1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.inline.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->print(" "); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 log_warning(gc)("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 log_warning(gc)("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 549 // Parallel task queues; these are shared for the 550 // concurrent and stop-world phases of CMS, but 551 // are not shared with parallel scavenge (ParNew). 552 { 553 uint i; 554 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 555 556 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 557 || ParallelRefProcEnabled) 558 && num_queues > 0) { 559 _task_queues = new OopTaskQueueSet(num_queues); 560 if (_task_queues == NULL) { 561 log_warning(gc)("task_queues allocation failure."); 562 return; 563 } 564 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 565 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 566 for (i = 0; i < num_queues; i++) { 567 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 568 if (q == NULL) { 569 log_warning(gc)("work_queue allocation failure."); 570 return; 571 } 572 _task_queues->register_queue(i, q); 573 } 574 for (i = 0; i < num_queues; i++) { 575 _task_queues->queue(i)->initialize(); 576 _hash_seed[i] = 17; // copied from ParNew 577 } 578 } 579 } 580 581 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 582 583 // Clip CMSBootstrapOccupancy between 0 and 100. 584 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 585 586 // Now tell CMS generations the identity of their collector 587 ConcurrentMarkSweepGeneration::set_collector(this); 588 589 // Create & start a CMS thread for this CMS collector 590 _cmsThread = ConcurrentMarkSweepThread::start(this); 591 assert(cmsThread() != NULL, "CMS Thread should have been created"); 592 assert(cmsThread()->collector() == this, 593 "CMS Thread should refer to this gen"); 594 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 595 596 // Support for parallelizing young gen rescan 597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 598 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 599 _young_gen = (ParNewGeneration*)gch->young_gen(); 600 if (gch->supports_inline_contig_alloc()) { 601 _top_addr = gch->top_addr(); 602 _end_addr = gch->end_addr(); 603 assert(_young_gen != NULL, "no _young_gen"); 604 _eden_chunk_index = 0; 605 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 606 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 607 } 608 609 // Support for parallelizing survivor space rescan 610 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 611 const size_t max_plab_samples = 612 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 613 614 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 615 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 616 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 617 _survivor_chunk_capacity = max_plab_samples; 618 for (uint i = 0; i < ParallelGCThreads; i++) { 619 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 620 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 621 assert(cur->end() == 0, "Should be 0"); 622 assert(cur->array() == vec, "Should be vec"); 623 assert(cur->capacity() == max_plab_samples, "Error"); 624 } 625 } 626 627 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 628 _gc_counters = new CollectorCounters("CMS", 1); 629 _completed_initialization = true; 630 _inter_sweep_timer.start(); // start of time 631 } 632 633 const char* ConcurrentMarkSweepGeneration::name() const { 634 return "concurrent mark-sweep generation"; 635 } 636 void ConcurrentMarkSweepGeneration::update_counters() { 637 if (UsePerfData) { 638 _space_counters->update_all(); 639 _gen_counters->update_all(); 640 } 641 } 642 643 // this is an optimized version of update_counters(). it takes the 644 // used value as a parameter rather than computing it. 645 // 646 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 647 if (UsePerfData) { 648 _space_counters->update_used(used); 649 _space_counters->update_capacity(); 650 _gen_counters->update_all(); 651 } 652 } 653 654 void ConcurrentMarkSweepGeneration::print() const { 655 Generation::print(); 656 cmsSpace()->print(); 657 } 658 659 #ifndef PRODUCT 660 void ConcurrentMarkSweepGeneration::print_statistics() { 661 cmsSpace()->printFLCensus(0); 662 } 663 #endif 664 665 size_t 666 ConcurrentMarkSweepGeneration::contiguous_available() const { 667 // dld proposes an improvement in precision here. If the committed 668 // part of the space ends in a free block we should add that to 669 // uncommitted size in the calculation below. Will make this 670 // change later, staying with the approximation below for the 671 // time being. -- ysr. 672 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 673 } 674 675 size_t 676 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 677 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 678 } 679 680 size_t ConcurrentMarkSweepGeneration::max_available() const { 681 return free() + _virtual_space.uncommitted_size(); 682 } 683 684 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 685 size_t available = max_available(); 686 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 687 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 688 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 689 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 690 return res; 691 } 692 693 // At a promotion failure dump information on block layout in heap 694 // (cms old generation). 695 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 696 LogHandle(gc, promotion) log; 697 if (log.is_trace()) { 698 ResourceMark rm; 699 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 700 } 701 } 702 703 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 704 // Clear the promotion information. These pointers can be adjusted 705 // along with all the other pointers into the heap but 706 // compaction is expected to be a rare event with 707 // a heap using cms so don't do it without seeing the need. 708 for (uint i = 0; i < ParallelGCThreads; i++) { 709 _par_gc_thread_states[i]->promo.reset(); 710 } 711 } 712 713 void ConcurrentMarkSweepGeneration::compute_new_size() { 714 assert_locked_or_safepoint(Heap_lock); 715 716 // If incremental collection failed, we just want to expand 717 // to the limit. 718 if (incremental_collection_failed()) { 719 clear_incremental_collection_failed(); 720 grow_to_reserved(); 721 return; 722 } 723 724 // The heap has been compacted but not reset yet. 725 // Any metric such as free() or used() will be incorrect. 726 727 CardGeneration::compute_new_size(); 728 729 // Reset again after a possible resizing 730 if (did_compact()) { 731 cmsSpace()->reset_after_compaction(); 732 } 733 } 734 735 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 736 assert_locked_or_safepoint(Heap_lock); 737 738 // If incremental collection failed, we just want to expand 739 // to the limit. 740 if (incremental_collection_failed()) { 741 clear_incremental_collection_failed(); 742 grow_to_reserved(); 743 return; 744 } 745 746 double free_percentage = ((double) free()) / capacity(); 747 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 748 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 749 750 // compute expansion delta needed for reaching desired free percentage 751 if (free_percentage < desired_free_percentage) { 752 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 753 assert(desired_capacity >= capacity(), "invalid expansion size"); 754 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 755 LogHandle(gc) log; 756 if (log.is_trace()) { 757 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 758 log.trace("From compute_new_size: "); 759 log.trace(" Free fraction %f", free_percentage); 760 log.trace(" Desired free fraction %f", desired_free_percentage); 761 log.trace(" Maximum free fraction %f", maximum_free_percentage); 762 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 763 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 764 GenCollectedHeap* gch = GenCollectedHeap::heap(); 765 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 766 size_t young_size = gch->young_gen()->capacity(); 767 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 768 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 769 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 770 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 771 } 772 // safe if expansion fails 773 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 774 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 775 } else { 776 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 777 assert(desired_capacity <= capacity(), "invalid expansion size"); 778 size_t shrink_bytes = capacity() - desired_capacity; 779 // Don't shrink unless the delta is greater than the minimum shrink we want 780 if (shrink_bytes >= MinHeapDeltaBytes) { 781 shrink_free_list_by(shrink_bytes); 782 } 783 } 784 } 785 786 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 787 return cmsSpace()->freelistLock(); 788 } 789 790 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 791 CMSSynchronousYieldRequest yr; 792 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 793 return have_lock_and_allocate(size, tlab); 794 } 795 796 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 797 bool tlab /* ignored */) { 798 assert_lock_strong(freelistLock()); 799 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 800 HeapWord* res = cmsSpace()->allocate(adjustedSize); 801 // Allocate the object live (grey) if the background collector has 802 // started marking. This is necessary because the marker may 803 // have passed this address and consequently this object will 804 // not otherwise be greyed and would be incorrectly swept up. 805 // Note that if this object contains references, the writing 806 // of those references will dirty the card containing this object 807 // allowing the object to be blackened (and its references scanned) 808 // either during a preclean phase or at the final checkpoint. 809 if (res != NULL) { 810 // We may block here with an uninitialized object with 811 // its mark-bit or P-bits not yet set. Such objects need 812 // to be safely navigable by block_start(). 813 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 814 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 815 collector()->direct_allocated(res, adjustedSize); 816 _direct_allocated_words += adjustedSize; 817 // allocation counters 818 NOT_PRODUCT( 819 _numObjectsAllocated++; 820 _numWordsAllocated += (int)adjustedSize; 821 ) 822 } 823 return res; 824 } 825 826 // In the case of direct allocation by mutators in a generation that 827 // is being concurrently collected, the object must be allocated 828 // live (grey) if the background collector has started marking. 829 // This is necessary because the marker may 830 // have passed this address and consequently this object will 831 // not otherwise be greyed and would be incorrectly swept up. 832 // Note that if this object contains references, the writing 833 // of those references will dirty the card containing this object 834 // allowing the object to be blackened (and its references scanned) 835 // either during a preclean phase or at the final checkpoint. 836 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 837 assert(_markBitMap.covers(start, size), "Out of bounds"); 838 if (_collectorState >= Marking) { 839 MutexLockerEx y(_markBitMap.lock(), 840 Mutex::_no_safepoint_check_flag); 841 // [see comments preceding SweepClosure::do_blk() below for details] 842 // 843 // Can the P-bits be deleted now? JJJ 844 // 845 // 1. need to mark the object as live so it isn't collected 846 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 847 // 3. need to mark the end of the object so marking, precleaning or sweeping 848 // can skip over uninitialized or unparsable objects. An allocated 849 // object is considered uninitialized for our purposes as long as 850 // its klass word is NULL. All old gen objects are parsable 851 // as soon as they are initialized.) 852 _markBitMap.mark(start); // object is live 853 _markBitMap.mark(start + 1); // object is potentially uninitialized? 854 _markBitMap.mark(start + size - 1); 855 // mark end of object 856 } 857 // check that oop looks uninitialized 858 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 859 } 860 861 void CMSCollector::promoted(bool par, HeapWord* start, 862 bool is_obj_array, size_t obj_size) { 863 assert(_markBitMap.covers(start), "Out of bounds"); 864 // See comment in direct_allocated() about when objects should 865 // be allocated live. 866 if (_collectorState >= Marking) { 867 // we already hold the marking bit map lock, taken in 868 // the prologue 869 if (par) { 870 _markBitMap.par_mark(start); 871 } else { 872 _markBitMap.mark(start); 873 } 874 // We don't need to mark the object as uninitialized (as 875 // in direct_allocated above) because this is being done with the 876 // world stopped and the object will be initialized by the 877 // time the marking, precleaning or sweeping get to look at it. 878 // But see the code for copying objects into the CMS generation, 879 // where we need to ensure that concurrent readers of the 880 // block offset table are able to safely navigate a block that 881 // is in flux from being free to being allocated (and in 882 // transition while being copied into) and subsequently 883 // becoming a bona-fide object when the copy/promotion is complete. 884 assert(SafepointSynchronize::is_at_safepoint(), 885 "expect promotion only at safepoints"); 886 887 if (_collectorState < Sweeping) { 888 // Mark the appropriate cards in the modUnionTable, so that 889 // this object gets scanned before the sweep. If this is 890 // not done, CMS generation references in the object might 891 // not get marked. 892 // For the case of arrays, which are otherwise precisely 893 // marked, we need to dirty the entire array, not just its head. 894 if (is_obj_array) { 895 // The [par_]mark_range() method expects mr.end() below to 896 // be aligned to the granularity of a bit's representation 897 // in the heap. In the case of the MUT below, that's a 898 // card size. 899 MemRegion mr(start, 900 (HeapWord*)round_to((intptr_t)(start + obj_size), 901 CardTableModRefBS::card_size /* bytes */)); 902 if (par) { 903 _modUnionTable.par_mark_range(mr); 904 } else { 905 _modUnionTable.mark_range(mr); 906 } 907 } else { // not an obj array; we can just mark the head 908 if (par) { 909 _modUnionTable.par_mark(start); 910 } else { 911 _modUnionTable.mark(start); 912 } 913 } 914 } 915 } 916 } 917 918 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 919 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 920 // allocate, copy and if necessary update promoinfo -- 921 // delegate to underlying space. 922 assert_lock_strong(freelistLock()); 923 924 #ifndef PRODUCT 925 if (GenCollectedHeap::heap()->promotion_should_fail()) { 926 return NULL; 927 } 928 #endif // #ifndef PRODUCT 929 930 oop res = _cmsSpace->promote(obj, obj_size); 931 if (res == NULL) { 932 // expand and retry 933 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 934 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 935 // Since this is the old generation, we don't try to promote 936 // into a more senior generation. 937 res = _cmsSpace->promote(obj, obj_size); 938 } 939 if (res != NULL) { 940 // See comment in allocate() about when objects should 941 // be allocated live. 942 assert(obj->is_oop(), "Will dereference klass pointer below"); 943 collector()->promoted(false, // Not parallel 944 (HeapWord*)res, obj->is_objArray(), obj_size); 945 // promotion counters 946 NOT_PRODUCT( 947 _numObjectsPromoted++; 948 _numWordsPromoted += 949 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 950 ) 951 } 952 return res; 953 } 954 955 956 // IMPORTANT: Notes on object size recognition in CMS. 957 // --------------------------------------------------- 958 // A block of storage in the CMS generation is always in 959 // one of three states. A free block (FREE), an allocated 960 // object (OBJECT) whose size() method reports the correct size, 961 // and an intermediate state (TRANSIENT) in which its size cannot 962 // be accurately determined. 963 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 964 // ----------------------------------------------------- 965 // FREE: klass_word & 1 == 1; mark_word holds block size 966 // 967 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 968 // obj->size() computes correct size 969 // 970 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 971 // 972 // STATE IDENTIFICATION: (64 bit+COOPS) 973 // ------------------------------------ 974 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 975 // 976 // OBJECT: klass_word installed; klass_word != 0; 977 // obj->size() computes correct size 978 // 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 980 // 981 // 982 // STATE TRANSITION DIAGRAM 983 // 984 // mut / parnew mut / parnew 985 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 986 // ^ | 987 // |------------------------ DEAD <------------------------------------| 988 // sweep mut 989 // 990 // While a block is in TRANSIENT state its size cannot be determined 991 // so readers will either need to come back later or stall until 992 // the size can be determined. Note that for the case of direct 993 // allocation, P-bits, when available, may be used to determine the 994 // size of an object that may not yet have been initialized. 995 996 // Things to support parallel young-gen collection. 997 oop 998 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 999 oop old, markOop m, 1000 size_t word_sz) { 1001 #ifndef PRODUCT 1002 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1003 return NULL; 1004 } 1005 #endif // #ifndef PRODUCT 1006 1007 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1008 PromotionInfo* promoInfo = &ps->promo; 1009 // if we are tracking promotions, then first ensure space for 1010 // promotion (including spooling space for saving header if necessary). 1011 // then allocate and copy, then track promoted info if needed. 1012 // When tracking (see PromotionInfo::track()), the mark word may 1013 // be displaced and in this case restoration of the mark word 1014 // occurs in the (oop_since_save_marks_)iterate phase. 1015 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1016 // Out of space for allocating spooling buffers; 1017 // try expanding and allocating spooling buffers. 1018 if (!expand_and_ensure_spooling_space(promoInfo)) { 1019 return NULL; 1020 } 1021 } 1022 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1023 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1024 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1025 if (obj_ptr == NULL) { 1026 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1027 if (obj_ptr == NULL) { 1028 return NULL; 1029 } 1030 } 1031 oop obj = oop(obj_ptr); 1032 OrderAccess::storestore(); 1033 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1034 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1035 // IMPORTANT: See note on object initialization for CMS above. 1036 // Otherwise, copy the object. Here we must be careful to insert the 1037 // klass pointer last, since this marks the block as an allocated object. 1038 // Except with compressed oops it's the mark word. 1039 HeapWord* old_ptr = (HeapWord*)old; 1040 // Restore the mark word copied above. 1041 obj->set_mark(m); 1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1044 OrderAccess::storestore(); 1045 1046 if (UseCompressedClassPointers) { 1047 // Copy gap missed by (aligned) header size calculation below 1048 obj->set_klass_gap(old->klass_gap()); 1049 } 1050 if (word_sz > (size_t)oopDesc::header_size()) { 1051 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1052 obj_ptr + oopDesc::header_size(), 1053 word_sz - oopDesc::header_size()); 1054 } 1055 1056 // Now we can track the promoted object, if necessary. We take care 1057 // to delay the transition from uninitialized to full object 1058 // (i.e., insertion of klass pointer) until after, so that it 1059 // atomically becomes a promoted object. 1060 if (promoInfo->tracking()) { 1061 promoInfo->track((PromotedObject*)obj, old->klass()); 1062 } 1063 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1064 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1065 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1066 1067 // Finally, install the klass pointer (this should be volatile). 1068 OrderAccess::storestore(); 1069 obj->set_klass(old->klass()); 1070 // We should now be able to calculate the right size for this object 1071 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1072 1073 collector()->promoted(true, // parallel 1074 obj_ptr, old->is_objArray(), word_sz); 1075 1076 NOT_PRODUCT( 1077 Atomic::inc_ptr(&_numObjectsPromoted); 1078 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1079 ) 1080 1081 return obj; 1082 } 1083 1084 void 1085 ConcurrentMarkSweepGeneration:: 1086 par_promote_alloc_done(int thread_num) { 1087 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1088 ps->lab.retire(thread_num); 1089 } 1090 1091 void 1092 ConcurrentMarkSweepGeneration:: 1093 par_oop_since_save_marks_iterate_done(int thread_num) { 1094 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1095 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1096 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1097 } 1098 1099 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1100 size_t size, 1101 bool tlab) 1102 { 1103 // We allow a STW collection only if a full 1104 // collection was requested. 1105 return full || should_allocate(size, tlab); // FIX ME !!! 1106 // This and promotion failure handling are connected at the 1107 // hip and should be fixed by untying them. 1108 } 1109 1110 bool CMSCollector::shouldConcurrentCollect() { 1111 if (_full_gc_requested) { 1112 log_trace(gc)("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1113 return true; 1114 } 1115 1116 FreelistLocker x(this); 1117 // ------------------------------------------------------------------ 1118 // Print out lots of information which affects the initiation of 1119 // a collection. 1120 LogHandle(gc) log; 1121 if (log.is_trace() && stats().valid()) { 1122 log.trace("CMSCollector shouldConcurrentCollect: "); 1123 ResourceMark rm; 1124 stats().print_on(log.debug_stream()); 1125 log.trace("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1126 log.trace("free=" SIZE_FORMAT, _cmsGen->free()); 1127 log.trace("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1128 log.trace("promotion_rate=%g", stats().promotion_rate()); 1129 log.trace("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1130 log.trace("occupancy=%3.7f", _cmsGen->occupancy()); 1131 log.trace("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1132 log.trace("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1133 log.trace("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1134 log.trace("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1135 } 1136 // ------------------------------------------------------------------ 1137 1138 // If the estimated time to complete a cms collection (cms_duration()) 1139 // is less than the estimated time remaining until the cms generation 1140 // is full, start a collection. 1141 if (!UseCMSInitiatingOccupancyOnly) { 1142 if (stats().valid()) { 1143 if (stats().time_until_cms_start() == 0.0) { 1144 return true; 1145 } 1146 } else { 1147 // We want to conservatively collect somewhat early in order 1148 // to try and "bootstrap" our CMS/promotion statistics; 1149 // this branch will not fire after the first successful CMS 1150 // collection because the stats should then be valid. 1151 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1152 log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1153 _cmsGen->occupancy(), _bootstrap_occupancy); 1154 return true; 1155 } 1156 } 1157 } 1158 1159 // Otherwise, we start a collection cycle if 1160 // old gen want a collection cycle started. Each may use 1161 // an appropriate criterion for making this decision. 1162 // XXX We need to make sure that the gen expansion 1163 // criterion dovetails well with this. XXX NEED TO FIX THIS 1164 if (_cmsGen->should_concurrent_collect()) { 1165 log_trace(gc)("CMS old gen initiated"); 1166 return true; 1167 } 1168 1169 // We start a collection if we believe an incremental collection may fail; 1170 // this is not likely to be productive in practice because it's probably too 1171 // late anyway. 1172 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1173 assert(gch->collector_policy()->is_generation_policy(), 1174 "You may want to check the correctness of the following"); 1175 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1176 log_trace(gc)("CMSCollector: collect because incremental collection will fail "); 1177 return true; 1178 } 1179 1180 if (MetaspaceGC::should_concurrent_collect()) { 1181 log_trace(gc)("CMSCollector: collect for metadata allocation "); 1182 return true; 1183 } 1184 1185 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1186 if (CMSTriggerInterval >= 0) { 1187 if (CMSTriggerInterval == 0) { 1188 // Trigger always 1189 return true; 1190 } 1191 1192 // Check the CMS time since begin (we do not check the stats validity 1193 // as we want to be able to trigger the first CMS cycle as well) 1194 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1195 if (stats().valid()) { 1196 log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1197 stats().cms_time_since_begin()); 1198 } else { 1199 log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)"); 1200 } 1201 return true; 1202 } 1203 } 1204 1205 return false; 1206 } 1207 1208 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1209 1210 // Clear _expansion_cause fields of constituent generations 1211 void CMSCollector::clear_expansion_cause() { 1212 _cmsGen->clear_expansion_cause(); 1213 } 1214 1215 // We should be conservative in starting a collection cycle. To 1216 // start too eagerly runs the risk of collecting too often in the 1217 // extreme. To collect too rarely falls back on full collections, 1218 // which works, even if not optimum in terms of concurrent work. 1219 // As a work around for too eagerly collecting, use the flag 1220 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1221 // giving the user an easily understandable way of controlling the 1222 // collections. 1223 // We want to start a new collection cycle if any of the following 1224 // conditions hold: 1225 // . our current occupancy exceeds the configured initiating occupancy 1226 // for this generation, or 1227 // . we recently needed to expand this space and have not, since that 1228 // expansion, done a collection of this generation, or 1229 // . the underlying space believes that it may be a good idea to initiate 1230 // a concurrent collection (this may be based on criteria such as the 1231 // following: the space uses linear allocation and linear allocation is 1232 // going to fail, or there is believed to be excessive fragmentation in 1233 // the generation, etc... or ... 1234 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1235 // the case of the old generation; see CR 6543076): 1236 // we may be approaching a point at which allocation requests may fail because 1237 // we will be out of sufficient free space given allocation rate estimates.] 1238 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1239 1240 assert_lock_strong(freelistLock()); 1241 if (occupancy() > initiating_occupancy()) { 1242 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1243 short_name(), occupancy(), initiating_occupancy()); 1244 return true; 1245 } 1246 if (UseCMSInitiatingOccupancyOnly) { 1247 return false; 1248 } 1249 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1250 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1251 return true; 1252 } 1253 return false; 1254 } 1255 1256 void ConcurrentMarkSweepGeneration::collect(bool full, 1257 bool clear_all_soft_refs, 1258 size_t size, 1259 bool tlab) 1260 { 1261 collector()->collect(full, clear_all_soft_refs, size, tlab); 1262 } 1263 1264 void CMSCollector::collect(bool full, 1265 bool clear_all_soft_refs, 1266 size_t size, 1267 bool tlab) 1268 { 1269 // The following "if" branch is present for defensive reasons. 1270 // In the current uses of this interface, it can be replaced with: 1271 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1272 // But I am not placing that assert here to allow future 1273 // generality in invoking this interface. 1274 if (GCLocker::is_active()) { 1275 // A consistency test for GCLocker 1276 assert(GCLocker::needs_gc(), "Should have been set already"); 1277 // Skip this foreground collection, instead 1278 // expanding the heap if necessary. 1279 // Need the free list locks for the call to free() in compute_new_size() 1280 compute_new_size(); 1281 return; 1282 } 1283 acquire_control_and_collect(full, clear_all_soft_refs); 1284 } 1285 1286 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1287 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1288 unsigned int gc_count = gch->total_full_collections(); 1289 if (gc_count == full_gc_count) { 1290 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1291 _full_gc_requested = true; 1292 _full_gc_cause = cause; 1293 CGC_lock->notify(); // nudge CMS thread 1294 } else { 1295 assert(gc_count > full_gc_count, "Error: causal loop"); 1296 } 1297 } 1298 1299 bool CMSCollector::is_external_interruption() { 1300 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1301 return GCCause::is_user_requested_gc(cause) || 1302 GCCause::is_serviceability_requested_gc(cause); 1303 } 1304 1305 void CMSCollector::report_concurrent_mode_interruption() { 1306 if (is_external_interruption()) { 1307 log_debug(gc)("Concurrent mode interrupted"); 1308 } else { 1309 log_debug(gc)("Concurrent mode failure"); 1310 _gc_tracer_cm->report_concurrent_mode_failure(); 1311 } 1312 } 1313 1314 1315 // The foreground and background collectors need to coordinate in order 1316 // to make sure that they do not mutually interfere with CMS collections. 1317 // When a background collection is active, 1318 // the foreground collector may need to take over (preempt) and 1319 // synchronously complete an ongoing collection. Depending on the 1320 // frequency of the background collections and the heap usage 1321 // of the application, this preemption can be seldom or frequent. 1322 // There are only certain 1323 // points in the background collection that the "collection-baton" 1324 // can be passed to the foreground collector. 1325 // 1326 // The foreground collector will wait for the baton before 1327 // starting any part of the collection. The foreground collector 1328 // will only wait at one location. 1329 // 1330 // The background collector will yield the baton before starting a new 1331 // phase of the collection (e.g., before initial marking, marking from roots, 1332 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1333 // of the loop which switches the phases. The background collector does some 1334 // of the phases (initial mark, final re-mark) with the world stopped. 1335 // Because of locking involved in stopping the world, 1336 // the foreground collector should not block waiting for the background 1337 // collector when it is doing a stop-the-world phase. The background 1338 // collector will yield the baton at an additional point just before 1339 // it enters a stop-the-world phase. Once the world is stopped, the 1340 // background collector checks the phase of the collection. If the 1341 // phase has not changed, it proceeds with the collection. If the 1342 // phase has changed, it skips that phase of the collection. See 1343 // the comments on the use of the Heap_lock in collect_in_background(). 1344 // 1345 // Variable used in baton passing. 1346 // _foregroundGCIsActive - Set to true by the foreground collector when 1347 // it wants the baton. The foreground clears it when it has finished 1348 // the collection. 1349 // _foregroundGCShouldWait - Set to true by the background collector 1350 // when it is running. The foreground collector waits while 1351 // _foregroundGCShouldWait is true. 1352 // CGC_lock - monitor used to protect access to the above variables 1353 // and to notify the foreground and background collectors. 1354 // _collectorState - current state of the CMS collection. 1355 // 1356 // The foreground collector 1357 // acquires the CGC_lock 1358 // sets _foregroundGCIsActive 1359 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1360 // various locks acquired in preparation for the collection 1361 // are released so as not to block the background collector 1362 // that is in the midst of a collection 1363 // proceeds with the collection 1364 // clears _foregroundGCIsActive 1365 // returns 1366 // 1367 // The background collector in a loop iterating on the phases of the 1368 // collection 1369 // acquires the CGC_lock 1370 // sets _foregroundGCShouldWait 1371 // if _foregroundGCIsActive is set 1372 // clears _foregroundGCShouldWait, notifies _CGC_lock 1373 // waits on _CGC_lock for _foregroundGCIsActive to become false 1374 // and exits the loop. 1375 // otherwise 1376 // proceed with that phase of the collection 1377 // if the phase is a stop-the-world phase, 1378 // yield the baton once more just before enqueueing 1379 // the stop-world CMS operation (executed by the VM thread). 1380 // returns after all phases of the collection are done 1381 // 1382 1383 void CMSCollector::acquire_control_and_collect(bool full, 1384 bool clear_all_soft_refs) { 1385 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1386 assert(!Thread::current()->is_ConcurrentGC_thread(), 1387 "shouldn't try to acquire control from self!"); 1388 1389 // Start the protocol for acquiring control of the 1390 // collection from the background collector (aka CMS thread). 1391 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1392 "VM thread should have CMS token"); 1393 // Remember the possibly interrupted state of an ongoing 1394 // concurrent collection 1395 CollectorState first_state = _collectorState; 1396 1397 // Signal to a possibly ongoing concurrent collection that 1398 // we want to do a foreground collection. 1399 _foregroundGCIsActive = true; 1400 1401 // release locks and wait for a notify from the background collector 1402 // releasing the locks in only necessary for phases which 1403 // do yields to improve the granularity of the collection. 1404 assert_lock_strong(bitMapLock()); 1405 // We need to lock the Free list lock for the space that we are 1406 // currently collecting. 1407 assert(haveFreelistLocks(), "Must be holding free list locks"); 1408 bitMapLock()->unlock(); 1409 releaseFreelistLocks(); 1410 { 1411 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1412 if (_foregroundGCShouldWait) { 1413 // We are going to be waiting for action for the CMS thread; 1414 // it had better not be gone (for instance at shutdown)! 1415 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1416 "CMS thread must be running"); 1417 // Wait here until the background collector gives us the go-ahead 1418 ConcurrentMarkSweepThread::clear_CMS_flag( 1419 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1420 // Get a possibly blocked CMS thread going: 1421 // Note that we set _foregroundGCIsActive true above, 1422 // without protection of the CGC_lock. 1423 CGC_lock->notify(); 1424 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1425 "Possible deadlock"); 1426 while (_foregroundGCShouldWait) { 1427 // wait for notification 1428 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1429 // Possibility of delay/starvation here, since CMS token does 1430 // not know to give priority to VM thread? Actually, i think 1431 // there wouldn't be any delay/starvation, but the proof of 1432 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1433 } 1434 ConcurrentMarkSweepThread::set_CMS_flag( 1435 ConcurrentMarkSweepThread::CMS_vm_has_token); 1436 } 1437 } 1438 // The CMS_token is already held. Get back the other locks. 1439 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1440 "VM thread should have CMS token"); 1441 getFreelistLocks(); 1442 bitMapLock()->lock_without_safepoint_check(); 1443 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1444 p2i(Thread::current()), first_state); 1445 log_debug(gc, state)(" gets control with state %d", _collectorState); 1446 1447 // Inform cms gen if this was due to partial collection failing. 1448 // The CMS gen may use this fact to determine its expansion policy. 1449 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1450 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1451 assert(!