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 warning("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 warning("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 warning("GC/CMS: _conc_workers allocation failure: " 531 "forcing -CMSConcurrentMTEnabled"); 532 CMSConcurrentMTEnabled = false; 533 } else { 534 _conc_workers->initialize_workers(); 535 } 536 } else { 537 CMSConcurrentMTEnabled = false; 538 } 539 } 540 if (!CMSConcurrentMTEnabled) { 541 ConcGCThreads = 0; 542 } else { 543 // Turn off CMSCleanOnEnter optimization temporarily for 544 // the MT case where it's not fixed yet; see 6178663. 545 CMSCleanOnEnter = false; 546 } 547 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 548 "Inconsistency"); 549 550 // Parallel task queues; these are shared for the 551 // concurrent and stop-world phases of CMS, but 552 // are not shared with parallel scavenge (ParNew). 553 { 554 uint i; 555 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 556 557 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 558 || ParallelRefProcEnabled) 559 && num_queues > 0) { 560 _task_queues = new OopTaskQueueSet(num_queues); 561 if (_task_queues == NULL) { 562 warning("task_queues allocation failure."); 563 return; 564 } 565 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 566 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 567 for (i = 0; i < num_queues; i++) { 568 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 569 if (q == NULL) { 570 warning("work_queue allocation failure."); 571 return; 572 } 573 _task_queues->register_queue(i, q); 574 } 575 for (i = 0; i < num_queues; i++) { 576 _task_queues->queue(i)->initialize(); 577 _hash_seed[i] = 17; // copied from ParNew 578 } 579 } 580 } 581 582 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 583 584 // Clip CMSBootstrapOccupancy between 0 and 100. 585 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 586 587 // Now tell CMS generations the identity of their collector 588 ConcurrentMarkSweepGeneration::set_collector(this); 589 590 // Create & start a CMS thread for this CMS collector 591 _cmsThread = ConcurrentMarkSweepThread::start(this); 592 assert(cmsThread() != NULL, "CMS Thread should have been created"); 593 assert(cmsThread()->collector() == this, 594 "CMS Thread should refer to this gen"); 595 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 596 597 // Support for parallelizing young gen rescan 598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 599 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 600 _young_gen = (ParNewGeneration*)gch->young_gen(); 601 if (gch->supports_inline_contig_alloc()) { 602 _top_addr = gch->top_addr(); 603 _end_addr = gch->end_addr(); 604 assert(_young_gen != NULL, "no _young_gen"); 605 _eden_chunk_index = 0; 606 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 607 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 608 } 609 610 // Support for parallelizing survivor space rescan 611 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 612 const size_t max_plab_samples = 613 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 614 615 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 616 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 617 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 618 _survivor_chunk_capacity = max_plab_samples; 619 for (uint i = 0; i < ParallelGCThreads; i++) { 620 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 621 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 622 assert(cur->end() == 0, "Should be 0"); 623 assert(cur->array() == vec, "Should be vec"); 624 assert(cur->capacity() == max_plab_samples, "Error"); 625 } 626 } 627 628 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 629 _gc_counters = new CollectorCounters("CMS stop-the-world full collections", 1); 630 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 631 _completed_initialization = true; 632 _inter_sweep_timer.start(); // start of time 633 } 634 635 const char* ConcurrentMarkSweepGeneration::name() const { 636 return "concurrent mark-sweep generation"; 637 } 638 void ConcurrentMarkSweepGeneration::update_counters() { 639 if (UsePerfData) { 640 _space_counters->update_all(); 641 _gen_counters->update_all(); 642 } 643 } 644 645 // this is an optimized version of update_counters(). it takes the 646 // used value as a parameter rather than computing it. 647 // 648 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 649 if (UsePerfData) { 650 _space_counters->update_used(used); 651 _space_counters->update_capacity(); 652 _gen_counters->update_all(); 653 } 654 } 655 656 void ConcurrentMarkSweepGeneration::print() const { 657 Generation::print(); 658 cmsSpace()->print(); 659 } 660 661 #ifndef PRODUCT 662 void ConcurrentMarkSweepGeneration::print_statistics() { 663 cmsSpace()->printFLCensus(0); 664 } 665 #endif 666 667 size_t 668 ConcurrentMarkSweepGeneration::contiguous_available() const { 669 // dld proposes an improvement in precision here. If the committed 670 // part of the space ends in a free block we should add that to 671 // uncommitted size in the calculation below. Will make this 672 // change later, staying with the approximation below for the 673 // time being. -- ysr. 674 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 675 } 676 677 size_t 678 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 679 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 680 } 681 682 size_t ConcurrentMarkSweepGeneration::max_available() const { 683 return free() + _virtual_space.uncommitted_size(); 684 } 685 686 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 687 size_t available = max_available(); 688 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 689 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 690 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 691 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 692 return res; 693 } 694 695 // At a promotion failure dump information on block layout in heap 696 // (cms old generation). 697 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 698 LogHandle(gc, promotion) log; 699 if (log.is_trace()) { 700 ResourceMark rm; 701 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 702 } 703 } 704 705 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 706 // Clear the promotion information. These pointers can be adjusted 707 // along with all the other pointers into the heap but 708 // compaction is expected to be a rare event with 709 // a heap using cms so don't do it without seeing the need. 710 for (uint i = 0; i < ParallelGCThreads; i++) { 711 _par_gc_thread_states[i]->promo.reset(); 712 } 713 } 714 715 void ConcurrentMarkSweepGeneration::compute_new_size() { 716 assert_locked_or_safepoint(Heap_lock); 717 718 // If incremental collection failed, we just want to expand 719 // to the limit. 720 if (incremental_collection_failed()) { 721 clear_incremental_collection_failed(); 722 grow_to_reserved(); 723 return; 724 } 725 726 // The heap has been compacted but not reset yet. 727 // Any metric such as free() or used() will be incorrect. 728 729 CardGeneration::compute_new_size(); 730 731 // Reset again after a possible resizing 732 if (did_compact()) { 733 cmsSpace()->reset_after_compaction(); 734 } 735 } 736 737 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 738 assert_locked_or_safepoint(Heap_lock); 739 740 // If incremental collection failed, we just want to expand 741 // to the limit. 742 if (incremental_collection_failed()) { 743 clear_incremental_collection_failed(); 744 grow_to_reserved(); 745 return; 746 } 747 748 double free_percentage = ((double) free()) / capacity(); 749 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 750 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 751 752 // compute expansion delta needed for reaching desired free percentage 753 if (free_percentage < desired_free_percentage) { 754 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 755 assert(desired_capacity >= capacity(), "invalid expansion size"); 756 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 757 LogHandle(gc) log; 758 if (log.is_trace()) { 759 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 760 log.trace("From compute_new_size: "); 761 log.trace(" Free fraction %f", free_percentage); 762 log.trace(" Desired free fraction %f", desired_free_percentage); 763 log.trace(" Maximum free fraction %f", maximum_free_percentage); 764 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 765 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 766 GenCollectedHeap* gch = GenCollectedHeap::heap(); 767 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 768 size_t young_size = gch->young_gen()->capacity(); 769 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 770 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 771 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 772 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 773 } 774 // safe if expansion fails 775 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 776 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 777 } else { 778 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 779 assert(desired_capacity <= capacity(), "invalid expansion size"); 780 size_t shrink_bytes = capacity() - desired_capacity; 781 // Don't shrink unless the delta is greater than the minimum shrink we want 782 if (shrink_bytes >= MinHeapDeltaBytes) { 783 shrink_free_list_by(shrink_bytes); 784 } 785 } 786 } 787 788 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 789 return cmsSpace()->freelistLock(); 790 } 791 792 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 793 CMSSynchronousYieldRequest yr; 794 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 795 return have_lock_and_allocate(size, tlab); 796 } 797 798 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 799 bool tlab /* ignored */) { 800 assert_lock_strong(freelistLock()); 801 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 802 HeapWord* res = cmsSpace()->allocate(adjustedSize); 803 // Allocate the object live (grey) if the background collector has 804 // started marking. This is necessary because the marker may 805 // have passed this address and consequently this object will 806 // not otherwise be greyed and would be incorrectly swept up. 807 // Note that if this object contains references, the writing 808 // of those references will dirty the card containing this object 809 // allowing the object to be blackened (and its references scanned) 810 // either during a preclean phase or at the final checkpoint. 811 if (res != NULL) { 812 // We may block here with an uninitialized object with 813 // its mark-bit or P-bits not yet set. Such objects need 814 // to be safely navigable by block_start(). 815 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 816 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 817 collector()->direct_allocated(res, adjustedSize); 818 _direct_allocated_words += adjustedSize; 819 // allocation counters 820 NOT_PRODUCT( 821 _numObjectsAllocated++; 822 _numWordsAllocated += (int)adjustedSize; 823 ) 824 } 825 return res; 826 } 827 828 // In the case of direct allocation by mutators in a generation that 829 // is being concurrently collected, the object must be allocated 830 // live (grey) if the background collector has started marking. 831 // This is necessary because the marker may 832 // have passed this address and consequently this object will 833 // not otherwise be greyed and would be incorrectly swept up. 834 // Note that if this object contains references, the writing 835 // of those references will dirty the card containing this object 836 // allowing the object to be blackened (and its references scanned) 837 // either during a preclean phase or at the final checkpoint. 838 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 839 assert(_markBitMap.covers(start, size), "Out of bounds"); 840 if (_collectorState >= Marking) { 841 MutexLockerEx y(_markBitMap.lock(), 842 Mutex::_no_safepoint_check_flag); 843 // [see comments preceding SweepClosure::do_blk() below for details] 844 // 845 // Can the P-bits be deleted now? JJJ 846 // 847 // 1. need to mark the object as live so it isn't collected 848 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 849 // 3. need to mark the end of the object so marking, precleaning or sweeping 850 // can skip over uninitialized or unparsable objects. An allocated 851 // object is considered uninitialized for our purposes as long as 852 // its klass word is NULL. All old gen objects are parsable 853 // as soon as they are initialized.) 854 _markBitMap.mark(start); // object is live 855 _markBitMap.mark(start + 1); // object is potentially uninitialized? 856 _markBitMap.mark(start + size - 1); 857 // mark end of object 858 } 859 // check that oop looks uninitialized 860 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 861 } 862 863 void CMSCollector::promoted(bool par, HeapWord* start, 864 bool is_obj_array, size_t obj_size) { 865 assert(_markBitMap.covers(start), "Out of bounds"); 866 // See comment in direct_allocated() about when objects should 867 // be allocated live. 868 if (_collectorState >= Marking) { 869 // we already hold the marking bit map lock, taken in 870 // the prologue 871 if (par) { 872 _markBitMap.par_mark(start); 873 } else { 874 _markBitMap.mark(start); 875 } 876 // We don't need to mark the object as uninitialized (as 877 // in direct_allocated above) because this is being done with the 878 // world stopped and the object will be initialized by the 879 // time the marking, precleaning or sweeping get to look at it. 880 // But see the code for copying objects into the CMS generation, 881 // where we need to ensure that concurrent readers of the 882 // block offset table are able to safely navigate a block that 883 // is in flux from being free to being allocated (and in 884 // transition while being copied into) and subsequently 885 // becoming a bona-fide object when the copy/promotion is complete. 886 assert(SafepointSynchronize::is_at_safepoint(), 887 "expect promotion only at safepoints"); 888 889 if (_collectorState < Sweeping) { 890 // Mark the appropriate cards in the modUnionTable, so that 891 // this object gets scanned before the sweep. If this is 892 // not done, CMS generation references in the object might 893 // not get marked. 894 // For the case of arrays, which are otherwise precisely 895 // marked, we need to dirty the entire array, not just its head. 896 if (is_obj_array) { 897 // The [par_]mark_range() method expects mr.end() below to 898 // be aligned to the granularity of a bit's representation 899 // in the heap. In the case of the MUT below, that's a 900 // card size. 901 MemRegion mr(start, 902 (HeapWord*)round_to((intptr_t)(start + obj_size), 903 CardTableModRefBS::card_size /* bytes */)); 904 if (par) { 905 _modUnionTable.par_mark_range(mr); 906 } else { 907 _modUnionTable.mark_range(mr); 908 } 909 } else { // not an obj array; we can just mark the head 910 if (par) { 911 _modUnionTable.par_mark(start); 912 } else { 913 _modUnionTable.mark(start); 914 } 915 } 916 } 917 } 918 } 919 920 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 921 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 922 // allocate, copy and if necessary update promoinfo -- 923 // delegate to underlying space. 924 assert_lock_strong(freelistLock()); 925 926 #ifndef PRODUCT 927 if (GenCollectedHeap::heap()->promotion_should_fail()) { 928 return NULL; 929 } 930 #endif // #ifndef PRODUCT 931 932 oop res = _cmsSpace->promote(obj, obj_size); 933 if (res == NULL) { 934 // expand and retry 935 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 936 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 937 // Since this is the old generation, we don't try to promote 938 // into a more senior generation. 939 res = _cmsSpace->promote(obj, obj_size); 940 } 941 if (res != NULL) { 942 // See comment in allocate() about when objects should 943 // be allocated live. 944 assert(obj->is_oop(), "Will dereference klass pointer below"); 945 collector()->promoted(false, // Not parallel 946 (HeapWord*)res, obj->is_objArray(), obj_size); 947 // promotion counters 948 NOT_PRODUCT( 949 _numObjectsPromoted++; 950 _numWordsPromoted += 951 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 952 ) 953 } 954 return res; 955 } 956 957 958 // IMPORTANT: Notes on object size recognition in CMS. 959 // --------------------------------------------------- 960 // A block of storage in the CMS generation is always in 961 // one of three states. A free block (FREE), an allocated 962 // object (OBJECT) whose size() method reports the correct size, 963 // and an intermediate state (TRANSIENT) in which its size cannot 964 // be accurately determined. 965 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 966 // ----------------------------------------------------- 967 // FREE: klass_word & 1 == 1; mark_word holds block size 968 // 969 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 970 // obj->size() computes correct size 971 // 972 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 973 // 974 // STATE IDENTIFICATION: (64 bit+COOPS) 975 // ------------------------------------ 976 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 977 // 978 // OBJECT: klass_word installed; klass_word != 0; 979 // obj->size() computes correct size 980 // 981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 982 // 983 // 984 // STATE TRANSITION DIAGRAM 985 // 986 // mut / parnew mut / parnew 987 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 988 // ^ | 989 // |------------------------ DEAD <------------------------------------| 990 // sweep mut 991 // 992 // While a block is in TRANSIENT state its size cannot be determined 993 // so readers will either need to come back later or stall until 994 // the size can be determined. Note that for the case of direct 995 // allocation, P-bits, when available, may be used to determine the 996 // size of an object that may not yet have been initialized. 997 998 // Things to support parallel young-gen collection. 999 oop 1000 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1001 oop old, markOop m, 1002 size_t word_sz) { 1003 #ifndef PRODUCT 1004 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1005 return NULL; 1006 } 1007 #endif // #ifndef PRODUCT 1008 1009 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1010 PromotionInfo* promoInfo = &ps->promo; 1011 // if we are tracking promotions, then first ensure space for 1012 // promotion (including spooling space for saving header if necessary). 1013 // then allocate and copy, then track promoted info if needed. 1014 // When tracking (see PromotionInfo::track()), the mark word may 1015 // be displaced and in this case restoration of the mark word 1016 // occurs in the (oop_since_save_marks_)iterate phase. 1017 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1018 // Out of space for allocating spooling buffers; 1019 // try expanding and allocating spooling buffers. 1020 if (!expand_and_ensure_spooling_space(promoInfo)) { 1021 return NULL; 1022 } 1023 } 1024 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1025 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1026 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1027 if (obj_ptr == NULL) { 1028 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1029 if (obj_ptr == NULL) { 1030 return NULL; 1031 } 1032 } 1033 oop obj = oop(obj_ptr); 1034 OrderAccess::storestore(); 1035 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1036 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1037 // IMPORTANT: See note on object initialization for CMS above. 1038 // Otherwise, copy the object. Here we must be careful to insert the 1039 // klass pointer last, since this marks the block as an allocated object. 1040 // Except with compressed oops it's the mark word. 1041 HeapWord* old_ptr = (HeapWord*)old; 1042 // Restore the mark word copied above. 1043 obj->set_mark(m); 1044 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1045 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1046 OrderAccess::storestore(); 1047 1048 if (UseCompressedClassPointers) { 1049 // Copy gap missed by (aligned) header size calculation below 1050 obj->set_klass_gap(old->klass_gap()); 1051 } 1052 if (word_sz > (size_t)oopDesc::header_size()) { 1053 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1054 obj_ptr + oopDesc::header_size(), 1055 word_sz - oopDesc::header_size()); 1056 } 1057 1058 // Now we can track the promoted object, if necessary. We take care 1059 // to delay the transition from uninitialized to full object 1060 // (i.e., insertion of klass pointer) until after, so that it 1061 // atomically becomes a promoted object. 1062 if (promoInfo->tracking()) { 1063 promoInfo->track((PromotedObject*)obj, old->klass()); 1064 } 1065 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1066 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1067 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1068 1069 // Finally, install the klass pointer (this should be volatile). 1070 OrderAccess::storestore(); 1071 obj->set_klass(old->klass()); 1072 // We should now be able to calculate the right size for this object 1073 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1074 1075 collector()->promoted(true, // parallel 1076 obj_ptr, old->is_objArray(), word_sz); 1077 1078 NOT_PRODUCT( 1079 Atomic::inc_ptr(&_numObjectsPromoted); 1080 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1081 ) 1082 1083 return obj; 1084 } 1085 1086 void 1087 ConcurrentMarkSweepGeneration:: 1088 par_promote_alloc_done(int thread_num) { 1089 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1090 ps->lab.retire(thread_num); 1091 } 1092 1093 void 1094 ConcurrentMarkSweepGeneration:: 1095 par_oop_since_save_marks_iterate_done(int thread_num) { 1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1097 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1098 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1099 } 1100 1101 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1102 size_t size, 1103 bool tlab) 1104 { 1105 // We allow a STW collection only if a full 1106 // collection was requested. 1107 return full || should_allocate(size, tlab); // FIX ME !!! 1108 // This and promotion failure handling are connected at the 1109 // hip and should be fixed by untying them. 1110 } 1111 1112 bool CMSCollector::shouldConcurrentCollect() { 1113 if (_full_gc_requested) { 1114 log_trace(gc)("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1115 return true; 1116 } 1117 1118 FreelistLocker x(this); 1119 // ------------------------------------------------------------------ 1120 // Print out lots of information which affects the initiation of 1121 // a collection. 1122 LogHandle(gc) log; 1123 if (log.is_trace() && stats().valid()) { 1124 log.trace("CMSCollector shouldConcurrentCollect: "); 1125 ResourceMark rm; 1126 stats().print_on(log.debug_stream()); 1127 log.trace("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1128 log.trace("free=" SIZE_FORMAT, _cmsGen->free()); 1129 log.trace("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1130 log.trace("promotion_rate=%g", stats().promotion_rate()); 1131 log.trace("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1132 log.trace("occupancy=%3.7f", _cmsGen->occupancy()); 1133 log.trace("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1134 log.trace("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1135 log.trace("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1136 log.trace("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1137 } 1138 // ------------------------------------------------------------------ 1139 1140 // If the estimated time to complete a cms collection (cms_duration()) 1141 // is less than the estimated time remaining until the cms generation 1142 // is full, start a collection. 1143 if (!UseCMSInitiatingOccupancyOnly) { 1144 if (stats().valid()) { 1145 if (stats().time_until_cms_start() == 0.0) { 1146 return true; 1147 } 1148 } else { 1149 // We want to conservatively collect somewhat early in order 1150 // to try and "bootstrap" our CMS/promotion statistics; 1151 // this branch will not fire after the first successful CMS 1152 // collection because the stats should then be valid. 1153 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1154 log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1155 _cmsGen->occupancy(), _bootstrap_occupancy); 1156 return true; 1157 } 1158 } 1159 } 1160 1161 // Otherwise, we start a collection cycle if 1162 // old gen want a collection cycle started. Each may use 1163 // an appropriate criterion for making this decision. 1164 // XXX We need to make sure that the gen expansion 1165 // criterion dovetails well with this. XXX NEED TO FIX THIS 1166 if (_cmsGen->should_concurrent_collect()) { 1167 log_trace(gc)("CMS old gen initiated"); 1168 return true; 1169 } 1170 1171 // We start a collection if we believe an incremental collection may fail; 1172 // this is not likely to be productive in practice because it's probably too 1173 // late anyway. 1174 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1175 assert(gch->collector_policy()->is_generation_policy(), 1176 "You may want to check the correctness of the following"); 1177 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1178 log_trace(gc)("CMSCollector: collect because incremental collection will fail "); 1179 return true; 1180 } 1181 1182 if (MetaspaceGC::should_concurrent_collect()) { 1183 log_trace(gc)("CMSCollector: collect for metadata allocation "); 1184 return true; 1185 } 1186 1187 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1188 if (CMSTriggerInterval >= 0) { 1189 if (CMSTriggerInterval == 0) { 1190 // Trigger always 1191 return true; 1192 } 1193 1194 // Check the CMS time since begin (we do not check the stats validity 1195 // as we want to be able to trigger the first CMS cycle as well) 1196 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1197 if (stats().valid()) { 1198 log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1199 stats().cms_time_since_begin()); 1200 } else { 1201 log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)"); 1202 } 1203 return true; 1204 } 1205 } 1206 1207 return false; 1208 } 1209 1210 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1211 1212 // Clear _expansion_cause fields of constituent generations 1213 void CMSCollector::clear_expansion_cause() { 1214 _cmsGen->clear_expansion_cause(); 1215 } 1216 1217 // We should be conservative in starting a collection cycle. To 1218 // start too eagerly runs the risk of collecting too often in the 1219 // extreme. To collect too rarely falls back on full collections, 1220 // which works, even if not optimum in terms of concurrent work. 1221 // As a work around for too eagerly collecting, use the flag 1222 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1223 // giving the user an easily understandable way of controlling the 1224 // collections. 1225 // We want to start a new collection cycle if any of the following 1226 // conditions hold: 1227 // . our current occupancy exceeds the configured initiating occupancy 1228 // for this generation, or 1229 // . we recently needed to expand this space and have not, since that 1230 // expansion, done a collection of this generation, or 1231 // . the underlying space believes that it may be a good idea to initiate 1232 // a concurrent collection (this may be based on criteria such as the 1233 // following: the space uses linear allocation and linear allocation is 1234 // going to fail, or there is believed to be excessive fragmentation in 1235 // the generation, etc... or ... 1236 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1237 // the case of the old generation; see CR 6543076): 1238 // we may be approaching a point at which allocation requests may fail because 1239 // we will be out of sufficient free space given allocation rate estimates.] 1240 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1241 1242 assert_lock_strong(freelistLock()); 1243 if (occupancy() > initiating_occupancy()) { 1244 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1245 short_name(), occupancy(), initiating_occupancy()); 1246 return true; 1247 } 1248 if (UseCMSInitiatingOccupancyOnly) { 1249 return false; 1250 } 1251 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1252 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1253 return true; 1254 } 1255 return false; 1256 } 1257 1258 void ConcurrentMarkSweepGeneration::collect(bool full, 1259 bool clear_all_soft_refs, 1260 size_t size, 1261 bool tlab) 1262 { 1263 collector()->collect(full, clear_all_soft_refs, size, tlab); 1264 } 1265 1266 void CMSCollector::collect(bool full, 1267 bool clear_all_soft_refs, 1268 size_t size, 1269 bool tlab) 1270 { 1271 // The following "if" branch is present for defensive reasons. 1272 // In the current uses of this interface, it can be replaced with: 1273 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1274 // But I am not placing that assert here to allow future 1275 // generality in invoking this interface. 1276 if (GCLocker::is_active()) { 1277 // A consistency test for GCLocker 1278 assert(GCLocker::needs_gc(), "Should have been set already"); 1279 // Skip this foreground collection, instead 1280 // expanding the heap if necessary. 1281 // Need the free list locks for the call to free() in compute_new_size() 1282 compute_new_size(); 1283 return; 1284 } 1285 acquire_control_and_collect(full, clear_all_soft_refs); 1286 } 1287 1288 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1289 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1290 unsigned int gc_count = gch->total_full_collections(); 1291 if (gc_count == full_gc_count) { 1292 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1293 _full_gc_requested = true; 1294 _full_gc_cause = cause; 1295 CGC_lock->notify(); // nudge CMS thread 1296 } else { 1297 assert(gc_count > full_gc_count, "Error: causal loop"); 1298 } 1299 } 1300 1301 bool CMSCollector::is_external_interruption() { 1302 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1303 return GCCause::is_user_requested_gc(cause) || 1304 GCCause::is_serviceability_requested_gc(cause); 1305 } 1306 1307 void CMSCollector::report_concurrent_mode_interruption() { 1308 if (is_external_interruption()) { 1309 log_debug(gc)("Concurrent mode interrupted"); 1310 } else { 1311 log_debug(gc)("Concurrent mode failure"); 1312 _gc_tracer_cm->report_concurrent_mode_failure(); 1313 } 1314 } 1315 1316 1317 // The foreground and background collectors need to coordinate in order 1318 // to make sure that they do not mutually interfere with CMS collections. 1319 // When a background collection is active, 1320 // the foreground collector may need to take over (preempt) and 1321 // synchronously complete an ongoing collection. Depending on the 1322 // frequency of the background collections and the heap usage 1323 // of the application, this preemption can be seldom or frequent. 1324 // There are only certain 1325 // points in the background collection that the "collection-baton" 1326 // can be passed to the foreground collector. 1327 // 1328 // The foreground collector will wait for the baton before 1329 // starting any part of the collection. The foreground collector 1330 // will only wait at one location. 1331 // 1332 // The background collector will yield the baton before starting a new 1333 // phase of the collection (e.g., before initial marking, marking from roots, 1334 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1335 // of the loop which switches the phases. The background collector does some 1336 // of the phases (initial mark, final re-mark) with the world stopped. 1337 // Because of locking involved in stopping the world, 1338 // the foreground collector should not block waiting for the background 1339 // collector when it is doing a stop-the-world phase. The background 1340 // collector will yield the baton at an additional point just before 1341 // it enters a stop-the-world phase. Once the world is stopped, the 1342 // background collector checks the phase of the collection. If the 1343 // phase has not changed, it proceeds with the collection. If the 1344 // phase has changed, it skips that phase of the collection. See 1345 // the comments on the use of the Heap_lock in collect_in_background(). 1346 // 1347 // Variable used in baton passing. 1348 // _foregroundGCIsActive - Set to true by the foreground collector when 1349 // it wants the baton. The foreground clears it when it has finished 1350 // the collection. 1351 // _foregroundGCShouldWait - Set to true by the background collector 1352 // when it is running. The foreground collector waits while 1353 // _foregroundGCShouldWait is true. 1354 // CGC_lock - monitor used to protect access to the above variables 1355 // and to notify the foreground and background collectors. 1356 // _collectorState - current state of the CMS collection. 1357 // 1358 // The foreground collector 1359 // acquires the CGC_lock 1360 // sets _foregroundGCIsActive 1361 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1362 // various locks acquired in preparation for the collection 1363 // are released so as not to block the background collector 1364 // that is in the midst of a collection 1365 // proceeds with the collection 1366 // clears _foregroundGCIsActive 1367 // returns 1368 // 1369 // The background collector in a loop iterating on the phases of the 1370 // collection 1371 // acquires the CGC_lock 1372 // sets _foregroundGCShouldWait 1373 // if _foregroundGCIsActive is set 1374 // clears _foregroundGCShouldWait, notifies _CGC_lock 1375 // waits on _CGC_lock for _foregroundGCIsActive to become false 1376 // and exits the loop. 1377 // otherwise 1378 // proceed with that phase of the collection 1379 // if the phase is a stop-the-world phase, 1380 // yield the baton once more just before enqueueing 1381 // the stop-world CMS operation (executed by the VM thread). 1382 // returns after all phases of the collection are done 1383 // 1384 1385 void CMSCollector::acquire_control_and_collect(bool full, 1386 bool clear_all_soft_refs) { 1387 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1388 assert(!Thread::current()->is_ConcurrentGC_thread(), 1389 "shouldn't try to acquire control from self!"); 1390 1391 // Start the protocol for acquiring control of the 1392 // collection from the background collector (aka CMS thread). 1393 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1394 "VM thread should have CMS token"); 1395 // Remember the possibly interrupted state of an ongoing 1396 // concurrent collection 1397 CollectorState first_state = _collectorState; 1398 1399 // Signal to a possibly ongoing concurrent collection that 1400 // we want to do a foreground collection. 1401 _foregroundGCIsActive = true; 1402 1403 // release locks and wait for a notify from the background collector 1404 // releasing the locks in only necessary for phases which 1405 // do yields to improve the granularity of the collection. 1406 assert_lock_strong(bitMapLock()); 1407 // We need to lock the Free list lock for the space that we are 1408 // currently collecting. 1409 assert(haveFreelistLocks(), "Must be holding free list locks"); 1410 bitMapLock()->unlock(); 1411 releaseFreelistLocks(); 1412 { 1413 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1414 if (_foregroundGCShouldWait) { 1415 // We are going to be waiting for action for the CMS thread; 1416 // it had better not be gone (for instance at shutdown)! 1417 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1418 "CMS thread must be running"); 1419 // Wait here until the background collector gives us the go-ahead 1420 ConcurrentMarkSweepThread::clear_CMS_flag( 1421 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1422 // Get a possibly blocked CMS thread going: 1423 // Note that we set _foregroundGCIsActive true above, 1424 // without protection of the CGC_lock. 1425 CGC_lock->notify(); 1426 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1427 "Possible deadlock"); 1428 while (_foregroundGCShouldWait) { 1429 // wait for notification 1430 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1431 // Possibility of delay/starvation here, since CMS token does 1432 // not know to give priority to VM thread? Actually, i think 1433 // there wouldn't be any delay/starvation, but the proof of 1434 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1435 } 1436 ConcurrentMarkSweepThread::set_CMS_flag( 1437 ConcurrentMarkSweepThread::CMS_vm_has_token); 1438 } 1439 } 1440 // The CMS_token is already held. Get back the other locks. 