1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsHeap.hpp" 33 #include "gc/cms/cmsOopClosures.inline.hpp" 34 #include "gc/cms/compactibleFreeListSpace.hpp" 35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc/cms/concurrentMarkSweepThread.hpp" 37 #include "gc/cms/parNewGeneration.hpp" 38 #include "gc/cms/vmCMSOperations.hpp" 39 #include "gc/serial/genMarkSweep.hpp" 40 #include "gc/serial/tenuredGeneration.hpp" 41 #include "gc/shared/adaptiveSizePolicy.hpp" 42 #include "gc/shared/cardGeneration.inline.hpp" 43 #include "gc/shared/cardTableRS.hpp" 44 #include "gc/shared/collectedHeap.inline.hpp" 45 #include "gc/shared/collectorCounters.hpp" 46 #include "gc/shared/collectorPolicy.hpp" 47 #include "gc/shared/gcLocker.inline.hpp" 48 #include "gc/shared/gcPolicyCounters.hpp" 49 #include "gc/shared/gcTimer.hpp" 50 #include "gc/shared/gcTrace.hpp" 51 #include "gc/shared/gcTraceTime.inline.hpp" 52 #include "gc/shared/genCollectedHeap.hpp" 53 #include "gc/shared/genOopClosures.inline.hpp" 54 #include "gc/shared/isGCActiveMark.hpp" 55 #include "gc/shared/referencePolicy.hpp" 56 #include "gc/shared/strongRootsScope.hpp" 57 #include "gc/shared/taskqueue.inline.hpp" 58 #include "gc/shared/weakProcessor.hpp" 59 #include "logging/log.hpp" 60 #include "logging/logStream.hpp" 61 #include "memory/allocation.hpp" 62 #include "memory/iterator.inline.hpp" 63 #include "memory/padded.hpp" 64 #include "memory/resourceArea.hpp" 65 #include "oops/oop.inline.hpp" 66 #include "prims/jvmtiExport.hpp" 67 #include "runtime/atomic.hpp" 68 #include "runtime/globals_extension.hpp" 69 #include "runtime/handles.inline.hpp" 70 #include "runtime/java.hpp" 71 #include "runtime/orderAccess.inline.hpp" 72 #include "runtime/timer.hpp" 73 #include "runtime/vmThread.hpp" 74 #include "services/memoryService.hpp" 75 #include "services/runtimeService.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/stack.inline.hpp" 78 79 // statics 80 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 81 bool CMSCollector::_full_gc_requested = false; 82 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 83 84 ////////////////////////////////////////////////////////////////// 85 // In support of CMS/VM thread synchronization 86 ////////////////////////////////////////////////////////////////// 87 // We split use of the CGC_lock into 2 "levels". 88 // The low-level locking is of the usual CGC_lock monitor. We introduce 89 // a higher level "token" (hereafter "CMS token") built on top of the 90 // low level monitor (hereafter "CGC lock"). 91 // The token-passing protocol gives priority to the VM thread. The 92 // CMS-lock doesn't provide any fairness guarantees, but clients 93 // should ensure that it is only held for very short, bounded 94 // durations. 95 // 96 // When either of the CMS thread or the VM thread is involved in 97 // collection operations during which it does not want the other 98 // thread to interfere, it obtains the CMS token. 99 // 100 // If either thread tries to get the token while the other has 101 // it, that thread waits. However, if the VM thread and CMS thread 102 // both want the token, then the VM thread gets priority while the 103 // CMS thread waits. This ensures, for instance, that the "concurrent" 104 // phases of the CMS thread's work do not block out the VM thread 105 // for long periods of time as the CMS thread continues to hog 106 // the token. (See bug 4616232). 107 // 108 // The baton-passing functions are, however, controlled by the 109 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 110 // and here the low-level CMS lock, not the high level token, 111 // ensures mutual exclusion. 112 // 113 // Two important conditions that we have to satisfy: 114 // 1. if a thread does a low-level wait on the CMS lock, then it 115 // relinquishes the CMS token if it were holding that token 116 // when it acquired the low-level CMS lock. 117 // 2. any low-level notifications on the low-level lock 118 // should only be sent when a thread has relinquished the token. 119 // 120 // In the absence of either property, we'd have potential deadlock. 121 // 122 // We protect each of the CMS (concurrent and sequential) phases 123 // with the CMS _token_, not the CMS _lock_. 124 // 125 // The only code protected by CMS lock is the token acquisition code 126 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 127 // baton-passing code. 128 // 129 // Unfortunately, i couldn't come up with a good abstraction to factor and 130 // hide the naked CGC_lock manipulation in the baton-passing code 131 // further below. That's something we should try to do. Also, the proof 132 // of correctness of this 2-level locking scheme is far from obvious, 133 // and potentially quite slippery. We have an uneasy suspicion, for instance, 134 // that there may be a theoretical possibility of delay/starvation in the 135 // low-level lock/wait/notify scheme used for the baton-passing because of 136 // potential interference with the priority scheme embodied in the 137 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 138 // invocation further below and marked with "XXX 20011219YSR". 139 // Indeed, as we note elsewhere, this may become yet more slippery 140 // in the presence of multiple CMS and/or multiple VM threads. XXX 141 142 class CMSTokenSync: public StackObj { 143 private: 144 bool _is_cms_thread; 145 public: 146 CMSTokenSync(bool is_cms_thread): 147 _is_cms_thread(is_cms_thread) { 148 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 149 "Incorrect argument to constructor"); 150 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 151 } 152 153 ~CMSTokenSync() { 154 assert(_is_cms_thread ? 155 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 156 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 157 "Incorrect state"); 158 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 159 } 160 }; 161 162 // Convenience class that does a CMSTokenSync, and then acquires 163 // upto three locks. 164 class CMSTokenSyncWithLocks: public CMSTokenSync { 165 private: 166 // Note: locks are acquired in textual declaration order 167 // and released in the opposite order 168 MutexLockerEx _locker1, _locker2, _locker3; 169 public: 170 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 171 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 172 CMSTokenSync(is_cms_thread), 173 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 174 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 175 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 176 { } 177 }; 178 179 180 ////////////////////////////////////////////////////////////////// 181 // Concurrent Mark-Sweep Generation ///////////////////////////// 182 ////////////////////////////////////////////////////////////////// 183 184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 185 186 // This struct contains per-thread things necessary to support parallel 187 // young-gen collection. 188 class CMSParGCThreadState: public CHeapObj<mtGC> { 189 public: 190 CompactibleFreeListSpaceLAB lab; 191 PromotionInfo promo; 192 193 // Constructor. 194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 195 promo.setSpace(cfls); 196 } 197 }; 198 199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 200 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 201 CardGeneration(rs, initial_byte_size, ct), 202 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 203 _did_compact(false) 204 { 205 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 206 HeapWord* end = (HeapWord*) _virtual_space.high(); 207 208 _direct_allocated_words = 0; 209 NOT_PRODUCT( 210 _numObjectsPromoted = 0; 211 _numWordsPromoted = 0; 212 _numObjectsAllocated = 0; 213 _numWordsAllocated = 0; 214 ) 215 216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 217 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 218 _cmsSpace->_old_gen = this; 219 220 _gc_stats = new CMSGCStats(); 221 222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 223 // offsets match. The ability to tell free chunks from objects 224 // depends on this property. 225 debug_only( 226 FreeChunk* junk = NULL; 227 assert(UseCompressedClassPointers || 228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 229 "Offset of FreeChunk::_prev within FreeChunk must match" 230 " that of OopDesc::_klass within OopDesc"); 231 ) 232 233 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 234 for (uint i = 0; i < ParallelGCThreads; i++) { 235 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 236 } 237 238 _incremental_collection_failed = false; 239 // The "dilatation_factor" is the expansion that can occur on 240 // account of the fact that the minimum object size in the CMS 241 // generation may be larger than that in, say, a contiguous young 242 // generation. 243 // Ideally, in the calculation below, we'd compute the dilatation 244 // factor as: MinChunkSize/(promoting_gen's min object size) 245 // Since we do not have such a general query interface for the 246 // promoting generation, we'll instead just use the minimum 247 // object size (which today is a header's worth of space); 248 // note that all arithmetic is in units of HeapWords. 249 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 250 assert(_dilatation_factor >= 1.0, "from previous assert"); 251 } 252 253 254 // The field "_initiating_occupancy" represents the occupancy percentage 255 // at which we trigger a new collection cycle. Unless explicitly specified 256 // via CMSInitiatingOccupancyFraction (argument "io" below), it 257 // is calculated by: 258 // 259 // Let "f" be MinHeapFreeRatio in 260 // 261 // _initiating_occupancy = 100-f + 262 // f * (CMSTriggerRatio/100) 263 // where CMSTriggerRatio is the argument "tr" below. 264 // 265 // That is, if we assume the heap is at its desired maximum occupancy at the 266 // end of a collection, we let CMSTriggerRatio of the (purported) free 267 // space be allocated before initiating a new collection cycle. 268 // 269 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 270 assert(io <= 100 && tr <= 100, "Check the arguments"); 271 if (io >= 0) { 272 _initiating_occupancy = (double)io / 100.0; 273 } else { 274 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 275 (double)(tr * MinHeapFreeRatio) / 100.0) 276 / 100.0; 277 } 278 } 279 280 void ConcurrentMarkSweepGeneration::ref_processor_init() { 281 assert(collector() != NULL, "no collector"); 282 collector()->ref_processor_init(); 283 } 284 285 void CMSCollector::ref_processor_init() { 286 if (_ref_processor == NULL) { 287 // Allocate and initialize a reference processor 288 _ref_processor = 289 new ReferenceProcessor(_span, // span 290 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 291 ParallelGCThreads, // mt processing degree 292 _cmsGen->refs_discovery_is_mt(), // mt discovery 293 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 294 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 295 &_is_alive_closure); // closure for liveness info 296 // Initialize the _ref_processor field of CMSGen 297 _cmsGen->set_ref_processor(_ref_processor); 298 299 } 300 } 301 302 AdaptiveSizePolicy* CMSCollector::size_policy() { 303 return CMSHeap::heap()->size_policy(); 304 } 305 306 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 307 308 const char* gen_name = "old"; 309 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 310 // Generation Counters - generation 1, 1 subspace 311 _gen_counters = new GenerationCounters(gen_name, 1, 1, 312 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 313 314 _space_counters = new GSpaceCounters(gen_name, 0, 315 _virtual_space.reserved_size(), 316 this, _gen_counters); 317 } 318 319 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 320 _cms_gen(cms_gen) 321 { 322 assert(alpha <= 100, "bad value"); 323 _saved_alpha = alpha; 324 325 // Initialize the alphas to the bootstrap value of 100. 326 _gc0_alpha = _cms_alpha = 100; 327 328 _cms_begin_time.update(); 329 _cms_end_time.update(); 330 331 _gc0_duration = 0.0; 332 _gc0_period = 0.0; 333 _gc0_promoted = 0; 334 335 _cms_duration = 0.0; 336 _cms_period = 0.0; 337 _cms_allocated = 0; 338 339 _cms_used_at_gc0_begin = 0; 340 _cms_used_at_gc0_end = 0; 341 _allow_duty_cycle_reduction = false; 342 _valid_bits = 0; 343 } 344 345 double CMSStats::cms_free_adjustment_factor(size_t free) const { 346 // TBD: CR 6909490 347 return 1.0; 348 } 349 350 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 351 } 352 353 // If promotion failure handling is on use 354 // the padded average size of the promotion for each 355 // young generation collection. 356 double CMSStats::time_until_cms_gen_full() const { 357 size_t cms_free = _cms_gen->cmsSpace()->free(); 358 CMSHeap* heap = CMSHeap::heap(); 359 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 360 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 361 if (cms_free > expected_promotion) { 362 // Start a cms collection if there isn't enough space to promote 363 // for the next young collection. Use the padded average as 364 // a safety factor. 365 cms_free -= expected_promotion; 366 367 // Adjust by the safety factor. 368 double cms_free_dbl = (double)cms_free; 369 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 370 // Apply a further correction factor which tries to adjust 371 // for recent occurance of concurrent mode failures. 372 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 373 cms_free_dbl = cms_free_dbl * cms_adjustment; 374 375 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 376 cms_free, expected_promotion); 377 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 378 // Add 1 in case the consumption rate goes to zero. 379 return cms_free_dbl / (cms_consumption_rate() + 1.0); 380 } 381 return 0.0; 382 } 383 384 // Compare the duration of the cms collection to the 385 // time remaining before the cms generation is empty. 386 // Note that the time from the start of the cms collection 387 // to the start of the cms sweep (less than the total 388 // duration of the cms collection) can be used. This 389 // has been tried and some applications experienced 390 // promotion failures early in execution. This was 391 // possibly because the averages were not accurate 392 // enough at the beginning. 393 double CMSStats::time_until_cms_start() const { 394 // We add "gc0_period" to the "work" calculation 395 // below because this query is done (mostly) at the 396 // end of a scavenge, so we need to conservatively 397 // account for that much possible delay 398 // in the query so as to avoid concurrent mode failures 399 // due to starting the collection just a wee bit too 400 // late. 401 double work = cms_duration() + gc0_period(); 402 double deadline = time_until_cms_gen_full(); 403 // If a concurrent mode failure occurred recently, we want to be 404 // more conservative and halve our expected time_until_cms_gen_full() 405 if (work > deadline) { 406 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 407 cms_duration(), gc0_period(), time_until_cms_gen_full()); 408 return 0.0; 409 } 410 return work - deadline; 411 } 412 413 #ifndef PRODUCT 414 void CMSStats::print_on(outputStream *st) const { 415 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 416 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 417 gc0_duration(), gc0_period(), gc0_promoted()); 418 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 419 cms_duration(), cms_period(), cms_allocated()); 420 st->print(",cms_since_beg=%g,cms_since_end=%g", 421 cms_time_since_begin(), cms_time_since_end()); 422 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 423 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 424 425 if (valid()) { 426 st->print(",promo_rate=%g,cms_alloc_rate=%g", 427 promotion_rate(), cms_allocation_rate()); 428 st->print(",cms_consumption_rate=%g,time_until_full=%g", 429 cms_consumption_rate(), time_until_cms_gen_full()); 430 } 431 st->cr(); 432 } 433 #endif // #ifndef PRODUCT 434 435 CMSCollector::CollectorState CMSCollector::_collectorState = 436 CMSCollector::Idling; 437 bool CMSCollector::_foregroundGCIsActive = false; 438 bool CMSCollector::_foregroundGCShouldWait = false; 439 440 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 441 CardTableRS* ct, 442 ConcurrentMarkSweepPolicy* cp): 443 _cmsGen(cmsGen), 444 _ct(ct), 445 _ref_processor(NULL), // will be set later 446 _conc_workers(NULL), // may be set later 447 _abort_preclean(false), 448 _start_sampling(false), 449 _between_prologue_and_epilogue(false), 450 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 451 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 452 -1 /* lock-free */, "No_lock" /* dummy */), 453 _modUnionClosurePar(&_modUnionTable), 454 // Adjust my span to cover old (cms) gen 455 _span(cmsGen->reserved()), 456 // Construct the is_alive_closure with _span & markBitMap 457 _is_alive_closure(_span, &_markBitMap), 458 _restart_addr(NULL), 459 _overflow_list(NULL), 460 _stats(cmsGen), 461 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 462 //verify that this lock should be acquired with safepoint check. 463 Monitor::_safepoint_check_sometimes)), 464 _eden_chunk_array(NULL), // may be set in ctor body 465 _eden_chunk_capacity(0), // -- ditto -- 466 _eden_chunk_index(0), // -- ditto -- 467 _survivor_plab_array(NULL), // -- ditto -- 468 _survivor_chunk_array(NULL), // -- ditto -- 469 _survivor_chunk_capacity(0), // -- ditto -- 470 _survivor_chunk_index(0), // -- ditto -- 471 _ser_pmc_preclean_ovflw(0), 472 _ser_kac_preclean_ovflw(0), 473 _ser_pmc_remark_ovflw(0), 474 _par_pmc_remark_ovflw(0), 475 _ser_kac_ovflw(0), 476 _par_kac_ovflw(0), 477 #ifndef PRODUCT 478 _num_par_pushes(0), 479 #endif 480 _collection_count_start(0), 481 _verifying(false), 482 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 483 _completed_initialization(false), 484 _collector_policy(cp), 485 _should_unload_classes(CMSClassUnloadingEnabled), 486 _concurrent_cycles_since_last_unload(0), 487 _roots_scanning_options(GenCollectedHeap::SO_None), 488 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 489 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 490 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 491 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 492 _cms_start_registered(false) 493 { 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 log_warning(gc)("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 log_warning(gc)("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 549 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 550 551 // Parallel task queues; these are shared for the 552 // concurrent and stop-world phases of CMS, but 553 // are not shared with parallel scavenge (ParNew). 554 { 555 uint i; 556 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 557 558 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 559 || ParallelRefProcEnabled) 560 && num_queues > 0) { 561 _task_queues = new OopTaskQueueSet(num_queues); 562 if (_task_queues == NULL) { 563 log_warning(gc)("task_queues allocation failure."); 564 return; 565 } 566 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 567 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 568 for (i = 0; i < num_queues; i++) { 569 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 570 if (q == NULL) { 571 log_warning(gc)("work_queue allocation failure."); 572 return; 573 } 574 _task_queues->register_queue(i, q); 575 } 576 for (i = 0; i < num_queues; i++) { 577 _task_queues->queue(i)->initialize(); 578 _hash_seed[i] = 17; // copied from ParNew 579 } 580 } 581 } 582 583 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 584 585 // Clip CMSBootstrapOccupancy between 0 and 100. 586 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 587 588 // Now tell CMS generations the identity of their collector 589 ConcurrentMarkSweepGeneration::set_collector(this); 590 591 // Create & start a CMS thread for this CMS collector 592 _cmsThread = ConcurrentMarkSweepThread::start(this); 593 assert(cmsThread() != NULL, "CMS Thread should have been created"); 594 assert(cmsThread()->collector() == this, 595 "CMS Thread should refer to this gen"); 596 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 597 598 // Support for parallelizing young gen rescan 599 CMSHeap* heap = CMSHeap::heap(); 600 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 601 _young_gen = (ParNewGeneration*)heap->young_gen(); 602 if (heap->supports_inline_contig_alloc()) { 603 _top_addr = heap->top_addr(); 604 _end_addr = heap->end_addr(); 605 assert(_young_gen != NULL, "no _young_gen"); 606 _eden_chunk_index = 0; 607 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 608 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 609 } 610 611 // Support for parallelizing survivor space rescan 612 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 613 const size_t max_plab_samples = 614 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 615 616 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 617 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 618 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 619 _survivor_chunk_capacity = max_plab_samples; 620 for (uint i = 0; i < ParallelGCThreads; i++) { 621 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 622 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 623 assert(cur->end() == 0, "Should be 0"); 624 assert(cur->array() == vec, "Should be vec"); 625 assert(cur->capacity() == max_plab_samples, "Error"); 626 } 627 } 628 629 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 630 _gc_counters = new CollectorCounters("CMS", 1); 631 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 632 _completed_initialization = true; 633 _inter_sweep_timer.start(); // start of time 634 } 635 636 const char* ConcurrentMarkSweepGeneration::name() const { 637 return "concurrent mark-sweep generation"; 638 } 639 void ConcurrentMarkSweepGeneration::update_counters() { 640 if (UsePerfData) { 641 _space_counters->update_all(); 642 _gen_counters->update_all(); 643 } 644 } 645 646 // this is an optimized version of update_counters(). it takes the 647 // used value as a parameter rather than computing it. 648 // 649 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 650 if (UsePerfData) { 651 _space_counters->update_used(used); 652 _space_counters->update_capacity(); 653 _gen_counters->update_all(); 654 } 655 } 656 657 void ConcurrentMarkSweepGeneration::print() const { 658 Generation::print(); 659 cmsSpace()->print(); 660 } 661 662 #ifndef PRODUCT 663 void ConcurrentMarkSweepGeneration::print_statistics() { 664 cmsSpace()->printFLCensus(0); 665 } 666 #endif 667 668 size_t 669 ConcurrentMarkSweepGeneration::contiguous_available() const { 670 // dld proposes an improvement in precision here. If the committed 671 // part of the space ends in a free block we should add that to 672 // uncommitted size in the calculation below. Will make this 673 // change later, staying with the approximation below for the 674 // time being. -- ysr. 675 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 676 } 677 678 size_t 679 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 680 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 681 } 682 683 size_t ConcurrentMarkSweepGeneration::max_available() const { 684 return free() + _virtual_space.uncommitted_size(); 685 } 686 687 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 688 size_t available = max_available(); 689 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 690 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 691 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 692 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 693 return res; 694 } 695 696 // At a promotion failure dump information on block layout in heap 697 // (cms old generation). 698 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 699 Log(gc, promotion) log; 700 if (log.is_trace()) { 701 LogStream ls(log.trace()); 702 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 703 } 704 } 705 706 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 707 // Clear the promotion information. These pointers can be adjusted 708 // along with all the other pointers into the heap but 709 // compaction is expected to be a rare event with 710 // a heap using cms so don't do it without seeing the need. 711 for (uint i = 0; i < ParallelGCThreads; i++) { 712 _par_gc_thread_states[i]->promo.reset(); 713 } 714 } 715 716 void ConcurrentMarkSweepGeneration::compute_new_size() { 717 assert_locked_or_safepoint(Heap_lock); 718 719 // If incremental collection failed, we just want to expand 720 // to the limit. 721 if (incremental_collection_failed()) { 722 clear_incremental_collection_failed(); 723 grow_to_reserved(); 724 return; 725 } 726 727 // The heap has been compacted but not reset yet. 728 // Any metric such as free() or used() will be incorrect. 729 730 CardGeneration::compute_new_size(); 731 732 // Reset again after a possible resizing 733 if (did_compact()) { 734 cmsSpace()->reset_after_compaction(); 735 } 736 } 737 738 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 739 assert_locked_or_safepoint(Heap_lock); 740 741 // If incremental collection failed, we just want to expand 742 // to the limit. 743 if (incremental_collection_failed()) { 744 clear_incremental_collection_failed(); 745 grow_to_reserved(); 746 return; 747 } 748 749 double free_percentage = ((double) free()) / capacity(); 750 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 751 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 752 753 // compute expansion delta needed for reaching desired free percentage 754 if (free_percentage < desired_free_percentage) { 755 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 756 assert(desired_capacity >= capacity(), "invalid expansion size"); 757 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 758 Log(gc) log; 759 if (log.is_trace()) { 760 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 761 log.trace("From compute_new_size: "); 762 log.trace(" Free fraction %f", free_percentage); 763 log.trace(" Desired free fraction %f", desired_free_percentage); 764 log.trace(" Maximum free fraction %f", maximum_free_percentage); 765 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 766 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 767 CMSHeap* heap = CMSHeap::heap(); 768 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation"); 769 size_t young_size = heap->young_gen()->capacity(); 770 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 771 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 772 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 773 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 774 } 775 // safe if expansion fails 776 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 777 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 778 } else { 779 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 780 assert(desired_capacity <= capacity(), "invalid expansion size"); 781 size_t shrink_bytes = capacity() - desired_capacity; 782 // Don't shrink unless the delta is greater than the minimum shrink we want 783 if (shrink_bytes >= MinHeapDeltaBytes) { 784 shrink_free_list_by(shrink_bytes); 785 } 786 } 787 } 788 789 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 790 return cmsSpace()->freelistLock(); 791 } 792 793 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 794 CMSSynchronousYieldRequest yr; 795 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 796 return have_lock_and_allocate(size, tlab); 797 } 798 799 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 800 bool tlab /* ignored */) { 801 assert_lock_strong(freelistLock()); 802 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 803 HeapWord* res = cmsSpace()->allocate(adjustedSize); 804 // Allocate the object live (grey) if the background collector has 805 // started marking. This is necessary because the marker may 806 // have passed this address and consequently this object will 807 // not otherwise be greyed and would be incorrectly swept up. 808 // Note that if this object contains references, the writing 809 // of those references will dirty the card containing this object 810 // allowing the object to be blackened (and its references scanned) 811 // either during a preclean phase or at the final checkpoint. 812 if (res != NULL) { 813 // We may block here with an uninitialized object with 814 // its mark-bit or P-bits not yet set. Such objects need 815 // to be safely navigable by block_start(). 816 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 817 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 818 collector()->direct_allocated(res, adjustedSize); 819 _direct_allocated_words += adjustedSize; 820 // allocation counters 821 NOT_PRODUCT( 822 _numObjectsAllocated++; 823 _numWordsAllocated += (int)adjustedSize; 824 ) 825 } 826 return res; 827 } 828 829 // In the case of direct allocation by mutators in a generation that 830 // is being concurrently collected, the object must be allocated 831 // live (grey) if the background collector has started marking. 832 // This is necessary because the marker may 833 // have passed this address and consequently this object will 834 // not otherwise be greyed and would be incorrectly swept up. 835 // Note that if this object contains references, the writing 836 // of those references will dirty the card containing this object 837 // allowing the object to be blackened (and its references scanned) 838 // either during a preclean phase or at the final checkpoint. 839 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 840 assert(_markBitMap.covers(start, size), "Out of bounds"); 841 if (_collectorState >= Marking) { 842 MutexLockerEx y(_markBitMap.lock(), 843 Mutex::_no_safepoint_check_flag); 844 // [see comments preceding SweepClosure::do_blk() below for details] 845 // 846 // Can the P-bits be deleted now? JJJ 847 // 848 // 1. need to mark the object as live so it isn't collected 849 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 850 // 3. need to mark the end of the object so marking, precleaning or sweeping 851 // can skip over uninitialized or unparsable objects. An allocated 852 // object is considered uninitialized for our purposes as long as 853 // its klass word is NULL. All old gen objects are parsable 854 // as soon as they are initialized.) 855 _markBitMap.mark(start); // object is live 856 _markBitMap.mark(start + 1); // object is potentially uninitialized? 857 _markBitMap.mark(start + size - 1); 858 // mark end of object 859 } 860 // check that oop looks uninitialized 861 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 862 } 863 864 void CMSCollector::promoted(bool par, HeapWord* start, 865 bool is_obj_array, size_t obj_size) { 866 assert(_markBitMap.covers(start), "Out of bounds"); 867 // See comment in direct_allocated() about when objects should 868 // be allocated live. 869 if (_collectorState >= Marking) { 870 // we already hold the marking bit map lock, taken in 871 // the prologue 872 if (par) { 873 _markBitMap.par_mark(start); 874 } else { 875 _markBitMap.mark(start); 876 } 877 // We don't need to mark the object as uninitialized (as 878 // in direct_allocated above) because this is being done with the 879 // world stopped and the object will be initialized by the 880 // time the marking, precleaning or sweeping get to look at it. 881 // But see the code for copying objects into the CMS generation, 882 // where we need to ensure that concurrent readers of the 883 // block offset table are able to safely navigate a block that 884 // is in flux from being free to being allocated (and in 885 // transition while being copied into) and subsequently 886 // becoming a bona-fide object when the copy/promotion is complete. 887 assert(SafepointSynchronize::is_at_safepoint(), 888 "expect promotion only at safepoints"); 889 890 if (_collectorState < Sweeping) { 891 // Mark the appropriate cards in the modUnionTable, so that 892 // this object gets scanned before the sweep. If this is 893 // not done, CMS generation references in the object might 894 // not get marked. 895 // For the case of arrays, which are otherwise precisely 896 // marked, we need to dirty the entire array, not just its head. 897 if (is_obj_array) { 898 // The [par_]mark_range() method expects mr.end() below to 899 // be aligned to the granularity of a bit's representation 900 // in the heap. In the case of the MUT below, that's a 901 // card size. 902 MemRegion mr(start, 903 align_up(start + obj_size, 904 CardTable::card_size /* bytes */)); 905 if (par) { 906 _modUnionTable.par_mark_range(mr); 907 } else { 908 _modUnionTable.mark_range(mr); 909 } 910 } else { // not an obj array; we can just mark the head 911 if (par) { 912 _modUnionTable.par_mark(start); 913 } else { 914 _modUnionTable.mark(start); 915 } 916 } 917 } 918 } 919 } 920 921 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 922 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 923 // allocate, copy and if necessary update promoinfo -- 924 // delegate to underlying space. 925 assert_lock_strong(freelistLock()); 926 927 #ifndef PRODUCT 928 if (CMSHeap::heap()->promotion_should_fail()) { 929 return NULL; 930 } 931 #endif // #ifndef PRODUCT 932 933 oop res = _cmsSpace->promote(obj, obj_size); 934 if (res == NULL) { 935 // expand and retry 936 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 937 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 938 // Since this is the old generation, we don't try to promote 939 // into a more senior generation. 940 res = _cmsSpace->promote(obj, obj_size); 941 } 942 if (res != NULL) { 943 // See comment in allocate() about when objects should 944 // be allocated live. 945 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 946 collector()->promoted(false, // Not parallel 947 (HeapWord*)res, obj->is_objArray(), obj_size); 948 // promotion counters 949 NOT_PRODUCT( 950 _numObjectsPromoted++; 951 _numWordsPromoted += 952 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 953 ) 954 } 955 return res; 956 } 957 958 959 // IMPORTANT: Notes on object size recognition in CMS. 960 // --------------------------------------------------- 961 // A block of storage in the CMS generation is always in 962 // one of three states. A free block (FREE), an allocated 963 // object (OBJECT) whose size() method reports the correct size, 964 // and an intermediate state (TRANSIENT) in which its size cannot 965 // be accurately determined. 966 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 967 // ----------------------------------------------------- 968 // FREE: klass_word & 1 == 1; mark_word holds block size 969 // 970 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 971 // obj->size() computes correct size 972 // 973 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 974 // 975 // STATE IDENTIFICATION: (64 bit+COOPS) 976 // ------------------------------------ 977 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 978 // 979 // OBJECT: klass_word installed; klass_word != 0; 980 // obj->size() computes correct size 981 // 982 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 983 // 984 // 985 // STATE TRANSITION DIAGRAM 986 // 987 // mut / parnew mut / parnew 988 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 989 // ^ | 990 // |------------------------ DEAD <------------------------------------| 991 // sweep mut 992 // 993 // While a block is in TRANSIENT state its size cannot be determined 994 // so readers will either need to come back later or stall until 995 // the size can be determined. Note that for the case of direct 996 // allocation, P-bits, when available, may be used to determine the 997 // size of an object that may not yet have been initialized. 998 999 // Things to support parallel young-gen collection. 1000 oop 1001 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1002 oop old, markOop m, 1003 size_t word_sz) { 1004 #ifndef PRODUCT 1005 if (CMSHeap::heap()->promotion_should_fail()) { 1006 return NULL; 1007 } 1008 #endif // #ifndef PRODUCT 1009 1010 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1011 PromotionInfo* promoInfo = &ps->promo; 1012 // if we are tracking promotions, then first ensure space for 1013 // promotion (including spooling space for saving header if necessary). 1014 // then allocate and copy, then track promoted info if needed. 1015 // When tracking (see PromotionInfo::track()), the mark word may 1016 // be displaced and in this case restoration of the mark word 1017 // occurs in the (oop_since_save_marks_)iterate phase. 1018 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1019 // Out of space for allocating spooling buffers; 1020 // try expanding and allocating spooling buffers. 1021 if (!expand_and_ensure_spooling_space(promoInfo)) { 1022 return NULL; 1023 } 1024 } 1025 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1026 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1027 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1028 if (obj_ptr == NULL) { 1029 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1030 if (obj_ptr == NULL) { 1031 return NULL; 1032 } 1033 } 1034 oop obj = oop(obj_ptr); 1035 OrderAccess::storestore(); 1036 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1037 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1038 // IMPORTANT: See note on object initialization for CMS above. 