections 5540 // are being measured for gc overhead limits, clean the "near" flag 5541 // and count. 5542 size_policy()->reset_gc_overhead_limit_count(); 5543 _collectorState = Idling; 5544 } 5545 5546 register_gc_end(); 5547 } 5548 5549 // Same as above but for STW paths 5550 void CMSCollector::reset_stw() { 5551 // already have the lock 5552 assert(_collectorState == Resetting, "just checking"); 5553 assert_lock_strong(bitMapLock()); 5554 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5555 _markBitMap.clear_all(); 5556 _collectorState = Idling; 5557 register_gc_end(); 5558 } 5559 5560 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5561 GCTraceCPUTime tcpu; 5562 TraceCollectorStats tcs(counters()); 5563 5564 switch (op) { 5565 case CMS_op_checkpointRootsInitial: { 5566 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5567 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5568 checkpointRootsInitial(); 5569 break; 5570 } 5571 case CMS_op_checkpointRootsFinal: { 5572 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5573 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5574 checkpointRootsFinal(); 5575 break; 5576 } 5577 default: 5578 fatal("No such CMS_op"); 5579 } 5580 } 5581 5582 #ifndef PRODUCT 5583 size_t const CMSCollector::skip_header_HeapWords() { 5584 return FreeChunk::header_size(); 5585 } 5586 5587 // Try and collect here conditions that should hold when 5588 // CMS thread is exiting. The idea is that the foreground GC 5589 // thread should not be blocked if it wants to terminate 5590 // the CMS thread and yet continue to run the VM for a while 5591 // after that. 5592 void CMSCollector::verify_ok_to_terminate() const { 5593 assert(Thread::current()->is_ConcurrentGC_thread(), 5594 "should be called by CMS thread"); 5595 assert(!_foregroundGCShouldWait, "should be false"); 5596 // We could check here that all the various low-level locks 5597 // are not held by the CMS thread, but that is overkill; see 5598 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5599 // is checked. 5600 } 5601 #endif 5602 5603 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5604 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5605 "missing Printezis mark?"); 5606 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5607 size_t size = pointer_delta(nextOneAddr + 1, addr); 5608 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5609 "alignment problem"); 5610 assert(size >= 3, "Necessary for Printezis marks to work"); 5611 return size; 5612 } 5613 5614 // A variant of the above (block_size_using_printezis_bits()) except 5615 // that we return 0 if the P-bits are not yet set. 5616 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5617 if (_markBitMap.isMarked(addr + 1)) { 5618 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5619 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5620 size_t size = pointer_delta(nextOneAddr + 1, addr); 5621 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5622 "alignment problem"); 5623 assert(size >= 3, "Necessary for Printezis marks to work"); 5624 return size; 5625 } 5626 return 0; 5627 } 5628 5629 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5630 size_t sz = 0; 5631 oop p = (oop)addr; 5632 if (p->klass_or_null_acquire() != NULL) { 5633 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5634 } else { 5635 sz = block_size_using_printezis_bits(addr); 5636 } 5637 assert(sz > 0, "size must be nonzero"); 5638 HeapWord* next_block = addr + sz; 5639 HeapWord* next_card = align_up(next_block, CardTableModRefBS::card_size); 5640 assert(align_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5641 align_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5642 "must be different cards"); 5643 return next_card; 5644 } 5645 5646 5647 // CMS Bit Map Wrapper ///////////////////////////////////////// 5648 5649 // Construct a CMS bit map infrastructure, but don't create the 5650 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5651 // further below. 5652 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5653 _bm(), 5654 _shifter(shifter), 5655 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5656 Monitor::_safepoint_check_sometimes) : NULL) 5657 { 5658 _bmStartWord = 0; 5659 _bmWordSize = 0; 5660 } 5661 5662 bool CMSBitMap::allocate(MemRegion mr) { 5663 _bmStartWord = mr.start(); 5664 _bmWordSize = mr.word_size(); 5665 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5666 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5667 if (!brs.is_reserved()) { 5668 log_warning(gc)("CMS bit map allocation failure"); 5669 return false; 5670 } 5671 // For now we'll just commit all of the bit map up front. 5672 // Later on we'll try to be more parsimonious with swap. 5673 if (!_virtual_space.initialize(brs, brs.size())) { 5674 log_warning(gc)("CMS bit map backing store failure"); 5675 return false; 5676 } 5677 assert(_virtual_space.committed_size() == brs.size(), 5678 "didn't reserve backing store for all of CMS bit map?"); 5679 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5680 _bmWordSize, "inconsistency in bit map sizing"); 5681 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5682 5683 // bm.clear(); // can we rely on getting zero'd memory? verify below 5684 assert(isAllClear(), 5685 "Expected zero'd memory from ReservedSpace constructor"); 5686 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5687 "consistency check"); 5688 return true; 5689 } 5690 5691 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5692 HeapWord *next_addr, *end_addr, *last_addr; 5693 assert_locked(); 5694 assert(covers(mr), "out-of-range error"); 5695 // XXX assert that start and end are appropriately aligned 5696 for (next_addr = mr.start(), end_addr = mr.end(); 5697 next_addr < end_addr; next_addr = last_addr) { 5698 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5699 last_addr = dirty_region.end(); 5700 if (!dirty_region.is_empty()) { 5701 cl->do_MemRegion(dirty_region); 5702 } else { 5703 assert(last_addr == end_addr, "program logic"); 5704 return; 5705 } 5706 } 5707 } 5708 5709 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5710 _bm.print_on_error(st, prefix); 5711 } 5712 5713 #ifndef PRODUCT 5714 void CMSBitMap::assert_locked() const { 5715 CMSLockVerifier::assert_locked(lock()); 5716 } 5717 5718 bool CMSBitMap::covers(MemRegion mr) const { 5719 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5720 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5721 "size inconsistency"); 5722 return (mr.start() >= _bmStartWord) && 5723 (mr.end() <= endWord()); 5724 } 5725 5726 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5727 return (start >= _bmStartWord && (start + size) <= endWord()); 5728 } 5729 5730 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5731 // verify that there are no 1 bits in the interval [left, right) 5732 FalseBitMapClosure falseBitMapClosure; 5733 iterate(&falseBitMapClosure, left, right); 5734 } 5735 5736 void CMSBitMap::region_invariant(MemRegion mr) 5737 { 5738 assert_locked(); 5739 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5740 assert(!mr.is_empty(), "unexpected empty region"); 5741 assert(covers(mr), "mr should be covered by bit map"); 5742 // convert address range into offset range 5743 size_t start_ofs = heapWordToOffset(mr.start()); 5744 // Make sure that end() is appropriately aligned 5745 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5746 "Misaligned mr.end()"); 5747 size_t end_ofs = heapWordToOffset(mr.end()); 5748 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5749 } 5750 5751 #endif 5752 5753 bool CMSMarkStack::allocate(size_t size) { 5754 // allocate a stack of the requisite depth 5755 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5756 size * sizeof(oop))); 5757 if (!rs.is_reserved()) { 5758 log_warning(gc)("CMSMarkStack allocation failure"); 5759 return false; 5760 } 5761 if (!_virtual_space.initialize(rs, rs.size())) { 5762 log_warning(gc)("CMSMarkStack backing store failure"); 5763 return false; 5764 } 5765 assert(_virtual_space.committed_size() == rs.size(), 5766 "didn't reserve backing store for all of CMS stack?"); 5767 _base = (oop*)(_virtual_space.low()); 5768 _index = 0; 5769 _capacity = size; 5770 NOT_PRODUCT(_max_depth = 0); 5771 return true; 5772 } 5773 5774 // XXX FIX ME !!! In the MT case we come in here holding a 5775 // leaf lock. For printing we need to take a further lock 5776 // which has lower rank. We need to recalibrate the two 5777 // lock-ranks involved in order to be able to print the 5778 // messages below. (Or defer the printing to the caller. 5779 // For now we take the expedient path of just disabling the 5780 // messages for the problematic case.) 5781 void CMSMarkStack::expand() { 5782 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5783 if (_capacity == MarkStackSizeMax) { 5784 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5785 // We print a warning message only once per CMS cycle. 5786 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5787 } 5788 return; 5789 } 5790 // Double capacity if possible 5791 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5792 // Do not give up existing stack until we have managed to 5793 // get the double capacity that we desired. 5794 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5795 new_capacity * sizeof(oop))); 5796 if (rs.is_reserved()) { 5797 // Release the backing store associated with old stack 5798 _virtual_space.release(); 5799 // Reinitialize virtual space for new stack 5800 if (!_virtual_space.initialize(rs, rs.size())) { 5801 fatal("Not enough swap for expanded marking stack"); 5802 } 5803 _base = (oop*)(_virtual_space.low()); 5804 _index = 0; 5805 _capacity = new_capacity; 5806 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5807 // Failed to double capacity, continue; 5808 // we print a detail message only once per CMS cycle. 5809 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5810 _capacity / K, new_capacity / K); 5811 } 5812 } 5813 5814 5815 // Closures 5816 // XXX: there seems to be a lot of code duplication here; 5817 // should refactor and consolidate common code. 5818 5819 // This closure is used to mark refs into the CMS generation in 5820 // the CMS bit map. Called at the first checkpoint. This closure 5821 // assumes that we do not need to re-mark dirty cards; if the CMS 5822 // generation on which this is used is not an oldest 5823 // generation then this will lose younger_gen cards! 5824 5825 MarkRefsIntoClosure::MarkRefsIntoClosure( 5826 MemRegion span, CMSBitMap* bitMap): 5827 _span(span), 5828 _bitMap(bitMap) 5829 { 5830 assert(ref_processor() == NULL, "deliberately left NULL"); 5831 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5832 } 5833 5834 void MarkRefsIntoClosure::do_oop(oop obj) { 5835 // if p points into _span, then mark corresponding bit in _markBitMap 5836 assert(oopDesc::is_oop(obj), "expected an oop"); 5837 HeapWord* addr = (HeapWord*)obj; 5838 if (_span.contains(addr)) { 5839 // this should be made more efficient 5840 _bitMap->mark(addr); 5841 } 5842 } 5843 5844 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5845 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5846 5847 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5848 MemRegion span, CMSBitMap* bitMap): 5849 _span(span), 5850 _bitMap(bitMap) 5851 { 5852 assert(ref_processor() == NULL, "deliberately left NULL"); 5853 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5854 } 5855 5856 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5857 // if p points into _span, then mark corresponding bit in _markBitMap 5858 assert(oopDesc::is_oop(obj), "expected an oop"); 5859 HeapWord* addr = (HeapWord*)obj; 5860 if (_span.contains(addr)) { 5861 // this should be made more efficient 5862 _bitMap->par_mark(addr); 5863 } 5864 } 5865 5866 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5867 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5868 5869 // A variant of the above, used for CMS marking verification. 5870 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5871 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5872 _span(span), 5873 _verification_bm(verification_bm), 5874 _cms_bm(cms_bm) 5875 { 5876 assert(ref_processor() == NULL, "deliberately left NULL"); 5877 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5878 } 5879 5880 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5881 // if p points into _span, then mark corresponding bit in _markBitMap 5882 assert(oopDesc::is_oop(obj), "expected an oop"); 5883 HeapWord* addr = (HeapWord*)obj; 5884 if (_span.contains(addr)) { 5885 _verification_bm->mark(addr); 5886 if (!