1441 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1442 "VM thread should have CMS token"); 1443 getFreelistLocks(); 1444 bitMapLock()->lock_without_safepoint_check(); 1445 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1446 p2i(Thread::current()), first_state); 1447 log_debug(gc, state)(" gets control with state %d", _collectorState); 1448 1449 // Inform cms gen if this was due to partial collection failing. 1450 // The CMS gen may use this fact to determine its expansion policy. 1451 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1452 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1453 assert(!_cmsGen->incremental_collection_failed(), 1454 "Should have been noticed, reacted to and cleared"); 1455 _cmsGen->set_incremental_collection_failed(); 1456 } 1457 1458 if (first_state > Idling) { 1459 report_concurrent_mode_interruption(); 1460 } 1461 1462 set_did_compact(true); 1463 1464 // If the collection is being acquired from the background 1465 // collector, there may be references on the discovered 1466 // references lists. Abandon those references, since some 1467 // of them may have become unreachable after concurrent 1468 // discovery; the STW compacting collector will redo discovery 1469 // more precisely, without being subject to floating garbage. 1470 // Leaving otherwise unreachable references in the discovered 1471 // lists would require special handling. 1472 ref_processor()->disable_discovery(); 1473 ref_processor()->abandon_partial_discovery(); 1474 ref_processor()->verify_no_references_recorded(); 1475 1476 if (first_state > Idling) { 1477 save_heap_summary(); 1478 } 1479 1480 do_compaction_work(clear_all_soft_refs); 1481 1482 // Has the GC time limit been exceeded? 1483 size_t max_eden_size = _young_gen->max_eden_size(); 1484 GCCause::Cause gc_cause = gch->gc_cause(); 1485 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1486 _young_gen->eden()->used(), 1487 _cmsGen->max_capacity(), 1488 max_eden_size, 1489 full, 1490 gc_cause, 1491 gch->collector_policy()); 1492 1493 // Reset the expansion cause, now that we just completed 1494 // a collection cycle. 1495 clear_expansion_cause(); 1496 _foregroundGCIsActive = false; 1497 return; 1498 } 1499 1500 // Resize the tenured generation 1501 // after obtaining the free list locks for the 1502 // two generations. 1503 void CMSCollector::compute_new_size() { 1504 assert_locked_or_safepoint(Heap_lock); 1505 FreelistLocker z(this); 1506 MetaspaceGC::compute_new_size(); 1507 _cmsGen->compute_new_size_free_list(); 1508 } 1509 1510 // A work method used by the foreground collector to do 1511 // a mark-sweep-compact. 1512 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1513 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1514 1515 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1516 gc_timer->register_gc_start(); 1517 1518 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1519 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1520 1521 gch->pre_full_gc_dump(gc_timer); 1522 1523 GCTraceTime(Trace, gc) t("CMS:MSC"); 1524 1525 // Temporarily widen the span of the weak reference processing to 1526 // the entire heap. 1527 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1528 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1529 // Temporarily, clear the "is_alive_non_header" field of the 1530 // reference processor. 1531 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1532 // Temporarily make reference _processing_ single threaded (non-MT). 1533 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1534 // Temporarily make refs discovery atomic 1535 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1536 // Temporarily make reference _discovery_ single threaded (non-MT) 1537 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1538 1539 ref_processor()->set_enqueuing_is_done(false); 1540 ref_processor()->enable_discovery(); 1541 ref_processor()->setup_policy(clear_all_soft_refs); 1542 // If an asynchronous collection finishes, the _modUnionTable is 1543 // all clear. If we are assuming the collection from an asynchronous 1544 // collection, clear the _modUnionTable. 1545 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1546 "_modUnionTable should be clear if the baton was not passed"); 1547 _modUnionTable.clear_all(); 1548 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1549 "mod union for klasses should be clear if the baton was passed"); 1550 _ct->klass_rem_set()->clear_mod_union(); 1551 1552 // We must adjust the allocation statistics being maintained 1553 // in the free list space. We do so by reading and clearing 1554 // the sweep timer and updating the block flux rate estimates below. 1555 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1556 if (_inter_sweep_timer.is_active()) { 1557 _inter_sweep_timer.stop(); 1558 // Note that we do not use this sample to update the _inter_sweep_estimate. 1559 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1560 _inter_sweep_estimate.padded_average(), 1561 _intra_sweep_estimate.padded_average()); 1562 } 1563 1564 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1565 #ifdef ASSERT 1566 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1567 size_t free_size = cms_space->free(); 1568 assert(free_size == 1569 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1570 * HeapWordSize, 1571 "All the free space should be compacted into one chunk at top"); 1572 assert(cms_space->dictionary()->total_chunk_size( 1573 debug_only(cms_space->freelistLock())) == 0 || 1574 cms_space->totalSizeInIndexedFreeLists() == 0, 1575 "All the free space should be in a single chunk"); 1576 size_t num = cms_space->totalCount(); 1577 assert((free_size == 0 && num == 0) || 1578 (free_size > 0 && (num == 1 || num == 2)), 1579 "There should be at most 2 free chunks after compaction"); 1580 #endif // ASSERT 1581 _collectorState = Resetting; 1582 assert(_restart_addr == NULL, 1583 "Should have been NULL'd before baton was passed"); 1584 reset_stw(); 1585 _cmsGen->reset_after_compaction(); 1586 _concurrent_cycles_since_last_unload = 0; 1587 1588 // Clear any data recorded in the PLAB chunk arrays. 1589 if (_survivor_plab_array != NULL) { 1590 reset_survivor_plab_arrays(); 1591 } 1592 1593 // Adjust the per-size allocation stats for the next epoch. 1594 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1595 // Restart the "inter sweep timer" for the next epoch. 1596 _inter_sweep_timer.reset(); 1597 _inter_sweep_timer.start(); 1598 1599 gch->post_full_gc_dump(gc_timer); 1600 1601 gc_timer->register_gc_end(); 1602 1603 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1604 1605 // For a mark-sweep-compact, compute_new_size() will be called 1606 // in the heap's do_collection() method. 1607 } 1608 1609 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1610 LogHandle(gc, heap) log; 1611 if (!log.is_trace()) { 1612 return; 1613 } 1614 1615 ContiguousSpace* eden_space = _young_gen->eden(); 1616 ContiguousSpace* from_space = _young_gen->from(); 1617 ContiguousSpace* to_space = _young_gen->to(); 1618 // Eden 1619 if (_eden_chunk_array != NULL) { 1620 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1621 p2i(eden_space->bottom()), p2i(eden_space->top()), 1622 p2i(eden_space->end()), eden_space->capacity()); 1623 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1624 _eden_chunk_index, _eden_chunk_capacity); 1625 for (size_t i = 0; i < _eden_chunk_index; i++) { 1626 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1627 } 1628 } 1629 // Survivor 1630 if (_survivor_chunk_array != NULL) { 1631 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1632 p2i(from_space->bottom()), p2i(from_space->top()), 1633 p2i(from_space->end()), from_space->capacity()); 1634 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1635 _survivor_chunk_index, _survivor_chunk_capacity); 1636 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1637 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1638 } 1639 } 1640 } 1641 1642 void CMSCollector::getFreelistLocks() const { 1643 // Get locks for all free lists in all generations that this 1644 // collector is responsible for 1645 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1646 } 1647 1648 void CMSCollector::releaseFreelistLocks() const { 1649 // Release locks for all free lists in all generations that this 1650 // collector is responsible for 1651 _cmsGen->freelistLock()->unlock(); 1652 } 1653 1654 bool CMSCollector::haveFreelistLocks() const { 1655 // Check locks for all free lists in all generations that this 1656 // collector is responsible for 1657 assert_lock_strong(_cmsGen->freelistLock()); 1658 PRODUCT_ONLY(ShouldNotReachHere()); 1659 return true; 1660 } 1661 1662 // A utility class that is used by the CMS collector to 1663 // temporarily "release" the foreground collector from its 1664 // usual obligation to wait for the background collector to 1665 // complete an ongoing phase before proceeding. 1666 class ReleaseForegroundGC: public StackObj { 1667 private: 1668 CMSCollector* _c; 1669 public: 1670 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1671 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1672 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1673 // allow a potentially blocked foreground collector to proceed 1674 _c->_foregroundGCShouldWait = false; 1675 if (_c->_foregroundGCIsActive) { 1676 CGC_lock->notify(); 1677 } 1678 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1679 "Possible deadlock"); 1680 } 1681 1682 ~ReleaseForegroundGC() { 1683 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1685 _c->_foregroundGCShouldWait = true; 1686 } 1687 }; 1688 1689 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1690 assert(Thread::current()->is_ConcurrentGC_thread(), 1691 "A CMS asynchronous collection is only allowed on a CMS thread."); 1692 1693 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1694 { 1695 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1696 MutexLockerEx hl(Heap_lock, safepoint_check); 1697 FreelistLocker fll(this); 1698 MutexLockerEx x(CGC_lock, safepoint_check); 1699 if (_foregroundGCIsActive) { 1700 // The foreground collector is. Skip this 1701 // background collection. 1702 assert(!_foregroundGCShouldWait, "Should be clear"); 1703 return; 1704 } else { 1705 assert(_collectorState == Idling, "Should be idling before start."); 1706 _collectorState = InitialMarking; 1707 register_gc_start(cause); 1708 // Reset the expansion cause, now that we are about to begin 1709 // a new cycle. 1710 clear_expansion_cause(); 1711 1712 // Clear the MetaspaceGC flag since a concurrent collection 1713 // is starting but also clear it after the collection. 1714 MetaspaceGC::set_should_concurrent_collect(false); 1715 } 1716 // Decide if we want to enable class unloading as part of the 1717 // ensuing concurrent GC cycle. 1718 update_should_unload_classes(); 1719 _full_gc_requested = false; // acks all outstanding full gc requests 1720 _full_gc_cause = GCCause::_no_gc; 1721 // Signal that we are about to start a collection 1722 gch->increment_total_full_collections(); // ... starting a collection cycle 1723 _collection_count_start = gch->total_full_collections(); 1724 } 1725 1726 size_t prev_used = _cmsGen->used(); 1727 1728 // The change of the collection state is normally done at this level; 1729 // the exceptions are phases that are executed while the world is 1730 // stopped. For those phases the change of state is done while the 1731 // world is stopped. For baton passing purposes this allows the 1732 // background collector to finish the phase and change state atomically. 1733 // The foreground collector cannot wait on a phase that is done 1734 // while the world is stopped because the foreground collector already 1735 // has the world stopped and would deadlock. 1736 while (_collectorState != Idling) { 1737 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1738 p2i(Thread::current()), _collectorState); 1739 // The foreground collector 1740 // holds the Heap_lock throughout its collection. 1741 // holds the CMS token (but not the lock) 1742 // except while it is waiting for the background collector to yield. 1743 // 1744 // The foreground collector should be blocked (not for long) 1745 // if the background collector is about to start a phase 1746 // executed with world stopped. If the background 1747 // collector has already started such a phase, the 1748 // foreground collector is blocked waiting for the 1749 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1750 // are executed in the VM thread. 1751 // 1752 // The locking order is 1753 // PendingListLock (PLL) -- if applicable (FinalMarking) 1754 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1755 // CMS token (claimed in 1756 // stop_world_and_do() --> 1757 // safepoint_synchronize() --> 1758 // CMSThread::synchronize()) 1759 1760 { 1761 // Check if the FG collector wants us to yield. 1762 CMSTokenSync x(true); // is cms thread 1763 if (waitForForegroundGC()) { 1764 // We yielded to a foreground GC, nothing more to be 1765 // done this round. 1766 assert(_foregroundGCShouldWait == false, "We set it to false in " 1767 "waitForForegroundGC()"); 1768 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1769 p2i(Thread::current()), _collectorState); 1770 return; 1771 } else { 1772 // The background collector can run but check to see if the 1773 // foreground collector has done a collection while the 1774 // background collector was waiting to get the CGC_lock 1775 // above. If yes, break so that _foregroundGCShouldWait 1776 // is cleared before returning. 1777 if (_collectorState == Idling) { 1778 break; 1779 } 1780 } 1781 } 1782 1783 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1784 "should be waiting"); 1785 1786 switch (_collectorState) { 1787 case InitialMarking: 1788 { 1789 ReleaseForegroundGC x(this); 1790 stats().record_cms_begin(); 1791 VM_CMS_Initial_Mark initial_mark_op(this); 1792 VMThread::execute(&initial_mark_op); 1793 } 1794 // The collector state may be any legal state at this point 1795 // since the background collector may have yielded to the 1796 // foreground collector. 1797 break; 1798 case Marking: 1799 // initial marking in checkpointRootsInitialWork has been completed 1800 if (markFromRoots()) { // we were successful 1801 assert(_collectorState == Precleaning, "Collector state should " 1802 "have changed"); 1803 } else { 1804 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1805 } 1806 break; 1807 case Precleaning: 1808 // marking from roots in markFromRoots has been completed 1809 preclean(); 1810 assert(_collectorState == AbortablePreclean || 1811 _collectorState == FinalMarking, 1812 "Collector state should have changed"); 1813 break; 1814 case AbortablePreclean: 1815 abortable_preclean(); 1816 assert(_collectorState == FinalMarking, "Collector state should " 1817 "have changed"); 1818 break; 1819 case FinalMarking: 1820 { 1821 ReleaseForegroundGC x(this); 1822 1823 VM_CMS_Final_Remark final_remark_op(this); 1824 VMThread::execute(&final_remark_op); 1825 } 1826 assert(_foregroundGCShouldWait, "block post-condition"); 1827 break; 1828 case Sweeping: 1829 // final marking in checkpointRootsFinal has been completed 1830 sweep(); 1831 assert(_collectorState == Resizing, "Collector state change " 1832 "to Resizing must be done under the free_list_lock"); 1833 1834 case Resizing: { 1835 // Sweeping has been completed... 1836 // At this point the background collection has completed. 1837 // Don't move the call to compute_new_size() down 1838 // into code that might be executed if the background 1839 // collection was preempted. 1840 { 1841 ReleaseForegroundGC x(this); // unblock FG collection 1842 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1843 CMSTokenSync z(true); // not strictly needed. 1844 if (_collectorState == Resizing) { 1845 compute_new_size(); 1846 save_heap_summary(); 1847 _collectorState = Resetting; 1848 } else { 1849 assert(_collectorState == Idling, "The state should only change" 1850 " because the foreground collector has finished the collection"); 1851 } 1852 } 1853 break; 1854 } 1855 case Resetting: 1856 // CMS heap resizing has been completed 1857 reset_concurrent(); 1858 assert(_collectorState == Idling, "Collector state should " 1859 "have changed"); 1860 1861 MetaspaceGC::set_should_concurrent_collect(false); 1862 1863 stats().record_cms_end(); 1864 // Don't move the concurrent_phases_end() and compute_new_size() 1865 // calls to here because a preempted background collection 1866 // has it's state set to "Resetting". 1867 break; 1868 case Idling: 1869 default: 1870 ShouldNotReachHere(); 1871 break; 1872 } 1873 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1874 p2i(Thread::current()), _collectorState); 1875 assert(_foregroundGCShouldWait, "block post-condition"); 1876 } 1877 1878 // Should this be in gc_epilogue? 1879 collector_policy()->counters()->update_counters(); 1880 1881 { 1882 // Clear _foregroundGCShouldWait and, in the event that the 1883 // foreground collector is waiting, notify it, before 1884 // returning. 1885 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1886 _foregroundGCShouldWait = false; 1887 if (_foregroundGCIsActive) { 1888 CGC_lock->notify(); 1889 } 1890 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1891 "Possible deadlock"); 1892 } 1893 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1894 p2i(Thread::current()), _collectorState); 1895 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1896 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1897 } 1898 1899 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1900 _cms_start_registered = true; 1901 _gc_timer_cm->register_gc_start(); 1902 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1903 } 1904 1905 void CMSCollector::register_gc_end() { 1906 if (_cms_start_registered) { 1907 report_heap_summary(GCWhen::AfterGC); 1908 1909 _gc_timer_cm->register_gc_end(); 1910 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1911 _cms_start_registered = false; 1912 } 1913 } 1914 1915 void CMSCollector::save_heap_summary() { 1916 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1917 _last_heap_summary = gch->create_heap_summary(); 1918 _last_metaspace_summary = gch->create_metaspace_summary(); 1919 } 1920 1921 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1922 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1923 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1924 } 1925 1926 bool CMSCollector::waitForForegroundGC() { 1927 bool res = false; 1928 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1929 "CMS thread should have CMS token"); 1930 // Block the foreground collector until the 1931 // background collectors decides whether to 1932 // yield. 1933 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1934 _foregroundGCShouldWait = true; 1935 if (_foregroundGCIsActive) { 1936 // The background collector yields to the 1937 // foreground collector and returns a value 1938 // indicating that it has yielded. The foreground 1939 // collector can proceed. 1940 res = true; 1941 _foregroundGCShouldWait = false; 1942 ConcurrentMarkSweepThread::clear_CMS_flag( 1943 ConcurrentMarkSweepThread::CMS_cms_has_token); 1944 ConcurrentMarkSweepThread::set_CMS_flag( 1945 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1946 // Get a possibly blocked foreground thread going 1947 CGC_lock->notify(); 1948 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1949 p2i(Thread::current()), _collectorState); 1950 while (_foregroundGCIsActive) { 1951 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1952 } 1953 ConcurrentMarkSweepThread::set_CMS_flag( 1954 ConcurrentMarkSweepThread::CMS_cms_has_token); 1955 ConcurrentMarkSweepThread::clear_CMS_flag( 1956 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1957 } 1958 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1959 p2i(Thread::current()), _collectorState); 1960 return res; 1961 } 1962 1963 // Because of the need to lock the free lists and other structures in 1964 // the collector, common to all the generations that the collector is 1965 // collecting, we need the gc_prologues of individual CMS generations 1966 // delegate to their collector. It may have been simpler had the 1967 // current infrastructure allowed one to call a prologue on a 1968 // collector. In the absence of that we have the generation's 1969 // prologue delegate to the collector, which delegates back 1970 // some "local" work to a worker method in the individual generations 1971 // that it's responsible for collecting, while itself doing any 1972 // work common to all generations it's responsible for. A similar 1973 // comment applies to the gc_epilogue()'s. 1974 // The role of the variable _between_prologue_and_epilogue is to 1975 // enforce the invocation protocol. 1976 void CMSCollector::gc_prologue(bool full) { 1977 // Call gc_prologue_work() for the CMSGen 1978 // we are responsible for. 1979 1980 // The following locking discipline assumes that we are only called 1981 // when the world is stopped. 1982 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1983 1984 // The CMSCollector prologue must call the gc_prologues for the 1985 // "generations" that it's responsible 1986 // for. 1987 1988 assert( Thread::current()->is_VM_thread() 1989 || ( CMSScavengeBeforeRemark 1990 && Thread::current()->is_ConcurrentGC_thread()), 1991 "Incorrect thread type for prologue execution"); 1992 1993 if (_between_prologue_and_epilogue) { 1994 // We have already been invoked; this is a gc_prologue delegation 1995 // from yet another CMS generation that we are responsible for, just 1996 // ignore it since all relevant work has already been done. 1997 return; 1998 } 1999 2000 // set a bit saying prologue has been called; cleared in epilogue 2001 _between_prologue_and_epilogue = true; 2002 // Claim locks for common data structures, then call gc_prologue_work() 2003 // for each CMSGen. 2004 2005 getFreelistLocks(); // gets free list locks on constituent spaces 2006 bitMapLock()->lock_without_safepoint_check(); 2007 2008 // Should call gc_prologue_work() for all cms gens we are responsible for 2009 bool duringMarking = _collectorState >= Marking 2010 && _collectorState < Sweeping; 2011 2012 // The young collections clear the modified oops state, which tells if 2013 // there are any modified oops in the class. The remark phase also needs 2014 // that information. Tell the young collection to save the union of all 2015 // modified klasses. 2016 if (duringMarking) { 2017 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2018 } 2019 2020 bool registerClosure = duringMarking; 2021 2022 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2023 2024 if (!full) { 2025 stats().record_gc0_begin(); 2026 } 2027 } 2028 2029 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2030 2031 _capacity_at_prologue = capacity(); 2032 _used_at_prologue = used(); 2033 2034 // Delegate to CMScollector which knows how to coordinate between 2035 // this and any other CMS generations that it is responsible for 2036 // collecting. 2037 collector()->gc_prologue(full); 2038 } 2039 2040 // This is a "private" interface for use by this generation's CMSCollector. 2041 // Not to be called directly by any other entity (for instance, 2042 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2043 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2044 bool registerClosure, ModUnionClosure* modUnionClosure) { 2045 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2046 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2047 "Should be NULL"); 2048 if (registerClosure) { 2049 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2050 } 2051 cmsSpace()->gc_prologue(); 2052 // Clear stat counters 2053 NOT_PRODUCT( 2054 assert(_numObjectsPromoted == 0, "check"); 2055 assert(_numWordsPromoted == 0, "check"); 2056 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2057 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2058 _numObjectsAllocated = 0; 2059 _numWordsAllocated = 0; 2060 ) 2061 } 2062 2063 void CMSCollector::gc_epilogue(bool full) { 2064 // The following locking discipline assumes that we are only called 2065 // when the world is stopped. 2066 assert(SafepointSynchronize::is_at_safepoint(), 2067 "world is stopped assumption"); 2068 2069 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2070 // if linear allocation blocks need to be appropriately marked to allow the 2071 // the blocks to be parsable. We also check here whether we need to nudge the 2072 // CMS collector thread to start a new cycle (if it's not already active). 2073 assert( Thread::current()->is_VM_thread() 2074 || ( CMSScavengeBeforeRemark 2075 && Thread::current()->is_ConcurrentGC_thread()), 2076 "Incorrect thread type for epilogue execution"); 2077 2078 if (!_between_prologue_and_epilogue) { 2079 // We have already been invoked; this is a gc_epilogue delegation 2080 // from yet another CMS generation that we are responsible for, just 2081 // ignore it since all relevant work has already been done. 2082 return; 2083 } 2084 assert(haveFreelistLocks(), "must have freelist locks"); 2085 assert_lock_strong(bitMapLock()); 2086 2087 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2088 2089 _cmsGen->gc_epilogue_work(full); 2090 2091 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2092 // in case sampling was not already enabled, enable it 2093 _start_sampling = true; 2094 } 2095 // reset _eden_chunk_array so sampling starts afresh 2096 _eden_chunk_index = 0; 2097 2098 size_t cms_used = _cmsGen->cmsSpace()->used(); 2099 2100 // update performance counters - this uses a special version of 2101 // update_counters() that allows the utilization to be passed as a 2102 // parameter, avoiding multiple calls to used(). 2103 // 2104 _cmsGen->update_counters(cms_used); 2105 2106 bitMapLock()->unlock(); 2107 releaseFreelistLocks(); 2108 2109 if (!CleanChunkPoolAsync) { 2110 Chunk::clean_chunk_pool(); 2111 } 2112 2113 set_did_compact(false); 2114 _between_prologue_and_epilogue = false; // ready for next cycle 2115 } 2116 2117 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2118 collector()->gc_epilogue(full); 2119 2120 // Also reset promotion tracking in par gc thread states. 2121 for (uint i = 0; i < ParallelGCThreads; i++) { 2122 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2123 } 2124 } 2125 2126 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2127 assert(!incremental_collection_failed(), "Should have been cleared"); 2128 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2129 cmsSpace()->gc_epilogue(); 2130 // Print stat counters 2131 NOT_PRODUCT( 2132 assert(_numObjectsAllocated == 0, "check"); 2133 assert(_numWordsAllocated == 0, "check"); 2134 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2135 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2136 _numObjectsPromoted = 0; 2137 _numWordsPromoted = 0; 2138 ) 2139 2140 // Call down the chain in contiguous_available needs the freelistLock 2141 // so print this out before releasing the freeListLock. 2142 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2143 } 2144 2145 #ifndef PRODUCT 2146 bool CMSCollector::have_cms_token() { 2147 Thread* thr = Thread::current(); 2148 if (thr->is_VM_thread()) { 2149 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2150 } else if (thr->is_ConcurrentGC_thread()) { 2151 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2152 } else if (thr->is_GC_task_thread()) { 2153 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2154 ParGCRareEvent_lock->owned_by_self(); 2155 } 2156 return false; 2157 } 2158 2159 // Check reachability of the given heap address in CMS generation, 2160 // treating all other generations as roots. 2161 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2162 // We could "guarantee" below, rather than assert, but I'll 2163 // leave these as "asserts" so that an adventurous debugger 2164 // could try this in the product build provided some subset of 2165 // the conditions were met, provided they were interested in the 2166 // results and knew that the computation below wouldn't interfere 2167 // with other concurrent computations mutating the structures 2168 // being read or written. 2169 assert(SafepointSynchronize::is_at_safepoint(), 2170 "Else mutations in object graph will make answer suspect"); 2171 assert(have_cms_token(), "Should hold cms token"); 2172 assert(haveFreelistLocks(), "must hold free list locks"); 2173 assert_lock_strong(bitMapLock()); 2174 2175 // Clear the marking bit map array before starting, but, just 2176 // for kicks, first report if the given address is already marked 2177 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2178 _markBitMap.isMarked(addr) ? "" : " not"); 2179 2180 if (verify_after_remark()) { 2181 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2182 bool result = verification_mark_bm()->isMarked(addr); 2183 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2184 result ? "IS" : "is NOT"); 2185 return result; 2186 } else { 2187 tty->print_cr("Could not compute result"); 2188 return false; 2189 } 2190 } 2191 #endif 2192 2193 void 2194 CMSCollector::print_on_error(outputStream* st) { 2195 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2196 if (collector != NULL) { 2197 CMSBitMap* bitmap = &collector->_markBitMap; 2198 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2199 bitmap->print_on_error(st, " Bits: "); 2200 2201 st->cr(); 2202 2203 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2204 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2205 mut_bitmap->print_on_error(st, " Bits: "); 2206 } 2207 } 2208 2209 //////////////////////////////////////////////////////// 2210 // CMS Verification Support 2211 //////////////////////////////////////////////////////// 2212 // Following the remark phase, the following invariant 2213 // should hold -- each object in the CMS heap which is 2214 // marked in markBitMap() should be marked in the verification_mark_bm(). 2215 2216 class VerifyMarkedClosure: public BitMapClosure { 2217 CMSBitMap* _marks; 2218 bool _failed; 2219 2220 public: 2221 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2222 2223 bool do_bit(size_t offset) { 2224 HeapWord* addr = _marks->offsetToHeapWord(offset); 2225 if (!_marks->isMarked(addr)) { 2226 LogHandle(gc, verify) log; 2227 ResourceMark rm; 2228 oop(addr)->print_on(log.error_stream()); 2229 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2230 _failed = true; 2231 } 2232 return true; 2233 } 2234 2235 bool failed() { return _failed; } 2236 }; 2237 2238 bool CMSCollector::verify_after_remark() { 2239 GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking."); 2240 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2241 static bool init = false; 2242 2243 assert(SafepointSynchronize::is_at_safepoint(), 2244 "Else mutations in object graph will make answer suspect"); 2245 assert(have_cms_token(), 2246 "Else there may be mutual interference in use of " 2247 " verification data structures"); 2248 assert(_collectorState > Marking && _collectorState <= Sweeping, 2249 "Else marking info checked here may be obsolete"); 2250 assert(haveFreelistLocks(), "must hold free list locks"); 2251 assert_lock_strong(bitMapLock()); 2252 2253 2254 // Allocate marking bit map if not already allocated 2255 if (!init) { // first time 2256 if (!verification_mark_bm()->allocate(_span)) { 2257 return false; 2258 } 2259 init = true; 2260 } 2261 2262 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2263 2264 // Turn off refs discovery -- so we will be tracing through refs. 2265 // This is as intended, because by this time 2266 // GC must already have cleared any refs that need to be cleared, 2267 // and traced those that need to be marked; moreover, 2268 // the marking done here is not going to interfere in any 2269 // way with the marking information used by GC. 2270 NoRefDiscovery no_discovery(ref_processor()); 2271 2272 #if defined(COMPILER2) || INCLUDE_JVMCI 2273 DerivedPointerTableDeactivate dpt_deact; 2274 #endif 2275 2276 // Clear any marks from a previous round 2277 verification_mark_bm()->clear_all(); 2278 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2279 verify_work_stacks_empty(); 2280 2281 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2282 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2283 // Update the saved marks which may affect the root scans. 2284 gch->save_marks(); 2285 2286 if (CMSRemarkVerifyVariant == 1) { 2287 // In this first variant of verification, we complete 2288 // all marking, then check if the new marks-vector is 2289 // a subset of the CMS marks-vector. 2290 verify_after_remark_work_1(); 2291 } else if (CMSRemarkVerifyVariant == 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 } else { 2299 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2300 CMSRemarkVerifyVariant); 2301 } 2302 return true; 2303 } 2304 2305 void CMSCollector::verify_after_remark_work_1() { 2306 ResourceMark rm; 2307 HandleMark hm; 2308 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2309 2310 // Get a clear set of claim bits for the roots processing to work with. 2311 ClassLoaderDataGraph::clear_claimed_marks(); 2312 2313 // Mark from roots one level into CMS 2314 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2315 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2316 2317 { 2318 StrongRootsScope srs(1); 2319 2320 gch->gen_process_roots(&srs, 2321 GenCollectedHeap::OldGen, 2322 true, // young gen as roots 2323 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2324 should_unload_classes(), 2325 ¬Older, 2326 NULL, 2327 NULL); 2328 } 2329 2330 // Now mark from the roots 2331 MarkFromRootsClosure markFromRootsClosure(this, _span, 2332 verification_mark_bm(), verification_mark_stack(), 2333 false /* don't yield */, true /* verifying */); 2334 assert(_restart_addr == NULL, "Expected pre-condition"); 2335 verification_mark_bm()->iterate(&markFromRootsClosure); 2336 while (_restart_addr != NULL) { 2337 // Deal with stack overflow: by restarting at the indicated 2338 // address. 2339 HeapWord* ra = _restart_addr; 2340 markFromRootsClosure.reset(ra); 2341 _restart_addr = NULL; 2342 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2343 } 2344 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2345 verify_work_stacks_empty(); 2346 2347 // Marking completed -- now verify that each bit marked in 2348 // verification_mark_bm() is also marked in markBitMap(); flag all 2349 // errors by printing corresponding objects. 2350 VerifyMarkedClosure vcl(markBitMap()); 2351 verification_mark_bm()->iterate(&vcl); 2352 if (vcl.failed()) { 2353 LogHandle(gc, verify) log; 2354 log.error("Failed marking verification after remark"); 2355 ResourceMark rm; 2356 gch->print_on(log.error_stream()); 2357 fatal("CMS: failed marking verification after remark"); 2358 } 2359 } 2360 2361 class VerifyKlassOopsKlassClosure : public KlassClosure { 2362 class VerifyKlassOopsClosure : public OopClosure { 2363 CMSBitMap* _bitmap; 2364 public: 2365 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2366 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2367 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2368 } _oop_closure; 2369 public: 2370 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2371 void do_klass(Klass* k) { 2372 k->oops_do(&_oop_closure); 2373 } 2374 }; 2375 2376 void CMSCollector::verify_after_remark_work_2() { 2377 ResourceMark rm; 2378 HandleMark hm; 2379 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2380 2381 // Get a clear set of claim bits for the roots processing to work with. 2382 ClassLoaderDataGraph::clear_claimed_marks(); 2383 2384 // Mark from roots one level into CMS 2385 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2386 markBitMap()); 2387 CLDToOopClosure cld_closure(¬Older, true); 2388 2389 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2390 2391 { 2392 StrongRootsScope srs(1); 2393 2394 gch->gen_process_roots(&srs, 2395 GenCollectedHeap::OldGen, 2396 true, // young gen as roots 2397 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2398 should_unload_classes(), 2399 ¬Older, 2400 NULL, 2401 &cld_closure); 2402 } 2403 2404 // Now mark from the roots 2405 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2406 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2407 assert(_restart_addr == NULL, "Expected pre-condition"); 2408 verification_mark_bm()->iterate(&markFromRootsClosure); 2409 while (_restart_addr != NULL) { 2410 // Deal with stack overflow: by restarting at the indicated 2411 // address. 2412 HeapWord* ra = _restart_addr; 2413 markFromRootsClosure.reset(ra); 2414 _restart_addr = NULL; 2415 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2416 } 2417 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2418 verify_work_stacks_empty(); 2419 2420 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2421 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2422 2423 // Marking completed -- now verify that each bit marked in 2424 // verification_mark_bm() is also marked in markBitMap(); flag all 2425 // errors by printing corresponding objects. 