1039 // Otherwise, copy the object. Here we must be careful to insert the 1040 // klass pointer last, since this marks the block as an allocated object. 1041 // Except with compressed oops it's the mark word. 1042 HeapWord* old_ptr = (HeapWord*)old; 1043 // Restore the mark word copied above. 1044 obj->set_mark(m); 1045 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1046 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1047 OrderAccess::storestore(); 1048 1049 if (UseCompressedClassPointers) { 1050 // Copy gap missed by (aligned) header size calculation below 1051 obj->set_klass_gap(old->klass_gap()); 1052 } 1053 if (word_sz > (size_t)oopDesc::header_size()) { 1054 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1055 obj_ptr + oopDesc::header_size(), 1056 word_sz - oopDesc::header_size()); 1057 } 1058 1059 // Now we can track the promoted object, if necessary. We take care 1060 // to delay the transition from uninitialized to full object 1061 // (i.e., insertion of klass pointer) until after, so that it 1062 // atomically becomes a promoted object. 1063 if (promoInfo->tracking()) { 1064 promoInfo->track((PromotedObject*)obj, old->klass()); 1065 } 1066 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1067 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1068 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1069 1070 // Finally, install the klass pointer (this should be volatile). 1071 OrderAccess::storestore(); 1072 obj->set_klass(old->klass()); 1073 // We should now be able to calculate the right size for this object 1074 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1075 1076 collector()->promoted(true, // parallel 1077 obj_ptr, old->is_objArray(), word_sz); 1078 1079 NOT_PRODUCT( 1080 Atomic::inc(&_numObjectsPromoted); 1081 Atomic::add(alloc_sz, &_numWordsPromoted); 1082 ) 1083 1084 return obj; 1085 } 1086 1087 void 1088 ConcurrentMarkSweepGeneration:: 1089 par_promote_alloc_done(int thread_num) { 1090 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1091 ps->lab.retire(thread_num); 1092 } 1093 1094 void 1095 ConcurrentMarkSweepGeneration:: 1096 par_oop_since_save_marks_iterate_done(int thread_num) { 1097 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1098 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1099 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1100 1101 // Because card-scanning has been completed, subsequent phases 1102 // (e.g., reference processing) will not need to recognize which 1103 // objects have been promoted during this GC. So, we can now disable 1104 // promotion tracking. 1105 ps->promo.stopTrackingPromotions(); 1106 } 1107 1108 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1109 size_t size, 1110 bool tlab) 1111 { 1112 // We allow a STW collection only if a full 1113 // collection was requested. 1114 return full || should_allocate(size, tlab); // FIX ME !!! 1115 // This and promotion failure handling are connected at the 1116 // hip and should be fixed by untying them. 1117 } 1118 1119 bool CMSCollector::shouldConcurrentCollect() { 1120 LogTarget(Trace, gc) log; 1121 1122 if (_full_gc_requested) { 1123 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1124 return true; 1125 } 1126 1127 FreelistLocker x(this); 1128 // ------------------------------------------------------------------ 1129 // Print out lots of information which affects the initiation of 1130 // a collection. 1131 if (log.is_enabled() && stats().valid()) { 1132 log.print("CMSCollector shouldConcurrentCollect: "); 1133 1134 LogStream out(log); 1135 stats().print_on(&out); 1136 1137 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1138 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1139 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1140 log.print("promotion_rate=%g", stats().promotion_rate()); 1141 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1142 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1143 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1144 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1145 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1146 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1147 } 1148 // ------------------------------------------------------------------ 1149 1150 // If the estimated time to complete a cms collection (cms_duration()) 1151 // is less than the estimated time remaining until the cms generation 1152 // is full, start a collection. 1153 if (!UseCMSInitiatingOccupancyOnly) { 1154 if (stats().valid()) { 1155 if (stats().time_until_cms_start() == 0.0) { 1156 return true; 1157 } 1158 } else { 1159 // We want to conservatively collect somewhat early in order 1160 // to try and "bootstrap" our CMS/promotion statistics; 1161 // this branch will not fire after the first successful CMS 1162 // collection because the stats should then be valid. 1163 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1164 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1165 _cmsGen->occupancy(), _bootstrap_occupancy); 1166 return true; 1167 } 1168 } 1169 } 1170 1171 // Otherwise, we start a collection cycle if 1172 // old gen want a collection cycle started. Each may use 1173 // an appropriate criterion for making this decision. 1174 // XXX We need to make sure that the gen expansion 1175 // criterion dovetails well with this. XXX NEED TO FIX THIS 1176 if (_cmsGen->should_concurrent_collect()) { 1177 log.print("CMS old gen initiated"); 1178 return true; 1179 } 1180 1181 // We start a collection if we believe an incremental collection may fail; 1182 // this is not likely to be productive in practice because it's probably too 1183 // late anyway. 1184 CMSHeap* heap = CMSHeap::heap(); 1185 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1186 log.print("CMSCollector: collect because incremental collection will fail "); 1187 return true; 1188 } 1189 1190 if (MetaspaceGC::should_concurrent_collect()) { 1191 log.print("CMSCollector: collect for metadata allocation "); 1192 return true; 1193 } 1194 1195 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1196 if (CMSTriggerInterval >= 0) { 1197 if (CMSTriggerInterval == 0) { 1198 // Trigger always 1199 return true; 1200 } 1201 1202 // Check the CMS time since begin (we do not check the stats validity 1203 // as we want to be able to trigger the first CMS cycle as well) 1204 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1205 if (stats().valid()) { 1206 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1207 stats().cms_time_since_begin()); 1208 } else { 1209 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1210 } 1211 return true; 1212 } 1213 } 1214 1215 return false; 1216 } 1217 1218 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1219 1220 // Clear _expansion_cause fields of constituent generations 1221 void CMSCollector::clear_expansion_cause() { 1222 _cmsGen->clear_expansion_cause(); 1223 } 1224 1225 // We should be conservative in starting a collection cycle. To 1226 // start too eagerly runs the risk of collecting too often in the 1227 // extreme. To collect too rarely falls back on full collections, 1228 // which works, even if not optimum in terms of concurrent work. 1229 // As a work around for too eagerly collecting, use the flag 1230 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1231 // giving the user an easily understandable way of controlling the 1232 // collections. 1233 // We want to start a new collection cycle if any of the following 1234 // conditions hold: 1235 // . our current occupancy exceeds the configured initiating occupancy 1236 // for this generation, or 1237 // . we recently needed to expand this space and have not, since that 1238 // expansion, done a collection of this generation, or 1239 // . the underlying space believes that it may be a good idea to initiate 1240 // a concurrent collection (this may be based on criteria such as the 1241 // following: the space uses linear allocation and linear allocation is 1242 // going to fail, or there is believed to be excessive fragmentation in 1243 // the generation, etc... or ... 1244 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1245 // the case of the old generation; see CR 6543076): 1246 // we may be approaching a point at which allocation requests may fail because 1247 // we will be out of sufficient free space given allocation rate estimates.] 1248 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1249 1250 assert_lock_strong(freelistLock()); 1251 if (occupancy() > initiating_occupancy()) { 1252 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1253 short_name(), occupancy(), initiating_occupancy()); 1254 return true; 1255 } 1256 if (UseCMSInitiatingOccupancyOnly) { 1257 return false; 1258 } 1259 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1260 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1261 return true; 1262 } 1263 return false; 1264 } 1265 1266 void ConcurrentMarkSweepGeneration::collect(bool full, 1267 bool clear_all_soft_refs, 1268 size_t size, 1269 bool tlab) 1270 { 1271 collector()->collect(full, clear_all_soft_refs, size, tlab); 1272 } 1273 1274 void CMSCollector::collect(bool full, 1275 bool clear_all_soft_refs, 1276 size_t size, 1277 bool tlab) 1278 { 1279 // The following "if" branch is present for defensive reasons. 1280 // In the current uses of this interface, it can be replaced with: 1281 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1282 // But I am not placing that assert here to allow future 1283 // generality in invoking this interface. 1284 if (GCLocker::is_active()) { 1285 // A consistency test for GCLocker 1286 assert(GCLocker::needs_gc(), "Should have been set already"); 1287 // Skip this foreground collection, instead 1288 // expanding the heap if necessary. 1289 // Need the free list locks for the call to free() in compute_new_size() 1290 compute_new_size(); 1291 return; 1292 } 1293 acquire_control_and_collect(full, clear_all_soft_refs); 1294 } 1295 1296 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1297 CMSHeap* heap = CMSHeap::heap(); 1298 unsigned int gc_count = heap->total_full_collections(); 1299 if (gc_count == full_gc_count) { 1300 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1301 _full_gc_requested = true; 1302 _full_gc_cause = cause; 1303 CGC_lock->notify(); // nudge CMS thread 1304 } else { 1305 assert(gc_count > full_gc_count, "Error: causal loop"); 1306 } 1307 } 1308 1309 bool CMSCollector::is_external_interruption() { 1310 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1311 return GCCause::is_user_requested_gc(cause) || 1312 GCCause::is_serviceability_requested_gc(cause); 1313 } 1314 1315 void CMSCollector::report_concurrent_mode_interruption() { 1316 if (is_external_interruption()) { 1317 log_debug(gc)("Concurrent mode interrupted"); 1318 } else { 1319 log_debug(gc)("Concurrent mode failure"); 1320 _gc_tracer_cm->report_concurrent_mode_failure(); 1321 } 1322 } 1323 1324 1325 // The foreground and background collectors need to coordinate in order 1326 // to make sure that they do not mutually interfere with CMS collections. 1327 // When a background collection is active, 1328 // the foreground collector may need to take over (preempt) and 1329 // synchronously complete an ongoing collection. Depending on the 1330 // frequency of the background collections and the heap usage 1331 // of the application, this preemption can be seldom or frequent. 1332 // There are only certain 1333 // points in the background collection that the "collection-baton" 1334 // can be passed to the foreground collector. 1335 // 1336 // The foreground collector will wait for the baton before 1337 // starting any part of the collection. The foreground collector 1338 // will only wait at one location. 1339 // 1340 // The background collector will yield the baton before starting a new 1341 // phase of the collection (e.g., before initial marking, marking from roots, 1342 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1343 // of the loop which switches the phases. The background collector does some 1344 // of the phases (initial mark, final re-mark) with the world stopped. 1345 // Because of locking involved in stopping the world, 1346 // the foreground collector should not block waiting for the background 1347 // collector when it is doing a stop-the-world phase. The background 1348 // collector will yield the baton at an additional point just before 1349 // it enters a stop-the-world phase. Once the world is stopped, the 1350 // background collector checks the phase of the collection. If the 1351 // phase has not changed, it proceeds with the collection. If the 1352 // phase has changed, it skips that phase of the collection. See 1353 // the comments on the use of the Heap_lock in collect_in_background(). 1354 // 1355 // Variable used in baton passing. 1356 // _foregroundGCIsActive - Set to true by the foreground collector when 1357 // it wants the baton. The foreground clears it when it has finished 1358 // the collection. 1359 // _foregroundGCShouldWait - Set to true by the background collector 1360 // when it is running. The foreground collector waits while 1361 // _foregroundGCShouldWait is true. 1362 // CGC_lock - monitor used to protect access to the above variables 1363 // and to notify the foreground and background collectors. 1364 // _collectorState - current state of the CMS collection. 1365 // 1366 // The foreground collector 1367 // acquires the CGC_lock 1368 // sets _foregroundGCIsActive 1369 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1370 // various locks acquired in preparation for the collection 1371 // are released so as not to block the background collector 1372 // that is in the midst of a collection 1373 // proceeds with the collection 1374 // clears _foregroundGCIsActive 1375 // returns 1376 // 1377 // The background collector in a loop iterating on the phases of the 1378 // collection 1379 // acquires the CGC_lock 1380 // sets _foregroundGCShouldWait 1381 // if _foregroundGCIsActive is set 1382 // clears _foregroundGCShouldWait, notifies _CGC_lock 1383 // waits on _CGC_lock for _foregroundGCIsActive to become false 1384 // and exits the loop. 1385 // otherwise 1386 // proceed with that phase of the collection 1387 // if the phase is a stop-the-world phase, 1388 // yield the baton once more just before enqueueing 1389 // the stop-world CMS operation (executed by the VM thread). 1390 // returns after all phases of the collection are done 1391 // 1392 1393 void CMSCollector::acquire_control_and_collect(bool full, 1394 bool clear_all_soft_refs) { 1395 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1396 assert(!Thread::current()->is_ConcurrentGC_thread(), 1397 "shouldn't try to acquire control from self!"); 1398 1399 // Start the protocol for acquiring control of the 1400 // collection from the background collector (aka CMS thread). 1401 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1402 "VM thread should have CMS token"); 1403 // Remember the possibly interrupted state of an ongoing 1404 // concurrent collection 1405 CollectorState first_state = _collectorState; 1406 1407 // Signal to a possibly ongoing concurrent collection that 1408 // we want to do a foreground collection. 1409 _foregroundGCIsActive = true; 1410 1411 // release locks and wait for a notify from the background collector 1412 // releasing the locks in only necessary for phases which 1413 // do yields to improve the granularity of the collection. 1414 assert_lock_strong(bitMapLock()); 1415 // We need to lock the Free list lock for the space that we are 1416 // currently collecting. 1417 assert(haveFreelistLocks(), "Must be holding free list locks"); 1418 bitMapLock()->unlock(); 1419 releaseFreelistLocks(); 1420 { 1421 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1422 if (_foregroundGCShouldWait) { 1423 // We are going to be waiting for action for the CMS thread; 1424 // it had better not be gone (for instance at shutdown)! 1425 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1426 "CMS thread must be running"); 1427 // Wait here until the background collector gives us the go-ahead 1428 ConcurrentMarkSweepThread::clear_CMS_flag( 1429 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1430 // Get a possibly blocked CMS thread going: 1431 // Note that we set _foregroundGCIsActive true above, 1432 // without protection of the CGC_lock. 1433 CGC_lock->notify(); 1434 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1435 "Possible deadlock"); 1436 while (_foregroundGCShouldWait) { 1437 // wait for notification 1438 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1439 // Possibility of delay/starvation here, since CMS token does 1440 // not know to give priority to VM thread? Actually, i think 1441 // there wouldn't be any delay/starvation, but the proof of 1442 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1443 } 1444 ConcurrentMarkSweepThread::set_CMS_flag( 1445 ConcurrentMarkSweepThread::CMS_vm_has_token); 1446 } 1447 } 1448 // The CMS_token is already held. Get back the other locks. 1449 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1450 "VM thread should have CMS token"); 1451 getFreelistLocks(); 1452 bitMapLock()->lock_without_safepoint_check(); 1453 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1454 p2i(Thread::current()), first_state); 1455 log_debug(gc, state)(" gets control with state %d", _collectorState); 1456 1457 // Inform cms gen if this was due to partial collection failing. 1458 // The CMS gen may use this fact to determine its expansion policy. 1459 CMSHeap* heap = CMSHeap::heap(); 1460 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1461 assert(!_cmsGen->incremental_collection_failed(), 1462 "Should have been noticed, reacted to and cleared"); 1463 _cmsGen->set_incremental_collection_failed(); 1464 } 1465 1466 if (first_state > Idling) { 1467 report_concurrent_mode_interruption(); 1468 } 1469 1470 set_did_compact(true); 1471 1472 // If the collection is being acquired from the background 1473 // collector, there may be references on the discovered 1474 // references lists. Abandon those references, since some 1475 // of them may have become unreachable after concurrent 1476 // discovery; the STW compacting collector will redo discovery 1477 // more precisely, without being subject to floating garbage. 1478 // Leaving otherwise unreachable references in the discovered 1479 // lists would require special handling. 1480 ref_processor()->disable_discovery(); 1481 ref_processor()->abandon_partial_discovery(); 1482 ref_processor()->verify_no_references_recorded(); 1483 1484 if (first_state > Idling) { 1485 save_heap_summary(); 1486 } 1487 1488 do_compaction_work(clear_all_soft_refs); 1489 1490 // Has the GC time limit been exceeded? 1491 size_t max_eden_size = _young_gen->max_eden_size(); 1492 GCCause::Cause gc_cause = heap->gc_cause(); 1493 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1494 _young_gen->eden()->used(), 1495 _cmsGen->max_capacity(), 1496 max_eden_size, 1497 full, 1498 gc_cause, 1499 heap->soft_ref_policy()); 1500 1501 // Reset the expansion cause, now that we just completed 1502 // a collection cycle. 1503 clear_expansion_cause(); 1504 _foregroundGCIsActive = false; 1505 return; 1506 } 1507 1508 // Resize the tenured generation 1509 // after obtaining the free list locks for the 1510 // two generations. 1511 void CMSCollector::compute_new_size() { 1512 assert_locked_or_safepoint(Heap_lock); 1513 FreelistLocker z(this); 1514 MetaspaceGC::compute_new_size(); 1515 _cmsGen->compute_new_size_free_list(); 1516 } 1517 1518 // A work method used by the foreground collector to do 1519 // a mark-sweep-compact. 1520 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1521 CMSHeap* heap = CMSHeap::heap(); 1522 1523 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1524 gc_timer->register_gc_start(); 1525 1526 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1527 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1528 1529 heap->pre_full_gc_dump(gc_timer); 1530 1531 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1532 1533 // Temporarily widen the span of the weak reference processing to 1534 // the entire heap. 1535 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1536 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1537 // Temporarily, clear the "is_alive_non_header" field of the 1538 // reference processor. 1539 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1540 // Temporarily make reference _processing_ single threaded (non-MT). 1541 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1542 // Temporarily make refs discovery atomic 1543 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1544 // Temporarily make reference _discovery_ single threaded (non-MT) 1545 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1546 1547 ref_processor()->set_enqueuing_is_done(false); 1548 ref_processor()->enable_discovery(); 1549 ref_processor()->setup_policy(clear_all_soft_refs); 1550 // If an asynchronous collection finishes, the _modUnionTable is 1551 // all clear. If we are assuming the collection from an asynchronous 1552 // collection, clear the _modUnionTable. 1553 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1554 "_modUnionTable should be clear if the baton was not passed"); 1555 _modUnionTable.clear_all(); 1556 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1557 "mod union for klasses should be clear if the baton was passed"); 1558 _ct->cld_rem_set()->clear_mod_union(); 1559 1560 1561 // We must adjust the allocation statistics being maintained 1562 // in the free list space. We do so by reading and clearing 1563 // the sweep timer and updating the block flux rate estimates below. 1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1565 if (_inter_sweep_timer.is_active()) { 1566 _inter_sweep_timer.stop(); 1567 // Note that we do not use this sample to update the _inter_sweep_estimate. 1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1569 _inter_sweep_estimate.padded_average(), 1570 _intra_sweep_estimate.padded_average()); 1571 } 1572 1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1574 #ifdef ASSERT 1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1576 size_t free_size = cms_space->free(); 1577 assert(free_size == 1578 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1579 * HeapWordSize, 1580 "All the free space should be compacted into one chunk at top"); 1581 assert(cms_space->dictionary()->total_chunk_size( 1582 debug_only(cms_space->freelistLock())) == 0 || 1583 cms_space->totalSizeInIndexedFreeLists() == 0, 1584 "All the free space should be in a single chunk"); 1585 size_t num = cms_space->totalCount(); 1586 assert((free_size == 0 && num == 0) || 1587 (free_size > 0 && (num == 1 || num == 2)), 1588 "There should be at most 2 free chunks after compaction"); 1589 #endif // ASSERT 1590 _collectorState = Resetting; 1591 assert(_restart_addr == NULL, 1592 "Should have been NULL'd before baton was passed"); 1593 reset_stw(); 1594 _cmsGen->reset_after_compaction(); 1595 _concurrent_cycles_since_last_unload = 0; 1596 1597 // Clear any data recorded in the PLAB chunk arrays. 1598 if (_survivor_plab_array != NULL) { 1599 reset_survivor_plab_arrays(); 1600 } 1601 1602 // Adjust the per-size allocation stats for the next epoch. 1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1604 // Restart the "inter sweep timer" for the next epoch. 1605 _inter_sweep_timer.reset(); 1606 _inter_sweep_timer.start(); 1607 1608 // No longer a need to do a concurrent collection for Metaspace. 1609 MetaspaceGC::set_should_concurrent_collect(false); 1610 1611 heap->post_full_gc_dump(gc_timer); 1612 1613 gc_timer->register_gc_end(); 1614 1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1616 1617 // For a mark-sweep-compact, compute_new_size() will be called 1618 // in the heap's do_collection() method. 1619 } 1620 1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1622 Log(gc, heap) log; 1623 if (!log.is_trace()) { 1624 return; 1625 } 1626 1627 ContiguousSpace* eden_space = _young_gen->eden(); 1628 ContiguousSpace* from_space = _young_gen->from(); 1629 ContiguousSpace* to_space = _young_gen->to(); 1630 // Eden 1631 if (_eden_chunk_array != NULL) { 1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1633 p2i(eden_space->bottom()), p2i(eden_space->top()), 1634 p2i(eden_space->end()), eden_space->capacity()); 1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1636 _eden_chunk_index, _eden_chunk_capacity); 1637 for (size_t i = 0; i < _eden_chunk_index; i++) { 1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1639 } 1640 } 1641 // Survivor 1642 if (_survivor_chunk_array != NULL) { 1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1644 p2i(from_space->bottom()), p2i(from_space->top()), 1645 p2i(from_space->end()), from_space->capacity()); 1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1647 _survivor_chunk_index, _survivor_chunk_capacity); 1648 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1650 } 1651 } 1652 } 1653 1654 void CMSCollector::getFreelistLocks() const { 1655 // Get locks for all free lists in all generations that this 1656 // collector is responsible for 1657 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1658 } 1659 1660 void CMSCollector::releaseFreelistLocks() const { 1661 // Release locks for all free lists in all generations that this 1662 // collector is responsible for 1663 _cmsGen->freelistLock()->unlock(); 1664 } 1665 1666 bool CMSCollector::haveFreelistLocks() const { 1667 // Check locks for all free lists in all generations that this 1668 // collector is responsible for 1669 assert_lock_strong(_cmsGen->freelistLock()); 1670 PRODUCT_ONLY(ShouldNotReachHere()); 1671 return true; 1672 } 1673 1674 // A utility class that is used by the CMS collector to 1675 // temporarily "release" the foreground collector from its 1676 // usual obligation to wait for the background collector to 1677 // complete an ongoing phase before proceeding. 1678 class ReleaseForegroundGC: public StackObj { 1679 private: 1680 CMSCollector* _c; 1681 public: 1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1685 // allow a potentially blocked foreground collector to proceed 1686 _c->_foregroundGCShouldWait = false; 1687 if (_c->_foregroundGCIsActive) { 1688 CGC_lock->notify(); 1689 } 1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1691 "Possible deadlock"); 1692 } 1693 1694 ~ReleaseForegroundGC() { 1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1697 _c->_foregroundGCShouldWait = true; 1698 } 1699 }; 1700 1701 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1702 assert(Thread::current()->is_ConcurrentGC_thread(), 1703 "A CMS asynchronous collection is only allowed on a CMS thread."); 1704 1705 CMSHeap* heap = CMSHeap::heap(); 1706 { 1707 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1708 MutexLockerEx hl(Heap_lock, safepoint_check); 1709 FreelistLocker fll(this); 1710 MutexLockerEx x(CGC_lock, safepoint_check); 1711 if (_foregroundGCIsActive) { 1712 // The foreground collector is. Skip this 1713 // background collection. 1714 assert(!_foregroundGCShouldWait, "Should be clear"); 1715 return; 1716 } else { 1717 assert(_collectorState == Idling, "Should be idling before start."); 1718 _collectorState = InitialMarking; 1719 register_gc_start(cause); 1720 // Reset the expansion cause, now that we are about to begin 1721 // a new cycle. 1722 clear_expansion_cause(); 1723 1724 // Clear the MetaspaceGC flag since a concurrent collection 1725 // is starting but also clear it after the collection. 1726 MetaspaceGC::set_should_concurrent_collect(false); 1727 } 1728 // Decide if we want to enable class unloading as part of the 1729 // ensuing concurrent GC cycle. 1730 update_should_unload_classes(); 1731 _full_gc_requested = false; // acks all outstanding full gc requests 1732 _full_gc_cause = GCCause::_no_gc; 1733 // Signal that we are about to start a collection 1734 heap->increment_total_full_collections(); // ... starting a collection cycle 1735 _collection_count_start = heap->total_full_collections(); 1736 } 1737 1738 size_t prev_used = _cmsGen->used(); 1739 1740 // The change of the collection state is normally done at this level; 1741 // the exceptions are phases that are executed while the world is 1742 // stopped. For those phases the change of state is done while the 1743 // world is stopped. For baton passing purposes this allows the 1744 // background collector to finish the phase and change state atomically. 1745 // The foreground collector cannot wait on a phase that is done 1746 // while the world is stopped because the foreground collector already 1747 // has the world stopped and would deadlock. 1748 while (_collectorState != Idling) { 1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1750 p2i(Thread::current()), _collectorState); 1751 // The foreground collector 1752 // holds the Heap_lock throughout its collection. 1753 // holds the CMS token (but not the lock) 1754 // except while it is waiting for the background collector to yield. 1755 // 1756 // The foreground collector should be blocked (not for long) 1757 // if the background collector is about to start a phase 1758 // executed with world stopped. If the background 1759 // collector has already started such a phase, the 1760 // foreground collector is blocked waiting for the 1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1762 // are executed in the VM thread. 1763 // 1764 // The locking order is 1765 // PendingListLock (PLL) -- if applicable (FinalMarking) 1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1767 // CMS token (claimed in 1768 // stop_world_and_do() --> 1769 // safepoint_synchronize() --> 1770 // CMSThread::synchronize()) 1771 1772 { 1773 // Check if the FG collector wants us to yield. 1774 CMSTokenSync x(true); // is cms thread 1775 if (waitForForegroundGC()) { 1776 // We yielded to a foreground GC, nothing more to be 1777 // done this round. 1778 assert(_foregroundGCShouldWait == false, "We set it to false in " 1779 "waitForForegroundGC()"); 1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1781 p2i(Thread::current()), _collectorState); 1782 return; 1783 } else { 1784 // The background collector can run but check to see if the 1785 // foreground collector has done a collection while the 1786 // background collector was waiting to get the CGC_lock 1787 // above. If yes, break so that _foregroundGCShouldWait 1788 // is cleared before returning. 1789 if (_collectorState == Idling) { 1790 break; 1791 } 1792 } 1793 } 1794 1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1796 "should be waiting"); 1797 1798 switch (_collectorState) { 1799 case InitialMarking: 1800 { 1801 ReleaseForegroundGC x(this); 1802 stats().record_cms_begin(); 1803 VM_CMS_Initial_Mark initial_mark_op(this); 1804 VMThread::execute(&initial_mark_op); 1805 } 1806 // The collector state may be any legal state at this point 1807 // since the background collector may have yielded to the 1808 // foreground collector. 1809 break; 1810 case Marking: 1811 // initial marking in checkpointRootsInitialWork has been completed 1812 if (markFromRoots()) { // we were successful 1813 assert(_collectorState == Precleaning, "Collector state should " 1814 "have changed"); 1815 } else { 1816 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1817 } 1818 break; 1819 case Precleaning: 1820 // marking from roots in markFromRoots has been completed 1821 preclean(); 1822 assert(_collectorState == AbortablePreclean || 1823 _collectorState == FinalMarking, 1824 "Collector state should have changed"); 1825 break; 1826 case AbortablePreclean: 1827 abortable_preclean(); 1828 assert(_collectorState == FinalMarking, "Collector state should " 1829 "have changed"); 1830 break; 1831 case FinalMarking: 1832 { 1833 ReleaseForegroundGC x(this); 1834 1835 VM_CMS_Final_Remark final_remark_op(this); 1836 VMThread::execute(&final_remark_op); 1837 } 1838 assert(_foregroundGCShouldWait, "block post-condition"); 1839 break; 1840 case Sweeping: 1841 // final marking in checkpointRootsFinal has been completed 1842 sweep(); 1843 assert(_collectorState == Resizing, "Collector state change " 1844 "to Resizing must be done under the free_list_lock"); 1845 1846 case Resizing: { 1847 // Sweeping has been completed... 1848 // At this point the background collection has completed. 1849 // Don't move the call to compute_new_size() down 1850 // into code that might be executed if the background 1851 // collection was preempted. 1852 { 1853 ReleaseForegroundGC x(this); // unblock FG collection 1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1855 CMSTokenSync z(true); // not strictly needed. 1856 if (_collectorState == Resizing) { 1857 compute_new_size(); 1858 save_heap_summary(); 1859 _collectorState = Resetting; 1860 } else { 1861 assert(_collectorState == Idling, "The state should only change" 1862 " because the foreground collector has finished the collection"); 1863 } 1864 } 1865 break; 1866 } 1867 case Resetting: 1868 // CMS heap resizing has been completed 1869 reset_concurrent(); 1870 assert(_collectorState == Idling, "Collector state should " 1871 "have changed"); 1872 1873 MetaspaceGC::set_should_concurrent_collect(false); 1874 1875 stats().record_cms_end(); 1876 // Don't move the concurrent_phases_end() and compute_new_size() 1877 // calls to here because a preempted background collection 1878 // has it's state set to "Resetting". 1879 break; 1880 case Idling: 1881 default: 1882 ShouldNotReachHere(); 1883 break; 1884 } 1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1886 p2i(Thread::current()), _collectorState); 1887 assert(_foregroundGCShouldWait, "block post-condition"); 1888 } 1889 1890 // Should this be in gc_epilogue? 1891 heap->counters()->update_counters(); 1892 1893 { 1894 // Clear _foregroundGCShouldWait and, in the event that the 1895 // foreground collector is waiting, notify it, before 1896 // returning. 1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1898 _foregroundGCShouldWait = false; 1899 if (_foregroundGCIsActive) { 1900 CGC_lock->notify(); 1901 } 1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1903 "Possible deadlock"); 1904 } 1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1906 p2i(Thread::current()), _collectorState); 1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1909 } 1910 1911 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1912 _cms_start_registered = true; 1913 _gc_timer_cm->register_gc_start(); 1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1915 } 1916 1917 void CMSCollector::register_gc_end() { 1918 if (_cms_start_registered) { 1919 report_heap_summary(GCWhen::AfterGC); 1920 1921 _gc_timer_cm->register_gc_end(); 1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1923 _cms_start_registered = false; 1924 } 1925 } 1926 1927 void CMSCollector::save_heap_summary() { 1928 CMSHeap* heap = CMSHeap::heap(); 1929 _last_heap_summary = heap->create_heap_summary(); 1930 _last_metaspace_summary = heap->create_metaspace_summary(); 1931 } 1932 1933 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1936 } 1937 1938 bool CMSCollector::waitForForegroundGC() { 1939 bool res = false; 1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1941 "CMS thread should have CMS token"); 1942 // Block the foreground collector until the 1943 // background collectors decides whether to 1944 // yield. 1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1946 _foregroundGCShouldWait = true; 1947 if (_foregroundGCIsActive) { 1948 // The background collector yields to the 1949 // foreground collector and returns a value 1950 // indicating that it has yielded. The foreground 1951 // collector can proceed. 1952 res = true; 1953 _foregroundGCShouldWait = false; 1954 ConcurrentMarkSweepThread::clear_CMS_flag( 1955 ConcurrentMarkSweepThread::CMS_cms_has_token); 1956 ConcurrentMarkSweepThread::set_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1958 // Get a possibly blocked foreground thread going 1959 CGC_lock->notify(); 1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1961 p2i(Thread::current()), _collectorState); 1962 while (_foregroundGCIsActive) { 1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1964 } 1965 ConcurrentMarkSweepThread::set_CMS_flag( 1966 ConcurrentMarkSweepThread::CMS_cms_has_token); 1967 ConcurrentMarkSweepThread::clear_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1969 } 1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 return res; 1973 } 1974 1975 // Because of the need to lock the free lists and other structures in 1976 // the collector, common to all the generations that the collector is 1977 // collecting, we need the gc_prologues of individual CMS generations 1978 // delegate to their collector. It may have been simpler had the 1979 // current infrastructure allowed one to call a prologue on a 1980 // collector. In the absence of that we have the generation's 1981 // prologue delegate to the collector, which delegates back 1982 // some "local" work to a worker method in the individual generations 1983 // that it's responsible for collecting, while itself doing any 1984 // work common to all generations it's responsible for. A similar 1985 // comment applies to the gc_epilogue()'s. 1986 // The role of the variable _between_prologue_and_epilogue is to 1987 // enforce the invocation protocol. 