1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.inline.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->cr(); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 log_warning(gc)("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 log_warning(gc)("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 549 // Parallel task queues; these are shared for the 550 // concurrent and stop-world phases of CMS, but 551 // are not shared with parallel scavenge (ParNew). 552 { 553 uint i; 554 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 555 556 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 557 || ParallelRefProcEnabled) 558 && num_queues > 0) { 559 _task_queues = new OopTaskQueueSet(num_queues); 560 if (_task_queues == NULL) { 561 log_warning(gc)("task_queues allocation failure."); 562 return; 563 } 564 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 565 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 566 for (i = 0; i < num_queues; i++) { 567 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 568 if (q == NULL) { 569 log_warning(gc)("work_queue allocation failure."); 570 return; 571 } 572 _task_queues->register_queue(i, q); 573 } 574 for (i = 0; i < num_queues; i++) { 575 _task_queues->queue(i)->initialize(); 576 _hash_seed[i] = 17; // copied from ParNew 577 } 578 } 579 } 580 581 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 582 583 // Clip CMSBootstrapOccupancy between 0 and 100. 584 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 585 586 // Now tell CMS generations the identity of their collector 587 ConcurrentMarkSweepGeneration::set_collector(this); 588 589 // Create & start a CMS thread for this CMS collector 590 _cmsThread = ConcurrentMarkSweepThread::start(this); 591 assert(cmsThread() != NULL, "CMS Thread should have been created"); 592 assert(cmsThread()->collector() == this, 593 "CMS Thread should refer to this gen"); 594 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 595 596 // Support for parallelizing young gen rescan 597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 598 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 599 _young_gen = (ParNewGeneration*)gch->young_gen(); 600 if (gch->supports_inline_contig_alloc()) { 601 _top_addr = gch->top_addr(); 602 _end_addr = gch->end_addr(); 603 assert(_young_gen != NULL, "no _young_gen"); 604 _eden_chunk_index = 0; 605 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 606 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 607 } 608 609 // Support for parallelizing survivor space rescan 610 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 611 const size_t max_plab_samples = 612 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 613 614 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 615 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 616 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 617 _survivor_chunk_capacity = max_plab_samples; 618 for (uint i = 0; i < ParallelGCThreads; i++) { 619 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 620 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 621 assert(cur->end() == 0, "Should be 0"); 622 assert(cur->array() == vec, "Should be vec"); 623 assert(cur->capacity() == max_plab_samples, "Error"); 624 } 625 } 626 627 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 628 _gc_counters = new CollectorCounters("CMS", 1); 629 _completed_initialization = true; 630 _inter_sweep_timer.start(); // start of time 631 } 632 633 const char* ConcurrentMarkSweepGeneration::name() const { 634 return "concurrent mark-sweep generation"; 635 } 636 void ConcurrentMarkSweepGeneration::update_counters() { 637 if (UsePerfData) { 638 _space_counters->update_all(); 639 _gen_counters->update_all(); 640 } 641 } 642 643 // this is an optimized version of update_counters(). it takes the 644 // used value as a parameter rather than computing it. 645 // 646 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 647 if (UsePerfData) { 648 _space_counters->update_used(used); 649 _space_counters->update_capacity(); 650 _gen_counters->update_all(); 651 } 652 } 653 654 void ConcurrentMarkSweepGeneration::print() const { 655 Generation::print(); 656 cmsSpace()->print(); 657 } 658 659 #ifndef PRODUCT 660 void ConcurrentMarkSweepGeneration::print_statistics() { 661 cmsSpace()->printFLCensus(0); 662 } 663 #endif 664 665 size_t 666 ConcurrentMarkSweepGeneration::contiguous_available() const { 667 // dld proposes an improvement in precision here. If the committed 668 // part of the space ends in a free block we should add that to 669 // uncommitted size in the calculation below. Will make this 670 // change later, staying with the approximation below for the 671 // time being. -- ysr. 672 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 673 } 674 675 size_t 676 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 677 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 678 } 679 680 size_t ConcurrentMarkSweepGeneration::max_available() const { 681 return free() + _virtual_space.uncommitted_size(); 682 } 683 684 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 685 size_t available = max_available(); 686 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 687 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 688 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 689 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 690 return res; 691 } 692 693 // At a promotion failure dump information on block layout in heap 694 // (cms old generation). 695 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 696 Log(gc, promotion) log; 697 if (log.is_trace()) { 698 ResourceMark rm; 699 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 700 } 701 } 702 703 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 704 // Clear the promotion information. These pointers can be adjusted 705 // along with all the other pointers into the heap but 706 // compaction is expected to be a rare event with 707 // a heap using cms so don't do it without seeing the need. 708 for (uint i = 0; i < ParallelGCThreads; i++) { 709 _par_gc_thread_states[i]->promo.reset(); 710 } 711 } 712 713 void ConcurrentMarkSweepGeneration::compute_new_size() { 714 assert_locked_or_safepoint(Heap_lock); 715 716 // If incremental collection failed, we just want to expand 717 // to the limit. 718 if (incremental_collection_failed()) { 719 clear_incremental_collection_failed(); 720 grow_to_reserved(); 721 return; 722 } 723 724 // The heap has been compacted but not reset yet. 725 // Any metric such as free() or used() will be incorrect. 726 727 CardGeneration::compute_new_size(); 728 729 // Reset again after a possible resizing 730 if (did_compact()) { 731 cmsSpace()->reset_after_compaction(); 732 } 733 } 734 735 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 736 assert_locked_or_safepoint(Heap_lock); 737 738 // If incremental collection failed, we just want to expand 739 // to the limit. 