_cmsGen->incremental_collection_failed(), 1452 "Should have been noticed, reacted to and cleared"); 1453 _cmsGen->set_incremental_collection_failed(); 1454 } 1455 1456 if (first_state > Idling) { 1457 report_concurrent_mode_interruption(); 1458 } 1459 1460 set_did_compact(true); 1461 1462 // If the collection is being acquired from the background 1463 // collector, there may be references on the discovered 1464 // references lists. Abandon those references, since some 1465 // of them may have become unreachable after concurrent 1466 // discovery; the STW compacting collector will redo discovery 1467 // more precisely, without being subject to floating garbage. 1468 // Leaving otherwise unreachable references in the discovered 1469 // lists would require special handling. 1470 ref_processor()->disable_discovery(); 1471 ref_processor()->abandon_partial_discovery(); 1472 ref_processor()->verify_no_references_recorded(); 1473 1474 if (first_state > Idling) { 1475 save_heap_summary(); 1476 } 1477 1478 do_compaction_work(clear_all_soft_refs); 1479 1480 // Has the GC time limit been exceeded? 1481 size_t max_eden_size = _young_gen->max_eden_size(); 1482 GCCause::Cause gc_cause = gch->gc_cause(); 1483 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1484 _young_gen->eden()->used(), 1485 _cmsGen->max_capacity(), 1486 max_eden_size, 1487 full, 1488 gc_cause, 1489 gch->collector_policy()); 1490 1491 // Reset the expansion cause, now that we just completed 1492 // a collection cycle. 1493 clear_expansion_cause(); 1494 _foregroundGCIsActive = false; 1495 return; 1496 } 1497 1498 // Resize the tenured generation 1499 // after obtaining the free list locks for the 1500 // two generations. 1501 void CMSCollector::compute_new_size() { 1502 assert_locked_or_safepoint(Heap_lock); 1503 FreelistLocker z(this); 1504 MetaspaceGC::compute_new_size(); 1505 _cmsGen->compute_new_size_free_list(); 1506 } 1507 1508 // A work method used by the foreground collector to do 1509 // a mark-sweep-compact. 1510 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1511 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1512 1513 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1514 gc_timer->register_gc_start(); 1515 1516 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1517 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1518 1519 gch->pre_full_gc_dump(gc_timer); 1520 1521 GCTraceTime(Trace, gc) t("CMS:MSC"); 1522 1523 // Temporarily widen the span of the weak reference processing to 1524 // the entire heap. 1525 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1526 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1527 // Temporarily, clear the "is_alive_non_header" field of the 1528 // reference processor. 1529 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1530 // Temporarily make reference _processing_ single threaded (non-MT). 1531 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1532 // Temporarily make refs discovery atomic 1533 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1534 // Temporarily make reference _discovery_ single threaded (non-MT) 1535 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1536 1537 ref_processor()->set_enqueuing_is_done(false); 1538 ref_processor()->enable_discovery(); 1539 ref_processor()->setup_policy(clear_all_soft_refs); 1540 // If an asynchronous collection finishes, the _modUnionTable is 1541 // all clear. If we are assuming the collection from an asynchronous 1542 // collection, clear the _modUnionTable. 1543 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1544 "_modUnionTable should be clear if the baton was not passed"); 1545 _modUnionTable.clear_all(); 1546 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1547 "mod union for klasses should be clear if the baton was passed"); 1548 _ct->klass_rem_set()->clear_mod_union(); 1549 1550 // We must adjust the allocation statistics being maintained 1551 // in the free list space. We do so by reading and clearing 1552 // the sweep timer and updating the block flux rate estimates below. 1553 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1554 if (_inter_sweep_timer.is_active()) { 1555 _inter_sweep_timer.stop(); 1556 // Note that we do not use this sample to update the _inter_sweep_estimate. 1557 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1558 _inter_sweep_estimate.padded_average(), 1559 _intra_sweep_estimate.padded_average()); 1560 } 1561 1562 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1563 #ifdef ASSERT 1564 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1565 size_t free_size = cms_space->free(); 1566 assert(free_size == 1567 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1568 * HeapWordSize, 1569 "All the free space should be compacted into one chunk at top"); 1570 assert(cms_space->dictionary()->total_chunk_size( 1571 debug_only(cms_space->freelistLock())) == 0 || 1572 cms_space->totalSizeInIndexedFreeLists() == 0, 1573 "All the free space should be in a single chunk"); 1574 size_t num = cms_space->totalCount(); 1575 assert((free_size == 0 && num == 0) || 1576 (free_size > 0 && (num == 1 || num == 2)), 1577 "There should be at most 2 free chunks after compaction"); 1578 #endif // ASSERT 1579 _collectorState = Resetting; 1580 assert(_restart_addr == NULL, 1581 "Should have been NULL'd before baton was passed"); 1582 reset_stw(); 1583 _cmsGen->reset_after_compaction(); 1584 _concurrent_cycles_since_last_unload = 0; 1585 1586 // Clear any data recorded in the PLAB chunk arrays. 1587 if (_survivor_plab_array != NULL) { 1588 reset_survivor_plab_arrays(); 1589 } 1590 1591 // Adjust the per-size allocation stats for the next epoch. 1592 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1593 // Restart the "inter sweep timer" for the next epoch. 1594 _inter_sweep_timer.reset(); 1595 _inter_sweep_timer.start(); 1596 1597 gch->post_full_gc_dump(gc_timer); 1598 1599 gc_timer->register_gc_end(); 1600 1601 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1602 1603 // For a mark-sweep-compact, compute_new_size() will be called 1604 // in the heap's do_collection() method. 1605 } 1606 1607 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1608 LogHandle(gc, heap) log; 1609 if (!log.is_trace()) { 1610 return; 1611 } 1612 1613 ContiguousSpace* eden_space = _young_gen->eden(); 1614 ContiguousSpace* from_space = _young_gen->from(); 1615 ContiguousSpace* to_space = _young_gen->to(); 1616 // Eden 1617 if (_eden_chunk_array != NULL) { 1618 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1619 p2i(eden_space->bottom()), p2i(eden_space->top()), 1620 p2i(eden_space->end()), eden_space->capacity()); 1621 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1622 _eden_chunk_index, _eden_chunk_capacity); 1623 for (size_t i = 0; i < _eden_chunk_index; i++) { 1624 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1625 } 1626 } 1627 // Survivor 1628 if (_survivor_chunk_array != NULL) { 1629 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1630 p2i(from_space->bottom()), p2i(from_space->top()), 1631 p2i(from_space->end()), from_space->capacity()); 1632 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1633 _survivor_chunk_index, _survivor_chunk_capacity); 1634 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1635 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1636 } 1637 } 1638 } 1639 1640 void CMSCollector::getFreelistLocks() const { 1641 // Get locks for all free lists in all generations that this 1642 // collector is responsible for 1643 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1644 } 1645 1646 void CMSCollector::releaseFreelistLocks() const { 1647 // Release locks for all free lists in all generations that this 1648 // collector is responsible for 1649 _cmsGen->freelistLock()->unlock(); 1650 } 1651 1652 bool CMSCollector::haveFreelistLocks() const { 1653 // Check locks for all free lists in all generations that this 1654 // collector is responsible for 1655 assert_lock_strong(_cmsGen->freelistLock()); 1656 PRODUCT_ONLY(ShouldNotReachHere()); 1657 return true; 1658 } 1659 1660 // A utility class that is used by the CMS collector to 1661 // temporarily "release" the foreground collector from its 1662 // usual obligation to wait for the background collector to 1663 // complete an ongoing phase before proceeding. 1664 class ReleaseForegroundGC: public StackObj { 1665 private: 1666 CMSCollector* _c; 1667 public: 1668 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1669 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1670 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1671 // allow a potentially blocked foreground collector to proceed 1672 _c->_foregroundGCShouldWait = false; 1673 if (_c->_foregroundGCIsActive) { 1674 CGC_lock->notify(); 1675 } 1676 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1677 "Possible deadlock"); 1678 } 1679 1680 ~ReleaseForegroundGC() { 1681 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1682 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1683 _c->_foregroundGCShouldWait = true; 1684 } 1685 }; 1686 1687 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1688 assert(Thread::current()->is_ConcurrentGC_thread(), 1689 "A CMS asynchronous collection is only allowed on a CMS thread."); 1690 1691 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1692 { 1693 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1694 MutexLockerEx hl(Heap_lock, safepoint_check); 1695 FreelistLocker fll(this); 1696 MutexLockerEx x(CGC_lock, safepoint_check); 1697 if (_foregroundGCIsActive) { 1698 // The foreground collector is. Skip this 1699 // background collection. 1700 assert(!_foregroundGCShouldWait, "Should be clear"); 1701 return; 1702 } else { 1703 assert(_collectorState == Idling, "Should be idling before start."); 1704 _collectorState = InitialMarking; 1705 register_gc_start(cause); 1706 // Reset the expansion cause, now that we are about to begin 1707 // a new cycle. 1708 clear_expansion_cause(); 1709 1710 // Clear the MetaspaceGC flag since a concurrent collection 1711 // is starting but also clear it after the collection. 1712 MetaspaceGC::set_should_concurrent_collect(false); 1713 } 1714 // Decide if we want to enable class unloading as part of the 1715 // ensuing concurrent GC cycle. 1716 update_should_unload_classes(); 1717 _full_gc_requested = false; // acks all outstanding full gc requests 1718 _full_gc_cause = GCCause::_no_gc; 1719 // Signal that we are about to start a collection 1720 gch->increment_total_full_collections(); // ... starting a collection cycle 1721 _collection_count_start = gch->total_full_collections(); 1722 } 1723 1724 size_t prev_used = _cmsGen->used(); 1725 1726 // The change of the collection state is normally done at this level; 1727 // the exceptions are phases that are executed while the world is 1728 // stopped. For those phases the change of state is done while the 1729 // world is stopped. For baton passing purposes this allows the 1730 // background collector to finish the phase and change state atomically. 1731 // The foreground collector cannot wait on a phase that is done 1732 // while the world is stopped because the foreground collector already 1733 // has the world stopped and would deadlock. 1734 while (_collectorState != Idling) { 1735 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1736 p2i(Thread::current()), _collectorState); 1737 // The foreground collector 1738 // holds the Heap_lock throughout its collection. 1739 // holds the CMS token (but not the lock) 1740 // except while it is waiting for the background collector to yield. 1741 // 1742 // The foreground collector should be blocked (not for long) 1743 // if the background collector is about to start a phase 1744 // executed with world stopped. If the background 1745 // collector has already started such a phase, the 1746 // foreground collector is blocked waiting for the 1747 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1748 // are executed in the VM thread. 1749 // 1750 // The locking order is 1751 // PendingListLock (PLL) -- if applicable (FinalMarking) 1752 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1753 // CMS token (claimed in 1754 // stop_world_and_do() --> 1755 // safepoint_synchronize() --> 1756 // CMSThread::synchronize()) 1757 1758 { 1759 // Check if the FG collector wants us to yield. 1760 CMSTokenSync x(true); // is cms thread 1761 if (waitForForegroundGC()) { 1762 // We yielded to a foreground GC, nothing more to be 1763 // done this round. 1764 assert(_foregroundGCShouldWait == false, "We set it to false in " 1765 "waitForForegroundGC()"); 1766 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1767 p2i(Thread::current()), _collectorState); 1768 return; 1769 } else { 1770 // The background collector can run but check to see if the 1771 // foreground collector has done a collection while the 1772 // background collector was waiting to get the CGC_lock 1773 // above. If yes, break so that _foregroundGCShouldWait 1774 // is cleared before returning. 1775 if (_collectorState == Idling) { 1776 break; 1777 } 1778 } 1779 } 1780 1781 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1782 "should be waiting"); 1783 1784 switch (_collectorState) { 1785 case InitialMarking: 1786 { 1787 ReleaseForegroundGC x(this); 1788 stats().record_cms_begin(); 1789 VM_CMS_Initial_Mark initial_mark_op(this); 1790 VMThread::execute(&initial_mark_op); 1791 } 1792 // The collector state may be any legal state at this point 1793 // since the background collector may have yielded to the 1794 // foreground collector. 1795 break; 1796 case Marking: 1797 // initial marking in checkpointRootsInitialWork has been completed 1798 if (markFromRoots()) { // we were successful 1799 assert(_collectorState == Precleaning, "Collector state should " 1800 "have changed"); 1801 } else { 1802 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1803 } 1804 break; 1805 case Precleaning: 1806 // marking from roots in markFromRoots has been completed 1807 preclean(); 1808 assert(_collectorState == AbortablePreclean || 1809 _collectorState == FinalMarking, 1810 "Collector state should have changed"); 1811 break; 1812 case AbortablePreclean: 1813 abortable_preclean(); 1814 assert(_collectorState == FinalMarking, "Collector state should " 1815 "have changed"); 1816 break; 1817 case FinalMarking: 1818 { 1819 ReleaseForegroundGC x(this); 1820 1821 VM_CMS_Final_Remark final_remark_op(this); 1822 VMThread::execute(&final_remark_op); 1823 } 1824 assert(_foregroundGCShouldWait, "block post-condition"); 1825 break; 1826 case Sweeping: 1827 // final marking in checkpointRootsFinal has been completed 1828 sweep(); 1829 assert(_collectorState == Resizing, "Collector state change " 1830 "to Resizing must be done under the free_list_lock"); 1831 1832 case Resizing: { 1833 // Sweeping has been completed... 1834 // At this point the background collection has completed. 1835 // Don't move the call to compute_new_size() down 1836 // into code that might be executed if the background 1837 // collection was preempted. 1838 { 1839 ReleaseForegroundGC x(this); // unblock FG collection 1840 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1841 CMSTokenSync z(true); // not strictly needed. 1842 if (_collectorState == Resizing) { 1843 compute_new_size(); 1844 save_heap_summary(); 1845 _collectorState = Resetting; 1846 } else { 1847 assert(_collectorState == Idling, "The state should only change" 1848 " because the foreground collector has finished the collection"); 1849 } 1850 } 1851 break; 1852 } 1853 case Resetting: 1854 // CMS heap resizing has been completed 1855 reset_concurrent(); 1856 assert(_collectorState == Idling, "Collector state should " 1857 "have changed"); 1858 1859 MetaspaceGC::set_should_concurrent_collect(false); 1860 1861 stats().record_cms_end(); 1862 // Don't move the concurrent_phases_end() and compute_new_size() 1863 // calls to here because a preempted background collection 1864 // has it's state set to "Resetting". 1865 break; 1866 case Idling: 1867 default: 1868 ShouldNotReachHere(); 1869 break; 1870 } 1871 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1872 p2i(Thread::current()), _collectorState); 1873 assert(_foregroundGCShouldWait, "block post-condition"); 1874 } 1875 1876 // Should this be in gc_epilogue? 1877 collector_policy()->counters()->update_counters(); 1878 1879 { 1880 // Clear _foregroundGCShouldWait and, in the event that the 1881 // foreground collector is waiting, notify it, before 1882 // returning. 1883 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1884 _foregroundGCShouldWait = false; 1885 if (_foregroundGCIsActive) { 1886 CGC_lock->notify(); 1887 } 1888 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1889 "Possible deadlock"); 1890 } 1891 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1892 p2i(Thread::current()), _collectorState); 1893 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1894 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1895 } 1896 1897 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1898 _cms_start_registered = true; 1899 _gc_timer_cm->register_gc_start(); 1900 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1901 } 1902 1903 void CMSCollector::register_gc_end() { 1904 if (_cms_start_registered) { 1905 report_heap_summary(GCWhen::AfterGC); 1906 1907 _gc_timer_cm->register_gc_end(); 1908 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1909 _cms_start_registered = false; 1910 } 1911 } 1912 1913 void CMSCollector::save_heap_summary() { 1914 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1915 _last_heap_summary = gch->create_heap_summary(); 1916 _last_metaspace_summary = gch->create_metaspace_summary(); 1917 } 1918 1919 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1920 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1921 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1922 } 1923 1924 bool CMSCollector::waitForForegroundGC() { 1925 bool res = false; 1926 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1927 "CMS thread should have CMS token"); 1928 // Block the foreground collector until the 1929 // background collectors decides whether to 1930 // yield. 1931 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1932 _foregroundGCShouldWait = true; 1933 if (_foregroundGCIsActive) { 1934 // The background collector yields to the 1935 // foreground collector and returns a value 1936 // indicating that it has yielded. The foreground 1937 // collector can proceed. 1938 res = true; 1939 _foregroundGCShouldWait = false; 1940 ConcurrentMarkSweepThread::clear_CMS_flag( 1941 ConcurrentMarkSweepThread::CMS_cms_has_token); 1942 ConcurrentMarkSweepThread::set_CMS_flag( 1943 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1944 // Get a possibly blocked foreground thread going 1945 CGC_lock->notify(); 1946 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1947 p2i(Thread::current()), _collectorState); 1948 while (_foregroundGCIsActive) { 1949 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1950 } 1951 ConcurrentMarkSweepThread::set_CMS_flag( 1952 ConcurrentMarkSweepThread::CMS_cms_has_token); 1953 ConcurrentMarkSweepThread::clear_CMS_flag( 1954 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1955 } 1956 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1957 p2i(Thread::current()), _collectorState); 1958 return res; 1959 } 1960 1961 // Because of the need to lock the free lists and other structures in 1962 // the collector, common to all the generations that the collector is 1963 // collecting, we need the gc_prologues of individual CMS generations 1964 // delegate to their collector. It may have been simpler had the 1965 // current infrastructure allowed one to call a prologue on a 1966 // collector. In the absence of that we have the generation's 1967 // prologue delegate to the collector, which delegates back 1968 // some "local" work to a worker method in the individual generations 1969 // that it's responsible for collecting, while itself doing any 1970 // work common to all generations it's responsible for. A similar 1971 // comment applies to the gc_epilogue()'s. 1972 // The role of the variable _between_prologue_and_epilogue is to 1973 // enforce the invocation protocol. 1974 void CMSCollector::gc_prologue(bool full) { 1975 // Call gc_prologue_work() for the CMSGen 1976 // we are responsible for. 1977 1978 // The following locking discipline assumes that we are only called 1979 // when the world is stopped. 1980 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1981 1982 // The CMSCollector prologue must call the gc_prologues for the 1983 // "generations" that it's responsible 1984 // for. 1985 1986 assert( Thread::current()->is_VM_thread() 1987 || ( CMSScavengeBeforeRemark 1988 && Thread::current()->is_ConcurrentGC_thread()), 1989 "Incorrect thread type for prologue execution"); 1990 1991 if (_between_prologue_and_epilogue) { 1992 // We have already been invoked; this is a gc_prologue delegation 1993 // from yet another CMS generation that we are responsible for, just 1994 // ignore it since all relevant work has already been done. 1995 return; 1996 } 1997 1998 // set a bit saying prologue has been called; cleared in epilogue 1999 _between_prologue_and_epilogue = true; 2000 // Claim locks for common data structures, then call gc_prologue_work() 2001 // for each CMSGen. 2002 2003 getFreelistLocks(); // gets free list locks on constituent spaces 2004 bitMapLock()->lock_without_safepoint_check(); 2005 2006 // Should call gc_prologue_work() for all cms gens we are responsible for 2007 bool duringMarking = _collectorState >= Marking 2008 && _collectorState < Sweeping; 2009 2010 // The young collections clear the modified oops state, which tells if 2011 // there are any modified oops in the class. The remark phase also needs 2012 // that information. Tell the young collection to save the union of all 2013 // modified klasses. 2014 if (duringMarking) { 2015 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2016 } 2017 2018 bool registerClosure = duringMarking; 2019 2020 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2021 2022 if (!full) { 2023 stats().record_gc0_begin(); 2024 } 2025 } 2026 2027 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2028 2029 _capacity_at_prologue = capacity(); 2030 _used_at_prologue = used(); 2031 2032 // Delegate to CMScollector which knows how to coordinate between 2033 // this and any other CMS generations that it is responsible for 2034 // collecting. 2035 collector()->gc_prologue(full); 2036 } 2037 2038 // This is a "private" interface for use by this generation's CMSCollector. 2039 // Not to be called directly by any other entity (for instance, 2040 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2041 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2042 bool registerClosure, ModUnionClosure* modUnionClosure) { 2043 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2044 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2045 "Should be NULL"); 2046 if (registerClosure) { 2047 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2048 } 2049 cmsSpace()->gc_prologue(); 2050 // Clear stat counters 2051 NOT_PRODUCT( 2052 assert(_numObjectsPromoted == 0, "check"); 2053 assert(_numWordsPromoted == 0, "check"); 2054 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2055 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2056 _numObjectsAllocated = 0; 2057 _numWordsAllocated = 0; 2058 ) 2059 } 2060 2061 void CMSCollector::gc_epilogue(bool full) { 2062 // The following locking discipline assumes that we are only called 2063 // when the world is stopped. 2064 assert(SafepointSynchronize::is_at_safepoint(), 2065 "world is stopped assumption"); 2066 2067 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2068 // if linear allocation blocks need to be appropriately marked to allow the 2069 // the blocks to be parsable. We also check here whether we need to nudge the 2070 // CMS collector thread to start a new cycle (if it's not already active). 2071 assert( Thread::current()->is_VM_thread() 2072 || ( CMSScavengeBeforeRemark 2073 && Thread::current()->is_ConcurrentGC_thread()), 2074 "Incorrect thread type for epilogue execution"); 2075 2076 if (!_between_prologue_and_epilogue) { 2077 // We have already been invoked; this is a gc_epilogue delegation 2078 // from yet another CMS generation that we are responsible for, just 2079 // ignore it since all relevant work has already been done. 2080 return; 2081 } 2082 assert(haveFreelistLocks(), "must have freelist locks"); 2083 assert_lock_strong(bitMapLock()); 2084 2085 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2086 2087 _cmsGen->gc_epilogue_work(full); 2088 2089 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2090 // in case sampling was not already enabled, enable it 2091 _start_sampling = true; 2092 } 2093 // reset _eden_chunk_array so sampling starts afresh 2094 _eden_chunk_index = 0; 2095 2096 size_t cms_used = _cmsGen->cmsSpace()->used(); 2097 2098 // update performance counters - this uses a special version of 2099 // update_counters() that allows the utilization to be passed as a 2100 // parameter, avoiding multiple calls to used(). 2101 // 2102 _cmsGen->update_counters(cms_used); 2103 2104 bitMapLock()->unlock(); 2105 releaseFreelistLocks(); 2106 2107 if (!CleanChunkPoolAsync) { 2108 Chunk::clean_chunk_pool(); 2109 } 2110 2111 set_did_compact(false); 2112 _between_prologue_and_epilogue = false; // ready for next cycle 2113 } 2114 2115 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2116 collector()->gc_epilogue(full); 2117 2118 // Also reset promotion tracking in par gc thread states. 2119 for (uint i = 0; i < ParallelGCThreads; i++) { 2120 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2121 } 2122 } 2123 2124 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2125 assert(!incremental_collection_failed(), "Should have been cleared"); 2126 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2127 cmsSpace()->gc_epilogue(); 2128 // Print stat counters 2129 NOT_PRODUCT( 2130 assert(_numObjectsAllocated == 0, "check"); 2131 assert(_numWordsAllocated == 0, "check"); 2132 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2133 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2134 _numObjectsPromoted = 0; 2135 _numWordsPromoted = 0; 2136 ) 2137 2138 // Call down the chain in contiguous_available needs the freelistLock 2139 // so print this out before releasing the freeListLock. 2140 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2141 } 2142 2143 #ifndef PRODUCT 2144 bool CMSCollector::have_cms_token() { 2145 Thread* thr = Thread::current(); 2146 if (thr->is_VM_thread()) { 2147 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2148 } else if (thr->is_ConcurrentGC_thread()) { 2149 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2150 } else if (thr->is_GC_task_thread()) { 2151 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2152 ParGCRareEvent_lock->owned_by_self(); 2153 } 2154 return false; 2155 } 2156 2157 // Check reachability of the given heap address in CMS generation, 2158 // treating all other generations as roots. 2159 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2160 // We could "guarantee" below, rather than assert, but I'll 2161 // leave these as "asserts" so that an adventurous debugger 2162 // could try this in the product build provided some subset of 2163 // the conditions were met, provided they were interested in the 2164 // results and knew that the computation below wouldn't interfere 2165 // with other concurrent computations mutating the structures 2166 // being read or written. 2167 assert(SafepointSynchronize::is_at_safepoint(), 2168 "Else mutations in object graph will make answer suspect"); 2169 assert(have_cms_token(), "Should hold cms token"); 2170 assert(haveFreelistLocks(), "must hold free list locks"); 2171 assert_lock_strong(bitMapLock()); 2172 2173 // Clear the marking bit map array before starting, but, just 2174 // for kicks, first report if the given address is already marked 2175 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2176 _markBitMap.isMarked(addr) ? "" : " not"); 2177 2178 if (verify_after_remark()) { 2179 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2180 bool result = verification_mark_bm()->isMarked(addr); 2181 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2182 result ? "IS" : "is NOT"); 2183 return result; 2184 } else { 2185 tty->print_cr("Could not compute result"); 2186 return false; 2187 } 2188 } 2189 #endif 2190 2191 void 2192 CMSCollector::print_on_error(outputStream* st) { 2193 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2194 if (collector != NULL) { 2195 CMSBitMap* bitmap = &collector->_markBitMap; 2196 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2197 bitmap->print_on_error(st, " Bits: "); 2198 2199 st->cr(); 2200 2201 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2202 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2203 mut_bitmap->print_on_error(st, " Bits: "); 2204 } 2205 } 2206 2207 //////////////////////////////////////////////////////// 2208 // CMS Verification Support 2209 //////////////////////////////////////////////////////// 2210 // Following the remark phase, the following invariant 2211 // should hold -- each object in the CMS heap which is 2212 // marked in markBitMap() should be marked in the verification_mark_bm(). 2213 2214 class VerifyMarkedClosure: public BitMapClosure { 2215 CMSBitMap* _marks; 2216 bool _failed; 2217 2218 public: 2219 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2220 2221 bool do_bit(size_t offset) { 2222 HeapWord* addr = _marks->offsetToHeapWord(offset); 2223 if (!_marks->isMarked(addr)) { 2224 LogHandle(gc, verify) log; 2225 ResourceMark rm; 2226 oop(addr)->print_on(log.error_stream()); 2227 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2228 _failed = true; 2229 } 2230 return true; 2231 } 2232 2233 bool failed() { return _failed; } 2234 }; 2235 2236 bool CMSCollector::verify_after_remark() { 2237 GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking."); 2238 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2239 static bool init = false; 2240 2241 assert(SafepointSynchronize::is_at_safepoint(), 2242 "Else mutations in object graph will make answer suspect"); 2243 assert(have_cms_token(), 2244 "Else there may be mutual interference in use of " 2245 " verification data structures"); 2246 assert(_collectorState > Marking && _collectorState <= Sweeping, 2247 "Else marking info checked here may be obsolete"); 2248 assert(haveFreelistLocks(), "must hold free list locks"); 2249 assert_lock_strong(bitMapLock()); 2250 2251 2252 // Allocate marking bit map if not already allocated 2253 if (!init) { // first time 2254 if (!verification_mark_bm()->allocate(_span)) { 2255 return false; 2256 } 2257 init = true; 2258 } 2259 2260 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2261 2262 // Turn off refs discovery -- so we will be tracing through refs. 2263 // This is as intended, because by this time 2264 // GC must already have cleared any refs that need to be cleared, 2265 // and traced those that need to be marked; moreover, 2266 // the marking done here is not going to interfere in any 2267 // way with the marking information used by GC. 2268 NoRefDiscovery no_discovery(ref_processor()); 2269 2270 #if defined(COMPILER2) || INCLUDE_JVMCI 2271 DerivedPointerTableDeactivate dpt_deact; 2272 #endif 2273 2274 // Clear any marks from a previous round 2275 verification_mark_bm()->clear_all(); 2276 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2277 verify_work_stacks_empty(); 2278 2279 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2280 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2281 // Update the saved marks which may affect the root scans. 2282 gch->save_marks(); 2283 2284 if (CMSRemarkVerifyVariant == 1) { 2285 // In this first variant of verification, we complete 2286 // all marking, then check if the new marks-vector is 2287 // a subset of the CMS marks-vector. 2288 verify_after_remark_work_1(); 2289 } else { 2290 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2291 // In this second variant of verification, we flag an error 2292 // (i.e. an object reachable in the new marks-vector not reachable 2293 // in the CMS marks-vector) immediately, also indicating the 2294 // identify of an object (A) that references the unmarked object (B) -- 2295 // presumably, a mutation to A failed to be picked up by preclean/remark? 2296 verify_after_remark_work_2(); 2297 } 2298 2299 return true; 2300 } 2301 2302 void CMSCollector::verify_after_remark_work_1() { 2303 ResourceMark rm; 2304 HandleMark hm; 2305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2306 2307 // Get a clear set of claim bits for the roots processing to work with. 2308 ClassLoaderDataGraph::clear_claimed_marks(); 2309 2310 // Mark from roots one level into CMS 2311 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2312 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2313 2314 { 2315 StrongRootsScope srs(1); 2316 2317 gch->gen_process_roots(&srs, 2318 GenCollectedHeap::OldGen, 2319 true, // young gen as roots 2320 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2321 should_unload_classes(), 2322 ¬Older, 2323 NULL, 2324 NULL); 2325 } 2326 2327 // Now mark from the roots 2328 MarkFromRootsClosure markFromRootsClosure(this, _span, 2329 verification_mark_bm(), verification_mark_stack(), 2330 false /* don't yield */, true /* verifying */); 2331 assert(_restart_addr == NULL, "Expected pre-condition"); 2332 verification_mark_bm()->iterate(&markFromRootsClosure); 2333 while (_restart_addr != NULL) { 2334 // Deal with stack overflow: by restarting at the indicated 2335 // address. 2336 HeapWord* ra = _restart_addr; 2337 markFromRootsClosure.reset(ra); 2338 _restart_addr = NULL; 2339 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2340 } 2341 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2342 verify_work_stacks_empty(); 2343 2344 // Marking completed -- now verify that each bit marked in 2345 // verification_mark_bm() is also marked in markBitMap(); flag all 2346 // errors by printing corresponding objects. 2347 VerifyMarkedClosure vcl(markBitMap()); 2348 verification_mark_bm()->iterate(&vcl); 2349 if (vcl.failed()) { 2350 LogHandle(gc, verify) log; 2351 log.error("Failed marking verification after remark"); 2352 ResourceMark rm; 2353 gch->print_on(log.error_stream()); 2354 fatal("CMS: failed marking verification after remark"); 2355 } 2356 } 2357 2358 class VerifyKlassOopsKlassClosure : public KlassClosure { 2359 class VerifyKlassOopsClosure : public OopClosure { 2360 CMSBitMap* _bitmap; 2361 public: 2362 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2363 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2364 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2365 } _oop_closure; 2366 public: 2367 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2368 void do_klass(Klass* k) { 2369 k->oops_do(&_oop_closure); 2370 } 2371 }; 2372 2373 void CMSCollector::verify_after_remark_work_2() { 2374 ResourceMark rm; 2375 HandleMark hm; 2376 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2377 2378 // Get a clear set of claim bits for the roots processing to work with. 2379 ClassLoaderDataGraph::clear_claimed_marks(); 2380 2381 // Mark from roots one level into CMS 2382 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2383 markBitMap()); 2384 CLDToOopClosure cld_closure(¬Older, true); 2385 2386 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2387 2388 { 2389 StrongRootsScope srs(1); 2390 2391 gch->gen_process_roots(&srs, 2392 GenCollectedHeap::OldGen, 2393 true, // young gen as roots 2394 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2395 should_unload_classes(), 2396 ¬Older, 2397 NULL, 2398 &cld_closure); 2399 } 2400 2401 // Now mark from the roots 2402 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2403 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2404 assert(_restart_addr == NULL, "Expected pre-condition"); 2405 verification_mark_bm()->iterate(&markFromRootsClosure); 2406 while (_restart_addr != NULL) { 2407 // Deal with stack overflow: by restarting at the indicated 2408 // address. 2409 HeapWord* ra = _restart_addr; 2410 markFromRootsClosure.reset(ra); 2411 _restart_addr = NULL; 2412 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2413 } 2414 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2415 verify_work_stacks_empty(); 2416 2417 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2418 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2419 2420 // Marking completed -- now verify that each bit marked in 2421 // verification_mark_bm() is also marked in markBitMap(); flag all 2422 // errors by printing corresponding objects. 