1988 void CMSCollector::gc_prologue(bool full) { 1989 // Call gc_prologue_work() for the CMSGen 1990 // we are responsible for. 1991 1992 // The following locking discipline assumes that we are only called 1993 // when the world is stopped. 1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1995 1996 // The CMSCollector prologue must call the gc_prologues for the 1997 // "generations" that it's responsible 1998 // for. 1999 2000 assert( Thread::current()->is_VM_thread() 2001 || ( CMSScavengeBeforeRemark 2002 && Thread::current()->is_ConcurrentGC_thread()), 2003 "Incorrect thread type for prologue execution"); 2004 2005 if (_between_prologue_and_epilogue) { 2006 // We have already been invoked; this is a gc_prologue delegation 2007 // from yet another CMS generation that we are responsible for, just 2008 // ignore it since all relevant work has already been done. 2009 return; 2010 } 2011 2012 // set a bit saying prologue has been called; cleared in epilogue 2013 _between_prologue_and_epilogue = true; 2014 // Claim locks for common data structures, then call gc_prologue_work() 2015 // for each CMSGen. 2016 2017 getFreelistLocks(); // gets free list locks on constituent spaces 2018 bitMapLock()->lock_without_safepoint_check(); 2019 2020 // Should call gc_prologue_work() for all cms gens we are responsible for 2021 bool duringMarking = _collectorState >= Marking 2022 && _collectorState < Sweeping; 2023 2024 // The young collections clear the modified oops state, which tells if 2025 // there are any modified oops in the class. The remark phase also needs 2026 // that information. Tell the young collection to save the union of all 2027 // modified klasses. 2028 if (duringMarking) { 2029 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2030 } 2031 2032 bool registerClosure = duringMarking; 2033 2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2035 2036 if (!full) { 2037 stats().record_gc0_begin(); 2038 } 2039 } 2040 2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2042 2043 _capacity_at_prologue = capacity(); 2044 _used_at_prologue = used(); 2045 2046 // We enable promotion tracking so that card-scanning can recognize 2047 // which objects have been promoted during this GC and skip them. 2048 for (uint i = 0; i < ParallelGCThreads; i++) { 2049 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2050 } 2051 2052 // Delegate to CMScollector which knows how to coordinate between 2053 // this and any other CMS generations that it is responsible for 2054 // collecting. 2055 collector()->gc_prologue(full); 2056 } 2057 2058 // This is a "private" interface for use by this generation's CMSCollector. 2059 // Not to be called directly by any other entity (for instance, 2060 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2062 bool registerClosure, ModUnionClosure* modUnionClosure) { 2063 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2065 "Should be NULL"); 2066 if (registerClosure) { 2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2068 } 2069 cmsSpace()->gc_prologue(); 2070 // Clear stat counters 2071 NOT_PRODUCT( 2072 assert(_numObjectsPromoted == 0, "check"); 2073 assert(_numWordsPromoted == 0, "check"); 2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2076 _numObjectsAllocated = 0; 2077 _numWordsAllocated = 0; 2078 ) 2079 } 2080 2081 void CMSCollector::gc_epilogue(bool full) { 2082 // The following locking discipline assumes that we are only called 2083 // when the world is stopped. 2084 assert(SafepointSynchronize::is_at_safepoint(), 2085 "world is stopped assumption"); 2086 2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2088 // if linear allocation blocks need to be appropriately marked to allow the 2089 // the blocks to be parsable. We also check here whether we need to nudge the 2090 // CMS collector thread to start a new cycle (if it's not already active). 2091 assert( Thread::current()->is_VM_thread() 2092 || ( CMSScavengeBeforeRemark 2093 && Thread::current()->is_ConcurrentGC_thread()), 2094 "Incorrect thread type for epilogue execution"); 2095 2096 if (!_between_prologue_and_epilogue) { 2097 // We have already been invoked; this is a gc_epilogue delegation 2098 // from yet another CMS generation that we are responsible for, just 2099 // ignore it since all relevant work has already been done. 2100 return; 2101 } 2102 assert(haveFreelistLocks(), "must have freelist locks"); 2103 assert_lock_strong(bitMapLock()); 2104 2105 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2106 2107 _cmsGen->gc_epilogue_work(full); 2108 2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2110 // in case sampling was not already enabled, enable it 2111 _start_sampling = true; 2112 } 2113 // reset _eden_chunk_array so sampling starts afresh 2114 _eden_chunk_index = 0; 2115 2116 size_t cms_used = _cmsGen->cmsSpace()->used(); 2117 2118 // update performance counters - this uses a special version of 2119 // update_counters() that allows the utilization to be passed as a 2120 // parameter, avoiding multiple calls to used(). 2121 // 2122 _cmsGen->update_counters(cms_used); 2123 2124 bitMapLock()->unlock(); 2125 releaseFreelistLocks(); 2126 2127 if (!CleanChunkPoolAsync) { 2128 Chunk::clean_chunk_pool(); 2129 } 2130 2131 set_did_compact(false); 2132 _between_prologue_and_epilogue = false; // ready for next cycle 2133 } 2134 2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2136 collector()->gc_epilogue(full); 2137 2138 // When using ParNew, promotion tracking should have already been 2139 // disabled. However, the prologue (which enables promotion 2140 // tracking) and epilogue are called irrespective of the type of 2141 // GC. So they will also be called before and after Full GCs, during 2142 // which promotion tracking will not be explicitly disabled. So, 2143 // it's safer to also disable it here too (to be symmetric with 2144 // enabling it in the prologue). 2145 for (uint i = 0; i < ParallelGCThreads; i++) { 2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2147 } 2148 } 2149 2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2151 assert(!incremental_collection_failed(), "Should have been cleared"); 2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2153 cmsSpace()->gc_epilogue(); 2154 // Print stat counters 2155 NOT_PRODUCT( 2156 assert(_numObjectsAllocated == 0, "check"); 2157 assert(_numWordsAllocated == 0, "check"); 2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2160 _numObjectsPromoted = 0; 2161 _numWordsPromoted = 0; 2162 ) 2163 2164 // Call down the chain in contiguous_available needs the freelistLock 2165 // so print this out before releasing the freeListLock. 2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2167 } 2168 2169 #ifndef PRODUCT 2170 bool CMSCollector::have_cms_token() { 2171 Thread* thr = Thread::current(); 2172 if (thr->is_VM_thread()) { 2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2174 } else if (thr->is_ConcurrentGC_thread()) { 2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2176 } else if (thr->is_GC_task_thread()) { 2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2178 ParGCRareEvent_lock->owned_by_self(); 2179 } 2180 return false; 2181 } 2182 2183 // Check reachability of the given heap address in CMS generation, 2184 // treating all other generations as roots. 2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2186 // We could "guarantee" below, rather than assert, but I'll 2187 // leave these as "asserts" so that an adventurous debugger 2188 // could try this in the product build provided some subset of 2189 // the conditions were met, provided they were interested in the 2190 // results and knew that the computation below wouldn't interfere 2191 // with other concurrent computations mutating the structures 2192 // being read or written. 2193 assert(SafepointSynchronize::is_at_safepoint(), 2194 "Else mutations in object graph will make answer suspect"); 2195 assert(have_cms_token(), "Should hold cms token"); 2196 assert(haveFreelistLocks(), "must hold free list locks"); 2197 assert_lock_strong(bitMapLock()); 2198 2199 // Clear the marking bit map array before starting, but, just 2200 // for kicks, first report if the given address is already marked 2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2202 _markBitMap.isMarked(addr) ? "" : " not"); 2203 2204 if (verify_after_remark()) { 2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2206 bool result = verification_mark_bm()->isMarked(addr); 2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2208 result ? "IS" : "is NOT"); 2209 return result; 2210 } else { 2211 tty->print_cr("Could not compute result"); 2212 return false; 2213 } 2214 } 2215 #endif 2216 2217 void 2218 CMSCollector::print_on_error(outputStream* st) { 2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2220 if (collector != NULL) { 2221 CMSBitMap* bitmap = &collector->_markBitMap; 2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2223 bitmap->print_on_error(st, " Bits: "); 2224 2225 st->cr(); 2226 2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2229 mut_bitmap->print_on_error(st, " Bits: "); 2230 } 2231 } 2232 2233 //////////////////////////////////////////////////////// 2234 // CMS Verification Support 2235 //////////////////////////////////////////////////////// 2236 // Following the remark phase, the following invariant 2237 // should hold -- each object in the CMS heap which is 2238 // marked in markBitMap() should be marked in the verification_mark_bm(). 2239 2240 class VerifyMarkedClosure: public BitMapClosure { 2241 CMSBitMap* _marks; 2242 bool _failed; 2243 2244 public: 2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2246 2247 bool do_bit(size_t offset) { 2248 HeapWord* addr = _marks->offsetToHeapWord(offset); 2249 if (!_marks->isMarked(addr)) { 2250 Log(gc, verify) log; 2251 ResourceMark rm; 2252 LogStream ls(log.error()); 2253 oop(addr)->print_on(&ls); 2254 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2255 _failed = true; 2256 } 2257 return true; 2258 } 2259 2260 bool failed() { return _failed; } 2261 }; 2262 2263 bool CMSCollector::verify_after_remark() { 2264 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2265 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2266 static bool init = false; 2267 2268 assert(SafepointSynchronize::is_at_safepoint(), 2269 "Else mutations in object graph will make answer suspect"); 2270 assert(have_cms_token(), 2271 "Else there may be mutual interference in use of " 2272 " verification data structures"); 2273 assert(_collectorState > Marking && _collectorState <= Sweeping, 2274 "Else marking info checked here may be obsolete"); 2275 assert(haveFreelistLocks(), "must hold free list locks"); 2276 assert_lock_strong(bitMapLock()); 2277 2278 2279 // Allocate marking bit map if not already allocated 2280 if (!init) { // first time 2281 if (!verification_mark_bm()->allocate(_span)) { 2282 return false; 2283 } 2284 init = true; 2285 } 2286 2287 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2288 2289 // Turn off refs discovery -- so we will be tracing through refs. 2290 // This is as intended, because by this time 2291 // GC must already have cleared any refs that need to be cleared, 2292 // and traced those that need to be marked; moreover, 2293 // the marking done here is not going to interfere in any 2294 // way with the marking information used by GC. 2295 NoRefDiscovery no_discovery(ref_processor()); 2296 2297 #if COMPILER2_OR_JVMCI 2298 DerivedPointerTableDeactivate dpt_deact; 2299 #endif 2300 2301 // Clear any marks from a previous round 2302 verification_mark_bm()->clear_all(); 2303 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2304 verify_work_stacks_empty(); 2305 2306 CMSHeap* heap = CMSHeap::heap(); 2307 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2308 // Update the saved marks which may affect the root scans. 2309 heap->save_marks(); 2310 2311 if (CMSRemarkVerifyVariant == 1) { 2312 // In this first variant of verification, we complete 2313 // all marking, then check if the new marks-vector is 2314 // a subset of the CMS marks-vector. 2315 verify_after_remark_work_1(); 2316 } else { 2317 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2318 // In this second variant of verification, we flag an error 2319 // (i.e. an object reachable in the new marks-vector not reachable 2320 // in the CMS marks-vector) immediately, also indicating the 2321 // identify of an object (A) that references the unmarked object (B) -- 2322 // presumably, a mutation to A failed to be picked up by preclean/remark? 2323 verify_after_remark_work_2(); 2324 } 2325 2326 return true; 2327 } 2328 2329 void CMSCollector::verify_after_remark_work_1() { 2330 ResourceMark rm; 2331 HandleMark hm; 2332 CMSHeap* heap = CMSHeap::heap(); 2333 2334 // Get a clear set of claim bits for the roots processing to work with. 2335 ClassLoaderDataGraph::clear_claimed_marks(); 2336 2337 // Mark from roots one level into CMS 2338 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2339 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2340 2341 { 2342 StrongRootsScope srs(1); 2343 2344 heap->cms_process_roots(&srs, 2345 true, // young gen as roots 2346 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2347 should_unload_classes(), 2348 ¬Older, 2349 NULL); 2350 } 2351 2352 // Now mark from the roots 2353 MarkFromRootsClosure markFromRootsClosure(this, _span, 2354 verification_mark_bm(), verification_mark_stack(), 2355 false /* don't yield */, true /* verifying */); 2356 assert(_restart_addr == NULL, "Expected pre-condition"); 2357 verification_mark_bm()->iterate(&markFromRootsClosure); 2358 while (_restart_addr != NULL) { 2359 // Deal with stack overflow: by restarting at the indicated 2360 // address. 2361 HeapWord* ra = _restart_addr; 2362 markFromRootsClosure.reset(ra); 2363 _restart_addr = NULL; 2364 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2365 } 2366 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2367 verify_work_stacks_empty(); 2368 2369 // Marking completed -- now verify that each bit marked in 2370 // verification_mark_bm() is also marked in markBitMap(); flag all 2371 // errors by printing corresponding objects. 2372 VerifyMarkedClosure vcl(markBitMap()); 2373 verification_mark_bm()->iterate(&vcl); 2374 if (vcl.failed()) { 2375 Log(gc, verify) log; 2376 log.error("Failed marking verification after remark"); 2377 ResourceMark rm; 2378 LogStream ls(log.error()); 2379 heap->print_on(&ls); 2380 fatal("CMS: failed marking verification after remark"); 2381 } 2382 } 2383 2384 class VerifyCLDOopsCLDClosure : public CLDClosure { 2385 class VerifyCLDOopsClosure : public OopClosure { 2386 CMSBitMap* _bitmap; 2387 public: 2388 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2389 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2390 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2391 } _oop_closure; 2392 public: 2393 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2394 void do_cld(ClassLoaderData* cld) { 2395 cld->oops_do(&_oop_closure, false, false); 2396 } 2397 }; 2398 2399 void CMSCollector::verify_after_remark_work_2() { 2400 ResourceMark rm; 2401 HandleMark hm; 2402 CMSHeap* heap = CMSHeap::heap(); 2403 2404 // Get a clear set of claim bits for the roots processing to work with. 2405 ClassLoaderDataGraph::clear_claimed_marks(); 2406 2407 // Mark from roots one level into CMS 2408 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2409 markBitMap()); 2410 CLDToOopClosure cld_closure(¬Older, true); 2411 2412 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2413 2414 { 2415 StrongRootsScope srs(1); 2416 2417 heap->cms_process_roots(&srs, 2418 true, // young gen as roots 2419 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2420 should_unload_classes(), 2421 ¬Older, 2422 &cld_closure); 2423 } 2424 2425 // Now mark from the roots 2426 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2427 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2428 assert(_restart_addr == NULL, "Expected pre-condition"); 2429 verification_mark_bm()->iterate(&markFromRootsClosure); 2430 while (_restart_addr != NULL) { 2431 // Deal with stack overflow: by restarting at the indicated 2432 // address. 2433 HeapWord* ra = _restart_addr; 2434 markFromRootsClosure.reset(ra); 2435 _restart_addr = NULL; 2436 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2437 } 2438 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2439 verify_work_stacks_empty(); 2440 2441 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2442 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2443 2444 // Marking completed -- now verify that each bit marked in 2445 // verification_mark_bm() is also marked in markBitMap(); flag all 2446 // errors by printing corresponding objects. 2447 VerifyMarkedClosure vcl(markBitMap()); 2448 verification_mark_bm()->iterate(&vcl); 2449 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2450 } 2451 2452 void ConcurrentMarkSweepGeneration::save_marks() { 2453 // delegate to CMS space 2454 cmsSpace()->save_marks(); 2455 } 2456 2457 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2458 return cmsSpace()->no_allocs_since_save_marks(); 2459 } 2460 2461 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2462 \ 2463 void ConcurrentMarkSweepGeneration:: \ 2464 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2465 cl->set_generation(this); \ 2466 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2467 cl->reset_generation(); \ 2468 save_marks(); \ 2469 } 2470 2471 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2472 2473 void 2474 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2475 if (freelistLock()->owned_by_self()) { 2476 Generation::oop_iterate(cl); 2477 } else { 2478 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2479 Generation::oop_iterate(cl); 2480 } 2481 } 2482 2483 void 2484 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2485 if (freelistLock()->owned_by_self()) { 2486 Generation::object_iterate(cl); 2487 } else { 2488 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2489 Generation::object_iterate(cl); 2490 } 2491 } 2492 2493 void 2494 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2495 if (freelistLock()->owned_by_self()) { 2496 Generation::safe_object_iterate(cl); 2497 } else { 2498 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2499 Generation::safe_object_iterate(cl); 2500 } 2501 } 2502 2503 void 2504 ConcurrentMarkSweepGeneration::post_compact() { 2505 } 2506 2507 void 2508 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2509 // Fix the linear allocation blocks to look like free blocks. 2510 2511 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2512 // are not called when the heap is verified during universe initialization and 2513 // at vm shutdown. 2514 if (freelistLock()->owned_by_self()) { 2515 cmsSpace()->prepare_for_verify(); 2516 } else { 2517 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2518 cmsSpace()->prepare_for_verify(); 2519 } 2520 } 2521 2522 void 2523 ConcurrentMarkSweepGeneration::verify() { 2524 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2525 // are not called when the heap is verified during universe initialization and 2526 // at vm shutdown. 2527 if (freelistLock()->owned_by_self()) { 2528 cmsSpace()->verify(); 2529 } else { 2530 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2531 cmsSpace()->verify(); 2532 } 2533 } 2534 2535 void CMSCollector::verify() { 2536 _cmsGen->verify(); 2537 } 2538 2539 #ifndef PRODUCT 2540 bool CMSCollector::overflow_list_is_empty() const { 2541 assert(_num_par_pushes >= 0, "Inconsistency"); 2542 if (_overflow_list == NULL) { 2543 assert(_num_par_pushes == 0, "Inconsistency"); 2544 } 2545 return _overflow_list == NULL; 2546 } 2547 2548 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2549 // merely consolidate assertion checks that appear to occur together frequently. 2550 void CMSCollector::verify_work_stacks_empty() const { 2551 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2552 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2553 } 2554 2555 void CMSCollector::verify_overflow_empty() const { 2556 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2557 assert(no_preserved_marks(), "No preserved marks"); 2558 } 2559 #endif // PRODUCT 2560 2561 // Decide if we want to enable class unloading as part of the 2562 // ensuing concurrent GC cycle. We will collect and 2563 // unload classes if it's the case that: 2564 // (a) class unloading is enabled at the command line, and 2565 // (b) old gen is getting really full 2566 // NOTE: Provided there is no change in the state of the heap between 2567 // calls to this method, it should have idempotent results. Moreover, 2568 // its results should be monotonically increasing (i.e. going from 0 to 1, 2569 // but not 1 to 0) between successive calls between which the heap was 2570 // not collected. For the implementation below, it must thus rely on 2571 // the property that concurrent_cycles_since_last_unload() 2572 // will not decrease unless a collection cycle happened and that 2573 // _cmsGen->is_too_full() are 2574 // themselves also monotonic in that sense. See check_monotonicity() 2575 // below. 2576 void CMSCollector::update_should_unload_classes() { 2577 _should_unload_classes = false; 2578 if (CMSClassUnloadingEnabled) { 2579 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2580 CMSClassUnloadingMaxInterval) 2581 || _cmsGen->is_too_full(); 2582 } 2583 } 2584 2585 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2586 bool res = should_concurrent_collect(); 2587 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2588 return res; 2589 } 2590 2591 void CMSCollector::setup_cms_unloading_and_verification_state() { 2592 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2593 || VerifyBeforeExit; 2594 const int rso = GenCollectedHeap::SO_AllCodeCache; 2595 2596 // We set the proper root for this CMS cycle here. 2597 if (should_unload_classes()) { // Should unload classes this cycle 2598 remove_root_scanning_option(rso); // Shrink the root set appropriately 2599 set_verifying(should_verify); // Set verification state for this cycle 2600 return; // Nothing else needs to be done at this time 2601 } 2602 2603 // Not unloading classes this cycle 2604 assert(!should_unload_classes(), "Inconsistency!"); 2605 2606 // If we are not unloading classes then add SO_AllCodeCache to root 2607 // scanning options. 2608 add_root_scanning_option(rso); 2609 2610 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2611 set_verifying(true); 2612 } else if (verifying() && !should_verify) { 2613 // We were verifying, but some verification flags got disabled. 2614 set_verifying(false); 2615 // Exclude symbols, strings and code cache elements from root scanning to 2616 // reduce IM and RM pauses. 2617 remove_root_scanning_option(rso); 2618 } 2619 } 2620 2621 2622 #ifndef PRODUCT 2623 HeapWord* CMSCollector::block_start(const void* p) const { 2624 const HeapWord* addr = (HeapWord*)p; 2625 if (_span.contains(p)) { 2626 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2627 return _cmsGen->cmsSpace()->block_start(p); 2628 } 2629 } 2630 return NULL; 2631 } 2632 #endif 2633 2634 HeapWord* 2635 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2636 bool tlab, 2637 bool parallel) { 2638 CMSSynchronousYieldRequest yr; 2639 assert(!tlab, "Can't deal with TLAB allocation"); 2640 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2641 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2642 if (GCExpandToAllocateDelayMillis > 0) { 2643 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2644 } 2645 return have_lock_and_allocate(word_size, tlab); 2646 } 2647 2648 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2649 size_t bytes, 2650 size_t expand_bytes, 2651 CMSExpansionCause::Cause cause) 2652 { 2653 2654 bool success = expand(bytes, expand_bytes); 2655 2656 // remember why we expanded; this information is used 2657 // by shouldConcurrentCollect() when making decisions on whether to start 2658 // a new CMS cycle. 2659 if (success) { 2660 set_expansion_cause(cause); 2661 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2662 } 2663 } 2664 2665 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2666 HeapWord* res = NULL; 2667 MutexLocker x(ParGCRareEvent_lock); 2668 while (true) { 2669 // Expansion by some other thread might make alloc OK now: 2670 res = ps->lab.alloc(word_sz); 2671 if (res != NULL) return res; 2672 // If there's not enough expansion space available, give up. 2673 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2674 return NULL; 2675 } 2676 // Otherwise, we try expansion. 2677 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2678 // Now go around the loop and try alloc again; 2679 // A competing par_promote might beat us to the expansion space, 2680 // so we may go around the loop again if promotion fails again. 2681 if (GCExpandToAllocateDelayMillis > 0) { 2682 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2683 } 2684 } 2685 } 2686 2687 2688 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2689 PromotionInfo* promo) { 2690 MutexLocker x(ParGCRareEvent_lock); 2691 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2692 while (true) { 2693 // Expansion by some other thread might make alloc OK now: 2694 if (promo->ensure_spooling_space()) { 2695 assert(promo->has_spooling_space(), 2696 "Post-condition of successful ensure_spooling_space()"); 2697 return true; 2698 } 2699 // If there's not enough expansion space available, give up. 2700 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2701 return false; 2702 } 2703 // Otherwise, we try expansion. 2704 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2705 // Now go around the loop and try alloc again; 2706 // A competing allocation might beat us to the expansion space, 2707 // so we may go around the loop again if allocation fails again. 2708 if (GCExpandToAllocateDelayMillis > 0) { 2709 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2710 } 2711 } 2712 } 2713 2714 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2715 // Only shrink if a compaction was done so that all the free space 2716 // in the generation is in a contiguous block at the end. 2717 if (did_compact()) { 2718 CardGeneration::shrink(bytes); 2719 } 2720 } 2721 2722 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2723 assert_locked_or_safepoint(Heap_lock); 2724 } 2725 2726 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2727 assert_locked_or_safepoint(Heap_lock); 2728 assert_lock_strong(freelistLock()); 2729 log_trace(gc)("Shrinking of CMS not yet implemented"); 2730 return; 2731 } 2732 2733 2734 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2735 // phases. 2736 class CMSPhaseAccounting: public StackObj { 2737 public: 2738 CMSPhaseAccounting(CMSCollector *collector, 2739 const char *title); 2740 ~CMSPhaseAccounting(); 2741 2742 private: 2743 CMSCollector *_collector; 2744 const char *_title; 2745 GCTraceConcTime(Info, gc) _trace_time; 2746 2747 public: 2748 // Not MT-safe; so do not pass around these StackObj's 2749 // where they may be accessed by other threads. 2750 double wallclock_millis() { 2751 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2752 } 2753 }; 2754 2755 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2756 const char *title) : 2757 _collector(collector), _title(title), _trace_time(title) { 2758 2759 _collector->resetYields(); 2760 _collector->resetTimer(); 2761 _collector->startTimer(); 2762 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2763 } 2764 2765 CMSPhaseAccounting::~CMSPhaseAccounting() { 2766 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2767 _collector->stopTimer(); 2768 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2769 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2770 } 2771 2772 // CMS work 2773 2774 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2775 class CMSParMarkTask : public AbstractGangTask { 2776 protected: 2777 CMSCollector* _collector; 2778 uint _n_workers; 2779 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2780 AbstractGangTask(name), 2781 _collector(collector), 2782 _n_workers(n_workers) {} 2783 // Work method in support of parallel rescan ... of young gen spaces 2784 void do_young_space_rescan(OopsInGenClosure* cl, 2785 ContiguousSpace* space, 2786 HeapWord** chunk_array, size_t chunk_top); 2787 void work_on_young_gen_roots(OopsInGenClosure* cl); 2788 }; 2789 2790 // Parallel initial mark task 2791 class CMSParInitialMarkTask: public CMSParMarkTask { 2792 StrongRootsScope* _strong_roots_scope; 2793 public: 2794 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2795 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2796 _strong_roots_scope(strong_roots_scope) {} 2797 void work(uint worker_id); 2798 }; 2799 2800 // Checkpoint the roots into this generation from outside 2801 // this generation. [Note this initial checkpoint need only 2802 // be approximate -- we'll do a catch up phase subsequently.] 2803 void CMSCollector::checkpointRootsInitial() { 2804 assert(_collectorState == InitialMarking, "Wrong collector state"); 2805 check_correct_thread_executing(); 2806 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2807 2808 save_heap_summary(); 2809 report_heap_summary(GCWhen::BeforeGC); 2810 2811 ReferenceProcessor* rp = ref_processor(); 2812 assert(_restart_addr == NULL, "Control point invariant"); 2813 { 2814 // acquire locks for subsequent manipulations 2815 MutexLockerEx x(bitMapLock(), 2816 Mutex::_no_safepoint_check_flag); 2817 checkpointRootsInitialWork(); 2818 // enable ("weak") refs discovery 2819 rp->enable_discovery(); 2820 _collectorState = Marking; 2821 } 2822 } 2823 2824 void CMSCollector::checkpointRootsInitialWork() { 2825 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2826 assert(_collectorState == InitialMarking, "just checking"); 2827 2828 // Already have locks. 2829 assert_lock_strong(bitMapLock()); 2830 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2831 2832 // Setup the verification and class unloading state for this 2833 // CMS collection cycle. 2834 setup_cms_unloading_and_verification_state(); 2835 2836 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2837 2838 // Reset all the PLAB chunk arrays if necessary. 2839 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2840 reset_survivor_plab_arrays(); 2841 } 2842 2843 ResourceMark rm; 2844 HandleMark hm; 2845 2846 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2847 CMSHeap* heap = CMSHeap::heap(); 2848 2849 verify_work_stacks_empty(); 2850 verify_overflow_empty(); 2851 2852 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2853 // Update the saved marks which may affect the root scans. 2854 heap->save_marks(); 2855 2856 // weak reference processing has not started yet. 2857 ref_processor()->set_enqueuing_is_done(false); 2858 2859 // Need to remember all newly created CLDs, 2860 // so that we can guarantee that the remark finds them. 2861 ClassLoaderDataGraph::remember_new_clds(true); 2862 2863 // Whenever a CLD is found, it will be claimed before proceeding to mark 2864 // the klasses. The claimed marks need to be cleared before marking starts. 2865 ClassLoaderDataGraph::clear_claimed_marks(); 2866 2867 print_eden_and_survivor_chunk_arrays(); 2868 2869 { 2870 #if COMPILER2_OR_JVMCI 2871 DerivedPointerTableDeactivate dpt_deact; 2872 #endif 2873 if (CMSParallelInitialMarkEnabled) { 2874 // The parallel version. 2875 WorkGang* workers = heap->workers(); 2876 assert(workers != NULL, "Need parallel worker threads."); 2877 uint n_workers = workers->active_workers(); 2878 2879 StrongRootsScope srs(n_workers); 2880 2881 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2882 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2883 // If the total workers is greater than 1, then multiple workers 2884 // may be used at some time and the initialization has been set 2885 // such that the single threaded path cannot be used. 2886 if (workers->total_workers() > 1) { 2887 workers->run_task(&tsk); 2888 } else { 2889 tsk.work(0); 2890 } 2891 } else { 2892 // The serial version. 2893 CLDToOopClosure cld_closure(¬Older, true); 2894 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2895 2896 StrongRootsScope srs(1); 2897 2898 heap->cms_process_roots(&srs, 2899 true, // young gen as roots 2900 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2901 should_unload_classes(), 2902 ¬Older, 2903 &cld_closure); 2904 } 2905 } 2906 2907 // Clear mod-union table; it will be dirtied in the prologue of 2908 // CMS generation per each young generation collection. 2909 2910 assert(_modUnionTable.isAllClear(), 2911 "Was cleared in most recent final checkpoint phase" 2912 " or no bits are set in the gc_prologue before the start of the next " 2913 "subsequent marking phase."); 2914 2915 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2916 2917 // Save the end of the used_region of the constituent generations 2918 // to be used to limit the extent of sweep in each generation. 2919 save_sweep_limits(); 2920 verify_overflow_empty(); 2921 } 2922 2923 bool CMSCollector::markFromRoots() { 2924 // we might be tempted to assert that: 2925 // assert(!SafepointSynchronize::is_at_safepoint(), 2926 // "inconsistent argument?"); 2927 // However that wouldn't be right, because it's possible that 2928 // a safepoint is indeed in progress as a young generation 2929 // stop-the-world GC happens even as we mark in this generation. 2930 assert(_collectorState == Marking, "inconsistent state?"); 2931 check_correct_thread_executing(); 2932 verify_overflow_empty(); 2933 2934 // Weak ref discovery note: We may be discovering weak 2935 // refs in this generation concurrent (but interleaved) with 2936 // weak ref discovery by the young generation collector. 2937 2938 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2939 GCTraceCPUTime tcpu; 2940 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2941 bool res = markFromRootsWork(); 2942 if (res) { 2943 _collectorState = Precleaning; 2944 } else { // We failed and a foreground collection wants to take over 2945 assert(_foregroundGCIsActive, "internal state inconsistency"); 2946 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2947 log_debug(gc)("bailing out to foreground collection"); 2948 } 2949 verify_overflow_empty(); 2950 return res; 2951 } 2952 2953 bool CMSCollector::markFromRootsWork() { 2954 // iterate over marked bits in bit map, doing a full scan and mark 2955 // from these roots using the following algorithm: 2956 // . if oop is to the right of the current scan pointer, 2957 // mark corresponding bit (we'll process it later) 2958 // . else (oop is to left of current scan pointer) 2959 // push oop on marking stack 2960 // . drain the marking stack 2961 2962 // Note that when we do a marking step we need to hold the 2963 // bit map lock -- recall that direct allocation (by mutators) 2964 // and promotion (by the young generation collector) is also 2965 // marking the bit map. [the so-called allocate live policy.] 2966 // Because the implementation of bit map marking is not 2967 // robust wrt simultaneous marking of bits in the same word, 2968 // we need to make sure that there is no such interference 2969 // between concurrent such updates. 2970 2971 // already have locks 2972 assert_lock_strong(bitMapLock()); 2973 2974 verify_work_stacks_empty(); 2975 verify_overflow_empty(); 2976 bool result = false; 2977 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2978 result = do_marking_mt(); 2979 } else { 2980 result = do_marking_st(); 2981 } 2982 return result; 2983 } 2984 2985 // Forward decl 2986 class CMSConcMarkingTask; 2987 2988 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2989 CMSCollector* _collector; 2990 CMSConcMarkingTask* _task; 2991 public: 2992 virtual void yield(); 2993 2994 // "n_threads" is the number of threads to be terminated. 2995 // "queue_set" is a set of work queues of other threads. 2996 // "collector" is the CMS collector associated with this task terminator. 2997 // "yield" indicates whether we need the gang as a whole to yield. 2998 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2999 ParallelTaskTerminator(n_threads, queue_set), 3000 _collector(collector) { } 3001 3002 void set_task(CMSConcMarkingTask* task) { 3003 _task = task; 3004 } 3005 }; 3006 3007 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3008 CMSConcMarkingTask* _task; 3009 public: 3010 bool should_exit_termination(); 3011 void set_task(CMSConcMarkingTask* task) { 3012 _task = task; 3013 } 3014 }; 3015 3016 // MT Concurrent Marking Task 3017 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3018 CMSCollector* _collector; 3019 uint _n_workers; // requested/desired # workers 3020 bool _result; 3021 CompactibleFreeListSpace* _cms_space; 3022 char _pad_front[64]; // padding to ... 