_cms_bm->isMarked(addr)) { 5887 Log(gc, verify) log; 5888 ResourceMark rm; 5889 LogStream ls(log.error()); 5890 oop(addr)->print_on(&ls); 5891 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5892 fatal("... aborting"); 5893 } 5894 } 5895 } 5896 5897 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5898 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5899 5900 ////////////////////////////////////////////////// 5901 // MarkRefsIntoAndScanClosure 5902 ////////////////////////////////////////////////// 5903 5904 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5905 ReferenceProcessor* rp, 5906 CMSBitMap* bit_map, 5907 CMSBitMap* mod_union_table, 5908 CMSMarkStack* mark_stack, 5909 CMSCollector* collector, 5910 bool should_yield, 5911 bool concurrent_precleaning): 5912 _collector(collector), 5913 _span(span), 5914 _bit_map(bit_map), 5915 _mark_stack(mark_stack), 5916 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5917 mark_stack, concurrent_precleaning), 5918 _yield(should_yield), 5919 _concurrent_precleaning(concurrent_precleaning), 5920 _freelistLock(NULL) 5921 { 5922 // FIXME: Should initialize in base class constructor. 5923 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5924 set_ref_processor_internal(rp); 5925 } 5926 5927 // This closure is used to mark refs into the CMS generation at the 5928 // second (final) checkpoint, and to scan and transitively follow 5929 // the unmarked oops. It is also used during the concurrent precleaning 5930 // phase while scanning objects on dirty cards in the CMS generation. 5931 // The marks are made in the marking bit map and the marking stack is 5932 // used for keeping the (newly) grey objects during the scan. 5933 // The parallel version (Par_...) appears further below. 5934 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5935 if (obj != NULL) { 5936 assert(oopDesc::is_oop(obj), "expected an oop"); 5937 HeapWord* addr = (HeapWord*)obj; 5938 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5939 assert(_collector->overflow_list_is_empty(), 5940 "overflow list should be empty"); 5941 if (_span.contains(addr) && 5942 !_bit_map->isMarked(addr)) { 5943 // mark bit map (object is now grey) 5944 _bit_map->mark(addr); 5945 // push on marking stack (stack should be empty), and drain the 5946 // stack by applying this closure to the oops in the oops popped 5947 // from the stack (i.e. blacken the grey objects) 5948 bool res = _mark_stack->push(obj); 5949 assert(res, "Should have space to push on empty stack"); 5950 do { 5951 oop new_oop = _mark_stack->pop(); 5952 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5953 assert(_bit_map->isMarked((HeapWord*)new_oop), 5954 "only grey objects on this stack"); 5955 // iterate over the oops in this oop, marking and pushing 5956 // the ones in CMS heap (i.e. in _span). 5957 new_oop->oop_iterate(&_pushAndMarkClosure); 5958 // check if it's time to yield 5959 do_yield_check(); 5960 } while (!_mark_stack->isEmpty() || 5961 (!_concurrent_precleaning && take_from_overflow_list())); 5962 // if marking stack is empty, and we are not doing this 5963 // during precleaning, then check the overflow list 5964 } 5965 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5966 assert(_collector->overflow_list_is_empty(), 5967 "overflow list was drained above"); 5968 5969 assert(_collector->no_preserved_marks(), 5970 "All preserved marks should have been restored above"); 5971 } 5972 } 5973 5974 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5975 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5976 5977 void MarkRefsIntoAndScanClosure::do_yield_work() { 5978 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5979 "CMS thread should hold CMS token"); 5980 assert_lock_strong(_freelistLock); 5981 assert_lock_strong(_bit_map->lock()); 5982 // relinquish the free_list_lock and bitMaplock() 5983 _bit_map->lock()->unlock(); 5984 _freelistLock->unlock(); 5985 ConcurrentMarkSweepThread::desynchronize(true); 5986 _collector->stopTimer(); 5987 _collector->incrementYields(); 5988 5989 // See the comment in coordinator_yield() 5990 for (unsigned i = 0; 5991 i < CMSYieldSleepCount && 5992 ConcurrentMarkSweepThread::should_yield() && 5993 !CMSCollector::foregroundGCIsActive(); 5994 ++i) { 5995 os::sleep(Thread::current(), 1, false); 5996 } 5997 5998 ConcurrentMarkSweepThread::synchronize(true); 5999 _freelistLock->lock_without_safepoint_check(); 6000 _bit_map->lock()->lock_without_safepoint_check(); 6001 _collector->startTimer(); 6002 } 6003 6004 /////////////////////////////////////////////////////////// 6005 // ParMarkRefsIntoAndScanClosure: a parallel version of 6006 // MarkRefsIntoAndScanClosure 6007 /////////////////////////////////////////////////////////// 6008 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6009 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6010 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6011 _span(span), 6012 _bit_map(bit_map), 6013 _work_queue(work_queue), 6014 _low_water_mark(MIN2((work_queue->max_elems()/4), 6015 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6016 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6017 { 6018 // FIXME: Should initialize in base class constructor. 6019 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6020 set_ref_processor_internal(rp); 6021 } 6022 6023 // This closure is used to mark refs into the CMS generation at the 6024 // second (final) checkpoint, and to scan and transitively follow 6025 // the unmarked oops. The marks are made in the marking bit map and 6026 // the work_queue is used for keeping the (newly) grey objects during 6027 // the scan phase whence they are also available for stealing by parallel 6028 // threads. Since the marking bit map is shared, updates are 6029 // synchronized (via CAS). 6030 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6031 if (obj != NULL) { 6032 // Ignore mark word because this could be an already marked oop 6033 // that may be chained at the end of the overflow list. 6034 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6035 HeapWord* addr = (HeapWord*)obj; 6036 if (_span.contains(addr) && 6037 !_bit_map->isMarked(addr)) { 6038 // mark bit map (object will become grey): 6039 // It is possible for several threads to be 6040 // trying to "claim" this object concurrently; 6041 // the unique thread that succeeds in marking the 6042 // object first will do the subsequent push on 6043 // to the work queue (or overflow list). 6044 if (_bit_map->par_mark(addr)) { 6045 // push on work_queue (which may not be empty), and trim the 6046 // queue to an appropriate length by applying this closure to 6047 // the oops in the oops popped from the stack (i.e. blacken the 6048 // grey objects) 6049 bool res = _work_queue->push(obj); 6050 assert(res, "Low water mark should be less than capacity?"); 6051 trim_queue(_low_water_mark); 6052 } // Else, another thread claimed the object 6053 } 6054 } 6055 } 6056 6057 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6058 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6059 6060 // This closure is used to rescan the marked objects on the dirty cards 6061 // in the mod union table and the card table proper. 6062 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6063 oop p, MemRegion mr) { 6064 6065 size_t size = 0; 6066 HeapWord* addr = (HeapWord*)p; 6067 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6068 assert(_span.contains(addr), "we are scanning the CMS generation"); 6069 // check if it's time to yield 6070 if (do_yield_check()) { 6071 // We yielded for some foreground stop-world work, 6072 // and we have been asked to abort this ongoing preclean cycle. 6073 return 0; 6074 } 6075 if (_bitMap->isMarked(addr)) { 6076 // it's marked; is it potentially uninitialized? 6077 if (p->klass_or_null_acquire() != NULL) { 6078 // an initialized object; ignore mark word in verification below 6079 // since we are running concurrent with mutators 6080 assert(oopDesc::is_oop(p, true), "should be an oop"); 6081 if (p->is_objArray()) { 6082 // objArrays are precisely marked; restrict scanning 6083 // to dirty cards only. 6084 size = CompactibleFreeListSpace::adjustObjectSize( 6085 p->oop_iterate_size(_scanningClosure, mr)); 6086 } else { 6087 // A non-array may have been imprecisely marked; we need 6088 // to scan object in its entirety. 6089 size = CompactibleFreeListSpace::adjustObjectSize( 6090 p->oop_iterate_size(_scanningClosure)); 6091 } 6092 #ifdef ASSERT 6093 size_t direct_size = 6094 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6095 assert(size == direct_size, "Inconsistency in size"); 6096 assert(size >= 3, "Necessary for Printezis marks to work"); 6097 HeapWord* start_pbit = addr + 1; 6098 HeapWord* end_pbit = addr + size - 1; 6099 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6100 "inconsistent Printezis mark"); 6101 // Verify inner mark bits (between Printezis bits) are clear, 6102 // but don't repeat if there are multiple dirty regions for 6103 // the same object, to avoid potential O(N^2) performance. 6104 if (addr != _last_scanned_object) { 6105 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6106 _last_scanned_object = addr; 6107 } 6108 #endif // ASSERT 6109 } else { 6110 // An uninitialized object. 6111 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6112 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6113 size = pointer_delta(nextOneAddr + 1, addr); 6114 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6115 "alignment problem"); 6116 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6117 // will dirty the card when the klass pointer is installed in the 6118 // object (signaling the completion of initialization). 6119 } 6120 } else { 6121 // Either a not yet marked object or an uninitialized object 6122 if (p->klass_or_null_acquire() == NULL) { 6123 // An uninitialized object, skip to the next card, since 6124 // we may not be able to read its P-bits yet. 6125 assert(size == 0, "Initial value"); 6126 } else { 6127 // An object not (yet) reached by marking: we merely need to 6128 // compute its size so as to go look at the next block. 6129 assert(oopDesc::is_oop(p, true), "should be an oop"); 6130 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6131 } 6132 } 6133 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6134 return size; 6135 } 6136 6137 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6138 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6139 "CMS thread should hold CMS token"); 6140 assert_lock_strong(_freelistLock); 6141 assert_lock_strong(_bitMap->lock()); 6142 // relinquish the free_list_lock and bitMaplock() 6143 _bitMap->lock()->unlock(); 6144 _freelistLock->unlock(); 6145 ConcurrentMarkSweepThread::desynchronize(true); 6146 _collector->stopTimer(); 6147 _collector->incrementYields(); 6148 6149 // See the comment in coordinator_yield() 6150 for (unsigned i = 0; i < CMSYieldSleepCount && 6151 ConcurrentMarkSweepThread::should_yield() && 6152 !CMSCollector::foregroundGCIsActive(); ++i) { 6153 os::sleep(Thread::current(), 1, false); 6154 } 6155 6156 ConcurrentMarkSweepThread::synchronize(true); 6157 _freelistLock->lock_without_safepoint_check(); 6158 _bitMap->lock()->lock_without_safepoint_check(); 6159 _collector->startTimer(); 6160 } 6161 6162 6163 ////////////////////////////////////////////////////////////////// 6164 // SurvivorSpacePrecleanClosure 6165 ////////////////////////////////////////////////////////////////// 6166 // This (single-threaded) closure is used to preclean the oops in 6167 // the survivor spaces. 6168 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6169 6170 HeapWord* addr = (HeapWord*)p; 6171 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6172 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6173 assert(p->klass_or_null() != NULL, "object should be initialized"); 6174 // an initialized object; ignore mark word in verification below 6175 // since we are running concurrent with mutators 6176 assert(oopDesc::is_oop(p, true), "should be an oop"); 6177 // Note that we do not yield while we iterate over 6178 // the interior oops of p, pushing the relevant ones 6179 // on our marking stack. 6180 size_t size = p->oop_iterate_size(_scanning_closure); 6181 do_yield_check(); 6182 // Observe that below, we do not abandon the preclean 6183 // phase as soon as we should; rather we empty the 6184 // marking stack before returning. This is to satisfy 6185 // some existing assertions. In general, it may be a 6186 // good idea to abort immediately and complete the marking 6187 // from the grey objects at a later time. 6188 while (!_mark_stack->isEmpty()) { 6189 oop new_oop = _mark_stack->pop(); 6190 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6191 assert(_bit_map->isMarked((HeapWord*)new_oop), 6192 "only grey objects on this stack"); 6193 // iterate over the oops in this oop, marking and pushing 6194 // the ones in CMS heap (i.e. in _span). 6195 new_oop->oop_iterate(_scanning_closure); 6196 // check if it's time to yield 6197 do_yield_check(); 6198 } 6199 unsigned int after_count = 6200 CMSHeap::heap()->total_collections(); 6201 bool abort = (_before_count != after_count) || 6202 _collector->should_abort_preclean(); 6203 return abort ? 