2423 VerifyMarkedClosure vcl(markBitMap()); 2424 verification_mark_bm()->iterate(&vcl); 2425 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2426 } 2427 2428 void ConcurrentMarkSweepGeneration::save_marks() { 2429 // delegate to CMS space 2430 cmsSpace()->save_marks(); 2431 for (uint i = 0; i < ParallelGCThreads; i++) { 2432 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2433 } 2434 } 2435 2436 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2437 return cmsSpace()->no_allocs_since_save_marks(); 2438 } 2439 2440 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2441 \ 2442 void ConcurrentMarkSweepGeneration:: \ 2443 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2444 cl->set_generation(this); \ 2445 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2446 cl->reset_generation(); \ 2447 save_marks(); \ 2448 } 2449 2450 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2451 2452 void 2453 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2454 if (freelistLock()->owned_by_self()) { 2455 Generation::oop_iterate(cl); 2456 } else { 2457 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2458 Generation::oop_iterate(cl); 2459 } 2460 } 2461 2462 void 2463 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2464 if (freelistLock()->owned_by_self()) { 2465 Generation::object_iterate(cl); 2466 } else { 2467 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2468 Generation::object_iterate(cl); 2469 } 2470 } 2471 2472 void 2473 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2474 if (freelistLock()->owned_by_self()) { 2475 Generation::safe_object_iterate(cl); 2476 } else { 2477 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2478 Generation::safe_object_iterate(cl); 2479 } 2480 } 2481 2482 void 2483 ConcurrentMarkSweepGeneration::post_compact() { 2484 } 2485 2486 void 2487 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2488 // Fix the linear allocation blocks to look like free blocks. 2489 2490 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2491 // are not called when the heap is verified during universe initialization and 2492 // at vm shutdown. 2493 if (freelistLock()->owned_by_self()) { 2494 cmsSpace()->prepare_for_verify(); 2495 } else { 2496 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2497 cmsSpace()->prepare_for_verify(); 2498 } 2499 } 2500 2501 void 2502 ConcurrentMarkSweepGeneration::verify() { 2503 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2504 // are not called when the heap is verified during universe initialization and 2505 // at vm shutdown. 2506 if (freelistLock()->owned_by_self()) { 2507 cmsSpace()->verify(); 2508 } else { 2509 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2510 cmsSpace()->verify(); 2511 } 2512 } 2513 2514 void CMSCollector::verify() { 2515 _cmsGen->verify(); 2516 } 2517 2518 #ifndef PRODUCT 2519 bool CMSCollector::overflow_list_is_empty() const { 2520 assert(_num_par_pushes >= 0, "Inconsistency"); 2521 if (_overflow_list == NULL) { 2522 assert(_num_par_pushes == 0, "Inconsistency"); 2523 } 2524 return _overflow_list == NULL; 2525 } 2526 2527 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2528 // merely consolidate assertion checks that appear to occur together frequently. 2529 void CMSCollector::verify_work_stacks_empty() const { 2530 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2531 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2532 } 2533 2534 void CMSCollector::verify_overflow_empty() const { 2535 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2536 assert(no_preserved_marks(), "No preserved marks"); 2537 } 2538 #endif // PRODUCT 2539 2540 // Decide if we want to enable class unloading as part of the 2541 // ensuing concurrent GC cycle. We will collect and 2542 // unload classes if it's the case that: 2543 // (1) an explicit gc request has been made and the flag 2544 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2545 // (2) (a) class unloading is enabled at the command line, and 2546 // (b) old gen is getting really full 2547 // NOTE: Provided there is no change in the state of the heap between 2548 // calls to this method, it should have idempotent results. Moreover, 2549 // its results should be monotonically increasing (i.e. going from 0 to 1, 2550 // but not 1 to 0) between successive calls between which the heap was 2551 // not collected. For the implementation below, it must thus rely on 2552 // the property that concurrent_cycles_since_last_unload() 2553 // will not decrease unless a collection cycle happened and that 2554 // _cmsGen->is_too_full() are 2555 // themselves also monotonic in that sense. See check_monotonicity() 2556 // below. 2557 void CMSCollector::update_should_unload_classes() { 2558 _should_unload_classes = false; 2559 // Condition 1 above 2560 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2561 _should_unload_classes = true; 2562 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2563 // Disjuncts 2.b.(i,ii,iii) above 2564 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2565 CMSClassUnloadingMaxInterval) 2566 || _cmsGen->is_too_full(); 2567 } 2568 } 2569 2570 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2571 bool res = should_concurrent_collect(); 2572 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2573 return res; 2574 } 2575 2576 void CMSCollector::setup_cms_unloading_and_verification_state() { 2577 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2578 || VerifyBeforeExit; 2579 const int rso = GenCollectedHeap::SO_AllCodeCache; 2580 2581 // We set the proper root for this CMS cycle here. 2582 if (should_unload_classes()) { // Should unload classes this cycle 2583 remove_root_scanning_option(rso); // Shrink the root set appropriately 2584 set_verifying(should_verify); // Set verification state for this cycle 2585 return; // Nothing else needs to be done at this time 2586 } 2587 2588 // Not unloading classes this cycle 2589 assert(!should_unload_classes(), "Inconsistency!"); 2590 2591 // If we are not unloading classes then add SO_AllCodeCache to root 2592 // scanning options. 2593 add_root_scanning_option(rso); 2594 2595 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2596 set_verifying(true); 2597 } else if (verifying() && !should_verify) { 2598 // We were verifying, but some verification flags got disabled. 2599 set_verifying(false); 2600 // Exclude symbols, strings and code cache elements from root scanning to 2601 // reduce IM and RM pauses. 2602 remove_root_scanning_option(rso); 2603 } 2604 } 2605 2606 2607 #ifndef PRODUCT 2608 HeapWord* CMSCollector::block_start(const void* p) const { 2609 const HeapWord* addr = (HeapWord*)p; 2610 if (_span.contains(p)) { 2611 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2612 return _cmsGen->cmsSpace()->block_start(p); 2613 } 2614 } 2615 return NULL; 2616 } 2617 #endif 2618 2619 HeapWord* 2620 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2621 bool tlab, 2622 bool parallel) { 2623 CMSSynchronousYieldRequest yr; 2624 assert(!tlab, "Can't deal with TLAB allocation"); 2625 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2626 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2627 if (GCExpandToAllocateDelayMillis > 0) { 2628 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2629 } 2630 return have_lock_and_allocate(word_size, tlab); 2631 } 2632 2633 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2634 size_t bytes, 2635 size_t expand_bytes, 2636 CMSExpansionCause::Cause cause) 2637 { 2638 2639 bool success = expand(bytes, expand_bytes); 2640 2641 // remember why we expanded; this information is used 2642 // by shouldConcurrentCollect() when making decisions on whether to start 2643 // a new CMS cycle. 2644 if (success) { 2645 set_expansion_cause(cause); 2646 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2647 } 2648 } 2649 2650 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2651 HeapWord* res = NULL; 2652 MutexLocker x(ParGCRareEvent_lock); 2653 while (true) { 2654 // Expansion by some other thread might make alloc OK now: 2655 res = ps->lab.alloc(word_sz); 2656 if (res != NULL) return res; 2657 // If there's not enough expansion space available, give up. 2658 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2659 return NULL; 2660 } 2661 // Otherwise, we try expansion. 2662 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2663 // Now go around the loop and try alloc again; 2664 // A competing par_promote might beat us to the expansion space, 2665 // so we may go around the loop again if promotion fails again. 2666 if (GCExpandToAllocateDelayMillis > 0) { 2667 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2668 } 2669 } 2670 } 2671 2672 2673 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2674 PromotionInfo* promo) { 2675 MutexLocker x(ParGCRareEvent_lock); 2676 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2677 while (true) { 2678 // Expansion by some other thread might make alloc OK now: 2679 if (promo->ensure_spooling_space()) { 2680 assert(promo->has_spooling_space(), 2681 "Post-condition of successful ensure_spooling_space()"); 2682 return true; 2683 } 2684 // If there's not enough expansion space available, give up. 2685 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2686 return false; 2687 } 2688 // Otherwise, we try expansion. 2689 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2690 // Now go around the loop and try alloc again; 2691 // A competing allocation might beat us to the expansion space, 2692 // so we may go around the loop again if allocation fails again. 2693 if (GCExpandToAllocateDelayMillis > 0) { 2694 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2695 } 2696 } 2697 } 2698 2699 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2700 // Only shrink if a compaction was done so that all the free space 2701 // in the generation is in a contiguous block at the end. 2702 if (did_compact()) { 2703 CardGeneration::shrink(bytes); 2704 } 2705 } 2706 2707 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2708 assert_locked_or_safepoint(Heap_lock); 2709 } 2710 2711 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2712 assert_locked_or_safepoint(Heap_lock); 2713 assert_lock_strong(freelistLock()); 2714 log_trace(gc)("Shrinking of CMS not yet implemented"); 2715 return; 2716 } 2717 2718 2719 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2720 // phases. 2721 class CMSPhaseAccounting: public StackObj { 2722 public: 2723 CMSPhaseAccounting(CMSCollector *collector, 2724 const char *title); 2725 ~CMSPhaseAccounting(); 2726 2727 private: 2728 CMSCollector *_collector; 2729 const char *_title; 2730 GCTraceConcTime(Info, gc) _trace_time; 2731 2732 public: 2733 // Not MT-safe; so do not pass around these StackObj's 2734 // where they may be accessed by other threads. 2735 double wallclock_millis() { 2736 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2737 } 2738 }; 2739 2740 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2741 const char *title) : 2742 _collector(collector), _title(title), _trace_time(title) { 2743 2744 _collector->resetYields(); 2745 _collector->resetTimer(); 2746 _collector->startTimer(); 2747 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2748 } 2749 2750 CMSPhaseAccounting::~CMSPhaseAccounting() { 2751 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2752 _collector->stopTimer(); 2753 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2754 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2755 } 2756 2757 // CMS work 2758 2759 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2760 class CMSParMarkTask : public AbstractGangTask { 2761 protected: 2762 CMSCollector* _collector; 2763 uint _n_workers; 2764 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2765 AbstractGangTask(name), 2766 _collector(collector), 2767 _n_workers(n_workers) {} 2768 // Work method in support of parallel rescan ... of young gen spaces 2769 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2770 ContiguousSpace* space, 2771 HeapWord** chunk_array, size_t chunk_top); 2772 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2773 }; 2774 2775 // Parallel initial mark task 2776 class CMSParInitialMarkTask: public CMSParMarkTask { 2777 StrongRootsScope* _strong_roots_scope; 2778 public: 2779 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2780 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2781 _strong_roots_scope(strong_roots_scope) {} 2782 void work(uint worker_id); 2783 }; 2784 2785 // Checkpoint the roots into this generation from outside 2786 // this generation. [Note this initial checkpoint need only 2787 // be approximate -- we'll do a catch up phase subsequently.] 2788 void CMSCollector::checkpointRootsInitial() { 2789 assert(_collectorState == InitialMarking, "Wrong collector state"); 2790 check_correct_thread_executing(); 2791 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2792 2793 save_heap_summary(); 2794 report_heap_summary(GCWhen::BeforeGC); 2795 2796 ReferenceProcessor* rp = ref_processor(); 2797 assert(_restart_addr == NULL, "Control point invariant"); 2798 { 2799 // acquire locks for subsequent manipulations 2800 MutexLockerEx x(bitMapLock(), 2801 Mutex::_no_safepoint_check_flag); 2802 checkpointRootsInitialWork(); 2803 // enable ("weak") refs discovery 2804 rp->enable_discovery(); 2805 _collectorState = Marking; 2806 } 2807 } 2808 2809 void CMSCollector::checkpointRootsInitialWork() { 2810 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2811 assert(_collectorState == InitialMarking, "just checking"); 2812 2813 // Already have locks. 2814 assert_lock_strong(bitMapLock()); 2815 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2816 2817 // Setup the verification and class unloading state for this 2818 // CMS collection cycle. 2819 setup_cms_unloading_and_verification_state(); 2820 2821 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2822 2823 // Reset all the PLAB chunk arrays if necessary. 2824 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2825 reset_survivor_plab_arrays(); 2826 } 2827 2828 ResourceMark rm; 2829 HandleMark hm; 2830 2831 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2832 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2833 2834 verify_work_stacks_empty(); 2835 verify_overflow_empty(); 2836 2837 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2838 // Update the saved marks which may affect the root scans. 2839 gch->save_marks(); 2840 2841 // weak reference processing has not started yet. 2842 ref_processor()->set_enqueuing_is_done(false); 2843 2844 // Need to remember all newly created CLDs, 2845 // so that we can guarantee that the remark finds them. 2846 ClassLoaderDataGraph::remember_new_clds(true); 2847 2848 // Whenever a CLD is found, it will be claimed before proceeding to mark 2849 // the klasses. The claimed marks need to be cleared before marking starts. 2850 ClassLoaderDataGraph::clear_claimed_marks(); 2851 2852 print_eden_and_survivor_chunk_arrays(); 2853 2854 { 2855 #if defined(COMPILER2) || INCLUDE_JVMCI 2856 DerivedPointerTableDeactivate dpt_deact; 2857 #endif 2858 if (CMSParallelInitialMarkEnabled) { 2859 // The parallel version. 2860 WorkGang* workers = gch->workers(); 2861 assert(workers != NULL, "Need parallel worker threads."); 2862 uint n_workers = workers->active_workers(); 2863 2864 StrongRootsScope srs(n_workers); 2865 2866 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2867 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2868 if (n_workers > 1) { 2869 workers->run_task(&tsk); 2870 } else { 2871 tsk.work(0); 2872 } 2873 } else { 2874 // The serial version. 2875 CLDToOopClosure cld_closure(¬Older, true); 2876 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2877 2878 StrongRootsScope srs(1); 2879 2880 gch->gen_process_roots(&srs, 2881 GenCollectedHeap::OldGen, 2882 true, // young gen as roots 2883 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2884 should_unload_classes(), 2885 ¬Older, 2886 NULL, 2887 &cld_closure); 2888 } 2889 } 2890 2891 // Clear mod-union table; it will be dirtied in the prologue of 2892 // CMS generation per each young generation collection. 2893 2894 assert(_modUnionTable.isAllClear(), 2895 "Was cleared in most recent final checkpoint phase" 2896 " or no bits are set in the gc_prologue before the start of the next " 2897 "subsequent marking phase."); 2898 2899 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2900 2901 // Save the end of the used_region of the constituent generations 2902 // to be used to limit the extent of sweep in each generation. 2903 save_sweep_limits(); 2904 verify_overflow_empty(); 2905 } 2906 2907 bool CMSCollector::markFromRoots() { 2908 // we might be tempted to assert that: 2909 // assert(!SafepointSynchronize::is_at_safepoint(), 2910 // "inconsistent argument?"); 2911 // However that wouldn't be right, because it's possible that 2912 // a safepoint is indeed in progress as a young generation 2913 // stop-the-world GC happens even as we mark in this generation. 2914 assert(_collectorState == Marking, "inconsistent state?"); 2915 check_correct_thread_executing(); 2916 verify_overflow_empty(); 2917 2918 // Weak ref discovery note: We may be discovering weak 2919 // refs in this generation concurrent (but interleaved) with 2920 // weak ref discovery by the young generation collector. 2921 2922 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2923 GCTraceCPUTime tcpu; 2924 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2925 bool res = markFromRootsWork(); 2926 if (res) { 2927 _collectorState = Precleaning; 2928 } else { // We failed and a foreground collection wants to take over 2929 assert(_foregroundGCIsActive, "internal state inconsistency"); 2930 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2931 log_debug(gc)("bailing out to foreground collection"); 2932 } 2933 verify_overflow_empty(); 2934 return res; 2935 } 2936 2937 bool CMSCollector::markFromRootsWork() { 2938 // iterate over marked bits in bit map, doing a full scan and mark 2939 // from these roots using the following algorithm: 2940 // . if oop is to the right of the current scan pointer, 2941 // mark corresponding bit (we'll process it later) 2942 // . else (oop is to left of current scan pointer) 2943 // push oop on marking stack 2944 // . drain the marking stack 2945 2946 // Note that when we do a marking step we need to hold the 2947 // bit map lock -- recall that direct allocation (by mutators) 2948 // and promotion (by the young generation collector) is also 2949 // marking the bit map. [the so-called allocate live policy.] 2950 // Because the implementation of bit map marking is not 2951 // robust wrt simultaneous marking of bits in the same word, 2952 // we need to make sure that there is no such interference 2953 // between concurrent such updates. 2954 2955 // already have locks 2956 assert_lock_strong(bitMapLock()); 2957 2958 verify_work_stacks_empty(); 2959 verify_overflow_empty(); 2960 bool result = false; 2961 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2962 result = do_marking_mt(); 2963 } else { 2964 result = do_marking_st(); 2965 } 2966 return result; 2967 } 2968 2969 // Forward decl 2970 class CMSConcMarkingTask; 2971 2972 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2973 CMSCollector* _collector; 2974 CMSConcMarkingTask* _task; 2975 public: 2976 virtual void yield(); 2977 2978 // "n_threads" is the number of threads to be terminated. 2979 // "queue_set" is a set of work queues of other threads. 2980 // "collector" is the CMS collector associated with this task terminator. 2981 // "yield" indicates whether we need the gang as a whole to yield. 2982 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2983 ParallelTaskTerminator(n_threads, queue_set), 2984 _collector(collector) { } 2985 2986 void set_task(CMSConcMarkingTask* task) { 2987 _task = task; 2988 } 2989 }; 2990 2991 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2992 CMSConcMarkingTask* _task; 2993 public: 2994 bool should_exit_termination(); 2995 void set_task(CMSConcMarkingTask* task) { 2996 _task = task; 2997 } 2998 }; 2999 3000 // MT Concurrent Marking Task 3001 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3002 CMSCollector* _collector; 3003 uint _n_workers; // requested/desired # workers 3004 bool _result; 3005 CompactibleFreeListSpace* _cms_space; 3006 char _pad_front[64]; // padding to ... 3007 HeapWord* _global_finger; // ... avoid sharing cache line 3008 char _pad_back[64]; 3009 HeapWord* _restart_addr; 3010 3011 // Exposed here for yielding support 3012 Mutex* const _bit_map_lock; 3013 3014 // The per thread work queues, available here for stealing 3015 OopTaskQueueSet* _task_queues; 3016 3017 // Termination (and yielding) support 3018 CMSConcMarkingTerminator _term; 3019 CMSConcMarkingTerminatorTerminator _term_term; 3020 3021 public: 3022 CMSConcMarkingTask(CMSCollector* collector, 3023 CompactibleFreeListSpace* cms_space, 3024 YieldingFlexibleWorkGang* workers, 3025 OopTaskQueueSet* task_queues): 3026 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3027 _collector(collector), 3028 _cms_space(cms_space), 3029 _n_workers(0), _result(true), 3030 _task_queues(task_queues), 3031 _term(_n_workers, task_queues, _collector), 3032 _bit_map_lock(collector->bitMapLock()) 3033 { 3034 _requested_size = _n_workers; 3035 _term.set_task(this); 3036 _term_term.set_task(this); 3037 _restart_addr = _global_finger = _cms_space->bottom(); 3038 } 3039 3040 3041 OopTaskQueueSet* task_queues() { return _task_queues; } 3042 3043 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3044 3045 HeapWord** global_finger_addr() { return &_global_finger; } 3046 3047 CMSConcMarkingTerminator* terminator() { return &_term; } 3048 3049 virtual void set_for_termination(uint active_workers) { 3050 terminator()->reset_for_reuse(active_workers); 3051 } 3052 3053 void work(uint worker_id); 3054 bool should_yield() { 3055 return ConcurrentMarkSweepThread::should_yield() 3056 && !_collector->foregroundGCIsActive(); 3057 } 3058 3059 virtual void coordinator_yield(); // stuff done by coordinator 3060 bool result() { return _result; } 3061 3062 void reset(HeapWord* ra) { 3063 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3064 _restart_addr = _global_finger = ra; 3065 _term.reset_for_reuse(); 3066 } 3067 3068 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3069 OopTaskQueue* work_q); 3070 3071 private: 3072 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3073 void do_work_steal(int i); 3074 void bump_global_finger(HeapWord* f); 3075 }; 3076 3077 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3078 assert(_task != NULL, "Error"); 3079 return _task->yielding(); 3080 // Note that we do not need the disjunct || _task->should_yield() above 3081 // because we want terminating threads to yield only if the task 3082 // is already in the midst of yielding, which happens only after at least one 3083 // thread has yielded. 3084 } 3085 3086 void CMSConcMarkingTerminator::yield() { 3087 if (_task->should_yield()) { 3088 _task->yield(); 3089 } else { 3090 ParallelTaskTerminator::yield(); 3091 } 3092 } 3093 3094 //////////////////////////////////////////////////////////////// 3095 // Concurrent Marking Algorithm Sketch 3096 //////////////////////////////////////////////////////////////// 3097 // Until all tasks exhausted (both spaces): 3098 // -- claim next available chunk 3099 // -- bump global finger via CAS 3100 // -- find first object that starts in this chunk 3101 // and start scanning bitmap from that position 3102 // -- scan marked objects for oops 3103 // -- CAS-mark target, and if successful: 3104 // . if target oop is above global finger (volatile read) 3105 // nothing to do 3106 // . if target oop is in chunk and above local finger 3107 // then nothing to do 3108 // . else push on work-queue 3109 // -- Deal with possible overflow issues: 3110 // . local work-queue overflow causes stuff to be pushed on 3111 // global (common) overflow queue 3112 // . always first empty local work queue 3113 // . then get a batch of oops from global work queue if any 3114 // . then do work stealing 3115 // -- When all tasks claimed (both spaces) 3116 // and local work queue empty, 3117 // then in a loop do: 3118 // . check global overflow stack; steal a batch of oops and trace 3119 // . try to steal from other threads oif GOS is empty 3120 // . if neither is available, offer termination 3121 // -- Terminate and return result 3122 // 3123 void CMSConcMarkingTask::work(uint worker_id) { 3124 elapsedTimer _timer; 3125 ResourceMark rm; 3126 HandleMark hm; 3127 3128 DEBUG_ONLY(_collector->verify_overflow_empty();) 3129 3130 // Before we begin work, our work queue should be empty 3131 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3132 // Scan the bitmap covering _cms_space, tracing through grey objects. 3133 _timer.start(); 3134 do_scan_and_mark(worker_id, _cms_space); 3135 _timer.stop(); 3136 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3137 3138 // ... do work stealing 3139 _timer.reset(); 3140 _timer.start(); 3141 do_work_steal(worker_id); 3142 _timer.stop(); 3143 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3144 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3145 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3146 // Note that under the current task protocol, the 3147 // following assertion is true even of the spaces 3148 // expanded since the completion of the concurrent 3149 // marking. XXX This will likely change under a strict 3150 // ABORT semantics. 3151 // After perm removal the comparison was changed to 3152 // greater than or equal to from strictly greater than. 3153 // Before perm removal the highest address sweep would 3154 // have been at the end of perm gen but now is at the 3155 // end of the tenured gen. 3156 assert(_global_finger >= _cms_space->end(), 3157 "All tasks have been completed"); 3158 DEBUG_ONLY(_collector->verify_overflow_empty();) 3159 } 3160 3161 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3162 HeapWord* read = _global_finger; 3163 HeapWord* cur = read; 3164 while (f > read) { 3165 cur = read; 3166 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3167 if (cur == read) { 3168 // our cas succeeded 3169 assert(_global_finger >= f, "protocol consistency"); 3170 break; 3171 } 3172 } 3173 } 3174 3175 // This is really inefficient, and should be redone by 3176 // using (not yet available) block-read and -write interfaces to the 3177 // stack and the work_queue. XXX FIX ME !!! 3178 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3179 OopTaskQueue* work_q) { 3180 // Fast lock-free check 3181 if (ovflw_stk->length() == 0) { 3182 return false; 3183 } 3184 assert(work_q->size() == 0, "Shouldn't steal"); 3185 MutexLockerEx ml(ovflw_stk->par_lock(), 3186 Mutex::_no_safepoint_check_flag); 3187 // Grab up to 1/4 the size of the work queue 3188 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3189 (size_t)ParGCDesiredObjsFromOverflowList); 3190 num = MIN2(num, ovflw_stk->length()); 3191 for (int i = (int) num; i > 0; i--) { 3192 oop cur = ovflw_stk->pop(); 3193 assert(cur != NULL, "Counted wrong?"); 3194 work_q->push(cur); 3195 } 3196 return num > 0; 3197 } 3198 3199 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3200 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3201 int n_tasks = pst->n_tasks(); 3202 // We allow that there may be no tasks to do here because 3203 // we are restarting after a stack overflow. 3204 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3205 uint nth_task = 0; 3206 3207 HeapWord* aligned_start = sp->bottom(); 3208 if (sp->used_region().contains(_restart_addr)) { 3209 // Align down to a card boundary for the start of 0th task 3210 // for this space. 3211 aligned_start = 3212 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3213 CardTableModRefBS::card_size); 3214 } 3215 3216 size_t chunk_size = sp->marking_task_size(); 3217 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3218 // Having claimed the nth task in this space, 3219 // compute the chunk that it corresponds to: 3220 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3221 aligned_start + (nth_task+1)*chunk_size); 3222 // Try and bump the global finger via a CAS; 3223 // note that we need to do the global finger bump 3224 // _before_ taking the intersection below, because 3225 // the task corresponding to that region will be 3226 // deemed done even if the used_region() expands 3227 // because of allocation -- as it almost certainly will 3228 // during start-up while the threads yield in the 3229 // closure below. 3230 HeapWord* finger = span.end(); 3231 bump_global_finger(finger); // atomically 3232 // There are null tasks here corresponding to chunks 3233 // beyond the "top" address of the space. 3234 span = span.intersection(sp->used_region()); 3235 if (!span.is_empty()) { // Non-null task 3236 HeapWord* prev_obj; 3237 assert(!span.contains(_restart_addr) || nth_task == 0, 3238 "Inconsistency"); 3239 if (nth_task == 0) { 3240 // For the 0th task, we'll not need to compute a block_start. 3241 if (span.contains(_restart_addr)) { 3242 // In the case of a restart because of stack overflow, 3243 // we might additionally skip a chunk prefix. 3244 prev_obj = _restart_addr; 3245 } else { 3246 prev_obj = span.start(); 3247 } 3248 } else { 3249 // We want to skip the first object because 3250 // the protocol is to scan any object in its entirety 3251 // that _starts_ in this span; a fortiori, any 3252 // object starting in an earlier span is scanned 3253 // as part of an earlier claimed task. 3254 // Below we use the "careful" version of block_start 3255 // so we do not try to navigate uninitialized objects. 3256 prev_obj = sp->block_start_careful(span.start()); 3257 // Below we use a variant of block_size that uses the 3258 // Printezis bits to avoid waiting for allocated 3259 // objects to become initialized/parsable. 3260 while (prev_obj < span.start()) { 3261 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3262 if (sz > 0) { 3263 prev_obj += sz; 3264 } else { 3265 // In this case we may end up doing a bit of redundant 3266 // scanning, but that appears unavoidable, short of 3267 // locking the free list locks; see bug 6324141. 3268 break; 3269 } 3270 } 3271 } 3272 if (prev_obj < span.end()) { 3273 MemRegion my_span = MemRegion(prev_obj, span.end()); 3274 // Do the marking work within a non-empty span -- 3275 // the last argument to the constructor indicates whether the 3276 // iteration should be incremental with periodic yields. 3277 ParMarkFromRootsClosure cl(this, _collector, my_span, 3278 &_collector->_markBitMap, 3279 work_queue(i), 3280 &_collector->_markStack); 3281 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3282 } // else nothing to do for this task 3283 } // else nothing to do for this task 3284 } 3285 // We'd be tempted to assert here that since there are no 3286 // more tasks left to claim in this space, the global_finger 3287 // must exceed space->top() and a fortiori space->end(). However, 3288 // that would not quite be correct because the bumping of 3289 // global_finger occurs strictly after the claiming of a task, 3290 // so by the time we reach here the global finger may not yet 3291 // have been bumped up by the thread that claimed the last 3292 // task. 3293 pst->all_tasks_completed(); 3294 } 3295 3296 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3297 private: 3298 CMSCollector* _collector; 3299 CMSConcMarkingTask* _task; 3300 MemRegion _span; 3301 CMSBitMap* _bit_map; 3302 CMSMarkStack* _overflow_stack; 3303 OopTaskQueue* _work_queue; 3304 protected: 3305 DO_OOP_WORK_DEFN 3306 public: 3307 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3308 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3309 MetadataAwareOopClosure(collector->ref_processor()), 3310 _collector(collector), 3311 _task(task), 3312 _span(collector->_span), 3313 _work_queue(work_queue), 3314 _bit_map(bit_map), 3315 _overflow_stack(overflow_stack) 3316 { } 3317 virtual void do_oop(oop* p); 3318 virtual void do_oop(narrowOop* p); 3319 3320 void trim_queue(size_t max); 3321 void handle_stack_overflow(HeapWord* lost); 3322 void do_yield_check() { 3323 if (_task->should_yield()) { 3324 _task->yield(); 3325 } 3326 } 3327 }; 3328 3329 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3330 3331 // Grey object scanning during work stealing phase -- 3332 // the salient assumption here is that any references 3333 // that are in these stolen objects being scanned must 3334 // already have been initialized (else they would not have 3335 // been published), so we do not need to check for 3336 // uninitialized objects before pushing here. 3337 void ParConcMarkingClosure::do_oop(oop obj) { 3338 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3339 HeapWord* addr = (HeapWord*)obj; 3340 // Check if oop points into the CMS generation 3341 // and is not marked 3342 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3343 // a white object ... 3344 // If we manage to "claim" the object, by being the 3345 // first thread to mark it, then we push it on our 3346 // marking stack 3347 if (_bit_map->par_mark(addr)) { // ... now grey 3348 // push on work queue (grey set) 3349 bool simulate_overflow = false; 3350 NOT_PRODUCT( 3351 if (CMSMarkStackOverflowALot && 3352 _collector->simulate_overflow()) { 3353 // simulate a stack overflow 3354 simulate_overflow = true; 3355 } 3356 ) 3357 if (simulate_overflow || 3358 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3359 // stack overflow 3360 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3361 // We cannot assert that the overflow stack is full because 3362 // it may have been emptied since. 3363 assert(simulate_overflow || 3364 _work_queue->size() == _work_queue->max_elems(), 3365 "Else push should have succeeded"); 3366 handle_stack_overflow(addr); 3367 } 3368 } // Else, some other thread got there first 3369 do_yield_check(); 3370 } 3371 } 3372 3373 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3374 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3375 3376 void ParConcMarkingClosure::trim_queue(size_t max) { 3377 while (_work_queue->size() > max) { 3378 oop new_oop; 3379 if (_work_queue->pop_local(new_oop)) { 3380 assert(new_oop->is_oop(), "Should be an oop"); 3381 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3382 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3383 new_oop->oop_iterate(this); // do_oop() above 3384 do_yield_check(); 3385 } 3386 } 3387 } 3388 3389 // Upon stack overflow, we discard (part of) the stack, 3390 // remembering the least address amongst those discarded 3391 // in CMSCollector's _restart_address. 3392 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3393 // We need to do this under a mutex to prevent other 3394 // workers from interfering with the work done below. 3395 MutexLockerEx ml(_overflow_stack->par_lock(), 3396 Mutex::_no_safepoint_check_flag); 3397 // Remember the least grey address discarded 3398 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3399 _collector->lower_restart_addr(ra); 3400 _overflow_stack->reset(); // discard stack contents 3401 _overflow_stack->expand(); // expand the stack if possible 3402 } 3403 3404 3405 void CMSConcMarkingTask::do_work_steal(int i) { 3406 OopTaskQueue* work_q = work_queue(i); 3407 oop obj_to_scan; 3408 CMSBitMap* bm = &(_collector->_markBitMap); 3409 CMSMarkStack* ovflw = &(_collector->_markStack); 3410 int* seed = _collector->hash_seed(i); 3411 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3412 while (true) { 3413 cl.trim_queue(0); 3414 assert(work_q->size() == 0, "Should have been emptied above"); 3415 if (get_work_from_overflow_stack(ovflw, work_q)) { 3416 // Can't assert below because the work obtained from the 3417 // overflow stack may already have been stolen from us. 3418 // assert(work_q->size() > 0, "Work from overflow stack"); 3419 continue; 3420 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3421 assert(obj_to_scan->is_oop(), "Should be an oop"); 3422 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3423 obj_to_scan->oop_iterate(&cl); 3424 } else if (terminator()->offer_termination(&_term_term)) { 3425 assert(work_q->size() == 0, "Impossible!"); 3426 break; 3427 } else if (yielding() || should_yield()) { 3428 yield(); 3429 } 3430 } 3431 } 3432 3433 // This is run by the CMS (coordinator) thread. 