740 if (incremental_collection_failed()) { 741 clear_incremental_collection_failed(); 742 grow_to_reserved(); 743 return; 744 } 745 746 double free_percentage = ((double) free()) / capacity(); 747 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 748 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 749 750 // compute expansion delta needed for reaching desired free percentage 751 if (free_percentage < desired_free_percentage) { 752 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 753 assert(desired_capacity >= capacity(), "invalid expansion size"); 754 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 755 Log(gc) log; 756 if (log.is_trace()) { 757 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 758 log.trace("From compute_new_size: "); 759 log.trace(" Free fraction %f", free_percentage); 760 log.trace(" Desired free fraction %f", desired_free_percentage); 761 log.trace(" Maximum free fraction %f", maximum_free_percentage); 762 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 763 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 764 GenCollectedHeap* gch = GenCollectedHeap::heap(); 765 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 766 size_t young_size = gch->young_gen()->capacity(); 767 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 768 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 769 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 770 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 771 } 772 // safe if expansion fails 773 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 774 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 775 } else { 776 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 777 assert(desired_capacity <= capacity(), "invalid expansion size"); 778 size_t shrink_bytes = capacity() - desired_capacity; 779 // Don't shrink unless the delta is greater than the minimum shrink we want 780 if (shrink_bytes >= MinHeapDeltaBytes) { 781 shrink_free_list_by(shrink_bytes); 782 } 783 } 784 } 785 786 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 787 return cmsSpace()->freelistLock(); 788 } 789 790 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 791 CMSSynchronousYieldRequest yr; 792 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 793 return have_lock_and_allocate(size, tlab); 794 } 795 796 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 797 bool tlab /* ignored */) { 798 assert_lock_strong(freelistLock()); 799 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 800 HeapWord* res = cmsSpace()->allocate(adjustedSize); 801 // Allocate the object live (grey) if the background collector has 802 // started marking. This is necessary because the marker may 803 // have passed this address and consequently this object will 804 // not otherwise be greyed and would be incorrectly swept up. 805 // Note that if this object contains references, the writing 806 // of those references will dirty the card containing this object 807 // allowing the object to be blackened (and its references scanned) 808 // either during a preclean phase or at the final checkpoint. 809 if (res != NULL) { 810 // We may block here with an uninitialized object with 811 // its mark-bit or P-bits not yet set. Such objects need 812 // to be safely navigable by block_start(). 813 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 814 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 815 collector()->direct_allocated(res, adjustedSize); 816 _direct_allocated_words += adjustedSize; 817 // allocation counters 818 NOT_PRODUCT( 819 _numObjectsAllocated++; 820 _numWordsAllocated += (int)adjustedSize; 821 ) 822 } 823 return res; 824 } 825 826 // In the case of direct allocation by mutators in a generation that 827 // is being concurrently collected, the object must be allocated 828 // live (grey) if the background collector has started marking. 829 // This is necessary because the marker may 830 // have passed this address and consequently this object will 831 // not otherwise be greyed and would be incorrectly swept up. 832 // Note that if this object contains references, the writing 833 // of those references will dirty the card containing this object 834 // allowing the object to be blackened (and its references scanned) 835 // either during a preclean phase or at the final checkpoint. 836 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 837 assert(_markBitMap.covers(start, size), "Out of bounds"); 838 if (_collectorState >= Marking) { 839 MutexLockerEx y(_markBitMap.lock(), 840 Mutex::_no_safepoint_check_flag); 841 // [see comments preceding SweepClosure::do_blk() below for details] 842 // 843 // Can the P-bits be deleted now? JJJ 844 // 845 // 1. need to mark the object as live so it isn't collected 846 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 847 // 3. need to mark the end of the object so marking, precleaning or sweeping 848 // can skip over uninitialized or unparsable objects. An allocated 849 // object is considered uninitialized for our purposes as long as 850 // its klass word is NULL. All old gen objects are parsable 851 // as soon as they are initialized.) 852 _markBitMap.mark(start); // object is live 853 _markBitMap.mark(start + 1); // object is potentially uninitialized? 854 _markBitMap.mark(start + size - 1); 855 // mark end of object 856 } 857 // check that oop looks uninitialized 858 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 859 } 860 861 void CMSCollector::promoted(bool par, HeapWord* start, 862 bool is_obj_array, size_t obj_size) { 863 assert(_markBitMap.covers(start), "Out of bounds"); 864 // See comment in direct_allocated() about when objects should 865 // be allocated live. 866 if (_collectorState >= Marking) { 867 // we already hold the marking bit map lock, taken in 868 // the prologue 869 if (par) { 870 _markBitMap.par_mark(start); 871 } else { 872 _markBitMap.mark(start); 873 } 874 // We don't need to mark the object as uninitialized (as 875 // in direct_allocated above) because this is being done with the 876 // world stopped and the object will be initialized by the 877 // time the marking, precleaning or sweeping get to look at it. 878 // But see the code for copying objects into the CMS generation, 879 // where we need to ensure that concurrent readers of the 880 // block offset table are able to safely navigate a block that 881 // is in flux from being free to being allocated (and in 882 // transition while being copied into) and subsequently 883 // becoming a bona-fide object when the copy/promotion is complete. 884 assert(SafepointSynchronize::is_at_safepoint(), 885 "expect promotion only at safepoints"); 886 887 if (_collectorState < Sweeping) { 888 // Mark the appropriate cards in the modUnionTable, so that 889 // this object gets scanned before the sweep. If this is 890 // not done, CMS generation references in the object might 891 // not get marked. 892 // For the case of arrays, which are otherwise precisely 893 // marked, we need to dirty the entire array, not just its head. 894 if (is_obj_array) { 895 // The [par_]mark_range() method expects mr.end() below to 896 // be aligned to the granularity of a bit's representation 897 // in the heap. In the case of the MUT below, that's a 898 // card size. 899 MemRegion mr(start, 900 (HeapWord*)round_to((intptr_t)(start + obj_size), 901 CardTableModRefBS::card_size /* bytes */)); 902 if (par) { 903 _modUnionTable.par_mark_range(mr); 904 } else { 905 _modUnionTable.mark_range(mr); 906 } 907 } else { // not an obj array; we can just mark the head 908 if (par) { 909 _modUnionTable.par_mark(start); 910 } else { 911 _modUnionTable.mark(start); 912 } 913 } 914 } 915 } 916 } 917 918 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 919 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 920 // allocate, copy and if necessary update promoinfo -- 921 // delegate to underlying space. 922 assert_lock_strong(freelistLock()); 923 924 #ifndef PRODUCT 925 if (GenCollectedHeap::heap()->promotion_should_fail()) { 926 return NULL; 927 } 928 #endif // #ifndef PRODUCT 929 930 oop res = _cmsSpace->promote(obj, obj_size); 931 if (res == NULL) { 932 // expand and retry 933 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 934 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 935 // Since this is the old generation, we don't try to promote 936 // into a more senior generation. 937 res = _cmsSpace->promote(obj, obj_size); 938 } 939 if (res != NULL) { 940 // See comment in allocate() about when objects should 941 // be allocated live. 942 assert(obj->is_oop(), "Will dereference klass pointer below"); 943 collector()->promoted(false, // Not parallel 944 (HeapWord*)res, obj->is_objArray(), obj_size); 945 // promotion counters 946 NOT_PRODUCT( 947 _numObjectsPromoted++; 948 _numWordsPromoted += 949 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 950 ) 951 } 952 return res; 953 } 954 955 956 // IMPORTANT: Notes on object size recognition in CMS. 957 // --------------------------------------------------- 958 // A block of storage in the CMS generation is always in 959 // one of three states. A free block (FREE), an allocated 960 // object (OBJECT) whose size() method reports the correct size, 961 // and an intermediate state (TRANSIENT) in which its size cannot 962 // be accurately determined. 