2426 VerifyMarkedClosure vcl(markBitMap()); 2427 verification_mark_bm()->iterate(&vcl); 2428 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2429 } 2430 2431 void ConcurrentMarkSweepGeneration::save_marks() { 2432 // delegate to CMS space 2433 cmsSpace()->save_marks(); 2434 for (uint i = 0; i < ParallelGCThreads; i++) { 2435 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2436 } 2437 } 2438 2439 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2440 return cmsSpace()->no_allocs_since_save_marks(); 2441 } 2442 2443 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2444 \ 2445 void ConcurrentMarkSweepGeneration:: \ 2446 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2447 cl->set_generation(this); \ 2448 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2449 cl->reset_generation(); \ 2450 save_marks(); \ 2451 } 2452 2453 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2454 2455 void 2456 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2457 if (freelistLock()->owned_by_self()) { 2458 Generation::oop_iterate(cl); 2459 } else { 2460 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2461 Generation::oop_iterate(cl); 2462 } 2463 } 2464 2465 void 2466 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2467 if (freelistLock()->owned_by_self()) { 2468 Generation::object_iterate(cl); 2469 } else { 2470 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2471 Generation::object_iterate(cl); 2472 } 2473 } 2474 2475 void 2476 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2477 if (freelistLock()->owned_by_self()) { 2478 Generation::safe_object_iterate(cl); 2479 } else { 2480 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2481 Generation::safe_object_iterate(cl); 2482 } 2483 } 2484 2485 void 2486 ConcurrentMarkSweepGeneration::post_compact() { 2487 } 2488 2489 void 2490 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2491 // Fix the linear allocation blocks to look like free blocks. 2492 2493 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2494 // are not called when the heap is verified during universe initialization and 2495 // at vm shutdown. 2496 if (freelistLock()->owned_by_self()) { 2497 cmsSpace()->prepare_for_verify(); 2498 } else { 2499 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2500 cmsSpace()->prepare_for_verify(); 2501 } 2502 } 2503 2504 void 2505 ConcurrentMarkSweepGeneration::verify() { 2506 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2507 // are not called when the heap is verified during universe initialization and 2508 // at vm shutdown. 2509 if (freelistLock()->owned_by_self()) { 2510 cmsSpace()->verify(); 2511 } else { 2512 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2513 cmsSpace()->verify(); 2514 } 2515 } 2516 2517 void CMSCollector::verify() { 2518 _cmsGen->verify(); 2519 } 2520 2521 #ifndef PRODUCT 2522 bool CMSCollector::overflow_list_is_empty() const { 2523 assert(_num_par_pushes >= 0, "Inconsistency"); 2524 if (_overflow_list == NULL) { 2525 assert(_num_par_pushes == 0, "Inconsistency"); 2526 } 2527 return _overflow_list == NULL; 2528 } 2529 2530 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2531 // merely consolidate assertion checks that appear to occur together frequently. 2532 void CMSCollector::verify_work_stacks_empty() const { 2533 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2534 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2535 } 2536 2537 void CMSCollector::verify_overflow_empty() const { 2538 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2539 assert(no_preserved_marks(), "No preserved marks"); 2540 } 2541 #endif // PRODUCT 2542 2543 // Decide if we want to enable class unloading as part of the 2544 // ensuing concurrent GC cycle. We will collect and 2545 // unload classes if it's the case that: 2546 // (1) an explicit gc request has been made and the flag 2547 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2548 // (2) (a) class unloading is enabled at the command line, and 2549 // (b) old gen is getting really full 2550 // NOTE: Provided there is no change in the state of the heap between 2551 // calls to this method, it should have idempotent results. Moreover, 2552 // its results should be monotonically increasing (i.e. going from 0 to 1, 2553 // but not 1 to 0) between successive calls between which the heap was 2554 // not collected. For the implementation below, it must thus rely on 2555 // the property that concurrent_cycles_since_last_unload() 2556 // will not decrease unless a collection cycle happened and that 2557 // _cmsGen->is_too_full() are 2558 // themselves also monotonic in that sense. See check_monotonicity() 2559 // below. 2560 void CMSCollector::update_should_unload_classes() { 2561 _should_unload_classes = false; 2562 // Condition 1 above 2563 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2564 _should_unload_classes = true; 2565 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2566 // Disjuncts 2.b.(i,ii,iii) above 2567 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2568 CMSClassUnloadingMaxInterval) 2569 || _cmsGen->is_too_full(); 2570 } 2571 } 2572 2573 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2574 bool res = should_concurrent_collect(); 2575 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2576 return res; 2577 } 2578 2579 void CMSCollector::setup_cms_unloading_and_verification_state() { 2580 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2581 || VerifyBeforeExit; 2582 const int rso = GenCollectedHeap::SO_AllCodeCache; 2583 2584 // We set the proper root for this CMS cycle here. 2585 if (should_unload_classes()) { // Should unload classes this cycle 2586 remove_root_scanning_option(rso); // Shrink the root set appropriately 2587 set_verifying(should_verify); // Set verification state for this cycle 2588 return; // Nothing else needs to be done at this time 2589 } 2590 2591 // Not unloading classes this cycle 2592 assert(!should_unload_classes(), "Inconsistency!"); 2593 2594 // If we are not unloading classes then add SO_AllCodeCache to root 2595 // scanning options. 2596 add_root_scanning_option(rso); 2597 2598 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2599 set_verifying(true); 2600 } else if (verifying() && !should_verify) { 2601 // We were verifying, but some verification flags got disabled. 2602 set_verifying(false); 2603 // Exclude symbols, strings and code cache elements from root scanning to 2604 // reduce IM and RM pauses. 2605 remove_root_scanning_option(rso); 2606 } 2607 } 2608 2609 2610 #ifndef PRODUCT 2611 HeapWord* CMSCollector::block_start(const void* p) const { 2612 const HeapWord* addr = (HeapWord*)p; 2613 if (_span.contains(p)) { 2614 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2615 return _cmsGen->cmsSpace()->block_start(p); 2616 } 2617 } 2618 return NULL; 2619 } 2620 #endif 2621 2622 HeapWord* 2623 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2624 bool tlab, 2625 bool parallel) { 2626 CMSSynchronousYieldRequest yr; 2627 assert(!tlab, "Can't deal with TLAB allocation"); 2628 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2629 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2630 if (GCExpandToAllocateDelayMillis > 0) { 2631 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2632 } 2633 return have_lock_and_allocate(word_size, tlab); 2634 } 2635 2636 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2637 size_t bytes, 2638 size_t expand_bytes, 2639 CMSExpansionCause::Cause cause) 2640 { 2641 2642 bool success = expand(bytes, expand_bytes); 2643 2644 // remember why we expanded; this information is used 2645 // by shouldConcurrentCollect() when making decisions on whether to start 2646 // a new CMS cycle. 2647 if (success) { 2648 set_expansion_cause(cause); 2649 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2650 } 2651 } 2652 2653 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2654 HeapWord* res = NULL; 2655 MutexLocker x(ParGCRareEvent_lock); 2656 while (true) { 2657 // Expansion by some other thread might make alloc OK now: 2658 res = ps->lab.alloc(word_sz); 2659 if (res != NULL) return res; 2660 // If there's not enough expansion space available, give up. 2661 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2662 return NULL; 2663 } 2664 // Otherwise, we try expansion. 2665 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2666 // Now go around the loop and try alloc again; 2667 // A competing par_promote might beat us to the expansion space, 2668 // so we may go around the loop again if promotion fails again. 2669 if (GCExpandToAllocateDelayMillis > 0) { 2670 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2671 } 2672 } 2673 } 2674 2675 2676 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2677 PromotionInfo* promo) { 2678 MutexLocker x(ParGCRareEvent_lock); 2679 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2680 while (true) { 2681 // Expansion by some other thread might make alloc OK now: 2682 if (promo->ensure_spooling_space()) { 2683 assert(promo->has_spooling_space(), 2684 "Post-condition of successful ensure_spooling_space()"); 2685 return true; 2686 } 2687 // If there's not enough expansion space available, give up. 2688 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2689 return false; 2690 } 2691 // Otherwise, we try expansion. 2692 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2693 // Now go around the loop and try alloc again; 2694 // A competing allocation might beat us to the expansion space, 2695 // so we may go around the loop again if allocation fails again. 2696 if (GCExpandToAllocateDelayMillis > 0) { 2697 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2698 } 2699 } 2700 } 2701 2702 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2703 // Only shrink if a compaction was done so that all the free space 2704 // in the generation is in a contiguous block at the end. 2705 if (did_compact()) { 2706 CardGeneration::shrink(bytes); 2707 } 2708 } 2709 2710 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2711 assert_locked_or_safepoint(Heap_lock); 2712 } 2713 2714 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2715 assert_locked_or_safepoint(Heap_lock); 2716 assert_lock_strong(freelistLock()); 2717 log_trace(gc)("Shrinking of CMS not yet implemented"); 2718 return; 2719 } 2720 2721 2722 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2723 // phases. 2724 class CMSPhaseAccounting: public StackObj { 2725 public: 2726 CMSPhaseAccounting(CMSCollector *collector, 2727 const char *title); 2728 ~CMSPhaseAccounting(); 2729 2730 private: 2731 CMSCollector *_collector; 2732 const char *_title; 2733 GCTraceConcTime(Info, gc) _trace_time; 2734 2735 public: 2736 // Not MT-safe; so do not pass around these StackObj's 2737 // where they may be accessed by other threads. 2738 double wallclock_millis() { 2739 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2740 } 2741 }; 2742 2743 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2744 const char *title) : 2745 _collector(collector), _title(title), _trace_time(title) { 2746 2747 _collector->resetYields(); 2748 _collector->resetTimer(); 2749 _collector->startTimer(); 2750 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2751 } 2752 2753 CMSPhaseAccounting::~CMSPhaseAccounting() { 2754 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2755 _collector->stopTimer(); 2756 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2757 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2758 } 2759 2760 // CMS work 2761 2762 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2763 class CMSParMarkTask : public AbstractGangTask { 2764 protected: 2765 CMSCollector* _collector; 2766 uint _n_workers; 2767 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2768 AbstractGangTask(name), 2769 _collector(collector), 2770 _n_workers(n_workers) {} 2771 // Work method in support of parallel rescan ... of young gen spaces 2772 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2773 ContiguousSpace* space, 2774 HeapWord** chunk_array, size_t chunk_top); 2775 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2776 }; 2777 2778 // Parallel initial mark task 2779 class CMSParInitialMarkTask: public CMSParMarkTask { 2780 StrongRootsScope* _strong_roots_scope; 2781 public: 2782 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2783 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2784 _strong_roots_scope(strong_roots_scope) {} 2785 void work(uint worker_id); 2786 }; 2787 2788 // Checkpoint the roots into this generation from outside 2789 // this generation. [Note this initial checkpoint need only 2790 // be approximate -- we'll do a catch up phase subsequently.] 2791 void CMSCollector::checkpointRootsInitial() { 2792 assert(_collectorState == InitialMarking, "Wrong collector state"); 2793 check_correct_thread_executing(); 2794 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2795 2796 save_heap_summary(); 2797 report_heap_summary(GCWhen::BeforeGC); 2798 2799 ReferenceProcessor* rp = ref_processor(); 2800 assert(_restart_addr == NULL, "Control point invariant"); 2801 { 2802 // acquire locks for subsequent manipulations 2803 MutexLockerEx x(bitMapLock(), 2804 Mutex::_no_safepoint_check_flag); 2805 checkpointRootsInitialWork(); 2806 // enable ("weak") refs discovery 2807 rp->enable_discovery(); 2808 _collectorState = Marking; 2809 } 2810 } 2811 2812 void CMSCollector::checkpointRootsInitialWork() { 2813 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2814 assert(_collectorState == InitialMarking, "just checking"); 2815 2816 // Already have locks. 2817 assert_lock_strong(bitMapLock()); 2818 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2819 2820 // Setup the verification and class unloading state for this 2821 // CMS collection cycle. 2822 setup_cms_unloading_and_verification_state(); 2823 2824 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2825 2826 // Reset all the PLAB chunk arrays if necessary. 2827 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2828 reset_survivor_plab_arrays(); 2829 } 2830 2831 ResourceMark rm; 2832 HandleMark hm; 2833 2834 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2835 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2836 2837 verify_work_stacks_empty(); 2838 verify_overflow_empty(); 2839 2840 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2841 // Update the saved marks which may affect the root scans. 2842 gch->save_marks(); 2843 2844 // weak reference processing has not started yet. 2845 ref_processor()->set_enqueuing_is_done(false); 2846 2847 // Need to remember all newly created CLDs, 2848 // so that we can guarantee that the remark finds them. 2849 ClassLoaderDataGraph::remember_new_clds(true); 2850 2851 // Whenever a CLD is found, it will be claimed before proceeding to mark 2852 // the klasses. The claimed marks need to be cleared before marking starts. 2853 ClassLoaderDataGraph::clear_claimed_marks(); 2854 2855 print_eden_and_survivor_chunk_arrays(); 2856 2857 { 2858 #if defined(COMPILER2) || INCLUDE_JVMCI 2859 DerivedPointerTableDeactivate dpt_deact; 2860 #endif 2861 if (CMSParallelInitialMarkEnabled) { 2862 // The parallel version. 2863 WorkGang* workers = gch->workers(); 2864 assert(workers != NULL, "Need parallel worker threads."); 2865 uint n_workers = workers->active_workers(); 2866 2867 StrongRootsScope srs(n_workers); 2868 2869 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2870 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2871 if (n_workers > 1) { 2872 workers->run_task(&tsk); 2873 } else { 2874 tsk.work(0); 2875 } 2876 } else { 2877 // The serial version. 2878 CLDToOopClosure cld_closure(¬Older, true); 2879 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2880 2881 StrongRootsScope srs(1); 2882 2883 gch->gen_process_roots(&srs, 2884 GenCollectedHeap::OldGen, 2885 true, // young gen as roots 2886 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2887 should_unload_classes(), 2888 ¬Older, 2889 NULL, 2890 &cld_closure); 2891 } 2892 } 2893 2894 // Clear mod-union table; it will be dirtied in the prologue of 2895 // CMS generation per each young generation collection. 2896 2897 assert(_modUnionTable.isAllClear(), 2898 "Was cleared in most recent final checkpoint phase" 2899 " or no bits are set in the gc_prologue before the start of the next " 2900 "subsequent marking phase."); 2901 2902 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2903 2904 // Save the end of the used_region of the constituent generations 2905 // to be used to limit the extent of sweep in each generation. 2906 save_sweep_limits(); 2907 verify_overflow_empty(); 2908 } 2909 2910 bool CMSCollector::markFromRoots() { 2911 // we might be tempted to assert that: 2912 // assert(!SafepointSynchronize::is_at_safepoint(), 2913 // "inconsistent argument?"); 2914 // However that wouldn't be right, because it's possible that 2915 // a safepoint is indeed in progress as a young generation 2916 // stop-the-world GC happens even as we mark in this generation. 2917 assert(_collectorState == Marking, "inconsistent state?"); 2918 check_correct_thread_executing(); 2919 verify_overflow_empty(); 2920 2921 // Weak ref discovery note: We may be discovering weak 2922 // refs in this generation concurrent (but interleaved) with 2923 // weak ref discovery by the young generation collector. 2924 2925 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2926 GCTraceCPUTime tcpu; 2927 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2928 bool res = markFromRootsWork(); 2929 if (res) { 2930 _collectorState = Precleaning; 2931 } else { // We failed and a foreground collection wants to take over 2932 assert(_foregroundGCIsActive, "internal state inconsistency"); 2933 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2934 log_debug(gc)("bailing out to foreground collection"); 2935 } 2936 verify_overflow_empty(); 2937 return res; 2938 } 2939 2940 bool CMSCollector::markFromRootsWork() { 2941 // iterate over marked bits in bit map, doing a full scan and mark 2942 // from these roots using the following algorithm: 2943 // . if oop is to the right of the current scan pointer, 2944 // mark corresponding bit (we'll process it later) 2945 // . else (oop is to left of current scan pointer) 2946 // push oop on marking stack 2947 // . drain the marking stack 2948 2949 // Note that when we do a marking step we need to hold the 2950 // bit map lock -- recall that direct allocation (by mutators) 2951 // and promotion (by the young generation collector) is also 2952 // marking the bit map. [the so-called allocate live policy.] 2953 // Because the implementation of bit map marking is not 2954 // robust wrt simultaneous marking of bits in the same word, 2955 // we need to make sure that there is no such interference 2956 // between concurrent such updates. 2957 2958 // already have locks 2959 assert_lock_strong(bitMapLock()); 2960 2961 verify_work_stacks_empty(); 2962 verify_overflow_empty(); 2963 bool result = false; 2964 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2965 result = do_marking_mt(); 2966 } else { 2967 result = do_marking_st(); 2968 } 2969 return result; 2970 } 2971 2972 // Forward decl 2973 class CMSConcMarkingTask; 2974 2975 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2976 CMSCollector* _collector; 2977 CMSConcMarkingTask* _task; 2978 public: 2979 virtual void yield(); 2980 2981 // "n_threads" is the number of threads to be terminated. 2982 // "queue_set" is a set of work queues of other threads. 2983 // "collector" is the CMS collector associated with this task terminator. 2984 // "yield" indicates whether we need the gang as a whole to yield. 2985 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2986 ParallelTaskTerminator(n_threads, queue_set), 2987 _collector(collector) { } 2988 2989 void set_task(CMSConcMarkingTask* task) { 2990 _task = task; 2991 } 2992 }; 2993 2994 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2995 CMSConcMarkingTask* _task; 2996 public: 2997 bool should_exit_termination(); 2998 void set_task(CMSConcMarkingTask* task) { 2999 _task = task; 3000 } 3001 }; 3002 3003 // MT Concurrent Marking Task 3004 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3005 CMSCollector* _collector; 3006 uint _n_workers; // requested/desired # workers 3007 bool _result; 3008 CompactibleFreeListSpace* _cms_space; 3009 char _pad_front[64]; // padding to ... 3010 HeapWord* _global_finger; // ... avoid sharing cache line 3011 char _pad_back[64]; 3012 HeapWord* _restart_addr; 3013 3014 // Exposed here for yielding support 3015 Mutex* const _bit_map_lock; 3016 3017 // The per thread work queues, available here for stealing 3018 OopTaskQueueSet* _task_queues; 3019 3020 // Termination (and yielding) support 3021 CMSConcMarkingTerminator _term; 3022 CMSConcMarkingTerminatorTerminator _term_term; 3023 3024 public: 3025 CMSConcMarkingTask(CMSCollector* collector, 3026 CompactibleFreeListSpace* cms_space, 3027 YieldingFlexibleWorkGang* workers, 3028 OopTaskQueueSet* task_queues): 3029 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3030 _collector(collector), 3031 _cms_space(cms_space), 3032 _n_workers(0), _result(true), 3033 _task_queues(task_queues), 3034 _term(_n_workers, task_queues, _collector), 3035 _bit_map_lock(collector->bitMapLock()) 3036 { 3037 _requested_size = _n_workers; 3038 _term.set_task(this); 3039 _term_term.set_task(this); 3040 _restart_addr = _global_finger = _cms_space->bottom(); 3041 } 3042 3043 3044 OopTaskQueueSet* task_queues() { return _task_queues; } 3045 3046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3047 3048 HeapWord** global_finger_addr() { return &_global_finger; } 3049 3050 CMSConcMarkingTerminator* terminator() { return &_term; } 3051 3052 virtual void set_for_termination(uint active_workers) { 3053 terminator()->reset_for_reuse(active_workers); 3054 } 3055 3056 void work(uint worker_id); 3057 bool should_yield() { 3058 return ConcurrentMarkSweepThread::should_yield() 3059 && !_collector->foregroundGCIsActive(); 3060 } 3061 3062 virtual void coordinator_yield(); // stuff done by coordinator 3063 bool result() { return _result; } 3064 3065 void reset(HeapWord* ra) { 3066 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3067 _restart_addr = _global_finger = ra; 3068 _term.reset_for_reuse(); 3069 } 3070 3071 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3072 OopTaskQueue* work_q); 3073 3074 private: 3075 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3076 void do_work_steal(int i); 3077 void bump_global_finger(HeapWord* f); 3078 }; 3079 3080 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3081 assert(_task != NULL, "Error"); 3082 return _task->yielding(); 3083 // Note that we do not need the disjunct || _task->should_yield() above 3084 // because we want terminating threads to yield only if the task 3085 // is already in the midst of yielding, which happens only after at least one 3086 // thread has yielded. 3087 } 3088 3089 void CMSConcMarkingTerminator::yield() { 3090 if (_task->should_yield()) { 3091 _task->yield(); 3092 } else { 3093 ParallelTaskTerminator::yield(); 3094 } 3095 } 3096 3097 //////////////////////////////////////////////////////////////// 3098 // Concurrent Marking Algorithm Sketch 3099 //////////////////////////////////////////////////////////////// 3100 // Until all tasks exhausted (both spaces): 3101 // -- claim next available chunk 3102 // -- bump global finger via CAS 3103 // -- find first object that starts in this chunk 3104 // and start scanning bitmap from that position 3105 // -- scan marked objects for oops 3106 // -- CAS-mark target, and if successful: 3107 // . if target oop is above global finger (volatile read) 3108 // nothing to do 3109 // . if target oop is in chunk and above local finger 3110 // then nothing to do 3111 // . else push on work-queue 3112 // -- Deal with possible overflow issues: 3113 // . local work-queue overflow causes stuff to be pushed on 3114 // global (common) overflow queue 3115 // . always first empty local work queue 3116 // . then get a batch of oops from global work queue if any 3117 // . then do work stealing 3118 // -- When all tasks claimed (both spaces) 3119 // and local work queue empty, 3120 // then in a loop do: 3121 // . check global overflow stack; steal a batch of oops and trace 3122 // . try to steal from other threads oif GOS is empty 3123 // . if neither is available, offer termination 3124 // -- Terminate and return result 3125 // 3126 void CMSConcMarkingTask::work(uint worker_id) { 3127 elapsedTimer _timer; 3128 ResourceMark rm; 3129 HandleMark hm; 3130 3131 DEBUG_ONLY(_collector->verify_overflow_empty();) 3132 3133 // Before we begin work, our work queue should be empty 3134 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3135 // Scan the bitmap covering _cms_space, tracing through grey objects. 3136 _timer.start(); 3137 do_scan_and_mark(worker_id, _cms_space); 3138 _timer.stop(); 3139 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3140 3141 // ... do work stealing 3142 _timer.reset(); 3143 _timer.start(); 3144 do_work_steal(worker_id); 3145 _timer.stop(); 3146 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3147 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3148 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3149 // Note that under the current task protocol, the 3150 // following assertion is true even of the spaces 3151 // expanded since the completion of the concurrent 3152 // marking. XXX This will likely change under a strict 3153 // ABORT semantics. 3154 // After perm removal the comparison was changed to 3155 // greater than or equal to from strictly greater than. 3156 // Before perm removal the highest address sweep would 3157 // have been at the end of perm gen but now is at the 3158 // end of the tenured gen. 3159 assert(_global_finger >= _cms_space->end(), 3160 "All tasks have been completed"); 3161 DEBUG_ONLY(_collector->verify_overflow_empty();) 3162 } 3163 3164 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3165 HeapWord* read = _global_finger; 3166 HeapWord* cur = read; 3167 while (f > read) { 3168 cur = read; 3169 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3170 if (cur == read) { 3171 // our cas succeeded 3172 assert(_global_finger >= f, "protocol consistency"); 3173 break; 3174 } 3175 } 3176 } 3177 3178 // This is really inefficient, and should be redone by 3179 // using (not yet available) block-read and -write interfaces to the 3180 // stack and the work_queue. XXX FIX ME !!! 3181 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3182 OopTaskQueue* work_q) { 3183 // Fast lock-free check 3184 if (ovflw_stk->length() == 0) { 3185 return false; 3186 } 3187 assert(work_q->size() == 0, "Shouldn't steal"); 3188 MutexLockerEx ml(ovflw_stk->par_lock(), 3189 Mutex::_no_safepoint_check_flag); 3190 // Grab up to 1/4 the size of the work queue 3191 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3192 (size_t)ParGCDesiredObjsFromOverflowList); 3193 num = MIN2(num, ovflw_stk->length()); 3194 for (int i = (int) num; i > 0; i--) { 3195 oop cur = ovflw_stk->pop(); 3196 assert(cur != NULL, "Counted wrong?"); 3197 work_q->push(cur); 3198 } 3199 return num > 0; 3200 } 3201 3202 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3203 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3204 int n_tasks = pst->n_tasks(); 3205 // We allow that there may be no tasks to do here because 3206 // we are restarting after a stack overflow. 3207 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3208 uint nth_task = 0; 3209 3210 HeapWord* aligned_start = sp->bottom(); 3211 if (sp->used_region().contains(_restart_addr)) { 3212 // Align down to a card boundary for the start of 0th task 3213 // for this space. 3214 aligned_start = 3215 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3216 CardTableModRefBS::card_size); 3217 } 3218 3219 size_t chunk_size = sp->marking_task_size(); 3220 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3221 // Having claimed the nth task in this space, 3222 // compute the chunk that it corresponds to: 3223 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3224 aligned_start + (nth_task+1)*chunk_size); 3225 // Try and bump the global finger via a CAS; 3226 // note that we need to do the global finger bump 3227 // _before_ taking the intersection below, because 3228 // the task corresponding to that region will be 3229 // deemed done even if the used_region() expands 3230 // because of allocation -- as it almost certainly will 3231 // during start-up while the threads yield in the 3232 // closure below. 3233 HeapWord* finger = span.end(); 3234 bump_global_finger(finger); // atomically 3235 // There are null tasks here corresponding to chunks 3236 // beyond the "top" address of the space. 3237 span = span.intersection(sp->used_region()); 3238 if (!span.is_empty()) { // Non-null task 3239 HeapWord* prev_obj; 3240 assert(!span.contains(_restart_addr) || nth_task == 0, 3241 "Inconsistency"); 3242 if (nth_task == 0) { 3243 // For the 0th task, we'll not need to compute a block_start. 3244 if (span.contains(_restart_addr)) { 3245 // In the case of a restart because of stack overflow, 3246 // we might additionally skip a chunk prefix. 3247 prev_obj = _restart_addr; 3248 } else { 3249 prev_obj = span.start(); 3250 } 3251 } else { 3252 // We want to skip the first object because 3253 // the protocol is to scan any object in its entirety 3254 // that _starts_ in this span; a fortiori, any 3255 // object starting in an earlier span is scanned 3256 // as part of an earlier claimed task. 3257 // Below we use the "careful" version of block_start 3258 // so we do not try to navigate uninitialized objects. 3259 prev_obj = sp->block_start_careful(span.start()); 3260 // Below we use a variant of block_size that uses the 3261 // Printezis bits to avoid waiting for allocated 3262 // objects to become initialized/parsable. 3263 while (prev_obj < span.start()) { 3264 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3265 if (sz > 0) { 3266 prev_obj += sz; 3267 } else { 3268 // In this case we may end up doing a bit of redundant 3269 // scanning, but that appears unavoidable, short of 3270 // locking the free list locks; see bug 6324141. 3271 break; 3272 } 3273 } 3274 } 3275 if (prev_obj < span.end()) { 3276 MemRegion my_span = MemRegion(prev_obj, span.end()); 3277 // Do the marking work within a non-empty span -- 3278 // the last argument to the constructor indicates whether the 3279 // iteration should be incremental with periodic yields. 3280 ParMarkFromRootsClosure cl(this, _collector, my_span, 3281 &_collector->_markBitMap, 3282 work_queue(i), 3283 &_collector->_markStack); 3284 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3285 } // else nothing to do for this task 3286 } // else nothing to do for this task 3287 } 3288 // We'd be tempted to assert here that since there are no 3289 // more tasks left to claim in this space, the global_finger 3290 // must exceed space->top() and a fortiori space->end(). However, 3291 // that would not quite be correct because the bumping of 3292 // global_finger occurs strictly after the claiming of a task, 3293 // so by the time we reach here the global finger may not yet 3294 // have been bumped up by the thread that claimed the last 3295 // task. 3296 pst->all_tasks_completed(); 3297 } 3298 3299 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3300 private: 3301 CMSCollector* _collector; 3302 CMSConcMarkingTask* _task; 3303 MemRegion _span; 3304 CMSBitMap* _bit_map; 3305 CMSMarkStack* _overflow_stack; 3306 OopTaskQueue* _work_queue; 3307 protected: 3308 DO_OOP_WORK_DEFN 3309 public: 3310 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3311 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3312 MetadataAwareOopClosure(collector->ref_processor()), 3313 _collector(collector), 3314 _task(task), 3315 _span(collector->_span), 3316 _work_queue(work_queue), 3317 _bit_map(bit_map), 3318 _overflow_stack(overflow_stack) 3319 { } 3320 virtual void do_oop(oop* p); 3321 virtual void do_oop(narrowOop* p); 3322 3323 void trim_queue(size_t max); 3324 void handle_stack_overflow(HeapWord* lost); 3325 void do_yield_check() { 3326 if (_task->should_yield()) { 3327 _task->yield(); 3328 } 3329 } 3330 }; 3331 3332 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3333 3334 // Grey object scanning during work stealing phase -- 3335 // the salient assumption here is that any references 3336 // that are in these stolen objects being scanned must 3337 // already have been initialized (else they would not have 3338 // been published), so we do not need to check for 3339 // uninitialized objects before pushing here. 3340 void ParConcMarkingClosure::do_oop(oop obj) { 3341 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3342 HeapWord* addr = (HeapWord*)obj; 3343 // Check if oop points into the CMS generation 3344 // and is not marked 3345 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3346 // a white object ... 3347 // If we manage to "claim" the object, by being the 3348 // first thread to mark it, then we push it on our 3349 // marking stack 3350 if (_bit_map->par_mark(addr)) { // ... now grey 3351 // push on work queue (grey set) 3352 bool simulate_overflow = false; 3353 NOT_PRODUCT( 3354 if (CMSMarkStackOverflowALot && 3355 _collector->simulate_overflow()) { 3356 // simulate a stack overflow 3357 simulate_overflow = true; 3358 } 3359 ) 3360 if (simulate_overflow || 3361 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3362 // stack overflow 3363 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3364 // We cannot assert that the overflow stack is full because 3365 // it may have been emptied since. 3366 assert(simulate_overflow || 3367 _work_queue->size() == _work_queue->max_elems(), 3368 "Else push should have succeeded"); 3369 handle_stack_overflow(addr); 3370 } 3371 } // Else, some other thread got there first 3372 do_yield_check(); 3373 } 3374 } 3375 3376 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3377 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3378 3379 void ParConcMarkingClosure::trim_queue(size_t max) { 3380 while (_work_queue->size() > max) { 3381 oop new_oop; 3382 if (_work_queue->pop_local(new_oop)) { 3383 assert(new_oop->is_oop(), "Should be an oop"); 3384 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3385 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3386 new_oop->oop_iterate(this); // do_oop() above 3387 do_yield_check(); 3388 } 3389 } 3390 } 3391 3392 // Upon stack overflow, we discard (part of) the stack, 3393 // remembering the least address amongst those discarded 3394 // in CMSCollector's _restart_address. 3395 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3396 // We need to do this under a mutex to prevent other 3397 // workers from interfering with the work done below. 3398 MutexLockerEx ml(_overflow_stack->par_lock(), 3399 Mutex::_no_safepoint_check_flag); 3400 // Remember the least grey address discarded 3401 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3402 _collector->lower_restart_addr(ra); 3403 _overflow_stack->reset(); // discard stack contents 3404 _overflow_stack->expand(); // expand the stack if possible 3405 } 3406 3407 3408 void CMSConcMarkingTask::do_work_steal(int i) { 3409 OopTaskQueue* work_q = work_queue(i); 3410 oop obj_to_scan; 3411 CMSBitMap* bm = &(_collector->_markBitMap); 3412 CMSMarkStack* ovflw = &(_collector->_markStack); 3413 int* seed = _collector->hash_seed(i); 3414 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3415 while (true) { 3416 cl.trim_queue(0); 3417 assert(work_q->size() == 0, "Should have been emptied above"); 3418 if (get_work_from_overflow_stack(ovflw, work_q)) { 3419 // Can't assert below because the work obtained from the 3420 // overflow stack may already have been stolen from us. 3421 // assert(work_q->size() > 0, "Work from overflow stack"); 3422 continue; 3423 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3424 assert(obj_to_scan->is_oop(), "Should be an oop"); 3425 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3426 obj_to_scan->oop_iterate(&cl); 3427 } else if (terminator()->offer_termination(&_term_term)) { 3428 assert(work_q->size() == 0, "Impossible!"); 3429 break; 3430 } else if (yielding() || should_yield()) { 3431 yield(); 3432 } 3433 } 3434 } 3435 3436 // This is run by the CMS (coordinator) thread. 