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