963 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 964 // ----------------------------------------------------- 965 // FREE: klass_word & 1 == 1; mark_word holds block size 966 // 967 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 968 // obj->size() computes correct size 969 // 970 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 971 // 972 // STATE IDENTIFICATION: (64 bit+COOPS) 973 // ------------------------------------ 974 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 975 // 976 // OBJECT: klass_word installed; klass_word != 0; 977 // obj->size() computes correct size 978 // 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 980 // 981 // 982 // STATE TRANSITION DIAGRAM 983 // 984 // mut / parnew mut / parnew 985 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 986 // ^ | 987 // |------------------------ DEAD <------------------------------------| 988 // sweep mut 989 // 990 // While a block is in TRANSIENT state its size cannot be determined 991 // so readers will either need to come back later or stall until 992 // the size can be determined. Note that for the case of direct 993 // allocation, P-bits, when available, may be used to determine the 994 // size of an object that may not yet have been initialized. 995 996 // Things to support parallel young-gen collection. 997 oop 998 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 999 oop old, markOop m, 1000 size_t word_sz) { 1001 #ifndef PRODUCT 1002 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1003 return NULL; 1004 } 1005 #endif // #ifndef PRODUCT 1006 1007 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1008 PromotionInfo* promoInfo = &ps->promo; 1009 // if we are tracking promotions, then first ensure space for 1010 // promotion (including spooling space for saving header if necessary). 1011 // then allocate and copy, then track promoted info if needed. 1012 // When tracking (see PromotionInfo::track()), the mark word may 1013 // be displaced and in this case restoration of the mark word 1014 // occurs in the (oop_since_save_marks_)iterate phase. 1015 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1016 // Out of space for allocating spooling buffers; 1017 // try expanding and allocating spooling buffers. 1018 if (!expand_and_ensure_spooling_space(promoInfo)) { 1019 return NULL; 1020 } 1021 } 1022 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1023 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1024 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1025 if (obj_ptr == NULL) { 1026 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1027 if (obj_ptr == NULL) { 1028 return NULL; 1029 } 1030 } 1031 oop obj = oop(obj_ptr); 1032 OrderAccess::storestore(); 1033 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1034 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1035 // IMPORTANT: See note on object initialization for CMS above. 1036 // Otherwise, copy the object. Here we must be careful to insert the 1037 // klass pointer last, since this marks the block as an allocated object. 1038 // Except with compressed oops it's the mark word. 1039 HeapWord* old_ptr = (HeapWord*)old; 1040 // Restore the mark word copied above. 1041 obj->set_mark(m); 1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1044 OrderAccess::storestore(); 1045 1046 if (UseCompressedClassPointers) { 1047 // Copy gap missed by (aligned) header size calculation below 1048 obj->set_klass_gap(old->klass_gap()); 1049 } 1050 if (word_sz > (size_t)oopDesc::header_size()) { 1051 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1052 obj_ptr + oopDesc::header_size(), 1053 word_sz - oopDesc::header_size()); 1054 } 1055 1056 // Now we can track the promoted object, if necessary. We take care 1057 // to delay the transition from uninitialized to full object 1058 // (i.e., insertion of klass pointer) until after, so that it 1059 // atomically becomes a promoted object. 1060 if (promoInfo->tracking()) { 1061 promoInfo->track((PromotedObject*)obj, old->klass()); 1062 } 1063 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1064 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1065 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1066 1067 // Finally, install the klass pointer (this should be volatile). 1068 OrderAccess::storestore(); 1069 obj->set_klass(old->klass()); 1070 // We should now be able to calculate the right size for this object 1071 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1072 1073 collector()->promoted(true, // parallel 1074 obj_ptr, old->is_objArray(), word_sz); 1075 1076 NOT_PRODUCT( 1077 Atomic::inc_ptr(&_numObjectsPromoted); 1078 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1079 ) 1080 1081 return obj; 1082 } 1083 1084 void 1085 ConcurrentMarkSweepGeneration:: 1086 par_promote_alloc_done(int thread_num) { 1087 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1088 ps->lab.retire(thread_num); 1089 } 1090 1091 void 1092 ConcurrentMarkSweepGeneration:: 1093 par_oop_since_save_marks_iterate_done(int thread_num) { 1094 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1095 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1096 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1097 1098 // Because card-scanning has been completed, subsequent phases 1099 // (e.g., reference processing) will not need to recognize which 1100 // objects have been promoted during this GC. So, we can now disable 1101 // promotion tracking. 1102 ps->promo.stopTrackingPromotions(); 1103 } 1104 1105 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1106 size_t size, 1107 bool tlab) 1108 { 1109 // We allow a STW collection only if a full 1110 // collection was requested. 1111 return full || should_allocate(size, tlab); // FIX ME !!! 1112 // This and promotion failure handling are connected at the 1113 // hip and should be fixed by untying them. 1114 } 1115 1116 bool CMSCollector::shouldConcurrentCollect() { 1117 LogTarget(Trace, gc) log; 1118 1119 if (_full_gc_requested) { 1120 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1121 return true; 1122 } 1123 1124 FreelistLocker x(this); 1125 // ------------------------------------------------------------------ 1126 // Print out lots of information which affects the initiation of 1127 // a collection. 1128 if (log.is_enabled() && stats().valid()) { 1129 log.print("CMSCollector shouldConcurrentCollect: "); 1130 1131 LogStream out(log); 1132 stats().print_on(&out); 1133 1134 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1135 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1136 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1137 log.print("promotion_rate=%g", stats().promotion_rate()); 1138 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1139 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1140 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1141 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1142 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1143 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1144 } 1145 // ------------------------------------------------------------------ 1146 1147 // If the estimated time to complete a cms collection (cms_duration()) 1148 // is less than the estimated time remaining until the cms generation 1149 // is full, start a collection. 1150 if (!UseCMSInitiatingOccupancyOnly) { 1151 if (stats().valid()) { 1152 if (stats().time_until_cms_start() == 0.0) { 1153 return true; 1154 } 1155 } else { 1156 // We want to conservatively collect somewhat early in order 1157 // to try and "bootstrap" our CMS/promotion statistics; 1158 // this branch will not fire after the first successful CMS 1159 // collection because the stats should then be valid. 1160 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1161 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1162 _cmsGen->occupancy(), _bootstrap_occupancy); 1163 return true; 1164 } 1165 } 1166 } 1167 1168 // Otherwise, we start a collection cycle if 1169 // old gen want a collection cycle started. Each may use 1170 // an appropriate criterion for making this decision. 