0 : size; 6204 } 6205 6206 void SurvivorSpacePrecleanClosure::do_yield_work() { 6207 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6208 "CMS thread should hold CMS token"); 6209 assert_lock_strong(_bit_map->lock()); 6210 // Relinquish the bit map lock 6211 _bit_map->lock()->unlock(); 6212 ConcurrentMarkSweepThread::desynchronize(true); 6213 _collector->stopTimer(); 6214 _collector->incrementYields(); 6215 6216 // See the comment in coordinator_yield() 6217 for (unsigned i = 0; i < CMSYieldSleepCount && 6218 ConcurrentMarkSweepThread::should_yield() && 6219 !CMSCollector::foregroundGCIsActive(); ++i) { 6220 os::sleep(Thread::current(), 1, false); 6221 } 6222 6223 ConcurrentMarkSweepThread::synchronize(true); 6224 _bit_map->lock()->lock_without_safepoint_check(); 6225 _collector->startTimer(); 6226 } 6227 6228 // This closure is used to rescan the marked objects on the dirty cards 6229 // in the mod union table and the card table proper. In the parallel 6230 // case, although the bitMap is shared, we do a single read so the 6231 // isMarked() query is "safe". 6232 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6233 // Ignore mark word because we are running concurrent with mutators 6234 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6235 HeapWord* addr = (HeapWord*)p; 6236 assert(_span.contains(addr), "we are scanning the CMS generation"); 6237 bool is_obj_array = false; 6238 #ifdef ASSERT 6239 if (!_parallel) { 6240 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6241 assert(_collector->overflow_list_is_empty(), 6242 "overflow list should be empty"); 6243 6244 } 6245 #endif // ASSERT 6246 if (_bit_map->isMarked(addr)) { 6247 // Obj arrays are precisely marked, non-arrays are not; 6248 // so we scan objArrays precisely and non-arrays in their 6249 // entirety. 6250 if (p->is_objArray()) { 6251 is_obj_array = true; 6252 if (_parallel) { 6253 p->oop_iterate(_par_scan_closure, mr); 6254 } else { 6255 p->oop_iterate(_scan_closure, mr); 6256 } 6257 } else { 6258 if (_parallel) { 6259 p->oop_iterate(_par_scan_closure); 6260 } else { 6261 p->oop_iterate(_scan_closure); 6262 } 6263 } 6264 } 6265 #ifdef ASSERT 6266 if (!_parallel) { 6267 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6268 assert(_collector->overflow_list_is_empty(), 6269 "overflow list should be empty"); 6270 6271 } 6272 #endif // ASSERT 6273 return is_obj_array; 6274 } 6275 6276 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6277 MemRegion span, 6278 CMSBitMap* bitMap, CMSMarkStack* markStack, 6279 bool should_yield, bool verifying): 6280 _collector(collector), 6281 _span(span), 6282 _bitMap(bitMap), 6283 _mut(&collector->_modUnionTable), 6284 _markStack(markStack), 6285 _yield(should_yield), 6286 _skipBits(0) 6287 { 6288 assert(_markStack->isEmpty(), "stack should be empty"); 6289 _finger = _bitMap->startWord(); 6290 _threshold = _finger; 6291 assert(_collector->_restart_addr == NULL, "Sanity check"); 6292 assert(_span.contains(_finger), "Out of bounds _finger?"); 6293 DEBUG_ONLY(_verifying = verifying;) 6294 } 6295 6296 void MarkFromRootsClosure::reset(HeapWord* addr) { 6297 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6298 assert(_span.contains(addr), "Out of bounds _finger?"); 6299 _finger = addr; 6300 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6301 } 6302 6303 // Should revisit to see if this should be restructured for 6304 // greater efficiency. 6305 bool MarkFromRootsClosure::do_bit(size_t offset) { 6306 if (_skipBits > 0) { 6307 _skipBits--; 6308 return true; 6309 } 6310 // convert offset into a HeapWord* 6311 HeapWord* addr = _bitMap->startWord() + offset; 6312 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6313 "address out of range"); 6314 assert(_bitMap->isMarked(addr), "tautology"); 6315 if (_bitMap->isMarked(addr+1)) { 6316 // this is an allocated but not yet initialized object 6317 assert(_skipBits == 0, "tautology"); 6318 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6319 oop p = oop(addr); 6320 if (p->klass_or_null_acquire() == NULL) { 6321 DEBUG_ONLY(if (!_verifying) {) 6322 // We re-dirty the cards on which this object lies and increase 6323 // the _threshold so that we'll come back to scan this object 6324 // during the preclean or remark phase. (CMSCleanOnEnter) 6325 if (CMSCleanOnEnter) { 6326 size_t sz = _collector->block_size_using_printezis_bits(addr); 6327 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6328 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6329 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6330 // Bump _threshold to end_card_addr; note that 6331 // _threshold cannot possibly exceed end_card_addr, anyhow. 6332 // This prevents future clearing of the card as the scan proceeds 6333 // to the right. 6334 assert(_threshold <= end_card_addr, 6335 "Because we are just scanning into this object"); 6336 if (_threshold < end_card_addr) { 6337 _threshold = end_card_addr; 6338 } 6339 if (p->klass_or_null_acquire() != NULL) { 6340 // Redirty the range of cards... 6341 _mut->mark_range(redirty_range); 6342 } // ...else the setting of klass will dirty the card anyway. 6343 } 6344 DEBUG_ONLY(}) 6345 return true; 6346 } 6347 } 6348 scanOopsInOop(addr); 6349 return true; 6350 } 6351 6352 // We take a break if we've been at this for a while, 6353 // so as to avoid monopolizing the locks involved. 6354 void MarkFromRootsClosure::do_yield_work() { 6355 // First give up the locks, then yield, then re-lock 6356 // We should probably use a constructor/destructor idiom to 6357 // do this unlock/lock or modify the MutexUnlocker class to 6358 // serve our purpose. XXX 6359 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6360 "CMS thread should hold CMS token"); 6361 assert_lock_strong(_bitMap->lock()); 6362 _bitMap->lock()->unlock(); 6363 ConcurrentMarkSweepThread::desynchronize(true); 6364 _collector->stopTimer(); 6365 _collector->incrementYields(); 6366 6367 // See the comment in coordinator_yield() 6368 for (unsigned i = 0; i < CMSYieldSleepCount && 6369 ConcurrentMarkSweepThread::should_yield() && 6370 !CMSCollector::foregroundGCIsActive(); ++i) { 6371 os::sleep(Thread::current(), 1, false); 6372 } 6373 6374 ConcurrentMarkSweepThread::synchronize(true); 6375 _bitMap->lock()->lock_without_safepoint_check(); 6376 _collector->startTimer(); 6377 } 6378 6379 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6380 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6381 assert(_markStack->isEmpty(), 6382 "should drain stack to limit stack usage"); 6383 // convert ptr to an oop preparatory to scanning 6384 oop obj = oop(ptr); 6385 // Ignore mark word in verification below, since we 6386 // may be running concurrent with mutators. 6387 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6388 assert(_finger <= ptr, "_finger runneth ahead"); 6389 // advance the finger to right end of this object 6390 _finger = ptr + obj->size(); 6391 assert(_finger > ptr, "we just incremented it above"); 6392 // On large heaps, it may take us some time to get through 6393 // the marking phase. During 6394 // this time it's possible that a lot of mutations have 6395 // accumulated in the card table and the mod union table -- 6396 // these mutation records are redundant until we have 6397 // actually traced into the corresponding card. 6398 // Here, we check whether advancing the finger would make 6399 // us cross into a new card, and if so clear corresponding 6400 // cards in the MUT (preclean them in the card-table in the 6401 // future). 6402 6403 DEBUG_ONLY(if (!_verifying) {) 6404 // The clean-on-enter optimization is disabled by default, 6405 // until we fix 6178663. 6406 if (CMSCleanOnEnter && (_finger > _threshold)) { 6407 // [_threshold, _finger) represents the interval 6408 // of cards to be cleared in MUT (or precleaned in card table). 6409 // The set of cards to be cleared is all those that overlap 6410 // with the interval [_threshold, _finger); note that 6411 // _threshold is always kept card-aligned but _finger isn't 6412 // always card-aligned. 6413 HeapWord* old_threshold = _threshold; 6414 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6415 "_threshold should always be card-aligned"); 6416 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6417 MemRegion mr(old_threshold, _threshold); 6418 assert(!mr.is_empty(), "Control point invariant"); 6419 assert(_span.contains(mr), "Should clear within span"); 6420 _mut->clear_range(mr); 6421 } 6422 DEBUG_ONLY(}) 6423 // Note: the finger doesn't advance while we drain 6424 // the stack below. 6425 PushOrMarkClosure pushOrMarkClosure(_collector, 6426 _span, _bitMap, _markStack, 6427 _finger, this); 6428 bool res = _markStack->push(obj); 6429 assert(res, "Empty non-zero size stack should have space for single push"); 6430 while (!_markStack->isEmpty()) { 6431 oop new_oop = _markStack->pop(); 6432 // Skip verifying header mark word below because we are 6433 // running concurrent with mutators. 6434 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6435 // now scan this oop's oops 6436 new_oop->oop_iterate(&pushOrMarkClosure); 6437 do_yield_check(); 6438 } 6439 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6440 } 6441 6442 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6443 CMSCollector* collector, MemRegion span, 6444 CMSBitMap* bit_map, 6445 OopTaskQueue* work_queue, 6446 CMSMarkStack* overflow_stack): 6447 _collector(collector), 6448 _whole_span(collector->_span), 6449 _span(span), 6450 _bit_map(bit_map), 6451 _mut(&collector->_modUnionTable), 6452 _work_queue(work_queue), 6453 _overflow_stack(overflow_stack), 6454 _skip_bits(0), 6455 _task(task) 6456 { 6457 assert(_work_queue->size() == 0, "work_queue should be empty"); 6458 _finger = span.start(); 6459 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6460 assert(_span.contains(_finger), "Out of bounds _finger?"); 6461 } 6462 6463 // Should revisit to see if this should be restructured for 6464 // greater efficiency. 6465 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6466 if (_skip_bits > 0) { 6467 _skip_bits--; 6468 return true; 6469 } 6470 // convert offset into a HeapWord* 6471 HeapWord* addr = _bit_map->startWord() + offset; 6472 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6473 "address out of range"); 6474 assert(_bit_map->isMarked(addr), "tautology"); 6475 if (_bit_map->isMarked(addr+1)) { 6476 // this is an allocated object that might not yet be initialized 6477 assert(_skip_bits == 0, "tautology"); 6478 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6479 oop p = oop(addr); 6480 if (p->klass_or_null_acquire() == NULL) { 6481 // in the case of Clean-on-Enter optimization, redirty card 6482 // and avoid clearing card by increasing the threshold. 6483 return true; 6484 } 6485 } 6486 scan_oops_in_oop(addr); 6487 return true; 6488 } 6489 6490 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6491 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6492 // Should we assert that our work queue is empty or 6493 // below some drain limit? 6494 assert(_work_queue->size() == 0, 6495 "should drain stack to limit stack usage"); 6496 // convert ptr to an oop preparatory to scanning 6497 oop obj = oop(ptr); 6498 // Ignore mark word in verification below, since we 6499 // may be running concurrent with mutators. 6500 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6501 assert(_finger <= ptr, "_finger runneth ahead"); 6502 // advance the finger to right end of this object 6503 _finger = ptr + obj->size(); 6504 assert(_finger > ptr, "we just incremented it above"); 6505 // On large heaps, it may take us some time to get through 6506 // the marking phase. During 6507 // this time it's possible that a lot of mutations have 6508 // accumulated in the card table and the mod union table -- 6509 // these mutation records are redundant until we have 6510 // actually traced into the corresponding card. 6511 // Here, we check whether advancing the finger would make 6512 // us cross into a new card, and if so clear corresponding 6513 // cards in the MUT (preclean them in the card-table in the 6514 // future). 6515 6516 // The clean-on-enter optimization is disabled by default, 6517 // until we fix 6178663. 6518 if (CMSCleanOnEnter && (_finger > _threshold)) { 6519 // [_threshold, _finger) represents the interval 6520 // of cards to be cleared in MUT (or precleaned in card table). 6521 // The set of cards to be cleared is all those that overlap 6522 // with the interval [_threshold, _finger); note that 6523 // _threshold is always kept card-aligned but _finger isn't 6524 // always card-aligned. 