3023 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3024 char _pad_back[64]; 3025 HeapWord* _restart_addr; 3026 3027 // Exposed here for yielding support 3028 Mutex* const _bit_map_lock; 3029 3030 // The per thread work queues, available here for stealing 3031 OopTaskQueueSet* _task_queues; 3032 3033 // Termination (and yielding) support 3034 CMSConcMarkingTerminator _term; 3035 CMSConcMarkingTerminatorTerminator _term_term; 3036 3037 public: 3038 CMSConcMarkingTask(CMSCollector* collector, 3039 CompactibleFreeListSpace* cms_space, 3040 YieldingFlexibleWorkGang* workers, 3041 OopTaskQueueSet* task_queues): 3042 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3043 _collector(collector), 3044 _cms_space(cms_space), 3045 _n_workers(0), _result(true), 3046 _task_queues(task_queues), 3047 _term(_n_workers, task_queues, _collector), 3048 _bit_map_lock(collector->bitMapLock()) 3049 { 3050 _requested_size = _n_workers; 3051 _term.set_task(this); 3052 _term_term.set_task(this); 3053 _restart_addr = _global_finger = _cms_space->bottom(); 3054 } 3055 3056 3057 OopTaskQueueSet* task_queues() { return _task_queues; } 3058 3059 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3060 3061 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3062 3063 CMSConcMarkingTerminator* terminator() { return &_term; } 3064 3065 virtual void set_for_termination(uint active_workers) { 3066 terminator()->reset_for_reuse(active_workers); 3067 } 3068 3069 void work(uint worker_id); 3070 bool should_yield() { 3071 return ConcurrentMarkSweepThread::should_yield() 3072 && !_collector->foregroundGCIsActive(); 3073 } 3074 3075 virtual void coordinator_yield(); // stuff done by coordinator 3076 bool result() { return _result; } 3077 3078 void reset(HeapWord* ra) { 3079 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3080 _restart_addr = _global_finger = ra; 3081 _term.reset_for_reuse(); 3082 } 3083 3084 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3085 OopTaskQueue* work_q); 3086 3087 private: 3088 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3089 void do_work_steal(int i); 3090 void bump_global_finger(HeapWord* f); 3091 }; 3092 3093 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3094 assert(_task != NULL, "Error"); 3095 return _task->yielding(); 3096 // Note that we do not need the disjunct || _task->should_yield() above 3097 // because we want terminating threads to yield only if the task 3098 // is already in the midst of yielding, which happens only after at least one 3099 // thread has yielded. 3100 } 3101 3102 void CMSConcMarkingTerminator::yield() { 3103 if (_task->should_yield()) { 3104 _task->yield(); 3105 } else { 3106 ParallelTaskTerminator::yield(); 3107 } 3108 } 3109 3110 //////////////////////////////////////////////////////////////// 3111 // Concurrent Marking Algorithm Sketch 3112 //////////////////////////////////////////////////////////////// 3113 // Until all tasks exhausted (both spaces): 3114 // -- claim next available chunk 3115 // -- bump global finger via CAS 3116 // -- find first object that starts in this chunk 3117 // and start scanning bitmap from that position 3118 // -- scan marked objects for oops 3119 // -- CAS-mark target, and if successful: 3120 // . if target oop is above global finger (volatile read) 3121 // nothing to do 3122 // . if target oop is in chunk and above local finger 3123 // then nothing to do 3124 // . else push on work-queue 3125 // -- Deal with possible overflow issues: 3126 // . local work-queue overflow causes stuff to be pushed on 3127 // global (common) overflow queue 3128 // . always first empty local work queue 3129 // . then get a batch of oops from global work queue if any 3130 // . then do work stealing 3131 // -- When all tasks claimed (both spaces) 3132 // and local work queue empty, 3133 // then in a loop do: 3134 // . check global overflow stack; steal a batch of oops and trace 3135 // . try to steal from other threads oif GOS is empty 3136 // . if neither is available, offer termination 3137 // -- Terminate and return result 3138 // 3139 void CMSConcMarkingTask::work(uint worker_id) { 3140 elapsedTimer _timer; 3141 ResourceMark rm; 3142 HandleMark hm; 3143 3144 DEBUG_ONLY(_collector->verify_overflow_empty();) 3145 3146 // Before we begin work, our work queue should be empty 3147 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3148 // Scan the bitmap covering _cms_space, tracing through grey objects. 3149 _timer.start(); 3150 do_scan_and_mark(worker_id, _cms_space); 3151 _timer.stop(); 3152 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3153 3154 // ... do work stealing 3155 _timer.reset(); 3156 _timer.start(); 3157 do_work_steal(worker_id); 3158 _timer.stop(); 3159 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3160 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3161 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3162 // Note that under the current task protocol, the 3163 // following assertion is true even of the spaces 3164 // expanded since the completion of the concurrent 3165 // marking. XXX This will likely change under a strict 3166 // ABORT semantics. 3167 // After perm removal the comparison was changed to 3168 // greater than or equal to from strictly greater than. 3169 // Before perm removal the highest address sweep would 3170 // have been at the end of perm gen but now is at the 3171 // end of the tenured gen. 3172 assert(_global_finger >= _cms_space->end(), 3173 "All tasks have been completed"); 3174 DEBUG_ONLY(_collector->verify_overflow_empty();) 3175 } 3176 3177 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3178 HeapWord* read = _global_finger; 3179 HeapWord* cur = read; 3180 while (f > read) { 3181 cur = read; 3182 read = Atomic::cmpxchg(f, &_global_finger, cur); 3183 if (cur == read) { 3184 // our cas succeeded 3185 assert(_global_finger >= f, "protocol consistency"); 3186 break; 3187 } 3188 } 3189 } 3190 3191 // This is really inefficient, and should be redone by 3192 // using (not yet available) block-read and -write interfaces to the 3193 // stack and the work_queue. XXX FIX ME !!! 3194 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3195 OopTaskQueue* work_q) { 3196 // Fast lock-free check 3197 if (ovflw_stk->length() == 0) { 3198 return false; 3199 } 3200 assert(work_q->size() == 0, "Shouldn't steal"); 3201 MutexLockerEx ml(ovflw_stk->par_lock(), 3202 Mutex::_no_safepoint_check_flag); 3203 // Grab up to 1/4 the size of the work queue 3204 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3205 (size_t)ParGCDesiredObjsFromOverflowList); 3206 num = MIN2(num, ovflw_stk->length()); 3207 for (int i = (int) num; i > 0; i--) { 3208 oop cur = ovflw_stk->pop(); 3209 assert(cur != NULL, "Counted wrong?"); 3210 work_q->push(cur); 3211 } 3212 return num > 0; 3213 } 3214 3215 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3216 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3217 int n_tasks = pst->n_tasks(); 3218 // We allow that there may be no tasks to do here because 3219 // we are restarting after a stack overflow. 3220 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3221 uint nth_task = 0; 3222 3223 HeapWord* aligned_start = sp->bottom(); 3224 if (sp->used_region().contains(_restart_addr)) { 3225 // Align down to a card boundary for the start of 0th task 3226 // for this space. 3227 aligned_start = align_down(_restart_addr, CardTable::card_size); 3228 } 3229 3230 size_t chunk_size = sp->marking_task_size(); 3231 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3232 // Having claimed the nth task in this space, 3233 // compute the chunk that it corresponds to: 3234 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3235 aligned_start + (nth_task+1)*chunk_size); 3236 // Try and bump the global finger via a CAS; 3237 // note that we need to do the global finger bump 3238 // _before_ taking the intersection below, because 3239 // the task corresponding to that region will be 3240 // deemed done even if the used_region() expands 3241 // because of allocation -- as it almost certainly will 3242 // during start-up while the threads yield in the 3243 // closure below. 3244 HeapWord* finger = span.end(); 3245 bump_global_finger(finger); // atomically 3246 // There are null tasks here corresponding to chunks 3247 // beyond the "top" address of the space. 3248 span = span.intersection(sp->used_region()); 3249 if (!span.is_empty()) { // Non-null task 3250 HeapWord* prev_obj; 3251 assert(!span.contains(_restart_addr) || nth_task == 0, 3252 "Inconsistency"); 3253 if (nth_task == 0) { 3254 // For the 0th task, we'll not need to compute a block_start. 3255 if (span.contains(_restart_addr)) { 3256 // In the case of a restart because of stack overflow, 3257 // we might additionally skip a chunk prefix. 3258 prev_obj = _restart_addr; 3259 } else { 3260 prev_obj = span.start(); 3261 } 3262 } else { 3263 // We want to skip the first object because 3264 // the protocol is to scan any object in its entirety 3265 // that _starts_ in this span; a fortiori, any 3266 // object starting in an earlier span is scanned 3267 // as part of an earlier claimed task. 3268 // Below we use the "careful" version of block_start 3269 // so we do not try to navigate uninitialized objects. 3270 prev_obj = sp->block_start_careful(span.start()); 3271 // Below we use a variant of block_size that uses the 3272 // Printezis bits to avoid waiting for allocated 3273 // objects to become initialized/parsable. 3274 while (prev_obj < span.start()) { 3275 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3276 if (sz > 0) { 3277 prev_obj += sz; 3278 } else { 3279 // In this case we may end up doing a bit of redundant 3280 // scanning, but that appears unavoidable, short of 3281 // locking the free list locks; see bug 6324141. 3282 break; 3283 } 3284 } 3285 } 3286 if (prev_obj < span.end()) { 3287 MemRegion my_span = MemRegion(prev_obj, span.end()); 3288 // Do the marking work within a non-empty span -- 3289 // the last argument to the constructor indicates whether the 3290 // iteration should be incremental with periodic yields. 3291 ParMarkFromRootsClosure cl(this, _collector, my_span, 3292 &_collector->_markBitMap, 3293 work_queue(i), 3294 &_collector->_markStack); 3295 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3296 } // else nothing to do for this task 3297 } // else nothing to do for this task 3298 } 3299 // We'd be tempted to assert here that since there are no 3300 // more tasks left to claim in this space, the global_finger 3301 // must exceed space->top() and a fortiori space->end(). However, 3302 // that would not quite be correct because the bumping of 3303 // global_finger occurs strictly after the claiming of a task, 3304 // so by the time we reach here the global finger may not yet 3305 // have been bumped up by the thread that claimed the last 3306 // task. 3307 pst->all_tasks_completed(); 3308 } 3309 3310 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3311 private: 3312 CMSCollector* _collector; 3313 CMSConcMarkingTask* _task; 3314 MemRegion _span; 3315 CMSBitMap* _bit_map; 3316 CMSMarkStack* _overflow_stack; 3317 OopTaskQueue* _work_queue; 3318 protected: 3319 DO_OOP_WORK_DEFN 3320 public: 3321 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3322 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3323 MetadataAwareOopClosure(collector->ref_processor()), 3324 _collector(collector), 3325 _task(task), 3326 _span(collector->_span), 3327 _work_queue(work_queue), 3328 _bit_map(bit_map), 3329 _overflow_stack(overflow_stack) 3330 { } 3331 virtual void do_oop(oop* p); 3332 virtual void do_oop(narrowOop* p); 3333 3334 void trim_queue(size_t max); 3335 void handle_stack_overflow(HeapWord* lost); 3336 void do_yield_check() { 3337 if (_task->should_yield()) { 3338 _task->yield(); 3339 } 3340 } 3341 }; 3342 3343 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3344 3345 // Grey object scanning during work stealing phase -- 3346 // the salient assumption here is that any references 3347 // that are in these stolen objects being scanned must 3348 // already have been initialized (else they would not have 3349 // been published), so we do not need to check for 3350 // uninitialized objects before pushing here. 3351 void ParConcMarkingClosure::do_oop(oop obj) { 3352 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3353 HeapWord* addr = (HeapWord*)obj; 3354 // Check if oop points into the CMS generation 3355 // and is not marked 3356 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3357 // a white object ... 3358 // If we manage to "claim" the object, by being the 3359 // first thread to mark it, then we push it on our 3360 // marking stack 3361 if (_bit_map->par_mark(addr)) { // ... now grey 3362 // push on work queue (grey set) 3363 bool simulate_overflow = false; 3364 NOT_PRODUCT( 3365 if (CMSMarkStackOverflowALot && 3366 _collector->simulate_overflow()) { 3367 // simulate a stack overflow 3368 simulate_overflow = true; 3369 } 3370 ) 3371 if (simulate_overflow || 3372 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3373 // stack overflow 3374 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3375 // We cannot assert that the overflow stack is full because 3376 // it may have been emptied since. 3377 assert(simulate_overflow || 3378 _work_queue->size() == _work_queue->max_elems(), 3379 "Else push should have succeeded"); 3380 handle_stack_overflow(addr); 3381 } 3382 } // Else, some other thread got there first 3383 do_yield_check(); 3384 } 3385 } 3386 3387 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3388 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3389 3390 void ParConcMarkingClosure::trim_queue(size_t max) { 3391 while (_work_queue->size() > max) { 3392 oop new_oop; 3393 if (_work_queue->pop_local(new_oop)) { 3394 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3395 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3396 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3397 new_oop->oop_iterate(this); // do_oop() above 3398 do_yield_check(); 3399 } 3400 } 3401 } 3402 3403 // Upon stack overflow, we discard (part of) the stack, 3404 // remembering the least address amongst those discarded 3405 // in CMSCollector's _restart_address. 3406 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3407 // We need to do this under a mutex to prevent other 3408 // workers from interfering with the work done below. 3409 MutexLockerEx ml(_overflow_stack->par_lock(), 3410 Mutex::_no_safepoint_check_flag); 3411 // Remember the least grey address discarded 3412 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3413 _collector->lower_restart_addr(ra); 3414 _overflow_stack->reset(); // discard stack contents 3415 _overflow_stack->expand(); // expand the stack if possible 3416 } 3417 3418 3419 void CMSConcMarkingTask::do_work_steal(int i) { 3420 OopTaskQueue* work_q = work_queue(i); 3421 oop obj_to_scan; 3422 CMSBitMap* bm = &(_collector->_markBitMap); 3423 CMSMarkStack* ovflw = &(_collector->_markStack); 3424 int* seed = _collector->hash_seed(i); 3425 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3426 while (true) { 3427 cl.trim_queue(0); 3428 assert(work_q->size() == 0, "Should have been emptied above"); 3429 if (get_work_from_overflow_stack(ovflw, work_q)) { 3430 // Can't assert below because the work obtained from the 3431 // overflow stack may already have been stolen from us. 3432 // assert(work_q->size() > 0, "Work from overflow stack"); 3433 continue; 3434 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3435 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3436 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3437 obj_to_scan->oop_iterate(&cl); 3438 } else if (terminator()->offer_termination(&_term_term)) { 3439 assert(work_q->size() == 0, "Impossible!"); 3440 break; 3441 } else if (yielding() || should_yield()) { 3442 yield(); 3443 } 3444 } 3445 } 3446 3447 // This is run by the CMS (coordinator) thread. 3448 void CMSConcMarkingTask::coordinator_yield() { 3449 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3450 "CMS thread should hold CMS token"); 3451 // First give up the locks, then yield, then re-lock 3452 // We should probably use a constructor/destructor idiom to 3453 // do this unlock/lock or modify the MutexUnlocker class to 3454 // serve our purpose. XXX 3455 assert_lock_strong(_bit_map_lock); 3456 _bit_map_lock->unlock(); 3457 ConcurrentMarkSweepThread::desynchronize(true); 3458 _collector->stopTimer(); 3459 _collector->incrementYields(); 3460 3461 // It is possible for whichever thread initiated the yield request 3462 // not to get a chance to wake up and take the bitmap lock between 3463 // this thread releasing it and reacquiring it. So, while the 3464 // should_yield() flag is on, let's sleep for a bit to give the 3465 // other thread a chance to wake up. The limit imposed on the number 3466 // of iterations is defensive, to avoid any unforseen circumstances 3467 // putting us into an infinite loop. Since it's always been this 3468 // (coordinator_yield()) method that was observed to cause the 3469 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3470 // which is by default non-zero. For the other seven methods that 3471 // also perform the yield operation, as are using a different 3472 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3473 // can enable the sleeping for those methods too, if necessary. 3474 // See 6442774. 3475 // 3476 // We really need to reconsider the synchronization between the GC 3477 // thread and the yield-requesting threads in the future and we 3478 // should really use wait/notify, which is the recommended 3479 // way of doing this type of interaction. Additionally, we should 3480 // consolidate the eight methods that do the yield operation and they 3481 // are almost identical into one for better maintainability and 3482 // readability. See 6445193. 3483 // 3484 // Tony 2006.06.29 3485 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3486 ConcurrentMarkSweepThread::should_yield() && 3487 !CMSCollector::foregroundGCIsActive(); ++i) { 3488 os::sleep(Thread::current(), 1, false); 3489 } 3490 3491 ConcurrentMarkSweepThread::synchronize(true); 3492 _bit_map_lock->lock_without_safepoint_check(); 3493 _collector->startTimer(); 3494 } 3495 3496 bool CMSCollector::do_marking_mt() { 3497 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3498 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3499 conc_workers()->active_workers(), 3500 Threads::number_of_non_daemon_threads()); 3501 num_workers = conc_workers()->update_active_workers(num_workers); 3502 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3503 3504 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3505 3506 CMSConcMarkingTask tsk(this, 3507 cms_space, 3508 conc_workers(), 3509 task_queues()); 3510 3511 // Since the actual number of workers we get may be different 3512 // from the number we requested above, do we need to do anything different 3513 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3514 // class?? XXX 3515 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3516 3517 // Refs discovery is already non-atomic. 3518 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3519 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3520 conc_workers()->start_task(&tsk); 3521 while (tsk.yielded()) { 3522 tsk.coordinator_yield(); 3523 conc_workers()->continue_task(&tsk); 3524 } 3525 // If the task was aborted, _restart_addr will be non-NULL 3526 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3527 while (_restart_addr != NULL) { 3528 // XXX For now we do not make use of ABORTED state and have not 3529 // yet implemented the right abort semantics (even in the original 3530 // single-threaded CMS case). That needs some more investigation 3531 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3532 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3533 // If _restart_addr is non-NULL, a marking stack overflow 3534 // occurred; we need to do a fresh marking iteration from the 3535 // indicated restart address. 3536 if (_foregroundGCIsActive) { 3537 // We may be running into repeated stack overflows, having 3538 // reached the limit of the stack size, while making very 3539 // slow forward progress. It may be best to bail out and 3540 // let the foreground collector do its job. 3541 // Clear _restart_addr, so that foreground GC 3542 // works from scratch. This avoids the headache of 3543 // a "rescan" which would otherwise be needed because 3544 // of the dirty mod union table & card table. 3545 _restart_addr = NULL; 3546 return false; 3547 } 3548 // Adjust the task to restart from _restart_addr 3549 tsk.reset(_restart_addr); 3550 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3551 _restart_addr); 3552 _restart_addr = NULL; 3553 // Get the workers going again 3554 conc_workers()->start_task(&tsk); 3555 while (tsk.yielded()) { 3556 tsk.coordinator_yield(); 3557 conc_workers()->continue_task(&tsk); 3558 } 3559 } 3560 assert(tsk.completed(), "Inconsistency"); 3561 assert(tsk.result() == true, "Inconsistency"); 3562 return true; 3563 } 3564 3565 bool CMSCollector::do_marking_st() { 3566 ResourceMark rm; 3567 HandleMark hm; 3568 3569 // Temporarily make refs discovery single threaded (non-MT) 3570 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3571 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3572 &_markStack, CMSYield); 3573 // the last argument to iterate indicates whether the iteration 3574 // should be incremental with periodic yields. 3575 _markBitMap.iterate(&markFromRootsClosure); 3576 // If _restart_addr is non-NULL, a marking stack overflow 3577 // occurred; we need to do a fresh iteration from the 3578 // indicated restart address. 3579 while (_restart_addr != NULL) { 3580 if (_foregroundGCIsActive) { 3581 // We may be running into repeated stack overflows, having 3582 // reached the limit of the stack size, while making very 3583 // slow forward progress. It may be best to bail out and 3584 // let the foreground collector do its job. 3585 // Clear _restart_addr, so that foreground GC 3586 // works from scratch. This avoids the headache of 3587 // a "rescan" which would otherwise be needed because 3588 // of the dirty mod union table & card table. 3589 _restart_addr = NULL; 3590 return false; // indicating failure to complete marking 3591 } 3592 // Deal with stack overflow: 3593 // we restart marking from _restart_addr 3594 HeapWord* ra = _restart_addr; 3595 markFromRootsClosure.reset(ra); 3596 _restart_addr = NULL; 3597 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3598 } 3599 return true; 3600 } 3601 3602 void CMSCollector::preclean() { 3603 check_correct_thread_executing(); 3604 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3605 verify_work_stacks_empty(); 3606 verify_overflow_empty(); 3607 _abort_preclean = false; 3608 if (CMSPrecleaningEnabled) { 3609 if (!CMSEdenChunksRecordAlways) { 3610 _eden_chunk_index = 0; 3611 } 3612 size_t used = get_eden_used(); 3613 size_t capacity = get_eden_capacity(); 3614 // Don't start sampling unless we will get sufficiently 3615 // many samples. 3616 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3617 * CMSScheduleRemarkEdenPenetration)) { 3618 _start_sampling = true; 3619 } else { 3620 _start_sampling = false; 3621 } 3622 GCTraceCPUTime tcpu; 3623 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3624 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3625 } 3626 CMSTokenSync x(true); // is cms thread 3627 if (CMSPrecleaningEnabled) { 3628 sample_eden(); 3629 _collectorState = AbortablePreclean; 3630 } else { 3631 _collectorState = FinalMarking; 3632 } 3633 verify_work_stacks_empty(); 3634 verify_overflow_empty(); 3635 } 3636 3637 // Try and schedule the remark such that young gen 3638 // occupancy is CMSScheduleRemarkEdenPenetration %. 3639 void CMSCollector::abortable_preclean() { 3640 check_correct_thread_executing(); 3641 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3642 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3643 3644 // If Eden's current occupancy is below this threshold, 3645 // immediately schedule the remark; else preclean 3646 // past the next scavenge in an effort to 3647 // schedule the pause as described above. By choosing 3648 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3649 // we will never do an actual abortable preclean cycle. 3650 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3651 GCTraceCPUTime tcpu; 3652 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3653 // We need more smarts in the abortable preclean 3654 // loop below to deal with cases where allocation 3655 // in young gen is very very slow, and our precleaning 3656 // is running a losing race against a horde of 3657 // mutators intent on flooding us with CMS updates 3658 // (dirty cards). 3659 // One, admittedly dumb, strategy is to give up 3660 // after a certain number of abortable precleaning loops 3661 // or after a certain maximum time. We want to make 3662 // this smarter in the next iteration. 3663 // XXX FIX ME!!! YSR 3664 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3665 while (!(should_abort_preclean() || 3666 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3667 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3668 cumworkdone += workdone; 3669 loops++; 3670 // Voluntarily terminate abortable preclean phase if we have 3671 // been at it for too long. 3672 if ((CMSMaxAbortablePrecleanLoops != 0) && 3673 loops >= CMSMaxAbortablePrecleanLoops) { 3674 log_debug(gc)(" CMS: abort preclean due to loops "); 3675 break; 3676 } 3677 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3678 log_debug(gc)(" CMS: abort preclean due to time "); 3679 break; 3680 } 3681 // If we are doing little work each iteration, we should 3682 // take a short break. 3683 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3684 // Sleep for some time, waiting for work to accumulate 3685 stopTimer(); 3686 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3687 startTimer(); 3688 waited++; 3689 } 3690 } 3691 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3692 loops, waited, cumworkdone); 3693 } 3694 CMSTokenSync x(true); // is cms thread 3695 if (_collectorState != Idling) { 3696 assert(_collectorState == AbortablePreclean, 3697 "Spontaneous state transition?"); 3698 _collectorState = FinalMarking; 3699 } // Else, a foreground collection completed this CMS cycle. 3700 return; 3701 } 3702 3703 // Respond to an Eden sampling opportunity 3704 void CMSCollector::sample_eden() { 3705 // Make sure a young gc cannot sneak in between our 3706 // reading and recording of a sample. 3707 assert(Thread::current()->is_ConcurrentGC_thread(), 3708 "Only the cms thread may collect Eden samples"); 3709 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3710 "Should collect samples while holding CMS token"); 3711 if (!_start_sampling) { 3712 return; 3713 } 3714 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3715 // is populated by the young generation. 3716 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3717 if (_eden_chunk_index < _eden_chunk_capacity) { 3718 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3719 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3720 "Unexpected state of Eden"); 3721 // We'd like to check that what we just sampled is an oop-start address; 3722 // however, we cannot do that here since the object may not yet have been 3723 // initialized. So we'll instead do the check when we _use_ this sample 3724 // later. 3725 if (_eden_chunk_index == 0 || 3726 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3727 _eden_chunk_array[_eden_chunk_index-1]) 3728 >= CMSSamplingGrain)) { 3729 _eden_chunk_index++; // commit sample 3730 } 3731 } 3732 } 3733 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3734 size_t used = get_eden_used(); 3735 size_t capacity = get_eden_capacity(); 3736 assert(used <= capacity, "Unexpected state of Eden"); 3737 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3738 _abort_preclean = true; 3739 } 3740 } 3741 } 3742 3743 3744 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3745 assert(_collectorState == Precleaning || 3746 _collectorState == AbortablePreclean, "incorrect state"); 3747 ResourceMark rm; 3748 HandleMark hm; 3749 3750 // Precleaning is currently not MT but the reference processor 3751 // may be set for MT. Disable it temporarily here. 3752 ReferenceProcessor* rp = ref_processor(); 3753 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3754 3755 // Do one pass of scrubbing the discovered reference lists 3756 // to remove any reference objects with strongly-reachable 3757 // referents. 3758 if (clean_refs) { 3759 CMSPrecleanRefsYieldClosure yield_cl(this); 3760 assert(rp->span().equals(_span), "Spans should be equal"); 3761 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3762 &_markStack, true /* preclean */); 3763 CMSDrainMarkingStackClosure complete_trace(this, 3764 _span, &_markBitMap, &_markStack, 3765 &keep_alive, true /* preclean */); 3766 3767 // We don't want this step to interfere with a young 3768 // collection because we don't want to take CPU 3769 // or memory bandwidth away from the young GC threads 3770 // (which may be as many as there are CPUs). 3771 // Note that we don't need to protect ourselves from 3772 // interference with mutators because they can't 3773 // manipulate the discovered reference lists nor affect 3774 // the computed reachability of the referents, the 3775 // only properties manipulated by the precleaning 3776 // of these reference lists. 3777 stopTimer(); 3778 CMSTokenSyncWithLocks x(true /* is cms thread */, 3779 bitMapLock()); 3780 startTimer(); 3781 sample_eden(); 3782 3783 // The following will yield to allow foreground 3784 // collection to proceed promptly. XXX YSR: 3785 // The code in this method may need further 3786 // tweaking for better performance and some restructuring 3787 // for cleaner interfaces. 3788 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3789 rp->preclean_discovered_references( 3790 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3791 gc_timer); 3792 } 3793 3794 if (clean_survivor) { // preclean the active survivor space(s) 3795 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3796 &_markBitMap, &_modUnionTable, 3797 &_markStack, true /* precleaning phase */); 3798 stopTimer(); 3799 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3800 bitMapLock()); 3801 startTimer(); 3802 unsigned int before_count = 3803 CMSHeap::heap()->total_collections(); 3804 SurvivorSpacePrecleanClosure 3805 sss_cl(this, _span, &_markBitMap, &_markStack, 3806 &pam_cl, before_count, CMSYield); 3807 _young_gen->from()->object_iterate_careful(&sss_cl); 3808 _young_gen->to()->object_iterate_careful(&sss_cl); 3809 } 3810 MarkRefsIntoAndScanClosure 3811 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3812 &_markStack, this, CMSYield, 3813 true /* precleaning phase */); 3814 // CAUTION: The following closure has persistent state that may need to 3815 // be reset upon a decrease in the sequence of addresses it 3816 // processes. 3817 ScanMarkedObjectsAgainCarefullyClosure 3818 smoac_cl(this, _span, 3819 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3820 3821 // Preclean dirty cards in ModUnionTable and CardTable using 3822 // appropriate convergence criterion; 3823 // repeat CMSPrecleanIter times unless we find that 3824 // we are losing. 3825 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3826 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3827 "Bad convergence multiplier"); 3828 assert(CMSPrecleanThreshold >= 100, 3829 "Unreasonably low CMSPrecleanThreshold"); 3830 3831 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3832 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3833 numIter < CMSPrecleanIter; 3834 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3835 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3836 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3837 // Either there are very few dirty cards, so re-mark 3838 // pause will be small anyway, or our pre-cleaning isn't 3839 // that much faster than the rate at which cards are being 3840 // dirtied, so we might as well stop and re-mark since 3841 // precleaning won't improve our re-mark time by much. 3842 if (curNumCards <= CMSPrecleanThreshold || 3843 (numIter > 0 && 3844 (curNumCards * CMSPrecleanDenominator > 3845 lastNumCards * CMSPrecleanNumerator))) { 3846 numIter++; 3847 cumNumCards += curNumCards; 3848 break; 3849 } 3850 } 3851 3852 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3853 3854 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3855 cumNumCards += curNumCards; 3856 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3857 curNumCards, cumNumCards, numIter); 3858 return cumNumCards; // as a measure of useful work done 3859 } 3860 3861 // PRECLEANING NOTES: 3862 // Precleaning involves: 3863 // . reading the bits of the modUnionTable and clearing the set bits. 3864 // . For the cards corresponding to the set bits, we scan the 3865 // objects on those cards. This means we need the free_list_lock 3866 // so that we can safely iterate over the CMS space when scanning 3867 // for oops. 3868 // . When we scan the objects, we'll be both reading and setting 3869 // marks in the marking bit map, so we'll need the marking bit map. 3870 // . For protecting _collector_state transitions, we take the CGC_lock. 3871 // Note that any races in the reading of of card table entries by the 3872 // CMS thread on the one hand and the clearing of those entries by the 3873 // VM thread or the setting of those entries by the mutator threads on the 3874 // other are quite benign. However, for efficiency it makes sense to keep 3875 // the VM thread from racing with the CMS thread while the latter is 3876 // dirty card info to the modUnionTable. We therefore also use the 3877 // CGC_lock to protect the reading of the card table and the mod union 3878 // table by the CM thread. 3879 // . We run concurrently with mutator updates, so scanning 3880 // needs to be done carefully -- we should not try to scan 3881 // potentially uninitialized objects. 3882 // 3883 // Locking strategy: While holding the CGC_lock, we scan over and 3884 // reset a maximal dirty range of the mod union / card tables, then lock 3885 // the free_list_lock and bitmap lock to do a full marking, then 3886 // release these locks; and repeat the cycle. This allows for a 3887 // certain amount of fairness in the sharing of these locks between 3888 // the CMS collector on the one hand, and the VM thread and the 3889 // mutators on the other. 3890 3891 // NOTE: preclean_mod_union_table() and preclean_card_table() 3892 // further below are largely identical; if you need to modify 3893 // one of these methods, please check the other method too. 3894 3895 size_t CMSCollector::preclean_mod_union_table( 3896 ConcurrentMarkSweepGeneration* old_gen, 3897 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3898 verify_work_stacks_empty(); 3899 verify_overflow_empty(); 3900 3901 // strategy: starting with the first card, accumulate contiguous 3902 // ranges of dirty cards; clear these cards, then scan the region 3903 // covered by these cards. 3904 3905 // Since all of the MUT is committed ahead, we can just use 3906 // that, in case the generations expand while we are precleaning. 3907 // It might also be fine to just use the committed part of the 3908 // generation, but we might potentially miss cards when the 3909 // generation is rapidly expanding while we are in the midst 3910 // of precleaning. 3911 HeapWord* startAddr = old_gen->reserved().start(); 3912 HeapWord* endAddr = old_gen->reserved().end(); 3913 3914 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3915 3916 size_t numDirtyCards, cumNumDirtyCards; 3917 HeapWord *nextAddr, *lastAddr; 3918 for (cumNumDirtyCards = numDirtyCards = 0, 3919 nextAddr = lastAddr = startAddr; 3920 nextAddr < endAddr; 3921 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3922 3923 ResourceMark rm; 3924 HandleMark hm; 3925 3926 MemRegion dirtyRegion; 3927 { 3928 stopTimer(); 3929 // Potential yield point 3930 CMSTokenSync ts(true); 3931 startTimer(); 3932 sample_eden(); 3933 // Get dirty region starting at nextOffset (inclusive), 3934 // simultaneously clearing it. 3935 dirtyRegion = 3936 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3937 assert(dirtyRegion.start() >= nextAddr, 3938 "returned region inconsistent?"); 3939 } 3940 // Remember where the next search should begin. 3941 // The returned region (if non-empty) is a right open interval, 3942 // so lastOffset is obtained from the right end of that 3943 // interval. 3944 lastAddr = dirtyRegion.end(); 3945 // Should do something more transparent and less hacky XXX 3946 numDirtyCards = 3947 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3948 3949 // We'll scan the cards in the dirty region (with periodic 3950 // yields for foreground GC as needed). 3951 if (!dirtyRegion.is_empty()) { 3952 assert(numDirtyCards > 0, "consistency check"); 3953 HeapWord* stop_point = NULL; 3954 stopTimer(); 3955 // Potential yield point 3956 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3957 bitMapLock()); 3958 startTimer(); 3959 { 3960 verify_work_stacks_empty(); 3961 verify_overflow_empty(); 3962 sample_eden(); 3963 stop_point = 3964 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3965 } 3966 if (stop_point != NULL) { 3967 // The careful iteration stopped early either because it found an 3968 // uninitialized object, or because we were in the midst of an 3969 // "abortable preclean", which should now be aborted. Redirty 3970 // the bits corresponding to the partially-scanned or unscanned 3971 // cards. We'll either restart at the next block boundary or 3972 // abort the preclean. 