3434 void CMSConcMarkingTask::coordinator_yield() { 3435 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3436 "CMS thread should hold CMS token"); 3437 // First give up the locks, then yield, then re-lock 3438 // We should probably use a constructor/destructor idiom to 3439 // do this unlock/lock or modify the MutexUnlocker class to 3440 // serve our purpose. XXX 3441 assert_lock_strong(_bit_map_lock); 3442 _bit_map_lock->unlock(); 3443 ConcurrentMarkSweepThread::desynchronize(true); 3444 _collector->stopTimer(); 3445 _collector->incrementYields(); 3446 3447 // It is possible for whichever thread initiated the yield request 3448 // not to get a chance to wake up and take the bitmap lock between 3449 // this thread releasing it and reacquiring it. So, while the 3450 // should_yield() flag is on, let's sleep for a bit to give the 3451 // other thread a chance to wake up. The limit imposed on the number 3452 // of iterations is defensive, to avoid any unforseen circumstances 3453 // putting us into an infinite loop. Since it's always been this 3454 // (coordinator_yield()) method that was observed to cause the 3455 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3456 // which is by default non-zero. For the other seven methods that 3457 // also perform the yield operation, as are using a different 3458 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3459 // can enable the sleeping for those methods too, if necessary. 3460 // See 6442774. 3461 // 3462 // We really need to reconsider the synchronization between the GC 3463 // thread and the yield-requesting threads in the future and we 3464 // should really use wait/notify, which is the recommended 3465 // way of doing this type of interaction. Additionally, we should 3466 // consolidate the eight methods that do the yield operation and they 3467 // are almost identical into one for better maintainability and 3468 // readability. See 6445193. 3469 // 3470 // Tony 2006.06.29 3471 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3472 ConcurrentMarkSweepThread::should_yield() && 3473 !CMSCollector::foregroundGCIsActive(); ++i) { 3474 os::sleep(Thread::current(), 1, false); 3475 } 3476 3477 ConcurrentMarkSweepThread::synchronize(true); 3478 _bit_map_lock->lock_without_safepoint_check(); 3479 _collector->startTimer(); 3480 } 3481 3482 bool CMSCollector::do_marking_mt() { 3483 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3484 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3485 conc_workers()->active_workers(), 3486 Threads::number_of_non_daemon_threads()); 3487 conc_workers()->set_active_workers(num_workers); 3488 3489 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3490 3491 CMSConcMarkingTask tsk(this, 3492 cms_space, 3493 conc_workers(), 3494 task_queues()); 3495 3496 // Since the actual number of workers we get may be different 3497 // from the number we requested above, do we need to do anything different 3498 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3499 // class?? XXX 3500 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3501 3502 // Refs discovery is already non-atomic. 3503 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3504 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3505 conc_workers()->start_task(&tsk); 3506 while (tsk.yielded()) { 3507 tsk.coordinator_yield(); 3508 conc_workers()->continue_task(&tsk); 3509 } 3510 // If the task was aborted, _restart_addr will be non-NULL 3511 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3512 while (_restart_addr != NULL) { 3513 // XXX For now we do not make use of ABORTED state and have not 3514 // yet implemented the right abort semantics (even in the original 3515 // single-threaded CMS case). That needs some more investigation 3516 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3517 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3518 // If _restart_addr is non-NULL, a marking stack overflow 3519 // occurred; we need to do a fresh marking iteration from the 3520 // indicated restart address. 3521 if (_foregroundGCIsActive) { 3522 // We may be running into repeated stack overflows, having 3523 // reached the limit of the stack size, while making very 3524 // slow forward progress. It may be best to bail out and 3525 // let the foreground collector do its job. 3526 // Clear _restart_addr, so that foreground GC 3527 // works from scratch. This avoids the headache of 3528 // a "rescan" which would otherwise be needed because 3529 // of the dirty mod union table & card table. 3530 _restart_addr = NULL; 3531 return false; 3532 } 3533 // Adjust the task to restart from _restart_addr 3534 tsk.reset(_restart_addr); 3535 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3536 _restart_addr); 3537 _restart_addr = NULL; 3538 // Get the workers going again 3539 conc_workers()->start_task(&tsk); 3540 while (tsk.yielded()) { 3541 tsk.coordinator_yield(); 3542 conc_workers()->continue_task(&tsk); 3543 } 3544 } 3545 assert(tsk.completed(), "Inconsistency"); 3546 assert(tsk.result() == true, "Inconsistency"); 3547 return true; 3548 } 3549 3550 bool CMSCollector::do_marking_st() { 3551 ResourceMark rm; 3552 HandleMark hm; 3553 3554 // Temporarily make refs discovery single threaded (non-MT) 3555 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3556 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3557 &_markStack, CMSYield); 3558 // the last argument to iterate indicates whether the iteration 3559 // should be incremental with periodic yields. 3560 _markBitMap.iterate(&markFromRootsClosure); 3561 // If _restart_addr is non-NULL, a marking stack overflow 3562 // occurred; we need to do a fresh iteration from the 3563 // indicated restart address. 3564 while (_restart_addr != NULL) { 3565 if (_foregroundGCIsActive) { 3566 // We may be running into repeated stack overflows, having 3567 // reached the limit of the stack size, while making very 3568 // slow forward progress. It may be best to bail out and 3569 // let the foreground collector do its job. 3570 // Clear _restart_addr, so that foreground GC 3571 // works from scratch. This avoids the headache of 3572 // a "rescan" which would otherwise be needed because 3573 // of the dirty mod union table & card table. 3574 _restart_addr = NULL; 3575 return false; // indicating failure to complete marking 3576 } 3577 // Deal with stack overflow: 3578 // we restart marking from _restart_addr 3579 HeapWord* ra = _restart_addr; 3580 markFromRootsClosure.reset(ra); 3581 _restart_addr = NULL; 3582 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3583 } 3584 return true; 3585 } 3586 3587 void CMSCollector::preclean() { 3588 check_correct_thread_executing(); 3589 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3590 verify_work_stacks_empty(); 3591 verify_overflow_empty(); 3592 _abort_preclean = false; 3593 if (CMSPrecleaningEnabled) { 3594 if (!CMSEdenChunksRecordAlways) { 3595 _eden_chunk_index = 0; 3596 } 3597 size_t used = get_eden_used(); 3598 size_t capacity = get_eden_capacity(); 3599 // Don't start sampling unless we will get sufficiently 3600 // many samples. 3601 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3602 * CMSScheduleRemarkEdenPenetration)) { 3603 _start_sampling = true; 3604 } else { 3605 _start_sampling = false; 3606 } 3607 GCTraceCPUTime tcpu; 3608 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3609 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3610 } 3611 CMSTokenSync x(true); // is cms thread 3612 if (CMSPrecleaningEnabled) { 3613 sample_eden(); 3614 _collectorState = AbortablePreclean; 3615 } else { 3616 _collectorState = FinalMarking; 3617 } 3618 verify_work_stacks_empty(); 3619 verify_overflow_empty(); 3620 } 3621 3622 // Try and schedule the remark such that young gen 3623 // occupancy is CMSScheduleRemarkEdenPenetration %. 3624 void CMSCollector::abortable_preclean() { 3625 check_correct_thread_executing(); 3626 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3627 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3628 3629 // If Eden's current occupancy is below this threshold, 3630 // immediately schedule the remark; else preclean 3631 // past the next scavenge in an effort to 3632 // schedule the pause as described above. By choosing 3633 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3634 // we will never do an actual abortable preclean cycle. 3635 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3636 GCTraceCPUTime tcpu; 3637 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3638 // We need more smarts in the abortable preclean 3639 // loop below to deal with cases where allocation 3640 // in young gen is very very slow, and our precleaning 3641 // is running a losing race against a horde of 3642 // mutators intent on flooding us with CMS updates 3643 // (dirty cards). 3644 // One, admittedly dumb, strategy is to give up 3645 // after a certain number of abortable precleaning loops 3646 // or after a certain maximum time. We want to make 3647 // this smarter in the next iteration. 3648 // XXX FIX ME!!! YSR 3649 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3650 while (!(should_abort_preclean() || 3651 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3652 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3653 cumworkdone += workdone; 3654 loops++; 3655 // Voluntarily terminate abortable preclean phase if we have 3656 // been at it for too long. 3657 if ((CMSMaxAbortablePrecleanLoops != 0) && 3658 loops >= CMSMaxAbortablePrecleanLoops) { 3659 log_debug(gc)(" CMS: abort preclean due to loops "); 3660 break; 3661 } 3662 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3663 log_debug(gc)(" CMS: abort preclean due to time "); 3664 break; 3665 } 3666 // If we are doing little work each iteration, we should 3667 // take a short break. 3668 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3669 // Sleep for some time, waiting for work to accumulate 3670 stopTimer(); 3671 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3672 startTimer(); 3673 waited++; 3674 } 3675 } 3676 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3677 loops, waited, cumworkdone); 3678 } 3679 CMSTokenSync x(true); // is cms thread 3680 if (_collectorState != Idling) { 3681 assert(_collectorState == AbortablePreclean, 3682 "Spontaneous state transition?"); 3683 _collectorState = FinalMarking; 3684 } // Else, a foreground collection completed this CMS cycle. 3685 return; 3686 } 3687 3688 // Respond to an Eden sampling opportunity 3689 void CMSCollector::sample_eden() { 3690 // Make sure a young gc cannot sneak in between our 3691 // reading and recording of a sample. 3692 assert(Thread::current()->is_ConcurrentGC_thread(), 3693 "Only the cms thread may collect Eden samples"); 3694 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3695 "Should collect samples while holding CMS token"); 3696 if (!_start_sampling) { 3697 return; 3698 } 3699 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3700 // is populated by the young generation. 3701 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3702 if (_eden_chunk_index < _eden_chunk_capacity) { 3703 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3704 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3705 "Unexpected state of Eden"); 3706 // We'd like to check that what we just sampled is an oop-start address; 3707 // however, we cannot do that here since the object may not yet have been 3708 // initialized. So we'll instead do the check when we _use_ this sample 3709 // later. 3710 if (_eden_chunk_index == 0 || 3711 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3712 _eden_chunk_array[_eden_chunk_index-1]) 3713 >= CMSSamplingGrain)) { 3714 _eden_chunk_index++; // commit sample 3715 } 3716 } 3717 } 3718 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3719 size_t used = get_eden_used(); 3720 size_t capacity = get_eden_capacity(); 3721 assert(used <= capacity, "Unexpected state of Eden"); 3722 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3723 _abort_preclean = true; 3724 } 3725 } 3726 } 3727 3728 3729 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3730 assert(_collectorState == Precleaning || 3731 _collectorState == AbortablePreclean, "incorrect state"); 3732 ResourceMark rm; 3733 HandleMark hm; 3734 3735 // Precleaning is currently not MT but the reference processor 3736 // may be set for MT. Disable it temporarily here. 3737 ReferenceProcessor* rp = ref_processor(); 3738 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3739 3740 // Do one pass of scrubbing the discovered reference lists 3741 // to remove any reference objects with strongly-reachable 3742 // referents. 3743 if (clean_refs) { 3744 CMSPrecleanRefsYieldClosure yield_cl(this); 3745 assert(rp->span().equals(_span), "Spans should be equal"); 3746 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3747 &_markStack, true /* preclean */); 3748 CMSDrainMarkingStackClosure complete_trace(this, 3749 _span, &_markBitMap, &_markStack, 3750 &keep_alive, true /* preclean */); 3751 3752 // We don't want this step to interfere with a young 3753 // collection because we don't want to take CPU 3754 // or memory bandwidth away from the young GC threads 3755 // (which may be as many as there are CPUs). 3756 // Note that we don't need to protect ourselves from 3757 // interference with mutators because they can't 3758 // manipulate the discovered reference lists nor affect 3759 // the computed reachability of the referents, the 3760 // only properties manipulated by the precleaning 3761 // of these reference lists. 3762 stopTimer(); 3763 CMSTokenSyncWithLocks x(true /* is cms thread */, 3764 bitMapLock()); 3765 startTimer(); 3766 sample_eden(); 3767 3768 // The following will yield to allow foreground 3769 // collection to proceed promptly. XXX YSR: 3770 // The code in this method may need further 3771 // tweaking for better performance and some restructuring 3772 // for cleaner interfaces. 3773 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3774 rp->preclean_discovered_references( 3775 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3776 gc_timer); 3777 } 3778 3779 if (clean_survivor) { // preclean the active survivor space(s) 3780 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3781 &_markBitMap, &_modUnionTable, 3782 &_markStack, true /* precleaning phase */); 3783 stopTimer(); 3784 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3785 bitMapLock()); 3786 startTimer(); 3787 unsigned int before_count = 3788 GenCollectedHeap::heap()->total_collections(); 3789 SurvivorSpacePrecleanClosure 3790 sss_cl(this, _span, &_markBitMap, &_markStack, 3791 &pam_cl, before_count, CMSYield); 3792 _young_gen->from()->object_iterate_careful(&sss_cl); 3793 _young_gen->to()->object_iterate_careful(&sss_cl); 3794 } 3795 MarkRefsIntoAndScanClosure 3796 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3797 &_markStack, this, CMSYield, 3798 true /* precleaning phase */); 3799 // CAUTION: The following closure has persistent state that may need to 3800 // be reset upon a decrease in the sequence of addresses it 3801 // processes. 3802 ScanMarkedObjectsAgainCarefullyClosure 3803 smoac_cl(this, _span, 3804 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3805 3806 // Preclean dirty cards in ModUnionTable and CardTable using 3807 // appropriate convergence criterion; 3808 // repeat CMSPrecleanIter times unless we find that 3809 // we are losing. 3810 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3811 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3812 "Bad convergence multiplier"); 3813 assert(CMSPrecleanThreshold >= 100, 3814 "Unreasonably low CMSPrecleanThreshold"); 3815 3816 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3817 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3818 numIter < CMSPrecleanIter; 3819 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3820 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3821 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3822 // Either there are very few dirty cards, so re-mark 3823 // pause will be small anyway, or our pre-cleaning isn't 3824 // that much faster than the rate at which cards are being 3825 // dirtied, so we might as well stop and re-mark since 3826 // precleaning won't improve our re-mark time by much. 3827 if (curNumCards <= CMSPrecleanThreshold || 3828 (numIter > 0 && 3829 (curNumCards * CMSPrecleanDenominator > 3830 lastNumCards * CMSPrecleanNumerator))) { 3831 numIter++; 3832 cumNumCards += curNumCards; 3833 break; 3834 } 3835 } 3836 3837 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3838 3839 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3840 cumNumCards += curNumCards; 3841 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3842 curNumCards, cumNumCards, numIter); 3843 return cumNumCards; // as a measure of useful work done 3844 } 3845 3846 // PRECLEANING NOTES: 3847 // Precleaning involves: 3848 // . reading the bits of the modUnionTable and clearing the set bits. 3849 // . For the cards corresponding to the set bits, we scan the 3850 // objects on those cards. This means we need the free_list_lock 3851 // so that we can safely iterate over the CMS space when scanning 3852 // for oops. 3853 // . When we scan the objects, we'll be both reading and setting 3854 // marks in the marking bit map, so we'll need the marking bit map. 3855 // . For protecting _collector_state transitions, we take the CGC_lock. 3856 // Note that any races in the reading of of card table entries by the 3857 // CMS thread on the one hand and the clearing of those entries by the 3858 // VM thread or the setting of those entries by the mutator threads on the 3859 // other are quite benign. However, for efficiency it makes sense to keep 3860 // the VM thread from racing with the CMS thread while the latter is 3861 // dirty card info to the modUnionTable. We therefore also use the 3862 // CGC_lock to protect the reading of the card table and the mod union 3863 // table by the CM thread. 3864 // . We run concurrently with mutator updates, so scanning 3865 // needs to be done carefully -- we should not try to scan 3866 // potentially uninitialized objects. 3867 // 3868 // Locking strategy: While holding the CGC_lock, we scan over and 3869 // reset a maximal dirty range of the mod union / card tables, then lock 3870 // the free_list_lock and bitmap lock to do a full marking, then 3871 // release these locks; and repeat the cycle. This allows for a 3872 // certain amount of fairness in the sharing of these locks between 3873 // the CMS collector on the one hand, and the VM thread and the 3874 // mutators on the other. 3875 3876 // NOTE: preclean_mod_union_table() and preclean_card_table() 3877 // further below are largely identical; if you need to modify 3878 // one of these methods, please check the other method too. 3879 3880 size_t CMSCollector::preclean_mod_union_table( 3881 ConcurrentMarkSweepGeneration* old_gen, 3882 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3883 verify_work_stacks_empty(); 3884 verify_overflow_empty(); 3885 3886 // strategy: starting with the first card, accumulate contiguous 3887 // ranges of dirty cards; clear these cards, then scan the region 3888 // covered by these cards. 3889 3890 // Since all of the MUT is committed ahead, we can just use 3891 // that, in case the generations expand while we are precleaning. 3892 // It might also be fine to just use the committed part of the 3893 // generation, but we might potentially miss cards when the 3894 // generation is rapidly expanding while we are in the midst 3895 // of precleaning. 3896 HeapWord* startAddr = old_gen->reserved().start(); 3897 HeapWord* endAddr = old_gen->reserved().end(); 3898 3899 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3900 3901 size_t numDirtyCards, cumNumDirtyCards; 3902 HeapWord *nextAddr, *lastAddr; 3903 for (cumNumDirtyCards = numDirtyCards = 0, 3904 nextAddr = lastAddr = startAddr; 3905 nextAddr < endAddr; 3906 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3907 3908 ResourceMark rm; 3909 HandleMark hm; 3910 3911 MemRegion dirtyRegion; 3912 { 3913 stopTimer(); 3914 // Potential yield point 3915 CMSTokenSync ts(true); 3916 startTimer(); 3917 sample_eden(); 3918 // Get dirty region starting at nextOffset (inclusive), 3919 // simultaneously clearing it. 3920 dirtyRegion = 3921 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3922 assert(dirtyRegion.start() >= nextAddr, 3923 "returned region inconsistent?"); 3924 } 3925 // Remember where the next search should begin. 3926 // The returned region (if non-empty) is a right open interval, 3927 // so lastOffset is obtained from the right end of that 3928 // interval. 3929 lastAddr = dirtyRegion.end(); 3930 // Should do something more transparent and less hacky XXX 3931 numDirtyCards = 3932 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3933 3934 // We'll scan the cards in the dirty region (with periodic 3935 // yields for foreground GC as needed). 3936 if (!dirtyRegion.is_empty()) { 3937 assert(numDirtyCards > 0, "consistency check"); 3938 HeapWord* stop_point = NULL; 3939 stopTimer(); 3940 // Potential yield point 3941 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3942 bitMapLock()); 3943 startTimer(); 3944 { 3945 verify_work_stacks_empty(); 3946 verify_overflow_empty(); 3947 sample_eden(); 3948 stop_point = 3949 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3950 } 3951 if (stop_point != NULL) { 3952 // The careful iteration stopped early either because it found an 3953 // uninitialized object, or because we were in the midst of an 3954 // "abortable preclean", which should now be aborted. Redirty 3955 // the bits corresponding to the partially-scanned or unscanned 3956 // cards. We'll either restart at the next block boundary or 3957 // abort the preclean. 3958 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3959 "Should only be AbortablePreclean."); 3960 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3961 if (should_abort_preclean()) { 3962 break; // out of preclean loop 3963 } else { 3964 // Compute the next address at which preclean should pick up; 3965 // might need bitMapLock in order to read P-bits. 3966 lastAddr = next_card_start_after_block(stop_point); 3967 } 3968 } 3969 } else { 3970 assert(lastAddr == endAddr, "consistency check"); 3971 assert(numDirtyCards == 0, "consistency check"); 3972 break; 3973 } 3974 } 3975 verify_work_stacks_empty(); 3976 verify_overflow_empty(); 3977 return cumNumDirtyCards; 3978 } 3979 3980 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3981 // below are largely identical; if you need to modify 3982 // one of these methods, please check the other method too. 3983 3984 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3985 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3986 // strategy: it's similar to precleamModUnionTable above, in that 3987 // we accumulate contiguous ranges of dirty cards, mark these cards 3988 // precleaned, then scan the region covered by these cards. 3989 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3990 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3991 3992 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3993 3994 size_t numDirtyCards, cumNumDirtyCards; 3995 HeapWord *lastAddr, *nextAddr; 3996 3997 for (cumNumDirtyCards = numDirtyCards = 0, 3998 nextAddr = lastAddr = startAddr; 3999 nextAddr < endAddr; 4000 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4001 4002 ResourceMark rm; 4003 HandleMark hm; 4004 4005 MemRegion dirtyRegion; 4006 { 4007 // See comments in "Precleaning notes" above on why we 4008 // do this locking. XXX Could the locking overheads be 4009 // too high when dirty cards are sparse? [I don't think so.] 4010 stopTimer(); 4011 CMSTokenSync x(true); // is cms thread 4012 startTimer(); 4013 sample_eden(); 4014 // Get and clear dirty region from card table 4015 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4016 MemRegion(nextAddr, endAddr), 4017 true, 4018 CardTableModRefBS::precleaned_card_val()); 4019 4020 assert(dirtyRegion.start() >= nextAddr, 4021 "returned region inconsistent?"); 4022 } 4023 lastAddr = dirtyRegion.end(); 4024 numDirtyCards = 4025 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4026 4027 if (!dirtyRegion.is_empty()) { 4028 stopTimer(); 4029 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4030 startTimer(); 4031 sample_eden(); 4032 verify_work_stacks_empty(); 4033 verify_overflow_empty(); 4034 HeapWord* stop_point = 4035 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4036 if (stop_point != NULL) { 4037 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4038 "Should only be AbortablePreclean."); 4039 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4040 if (should_abort_preclean()) { 4041 break; // out of preclean loop 4042 } else { 4043 // Compute the next address at which preclean should pick up. 4044 lastAddr = next_card_start_after_block(stop_point); 4045 } 4046 } 4047 } else { 4048 break; 4049 } 4050 } 4051 verify_work_stacks_empty(); 4052 verify_overflow_empty(); 4053 return cumNumDirtyCards; 4054 } 4055 4056 class PrecleanKlassClosure : public KlassClosure { 4057 KlassToOopClosure _cm_klass_closure; 4058 public: 4059 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4060 void do_klass(Klass* k) { 4061 if (k->has_accumulated_modified_oops()) { 4062 k->clear_accumulated_modified_oops(); 4063 4064 _cm_klass_closure.do_klass(k); 4065 } 4066 } 4067 }; 4068 4069 // The freelist lock is needed to prevent asserts, is it really needed? 4070 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4071 4072 cl->set_freelistLock(freelistLock); 4073 4074 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4075 4076 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4077 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4078 PrecleanKlassClosure preclean_klass_closure(cl); 4079 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4080 4081 verify_work_stacks_empty(); 4082 verify_overflow_empty(); 4083 } 4084 4085 void CMSCollector::checkpointRootsFinal() { 4086 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4087 check_correct_thread_executing(); 4088 // world is stopped at this checkpoint 4089 assert(SafepointSynchronize::is_at_safepoint(), 4090 "world should be stopped"); 4091 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4092 4093 verify_work_stacks_empty(); 4094 verify_overflow_empty(); 4095 4096 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4097 _young_gen->used() / K, _young_gen->capacity() / K); 4098 { 4099 if (CMSScavengeBeforeRemark) { 4100 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4101 // Temporarily set flag to false, GCH->do_collection will 4102 // expect it to be false and set to true 4103 FlagSetting fl(gch->_is_gc_active, false); 4104 4105 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4106 4107 gch->do_collection(true, // full (i.e. force, see below) 4108 false, // !clear_all_soft_refs 4109 0, // size 4110 false, // is_tlab 4111 GenCollectedHeap::YoungGen // type 4112 ); 4113 } 4114 FreelistLocker x(this); 4115 MutexLockerEx y(bitMapLock(), 4116 Mutex::_no_safepoint_check_flag); 4117 checkpointRootsFinalWork(); 4118 } 4119 verify_work_stacks_empty(); 4120 verify_overflow_empty(); 4121 } 4122 4123 void CMSCollector::checkpointRootsFinalWork() { 4124 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4125 4126 assert(haveFreelistLocks(), "must have free list locks"); 4127 assert_lock_strong(bitMapLock()); 4128 4129 ResourceMark rm; 4130 HandleMark hm; 4131 4132 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4133 4134 if (should_unload_classes()) { 4135 CodeCache::gc_prologue(); 4136 } 4137 assert(haveFreelistLocks(), "must have free list locks"); 4138 assert_lock_strong(bitMapLock()); 4139 4140 // We might assume that we need not fill TLAB's when 4141 // CMSScavengeBeforeRemark is set, because we may have just done 4142 // a scavenge which would have filled all TLAB's -- and besides 4143 // Eden would be empty. This however may not always be the case -- 4144 // for instance although we asked for a scavenge, it may not have 4145 // happened because of a JNI critical section. We probably need 4146 // a policy for deciding whether we can in that case wait until 4147 // the critical section releases and then do the remark following 4148 // the scavenge, and skip it here. In the absence of that policy, 4149 // or of an indication of whether the scavenge did indeed occur, 4150 // we cannot rely on TLAB's having been filled and must do 4151 // so here just in case a scavenge did not happen. 4152 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4153 // Update the saved marks which may affect the root scans. 4154 gch->save_marks(); 4155 4156 print_eden_and_survivor_chunk_arrays(); 4157 4158 { 4159 #if defined(COMPILER2) || INCLUDE_JVMCI 4160 DerivedPointerTableDeactivate dpt_deact; 4161 #endif 4162 4163 // Note on the role of the mod union table: 4164 // Since the marker in "markFromRoots" marks concurrently with 4165 // mutators, it is possible for some reachable objects not to have been 4166 // scanned. For instance, an only reference to an object A was 4167 // placed in object B after the marker scanned B. Unless B is rescanned, 4168 // A would be collected. Such updates to references in marked objects 4169 // are detected via the mod union table which is the set of all cards 4170 // dirtied since the first checkpoint in this GC cycle and prior to 4171 // the most recent young generation GC, minus those cleaned up by the 4172 // concurrent precleaning. 4173 if (CMSParallelRemarkEnabled) { 4174 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4175 do_remark_parallel(); 4176 } else { 4177 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4178 do_remark_non_parallel(); 4179 } 4180 } 4181 verify_work_stacks_empty(); 4182 verify_overflow_empty(); 4183 4184 { 4185 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4186 refProcessingWork(); 4187 } 4188 verify_work_stacks_empty(); 4189 verify_overflow_empty(); 4190 4191 if (should_unload_classes()) { 4192 CodeCache::gc_epilogue(); 4193 } 4194 JvmtiExport::gc_epilogue(); 4195 4196 // If we encountered any (marking stack / work queue) overflow 4197 // events during the current CMS cycle, take appropriate 4198 // remedial measures, where possible, so as to try and avoid 4199 // recurrence of that condition. 4200 assert(_markStack.isEmpty(), "No grey objects"); 4201 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4202 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4203 if (ser_ovflw > 0) { 4204 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4205 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4206 _markStack.expand(); 4207 _ser_pmc_remark_ovflw = 0; 4208 _ser_pmc_preclean_ovflw = 0; 4209 _ser_kac_preclean_ovflw = 0; 4210 _ser_kac_ovflw = 0; 4211 } 4212 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4213 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4214 _par_pmc_remark_ovflw, _par_kac_ovflw); 4215 _par_pmc_remark_ovflw = 0; 4216 _par_kac_ovflw = 0; 4217 } 4218 if (_markStack._hit_limit > 0) { 4219 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4220 _markStack._hit_limit); 4221 } 4222 if (_markStack._failed_double > 0) { 4223 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4224 _markStack._failed_double, _markStack.capacity()); 4225 } 4226 _markStack._hit_limit = 0; 4227 _markStack._failed_double = 0; 4228 4229 if ((VerifyAfterGC || VerifyDuringGC) && 4230 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4231 verify_after_remark(); 4232 } 4233 4234 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4235 4236 // Change under the freelistLocks. 4237 _collectorState = Sweeping; 4238 // Call isAllClear() under bitMapLock 4239 assert(_modUnionTable.isAllClear(), 4240 "Should be clear by end of the final marking"); 4241 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4242 "Should be clear by end of the final marking"); 4243 } 4244 4245 void CMSParInitialMarkTask::work(uint worker_id) { 4246 elapsedTimer _timer; 4247 ResourceMark rm; 4248 HandleMark hm; 4249 4250 // ---------- scan from roots -------------- 4251 _timer.start(); 4252 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4253 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4254 4255 // ---------- young gen roots -------------- 4256 { 4257 work_on_young_gen_roots(worker_id, &par_mri_cl); 4258 _timer.stop(); 4259 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4260 } 4261 4262 // ---------- remaining roots -------------- 4263 _timer.reset(); 4264 _timer.start(); 4265 4266 CLDToOopClosure cld_closure(&par_mri_cl, true); 4267 4268 gch->gen_process_roots(_strong_roots_scope, 4269 GenCollectedHeap::OldGen, 4270 false, // yg was scanned above 4271 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4272 _collector->should_unload_classes(), 4273 &par_mri_cl, 4274 NULL, 4275 &cld_closure); 4276 assert(_collector->should_unload_classes() 4277 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4278 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4279 _timer.stop(); 4280 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4281 } 4282 4283 // Parallel remark task 4284 class CMSParRemarkTask: public CMSParMarkTask { 4285 CompactibleFreeListSpace* _cms_space; 4286 4287 // The per-thread work queues, available here for stealing. 4288 OopTaskQueueSet* _task_queues; 4289 ParallelTaskTerminator _term; 4290 StrongRootsScope* _strong_roots_scope; 4291 4292 public: 4293 // A value of 0 passed to n_workers will cause the number of 4294 // workers to be taken from the active workers in the work gang. 4295 CMSParRemarkTask(CMSCollector* collector, 4296 CompactibleFreeListSpace* cms_space, 4297 uint n_workers, WorkGang* workers, 4298 OopTaskQueueSet* task_queues, 4299 StrongRootsScope* strong_roots_scope): 4300 CMSParMarkTask("Rescan roots and grey objects in parallel", 4301 collector, n_workers), 4302 _cms_space(cms_space), 4303 _task_queues(task_queues), 4304 _term(n_workers, task_queues), 4305 _strong_roots_scope(strong_roots_scope) { } 4306 4307 OopTaskQueueSet* task_queues() { return _task_queues; } 4308 4309 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4310 4311 ParallelTaskTerminator* terminator() { return &_term; } 4312 uint n_workers() { return _n_workers; } 4313 4314 void work(uint worker_id); 4315 4316 private: 4317 // ... of dirty cards in old space 4318 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4319 ParMarkRefsIntoAndScanClosure* cl); 4320 4321 // ... work stealing for the above 4322 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4323 }; 4324 4325 class RemarkKlassClosure : public KlassClosure { 4326 KlassToOopClosure _cm_klass_closure; 4327 public: 4328 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4329 void do_klass(Klass* k) { 4330 // Check if we have modified any oops in the Klass during the concurrent marking. 4331 if (k->has_accumulated_modified_oops()) { 4332 k->clear_accumulated_modified_oops(); 4333 4334 // We could have transfered the current modified marks to the accumulated marks, 4335 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4336 } else if (k->has_modified_oops()) { 4337 // Don't clear anything, this info is needed by the next young collection. 4338 } else { 4339 // No modified oops in the Klass. 4340 return; 4341 } 4342 4343 // The klass has modified fields, need to scan the klass. 4344 _cm_klass_closure.do_klass(k); 4345 } 4346 }; 4347 4348 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4349 ParNewGeneration* young_gen = _collector->_young_gen; 4350 ContiguousSpace* eden_space = young_gen->eden(); 4351 ContiguousSpace* from_space = young_gen->from(); 4352 ContiguousSpace* to_space = young_gen->to(); 4353 4354 HeapWord** eca = _collector->_eden_chunk_array; 4355 size_t ect = _collector->_eden_chunk_index; 4356 HeapWord** sca = _collector->_survivor_chunk_array; 4357 size_t sct = _collector->_survivor_chunk_index; 4358 4359 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4360 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4361 4362 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4363 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4364 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4365 } 4366 4367 // work_queue(i) is passed to the closure 4368 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4369 // also is passed to do_dirty_card_rescan_tasks() and to 4370 // do_work_steal() to select the i-th task_queue. 4371 4372 void CMSParRemarkTask::work(uint worker_id) { 4373 elapsedTimer _timer; 4374 ResourceMark rm; 4375 HandleMark hm; 4376 4377 // ---------- rescan from roots -------------- 4378 _timer.start(); 4379 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4380 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4381 _collector->_span, _collector->ref_processor(), 4382 &(_collector->_markBitMap), 4383 work_queue(worker_id)); 4384 4385 // Rescan young gen roots first since these are likely 4386 // coarsely partitioned and may, on that account, constitute 4387 // the critical path; thus, it's best to start off that 4388 // work first. 