1171 // XXX We need to make sure that the gen expansion 1172 // criterion dovetails well with this. XXX NEED TO FIX THIS 1173 if (_cmsGen->should_concurrent_collect()) { 1174 log.print("CMS old gen initiated"); 1175 return true; 1176 } 1177 1178 // We start a collection if we believe an incremental collection may fail; 1179 // this is not likely to be productive in practice because it's probably too 1180 // late anyway. 1181 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1182 assert(gch->collector_policy()->is_generation_policy(), 1183 "You may want to check the correctness of the following"); 1184 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1185 log.print("CMSCollector: collect because incremental collection will fail "); 1186 return true; 1187 } 1188 1189 if (MetaspaceGC::should_concurrent_collect()) { 1190 log.print("CMSCollector: collect for metadata allocation "); 1191 return true; 1192 } 1193 1194 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1195 if (CMSTriggerInterval >= 0) { 1196 if (CMSTriggerInterval == 0) { 1197 // Trigger always 1198 return true; 1199 } 1200 1201 // Check the CMS time since begin (we do not check the stats validity 1202 // as we want to be able to trigger the first CMS cycle as well) 1203 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1204 if (stats().valid()) { 1205 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1206 stats().cms_time_since_begin()); 1207 } else { 1208 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1209 } 1210 return true; 1211 } 1212 } 1213 1214 return false; 1215 } 1216 1217 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1218 1219 // Clear _expansion_cause fields of constituent generations 1220 void CMSCollector::clear_expansion_cause() { 1221 _cmsGen->clear_expansion_cause(); 1222 } 1223 1224 // We should be conservative in starting a collection cycle. To 1225 // start too eagerly runs the risk of collecting too often in the 1226 // extreme. To collect too rarely falls back on full collections, 1227 // which works, even if not optimum in terms of concurrent work. 1228 // As a work around for too eagerly collecting, use the flag 1229 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1230 // giving the user an easily understandable way of controlling the 1231 // collections. 1232 // We want to start a new collection cycle if any of the following 1233 // conditions hold: 1234 // . our current occupancy exceeds the configured initiating occupancy 1235 // for this generation, or 1236 // . we recently needed to expand this space and have not, since that 1237 // expansion, done a collection of this generation, or 1238 // . the underlying space believes that it may be a good idea to initiate 1239 // a concurrent collection (this may be based on criteria such as the 1240 // following: the space uses linear allocation and linear allocation is 1241 // going to fail, or there is believed to be excessive fragmentation in 1242 // the generation, etc... or ... 1243 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1244 // the case of the old generation; see CR 6543076): 1245 // we may be approaching a point at which allocation requests may fail because 1246 // we will be out of sufficient free space given allocation rate estimates.] 1247 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1248 1249 assert_lock_strong(freelistLock()); 1250 if (occupancy() > initiating_occupancy()) { 1251 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1252 short_name(), occupancy(), initiating_occupancy()); 1253 return true; 1254 } 1255 if (UseCMSInitiatingOccupancyOnly) { 1256 return false; 1257 } 1258 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1259 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1260 return true; 1261 } 1262 return false; 1263 } 1264 1265 void ConcurrentMarkSweepGeneration::collect(bool full, 1266 bool clear_all_soft_refs, 1267 size_t size, 1268 bool tlab) 1269 { 1270 collector()->collect(full, clear_all_soft_refs, size, tlab); 1271 } 1272 1273 void CMSCollector::collect(bool full, 1274 bool clear_all_soft_refs, 1275 size_t size, 1276 bool tlab) 1277 { 1278 // The following "if" branch is present for defensive reasons. 1279 // In the current uses of this interface, it can be replaced with: 1280 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1281 // But I am not placing that assert here to allow future 1282 // generality in invoking this interface. 1283 if (GCLocker::is_active()) { 1284 // A consistency test for GCLocker 1285 assert(GCLocker::needs_gc(), "Should have been set already"); 1286 // Skip this foreground collection, instead 1287 // expanding the heap if necessary. 1288 // Need the free list locks for the call to free() in compute_new_size() 1289 compute_new_size(); 1290 return; 1291 } 1292 acquire_control_and_collect(full, clear_all_soft_refs); 1293 } 1294 1295 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1296 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1297 unsigned int gc_count = gch->total_full_collections(); 1298 if (gc_count == full_gc_count) { 1299 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1300 _full_gc_requested = true; 1301 _full_gc_cause = cause; 1302 CGC_lock->notify(); // nudge CMS thread 1303 } else { 1304 assert(gc_count > full_gc_count, "Error: causal loop"); 1305 } 1306 } 1307 1308 bool CMSCollector::is_external_interruption() { 1309 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1310 return GCCause::is_user_requested_gc(cause) || 1311 GCCause::is_serviceability_requested_gc(cause); 1312 } 1313 1314 void CMSCollector::report_concurrent_mode_interruption() { 1315 if (is_external_interruption()) { 1316 log_debug(gc)("Concurrent mode interrupted"); 1317 } else { 1318 log_debug(gc)("Concurrent mode failure"); 1319 _gc_tracer_cm->report_concurrent_mode_failure(); 1320 } 1321 } 1322 1323 1324 // The foreground and background collectors need to coordinate in order 1325 // to make sure that they do not mutually interfere with CMS collections. 1326 // When a background collection is active, 1327 // the foreground collector may need to take over (preempt) and 1328 // synchronously complete an ongoing collection. Depending on the 1329 // frequency of the background collections and the heap usage 1330 // of the application, this preemption can be seldom or frequent. 1331 // There are only certain 1332 // points in the background collection that the "collection-baton" 1333 // can be passed to the foreground collector. 1334 // 1335 // The foreground collector will wait for the baton before 1336 // starting any part of the collection. The foreground collector 1337 // will only wait at one location. 1338 // 1339 // The background collector will yield the baton before starting a new 1340 // phase of the collection (e.g., before initial marking, marking from roots, 1341 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1342 // of the loop which switches the phases. The background collector does some 1343 // of the phases (initial mark, final re-mark) with the world stopped. 1344 // Because of locking involved in stopping the world, 1345 // the foreground collector should not block waiting for the background 1346 // collector when it is doing a stop-the-world phase. The background 1347 // collector will yield the baton at an additional point just before 1348 // it enters a stop-the-world phase. Once the world is stopped, the 1349 // background collector checks the phase of the collection. If the 1350 // phase has not changed, it proceeds with the collection. If the 1351 // phase has changed, it skips that phase of the collection. See 1352 // the comments on the use of the Heap_lock in collect_in_background(). 1353 // 1354 // Variable used in baton passing. 1355 // _foregroundGCIsActive - Set to true by the foreground collector when 1356 // it wants the baton. The foreground clears it when it has finished 1357 // the collection. 1358 // _foregroundGCShouldWait - Set to true by the background collector 1359 // when it is running. The foreground collector waits while 1360 // _foregroundGCShouldWait is true. 1361 // CGC_lock - monitor used to protect access to the above variables 1362 // and to notify the foreground and background collectors. 