3973 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3974 "Should only be AbortablePreclean."); 3975 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3976 if (should_abort_preclean()) { 3977 break; // out of preclean loop 3978 } else { 3979 // Compute the next address at which preclean should pick up; 3980 // might need bitMapLock in order to read P-bits. 3981 lastAddr = next_card_start_after_block(stop_point); 3982 } 3983 } 3984 } else { 3985 assert(lastAddr == endAddr, "consistency check"); 3986 assert(numDirtyCards == 0, "consistency check"); 3987 break; 3988 } 3989 } 3990 verify_work_stacks_empty(); 3991 verify_overflow_empty(); 3992 return cumNumDirtyCards; 3993 } 3994 3995 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3996 // below are largely identical; if you need to modify 3997 // one of these methods, please check the other method too. 3998 3999 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4000 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4001 // strategy: it's similar to precleamModUnionTable above, in that 4002 // we accumulate contiguous ranges of dirty cards, mark these cards 4003 // precleaned, then scan the region covered by these cards. 4004 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4005 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4006 4007 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4008 4009 size_t numDirtyCards, cumNumDirtyCards; 4010 HeapWord *lastAddr, *nextAddr; 4011 4012 for (cumNumDirtyCards = numDirtyCards = 0, 4013 nextAddr = lastAddr = startAddr; 4014 nextAddr < endAddr; 4015 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4016 4017 ResourceMark rm; 4018 HandleMark hm; 4019 4020 MemRegion dirtyRegion; 4021 { 4022 // See comments in "Precleaning notes" above on why we 4023 // do this locking. XXX Could the locking overheads be 4024 // too high when dirty cards are sparse? [I don't think so.] 4025 stopTimer(); 4026 CMSTokenSync x(true); // is cms thread 4027 startTimer(); 4028 sample_eden(); 4029 // Get and clear dirty region from card table 4030 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4031 true, 4032 CardTable::precleaned_card_val()); 4033 4034 assert(dirtyRegion.start() >= nextAddr, 4035 "returned region inconsistent?"); 4036 } 4037 lastAddr = dirtyRegion.end(); 4038 numDirtyCards = 4039 dirtyRegion.word_size()/CardTable::card_size_in_words; 4040 4041 if (!dirtyRegion.is_empty()) { 4042 stopTimer(); 4043 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4044 startTimer(); 4045 sample_eden(); 4046 verify_work_stacks_empty(); 4047 verify_overflow_empty(); 4048 HeapWord* stop_point = 4049 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4050 if (stop_point != NULL) { 4051 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4052 "Should only be AbortablePreclean."); 4053 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4054 if (should_abort_preclean()) { 4055 break; // out of preclean loop 4056 } else { 4057 // Compute the next address at which preclean should pick up. 4058 lastAddr = next_card_start_after_block(stop_point); 4059 } 4060 } 4061 } else { 4062 break; 4063 } 4064 } 4065 verify_work_stacks_empty(); 4066 verify_overflow_empty(); 4067 return cumNumDirtyCards; 4068 } 4069 4070 class PrecleanCLDClosure : public CLDClosure { 4071 MetadataAwareOopsInGenClosure* _cm_closure; 4072 public: 4073 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4074 void do_cld(ClassLoaderData* cld) { 4075 if (cld->has_accumulated_modified_oops()) { 4076 cld->clear_accumulated_modified_oops(); 4077 4078 _cm_closure->do_cld(cld); 4079 } 4080 } 4081 }; 4082 4083 // The freelist lock is needed to prevent asserts, is it really needed? 4084 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4085 4086 cl->set_freelistLock(freelistLock); 4087 4088 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4089 4090 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4091 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4092 PrecleanCLDClosure preclean_closure(cl); 4093 ClassLoaderDataGraph::cld_do(&preclean_closure); 4094 4095 verify_work_stacks_empty(); 4096 verify_overflow_empty(); 4097 } 4098 4099 void CMSCollector::checkpointRootsFinal() { 4100 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4101 check_correct_thread_executing(); 4102 // world is stopped at this checkpoint 4103 assert(SafepointSynchronize::is_at_safepoint(), 4104 "world should be stopped"); 4105 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4106 4107 verify_work_stacks_empty(); 4108 verify_overflow_empty(); 4109 4110 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4111 _young_gen->used() / K, _young_gen->capacity() / K); 4112 { 4113 if (CMSScavengeBeforeRemark) { 4114 CMSHeap* heap = CMSHeap::heap(); 4115 // Temporarily set flag to false, GCH->do_collection will 4116 // expect it to be false and set to true 4117 FlagSetting fl(heap->_is_gc_active, false); 4118 4119 heap->do_collection(true, // full (i.e. force, see below) 4120 false, // !clear_all_soft_refs 4121 0, // size 4122 false, // is_tlab 4123 GenCollectedHeap::YoungGen // type 4124 ); 4125 } 4126 FreelistLocker x(this); 4127 MutexLockerEx y(bitMapLock(), 4128 Mutex::_no_safepoint_check_flag); 4129 checkpointRootsFinalWork(); 4130 } 4131 verify_work_stacks_empty(); 4132 verify_overflow_empty(); 4133 } 4134 4135 void CMSCollector::checkpointRootsFinalWork() { 4136 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4137 4138 assert(haveFreelistLocks(), "must have free list locks"); 4139 assert_lock_strong(bitMapLock()); 4140 4141 ResourceMark rm; 4142 HandleMark hm; 4143 4144 CMSHeap* heap = CMSHeap::heap(); 4145 4146 if (should_unload_classes()) { 4147 CodeCache::gc_prologue(); 4148 } 4149 assert(haveFreelistLocks(), "must have free list locks"); 4150 assert_lock_strong(bitMapLock()); 4151 4152 // We might assume that we need not fill TLAB's when 4153 // CMSScavengeBeforeRemark is set, because we may have just done 4154 // a scavenge which would have filled all TLAB's -- and besides 4155 // Eden would be empty. This however may not always be the case -- 4156 // for instance although we asked for a scavenge, it may not have 4157 // happened because of a JNI critical section. We probably need 4158 // a policy for deciding whether we can in that case wait until 4159 // the critical section releases and then do the remark following 4160 // the scavenge, and skip it here. In the absence of that policy, 4161 // or of an indication of whether the scavenge did indeed occur, 4162 // we cannot rely on TLAB's having been filled and must do 4163 // so here just in case a scavenge did not happen. 4164 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4165 // Update the saved marks which may affect the root scans. 4166 heap->save_marks(); 4167 4168 print_eden_and_survivor_chunk_arrays(); 4169 4170 { 4171 #if COMPILER2_OR_JVMCI 4172 DerivedPointerTableDeactivate dpt_deact; 4173 #endif 4174 4175 // Note on the role of the mod union table: 4176 // Since the marker in "markFromRoots" marks concurrently with 4177 // mutators, it is possible for some reachable objects not to have been 4178 // scanned. For instance, an only reference to an object A was 4179 // placed in object B after the marker scanned B. Unless B is rescanned, 4180 // A would be collected. Such updates to references in marked objects 4181 // are detected via the mod union table which is the set of all cards 4182 // dirtied since the first checkpoint in this GC cycle and prior to 4183 // the most recent young generation GC, minus those cleaned up by the 4184 // concurrent precleaning. 4185 if (CMSParallelRemarkEnabled) { 4186 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4187 do_remark_parallel(); 4188 } else { 4189 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4190 do_remark_non_parallel(); 4191 } 4192 } 4193 verify_work_stacks_empty(); 4194 verify_overflow_empty(); 4195 4196 { 4197 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4198 refProcessingWork(); 4199 } 4200 verify_work_stacks_empty(); 4201 verify_overflow_empty(); 4202 4203 if (should_unload_classes()) { 4204 CodeCache::gc_epilogue(); 4205 } 4206 JvmtiExport::gc_epilogue(); 4207 4208 // If we encountered any (marking stack / work queue) overflow 4209 // events during the current CMS cycle, take appropriate 4210 // remedial measures, where possible, so as to try and avoid 4211 // recurrence of that condition. 4212 assert(_markStack.isEmpty(), "No grey objects"); 4213 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4214 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4215 if (ser_ovflw > 0) { 4216 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4217 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4218 _markStack.expand(); 4219 _ser_pmc_remark_ovflw = 0; 4220 _ser_pmc_preclean_ovflw = 0; 4221 _ser_kac_preclean_ovflw = 0; 4222 _ser_kac_ovflw = 0; 4223 } 4224 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4225 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4226 _par_pmc_remark_ovflw, _par_kac_ovflw); 4227 _par_pmc_remark_ovflw = 0; 4228 _par_kac_ovflw = 0; 4229 } 4230 if (_markStack._hit_limit > 0) { 4231 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4232 _markStack._hit_limit); 4233 } 4234 if (_markStack._failed_double > 0) { 4235 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4236 _markStack._failed_double, _markStack.capacity()); 4237 } 4238 _markStack._hit_limit = 0; 4239 _markStack._failed_double = 0; 4240 4241 if ((VerifyAfterGC || VerifyDuringGC) && 4242 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4243 verify_after_remark(); 4244 } 4245 4246 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4247 4248 // Change under the freelistLocks. 4249 _collectorState = Sweeping; 4250 // Call isAllClear() under bitMapLock 4251 assert(_modUnionTable.isAllClear(), 4252 "Should be clear by end of the final marking"); 4253 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4254 "Should be clear by end of the final marking"); 4255 } 4256 4257 void CMSParInitialMarkTask::work(uint worker_id) { 4258 elapsedTimer _timer; 4259 ResourceMark rm; 4260 HandleMark hm; 4261 4262 // ---------- scan from roots -------------- 4263 _timer.start(); 4264 CMSHeap* heap = CMSHeap::heap(); 4265 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4266 4267 // ---------- young gen roots -------------- 4268 { 4269 work_on_young_gen_roots(&par_mri_cl); 4270 _timer.stop(); 4271 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4272 } 4273 4274 // ---------- remaining roots -------------- 4275 _timer.reset(); 4276 _timer.start(); 4277 4278 CLDToOopClosure cld_closure(&par_mri_cl, true); 4279 4280 heap->cms_process_roots(_strong_roots_scope, 4281 false, // yg was scanned above 4282 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4283 _collector->should_unload_classes(), 4284 &par_mri_cl, 4285 &cld_closure); 4286 assert(_collector->should_unload_classes() 4287 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4288 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4289 _timer.stop(); 4290 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4291 } 4292 4293 // Parallel remark task 4294 class CMSParRemarkTask: public CMSParMarkTask { 4295 CompactibleFreeListSpace* _cms_space; 4296 4297 // The per-thread work queues, available here for stealing. 4298 OopTaskQueueSet* _task_queues; 4299 ParallelTaskTerminator _term; 4300 StrongRootsScope* _strong_roots_scope; 4301 4302 public: 4303 // A value of 0 passed to n_workers will cause the number of 4304 // workers to be taken from the active workers in the work gang. 4305 CMSParRemarkTask(CMSCollector* collector, 4306 CompactibleFreeListSpace* cms_space, 4307 uint n_workers, WorkGang* workers, 4308 OopTaskQueueSet* task_queues, 4309 StrongRootsScope* strong_roots_scope): 4310 CMSParMarkTask("Rescan roots and grey objects in parallel", 4311 collector, n_workers), 4312 _cms_space(cms_space), 4313 _task_queues(task_queues), 4314 _term(n_workers, task_queues), 4315 _strong_roots_scope(strong_roots_scope) { } 4316 4317 OopTaskQueueSet* task_queues() { return _task_queues; } 4318 4319 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4320 4321 ParallelTaskTerminator* terminator() { return &_term; } 4322 uint n_workers() { return _n_workers; } 4323 4324 void work(uint worker_id); 4325 4326 private: 4327 // ... of dirty cards in old space 4328 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4329 ParMarkRefsIntoAndScanClosure* cl); 4330 4331 // ... work stealing for the above 4332 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4333 }; 4334 4335 class RemarkCLDClosure : public CLDClosure { 4336 CLDToOopClosure _cm_closure; 4337 public: 4338 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4339 void do_cld(ClassLoaderData* cld) { 4340 // Check if we have modified any oops in the CLD during the concurrent marking. 4341 if (cld->has_accumulated_modified_oops()) { 4342 cld->clear_accumulated_modified_oops(); 4343 4344 // We could have transfered the current modified marks to the accumulated marks, 4345 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4346 } else if (cld->has_modified_oops()) { 4347 // Don't clear anything, this info is needed by the next young collection. 4348 } else { 4349 // No modified oops in the ClassLoaderData. 4350 return; 4351 } 4352 4353 // The klass has modified fields, need to scan the klass. 4354 _cm_closure.do_cld(cld); 4355 } 4356 }; 4357 4358 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4359 ParNewGeneration* young_gen = _collector->_young_gen; 4360 ContiguousSpace* eden_space = young_gen->eden(); 4361 ContiguousSpace* from_space = young_gen->from(); 4362 ContiguousSpace* to_space = young_gen->to(); 4363 4364 HeapWord** eca = _collector->_eden_chunk_array; 4365 size_t ect = _collector->_eden_chunk_index; 4366 HeapWord** sca = _collector->_survivor_chunk_array; 4367 size_t sct = _collector->_survivor_chunk_index; 4368 4369 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4370 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4371 4372 do_young_space_rescan(cl, to_space, NULL, 0); 4373 do_young_space_rescan(cl, from_space, sca, sct); 4374 do_young_space_rescan(cl, eden_space, eca, ect); 4375 } 4376 4377 // work_queue(i) is passed to the closure 4378 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4379 // also is passed to do_dirty_card_rescan_tasks() and to 4380 // do_work_steal() to select the i-th task_queue. 4381 4382 void CMSParRemarkTask::work(uint worker_id) { 4383 elapsedTimer _timer; 4384 ResourceMark rm; 4385 HandleMark hm; 4386 4387 // ---------- rescan from roots -------------- 4388 _timer.start(); 4389 CMSHeap* heap = CMSHeap::heap(); 4390 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4391 _collector->_span, _collector->ref_processor(), 4392 &(_collector->_markBitMap), 4393 work_queue(worker_id)); 4394 4395 // Rescan young gen roots first since these are likely 4396 // coarsely partitioned and may, on that account, constitute 4397 // the critical path; thus, it's best to start off that 4398 // work first. 4399 // ---------- young gen roots -------------- 4400 { 4401 work_on_young_gen_roots(&par_mrias_cl); 4402 _timer.stop(); 4403 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4404 } 4405 4406 // ---------- remaining roots -------------- 4407 _timer.reset(); 4408 _timer.start(); 4409 heap->cms_process_roots(_strong_roots_scope, 4410 false, // yg was scanned above 4411 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4412 _collector->should_unload_classes(), 4413 &par_mrias_cl, 4414 NULL); // The dirty klasses will be handled below 4415 4416 assert(_collector->should_unload_classes() 4417 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4418 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4419 _timer.stop(); 4420 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4421 4422 // ---------- unhandled CLD scanning ---------- 4423 if (worker_id == 0) { // Single threaded at the moment. 4424 _timer.reset(); 4425 _timer.start(); 4426 4427 // Scan all new class loader data objects and new dependencies that were 4428 // introduced during concurrent marking. 4429 ResourceMark rm; 4430 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4431 for (int i = 0; i < array->length(); i++) { 4432 par_mrias_cl.do_cld_nv(array->at(i)); 4433 } 4434 4435 // We don't need to keep track of new CLDs anymore. 4436 ClassLoaderDataGraph::remember_new_clds(false); 4437 4438 _timer.stop(); 4439 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4440 } 4441 4442 // We might have added oops to ClassLoaderData::_handles during the 4443 // concurrent marking phase. These oops do not always point to newly allocated objects 4444 // that are guaranteed to be kept alive. Hence, 4445 // we do have to revisit the _handles block during the remark phase. 4446 4447 // ---------- dirty CLD scanning ---------- 4448 if (worker_id == 0) { // Single threaded at the moment. 4449 _timer.reset(); 4450 _timer.start(); 4451 4452 // Scan all classes that was dirtied during the concurrent marking phase. 4453 RemarkCLDClosure remark_closure(&par_mrias_cl); 4454 ClassLoaderDataGraph::cld_do(&remark_closure); 4455 4456 _timer.stop(); 4457 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4458 } 4459 4460 4461 // ---------- rescan dirty cards ------------ 4462 _timer.reset(); 4463 _timer.start(); 4464 4465 // Do the rescan tasks for each of the two spaces 4466 // (cms_space) in turn. 4467 // "worker_id" is passed to select the task_queue for "worker_id" 4468 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4469 _timer.stop(); 4470 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4471 4472 // ---------- steal work from other threads ... 4473 // ---------- ... and drain overflow list. 4474 _timer.reset(); 4475 _timer.start(); 4476 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4477 _timer.stop(); 4478 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4479 } 4480 4481 void 4482 CMSParMarkTask::do_young_space_rescan( 4483 OopsInGenClosure* cl, ContiguousSpace* space, 4484 HeapWord** chunk_array, size_t chunk_top) { 4485 // Until all tasks completed: 4486 // . claim an unclaimed task 4487 // . compute region boundaries corresponding to task claimed 4488 // using chunk_array 4489 // . par_oop_iterate(cl) over that region 4490 4491 ResourceMark rm; 4492 HandleMark hm; 4493 4494 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4495 4496 uint nth_task = 0; 4497 uint n_tasks = pst->n_tasks(); 4498 4499 if (n_tasks > 0) { 4500 assert(pst->valid(), "Uninitialized use?"); 4501 HeapWord *start, *end; 4502 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4503 // We claimed task # nth_task; compute its boundaries. 4504 if (chunk_top == 0) { // no samples were taken 4505 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4506 start = space->bottom(); 4507 end = space->top(); 4508 } else if (nth_task == 0) { 4509 start = space->bottom(); 4510 end = chunk_array[nth_task]; 4511 } else if (nth_task < (uint)chunk_top) { 4512 assert(nth_task >= 1, "Control point invariant"); 4513 start = chunk_array[nth_task - 1]; 4514 end = chunk_array[nth_task]; 4515 } else { 4516 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4517 start = chunk_array[chunk_top - 1]; 4518 end = space->top(); 4519 } 4520 MemRegion mr(start, end); 4521 // Verify that mr is in space 4522 assert(mr.is_empty() || space->used_region().contains(mr), 4523 "Should be in space"); 4524 // Verify that "start" is an object boundary 4525 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4526 "Should be an oop"); 4527 space->par_oop_iterate(mr, cl); 4528 } 4529 pst->all_tasks_completed(); 4530 } 4531 } 4532 4533 void 4534 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4535 CompactibleFreeListSpace* sp, int i, 4536 ParMarkRefsIntoAndScanClosure* cl) { 4537 // Until all tasks completed: 4538 // . claim an unclaimed task 4539 // . compute region boundaries corresponding to task claimed 4540 // . transfer dirty bits ct->mut for that region 4541 // . apply rescanclosure to dirty mut bits for that region 4542 4543 ResourceMark rm; 4544 HandleMark hm; 4545 4546 OopTaskQueue* work_q = work_queue(i); 4547 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4548 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4549 // CAUTION: This closure has state that persists across calls to 4550 // the work method dirty_range_iterate_clear() in that it has 4551 // embedded in it a (subtype of) UpwardsObjectClosure. The 4552 // use of that state in the embedded UpwardsObjectClosure instance 4553 // assumes that the cards are always iterated (even if in parallel 4554 // by several threads) in monotonically increasing order per each 4555 // thread. This is true of the implementation below which picks 4556 // card ranges (chunks) in monotonically increasing order globally 4557 // and, a-fortiori, in monotonically increasing order per thread 4558 // (the latter order being a subsequence of the former). 4559 // If the work code below is ever reorganized into a more chaotic 4560 // work-partitioning form than the current "sequential tasks" 4561 // paradigm, the use of that persistent state will have to be 4562 // revisited and modified appropriately. See also related 4563 // bug 4756801 work on which should examine this code to make 4564 // sure that the changes there do not run counter to the 4565 // assumptions made here and necessary for correctness and 4566 // efficiency. Note also that this code might yield inefficient 4567 // behavior in the case of very large objects that span one or 4568 // more work chunks. Such objects would potentially be scanned 4569 // several times redundantly. Work on 4756801 should try and 4570 // address that performance anomaly if at all possible. XXX 4571 MemRegion full_span = _collector->_span; 4572 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4573 MarkFromDirtyCardsClosure 4574 greyRescanClosure(_collector, full_span, // entire span of interest 4575 sp, bm, work_q, cl); 4576 4577 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4578 assert(pst->valid(), "Uninitialized use?"); 4579 uint nth_task = 0; 4580 const int alignment = CardTable::card_size * BitsPerWord; 4581 MemRegion span = sp->used_region(); 4582 HeapWord* start_addr = span.start(); 4583 HeapWord* end_addr = align_up(span.end(), alignment); 4584 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4585 assert(is_aligned(start_addr, alignment), "Check alignment"); 4586 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4587 4588 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4589 // Having claimed the nth_task, compute corresponding mem-region, 4590 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4591 // The alignment restriction ensures that we do not need any 4592 // synchronization with other gang-workers while setting or 4593 // clearing bits in thus chunk of the MUT. 4594 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4595 start_addr + (nth_task+1)*chunk_size); 4596 // The last chunk's end might be way beyond end of the 4597 // used region. In that case pull back appropriately. 4598 if (this_span.end() > end_addr) { 4599 this_span.set_end(end_addr); 4600 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4601 } 4602 // Iterate over the dirty cards covering this chunk, marking them 4603 // precleaned, and setting the corresponding bits in the mod union 4604 // table. Since we have been careful to partition at Card and MUT-word 4605 // boundaries no synchronization is needed between parallel threads. 4606 _collector->_ct->dirty_card_iterate(this_span, 4607 &modUnionClosure); 4608 4609 // Having transferred these marks into the modUnionTable, 4610 // rescan the marked objects on the dirty cards in the modUnionTable. 4611 // Even if this is at a synchronous collection, the initial marking 4612 // may have been done during an asynchronous collection so there 4613 // may be dirty bits in the mod-union table. 4614 _collector->_modUnionTable.dirty_range_iterate_clear( 4615 this_span, &greyRescanClosure); 4616 _collector->_modUnionTable.verifyNoOneBitsInRange( 4617 this_span.start(), 4618 this_span.end()); 4619 } 4620 pst->all_tasks_completed(); // declare that i am done 4621 } 4622 4623 // . see if we can share work_queues with ParNew? XXX 4624 void 4625 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4626 int* seed) { 4627 OopTaskQueue* work_q = work_queue(i); 4628 NOT_PRODUCT(int num_steals = 0;) 4629 oop obj_to_scan; 4630 CMSBitMap* bm = &(_collector->_markBitMap); 4631 4632 while (true) { 4633 // Completely finish any left over work from (an) earlier round(s) 4634 cl->trim_queue(0); 4635 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4636 (size_t)ParGCDesiredObjsFromOverflowList); 4637 // Now check if there's any work in the overflow list 4638 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4639 // only affects the number of attempts made to get work from the 4640 // overflow list and does not affect the number of workers. Just 4641 // pass ParallelGCThreads so this behavior is unchanged. 4642 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4643 work_q, 4644 ParallelGCThreads)) { 4645 // found something in global overflow list; 4646 // not yet ready to go stealing work from others. 4647 // We'd like to assert(work_q->size() != 0, ...) 4648 // because we just took work from the overflow list, 4649 // but of course we can't since all of that could have 4650 // been already stolen from us. 4651 // "He giveth and He taketh away." 4652 continue; 4653 } 4654 // Verify that we have no work before we resort to stealing 4655 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4656 // Try to steal from other queues that have work 4657 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4658 NOT_PRODUCT(num_steals++;) 4659 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4660 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4661 // Do scanning work 4662 obj_to_scan->oop_iterate(cl); 4663 // Loop around, finish this work, and try to steal some more 4664 } else if (terminator()->offer_termination()) { 4665 break; // nirvana from the infinite cycle 4666 } 4667 } 4668 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4669 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4670 "Else our work is not yet done"); 4671 } 4672 4673 // Record object boundaries in _eden_chunk_array by sampling the eden 4674 // top in the slow-path eden object allocation code path and record 4675 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4676 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4677 // sampling in sample_eden() that activates during the part of the 4678 // preclean phase. 4679 void CMSCollector::sample_eden_chunk() { 4680 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4681 if (_eden_chunk_lock->try_lock()) { 4682 // Record a sample. This is the critical section. The contents 4683 // of the _eden_chunk_array have to be non-decreasing in the 4684 // address order. 4685 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4686 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4687 "Unexpected state of Eden"); 4688 if (_eden_chunk_index == 0 || 4689 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4690 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4691 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4692 _eden_chunk_index++; // commit sample 4693 } 4694 _eden_chunk_lock->unlock(); 4695 } 4696 } 4697 } 4698 4699 // Return a thread-local PLAB recording array, as appropriate. 4700 void* CMSCollector::get_data_recorder(int thr_num) { 4701 if (_survivor_plab_array != NULL && 4702 (CMSPLABRecordAlways || 4703 (_collectorState > Marking && _collectorState < FinalMarking))) { 4704 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4705 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4706 ca->reset(); // clear it so that fresh data is recorded 4707 return (void*) ca; 4708 } else { 4709 return NULL; 4710 } 4711 } 4712 4713 // Reset all the thread-local PLAB recording arrays 4714 void CMSCollector::reset_survivor_plab_arrays() { 4715 for (uint i = 0; i < ParallelGCThreads; i++) { 4716 _survivor_plab_array[i].reset(); 4717 } 4718 } 4719 4720 // Merge the per-thread plab arrays into the global survivor chunk 4721 // array which will provide the partitioning of the survivor space 4722 // for CMS initial scan and rescan. 4723 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4724 int no_of_gc_threads) { 4725 assert(_survivor_plab_array != NULL, "Error"); 4726 assert(_survivor_chunk_array != NULL, "Error"); 4727 assert(_collectorState == FinalMarking || 4728 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 _cursor[j] = 0; 4731 } 4732 HeapWord* top = surv->top(); 4733 size_t i; 4734 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4735 HeapWord* min_val = top; // Higher than any PLAB address 4736 uint min_tid = 0; // position of min_val this round 4737 for (int j = 0; j < no_of_gc_threads; j++) { 4738 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4739 if (_cursor[j] == cur_sca->end()) { 4740 continue; 4741 } 4742 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4743 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4744 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4745 if (cur_val < min_val) { 4746 min_tid = j; 4747 min_val = cur_val; 4748 } else { 4749 assert(cur_val < top, "All recorded addresses should be less"); 4750 } 4751 } 4752 // At this point min_val and min_tid are respectively 4753 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4754 // and the thread (j) that witnesses that address. 4755 // We record this address in the _survivor_chunk_array[i] 4756 // and increment _cursor[min_tid] prior to the next round i. 4757 if (min_val == top) { 4758 break; 4759 } 4760 _survivor_chunk_array[i] = min_val; 4761 _cursor[min_tid]++; 4762 } 4763 // We are all done; record the size of the _survivor_chunk_array 4764 _survivor_chunk_index = i; // exclusive: [0, i) 4765 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4766 // Verify that we used up all the recorded entries 4767 #ifdef ASSERT 4768 size_t total = 0; 4769 for (int j = 0; j < no_of_gc_threads; j++) { 4770 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4771 total += _cursor[j]; 4772 } 4773 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4774 // Check that the merged array is in sorted order 4775 if (total > 0) { 4776 for (size_t i = 0; i < total - 1; i++) { 4777 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4778 i, p2i(_survivor_chunk_array[i])); 4779 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4780 "Not sorted"); 4781 } 4782 } 4783 #endif // ASSERT 4784 } 4785 4786 // Set up the space's par_seq_tasks structure for work claiming 4787 // for parallel initial scan and rescan of young gen. 4788 // See ParRescanTask where this is currently used. 4789 void 4790 CMSCollector:: 4791 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4792 assert(n_threads > 0, "Unexpected n_threads argument"); 4793 4794 // Eden space 4795 if (!_young_gen->eden()->is_empty()) { 4796 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4797 assert(!pst->valid(), "Clobbering existing data?"); 4798 // Each valid entry in [0, _eden_chunk_index) represents a task. 4799 size_t n_tasks = _eden_chunk_index + 1; 4800 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4801 // Sets the condition for completion of the subtask (how many threads 4802 // need to finish in order to be done). 4803 pst->set_n_threads(n_threads); 4804 pst->set_n_tasks((int)n_tasks); 4805 } 4806 4807 // Merge the survivor plab arrays into _survivor_chunk_array 4808 if (_survivor_plab_array != NULL) { 4809 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4810 } else { 4811 assert(_survivor_chunk_index == 0, "Error"); 4812 } 4813 4814 // To space 4815 { 4816 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 // Sets the condition for completion of the subtask (how many threads 4819 // need to finish in order to be done). 4820 pst->set_n_threads(n_threads); 4821 pst->set_n_tasks(1); 4822 assert(pst->valid(), "Error"); 4823 } 4824 4825 // From space 4826 { 4827 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4828 assert(!pst->valid(), "Clobbering existing data?"); 4829 size_t n_tasks = _survivor_chunk_index + 1; 4830 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4831 // Sets the condition for completion of the subtask (how many threads 4832 // need to finish in order to be done). 4833 pst->set_n_threads(n_threads); 4834 pst->set_n_tasks((int)n_tasks); 4835 assert(pst->valid(), "Error"); 4836 } 4837 } 4838 4839 // Parallel version of remark 4840 void CMSCollector::do_remark_parallel() { 4841 CMSHeap* heap = CMSHeap::heap(); 4842 WorkGang* workers = heap->workers(); 4843 assert(workers != NULL, "Need parallel worker threads."); 4844 // Choose to use the number of GC workers most recently set 4845 // into "active_workers". 4846 uint n_workers = workers->active_workers(); 4847 4848 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4849 4850 StrongRootsScope srs(n_workers); 4851 4852 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4853 4854 // We won't be iterating over the cards in the card table updating 4855 // the younger_gen cards, so we shouldn't call the following else 4856 // the verification code as well as subsequent younger_refs_iterate 4857 // code would get confused. XXX 4858 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4859 4860 // The young gen rescan work will not be done as part of 4861 // process_roots (which currently doesn't know how to 4862 // parallelize such a scan), but rather will be broken up into 4863 // a set of parallel tasks (via the sampling that the [abortable] 4864 // preclean phase did of eden, plus the [two] tasks of 4865 // scanning the [two] survivor spaces. Further fine-grain 4866 // parallelization of the scanning of the survivor spaces 4867 // themselves, and of precleaning of the young gen itself 4868 // is deferred to the future. 4869 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4870 4871 // The dirty card rescan work is broken up into a "sequence" 4872 // of parallel tasks (per constituent space) that are dynamically 4873 // claimed by the parallel threads. 4874 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4875 4876 // It turns out that even when we're using 1 thread, doing the work in a 4877 // separate thread causes wide variance in run times. We can't help this 4878 // in the multi-threaded case, but we special-case n=1 here to get 4879 // repeatable measurements of the 1-thread overhead of the parallel code. 4880 if (n_workers > 1) { 4881 // Make refs discovery MT-safe, if it isn't already: it may not 4882 // necessarily be so, since it's possible that we are doing 4883 // ST marking. 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4885 workers->run_task(&tsk); 4886 } else { 4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4888 tsk.work(0); 4889 } 4890 4891 // restore, single-threaded for now, any preserved marks 4892 // as a result of work_q overflow 4893 restore_preserved_marks_if_any(); 4894 } 4895 4896 // Non-parallel version of remark 4897 void CMSCollector::do_remark_non_parallel() { 4898 ResourceMark rm; 4899 HandleMark hm; 4900 CMSHeap* heap = CMSHeap::heap(); 4901 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4902 4903 MarkRefsIntoAndScanClosure 4904 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4905 &_markStack, this, 4906 false /* should_yield */, false /* not precleaning */); 4907 MarkFromDirtyCardsClosure 4908 markFromDirtyCardsClosure(this, _span, 4909 NULL, // space is set further below 4910 &_markBitMap, &_markStack, &mrias_cl); 4911 { 4912 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4913 // Iterate over the dirty cards, setting the corresponding bits in the 4914 // mod union table. 4915 { 4916 ModUnionClosure modUnionClosure(&_modUnionTable); 4917 _ct->dirty_card_iterate(_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 = CardTable::card_size * BitsPerWord; 4925 { 4926 // ... First handle dirty cards in CMS gen 4927 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4928 MemRegion ur = _cmsGen->used_region(); 4929 HeapWord* lb = ur.start(); 4930 HeapWord* ub = align_up(ur.end(), alignment); 4931 MemRegion cms_span(lb, ub); 4932 _modUnionTable.dirty_range_iterate_clear(cms_span, 4933 &markFromDirtyCardsClosure); 4934 verify_work_stacks_empty(); 4935 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4936 } 4937 } 4938 if (VerifyDuringGC && 4939 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4940 HandleMark hm; // Discard invalid handles created during verification 4941 Universe::verify(); 4942 } 4943 { 4944 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4945 4946 verify_work_stacks_empty(); 4947 4948 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4949 StrongRootsScope srs(1); 4950 4951 heap->cms_process_roots(&srs, 4952 true, // young gen as roots 4953 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4954 should_unload_classes(), 4955 &mrias_cl, 4956 NULL); // The dirty klasses will be handled below 4957 4958 assert(should_unload_classes() 4959 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4960 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4961 } 4962 4963 { 4964 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4965 4966 verify_work_stacks_empty(); 4967 4968 // Scan all class loader data objects that might have been introduced 4969 // during concurrent marking. 