4389 // ---------- young gen roots -------------- 4390 { 4391 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4392 _timer.stop(); 4393 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4394 } 4395 4396 // ---------- remaining roots -------------- 4397 _timer.reset(); 4398 _timer.start(); 4399 gch->gen_process_roots(_strong_roots_scope, 4400 GenCollectedHeap::OldGen, 4401 false, // yg was scanned above 4402 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4403 _collector->should_unload_classes(), 4404 &par_mrias_cl, 4405 NULL, 4406 NULL); // The dirty klasses will be handled below 4407 4408 assert(_collector->should_unload_classes() 4409 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4410 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4411 _timer.stop(); 4412 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4413 4414 // ---------- unhandled CLD scanning ---------- 4415 if (worker_id == 0) { // Single threaded at the moment. 4416 _timer.reset(); 4417 _timer.start(); 4418 4419 // Scan all new class loader data objects and new dependencies that were 4420 // introduced during concurrent marking. 4421 ResourceMark rm; 4422 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4423 for (int i = 0; i < array->length(); i++) { 4424 par_mrias_cl.do_cld_nv(array->at(i)); 4425 } 4426 4427 // We don't need to keep track of new CLDs anymore. 4428 ClassLoaderDataGraph::remember_new_clds(false); 4429 4430 _timer.stop(); 4431 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4432 } 4433 4434 // ---------- dirty klass scanning ---------- 4435 if (worker_id == 0) { // Single threaded at the moment. 4436 _timer.reset(); 4437 _timer.start(); 4438 4439 // Scan all classes that was dirtied during the concurrent marking phase. 4440 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4441 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4442 4443 _timer.stop(); 4444 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4445 } 4446 4447 // We might have added oops to ClassLoaderData::_handles during the 4448 // concurrent marking phase. These oops point to newly allocated objects 4449 // that are guaranteed to be kept alive. Either by the direct allocation 4450 // code, or when the young collector processes the roots. Hence, 4451 // we don't have to revisit the _handles block during the remark phase. 4452 4453 // ---------- rescan dirty cards ------------ 4454 _timer.reset(); 4455 _timer.start(); 4456 4457 // Do the rescan tasks for each of the two spaces 4458 // (cms_space) in turn. 4459 // "worker_id" is passed to select the task_queue for "worker_id" 4460 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4461 _timer.stop(); 4462 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4463 4464 // ---------- steal work from other threads ... 4465 // ---------- ... and drain overflow list. 4466 _timer.reset(); 4467 _timer.start(); 4468 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4469 _timer.stop(); 4470 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4471 } 4472 4473 // Note that parameter "i" is not used. 4474 void 4475 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4476 OopsInGenClosure* cl, ContiguousSpace* space, 4477 HeapWord** chunk_array, size_t chunk_top) { 4478 // Until all tasks completed: 4479 // . claim an unclaimed task 4480 // . compute region boundaries corresponding to task claimed 4481 // using chunk_array 4482 // . par_oop_iterate(cl) over that region 4483 4484 ResourceMark rm; 4485 HandleMark hm; 4486 4487 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4488 4489 uint nth_task = 0; 4490 uint n_tasks = pst->n_tasks(); 4491 4492 if (n_tasks > 0) { 4493 assert(pst->valid(), "Uninitialized use?"); 4494 HeapWord *start, *end; 4495 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4496 // We claimed task # nth_task; compute its boundaries. 4497 if (chunk_top == 0) { // no samples were taken 4498 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4499 start = space->bottom(); 4500 end = space->top(); 4501 } else if (nth_task == 0) { 4502 start = space->bottom(); 4503 end = chunk_array[nth_task]; 4504 } else if (nth_task < (uint)chunk_top) { 4505 assert(nth_task >= 1, "Control point invariant"); 4506 start = chunk_array[nth_task - 1]; 4507 end = chunk_array[nth_task]; 4508 } else { 4509 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4510 start = chunk_array[chunk_top - 1]; 4511 end = space->top(); 4512 } 4513 MemRegion mr(start, end); 4514 // Verify that mr is in space 4515 assert(mr.is_empty() || space->used_region().contains(mr), 4516 "Should be in space"); 4517 // Verify that "start" is an object boundary 4518 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4519 "Should be an oop"); 4520 space->par_oop_iterate(mr, cl); 4521 } 4522 pst->all_tasks_completed(); 4523 } 4524 } 4525 4526 void 4527 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4528 CompactibleFreeListSpace* sp, int i, 4529 ParMarkRefsIntoAndScanClosure* cl) { 4530 // Until all tasks completed: 4531 // . claim an unclaimed task 4532 // . compute region boundaries corresponding to task claimed 4533 // . transfer dirty bits ct->mut for that region 4534 // . apply rescanclosure to dirty mut bits for that region 4535 4536 ResourceMark rm; 4537 HandleMark hm; 4538 4539 OopTaskQueue* work_q = work_queue(i); 4540 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4541 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4542 // CAUTION: This closure has state that persists across calls to 4543 // the work method dirty_range_iterate_clear() in that it has 4544 // embedded in it a (subtype of) UpwardsObjectClosure. The 4545 // use of that state in the embedded UpwardsObjectClosure instance 4546 // assumes that the cards are always iterated (even if in parallel 4547 // by several threads) in monotonically increasing order per each 4548 // thread. This is true of the implementation below which picks 4549 // card ranges (chunks) in monotonically increasing order globally 4550 // and, a-fortiori, in monotonically increasing order per thread 4551 // (the latter order being a subsequence of the former). 4552 // If the work code below is ever reorganized into a more chaotic 4553 // work-partitioning form than the current "sequential tasks" 4554 // paradigm, the use of that persistent state will have to be 4555 // revisited and modified appropriately. See also related 4556 // bug 4756801 work on which should examine this code to make 4557 // sure that the changes there do not run counter to the 4558 // assumptions made here and necessary for correctness and 4559 // efficiency. Note also that this code might yield inefficient 4560 // behavior in the case of very large objects that span one or 4561 // more work chunks. Such objects would potentially be scanned 4562 // several times redundantly. Work on 4756801 should try and 4563 // address that performance anomaly if at all possible. XXX 4564 MemRegion full_span = _collector->_span; 4565 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4566 MarkFromDirtyCardsClosure 4567 greyRescanClosure(_collector, full_span, // entire span of interest 4568 sp, bm, work_q, cl); 4569 4570 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4571 assert(pst->valid(), "Uninitialized use?"); 4572 uint nth_task = 0; 4573 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4574 MemRegion span = sp->used_region(); 4575 HeapWord* start_addr = span.start(); 4576 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4577 alignment); 4578 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4579 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4580 start_addr, "Check alignment"); 4581 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4582 chunk_size, "Check alignment"); 4583 4584 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4585 // Having claimed the nth_task, compute corresponding mem-region, 4586 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4587 // The alignment restriction ensures that we do not need any 4588 // synchronization with other gang-workers while setting or 4589 // clearing bits in thus chunk of the MUT. 4590 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4591 start_addr + (nth_task+1)*chunk_size); 4592 // The last chunk's end might be way beyond end of the 4593 // used region. In that case pull back appropriately. 4594 if (this_span.end() > end_addr) { 4595 this_span.set_end(end_addr); 4596 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4597 } 4598 // Iterate over the dirty cards covering this chunk, marking them 4599 // precleaned, and setting the corresponding bits in the mod union 4600 // table. Since we have been careful to partition at Card and MUT-word 4601 // boundaries no synchronization is needed between parallel threads. 4602 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4603 &modUnionClosure); 4604 4605 // Having transferred these marks into the modUnionTable, 4606 // rescan the marked objects on the dirty cards in the modUnionTable. 4607 // Even if this is at a synchronous collection, the initial marking 4608 // may have been done during an asynchronous collection so there 4609 // may be dirty bits in the mod-union table. 4610 _collector->_modUnionTable.dirty_range_iterate_clear( 4611 this_span, &greyRescanClosure); 4612 _collector->_modUnionTable.verifyNoOneBitsInRange( 4613 this_span.start(), 4614 this_span.end()); 4615 } 4616 pst->all_tasks_completed(); // declare that i am done 4617 } 4618 4619 // . see if we can share work_queues with ParNew? XXX 4620 void 4621 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4622 int* seed) { 4623 OopTaskQueue* work_q = work_queue(i); 4624 NOT_PRODUCT(int num_steals = 0;) 4625 oop obj_to_scan; 4626 CMSBitMap* bm = &(_collector->_markBitMap); 4627 4628 while (true) { 4629 // Completely finish any left over work from (an) earlier round(s) 4630 cl->trim_queue(0); 4631 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4632 (size_t)ParGCDesiredObjsFromOverflowList); 4633 // Now check if there's any work in the overflow list 4634 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4635 // only affects the number of attempts made to get work from the 4636 // overflow list and does not affect the number of workers. Just 4637 // pass ParallelGCThreads so this behavior is unchanged. 4638 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4639 work_q, 4640 ParallelGCThreads)) { 4641 // found something in global overflow list; 4642 // not yet ready to go stealing work from others. 4643 // We'd like to assert(work_q->size() != 0, ...) 4644 // because we just took work from the overflow list, 4645 // but of course we can't since all of that could have 4646 // been already stolen from us. 4647 // "He giveth and He taketh away." 4648 continue; 4649 } 4650 // Verify that we have no work before we resort to stealing 4651 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4652 // Try to steal from other queues that have work 4653 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4654 NOT_PRODUCT(num_steals++;) 4655 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4656 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4657 // Do scanning work 4658 obj_to_scan->oop_iterate(cl); 4659 // Loop around, finish this work, and try to steal some more 4660 } else if (terminator()->offer_termination()) { 4661 break; // nirvana from the infinite cycle 4662 } 4663 } 4664 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4665 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4666 "Else our work is not yet done"); 4667 } 4668 4669 // Record object boundaries in _eden_chunk_array by sampling the eden 4670 // top in the slow-path eden object allocation code path and record 4671 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4672 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4673 // sampling in sample_eden() that activates during the part of the 4674 // preclean phase. 4675 void CMSCollector::sample_eden_chunk() { 4676 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4677 if (_eden_chunk_lock->try_lock()) { 4678 // Record a sample. This is the critical section. The contents 4679 // of the _eden_chunk_array have to be non-decreasing in the 4680 // address order. 4681 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4682 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4683 "Unexpected state of Eden"); 4684 if (_eden_chunk_index == 0 || 4685 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4686 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4687 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4688 _eden_chunk_index++; // commit sample 4689 } 4690 _eden_chunk_lock->unlock(); 4691 } 4692 } 4693 } 4694 4695 // Return a thread-local PLAB recording array, as appropriate. 4696 void* CMSCollector::get_data_recorder(int thr_num) { 4697 if (_survivor_plab_array != NULL && 4698 (CMSPLABRecordAlways || 4699 (_collectorState > Marking && _collectorState < FinalMarking))) { 4700 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4701 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4702 ca->reset(); // clear it so that fresh data is recorded 4703 return (void*) ca; 4704 } else { 4705 return NULL; 4706 } 4707 } 4708 4709 // Reset all the thread-local PLAB recording arrays 4710 void CMSCollector::reset_survivor_plab_arrays() { 4711 for (uint i = 0; i < ParallelGCThreads; i++) { 4712 _survivor_plab_array[i].reset(); 4713 } 4714 } 4715 4716 // Merge the per-thread plab arrays into the global survivor chunk 4717 // array which will provide the partitioning of the survivor space 4718 // for CMS initial scan and rescan. 4719 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4720 int no_of_gc_threads) { 4721 assert(_survivor_plab_array != NULL, "Error"); 4722 assert(_survivor_chunk_array != NULL, "Error"); 4723 assert(_collectorState == FinalMarking || 4724 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4725 for (int j = 0; j < no_of_gc_threads; j++) { 4726 _cursor[j] = 0; 4727 } 4728 HeapWord* top = surv->top(); 4729 size_t i; 4730 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4731 HeapWord* min_val = top; // Higher than any PLAB address 4732 uint min_tid = 0; // position of min_val this round 4733 for (int j = 0; j < no_of_gc_threads; j++) { 4734 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4735 if (_cursor[j] == cur_sca->end()) { 4736 continue; 4737 } 4738 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4739 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4740 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4741 if (cur_val < min_val) { 4742 min_tid = j; 4743 min_val = cur_val; 4744 } else { 4745 assert(cur_val < top, "All recorded addresses should be less"); 4746 } 4747 } 4748 // At this point min_val and min_tid are respectively 4749 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4750 // and the thread (j) that witnesses that address. 4751 // We record this address in the _survivor_chunk_array[i] 4752 // and increment _cursor[min_tid] prior to the next round i. 4753 if (min_val == top) { 4754 break; 4755 } 4756 _survivor_chunk_array[i] = min_val; 4757 _cursor[min_tid]++; 4758 } 4759 // We are all done; record the size of the _survivor_chunk_array 4760 _survivor_chunk_index = i; // exclusive: [0, i) 4761 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4762 // Verify that we used up all the recorded entries 4763 #ifdef ASSERT 4764 size_t total = 0; 4765 for (int j = 0; j < no_of_gc_threads; j++) { 4766 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4767 total += _cursor[j]; 4768 } 4769 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4770 // Check that the merged array is in sorted order 4771 if (total > 0) { 4772 for (size_t i = 0; i < total - 1; i++) { 4773 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4774 i, p2i(_survivor_chunk_array[i])); 4775 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4776 "Not sorted"); 4777 } 4778 } 4779 #endif // ASSERT 4780 } 4781 4782 // Set up the space's par_seq_tasks structure for work claiming 4783 // for parallel initial scan and rescan of young gen. 4784 // See ParRescanTask where this is currently used. 4785 void 4786 CMSCollector:: 4787 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4788 assert(n_threads > 0, "Unexpected n_threads argument"); 4789 4790 // Eden space 4791 if (!_young_gen->eden()->is_empty()) { 4792 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4793 assert(!pst->valid(), "Clobbering existing data?"); 4794 // Each valid entry in [0, _eden_chunk_index) represents a task. 4795 size_t n_tasks = _eden_chunk_index + 1; 4796 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4797 // Sets the condition for completion of the subtask (how many threads 4798 // need to finish in order to be done). 4799 pst->set_n_threads(n_threads); 4800 pst->set_n_tasks((int)n_tasks); 4801 } 4802 4803 // Merge the survivor plab arrays into _survivor_chunk_array 4804 if (_survivor_plab_array != NULL) { 4805 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4806 } else { 4807 assert(_survivor_chunk_index == 0, "Error"); 4808 } 4809 4810 // To space 4811 { 4812 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4813 assert(!pst->valid(), "Clobbering existing data?"); 4814 // Sets the condition for completion of the subtask (how many threads 4815 // need to finish in order to be done). 4816 pst->set_n_threads(n_threads); 4817 pst->set_n_tasks(1); 4818 assert(pst->valid(), "Error"); 4819 } 4820 4821 // From space 4822 { 4823 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4824 assert(!pst->valid(), "Clobbering existing data?"); 4825 size_t n_tasks = _survivor_chunk_index + 1; 4826 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4827 // Sets the condition for completion of the subtask (how many threads 4828 // need to finish in order to be done). 4829 pst->set_n_threads(n_threads); 4830 pst->set_n_tasks((int)n_tasks); 4831 assert(pst->valid(), "Error"); 4832 } 4833 } 4834 4835 // Parallel version of remark 4836 void CMSCollector::do_remark_parallel() { 4837 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4838 WorkGang* workers = gch->workers(); 4839 assert(workers != NULL, "Need parallel worker threads."); 4840 // Choose to use the number of GC workers most recently set 4841 // into "active_workers". 4842 uint n_workers = workers->active_workers(); 4843 4844 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4845 4846 StrongRootsScope srs(n_workers); 4847 4848 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4849 4850 // We won't be iterating over the cards in the card table updating 4851 // the younger_gen cards, so we shouldn't call the following else 4852 // the verification code as well as subsequent younger_refs_iterate 4853 // code would get confused. XXX 4854 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4855 4856 // The young gen rescan work will not be done as part of 4857 // process_roots (which currently doesn't know how to 4858 // parallelize such a scan), but rather will be broken up into 4859 // a set of parallel tasks (via the sampling that the [abortable] 4860 // preclean phase did of eden, plus the [two] tasks of 4861 // scanning the [two] survivor spaces. Further fine-grain 4862 // parallelization of the scanning of the survivor spaces 4863 // themselves, and of precleaning of the young gen itself 4864 // is deferred to the future. 4865 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4866 4867 // The dirty card rescan work is broken up into a "sequence" 4868 // of parallel tasks (per constituent space) that are dynamically 4869 // claimed by the parallel threads. 4870 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4871 4872 // It turns out that even when we're using 1 thread, doing the work in a 4873 // separate thread causes wide variance in run times. We can't help this 4874 // in the multi-threaded case, but we special-case n=1 here to get 4875 // repeatable measurements of the 1-thread overhead of the parallel code. 4876 if (n_workers > 1) { 4877 // Make refs discovery MT-safe, if it isn't already: it may not 4878 // necessarily be so, since it's possible that we are doing 4879 // ST marking. 4880 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4881 workers->run_task(&tsk); 4882 } else { 4883 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4884 tsk.work(0); 4885 } 4886 4887 // restore, single-threaded for now, any preserved marks 4888 // as a result of work_q overflow 4889 restore_preserved_marks_if_any(); 4890 } 4891 4892 // Non-parallel version of remark 4893 void CMSCollector::do_remark_non_parallel() { 4894 ResourceMark rm; 4895 HandleMark hm; 4896 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4897 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4898 4899 MarkRefsIntoAndScanClosure 4900 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4901 &_markStack, this, 4902 false /* should_yield */, false /* not precleaning */); 4903 MarkFromDirtyCardsClosure 4904 markFromDirtyCardsClosure(this, _span, 4905 NULL, // space is set further below 4906 &_markBitMap, &_markStack, &mrias_cl); 4907 { 4908 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4909 // Iterate over the dirty cards, setting the corresponding bits in the 4910 // mod union table. 4911 { 4912 ModUnionClosure modUnionClosure(&_modUnionTable); 4913 _ct->ct_bs()->dirty_card_iterate( 4914 _cmsGen->used_region(), 4915 &modUnionClosure); 4916 } 4917 // Having transferred these marks into the modUnionTable, we just need 4918 // to rescan the marked objects on the dirty cards in the modUnionTable. 4919 // The initial marking may have been done during an asynchronous 4920 // collection so there may be dirty bits in the mod-union table. 4921 const int alignment = 4922 CardTableModRefBS::card_size * BitsPerWord; 4923 { 4924 // ... First handle dirty cards in CMS gen 4925 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4926 MemRegion ur = _cmsGen->used_region(); 4927 HeapWord* lb = ur.start(); 4928 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4929 MemRegion cms_span(lb, ub); 4930 _modUnionTable.dirty_range_iterate_clear(cms_span, 4931 &markFromDirtyCardsClosure); 4932 verify_work_stacks_empty(); 4933 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4934 } 4935 } 4936 if (VerifyDuringGC && 4937 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4938 HandleMark hm; // Discard invalid handles created during verification 4939 Universe::verify(); 4940 } 4941 { 4942 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4943 4944 verify_work_stacks_empty(); 4945 4946 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4947 StrongRootsScope srs(1); 4948 4949 gch->gen_process_roots(&srs, 4950 GenCollectedHeap::OldGen, 4951 true, // young gen as roots 4952 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4953 should_unload_classes(), 4954 &mrias_cl, 4955 NULL, 4956 NULL); // The dirty klasses will be handled below 4957 4958 assert(should_unload_classes() 4959 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4960 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4961 } 4962 4963 { 4964 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4965 4966 verify_work_stacks_empty(); 4967 4968 // Scan all class loader data objects that might have been introduced 4969 // during concurrent marking. 4970 ResourceMark rm; 4971 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4972 for (int i = 0; i < array->length(); i++) { 4973 mrias_cl.do_cld_nv(array->at(i)); 4974 } 4975 4976 // We don't need to keep track of new CLDs anymore. 4977 ClassLoaderDataGraph::remember_new_clds(false); 4978 4979 verify_work_stacks_empty(); 4980 } 4981 4982 { 4983 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4984 4985 verify_work_stacks_empty(); 4986 4987 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4988 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4989 4990 verify_work_stacks_empty(); 4991 } 4992 4993 // We might have added oops to ClassLoaderData::_handles during the 4994 // concurrent marking phase. These oops point to newly allocated objects 4995 // that are guaranteed to be kept alive. Either by the direct allocation 4996 // code, or when the young collector processes the roots. Hence, 4997 // we don't have to revisit the _handles block during the remark phase. 4998 4999 verify_work_stacks_empty(); 5000 // Restore evacuated mark words, if any, used for overflow list links 5001 restore_preserved_marks_if_any(); 5002 5003 verify_overflow_empty(); 5004 } 5005 5006 //////////////////////////////////////////////////////// 5007 // Parallel Reference Processing Task Proxy Class 5008 //////////////////////////////////////////////////////// 5009 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5010 OopTaskQueueSet* _queues; 5011 ParallelTaskTerminator _terminator; 5012 public: 5013 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5014 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5015 ParallelTaskTerminator* terminator() { return &_terminator; } 5016 OopTaskQueueSet* queues() { return _queues; } 5017 }; 5018 5019 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5020 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5021 CMSCollector* _collector; 5022 CMSBitMap* _mark_bit_map; 5023 const MemRegion _span; 5024 ProcessTask& _task; 5025 5026 public: 5027 CMSRefProcTaskProxy(ProcessTask& task, 5028 CMSCollector* collector, 5029 const MemRegion& span, 5030 CMSBitMap* mark_bit_map, 5031 AbstractWorkGang* workers, 5032 OopTaskQueueSet* task_queues): 5033 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5034 task_queues, 5035 workers->active_workers()), 5036 _task(task), 5037 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5038 { 5039 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5040 "Inconsistency in _span"); 5041 } 5042 5043 OopTaskQueueSet* task_queues() { return queues(); } 5044 5045 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5046 5047 void do_work_steal(int i, 5048 CMSParDrainMarkingStackClosure* drain, 5049 CMSParKeepAliveClosure* keep_alive, 5050 int* seed); 5051 5052 virtual void work(uint worker_id); 5053 }; 5054 5055 void CMSRefProcTaskProxy::work(uint worker_id) { 5056 ResourceMark rm; 5057 HandleMark hm; 5058 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5059 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5060 _mark_bit_map, 5061 work_queue(worker_id)); 5062 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5063 _mark_bit_map, 5064 work_queue(worker_id)); 5065 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5066 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5067 if (_task.marks_oops_alive()) { 5068 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5069 _collector->hash_seed(worker_id)); 5070 } 5071 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5072 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5073 } 5074 5075 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5076 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5077 EnqueueTask& _task; 5078 5079 public: 5080 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5081 : AbstractGangTask("Enqueue reference objects in parallel"), 5082 _task(task) 5083 { } 5084 5085 virtual void work(uint worker_id) 5086 { 5087 _task.work(worker_id); 5088 } 5089 }; 5090 5091 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5092 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5093 _span(span), 5094 _bit_map(bit_map), 5095 _work_queue(work_queue), 5096 _mark_and_push(collector, span, bit_map, work_queue), 5097 _low_water_mark(MIN2((work_queue->max_elems()/4), 5098 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5099 { } 5100 5101 // . see if we can share work_queues with ParNew? XXX 5102 void CMSRefProcTaskProxy::do_work_steal(int i, 5103 CMSParDrainMarkingStackClosure* drain, 5104 CMSParKeepAliveClosure* keep_alive, 5105 int* seed) { 5106 OopTaskQueue* work_q = work_queue(i); 5107 NOT_PRODUCT(int num_steals = 0;) 5108 oop obj_to_scan; 5109 5110 while (true) { 5111 // Completely finish any left over work from (an) earlier round(s) 5112 drain->trim_queue(0); 5113 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5114 (size_t)ParGCDesiredObjsFromOverflowList); 5115 // Now check if there's any work in the overflow list 5116 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5117 // only affects the number of attempts made to get work from the 5118 // overflow list and does not affect the number of workers. Just 5119 // pass ParallelGCThreads so this behavior is unchanged. 5120 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5121 work_q, 5122 ParallelGCThreads)) { 5123 // Found something in global overflow list; 5124 // not yet ready to go stealing work from others. 5125 // We'd like to assert(work_q->size() != 0, ...) 5126 // because we just took work from the overflow list, 5127 // but of course we can't, since all of that might have 5128 // been already stolen from us. 5129 continue; 5130 } 5131 // Verify that we have no work before we resort to stealing 5132 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5133 // Try to steal from other queues that have work 5134 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5135 NOT_PRODUCT(num_steals++;) 5136 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5137 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5138 // Do scanning work 5139 obj_to_scan->oop_iterate(keep_alive); 5140 // Loop around, finish this work, and try to steal some more 5141 } else if (terminator()->offer_termination()) { 5142 break; // nirvana from the infinite cycle 5143 } 5144 } 5145 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5146 } 5147 5148 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5149 { 5150 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5151 WorkGang* workers = gch->workers(); 5152 assert(workers != NULL, "Need parallel worker threads."); 5153 CMSRefProcTaskProxy rp_task(task, &_collector, 5154 _collector.ref_processor()->span(), 5155 _collector.markBitMap(), 5156 workers, _collector.task_queues()); 5157 workers->run_task(&rp_task); 5158 } 5159 5160 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5161 { 5162 5163 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5164 WorkGang* workers = gch->workers(); 5165 assert(workers != NULL, "Need parallel worker threads."); 5166 CMSRefEnqueueTaskProxy enq_task(task); 5167 workers->run_task(&enq_task); 5168 } 5169 5170 void CMSCollector::refProcessingWork() { 5171 ResourceMark rm; 5172 HandleMark hm; 5173 5174 ReferenceProcessor* rp = ref_processor(); 5175 assert(rp->span().equals(_span), "Spans should be equal"); 5176 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5177 // Process weak references. 5178 rp->setup_policy(false); 5179 verify_work_stacks_empty(); 5180 5181 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5182 &_markStack, false /* !preclean */); 5183 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5184 _span, &_markBitMap, &_markStack, 5185 &cmsKeepAliveClosure, false /* !preclean */); 5186 { 5187 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5188 5189 ReferenceProcessorStats stats; 5190 if (rp->processing_is_mt()) { 5191 // Set the degree of MT here. If the discovery is done MT, there 5192 // may have been a different number of threads doing the discovery 5193 // and a different number of discovered lists may have Ref objects. 5194 // That is OK as long as the Reference lists are balanced (see 5195 // balance_all_queues() and balance_queues()). 5196 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5197 uint active_workers = ParallelGCThreads; 5198 WorkGang* workers = gch->workers(); 5199 if (workers != NULL) { 5200 active_workers = workers->active_workers(); 5201 // The expectation is that active_workers will have already 5202 // been set to a reasonable value. If it has not been set, 5203 // investigate. 5204 assert(active_workers > 0, "Should have been set during scavenge"); 5205 } 5206 rp->set_active_mt_degree(active_workers); 5207 CMSRefProcTaskExecutor task_executor(*this); 5208 stats = rp->process_discovered_references(&_is_alive_closure, 5209 &cmsKeepAliveClosure, 5210 &cmsDrainMarkingStackClosure, 5211 &task_executor, 5212 _gc_timer_cm); 5213 } else { 5214 stats = rp->process_discovered_references(&_is_alive_closure, 5215 &cmsKeepAliveClosure, 5216 &cmsDrainMarkingStackClosure, 5217 NULL, 5218 _gc_timer_cm); 5219 } 5220 _gc_tracer_cm->report_gc_reference_stats(stats); 5221 5222 } 5223 5224 // This is the point where the entire marking should have completed. 5225 verify_work_stacks_empty(); 5226 5227 if (should_unload_classes()) { 5228 { 5229 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5230 5231 // Unload classes and purge the SystemDictionary. 5232 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5233 5234 // Unload nmethods. 5235 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5236 5237 // Prune dead klasses from subklass/sibling/implementor lists. 5238 Klass::clean_weak_klass_links(&_is_alive_closure); 5239 } 5240 5241 { 5242 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5243 // Clean up unreferenced symbols in symbol table. 5244 SymbolTable::unlink(); 5245 } 5246 5247 { 5248 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5249 // Delete entries for dead interned strings. 5250 StringTable::unlink(&_is_alive_closure); 5251 } 5252 } 5253 5254 5255 // Restore any preserved marks as a result of mark stack or 5256 // work queue overflow 5257 restore_preserved_marks_if_any(); // done single-threaded for now 5258 5259 rp->set_enqueuing_is_done(true); 5260 if (rp->processing_is_mt()) { 5261 rp->balance_all_queues(); 5262 CMSRefProcTaskExecutor task_executor(*this); 5263 rp->enqueue_discovered_references(&task_executor); 5264 } else { 5265 rp->enqueue_discovered_references(NULL); 5266 } 5267 rp->verify_no_references_recorded(); 5268 assert(!rp->discovery_enabled(), "should have been disabled"); 5269 } 5270 5271 #ifndef PRODUCT 5272 void CMSCollector::check_correct_thread_executing() { 5273 Thread* t = Thread::current(); 5274 // Only the VM thread or the CMS thread should be here. 5275 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5276 "Unexpected thread type"); 5277 // If this is the vm thread, the foreground process 5278 // should not be waiting. Note that _foregroundGCIsActive is 5279 // true while the foreground collector is waiting. 5280 if (_foregroundGCShouldWait) { 5281 // We cannot be the VM thread 5282 assert(t->is_ConcurrentGC_thread(), 5283 "Should be CMS thread"); 5284 } else { 5285 // We can be the CMS thread only if we are in a stop-world 5286 // phase of CMS collection. 5287 if (t->is_ConcurrentGC_thread()) { 5288 assert(_collectorState == InitialMarking || 5289 _collectorState == FinalMarking, 5290 "Should be a stop-world phase"); 5291 // The CMS thread should be holding the CMS_token. 5292 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5293 "Potential interference with concurrently " 5294 "executing VM thread"); 5295 } 5296 } 5297 } 5298 #endif 5299 5300 void CMSCollector::sweep() { 5301 assert(_collectorState == Sweeping, "just checking"); 5302 check_correct_thread_executing(); 5303 verify_work_stacks_empty(); 5304 verify_overflow_empty(); 5305 increment_sweep_count(); 5306 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5307 5308 _inter_sweep_timer.stop(); 5309 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5310 5311 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5312 _intra_sweep_timer.reset(); 5313 _intra_sweep_timer.start(); 5314 { 5315 GCTraceCPUTime tcpu; 5316 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5317 // First sweep the old gen 5318 { 5319 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5320 bitMapLock()); 5321 sweepWork(_cmsGen); 5322 } 5323 5324 // Update Universe::_heap_*_at_gc figures. 5325 // We need all the free list locks to make the abstract state 5326 // transition from Sweeping to Resetting. See detailed note 5327 // further below. 5328 { 5329 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5330 // Update heap occupancy information which is used as 5331 // input to soft ref clearing policy at the next gc. 5332 Universe::update_heap_info_at_gc(); 5333 _collectorState = Resizing; 5334 } 5335 } 5336 verify_work_stacks_empty(); 5337 verify_overflow_empty(); 5338 5339 if (should_unload_classes()) { 5340 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5341 // requires that the virtual spaces are stable and not deleted. 5342 ClassLoaderDataGraph::set_should_purge(true); 5343 } 5344 5345 _intra_sweep_timer.stop(); 5346 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5347 5348 _inter_sweep_timer.reset(); 5349 _inter_sweep_timer.start(); 5350 5351 // We need to use a monotonically non-decreasing time in ms 5352 // or we will see time-warp warnings and os::javaTimeMillis() 5353 // does not guarantee monotonicity. 