6525 HeapWord* old_threshold = _threshold; 6526 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6527 "_threshold should always be card-aligned"); 6528 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6529 MemRegion mr(old_threshold, _threshold); 6530 assert(!mr.is_empty(), "Control point invariant"); 6531 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6532 _mut->clear_range(mr); 6533 } 6534 6535 // Note: the local finger doesn't advance while we drain 6536 // the stack below, but the global finger sure can and will. 6537 HeapWord* volatile* gfa = _task->global_finger_addr(); 6538 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6539 _span, _bit_map, 6540 _work_queue, 6541 _overflow_stack, 6542 _finger, 6543 gfa, this); 6544 bool res = _work_queue->push(obj); // overflow could occur here 6545 assert(res, "Will hold once we use workqueues"); 6546 while (true) { 6547 oop new_oop; 6548 if (!_work_queue->pop_local(new_oop)) { 6549 // We emptied our work_queue; check if there's stuff that can 6550 // be gotten from the overflow stack. 6551 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6552 _overflow_stack, _work_queue)) { 6553 do_yield_check(); 6554 continue; 6555 } else { // done 6556 break; 6557 } 6558 } 6559 // Skip verifying header mark word below because we are 6560 // running concurrent with mutators. 6561 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6562 // now scan this oop's oops 6563 new_oop->oop_iterate(&pushOrMarkClosure); 6564 do_yield_check(); 6565 } 6566 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6567 } 6568 6569 // Yield in response to a request from VM Thread or 6570 // from mutators. 6571 void ParMarkFromRootsClosure::do_yield_work() { 6572 assert(_task != NULL, "sanity"); 6573 _task->yield(); 6574 } 6575 6576 // A variant of the above used for verifying CMS marking work. 6577 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6578 MemRegion span, 6579 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6580 CMSMarkStack* mark_stack): 6581 _collector(collector), 6582 _span(span), 6583 _verification_bm(verification_bm), 6584 _cms_bm(cms_bm), 6585 _mark_stack(mark_stack), 6586 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6587 mark_stack) 6588 { 6589 assert(_mark_stack->isEmpty(), "stack should be empty"); 6590 _finger = _verification_bm->startWord(); 6591 assert(_collector->_restart_addr == NULL, "Sanity check"); 6592 assert(_span.contains(_finger), "Out of bounds _finger?"); 6593 } 6594 6595 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6596 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6597 assert(_span.contains(addr), "Out of bounds _finger?"); 6598 _finger = addr; 6599 } 6600 6601 // Should revisit to see if this should be restructured for 6602 // greater efficiency. 6603 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6604 // convert offset into a HeapWord* 6605 HeapWord* addr = _verification_bm->startWord() + offset; 6606 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6607 "address out of range"); 6608 assert(_verification_bm->isMarked(addr), "tautology"); 6609 assert(_cms_bm->isMarked(addr), "tautology"); 6610 6611 assert(_mark_stack->isEmpty(), 6612 "should drain stack to limit stack usage"); 6613 // convert addr to an oop preparatory to scanning 6614 oop obj = oop(addr); 6615 assert(oopDesc::is_oop(obj), "should be an oop"); 6616 assert(_finger <= addr, "_finger runneth ahead"); 6617 // advance the finger to right end of this object 6618 _finger = addr + obj->size(); 6619 assert(_finger > addr, "we just incremented it above"); 6620 // Note: the finger doesn't advance while we drain 6621 // the stack below. 6622 bool res = _mark_stack->push(obj); 6623 assert(res, "Empty non-zero size stack should have space for single push"); 6624 while (!_mark_stack->isEmpty()) { 6625 oop new_oop = _mark_stack->pop(); 6626 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6627 // now scan this oop's oops 6628 new_oop->oop_iterate(&_pam_verify_closure); 6629 } 6630 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6631 return true; 6632 } 6633 6634 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6635 CMSCollector* collector, MemRegion span, 6636 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6637 CMSMarkStack* mark_stack): 6638 MetadataAwareOopClosure(collector->ref_processor()), 6639 _collector(collector), 6640 _span(span), 6641 _verification_bm(verification_bm), 6642 _cms_bm(cms_bm), 6643 _mark_stack(mark_stack) 6644 { } 6645 6646 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6647 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6648 6649 // Upon stack overflow, we discard (part of) the stack, 6650 // remembering the least address amongst those discarded 6651 // in CMSCollector's _restart_address. 6652 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6653 // Remember the least grey address discarded 6654 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6655 _collector->lower_restart_addr(ra); 6656 _mark_stack->reset(); // discard stack contents 6657 _mark_stack->expand(); // expand the stack if possible 6658 } 6659 6660 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6661 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6662 HeapWord* addr = (HeapWord*)obj; 6663 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6664 // Oop lies in _span and isn't yet grey or black 6665 _verification_bm->mark(addr); // now grey 6666 if (!_cms_bm->isMarked(addr)) { 6667 Log(gc, verify) log; 6668 ResourceMark rm; 6669 LogStream ls(log.error()); 6670 oop(addr)->print_on(&ls); 6671 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6672 fatal("... aborting"); 6673 } 6674 6675 if (!_mark_stack->push(obj)) { // stack overflow 6676 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6677 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6678 handle_stack_overflow(addr); 6679 } 6680 // anything including and to the right of _finger 6681 // will be scanned as we iterate over the remainder of the 6682 // bit map 6683 } 6684 } 6685 6686 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6687 MemRegion span, 6688 CMSBitMap* bitMap, CMSMarkStack* markStack, 6689 HeapWord* finger, MarkFromRootsClosure* parent) : 6690 MetadataAwareOopClosure(collector->ref_processor()), 6691 _collector(collector), 6692 _span(span), 6693 _bitMap(bitMap), 6694 _markStack(markStack), 6695 _finger(finger), 6696 _parent(parent) 6697 { } 6698 6699 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6700 MemRegion span, 6701 CMSBitMap* bit_map, 6702 OopTaskQueue* work_queue, 6703 CMSMarkStack* overflow_stack, 6704 HeapWord* finger, 6705 HeapWord* volatile* global_finger_addr, 6706 ParMarkFromRootsClosure* parent) : 6707 MetadataAwareOopClosure(collector->ref_processor()), 6708 _collector(collector), 6709 _whole_span(collector->_span), 6710 _span(span), 6711 _bit_map(bit_map), 6712 _work_queue(work_queue), 6713 _overflow_stack(overflow_stack), 6714 _finger(finger), 6715 _global_finger_addr(global_finger_addr), 6716 _parent(parent) 6717 { } 6718 6719 // Assumes thread-safe access by callers, who are 6720 // responsible for mutual exclusion. 6721 void CMSCollector::lower_restart_addr(HeapWord* low) { 6722 assert(_span.contains(low), "Out of bounds addr"); 6723 if (_restart_addr == NULL) { 6724 _restart_addr = low; 6725 } else { 6726 _restart_addr = MIN2(_restart_addr, low); 6727 } 6728 } 6729 6730 // Upon stack overflow, we discard (part of) the stack, 6731 // remembering the least address amongst those discarded 6732 // in CMSCollector's _restart_address. 6733 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6734 // Remember the least grey address discarded 6735 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6736 _collector->lower_restart_addr(ra); 6737 _markStack->reset(); // discard stack contents 6738 _markStack->expand(); // expand the stack if possible 6739 } 6740 6741 // Upon stack overflow, we discard (part of) the stack, 6742 // remembering the least address amongst those discarded 6743 // in CMSCollector's _restart_address. 6744 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6745 // We need to do this under a mutex to prevent other 6746 // workers from interfering with the work done below. 6747 MutexLockerEx ml(_overflow_stack->par_lock(), 6748 Mutex::_no_safepoint_check_flag); 6749 // Remember the least grey address discarded 6750 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6751 _collector->lower_restart_addr(ra); 6752 _overflow_stack->reset(); // discard stack contents 6753 _overflow_stack->expand(); // expand the stack if possible 6754 } 6755 6756 void PushOrMarkClosure::do_oop(oop obj) { 6757 // Ignore mark word because we are running concurrent with mutators. 6758 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6759 HeapWord* addr = (HeapWord*)obj; 6760 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6761 // Oop lies in _span and isn't yet grey or black 6762 _bitMap->mark(addr); // now grey 6763 if (addr < _finger) { 6764 // the bit map iteration has already either passed, or 6765 // sampled, this bit in the bit map; we'll need to 6766 // use the marking stack to scan this oop's oops. 6767 bool simulate_overflow = false; 6768 NOT_PRODUCT( 6769 if (CMSMarkStackOverflowALot && 6770 _collector->simulate_overflow()) { 6771 // simulate a stack overflow 6772 simulate_overflow = true; 6773 } 6774 ) 6775 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6776 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6777 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6778 handle_stack_overflow(addr); 6779 } 6780 } 6781 // anything including and to the right of _finger 6782 // will be scanned as we iterate over the remainder of the 6783 // bit map 6784 do_yield_check(); 6785 } 6786 } 6787 6788 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6789 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6790 6791 void ParPushOrMarkClosure::do_oop(oop obj) { 6792 // Ignore mark word because we are running concurrent with mutators. 6793 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6794 HeapWord* addr = (HeapWord*)obj; 6795 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6796 // Oop lies in _span and isn't yet grey or black 6797 // We read the global_finger (volatile read) strictly after marking oop 6798 bool res = _bit_map->par_mark(addr); // now grey 6799 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6800 // Should we push this marked oop on our stack? 6801 // -- if someone else marked it, nothing to do 6802 // -- if target oop is above global finger nothing to do 6803 // -- if target oop is in chunk and above local finger 6804 // then nothing to do 6805 // -- else push on work queue 6806 if ( !res // someone else marked it, they will deal with it 6807 || (addr >= *gfa) // will be scanned in a later task 6808 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6809 return; 6810 } 6811 // the bit map iteration has already either passed, or 6812 // sampled, this bit in the bit map; we'll need to 6813 // use the marking stack to scan this oop's oops. 6814 bool simulate_overflow = false; 6815 NOT_PRODUCT( 6816 if (CMSMarkStackOverflowALot && 6817 _collector->simulate_overflow()) { 6818 // simulate a stack overflow 6819 simulate_overflow = true; 6820 } 6821 ) 6822 if (simulate_overflow || 6823 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6824 // stack overflow 6825 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6826 // We cannot assert that the overflow stack is full because 6827 // it may have been emptied since. 6828 assert(simulate_overflow || 6829 _work_queue->size() == _work_queue->max_elems(), 6830 "Else push should have succeeded"); 6831 handle_stack_overflow(addr); 6832 } 6833 do_yield_check(); 6834 } 6835 } 6836 6837 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6838 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6839 6840 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6841 MemRegion span, 6842 ReferenceProcessor* rp, 6843 CMSBitMap* bit_map, 6844 CMSBitMap* mod_union_table, 6845 CMSMarkStack* mark_stack, 6846 bool concurrent_precleaning): 6847 MetadataAwareOopClosure(rp), 6848 _collector(collector), 6849 _span(span), 6850 _bit_map(bit_map), 6851 _mod_union_table(mod_union_table), 6852 _mark_stack(mark_stack), 6853 _concurrent_precleaning(concurrent_precleaning) 6854 { 6855 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6856 } 6857 6858 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6859 // the non-parallel version (the parallel version appears further below.) 6860 void PushAndMarkClosure::do_oop(oop obj) { 6861 // Ignore mark word verification. If during concurrent precleaning, 6862 // the object monitor may be locked. If during the checkpoint 6863 // phases, the object may already have been reached by a different 6864 // path and may be at the end of the global overflow list (so 6865 // the mark word may be NULL). 6866 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6867 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6868 HeapWord* addr = (HeapWord*)obj; 6869 // Check if oop points into the CMS generation 6870 // and is not marked 6871 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6872 // a white object ... 