1363 // _collectorState - current state of the CMS collection. 1364 // 1365 // The foreground collector 1366 // acquires the CGC_lock 1367 // sets _foregroundGCIsActive 1368 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1369 // various locks acquired in preparation for the collection 1370 // are released so as not to block the background collector 1371 // that is in the midst of a collection 1372 // proceeds with the collection 1373 // clears _foregroundGCIsActive 1374 // returns 1375 // 1376 // The background collector in a loop iterating on the phases of the 1377 // collection 1378 // acquires the CGC_lock 1379 // sets _foregroundGCShouldWait 1380 // if _foregroundGCIsActive is set 1381 // clears _foregroundGCShouldWait, notifies _CGC_lock 1382 // waits on _CGC_lock for _foregroundGCIsActive to become false 1383 // and exits the loop. 1384 // otherwise 1385 // proceed with that phase of the collection 1386 // if the phase is a stop-the-world phase, 1387 // yield the baton once more just before enqueueing 1388 // the stop-world CMS operation (executed by the VM thread). 1389 // returns after all phases of the collection are done 1390 // 1391 1392 void CMSCollector::acquire_control_and_collect(bool full, 1393 bool clear_all_soft_refs) { 1394 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1395 assert(!Thread::current()->is_ConcurrentGC_thread(), 1396 "shouldn't try to acquire control from self!"); 1397 1398 // Start the protocol for acquiring control of the 1399 // collection from the background collector (aka CMS thread). 1400 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1401 "VM thread should have CMS token"); 1402 // Remember the possibly interrupted state of an ongoing 1403 // concurrent collection 1404 CollectorState first_state = _collectorState; 1405 1406 // Signal to a possibly ongoing concurrent collection that 1407 // we want to do a foreground collection. 1408 _foregroundGCIsActive = true; 1409 1410 // release locks and wait for a notify from the background collector 1411 // releasing the locks in only necessary for phases which 1412 // do yields to improve the granularity of the collection. 1413 assert_lock_strong(bitMapLock()); 1414 // We need to lock the Free list lock for the space that we are 1415 // currently collecting. 1416 assert(haveFreelistLocks(), "Must be holding free list locks"); 1417 bitMapLock()->unlock(); 1418 releaseFreelistLocks(); 1419 { 1420 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1421 if (_foregroundGCShouldWait) { 1422 // We are going to be waiting for action for the CMS thread; 1423 // it had better not be gone (for instance at shutdown)! 1424 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1425 "CMS thread must be running"); 1426 // Wait here until the background collector gives us the go-ahead 1427 ConcurrentMarkSweepThread::clear_CMS_flag( 1428 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1429 // Get a possibly blocked CMS thread going: 1430 // Note that we set _foregroundGCIsActive true above, 1431 // without protection of the CGC_lock. 1432 CGC_lock->notify(); 1433 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1434 "Possible deadlock"); 1435 while (_foregroundGCShouldWait) { 1436 // wait for notification 1437 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1438 // Possibility of delay/starvation here, since CMS token does 1439 // not know to give priority to VM thread? Actually, i think 1440 // there wouldn't be any delay/starvation, but the proof of 1441 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1442 } 1443 ConcurrentMarkSweepThread::set_CMS_flag( 1444 ConcurrentMarkSweepThread::CMS_vm_has_token); 1445 } 1446 } 1447 // The CMS_token is already held. Get back the other locks. 1448 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1449 "VM thread should have CMS token"); 1450 getFreelistLocks(); 1451 bitMapLock()->lock_without_safepoint_check(); 1452 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1453 p2i(Thread::current()), first_state); 1454 log_debug(gc, state)(" gets control with state %d", _collectorState); 1455 1456 // Inform cms gen if this was due to partial collection failing. 1457 // The CMS gen may use this fact to determine its expansion policy. 1458 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1459 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1460 assert(!_cmsGen->incremental_collection_failed(), 1461 "Should have been noticed, reacted to and cleared"); 1462 _cmsGen->set_incremental_collection_failed(); 1463 } 1464 1465 if (first_state > Idling) { 1466 report_concurrent_mode_interruption(); 1467 } 1468 1469 set_did_compact(true); 1470 1471 // If the collection is being acquired from the background 1472 // collector, there may be references on the discovered 1473 // references lists. Abandon those references, since some 1474 // of them may have become unreachable after concurrent 1475 // discovery; the STW compacting collector will redo discovery 1476 // more precisely, without being subject to floating garbage. 1477 // Leaving otherwise unreachable references in the discovered 1478 // lists would require special handling. 1479 ref_processor()->disable_discovery(); 1480 ref_processor()->abandon_partial_discovery(); 1481 ref_processor()->verify_no_references_recorded(); 1482 1483 if (first_state > Idling) { 1484 save_heap_summary(); 1485 } 1486 1487 do_compaction_work(clear_all_soft_refs); 1488 1489 // Has the GC time limit been exceeded? 1490 size_t max_eden_size = _young_gen->max_eden_size(); 1491 GCCause::Cause gc_cause = gch->gc_cause(); 1492 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1493 _young_gen->eden()->used(), 1494 _cmsGen->max_capacity(), 1495 max_eden_size, 1496 full, 1497 gc_cause, 1498 gch->collector_policy()); 1499 1500 // Reset the expansion cause, now that we just completed 1501 // a collection cycle. 1502 clear_expansion_cause(); 1503 _foregroundGCIsActive = false; 1504 return; 1505 } 1506 1507 // Resize the tenured generation 1508 // after obtaining the free list locks for the 1509 // two generations. 1510 void CMSCollector::compute_new_size() { 1511 assert_locked_or_safepoint(Heap_lock); 1512 FreelistLocker z(this); 1513 MetaspaceGC::compute_new_size(); 1514 _cmsGen->compute_new_size_free_list(); 1515 } 1516 1517 // A work method used by the foreground collector to do 1518 // a mark-sweep-compact. 1519 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1520 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1521 1522 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1523 gc_timer->register_gc_start(); 1524 1525 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1526 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1527 1528 gch->pre_full_gc_dump(gc_timer); 1529 1530 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1531 1532 // Temporarily widen the span of the weak reference processing to 1533 // the entire heap. 1534 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1535 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1536 // Temporarily, clear the "is_alive_non_header" field of the 1537 // reference processor. 1538 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1539 // Temporarily make reference _processing_ single threaded (non-MT). 1540 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1541 // Temporarily make refs discovery atomic 1542 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1543 // Temporarily make reference _discovery_ single threaded (non-MT) 1544 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1545 1546 ref_processor()->set_enqueuing_is_done(false); 1547 ref_processor()->enable_discovery(); 1548 ref_processor()->setup_policy(clear_all_soft_refs); 1549 // If an asynchronous collection finishes, the _modUnionTable is 1550 // all clear. If we are assuming the collection from an asynchronous 1551 // collection, clear the _modUnionTable. 