5354 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5355 update_time_of_last_gc(now); 5356 5357 // NOTE on abstract state transitions: 5358 // Mutators allocate-live and/or mark the mod-union table dirty 5359 // based on the state of the collection. The former is done in 5360 // the interval [Marking, Sweeping] and the latter in the interval 5361 // [Marking, Sweeping). Thus the transitions into the Marking state 5362 // and out of the Sweeping state must be synchronously visible 5363 // globally to the mutators. 5364 // The transition into the Marking state happens with the world 5365 // stopped so the mutators will globally see it. Sweeping is 5366 // done asynchronously by the background collector so the transition 5367 // from the Sweeping state to the Resizing state must be done 5368 // under the freelistLock (as is the check for whether to 5369 // allocate-live and whether to dirty the mod-union table). 5370 assert(_collectorState == Resizing, "Change of collector state to" 5371 " Resizing must be done under the freelistLocks (plural)"); 5372 5373 // Now that sweeping has been completed, we clear 5374 // the incremental_collection_failed flag, 5375 // thus inviting a younger gen collection to promote into 5376 // this generation. If such a promotion may still fail, 5377 // the flag will be set again when a young collection is 5378 // attempted. 5379 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5380 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5381 gch->update_full_collections_completed(_collection_count_start); 5382 } 5383 5384 // FIX ME!!! Looks like this belongs in CFLSpace, with 5385 // CMSGen merely delegating to it. 5386 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5387 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5388 HeapWord* minAddr = _cmsSpace->bottom(); 5389 HeapWord* largestAddr = 5390 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5391 if (largestAddr == NULL) { 5392 // The dictionary appears to be empty. In this case 5393 // try to coalesce at the end of the heap. 5394 largestAddr = _cmsSpace->end(); 5395 } 5396 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5397 size_t nearLargestOffset = 5398 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5399 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5400 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5401 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5402 } 5403 5404 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5405 return addr >= _cmsSpace->nearLargestChunk(); 5406 } 5407 5408 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5409 return _cmsSpace->find_chunk_at_end(); 5410 } 5411 5412 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5413 bool full) { 5414 // If the young generation has been collected, gather any statistics 5415 // that are of interest at this point. 5416 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5417 if (!full && current_is_young) { 5418 // Gather statistics on the young generation collection. 5419 collector()->stats().record_gc0_end(used()); 5420 } 5421 } 5422 5423 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5424 // We iterate over the space(s) underlying this generation, 5425 // checking the mark bit map to see if the bits corresponding 5426 // to specific blocks are marked or not. Blocks that are 5427 // marked are live and are not swept up. All remaining blocks 5428 // are swept up, with coalescing on-the-fly as we sweep up 5429 // contiguous free and/or garbage blocks: 5430 // We need to ensure that the sweeper synchronizes with allocators 5431 // and stop-the-world collectors. In particular, the following 5432 // locks are used: 5433 // . CMS token: if this is held, a stop the world collection cannot occur 5434 // . freelistLock: if this is held no allocation can occur from this 5435 // generation by another thread 5436 // . bitMapLock: if this is held, no other thread can access or update 5437 // 5438 5439 // Note that we need to hold the freelistLock if we use 5440 // block iterate below; else the iterator might go awry if 5441 // a mutator (or promotion) causes block contents to change 5442 // (for instance if the allocator divvies up a block). 5443 // If we hold the free list lock, for all practical purposes 5444 // young generation GC's can't occur (they'll usually need to 5445 // promote), so we might as well prevent all young generation 5446 // GC's while we do a sweeping step. For the same reason, we might 5447 // as well take the bit map lock for the entire duration 5448 5449 // check that we hold the requisite locks 5450 assert(have_cms_token(), "Should hold cms token"); 5451 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5452 assert_lock_strong(old_gen->freelistLock()); 5453 assert_lock_strong(bitMapLock()); 5454 5455 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5456 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5457 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5458 _inter_sweep_estimate.padded_average(), 5459 _intra_sweep_estimate.padded_average()); 5460 old_gen->setNearLargestChunk(); 5461 5462 { 5463 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5464 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5465 // We need to free-up/coalesce garbage/blocks from a 5466 // co-terminal free run. This is done in the SweepClosure 5467 // destructor; so, do not remove this scope, else the 5468 // end-of-sweep-census below will be off by a little bit. 5469 } 5470 old_gen->cmsSpace()->sweep_completed(); 5471 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5472 if (should_unload_classes()) { // unloaded classes this cycle, 5473 _concurrent_cycles_since_last_unload = 0; // ... reset count 5474 } else { // did not unload classes, 5475 _concurrent_cycles_since_last_unload++; // ... increment count 5476 } 5477 } 5478 5479 // Reset CMS data structures (for now just the marking bit map) 5480 // preparatory for the next cycle. 5481 void CMSCollector::reset_concurrent() { 5482 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5483 5484 // If the state is not "Resetting", the foreground thread 5485 // has done a collection and the resetting. 5486 if (_collectorState != Resetting) { 5487 assert(_collectorState == Idling, "The state should only change" 5488 " because the foreground collector has finished the collection"); 5489 return; 5490 } 5491 5492 { 5493 // Clear the mark bitmap (no grey objects to start with) 5494 // for the next cycle. 5495 GCTraceCPUTime tcpu; 5496 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5497 5498 HeapWord* curAddr = _markBitMap.startWord(); 5499 while (curAddr < _markBitMap.endWord()) { 5500 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5501 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5502 _markBitMap.clear_large_range(chunk); 5503 if (ConcurrentMarkSweepThread::should_yield() && 5504 !foregroundGCIsActive() && 5505 CMSYield) { 5506 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5507 "CMS thread should hold CMS token"); 5508 assert_lock_strong(bitMapLock()); 5509 bitMapLock()->unlock(); 5510 ConcurrentMarkSweepThread::desynchronize(true); 5511 stopTimer(); 5512 incrementYields(); 5513 5514 // See the comment in coordinator_yield() 5515 for (unsigned i = 0; i < CMSYieldSleepCount && 5516 ConcurrentMarkSweepThread::should_yield() && 5517 !CMSCollector::foregroundGCIsActive(); ++i) { 5518 os::sleep(Thread::current(), 1, false); 5519 } 5520 5521 ConcurrentMarkSweepThread::synchronize(true); 5522 bitMapLock()->lock_without_safepoint_check(); 5523 startTimer(); 5524 } 5525 curAddr = chunk.end(); 5526 } 5527 // A successful mostly concurrent collection has been done. 5528 // Because only the full (i.e., concurrent mode failure) collections 5529 // are being measured for gc overhead limits, clean the "near" flag 5530 // and count. 5531 size_policy()->reset_gc_overhead_limit_count(); 5532 _collectorState = Idling; 5533 } 5534 5535 register_gc_end(); 5536 } 5537 5538 // Same as above but for STW paths 5539 void CMSCollector::reset_stw() { 5540 // already have the lock 5541 assert(_collectorState == Resetting, "just checking"); 5542 assert_lock_strong(bitMapLock()); 5543 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5544 _markBitMap.clear_all(); 5545 _collectorState = Idling; 5546 register_gc_end(); 5547 } 5548 5549 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5550 GCTraceCPUTime tcpu; 5551 TraceCollectorStats tcs(counters()); 5552 5553 switch (op) { 5554 case CMS_op_checkpointRootsInitial: { 5555 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5556 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5557 checkpointRootsInitial(); 5558 break; 5559 } 5560 case CMS_op_checkpointRootsFinal: { 5561 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5562 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5563 checkpointRootsFinal(); 5564 break; 5565 } 5566 default: 5567 fatal("No such CMS_op"); 5568 } 5569 } 5570 5571 #ifndef PRODUCT 5572 size_t const CMSCollector::skip_header_HeapWords() { 5573 return FreeChunk::header_size(); 5574 } 5575 5576 // Try and collect here conditions that should hold when 5577 // CMS thread is exiting. The idea is that the foreground GC 5578 // thread should not be blocked if it wants to terminate 5579 // the CMS thread and yet continue to run the VM for a while 5580 // after that. 5581 void CMSCollector::verify_ok_to_terminate() const { 5582 assert(Thread::current()->is_ConcurrentGC_thread(), 5583 "should be called by CMS thread"); 5584 assert(!_foregroundGCShouldWait, "should be false"); 5585 // We could check here that all the various low-level locks 5586 // are not held by the CMS thread, but that is overkill; see 5587 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5588 // is checked. 5589 } 5590 #endif 5591 5592 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5593 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5594 "missing Printezis mark?"); 5595 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5596 size_t size = pointer_delta(nextOneAddr + 1, addr); 5597 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5598 "alignment problem"); 5599 assert(size >= 3, "Necessary for Printezis marks to work"); 5600 return size; 5601 } 5602 5603 // A variant of the above (block_size_using_printezis_bits()) except 5604 // that we return 0 if the P-bits are not yet set. 5605 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5606 if (_markBitMap.isMarked(addr + 1)) { 5607 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5608 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5609 size_t size = pointer_delta(nextOneAddr + 1, addr); 5610 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5611 "alignment problem"); 5612 assert(size >= 3, "Necessary for Printezis marks to work"); 5613 return size; 5614 } 5615 return 0; 5616 } 5617 5618 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5619 size_t sz = 0; 5620 oop p = (oop)addr; 5621 if (p->klass_or_null() != NULL) { 5622 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5623 } else { 5624 sz = block_size_using_printezis_bits(addr); 5625 } 5626 assert(sz > 0, "size must be nonzero"); 5627 HeapWord* next_block = addr + sz; 5628 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5629 CardTableModRefBS::card_size); 5630 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5631 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5632 "must be different cards"); 5633 return next_card; 5634 } 5635 5636 5637 // CMS Bit Map Wrapper ///////////////////////////////////////// 5638 5639 // Construct a CMS bit map infrastructure, but don't create the 5640 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5641 // further below. 5642 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5643 _bm(), 5644 _shifter(shifter), 5645 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5646 Monitor::_safepoint_check_sometimes) : NULL) 5647 { 5648 _bmStartWord = 0; 5649 _bmWordSize = 0; 5650 } 5651 5652 bool CMSBitMap::allocate(MemRegion mr) { 5653 _bmStartWord = mr.start(); 5654 _bmWordSize = mr.word_size(); 5655 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5656 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5657 if (!brs.is_reserved()) { 5658 log_warning(gc)("CMS bit map allocation failure"); 5659 return false; 5660 } 5661 // For now we'll just commit all of the bit map up front. 5662 // Later on we'll try to be more parsimonious with swap. 5663 if (!_virtual_space.initialize(brs, brs.size())) { 5664 log_warning(gc)("CMS bit map backing store failure"); 5665 return false; 5666 } 5667 assert(_virtual_space.committed_size() == brs.size(), 5668 "didn't reserve backing store for all of CMS bit map?"); 5669 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5670 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5671 _bmWordSize, "inconsistency in bit map sizing"); 5672 _bm.set_size(_bmWordSize >> _shifter); 5673 5674 // bm.clear(); // can we rely on getting zero'd memory? verify below 5675 assert(isAllClear(), 5676 "Expected zero'd memory from ReservedSpace constructor"); 5677 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5678 "consistency check"); 5679 return true; 5680 } 5681 5682 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5683 HeapWord *next_addr, *end_addr, *last_addr; 5684 assert_locked(); 5685 assert(covers(mr), "out-of-range error"); 5686 // XXX assert that start and end are appropriately aligned 5687 for (next_addr = mr.start(), end_addr = mr.end(); 5688 next_addr < end_addr; next_addr = last_addr) { 5689 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5690 last_addr = dirty_region.end(); 5691 if (!dirty_region.is_empty()) { 5692 cl->do_MemRegion(dirty_region); 5693 } else { 5694 assert(last_addr == end_addr, "program logic"); 5695 return; 5696 } 5697 } 5698 } 5699 5700 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5701 _bm.print_on_error(st, prefix); 5702 } 5703 5704 #ifndef PRODUCT 5705 void CMSBitMap::assert_locked() const { 5706 CMSLockVerifier::assert_locked(lock()); 5707 } 5708 5709 bool CMSBitMap::covers(MemRegion mr) const { 5710 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5711 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5712 "size inconsistency"); 5713 return (mr.start() >= _bmStartWord) && 5714 (mr.end() <= endWord()); 5715 } 5716 5717 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5718 return (start >= _bmStartWord && (start + size) <= endWord()); 5719 } 5720 5721 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5722 // verify that there are no 1 bits in the interval [left, right) 5723 FalseBitMapClosure falseBitMapClosure; 5724 iterate(&falseBitMapClosure, left, right); 5725 } 5726 5727 void CMSBitMap::region_invariant(MemRegion mr) 5728 { 5729 assert_locked(); 5730 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5731 assert(!mr.is_empty(), "unexpected empty region"); 5732 assert(covers(mr), "mr should be covered by bit map"); 5733 // convert address range into offset range 5734 size_t start_ofs = heapWordToOffset(mr.start()); 5735 // Make sure that end() is appropriately aligned 5736 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5737 (1 << (_shifter+LogHeapWordSize))), 5738 "Misaligned mr.end()"); 5739 size_t end_ofs = heapWordToOffset(mr.end()); 5740 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5741 } 5742 5743 #endif 5744 5745 bool CMSMarkStack::allocate(size_t size) { 5746 // allocate a stack of the requisite depth 5747 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5748 size * sizeof(oop))); 5749 if (!rs.is_reserved()) { 5750 log_warning(gc)("CMSMarkStack allocation failure"); 5751 return false; 5752 } 5753 if (!_virtual_space.initialize(rs, rs.size())) { 5754 log_warning(gc)("CMSMarkStack backing store failure"); 5755 return false; 5756 } 5757 assert(_virtual_space.committed_size() == rs.size(), 5758 "didn't reserve backing store for all of CMS stack?"); 5759 _base = (oop*)(_virtual_space.low()); 5760 _index = 0; 5761 _capacity = size; 5762 NOT_PRODUCT(_max_depth = 0); 5763 return true; 5764 } 5765 5766 // XXX FIX ME !!! In the MT case we come in here holding a 5767 // leaf lock. For printing we need to take a further lock 5768 // which has lower rank. We need to recalibrate the two 5769 // lock-ranks involved in order to be able to print the 5770 // messages below. (Or defer the printing to the caller. 5771 // For now we take the expedient path of just disabling the 5772 // messages for the problematic case.) 5773 void CMSMarkStack::expand() { 5774 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5775 if (_capacity == MarkStackSizeMax) { 5776 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5777 // We print a warning message only once per CMS cycle. 5778 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5779 } 5780 return; 5781 } 5782 // Double capacity if possible 5783 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5784 // Do not give up existing stack until we have managed to 5785 // get the double capacity that we desired. 5786 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5787 new_capacity * sizeof(oop))); 5788 if (rs.is_reserved()) { 5789 // Release the backing store associated with old stack 5790 _virtual_space.release(); 5791 // Reinitialize virtual space for new stack 5792 if (!_virtual_space.initialize(rs, rs.size())) { 5793 fatal("Not enough swap for expanded marking stack"); 5794 } 5795 _base = (oop*)(_virtual_space.low()); 5796 _index = 0; 5797 _capacity = new_capacity; 5798 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5799 // Failed to double capacity, continue; 5800 // we print a detail message only once per CMS cycle. 5801 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5802 _capacity / K, new_capacity / K); 5803 } 5804 } 5805 5806 5807 // Closures 5808 // XXX: there seems to be a lot of code duplication here; 5809 // should refactor and consolidate common code. 5810 5811 // This closure is used to mark refs into the CMS generation in 5812 // the CMS bit map. Called at the first checkpoint. This closure 5813 // assumes that we do not need to re-mark dirty cards; if the CMS 5814 // generation on which this is used is not an oldest 5815 // generation then this will lose younger_gen cards! 5816 5817 MarkRefsIntoClosure::MarkRefsIntoClosure( 5818 MemRegion span, CMSBitMap* bitMap): 5819 _span(span), 5820 _bitMap(bitMap) 5821 { 5822 assert(ref_processor() == NULL, "deliberately left NULL"); 5823 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5824 } 5825 5826 void MarkRefsIntoClosure::do_oop(oop obj) { 5827 // if p points into _span, then mark corresponding bit in _markBitMap 5828 assert(obj->is_oop(), "expected an oop"); 5829 HeapWord* addr = (HeapWord*)obj; 5830 if (_span.contains(addr)) { 5831 // this should be made more efficient 5832 _bitMap->mark(addr); 5833 } 5834 } 5835 5836 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5837 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5838 5839 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5840 MemRegion span, CMSBitMap* bitMap): 5841 _span(span), 5842 _bitMap(bitMap) 5843 { 5844 assert(ref_processor() == NULL, "deliberately left NULL"); 5845 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5846 } 5847 5848 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5849 // if p points into _span, then mark corresponding bit in _markBitMap 5850 assert(obj->is_oop(), "expected an oop"); 5851 HeapWord* addr = (HeapWord*)obj; 5852 if (_span.contains(addr)) { 5853 // this should be made more efficient 5854 _bitMap->par_mark(addr); 5855 } 5856 } 5857 5858 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5859 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5860 5861 // A variant of the above, used for CMS marking verification. 5862 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5863 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5864 _span(span), 5865 _verification_bm(verification_bm), 5866 _cms_bm(cms_bm) 5867 { 5868 assert(ref_processor() == NULL, "deliberately left NULL"); 5869 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5870 } 5871 5872 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5873 // if p points into _span, then mark corresponding bit in _markBitMap 5874 assert(obj->is_oop(), "expected an oop"); 5875 HeapWord* addr = (HeapWord*)obj; 5876 if (_span.contains(addr)) { 5877 _verification_bm->mark(addr); 5878 if (!_cms_bm->isMarked(addr)) { 5879 LogHandle(gc, verify) log; 5880 ResourceMark rm; 5881 oop(addr)->print_on(log.error_stream()); 5882 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5883 fatal("... aborting"); 5884 } 5885 } 5886 } 5887 5888 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5889 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5890 5891 ////////////////////////////////////////////////// 5892 // MarkRefsIntoAndScanClosure 5893 ////////////////////////////////////////////////// 5894 5895 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5896 ReferenceProcessor* rp, 5897 CMSBitMap* bit_map, 5898 CMSBitMap* mod_union_table, 5899 CMSMarkStack* mark_stack, 5900 CMSCollector* collector, 5901 bool should_yield, 5902 bool concurrent_precleaning): 5903 _collector(collector), 5904 _span(span), 5905 _bit_map(bit_map), 5906 _mark_stack(mark_stack), 5907 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5908 mark_stack, concurrent_precleaning), 5909 _yield(should_yield), 5910 _concurrent_precleaning(concurrent_precleaning), 5911 _freelistLock(NULL) 5912 { 5913 // FIXME: Should initialize in base class constructor. 5914 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5915 set_ref_processor_internal(rp); 5916 } 5917 5918 // This closure is used to mark refs into the CMS generation at the 5919 // second (final) checkpoint, and to scan and transitively follow 5920 // the unmarked oops. It is also used during the concurrent precleaning 5921 // phase while scanning objects on dirty cards in the CMS generation. 5922 // The marks are made in the marking bit map and the marking stack is 5923 // used for keeping the (newly) grey objects during the scan. 5924 // The parallel version (Par_...) appears further below. 5925 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5926 if (obj != NULL) { 5927 assert(obj->is_oop(), "expected an oop"); 5928 HeapWord* addr = (HeapWord*)obj; 5929 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5930 assert(_collector->overflow_list_is_empty(), 5931 "overflow list should be empty"); 5932 if (_span.contains(addr) && 5933 !_bit_map->isMarked(addr)) { 5934 // mark bit map (object is now grey) 5935 _bit_map->mark(addr); 5936 // push on marking stack (stack should be empty), and drain the 5937 // stack by applying this closure to the oops in the oops popped 5938 // from the stack (i.e. blacken the grey objects) 5939 bool res = _mark_stack->push(obj); 5940 assert(res, "Should have space to push on empty stack"); 5941 do { 5942 oop new_oop = _mark_stack->pop(); 5943 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5944 assert(_bit_map->isMarked((HeapWord*)new_oop), 5945 "only grey objects on this stack"); 5946 // iterate over the oops in this oop, marking and pushing 5947 // the ones in CMS heap (i.e. in _span). 5948 new_oop->oop_iterate(&_pushAndMarkClosure); 5949 // check if it's time to yield 5950 do_yield_check(); 5951 } while (!_mark_stack->isEmpty() || 5952 (!_concurrent_precleaning && take_from_overflow_list())); 5953 // if marking stack is empty, and we are not doing this 5954 // during precleaning, then check the overflow list 5955 } 5956 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5957 assert(_collector->overflow_list_is_empty(), 5958 "overflow list was drained above"); 5959 5960 assert(_collector->no_preserved_marks(), 5961 "All preserved marks should have been restored above"); 5962 } 5963 } 5964 5965 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5966 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5967 5968 void MarkRefsIntoAndScanClosure::do_yield_work() { 5969 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5970 "CMS thread should hold CMS token"); 5971 assert_lock_strong(_freelistLock); 5972 assert_lock_strong(_bit_map->lock()); 5973 // relinquish the free_list_lock and bitMaplock() 5974 _bit_map->lock()->unlock(); 5975 _freelistLock->unlock(); 5976 ConcurrentMarkSweepThread::desynchronize(true); 5977 _collector->stopTimer(); 5978 _collector->incrementYields(); 5979 5980 // See the comment in coordinator_yield() 5981 for (unsigned i = 0; 5982 i < CMSYieldSleepCount && 5983 ConcurrentMarkSweepThread::should_yield() && 5984 !CMSCollector::foregroundGCIsActive(); 5985 ++i) { 5986 os::sleep(Thread::current(), 1, false); 5987 } 5988 5989 ConcurrentMarkSweepThread::synchronize(true); 5990 _freelistLock->lock_without_safepoint_check(); 5991 _bit_map->lock()->lock_without_safepoint_check(); 5992 _collector->startTimer(); 5993 } 5994 5995 /////////////////////////////////////////////////////////// 5996 // ParMarkRefsIntoAndScanClosure: a parallel version of 5997 // MarkRefsIntoAndScanClosure 5998 /////////////////////////////////////////////////////////// 5999 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6000 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6001 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6002 _span(span), 6003 _bit_map(bit_map), 6004 _work_queue(work_queue), 6005 _low_water_mark(MIN2((work_queue->max_elems()/4), 6006 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6007 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6008 { 6009 // FIXME: Should initialize in base class constructor. 6010 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6011 set_ref_processor_internal(rp); 6012 } 6013 6014 // This closure is used to mark refs into the CMS generation at the 6015 // second (final) checkpoint, and to scan and transitively follow 6016 // the unmarked oops. The marks are made in the marking bit map and 6017 // the work_queue is used for keeping the (newly) grey objects during 6018 // the scan phase whence they are also available for stealing by parallel 6019 // threads. Since the marking bit map is shared, updates are 6020 // synchronized (via CAS). 6021 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6022 if (obj != NULL) { 6023 // Ignore mark word because this could be an already marked oop 6024 // that may be chained at the end of the overflow list. 6025 assert(obj->is_oop(true), "expected an oop"); 6026 HeapWord* addr = (HeapWord*)obj; 6027 if (_span.contains(addr) && 6028 !_bit_map->isMarked(addr)) { 6029 // mark bit map (object will become grey): 6030 // It is possible for several threads to be 6031 // trying to "claim" this object concurrently; 6032 // the unique thread that succeeds in marking the 6033 // object first will do the subsequent push on 6034 // to the work queue (or overflow list). 6035 if (_bit_map->par_mark(addr)) { 6036 // push on work_queue (which may not be empty), and trim the 6037 // queue to an appropriate length by applying this closure to 6038 // the oops in the oops popped from the stack (i.e. blacken the 6039 // grey objects) 6040 bool res = _work_queue->push(obj); 6041 assert(res, "Low water mark should be less than capacity?"); 6042 trim_queue(_low_water_mark); 6043 } // Else, another thread claimed the object 6044 } 6045 } 6046 } 6047 6048 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6049 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6050 6051 // This closure is used to rescan the marked objects on the dirty cards 6052 // in the mod union table and the card table proper. 6053 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6054 oop p, MemRegion mr) { 6055 6056 size_t size = 0; 6057 HeapWord* addr = (HeapWord*)p; 6058 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6059 assert(_span.contains(addr), "we are scanning the CMS generation"); 6060 // check if it's time to yield 6061 if (do_yield_check()) { 6062 // We yielded for some foreground stop-world work, 6063 // and we have been asked to abort this ongoing preclean cycle. 6064 return 0; 6065 } 6066 if (_bitMap->isMarked(addr)) { 6067 // it's marked; is it potentially uninitialized? 6068 if (p->klass_or_null() != NULL) { 6069 // an initialized object; ignore mark word in verification below 6070 // since we are running concurrent with mutators 6071 assert(p->is_oop(true), "should be an oop"); 6072 if (p->is_objArray()) { 6073 // objArrays are precisely marked; restrict scanning 6074 // to dirty cards only. 6075 size = CompactibleFreeListSpace::adjustObjectSize( 6076 p->oop_iterate_size(_scanningClosure, mr)); 6077 } else { 6078 // A non-array may have been imprecisely marked; we need 6079 // to scan object in its entirety. 6080 size = CompactibleFreeListSpace::adjustObjectSize( 6081 p->oop_iterate_size(_scanningClosure)); 6082 } 6083 #ifdef ASSERT 6084 size_t direct_size = 6085 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6086 assert(size == direct_size, "Inconsistency in size"); 6087 assert(size >= 3, "Necessary for Printezis marks to work"); 6088 if (!_bitMap->isMarked(addr+1)) { 6089 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6090 } else { 6091 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6092 assert(_bitMap->isMarked(addr+size-1), 6093 "inconsistent Printezis mark"); 6094 } 6095 #endif // ASSERT 6096 } else { 6097 // An uninitialized object. 6098 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6099 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6100 size = pointer_delta(nextOneAddr + 1, addr); 6101 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6102 "alignment problem"); 6103 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6104 // will dirty the card when the klass pointer is installed in the 6105 // object (signaling the completion of initialization). 6106 } 6107 } else { 6108 // Either a not yet marked object or an uninitialized object 6109 if (p->klass_or_null() == NULL) { 6110 // An uninitialized object, skip to the next card, since 6111 // we may not be able to read its P-bits yet. 6112 assert(size == 0, "Initial value"); 6113 } else { 6114 // An object not (yet) reached by marking: we merely need to 6115 // compute its size so as to go look at the next block. 6116 assert(p->is_oop(true), "should be an oop"); 6117 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6118 } 6119 } 6120 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6121 return size; 6122 } 6123 6124 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6125 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6126 "CMS thread should hold CMS token"); 6127 assert_lock_strong(_freelistLock); 6128 assert_lock_strong(_bitMap->lock()); 6129 // relinquish the free_list_lock and bitMaplock() 6130 _bitMap->lock()->unlock(); 6131 _freelistLock->unlock(); 6132 ConcurrentMarkSweepThread::desynchronize(true); 6133 _collector->stopTimer(); 6134 _collector->incrementYields(); 6135 6136 // See the comment in coordinator_yield() 6137 for (unsigned i = 0; i < CMSYieldSleepCount && 6138 ConcurrentMarkSweepThread::should_yield() && 6139 !CMSCollector::foregroundGCIsActive(); ++i) { 6140 os::sleep(Thread::current(), 1, false); 6141 } 6142 6143 ConcurrentMarkSweepThread::synchronize(true); 6144 _freelistLock->lock_without_safepoint_check(); 6145 _bitMap->lock()->lock_without_safepoint_check(); 6146 _collector->startTimer(); 6147 } 6148 6149 6150 ////////////////////////////////////////////////////////////////// 6151 // SurvivorSpacePrecleanClosure 6152 ////////////////////////////////////////////////////////////////// 6153 // This (single-threaded) closure is used to preclean the oops in 6154 // the survivor spaces. 6155 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6156 6157 HeapWord* addr = (HeapWord*)p; 6158 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6159 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6160 assert(p->klass_or_null() != NULL, "object should be initialized"); 6161 // an initialized object; ignore mark word in verification below 6162 // since we are running concurrent with mutators 6163 assert(p->is_oop(true), "should be an oop"); 6164 // Note that we do not yield while we iterate over 6165 // the interior oops of p, pushing the relevant ones 6166 // on our marking stack. 6167 size_t size = p->oop_iterate_size(_scanning_closure); 6168 do_yield_check(); 6169 // Observe that below, we do not abandon the preclean 6170 // phase as soon as we should; rather we empty the 6171 // marking stack before returning. This is to satisfy 6172 // some existing assertions. In general, it may be a 6173 // good idea to abort immediately and complete the marking 6174 // from the grey objects at a later time. 6175 while (!_mark_stack->isEmpty()) { 6176 oop new_oop = _mark_stack->pop(); 6177 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6178 assert(_bit_map->isMarked((HeapWord*)new_oop), 6179 "only grey objects on this stack"); 6180 // iterate over the oops in this oop, marking and pushing 6181 // the ones in CMS heap (i.e. in _span). 6182 new_oop->oop_iterate(_scanning_closure); 6183 // check if it's time to yield 6184 do_yield_check(); 6185 } 6186 unsigned int after_count = 6187 GenCollectedHeap::heap()->total_collections(); 6188 bool abort = (_before_count != after_count) || 6189 _collector->should_abort_preclean(); 6190 return abort ? 0 : size; 6191 } 6192 6193 void SurvivorSpacePrecleanClosure::do_yield_work() { 6194 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6195 "CMS thread should hold CMS token"); 6196 assert_lock_strong(_bit_map->lock()); 6197 // Relinquish the bit map lock 6198 _bit_map->lock()->unlock(); 6199 ConcurrentMarkSweepThread::desynchronize(true); 6200 _collector->stopTimer(); 6201 _collector->incrementYields(); 6202 6203 // See the comment in coordinator_yield() 6204 for (unsigned i = 0; i < CMSYieldSleepCount && 6205 ConcurrentMarkSweepThread::should_yield() && 6206 !CMSCollector::foregroundGCIsActive(); ++i) { 6207 os::sleep(Thread::current(), 1, false); 6208 } 6209 6210 ConcurrentMarkSweepThread::synchronize(true); 6211 _bit_map->lock()->lock_without_safepoint_check(); 6212 _collector->startTimer(); 6213 } 6214 6215 // This closure is used to rescan the marked objects on the dirty cards 6216 // in the mod union table and the card table proper. In the parallel 6217 // case, although the bitMap is shared, we do a single read so the 6218 // isMarked() query is "safe". 6219 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6220 // Ignore mark word because we are running concurrent with mutators 6221 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6222 HeapWord* addr = (HeapWord*)p; 6223 assert(_span.contains(addr), "we are scanning the CMS generation"); 6224 bool is_obj_array = false; 6225 #ifdef ASSERT 6226 if (!_parallel) { 6227 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6228 assert(_collector->overflow_list_is_empty(), 6229 "overflow list should be empty"); 6230 6231 } 6232 #endif // ASSERT 6233 if (_bit_map->isMarked(addr)) { 6234 // Obj arrays are precisely marked, non-arrays are not; 6235 // so we scan objArrays precisely and non-arrays in their 6236 // entirety. 6237 if (p->is_objArray()) { 6238 is_obj_array = true; 6239 if (_parallel) { 6240 p->oop_iterate(_par_scan_closure, mr); 6241 } else { 6242 p->oop_iterate(_scan_closure, mr); 6243 } 6244 } else { 6245 if (_parallel) { 6246 p->oop_iterate(_par_scan_closure); 6247 } else { 6248 p->oop_iterate(_scan_closure); 6249 } 6250 } 6251 } 6252 #ifdef ASSERT 6253 if (!_parallel) { 6254 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6255 assert(_collector->overflow_list_is_empty(), 6256 "overflow list should be empty"); 6257 6258 } 6259 #endif // ASSERT 6260 return is_obj_array; 6261 } 6262 6263 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6264 MemRegion span, 6265 CMSBitMap* bitMap, CMSMarkStack* markStack, 6266 bool should_yield, bool verifying): 6267 _collector(collector), 6268 _span(span), 6269 _bitMap(bitMap), 6270 _mut(&collector->_modUnionTable), 6271 _markStack(markStack), 6272 _yield(should_yield), 6273 _skipBits(0) 6274 { 6275 assert(_markStack->isEmpty(), "stack should be empty"); 6276 _finger = _bitMap->startWord(); 6277 _threshold = _finger; 6278 assert(_collector->_restart_addr == NULL, "Sanity check"); 6279 assert(_span.contains(_finger), "Out of bounds _finger?"); 6280 DEBUG_ONLY(_verifying = verifying;) 6281 } 6282 6283 void MarkFromRootsClosure::reset(HeapWord* addr) { 6284 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6285 assert(_span.contains(addr), "Out of bounds _finger?"); 6286 _finger = addr; 6287 _threshold = (HeapWord*)round_to( 6288 (intptr_t)_finger, CardTableModRefBS::card_size); 6289 } 6290 6291 // Should revisit to see if this should be restructured for 6292 // greater efficiency. 