6873 _bit_map->mark(addr); // ... now grey 6874 // push on the marking stack (grey set) 6875 bool simulate_overflow = false; 6876 NOT_PRODUCT( 6877 if (CMSMarkStackOverflowALot && 6878 _collector->simulate_overflow()) { 6879 // simulate a stack overflow 6880 simulate_overflow = true; 6881 } 6882 ) 6883 if (simulate_overflow || !_mark_stack->push(obj)) { 6884 if (_concurrent_precleaning) { 6885 // During precleaning we can just dirty the appropriate card(s) 6886 // in the mod union table, thus ensuring that the object remains 6887 // in the grey set and continue. In the case of object arrays 6888 // we need to dirty all of the cards that the object spans, 6889 // since the rescan of object arrays will be limited to the 6890 // dirty cards. 6891 // Note that no one can be interfering with us in this action 6892 // of dirtying the mod union table, so no locking or atomics 6893 // are required. 6894 if (obj->is_objArray()) { 6895 size_t sz = obj->size(); 6896 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6897 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6898 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6899 _mod_union_table->mark_range(redirty_range); 6900 } else { 6901 _mod_union_table->mark(addr); 6902 } 6903 _collector->_ser_pmc_preclean_ovflw++; 6904 } else { 6905 // During the remark phase, we need to remember this oop 6906 // in the overflow list. 6907 _collector->push_on_overflow_list(obj); 6908 _collector->_ser_pmc_remark_ovflw++; 6909 } 6910 } 6911 } 6912 } 6913 6914 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6915 MemRegion span, 6916 ReferenceProcessor* rp, 6917 CMSBitMap* bit_map, 6918 OopTaskQueue* work_queue): 6919 MetadataAwareOopClosure(rp), 6920 _collector(collector), 6921 _span(span), 6922 _bit_map(bit_map), 6923 _work_queue(work_queue) 6924 { 6925 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6926 } 6927 6928 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6929 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6930 6931 // Grey object rescan during second checkpoint phase -- 6932 // the parallel version. 6933 void ParPushAndMarkClosure::do_oop(oop obj) { 6934 // In the assert below, we ignore the mark word because 6935 // this oop may point to an already visited object that is 6936 // on the overflow stack (in which case the mark word has 6937 // been hijacked for chaining into the overflow stack -- 6938 // if this is the last object in the overflow stack then 6939 // its mark word will be NULL). Because this object may 6940 // have been subsequently popped off the global overflow 6941 // stack, and the mark word possibly restored to the prototypical 6942 // value, by the time we get to examined this failing assert in 6943 // the debugger, is_oop_or_null(false) may subsequently start 6944 // to hold. 6945 assert(oopDesc::is_oop_or_null(obj, true), 6946 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6947 HeapWord* addr = (HeapWord*)obj; 6948 // Check if oop points into the CMS generation 6949 // and is not marked 6950 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6951 // a white object ... 6952 // If we manage to "claim" the object, by being the 6953 // first thread to mark it, then we push it on our 6954 // marking stack 6955 if (_bit_map->par_mark(addr)) { // ... now grey 6956 // push on work queue (grey set) 6957 bool simulate_overflow = false; 6958 NOT_PRODUCT( 6959 if (CMSMarkStackOverflowALot && 6960 _collector->par_simulate_overflow()) { 6961 // simulate a stack overflow 6962 simulate_overflow = true; 6963 } 6964 ) 6965 if (simulate_overflow || !_work_queue->push(obj)) { 6966 _collector->par_push_on_overflow_list(obj); 6967 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6968 } 6969 } // Else, some other thread got there first 6970 } 6971 } 6972 6973 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6974 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6975 6976 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6977 Mutex* bml = _collector->bitMapLock(); 6978 assert_lock_strong(bml); 6979 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6980 "CMS thread should hold CMS token"); 6981 6982 bml->unlock(); 6983 ConcurrentMarkSweepThread::desynchronize(true); 6984 6985 _collector->stopTimer(); 6986 _collector->incrementYields(); 6987 6988 // See the comment in coordinator_yield() 6989 for (unsigned i = 0; i < CMSYieldSleepCount && 6990 ConcurrentMarkSweepThread::should_yield() && 6991 !CMSCollector::foregroundGCIsActive(); ++i) { 6992 os::sleep(Thread::current(), 1, false); 6993 } 6994 6995 ConcurrentMarkSweepThread::synchronize(true); 6996 bml->lock(); 6997 6998 _collector->startTimer(); 6999 } 7000 7001 bool CMSPrecleanRefsYieldClosure::should_return() { 7002 if (ConcurrentMarkSweepThread::should_yield()) { 7003 do_yield_work(); 7004 } 7005 return _collector->foregroundGCIsActive(); 7006 } 7007 7008 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7009 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7010 "mr should be aligned to start at a card boundary"); 7011 // We'd like to assert: 7012 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7013 // "mr should be a range of cards"); 7014 // However, that would be too strong in one case -- the last 7015 // partition ends at _unallocated_block which, in general, can be 7016 // an arbitrary boundary, not necessarily card aligned. 7017 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7018 _space->object_iterate_mem(mr, &_scan_cl); 7019 } 7020 7021 SweepClosure::SweepClosure(CMSCollector* collector, 7022 ConcurrentMarkSweepGeneration* g, 7023 CMSBitMap* bitMap, bool should_yield) : 7024 _collector(collector), 7025 _g(g), 7026 _sp(g->cmsSpace()), 7027 _limit(_sp->sweep_limit()), 7028 _freelistLock(_sp->freelistLock()), 7029 _bitMap(bitMap), 7030 _yield(should_yield), 7031 _inFreeRange(false), // No free range at beginning of sweep 7032 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7033 _lastFreeRangeCoalesced(false), 7034 _freeFinger(g->used_region().start()) 7035 { 7036 NOT_PRODUCT( 7037 _numObjectsFreed = 0; 7038 _numWordsFreed = 0; 7039 _numObjectsLive = 0; 7040 _numWordsLive = 0; 7041 _numObjectsAlreadyFree = 0; 7042 _numWordsAlreadyFree = 0; 7043 _last_fc = NULL; 7044 7045 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7046 _sp->dictionary()->initialize_dict_returned_bytes(); 7047 ) 7048 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7049 "sweep _limit out of bounds"); 7050 log_develop_trace(gc, sweep)("===================="); 7051 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7052 } 7053 7054 void SweepClosure::print_on(outputStream* st) const { 7055 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7056 p2i(_sp->bottom()), p2i(_sp->end())); 7057 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7058 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7059 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7060 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7061 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7062 } 7063 7064 #ifndef PRODUCT 7065 // Assertion checking only: no useful work in product mode -- 7066 // however, if any of the flags below become product flags, 7067 // you may need to review this code to see if it needs to be 7068 // enabled in product mode. 7069 SweepClosure::~SweepClosure() { 7070 assert_lock_strong(_freelistLock); 7071 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7072 "sweep _limit out of bounds"); 7073 if (inFreeRange()) { 7074 Log(gc, sweep) log; 7075 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7076 ResourceMark rm; 7077 LogStream ls(log.error()); 7078 print_on(&ls); 7079 ShouldNotReachHere(); 7080 } 7081 7082 if (log_is_enabled(Debug, gc, sweep)) { 7083 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7084 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7085 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7086 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7087 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7088 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7089 } 7090 7091 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7092 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7093 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7094 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7095 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7096 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7097 } 7098 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7099 log_develop_trace(gc, sweep)("================"); 7100 } 7101 #endif // PRODUCT 7102 7103 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7104 bool freeRangeInFreeLists) { 7105 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7106 p2i(freeFinger), freeRangeInFreeLists); 7107 assert(!inFreeRange(), "Trampling existing free range"); 7108 set_inFreeRange(true); 7109 set_lastFreeRangeCoalesced(false); 7110 7111 set_freeFinger(freeFinger); 7112 set_freeRangeInFreeLists(freeRangeInFreeLists); 7113 if (CMSTestInFreeList) { 7114 if (freeRangeInFreeLists) { 7115 FreeChunk* fc = (FreeChunk*) freeFinger; 7116 assert(fc->is_free(), "A chunk on the free list should be free."); 7117 assert(fc->size() > 0, "Free range should have a size"); 7118 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7119 } 7120 } 7121 } 7122 7123 // Note that the sweeper runs concurrently with mutators. Thus, 7124 // it is possible for direct allocation in this generation to happen 7125 // in the middle of the sweep. Note that the sweeper also coalesces 7126 // contiguous free blocks. Thus, unless the sweeper and the allocator 7127 // synchronize appropriately freshly allocated blocks may get swept up. 7128 // This is accomplished by the sweeper locking the free lists while 7129 // it is sweeping. Thus blocks that are determined to be free are 7130 // indeed free. There is however one additional complication: 7131 // blocks that have been allocated since the final checkpoint and 7132 // mark, will not have been marked and so would be treated as 7133 // unreachable and swept up. To prevent this, the allocator marks 7134 // the bit map when allocating during the sweep phase. This leads, 7135 // however, to a further complication -- objects may have been allocated 7136 // but not yet initialized -- in the sense that the header isn't yet 7137 // installed. The sweeper can not then determine the size of the block 7138 // in order to skip over it. To deal with this case, we use a technique 7139 // (due to Printezis) to encode such uninitialized block sizes in the 7140 // bit map. Since the bit map uses a bit per every HeapWord, but the 7141 // CMS generation has a minimum object size of 3 HeapWords, it follows 7142 // that "normal marks" won't be adjacent in the bit map (there will 7143 // always be at least two 0 bits between successive 1 bits). We make use 7144 // of these "unused" bits to represent uninitialized blocks -- the bit 7145 // corresponding to the start of the uninitialized object and the next 7146 // bit are both set. Finally, a 1 bit marks the end of the object that 7147 // started with the two consecutive 1 bits to indicate its potentially 7148 // uninitialized state. 7149 7150 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7151 FreeChunk* fc = (FreeChunk*)addr; 7152 size_t res; 7153 7154 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7155 // than "addr == _limit" because although _limit was a block boundary when 7156 // we started the sweep, it may no longer be one because heap expansion 7157 // may have caused us to coalesce the block ending at the address _limit 7158 // with a newly expanded chunk (this happens when _limit was set to the 7159 // previous _end of the space), so we may have stepped past _limit: 7160 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7161 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7162 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7163 "sweep _limit out of bounds"); 7164 assert(addr < _sp->end(), "addr out of bounds"); 7165 // Flush any free range we might be holding as a single 7166 // coalesced chunk to the appropriate free list. 7167 if (inFreeRange()) { 7168 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7169 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7170 flush_cur_free_chunk(freeFinger(), 7171 pointer_delta(addr, freeFinger())); 7172 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7173 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7174 lastFreeRangeCoalesced() ? 