3437 void CMSConcMarkingTask::coordinator_yield() { 3438 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3439 "CMS thread should hold CMS token"); 3440 // First give up the locks, then yield, then re-lock 3441 // We should probably use a constructor/destructor idiom to 3442 // do this unlock/lock or modify the MutexUnlocker class to 3443 // serve our purpose. XXX 3444 assert_lock_strong(_bit_map_lock); 3445 _bit_map_lock->unlock(); 3446 ConcurrentMarkSweepThread::desynchronize(true); 3447 _collector->stopTimer(); 3448 _collector->incrementYields(); 3449 3450 // It is possible for whichever thread initiated the yield request 3451 // not to get a chance to wake up and take the bitmap lock between 3452 // this thread releasing it and reacquiring it. So, while the 3453 // should_yield() flag is on, let's sleep for a bit to give the 3454 // other thread a chance to wake up. The limit imposed on the number 3455 // of iterations is defensive, to avoid any unforseen circumstances 3456 // putting us into an infinite loop. Since it's always been this 3457 // (coordinator_yield()) method that was observed to cause the 3458 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3459 // which is by default non-zero. For the other seven methods that 3460 // also perform the yield operation, as are using a different 3461 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3462 // can enable the sleeping for those methods too, if necessary. 3463 // See 6442774. 3464 // 3465 // We really need to reconsider the synchronization between the GC 3466 // thread and the yield-requesting threads in the future and we 3467 // should really use wait/notify, which is the recommended 3468 // way of doing this type of interaction. Additionally, we should 3469 // consolidate the eight methods that do the yield operation and they 3470 // are almost identical into one for better maintainability and 3471 // readability. See 6445193. 3472 // 3473 // Tony 2006.06.29 3474 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3475 ConcurrentMarkSweepThread::should_yield() && 3476 !CMSCollector::foregroundGCIsActive(); ++i) { 3477 os::sleep(Thread::current(), 1, false); 3478 } 3479 3480 ConcurrentMarkSweepThread::synchronize(true); 3481 _bit_map_lock->lock_without_safepoint_check(); 3482 _collector->startTimer(); 3483 } 3484 3485 bool CMSCollector::do_marking_mt() { 3486 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3487 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3488 conc_workers()->active_workers(), 3489 Threads::number_of_non_daemon_threads()); 3490 conc_workers()->set_active_workers(num_workers); 3491 3492 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3493 3494 CMSConcMarkingTask tsk(this, 3495 cms_space, 3496 conc_workers(), 3497 task_queues()); 3498 3499 // Since the actual number of workers we get may be different 3500 // from the number we requested above, do we need to do anything different 3501 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3502 // class?? XXX 3503 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3504 3505 // Refs discovery is already non-atomic. 3506 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3507 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3508 conc_workers()->start_task(&tsk); 3509 while (tsk.yielded()) { 3510 tsk.coordinator_yield(); 3511 conc_workers()->continue_task(&tsk); 3512 } 3513 // If the task was aborted, _restart_addr will be non-NULL 3514 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3515 while (_restart_addr != NULL) { 3516 // XXX For now we do not make use of ABORTED state and have not 3517 // yet implemented the right abort semantics (even in the original 3518 // single-threaded CMS case). That needs some more investigation 3519 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3520 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3521 // If _restart_addr is non-NULL, a marking stack overflow 3522 // occurred; we need to do a fresh marking iteration from the 3523 // indicated restart address. 3524 if (_foregroundGCIsActive) { 3525 // We may be running into repeated stack overflows, having 3526 // reached the limit of the stack size, while making very 3527 // slow forward progress. It may be best to bail out and 3528 // let the foreground collector do its job. 3529 // Clear _restart_addr, so that foreground GC 3530 // works from scratch. This avoids the headache of 3531 // a "rescan" which would otherwise be needed because 3532 // of the dirty mod union table & card table. 3533 _restart_addr = NULL; 3534 return false; 3535 } 3536 // Adjust the task to restart from _restart_addr 3537 tsk.reset(_restart_addr); 3538 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3539 _restart_addr); 3540 _restart_addr = NULL; 3541 // Get the workers going again 3542 conc_workers()->start_task(&tsk); 3543 while (tsk.yielded()) { 3544 tsk.coordinator_yield(); 3545 conc_workers()->continue_task(&tsk); 3546 } 3547 } 3548 assert(tsk.completed(), "Inconsistency"); 3549 assert(tsk.result() == true, "Inconsistency"); 3550 return true; 3551 } 3552 3553 bool CMSCollector::do_marking_st() { 3554 ResourceMark rm; 3555 HandleMark hm; 3556 3557 // Temporarily make refs discovery single threaded (non-MT) 3558 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3559 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3560 &_markStack, CMSYield); 3561 // the last argument to iterate indicates whether the iteration 3562 // should be incremental with periodic yields. 3563 _markBitMap.iterate(&markFromRootsClosure); 3564 // If _restart_addr is non-NULL, a marking stack overflow 3565 // occurred; we need to do a fresh iteration from the 3566 // indicated restart address. 3567 while (_restart_addr != NULL) { 3568 if (_foregroundGCIsActive) { 3569 // We may be running into repeated stack overflows, having 3570 // reached the limit of the stack size, while making very 3571 // slow forward progress. It may be best to bail out and 3572 // let the foreground collector do its job. 3573 // Clear _restart_addr, so that foreground GC 3574 // works from scratch. This avoids the headache of 3575 // a "rescan" which would otherwise be needed because 3576 // of the dirty mod union table & card table. 3577 _restart_addr = NULL; 3578 return false; // indicating failure to complete marking 3579 } 3580 // Deal with stack overflow: 3581 // we restart marking from _restart_addr 3582 HeapWord* ra = _restart_addr; 3583 markFromRootsClosure.reset(ra); 3584 _restart_addr = NULL; 3585 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3586 } 3587 return true; 3588 } 3589 3590 void CMSCollector::preclean() { 3591 check_correct_thread_executing(); 3592 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3593 verify_work_stacks_empty(); 3594 verify_overflow_empty(); 3595 _abort_preclean = false; 3596 if (CMSPrecleaningEnabled) { 3597 if (!CMSEdenChunksRecordAlways) { 3598 _eden_chunk_index = 0; 3599 } 3600 size_t used = get_eden_used(); 3601 size_t capacity = get_eden_capacity(); 3602 // Don't start sampling unless we will get sufficiently 3603 // many samples. 3604 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3605 * CMSScheduleRemarkEdenPenetration)) { 3606 _start_sampling = true; 3607 } else { 3608 _start_sampling = false; 3609 } 3610 GCTraceCPUTime tcpu; 3611 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3612 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3613 } 3614 CMSTokenSync x(true); // is cms thread 3615 if (CMSPrecleaningEnabled) { 3616 sample_eden(); 3617 _collectorState = AbortablePreclean; 3618 } else { 3619 _collectorState = FinalMarking; 3620 } 3621 verify_work_stacks_empty(); 3622 verify_overflow_empty(); 3623 } 3624 3625 // Try and schedule the remark such that young gen 3626 // occupancy is CMSScheduleRemarkEdenPenetration %. 3627 void CMSCollector::abortable_preclean() { 3628 check_correct_thread_executing(); 3629 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3630 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3631 3632 // If Eden's current occupancy is below this threshold, 3633 // immediately schedule the remark; else preclean 3634 // past the next scavenge in an effort to 3635 // schedule the pause as described above. By choosing 3636 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3637 // we will never do an actual abortable preclean cycle. 3638 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3639 GCTraceCPUTime tcpu; 3640 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3641 // We need more smarts in the abortable preclean 3642 // loop below to deal with cases where allocation 3643 // in young gen is very very slow, and our precleaning 3644 // is running a losing race against a horde of 3645 // mutators intent on flooding us with CMS updates 3646 // (dirty cards). 3647 // One, admittedly dumb, strategy is to give up 3648 // after a certain number of abortable precleaning loops 3649 // or after a certain maximum time. We want to make 3650 // this smarter in the next iteration. 3651 // XXX FIX ME!!! YSR 3652 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3653 while (!(should_abort_preclean() || 3654 ConcurrentMarkSweepThread::should_terminate())) { 3655 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3656 cumworkdone += workdone; 3657 loops++; 3658 // Voluntarily terminate abortable preclean phase if we have 3659 // been at it for too long. 3660 if ((CMSMaxAbortablePrecleanLoops != 0) && 3661 loops >= CMSMaxAbortablePrecleanLoops) { 3662 log_debug(gc)(" CMS: abort preclean due to loops "); 3663 break; 3664 } 3665 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3666 log_debug(gc)(" CMS: abort preclean due to time "); 3667 break; 3668 } 3669 // If we are doing little work each iteration, we should 3670 // take a short break. 3671 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3672 // Sleep for some time, waiting for work to accumulate 3673 stopTimer(); 3674 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3675 startTimer(); 3676 waited++; 3677 } 3678 } 3679 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3680 loops, waited, cumworkdone); 3681 } 3682 CMSTokenSync x(true); // is cms thread 3683 if (_collectorState != Idling) { 3684 assert(_collectorState == AbortablePreclean, 3685 "Spontaneous state transition?"); 3686 _collectorState = FinalMarking; 3687 } // Else, a foreground collection completed this CMS cycle. 3688 return; 3689 } 3690 3691 // Respond to an Eden sampling opportunity 3692 void CMSCollector::sample_eden() { 3693 // Make sure a young gc cannot sneak in between our 3694 // reading and recording of a sample. 3695 assert(Thread::current()->is_ConcurrentGC_thread(), 3696 "Only the cms thread may collect Eden samples"); 3697 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3698 "Should collect samples while holding CMS token"); 3699 if (!_start_sampling) { 3700 return; 3701 } 3702 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3703 // is populated by the young generation. 3704 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3705 if (_eden_chunk_index < _eden_chunk_capacity) { 3706 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3707 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3708 "Unexpected state of Eden"); 3709 // We'd like to check that what we just sampled is an oop-start address; 3710 // however, we cannot do that here since the object may not yet have been 3711 // initialized. So we'll instead do the check when we _use_ this sample 3712 // later. 3713 if (_eden_chunk_index == 0 || 3714 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3715 _eden_chunk_array[_eden_chunk_index-1]) 3716 >= CMSSamplingGrain)) { 3717 _eden_chunk_index++; // commit sample 3718 } 3719 } 3720 } 3721 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3722 size_t used = get_eden_used(); 3723 size_t capacity = get_eden_capacity(); 3724 assert(used <= capacity, "Unexpected state of Eden"); 3725 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3726 _abort_preclean = true; 3727 } 3728 } 3729 } 3730 3731 3732 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3733 assert(_collectorState == Precleaning || 3734 _collectorState == AbortablePreclean, "incorrect state"); 3735 ResourceMark rm; 3736 HandleMark hm; 3737 3738 // Precleaning is currently not MT but the reference processor 3739 // may be set for MT. Disable it temporarily here. 3740 ReferenceProcessor* rp = ref_processor(); 3741 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3742 3743 // Do one pass of scrubbing the discovered reference lists 3744 // to remove any reference objects with strongly-reachable 3745 // referents. 3746 if (clean_refs) { 3747 CMSPrecleanRefsYieldClosure yield_cl(this); 3748 assert(rp->span().equals(_span), "Spans should be equal"); 3749 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3750 &_markStack, true /* preclean */); 3751 CMSDrainMarkingStackClosure complete_trace(this, 3752 _span, &_markBitMap, &_markStack, 3753 &keep_alive, true /* preclean */); 3754 3755 // We don't want this step to interfere with a young 3756 // collection because we don't want to take CPU 3757 // or memory bandwidth away from the young GC threads 3758 // (which may be as many as there are CPUs). 3759 // Note that we don't need to protect ourselves from 3760 // interference with mutators because they can't 3761 // manipulate the discovered reference lists nor affect 3762 // the computed reachability of the referents, the 3763 // only properties manipulated by the precleaning 3764 // of these reference lists. 3765 stopTimer(); 3766 CMSTokenSyncWithLocks x(true /* is cms thread */, 3767 bitMapLock()); 3768 startTimer(); 3769 sample_eden(); 3770 3771 // The following will yield to allow foreground 3772 // collection to proceed promptly. XXX YSR: 3773 // The code in this method may need further 3774 // tweaking for better performance and some restructuring 3775 // for cleaner interfaces. 3776 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3777 rp->preclean_discovered_references( 3778 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3779 gc_timer); 3780 } 3781 3782 if (clean_survivor) { // preclean the active survivor space(s) 3783 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3784 &_markBitMap, &_modUnionTable, 3785 &_markStack, true /* precleaning phase */); 3786 stopTimer(); 3787 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3788 bitMapLock()); 3789 startTimer(); 3790 unsigned int before_count = 3791 GenCollectedHeap::heap()->total_collections(); 3792 SurvivorSpacePrecleanClosure 3793 sss_cl(this, _span, &_markBitMap, &_markStack, 3794 &pam_cl, before_count, CMSYield); 3795 _young_gen->from()->object_iterate_careful(&sss_cl); 3796 _young_gen->to()->object_iterate_careful(&sss_cl); 3797 } 3798 MarkRefsIntoAndScanClosure 3799 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3800 &_markStack, this, CMSYield, 3801 true /* precleaning phase */); 3802 // CAUTION: The following closure has persistent state that may need to 3803 // be reset upon a decrease in the sequence of addresses it 3804 // processes. 3805 ScanMarkedObjectsAgainCarefullyClosure 3806 smoac_cl(this, _span, 3807 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3808 3809 // Preclean dirty cards in ModUnionTable and CardTable using 3810 // appropriate convergence criterion; 3811 // repeat CMSPrecleanIter times unless we find that 3812 // we are losing. 3813 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3814 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3815 "Bad convergence multiplier"); 3816 assert(CMSPrecleanThreshold >= 100, 3817 "Unreasonably low CMSPrecleanThreshold"); 3818 3819 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3820 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3821 numIter < CMSPrecleanIter; 3822 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3823 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3824 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3825 // Either there are very few dirty cards, so re-mark 3826 // pause will be small anyway, or our pre-cleaning isn't 3827 // that much faster than the rate at which cards are being 3828 // dirtied, so we might as well stop and re-mark since 3829 // precleaning won't improve our re-mark time by much. 3830 if (curNumCards <= CMSPrecleanThreshold || 3831 (numIter > 0 && 3832 (curNumCards * CMSPrecleanDenominator > 3833 lastNumCards * CMSPrecleanNumerator))) { 3834 numIter++; 3835 cumNumCards += curNumCards; 3836 break; 3837 } 3838 } 3839 3840 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3841 3842 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3843 cumNumCards += curNumCards; 3844 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3845 curNumCards, cumNumCards, numIter); 3846 return cumNumCards; // as a measure of useful work done 3847 } 3848 3849 // PRECLEANING NOTES: 3850 // Precleaning involves: 3851 // . reading the bits of the modUnionTable and clearing the set bits. 3852 // . For the cards corresponding to the set bits, we scan the 3853 // objects on those cards. This means we need the free_list_lock 3854 // so that we can safely iterate over the CMS space when scanning 3855 // for oops. 3856 // . When we scan the objects, we'll be both reading and setting 3857 // marks in the marking bit map, so we'll need the marking bit map. 3858 // . For protecting _collector_state transitions, we take the CGC_lock. 3859 // Note that any races in the reading of of card table entries by the 3860 // CMS thread on the one hand and the clearing of those entries by the 3861 // VM thread or the setting of those entries by the mutator threads on the 3862 // other are quite benign. However, for efficiency it makes sense to keep 3863 // the VM thread from racing with the CMS thread while the latter is 3864 // dirty card info to the modUnionTable. We therefore also use the 3865 // CGC_lock to protect the reading of the card table and the mod union 3866 // table by the CM thread. 3867 // . We run concurrently with mutator updates, so scanning 3868 // needs to be done carefully -- we should not try to scan 3869 // potentially uninitialized objects. 3870 // 3871 // Locking strategy: While holding the CGC_lock, we scan over and 3872 // reset a maximal dirty range of the mod union / card tables, then lock 3873 // the free_list_lock and bitmap lock to do a full marking, then 3874 // release these locks; and repeat the cycle. This allows for a 3875 // certain amount of fairness in the sharing of these locks between 3876 // the CMS collector on the one hand, and the VM thread and the 3877 // mutators on the other. 3878 3879 // NOTE: preclean_mod_union_table() and preclean_card_table() 3880 // further below are largely identical; if you need to modify 3881 // one of these methods, please check the other method too. 3882 3883 size_t CMSCollector::preclean_mod_union_table( 3884 ConcurrentMarkSweepGeneration* old_gen, 3885 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3886 verify_work_stacks_empty(); 3887 verify_overflow_empty(); 3888 3889 // strategy: starting with the first card, accumulate contiguous 3890 // ranges of dirty cards; clear these cards, then scan the region 3891 // covered by these cards. 3892 3893 // Since all of the MUT is committed ahead, we can just use 3894 // that, in case the generations expand while we are precleaning. 3895 // It might also be fine to just use the committed part of the 3896 // generation, but we might potentially miss cards when the 3897 // generation is rapidly expanding while we are in the midst 3898 // of precleaning. 3899 HeapWord* startAddr = old_gen->reserved().start(); 3900 HeapWord* endAddr = old_gen->reserved().end(); 3901 3902 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3903 3904 size_t numDirtyCards, cumNumDirtyCards; 3905 HeapWord *nextAddr, *lastAddr; 3906 for (cumNumDirtyCards = numDirtyCards = 0, 3907 nextAddr = lastAddr = startAddr; 3908 nextAddr < endAddr; 3909 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3910 3911 ResourceMark rm; 3912 HandleMark hm; 3913 3914 MemRegion dirtyRegion; 3915 { 3916 stopTimer(); 3917 // Potential yield point 3918 CMSTokenSync ts(true); 3919 startTimer(); 3920 sample_eden(); 3921 // Get dirty region starting at nextOffset (inclusive), 3922 // simultaneously clearing it. 3923 dirtyRegion = 3924 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3925 assert(dirtyRegion.start() >= nextAddr, 3926 "returned region inconsistent?"); 3927 } 3928 // Remember where the next search should begin. 3929 // The returned region (if non-empty) is a right open interval, 3930 // so lastOffset is obtained from the right end of that 3931 // interval. 3932 lastAddr = dirtyRegion.end(); 3933 // Should do something more transparent and less hacky XXX 3934 numDirtyCards = 3935 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3936 3937 // We'll scan the cards in the dirty region (with periodic 3938 // yields for foreground GC as needed). 3939 if (!dirtyRegion.is_empty()) { 3940 assert(numDirtyCards > 0, "consistency check"); 3941 HeapWord* stop_point = NULL; 3942 stopTimer(); 3943 // Potential yield point 3944 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3945 bitMapLock()); 3946 startTimer(); 3947 { 3948 verify_work_stacks_empty(); 3949 verify_overflow_empty(); 3950 sample_eden(); 3951 stop_point = 3952 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3953 } 3954 if (stop_point != NULL) { 3955 // The careful iteration stopped early either because it found an 3956 // uninitialized object, or because we were in the midst of an 3957 // "abortable preclean", which should now be aborted. Redirty 3958 // the bits corresponding to the partially-scanned or unscanned 3959 // cards. We'll either restart at the next block boundary or 3960 // abort the preclean. 3961 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3962 "Should only be AbortablePreclean."); 3963 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3964 if (should_abort_preclean()) { 3965 break; // out of preclean loop 3966 } else { 3967 // Compute the next address at which preclean should pick up; 3968 // might need bitMapLock in order to read P-bits. 3969 lastAddr = next_card_start_after_block(stop_point); 3970 } 3971 } 3972 } else { 3973 assert(lastAddr == endAddr, "consistency check"); 3974 assert(numDirtyCards == 0, "consistency check"); 3975 break; 3976 } 3977 } 3978 verify_work_stacks_empty(); 3979 verify_overflow_empty(); 3980 return cumNumDirtyCards; 3981 } 3982 3983 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3984 // below are largely identical; if you need to modify 3985 // one of these methods, please check the other method too. 3986 3987 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3988 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3989 // strategy: it's similar to precleamModUnionTable above, in that 3990 // we accumulate contiguous ranges of dirty cards, mark these cards 3991 // precleaned, then scan the region covered by these cards. 3992 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3993 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3994 3995 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3996 3997 size_t numDirtyCards, cumNumDirtyCards; 3998 HeapWord *lastAddr, *nextAddr; 3999 4000 for (cumNumDirtyCards = numDirtyCards = 0, 4001 nextAddr = lastAddr = startAddr; 4002 nextAddr < endAddr; 4003 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4004 4005 ResourceMark rm; 4006 HandleMark hm; 4007 4008 MemRegion dirtyRegion; 4009 { 4010 // See comments in "Precleaning notes" above on why we 4011 // do this locking. XXX Could the locking overheads be 4012 // too high when dirty cards are sparse? [I don't think so.] 4013 stopTimer(); 4014 CMSTokenSync x(true); // is cms thread 4015 startTimer(); 4016 sample_eden(); 4017 // Get and clear dirty region from card table 4018 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4019 MemRegion(nextAddr, endAddr), 4020 true, 4021 CardTableModRefBS::precleaned_card_val()); 4022 4023 assert(dirtyRegion.start() >= nextAddr, 4024 "returned region inconsistent?"); 4025 } 4026 lastAddr = dirtyRegion.end(); 4027 numDirtyCards = 4028 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4029 4030 if (!dirtyRegion.is_empty()) { 4031 stopTimer(); 4032 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4033 startTimer(); 4034 sample_eden(); 4035 verify_work_stacks_empty(); 4036 verify_overflow_empty(); 4037 HeapWord* stop_point = 4038 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4039 if (stop_point != NULL) { 4040 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4041 "Should only be AbortablePreclean."); 4042 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4043 if (should_abort_preclean()) { 4044 break; // out of preclean loop 4045 } else { 4046 // Compute the next address at which preclean should pick up. 4047 lastAddr = next_card_start_after_block(stop_point); 4048 } 4049 } 4050 } else { 4051 break; 4052 } 4053 } 4054 verify_work_stacks_empty(); 4055 verify_overflow_empty(); 4056 return cumNumDirtyCards; 4057 } 4058 4059 class PrecleanKlassClosure : public KlassClosure { 4060 KlassToOopClosure _cm_klass_closure; 4061 public: 4062 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4063 void do_klass(Klass* k) { 4064 if (k->has_accumulated_modified_oops()) { 4065 k->clear_accumulated_modified_oops(); 4066 4067 _cm_klass_closure.do_klass(k); 4068 } 4069 } 4070 }; 4071 4072 // The freelist lock is needed to prevent asserts, is it really needed? 4073 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4074 4075 cl->set_freelistLock(freelistLock); 4076 4077 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4078 4079 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4080 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4081 PrecleanKlassClosure preclean_klass_closure(cl); 4082 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4083 4084 verify_work_stacks_empty(); 4085 verify_overflow_empty(); 4086 } 4087 4088 void CMSCollector::checkpointRootsFinal() { 4089 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4090 check_correct_thread_executing(); 4091 // world is stopped at this checkpoint 4092 assert(SafepointSynchronize::is_at_safepoint(), 4093 "world should be stopped"); 4094 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4095 4096 verify_work_stacks_empty(); 4097 verify_overflow_empty(); 4098 4099 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4100 _young_gen->used() / K, _young_gen->capacity() / K); 4101 { 4102 if (CMSScavengeBeforeRemark) { 4103 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4104 // Temporarily set flag to false, GCH->do_collection will 4105 // expect it to be false and set to true 4106 FlagSetting fl(gch->_is_gc_active, false); 4107 4108 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4109 4110 gch->do_collection(true, // full (i.e. force, see below) 4111 false, // !clear_all_soft_refs 4112 0, // size 4113 false, // is_tlab 4114 GenCollectedHeap::YoungGen // type 4115 ); 4116 } 4117 FreelistLocker x(this); 4118 MutexLockerEx y(bitMapLock(), 4119 Mutex::_no_safepoint_check_flag); 4120 checkpointRootsFinalWork(); 4121 } 4122 verify_work_stacks_empty(); 4123 verify_overflow_empty(); 4124 } 4125 4126 void CMSCollector::checkpointRootsFinalWork() { 4127 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4128 4129 assert(haveFreelistLocks(), "must have free list locks"); 4130 assert_lock_strong(bitMapLock()); 4131 4132 ResourceMark rm; 4133 HandleMark hm; 4134 4135 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4136 4137 if (should_unload_classes()) { 4138 CodeCache::gc_prologue(); 4139 } 4140 assert(haveFreelistLocks(), "must have free list locks"); 4141 assert_lock_strong(bitMapLock()); 4142 4143 // We might assume that we need not fill TLAB's when 4144 // CMSScavengeBeforeRemark is set, because we may have just done 4145 // a scavenge which would have filled all TLAB's -- and besides 4146 // Eden would be empty. This however may not always be the case -- 4147 // for instance although we asked for a scavenge, it may not have 4148 // happened because of a JNI critical section. We probably need 4149 // a policy for deciding whether we can in that case wait until 4150 // the critical section releases and then do the remark following 4151 // the scavenge, and skip it here. In the absence of that policy, 4152 // or of an indication of whether the scavenge did indeed occur, 4153 // we cannot rely on TLAB's having been filled and must do 4154 // so here just in case a scavenge did not happen. 4155 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4156 // Update the saved marks which may affect the root scans. 4157 gch->save_marks(); 4158 4159 print_eden_and_survivor_chunk_arrays(); 4160 4161 { 4162 #if defined(COMPILER2) || INCLUDE_JVMCI 4163 DerivedPointerTableDeactivate dpt_deact; 4164 #endif 4165 4166 // Note on the role of the mod union table: 4167 // Since the marker in "markFromRoots" marks concurrently with 4168 // mutators, it is possible for some reachable objects not to have been 4169 // scanned. For instance, an only reference to an object A was 4170 // placed in object B after the marker scanned B. Unless B is rescanned, 4171 // A would be collected. Such updates to references in marked objects 4172 // are detected via the mod union table which is the set of all cards 4173 // dirtied since the first checkpoint in this GC cycle and prior to 4174 // the most recent young generation GC, minus those cleaned up by the 4175 // concurrent precleaning. 4176 if (CMSParallelRemarkEnabled) { 4177 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4178 do_remark_parallel(); 4179 } else { 4180 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4181 do_remark_non_parallel(); 4182 } 4183 } 4184 verify_work_stacks_empty(); 4185 verify_overflow_empty(); 4186 4187 { 4188 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4189 refProcessingWork(); 4190 } 4191 verify_work_stacks_empty(); 4192 verify_overflow_empty(); 4193 4194 if (should_unload_classes()) { 4195 CodeCache::gc_epilogue(); 4196 } 4197 JvmtiExport::gc_epilogue(); 4198 4199 // If we encountered any (marking stack / work queue) overflow 4200 // events during the current CMS cycle, take appropriate 4201 // remedial measures, where possible, so as to try and avoid 4202 // recurrence of that condition. 4203 assert(_markStack.isEmpty(), "No grey objects"); 4204 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4205 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4206 if (ser_ovflw > 0) { 4207 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4208 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4209 _markStack.expand(); 4210 _ser_pmc_remark_ovflw = 0; 4211 _ser_pmc_preclean_ovflw = 0; 4212 _ser_kac_preclean_ovflw = 0; 4213 _ser_kac_ovflw = 0; 4214 } 4215 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4216 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4217 _par_pmc_remark_ovflw, _par_kac_ovflw); 4218 _par_pmc_remark_ovflw = 0; 4219 _par_kac_ovflw = 0; 4220 } 4221 if (_markStack._hit_limit > 0) { 4222 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4223 _markStack._hit_limit); 4224 } 4225 if (_markStack._failed_double > 0) { 4226 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4227 _markStack._failed_double, _markStack.capacity()); 4228 } 4229 _markStack._hit_limit = 0; 4230 _markStack._failed_double = 0; 4231 4232 if ((VerifyAfterGC || VerifyDuringGC) && 4233 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4234 verify_after_remark(); 4235 } 4236 4237 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4238 4239 // Change under the freelistLocks. 4240 _collectorState = Sweeping; 4241 // Call isAllClear() under bitMapLock 4242 assert(_modUnionTable.isAllClear(), 4243 "Should be clear by end of the final marking"); 4244 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4245 "Should be clear by end of the final marking"); 4246 } 4247 4248 void CMSParInitialMarkTask::work(uint worker_id) { 4249 elapsedTimer _timer; 4250 ResourceMark rm; 4251 HandleMark hm; 4252 4253 // ---------- scan from roots -------------- 4254 _timer.start(); 4255 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4256 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4257 4258 // ---------- young gen roots -------------- 4259 { 4260 work_on_young_gen_roots(worker_id, &par_mri_cl); 4261 _timer.stop(); 4262 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4263 } 4264 4265 // ---------- remaining roots -------------- 4266 _timer.reset(); 4267 _timer.start(); 4268 4269 CLDToOopClosure cld_closure(&par_mri_cl, true); 4270 4271 gch->gen_process_roots(_strong_roots_scope, 4272 GenCollectedHeap::OldGen, 4273 false, // yg was scanned above 4274 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4275 _collector->should_unload_classes(), 4276 &par_mri_cl, 4277 NULL, 4278 &cld_closure); 4279 assert(_collector->should_unload_classes() 4280 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4281 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4282 _timer.stop(); 4283 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4284 } 4285 4286 // Parallel remark task 4287 class CMSParRemarkTask: public CMSParMarkTask { 4288 CompactibleFreeListSpace* _cms_space; 4289 4290 // The per-thread work queues, available here for stealing. 4291 OopTaskQueueSet* _task_queues; 4292 ParallelTaskTerminator _term; 4293 StrongRootsScope* _strong_roots_scope; 4294 4295 public: 4296 // A value of 0 passed to n_workers will cause the number of 4297 // workers to be taken from the active workers in the work gang. 4298 CMSParRemarkTask(CMSCollector* collector, 4299 CompactibleFreeListSpace* cms_space, 4300 uint n_workers, WorkGang* workers, 4301 OopTaskQueueSet* task_queues, 4302 StrongRootsScope* strong_roots_scope): 4303 CMSParMarkTask("Rescan roots and grey objects in parallel", 4304 collector, n_workers), 4305 _cms_space(cms_space), 4306 _task_queues(task_queues), 4307 _term(n_workers, task_queues), 4308 _strong_roots_scope(strong_roots_scope) { } 4309 4310 OopTaskQueueSet* task_queues() { return _task_queues; } 4311 4312 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4313 4314 ParallelTaskTerminator* terminator() { return &_term; } 4315 uint n_workers() { return _n_workers; } 4316 4317 void work(uint worker_id); 4318 4319 private: 4320 // ... of dirty cards in old space 4321 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4322 ParMarkRefsIntoAndScanClosure* cl); 4323 4324 // ... work stealing for the above 4325 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4326 }; 4327 4328 class RemarkKlassClosure : public KlassClosure { 4329 KlassToOopClosure _cm_klass_closure; 4330 public: 4331 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4332 void do_klass(Klass* k) { 4333 // Check if we have modified any oops in the Klass during the concurrent marking. 4334 if (k->has_accumulated_modified_oops()) { 4335 k->clear_accumulated_modified_oops(); 4336 4337 // We could have transfered the current modified marks to the accumulated marks, 4338 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4339 } else if (k->has_modified_oops()) { 4340 // Don't clear anything, this info is needed by the next young collection. 4341 } else { 4342 // No modified oops in the Klass. 4343 return; 4344 } 4345 4346 // The klass has modified fields, need to scan the klass. 4347 _cm_klass_closure.do_klass(k); 4348 } 4349 }; 4350 4351 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4352 ParNewGeneration* young_gen = _collector->_young_gen; 4353 ContiguousSpace* eden_space = young_gen->eden(); 4354 ContiguousSpace* from_space = young_gen->from(); 4355 ContiguousSpace* to_space = young_gen->to(); 4356 4357 HeapWord** eca = _collector->_eden_chunk_array; 4358 size_t ect = _collector->_eden_chunk_index; 4359 HeapWord** sca = _collector->_survivor_chunk_array; 4360 size_t sct = _collector->_survivor_chunk_index; 4361 4362 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4363 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4364 4365 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4366 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4367 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4368 } 4369 4370 // work_queue(i) is passed to the closure 4371 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4372 // also is passed to do_dirty_card_rescan_tasks() and to 4373 // do_work_steal() to select the i-th task_queue. 4374 4375 void CMSParRemarkTask::work(uint worker_id) { 4376 elapsedTimer _timer; 4377 ResourceMark rm; 4378 HandleMark hm; 4379 4380 // ---------- rescan from roots -------------- 4381 _timer.start(); 4382 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4383 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4384 _collector->_span, _collector->ref_processor(), 4385 &(_collector->_markBitMap), 4386 work_queue(worker_id)); 4387 4388 // Rescan young gen roots first since these are likely 4389 // coarsely partitioned and may, on that account, constitute 4390 // the critical path; thus, it's best to start off that 4391 // work first. 