4970 ResourceMark rm; 4971 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4972 for (int i = 0; i < array->length(); i++) { 4973 mrias_cl.do_cld_nv(array->at(i)); 4974 } 4975 4976 // We don't need to keep track of new CLDs anymore. 4977 ClassLoaderDataGraph::remember_new_clds(false); 4978 4979 verify_work_stacks_empty(); 4980 } 4981 4982 // We might have added oops to ClassLoaderData::_handles during the 4983 // concurrent marking phase. These oops do not point to newly allocated objects 4984 // that are guaranteed to be kept alive. Hence, 4985 // we do have to revisit the _handles block during the remark phase. 4986 { 4987 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4988 4989 verify_work_stacks_empty(); 4990 4991 RemarkCLDClosure remark_closure(&mrias_cl); 4992 ClassLoaderDataGraph::cld_do(&remark_closure); 4993 4994 verify_work_stacks_empty(); 4995 } 4996 4997 verify_work_stacks_empty(); 4998 // Restore evacuated mark words, if any, used for overflow list links 4999 restore_preserved_marks_if_any(); 5000 5001 verify_overflow_empty(); 5002 } 5003 5004 //////////////////////////////////////////////////////// 5005 // Parallel Reference Processing Task Proxy Class 5006 //////////////////////////////////////////////////////// 5007 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5008 OopTaskQueueSet* _queues; 5009 ParallelTaskTerminator _terminator; 5010 public: 5011 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5012 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5013 ParallelTaskTerminator* terminator() { return &_terminator; } 5014 OopTaskQueueSet* queues() { return _queues; } 5015 }; 5016 5017 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5018 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5019 CMSCollector* _collector; 5020 CMSBitMap* _mark_bit_map; 5021 const MemRegion _span; 5022 ProcessTask& _task; 5023 5024 public: 5025 CMSRefProcTaskProxy(ProcessTask& task, 5026 CMSCollector* collector, 5027 const MemRegion& span, 5028 CMSBitMap* mark_bit_map, 5029 AbstractWorkGang* workers, 5030 OopTaskQueueSet* task_queues): 5031 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5032 task_queues, 5033 workers->active_workers()), 5034 _task(task), 5035 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5036 { 5037 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5038 "Inconsistency in _span"); 5039 } 5040 5041 OopTaskQueueSet* task_queues() { return queues(); } 5042 5043 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5044 5045 void do_work_steal(int i, 5046 CMSParDrainMarkingStackClosure* drain, 5047 CMSParKeepAliveClosure* keep_alive, 5048 int* seed); 5049 5050 virtual void work(uint worker_id); 5051 }; 5052 5053 void CMSRefProcTaskProxy::work(uint worker_id) { 5054 ResourceMark rm; 5055 HandleMark hm; 5056 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5057 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5058 _mark_bit_map, 5059 work_queue(worker_id)); 5060 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5061 _mark_bit_map, 5062 work_queue(worker_id)); 5063 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5064 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5065 if (_task.marks_oops_alive()) { 5066 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5067 _collector->hash_seed(worker_id)); 5068 } 5069 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5070 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5071 } 5072 5073 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5074 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5075 EnqueueTask& _task; 5076 5077 public: 5078 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5079 : AbstractGangTask("Enqueue reference objects in parallel"), 5080 _task(task) 5081 { } 5082 5083 virtual void work(uint worker_id) 5084 { 5085 _task.work(worker_id); 5086 } 5087 }; 5088 5089 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5090 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5091 _span(span), 5092 _bit_map(bit_map), 5093 _work_queue(work_queue), 5094 _mark_and_push(collector, span, bit_map, work_queue), 5095 _low_water_mark(MIN2((work_queue->max_elems()/4), 5096 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5097 { } 5098 5099 // . see if we can share work_queues with ParNew? XXX 5100 void CMSRefProcTaskProxy::do_work_steal(int i, 5101 CMSParDrainMarkingStackClosure* drain, 5102 CMSParKeepAliveClosure* keep_alive, 5103 int* seed) { 5104 OopTaskQueue* work_q = work_queue(i); 5105 NOT_PRODUCT(int num_steals = 0;) 5106 oop obj_to_scan; 5107 5108 while (true) { 5109 // Completely finish any left over work from (an) earlier round(s) 5110 drain->trim_queue(0); 5111 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5112 (size_t)ParGCDesiredObjsFromOverflowList); 5113 // Now check if there's any work in the overflow list 5114 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5115 // only affects the number of attempts made to get work from the 5116 // overflow list and does not affect the number of workers. Just 5117 // pass ParallelGCThreads so this behavior is unchanged. 5118 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5119 work_q, 5120 ParallelGCThreads)) { 5121 // Found something in global overflow list; 5122 // not yet ready to go stealing work from others. 5123 // We'd like to assert(work_q->size() != 0, ...) 5124 // because we just took work from the overflow list, 5125 // but of course we can't, since all of that might have 5126 // been already stolen from us. 5127 continue; 5128 } 5129 // Verify that we have no work before we resort to stealing 5130 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5131 // Try to steal from other queues that have work 5132 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5133 NOT_PRODUCT(num_steals++;) 5134 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5135 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5136 // Do scanning work 5137 obj_to_scan->oop_iterate(keep_alive); 5138 // Loop around, finish this work, and try to steal some more 5139 } else if (terminator()->offer_termination()) { 5140 break; // nirvana from the infinite cycle 5141 } 5142 } 5143 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5144 } 5145 5146 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5147 { 5148 CMSHeap* heap = CMSHeap::heap(); 5149 WorkGang* workers = heap->workers(); 5150 assert(workers != NULL, "Need parallel worker threads."); 5151 CMSRefProcTaskProxy rp_task(task, &_collector, 5152 _collector.ref_processor()->span(), 5153 _collector.markBitMap(), 5154 workers, _collector.task_queues()); 5155 workers->run_task(&rp_task); 5156 } 5157 5158 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5159 { 5160 5161 CMSHeap* heap = CMSHeap::heap(); 5162 WorkGang* workers = heap->workers(); 5163 assert(workers != NULL, "Need parallel worker threads."); 5164 CMSRefEnqueueTaskProxy enq_task(task); 5165 workers->run_task(&enq_task); 5166 } 5167 5168 void CMSCollector::refProcessingWork() { 5169 ResourceMark rm; 5170 HandleMark hm; 5171 5172 ReferenceProcessor* rp = ref_processor(); 5173 assert(rp->span().equals(_span), "Spans should be equal"); 5174 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5175 // Process weak references. 5176 rp->setup_policy(false); 5177 verify_work_stacks_empty(); 5178 5179 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q()); 5180 { 5181 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5182 5183 // Setup keep_alive and complete closures. 5184 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5185 &_markStack, false /* !preclean */); 5186 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5187 _span, &_markBitMap, &_markStack, 5188 &cmsKeepAliveClosure, false /* !preclean */); 5189 5190 ReferenceProcessorStats stats; 5191 if (rp->processing_is_mt()) { 5192 // Set the degree of MT here. If the discovery is done MT, there 5193 // may have been a different number of threads doing the discovery 5194 // and a different number of discovered lists may have Ref objects. 5195 // That is OK as long as the Reference lists are balanced (see 5196 // balance_all_queues() and balance_queues()). 5197 CMSHeap* heap = CMSHeap::heap(); 5198 uint active_workers = ParallelGCThreads; 5199 WorkGang* workers = heap->workers(); 5200 if (workers != NULL) { 5201 active_workers = workers->active_workers(); 5202 // The expectation is that active_workers will have already 5203 // been set to a reasonable value. If it has not been set, 5204 // investigate. 5205 assert(active_workers > 0, "Should have been set during scavenge"); 5206 } 5207 rp->set_active_mt_degree(active_workers); 5208 CMSRefProcTaskExecutor task_executor(*this); 5209 stats = rp->process_discovered_references(&_is_alive_closure, 5210 &cmsKeepAliveClosure, 5211 &cmsDrainMarkingStackClosure, 5212 &task_executor, 5213 &pt); 5214 } else { 5215 stats = rp->process_discovered_references(&_is_alive_closure, 5216 &cmsKeepAliveClosure, 5217 &cmsDrainMarkingStackClosure, 5218 NULL, 5219 &pt); 5220 } 5221 _gc_tracer_cm->report_gc_reference_stats(stats); 5222 pt.print_all_references(); 5223 } 5224 5225 // This is the point where the entire marking should have completed. 5226 verify_work_stacks_empty(); 5227 5228 { 5229 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5230 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5231 } 5232 5233 if (should_unload_classes()) { 5234 { 5235 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5236 5237 // Unload classes and purge the SystemDictionary. 5238 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5239 5240 // Unload nmethods. 5241 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5242 5243 // Prune dead klasses from subklass/sibling/implementor lists. 5244 Klass::clean_weak_klass_links(&_is_alive_closure); 5245 } 5246 5247 { 5248 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5249 // Clean up unreferenced symbols in symbol table. 5250 SymbolTable::unlink(); 5251 } 5252 5253 { 5254 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5255 // Delete entries for dead interned strings. 5256 StringTable::unlink(&_is_alive_closure); 5257 } 5258 } 5259 5260 // Restore any preserved marks as a result of mark stack or 5261 // work queue overflow 5262 restore_preserved_marks_if_any(); // done single-threaded for now 5263 5264 rp->set_enqueuing_is_done(true); 5265 if (rp->processing_is_mt()) { 5266 rp->balance_all_queues(); 5267 CMSRefProcTaskExecutor task_executor(*this); 5268 rp->enqueue_discovered_references(&task_executor, &pt); 5269 } else { 5270 rp->enqueue_discovered_references(NULL, &pt); 5271 } 5272 rp->verify_no_references_recorded(); 5273 pt.print_enqueue_phase(); 5274 assert(!rp->discovery_enabled(), "should have been disabled"); 5275 } 5276 5277 #ifndef PRODUCT 5278 void CMSCollector::check_correct_thread_executing() { 5279 Thread* t = Thread::current(); 5280 // Only the VM thread or the CMS thread should be here. 5281 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5282 "Unexpected thread type"); 5283 // If this is the vm thread, the foreground process 5284 // should not be waiting. Note that _foregroundGCIsActive is 5285 // true while the foreground collector is waiting. 5286 if (_foregroundGCShouldWait) { 5287 // We cannot be the VM thread 5288 assert(t->is_ConcurrentGC_thread(), 5289 "Should be CMS thread"); 5290 } else { 5291 // We can be the CMS thread only if we are in a stop-world 5292 // phase of CMS collection. 5293 if (t->is_ConcurrentGC_thread()) { 5294 assert(_collectorState == InitialMarking || 5295 _collectorState == FinalMarking, 5296 "Should be a stop-world phase"); 5297 // The CMS thread should be holding the CMS_token. 5298 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5299 "Potential interference with concurrently " 5300 "executing VM thread"); 5301 } 5302 } 5303 } 5304 #endif 5305 5306 void CMSCollector::sweep() { 5307 assert(_collectorState == Sweeping, "just checking"); 5308 check_correct_thread_executing(); 5309 verify_work_stacks_empty(); 5310 verify_overflow_empty(); 5311 increment_sweep_count(); 5312 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5313 5314 _inter_sweep_timer.stop(); 5315 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5316 5317 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5318 _intra_sweep_timer.reset(); 5319 _intra_sweep_timer.start(); 5320 { 5321 GCTraceCPUTime tcpu; 5322 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5323 // First sweep the old gen 5324 { 5325 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5326 bitMapLock()); 5327 sweepWork(_cmsGen); 5328 } 5329 5330 // Update Universe::_heap_*_at_gc figures. 5331 // We need all the free list locks to make the abstract state 5332 // transition from Sweeping to Resetting. See detailed note 5333 // further below. 5334 { 5335 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5336 // Update heap occupancy information which is used as 5337 // input to soft ref clearing policy at the next gc. 5338 Universe::update_heap_info_at_gc(); 5339 _collectorState = Resizing; 5340 } 5341 } 5342 verify_work_stacks_empty(); 5343 verify_overflow_empty(); 5344 5345 if (should_unload_classes()) { 5346 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5347 // requires that the virtual spaces are stable and not deleted. 5348 ClassLoaderDataGraph::set_should_purge(true); 5349 } 5350 5351 _intra_sweep_timer.stop(); 5352 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5353 5354 _inter_sweep_timer.reset(); 5355 _inter_sweep_timer.start(); 5356 5357 // We need to use a monotonically non-decreasing time in ms 5358 // or we will see time-warp warnings and os::javaTimeMillis() 5359 // does not guarantee monotonicity. 5360 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5361 update_time_of_last_gc(now); 5362 5363 // NOTE on abstract state transitions: 5364 // Mutators allocate-live and/or mark the mod-union table dirty 5365 // based on the state of the collection. The former is done in 5366 // the interval [Marking, Sweeping] and the latter in the interval 5367 // [Marking, Sweeping). Thus the transitions into the Marking state 5368 // and out of the Sweeping state must be synchronously visible 5369 // globally to the mutators. 5370 // The transition into the Marking state happens with the world 5371 // stopped so the mutators will globally see it. Sweeping is 5372 // done asynchronously by the background collector so the transition 5373 // from the Sweeping state to the Resizing state must be done 5374 // under the freelistLock (as is the check for whether to 5375 // allocate-live and whether to dirty the mod-union table). 5376 assert(_collectorState == Resizing, "Change of collector state to" 5377 " Resizing must be done under the freelistLocks (plural)"); 5378 5379 // Now that sweeping has been completed, we clear 5380 // the incremental_collection_failed flag, 5381 // thus inviting a younger gen collection to promote into 5382 // this generation. If such a promotion may still fail, 5383 // the flag will be set again when a young collection is 5384 // attempted. 5385 CMSHeap* heap = CMSHeap::heap(); 5386 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5387 heap->update_full_collections_completed(_collection_count_start); 5388 } 5389 5390 // FIX ME!!! Looks like this belongs in CFLSpace, with 5391 // CMSGen merely delegating to it. 5392 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5393 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5394 HeapWord* minAddr = _cmsSpace->bottom(); 5395 HeapWord* largestAddr = 5396 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5397 if (largestAddr == NULL) { 5398 // The dictionary appears to be empty. In this case 5399 // try to coalesce at the end of the heap. 5400 largestAddr = _cmsSpace->end(); 5401 } 5402 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5403 size_t nearLargestOffset = 5404 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5405 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5406 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5407 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5408 } 5409 5410 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5411 return addr >= _cmsSpace->nearLargestChunk(); 5412 } 5413 5414 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5415 return _cmsSpace->find_chunk_at_end(); 5416 } 5417 5418 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5419 bool full) { 5420 // If the young generation has been collected, gather any statistics 5421 // that are of interest at this point. 5422 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5423 if (!full && current_is_young) { 5424 // Gather statistics on the young generation collection. 5425 collector()->stats().record_gc0_end(used()); 5426 } 5427 } 5428 5429 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5430 // We iterate over the space(s) underlying this generation, 5431 // checking the mark bit map to see if the bits corresponding 5432 // to specific blocks are marked or not. Blocks that are 5433 // marked are live and are not swept up. All remaining blocks 5434 // are swept up, with coalescing on-the-fly as we sweep up 5435 // contiguous free and/or garbage blocks: 5436 // We need to ensure that the sweeper synchronizes with allocators 5437 // and stop-the-world collectors. In particular, the following 5438 // locks are used: 5439 // . CMS token: if this is held, a stop the world collection cannot occur 5440 // . freelistLock: if this is held no allocation can occur from this 5441 // generation by another thread 5442 // . bitMapLock: if this is held, no other thread can access or update 5443 // 5444 5445 // Note that we need to hold the freelistLock if we use 5446 // block iterate below; else the iterator might go awry if 5447 // a mutator (or promotion) causes block contents to change 5448 // (for instance if the allocator divvies up a block). 5449 // If we hold the free list lock, for all practical purposes 5450 // young generation GC's can't occur (they'll usually need to 5451 // promote), so we might as well prevent all young generation 5452 // GC's while we do a sweeping step. For the same reason, we might 5453 // as well take the bit map lock for the entire duration 5454 5455 // check that we hold the requisite locks 5456 assert(have_cms_token(), "Should hold cms token"); 5457 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5458 assert_lock_strong(old_gen->freelistLock()); 5459 assert_lock_strong(bitMapLock()); 5460 5461 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5462 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5463 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5464 _inter_sweep_estimate.padded_average(), 5465 _intra_sweep_estimate.padded_average()); 5466 old_gen->setNearLargestChunk(); 5467 5468 { 5469 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5470 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5471 // We need to free-up/coalesce garbage/blocks from a 5472 // co-terminal free run. This is done in the SweepClosure 5473 // destructor; so, do not remove this scope, else the 5474 // end-of-sweep-census below will be off by a little bit. 5475 } 5476 old_gen->cmsSpace()->sweep_completed(); 5477 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5478 if (should_unload_classes()) { // unloaded classes this cycle, 5479 _concurrent_cycles_since_last_unload = 0; // ... reset count 5480 } else { // did not unload classes, 5481 _concurrent_cycles_since_last_unload++; // ... increment count 5482 } 5483 } 5484 5485 // Reset CMS data structures (for now just the marking bit map) 5486 // preparatory for the next cycle. 5487 void CMSCollector::reset_concurrent() { 5488 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5489 5490 // If the state is not "Resetting", the foreground thread 5491 // has done a collection and the resetting. 5492 if (_collectorState != Resetting) { 5493 assert(_collectorState == Idling, "The state should only change" 5494 " because the foreground collector has finished the collection"); 5495 return; 5496 } 5497 5498 { 5499 // Clear the mark bitmap (no grey objects to start with) 5500 // for the next cycle. 5501 GCTraceCPUTime tcpu; 5502 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5503 5504 HeapWord* curAddr = _markBitMap.startWord(); 5505 while (curAddr < _markBitMap.endWord()) { 5506 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5507 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5508 _markBitMap.clear_large_range(chunk); 5509 if (ConcurrentMarkSweepThread::should_yield() && 5510 !foregroundGCIsActive() && 5511 CMSYield) { 5512 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5513 "CMS thread should hold CMS token"); 5514 assert_lock_strong(bitMapLock()); 5515 bitMapLock()->unlock(); 5516 ConcurrentMarkSweepThread::desynchronize(true); 5517 stopTimer(); 5518 incrementYields(); 5519 5520 // See the comment in coordinator_yield() 5521 for (unsigned i = 0; i < CMSYieldSleepCount && 5522 ConcurrentMarkSweepThread::should_yield() && 5523 !CMSCollector::foregroundGCIsActive(); ++i) { 5524 os::sleep(Thread::current(), 1, false); 5525 } 5526 5527 ConcurrentMarkSweepThread::synchronize(true); 5528 bitMapLock()->lock_without_safepoint_check(); 5529 startTimer(); 5530 } 5531 curAddr = chunk.end(); 5532 } 5533 // A successful mostly concurrent collection has been done. 5534 // Because only the full (i.e., concurrent mode failure) collections 5535 // are being measured for gc overhead limits, clean the "near" flag 5536 // and count. 5537 size_policy()->reset_gc_overhead_limit_count(); 5538 _collectorState = Idling; 5539 } 5540 5541 register_gc_end(); 5542 } 5543 5544 // Same as above but for STW paths 5545 void CMSCollector::reset_stw() { 5546 // already have the lock 5547 assert(_collectorState == Resetting, "just checking"); 5548 assert_lock_strong(bitMapLock()); 5549 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5550 _markBitMap.clear_all(); 5551 _collectorState = Idling; 5552 register_gc_end(); 5553 } 5554 5555 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5556 GCTraceCPUTime tcpu; 5557 TraceCollectorStats tcs_cgc(cgc_counters()); 5558 5559 switch (op) { 5560 case CMS_op_checkpointRootsInitial: { 5561 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5562 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5563 checkpointRootsInitial(); 5564 break; 5565 } 5566 case CMS_op_checkpointRootsFinal: { 5567 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5568 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5569 checkpointRootsFinal(); 5570 break; 5571 } 5572 default: 5573 fatal("No such CMS_op"); 5574 } 5575 } 5576 5577 #ifndef PRODUCT 5578 size_t const CMSCollector::skip_header_HeapWords() { 5579 return FreeChunk::header_size(); 5580 } 5581 5582 // Try and collect here conditions that should hold when 5583 // CMS thread is exiting. The idea is that the foreground GC 5584 // thread should not be blocked if it wants to terminate 5585 // the CMS thread and yet continue to run the VM for a while 5586 // after that. 5587 void CMSCollector::verify_ok_to_terminate() const { 5588 assert(Thread::current()->is_ConcurrentGC_thread(), 5589 "should be called by CMS thread"); 5590 assert(!_foregroundGCShouldWait, "should be false"); 5591 // We could check here that all the various low-level locks 5592 // are not held by the CMS thread, but that is overkill; see 5593 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5594 // is checked. 5595 } 5596 #endif 5597 5598 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5599 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5600 "missing Printezis mark?"); 5601 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5602 size_t size = pointer_delta(nextOneAddr + 1, addr); 5603 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5604 "alignment problem"); 5605 assert(size >= 3, "Necessary for Printezis marks to work"); 5606 return size; 5607 } 5608 5609 // A variant of the above (block_size_using_printezis_bits()) except 5610 // that we return 0 if the P-bits are not yet set. 5611 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5612 if (_markBitMap.isMarked(addr + 1)) { 5613 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5614 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5615 size_t size = pointer_delta(nextOneAddr + 1, addr); 5616 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5617 "alignment problem"); 5618 assert(size >= 3, "Necessary for Printezis marks to work"); 5619 return size; 5620 } 5621 return 0; 5622 } 5623 5624 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5625 size_t sz = 0; 5626 oop p = (oop)addr; 5627 if (p->klass_or_null_acquire() != NULL) { 5628 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5629 } else { 5630 sz = block_size_using_printezis_bits(addr); 5631 } 5632 assert(sz > 0, "size must be nonzero"); 5633 HeapWord* next_block = addr + sz; 5634 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5635 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5636 align_down((uintptr_t)next_card, CardTable::card_size), 5637 "must be different cards"); 5638 return next_card; 5639 } 5640 5641 5642 // CMS Bit Map Wrapper ///////////////////////////////////////// 5643 5644 // Construct a CMS bit map infrastructure, but don't create the 5645 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5646 // further below. 5647 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5648 _bm(), 5649 _shifter(shifter), 5650 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5651 Monitor::_safepoint_check_sometimes) : NULL) 5652 { 5653 _bmStartWord = 0; 5654 _bmWordSize = 0; 5655 } 5656 5657 bool CMSBitMap::allocate(MemRegion mr) { 5658 _bmStartWord = mr.start(); 5659 _bmWordSize = mr.word_size(); 5660 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5661 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5662 if (!brs.is_reserved()) { 5663 log_warning(gc)("CMS bit map allocation failure"); 5664 return false; 5665 } 5666 // For now we'll just commit all of the bit map up front. 5667 // Later on we'll try to be more parsimonious with swap. 5668 if (!_virtual_space.initialize(brs, brs.size())) { 5669 log_warning(gc)("CMS bit map backing store failure"); 5670 return false; 5671 } 5672 assert(_virtual_space.committed_size() == brs.size(), 5673 "didn't reserve backing store for all of CMS bit map?"); 5674 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5675 _bmWordSize, "inconsistency in bit map sizing"); 5676 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5677 5678 // bm.clear(); // can we rely on getting zero'd memory? verify below 5679 assert(isAllClear(), 5680 "Expected zero'd memory from ReservedSpace constructor"); 5681 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5682 "consistency check"); 5683 return true; 5684 } 5685 5686 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5687 HeapWord *next_addr, *end_addr, *last_addr; 5688 assert_locked(); 5689 assert(covers(mr), "out-of-range error"); 5690 // XXX assert that start and end are appropriately aligned 5691 for (next_addr = mr.start(), end_addr = mr.end(); 5692 next_addr < end_addr; next_addr = last_addr) { 5693 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5694 last_addr = dirty_region.end(); 5695 if (!dirty_region.is_empty()) { 5696 cl->do_MemRegion(dirty_region); 5697 } else { 5698 assert(last_addr == end_addr, "program logic"); 5699 return; 5700 } 5701 } 5702 } 5703 5704 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5705 _bm.print_on_error(st, prefix); 5706 } 5707 5708 #ifndef PRODUCT 5709 void CMSBitMap::assert_locked() const { 5710 CMSLockVerifier::assert_locked(lock()); 5711 } 5712 5713 bool CMSBitMap::covers(MemRegion mr) const { 5714 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5715 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5716 "size inconsistency"); 5717 return (mr.start() >= _bmStartWord) && 5718 (mr.end() <= endWord()); 5719 } 5720 5721 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5722 return (start >= _bmStartWord && (start + size) <= endWord()); 5723 } 5724 5725 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5726 // verify that there are no 1 bits in the interval [left, right) 5727 FalseBitMapClosure falseBitMapClosure; 5728 iterate(&falseBitMapClosure, left, right); 5729 } 5730 5731 void CMSBitMap::region_invariant(MemRegion mr) 5732 { 5733 assert_locked(); 5734 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5735 assert(!mr.is_empty(), "unexpected empty region"); 5736 assert(covers(mr), "mr should be covered by bit map"); 5737 // convert address range into offset range 5738 size_t start_ofs = heapWordToOffset(mr.start()); 5739 // Make sure that end() is appropriately aligned 5740 assert(mr.end() == align_up(mr.end(), (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 log_warning(gc)("CMSMarkStack allocation failure"); 5754 return false; 5755 } 5756 if (!_virtual_space.initialize(rs, rs.size())) { 5757 log_warning(gc)("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(oopDesc::is_oop(obj), "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(oopDesc::is_oop(obj), "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(oopDesc::is_oop(obj), "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 Log(gc, verify) log; 5883 ResourceMark rm; 5884 LogStream ls(log.error()); 5885 oop(addr)->print_on(&ls); 5886 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5887 fatal("... aborting"); 5888 } 5889 } 5890 } 5891 5892 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5893 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5894 5895 ////////////////////////////////////////////////// 5896 // MarkRefsIntoAndScanClosure 5897 ////////////////////////////////////////////////// 5898 5899 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5900 ReferenceProcessor* rp, 5901 CMSBitMap* bit_map, 5902 CMSBitMap* mod_union_table, 5903 CMSMarkStack* mark_stack, 5904 CMSCollector* collector, 5905 bool should_yield, 5906 bool concurrent_precleaning): 5907 _collector(collector), 5908 _span(span), 5909 _bit_map(bit_map), 5910 _mark_stack(mark_stack), 5911 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5912 mark_stack, concurrent_precleaning), 5913 _yield(should_yield), 5914 _concurrent_precleaning(concurrent_precleaning), 5915 _freelistLock(NULL) 5916 { 5917 // FIXME: Should initialize in base class constructor. 5918 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5919 set_ref_processor_internal(rp); 5920 } 5921 5922 // This closure is used to mark refs into the CMS generation at the 5923 // second (final) checkpoint, and to scan and transitively follow 5924 // the unmarked oops. It is also used during the concurrent precleaning 5925 // phase while scanning objects on dirty cards in the CMS generation. 5926 // The marks are made in the marking bit map and the marking stack is 5927 // used for keeping the (newly) grey objects during the scan. 5928 // The parallel version (Par_...) appears further below. 5929 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5930 if (obj != NULL) { 5931 assert(oopDesc::is_oop(obj), "expected an oop"); 5932 HeapWord* addr = (HeapWord*)obj; 5933 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5934 assert(_collector->overflow_list_is_empty(), 5935 "overflow list should be empty"); 5936 if (_span.contains(addr) && 5937 !_bit_map->isMarked(addr)) { 5938 // mark bit map (object is now grey) 5939 _bit_map->mark(addr); 5940 // push on marking stack (stack should be empty), and drain the 5941 // stack by applying this closure to the oops in the oops popped 5942 // from the stack (i.e. blacken the grey objects) 5943 bool res = _mark_stack->push(obj); 5944 assert(res, "Should have space to push on empty stack"); 5945 do { 5946 oop new_oop = _mark_stack->pop(); 5947 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5948 assert(_bit_map->isMarked((HeapWord*)new_oop), 5949 "only grey objects on this stack"); 5950 // iterate over the oops in this oop, marking and pushing 5951 // the ones in CMS heap (i.e. in _span). 5952 new_oop->oop_iterate(&_pushAndMarkClosure); 5953 // check if it's time to yield 5954 do_yield_check(); 5955 } while (!_mark_stack->isEmpty() || 5956 (!_concurrent_precleaning && take_from_overflow_list())); 5957 // if marking stack is empty, and we are not doing this 5958 // during precleaning, then check the overflow list 5959 } 5960 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5961 assert(_collector->overflow_list_is_empty(), 5962 "overflow list was drained above"); 5963 5964 assert(_collector->no_preserved_marks(), 5965 "All preserved marks should have been restored above"); 5966 } 5967 } 5968 5969 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5970 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5971 5972 void MarkRefsIntoAndScanClosure::do_yield_work() { 5973 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5974 "CMS thread should hold CMS token"); 5975 assert_lock_strong(_freelistLock); 5976 assert_lock_strong(_bit_map->lock()); 5977 // relinquish the free_list_lock and bitMaplock() 5978 _bit_map->lock()->unlock(); 5979 _freelistLock->unlock(); 5980 ConcurrentMarkSweepThread::desynchronize(true); 5981 _collector->stopTimer(); 5982 _collector->incrementYields(); 5983 5984 // See the comment in coordinator_yield() 5985 for (unsigned i = 0; 5986 i < CMSYieldSleepCount && 5987 ConcurrentMarkSweepThread::should_yield() && 5988 !CMSCollector::foregroundGCIsActive(); 5989 ++i) { 5990 os::sleep(Thread::current(), 1, false); 5991 } 5992 5993 ConcurrentMarkSweepThread::synchronize(true); 5994 _freelistLock->lock_without_safepoint_check(); 5995 _bit_map->lock()->lock_without_safepoint_check(); 5996 _collector->startTimer(); 5997 } 5998 5999 /////////////////////////////////////////////////////////// 6000 // ParMarkRefsIntoAndScanClosure: a parallel version of 6001 // MarkRefsIntoAndScanClosure 6002 /////////////////////////////////////////////////////////// 6003 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6004 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6005 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6006 _span(span), 6007 _bit_map(bit_map), 6008 _work_queue(work_queue), 6009 _low_water_mark(MIN2((work_queue->max_elems()/4), 6010 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6011 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6012 { 6013 // FIXME: Should initialize in base class constructor. 6014 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6015 set_ref_processor_internal(rp); 6016 } 6017 6018 // This closure is used to mark refs into the CMS generation at the 6019 // second (final) checkpoint, and to scan and transitively follow 6020 // the unmarked oops. The marks are made in the marking bit map and 6021 // the work_queue is used for keeping the (newly) grey objects during 6022 // the scan phase whence they are also available for stealing by parallel 6023 // threads. Since the marking bit map is shared, updates are 6024 // synchronized (via CAS). 6025 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6026 if (obj != NULL) { 6027 // Ignore mark word because this could be an already marked oop 6028 // that may be chained at the end of the overflow list. 6029 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6030 HeapWord* addr = (HeapWord*)obj; 6031 if (_span.contains(addr) && 6032 !_bit_map->isMarked(addr)) { 6033 // mark bit map (object will become grey): 6034 // It is possible for several threads to be 6035 // trying to "claim" this object concurrently; 6036 // the unique thread that succeeds in marking the 6037 // object first will do the subsequent push on 6038 // to the work queue (or overflow list). 6039 if (_bit_map->par_mark(addr)) { 6040 // push on work_queue (which may not be empty), and trim the 6041 // queue to an appropriate length by applying this closure to 6042 // the oops in the oops popped from the stack (i.e. blacken the 6043 // grey objects) 6044 bool res = _work_queue->push(obj); 6045 assert(res, "Low water mark should be less than capacity?"); 6046 trim_queue(_low_water_mark); 6047 } // Else, another thread claimed the object 6048 } 6049 } 6050 } 6051 6052 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6053 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6054 6055 // This closure is used to rescan the marked objects on the dirty cards 6056 // in the mod union table and the card table proper. 6057 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6058 oop p, MemRegion mr) { 6059 6060 size_t size = 0; 6061 HeapWord* addr = (HeapWord*)p; 6062 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6063 assert(_span.contains(addr), "we are scanning the CMS generation"); 6064 // check if it's time to yield 6065 if (do_yield_check()) { 6066 // We yielded for some foreground stop-world work, 6067 // and we have been asked to abort this ongoing preclean cycle. 