1 : 0); 7175 } 7176 7177 // help the iterator loop finish 7178 return pointer_delta(_sp->end(), addr); 7179 } 7180 7181 assert(addr < _limit, "sweep invariant"); 7182 // check if we should yield 7183 do_yield_check(addr); 7184 if (fc->is_free()) { 7185 // Chunk that is already free 7186 res = fc->size(); 7187 do_already_free_chunk(fc); 7188 debug_only(_sp->verifyFreeLists()); 7189 // If we flush the chunk at hand in lookahead_and_flush() 7190 // and it's coalesced with a preceding chunk, then the 7191 // process of "mangling" the payload of the coalesced block 7192 // will cause erasure of the size information from the 7193 // (erstwhile) header of all the coalesced blocks but the 7194 // first, so the first disjunct in the assert will not hold 7195 // in that specific case (in which case the second disjunct 7196 // will hold). 7197 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7198 "Otherwise the size info doesn't change at this step"); 7199 NOT_PRODUCT( 7200 _numObjectsAlreadyFree++; 7201 _numWordsAlreadyFree += res; 7202 ) 7203 NOT_PRODUCT(_last_fc = fc;) 7204 } else if (!_bitMap->isMarked(addr)) { 7205 // Chunk is fresh garbage 7206 res = do_garbage_chunk(fc); 7207 debug_only(_sp->verifyFreeLists()); 7208 NOT_PRODUCT( 7209 _numObjectsFreed++; 7210 _numWordsFreed += res; 7211 ) 7212 } else { 7213 // Chunk that is alive. 7214 res = do_live_chunk(fc); 7215 debug_only(_sp->verifyFreeLists()); 7216 NOT_PRODUCT( 7217 _numObjectsLive++; 7218 _numWordsLive += res; 7219 ) 7220 } 7221 return res; 7222 } 7223 7224 // For the smart allocation, record following 7225 // split deaths - a free chunk is removed from its free list because 7226 // it is being split into two or more chunks. 7227 // split birth - a free chunk is being added to its free list because 7228 // a larger free chunk has been split and resulted in this free chunk. 7229 // coal death - a free chunk is being removed from its free list because 7230 // it is being coalesced into a large free chunk. 7231 // coal birth - a free chunk is being added to its free list because 7232 // it was created when two or more free chunks where coalesced into 7233 // this free chunk. 7234 // 7235 // These statistics are used to determine the desired number of free 7236 // chunks of a given size. The desired number is chosen to be relative 7237 // to the end of a CMS sweep. The desired number at the end of a sweep 7238 // is the 7239 // count-at-end-of-previous-sweep (an amount that was enough) 7240 // - count-at-beginning-of-current-sweep (the excess) 7241 // + split-births (gains in this size during interval) 7242 // - split-deaths (demands on this size during interval) 7243 // where the interval is from the end of one sweep to the end of the 7244 // next. 7245 // 7246 // When sweeping the sweeper maintains an accumulated chunk which is 7247 // the chunk that is made up of chunks that have been coalesced. That 7248 // will be termed the left-hand chunk. A new chunk of garbage that 7249 // is being considered for coalescing will be referred to as the 7250 // right-hand chunk. 7251 // 7252 // When making a decision on whether to coalesce a right-hand chunk with 7253 // the current left-hand chunk, the current count vs. the desired count 7254 // of the left-hand chunk is considered. Also if the right-hand chunk 7255 // is near the large chunk at the end of the heap (see 7256 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7257 // left-hand chunk is coalesced. 7258 // 7259 // When making a decision about whether to split a chunk, the desired count 7260 // vs. the current count of the candidate to be split is also considered. 7261 // If the candidate is underpopulated (currently fewer chunks than desired) 7262 // a chunk of an overpopulated (currently more chunks than desired) size may 7263 // be chosen. The "hint" associated with a free list, if non-null, points 7264 // to a free list which may be overpopulated. 7265 // 7266 7267 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7268 const size_t size = fc->size(); 7269 // Chunks that cannot be coalesced are not in the 7270 // free lists. 7271 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7272 assert(_sp->verify_chunk_in_free_list(fc), 7273 "free chunk should be in free lists"); 7274 } 7275 // a chunk that is already free, should not have been 7276 // marked in the bit map 7277 HeapWord* const addr = (HeapWord*) fc; 7278 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7279 // Verify that the bit map has no bits marked between 7280 // addr and purported end of this block. 7281 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7282 7283 // Some chunks cannot be coalesced under any circumstances. 7284 // See the definition of cantCoalesce(). 7285 if (!fc->cantCoalesce()) { 7286 // This chunk can potentially be coalesced. 7287 // All the work is done in 7288 do_post_free_or_garbage_chunk(fc, size); 7289 // Note that if the chunk is not coalescable (the else arm 7290 // below), we unconditionally flush, without needing to do 7291 // a "lookahead," as we do below. 7292 if (inFreeRange()) lookahead_and_flush(fc, size); 7293 } else { 7294 // Code path common to both original and adaptive free lists. 7295 7296 // cant coalesce with previous block; this should be treated 7297 // as the end of a free run if any 7298 if (inFreeRange()) { 7299 // we kicked some butt; time to pick up the garbage 7300 assert(freeFinger() < addr, "freeFinger points too high"); 7301 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7302 } 7303 // else, nothing to do, just continue 7304 } 7305 } 7306 7307 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7308 // This is a chunk of garbage. It is not in any free list. 7309 // Add it to a free list or let it possibly be coalesced into 7310 // a larger chunk. 7311 HeapWord* const addr = (HeapWord*) fc; 7312 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7313 7314 // Verify that the bit map has no bits marked between 7315 // addr and purported end of just dead object. 7316 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7317 do_post_free_or_garbage_chunk(fc, size); 7318 7319 assert(_limit >= addr + size, 7320 "A freshly garbage chunk can't possibly straddle over _limit"); 7321 if (inFreeRange()) lookahead_and_flush(fc, size); 7322 return size; 7323 } 7324 7325 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7326 HeapWord* addr = (HeapWord*) fc; 7327 // The sweeper has just found a live object. Return any accumulated 7328 // left hand chunk to the free lists. 7329 if (inFreeRange()) { 7330 assert(freeFinger() < addr, "freeFinger points too high"); 7331 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7332 } 7333 7334 // This object is live: we'd normally expect this to be 7335 // an oop, and like to assert the following: 7336 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7337 // However, as we commented above, this may be an object whose 7338 // header hasn't yet been initialized. 7339 size_t size; 7340 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7341 if (_bitMap->isMarked(addr + 1)) { 7342 // Determine the size from the bit map, rather than trying to 7343 // compute it from the object header. 7344 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7345 size = pointer_delta(nextOneAddr + 1, addr); 7346 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7347 "alignment problem"); 7348 7349 #ifdef ASSERT 7350 if (oop(addr)->klass_or_null_acquire() != NULL) { 7351 // Ignore mark word because we are running concurrent with mutators 7352 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7353 assert(size == 7354 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7355 "P-mark and computed size do not agree"); 7356 } 7357 #endif 7358 7359 } else { 7360 // This should be an initialized object that's alive. 7361 assert(oop(addr)->klass_or_null_acquire() != NULL, 7362 "Should be an initialized object"); 7363 // Ignore mark word because we are running concurrent with mutators 7364 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7365 // Verify that the bit map has no bits marked between 7366 // addr and purported end of this block. 7367 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7368 assert(size >= 3, "Necessary for Printezis marks to work"); 7369 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7370 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7371 } 7372 return size; 7373 } 7374 7375 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7376 size_t chunkSize) { 7377 // do_post_free_or_garbage_chunk() should only be called in the case 7378 // of the adaptive free list allocator. 7379 const bool fcInFreeLists = fc->is_free(); 7380 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7381 if (CMSTestInFreeList && fcInFreeLists) { 7382 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7383 } 7384 7385 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7386 7387 HeapWord* const fc_addr = (HeapWord*) fc; 7388 7389 bool coalesce = false; 7390 const size_t left = pointer_delta(fc_addr, freeFinger()); 7391 const size_t right = chunkSize; 7392 switch (FLSCoalescePolicy) { 7393 // numeric value forms a coalition aggressiveness metric 7394 case 0: { // never coalesce 7395 coalesce = false; 7396 break; 7397 } 7398 case 1: { // coalesce if left & right chunks on overpopulated lists 7399 coalesce = _sp->coalOverPopulated(left) && 7400 _sp->coalOverPopulated(right); 7401 break; 7402 } 7403 case 2: { // coalesce if left chunk on overpopulated list (default) 7404 coalesce = _sp->coalOverPopulated(left); 7405 break; 7406 } 7407 case 3: { // coalesce if left OR right chunk on overpopulated list 7408 coalesce = _sp->coalOverPopulated(left) || 7409 _sp->coalOverPopulated(right); 7410 break; 7411 } 7412 case 4: { // always coalesce 7413 coalesce = true; 7414 break; 7415 } 7416 default: 7417 ShouldNotReachHere(); 7418 } 7419 7420 // Should the current free range be coalesced? 7421 // If the chunk is in a free range and either we decided to coalesce above 7422 // or the chunk is near the large block at the end of the heap 7423 // (isNearLargestChunk() returns true), then coalesce this chunk. 7424 const bool doCoalesce = inFreeRange() 7425 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7426 if (doCoalesce) { 7427 // Coalesce the current free range on the left with the new 7428 // chunk on the right. If either is on a free list, 7429 // it must be removed from the list and stashed in the closure. 7430 if (freeRangeInFreeLists()) { 7431 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7432 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7433 "Size of free range is inconsistent with chunk size."); 7434 if (CMSTestInFreeList) { 7435 assert(_sp->verify_chunk_in_free_list(ffc), 7436 "Chunk is not in free lists"); 7437 } 7438 _sp->coalDeath(ffc->size()); 7439 _sp->removeFreeChunkFromFreeLists(ffc); 7440 set_freeRangeInFreeLists(false); 7441 } 7442 if (fcInFreeLists) { 7443 _sp->coalDeath(chunkSize); 7444 assert(fc->size() == chunkSize, 7445 "The chunk has the wrong size or is not in the free lists"); 7446 _sp->removeFreeChunkFromFreeLists(fc); 7447 } 7448 set_lastFreeRangeCoalesced(true); 7449 print_free_block_coalesced(fc); 7450 } else { // not in a free range and/or should not coalesce 7451 // Return the current free range and start a new one. 7452 if (inFreeRange()) { 7453 // In a free range but cannot coalesce with the right hand chunk. 7454 // Put the current free range into the free lists. 7455 flush_cur_free_chunk(freeFinger(), 7456 pointer_delta(fc_addr, freeFinger())); 7457 } 7458 // Set up for new free range. Pass along whether the right hand 7459 // chunk is in the free lists. 7460 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7461 } 7462 } 7463 7464 // Lookahead flush: 7465 // If we are tracking a free range, and this is the last chunk that 7466 // we'll look at because its end crosses past _limit, we'll preemptively 7467 // flush it along with any free range we may be holding on to. Note that 7468 // this can be the case only for an already free or freshly garbage 7469 // chunk. If this block is an object, it can never straddle 7470 // over _limit. The "straddling" occurs when _limit is set at 7471 // the previous end of the space when this cycle started, and 7472 // a subsequent heap expansion caused the previously co-terminal 7473 // free block to be coalesced with the newly expanded portion, 7474 // thus rendering _limit a non-block-boundary making it dangerous 7475 // for the sweeper to step over and examine. 7476 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7477 assert(inFreeRange(), "Should only be called if currently in a free range."); 7478 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7479 assert(_sp->used_region().contains(eob - 1), 7480 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7481 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7482 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7483 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7484 if (eob >= _limit) { 7485 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7486 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7487 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7488 "[" PTR_FORMAT "," PTR_FORMAT ")", 7489 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7490 // Return the storage we are tracking back into the free lists. 