4392 // ---------- young gen roots -------------- 4393 { 4394 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4395 _timer.stop(); 4396 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4397 } 4398 4399 // ---------- remaining roots -------------- 4400 _timer.reset(); 4401 _timer.start(); 4402 gch->gen_process_roots(_strong_roots_scope, 4403 GenCollectedHeap::OldGen, 4404 false, // yg was scanned above 4405 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4406 _collector->should_unload_classes(), 4407 &par_mrias_cl, 4408 NULL, 4409 NULL); // The dirty klasses will be handled below 4410 4411 assert(_collector->should_unload_classes() 4412 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4413 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4414 _timer.stop(); 4415 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4416 4417 // ---------- unhandled CLD scanning ---------- 4418 if (worker_id == 0) { // Single threaded at the moment. 4419 _timer.reset(); 4420 _timer.start(); 4421 4422 // Scan all new class loader data objects and new dependencies that were 4423 // introduced during concurrent marking. 4424 ResourceMark rm; 4425 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4426 for (int i = 0; i < array->length(); i++) { 4427 par_mrias_cl.do_cld_nv(array->at(i)); 4428 } 4429 4430 // We don't need to keep track of new CLDs anymore. 4431 ClassLoaderDataGraph::remember_new_clds(false); 4432 4433 _timer.stop(); 4434 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4435 } 4436 4437 // ---------- dirty klass scanning ---------- 4438 if (worker_id == 0) { // Single threaded at the moment. 4439 _timer.reset(); 4440 _timer.start(); 4441 4442 // Scan all classes that was dirtied during the concurrent marking phase. 4443 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4444 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4445 4446 _timer.stop(); 4447 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4448 } 4449 4450 // We might have added oops to ClassLoaderData::_handles during the 4451 // concurrent marking phase. These oops point to newly allocated objects 4452 // that are guaranteed to be kept alive. Either by the direct allocation 4453 // code, or when the young collector processes the roots. Hence, 4454 // we don't have to revisit the _handles block during the remark phase. 4455 4456 // ---------- rescan dirty cards ------------ 4457 _timer.reset(); 4458 _timer.start(); 4459 4460 // Do the rescan tasks for each of the two spaces 4461 // (cms_space) in turn. 4462 // "worker_id" is passed to select the task_queue for "worker_id" 4463 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4464 _timer.stop(); 4465 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4466 4467 // ---------- steal work from other threads ... 4468 // ---------- ... and drain overflow list. 4469 _timer.reset(); 4470 _timer.start(); 4471 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4472 _timer.stop(); 4473 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4474 } 4475 4476 // Note that parameter "i" is not used. 4477 void 4478 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4479 OopsInGenClosure* cl, ContiguousSpace* space, 4480 HeapWord** chunk_array, size_t chunk_top) { 4481 // Until all tasks completed: 4482 // . claim an unclaimed task 4483 // . compute region boundaries corresponding to task claimed 4484 // using chunk_array 4485 // . par_oop_iterate(cl) over that region 4486 4487 ResourceMark rm; 4488 HandleMark hm; 4489 4490 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4491 4492 uint nth_task = 0; 4493 uint n_tasks = pst->n_tasks(); 4494 4495 if (n_tasks > 0) { 4496 assert(pst->valid(), "Uninitialized use?"); 4497 HeapWord *start, *end; 4498 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4499 // We claimed task # nth_task; compute its boundaries. 4500 if (chunk_top == 0) { // no samples were taken 4501 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4502 start = space->bottom(); 4503 end = space->top(); 4504 } else if (nth_task == 0) { 4505 start = space->bottom(); 4506 end = chunk_array[nth_task]; 4507 } else if (nth_task < (uint)chunk_top) { 4508 assert(nth_task >= 1, "Control point invariant"); 4509 start = chunk_array[nth_task - 1]; 4510 end = chunk_array[nth_task]; 4511 } else { 4512 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4513 start = chunk_array[chunk_top - 1]; 4514 end = space->top(); 4515 } 4516 MemRegion mr(start, end); 4517 // Verify that mr is in space 4518 assert(mr.is_empty() || space->used_region().contains(mr), 4519 "Should be in space"); 4520 // Verify that "start" is an object boundary 4521 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4522 "Should be an oop"); 4523 space->par_oop_iterate(mr, cl); 4524 } 4525 pst->all_tasks_completed(); 4526 } 4527 } 4528 4529 void 4530 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4531 CompactibleFreeListSpace* sp, int i, 4532 ParMarkRefsIntoAndScanClosure* cl) { 4533 // Until all tasks completed: 4534 // . claim an unclaimed task 4535 // . compute region boundaries corresponding to task claimed 4536 // . transfer dirty bits ct->mut for that region 4537 // . apply rescanclosure to dirty mut bits for that region 4538 4539 ResourceMark rm; 4540 HandleMark hm; 4541 4542 OopTaskQueue* work_q = work_queue(i); 4543 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4544 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4545 // CAUTION: This closure has state that persists across calls to 4546 // the work method dirty_range_iterate_clear() in that it has 4547 // embedded in it a (subtype of) UpwardsObjectClosure. The 4548 // use of that state in the embedded UpwardsObjectClosure instance 4549 // assumes that the cards are always iterated (even if in parallel 4550 // by several threads) in monotonically increasing order per each 4551 // thread. This is true of the implementation below which picks 4552 // card ranges (chunks) in monotonically increasing order globally 4553 // and, a-fortiori, in monotonically increasing order per thread 4554 // (the latter order being a subsequence of the former). 4555 // If the work code below is ever reorganized into a more chaotic 4556 // work-partitioning form than the current "sequential tasks" 4557 // paradigm, the use of that persistent state will have to be 4558 // revisited and modified appropriately. See also related 4559 // bug 4756801 work on which should examine this code to make 4560 // sure that the changes there do not run counter to the 4561 // assumptions made here and necessary for correctness and 4562 // efficiency. Note also that this code might yield inefficient 4563 // behavior in the case of very large objects that span one or 4564 // more work chunks. Such objects would potentially be scanned 4565 // several times redundantly. Work on 4756801 should try and 4566 // address that performance anomaly if at all possible. XXX 4567 MemRegion full_span = _collector->_span; 4568 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4569 MarkFromDirtyCardsClosure 4570 greyRescanClosure(_collector, full_span, // entire span of interest 4571 sp, bm, work_q, cl); 4572 4573 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4574 assert(pst->valid(), "Uninitialized use?"); 4575 uint nth_task = 0; 4576 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4577 MemRegion span = sp->used_region(); 4578 HeapWord* start_addr = span.start(); 4579 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4580 alignment); 4581 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4582 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4583 start_addr, "Check alignment"); 4584 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4585 chunk_size, "Check alignment"); 4586 4587 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4588 // Having claimed the nth_task, compute corresponding mem-region, 4589 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4590 // The alignment restriction ensures that we do not need any 4591 // synchronization with other gang-workers while setting or 4592 // clearing bits in thus chunk of the MUT. 4593 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4594 start_addr + (nth_task+1)*chunk_size); 4595 // The last chunk's end might be way beyond end of the 4596 // used region. In that case pull back appropriately. 4597 if (this_span.end() > end_addr) { 4598 this_span.set_end(end_addr); 4599 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4600 } 4601 // Iterate over the dirty cards covering this chunk, marking them 4602 // precleaned, and setting the corresponding bits in the mod union 4603 // table. Since we have been careful to partition at Card and MUT-word 4604 // boundaries no synchronization is needed between parallel threads. 4605 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4606 &modUnionClosure); 4607 4608 // Having transferred these marks into the modUnionTable, 4609 // rescan the marked objects on the dirty cards in the modUnionTable. 4610 // Even if this is at a synchronous collection, the initial marking 4611 // may have been done during an asynchronous collection so there 4612 // may be dirty bits in the mod-union table. 4613 _collector->_modUnionTable.dirty_range_iterate_clear( 4614 this_span, &greyRescanClosure); 4615 _collector->_modUnionTable.verifyNoOneBitsInRange( 4616 this_span.start(), 4617 this_span.end()); 4618 } 4619 pst->all_tasks_completed(); // declare that i am done 4620 } 4621 4622 // . see if we can share work_queues with ParNew? XXX 4623 void 4624 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4625 int* seed) { 4626 OopTaskQueue* work_q = work_queue(i); 4627 NOT_PRODUCT(int num_steals = 0;) 4628 oop obj_to_scan; 4629 CMSBitMap* bm = &(_collector->_markBitMap); 4630 4631 while (true) { 4632 // Completely finish any left over work from (an) earlier round(s) 4633 cl->trim_queue(0); 4634 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4635 (size_t)ParGCDesiredObjsFromOverflowList); 4636 // Now check if there's any work in the overflow list 4637 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4638 // only affects the number of attempts made to get work from the 4639 // overflow list and does not affect the number of workers. Just 4640 // pass ParallelGCThreads so this behavior is unchanged. 4641 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4642 work_q, 4643 ParallelGCThreads)) { 4644 // found something in global overflow list; 4645 // not yet ready to go stealing work from others. 4646 // We'd like to assert(work_q->size() != 0, ...) 4647 // because we just took work from the overflow list, 4648 // but of course we can't since all of that could have 4649 // been already stolen from us. 4650 // "He giveth and He taketh away." 4651 continue; 4652 } 4653 // Verify that we have no work before we resort to stealing 4654 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4655 // Try to steal from other queues that have work 4656 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4657 NOT_PRODUCT(num_steals++;) 4658 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4659 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4660 // Do scanning work 4661 obj_to_scan->oop_iterate(cl); 4662 // Loop around, finish this work, and try to steal some more 4663 } else if (terminator()->offer_termination()) { 4664 break; // nirvana from the infinite cycle 4665 } 4666 } 4667 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4668 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4669 "Else our work is not yet done"); 4670 } 4671 4672 // Record object boundaries in _eden_chunk_array by sampling the eden 4673 // top in the slow-path eden object allocation code path and record 4674 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4675 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4676 // sampling in sample_eden() that activates during the part of the 4677 // preclean phase. 4678 void CMSCollector::sample_eden_chunk() { 4679 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4680 if (_eden_chunk_lock->try_lock()) { 4681 // Record a sample. This is the critical section. The contents 4682 // of the _eden_chunk_array have to be non-decreasing in the 4683 // address order. 4684 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4685 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4686 "Unexpected state of Eden"); 4687 if (_eden_chunk_index == 0 || 4688 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4689 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4690 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4691 _eden_chunk_index++; // commit sample 4692 } 4693 _eden_chunk_lock->unlock(); 4694 } 4695 } 4696 } 4697 4698 // Return a thread-local PLAB recording array, as appropriate. 4699 void* CMSCollector::get_data_recorder(int thr_num) { 4700 if (_survivor_plab_array != NULL && 4701 (CMSPLABRecordAlways || 4702 (_collectorState > Marking && _collectorState < FinalMarking))) { 4703 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4704 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4705 ca->reset(); // clear it so that fresh data is recorded 4706 return (void*) ca; 4707 } else { 4708 return NULL; 4709 } 4710 } 4711 4712 // Reset all the thread-local PLAB recording arrays 4713 void CMSCollector::reset_survivor_plab_arrays() { 4714 for (uint i = 0; i < ParallelGCThreads; i++) { 4715 _survivor_plab_array[i].reset(); 4716 } 4717 } 4718 4719 // Merge the per-thread plab arrays into the global survivor chunk 4720 // array which will provide the partitioning of the survivor space 4721 // for CMS initial scan and rescan. 4722 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4723 int no_of_gc_threads) { 4724 assert(_survivor_plab_array != NULL, "Error"); 4725 assert(_survivor_chunk_array != NULL, "Error"); 4726 assert(_collectorState == FinalMarking || 4727 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4728 for (int j = 0; j < no_of_gc_threads; j++) { 4729 _cursor[j] = 0; 4730 } 4731 HeapWord* top = surv->top(); 4732 size_t i; 4733 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4734 HeapWord* min_val = top; // Higher than any PLAB address 4735 uint min_tid = 0; // position of min_val this round 4736 for (int j = 0; j < no_of_gc_threads; j++) { 4737 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4738 if (_cursor[j] == cur_sca->end()) { 4739 continue; 4740 } 4741 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4742 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4743 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4744 if (cur_val < min_val) { 4745 min_tid = j; 4746 min_val = cur_val; 4747 } else { 4748 assert(cur_val < top, "All recorded addresses should be less"); 4749 } 4750 } 4751 // At this point min_val and min_tid are respectively 4752 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4753 // and the thread (j) that witnesses that address. 4754 // We record this address in the _survivor_chunk_array[i] 4755 // and increment _cursor[min_tid] prior to the next round i. 4756 if (min_val == top) { 4757 break; 4758 } 4759 _survivor_chunk_array[i] = min_val; 4760 _cursor[min_tid]++; 4761 } 4762 // We are all done; record the size of the _survivor_chunk_array 4763 _survivor_chunk_index = i; // exclusive: [0, i) 4764 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4765 // Verify that we used up all the recorded entries 4766 #ifdef ASSERT 4767 size_t total = 0; 4768 for (int j = 0; j < no_of_gc_threads; j++) { 4769 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4770 total += _cursor[j]; 4771 } 4772 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4773 // Check that the merged array is in sorted order 4774 if (total > 0) { 4775 for (size_t i = 0; i < total - 1; i++) { 4776 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4777 i, p2i(_survivor_chunk_array[i])); 4778 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4779 "Not sorted"); 4780 } 4781 } 4782 #endif // ASSERT 4783 } 4784 4785 // Set up the space's par_seq_tasks structure for work claiming 4786 // for parallel initial scan and rescan of young gen. 4787 // See ParRescanTask where this is currently used. 4788 void 4789 CMSCollector:: 4790 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4791 assert(n_threads > 0, "Unexpected n_threads argument"); 4792 4793 // Eden space 4794 if (!_young_gen->eden()->is_empty()) { 4795 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4796 assert(!pst->valid(), "Clobbering existing data?"); 4797 // Each valid entry in [0, _eden_chunk_index) represents a task. 4798 size_t n_tasks = _eden_chunk_index + 1; 4799 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4800 // Sets the condition for completion of the subtask (how many threads 4801 // need to finish in order to be done). 4802 pst->set_n_threads(n_threads); 4803 pst->set_n_tasks((int)n_tasks); 4804 } 4805 4806 // Merge the survivor plab arrays into _survivor_chunk_array 4807 if (_survivor_plab_array != NULL) { 4808 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4809 } else { 4810 assert(_survivor_chunk_index == 0, "Error"); 4811 } 4812 4813 // To space 4814 { 4815 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4816 assert(!pst->valid(), "Clobbering existing data?"); 4817 // Sets the condition for completion of the subtask (how many threads 4818 // need to finish in order to be done). 4819 pst->set_n_threads(n_threads); 4820 pst->set_n_tasks(1); 4821 assert(pst->valid(), "Error"); 4822 } 4823 4824 // From space 4825 { 4826 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4827 assert(!pst->valid(), "Clobbering existing data?"); 4828 size_t n_tasks = _survivor_chunk_index + 1; 4829 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4830 // Sets the condition for completion of the subtask (how many threads 4831 // need to finish in order to be done). 4832 pst->set_n_threads(n_threads); 4833 pst->set_n_tasks((int)n_tasks); 4834 assert(pst->valid(), "Error"); 4835 } 4836 } 4837 4838 // Parallel version of remark 4839 void CMSCollector::do_remark_parallel() { 4840 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4841 WorkGang* workers = gch->workers(); 4842 assert(workers != NULL, "Need parallel worker threads."); 4843 // Choose to use the number of GC workers most recently set 4844 // into "active_workers". 4845 uint n_workers = workers->active_workers(); 4846 4847 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4848 4849 StrongRootsScope srs(n_workers); 4850 4851 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4852 4853 // We won't be iterating over the cards in the card table updating 4854 // the younger_gen cards, so we shouldn't call the following else 4855 // the verification code as well as subsequent younger_refs_iterate 4856 // code would get confused. XXX 4857 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4858 4859 // The young gen rescan work will not be done as part of 4860 // process_roots (which currently doesn't know how to 4861 // parallelize such a scan), but rather will be broken up into 4862 // a set of parallel tasks (via the sampling that the [abortable] 4863 // preclean phase did of eden, plus the [two] tasks of 4864 // scanning the [two] survivor spaces. Further fine-grain 4865 // parallelization of the scanning of the survivor spaces 4866 // themselves, and of precleaning of the young gen itself 4867 // is deferred to the future. 4868 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4869 4870 // The dirty card rescan work is broken up into a "sequence" 4871 // of parallel tasks (per constituent space) that are dynamically 4872 // claimed by the parallel threads. 4873 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4874 4875 // It turns out that even when we're using 1 thread, doing the work in a 4876 // separate thread causes wide variance in run times. We can't help this 4877 // in the multi-threaded case, but we special-case n=1 here to get 4878 // repeatable measurements of the 1-thread overhead of the parallel code. 4879 if (n_workers > 1) { 4880 // Make refs discovery MT-safe, if it isn't already: it may not 4881 // necessarily be so, since it's possible that we are doing 4882 // ST marking. 4883 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4884 workers->run_task(&tsk); 4885 } else { 4886 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4887 tsk.work(0); 4888 } 4889 4890 // restore, single-threaded for now, any preserved marks 4891 // as a result of work_q overflow 4892 restore_preserved_marks_if_any(); 4893 } 4894 4895 // Non-parallel version of remark 4896 void CMSCollector::do_remark_non_parallel() { 4897 ResourceMark rm; 4898 HandleMark hm; 4899 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4900 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4901 4902 MarkRefsIntoAndScanClosure 4903 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4904 &_markStack, this, 4905 false /* should_yield */, false /* not precleaning */); 4906 MarkFromDirtyCardsClosure 4907 markFromDirtyCardsClosure(this, _span, 4908 NULL, // space is set further below 4909 &_markBitMap, &_markStack, &mrias_cl); 4910 { 4911 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4912 // Iterate over the dirty cards, setting the corresponding bits in the 4913 // mod union table. 4914 { 4915 ModUnionClosure modUnionClosure(&_modUnionTable); 4916 _ct->ct_bs()->dirty_card_iterate( 4917 _cmsGen->used_region(), 4918 &modUnionClosure); 4919 } 4920 // Having transferred these marks into the modUnionTable, we just need 4921 // to rescan the marked objects on the dirty cards in the modUnionTable. 4922 // The initial marking may have been done during an asynchronous 4923 // collection so there may be dirty bits in the mod-union table. 4924 const int alignment = 4925 CardTableModRefBS::card_size * BitsPerWord; 4926 { 4927 // ... First handle dirty cards in CMS gen 4928 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4929 MemRegion ur = _cmsGen->used_region(); 4930 HeapWord* lb = ur.start(); 4931 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4932 MemRegion cms_span(lb, ub); 4933 _modUnionTable.dirty_range_iterate_clear(cms_span, 4934 &markFromDirtyCardsClosure); 4935 verify_work_stacks_empty(); 4936 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4937 } 4938 } 4939 if (VerifyDuringGC && 4940 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4941 HandleMark hm; // Discard invalid handles created during verification 4942 Universe::verify(); 4943 } 4944 { 4945 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4946 4947 verify_work_stacks_empty(); 4948 4949 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4950 StrongRootsScope srs(1); 4951 4952 gch->gen_process_roots(&srs, 4953 GenCollectedHeap::OldGen, 4954 true, // young gen as roots 4955 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4956 should_unload_classes(), 4957 &mrias_cl, 4958 NULL, 4959 NULL); // The dirty klasses will be handled below 4960 4961 assert(should_unload_classes() 4962 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4963 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4964 } 4965 4966 { 4967 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4968 4969 verify_work_stacks_empty(); 4970 4971 // Scan all class loader data objects that might have been introduced 4972 // during concurrent marking. 4973 ResourceMark rm; 4974 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4975 for (int i = 0; i < array->length(); i++) { 4976 mrias_cl.do_cld_nv(array->at(i)); 4977 } 4978 4979 // We don't need to keep track of new CLDs anymore. 4980 ClassLoaderDataGraph::remember_new_clds(false); 4981 4982 verify_work_stacks_empty(); 4983 } 4984 4985 { 4986 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4987 4988 verify_work_stacks_empty(); 4989 4990 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4991 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4992 4993 verify_work_stacks_empty(); 4994 } 4995 4996 // We might have added oops to ClassLoaderData::_handles during the 4997 // concurrent marking phase. These oops point to newly allocated objects 4998 // that are guaranteed to be kept alive. Either by the direct allocation 4999 // code, or when the young collector processes the roots. Hence, 5000 // we don't have to revisit the _handles block during the remark phase. 5001 5002 verify_work_stacks_empty(); 5003 // Restore evacuated mark words, if any, used for overflow list links 5004 restore_preserved_marks_if_any(); 5005 5006 verify_overflow_empty(); 5007 } 5008 5009 //////////////////////////////////////////////////////// 5010 // Parallel Reference Processing Task Proxy Class 5011 //////////////////////////////////////////////////////// 5012 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5013 OopTaskQueueSet* _queues; 5014 ParallelTaskTerminator _terminator; 5015 public: 5016 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5017 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5018 ParallelTaskTerminator* terminator() { return &_terminator; } 5019 OopTaskQueueSet* queues() { return _queues; } 5020 }; 5021 5022 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5023 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5024 CMSCollector* _collector; 5025 CMSBitMap* _mark_bit_map; 5026 const MemRegion _span; 5027 ProcessTask& _task; 5028 5029 public: 5030 CMSRefProcTaskProxy(ProcessTask& task, 5031 CMSCollector* collector, 5032 const MemRegion& span, 5033 CMSBitMap* mark_bit_map, 5034 AbstractWorkGang* workers, 5035 OopTaskQueueSet* task_queues): 5036 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5037 task_queues, 5038 workers->active_workers()), 5039 _task(task), 5040 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5041 { 5042 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5043 "Inconsistency in _span"); 5044 } 5045 5046 OopTaskQueueSet* task_queues() { return queues(); } 5047 5048 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5049 5050 void do_work_steal(int i, 5051 CMSParDrainMarkingStackClosure* drain, 5052 CMSParKeepAliveClosure* keep_alive, 5053 int* seed); 5054 5055 virtual void work(uint worker_id); 5056 }; 5057 5058 void CMSRefProcTaskProxy::work(uint worker_id) { 5059 ResourceMark rm; 5060 HandleMark hm; 5061 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5062 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5063 _mark_bit_map, 5064 work_queue(worker_id)); 5065 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5066 _mark_bit_map, 5067 work_queue(worker_id)); 5068 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5069 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5070 if (_task.marks_oops_alive()) { 5071 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5072 _collector->hash_seed(worker_id)); 5073 } 5074 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5075 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5076 } 5077 5078 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5079 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5080 EnqueueTask& _task; 5081 5082 public: 5083 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5084 : AbstractGangTask("Enqueue reference objects in parallel"), 5085 _task(task) 5086 { } 5087 5088 virtual void work(uint worker_id) 5089 { 5090 _task.work(worker_id); 5091 } 5092 }; 5093 5094 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5095 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5096 _span(span), 5097 _bit_map(bit_map), 5098 _work_queue(work_queue), 5099 _mark_and_push(collector, span, bit_map, work_queue), 5100 _low_water_mark(MIN2((work_queue->max_elems()/4), 5101 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5102 { } 5103 5104 // . see if we can share work_queues with ParNew? XXX 5105 void CMSRefProcTaskProxy::do_work_steal(int i, 5106 CMSParDrainMarkingStackClosure* drain, 5107 CMSParKeepAliveClosure* keep_alive, 5108 int* seed) { 5109 OopTaskQueue* work_q = work_queue(i); 5110 NOT_PRODUCT(int num_steals = 0;) 5111 oop obj_to_scan; 5112 5113 while (true) { 5114 // Completely finish any left over work from (an) earlier round(s) 5115 drain->trim_queue(0); 5116 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5117 (size_t)ParGCDesiredObjsFromOverflowList); 5118 // Now check if there's any work in the overflow list 5119 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5120 // only affects the number of attempts made to get work from the 5121 // overflow list and does not affect the number of workers. Just 5122 // pass ParallelGCThreads so this behavior is unchanged. 5123 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5124 work_q, 5125 ParallelGCThreads)) { 5126 // Found something in global overflow list; 5127 // not yet ready to go stealing work from others. 5128 // We'd like to assert(work_q->size() != 0, ...) 5129 // because we just took work from the overflow list, 5130 // but of course we can't, since all of that might have 5131 // been already stolen from us. 5132 continue; 5133 } 5134 // Verify that we have no work before we resort to stealing 5135 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5136 // Try to steal from other queues that have work 5137 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5138 NOT_PRODUCT(num_steals++;) 5139 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5140 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5141 // Do scanning work 5142 obj_to_scan->oop_iterate(keep_alive); 5143 // Loop around, finish this work, and try to steal some more 5144 } else if (terminator()->offer_termination()) { 5145 break; // nirvana from the infinite cycle 5146 } 5147 } 5148 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5149 } 5150 5151 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5152 { 5153 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5154 WorkGang* workers = gch->workers(); 5155 assert(workers != NULL, "Need parallel worker threads."); 5156 CMSRefProcTaskProxy rp_task(task, &_collector, 5157 _collector.ref_processor()->span(), 5158 _collector.markBitMap(), 5159 workers, _collector.task_queues()); 5160 workers->run_task(&rp_task); 5161 } 5162 5163 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5164 { 5165 5166 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5167 WorkGang* workers = gch->workers(); 5168 assert(workers != NULL, "Need parallel worker threads."); 5169 CMSRefEnqueueTaskProxy enq_task(task); 5170 workers->run_task(&enq_task); 5171 } 5172 5173 void CMSCollector::refProcessingWork() { 5174 ResourceMark rm; 5175 HandleMark hm; 5176 5177 ReferenceProcessor* rp = ref_processor(); 5178 assert(rp->span().equals(_span), "Spans should be equal"); 5179 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5180 // Process weak references. 5181 rp->setup_policy(false); 5182 verify_work_stacks_empty(); 5183 5184 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5185 &_markStack, false /* !preclean */); 5186 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5187 _span, &_markBitMap, &_markStack, 5188 &cmsKeepAliveClosure, false /* !preclean */); 5189 { 5190 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5191 5192 ReferenceProcessorStats stats; 5193 if (rp->processing_is_mt()) { 5194 // Set the degree of MT here. If the discovery is done MT, there 5195 // may have been a different number of threads doing the discovery 5196 // and a different number of discovered lists may have Ref objects. 5197 // That is OK as long as the Reference lists are balanced (see 5198 // balance_all_queues() and balance_queues()). 5199 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5200 uint active_workers = ParallelGCThreads; 5201 WorkGang* workers = gch->workers(); 5202 if (workers != NULL) { 5203 active_workers = workers->active_workers(); 5204 // The expectation is that active_workers will have already 5205 // been set to a reasonable value. If it has not been set, 5206 // investigate. 5207 assert(active_workers > 0, "Should have been set during scavenge"); 5208 } 5209 rp->set_active_mt_degree(active_workers); 5210 CMSRefProcTaskExecutor task_executor(*this); 5211 stats = rp->process_discovered_references(&_is_alive_closure, 5212 &cmsKeepAliveClosure, 5213 &cmsDrainMarkingStackClosure, 5214 &task_executor, 5215 _gc_timer_cm); 5216 } else { 5217 stats = rp->process_discovered_references(&_is_alive_closure, 5218 &cmsKeepAliveClosure, 5219 &cmsDrainMarkingStackClosure, 5220 NULL, 5221 _gc_timer_cm); 5222 } 5223 _gc_tracer_cm->report_gc_reference_stats(stats); 5224 5225 } 5226 5227 // This is the point where the entire marking should have completed. 5228 verify_work_stacks_empty(); 5229 5230 if (should_unload_classes()) { 5231 { 5232 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5233 5234 // Unload classes and purge the SystemDictionary. 5235 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5236 5237 // Unload nmethods. 5238 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5239 5240 // Prune dead klasses from subklass/sibling/implementor lists. 5241 Klass::clean_weak_klass_links(&_is_alive_closure); 5242 } 5243 5244 { 5245 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5246 // Clean up unreferenced symbols in symbol table. 5247 SymbolTable::unlink(); 5248 } 5249 5250 { 5251 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5252 // Delete entries for dead interned strings. 5253 StringTable::unlink(&_is_alive_closure); 5254 } 5255 } 5256 5257 5258 // Restore any preserved marks as a result of mark stack or 5259 // work queue overflow 5260 restore_preserved_marks_if_any(); // done single-threaded for now 5261 5262 rp->set_enqueuing_is_done(true); 5263 if (rp->processing_is_mt()) { 5264 rp->balance_all_queues(); 5265 CMSRefProcTaskExecutor task_executor(*this); 5266 rp->enqueue_discovered_references(&task_executor); 5267 } else { 5268 rp->enqueue_discovered_references(NULL); 5269 } 5270 rp->verify_no_references_recorded(); 5271 assert(!rp->discovery_enabled(), "should have been disabled"); 5272 } 5273 5274 #ifndef PRODUCT 5275 void CMSCollector::check_correct_thread_executing() { 5276 Thread* t = Thread::current(); 5277 // Only the VM thread or the CMS thread should be here. 5278 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5279 "Unexpected thread type"); 5280 // If this is the vm thread, the foreground process 5281 // should not be waiting. Note that _foregroundGCIsActive is 5282 // true while the foreground collector is waiting. 5283 if (_foregroundGCShouldWait) { 5284 // We cannot be the VM thread 5285 assert(t->is_ConcurrentGC_thread(), 5286 "Should be CMS thread"); 5287 } else { 5288 // We can be the CMS thread only if we are in a stop-world 5289 // phase of CMS collection. 5290 if (t->is_ConcurrentGC_thread()) { 5291 assert(_collectorState == InitialMarking || 5292 _collectorState == FinalMarking, 5293 "Should be a stop-world phase"); 5294 // The CMS thread should be holding the CMS_token. 5295 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5296 "Potential interference with concurrently " 5297 "executing VM thread"); 5298 } 5299 } 5300 } 5301 #endif 5302 5303 void CMSCollector::sweep() { 5304 assert(_collectorState == Sweeping, "just checking"); 5305 check_correct_thread_executing(); 5306 verify_work_stacks_empty(); 5307 verify_overflow_empty(); 5308 increment_sweep_count(); 5309 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5310 5311 _inter_sweep_timer.stop(); 5312 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5313 5314 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5315 _intra_sweep_timer.reset(); 5316 _intra_sweep_timer.start(); 5317 { 5318 GCTraceCPUTime tcpu; 5319 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5320 // First sweep the old gen 5321 { 5322 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5323 bitMapLock()); 5324 sweepWork(_cmsGen); 5325 } 5326 5327 // Update Universe::_heap_*_at_gc figures. 5328 // We need all the free list locks to make the abstract state 5329 // transition from Sweeping to Resetting. See detailed note 5330 // further below. 5331 { 5332 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5333 // Update heap occupancy information which is used as 5334 // input to soft ref clearing policy at the next gc. 5335 Universe::update_heap_info_at_gc(); 5336 _collectorState = Resizing; 5337 } 5338 } 5339 verify_work_stacks_empty(); 5340 verify_overflow_empty(); 5341 5342 if (should_unload_classes()) { 5343 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5344 // requires that the virtual spaces are stable and not deleted. 5345 ClassLoaderDataGraph::set_should_purge(true); 5346 } 5347 5348 _intra_sweep_timer.stop(); 5349 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5350 5351 _inter_sweep_timer.reset(); 5352 _inter_sweep_timer.start(); 5353 5354 // We need to use a monotonically non-decreasing time in ms 5355 // or we will see time-warp warnings and os::javaTimeMillis() 5356 // does not guarantee monotonicity. 