1552 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1553 "_modUnionTable should be clear if the baton was not passed"); 1554 _modUnionTable.clear_all(); 1555 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1556 "mod union for klasses should be clear if the baton was passed"); 1557 _ct->klass_rem_set()->clear_mod_union(); 1558 1559 // We must adjust the allocation statistics being maintained 1560 // in the free list space. We do so by reading and clearing 1561 // the sweep timer and updating the block flux rate estimates below. 1562 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1563 if (_inter_sweep_timer.is_active()) { 1564 _inter_sweep_timer.stop(); 1565 // Note that we do not use this sample to update the _inter_sweep_estimate. 1566 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1567 _inter_sweep_estimate.padded_average(), 1568 _intra_sweep_estimate.padded_average()); 1569 } 1570 1571 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1572 #ifdef ASSERT 1573 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1574 size_t free_size = cms_space->free(); 1575 assert(free_size == 1576 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1577 * HeapWordSize, 1578 "All the free space should be compacted into one chunk at top"); 1579 assert(cms_space->dictionary()->total_chunk_size( 1580 debug_only(cms_space->freelistLock())) == 0 || 1581 cms_space->totalSizeInIndexedFreeLists() == 0, 1582 "All the free space should be in a single chunk"); 1583 size_t num = cms_space->totalCount(); 1584 assert((free_size == 0 && num == 0) || 1585 (free_size > 0 && (num == 1 || num == 2)), 1586 "There should be at most 2 free chunks after compaction"); 1587 #endif // ASSERT 1588 _collectorState = Resetting; 1589 assert(_restart_addr == NULL, 1590 "Should have been NULL'd before baton was passed"); 1591 reset_stw(); 1592 _cmsGen->reset_after_compaction(); 1593 _concurrent_cycles_since_last_unload = 0; 1594 1595 // Clear any data recorded in the PLAB chunk arrays. 1596 if (_survivor_plab_array != NULL) { 1597 reset_survivor_plab_arrays(); 1598 } 1599 1600 // Adjust the per-size allocation stats for the next epoch. 1601 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1602 // Restart the "inter sweep timer" for the next epoch. 1603 _inter_sweep_timer.reset(); 1604 _inter_sweep_timer.start(); 1605 1606 gch->post_full_gc_dump(gc_timer); 1607 1608 gc_timer->register_gc_end(); 1609 1610 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1611 1612 // For a mark-sweep-compact, compute_new_size() will be called 1613 // in the heap's do_collection() method. 1614 } 1615 1616 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1617 Log(gc, heap) log; 1618 if (!log.is_trace()) { 1619 return; 1620 } 1621 1622 ContiguousSpace* eden_space = _young_gen->eden(); 1623 ContiguousSpace* from_space = _young_gen->from(); 1624 ContiguousSpace* to_space = _young_gen->to(); 1625 // Eden 1626 if (_eden_chunk_array != NULL) { 1627 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1628 p2i(eden_space->bottom()), p2i(eden_space->top()), 1629 p2i(eden_space->end()), eden_space->capacity()); 1630 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1631 _eden_chunk_index, _eden_chunk_capacity); 1632 for (size_t i = 0; i < _eden_chunk_index; i++) { 1633 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1634 } 1635 } 1636 // Survivor 1637 if (_survivor_chunk_array != NULL) { 1638 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1639 p2i(from_space->bottom()), p2i(from_space->top()), 1640 p2i(from_space->end()), from_space->capacity()); 1641 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1642 _survivor_chunk_index, _survivor_chunk_capacity); 1643 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1644 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1645 } 1646 } 1647 } 1648 1649 void CMSCollector::getFreelistLocks() const { 1650 // Get locks for all free lists in all generations that this 1651 // collector is responsible for 1652 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1653 } 1654 1655 void CMSCollector::releaseFreelistLocks() const { 1656 // Release locks for all free lists in all generations that this 1657 // collector is responsible for 1658 _cmsGen->freelistLock()->unlock(); 1659 } 1660 1661 bool CMSCollector::haveFreelistLocks() const { 1662 // Check locks for all free lists in all generations that this 1663 // collector is responsible for 1664 assert_lock_strong(_cmsGen->freelistLock()); 1665 PRODUCT_ONLY(ShouldNotReachHere()); 1666 return true; 1667 } 1668 1669 // A utility class that is used by the CMS collector to 1670 // temporarily "release" the foreground collector from its 1671 // usual obligation to wait for the background collector to 1672 // complete an ongoing phase before proceeding. 1673 class ReleaseForegroundGC: public StackObj { 1674 private: 1675 CMSCollector* _c; 1676 public: 1677 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1678 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1679 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1680 // allow a potentially blocked foreground collector to proceed 1681 _c->_foregroundGCShouldWait = false; 1682 if (_c->_foregroundGCIsActive) { 1683 CGC_lock->notify(); 1684 } 1685 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1686 "Possible deadlock"); 1687 } 1688 1689 ~ReleaseForegroundGC() { 1690 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1691 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1692 _c->_foregroundGCShouldWait = true; 1693 } 1694 }; 1695 1696 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1697 assert(Thread::current()->is_ConcurrentGC_thread(), 1698 "A CMS asynchronous collection is only allowed on a CMS thread."); 1699 1700 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1701 { 1702 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1703 MutexLockerEx hl(Heap_lock, safepoint_check); 1704 FreelistLocker fll(this); 1705 MutexLockerEx x(CGC_lock, safepoint_check); 1706 if (_foregroundGCIsActive) { 1707 // The foreground collector is. Skip this 1708 // background collection. 1709 assert(!_foregroundGCShouldWait, "Should be clear"); 1710 return; 1711 } else { 1712 assert(_collectorState == Idling, "Should be idling before start."); 1713 _collectorState = InitialMarking; 1714 register_gc_start(cause); 1715 // Reset the expansion cause, now that we are about to begin 1716 // a new cycle. 1717 clear_expansion_cause(); 1718 1719 // Clear the MetaspaceGC flag since a concurrent collection 1720 // is starting but also clear it after the collection. 1721 MetaspaceGC::set_should_concurrent_collect(false); 1722 } 1723 // Decide if we want to enable class unloading as part of the 1724 // ensuing concurrent GC cycle. 1725 update_should_unload_classes(); 1726 _full_gc_requested = false; // acks all outstanding full gc requests 1727 _full_gc_cause = GCCause::_no_gc; 1728 // Signal that we are about to start a collection 1729 gch->increment_total_full_collections(); // ... starting a collection cycle 1730 _collection_count_start = gch->total_full_collections(); 1731 } 1732 1733 size_t prev_used = _cmsGen->used(); 1734 1735 // The change of the collection state is normally done at this level; 1736 // the exceptions are phases that are executed while the world is 1737 // stopped. For those phases the change of state is done while the 1738 // world is stopped. For baton passing purposes this allows the 1739 // background collector to finish the phase and change state atomically. 1740 // The foreground collector cannot wait on a phase that is done 1741 // while the world is stopped because the foreground collector already 1742 // has the world stopped and would deadlock. 1743 while (_collectorState != Idling) { 1744 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1745 p2i(Thread::current()), _collectorState); 1746 // The foreground collector 1747 // holds the Heap_lock throughout its collection. 1748 // holds the CMS token (but not the lock) 1749 // except while it is waiting for the background collector to yield. 