7491 log_develop_trace(gc, sweep)("Flushing ... "); 7492 assert(freeFinger() < eob, "Error"); 7493 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7494 } 7495 } 7496 7497 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7498 assert(inFreeRange(), "Should only be called if currently in a free range."); 7499 assert(size > 0, 7500 "A zero sized chunk cannot be added to the free lists."); 7501 if (!freeRangeInFreeLists()) { 7502 if (CMSTestInFreeList) { 7503 FreeChunk* fc = (FreeChunk*) chunk; 7504 fc->set_size(size); 7505 assert(!_sp->verify_chunk_in_free_list(fc), 7506 "chunk should not be in free lists yet"); 7507 } 7508 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7509 // A new free range is going to be starting. The current 7510 // free range has not been added to the free lists yet or 7511 // was removed so add it back. 7512 // If the current free range was coalesced, then the death 7513 // of the free range was recorded. Record a birth now. 7514 if (lastFreeRangeCoalesced()) { 7515 _sp->coalBirth(size); 7516 } 7517 _sp->addChunkAndRepairOffsetTable(chunk, size, 7518 lastFreeRangeCoalesced()); 7519 } else { 7520 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7521 } 7522 set_inFreeRange(false); 7523 set_freeRangeInFreeLists(false); 7524 } 7525 7526 // We take a break if we've been at this for a while, 7527 // so as to avoid monopolizing the locks involved. 7528 void SweepClosure::do_yield_work(HeapWord* addr) { 7529 // Return current free chunk being used for coalescing (if any) 7530 // to the appropriate freelist. After yielding, the next 7531 // free block encountered will start a coalescing range of 7532 // free blocks. If the next free block is adjacent to the 7533 // chunk just flushed, they will need to wait for the next 7534 // sweep to be coalesced. 7535 if (inFreeRange()) { 7536 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7537 } 7538 7539 // First give up the locks, then yield, then re-lock. 7540 // We should probably use a constructor/destructor idiom to 7541 // do this unlock/lock or modify the MutexUnlocker class to 7542 // serve our purpose. XXX 7543 assert_lock_strong(_bitMap->lock()); 7544 assert_lock_strong(_freelistLock); 7545 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7546 "CMS thread should hold CMS token"); 7547 _bitMap->lock()->unlock(); 7548 _freelistLock->unlock(); 7549 ConcurrentMarkSweepThread::desynchronize(true); 7550 _collector->stopTimer(); 7551 _collector->incrementYields(); 7552 7553 // See the comment in coordinator_yield() 7554 for (unsigned i = 0; i < CMSYieldSleepCount && 7555 ConcurrentMarkSweepThread::should_yield() && 7556 !CMSCollector::foregroundGCIsActive(); ++i) { 7557 os::sleep(Thread::current(), 1, false); 7558 } 7559 7560 ConcurrentMarkSweepThread::synchronize(true); 7561 _freelistLock->lock(); 7562 _bitMap->lock()->lock_without_safepoint_check(); 7563 _collector->startTimer(); 7564 } 7565 7566 #ifndef PRODUCT 7567 // This is actually very useful in a product build if it can 7568 // be called from the debugger. Compile it into the product 7569 // as needed. 7570 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7571 return debug_cms_space->verify_chunk_in_free_list(fc); 7572 } 7573 #endif 7574 7575 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7576 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7577 p2i(fc), fc->size()); 7578 } 7579 7580 // CMSIsAliveClosure 7581 bool CMSIsAliveClosure::do_object_b(oop obj) { 7582 HeapWord* addr = (HeapWord*)obj; 7583 return addr != NULL && 7584 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7585 } 7586 7587 7588 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7589 MemRegion span, 7590 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7591 bool cpc): 7592 _collector(collector), 7593 _span(span), 7594 _bit_map(bit_map), 7595 _mark_stack(mark_stack), 7596 _concurrent_precleaning(cpc) { 7597 assert(!_span.is_empty(), "Empty span could spell trouble"); 7598 } 7599 7600 7601 // CMSKeepAliveClosure: the serial version 7602 void CMSKeepAliveClosure::do_oop(oop obj) { 7603 HeapWord* addr = (HeapWord*)obj; 7604 if (_span.contains(addr) && 7605 !_bit_map->isMarked(addr)) { 7606 _bit_map->mark(addr); 7607 bool simulate_overflow = false; 7608 NOT_PRODUCT( 7609 if (CMSMarkStackOverflowALot && 7610 _collector->simulate_overflow()) { 7611 // simulate a stack overflow 7612 simulate_overflow = true; 7613 } 7614 ) 7615 if (simulate_overflow || !_mark_stack->push(obj)) { 7616 if (_concurrent_precleaning) { 7617 // We dirty the overflown object and let the remark 7618 // phase deal with it. 7619 assert(_collector->overflow_list_is_empty(), "Error"); 7620 // In the case of object arrays, we need to dirty all of 7621 // the cards that the object spans. No locking or atomics 7622 // are needed since no one else can be mutating the mod union 7623 // table. 7624 if (obj->is_objArray()) { 7625 size_t sz = obj->size(); 7626 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 7627 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7628 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7629 _collector->_modUnionTable.mark_range(redirty_range); 7630 } else { 7631 _collector->_modUnionTable.mark(addr); 7632 } 7633 _collector->_ser_kac_preclean_ovflw++; 7634 } else { 7635 _collector->push_on_overflow_list(obj); 7636 _collector->_ser_kac_ovflw++; 7637 } 7638 } 7639 } 7640 } 7641 7642 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7643 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7644 7645 // CMSParKeepAliveClosure: a parallel version of the above. 7646 // The work queues are private to each closure (thread), 7647 // but (may be) available for stealing by other threads. 7648 void CMSParKeepAliveClosure::do_oop(oop obj) { 7649 HeapWord* addr = (HeapWord*)obj; 7650 if (_span.contains(addr) && 7651 !_bit_map->isMarked(addr)) { 7652 // In general, during recursive tracing, several threads 7653 // may be concurrently getting here; the first one to 7654 // "tag" it, claims it. 7655 if (_bit_map->par_mark(addr)) { 7656 bool res = _work_queue->push(obj); 7657 assert(res, "Low water mark should be much less than capacity"); 7658 // Do a recursive trim in the hope that this will keep 7659 // stack usage lower, but leave some oops for potential stealers 7660 trim_queue(_low_water_mark); 7661 } // Else, another thread got there first 7662 } 7663 } 7664 7665 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7666 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7667 7668 void CMSParKeepAliveClosure::trim_queue(uint max) { 7669 while (_work_queue->size() > max) { 7670 oop new_oop; 7671 if (_work_queue->pop_local(new_oop)) { 7672 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7673 assert(_bit_map->isMarked((HeapWord*)new_oop), 7674 "no white objects on this stack!"); 7675 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7676 // iterate over the oops in this oop, marking and pushing 7677 // the ones in CMS heap (i.e. in _span). 7678 new_oop->oop_iterate(&_mark_and_push); 7679 } 7680 } 7681 } 7682 7683 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7684 CMSCollector* collector, 7685 MemRegion span, CMSBitMap* bit_map, 7686 OopTaskQueue* work_queue): 7687 _collector(collector), 7688 _span(span), 7689 _bit_map(bit_map), 7690 _work_queue(work_queue) { } 7691 7692 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7693 HeapWord* addr = (HeapWord*)obj; 7694 if (_span.contains(addr) && 7695 !_bit_map->isMarked(addr)) { 7696 if (_bit_map->par_mark(addr)) { 7697 bool simulate_overflow = false; 7698 NOT_PRODUCT( 7699 if (CMSMarkStackOverflowALot && 7700 _collector->par_simulate_overflow()) { 7701 // simulate a stack overflow 7702 simulate_overflow = true; 7703 } 7704 ) 7705 if (simulate_overflow || !_work_queue->push(obj)) { 7706 _collector->par_push_on_overflow_list(obj); 7707 _collector->_par_kac_ovflw++; 7708 } 7709 } // Else another thread got there already 7710 } 7711 } 7712 7713 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7714 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7715 7716 ////////////////////////////////////////////////////////////////// 7717 // CMSExpansionCause ///////////////////////////// 7718 ////////////////////////////////////////////////////////////////// 7719 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7720 switch (cause) { 7721 case _no_expansion: 7722 return "No expansion"; 7723 case _satisfy_free_ratio: 7724 return "Free ratio"; 7725 case _satisfy_promotion: 7726 return "Satisfy promotion"; 7727 case _satisfy_allocation: 7728 return "allocation"; 7729 case _allocate_par_lab: 7730 return "Par LAB"; 7731 case _allocate_par_spooling_space: 7732 return "Par Spooling Space"; 7733 case _adaptive_size_policy: 7734 return "Ergonomics"; 7735 default: 7736 return "unknown"; 7737 } 7738 } 7739 7740 void CMSDrainMarkingStackClosure::do_void() { 7741 // the max number to take from overflow list at a time 7742 const size_t num = _mark_stack->capacity()/4; 7743 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7744 "Overflow list should be NULL during concurrent phases"); 7745 while (!_mark_stack->isEmpty() || 7746 // if stack is empty, check the overflow list 7747 _collector->take_from_overflow_list(num, _mark_stack)) { 7748 oop obj = _mark_stack->pop(); 7749 HeapWord* addr = (HeapWord*)obj; 7750 assert(_span.contains(addr), "Should be within span"); 7751 assert(_bit_map->isMarked(addr), "Should be marked"); 7752 assert(oopDesc::is_oop(obj), "Should be an oop"); 7753 obj->oop_iterate(_keep_alive); 7754 } 7755 } 7756 7757 void CMSParDrainMarkingStackClosure::do_void() { 7758 // drain queue 7759 trim_queue(0); 7760 } 7761 7762 // Trim our work_queue so its length is below max at return 7763 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7764 while (_work_queue->size() > max) { 7765 oop new_oop; 7766 if (_work_queue->pop_local(new_oop)) { 7767 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7768 assert(_bit_map->isMarked((HeapWord*)new_oop), 7769 "no white objects on this stack!"); 7770 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7771 // iterate over the oops in this oop, marking and pushing 7772 // the ones in CMS heap (i.e. in _span). 7773 new_oop->oop_iterate(&_mark_and_push); 7774 } 7775 } 7776 } 7777 7778 //////////////////////////////////////////////////////////////////// 7779 // Support for Marking Stack Overflow list handling and related code 7780 //////////////////////////////////////////////////////////////////// 7781 // Much of the following code is similar in shape and spirit to the 7782 // code used in ParNewGC. We should try and share that code 7783 // as much as possible in the future. 7784 7785 #ifndef PRODUCT 7786 // Debugging support for CMSStackOverflowALot 7787 7788 // It's OK to call this multi-threaded; the worst thing 7789 // that can happen is that we'll get a bunch of closely 7790 // spaced simulated overflows, but that's OK, in fact 7791 // probably good as it would exercise the overflow code 7792 // under contention. 7793 bool CMSCollector::simulate_overflow() { 7794 if (_overflow_counter-- <= 0) { // just being defensive 7795 _overflow_counter = CMSMarkStackOverflowInterval; 7796 return true; 7797 } else { 7798 return false; 7799 } 7800 } 7801 7802 bool CMSCollector::par_simulate_overflow() { 7803 return simulate_overflow(); 7804 } 7805 #endif 7806 7807 // Single-threaded 7808 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7809 assert(stack->isEmpty(), "Expected precondition"); 7810 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7811 size_t i = num; 7812 oop cur = _overflow_list; 7813 const markOop proto = markOopDesc::prototype(); 7814 NOT_PRODUCT(ssize_t n = 0;) 7815 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7816 next = oop(cur->mark()); 7817 cur->set_mark(proto); // until proven otherwise 7818 assert(oopDesc::is_oop(cur), "Should be an oop"); 7819 bool res = stack->push(cur); 7820 assert(res, "Bit off more than can chew?"); 7821 NOT_PRODUCT(n++;) 7822 } 7823 _overflow_list = cur; 7824 #ifndef PRODUCT 7825 assert(_num_par_pushes >= n, "Too many pops?"); 7826 _num_par_pushes -=n; 7827 #endif 7828 return !stack->isEmpty(); 7829 } 7830 7831 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7832 // (MT-safe) Get a prefix of at most "num" from the list. 7833 // The overflow list is chained through the mark word of 7834 // each object in the list. We fetch the entire list, 7835 // break off a prefix of the right size and return the 7836 // remainder. If other threads try to take objects from 7837 // the overflow list at that time, they will wait for 7838 // some time to see if data becomes available. If (and 7839 // only if) another thread places one or more object(s) 7840 // on the global list before we have returned the suffix 7841 // to the global list, we will walk down our local list 7842 // to find its end and append the global list to 7843 // our suffix before returning it. This suffix walk can 7844 // prove to be expensive (quadratic in the amount of traffic) 7845 // when there are many objects in the overflow list and 7846 // there is much producer-consumer contention on the list. 7847 // *NOTE*: The overflow list manipulation code here and 7848 // in ParNewGeneration:: are very similar in shape, 7849 // except that in the ParNew case we use the old (from/eden) 7850 // copy of the object to thread the list via its klass word. 7851 // Because of the common code, if you make any changes in 7852 // the code below, please check the ParNew version to see if 7853 // similar changes might be needed. 