5357 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5358 update_time_of_last_gc(now); 5359 5360 // NOTE on abstract state transitions: 5361 // Mutators allocate-live and/or mark the mod-union table dirty 5362 // based on the state of the collection. The former is done in 5363 // the interval [Marking, Sweeping] and the latter in the interval 5364 // [Marking, Sweeping). Thus the transitions into the Marking state 5365 // and out of the Sweeping state must be synchronously visible 5366 // globally to the mutators. 5367 // The transition into the Marking state happens with the world 5368 // stopped so the mutators will globally see it. Sweeping is 5369 // done asynchronously by the background collector so the transition 5370 // from the Sweeping state to the Resizing state must be done 5371 // under the freelistLock (as is the check for whether to 5372 // allocate-live and whether to dirty the mod-union table). 5373 assert(_collectorState == Resizing, "Change of collector state to" 5374 " Resizing must be done under the freelistLocks (plural)"); 5375 5376 // Now that sweeping has been completed, we clear 5377 // the incremental_collection_failed flag, 5378 // thus inviting a younger gen collection to promote into 5379 // this generation. If such a promotion may still fail, 5380 // the flag will be set again when a young collection is 5381 // attempted. 5382 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5383 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5384 gch->update_full_collections_completed(_collection_count_start); 5385 } 5386 5387 // FIX ME!!! Looks like this belongs in CFLSpace, with 5388 // CMSGen merely delegating to it. 5389 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5390 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5391 HeapWord* minAddr = _cmsSpace->bottom(); 5392 HeapWord* largestAddr = 5393 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5394 if (largestAddr == NULL) { 5395 // The dictionary appears to be empty. In this case 5396 // try to coalesce at the end of the heap. 5397 largestAddr = _cmsSpace->end(); 5398 } 5399 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5400 size_t nearLargestOffset = 5401 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5402 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5403 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5404 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5405 } 5406 5407 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5408 return addr >= _cmsSpace->nearLargestChunk(); 5409 } 5410 5411 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5412 return _cmsSpace->find_chunk_at_end(); 5413 } 5414 5415 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5416 bool full) { 5417 // If the young generation has been collected, gather any statistics 5418 // that are of interest at this point. 5419 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5420 if (!full && current_is_young) { 5421 // Gather statistics on the young generation collection. 5422 collector()->stats().record_gc0_end(used()); 5423 } 5424 } 5425 5426 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5427 // We iterate over the space(s) underlying this generation, 5428 // checking the mark bit map to see if the bits corresponding 5429 // to specific blocks are marked or not. Blocks that are 5430 // marked are live and are not swept up. All remaining blocks 5431 // are swept up, with coalescing on-the-fly as we sweep up 5432 // contiguous free and/or garbage blocks: 5433 // We need to ensure that the sweeper synchronizes with allocators 5434 // and stop-the-world collectors. In particular, the following 5435 // locks are used: 5436 // . CMS token: if this is held, a stop the world collection cannot occur 5437 // . freelistLock: if this is held no allocation can occur from this 5438 // generation by another thread 5439 // . bitMapLock: if this is held, no other thread can access or update 5440 // 5441 5442 // Note that we need to hold the freelistLock if we use 5443 // block iterate below; else the iterator might go awry if 5444 // a mutator (or promotion) causes block contents to change 5445 // (for instance if the allocator divvies up a block). 5446 // If we hold the free list lock, for all practical purposes 5447 // young generation GC's can't occur (they'll usually need to 5448 // promote), so we might as well prevent all young generation 5449 // GC's while we do a sweeping step. For the same reason, we might 5450 // as well take the bit map lock for the entire duration 5451 5452 // check that we hold the requisite locks 5453 assert(have_cms_token(), "Should hold cms token"); 5454 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5455 assert_lock_strong(old_gen->freelistLock()); 5456 assert_lock_strong(bitMapLock()); 5457 5458 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5459 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5460 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5461 _inter_sweep_estimate.padded_average(), 5462 _intra_sweep_estimate.padded_average()); 5463 old_gen->setNearLargestChunk(); 5464 5465 { 5466 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5467 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5468 // We need to free-up/coalesce garbage/blocks from a 5469 // co-terminal free run. This is done in the SweepClosure 5470 // destructor; so, do not remove this scope, else the 5471 // end-of-sweep-census below will be off by a little bit. 5472 } 5473 old_gen->cmsSpace()->sweep_completed(); 5474 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5475 if (should_unload_classes()) { // unloaded classes this cycle, 5476 _concurrent_cycles_since_last_unload = 0; // ... reset count 5477 } else { // did not unload classes, 5478 _concurrent_cycles_since_last_unload++; // ... increment count 5479 } 5480 } 5481 5482 // Reset CMS data structures (for now just the marking bit map) 5483 // preparatory for the next cycle. 5484 void CMSCollector::reset_concurrent() { 5485 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5486 5487 // If the state is not "Resetting", the foreground thread 5488 // has done a collection and the resetting. 5489 if (_collectorState != Resetting) { 5490 assert(_collectorState == Idling, "The state should only change" 5491 " because the foreground collector has finished the collection"); 5492 return; 5493 } 5494 5495 { 5496 // Clear the mark bitmap (no grey objects to start with) 5497 // for the next cycle. 5498 GCTraceCPUTime tcpu; 5499 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5500 5501 HeapWord* curAddr = _markBitMap.startWord(); 5502 while (curAddr < _markBitMap.endWord()) { 5503 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5504 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5505 _markBitMap.clear_large_range(chunk); 5506 if (ConcurrentMarkSweepThread::should_yield() && 5507 !foregroundGCIsActive() && 5508 CMSYield) { 5509 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5510 "CMS thread should hold CMS token"); 5511 assert_lock_strong(bitMapLock()); 5512 bitMapLock()->unlock(); 5513 ConcurrentMarkSweepThread::desynchronize(true); 5514 stopTimer(); 5515 incrementYields(); 5516 5517 // See the comment in coordinator_yield() 5518 for (unsigned i = 0; i < CMSYieldSleepCount && 5519 ConcurrentMarkSweepThread::should_yield() && 5520 !CMSCollector::foregroundGCIsActive(); ++i) { 5521 os::sleep(Thread::current(), 1, false); 5522 } 5523 5524 ConcurrentMarkSweepThread::synchronize(true); 5525 bitMapLock()->lock_without_safepoint_check(); 5526 startTimer(); 5527 } 5528 curAddr = chunk.end(); 5529 } 5530 // A successful mostly concurrent collection has been done. 5531 // Because only the full (i.e., concurrent mode failure) collections 5532 // are being measured for gc overhead limits, clean the "near" flag 5533 // and count. 5534 size_policy()->reset_gc_overhead_limit_count(); 5535 _collectorState = Idling; 5536 } 5537 5538 register_gc_end(); 5539 } 5540 5541 // Same as above but for STW paths 5542 void CMSCollector::reset_stw() { 5543 // already have the lock 5544 assert(_collectorState == Resetting, "just checking"); 5545 assert_lock_strong(bitMapLock()); 5546 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5547 _markBitMap.clear_all(); 5548 _collectorState = Idling; 5549 register_gc_end(); 5550 } 5551 5552 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5553 GCTraceCPUTime tcpu; 5554 TraceCollectorStats tcs(cgc_counters()); 5555 5556 switch (op) { 5557 case CMS_op_checkpointRootsInitial: { 5558 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5559 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5560 checkpointRootsInitial(); 5561 break; 5562 } 5563 case CMS_op_checkpointRootsFinal: { 5564 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5565 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5566 checkpointRootsFinal(); 5567 break; 5568 } 5569 default: 5570 fatal("No such CMS_op"); 5571 } 5572 } 5573 5574 #ifndef PRODUCT 5575 size_t const CMSCollector::skip_header_HeapWords() { 5576 return FreeChunk::header_size(); 5577 } 5578 5579 // Try and collect here conditions that should hold when 5580 // CMS thread is exiting. The idea is that the foreground GC 5581 // thread should not be blocked if it wants to terminate 5582 // the CMS thread and yet continue to run the VM for a while 5583 // after that. 5584 void CMSCollector::verify_ok_to_terminate() const { 5585 assert(Thread::current()->is_ConcurrentGC_thread(), 5586 "should be called by CMS thread"); 5587 assert(!_foregroundGCShouldWait, "should be false"); 5588 // We could check here that all the various low-level locks 5589 // are not held by the CMS thread, but that is overkill; see 5590 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5591 // is checked. 5592 } 5593 #endif 5594 5595 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5596 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5597 "missing Printezis mark?"); 5598 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5599 size_t size = pointer_delta(nextOneAddr + 1, addr); 5600 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5601 "alignment problem"); 5602 assert(size >= 3, "Necessary for Printezis marks to work"); 5603 return size; 5604 } 5605 5606 // A variant of the above (block_size_using_printezis_bits()) except 5607 // that we return 0 if the P-bits are not yet set. 5608 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5609 if (_markBitMap.isMarked(addr + 1)) { 5610 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5611 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5612 size_t size = pointer_delta(nextOneAddr + 1, addr); 5613 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5614 "alignment problem"); 5615 assert(size >= 3, "Necessary for Printezis marks to work"); 5616 return size; 5617 } 5618 return 0; 5619 } 5620 5621 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5622 size_t sz = 0; 5623 oop p = (oop)addr; 5624 if (p->klass_or_null() != NULL) { 5625 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5626 } else { 5627 sz = block_size_using_printezis_bits(addr); 5628 } 5629 assert(sz > 0, "size must be nonzero"); 5630 HeapWord* next_block = addr + sz; 5631 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5632 CardTableModRefBS::card_size); 5633 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5634 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5635 "must be different cards"); 5636 return next_card; 5637 } 5638 5639 5640 // CMS Bit Map Wrapper ///////////////////////////////////////// 5641 5642 // Construct a CMS bit map infrastructure, but don't create the 5643 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5644 // further below. 5645 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5646 _bm(), 5647 _shifter(shifter), 5648 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5649 Monitor::_safepoint_check_sometimes) : NULL) 5650 { 5651 _bmStartWord = 0; 5652 _bmWordSize = 0; 5653 } 5654 5655 bool CMSBitMap::allocate(MemRegion mr) { 5656 _bmStartWord = mr.start(); 5657 _bmWordSize = mr.word_size(); 5658 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5659 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5660 if (!brs.is_reserved()) { 5661 warning("CMS bit map allocation failure"); 5662 return false; 5663 } 5664 // For now we'll just commit all of the bit map up front. 5665 // Later on we'll try to be more parsimonious with swap. 5666 if (!_virtual_space.initialize(brs, brs.size())) { 5667 warning("CMS bit map backing store failure"); 5668 return false; 5669 } 5670 assert(_virtual_space.committed_size() == brs.size(), 5671 "didn't reserve backing store for all of CMS bit map?"); 5672 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5673 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5674 _bmWordSize, "inconsistency in bit map sizing"); 5675 _bm.set_size(_bmWordSize >> _shifter); 5676 5677 // bm.clear(); // can we rely on getting zero'd memory? verify below 5678 assert(isAllClear(), 5679 "Expected zero'd memory from ReservedSpace constructor"); 5680 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5681 "consistency check"); 5682 return true; 5683 } 5684 5685 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5686 HeapWord *next_addr, *end_addr, *last_addr; 5687 assert_locked(); 5688 assert(covers(mr), "out-of-range error"); 5689 // XXX assert that start and end are appropriately aligned 5690 for (next_addr = mr.start(), end_addr = mr.end(); 5691 next_addr < end_addr; next_addr = last_addr) { 5692 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5693 last_addr = dirty_region.end(); 5694 if (!dirty_region.is_empty()) { 5695 cl->do_MemRegion(dirty_region); 5696 } else { 5697 assert(last_addr == end_addr, "program logic"); 5698 return; 5699 } 5700 } 5701 } 5702 5703 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5704 _bm.print_on_error(st, prefix); 5705 } 5706 5707 #ifndef PRODUCT 5708 void CMSBitMap::assert_locked() const { 5709 CMSLockVerifier::assert_locked(lock()); 5710 } 5711 5712 bool CMSBitMap::covers(MemRegion mr) const { 5713 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5714 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5715 "size inconsistency"); 5716 return (mr.start() >= _bmStartWord) && 5717 (mr.end() <= endWord()); 5718 } 5719 5720 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5721 return (start >= _bmStartWord && (start + size) <= endWord()); 5722 } 5723 5724 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5725 // verify that there are no 1 bits in the interval [left, right) 5726 FalseBitMapClosure falseBitMapClosure; 5727 iterate(&falseBitMapClosure, left, right); 5728 } 5729 5730 void CMSBitMap::region_invariant(MemRegion mr) 5731 { 5732 assert_locked(); 5733 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5734 assert(!mr.is_empty(), "unexpected empty region"); 5735 assert(covers(mr), "mr should be covered by bit map"); 5736 // convert address range into offset range 5737 size_t start_ofs = heapWordToOffset(mr.start()); 5738 // Make sure that end() is appropriately aligned 5739 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5740 (1 << (_shifter+LogHeapWordSize))), 5741 "Misaligned mr.end()"); 5742 size_t end_ofs = heapWordToOffset(mr.end()); 5743 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5744 } 5745 5746 #endif 5747 5748 bool CMSMarkStack::allocate(size_t size) { 5749 // allocate a stack of the requisite depth 5750 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5751 size * sizeof(oop))); 5752 if (!rs.is_reserved()) { 5753 warning("CMSMarkStack allocation failure"); 5754 return false; 5755 } 5756 if (!_virtual_space.initialize(rs, rs.size())) { 5757 warning("CMSMarkStack backing store failure"); 5758 return false; 5759 } 5760 assert(_virtual_space.committed_size() == rs.size(), 5761 "didn't reserve backing store for all of CMS stack?"); 5762 _base = (oop*)(_virtual_space.low()); 5763 _index = 0; 5764 _capacity = size; 5765 NOT_PRODUCT(_max_depth = 0); 5766 return true; 5767 } 5768 5769 // XXX FIX ME !!! In the MT case we come in here holding a 5770 // leaf lock. For printing we need to take a further lock 5771 // which has lower rank. We need to recalibrate the two 5772 // lock-ranks involved in order to be able to print the 5773 // messages below. (Or defer the printing to the caller. 5774 // For now we take the expedient path of just disabling the 5775 // messages for the problematic case.) 5776 void CMSMarkStack::expand() { 5777 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5778 if (_capacity == MarkStackSizeMax) { 5779 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5780 // We print a warning message only once per CMS cycle. 5781 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5782 } 5783 return; 5784 } 5785 // Double capacity if possible 5786 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5787 // Do not give up existing stack until we have managed to 5788 // get the double capacity that we desired. 5789 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5790 new_capacity * sizeof(oop))); 5791 if (rs.is_reserved()) { 5792 // Release the backing store associated with old stack 5793 _virtual_space.release(); 5794 // Reinitialize virtual space for new stack 5795 if (!_virtual_space.initialize(rs, rs.size())) { 5796 fatal("Not enough swap for expanded marking stack"); 5797 } 5798 _base = (oop*)(_virtual_space.low()); 5799 _index = 0; 5800 _capacity = new_capacity; 5801 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5802 // Failed to double capacity, continue; 5803 // we print a detail message only once per CMS cycle. 5804 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5805 _capacity / K, new_capacity / K); 5806 } 5807 } 5808 5809 5810 // Closures 5811 // XXX: there seems to be a lot of code duplication here; 5812 // should refactor and consolidate common code. 5813 5814 // This closure is used to mark refs into the CMS generation in 5815 // the CMS bit map. Called at the first checkpoint. This closure 5816 // assumes that we do not need to re-mark dirty cards; if the CMS 5817 // generation on which this is used is not an oldest 5818 // generation then this will lose younger_gen cards! 5819 5820 MarkRefsIntoClosure::MarkRefsIntoClosure( 5821 MemRegion span, CMSBitMap* bitMap): 5822 _span(span), 5823 _bitMap(bitMap) 5824 { 5825 assert(ref_processor() == NULL, "deliberately left NULL"); 5826 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5827 } 5828 5829 void MarkRefsIntoClosure::do_oop(oop obj) { 5830 // if p points into _span, then mark corresponding bit in _markBitMap 5831 assert(obj->is_oop(), "expected an oop"); 5832 HeapWord* addr = (HeapWord*)obj; 5833 if (_span.contains(addr)) { 5834 // this should be made more efficient 5835 _bitMap->mark(addr); 5836 } 5837 } 5838 5839 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5840 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5841 5842 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5843 MemRegion span, CMSBitMap* bitMap): 5844 _span(span), 5845 _bitMap(bitMap) 5846 { 5847 assert(ref_processor() == NULL, "deliberately left NULL"); 5848 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5849 } 5850 5851 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5852 // if p points into _span, then mark corresponding bit in _markBitMap 5853 assert(obj->is_oop(), "expected an oop"); 5854 HeapWord* addr = (HeapWord*)obj; 5855 if (_span.contains(addr)) { 5856 // this should be made more efficient 5857 _bitMap->par_mark(addr); 5858 } 5859 } 5860 5861 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5862 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5863 5864 // A variant of the above, used for CMS marking verification. 5865 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5866 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5867 _span(span), 5868 _verification_bm(verification_bm), 5869 _cms_bm(cms_bm) 5870 { 5871 assert(ref_processor() == NULL, "deliberately left NULL"); 5872 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5873 } 5874 5875 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5876 // if p points into _span, then mark corresponding bit in _markBitMap 5877 assert(obj->is_oop(), "expected an oop"); 5878 HeapWord* addr = (HeapWord*)obj; 5879 if (_span.contains(addr)) { 5880 _verification_bm->mark(addr); 5881 if (!_cms_bm->isMarked(addr)) { 5882 LogHandle(gc, verify) log; 5883 ResourceMark rm; 5884 oop(addr)->print_on(log.error_stream()); 5885 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5886 fatal("... aborting"); 5887 } 5888 } 5889 } 5890 5891 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5892 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5893 5894 ////////////////////////////////////////////////// 5895 // MarkRefsIntoAndScanClosure 5896 ////////////////////////////////////////////////// 5897 5898 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5899 ReferenceProcessor* rp, 5900 CMSBitMap* bit_map, 5901 CMSBitMap* mod_union_table, 5902 CMSMarkStack* mark_stack, 5903 CMSCollector* collector, 5904 bool should_yield, 5905 bool concurrent_precleaning): 5906 _collector(collector), 5907 _span(span), 5908 _bit_map(bit_map), 5909 _mark_stack(mark_stack), 5910 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5911 mark_stack, concurrent_precleaning), 5912 _yield(should_yield), 5913 _concurrent_precleaning(concurrent_precleaning), 5914 _freelistLock(NULL) 5915 { 5916 // FIXME: Should initialize in base class constructor. 5917 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5918 set_ref_processor_internal(rp); 5919 } 5920 5921 // This closure is used to mark refs into the CMS generation at the 5922 // second (final) checkpoint, and to scan and transitively follow 5923 // the unmarked oops. It is also used during the concurrent precleaning 5924 // phase while scanning objects on dirty cards in the CMS generation. 5925 // The marks are made in the marking bit map and the marking stack is 5926 // used for keeping the (newly) grey objects during the scan. 5927 // The parallel version (Par_...) appears further below. 5928 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5929 if (obj != NULL) { 5930 assert(obj->is_oop(), "expected an oop"); 5931 HeapWord* addr = (HeapWord*)obj; 5932 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5933 assert(_collector->overflow_list_is_empty(), 5934 "overflow list should be empty"); 5935 if (_span.contains(addr) && 5936 !_bit_map->isMarked(addr)) { 5937 // mark bit map (object is now grey) 5938 _bit_map->mark(addr); 5939 // push on marking stack (stack should be empty), and drain the 5940 // stack by applying this closure to the oops in the oops popped 5941 // from the stack (i.e. blacken the grey objects) 5942 bool res = _mark_stack->push(obj); 5943 assert(res, "Should have space to push on empty stack"); 5944 do { 5945 oop new_oop = _mark_stack->pop(); 5946 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5947 assert(_bit_map->isMarked((HeapWord*)new_oop), 5948 "only grey objects on this stack"); 5949 // iterate over the oops in this oop, marking and pushing 5950 // the ones in CMS heap (i.e. in _span). 5951 new_oop->oop_iterate(&_pushAndMarkClosure); 5952 // check if it's time to yield 5953 do_yield_check(); 5954 } while (!_mark_stack->isEmpty() || 5955 (!_concurrent_precleaning && take_from_overflow_list())); 5956 // if marking stack is empty, and we are not doing this 5957 // during precleaning, then check the overflow list 5958 } 5959 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5960 assert(_collector->overflow_list_is_empty(), 5961 "overflow list was drained above"); 5962 5963 assert(_collector->no_preserved_marks(), 5964 "All preserved marks should have been restored above"); 5965 } 5966 } 5967 5968 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5969 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5970 5971 void MarkRefsIntoAndScanClosure::do_yield_work() { 5972 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5973 "CMS thread should hold CMS token"); 5974 assert_lock_strong(_freelistLock); 5975 assert_lock_strong(_bit_map->lock()); 5976 // relinquish the free_list_lock and bitMaplock() 5977 _bit_map->lock()->unlock(); 5978 _freelistLock->unlock(); 5979 ConcurrentMarkSweepThread::desynchronize(true); 5980 _collector->stopTimer(); 5981 _collector->incrementYields(); 5982 5983 // See the comment in coordinator_yield() 5984 for (unsigned i = 0; 5985 i < CMSYieldSleepCount && 5986 ConcurrentMarkSweepThread::should_yield() && 5987 !CMSCollector::foregroundGCIsActive(); 5988 ++i) { 5989 os::sleep(Thread::current(), 1, false); 5990 } 5991 5992 ConcurrentMarkSweepThread::synchronize(true); 5993 _freelistLock->lock_without_safepoint_check(); 5994 _bit_map->lock()->lock_without_safepoint_check(); 5995 _collector->startTimer(); 5996 } 5997 5998 /////////////////////////////////////////////////////////// 5999 // ParMarkRefsIntoAndScanClosure: a parallel version of 6000 // MarkRefsIntoAndScanClosure 6001 /////////////////////////////////////////////////////////// 6002 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6003 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6004 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6005 _span(span), 6006 _bit_map(bit_map), 6007 _work_queue(work_queue), 6008 _low_water_mark(MIN2((work_queue->max_elems()/4), 6009 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6010 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6011 { 6012 // FIXME: Should initialize in base class constructor. 6013 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6014 set_ref_processor_internal(rp); 6015 } 6016 6017 // This closure is used to mark refs into the CMS generation at the 6018 // second (final) checkpoint, and to scan and transitively follow 6019 // the unmarked oops. The marks are made in the marking bit map and 6020 // the work_queue is used for keeping the (newly) grey objects during 6021 // the scan phase whence they are also available for stealing by parallel 6022 // threads. Since the marking bit map is shared, updates are 6023 // synchronized (via CAS). 6024 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6025 if (obj != NULL) { 6026 // Ignore mark word because this could be an already marked oop 6027 // that may be chained at the end of the overflow list. 6028 assert(obj->is_oop(true), "expected an oop"); 6029 HeapWord* addr = (HeapWord*)obj; 6030 if (_span.contains(addr) && 6031 !_bit_map->isMarked(addr)) { 6032 // mark bit map (object will become grey): 6033 // It is possible for several threads to be 6034 // trying to "claim" this object concurrently; 6035 // the unique thread that succeeds in marking the 6036 // object first will do the subsequent push on 6037 // to the work queue (or overflow list). 6038 if (_bit_map->par_mark(addr)) { 6039 // push on work_queue (which may not be empty), and trim the 6040 // queue to an appropriate length by applying this closure to 6041 // the oops in the oops popped from the stack (i.e. blacken the 6042 // grey objects) 6043 bool res = _work_queue->push(obj); 6044 assert(res, "Low water mark should be less than capacity?"); 6045 trim_queue(_low_water_mark); 6046 } // Else, another thread claimed the object 6047 } 6048 } 6049 } 6050 6051 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6052 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6053 6054 // This closure is used to rescan the marked objects on the dirty cards 6055 // in the mod union table and the card table proper. 6056 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6057 oop p, MemRegion mr) { 6058 6059 size_t size = 0; 6060 HeapWord* addr = (HeapWord*)p; 6061 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6062 assert(_span.contains(addr), "we are scanning the CMS generation"); 6063 // check if it's time to yield 6064 if (do_yield_check()) { 6065 // We yielded for some foreground stop-world work, 6066 // and we have been asked to abort this ongoing preclean cycle. 6067 return 0; 6068 } 6069 if (_bitMap->isMarked(addr)) { 6070 // it's marked; is it potentially uninitialized? 6071 if (p->klass_or_null() != NULL) { 6072 // an initialized object; ignore mark word in verification below 6073 // since we are running concurrent with mutators 6074 assert(p->is_oop(true), "should be an oop"); 6075 if (p->is_objArray()) { 6076 // objArrays are precisely marked; restrict scanning 6077 // to dirty cards only. 6078 size = CompactibleFreeListSpace::adjustObjectSize( 6079 p->oop_iterate_size(_scanningClosure, mr)); 6080 } else { 6081 // A non-array may have been imprecisely marked; we need 6082 // to scan object in its entirety. 6083 size = CompactibleFreeListSpace::adjustObjectSize( 6084 p->oop_iterate_size(_scanningClosure)); 6085 } 6086 #ifdef ASSERT 6087 size_t direct_size = 6088 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6089 assert(size == direct_size, "Inconsistency in size"); 6090 assert(size >= 3, "Necessary for Printezis marks to work"); 6091 if (!_bitMap->isMarked(addr+1)) { 6092 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6093 } else { 6094 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6095 assert(_bitMap->isMarked(addr+size-1), 6096 "inconsistent Printezis mark"); 6097 } 6098 #endif // ASSERT 6099 } else { 6100 // An uninitialized object. 6101 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6102 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6103 size = pointer_delta(nextOneAddr + 1, addr); 6104 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6105 "alignment problem"); 6106 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6107 // will dirty the card when the klass pointer is installed in the 6108 // object (signaling the completion of initialization). 6109 } 6110 } else { 6111 // Either a not yet marked object or an uninitialized object 6112 if (p->klass_or_null() == NULL) { 6113 // An uninitialized object, skip to the next card, since 6114 // we may not be able to read its P-bits yet. 6115 assert(size == 0, "Initial value"); 6116 } else { 6117 // An object not (yet) reached by marking: we merely need to 6118 // compute its size so as to go look at the next block. 6119 assert(p->is_oop(true), "should be an oop"); 6120 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6121 } 6122 } 6123 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6124 return size; 6125 } 6126 6127 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6128 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6129 "CMS thread should hold CMS token"); 6130 assert_lock_strong(_freelistLock); 6131 assert_lock_strong(_bitMap->lock()); 6132 // relinquish the free_list_lock and bitMaplock() 6133 _bitMap->lock()->unlock(); 6134 _freelistLock->unlock(); 6135 ConcurrentMarkSweepThread::desynchronize(true); 6136 _collector->stopTimer(); 6137 _collector->incrementYields(); 6138 6139 // See the comment in coordinator_yield() 6140 for (unsigned i = 0; i < CMSYieldSleepCount && 6141 ConcurrentMarkSweepThread::should_yield() && 6142 !CMSCollector::foregroundGCIsActive(); ++i) { 6143 os::sleep(Thread::current(), 1, false); 6144 } 6145 6146 ConcurrentMarkSweepThread::synchronize(true); 6147 _freelistLock->lock_without_safepoint_check(); 6148 _bitMap->lock()->lock_without_safepoint_check(); 6149 _collector->startTimer(); 6150 } 6151 6152 6153 ////////////////////////////////////////////////////////////////// 6154 // SurvivorSpacePrecleanClosure 6155 ////////////////////////////////////////////////////////////////// 6156 // This (single-threaded) closure is used to preclean the oops in 6157 // the survivor spaces. 6158 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6159 6160 HeapWord* addr = (HeapWord*)p; 6161 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6162 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6163 assert(p->klass_or_null() != NULL, "object should be initialized"); 6164 // an initialized object; ignore mark word in verification below 6165 // since we are running concurrent with mutators 6166 assert(p->is_oop(true), "should be an oop"); 6167 // Note that we do not yield while we iterate over 6168 // the interior oops of p, pushing the relevant ones 6169 // on our marking stack. 6170 size_t size = p->oop_iterate_size(_scanning_closure); 6171 do_yield_check(); 6172 // Observe that below, we do not abandon the preclean 6173 // phase as soon as we should; rather we empty the 6174 // marking stack before returning. This is to satisfy 6175 // some existing assertions. In general, it may be a 6176 // good idea to abort immediately and complete the marking 6177 // from the grey objects at a later time. 6178 while (!_mark_stack->isEmpty()) { 6179 oop new_oop = _mark_stack->pop(); 6180 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6181 assert(_bit_map->isMarked((HeapWord*)new_oop), 6182 "only grey objects on this stack"); 6183 // iterate over the oops in this oop, marking and pushing 6184 // the ones in CMS heap (i.e. in _span). 6185 new_oop->oop_iterate(_scanning_closure); 6186 // check if it's time to yield 6187 do_yield_check(); 6188 } 6189 unsigned int after_count = 6190 GenCollectedHeap::heap()->total_collections(); 6191 bool abort = (_before_count != after_count) || 6192 _collector->should_abort_preclean(); 6193 return abort ? 0 : size; 6194 } 6195 6196 void SurvivorSpacePrecleanClosure::do_yield_work() { 6197 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6198 "CMS thread should hold CMS token"); 6199 assert_lock_strong(_bit_map->lock()); 6200 // Relinquish the bit map lock 6201 _bit_map->lock()->unlock(); 6202 ConcurrentMarkSweepThread::desynchronize(true); 6203 _collector->stopTimer(); 6204 _collector->incrementYields(); 6205 6206 // See the comment in coordinator_yield() 6207 for (unsigned i = 0; i < CMSYieldSleepCount && 6208 ConcurrentMarkSweepThread::should_yield() && 6209 !CMSCollector::foregroundGCIsActive(); ++i) { 6210 os::sleep(Thread::current(), 1, false); 6211 } 6212 6213 ConcurrentMarkSweepThread::synchronize(true); 6214 _bit_map->lock()->lock_without_safepoint_check(); 6215 _collector->startTimer(); 6216 } 6217 6218 // This closure is used to rescan the marked objects on the dirty cards 6219 // in the mod union table and the card table proper. In the parallel 6220 // case, although the bitMap is shared, we do a single read so the 6221 // isMarked() query is "safe". 6222 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6223 // Ignore mark word because we are running concurrent with mutators 6224 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6225 HeapWord* addr = (HeapWord*)p; 6226 assert(_span.contains(addr), "we are scanning the CMS generation"); 6227 bool is_obj_array = false; 6228 #ifdef ASSERT 6229 if (!_parallel) { 6230 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6231 assert(_collector->overflow_list_is_empty(), 6232 "overflow list should be empty"); 6233 6234 } 6235 #endif // ASSERT 6236 if (_bit_map->isMarked(addr)) { 6237 // Obj arrays are precisely marked, non-arrays are not; 6238 // so we scan objArrays precisely and non-arrays in their 6239 // entirety. 6240 if (p->is_objArray()) { 6241 is_obj_array = true; 6242 if (_parallel) { 6243 p->oop_iterate(_par_scan_closure, mr); 6244 } else { 6245 p->oop_iterate(_scan_closure, mr); 6246 } 6247 } else { 6248 if (_parallel) { 6249 p->oop_iterate(_par_scan_closure); 6250 } else { 6251 p->oop_iterate(_scan_closure); 6252 } 6253 } 6254 } 6255 #ifdef ASSERT 6256 if (!_parallel) { 6257 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6258 assert(_collector->overflow_list_is_empty(), 6259 "overflow list should be empty"); 6260 6261 } 6262 #endif // ASSERT 6263 return is_obj_array; 6264 } 6265 6266 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6267 MemRegion span, 6268 CMSBitMap* bitMap, CMSMarkStack* markStack, 6269 bool should_yield, bool verifying): 6270 _collector(collector), 6271 _span(span), 6272 _bitMap(bitMap), 6273 _mut(&collector->_modUnionTable), 6274 _markStack(markStack), 6275 _yield(should_yield), 6276 _skipBits(0) 6277 { 6278 assert(_markStack->isEmpty(), "stack should be empty"); 6279 _finger = _bitMap->startWord(); 6280 _threshold = _finger; 6281 assert(_collector->_restart_addr == NULL, "Sanity check"); 6282 assert(_span.contains(_finger), "Out of bounds _finger?"); 6283 DEBUG_ONLY(_verifying = verifying;) 6284 } 6285 6286 void MarkFromRootsClosure::reset(HeapWord* addr) { 6287 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6288 assert(_span.contains(addr), "Out of bounds _finger?"); 6289 _finger = addr; 6290 _threshold = (HeapWord*)round_to( 6291 (intptr_t)_finger, CardTableModRefBS::card_size); 6292 } 6293 6294 // Should revisit to see if this should be restructured for 6295 // greater efficiency. 