6293 bool MarkFromRootsClosure::do_bit(size_t offset) { 6294 if (_skipBits > 0) { 6295 _skipBits--; 6296 return true; 6297 } 6298 // convert offset into a HeapWord* 6299 HeapWord* addr = _bitMap->startWord() + offset; 6300 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6301 "address out of range"); 6302 assert(_bitMap->isMarked(addr), "tautology"); 6303 if (_bitMap->isMarked(addr+1)) { 6304 // this is an allocated but not yet initialized object 6305 assert(_skipBits == 0, "tautology"); 6306 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6307 oop p = oop(addr); 6308 if (p->klass_or_null() == NULL) { 6309 DEBUG_ONLY(if (!_verifying) {) 6310 // We re-dirty the cards on which this object lies and increase 6311 // the _threshold so that we'll come back to scan this object 6312 // during the preclean or remark phase. (CMSCleanOnEnter) 6313 if (CMSCleanOnEnter) { 6314 size_t sz = _collector->block_size_using_printezis_bits(addr); 6315 HeapWord* end_card_addr = (HeapWord*)round_to( 6316 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6317 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6318 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6319 // Bump _threshold to end_card_addr; note that 6320 // _threshold cannot possibly exceed end_card_addr, anyhow. 6321 // This prevents future clearing of the card as the scan proceeds 6322 // to the right. 6323 assert(_threshold <= end_card_addr, 6324 "Because we are just scanning into this object"); 6325 if (_threshold < end_card_addr) { 6326 _threshold = end_card_addr; 6327 } 6328 if (p->klass_or_null() != NULL) { 6329 // Redirty the range of cards... 6330 _mut->mark_range(redirty_range); 6331 } // ...else the setting of klass will dirty the card anyway. 6332 } 6333 DEBUG_ONLY(}) 6334 return true; 6335 } 6336 } 6337 scanOopsInOop(addr); 6338 return true; 6339 } 6340 6341 // We take a break if we've been at this for a while, 6342 // so as to avoid monopolizing the locks involved. 6343 void MarkFromRootsClosure::do_yield_work() { 6344 // First give up the locks, then yield, then re-lock 6345 // We should probably use a constructor/destructor idiom to 6346 // do this unlock/lock or modify the MutexUnlocker class to 6347 // serve our purpose. XXX 6348 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6349 "CMS thread should hold CMS token"); 6350 assert_lock_strong(_bitMap->lock()); 6351 _bitMap->lock()->unlock(); 6352 ConcurrentMarkSweepThread::desynchronize(true); 6353 _collector->stopTimer(); 6354 _collector->incrementYields(); 6355 6356 // See the comment in coordinator_yield() 6357 for (unsigned i = 0; i < CMSYieldSleepCount && 6358 ConcurrentMarkSweepThread::should_yield() && 6359 !CMSCollector::foregroundGCIsActive(); ++i) { 6360 os::sleep(Thread::current(), 1, false); 6361 } 6362 6363 ConcurrentMarkSweepThread::synchronize(true); 6364 _bitMap->lock()->lock_without_safepoint_check(); 6365 _collector->startTimer(); 6366 } 6367 6368 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6369 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6370 assert(_markStack->isEmpty(), 6371 "should drain stack to limit stack usage"); 6372 // convert ptr to an oop preparatory to scanning 6373 oop obj = oop(ptr); 6374 // Ignore mark word in verification below, since we 6375 // may be running concurrent with mutators. 6376 assert(obj->is_oop(true), "should be an oop"); 6377 assert(_finger <= ptr, "_finger runneth ahead"); 6378 // advance the finger to right end of this object 6379 _finger = ptr + obj->size(); 6380 assert(_finger > ptr, "we just incremented it above"); 6381 // On large heaps, it may take us some time to get through 6382 // the marking phase. During 6383 // this time it's possible that a lot of mutations have 6384 // accumulated in the card table and the mod union table -- 6385 // these mutation records are redundant until we have 6386 // actually traced into the corresponding card. 6387 // Here, we check whether advancing the finger would make 6388 // us cross into a new card, and if so clear corresponding 6389 // cards in the MUT (preclean them in the card-table in the 6390 // future). 6391 6392 DEBUG_ONLY(if (!_verifying) {) 6393 // The clean-on-enter optimization is disabled by default, 6394 // until we fix 6178663. 6395 if (CMSCleanOnEnter && (_finger > _threshold)) { 6396 // [_threshold, _finger) represents the interval 6397 // of cards to be cleared in MUT (or precleaned in card table). 6398 // The set of cards to be cleared is all those that overlap 6399 // with the interval [_threshold, _finger); note that 6400 // _threshold is always kept card-aligned but _finger isn't 6401 // always card-aligned. 6402 HeapWord* old_threshold = _threshold; 6403 assert(old_threshold == (HeapWord*)round_to( 6404 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6405 "_threshold should always be card-aligned"); 6406 _threshold = (HeapWord*)round_to( 6407 (intptr_t)_finger, CardTableModRefBS::card_size); 6408 MemRegion mr(old_threshold, _threshold); 6409 assert(!mr.is_empty(), "Control point invariant"); 6410 assert(_span.contains(mr), "Should clear within span"); 6411 _mut->clear_range(mr); 6412 } 6413 DEBUG_ONLY(}) 6414 // Note: the finger doesn't advance while we drain 6415 // the stack below. 6416 PushOrMarkClosure pushOrMarkClosure(_collector, 6417 _span, _bitMap, _markStack, 6418 _finger, this); 6419 bool res = _markStack->push(obj); 6420 assert(res, "Empty non-zero size stack should have space for single push"); 6421 while (!_markStack->isEmpty()) { 6422 oop new_oop = _markStack->pop(); 6423 // Skip verifying header mark word below because we are 6424 // running concurrent with mutators. 6425 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6426 // now scan this oop's oops 6427 new_oop->oop_iterate(&pushOrMarkClosure); 6428 do_yield_check(); 6429 } 6430 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6431 } 6432 6433 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6434 CMSCollector* collector, MemRegion span, 6435 CMSBitMap* bit_map, 6436 OopTaskQueue* work_queue, 6437 CMSMarkStack* overflow_stack): 6438 _collector(collector), 6439 _whole_span(collector->_span), 6440 _span(span), 6441 _bit_map(bit_map), 6442 _mut(&collector->_modUnionTable), 6443 _work_queue(work_queue), 6444 _overflow_stack(overflow_stack), 6445 _skip_bits(0), 6446 _task(task) 6447 { 6448 assert(_work_queue->size() == 0, "work_queue should be empty"); 6449 _finger = span.start(); 6450 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6451 assert(_span.contains(_finger), "Out of bounds _finger?"); 6452 } 6453 6454 // Should revisit to see if this should be restructured for 6455 // greater efficiency. 6456 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6457 if (_skip_bits > 0) { 6458 _skip_bits--; 6459 return true; 6460 } 6461 // convert offset into a HeapWord* 6462 HeapWord* addr = _bit_map->startWord() + offset; 6463 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6464 "address out of range"); 6465 assert(_bit_map->isMarked(addr), "tautology"); 6466 if (_bit_map->isMarked(addr+1)) { 6467 // this is an allocated object that might not yet be initialized 6468 assert(_skip_bits == 0, "tautology"); 6469 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6470 oop p = oop(addr); 6471 if (p->klass_or_null() == NULL) { 6472 // in the case of Clean-on-Enter optimization, redirty card 6473 // and avoid clearing card by increasing the threshold. 6474 return true; 6475 } 6476 } 6477 scan_oops_in_oop(addr); 6478 return true; 6479 } 6480 6481 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6482 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6483 // Should we assert that our work queue is empty or 6484 // below some drain limit? 6485 assert(_work_queue->size() == 0, 6486 "should drain stack to limit stack usage"); 6487 // convert ptr to an oop preparatory to scanning 6488 oop obj = oop(ptr); 6489 // Ignore mark word in verification below, since we 6490 // may be running concurrent with mutators. 6491 assert(obj->is_oop(true), "should be an oop"); 6492 assert(_finger <= ptr, "_finger runneth ahead"); 6493 // advance the finger to right end of this object 6494 _finger = ptr + obj->size(); 6495 assert(_finger > ptr, "we just incremented it above"); 6496 // On large heaps, it may take us some time to get through 6497 // the marking phase. During 6498 // this time it's possible that a lot of mutations have 6499 // accumulated in the card table and the mod union table -- 6500 // these mutation records are redundant until we have 6501 // actually traced into the corresponding card. 6502 // Here, we check whether advancing the finger would make 6503 // us cross into a new card, and if so clear corresponding 6504 // cards in the MUT (preclean them in the card-table in the 6505 // future). 6506 6507 // The clean-on-enter optimization is disabled by default, 6508 // until we fix 6178663. 6509 if (CMSCleanOnEnter && (_finger > _threshold)) { 6510 // [_threshold, _finger) represents the interval 6511 // of cards to be cleared in MUT (or precleaned in card table). 6512 // The set of cards to be cleared is all those that overlap 6513 // with the interval [_threshold, _finger); note that 6514 // _threshold is always kept card-aligned but _finger isn't 6515 // always card-aligned. 6516 HeapWord* old_threshold = _threshold; 6517 assert(old_threshold == (HeapWord*)round_to( 6518 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6519 "_threshold should always be card-aligned"); 6520 _threshold = (HeapWord*)round_to( 6521 (intptr_t)_finger, CardTableModRefBS::card_size); 6522 MemRegion mr(old_threshold, _threshold); 6523 assert(!mr.is_empty(), "Control point invariant"); 6524 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6525 _mut->clear_range(mr); 6526 } 6527 6528 // Note: the local finger doesn't advance while we drain 6529 // the stack below, but the global finger sure can and will. 6530 HeapWord** gfa = _task->global_finger_addr(); 6531 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6532 _span, _bit_map, 6533 _work_queue, 6534 _overflow_stack, 6535 _finger, 6536 gfa, this); 6537 bool res = _work_queue->push(obj); // overflow could occur here 6538 assert(res, "Will hold once we use workqueues"); 6539 while (true) { 6540 oop new_oop; 6541 if (!_work_queue->pop_local(new_oop)) { 6542 // We emptied our work_queue; check if there's stuff that can 6543 // be gotten from the overflow stack. 6544 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6545 _overflow_stack, _work_queue)) { 6546 do_yield_check(); 6547 continue; 6548 } else { // done 6549 break; 6550 } 6551 } 6552 // Skip verifying header mark word below because we are 6553 // running concurrent with mutators. 6554 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6555 // now scan this oop's oops 6556 new_oop->oop_iterate(&pushOrMarkClosure); 6557 do_yield_check(); 6558 } 6559 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6560 } 6561 6562 // Yield in response to a request from VM Thread or 6563 // from mutators. 6564 void ParMarkFromRootsClosure::do_yield_work() { 6565 assert(_task != NULL, "sanity"); 6566 _task->yield(); 6567 } 6568 6569 // A variant of the above used for verifying CMS marking work. 6570 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6571 MemRegion span, 6572 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6573 CMSMarkStack* mark_stack): 6574 _collector(collector), 6575 _span(span), 6576 _verification_bm(verification_bm), 6577 _cms_bm(cms_bm), 6578 _mark_stack(mark_stack), 6579 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6580 mark_stack) 6581 { 6582 assert(_mark_stack->isEmpty(), "stack should be empty"); 6583 _finger = _verification_bm->startWord(); 6584 assert(_collector->_restart_addr == NULL, "Sanity check"); 6585 assert(_span.contains(_finger), "Out of bounds _finger?"); 6586 } 6587 6588 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6589 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6590 assert(_span.contains(addr), "Out of bounds _finger?"); 6591 _finger = addr; 6592 } 6593 6594 // Should revisit to see if this should be restructured for 6595 // greater efficiency. 6596 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6597 // convert offset into a HeapWord* 6598 HeapWord* addr = _verification_bm->startWord() + offset; 6599 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6600 "address out of range"); 6601 assert(_verification_bm->isMarked(addr), "tautology"); 6602 assert(_cms_bm->isMarked(addr), "tautology"); 6603 6604 assert(_mark_stack->isEmpty(), 6605 "should drain stack to limit stack usage"); 6606 // convert addr to an oop preparatory to scanning 6607 oop obj = oop(addr); 6608 assert(obj->is_oop(), "should be an oop"); 6609 assert(_finger <= addr, "_finger runneth ahead"); 6610 // advance the finger to right end of this object 6611 _finger = addr + obj->size(); 6612 assert(_finger > addr, "we just incremented it above"); 6613 // Note: the finger doesn't advance while we drain 6614 // the stack below. 6615 bool res = _mark_stack->push(obj); 6616 assert(res, "Empty non-zero size stack should have space for single push"); 6617 while (!_mark_stack->isEmpty()) { 6618 oop new_oop = _mark_stack->pop(); 6619 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6620 // now scan this oop's oops 6621 new_oop->oop_iterate(&_pam_verify_closure); 6622 } 6623 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6624 return true; 6625 } 6626 6627 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6628 CMSCollector* collector, MemRegion span, 6629 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6630 CMSMarkStack* mark_stack): 6631 MetadataAwareOopClosure(collector->ref_processor()), 6632 _collector(collector), 6633 _span(span), 6634 _verification_bm(verification_bm), 6635 _cms_bm(cms_bm), 6636 _mark_stack(mark_stack) 6637 { } 6638 6639 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6640 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6641 6642 // Upon stack overflow, we discard (part of) the stack, 6643 // remembering the least address amongst those discarded 6644 // in CMSCollector's _restart_address. 6645 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6646 // Remember the least grey address discarded 6647 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6648 _collector->lower_restart_addr(ra); 6649 _mark_stack->reset(); // discard stack contents 6650 _mark_stack->expand(); // expand the stack if possible 6651 } 6652 6653 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6654 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6655 HeapWord* addr = (HeapWord*)obj; 6656 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6657 // Oop lies in _span and isn't yet grey or black 6658 _verification_bm->mark(addr); // now grey 6659 if (!_cms_bm->isMarked(addr)) { 6660 LogHandle(gc, verify) log; 6661 ResourceMark rm; 6662 oop(addr)->print_on(log.error_stream()); 6663 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6664 fatal("... aborting"); 6665 } 6666 6667 if (!_mark_stack->push(obj)) { // stack overflow 6668 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6669 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6670 handle_stack_overflow(addr); 6671 } 6672 // anything including and to the right of _finger 6673 // will be scanned as we iterate over the remainder of the 6674 // bit map 6675 } 6676 } 6677 6678 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6679 MemRegion span, 6680 CMSBitMap* bitMap, CMSMarkStack* markStack, 6681 HeapWord* finger, MarkFromRootsClosure* parent) : 6682 MetadataAwareOopClosure(collector->ref_processor()), 6683 _collector(collector), 6684 _span(span), 6685 _bitMap(bitMap), 6686 _markStack(markStack), 6687 _finger(finger), 6688 _parent(parent) 6689 { } 6690 6691 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6692 MemRegion span, 6693 CMSBitMap* bit_map, 6694 OopTaskQueue* work_queue, 6695 CMSMarkStack* overflow_stack, 6696 HeapWord* finger, 6697 HeapWord** global_finger_addr, 6698 ParMarkFromRootsClosure* parent) : 6699 MetadataAwareOopClosure(collector->ref_processor()), 6700 _collector(collector), 6701 _whole_span(collector->_span), 6702 _span(span), 6703 _bit_map(bit_map), 6704 _work_queue(work_queue), 6705 _overflow_stack(overflow_stack), 6706 _finger(finger), 6707 _global_finger_addr(global_finger_addr), 6708 _parent(parent) 6709 { } 6710 6711 // Assumes thread-safe access by callers, who are 6712 // responsible for mutual exclusion. 6713 void CMSCollector::lower_restart_addr(HeapWord* low) { 6714 assert(_span.contains(low), "Out of bounds addr"); 6715 if (_restart_addr == NULL) { 6716 _restart_addr = low; 6717 } else { 6718 _restart_addr = MIN2(_restart_addr, low); 6719 } 6720 } 6721 6722 // Upon stack overflow, we discard (part of) the stack, 6723 // remembering the least address amongst those discarded 6724 // in CMSCollector's _restart_address. 6725 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6726 // Remember the least grey address discarded 6727 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6728 _collector->lower_restart_addr(ra); 6729 _markStack->reset(); // discard stack contents 6730 _markStack->expand(); // expand the stack if possible 6731 } 6732 6733 // Upon stack overflow, we discard (part of) the stack, 6734 // remembering the least address amongst those discarded 6735 // in CMSCollector's _restart_address. 6736 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6737 // We need to do this under a mutex to prevent other 6738 // workers from interfering with the work done below. 6739 MutexLockerEx ml(_overflow_stack->par_lock(), 6740 Mutex::_no_safepoint_check_flag); 6741 // Remember the least grey address discarded 6742 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6743 _collector->lower_restart_addr(ra); 6744 _overflow_stack->reset(); // discard stack contents 6745 _overflow_stack->expand(); // expand the stack if possible 6746 } 6747 6748 void PushOrMarkClosure::do_oop(oop obj) { 6749 // Ignore mark word because we are running concurrent with mutators. 6750 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6751 HeapWord* addr = (HeapWord*)obj; 6752 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6753 // Oop lies in _span and isn't yet grey or black 6754 _bitMap->mark(addr); // now grey 6755 if (addr < _finger) { 6756 // the bit map iteration has already either passed, or 6757 // sampled, this bit in the bit map; we'll need to 6758 // use the marking stack to scan this oop's oops. 6759 bool simulate_overflow = false; 6760 NOT_PRODUCT( 6761 if (CMSMarkStackOverflowALot && 6762 _collector->simulate_overflow()) { 6763 // simulate a stack overflow 6764 simulate_overflow = true; 6765 } 6766 ) 6767 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6768 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6769 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6770 handle_stack_overflow(addr); 6771 } 6772 } 6773 // anything including and to the right of _finger 6774 // will be scanned as we iterate over the remainder of the 6775 // bit map 6776 do_yield_check(); 6777 } 6778 } 6779 6780 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6781 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6782 6783 void ParPushOrMarkClosure::do_oop(oop obj) { 6784 // Ignore mark word because we are running concurrent with mutators. 6785 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6786 HeapWord* addr = (HeapWord*)obj; 6787 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6788 // Oop lies in _span and isn't yet grey or black 6789 // We read the global_finger (volatile read) strictly after marking oop 6790 bool res = _bit_map->par_mark(addr); // now grey 6791 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6792 // Should we push this marked oop on our stack? 6793 // -- if someone else marked it, nothing to do 6794 // -- if target oop is above global finger nothing to do 6795 // -- if target oop is in chunk and above local finger 6796 // then nothing to do 6797 // -- else push on work queue 6798 if ( !res // someone else marked it, they will deal with it 6799 || (addr >= *gfa) // will be scanned in a later task 6800 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6801 return; 6802 } 6803 // the bit map iteration has already either passed, or 6804 // sampled, this bit in the bit map; we'll need to 6805 // use the marking stack to scan this oop's oops. 6806 bool simulate_overflow = false; 6807 NOT_PRODUCT( 6808 if (CMSMarkStackOverflowALot && 6809 _collector->simulate_overflow()) { 6810 // simulate a stack overflow 6811 simulate_overflow = true; 6812 } 6813 ) 6814 if (simulate_overflow || 6815 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6816 // stack overflow 6817 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6818 // We cannot assert that the overflow stack is full because 6819 // it may have been emptied since. 6820 assert(simulate_overflow || 6821 _work_queue->size() == _work_queue->max_elems(), 6822 "Else push should have succeeded"); 6823 handle_stack_overflow(addr); 6824 } 6825 do_yield_check(); 6826 } 6827 } 6828 6829 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6830 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6831 6832 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6833 MemRegion span, 6834 ReferenceProcessor* rp, 6835 CMSBitMap* bit_map, 6836 CMSBitMap* mod_union_table, 6837 CMSMarkStack* mark_stack, 6838 bool concurrent_precleaning): 6839 MetadataAwareOopClosure(rp), 6840 _collector(collector), 6841 _span(span), 6842 _bit_map(bit_map), 6843 _mod_union_table(mod_union_table), 6844 _mark_stack(mark_stack), 6845 _concurrent_precleaning(concurrent_precleaning) 6846 { 6847 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6848 } 6849 6850 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6851 // the non-parallel version (the parallel version appears further below.) 6852 void PushAndMarkClosure::do_oop(oop obj) { 6853 // Ignore mark word verification. If during concurrent precleaning, 6854 // the object monitor may be locked. If during the checkpoint 6855 // phases, the object may already have been reached by a different 6856 // path and may be at the end of the global overflow list (so 6857 // the mark word may be NULL). 6858 assert(obj->is_oop_or_null(true /* ignore mark word */), 6859 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6860 HeapWord* addr = (HeapWord*)obj; 6861 // Check if oop points into the CMS generation 6862 // and is not marked 6863 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6864 // a white object ... 6865 _bit_map->mark(addr); // ... now grey 6866 // push on the marking stack (grey set) 6867 bool simulate_overflow = false; 6868 NOT_PRODUCT( 6869 if (CMSMarkStackOverflowALot && 6870 _collector->simulate_overflow()) { 6871 // simulate a stack overflow 6872 simulate_overflow = true; 6873 } 6874 ) 6875 if (simulate_overflow || !_mark_stack->push(obj)) { 6876 if (_concurrent_precleaning) { 6877 // During precleaning we can just dirty the appropriate card(s) 6878 // in the mod union table, thus ensuring that the object remains 6879 // in the grey set and continue. In the case of object arrays 6880 // we need to dirty all of the cards that the object spans, 6881 // since the rescan of object arrays will be limited to the 6882 // dirty cards. 6883 // Note that no one can be interfering with us in this action 6884 // of dirtying the mod union table, so no locking or atomics 6885 // are required. 6886 if (obj->is_objArray()) { 6887 size_t sz = obj->size(); 6888 HeapWord* end_card_addr = (HeapWord*)round_to( 6889 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6890 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6891 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6892 _mod_union_table->mark_range(redirty_range); 6893 } else { 6894 _mod_union_table->mark(addr); 6895 } 6896 _collector->_ser_pmc_preclean_ovflw++; 6897 } else { 6898 // During the remark phase, we need to remember this oop 6899 // in the overflow list. 6900 _collector->push_on_overflow_list(obj); 6901 _collector->_ser_pmc_remark_ovflw++; 6902 } 6903 } 6904 } 6905 } 6906 6907 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6908 MemRegion span, 6909 ReferenceProcessor* rp, 6910 CMSBitMap* bit_map, 6911 OopTaskQueue* work_queue): 6912 MetadataAwareOopClosure(rp), 6913 _collector(collector), 6914 _span(span), 6915 _bit_map(bit_map), 6916 _work_queue(work_queue) 6917 { 6918 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6919 } 6920 6921 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6922 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6923 6924 // Grey object rescan during second checkpoint phase -- 6925 // the parallel version. 6926 void ParPushAndMarkClosure::do_oop(oop obj) { 6927 // In the assert below, we ignore the mark word because 6928 // this oop may point to an already visited object that is 6929 // on the overflow stack (in which case the mark word has 6930 // been hijacked for chaining into the overflow stack -- 6931 // if this is the last object in the overflow stack then 6932 // its mark word will be NULL). Because this object may 6933 // have been subsequently popped off the global overflow 6934 // stack, and the mark word possibly restored to the prototypical 6935 // value, by the time we get to examined this failing assert in 6936 // the debugger, is_oop_or_null(false) may subsequently start 6937 // to hold. 6938 assert(obj->is_oop_or_null(true), 6939 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6940 HeapWord* addr = (HeapWord*)obj; 6941 // Check if oop points into the CMS generation 6942 // and is not marked 6943 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6944 // a white object ... 6945 // If we manage to "claim" the object, by being the 6946 // first thread to mark it, then we push it on our 6947 // marking stack 6948 if (_bit_map->par_mark(addr)) { // ... now grey 6949 // push on work queue (grey set) 6950 bool simulate_overflow = false; 6951 NOT_PRODUCT( 6952 if (CMSMarkStackOverflowALot && 6953 _collector->par_simulate_overflow()) { 6954 // simulate a stack overflow 6955 simulate_overflow = true; 6956 } 6957 ) 6958 if (simulate_overflow || !_work_queue->push(obj)) { 6959 _collector->par_push_on_overflow_list(obj); 6960 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6961 } 6962 } // Else, some other thread got there first 6963 } 6964 } 6965 6966 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6967 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6968 6969 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6970 Mutex* bml = _collector->bitMapLock(); 6971 assert_lock_strong(bml); 6972 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6973 "CMS thread should hold CMS token"); 6974 6975 bml->unlock(); 6976 ConcurrentMarkSweepThread::desynchronize(true); 6977 6978 _collector->stopTimer(); 6979 _collector->incrementYields(); 6980 6981 // See the comment in coordinator_yield() 6982 for (unsigned i = 0; i < CMSYieldSleepCount && 6983 ConcurrentMarkSweepThread::should_yield() && 6984 !CMSCollector::foregroundGCIsActive(); ++i) { 6985 os::sleep(Thread::current(), 1, false); 6986 } 6987 6988 ConcurrentMarkSweepThread::synchronize(true); 6989 bml->lock(); 6990 6991 _collector->startTimer(); 6992 } 6993 6994 bool CMSPrecleanRefsYieldClosure::should_return() { 6995 if (ConcurrentMarkSweepThread::should_yield()) { 6996 do_yield_work(); 6997 } 6998 return _collector->foregroundGCIsActive(); 6999 } 7000 7001 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7002 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7003 "mr should be aligned to start at a card boundary"); 7004 // We'd like to assert: 7005 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7006 // "mr should be a range of cards"); 7007 // However, that would be too strong in one case -- the last 7008 // partition ends at _unallocated_block which, in general, can be 7009 // an arbitrary boundary, not necessarily card aligned. 7010 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7011 _space->object_iterate_mem(mr, &_scan_cl); 7012 } 7013 7014 SweepClosure::SweepClosure(CMSCollector* collector, 7015 ConcurrentMarkSweepGeneration* g, 7016 CMSBitMap* bitMap, bool should_yield) : 7017 _collector(collector), 7018 _g(g), 7019 _sp(g->cmsSpace()), 7020 _limit(_sp->sweep_limit()), 7021 _freelistLock(_sp->freelistLock()), 7022 _bitMap(bitMap), 7023 _yield(should_yield), 7024 _inFreeRange(false), // No free range at beginning of sweep 7025 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7026 _lastFreeRangeCoalesced(false), 7027 _freeFinger(g->used_region().start()) 7028 { 7029 NOT_PRODUCT( 7030 _numObjectsFreed = 0; 7031 _numWordsFreed = 0; 7032 _numObjectsLive = 0; 7033 _numWordsLive = 0; 7034 _numObjectsAlreadyFree = 0; 7035 _numWordsAlreadyFree = 0; 7036 _last_fc = NULL; 7037 7038 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7039 _sp->dictionary()->initialize_dict_returned_bytes(); 7040 ) 7041 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7042 "sweep _limit out of bounds"); 7043 log_develop_trace(gc, sweep)("===================="); 7044 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7045 } 7046 7047 void SweepClosure::print_on(outputStream* st) const { 7048 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7049 p2i(_sp->bottom()), p2i(_sp->end())); 7050 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7051 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7052 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7053 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7054 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7055 } 7056 7057 #ifndef PRODUCT 7058 // Assertion checking only: no useful work in product mode -- 7059 // however, if any of the flags below become product flags, 7060 // you may need to review this code to see if it needs to be 7061 // enabled in product mode. 7062 SweepClosure::~SweepClosure() { 7063 assert_lock_strong(_freelistLock); 7064 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7065 "sweep _limit out of bounds"); 7066 if (inFreeRange()) { 7067 LogHandle(gc, sweep) log; 7068 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7069 ResourceMark rm; 7070 print_on(log.error_stream()); 7071 ShouldNotReachHere(); 7072 } 7073 7074 if (log_is_enabled(Debug, gc, sweep)) { 7075 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7076 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7077 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7078 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7079 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7080 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7081 } 7082 7083 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7084 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7085 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7086 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7087 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7088 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7089 } 7090 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7091 log_develop_trace(gc, sweep)("================"); 7092 } 7093 #endif // PRODUCT 7094 7095 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7096 bool freeRangeInFreeLists) { 7097 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7098 p2i(freeFinger), freeRangeInFreeLists); 7099 assert(!inFreeRange(), "Trampling existing free range"); 7100 set_inFreeRange(true); 7101 set_lastFreeRangeCoalesced(false); 7102 7103 set_freeFinger(freeFinger); 7104 set_freeRangeInFreeLists(freeRangeInFreeLists); 7105 if (CMSTestInFreeList) { 7106 if (freeRangeInFreeLists) { 7107 FreeChunk* fc = (FreeChunk*) freeFinger; 7108 assert(fc->is_free(), "A chunk on the free list should be free."); 7109 assert(fc->size() > 0, "Free range should have a size"); 7110 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7111 } 7112 } 7113 } 7114 7115 // Note that the sweeper runs concurrently with mutators. Thus, 7116 // it is possible for direct allocation in this generation to happen 7117 // in the middle of the sweep. Note that the sweeper also coalesces 7118 // contiguous free blocks. Thus, unless the sweeper and the allocator 7119 // synchronize appropriately freshly allocated blocks may get swept up. 7120 // This is accomplished by the sweeper locking the free lists while 7121 // it is sweeping. Thus blocks that are determined to be free are 7122 // indeed free. There is however one additional complication: 7123 // blocks that have been allocated since the final checkpoint and 7124 // mark, will not have been marked and so would be treated as 7125 // unreachable and swept up. To prevent this, the allocator marks 7126 // the bit map when allocating during the sweep phase. This leads, 7127 // however, to a further complication -- objects may have been allocated 7128 // but not yet initialized -- in the sense that the header isn't yet 7129 // installed. The sweeper can not then determine the size of the block 7130 // in order to skip over it. To deal with this case, we use a technique 7131 // (due to Printezis) to encode such uninitialized block sizes in the 7132 // bit map. Since the bit map uses a bit per every HeapWord, but the 7133 // CMS generation has a minimum object size of 3 HeapWords, it follows 7134 // that "normal marks" won't be adjacent in the bit map (there will 7135 // always be at least two 0 bits between successive 1 bits). We make use 7136 // of these "unused" bits to represent uninitialized blocks -- the bit 7137 // corresponding to the start of the uninitialized object and the next 7138 // bit are both set. Finally, a 1 bit marks the end of the object that 7139 // started with the two consecutive 1 bits to indicate its potentially 7140 // uninitialized state. 7141 7142 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7143 FreeChunk* fc = (FreeChunk*)addr; 7144 size_t res; 7145 7146 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7147 // than "addr == _limit" because although _limit was a block boundary when 7148 // we started the sweep, it may no longer be one because heap expansion 7149 // may have caused us to coalesce the block ending at the address _limit 7150 // with a newly expanded chunk (this happens when _limit was set to the 7151 // previous _end of the space), so we may have stepped past _limit: 7152 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7153 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7154 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7155 "sweep _limit out of bounds"); 7156 assert(addr < _sp->end(), "addr out of bounds"); 7157 // Flush any free range we might be holding as a single 7158 // coalesced chunk to the appropriate free list. 7159 if (inFreeRange()) { 7160 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7161 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7162 flush_cur_free_chunk(freeFinger(), 7163 pointer_delta(addr, freeFinger())); 7164 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7165 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7166 lastFreeRangeCoalesced() ? 1 : 0); 7167 } 7168 7169 // help the iterator loop finish 7170 return pointer_delta(_sp->end(), addr); 7171 } 7172 7173 assert(addr < _limit, "sweep invariant"); 7174 // check if we should yield 7175 do_yield_check(addr); 7176 if (fc->is_free()) { 7177 // Chunk that is already free 7178 res = fc->size(); 7179 do_already_free_chunk(fc); 7180 debug_only(_sp->verifyFreeLists()); 7181 // If we flush the chunk at hand in lookahead_and_flush() 7182 // and it's coalesced with a preceding chunk, then the 7183 // process of "mangling" the payload of the coalesced block 7184 // will cause erasure of the size information from the 7185 // (erstwhile) header of all the coalesced blocks but the 7186 // first, so the first disjunct in the assert will not hold 7187 // in that specific case (in which case the second disjunct 7188 // will hold). 7189 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7190 "Otherwise the size info doesn't change at this step"); 7191 NOT_PRODUCT( 7192 _numObjectsAlreadyFree++; 7193 _numWordsAlreadyFree += res; 7194 ) 7195 NOT_PRODUCT(_last_fc = fc;) 7196 } else if (!_bitMap->isMarked(addr)) { 7197 // Chunk is fresh garbage 7198 res = do_garbage_chunk(fc); 7199 debug_only(_sp->verifyFreeLists()); 7200 NOT_PRODUCT( 7201 _numObjectsFreed++; 7202 _numWordsFreed += res; 7203 ) 7204 } else { 7205 // Chunk that is alive. 7206 res = do_live_chunk(fc); 7207 debug_only(_sp->verifyFreeLists()); 7208 NOT_PRODUCT( 7209 _numObjectsLive++; 7210 _numWordsLive += res; 7211 ) 7212 } 7213 return res; 7214 } 7215 7216 // For the smart allocation, record following 7217 // split deaths - a free chunk is removed from its free list because 7218 // it is being split into two or more chunks. 7219 // split birth - a free chunk is being added to its free list because 7220 // a larger free chunk has been split and resulted in this free chunk. 7221 // coal death - a free chunk is being removed from its free list because 7222 // it is being coalesced into a large free chunk. 