1750 // 1751 // The foreground collector should be blocked (not for long) 1752 // if the background collector is about to start a phase 1753 // executed with world stopped. If the background 1754 // collector has already started such a phase, the 1755 // foreground collector is blocked waiting for the 1756 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1757 // are executed in the VM thread. 1758 // 1759 // The locking order is 1760 // PendingListLock (PLL) -- if applicable (FinalMarking) 1761 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1762 // CMS token (claimed in 1763 // stop_world_and_do() --> 1764 // safepoint_synchronize() --> 1765 // CMSThread::synchronize()) 1766 1767 { 1768 // Check if the FG collector wants us to yield. 1769 CMSTokenSync x(true); // is cms thread 1770 if (waitForForegroundGC()) { 1771 // We yielded to a foreground GC, nothing more to be 1772 // done this round. 1773 assert(_foregroundGCShouldWait == false, "We set it to false in " 1774 "waitForForegroundGC()"); 1775 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1776 p2i(Thread::current()), _collectorState); 1777 return; 1778 } else { 1779 // The background collector can run but check to see if the 1780 // foreground collector has done a collection while the 1781 // background collector was waiting to get the CGC_lock 1782 // above. If yes, break so that _foregroundGCShouldWait 1783 // is cleared before returning. 1784 if (_collectorState == Idling) { 1785 break; 1786 } 1787 } 1788 } 1789 1790 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1791 "should be waiting"); 1792 1793 switch (_collectorState) { 1794 case InitialMarking: 1795 { 1796 ReleaseForegroundGC x(this); 1797 stats().record_cms_begin(); 1798 VM_CMS_Initial_Mark initial_mark_op(this); 1799 VMThread::execute(&initial_mark_op); 1800 } 1801 // The collector state may be any legal state at this point 1802 // since the background collector may have yielded to the 1803 // foreground collector. 1804 break; 1805 case Marking: 1806 // initial marking in checkpointRootsInitialWork has been completed 1807 if (markFromRoots()) { // we were successful 1808 assert(_collectorState == Precleaning, "Collector state should " 1809 "have changed"); 1810 } else { 1811 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1812 } 1813 break; 1814 case Precleaning: 1815 // marking from roots in markFromRoots has been completed 1816 preclean(); 1817 assert(_collectorState == AbortablePreclean || 1818 _collectorState == FinalMarking, 1819 "Collector state should have changed"); 1820 break; 1821 case AbortablePreclean: 1822 abortable_preclean(); 1823 assert(_collectorState == FinalMarking, "Collector state should " 1824 "have changed"); 1825 break; 1826 case FinalMarking: 1827 { 1828 ReleaseForegroundGC x(this); 1829 1830 VM_CMS_Final_Remark final_remark_op(this); 1831 VMThread::execute(&final_remark_op); 1832 } 1833 assert(_foregroundGCShouldWait, "block post-condition"); 1834 break; 1835 case Sweeping: 1836 // final marking in checkpointRootsFinal has been completed 1837 sweep(); 1838 assert(_collectorState == Resizing, "Collector state change " 1839 "to Resizing must be done under the free_list_lock"); 1840 1841 case Resizing: { 1842 // Sweeping has been completed... 1843 // At this point the background collection has completed. 1844 // Don't move the call to compute_new_size() down 1845 // into code that might be executed if the background 1846 // collection was preempted. 1847 { 1848 ReleaseForegroundGC x(this); // unblock FG collection 1849 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1850 CMSTokenSync z(true); // not strictly needed. 1851 if (_collectorState == Resizing) { 1852 compute_new_size(); 1853 save_heap_summary(); 1854 _collectorState = Resetting; 1855 } else { 1856 assert(_collectorState == Idling, "The state should only change" 1857 " because the foreground collector has finished the collection"); 1858 } 1859 } 1860 break; 1861 } 1862 case Resetting: 1863 // CMS heap resizing has been completed 1864 reset_concurrent(); 1865 assert(_collectorState == Idling, "Collector state should " 1866 "have changed"); 1867 1868 MetaspaceGC::set_should_concurrent_collect(false); 1869 1870 stats().record_cms_end(); 1871 // Don't move the concurrent_phases_end() and compute_new_size() 1872 // calls to here because a preempted background collection 1873 // has it's state set to "Resetting". 1874 break; 1875 case Idling: 1876 default: 1877 ShouldNotReachHere(); 1878 break; 1879 } 1880 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1881 p2i(Thread::current()), _collectorState); 1882 assert(_foregroundGCShouldWait, "block post-condition"); 1883 } 1884 1885 // Should this be in gc_epilogue? 1886 collector_policy()->counters()->update_counters(); 1887 1888 { 1889 // Clear _foregroundGCShouldWait and, in the event that the 1890 // foreground collector is waiting, notify it, before 1891 // returning. 1892 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1893 _foregroundGCShouldWait = false; 1894 if (_foregroundGCIsActive) { 1895 CGC_lock->notify(); 1896 } 1897 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1898 "Possible deadlock"); 1899 } 1900 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1901 p2i(Thread::current()), _collectorState); 1902 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1903 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1904 } 1905 1906 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1907 _cms_start_registered = true; 1908 _gc_timer_cm->register_gc_start(); 1909 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1910 } 1911 1912 void CMSCollector::register_gc_end() { 1913 if (_cms_start_registered) { 1914 report_heap_summary(GCWhen::AfterGC); 1915 1916 _gc_timer_cm->register_gc_end(); 1917 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1918 _cms_start_registered = false; 1919 } 1920 } 1921 1922 void CMSCollector::save_heap_summary() { 1923 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1924 _last_heap_summary = gch->create_heap_summary(); 1925 _last_metaspace_summary = gch->create_metaspace_summary(); 1926 } 1927 1928 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1929 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1930 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1931 } 1932 1933 bool CMSCollector::waitForForegroundGC() { 1934 bool res = false; 1935 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1936 "CMS thread should have CMS token"); 1937 // Block the foreground collector until the 1938 // background collectors decides whether to 1939 // yield. 1940 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1941 _foregroundGCShouldWait = true; 1942 if (_foregroundGCIsActive) { 1943 // The background collector yields to the 1944 // foreground collector and returns a value 1945 // indicating that it has yielded. The foreground 1946 // collector can proceed. 1947 res = true; 1948 _foregroundGCShouldWait = false; 1949 ConcurrentMarkSweepThread::clear_CMS_flag( 1950 ConcurrentMarkSweepThread::CMS_cms_has_token); 1951 ConcurrentMarkSweepThread::set_CMS_flag( 1952 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1953 // Get a possibly blocked foreground thread going 1954 CGC_lock->notify(); 1955 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1956 p2i(Thread::current()), _collectorState); 1957 while (_foregroundGCIsActive) { 1958 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1959 } 1960 ConcurrentMarkSweepThread::set_CMS_flag( 1961 ConcurrentMarkSweepThread::CMS_cms_has_token); 1962 ConcurrentMarkSweepThread::clear_CMS_flag( 1963 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1964 } 1965 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1966 p2i(Thread::current()), _collectorState); 1967 return res; 1968 } 1969 1970 // Because of the need to lock the free lists and other structures in 1971 // the collector, common to all the generations that the collector is 1972 // collecting, we need the gc_prologues of individual CMS generations 1973 // delegate to their collector. It may have been simpler had the 1974 // current infrastructure allowed one to call a prologue on a 1975 // collector. In the absence of that we have the generation's 1976 // prologue delegate to the collector, which delegates back 1977 // some "local" work to a worker method in the individual generations 1978 // that it's responsible for collecting, while itself doing any 1979 // work common to all generations it's responsible for. A similar 1980 // comment applies to the gc_epilogue()'s. 1981 // The role of the variable _between_prologue_and_epilogue is to 1982 // enforce the invocation protocol. 