6296 bool MarkFromRootsClosure::do_bit(size_t offset) { 6297 if (_skipBits > 0) { 6298 _skipBits--; 6299 return true; 6300 } 6301 // convert offset into a HeapWord* 6302 HeapWord* addr = _bitMap->startWord() + offset; 6303 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6304 "address out of range"); 6305 assert(_bitMap->isMarked(addr), "tautology"); 6306 if (_bitMap->isMarked(addr+1)) { 6307 // this is an allocated but not yet initialized object 6308 assert(_skipBits == 0, "tautology"); 6309 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6310 oop p = oop(addr); 6311 if (p->klass_or_null() == NULL) { 6312 DEBUG_ONLY(if (!_verifying) {) 6313 // We re-dirty the cards on which this object lies and increase 6314 // the _threshold so that we'll come back to scan this object 6315 // during the preclean or remark phase. (CMSCleanOnEnter) 6316 if (CMSCleanOnEnter) { 6317 size_t sz = _collector->block_size_using_printezis_bits(addr); 6318 HeapWord* end_card_addr = (HeapWord*)round_to( 6319 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6320 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6321 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6322 // Bump _threshold to end_card_addr; note that 6323 // _threshold cannot possibly exceed end_card_addr, anyhow. 6324 // This prevents future clearing of the card as the scan proceeds 6325 // to the right. 6326 assert(_threshold <= end_card_addr, 6327 "Because we are just scanning into this object"); 6328 if (_threshold < end_card_addr) { 6329 _threshold = end_card_addr; 6330 } 6331 if (p->klass_or_null() != NULL) { 6332 // Redirty the range of cards... 6333 _mut->mark_range(redirty_range); 6334 } // ...else the setting of klass will dirty the card anyway. 6335 } 6336 DEBUG_ONLY(}) 6337 return true; 6338 } 6339 } 6340 scanOopsInOop(addr); 6341 return true; 6342 } 6343 6344 // We take a break if we've been at this for a while, 6345 // so as to avoid monopolizing the locks involved. 6346 void MarkFromRootsClosure::do_yield_work() { 6347 // First give up the locks, then yield, then re-lock 6348 // We should probably use a constructor/destructor idiom to 6349 // do this unlock/lock or modify the MutexUnlocker class to 6350 // serve our purpose. XXX 6351 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6352 "CMS thread should hold CMS token"); 6353 assert_lock_strong(_bitMap->lock()); 6354 _bitMap->lock()->unlock(); 6355 ConcurrentMarkSweepThread::desynchronize(true); 6356 _collector->stopTimer(); 6357 _collector->incrementYields(); 6358 6359 // See the comment in coordinator_yield() 6360 for (unsigned i = 0; i < CMSYieldSleepCount && 6361 ConcurrentMarkSweepThread::should_yield() && 6362 !CMSCollector::foregroundGCIsActive(); ++i) { 6363 os::sleep(Thread::current(), 1, false); 6364 } 6365 6366 ConcurrentMarkSweepThread::synchronize(true); 6367 _bitMap->lock()->lock_without_safepoint_check(); 6368 _collector->startTimer(); 6369 } 6370 6371 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6372 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6373 assert(_markStack->isEmpty(), 6374 "should drain stack to limit stack usage"); 6375 // convert ptr to an oop preparatory to scanning 6376 oop obj = oop(ptr); 6377 // Ignore mark word in verification below, since we 6378 // may be running concurrent with mutators. 6379 assert(obj->is_oop(true), "should be an oop"); 6380 assert(_finger <= ptr, "_finger runneth ahead"); 6381 // advance the finger to right end of this object 6382 _finger = ptr + obj->size(); 6383 assert(_finger > ptr, "we just incremented it above"); 6384 // On large heaps, it may take us some time to get through 6385 // the marking phase. During 6386 // this time it's possible that a lot of mutations have 6387 // accumulated in the card table and the mod union table -- 6388 // these mutation records are redundant until we have 6389 // actually traced into the corresponding card. 6390 // Here, we check whether advancing the finger would make 6391 // us cross into a new card, and if so clear corresponding 6392 // cards in the MUT (preclean them in the card-table in the 6393 // future). 6394 6395 DEBUG_ONLY(if (!_verifying) {) 6396 // The clean-on-enter optimization is disabled by default, 6397 // until we fix 6178663. 6398 if (CMSCleanOnEnter && (_finger > _threshold)) { 6399 // [_threshold, _finger) represents the interval 6400 // of cards to be cleared in MUT (or precleaned in card table). 6401 // The set of cards to be cleared is all those that overlap 6402 // with the interval [_threshold, _finger); note that 6403 // _threshold is always kept card-aligned but _finger isn't 6404 // always card-aligned. 6405 HeapWord* old_threshold = _threshold; 6406 assert(old_threshold == (HeapWord*)round_to( 6407 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6408 "_threshold should always be card-aligned"); 6409 _threshold = (HeapWord*)round_to( 6410 (intptr_t)_finger, CardTableModRefBS::card_size); 6411 MemRegion mr(old_threshold, _threshold); 6412 assert(!mr.is_empty(), "Control point invariant"); 6413 assert(_span.contains(mr), "Should clear within span"); 6414 _mut->clear_range(mr); 6415 } 6416 DEBUG_ONLY(}) 6417 // Note: the finger doesn't advance while we drain 6418 // the stack below. 6419 PushOrMarkClosure pushOrMarkClosure(_collector, 6420 _span, _bitMap, _markStack, 6421 _finger, this); 6422 bool res = _markStack->push(obj); 6423 assert(res, "Empty non-zero size stack should have space for single push"); 6424 while (!_markStack->isEmpty()) { 6425 oop new_oop = _markStack->pop(); 6426 // Skip verifying header mark word below because we are 6427 // running concurrent with mutators. 6428 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6429 // now scan this oop's oops 6430 new_oop->oop_iterate(&pushOrMarkClosure); 6431 do_yield_check(); 6432 } 6433 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6434 } 6435 6436 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6437 CMSCollector* collector, MemRegion span, 6438 CMSBitMap* bit_map, 6439 OopTaskQueue* work_queue, 6440 CMSMarkStack* overflow_stack): 6441 _collector(collector), 6442 _whole_span(collector->_span), 6443 _span(span), 6444 _bit_map(bit_map), 6445 _mut(&collector->_modUnionTable), 6446 _work_queue(work_queue), 6447 _overflow_stack(overflow_stack), 6448 _skip_bits(0), 6449 _task(task) 6450 { 6451 assert(_work_queue->size() == 0, "work_queue should be empty"); 6452 _finger = span.start(); 6453 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6454 assert(_span.contains(_finger), "Out of bounds _finger?"); 6455 } 6456 6457 // Should revisit to see if this should be restructured for 6458 // greater efficiency. 6459 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6460 if (_skip_bits > 0) { 6461 _skip_bits--; 6462 return true; 6463 } 6464 // convert offset into a HeapWord* 6465 HeapWord* addr = _bit_map->startWord() + offset; 6466 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6467 "address out of range"); 6468 assert(_bit_map->isMarked(addr), "tautology"); 6469 if (_bit_map->isMarked(addr+1)) { 6470 // this is an allocated object that might not yet be initialized 6471 assert(_skip_bits == 0, "tautology"); 6472 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6473 oop p = oop(addr); 6474 if (p->klass_or_null() == NULL) { 6475 // in the case of Clean-on-Enter optimization, redirty card 6476 // and avoid clearing card by increasing the threshold. 6477 return true; 6478 } 6479 } 6480 scan_oops_in_oop(addr); 6481 return true; 6482 } 6483 6484 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6485 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6486 // Should we assert that our work queue is empty or 6487 // below some drain limit? 6488 assert(_work_queue->size() == 0, 6489 "should drain stack to limit stack usage"); 6490 // convert ptr to an oop preparatory to scanning 6491 oop obj = oop(ptr); 6492 // Ignore mark word in verification below, since we 6493 // may be running concurrent with mutators. 6494 assert(obj->is_oop(true), "should be an oop"); 6495 assert(_finger <= ptr, "_finger runneth ahead"); 6496 // advance the finger to right end of this object 6497 _finger = ptr + obj->size(); 6498 assert(_finger > ptr, "we just incremented it above"); 6499 // On large heaps, it may take us some time to get through 6500 // the marking phase. During 6501 // this time it's possible that a lot of mutations have 6502 // accumulated in the card table and the mod union table -- 6503 // these mutation records are redundant until we have 6504 // actually traced into the corresponding card. 6505 // Here, we check whether advancing the finger would make 6506 // us cross into a new card, and if so clear corresponding 6507 // cards in the MUT (preclean them in the card-table in the 6508 // future). 6509 6510 // The clean-on-enter optimization is disabled by default, 6511 // until we fix 6178663. 6512 if (CMSCleanOnEnter && (_finger > _threshold)) { 6513 // [_threshold, _finger) represents the interval 6514 // of cards to be cleared in MUT (or precleaned in card table). 6515 // The set of cards to be cleared is all those that overlap 6516 // with the interval [_threshold, _finger); note that 6517 // _threshold is always kept card-aligned but _finger isn't 6518 // always card-aligned. 6519 HeapWord* old_threshold = _threshold; 6520 assert(old_threshold == (HeapWord*)round_to( 6521 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6522 "_threshold should always be card-aligned"); 6523 _threshold = (HeapWord*)round_to( 6524 (intptr_t)_finger, CardTableModRefBS::card_size); 6525 MemRegion mr(old_threshold, _threshold); 6526 assert(!mr.is_empty(), "Control point invariant"); 6527 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6528 _mut->clear_range(mr); 6529 } 6530 6531 // Note: the local finger doesn't advance while we drain 6532 // the stack below, but the global finger sure can and will. 6533 HeapWord** gfa = _task->global_finger_addr(); 6534 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6535 _span, _bit_map, 6536 _work_queue, 6537 _overflow_stack, 6538 _finger, 6539 gfa, this); 6540 bool res = _work_queue->push(obj); // overflow could occur here 6541 assert(res, "Will hold once we use workqueues"); 6542 while (true) { 6543 oop new_oop; 6544 if (!_work_queue->pop_local(new_oop)) { 6545 // We emptied our work_queue; check if there's stuff that can 6546 // be gotten from the overflow stack. 6547 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6548 _overflow_stack, _work_queue)) { 6549 do_yield_check(); 6550 continue; 6551 } else { // done 6552 break; 6553 } 6554 } 6555 // Skip verifying header mark word below because we are 6556 // running concurrent with mutators. 6557 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6558 // now scan this oop's oops 6559 new_oop->oop_iterate(&pushOrMarkClosure); 6560 do_yield_check(); 6561 } 6562 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6563 } 6564 6565 // Yield in response to a request from VM Thread or 6566 // from mutators. 6567 void ParMarkFromRootsClosure::do_yield_work() { 6568 assert(_task != NULL, "sanity"); 6569 _task->yield(); 6570 } 6571 6572 // A variant of the above used for verifying CMS marking work. 6573 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6574 MemRegion span, 6575 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6576 CMSMarkStack* mark_stack): 6577 _collector(collector), 6578 _span(span), 6579 _verification_bm(verification_bm), 6580 _cms_bm(cms_bm), 6581 _mark_stack(mark_stack), 6582 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6583 mark_stack) 6584 { 6585 assert(_mark_stack->isEmpty(), "stack should be empty"); 6586 _finger = _verification_bm->startWord(); 6587 assert(_collector->_restart_addr == NULL, "Sanity check"); 6588 assert(_span.contains(_finger), "Out of bounds _finger?"); 6589 } 6590 6591 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6592 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6593 assert(_span.contains(addr), "Out of bounds _finger?"); 6594 _finger = addr; 6595 } 6596 6597 // Should revisit to see if this should be restructured for 6598 // greater efficiency. 6599 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6600 // convert offset into a HeapWord* 6601 HeapWord* addr = _verification_bm->startWord() + offset; 6602 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6603 "address out of range"); 6604 assert(_verification_bm->isMarked(addr), "tautology"); 6605 assert(_cms_bm->isMarked(addr), "tautology"); 6606 6607 assert(_mark_stack->isEmpty(), 6608 "should drain stack to limit stack usage"); 6609 // convert addr to an oop preparatory to scanning 6610 oop obj = oop(addr); 6611 assert(obj->is_oop(), "should be an oop"); 6612 assert(_finger <= addr, "_finger runneth ahead"); 6613 // advance the finger to right end of this object 6614 _finger = addr + obj->size(); 6615 assert(_finger > addr, "we just incremented it above"); 6616 // Note: the finger doesn't advance while we drain 6617 // the stack below. 6618 bool res = _mark_stack->push(obj); 6619 assert(res, "Empty non-zero size stack should have space for single push"); 6620 while (!_mark_stack->isEmpty()) { 6621 oop new_oop = _mark_stack->pop(); 6622 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6623 // now scan this oop's oops 6624 new_oop->oop_iterate(&_pam_verify_closure); 6625 } 6626 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6627 return true; 6628 } 6629 6630 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6631 CMSCollector* collector, MemRegion span, 6632 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6633 CMSMarkStack* mark_stack): 6634 MetadataAwareOopClosure(collector->ref_processor()), 6635 _collector(collector), 6636 _span(span), 6637 _verification_bm(verification_bm), 6638 _cms_bm(cms_bm), 6639 _mark_stack(mark_stack) 6640 { } 6641 6642 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6643 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6644 6645 // Upon stack overflow, we discard (part of) the stack, 6646 // remembering the least address amongst those discarded 6647 // in CMSCollector's _restart_address. 6648 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6649 // Remember the least grey address discarded 6650 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6651 _collector->lower_restart_addr(ra); 6652 _mark_stack->reset(); // discard stack contents 6653 _mark_stack->expand(); // expand the stack if possible 6654 } 6655 6656 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6657 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6658 HeapWord* addr = (HeapWord*)obj; 6659 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6660 // Oop lies in _span and isn't yet grey or black 6661 _verification_bm->mark(addr); // now grey 6662 if (!_cms_bm->isMarked(addr)) { 6663 LogHandle(gc, verify) log; 6664 ResourceMark rm; 6665 oop(addr)->print_on(log.error_stream()); 6666 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6667 fatal("... aborting"); 6668 } 6669 6670 if (!_mark_stack->push(obj)) { // stack overflow 6671 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6672 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6673 handle_stack_overflow(addr); 6674 } 6675 // anything including and to the right of _finger 6676 // will be scanned as we iterate over the remainder of the 6677 // bit map 6678 } 6679 } 6680 6681 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6682 MemRegion span, 6683 CMSBitMap* bitMap, CMSMarkStack* markStack, 6684 HeapWord* finger, MarkFromRootsClosure* parent) : 6685 MetadataAwareOopClosure(collector->ref_processor()), 6686 _collector(collector), 6687 _span(span), 6688 _bitMap(bitMap), 6689 _markStack(markStack), 6690 _finger(finger), 6691 _parent(parent) 6692 { } 6693 6694 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6695 MemRegion span, 6696 CMSBitMap* bit_map, 6697 OopTaskQueue* work_queue, 6698 CMSMarkStack* overflow_stack, 6699 HeapWord* finger, 6700 HeapWord** global_finger_addr, 6701 ParMarkFromRootsClosure* parent) : 6702 MetadataAwareOopClosure(collector->ref_processor()), 6703 _collector(collector), 6704 _whole_span(collector->_span), 6705 _span(span), 6706 _bit_map(bit_map), 6707 _work_queue(work_queue), 6708 _overflow_stack(overflow_stack), 6709 _finger(finger), 6710 _global_finger_addr(global_finger_addr), 6711 _parent(parent) 6712 { } 6713 6714 // Assumes thread-safe access by callers, who are 6715 // responsible for mutual exclusion. 6716 void CMSCollector::lower_restart_addr(HeapWord* low) { 6717 assert(_span.contains(low), "Out of bounds addr"); 6718 if (_restart_addr == NULL) { 6719 _restart_addr = low; 6720 } else { 6721 _restart_addr = MIN2(_restart_addr, low); 6722 } 6723 } 6724 6725 // Upon stack overflow, we discard (part of) the stack, 6726 // remembering the least address amongst those discarded 6727 // in CMSCollector's _restart_address. 6728 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6729 // Remember the least grey address discarded 6730 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6731 _collector->lower_restart_addr(ra); 6732 _markStack->reset(); // discard stack contents 6733 _markStack->expand(); // expand the stack if possible 6734 } 6735 6736 // Upon stack overflow, we discard (part of) the stack, 6737 // remembering the least address amongst those discarded 6738 // in CMSCollector's _restart_address. 6739 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6740 // We need to do this under a mutex to prevent other 6741 // workers from interfering with the work done below. 6742 MutexLockerEx ml(_overflow_stack->par_lock(), 6743 Mutex::_no_safepoint_check_flag); 6744 // Remember the least grey address discarded 6745 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6746 _collector->lower_restart_addr(ra); 6747 _overflow_stack->reset(); // discard stack contents 6748 _overflow_stack->expand(); // expand the stack if possible 6749 } 6750 6751 void PushOrMarkClosure::do_oop(oop obj) { 6752 // Ignore mark word because we are running concurrent with mutators. 6753 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6754 HeapWord* addr = (HeapWord*)obj; 6755 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6756 // Oop lies in _span and isn't yet grey or black 6757 _bitMap->mark(addr); // now grey 6758 if (addr < _finger) { 6759 // the bit map iteration has already either passed, or 6760 // sampled, this bit in the bit map; we'll need to 6761 // use the marking stack to scan this oop's oops. 6762 bool simulate_overflow = false; 6763 NOT_PRODUCT( 6764 if (CMSMarkStackOverflowALot && 6765 _collector->simulate_overflow()) { 6766 // simulate a stack overflow 6767 simulate_overflow = true; 6768 } 6769 ) 6770 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6771 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6772 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6773 handle_stack_overflow(addr); 6774 } 6775 } 6776 // anything including and to the right of _finger 6777 // will be scanned as we iterate over the remainder of the 6778 // bit map 6779 do_yield_check(); 6780 } 6781 } 6782 6783 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6784 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6785 6786 void ParPushOrMarkClosure::do_oop(oop obj) { 6787 // Ignore mark word because we are running concurrent with mutators. 6788 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6789 HeapWord* addr = (HeapWord*)obj; 6790 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6791 // Oop lies in _span and isn't yet grey or black 6792 // We read the global_finger (volatile read) strictly after marking oop 6793 bool res = _bit_map->par_mark(addr); // now grey 6794 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6795 // Should we push this marked oop on our stack? 6796 // -- if someone else marked it, nothing to do 6797 // -- if target oop is above global finger nothing to do 6798 // -- if target oop is in chunk and above local finger 6799 // then nothing to do 6800 // -- else push on work queue 6801 if ( !res // someone else marked it, they will deal with it 6802 || (addr >= *gfa) // will be scanned in a later task 6803 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6804 return; 6805 } 6806 // the bit map iteration has already either passed, or 6807 // sampled, this bit in the bit map; we'll need to 6808 // use the marking stack to scan this oop's oops. 6809 bool simulate_overflow = false; 6810 NOT_PRODUCT( 6811 if (CMSMarkStackOverflowALot && 6812 _collector->simulate_overflow()) { 6813 // simulate a stack overflow 6814 simulate_overflow = true; 6815 } 6816 ) 6817 if (simulate_overflow || 6818 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6819 // stack overflow 6820 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6821 // We cannot assert that the overflow stack is full because 6822 // it may have been emptied since. 6823 assert(simulate_overflow || 6824 _work_queue->size() == _work_queue->max_elems(), 6825 "Else push should have succeeded"); 6826 handle_stack_overflow(addr); 6827 } 6828 do_yield_check(); 6829 } 6830 } 6831 6832 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6833 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6834 6835 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6836 MemRegion span, 6837 ReferenceProcessor* rp, 6838 CMSBitMap* bit_map, 6839 CMSBitMap* mod_union_table, 6840 CMSMarkStack* mark_stack, 6841 bool concurrent_precleaning): 6842 MetadataAwareOopClosure(rp), 6843 _collector(collector), 6844 _span(span), 6845 _bit_map(bit_map), 6846 _mod_union_table(mod_union_table), 6847 _mark_stack(mark_stack), 6848 _concurrent_precleaning(concurrent_precleaning) 6849 { 6850 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6851 } 6852 6853 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6854 // the non-parallel version (the parallel version appears further below.) 6855 void PushAndMarkClosure::do_oop(oop obj) { 6856 // Ignore mark word verification. If during concurrent precleaning, 6857 // the object monitor may be locked. If during the checkpoint 6858 // phases, the object may already have been reached by a different 6859 // path and may be at the end of the global overflow list (so 6860 // the mark word may be NULL). 6861 assert(obj->is_oop_or_null(true /* ignore mark word */), 6862 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6863 HeapWord* addr = (HeapWord*)obj; 6864 // Check if oop points into the CMS generation 6865 // and is not marked 6866 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6867 // a white object ... 6868 _bit_map->mark(addr); // ... now grey 6869 // push on the marking stack (grey set) 6870 bool simulate_overflow = false; 6871 NOT_PRODUCT( 6872 if (CMSMarkStackOverflowALot && 6873 _collector->simulate_overflow()) { 6874 // simulate a stack overflow 6875 simulate_overflow = true; 6876 } 6877 ) 6878 if (simulate_overflow || !_mark_stack->push(obj)) { 6879 if (_concurrent_precleaning) { 6880 // During precleaning we can just dirty the appropriate card(s) 6881 // in the mod union table, thus ensuring that the object remains 6882 // in the grey set and continue. In the case of object arrays 6883 // we need to dirty all of the cards that the object spans, 6884 // since the rescan of object arrays will be limited to the 6885 // dirty cards. 6886 // Note that no one can be interfering with us in this action 6887 // of dirtying the mod union table, so no locking or atomics 6888 // are required. 6889 if (obj->is_objArray()) { 6890 size_t sz = obj->size(); 6891 HeapWord* end_card_addr = (HeapWord*)round_to( 6892 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6893 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6894 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6895 _mod_union_table->mark_range(redirty_range); 6896 } else { 6897 _mod_union_table->mark(addr); 6898 } 6899 _collector->_ser_pmc_preclean_ovflw++; 6900 } else { 6901 // During the remark phase, we need to remember this oop 6902 // in the overflow list. 6903 _collector->push_on_overflow_list(obj); 6904 _collector->_ser_pmc_remark_ovflw++; 6905 } 6906 } 6907 } 6908 } 6909 6910 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6911 MemRegion span, 6912 ReferenceProcessor* rp, 6913 CMSBitMap* bit_map, 6914 OopTaskQueue* work_queue): 6915 MetadataAwareOopClosure(rp), 6916 _collector(collector), 6917 _span(span), 6918 _bit_map(bit_map), 6919 _work_queue(work_queue) 6920 { 6921 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6922 } 6923 6924 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6925 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6926 6927 // Grey object rescan during second checkpoint phase -- 6928 // the parallel version. 6929 void ParPushAndMarkClosure::do_oop(oop obj) { 6930 // In the assert below, we ignore the mark word because 6931 // this oop may point to an already visited object that is 6932 // on the overflow stack (in which case the mark word has 6933 // been hijacked for chaining into the overflow stack -- 6934 // if this is the last object in the overflow stack then 6935 // its mark word will be NULL). Because this object may 6936 // have been subsequently popped off the global overflow 6937 // stack, and the mark word possibly restored to the prototypical 6938 // value, by the time we get to examined this failing assert in 6939 // the debugger, is_oop_or_null(false) may subsequently start 6940 // to hold. 6941 assert(obj->is_oop_or_null(true), 6942 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6943 HeapWord* addr = (HeapWord*)obj; 6944 // Check if oop points into the CMS generation 6945 // and is not marked 6946 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6947 // a white object ... 6948 // If we manage to "claim" the object, by being the 6949 // first thread to mark it, then we push it on our 6950 // marking stack 6951 if (_bit_map->par_mark(addr)) { // ... now grey 6952 // push on work queue (grey set) 6953 bool simulate_overflow = false; 6954 NOT_PRODUCT( 6955 if (CMSMarkStackOverflowALot && 6956 _collector->par_simulate_overflow()) { 6957 // simulate a stack overflow 6958 simulate_overflow = true; 6959 } 6960 ) 6961 if (simulate_overflow || !_work_queue->push(obj)) { 6962 _collector->par_push_on_overflow_list(obj); 6963 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6964 } 6965 } // Else, some other thread got there first 6966 } 6967 } 6968 6969 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6970 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6971 6972 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6973 Mutex* bml = _collector->bitMapLock(); 6974 assert_lock_strong(bml); 6975 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6976 "CMS thread should hold CMS token"); 6977 6978 bml->unlock(); 6979 ConcurrentMarkSweepThread::desynchronize(true); 6980 6981 _collector->stopTimer(); 6982 _collector->incrementYields(); 6983 6984 // See the comment in coordinator_yield() 6985 for (unsigned i = 0; i < CMSYieldSleepCount && 6986 ConcurrentMarkSweepThread::should_yield() && 6987 !CMSCollector::foregroundGCIsActive(); ++i) { 6988 os::sleep(Thread::current(), 1, false); 6989 } 6990 6991 ConcurrentMarkSweepThread::synchronize(true); 6992 bml->lock(); 6993 6994 _collector->startTimer(); 6995 } 6996 6997 bool CMSPrecleanRefsYieldClosure::should_return() { 6998 if (ConcurrentMarkSweepThread::should_yield()) { 6999 do_yield_work(); 7000 } 7001 return _collector->foregroundGCIsActive(); 7002 } 7003 7004 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7005 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7006 "mr should be aligned to start at a card boundary"); 7007 // We'd like to assert: 7008 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7009 // "mr should be a range of cards"); 7010 // However, that would be too strong in one case -- the last 7011 // partition ends at _unallocated_block which, in general, can be 7012 // an arbitrary boundary, not necessarily card aligned. 7013 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7014 _space->object_iterate_mem(mr, &_scan_cl); 7015 } 7016 7017 SweepClosure::SweepClosure(CMSCollector* collector, 7018 ConcurrentMarkSweepGeneration* g, 7019 CMSBitMap* bitMap, bool should_yield) : 7020 _collector(collector), 7021 _g(g), 7022 _sp(g->cmsSpace()), 7023 _limit(_sp->sweep_limit()), 7024 _freelistLock(_sp->freelistLock()), 7025 _bitMap(bitMap), 7026 _yield(should_yield), 7027 _inFreeRange(false), // No free range at beginning of sweep 7028 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7029 _lastFreeRangeCoalesced(false), 7030 _freeFinger(g->used_region().start()) 7031 { 7032 NOT_PRODUCT( 7033 _numObjectsFreed = 0; 7034 _numWordsFreed = 0; 7035 _numObjectsLive = 0; 7036 _numWordsLive = 0; 7037 _numObjectsAlreadyFree = 0; 7038 _numWordsAlreadyFree = 0; 7039 _last_fc = NULL; 7040 7041 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7042 _sp->dictionary()->initialize_dict_returned_bytes(); 7043 ) 7044 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7045 "sweep _limit out of bounds"); 7046 log_develop_trace(gc, sweep)("===================="); 7047 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7048 } 7049 7050 void SweepClosure::print_on(outputStream* st) const { 7051 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7052 p2i(_sp->bottom()), p2i(_sp->end())); 7053 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7054 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7055 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7056 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7057 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7058 } 7059 7060 #ifndef PRODUCT 7061 // Assertion checking only: no useful work in product mode -- 7062 // however, if any of the flags below become product flags, 7063 // you may need to review this code to see if it needs to be 7064 // enabled in product mode. 7065 SweepClosure::~SweepClosure() { 7066 assert_lock_strong(_freelistLock); 7067 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7068 "sweep _limit out of bounds"); 7069 if (inFreeRange()) { 7070 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7071 print(); 7072 ShouldNotReachHere(); 7073 } 7074 7075 if (log_is_enabled(Debug, gc, sweep)) { 7076 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7077 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7078 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7079 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7080 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7081 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7082 } 7083 7084 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7085 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7086 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7087 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7088 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7089 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7090 } 7091 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7092 log_develop_trace(gc, sweep)("================"); 7093 } 7094 #endif // PRODUCT 7095 7096 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7097 bool freeRangeInFreeLists) { 7098 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7099 p2i(freeFinger), freeRangeInFreeLists); 7100 assert(!inFreeRange(), "Trampling existing free range"); 7101 set_inFreeRange(true); 7102 set_lastFreeRangeCoalesced(false); 7103 7104 set_freeFinger(freeFinger); 7105 set_freeRangeInFreeLists(freeRangeInFreeLists); 7106 if (CMSTestInFreeList) { 7107 if (freeRangeInFreeLists) { 7108 FreeChunk* fc = (FreeChunk*) freeFinger; 7109 assert(fc->is_free(), "A chunk on the free list should be free."); 7110 assert(fc->size() > 0, "Free range should have a size"); 7111 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7112 } 7113 } 7114 } 7115 7116 // Note that the sweeper runs concurrently with mutators. Thus, 7117 // it is possible for direct allocation in this generation to happen 7118 // in the middle of the sweep. Note that the sweeper also coalesces 7119 // contiguous free blocks. Thus, unless the sweeper and the allocator 7120 // synchronize appropriately freshly allocated blocks may get swept up. 7121 // This is accomplished by the sweeper locking the free lists while 7122 // it is sweeping. Thus blocks that are determined to be free are 7123 // indeed free. There is however one additional complication: 7124 // blocks that have been allocated since the final checkpoint and 7125 // mark, will not have been marked and so would be treated as 7126 // unreachable and swept up. To prevent this, the allocator marks 7127 // the bit map when allocating during the sweep phase. This leads, 7128 // however, to a further complication -- objects may have been allocated 7129 // but not yet initialized -- in the sense that the header isn't yet 7130 // installed. The sweeper can not then determine the size of the block 7131 // in order to skip over it. To deal with this case, we use a technique 7132 // (due to Printezis) to encode such uninitialized block sizes in the 7133 // bit map. Since the bit map uses a bit per every HeapWord, but the 7134 // CMS generation has a minimum object size of 3 HeapWords, it follows 7135 // that "normal marks" won't be adjacent in the bit map (there will 7136 // always be at least two 0 bits between successive 1 bits). We make use 7137 // of these "unused" bits to represent uninitialized blocks -- the bit 7138 // corresponding to the start of the uninitialized object and the next 7139 // bit are both set. Finally, a 1 bit marks the end of the object that 7140 // started with the two consecutive 1 bits to indicate its potentially 7141 // uninitialized state. 7142 7143 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7144 FreeChunk* fc = (FreeChunk*)addr; 7145 size_t res; 7146 7147 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7148 // than "addr == _limit" because although _limit was a block boundary when 7149 // we started the sweep, it may no longer be one because heap expansion 7150 // may have caused us to coalesce the block ending at the address _limit 7151 // with a newly expanded chunk (this happens when _limit was set to the 7152 // previous _end of the space), so we may have stepped past _limit: 7153 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7154 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7155 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7156 "sweep _limit out of bounds"); 7157 assert(addr < _sp->end(), "addr out of bounds"); 7158 // Flush any free range we might be holding as a single 7159 // coalesced chunk to the appropriate free list. 7160 if (inFreeRange()) { 7161 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7162 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7163 flush_cur_free_chunk(freeFinger(), 7164 pointer_delta(addr, freeFinger())); 7165 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7166 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7167 lastFreeRangeCoalesced() ? 1 : 0); 7168 } 7169 7170 // help the iterator loop finish 7171 return pointer_delta(_sp->end(), addr); 7172 } 7173 7174 assert(addr < _limit, "sweep invariant"); 7175 // check if we should yield 7176 do_yield_check(addr); 7177 if (fc->is_free()) { 7178 // Chunk that is already free 7179 res = fc->size(); 7180 do_already_free_chunk(fc); 7181 debug_only(_sp->verifyFreeLists()); 7182 // If we flush the chunk at hand in lookahead_and_flush() 7183 // and it's coalesced with a preceding chunk, then the 7184 // process of "mangling" the payload of the coalesced block 7185 // will cause erasure of the size information from the 7186 // (erstwhile) header of all the coalesced blocks but the 7187 // first, so the first disjunct in the assert will not hold 7188 // in that specific case (in which case the second disjunct 7189 // will hold). 7190 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7191 "Otherwise the size info doesn't change at this step"); 7192 NOT_PRODUCT( 7193 _numObjectsAlreadyFree++; 7194 _numWordsAlreadyFree += res; 7195 ) 7196 NOT_PRODUCT(_last_fc = fc;) 7197 } else if (!_bitMap->isMarked(addr)) { 7198 // Chunk is fresh garbage 7199 res = do_garbage_chunk(fc); 7200 debug_only(_sp->verifyFreeLists()); 7201 NOT_PRODUCT( 7202 _numObjectsFreed++; 7203 _numWordsFreed += res; 7204 ) 7205 } else { 7206 // Chunk that is alive. 7207 res = do_live_chunk(fc); 7208 debug_only(_sp->verifyFreeLists()); 7209 NOT_PRODUCT( 7210 _numObjectsLive++; 7211 _numWordsLive += res; 7212 ) 7213 } 7214 return res; 7215 } 7216 7217 // For the smart allocation, record following 7218 // split deaths - a free chunk is removed from its free list because 7219 // it is being split into two or more chunks. 7220 // split birth - a free chunk is being added to its free list because 7221 // a larger free chunk has been split and resulted in this free chunk. 7222 // coal death - a free chunk is being removed from its free list because 7223 // it is being coalesced into a large free chunk. 