6068 return 0; 6069 } 6070 if (_bitMap->isMarked(addr)) { 6071 // it's marked; is it potentially uninitialized? 6072 if (p->klass_or_null_acquire() != NULL) { 6073 // an initialized object; ignore mark word in verification below 6074 // since we are running concurrent with mutators 6075 assert(oopDesc::is_oop(p, true), "should be an oop"); 6076 if (p->is_objArray()) { 6077 // objArrays are precisely marked; restrict scanning 6078 // to dirty cards only. 6079 size = CompactibleFreeListSpace::adjustObjectSize( 6080 p->oop_iterate_size(_scanningClosure, mr)); 6081 } else { 6082 // A non-array may have been imprecisely marked; we need 6083 // to scan object in its entirety. 6084 size = CompactibleFreeListSpace::adjustObjectSize( 6085 p->oop_iterate_size(_scanningClosure)); 6086 } 6087 #ifdef ASSERT 6088 size_t direct_size = 6089 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6090 assert(size == direct_size, "Inconsistency in size"); 6091 assert(size >= 3, "Necessary for Printezis marks to work"); 6092 HeapWord* start_pbit = addr + 1; 6093 HeapWord* end_pbit = addr + size - 1; 6094 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6095 "inconsistent Printezis mark"); 6096 // Verify inner mark bits (between Printezis bits) are clear, 6097 // but don't repeat if there are multiple dirty regions for 6098 // the same object, to avoid potential O(N^2) performance. 6099 if (addr != _last_scanned_object) { 6100 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6101 _last_scanned_object = addr; 6102 } 6103 #endif // ASSERT 6104 } else { 6105 // An uninitialized object. 6106 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6107 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6108 size = pointer_delta(nextOneAddr + 1, addr); 6109 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6110 "alignment problem"); 6111 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6112 // will dirty the card when the klass pointer is installed in the 6113 // object (signaling the completion of initialization). 6114 } 6115 } else { 6116 // Either a not yet marked object or an uninitialized object 6117 if (p->klass_or_null_acquire() == NULL) { 6118 // An uninitialized object, skip to the next card, since 6119 // we may not be able to read its P-bits yet. 6120 assert(size == 0, "Initial value"); 6121 } else { 6122 // An object not (yet) reached by marking: we merely need to 6123 // compute its size so as to go look at the next block. 6124 assert(oopDesc::is_oop(p, true), "should be an oop"); 6125 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6126 } 6127 } 6128 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6129 return size; 6130 } 6131 6132 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6133 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6134 "CMS thread should hold CMS token"); 6135 assert_lock_strong(_freelistLock); 6136 assert_lock_strong(_bitMap->lock()); 6137 // relinquish the free_list_lock and bitMaplock() 6138 _bitMap->lock()->unlock(); 6139 _freelistLock->unlock(); 6140 ConcurrentMarkSweepThread::desynchronize(true); 6141 _collector->stopTimer(); 6142 _collector->incrementYields(); 6143 6144 // See the comment in coordinator_yield() 6145 for (unsigned i = 0; i < CMSYieldSleepCount && 6146 ConcurrentMarkSweepThread::should_yield() && 6147 !CMSCollector::foregroundGCIsActive(); ++i) { 6148 os::sleep(Thread::current(), 1, false); 6149 } 6150 6151 ConcurrentMarkSweepThread::synchronize(true); 6152 _freelistLock->lock_without_safepoint_check(); 6153 _bitMap->lock()->lock_without_safepoint_check(); 6154 _collector->startTimer(); 6155 } 6156 6157 6158 ////////////////////////////////////////////////////////////////// 6159 // SurvivorSpacePrecleanClosure 6160 ////////////////////////////////////////////////////////////////// 6161 // This (single-threaded) closure is used to preclean the oops in 6162 // the survivor spaces. 6163 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6164 6165 HeapWord* addr = (HeapWord*)p; 6166 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6167 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6168 assert(p->klass_or_null() != NULL, "object should be initialized"); 6169 // an initialized object; ignore mark word in verification below 6170 // since we are running concurrent with mutators 6171 assert(oopDesc::is_oop(p, true), "should be an oop"); 6172 // Note that we do not yield while we iterate over 6173 // the interior oops of p, pushing the relevant ones 6174 // on our marking stack. 6175 size_t size = p->oop_iterate_size(_scanning_closure); 6176 do_yield_check(); 6177 // Observe that below, we do not abandon the preclean 6178 // phase as soon as we should; rather we empty the 6179 // marking stack before returning. This is to satisfy 6180 // some existing assertions. In general, it may be a 6181 // good idea to abort immediately and complete the marking 6182 // from the grey objects at a later time. 6183 while (!_mark_stack->isEmpty()) { 6184 oop new_oop = _mark_stack->pop(); 6185 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6186 assert(_bit_map->isMarked((HeapWord*)new_oop), 6187 "only grey objects on this stack"); 6188 // iterate over the oops in this oop, marking and pushing 6189 // the ones in CMS heap (i.e. in _span). 6190 new_oop->oop_iterate(_scanning_closure); 6191 // check if it's time to yield 6192 do_yield_check(); 6193 } 6194 unsigned int after_count = 6195 CMSHeap::heap()->total_collections(); 6196 bool abort = (_before_count != after_count) || 6197 _collector->should_abort_preclean(); 6198 return abort ? 0 : size; 6199 } 6200 6201 void SurvivorSpacePrecleanClosure::do_yield_work() { 6202 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6203 "CMS thread should hold CMS token"); 6204 assert_lock_strong(_bit_map->lock()); 6205 // Relinquish the bit map lock 6206 _bit_map->lock()->unlock(); 6207 ConcurrentMarkSweepThread::desynchronize(true); 6208 _collector->stopTimer(); 6209 _collector->incrementYields(); 6210 6211 // See the comment in coordinator_yield() 6212 for (unsigned i = 0; i < CMSYieldSleepCount && 6213 ConcurrentMarkSweepThread::should_yield() && 6214 !CMSCollector::foregroundGCIsActive(); ++i) { 6215 os::sleep(Thread::current(), 1, false); 6216 } 6217 6218 ConcurrentMarkSweepThread::synchronize(true); 6219 _bit_map->lock()->lock_without_safepoint_check(); 6220 _collector->startTimer(); 6221 } 6222 6223 // This closure is used to rescan the marked objects on the dirty cards 6224 // in the mod union table and the card table proper. In the parallel 6225 // case, although the bitMap is shared, we do a single read so the 6226 // isMarked() query is "safe". 6227 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6228 // Ignore mark word because we are running concurrent with mutators 6229 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6230 HeapWord* addr = (HeapWord*)p; 6231 assert(_span.contains(addr), "we are scanning the CMS generation"); 6232 bool is_obj_array = false; 6233 #ifdef ASSERT 6234 if (!_parallel) { 6235 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6236 assert(_collector->overflow_list_is_empty(), 6237 "overflow list should be empty"); 6238 6239 } 6240 #endif // ASSERT 6241 if (_bit_map->isMarked(addr)) { 6242 // Obj arrays are precisely marked, non-arrays are not; 6243 // so we scan objArrays precisely and non-arrays in their 6244 // entirety. 6245 if (p->is_objArray()) { 6246 is_obj_array = true; 6247 if (_parallel) { 6248 p->oop_iterate(_par_scan_closure, mr); 6249 } else { 6250 p->oop_iterate(_scan_closure, mr); 6251 } 6252 } else { 6253 if (_parallel) { 6254 p->oop_iterate(_par_scan_closure); 6255 } else { 6256 p->oop_iterate(_scan_closure); 6257 } 6258 } 6259 } 6260 #ifdef ASSERT 6261 if (!_parallel) { 6262 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6263 assert(_collector->overflow_list_is_empty(), 6264 "overflow list should be empty"); 6265 6266 } 6267 #endif // ASSERT 6268 return is_obj_array; 6269 } 6270 6271 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6272 MemRegion span, 6273 CMSBitMap* bitMap, CMSMarkStack* markStack, 6274 bool should_yield, bool verifying): 6275 _collector(collector), 6276 _span(span), 6277 _bitMap(bitMap), 6278 _mut(&collector->_modUnionTable), 6279 _markStack(markStack), 6280 _yield(should_yield), 6281 _skipBits(0) 6282 { 6283 assert(_markStack->isEmpty(), "stack should be empty"); 6284 _finger = _bitMap->startWord(); 6285 _threshold = _finger; 6286 assert(_collector->_restart_addr == NULL, "Sanity check"); 6287 assert(_span.contains(_finger), "Out of bounds _finger?"); 6288 DEBUG_ONLY(_verifying = verifying;) 6289 } 6290 6291 void MarkFromRootsClosure::reset(HeapWord* addr) { 6292 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6293 assert(_span.contains(addr), "Out of bounds _finger?"); 6294 _finger = addr; 6295 _threshold = align_up(_finger, CardTable::card_size); 6296 } 6297 6298 // Should revisit to see if this should be restructured for 6299 // greater efficiency. 6300 bool MarkFromRootsClosure::do_bit(size_t offset) { 6301 if (_skipBits > 0) { 6302 _skipBits--; 6303 return true; 6304 } 6305 // convert offset into a HeapWord* 6306 HeapWord* addr = _bitMap->startWord() + offset; 6307 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6308 "address out of range"); 6309 assert(_bitMap->isMarked(addr), "tautology"); 6310 if (_bitMap->isMarked(addr+1)) { 6311 // this is an allocated but not yet initialized object 6312 assert(_skipBits == 0, "tautology"); 6313 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6314 oop p = oop(addr); 6315 if (p->klass_or_null_acquire() == NULL) { 6316 DEBUG_ONLY(if (!_verifying) {) 6317 // We re-dirty the cards on which this object lies and increase 6318 // the _threshold so that we'll come back to scan this object 6319 // during the preclean or remark phase. (CMSCleanOnEnter) 6320 if (CMSCleanOnEnter) { 6321 size_t sz = _collector->block_size_using_printezis_bits(addr); 6322 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6323 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6324 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6325 // Bump _threshold to end_card_addr; note that 6326 // _threshold cannot possibly exceed end_card_addr, anyhow. 6327 // This prevents future clearing of the card as the scan proceeds 6328 // to the right. 6329 assert(_threshold <= end_card_addr, 6330 "Because we are just scanning into this object"); 6331 if (_threshold < end_card_addr) { 6332 _threshold = end_card_addr; 6333 } 6334 if (p->klass_or_null_acquire() != NULL) { 6335 // Redirty the range of cards... 6336 _mut->mark_range(redirty_range); 6337 } // ...else the setting of klass will dirty the card anyway. 6338 } 6339 DEBUG_ONLY(}) 6340 return true; 6341 } 6342 } 6343 scanOopsInOop(addr); 6344 return true; 6345 } 6346 6347 // We take a break if we've been at this for a while, 6348 // so as to avoid monopolizing the locks involved. 6349 void MarkFromRootsClosure::do_yield_work() { 6350 // First give up the locks, then yield, then re-lock 6351 // We should probably use a constructor/destructor idiom to 6352 // do this unlock/lock or modify the MutexUnlocker class to 6353 // serve our purpose. XXX 6354 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6355 "CMS thread should hold CMS token"); 6356 assert_lock_strong(_bitMap->lock()); 6357 _bitMap->lock()->unlock(); 6358 ConcurrentMarkSweepThread::desynchronize(true); 6359 _collector->stopTimer(); 6360 _collector->incrementYields(); 6361 6362 // See the comment in coordinator_yield() 6363 for (unsigned i = 0; i < CMSYieldSleepCount && 6364 ConcurrentMarkSweepThread::should_yield() && 6365 !CMSCollector::foregroundGCIsActive(); ++i) { 6366 os::sleep(Thread::current(), 1, false); 6367 } 6368 6369 ConcurrentMarkSweepThread::synchronize(true); 6370 _bitMap->lock()->lock_without_safepoint_check(); 6371 _collector->startTimer(); 6372 } 6373 6374 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6375 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6376 assert(_markStack->isEmpty(), 6377 "should drain stack to limit stack usage"); 6378 // convert ptr to an oop preparatory to scanning 6379 oop obj = oop(ptr); 6380 // Ignore mark word in verification below, since we 6381 // may be running concurrent with mutators. 6382 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6383 assert(_finger <= ptr, "_finger runneth ahead"); 6384 // advance the finger to right end of this object 6385 _finger = ptr + obj->size(); 6386 assert(_finger > ptr, "we just incremented it above"); 6387 // On large heaps, it may take us some time to get through 6388 // the marking phase. During 6389 // this time it's possible that a lot of mutations have 6390 // accumulated in the card table and the mod union table -- 6391 // these mutation records are redundant until we have 6392 // actually traced into the corresponding card. 6393 // Here, we check whether advancing the finger would make 6394 // us cross into a new card, and if so clear corresponding 6395 // cards in the MUT (preclean them in the card-table in the 6396 // future). 6397 6398 DEBUG_ONLY(if (!_verifying) {) 6399 // The clean-on-enter optimization is disabled by default, 6400 // until we fix 6178663. 6401 if (CMSCleanOnEnter && (_finger > _threshold)) { 6402 // [_threshold, _finger) represents the interval 6403 // of cards to be cleared in MUT (or precleaned in card table). 6404 // The set of cards to be cleared is all those that overlap 6405 // with the interval [_threshold, _finger); note that 6406 // _threshold is always kept card-aligned but _finger isn't 6407 // always card-aligned. 6408 HeapWord* old_threshold = _threshold; 6409 assert(is_aligned(old_threshold, CardTable::card_size), 6410 "_threshold should always be card-aligned"); 6411 _threshold = align_up(_finger, CardTable::card_size); 6412 MemRegion mr(old_threshold, _threshold); 6413 assert(!mr.is_empty(), "Control point invariant"); 6414 assert(_span.contains(mr), "Should clear within span"); 6415 _mut->clear_range(mr); 6416 } 6417 DEBUG_ONLY(}) 6418 // Note: the finger doesn't advance while we drain 6419 // the stack below. 6420 PushOrMarkClosure pushOrMarkClosure(_collector, 6421 _span, _bitMap, _markStack, 6422 _finger, this); 6423 bool res = _markStack->push(obj); 6424 assert(res, "Empty non-zero size stack should have space for single push"); 6425 while (!_markStack->isEmpty()) { 6426 oop new_oop = _markStack->pop(); 6427 // Skip verifying header mark word below because we are 6428 // running concurrent with mutators. 6429 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6430 // now scan this oop's oops 6431 new_oop->oop_iterate(&pushOrMarkClosure); 6432 do_yield_check(); 6433 } 6434 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6435 } 6436 6437 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6438 CMSCollector* collector, MemRegion span, 6439 CMSBitMap* bit_map, 6440 OopTaskQueue* work_queue, 6441 CMSMarkStack* overflow_stack): 6442 _collector(collector), 6443 _whole_span(collector->_span), 6444 _span(span), 6445 _bit_map(bit_map), 6446 _mut(&collector->_modUnionTable), 6447 _work_queue(work_queue), 6448 _overflow_stack(overflow_stack), 6449 _skip_bits(0), 6450 _task(task) 6451 { 6452 assert(_work_queue->size() == 0, "work_queue should be empty"); 6453 _finger = span.start(); 6454 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6455 assert(_span.contains(_finger), "Out of bounds _finger?"); 6456 } 6457 6458 // Should revisit to see if this should be restructured for 6459 // greater efficiency. 6460 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6461 if (_skip_bits > 0) { 6462 _skip_bits--; 6463 return true; 6464 } 6465 // convert offset into a HeapWord* 6466 HeapWord* addr = _bit_map->startWord() + offset; 6467 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6468 "address out of range"); 6469 assert(_bit_map->isMarked(addr), "tautology"); 6470 if (_bit_map->isMarked(addr+1)) { 6471 // this is an allocated object that might not yet be initialized 6472 assert(_skip_bits == 0, "tautology"); 6473 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6474 oop p = oop(addr); 6475 if (p->klass_or_null_acquire() == NULL) { 6476 // in the case of Clean-on-Enter optimization, redirty card 6477 // and avoid clearing card by increasing the threshold. 6478 return true; 6479 } 6480 } 6481 scan_oops_in_oop(addr); 6482 return true; 6483 } 6484 6485 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6486 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6487 // Should we assert that our work queue is empty or 6488 // below some drain limit? 6489 assert(_work_queue->size() == 0, 6490 "should drain stack to limit stack usage"); 6491 // convert ptr to an oop preparatory to scanning 6492 oop obj = oop(ptr); 6493 // Ignore mark word in verification below, since we 6494 // may be running concurrent with mutators. 6495 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6496 assert(_finger <= ptr, "_finger runneth ahead"); 6497 // advance the finger to right end of this object 6498 _finger = ptr + obj->size(); 6499 assert(_finger > ptr, "we just incremented it above"); 6500 // On large heaps, it may take us some time to get through 6501 // the marking phase. During 6502 // this time it's possible that a lot of mutations have 6503 // accumulated in the card table and the mod union table -- 6504 // these mutation records are redundant until we have 6505 // actually traced into the corresponding card. 6506 // Here, we check whether advancing the finger would make 6507 // us cross into a new card, and if so clear corresponding 6508 // cards in the MUT (preclean them in the card-table in the 6509 // future). 6510 6511 // The clean-on-enter optimization is disabled by default, 6512 // until we fix 6178663. 6513 if (CMSCleanOnEnter && (_finger > _threshold)) { 6514 // [_threshold, _finger) represents the interval 6515 // of cards to be cleared in MUT (or precleaned in card table). 6516 // The set of cards to be cleared is all those that overlap 6517 // with the interval [_threshold, _finger); note that 6518 // _threshold is always kept card-aligned but _finger isn't 6519 // always card-aligned. 6520 HeapWord* old_threshold = _threshold; 6521 assert(is_aligned(old_threshold, CardTable::card_size), 6522 "_threshold should always be card-aligned"); 6523 _threshold = align_up(_finger, CardTable::card_size); 6524 MemRegion mr(old_threshold, _threshold); 6525 assert(!mr.is_empty(), "Control point invariant"); 6526 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6527 _mut->clear_range(mr); 6528 } 6529 6530 // Note: the local finger doesn't advance while we drain 6531 // the stack below, but the global finger sure can and will. 6532 HeapWord* volatile* gfa = _task->global_finger_addr(); 6533 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6534 _span, _bit_map, 6535 _work_queue, 6536 _overflow_stack, 6537 _finger, 6538 gfa, this); 6539 bool res = _work_queue->push(obj); // overflow could occur here 6540 assert(res, "Will hold once we use workqueues"); 6541 while (true) { 6542 oop new_oop; 6543 if (!_work_queue->pop_local(new_oop)) { 6544 // We emptied our work_queue; check if there's stuff that can 6545 // be gotten from the overflow stack. 6546 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6547 _overflow_stack, _work_queue)) { 6548 do_yield_check(); 6549 continue; 6550 } else { // done 6551 break; 6552 } 6553 } 6554 // Skip verifying header mark word below because we are 6555 // running concurrent with mutators. 6556 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6557 // now scan this oop's oops 6558 new_oop->oop_iterate(&pushOrMarkClosure); 6559 do_yield_check(); 6560 } 6561 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6562 } 6563 6564 // Yield in response to a request from VM Thread or 6565 // from mutators. 6566 void ParMarkFromRootsClosure::do_yield_work() { 6567 assert(_task != NULL, "sanity"); 6568 _task->yield(); 6569 } 6570 6571 // A variant of the above used for verifying CMS marking work. 6572 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6573 MemRegion span, 6574 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6575 CMSMarkStack* mark_stack): 6576 _collector(collector), 6577 _span(span), 6578 _verification_bm(verification_bm), 6579 _cms_bm(cms_bm), 6580 _mark_stack(mark_stack), 6581 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6582 mark_stack) 6583 { 6584 assert(_mark_stack->isEmpty(), "stack should be empty"); 6585 _finger = _verification_bm->startWord(); 6586 assert(_collector->_restart_addr == NULL, "Sanity check"); 6587 assert(_span.contains(_finger), "Out of bounds _finger?"); 6588 } 6589 6590 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6591 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6592 assert(_span.contains(addr), "Out of bounds _finger?"); 6593 _finger = addr; 6594 } 6595 6596 // Should revisit to see if this should be restructured for 6597 // greater efficiency. 6598 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6599 // convert offset into a HeapWord* 6600 HeapWord* addr = _verification_bm->startWord() + offset; 6601 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6602 "address out of range"); 6603 assert(_verification_bm->isMarked(addr), "tautology"); 6604 assert(_cms_bm->isMarked(addr), "tautology"); 6605 6606 assert(_mark_stack->isEmpty(), 6607 "should drain stack to limit stack usage"); 6608 // convert addr to an oop preparatory to scanning 6609 oop obj = oop(addr); 6610 assert(oopDesc::is_oop(obj), "should be an oop"); 6611 assert(_finger <= addr, "_finger runneth ahead"); 6612 // advance the finger to right end of this object 6613 _finger = addr + obj->size(); 6614 assert(_finger > addr, "we just incremented it above"); 6615 // Note: the finger doesn't advance while we drain 6616 // the stack below. 6617 bool res = _mark_stack->push(obj); 6618 assert(res, "Empty non-zero size stack should have space for single push"); 6619 while (!_mark_stack->isEmpty()) { 6620 oop new_oop = _mark_stack->pop(); 6621 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6622 // now scan this oop's oops 6623 new_oop->oop_iterate(&_pam_verify_closure); 6624 } 6625 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6626 return true; 6627 } 6628 6629 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6630 CMSCollector* collector, MemRegion span, 6631 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6632 CMSMarkStack* mark_stack): 6633 MetadataAwareOopClosure(collector->ref_processor()), 6634 _collector(collector), 6635 _span(span), 6636 _verification_bm(verification_bm), 6637 _cms_bm(cms_bm), 6638 _mark_stack(mark_stack) 6639 { } 6640 6641 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6642 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6643 6644 // Upon stack overflow, we discard (part of) the stack, 6645 // remembering the least address amongst those discarded 6646 // in CMSCollector's _restart_address. 6647 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6648 // Remember the least grey address discarded 6649 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6650 _collector->lower_restart_addr(ra); 6651 _mark_stack->reset(); // discard stack contents 6652 _mark_stack->expand(); // expand the stack if possible 6653 } 6654 6655 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6656 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6657 HeapWord* addr = (HeapWord*)obj; 6658 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6659 // Oop lies in _span and isn't yet grey or black 6660 _verification_bm->mark(addr); // now grey 6661 if (!_cms_bm->isMarked(addr)) { 6662 Log(gc, verify) log; 6663 ResourceMark rm; 6664 LogStream ls(log.error()); 6665 oop(addr)->print_on(&ls); 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* volatile* 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(oopDesc::is_oop_or_null(obj, 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(oopDesc::is_oop_or_null(obj, 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(oopDesc::is_oop_or_null(obj, 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 = align_up(addr + sz, CardTable::card_size); 6892 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6893 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6894 _mod_union_table->mark_range(redirty_range); 6895 } else { 6896 _mod_union_table->mark(addr); 6897 } 6898 _collector->_ser_pmc_preclean_ovflw++; 6899 } else { 6900 // During the remark phase, we need to remember this oop 6901 // in the overflow list. 6902 _collector->push_on_overflow_list(obj); 6903 _collector->_ser_pmc_remark_ovflw++; 6904 } 6905 } 6906 } 6907 } 6908 6909 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6910 MemRegion span, 6911 ReferenceProcessor* rp, 6912 CMSBitMap* bit_map, 6913 OopTaskQueue* work_queue): 6914 MetadataAwareOopClosure(rp), 6915 _collector(collector), 6916 _span(span), 6917 _bit_map(bit_map), 6918 _work_queue(work_queue) 6919 { 6920 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6921 } 6922 6923 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6924 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6925 6926 // Grey object rescan during second checkpoint phase -- 6927 // the parallel version. 6928 void ParPushAndMarkClosure::do_oop(oop obj) { 6929 // In the assert below, we ignore the mark word because 6930 // this oop may point to an already visited object that is 6931 // on the overflow stack (in which case the mark word has 6932 // been hijacked for chaining into the overflow stack -- 6933 // if this is the last object in the overflow stack then 6934 // its mark word will be NULL). Because this object may 6935 // have been subsequently popped off the global overflow 6936 // stack, and the mark word possibly restored to the prototypical 6937 // value, by the time we get to examined this failing assert in 6938 // the debugger, is_oop_or_null(false) may subsequently start 6939 // to hold. 6940 assert(oopDesc::is_oop_or_null(obj, true), 6941 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6942 HeapWord* addr = (HeapWord*)obj; 6943 // Check if oop points into the CMS generation 6944 // and is not marked 6945 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6946 // a white object ... 6947 // If we manage to "claim" the object, by being the 6948 // first thread to mark it, then we push it on our 6949 // marking stack 6950 if (_bit_map->par_mark(addr)) { // ... now grey 6951 // push on work queue (grey set) 6952 bool simulate_overflow = false; 6953 NOT_PRODUCT( 6954 if (CMSMarkStackOverflowALot && 6955 _collector->par_simulate_overflow()) { 6956 // simulate a stack overflow 6957 simulate_overflow = true; 6958 } 6959 ) 6960 if (simulate_overflow || !_work_queue->push(obj)) { 6961 _collector->par_push_on_overflow_list(obj); 6962 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6963 } 6964 } // Else, some other thread got there first 6965 } 6966 } 6967 6968 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6969 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6970 6971 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6972 Mutex* bml = _collector->bitMapLock(); 6973 assert_lock_strong(bml); 6974 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6975 "CMS thread should hold CMS token"); 6976 6977 bml->unlock(); 6978 ConcurrentMarkSweepThread::desynchronize(true); 6979 6980 _collector->stopTimer(); 6981 _collector->incrementYields(); 6982 6983 // See the comment in coordinator_yield() 6984 for (unsigned i = 0; i < CMSYieldSleepCount && 6985 ConcurrentMarkSweepThread::should_yield() && 6986 !CMSCollector::foregroundGCIsActive(); ++i) { 6987 os::sleep(Thread::current(), 1, false); 6988 } 6989 6990 ConcurrentMarkSweepThread::synchronize(true); 6991 bml->lock(); 6992 6993 _collector->startTimer(); 6994 } 6995 6996 bool CMSPrecleanRefsYieldClosure::should_return() { 6997 if (ConcurrentMarkSweepThread::should_yield()) { 6998 do_yield_work(); 6999 } 7000 return _collector->foregroundGCIsActive(); 7001 } 7002 7003 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7004 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7005 "mr should be aligned to start at a card boundary"); 7006 // We'd like to assert: 7007 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7008 // "mr should be a range of cards"); 7009 // However, that would be too strong in one case -- the last 7010 // partition ends at _unallocated_block which, in general, can be 7011 // an arbitrary boundary, not necessarily card aligned. 7012 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7013 _space->object_iterate_mem(mr, &_scan_cl); 7014 } 7015 7016 SweepClosure::SweepClosure(CMSCollector* collector, 7017 ConcurrentMarkSweepGeneration* g, 7018 CMSBitMap* bitMap, bool should_yield) : 7019 _collector(collector), 7020 _g(g), 7021 _sp(g->cmsSpace()), 7022 _limit(_sp->sweep_limit()), 7023 _freelistLock(_sp->freelistLock()), 7024 _bitMap(bitMap), 7025 _yield(should_yield), 7026 _inFreeRange(false), // No free range at beginning of sweep 7027 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7028 _lastFreeRangeCoalesced(false), 7029 _freeFinger(g->used_region().start()) 7030 { 7031 NOT_PRODUCT( 7032 _numObjectsFreed = 0; 7033 _numWordsFreed = 0; 7034 _numObjectsLive = 0; 7035 _numWordsLive = 0; 7036 _numObjectsAlreadyFree = 0; 7037 _numWordsAlreadyFree = 0; 7038 _last_fc = NULL; 7039 7040 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7041 _sp->dictionary()->initialize_dict_returned_bytes(); 7042 ) 7043 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7044 "sweep _limit out of bounds"); 7045 log_develop_trace(gc, sweep)("===================="); 7046 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7047 } 7048 7049 void SweepClosure::print_on(outputStream* st) const { 7050 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7051 p2i(_sp->bottom()), p2i(_sp->end())); 7052 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7053 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7054 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7055 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7056 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7057 } 7058 7059 #ifndef PRODUCT 7060 // Assertion checking only: no useful work in product mode -- 7061 // however, if any of the flags below become product flags, 7062 // you may need to review this code to see if it needs to be 7063 // enabled in product mode. 7064 SweepClosure::~SweepClosure() { 7065 assert_lock_strong(_freelistLock); 7066 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7067 "sweep _limit out of bounds"); 7068 if (inFreeRange()) { 7069 Log(gc, sweep) log; 7070 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7071 ResourceMark rm; 7072 LogStream ls(log.error()); 7073 print_on(&ls); 7074 ShouldNotReachHere(); 7075 } 7076 7077 if (log_is_enabled(Debug, gc, sweep)) { 7078 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7079 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7080 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7081 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7082 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7083 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7084 } 7085 7086 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7087 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7088 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7089 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7090 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7091 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7092 } 7093 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7094 log_develop_trace(gc, sweep)("================"); 7095 } 7096 #endif // PRODUCT 7097 7098 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7099 bool freeRangeInFreeLists) { 7100 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7101 p2i(freeFinger), freeRangeInFreeLists); 7102 assert(!inFreeRange(), "Trampling existing free range"); 7103 set_inFreeRange(true); 7104 set_lastFreeRangeCoalesced(false); 7105 7106 set_freeFinger(freeFinger); 7107 set_freeRangeInFreeLists(freeRangeInFreeLists); 7108 if (CMSTestInFreeList) { 7109 if (freeRangeInFreeLists) { 7110 FreeChunk* fc = (FreeChunk*) freeFinger; 7111 assert(fc->is_free(), "A chunk on the free list should be free."); 7112 assert(fc->size() > 0, "Free range should have a size"); 7113 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7114 } 7115 } 7116 } 7117 7118 // Note that the sweeper runs concurrently with mutators. Thus, 7119 // it is possible for direct allocation in this generation to happen 7120 // in the middle of the sweep. Note that the sweeper also coalesces 7121 // contiguous free blocks. Thus, unless the sweeper and the allocator 7122 // synchronize appropriately freshly allocated blocks may get swept up. 7123 // This is accomplished by the sweeper locking the free lists while 7124 // it is sweeping. Thus blocks that are determined to be free are 7125 // indeed free. There is however one additional complication: 7126 // blocks that have been allocated since the final checkpoint and 7127 // mark, will not have been marked and so would be treated as 7128 // unreachable and swept up. To prevent this, the allocator marks 7129 // the bit map when allocating during the sweep phase. This leads, 7130 // however, to a further complication -- objects may have been allocated 7131 // but not yet initialized -- in the sense that the header isn't yet 7132 // installed. The sweeper can not then determine the size of the block 7133 // in order to skip over it. To deal with this case, we use a technique 7134 // (due to Printezis) to encode such uninitialized block sizes in the 7135 // bit map. Since the bit map uses a bit per every HeapWord, but the 7136 // CMS generation has a minimum object size of 3 HeapWords, it follows 7137 // that "normal marks" won't be adjacent in the bit map (there will 7138 // always be at least two 0 bits between successive 1 bits). We make use 7139 // of these "unused" bits to represent uninitialized blocks -- the bit 7140 // corresponding to the start of the uninitialized object and the next 7141 // bit are both set. Finally, a 1 bit marks the end of the object that 7142 // started with the two consecutive 1 bits to indicate its potentially 7143 // uninitialized state. 7144 7145 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7146 FreeChunk* fc = (FreeChunk*)addr; 7147 size_t res; 7148 7149 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7150 // than "addr == _limit" because although _limit was a block boundary when 7151 // we started the sweep, it may no longer be one because heap expansion 7152 // may have caused us to coalesce the block ending at the address _limit 7153 // with a newly expanded chunk (this happens when _limit was set to the 7154 // previous _end of the space), so we may have stepped past _limit: 7155 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7156 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7157 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7158 "sweep _limit out of bounds"); 7159 assert(addr < _sp->end(), "addr out of bounds"); 7160 // Flush any free range we might be holding as a single 7161 // coalesced chunk to the appropriate free list. 7162 if (inFreeRange()) { 7163 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7164 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7165 flush_cur_free_chunk(freeFinger(), 7166 pointer_delta(addr, freeFinger())); 7167 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7168 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7169 lastFreeRangeCoalesced() ? 1 : 0); 7170 } 7171 7172 // help the iterator loop finish 7173 return pointer_delta(_sp->end(), addr); 7174 } 7175 7176 assert(addr < _limit, "sweep invariant"); 7177 // check if we should yield 7178 do_yield_check(addr); 7179 if (fc->is_free()) { 7180 // Chunk that is already free 7181 res = fc->size(); 7182 do_already_free_chunk(fc); 7183 debug_only(_sp->verifyFreeLists()); 7184 // If we flush the chunk at hand in lookahead_and_flush() 7185 // and it's coalesced with a preceding chunk, then the 7186 // process of "mangling" the payload of the coalesced block 7187 // will cause erasure of the size information from the 7188 // (erstwhile) header of all the coalesced blocks but the 7189 // first, so the first disjunct in the assert will not hold 7190 // in that specific case (in which case the second disjunct 7191 // will hold). 7192 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7193 "Otherwise the size info doesn't change at this step"); 7194 NOT_PRODUCT( 7195 _numObjectsAlreadyFree++; 7196 _numWordsAlreadyFree += res; 7197 ) 7198 NOT_PRODUCT(_last_fc = fc;) 7199 } else if (!_bitMap->isMarked(addr)) { 7200 // Chunk is fresh garbage 7201 res = do_garbage_chunk(fc); 7202 debug_only(_sp->verifyFreeLists()); 7203 NOT_PRODUCT( 7204 _numObjectsFreed++; 7205 _numWordsFreed += res; 7206 ) 7207 } else { 7208 // Chunk that is alive. 7209 res = do_live_chunk(fc); 7210 debug_only(_sp->verifyFreeLists()); 7211 NOT_PRODUCT( 7212 _numObjectsLive++; 7213 _numWordsLive += res; 7214 ) 7215 } 7216 return res; 7217 } 7218 7219 // For the smart allocation, record following 7220 // split deaths - a free chunk is removed from its free list because 7221 // it is being split into two or more chunks. 