7223 // coal birth - a free chunk is being added to its free list because 7224 // it was created when two or more free chunks where coalesced into 7225 // this free chunk. 7226 // 7227 // These statistics are used to determine the desired number of free 7228 // chunks of a given size. The desired number is chosen to be relative 7229 // to the end of a CMS sweep. The desired number at the end of a sweep 7230 // is the 7231 // count-at-end-of-previous-sweep (an amount that was enough) 7232 // - count-at-beginning-of-current-sweep (the excess) 7233 // + split-births (gains in this size during interval) 7234 // - split-deaths (demands on this size during interval) 7235 // where the interval is from the end of one sweep to the end of the 7236 // next. 7237 // 7238 // When sweeping the sweeper maintains an accumulated chunk which is 7239 // the chunk that is made up of chunks that have been coalesced. That 7240 // will be termed the left-hand chunk. A new chunk of garbage that 7241 // is being considered for coalescing will be referred to as the 7242 // right-hand chunk. 7243 // 7244 // When making a decision on whether to coalesce a right-hand chunk with 7245 // the current left-hand chunk, the current count vs. the desired count 7246 // of the left-hand chunk is considered. Also if the right-hand chunk 7247 // is near the large chunk at the end of the heap (see 7248 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7249 // left-hand chunk is coalesced. 7250 // 7251 // When making a decision about whether to split a chunk, the desired count 7252 // vs. the current count of the candidate to be split is also considered. 7253 // If the candidate is underpopulated (currently fewer chunks than desired) 7254 // a chunk of an overpopulated (currently more chunks than desired) size may 7255 // be chosen. The "hint" associated with a free list, if non-null, points 7256 // to a free list which may be overpopulated. 7257 // 7258 7259 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7260 const size_t size = fc->size(); 7261 // Chunks that cannot be coalesced are not in the 7262 // free lists. 7263 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7264 assert(_sp->verify_chunk_in_free_list(fc), 7265 "free chunk should be in free lists"); 7266 } 7267 // a chunk that is already free, should not have been 7268 // marked in the bit map 7269 HeapWord* const addr = (HeapWord*) fc; 7270 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7271 // Verify that the bit map has no bits marked between 7272 // addr and purported end of this block. 7273 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7274 7275 // Some chunks cannot be coalesced under any circumstances. 7276 // See the definition of cantCoalesce(). 7277 if (!fc->cantCoalesce()) { 7278 // This chunk can potentially be coalesced. 7279 // All the work is done in 7280 do_post_free_or_garbage_chunk(fc, size); 7281 // Note that if the chunk is not coalescable (the else arm 7282 // below), we unconditionally flush, without needing to do 7283 // a "lookahead," as we do below. 7284 if (inFreeRange()) lookahead_and_flush(fc, size); 7285 } else { 7286 // Code path common to both original and adaptive free lists. 7287 7288 // cant coalesce with previous block; this should be treated 7289 // as the end of a free run if any 7290 if (inFreeRange()) { 7291 // we kicked some butt; time to pick up the garbage 7292 assert(freeFinger() < addr, "freeFinger points too high"); 7293 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7294 } 7295 // else, nothing to do, just continue 7296 } 7297 } 7298 7299 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7300 // This is a chunk of garbage. It is not in any free list. 7301 // Add it to a free list or let it possibly be coalesced into 7302 // a larger chunk. 7303 HeapWord* const addr = (HeapWord*) fc; 7304 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7305 7306 // Verify that the bit map has no bits marked between 7307 // addr and purported end of just dead object. 7308 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7309 do_post_free_or_garbage_chunk(fc, size); 7310 7311 assert(_limit >= addr + size, 7312 "A freshly garbage chunk can't possibly straddle over _limit"); 7313 if (inFreeRange()) lookahead_and_flush(fc, size); 7314 return size; 7315 } 7316 7317 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7318 HeapWord* addr = (HeapWord*) fc; 7319 // The sweeper has just found a live object. Return any accumulated 7320 // left hand chunk to the free lists. 7321 if (inFreeRange()) { 7322 assert(freeFinger() < addr, "freeFinger points too high"); 7323 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7324 } 7325 7326 // This object is live: we'd normally expect this to be 7327 // an oop, and like to assert the following: 7328 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7329 // However, as we commented above, this may be an object whose 7330 // header hasn't yet been initialized. 7331 size_t size; 7332 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7333 if (_bitMap->isMarked(addr + 1)) { 7334 // Determine the size from the bit map, rather than trying to 7335 // compute it from the object header. 7336 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7337 size = pointer_delta(nextOneAddr + 1, addr); 7338 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7339 "alignment problem"); 7340 7341 #ifdef ASSERT 7342 if (oop(addr)->klass_or_null() != NULL) { 7343 // Ignore mark word because we are running concurrent with mutators 7344 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7345 assert(size == 7346 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7347 "P-mark and computed size do not agree"); 7348 } 7349 #endif 7350 7351 } else { 7352 // This should be an initialized object that's alive. 7353 assert(oop(addr)->klass_or_null() != NULL, 7354 "Should be an initialized object"); 7355 // Ignore mark word because we are running concurrent with mutators 7356 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7357 // Verify that the bit map has no bits marked between 7358 // addr and purported end of this block. 7359 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7360 assert(size >= 3, "Necessary for Printezis marks to work"); 7361 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7362 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7363 } 7364 return size; 7365 } 7366 7367 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7368 size_t chunkSize) { 7369 // do_post_free_or_garbage_chunk() should only be called in the case 7370 // of the adaptive free list allocator. 7371 const bool fcInFreeLists = fc->is_free(); 7372 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7373 if (CMSTestInFreeList && fcInFreeLists) { 7374 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7375 } 7376 7377 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7378 7379 HeapWord* const fc_addr = (HeapWord*) fc; 7380 7381 bool coalesce = false; 7382 const size_t left = pointer_delta(fc_addr, freeFinger()); 7383 const size_t right = chunkSize; 7384 switch (FLSCoalescePolicy) { 7385 // numeric value forms a coalition aggressiveness metric 7386 case 0: { // never coalesce 7387 coalesce = false; 7388 break; 7389 } 7390 case 1: { // coalesce if left & right chunks on overpopulated lists 7391 coalesce = _sp->coalOverPopulated(left) && 7392 _sp->coalOverPopulated(right); 7393 break; 7394 } 7395 case 2: { // coalesce if left chunk on overpopulated list (default) 7396 coalesce = _sp->coalOverPopulated(left); 7397 break; 7398 } 7399 case 3: { // coalesce if left OR right chunk on overpopulated list 7400 coalesce = _sp->coalOverPopulated(left) || 7401 _sp->coalOverPopulated(right); 7402 break; 7403 } 7404 case 4: { // always coalesce 7405 coalesce = true; 7406 break; 7407 } 7408 default: 7409 ShouldNotReachHere(); 7410 } 7411 7412 // Should the current free range be coalesced? 7413 // If the chunk is in a free range and either we decided to coalesce above 7414 // or the chunk is near the large block at the end of the heap 7415 // (isNearLargestChunk() returns true), then coalesce this chunk. 7416 const bool doCoalesce = inFreeRange() 7417 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7418 if (doCoalesce) { 7419 // Coalesce the current free range on the left with the new 7420 // chunk on the right. If either is on a free list, 7421 // it must be removed from the list and stashed in the closure. 7422 if (freeRangeInFreeLists()) { 7423 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7424 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7425 "Size of free range is inconsistent with chunk size."); 7426 if (CMSTestInFreeList) { 7427 assert(_sp->verify_chunk_in_free_list(ffc), 7428 "Chunk is not in free lists"); 7429 } 7430 _sp->coalDeath(ffc->size()); 7431 _sp->removeFreeChunkFromFreeLists(ffc); 7432 set_freeRangeInFreeLists(false); 7433 } 7434 if (fcInFreeLists) { 7435 _sp->coalDeath(chunkSize); 7436 assert(fc->size() == chunkSize, 7437 "The chunk has the wrong size or is not in the free lists"); 7438 _sp->removeFreeChunkFromFreeLists(fc); 7439 } 7440 set_lastFreeRangeCoalesced(true); 7441 print_free_block_coalesced(fc); 7442 } else { // not in a free range and/or should not coalesce 7443 // Return the current free range and start a new one. 7444 if (inFreeRange()) { 7445 // In a free range but cannot coalesce with the right hand chunk. 7446 // Put the current free range into the free lists. 7447 flush_cur_free_chunk(freeFinger(), 7448 pointer_delta(fc_addr, freeFinger())); 7449 } 7450 // Set up for new free range. Pass along whether the right hand 7451 // chunk is in the free lists. 7452 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7453 } 7454 } 7455 7456 // Lookahead flush: 7457 // If we are tracking a free range, and this is the last chunk that 7458 // we'll look at because its end crosses past _limit, we'll preemptively 7459 // flush it along with any free range we may be holding on to. Note that 7460 // this can be the case only for an already free or freshly garbage 7461 // chunk. If this block is an object, it can never straddle 7462 // over _limit. The "straddling" occurs when _limit is set at 7463 // the previous end of the space when this cycle started, and 7464 // a subsequent heap expansion caused the previously co-terminal 7465 // free block to be coalesced with the newly expanded portion, 7466 // thus rendering _limit a non-block-boundary making it dangerous 7467 // for the sweeper to step over and examine. 7468 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7469 assert(inFreeRange(), "Should only be called if currently in a free range."); 7470 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7471 assert(_sp->used_region().contains(eob - 1), 7472 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7473 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7474 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7475 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7476 if (eob >= _limit) { 7477 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7478 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7479 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7480 "[" PTR_FORMAT "," PTR_FORMAT ")", 7481 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7482 // Return the storage we are tracking back into the free lists. 7483 log_develop_trace(gc, sweep)("Flushing ... "); 7484 assert(freeFinger() < eob, "Error"); 7485 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7486 } 7487 } 7488 7489 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7490 assert(inFreeRange(), "Should only be called if currently in a free range."); 7491 assert(size > 0, 7492 "A zero sized chunk cannot be added to the free lists."); 7493 if (!freeRangeInFreeLists()) { 7494 if (CMSTestInFreeList) { 7495 FreeChunk* fc = (FreeChunk*) chunk; 7496 fc->set_size(size); 7497 assert(!_sp->verify_chunk_in_free_list(fc), 7498 "chunk should not be in free lists yet"); 7499 } 7500 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7501 // A new free range is going to be starting. The current 7502 // free range has not been added to the free lists yet or 7503 // was removed so add it back. 7504 // If the current free range was coalesced, then the death 7505 // of the free range was recorded. Record a birth now. 7506 if (lastFreeRangeCoalesced()) { 7507 _sp->coalBirth(size); 7508 } 7509 _sp->addChunkAndRepairOffsetTable(chunk, size, 7510 lastFreeRangeCoalesced()); 7511 } else { 7512 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7513 } 7514 set_inFreeRange(false); 7515 set_freeRangeInFreeLists(false); 7516 } 7517 7518 // We take a break if we've been at this for a while, 7519 // so as to avoid monopolizing the locks involved. 7520 void SweepClosure::do_yield_work(HeapWord* addr) { 7521 // Return current free chunk being used for coalescing (if any) 7522 // to the appropriate freelist. After yielding, the next 7523 // free block encountered will start a coalescing range of 7524 // free blocks. If the next free block is adjacent to the 7525 // chunk just flushed, they will need to wait for the next 7526 // sweep to be coalesced. 7527 if (inFreeRange()) { 7528 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7529 } 7530 7531 // First give up the locks, then yield, then re-lock. 7532 // We should probably use a constructor/destructor idiom to 7533 // do this unlock/lock or modify the MutexUnlocker class to 7534 // serve our purpose. XXX 7535 assert_lock_strong(_bitMap->lock()); 7536 assert_lock_strong(_freelistLock); 7537 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7538 "CMS thread should hold CMS token"); 7539 _bitMap->lock()->unlock(); 7540 _freelistLock->unlock(); 7541 ConcurrentMarkSweepThread::desynchronize(true); 7542 _collector->stopTimer(); 7543 _collector->incrementYields(); 7544 7545 // See the comment in coordinator_yield() 7546 for (unsigned i = 0; i < CMSYieldSleepCount && 7547 ConcurrentMarkSweepThread::should_yield() && 7548 !CMSCollector::foregroundGCIsActive(); ++i) { 7549 os::sleep(Thread::current(), 1, false); 7550 } 7551 7552 ConcurrentMarkSweepThread::synchronize(true); 7553 _freelistLock->lock(); 7554 _bitMap->lock()->lock_without_safepoint_check(); 7555 _collector->startTimer(); 7556 } 7557 7558 #ifndef PRODUCT 7559 // This is actually very useful in a product build if it can 7560 // be called from the debugger. Compile it into the product 7561 // as needed. 7562 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7563 return debug_cms_space->verify_chunk_in_free_list(fc); 7564 } 7565 #endif 7566 7567 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7568 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7569 p2i(fc), fc->size()); 7570 } 7571 7572 // CMSIsAliveClosure 7573 bool CMSIsAliveClosure::do_object_b(oop obj) { 7574 HeapWord* addr = (HeapWord*)obj; 7575 return addr != NULL && 7576 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7577 } 7578 7579 7580 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7581 MemRegion span, 7582 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7583 bool cpc): 7584 _collector(collector), 7585 _span(span), 7586 _bit_map(bit_map), 7587 _mark_stack(mark_stack), 7588 _concurrent_precleaning(cpc) { 7589 assert(!_span.is_empty(), "Empty span could spell trouble"); 7590 } 7591 7592 7593 // CMSKeepAliveClosure: the serial version 7594 void CMSKeepAliveClosure::do_oop(oop obj) { 7595 HeapWord* addr = (HeapWord*)obj; 7596 if (_span.contains(addr) && 7597 !_bit_map->isMarked(addr)) { 7598 _bit_map->mark(addr); 7599 bool simulate_overflow = false; 7600 NOT_PRODUCT( 7601 if (CMSMarkStackOverflowALot && 7602 _collector->simulate_overflow()) { 7603 // simulate a stack overflow 7604 simulate_overflow = true; 7605 } 7606 ) 7607 if (simulate_overflow || !_mark_stack->push(obj)) { 7608 if (_concurrent_precleaning) { 7609 // We dirty the overflown object and let the remark 7610 // phase deal with it. 7611 assert(_collector->overflow_list_is_empty(), "Error"); 7612 // In the case of object arrays, we need to dirty all of 7613 // the cards that the object spans. No locking or atomics 7614 // are needed since no one else can be mutating the mod union 7615 // table. 7616 if (obj->is_objArray()) { 7617 size_t sz = obj->size(); 7618 HeapWord* end_card_addr = 7619 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7620 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7621 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7622 _collector->_modUnionTable.mark_range(redirty_range); 7623 } else { 7624 _collector->_modUnionTable.mark(addr); 7625 } 7626 _collector->_ser_kac_preclean_ovflw++; 7627 } else { 7628 _collector->push_on_overflow_list(obj); 7629 _collector->_ser_kac_ovflw++; 7630 } 7631 } 7632 } 7633 } 7634 7635 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7636 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7637 7638 // CMSParKeepAliveClosure: a parallel version of the above. 7639 // The work queues are private to each closure (thread), 7640 // but (may be) available for stealing by other threads. 7641 void CMSParKeepAliveClosure::do_oop(oop obj) { 7642 HeapWord* addr = (HeapWord*)obj; 7643 if (_span.contains(addr) && 7644 !_bit_map->isMarked(addr)) { 7645 // In general, during recursive tracing, several threads 7646 // may be concurrently getting here; the first one to 7647 // "tag" it, claims it. 7648 if (_bit_map->par_mark(addr)) { 7649 bool res = _work_queue->push(obj); 7650 assert(res, "Low water mark should be much less than capacity"); 7651 // Do a recursive trim in the hope that this will keep 7652 // stack usage lower, but leave some oops for potential stealers 7653 trim_queue(_low_water_mark); 7654 } // Else, another thread got there first 7655 } 7656 } 7657 7658 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7659 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7660 7661 void CMSParKeepAliveClosure::trim_queue(uint max) { 7662 while (_work_queue->size() > max) { 7663 oop new_oop; 7664 if (_work_queue->pop_local(new_oop)) { 7665 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7666 assert(_bit_map->isMarked((HeapWord*)new_oop), 7667 "no white objects on this stack!"); 7668 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7669 // iterate over the oops in this oop, marking and pushing 7670 // the ones in CMS heap (i.e. in _span). 7671 new_oop->oop_iterate(&_mark_and_push); 7672 } 7673 } 7674 } 7675 7676 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7677 CMSCollector* collector, 7678 MemRegion span, CMSBitMap* bit_map, 7679 OopTaskQueue* work_queue): 7680 _collector(collector), 7681 _span(span), 7682 _bit_map(bit_map), 7683 _work_queue(work_queue) { } 7684 7685 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7686 HeapWord* addr = (HeapWord*)obj; 7687 if (_span.contains(addr) && 7688 !_bit_map->isMarked(addr)) { 7689 if (_bit_map->par_mark(addr)) { 7690 bool simulate_overflow = false; 7691 NOT_PRODUCT( 7692 if (CMSMarkStackOverflowALot && 7693 _collector->par_simulate_overflow()) { 7694 // simulate a stack overflow 7695 simulate_overflow = true; 7696 } 7697 ) 7698 if (simulate_overflow || !_work_queue->push(obj)) { 7699 _collector->par_push_on_overflow_list(obj); 7700 _collector->_par_kac_ovflw++; 7701 } 7702 } // Else another thread got there already 7703 } 7704 } 7705 7706 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7707 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7708 7709 ////////////////////////////////////////////////////////////////// 7710 // CMSExpansionCause ///////////////////////////// 7711 ////////////////////////////////////////////////////////////////// 7712 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7713 switch (cause) { 7714 case _no_expansion: 7715 return "No expansion"; 7716 case _satisfy_free_ratio: 7717 return "Free ratio"; 7718 case _satisfy_promotion: 7719 return "Satisfy promotion"; 7720 case _satisfy_allocation: 7721 return "allocation"; 7722 case _allocate_par_lab: 7723 return "Par LAB"; 7724 case _allocate_par_spooling_space: 7725 return "Par Spooling Space"; 7726 case _adaptive_size_policy: 7727 return "Ergonomics"; 7728 default: 7729 return "unknown"; 7730 } 7731 } 7732 7733 void CMSDrainMarkingStackClosure::do_void() { 7734 // the max number to take from overflow list at a time 7735 const size_t num = _mark_stack->capacity()/4; 7736 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7737 "Overflow list should be NULL during concurrent phases"); 7738 while (!_mark_stack->isEmpty() || 7739 // if stack is empty, check the overflow list 7740 _collector->take_from_overflow_list(num, _mark_stack)) { 7741 oop obj = _mark_stack->pop(); 7742 HeapWord* addr = (HeapWord*)obj; 7743 assert(_span.contains(addr), "Should be within span"); 7744 assert(_bit_map->isMarked(addr), "Should be marked"); 7745 assert(obj->is_oop(), "Should be an oop"); 7746 obj->oop_iterate(_keep_alive); 7747 } 7748 } 7749 7750 void CMSParDrainMarkingStackClosure::do_void() { 7751 // drain queue 7752 trim_queue(0); 7753 } 7754 7755 // Trim our work_queue so its length is below max at return 7756 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7757 while (_work_queue->size() > max) { 7758 oop new_oop; 7759 if (_work_queue->pop_local(new_oop)) { 7760 assert(new_oop->is_oop(), "Expected an oop"); 7761 assert(_bit_map->isMarked((HeapWord*)new_oop), 7762 "no white objects on this stack!"); 7763 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7764 // iterate over the oops in this oop, marking and pushing 7765 // the ones in CMS heap (i.e. in _span). 7766 new_oop->oop_iterate(&_mark_and_push); 7767 } 7768 } 7769 } 7770 7771 //////////////////////////////////////////////////////////////////// 7772 // Support for Marking Stack Overflow list handling and related code 7773 //////////////////////////////////////////////////////////////////// 7774 // Much of the following code is similar in shape and spirit to the 7775 // code used in ParNewGC. We should try and share that code 7776 // as much as possible in the future. 7777 7778 #ifndef PRODUCT 7779 // Debugging support for CMSStackOverflowALot 7780 7781 // It's OK to call this multi-threaded; the worst thing 7782 // that can happen is that we'll get a bunch of closely 7783 // spaced simulated overflows, but that's OK, in fact 7784 // probably good as it would exercise the overflow code 7785 // under contention. 7786 bool CMSCollector::simulate_overflow() { 7787 if (_overflow_counter-- <= 0) { // just being defensive 7788 _overflow_counter = CMSMarkStackOverflowInterval; 7789 return true; 7790 } else { 7791 return false; 7792 } 7793 } 7794 7795 bool CMSCollector::par_simulate_overflow() { 7796 return simulate_overflow(); 7797 } 7798 #endif 7799 7800 // Single-threaded 7801 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7802 assert(stack->isEmpty(), "Expected precondition"); 7803 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7804 size_t i = num; 7805 oop cur = _overflow_list; 7806 const markOop proto = markOopDesc::prototype(); 7807 NOT_PRODUCT(ssize_t n = 0;) 7808 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7809 next = oop(cur->mark()); 7810 cur->set_mark(proto); // until proven otherwise 7811 assert(cur->is_oop(), "Should be an oop"); 7812 bool res = stack->push(cur); 7813 assert(res, "Bit off more than can chew?"); 7814 NOT_PRODUCT(n++;) 7815 } 7816 _overflow_list = cur; 7817 #ifndef PRODUCT 7818 assert(_num_par_pushes >= n, "Too many pops?"); 7819 _num_par_pushes -=n; 7820 #endif 7821 return !stack->isEmpty(); 7822 } 7823 7824 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7825 // (MT-safe) Get a prefix of at most "num" from the list. 7826 // The overflow list is chained through the mark word of 7827 // each object in the list. We fetch the entire list, 7828 // break off a prefix of the right size and return the 7829 // remainder. If other threads try to take objects from 7830 // the overflow list at that time, they will wait for 7831 // some time to see if data becomes available. If (and 7832 // only if) another thread places one or more object(s) 7833 // on the global list before we have returned the suffix 7834 // to the global list, we will walk down our local list 7835 // to find its end and append the global list to 7836 // our suffix before returning it. This suffix walk can 7837 // prove to be expensive (quadratic in the amount of traffic) 7838 // when there are many objects in the overflow list and 7839 // there is much producer-consumer contention on the list. 7840 // *NOTE*: The overflow list manipulation code here and 7841 // in ParNewGeneration:: are very similar in shape, 7842 // except that in the ParNew case we use the old (from/eden) 7843 // copy of the object to thread the list via its klass word. 7844 // Because of the common code, if you make any changes in 7845 // the code below, please check the ParNew version to see if 7846 // similar changes might be needed. 7847 // CR 6797058 has been filed to consolidate the common code. 7848 bool CMSCollector::par_take_from_overflow_list(size_t num, 7849 OopTaskQueue* work_q, 7850 int no_of_gc_threads) { 7851 assert(work_q->size() == 0, "First empty local work queue"); 7852 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7853 if (_overflow_list == NULL) { 7854 return false; 7855 } 7856 // Grab the entire list; we'll put back a suffix 7857 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7858 Thread* tid = Thread::current(); 7859 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7860 // set to ParallelGCThreads. 7861 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7862 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7863 // If the list is busy, we spin for a short while, 7864 // sleeping between attempts to get the list. 7865 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7866 os::sleep(tid, sleep_time_millis, false); 7867 if (_overflow_list == NULL) { 7868 // Nothing left to take 7869 return false; 7870 } else if (_overflow_list != BUSY) { 7871 // Try and grab the prefix 7872 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7873 } 7874 } 7875 // If the list was found to be empty, or we spun long 7876 // enough, we give up and return empty-handed. If we leave 7877 // the list in the BUSY state below, it must be the case that 7878 // some other thread holds the overflow list and will set it 7879 // to a non-BUSY state in the future. 7880 if (prefix == NULL || prefix == BUSY) { 7881 // Nothing to take or waited long enough 7882 if (prefix == NULL) { 7883 // Write back the NULL in case we overwrote it with BUSY above 7884 // and it is still the same value. 7885 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7886 } 7887 return false; 7888 } 7889 assert(prefix != NULL && prefix != BUSY, "Error"); 7890 size_t i = num; 7891 oop cur = prefix; 7892 // Walk down the first "num" objects, unless we reach the end. 7893 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7894 if (cur->mark() == NULL) { 7895 // We have "num" or fewer elements in the list, so there 7896 // is nothing to return to the global list. 7897 // Write back the NULL in lieu of the BUSY we wrote 7898 // above, if it is still the same value. 7899 if (_overflow_list == BUSY) { 7900 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7901 } 7902 } else { 7903 // Chop off the suffix and return it to the global list. 7904 assert(cur->mark() != BUSY, "Error"); 7905 oop suffix_head = cur->mark(); // suffix will be put back on global list 7906 cur->set_mark(NULL); // break off suffix 7907 // It's possible that the list is still in the empty(busy) state 7908 // we left it in a short while ago; in that case we may be 7909 // able to place back the suffix without incurring the cost 7910 // of a walk down the list. 7911 oop observed_overflow_list = _overflow_list; 7912 oop cur_overflow_list = observed_overflow_list; 7913 bool attached = false; 7914 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7915 observed_overflow_list = 7916 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7917 if (cur_overflow_list == observed_overflow_list) { 7918 attached = true; 7919 break; 7920 } else cur_overflow_list = observed_overflow_list; 7921 } 7922 if (!attached) { 7923 // Too bad, someone else sneaked in (at least) an element; we'll need 7924 // to do a splice. Find tail of suffix so we can prepend suffix to global 7925 // list. 7926 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7927 oop suffix_tail = cur; 7928 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7929 "Tautology"); 7930 observed_overflow_list = _overflow_list; 7931 do { 7932 cur_overflow_list = observed_overflow_list; 7933 if (cur_overflow_list != BUSY) { 7934 // Do the splice ... 7935 suffix_tail->set_mark(markOop(cur_overflow_list)); 7936 } else { // cur_overflow_list == BUSY 7937 suffix_tail->set_mark(NULL); 7938 } 7939 // ... and try to place spliced list back on overflow_list ... 7940 observed_overflow_list = 7941 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7942 } while (cur_overflow_list != observed_overflow_list); 7943 // ... until we have succeeded in doing so. 7944 } 7945 } 7946 7947 // Push the prefix elements on work_q 7948 assert(prefix != NULL, "control point invariant"); 7949 const markOop proto = markOopDesc::prototype(); 7950 oop next; 7951 NOT_PRODUCT(ssize_t n = 0;) 7952 for (cur = prefix; cur != NULL; cur = next) { 7953 next = oop(cur->mark()); 7954 cur->set_mark(proto); // until proven otherwise 7955 assert(cur->is_oop(), "Should be an oop"); 7956 bool res = work_q->push(cur); 7957 assert(res, "Bit off more than we can chew?"); 7958 NOT_PRODUCT(n++;) 7959 } 7960 #ifndef PRODUCT 7961 assert(_num_par_pushes >= n, "Too many pops?"); 7962 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7963 #endif 7964 return true; 7965 } 7966 7967 // Single-threaded 7968 void CMSCollector::push_on_overflow_list(oop p) { 7969 NOT_PRODUCT(_num_par_pushes++;) 7970 assert(p->is_oop(), "Not an oop"); 7971 preserve_mark_if_necessary(p); 7972 p->set_mark((markOop)_overflow_list); 7973 _overflow_list = p; 7974 } 7975 7976 // Multi-threaded; use CAS to prepend to overflow list 7977 void CMSCollector::par_push_on_overflow_list(oop p) { 7978 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7979 assert(p->is_oop(), "Not an oop"); 7980 par_preserve_mark_if_necessary(p); 7981 oop observed_overflow_list = _overflow_list; 7982 oop cur_overflow_list; 7983 do { 7984 cur_overflow_list = observed_overflow_list; 7985 if (cur_overflow_list != BUSY) { 7986 p->set_mark(markOop(cur_overflow_list)); 7987 } else { 7988 p->set_mark(NULL); 7989 } 7990 observed_overflow_list = 7991 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7992 } while (cur_overflow_list != observed_overflow_list); 7993 } 7994 #undef BUSY 7995 7996 // Single threaded 7997 // General Note on GrowableArray: pushes may silently fail 7998 // because we are (temporarily) out of C-heap for expanding 7999 // the stack. The problem is quite ubiquitous and affects 8000 // a lot of code in the JVM. The prudent thing for GrowableArray 8001 // to do (for now) is to exit with an error. However, that may 8002 // be too draconian in some cases because the caller may be 8003 // able to recover without much harm. For such cases, we 8004 // should probably introduce a "soft_push" method which returns 8005 // an indication of success or failure with the assumption that 8006 // the caller may be able to recover from a failure; code in 8007 // the VM can then be changed, incrementally, to deal with such 8008 // failures where possible, thus, incrementally hardening the VM 8009 // in such low resource situations. 8010 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8011 _preserved_oop_stack.push(p); 8012 _preserved_mark_stack.push(m); 8013 assert(m == p->mark(), "Mark word changed"); 8014 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8015 "bijection"); 8016 } 8017 8018 // Single threaded 8019 void CMSCollector::preserve_mark_if_necessary(oop p) { 8020 markOop m = p->mark(); 8021 if (m->must_be_preserved(p)) { 8022 preserve_mark_work(p, m); 8023 } 8024 } 8025 8026 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8027 markOop m = p->mark(); 8028 if (m->must_be_preserved(p)) { 8029 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8030 // Even though we read the mark word without holding 8031 // the lock, we are assured that it will not change 8032 // because we "own" this oop, so no other thread can 8033 // be trying to push it on the overflow list; see 8034 // the assertion in preserve_mark_work() that checks 8035 // that m == p->mark(). 8036 preserve_mark_work(p, m); 8037 } 8038 } 8039 8040 // We should be able to do this multi-threaded, 8041 // a chunk of stack being a task (this is 8042 // correct because each oop only ever appears 8043 // once in the overflow list. However, it's 8044 // not very easy to completely overlap this with 8045 // other operations, so will generally not be done 8046 // until all work's been completed. Because we 8047 // expect the preserved oop stack (set) to be small, 8048 // it's probably fine to do this single-threaded. 8049 // We can explore cleverer concurrent/overlapped/parallel 8050 // processing of preserved marks if we feel the 8051 // need for this in the future. Stack overflow should 8052 // be so rare in practice and, when it happens, its 8053 // effect on performance so great that this will 8054 // likely just be in the noise anyway. 8055 void CMSCollector::restore_preserved_marks_if_any() { 8056 assert(SafepointSynchronize::is_at_safepoint(), 8057 "world should be stopped"); 8058 assert(Thread::current()->is_ConcurrentGC_thread() || 8059 Thread::current()->is_VM_thread(), 8060 "should be single-threaded"); 8061 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8062 "bijection"); 8063 8064 while (!_preserved_oop_stack.is_empty()) { 8065 oop p = _preserved_oop_stack.pop(); 8066 assert(p->is_oop(), "Should be an oop"); 8067 assert(_span.contains(p), "oop should be in _span"); 8068 assert(p->mark() == markOopDesc::prototype(), 8069 "Set when taken from overflow list"); 8070 markOop m = _preserved_mark_stack.pop(); 8071 p->set_mark(m); 8072 } 8073 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8074 "stacks were cleared above"); 8075 } 8076 8077 #ifndef PRODUCT 8078 bool CMSCollector::no_preserved_marks() const { 8079 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8080 } 8081 #endif 8082 8083 // Transfer some number of overflown objects to usual marking 8084 // stack. Return true if some objects were transferred. 8085 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8086 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8087 (size_t)ParGCDesiredObjsFromOverflowList); 8088 8089 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8090 assert(_collector->overflow_list_is_empty() || res, 8091 "If list is not empty, we should have taken something"); 8092 assert(!res || !_mark_stack->isEmpty(), 8093 "If we took something, it should now be on our stack"); 8094 return res; 8095 } 8096 8097 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8098 size_t res = _sp->block_size_no_stall(addr, _collector); 8099 if (_sp->block_is_obj(addr)) { 8100 if (_live_bit_map->isMarked(addr)) { 8101 // It can't have been dead in a previous cycle 8102 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8103 } else { 8104 _dead_bit_map->mark(addr); // mark the dead object 8105 } 8106 } 8107 // Could be 0, if the block size could not be computed without stalling. 8108 return res; 8109 } 8110 8111 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8112 8113 switch (phase) { 8114 case CMSCollector::InitialMarking: 8115 initialize(true /* fullGC */ , 8116 cause /* cause of the GC */, 8117 true /* recordGCBeginTime */, 8118 true /* recordPreGCUsage */, 8119 false /* recordPeakUsage */, 8120 false /* recordPostGCusage */, 8121 true /* recordAccumulatedGCTime */, 8122 false /* recordGCEndTime */, 8123 false /* countCollection */ ); 8124 break; 8125 8126 case CMSCollector::FinalMarking: 8127 initialize(true /* fullGC */ , 8128 cause /* cause of the GC */, 8129 false /* recordGCBeginTime */, 8130 false /* recordPreGCUsage */, 8131 false /* recordPeakUsage */, 8132 false /* recordPostGCusage */, 8133 true /* recordAccumulatedGCTime */, 8134 false /* recordGCEndTime */, 8135 false /* countCollection */ ); 8136 break; 8137 8138 case CMSCollector::Sweeping: 8139 initialize(true /* fullGC */ , 8140 cause /* cause of the GC */, 8141 false /* recordGCBeginTime */, 8142 false /* recordPreGCUsage */, 8143 true /* recordPeakUsage */, 8144 true /* recordPostGCusage */, 8145 false /* recordAccumulatedGCTime */, 8146 true /* recordGCEndTime */, 8147 true /* countCollection */ ); 8148 break; 8149 8150 default: 8151 ShouldNotReachHere(); 8152 } 8153 }