7224 // coal birth - a free chunk is being added to its free list because 7225 // it was created when two or more free chunks where coalesced into 7226 // this free chunk. 7227 // 7228 // These statistics are used to determine the desired number of free 7229 // chunks of a given size. The desired number is chosen to be relative 7230 // to the end of a CMS sweep. The desired number at the end of a sweep 7231 // is the 7232 // count-at-end-of-previous-sweep (an amount that was enough) 7233 // - count-at-beginning-of-current-sweep (the excess) 7234 // + split-births (gains in this size during interval) 7235 // - split-deaths (demands on this size during interval) 7236 // where the interval is from the end of one sweep to the end of the 7237 // next. 7238 // 7239 // When sweeping the sweeper maintains an accumulated chunk which is 7240 // the chunk that is made up of chunks that have been coalesced. That 7241 // will be termed the left-hand chunk. A new chunk of garbage that 7242 // is being considered for coalescing will be referred to as the 7243 // right-hand chunk. 7244 // 7245 // When making a decision on whether to coalesce a right-hand chunk with 7246 // the current left-hand chunk, the current count vs. the desired count 7247 // of the left-hand chunk is considered. Also if the right-hand chunk 7248 // is near the large chunk at the end of the heap (see 7249 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7250 // left-hand chunk is coalesced. 7251 // 7252 // When making a decision about whether to split a chunk, the desired count 7253 // vs. the current count of the candidate to be split is also considered. 7254 // If the candidate is underpopulated (currently fewer chunks than desired) 7255 // a chunk of an overpopulated (currently more chunks than desired) size may 7256 // be chosen. The "hint" associated with a free list, if non-null, points 7257 // to a free list which may be overpopulated. 7258 // 7259 7260 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7261 const size_t size = fc->size(); 7262 // Chunks that cannot be coalesced are not in the 7263 // free lists. 7264 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7265 assert(_sp->verify_chunk_in_free_list(fc), 7266 "free chunk should be in free lists"); 7267 } 7268 // a chunk that is already free, should not have been 7269 // marked in the bit map 7270 HeapWord* const addr = (HeapWord*) fc; 7271 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7272 // Verify that the bit map has no bits marked between 7273 // addr and purported end of this block. 7274 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7275 7276 // Some chunks cannot be coalesced under any circumstances. 7277 // See the definition of cantCoalesce(). 7278 if (!fc->cantCoalesce()) { 7279 // This chunk can potentially be coalesced. 7280 // All the work is done in 7281 do_post_free_or_garbage_chunk(fc, size); 7282 // Note that if the chunk is not coalescable (the else arm 7283 // below), we unconditionally flush, without needing to do 7284 // a "lookahead," as we do below. 7285 if (inFreeRange()) lookahead_and_flush(fc, size); 7286 } else { 7287 // Code path common to both original and adaptive free lists. 7288 7289 // cant coalesce with previous block; this should be treated 7290 // as the end of a free run if any 7291 if (inFreeRange()) { 7292 // we kicked some butt; time to pick up the garbage 7293 assert(freeFinger() < addr, "freeFinger points too high"); 7294 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7295 } 7296 // else, nothing to do, just continue 7297 } 7298 } 7299 7300 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7301 // This is a chunk of garbage. It is not in any free list. 7302 // Add it to a free list or let it possibly be coalesced into 7303 // a larger chunk. 7304 HeapWord* const addr = (HeapWord*) fc; 7305 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7306 7307 // Verify that the bit map has no bits marked between 7308 // addr and purported end of just dead object. 7309 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7310 do_post_free_or_garbage_chunk(fc, size); 7311 7312 assert(_limit >= addr + size, 7313 "A freshly garbage chunk can't possibly straddle over _limit"); 7314 if (inFreeRange()) lookahead_and_flush(fc, size); 7315 return size; 7316 } 7317 7318 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7319 HeapWord* addr = (HeapWord*) fc; 7320 // The sweeper has just found a live object. Return any accumulated 7321 // left hand chunk to the free lists. 7322 if (inFreeRange()) { 7323 assert(freeFinger() < addr, "freeFinger points too high"); 7324 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7325 } 7326 7327 // This object is live: we'd normally expect this to be 7328 // an oop, and like to assert the following: 7329 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7330 // However, as we commented above, this may be an object whose 7331 // header hasn't yet been initialized. 7332 size_t size; 7333 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7334 if (_bitMap->isMarked(addr + 1)) { 7335 // Determine the size from the bit map, rather than trying to 7336 // compute it from the object header. 7337 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7338 size = pointer_delta(nextOneAddr + 1, addr); 7339 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7340 "alignment problem"); 7341 7342 #ifdef ASSERT 7343 if (oop(addr)->klass_or_null() != NULL) { 7344 // Ignore mark word because we are running concurrent with mutators 7345 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7346 assert(size == 7347 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7348 "P-mark and computed size do not agree"); 7349 } 7350 #endif 7351 7352 } else { 7353 // This should be an initialized object that's alive. 7354 assert(oop(addr)->klass_or_null() != NULL, 7355 "Should be an initialized object"); 7356 // Ignore mark word because we are running concurrent with mutators 7357 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7358 // Verify that the bit map has no bits marked between 7359 // addr and purported end of this block. 7360 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7361 assert(size >= 3, "Necessary for Printezis marks to work"); 7362 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7363 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7364 } 7365 return size; 7366 } 7367 7368 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7369 size_t chunkSize) { 7370 // do_post_free_or_garbage_chunk() should only be called in the case 7371 // of the adaptive free list allocator. 7372 const bool fcInFreeLists = fc->is_free(); 7373 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7374 if (CMSTestInFreeList && fcInFreeLists) { 7375 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7376 } 7377 7378 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7379 7380 HeapWord* const fc_addr = (HeapWord*) fc; 7381 7382 bool coalesce = false; 7383 const size_t left = pointer_delta(fc_addr, freeFinger()); 7384 const size_t right = chunkSize; 7385 switch (FLSCoalescePolicy) { 7386 // numeric value forms a coalition aggressiveness metric 7387 case 0: { // never coalesce 7388 coalesce = false; 7389 break; 7390 } 7391 case 1: { // coalesce if left & right chunks on overpopulated lists 7392 coalesce = _sp->coalOverPopulated(left) && 7393 _sp->coalOverPopulated(right); 7394 break; 7395 } 7396 case 2: { // coalesce if left chunk on overpopulated list (default) 7397 coalesce = _sp->coalOverPopulated(left); 7398 break; 7399 } 7400 case 3: { // coalesce if left OR right chunk on overpopulated list 7401 coalesce = _sp->coalOverPopulated(left) || 7402 _sp->coalOverPopulated(right); 7403 break; 7404 } 7405 case 4: { // always coalesce 7406 coalesce = true; 7407 break; 7408 } 7409 default: 7410 ShouldNotReachHere(); 7411 } 7412 7413 // Should the current free range be coalesced? 7414 // If the chunk is in a free range and either we decided to coalesce above 7415 // or the chunk is near the large block at the end of the heap 7416 // (isNearLargestChunk() returns true), then coalesce this chunk. 7417 const bool doCoalesce = inFreeRange() 7418 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7419 if (doCoalesce) { 7420 // Coalesce the current free range on the left with the new 7421 // chunk on the right. If either is on a free list, 7422 // it must be removed from the list and stashed in the closure. 7423 if (freeRangeInFreeLists()) { 7424 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7425 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7426 "Size of free range is inconsistent with chunk size."); 7427 if (CMSTestInFreeList) { 7428 assert(_sp->verify_chunk_in_free_list(ffc), 7429 "Chunk is not in free lists"); 7430 } 7431 _sp->coalDeath(ffc->size()); 7432 _sp->removeFreeChunkFromFreeLists(ffc); 7433 set_freeRangeInFreeLists(false); 7434 } 7435 if (fcInFreeLists) { 7436 _sp->coalDeath(chunkSize); 7437 assert(fc->size() == chunkSize, 7438 "The chunk has the wrong size or is not in the free lists"); 7439 _sp->removeFreeChunkFromFreeLists(fc); 7440 } 7441 set_lastFreeRangeCoalesced(true); 7442 print_free_block_coalesced(fc); 7443 } else { // not in a free range and/or should not coalesce 7444 // Return the current free range and start a new one. 7445 if (inFreeRange()) { 7446 // In a free range but cannot coalesce with the right hand chunk. 7447 // Put the current free range into the free lists. 7448 flush_cur_free_chunk(freeFinger(), 7449 pointer_delta(fc_addr, freeFinger())); 7450 } 7451 // Set up for new free range. Pass along whether the right hand 7452 // chunk is in the free lists. 7453 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7454 } 7455 } 7456 7457 // Lookahead flush: 7458 // If we are tracking a free range, and this is the last chunk that 7459 // we'll look at because its end crosses past _limit, we'll preemptively 7460 // flush it along with any free range we may be holding on to. Note that 7461 // this can be the case only for an already free or freshly garbage 7462 // chunk. If this block is an object, it can never straddle 7463 // over _limit. The "straddling" occurs when _limit is set at 7464 // the previous end of the space when this cycle started, and 7465 // a subsequent heap expansion caused the previously co-terminal 7466 // free block to be coalesced with the newly expanded portion, 7467 // thus rendering _limit a non-block-boundary making it dangerous 7468 // for the sweeper to step over and examine. 7469 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7470 assert(inFreeRange(), "Should only be called if currently in a free range."); 7471 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7472 assert(_sp->used_region().contains(eob - 1), 7473 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7474 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7475 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7476 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7477 if (eob >= _limit) { 7478 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7479 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7480 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7481 "[" PTR_FORMAT "," PTR_FORMAT ")", 7482 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7483 // Return the storage we are tracking back into the free lists. 7484 log_develop_trace(gc, sweep)("Flushing ... "); 7485 assert(freeFinger() < eob, "Error"); 7486 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7487 } 7488 } 7489 7490 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7491 assert(inFreeRange(), "Should only be called if currently in a free range."); 7492 assert(size > 0, 7493 "A zero sized chunk cannot be added to the free lists."); 7494 if (!freeRangeInFreeLists()) { 7495 if (CMSTestInFreeList) { 7496 FreeChunk* fc = (FreeChunk*) chunk; 7497 fc->set_size(size); 7498 assert(!_sp->verify_chunk_in_free_list(fc), 7499 "chunk should not be in free lists yet"); 7500 } 7501 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7502 // A new free range is going to be starting. The current 7503 // free range has not been added to the free lists yet or 7504 // was removed so add it back. 7505 // If the current free range was coalesced, then the death 7506 // of the free range was recorded. Record a birth now. 7507 if (lastFreeRangeCoalesced()) { 7508 _sp->coalBirth(size); 7509 } 7510 _sp->addChunkAndRepairOffsetTable(chunk, size, 7511 lastFreeRangeCoalesced()); 7512 } else { 7513 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7514 } 7515 set_inFreeRange(false); 7516 set_freeRangeInFreeLists(false); 7517 } 7518 7519 // We take a break if we've been at this for a while, 7520 // so as to avoid monopolizing the locks involved. 7521 void SweepClosure::do_yield_work(HeapWord* addr) { 7522 // Return current free chunk being used for coalescing (if any) 7523 // to the appropriate freelist. After yielding, the next 7524 // free block encountered will start a coalescing range of 7525 // free blocks. If the next free block is adjacent to the 7526 // chunk just flushed, they will need to wait for the next 7527 // sweep to be coalesced. 7528 if (inFreeRange()) { 7529 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7530 } 7531 7532 // First give up the locks, then yield, then re-lock. 7533 // We should probably use a constructor/destructor idiom to 7534 // do this unlock/lock or modify the MutexUnlocker class to 7535 // serve our purpose. XXX 7536 assert_lock_strong(_bitMap->lock()); 7537 assert_lock_strong(_freelistLock); 7538 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7539 "CMS thread should hold CMS token"); 7540 _bitMap->lock()->unlock(); 7541 _freelistLock->unlock(); 7542 ConcurrentMarkSweepThread::desynchronize(true); 7543 _collector->stopTimer(); 7544 _collector->incrementYields(); 7545 7546 // See the comment in coordinator_yield() 7547 for (unsigned i = 0; i < CMSYieldSleepCount && 7548 ConcurrentMarkSweepThread::should_yield() && 7549 !CMSCollector::foregroundGCIsActive(); ++i) { 7550 os::sleep(Thread::current(), 1, false); 7551 } 7552 7553 ConcurrentMarkSweepThread::synchronize(true); 7554 _freelistLock->lock(); 7555 _bitMap->lock()->lock_without_safepoint_check(); 7556 _collector->startTimer(); 7557 } 7558 7559 #ifndef PRODUCT 7560 // This is actually very useful in a product build if it can 7561 // be called from the debugger. Compile it into the product 7562 // as needed. 7563 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7564 return debug_cms_space->verify_chunk_in_free_list(fc); 7565 } 7566 #endif 7567 7568 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7569 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7570 p2i(fc), fc->size()); 7571 } 7572 7573 // CMSIsAliveClosure 7574 bool CMSIsAliveClosure::do_object_b(oop obj) { 7575 HeapWord* addr = (HeapWord*)obj; 7576 return addr != NULL && 7577 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7578 } 7579 7580 7581 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7582 MemRegion span, 7583 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7584 bool cpc): 7585 _collector(collector), 7586 _span(span), 7587 _bit_map(bit_map), 7588 _mark_stack(mark_stack), 7589 _concurrent_precleaning(cpc) { 7590 assert(!_span.is_empty(), "Empty span could spell trouble"); 7591 } 7592 7593 7594 // CMSKeepAliveClosure: the serial version 7595 void CMSKeepAliveClosure::do_oop(oop obj) { 7596 HeapWord* addr = (HeapWord*)obj; 7597 if (_span.contains(addr) && 7598 !_bit_map->isMarked(addr)) { 7599 _bit_map->mark(addr); 7600 bool simulate_overflow = false; 7601 NOT_PRODUCT( 7602 if (CMSMarkStackOverflowALot && 7603 _collector->simulate_overflow()) { 7604 // simulate a stack overflow 7605 simulate_overflow = true; 7606 } 7607 ) 7608 if (simulate_overflow || !_mark_stack->push(obj)) { 7609 if (_concurrent_precleaning) { 7610 // We dirty the overflown object and let the remark 7611 // phase deal with it. 7612 assert(_collector->overflow_list_is_empty(), "Error"); 7613 // In the case of object arrays, we need to dirty all of 7614 // the cards that the object spans. No locking or atomics 7615 // are needed since no one else can be mutating the mod union 7616 // table. 7617 if (obj->is_objArray()) { 7618 size_t sz = obj->size(); 7619 HeapWord* end_card_addr = 7620 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7621 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7622 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7623 _collector->_modUnionTable.mark_range(redirty_range); 7624 } else { 7625 _collector->_modUnionTable.mark(addr); 7626 } 7627 _collector->_ser_kac_preclean_ovflw++; 7628 } else { 7629 _collector->push_on_overflow_list(obj); 7630 _collector->_ser_kac_ovflw++; 7631 } 7632 } 7633 } 7634 } 7635 7636 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7637 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7638 7639 // CMSParKeepAliveClosure: a parallel version of the above. 7640 // The work queues are private to each closure (thread), 7641 // but (may be) available for stealing by other threads. 7642 void CMSParKeepAliveClosure::do_oop(oop obj) { 7643 HeapWord* addr = (HeapWord*)obj; 7644 if (_span.contains(addr) && 7645 !_bit_map->isMarked(addr)) { 7646 // In general, during recursive tracing, several threads 7647 // may be concurrently getting here; the first one to 7648 // "tag" it, claims it. 7649 if (_bit_map->par_mark(addr)) { 7650 bool res = _work_queue->push(obj); 7651 assert(res, "Low water mark should be much less than capacity"); 7652 // Do a recursive trim in the hope that this will keep 7653 // stack usage lower, but leave some oops for potential stealers 7654 trim_queue(_low_water_mark); 7655 } // Else, another thread got there first 7656 } 7657 } 7658 7659 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7660 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7661 7662 void CMSParKeepAliveClosure::trim_queue(uint max) { 7663 while (_work_queue->size() > max) { 7664 oop new_oop; 7665 if (_work_queue->pop_local(new_oop)) { 7666 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7667 assert(_bit_map->isMarked((HeapWord*)new_oop), 7668 "no white objects on this stack!"); 7669 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7670 // iterate over the oops in this oop, marking and pushing 7671 // the ones in CMS heap (i.e. in _span). 7672 new_oop->oop_iterate(&_mark_and_push); 7673 } 7674 } 7675 } 7676 7677 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7678 CMSCollector* collector, 7679 MemRegion span, CMSBitMap* bit_map, 7680 OopTaskQueue* work_queue): 7681 _collector(collector), 7682 _span(span), 7683 _bit_map(bit_map), 7684 _work_queue(work_queue) { } 7685 7686 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7687 HeapWord* addr = (HeapWord*)obj; 7688 if (_span.contains(addr) && 7689 !_bit_map->isMarked(addr)) { 7690 if (_bit_map->par_mark(addr)) { 7691 bool simulate_overflow = false; 7692 NOT_PRODUCT( 7693 if (CMSMarkStackOverflowALot && 7694 _collector->par_simulate_overflow()) { 7695 // simulate a stack overflow 7696 simulate_overflow = true; 7697 } 7698 ) 7699 if (simulate_overflow || !_work_queue->push(obj)) { 7700 _collector->par_push_on_overflow_list(obj); 7701 _collector->_par_kac_ovflw++; 7702 } 7703 } // Else another thread got there already 7704 } 7705 } 7706 7707 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7708 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7709 7710 ////////////////////////////////////////////////////////////////// 7711 // CMSExpansionCause ///////////////////////////// 7712 ////////////////////////////////////////////////////////////////// 7713 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7714 switch (cause) { 7715 case _no_expansion: 7716 return "No expansion"; 7717 case _satisfy_free_ratio: 7718 return "Free ratio"; 7719 case _satisfy_promotion: 7720 return "Satisfy promotion"; 7721 case _satisfy_allocation: 7722 return "allocation"; 7723 case _allocate_par_lab: 7724 return "Par LAB"; 7725 case _allocate_par_spooling_space: 7726 return "Par Spooling Space"; 7727 case _adaptive_size_policy: 7728 return "Ergonomics"; 7729 default: 7730 return "unknown"; 7731 } 7732 } 7733 7734 void CMSDrainMarkingStackClosure::do_void() { 7735 // the max number to take from overflow list at a time 7736 const size_t num = _mark_stack->capacity()/4; 7737 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7738 "Overflow list should be NULL during concurrent phases"); 7739 while (!_mark_stack->isEmpty() || 7740 // if stack is empty, check the overflow list 7741 _collector->take_from_overflow_list(num, _mark_stack)) { 7742 oop obj = _mark_stack->pop(); 7743 HeapWord* addr = (HeapWord*)obj; 7744 assert(_span.contains(addr), "Should be within span"); 7745 assert(_bit_map->isMarked(addr), "Should be marked"); 7746 assert(obj->is_oop(), "Should be an oop"); 7747 obj->oop_iterate(_keep_alive); 7748 } 7749 } 7750 7751 void CMSParDrainMarkingStackClosure::do_void() { 7752 // drain queue 7753 trim_queue(0); 7754 } 7755 7756 // Trim our work_queue so its length is below max at return 7757 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7758 while (_work_queue->size() > max) { 7759 oop new_oop; 7760 if (_work_queue->pop_local(new_oop)) { 7761 assert(new_oop->is_oop(), "Expected an oop"); 7762 assert(_bit_map->isMarked((HeapWord*)new_oop), 7763 "no white objects on this stack!"); 7764 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7765 // iterate over the oops in this oop, marking and pushing 7766 // the ones in CMS heap (i.e. in _span). 7767 new_oop->oop_iterate(&_mark_and_push); 7768 } 7769 } 7770 } 7771 7772 //////////////////////////////////////////////////////////////////// 7773 // Support for Marking Stack Overflow list handling and related code 7774 //////////////////////////////////////////////////////////////////// 7775 // Much of the following code is similar in shape and spirit to the 7776 // code used in ParNewGC. We should try and share that code 7777 // as much as possible in the future. 7778 7779 #ifndef PRODUCT 7780 // Debugging support for CMSStackOverflowALot 7781 7782 // It's OK to call this multi-threaded; the worst thing 7783 // that can happen is that we'll get a bunch of closely 7784 // spaced simulated overflows, but that's OK, in fact 7785 // probably good as it would exercise the overflow code 7786 // under contention. 7787 bool CMSCollector::simulate_overflow() { 7788 if (_overflow_counter-- <= 0) { // just being defensive 7789 _overflow_counter = CMSMarkStackOverflowInterval; 7790 return true; 7791 } else { 7792 return false; 7793 } 7794 } 7795 7796 bool CMSCollector::par_simulate_overflow() { 7797 return simulate_overflow(); 7798 } 7799 #endif 7800 7801 // Single-threaded 7802 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7803 assert(stack->isEmpty(), "Expected precondition"); 7804 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7805 size_t i = num; 7806 oop cur = _overflow_list; 7807 const markOop proto = markOopDesc::prototype(); 7808 NOT_PRODUCT(ssize_t n = 0;) 7809 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7810 next = oop(cur->mark()); 7811 cur->set_mark(proto); // until proven otherwise 7812 assert(cur->is_oop(), "Should be an oop"); 7813 bool res = stack->push(cur); 7814 assert(res, "Bit off more than can chew?"); 7815 NOT_PRODUCT(n++;) 7816 } 7817 _overflow_list = cur; 7818 #ifndef PRODUCT 7819 assert(_num_par_pushes >= n, "Too many pops?"); 7820 _num_par_pushes -=n; 7821 #endif 7822 return !stack->isEmpty(); 7823 } 7824 7825 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7826 // (MT-safe) Get a prefix of at most "num" from the list. 7827 // The overflow list is chained through the mark word of 7828 // each object in the list. We fetch the entire list, 7829 // break off a prefix of the right size and return the 7830 // remainder. If other threads try to take objects from 7831 // the overflow list at that time, they will wait for 7832 // some time to see if data becomes available. If (and 7833 // only if) another thread places one or more object(s) 7834 // on the global list before we have returned the suffix 7835 // to the global list, we will walk down our local list 7836 // to find its end and append the global list to 7837 // our suffix before returning it. This suffix walk can 7838 // prove to be expensive (quadratic in the amount of traffic) 7839 // when there are many objects in the overflow list and 7840 // there is much producer-consumer contention on the list. 7841 // *NOTE*: The overflow list manipulation code here and 7842 // in ParNewGeneration:: are very similar in shape, 7843 // except that in the ParNew case we use the old (from/eden) 7844 // copy of the object to thread the list via its klass word. 7845 // Because of the common code, if you make any changes in 7846 // the code below, please check the ParNew version to see if 7847 // similar changes might be needed. 7848 // CR 6797058 has been filed to consolidate the common code. 7849 bool CMSCollector::par_take_from_overflow_list(size_t num, 7850 OopTaskQueue* work_q, 7851 int no_of_gc_threads) { 7852 assert(work_q->size() == 0, "First empty local work queue"); 7853 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7854 if (_overflow_list == NULL) { 7855 return false; 7856 } 7857 // Grab the entire list; we'll put back a suffix 7858 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7859 Thread* tid = Thread::current(); 7860 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7861 // set to ParallelGCThreads. 7862 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7863 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7864 // If the list is busy, we spin for a short while, 7865 // sleeping between attempts to get the list. 7866 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7867 os::sleep(tid, sleep_time_millis, false); 7868 if (_overflow_list == NULL) { 7869 // Nothing left to take 7870 return false; 7871 } else if (_overflow_list != BUSY) { 7872 // Try and grab the prefix 7873 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7874 } 7875 } 7876 // If the list was found to be empty, or we spun long 7877 // enough, we give up and return empty-handed. If we leave 7878 // the list in the BUSY state below, it must be the case that 7879 // some other thread holds the overflow list and will set it 7880 // to a non-BUSY state in the future. 7881 if (prefix == NULL || prefix == BUSY) { 7882 // Nothing to take or waited long enough 7883 if (prefix == NULL) { 7884 // Write back the NULL in case we overwrote it with BUSY above 7885 // and it is still the same value. 7886 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7887 } 7888 return false; 7889 } 7890 assert(prefix != NULL && prefix != BUSY, "Error"); 7891 size_t i = num; 7892 oop cur = prefix; 7893 // Walk down the first "num" objects, unless we reach the end. 7894 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7895 if (cur->mark() == NULL) { 7896 // We have "num" or fewer elements in the list, so there 7897 // is nothing to return to the global list. 7898 // Write back the NULL in lieu of the BUSY we wrote 7899 // above, if it is still the same value. 7900 if (_overflow_list == BUSY) { 7901 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7902 } 7903 } else { 7904 // Chop off the suffix and return it to the global list. 7905 assert(cur->mark() != BUSY, "Error"); 7906 oop suffix_head = cur->mark(); // suffix will be put back on global list 7907 cur->set_mark(NULL); // break off suffix 7908 // It's possible that the list is still in the empty(busy) state 7909 // we left it in a short while ago; in that case we may be 7910 // able to place back the suffix without incurring the cost 7911 // of a walk down the list. 7912 oop observed_overflow_list = _overflow_list; 7913 oop cur_overflow_list = observed_overflow_list; 7914 bool attached = false; 7915 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7916 observed_overflow_list = 7917 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7918 if (cur_overflow_list == observed_overflow_list) { 7919 attached = true; 7920 break; 7921 } else cur_overflow_list = observed_overflow_list; 7922 } 7923 if (!attached) { 7924 // Too bad, someone else sneaked in (at least) an element; we'll need 7925 // to do a splice. Find tail of suffix so we can prepend suffix to global 7926 // list. 7927 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7928 oop suffix_tail = cur; 7929 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7930 "Tautology"); 7931 observed_overflow_list = _overflow_list; 7932 do { 7933 cur_overflow_list = observed_overflow_list; 7934 if (cur_overflow_list != BUSY) { 7935 // Do the splice ... 7936 suffix_tail->set_mark(markOop(cur_overflow_list)); 7937 } else { // cur_overflow_list == BUSY 7938 suffix_tail->set_mark(NULL); 7939 } 7940 // ... and try to place spliced list back on overflow_list ... 7941 observed_overflow_list = 7942 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7943 } while (cur_overflow_list != observed_overflow_list); 7944 // ... until we have succeeded in doing so. 7945 } 7946 } 7947 7948 // Push the prefix elements on work_q 7949 assert(prefix != NULL, "control point invariant"); 7950 const markOop proto = markOopDesc::prototype(); 7951 oop next; 7952 NOT_PRODUCT(ssize_t n = 0;) 7953 for (cur = prefix; cur != NULL; cur = next) { 7954 next = oop(cur->mark()); 7955 cur->set_mark(proto); // until proven otherwise 7956 assert(cur->is_oop(), "Should be an oop"); 7957 bool res = work_q->push(cur); 7958 assert(res, "Bit off more than we can chew?"); 7959 NOT_PRODUCT(n++;) 7960 } 7961 #ifndef PRODUCT 7962 assert(_num_par_pushes >= n, "Too many pops?"); 7963 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7964 #endif 7965 return true; 7966 } 7967 7968 // Single-threaded 7969 void CMSCollector::push_on_overflow_list(oop p) { 7970 NOT_PRODUCT(_num_par_pushes++;) 7971 assert(p->is_oop(), "Not an oop"); 7972 preserve_mark_if_necessary(p); 7973 p->set_mark((markOop)_overflow_list); 7974 _overflow_list = p; 7975 } 7976 7977 // Multi-threaded; use CAS to prepend to overflow list 7978 void CMSCollector::par_push_on_overflow_list(oop p) { 7979 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7980 assert(p->is_oop(), "Not an oop"); 7981 par_preserve_mark_if_necessary(p); 7982 oop observed_overflow_list = _overflow_list; 7983 oop cur_overflow_list; 7984 do { 7985 cur_overflow_list = observed_overflow_list; 7986 if (cur_overflow_list != BUSY) { 7987 p->set_mark(markOop(cur_overflow_list)); 7988 } else { 7989 p->set_mark(NULL); 7990 } 7991 observed_overflow_list = 7992 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7993 } while (cur_overflow_list != observed_overflow_list); 7994 } 7995 #undef BUSY 7996 7997 // Single threaded 7998 // General Note on GrowableArray: pushes may silently fail 7999 // because we are (temporarily) out of C-heap for expanding 8000 // the stack. The problem is quite ubiquitous and affects 8001 // a lot of code in the JVM. The prudent thing for GrowableArray 8002 // to do (for now) is to exit with an error. However, that may 8003 // be too draconian in some cases because the caller may be 8004 // able to recover without much harm. For such cases, we 8005 // should probably introduce a "soft_push" method which returns 8006 // an indication of success or failure with the assumption that 8007 // the caller may be able to recover from a failure; code in 8008 // the VM can then be changed, incrementally, to deal with such 8009 // failures where possible, thus, incrementally hardening the VM 8010 // in such low resource situations. 8011 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8012 _preserved_oop_stack.push(p); 8013 _preserved_mark_stack.push(m); 8014 assert(m == p->mark(), "Mark word changed"); 8015 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8016 "bijection"); 8017 } 8018 8019 // Single threaded 8020 void CMSCollector::preserve_mark_if_necessary(oop p) { 8021 markOop m = p->mark(); 8022 if (m->must_be_preserved(p)) { 8023 preserve_mark_work(p, m); 8024 } 8025 } 8026 8027 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8028 markOop m = p->mark(); 8029 if (m->must_be_preserved(p)) { 8030 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8031 // Even though we read the mark word without holding 8032 // the lock, we are assured that it will not change 8033 // because we "own" this oop, so no other thread can 8034 // be trying to push it on the overflow list; see 8035 // the assertion in preserve_mark_work() that checks 8036 // that m == p->mark(). 8037 preserve_mark_work(p, m); 8038 } 8039 } 8040 8041 // We should be able to do this multi-threaded, 8042 // a chunk of stack being a task (this is 8043 // correct because each oop only ever appears 8044 // once in the overflow list. However, it's 8045 // not very easy to completely overlap this with 8046 // other operations, so will generally not be done 8047 // until all work's been completed. Because we 8048 // expect the preserved oop stack (set) to be small, 8049 // it's probably fine to do this single-threaded. 8050 // We can explore cleverer concurrent/overlapped/parallel 8051 // processing of preserved marks if we feel the 8052 // need for this in the future. Stack overflow should 8053 // be so rare in practice and, when it happens, its 8054 // effect on performance so great that this will 8055 // likely just be in the noise anyway. 8056 void CMSCollector::restore_preserved_marks_if_any() { 8057 assert(SafepointSynchronize::is_at_safepoint(), 8058 "world should be stopped"); 8059 assert(Thread::current()->is_ConcurrentGC_thread() || 8060 Thread::current()->is_VM_thread(), 8061 "should be single-threaded"); 8062 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8063 "bijection"); 8064 8065 while (!_preserved_oop_stack.is_empty()) { 8066 oop p = _preserved_oop_stack.pop(); 8067 assert(p->is_oop(), "Should be an oop"); 8068 assert(_span.contains(p), "oop should be in _span"); 8069 assert(p->mark() == markOopDesc::prototype(), 8070 "Set when taken from overflow list"); 8071 markOop m = _preserved_mark_stack.pop(); 8072 p->set_mark(m); 8073 } 8074 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8075 "stacks were cleared above"); 8076 } 8077 8078 #ifndef PRODUCT 8079 bool CMSCollector::no_preserved_marks() const { 8080 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8081 } 8082 #endif 8083 8084 // Transfer some number of overflown objects to usual marking 8085 // stack. Return true if some objects were transferred. 8086 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8087 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8088 (size_t)ParGCDesiredObjsFromOverflowList); 8089 8090 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8091 assert(_collector->overflow_list_is_empty() || res, 8092 "If list is not empty, we should have taken something"); 8093 assert(!res || !_mark_stack->isEmpty(), 8094 "If we took something, it should now be on our stack"); 8095 return res; 8096 } 8097 8098 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8099 size_t res = _sp->block_size_no_stall(addr, _collector); 8100 if (_sp->block_is_obj(addr)) { 8101 if (_live_bit_map->isMarked(addr)) { 8102 // It can't have been dead in a previous cycle 8103 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8104 } else { 8105 _dead_bit_map->mark(addr); // mark the dead object 8106 } 8107 } 8108 // Could be 0, if the block size could not be computed without stalling. 8109 return res; 8110 } 8111 8112 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8113 8114 switch (phase) { 8115 case CMSCollector::InitialMarking: 8116 initialize(true /* fullGC */ , 8117 cause /* cause of the GC */, 8118 true /* recordGCBeginTime */, 8119 true /* recordPreGCUsage */, 8120 false /* recordPeakUsage */, 8121 false /* recordPostGCusage */, 8122 true /* recordAccumulatedGCTime */, 8123 false /* recordGCEndTime */, 8124 false /* countCollection */ ); 8125 break; 8126 8127 case CMSCollector::FinalMarking: 8128 initialize(true /* fullGC */ , 8129 cause /* cause of the GC */, 8130 false /* recordGCBeginTime */, 8131 false /* recordPreGCUsage */, 8132 false /* recordPeakUsage */, 8133 false /* recordPostGCusage */, 8134 true /* recordAccumulatedGCTime */, 8135 false /* recordGCEndTime */, 8136 false /* countCollection */ ); 8137 break; 8138 8139 case CMSCollector::Sweeping: 8140 initialize(true /* fullGC */ , 8141 cause /* cause of the GC */, 8142 false /* recordGCBeginTime */, 8143 false /* recordPreGCUsage */, 8144 true /* recordPeakUsage */, 8145 true /* recordPostGCusage */, 8146 false /* recordAccumulatedGCTime */, 8147 true /* recordGCEndTime */, 8148 true /* countCollection */ ); 8149 break; 8150 8151 default: 8152 ShouldNotReachHere(); 8153 } 8154 }