7222 // split birth - a free chunk is being added to its free list because 7223 // a larger free chunk has been split and resulted in this free chunk. 7224 // coal death - a free chunk is being removed from its free list because 7225 // it is being coalesced into a large free chunk. 7226 // coal birth - a free chunk is being added to its free list because 7227 // it was created when two or more free chunks where coalesced into 7228 // this free chunk. 7229 // 7230 // These statistics are used to determine the desired number of free 7231 // chunks of a given size. The desired number is chosen to be relative 7232 // to the end of a CMS sweep. The desired number at the end of a sweep 7233 // is the 7234 // count-at-end-of-previous-sweep (an amount that was enough) 7235 // - count-at-beginning-of-current-sweep (the excess) 7236 // + split-births (gains in this size during interval) 7237 // - split-deaths (demands on this size during interval) 7238 // where the interval is from the end of one sweep to the end of the 7239 // next. 7240 // 7241 // When sweeping the sweeper maintains an accumulated chunk which is 7242 // the chunk that is made up of chunks that have been coalesced. That 7243 // will be termed the left-hand chunk. A new chunk of garbage that 7244 // is being considered for coalescing will be referred to as the 7245 // right-hand chunk. 7246 // 7247 // When making a decision on whether to coalesce a right-hand chunk with 7248 // the current left-hand chunk, the current count vs. the desired count 7249 // of the left-hand chunk is considered. Also if the right-hand chunk 7250 // is near the large chunk at the end of the heap (see 7251 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7252 // left-hand chunk is coalesced. 7253 // 7254 // When making a decision about whether to split a chunk, the desired count 7255 // vs. the current count of the candidate to be split is also considered. 7256 // If the candidate is underpopulated (currently fewer chunks than desired) 7257 // a chunk of an overpopulated (currently more chunks than desired) size may 7258 // be chosen. The "hint" associated with a free list, if non-null, points 7259 // to a free list which may be overpopulated. 7260 // 7261 7262 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7263 const size_t size = fc->size(); 7264 // Chunks that cannot be coalesced are not in the 7265 // free lists. 7266 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7267 assert(_sp->verify_chunk_in_free_list(fc), 7268 "free chunk should be in free lists"); 7269 } 7270 // a chunk that is already free, should not have been 7271 // marked in the bit map 7272 HeapWord* const addr = (HeapWord*) fc; 7273 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7274 // Verify that the bit map has no bits marked between 7275 // addr and purported end of this block. 7276 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7277 7278 // Some chunks cannot be coalesced under any circumstances. 7279 // See the definition of cantCoalesce(). 7280 if (!fc->cantCoalesce()) { 7281 // This chunk can potentially be coalesced. 7282 // All the work is done in 7283 do_post_free_or_garbage_chunk(fc, size); 7284 // Note that if the chunk is not coalescable (the else arm 7285 // below), we unconditionally flush, without needing to do 7286 // a "lookahead," as we do below. 7287 if (inFreeRange()) lookahead_and_flush(fc, size); 7288 } else { 7289 // Code path common to both original and adaptive free lists. 7290 7291 // cant coalesce with previous block; this should be treated 7292 // as the end of a free run if any 7293 if (inFreeRange()) { 7294 // we kicked some butt; time to pick up the garbage 7295 assert(freeFinger() < addr, "freeFinger points too high"); 7296 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7297 } 7298 // else, nothing to do, just continue 7299 } 7300 } 7301 7302 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7303 // This is a chunk of garbage. It is not in any free list. 7304 // Add it to a free list or let it possibly be coalesced into 7305 // a larger chunk. 7306 HeapWord* const addr = (HeapWord*) fc; 7307 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7308 7309 // Verify that the bit map has no bits marked between 7310 // addr and purported end of just dead object. 7311 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7312 do_post_free_or_garbage_chunk(fc, size); 7313 7314 assert(_limit >= addr + size, 7315 "A freshly garbage chunk can't possibly straddle over _limit"); 7316 if (inFreeRange()) lookahead_and_flush(fc, size); 7317 return size; 7318 } 7319 7320 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7321 HeapWord* addr = (HeapWord*) fc; 7322 // The sweeper has just found a live object. Return any accumulated 7323 // left hand chunk to the free lists. 7324 if (inFreeRange()) { 7325 assert(freeFinger() < addr, "freeFinger points too high"); 7326 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7327 } 7328 7329 // This object is live: we'd normally expect this to be 7330 // an oop, and like to assert the following: 7331 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7332 // However, as we commented above, this may be an object whose 7333 // header hasn't yet been initialized. 7334 size_t size; 7335 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7336 if (_bitMap->isMarked(addr + 1)) { 7337 // Determine the size from the bit map, rather than trying to 7338 // compute it from the object header. 7339 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7340 size = pointer_delta(nextOneAddr + 1, addr); 7341 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7342 "alignment problem"); 7343 7344 #ifdef ASSERT 7345 if (oop(addr)->klass_or_null_acquire() != NULL) { 7346 // Ignore mark word because we are running concurrent with mutators 7347 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7348 assert(size == 7349 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7350 "P-mark and computed size do not agree"); 7351 } 7352 #endif 7353 7354 } else { 7355 // This should be an initialized object that's alive. 7356 assert(oop(addr)->klass_or_null_acquire() != NULL, 7357 "Should be an initialized object"); 7358 // Ignore mark word because we are running concurrent with mutators 7359 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7360 // Verify that the bit map has no bits marked between 7361 // addr and purported end of this block. 7362 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7363 assert(size >= 3, "Necessary for Printezis marks to work"); 7364 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7365 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7366 } 7367 return size; 7368 } 7369 7370 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7371 size_t chunkSize) { 7372 // do_post_free_or_garbage_chunk() should only be called in the case 7373 // of the adaptive free list allocator. 7374 const bool fcInFreeLists = fc->is_free(); 7375 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7376 if (CMSTestInFreeList && fcInFreeLists) { 7377 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7378 } 7379 7380 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7381 7382 HeapWord* const fc_addr = (HeapWord*) fc; 7383 7384 bool coalesce = false; 7385 const size_t left = pointer_delta(fc_addr, freeFinger()); 7386 const size_t right = chunkSize; 7387 switch (FLSCoalescePolicy) { 7388 // numeric value forms a coalition aggressiveness metric 7389 case 0: { // never coalesce 7390 coalesce = false; 7391 break; 7392 } 7393 case 1: { // coalesce if left & right chunks on overpopulated lists 7394 coalesce = _sp->coalOverPopulated(left) && 7395 _sp->coalOverPopulated(right); 7396 break; 7397 } 7398 case 2: { // coalesce if left chunk on overpopulated list (default) 7399 coalesce = _sp->coalOverPopulated(left); 7400 break; 7401 } 7402 case 3: { // coalesce if left OR right chunk on overpopulated list 7403 coalesce = _sp->coalOverPopulated(left) || 7404 _sp->coalOverPopulated(right); 7405 break; 7406 } 7407 case 4: { // always coalesce 7408 coalesce = true; 7409 break; 7410 } 7411 default: 7412 ShouldNotReachHere(); 7413 } 7414 7415 // Should the current free range be coalesced? 7416 // If the chunk is in a free range and either we decided to coalesce above 7417 // or the chunk is near the large block at the end of the heap 7418 // (isNearLargestChunk() returns true), then coalesce this chunk. 7419 const bool doCoalesce = inFreeRange() 7420 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7421 if (doCoalesce) { 7422 // Coalesce the current free range on the left with the new 7423 // chunk on the right. If either is on a free list, 7424 // it must be removed from the list and stashed in the closure. 7425 if (freeRangeInFreeLists()) { 7426 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7427 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7428 "Size of free range is inconsistent with chunk size."); 7429 if (CMSTestInFreeList) { 7430 assert(_sp->verify_chunk_in_free_list(ffc), 7431 "Chunk is not in free lists"); 7432 } 7433 _sp->coalDeath(ffc->size()); 7434 _sp->removeFreeChunkFromFreeLists(ffc); 7435 set_freeRangeInFreeLists(false); 7436 } 7437 if (fcInFreeLists) { 7438 _sp->coalDeath(chunkSize); 7439 assert(fc->size() == chunkSize, 7440 "The chunk has the wrong size or is not in the free lists"); 7441 _sp->removeFreeChunkFromFreeLists(fc); 7442 } 7443 set_lastFreeRangeCoalesced(true); 7444 print_free_block_coalesced(fc); 7445 } else { // not in a free range and/or should not coalesce 7446 // Return the current free range and start a new one. 7447 if (inFreeRange()) { 7448 // In a free range but cannot coalesce with the right hand chunk. 7449 // Put the current free range into the free lists. 7450 flush_cur_free_chunk(freeFinger(), 7451 pointer_delta(fc_addr, freeFinger())); 7452 } 7453 // Set up for new free range. Pass along whether the right hand 7454 // chunk is in the free lists. 7455 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7456 } 7457 } 7458 7459 // Lookahead flush: 7460 // If we are tracking a free range, and this is the last chunk that 7461 // we'll look at because its end crosses past _limit, we'll preemptively 7462 // flush it along with any free range we may be holding on to. Note that 7463 // this can be the case only for an already free or freshly garbage 7464 // chunk. If this block is an object, it can never straddle 7465 // over _limit. The "straddling" occurs when _limit is set at 7466 // the previous end of the space when this cycle started, and 7467 // a subsequent heap expansion caused the previously co-terminal 7468 // free block to be coalesced with the newly expanded portion, 7469 // thus rendering _limit a non-block-boundary making it dangerous 7470 // for the sweeper to step over and examine. 7471 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7472 assert(inFreeRange(), "Should only be called if currently in a free range."); 7473 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7474 assert(_sp->used_region().contains(eob - 1), 7475 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7476 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7477 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7478 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7479 if (eob >= _limit) { 7480 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7481 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7482 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7483 "[" PTR_FORMAT "," PTR_FORMAT ")", 7484 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7485 // Return the storage we are tracking back into the free lists. 7486 log_develop_trace(gc, sweep)("Flushing ... "); 7487 assert(freeFinger() < eob, "Error"); 7488 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7489 } 7490 } 7491 7492 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7493 assert(inFreeRange(), "Should only be called if currently in a free range."); 7494 assert(size > 0, 7495 "A zero sized chunk cannot be added to the free lists."); 7496 if (!freeRangeInFreeLists()) { 7497 if (CMSTestInFreeList) { 7498 FreeChunk* fc = (FreeChunk*) chunk; 7499 fc->set_size(size); 7500 assert(!_sp->verify_chunk_in_free_list(fc), 7501 "chunk should not be in free lists yet"); 7502 } 7503 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7504 // A new free range is going to be starting. The current 7505 // free range has not been added to the free lists yet or 7506 // was removed so add it back. 7507 // If the current free range was coalesced, then the death 7508 // of the free range was recorded. Record a birth now. 7509 if (lastFreeRangeCoalesced()) { 7510 _sp->coalBirth(size); 7511 } 7512 _sp->addChunkAndRepairOffsetTable(chunk, size, 7513 lastFreeRangeCoalesced()); 7514 } else { 7515 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7516 } 7517 set_inFreeRange(false); 7518 set_freeRangeInFreeLists(false); 7519 } 7520 7521 // We take a break if we've been at this for a while, 7522 // so as to avoid monopolizing the locks involved. 7523 void SweepClosure::do_yield_work(HeapWord* addr) { 7524 // Return current free chunk being used for coalescing (if any) 7525 // to the appropriate freelist. After yielding, the next 7526 // free block encountered will start a coalescing range of 7527 // free blocks. If the next free block is adjacent to the 7528 // chunk just flushed, they will need to wait for the next 7529 // sweep to be coalesced. 7530 if (inFreeRange()) { 7531 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7532 } 7533 7534 // First give up the locks, then yield, then re-lock. 7535 // We should probably use a constructor/destructor idiom to 7536 // do this unlock/lock or modify the MutexUnlocker class to 7537 // serve our purpose. XXX 7538 assert_lock_strong(_bitMap->lock()); 7539 assert_lock_strong(_freelistLock); 7540 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7541 "CMS thread should hold CMS token"); 7542 _bitMap->lock()->unlock(); 7543 _freelistLock->unlock(); 7544 ConcurrentMarkSweepThread::desynchronize(true); 7545 _collector->stopTimer(); 7546 _collector->incrementYields(); 7547 7548 // See the comment in coordinator_yield() 7549 for (unsigned i = 0; i < CMSYieldSleepCount && 7550 ConcurrentMarkSweepThread::should_yield() && 7551 !CMSCollector::foregroundGCIsActive(); ++i) { 7552 os::sleep(Thread::current(), 1, false); 7553 } 7554 7555 ConcurrentMarkSweepThread::synchronize(true); 7556 _freelistLock->lock(); 7557 _bitMap->lock()->lock_without_safepoint_check(); 7558 _collector->startTimer(); 7559 } 7560 7561 #ifndef PRODUCT 7562 // This is actually very useful in a product build if it can 7563 // be called from the debugger. Compile it into the product 7564 // as needed. 7565 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7566 return debug_cms_space->verify_chunk_in_free_list(fc); 7567 } 7568 #endif 7569 7570 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7571 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7572 p2i(fc), fc->size()); 7573 } 7574 7575 // CMSIsAliveClosure 7576 bool CMSIsAliveClosure::do_object_b(oop obj) { 7577 HeapWord* addr = (HeapWord*)obj; 7578 return addr != NULL && 7579 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7580 } 7581 7582 7583 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7584 MemRegion span, 7585 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7586 bool cpc): 7587 _collector(collector), 7588 _span(span), 7589 _bit_map(bit_map), 7590 _mark_stack(mark_stack), 7591 _concurrent_precleaning(cpc) { 7592 assert(!_span.is_empty(), "Empty span could spell trouble"); 7593 } 7594 7595 7596 // CMSKeepAliveClosure: the serial version 7597 void CMSKeepAliveClosure::do_oop(oop obj) { 7598 HeapWord* addr = (HeapWord*)obj; 7599 if (_span.contains(addr) && 7600 !_bit_map->isMarked(addr)) { 7601 _bit_map->mark(addr); 7602 bool simulate_overflow = false; 7603 NOT_PRODUCT( 7604 if (CMSMarkStackOverflowALot && 7605 _collector->simulate_overflow()) { 7606 // simulate a stack overflow 7607 simulate_overflow = true; 7608 } 7609 ) 7610 if (simulate_overflow || !_mark_stack->push(obj)) { 7611 if (_concurrent_precleaning) { 7612 // We dirty the overflown object and let the remark 7613 // phase deal with it. 7614 assert(_collector->overflow_list_is_empty(), "Error"); 7615 // In the case of object arrays, we need to dirty all of 7616 // the cards that the object spans. No locking or atomics 7617 // are needed since no one else can be mutating the mod union 7618 // table. 7619 if (obj->is_objArray()) { 7620 size_t sz = obj->size(); 7621 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7622 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7623 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7624 _collector->_modUnionTable.mark_range(redirty_range); 7625 } else { 7626 _collector->_modUnionTable.mark(addr); 7627 } 7628 _collector->_ser_kac_preclean_ovflw++; 7629 } else { 7630 _collector->push_on_overflow_list(obj); 7631 _collector->_ser_kac_ovflw++; 7632 } 7633 } 7634 } 7635 } 7636 7637 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7638 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7639 7640 // CMSParKeepAliveClosure: a parallel version of the above. 7641 // The work queues are private to each closure (thread), 7642 // but (may be) available for stealing by other threads. 7643 void CMSParKeepAliveClosure::do_oop(oop obj) { 7644 HeapWord* addr = (HeapWord*)obj; 7645 if (_span.contains(addr) && 7646 !_bit_map->isMarked(addr)) { 7647 // In general, during recursive tracing, several threads 7648 // may be concurrently getting here; the first one to 7649 // "tag" it, claims it. 7650 if (_bit_map->par_mark(addr)) { 7651 bool res = _work_queue->push(obj); 7652 assert(res, "Low water mark should be much less than capacity"); 7653 // Do a recursive trim in the hope that this will keep 7654 // stack usage lower, but leave some oops for potential stealers 7655 trim_queue(_low_water_mark); 7656 } // Else, another thread got there first 7657 } 7658 } 7659 7660 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7661 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7662 7663 void CMSParKeepAliveClosure::trim_queue(uint max) { 7664 while (_work_queue->size() > max) { 7665 oop new_oop; 7666 if (_work_queue->pop_local(new_oop)) { 7667 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7668 assert(_bit_map->isMarked((HeapWord*)new_oop), 7669 "no white objects on this stack!"); 7670 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7671 // iterate over the oops in this oop, marking and pushing 7672 // the ones in CMS heap (i.e. in _span). 7673 new_oop->oop_iterate(&_mark_and_push); 7674 } 7675 } 7676 } 7677 7678 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7679 CMSCollector* collector, 7680 MemRegion span, CMSBitMap* bit_map, 7681 OopTaskQueue* work_queue): 7682 _collector(collector), 7683 _span(span), 7684 _bit_map(bit_map), 7685 _work_queue(work_queue) { } 7686 7687 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7688 HeapWord* addr = (HeapWord*)obj; 7689 if (_span.contains(addr) && 7690 !_bit_map->isMarked(addr)) { 7691 if (_bit_map->par_mark(addr)) { 7692 bool simulate_overflow = false; 7693 NOT_PRODUCT( 7694 if (CMSMarkStackOverflowALot && 7695 _collector->par_simulate_overflow()) { 7696 // simulate a stack overflow 7697 simulate_overflow = true; 7698 } 7699 ) 7700 if (simulate_overflow || !_work_queue->push(obj)) { 7701 _collector->par_push_on_overflow_list(obj); 7702 _collector->_par_kac_ovflw++; 7703 } 7704 } // Else another thread got there already 7705 } 7706 } 7707 7708 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7709 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7710 7711 ////////////////////////////////////////////////////////////////// 7712 // CMSExpansionCause ///////////////////////////// 7713 ////////////////////////////////////////////////////////////////// 7714 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7715 switch (cause) { 7716 case _no_expansion: 7717 return "No expansion"; 7718 case _satisfy_free_ratio: 7719 return "Free ratio"; 7720 case _satisfy_promotion: 7721 return "Satisfy promotion"; 7722 case _satisfy_allocation: 7723 return "allocation"; 7724 case _allocate_par_lab: 7725 return "Par LAB"; 7726 case _allocate_par_spooling_space: 7727 return "Par Spooling Space"; 7728 case _adaptive_size_policy: 7729 return "Ergonomics"; 7730 default: 7731 return "unknown"; 7732 } 7733 } 7734 7735 void CMSDrainMarkingStackClosure::do_void() { 7736 // the max number to take from overflow list at a time 7737 const size_t num = _mark_stack->capacity()/4; 7738 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7739 "Overflow list should be NULL during concurrent phases"); 7740 while (!_mark_stack->isEmpty() || 7741 // if stack is empty, check the overflow list 7742 _collector->take_from_overflow_list(num, _mark_stack)) { 7743 oop obj = _mark_stack->pop(); 7744 HeapWord* addr = (HeapWord*)obj; 7745 assert(_span.contains(addr), "Should be within span"); 7746 assert(_bit_map->isMarked(addr), "Should be marked"); 7747 assert(oopDesc::is_oop(obj), "Should be an oop"); 7748 obj->oop_iterate(_keep_alive); 7749 } 7750 } 7751 7752 void CMSParDrainMarkingStackClosure::do_void() { 7753 // drain queue 7754 trim_queue(0); 7755 } 7756 7757 // Trim our work_queue so its length is below max at return 7758 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7759 while (_work_queue->size() > max) { 7760 oop new_oop; 7761 if (_work_queue->pop_local(new_oop)) { 7762 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7763 assert(_bit_map->isMarked((HeapWord*)new_oop), 7764 "no white objects on this stack!"); 7765 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7766 // iterate over the oops in this oop, marking and pushing 7767 // the ones in CMS heap (i.e. in _span). 7768 new_oop->oop_iterate(&_mark_and_push); 7769 } 7770 } 7771 } 7772 7773 //////////////////////////////////////////////////////////////////// 7774 // Support for Marking Stack Overflow list handling and related code 7775 //////////////////////////////////////////////////////////////////// 7776 // Much of the following code is similar in shape and spirit to the 7777 // code used in ParNewGC. We should try and share that code 7778 // as much as possible in the future. 7779 7780 #ifndef PRODUCT 7781 // Debugging support for CMSStackOverflowALot 7782 7783 // It's OK to call this multi-threaded; the worst thing 7784 // that can happen is that we'll get a bunch of closely 7785 // spaced simulated overflows, but that's OK, in fact 7786 // probably good as it would exercise the overflow code 7787 // under contention. 7788 bool CMSCollector::simulate_overflow() { 7789 if (_overflow_counter-- <= 0) { // just being defensive 7790 _overflow_counter = CMSMarkStackOverflowInterval; 7791 return true; 7792 } else { 7793 return false; 7794 } 7795 } 7796 7797 bool CMSCollector::par_simulate_overflow() { 7798 return simulate_overflow(); 7799 } 7800 #endif 7801 7802 // Single-threaded 7803 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7804 assert(stack->isEmpty(), "Expected precondition"); 7805 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7806 size_t i = num; 7807 oop cur = _overflow_list; 7808 const markOop proto = markOopDesc::prototype(); 7809 NOT_PRODUCT(ssize_t n = 0;) 7810 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7811 next = oop(cur->mark()); 7812 cur->set_mark(proto); // until proven otherwise 7813 assert(oopDesc::is_oop(cur), "Should be an oop"); 7814 bool res = stack->push(cur); 7815 assert(res, "Bit off more than can chew?"); 7816 NOT_PRODUCT(n++;) 7817 } 7818 _overflow_list = cur; 7819 #ifndef PRODUCT 7820 assert(_num_par_pushes >= n, "Too many pops?"); 7821 _num_par_pushes -=n; 7822 #endif 7823 return !stack->isEmpty(); 7824 } 7825 7826 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7827 // (MT-safe) Get a prefix of at most "num" from the list. 7828 // The overflow list is chained through the mark word of 7829 // each object in the list. We fetch the entire list, 7830 // break off a prefix of the right size and return the 7831 // remainder. If other threads try to take objects from 7832 // the overflow list at that time, they will wait for 7833 // some time to see if data becomes available. If (and 7834 // only if) another thread places one or more object(s) 7835 // on the global list before we have returned the suffix 7836 // to the global list, we will walk down our local list 7837 // to find its end and append the global list to 7838 // our suffix before returning it. This suffix walk can 7839 // prove to be expensive (quadratic in the amount of traffic) 7840 // when there are many objects in the overflow list and 7841 // there is much producer-consumer contention on the list. 7842 // *NOTE*: The overflow list manipulation code here and 7843 // in ParNewGeneration:: are very similar in shape, 7844 // except that in the ParNew case we use the old (from/eden) 7845 // copy of the object to thread the list via its klass word. 7846 // Because of the common code, if you make any changes in 7847 // the code below, please check the ParNew version to see if 7848 // similar changes might be needed. 7849 // CR 6797058 has been filed to consolidate the common code. 7850 bool CMSCollector::par_take_from_overflow_list(size_t num, 7851 OopTaskQueue* work_q, 7852 int no_of_gc_threads) { 7853 assert(work_q->size() == 0, "First empty local work queue"); 7854 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7855 if (_overflow_list == NULL) { 7856 return false; 7857 } 7858 // Grab the entire list; we'll put back a suffix 7859 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7860 Thread* tid = Thread::current(); 7861 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7862 // set to ParallelGCThreads. 7863 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7864 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7865 // If the list is busy, we spin for a short while, 7866 // sleeping between attempts to get the list. 7867 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7868 os::sleep(tid, sleep_time_millis, false); 7869 if (_overflow_list == NULL) { 7870 // Nothing left to take 7871 return false; 7872 } else if (_overflow_list != BUSY) { 7873 // Try and grab the prefix 7874 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7875 } 7876 } 7877 // If the list was found to be empty, or we spun long 7878 // enough, we give up and return empty-handed. If we leave 7879 // the list in the BUSY state below, it must be the case that 7880 // some other thread holds the overflow list and will set it 7881 // to a non-BUSY state in the future. 7882 if (prefix == NULL || prefix == BUSY) { 7883 // Nothing to take or waited long enough 7884 if (prefix == NULL) { 7885 // Write back the NULL in case we overwrote it with BUSY above 7886 // and it is still the same value. 7887 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7888 } 7889 return false; 7890 } 7891 assert(prefix != NULL && prefix != BUSY, "Error"); 7892 size_t i = num; 7893 oop cur = prefix; 7894 // Walk down the first "num" objects, unless we reach the end. 7895 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7896 if (cur->mark() == NULL) { 7897 // We have "num" or fewer elements in the list, so there 7898 // is nothing to return to the global list. 7899 // Write back the NULL in lieu of the BUSY we wrote 7900 // above, if it is still the same value. 7901 if (_overflow_list == BUSY) { 7902 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7903 } 7904 } else { 7905 // Chop off the suffix and return it to the global list. 7906 assert(cur->mark() != BUSY, "Error"); 7907 oop suffix_head = cur->mark(); // suffix will be put back on global list 7908 cur->set_mark(NULL); // break off suffix 7909 // It's possible that the list is still in the empty(busy) state 7910 // we left it in a short while ago; in that case we may be 7911 // able to place back the suffix without incurring the cost 7912 // of a walk down the list. 7913 oop observed_overflow_list = _overflow_list; 7914 oop cur_overflow_list = observed_overflow_list; 7915 bool attached = false; 7916 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7917 observed_overflow_list = 7918 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7919 if (cur_overflow_list == observed_overflow_list) { 7920 attached = true; 7921 break; 7922 } else cur_overflow_list = observed_overflow_list; 7923 } 7924 if (!attached) { 7925 // Too bad, someone else sneaked in (at least) an element; we'll need 7926 // to do a splice. Find tail of suffix so we can prepend suffix to global 7927 // list. 7928 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7929 oop suffix_tail = cur; 7930 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7931 "Tautology"); 7932 observed_overflow_list = _overflow_list; 7933 do { 7934 cur_overflow_list = observed_overflow_list; 7935 if (cur_overflow_list != BUSY) { 7936 // Do the splice ... 7937 suffix_tail->set_mark(markOop(cur_overflow_list)); 7938 } else { // cur_overflow_list == BUSY 7939 suffix_tail->set_mark(NULL); 7940 } 7941 // ... and try to place spliced list back on overflow_list ... 7942 observed_overflow_list = 7943 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7944 } while (cur_overflow_list != observed_overflow_list); 7945 // ... until we have succeeded in doing so. 7946 } 7947 } 7948 7949 // Push the prefix elements on work_q 7950 assert(prefix != NULL, "control point invariant"); 7951 const markOop proto = markOopDesc::prototype(); 7952 oop next; 7953 NOT_PRODUCT(ssize_t n = 0;) 7954 for (cur = prefix; cur != NULL; cur = next) { 7955 next = oop(cur->mark()); 7956 cur->set_mark(proto); // until proven otherwise 7957 assert(oopDesc::is_oop(cur), "Should be an oop"); 7958 bool res = work_q->push(cur); 7959 assert(res, "Bit off more than we can chew?"); 7960 NOT_PRODUCT(n++;) 7961 } 7962 #ifndef PRODUCT 7963 assert(_num_par_pushes >= n, "Too many pops?"); 7964 Atomic::sub(n, &_num_par_pushes); 7965 #endif 7966 return true; 7967 } 7968 7969 // Single-threaded 7970 void CMSCollector::push_on_overflow_list(oop p) { 7971 NOT_PRODUCT(_num_par_pushes++;) 7972 assert(oopDesc::is_oop(p), "Not an oop"); 7973 preserve_mark_if_necessary(p); 7974 p->set_mark((markOop)_overflow_list); 7975 _overflow_list = p; 7976 } 7977 7978 // Multi-threaded; use CAS to prepend to overflow list 7979 void CMSCollector::par_push_on_overflow_list(oop p) { 7980 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7981 assert(oopDesc::is_oop(p), "Not an oop"); 7982 par_preserve_mark_if_necessary(p); 7983 oop observed_overflow_list = _overflow_list; 7984 oop cur_overflow_list; 7985 do { 7986 cur_overflow_list = observed_overflow_list; 7987 if (cur_overflow_list != BUSY) { 7988 p->set_mark(markOop(cur_overflow_list)); 7989 } else { 7990 p->set_mark(NULL); 7991 } 7992 observed_overflow_list = 7993 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7994 } while (cur_overflow_list != observed_overflow_list); 7995 } 7996 #undef BUSY 7997 7998 // Single threaded 7999 // General Note on GrowableArray: pushes may silently fail 8000 // because we are (temporarily) out of C-heap for expanding 8001 // the stack. The problem is quite ubiquitous and affects 8002 // a lot of code in the JVM. The prudent thing for GrowableArray 8003 // to do (for now) is to exit with an error. However, that may 8004 // be too draconian in some cases because the caller may be 8005 // able to recover without much harm. For such cases, we 8006 // should probably introduce a "soft_push" method which returns 8007 // an indication of success or failure with the assumption that 8008 // the caller may be able to recover from a failure; code in 8009 // the VM can then be changed, incrementally, to deal with such 8010 // failures where possible, thus, incrementally hardening the VM 8011 // in such low resource situations. 8012 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8013 _preserved_oop_stack.push(p); 8014 _preserved_mark_stack.push(m); 8015 assert(m == p->mark(), "Mark word changed"); 8016 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8017 "bijection"); 8018 } 8019 8020 // Single threaded 8021 void CMSCollector::preserve_mark_if_necessary(oop p) { 8022 markOop m = p->mark(); 8023 if (m->must_be_preserved(p)) { 8024 preserve_mark_work(p, m); 8025 } 8026 } 8027 8028 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8029 markOop m = p->mark(); 8030 if (m->must_be_preserved(p)) { 8031 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8032 // Even though we read the mark word without holding 8033 // the lock, we are assured that it will not change 8034 // because we "own" this oop, so no other thread can 8035 // be trying to push it on the overflow list; see 8036 // the assertion in preserve_mark_work() that checks 8037 // that m == p->mark(). 8038 preserve_mark_work(p, m); 8039 } 8040 } 8041 8042 // We should be able to do this multi-threaded, 8043 // a chunk of stack being a task (this is 8044 // correct because each oop only ever appears 8045 // once in the overflow list. However, it's 8046 // not very easy to completely overlap this with 8047 // other operations, so will generally not be done 8048 // until all work's been completed. Because we 8049 // expect the preserved oop stack (set) to be small, 8050 // it's probably fine to do this single-threaded. 8051 // We can explore cleverer concurrent/overlapped/parallel 8052 // processing of preserved marks if we feel the 8053 // need for this in the future. Stack overflow should 8054 // be so rare in practice and, when it happens, its 8055 // effect on performance so great that this will 8056 // likely just be in the noise anyway. 8057 void CMSCollector::restore_preserved_marks_if_any() { 8058 assert(SafepointSynchronize::is_at_safepoint(), 8059 "world should be stopped"); 8060 assert(Thread::current()->is_ConcurrentGC_thread() || 8061 Thread::current()->is_VM_thread(), 8062 "should be single-threaded"); 8063 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8064 "bijection"); 8065 8066 while (!_preserved_oop_stack.is_empty()) { 8067 oop p = _preserved_oop_stack.pop(); 8068 assert(oopDesc::is_oop(p), "Should be an oop"); 8069 assert(_span.contains(p), "oop should be in _span"); 8070 assert(p->mark() == markOopDesc::prototype(), 8071 "Set when taken from overflow list"); 8072 markOop m = _preserved_mark_stack.pop(); 8073 p->set_mark(m); 8074 } 8075 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8076 "stacks were cleared above"); 8077 } 8078 8079 #ifndef PRODUCT 8080 bool CMSCollector::no_preserved_marks() const { 8081 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8082 } 8083 #endif 8084 8085 // Transfer some number of overflown objects to usual marking 8086 // stack. Return true if some objects were transferred. 8087 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8088 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8089 (size_t)ParGCDesiredObjsFromOverflowList); 8090 8091 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8092 assert(_collector->overflow_list_is_empty() || res, 8093 "If list is not empty, we should have taken something"); 8094 assert(!res || !_mark_stack->isEmpty(), 8095 "If we took something, it should now be on our stack"); 8096 return res; 8097 } 8098 8099 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8100 size_t res = _sp->block_size_no_stall(addr, _collector); 8101 if (_sp->block_is_obj(addr)) { 8102 if (_live_bit_map->isMarked(addr)) { 8103 // It can't have been dead in a previous cycle 8104 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8105 } else { 8106 _dead_bit_map->mark(addr); // mark the dead object 8107 } 8108 } 8109 // Could be 0, if the block size could not be computed without stalling. 8110 return res; 8111 } 8112 8113 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8114 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8115 switch (phase) { 8116 case CMSCollector::InitialMarking: 8117 initialize(manager /* GC manager */ , 8118 cause /* cause of the GC */, 8119 true /* recordGCBeginTime */, 8120 true /* recordPreGCUsage */, 8121 false /* recordPeakUsage */, 8122 false /* recordPostGCusage */, 8123 true /* recordAccumulatedGCTime */, 8124 false /* recordGCEndTime */, 8125 false /* countCollection */ ); 8126 break; 8127 8128 case CMSCollector::FinalMarking: 8129 initialize(manager /* GC manager */ , 8130 cause /* cause of the GC */, 8131 false /* recordGCBeginTime */, 8132 false /* recordPreGCUsage */, 8133 false /* recordPeakUsage */, 8134 false /* recordPostGCusage */, 8135 true /* recordAccumulatedGCTime */, 8136 false /* recordGCEndTime */, 8137 false /* countCollection */ ); 8138 break; 8139 8140 case CMSCollector::Sweeping: 8141 initialize(manager /* GC manager */ , 8142 cause /* cause of the GC */, 8143 false /* recordGCBeginTime */, 8144 false /* recordPreGCUsage */, 8145 true /* recordPeakUsage */, 8146 true /* recordPostGCusage */, 8147 false /* recordAccumulatedGCTime */, 8148 true /* recordGCEndTime */, 8149 true /* countCollection */ ); 8150 break; 8151 8152 default: 8153 ShouldNotReachHere(); 8154 } 8155 }