7854 // CR 6797058 has been filed to consolidate the common code. 7855 bool CMSCollector::par_take_from_overflow_list(size_t num, 7856 OopTaskQueue* work_q, 7857 int no_of_gc_threads) { 7858 assert(work_q->size() == 0, "First empty local work queue"); 7859 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7860 if (_overflow_list == NULL) { 7861 return false; 7862 } 7863 // Grab the entire list; we'll put back a suffix 7864 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7865 Thread* tid = Thread::current(); 7866 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7867 // set to ParallelGCThreads. 7868 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7869 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7870 // If the list is busy, we spin for a short while, 7871 // sleeping between attempts to get the list. 7872 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7873 os::sleep(tid, sleep_time_millis, false); 7874 if (_overflow_list == NULL) { 7875 // Nothing left to take 7876 return false; 7877 } else if (_overflow_list != BUSY) { 7878 // Try and grab the prefix 7879 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7880 } 7881 } 7882 // If the list was found to be empty, or we spun long 7883 // enough, we give up and return empty-handed. If we leave 7884 // the list in the BUSY state below, it must be the case that 7885 // some other thread holds the overflow list and will set it 7886 // to a non-BUSY state in the future. 7887 if (prefix == NULL || prefix == BUSY) { 7888 // Nothing to take or waited long enough 7889 if (prefix == NULL) { 7890 // Write back the NULL in case we overwrote it with BUSY above 7891 // and it is still the same value. 7892 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7893 } 7894 return false; 7895 } 7896 assert(prefix != NULL && prefix != BUSY, "Error"); 7897 size_t i = num; 7898 oop cur = prefix; 7899 // Walk down the first "num" objects, unless we reach the end. 7900 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7901 if (cur->mark() == NULL) { 7902 // We have "num" or fewer elements in the list, so there 7903 // is nothing to return to the global list. 7904 // Write back the NULL in lieu of the BUSY we wrote 7905 // above, if it is still the same value. 7906 if (_overflow_list == BUSY) { 7907 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7908 } 7909 } else { 7910 // Chop off the suffix and return it to the global list. 7911 assert(cur->mark() != BUSY, "Error"); 7912 oop suffix_head = cur->mark(); // suffix will be put back on global list 7913 cur->set_mark(NULL); // break off suffix 7914 // It's possible that the list is still in the empty(busy) state 7915 // we left it in a short while ago; in that case we may be 7916 // able to place back the suffix without incurring the cost 7917 // of a walk down the list. 7918 oop observed_overflow_list = _overflow_list; 7919 oop cur_overflow_list = observed_overflow_list; 7920 bool attached = false; 7921 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7922 observed_overflow_list = 7923 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7924 if (cur_overflow_list == observed_overflow_list) { 7925 attached = true; 7926 break; 7927 } else cur_overflow_list = observed_overflow_list; 7928 } 7929 if (!attached) { 7930 // Too bad, someone else sneaked in (at least) an element; we'll need 7931 // to do a splice. Find tail of suffix so we can prepend suffix to global 7932 // list. 7933 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7934 oop suffix_tail = cur; 7935 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7936 "Tautology"); 7937 observed_overflow_list = _overflow_list; 7938 do { 7939 cur_overflow_list = observed_overflow_list; 7940 if (cur_overflow_list != BUSY) { 7941 // Do the splice ... 7942 suffix_tail->set_mark(markOop(cur_overflow_list)); 7943 } else { // cur_overflow_list == BUSY 7944 suffix_tail->set_mark(NULL); 7945 } 7946 // ... and try to place spliced list back on overflow_list ... 7947 observed_overflow_list = 7948 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7949 } while (cur_overflow_list != observed_overflow_list); 7950 // ... until we have succeeded in doing so. 7951 } 7952 } 7953 7954 // Push the prefix elements on work_q 7955 assert(prefix != NULL, "control point invariant"); 7956 const markOop proto = markOopDesc::prototype(); 7957 oop next; 7958 NOT_PRODUCT(ssize_t n = 0;) 7959 for (cur = prefix; cur != NULL; cur = next) { 7960 next = oop(cur->mark()); 7961 cur->set_mark(proto); // until proven otherwise 7962 assert(oopDesc::is_oop(cur), "Should be an oop"); 7963 bool res = work_q->push(cur); 7964 assert(res, "Bit off more than we can chew?"); 7965 NOT_PRODUCT(n++;) 7966 } 7967 #ifndef PRODUCT 7968 assert(_num_par_pushes >= n, "Too many pops?"); 7969 Atomic::sub(n, &_num_par_pushes); 7970 #endif 7971 return true; 7972 } 7973 7974 // Single-threaded 7975 void CMSCollector::push_on_overflow_list(oop p) { 7976 NOT_PRODUCT(_num_par_pushes++;) 7977 assert(oopDesc::is_oop(p), "Not an oop"); 7978 preserve_mark_if_necessary(p); 7979 p->set_mark((markOop)_overflow_list); 7980 _overflow_list = p; 7981 } 7982 7983 // Multi-threaded; use CAS to prepend to overflow list 7984 void CMSCollector::par_push_on_overflow_list(oop p) { 7985 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7986 assert(oopDesc::is_oop(p), "Not an oop"); 7987 par_preserve_mark_if_necessary(p); 7988 oop observed_overflow_list = _overflow_list; 7989 oop cur_overflow_list; 7990 do { 7991 cur_overflow_list = observed_overflow_list; 7992 if (cur_overflow_list != BUSY) { 7993 p->set_mark(markOop(cur_overflow_list)); 7994 } else { 7995 p->set_mark(NULL); 7996 } 7997 observed_overflow_list = 7998 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7999 } while (cur_overflow_list != observed_overflow_list); 8000 } 8001 #undef BUSY 8002 8003 // Single threaded 8004 // General Note on GrowableArray: pushes may silently fail 8005 // because we are (temporarily) out of C-heap for expanding 8006 // the stack. The problem is quite ubiquitous and affects 8007 // a lot of code in the JVM. The prudent thing for GrowableArray 8008 // to do (for now) is to exit with an error. However, that may 8009 // be too draconian in some cases because the caller may be 8010 // able to recover without much harm. For such cases, we 8011 // should probably introduce a "soft_push" method which returns 8012 // an indication of success or failure with the assumption that 8013 // the caller may be able to recover from a failure; code in 8014 // the VM can then be changed, incrementally, to deal with such 8015 // failures where possible, thus, incrementally hardening the VM 8016 // in such low resource situations. 8017 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8018 _preserved_oop_stack.push(p); 8019 _preserved_mark_stack.push(m); 8020 assert(m == p->mark(), "Mark word changed"); 8021 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8022 "bijection"); 8023 } 8024 8025 // Single threaded 8026 void CMSCollector::preserve_mark_if_necessary(oop p) { 8027 markOop m = p->mark(); 8028 if (m->must_be_preserved(p)) { 8029 preserve_mark_work(p, m); 8030 } 8031 } 8032 8033 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8034 markOop m = p->mark(); 8035 if (m->must_be_preserved(p)) { 8036 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8037 // Even though we read the mark word without holding 8038 // the lock, we are assured that it will not change 8039 // because we "own" this oop, so no other thread can 8040 // be trying to push it on the overflow list; see 8041 // the assertion in preserve_mark_work() that checks 8042 // that m == p->mark(). 8043 preserve_mark_work(p, m); 8044 } 8045 } 8046 8047 // We should be able to do this multi-threaded, 8048 // a chunk of stack being a task (this is 8049 // correct because each oop only ever appears 8050 // once in the overflow list. However, it's 8051 // not very easy to completely overlap this with 8052 // other operations, so will generally not be done 8053 // until all work's been completed. Because we 8054 // expect the preserved oop stack (set) to be small, 8055 // it's probably fine to do this single-threaded. 8056 // We can explore cleverer concurrent/overlapped/parallel 8057 // processing of preserved marks if we feel the 8058 // need for this in the future. Stack overflow should 8059 // be so rare in practice and, when it happens, its 8060 // effect on performance so great that this will 8061 // likely just be in the noise anyway. 8062 void CMSCollector::restore_preserved_marks_if_any() { 8063 assert(SafepointSynchronize::is_at_safepoint(), 8064 "world should be stopped"); 8065 assert(Thread::current()->is_ConcurrentGC_thread() || 8066 Thread::current()->is_VM_thread(), 8067 "should be single-threaded"); 8068 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8069 "bijection"); 8070 8071 while (!_preserved_oop_stack.is_empty()) { 8072 oop p = _preserved_oop_stack.pop(); 8073 assert(oopDesc::is_oop(p), "Should be an oop"); 8074 assert(_span.contains(p), "oop should be in _span"); 8075 assert(p->mark() == markOopDesc::prototype(), 8076 "Set when taken from overflow list"); 8077 markOop m = _preserved_mark_stack.pop(); 8078 p->set_mark(m); 8079 } 8080 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8081 "stacks were cleared above"); 8082 } 8083 8084 #ifndef PRODUCT 8085 bool CMSCollector::no_preserved_marks() const { 8086 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8087 } 8088 #endif 8089 8090 // Transfer some number of overflown objects to usual marking 8091 // stack. Return true if some objects were transferred. 8092 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8093 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8094 (size_t)ParGCDesiredObjsFromOverflowList); 8095 8096 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8097 assert(_collector->overflow_list_is_empty() || res, 8098 "If list is not empty, we should have taken something"); 8099 assert(!res || !_mark_stack->isEmpty(), 8100 "If we took something, it should now be on our stack"); 8101 return res; 8102 } 8103 8104 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8105 size_t res = _sp->block_size_no_stall(addr, _collector); 8106 if (_sp->block_is_obj(addr)) { 8107 if (_live_bit_map->isMarked(addr)) { 8108 // It can't have been dead in a previous cycle 8109 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8110 } else { 8111 _dead_bit_map->mark(addr); // mark the dead object 8112 } 8113 } 8114 // Could be 0, if the block size could not be computed without stalling. 8115 return res; 8116 } 8117 8118 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8119 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8120 switch (phase) { 8121 case CMSCollector::InitialMarking: 8122 initialize(manager /* GC manager */ , 8123 cause /* cause of the GC */, 8124 true /* recordGCBeginTime */, 8125 true /* recordPreGCUsage */, 8126 false /* recordPeakUsage */, 8127 false /* recordPostGCusage */, 8128 true /* recordAccumulatedGCTime */, 8129 false /* recordGCEndTime */, 8130 false /* countCollection */ ); 8131 break; 8132 8133 case CMSCollector::FinalMarking: 8134 initialize(manager /* GC manager */ , 8135 cause /* cause of the GC */, 8136 false /* recordGCBeginTime */, 8137 false /* recordPreGCUsage */, 8138 false /* recordPeakUsage */, 8139 false /* recordPostGCusage */, 8140 true /* recordAccumulatedGCTime */, 8141 false /* recordGCEndTime */, 8142 false /* countCollection */ ); 8143 break; 8144 8145 case CMSCollector::Sweeping: 8146 initialize(manager /* GC manager */ , 8147 cause /* cause of the GC */, 8148 false /* recordGCBeginTime */, 8149 false /* recordPreGCUsage */, 8150 true /* recordPeakUsage */, 8151 true /* recordPostGCusage */, 8152 false /* recordAccumulatedGCTime */, 8153 true /* recordGCEndTime */, 8154 true /* countCollection */ ); 8155 break; 8156 8157 default: 8158 ShouldNotReachHere(); 8159 } 8160 } --- EOF ---