1983 void CMSCollector::gc_prologue(bool full) { 1984 // Call gc_prologue_work() for the CMSGen 1985 // we are responsible for. 1986 1987 // The following locking discipline assumes that we are only called 1988 // when the world is stopped. 1989 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1990 1991 // The CMSCollector prologue must call the gc_prologues for the 1992 // "generations" that it's responsible 1993 // for. 1994 1995 assert( Thread::current()->is_VM_thread() 1996 || ( CMSScavengeBeforeRemark 1997 && Thread::current()->is_ConcurrentGC_thread()), 1998 "Incorrect thread type for prologue execution"); 1999 2000 if (_between_prologue_and_epilogue) { 2001 // We have already been invoked; this is a gc_prologue delegation 2002 // from yet another CMS generation that we are responsible for, just 2003 // ignore it since all relevant work has already been done. 2004 return; 2005 } 2006 2007 // set a bit saying prologue has been called; cleared in epilogue 2008 _between_prologue_and_epilogue = true; 2009 // Claim locks for common data structures, then call gc_prologue_work() 2010 // for each CMSGen. 2011 2012 getFreelistLocks(); // gets free list locks on constituent spaces 2013 bitMapLock()->lock_without_safepoint_check(); 2014 2015 // Should call gc_prologue_work() for all cms gens we are responsible for 2016 bool duringMarking = _collectorState >= Marking 2017 && _collectorState < Sweeping; 2018 2019 // The young collections clear the modified oops state, which tells if 2020 // there are any modified oops in the class. The remark phase also needs 2021 // that information. Tell the young collection to save the union of all 2022 // modified klasses. 2023 if (duringMarking) { 2024 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2025 } 2026 2027 bool registerClosure = duringMarking; 2028 2029 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2030 2031 if (!full) { 2032 stats().record_gc0_begin(); 2033 } 2034 } 2035 2036 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2037 2038 _capacity_at_prologue = capacity(); 2039 _used_at_prologue = used(); 2040 2041 // We enable promotion tracking so that card-scanning can recognize 2042 // which objects have been promoted during this GC and skip them. 2043 for (uint i = 0; i < ParallelGCThreads; i++) { 2044 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2045 } 2046 2047 // Delegate to CMScollector which knows how to coordinate between 2048 // this and any other CMS generations that it is responsible for 2049 // collecting. 2050 collector()->gc_prologue(full); 2051 } 2052 2053 // This is a "private" interface for use by this generation's CMSCollector. 2054 // Not to be called directly by any other entity (for instance, 2055 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2056 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2057 bool registerClosure, ModUnionClosure* modUnionClosure) { 2058 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2059 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2060 "Should be NULL"); 2061 if (registerClosure) { 2062 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2063 } 2064 cmsSpace()->gc_prologue(); 2065 // Clear stat counters 2066 NOT_PRODUCT( 2067 assert(_numObjectsPromoted == 0, "check"); 2068 assert(_numWordsPromoted == 0, "check"); 2069 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2070 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2071 _numObjectsAllocated = 0; 2072 _numWordsAllocated = 0; 2073 ) 2074 } 2075 2076 void CMSCollector::gc_epilogue(bool full) { 2077 // The following locking discipline assumes that we are only called 2078 // when the world is stopped. 2079 assert(SafepointSynchronize::is_at_safepoint(), 2080 "world is stopped assumption"); 2081 2082 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2083 // if linear allocation blocks need to be appropriately marked to allow the 2084 // the blocks to be parsable. We also check here whether we need to nudge the 2085 // CMS collector thread to start a new cycle (if it's not already active). 2086 assert( Thread::current()->is_VM_thread() 2087 || ( CMSScavengeBeforeRemark 2088 && Thread::current()->is_ConcurrentGC_thread()), 2089 "Incorrect thread type for epilogue execution"); 2090 2091 if (!_between_prologue_and_epilogue) { 2092 // We have already been invoked; this is a gc_epilogue delegation 2093 // from yet another CMS generation that we are responsible for, just 2094 // ignore it since all relevant work has already been done. 2095 return; 2096 } 2097 assert(haveFreelistLocks(), "must have freelist locks"); 2098 assert_lock_strong(bitMapLock()); 2099 2100 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2101 2102 _cmsGen->gc_epilogue_work(full); 2103 2104 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2105 // in case sampling was not already enabled, enable it 2106 _start_sampling = true; 2107 } 2108 // reset _eden_chunk_array so sampling starts afresh 2109 _eden_chunk_index = 0; 2110 2111 size_t cms_used = _cmsGen->cmsSpace()->used(); 2112 2113 // update performance counters - this uses a special version of 2114 // update_counters() that allows the utilization to be passed as a 2115 // parameter, avoiding multiple calls to used(). 2116 // 2117 _cmsGen->update_counters(cms_used); 2118 2119 bitMapLock()->unlock(); 2120 releaseFreelistLocks(); 2121 2122 if (!CleanChunkPoolAsync) { 2123 Chunk::clean_chunk_pool(); 2124 } 2125 2126 set_did_compact(false); 2127 _between_prologue_and_epilogue = false; // ready for next cycle 2128 } 2129 2130 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2131 collector()->gc_epilogue(full); 2132 2133 // When using ParNew, promotion tracking should have already been 2134 // disabled. However, the prologue (which enables promotion 2135 // tracking) and epilogue are called irrespective of the type of 2136 // GC. So they will also be called before and after Full GCs, during 2137 // which promotion tracking will not be explicitly disabled. So, 2138 // it's safer to also disable it here too (to be symmetric with 2139 // enabling it in the prologue). 2140 for (uint i = 0; i < ParallelGCThreads; i++) { 2141 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2142 } 2143 } 2144 2145 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2146 assert(!incremental_collection_failed(), "Should have been cleared"); 2147 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2148 cmsSpace()->gc_epilogue(); 2149 // Print stat counters 2150 NOT_PRODUCT( 2151 assert(_numObjectsAllocated == 0, "check"); 2152 assert(_numWordsAllocated == 0, "check"); 2153 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2154 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2155 _numObjectsPromoted = 0; 2156 _numWordsPromoted = 0; 2157 ) 2158 2159 // Call down the chain in contiguous_available needs the freelistLock 2160 // so print this out before releasing the freeListLock. 2161 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2162 } 2163 2164 #ifndef PRODUCT 2165 bool CMSCollector::have_cms_token() { 2166 Thread* thr = Thread::current(); 2167 if (thr->is_VM_thread()) { 2168 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2169 } else if (thr->is_ConcurrentGC_thread()) { 2170 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2171 } else if (thr->is_GC_task_thread()) { 2172 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2173 ParGCRareEvent_lock->owned_by_self(); 2174 } 2175 return false; 2176 } 2177 2178 // Check reachability of the given heap address in CMS generation, 2179 // treating all other generations as roots. 2180 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2181 // We could "guarantee" below, rather than assert, but I'll 2182 // leave these as "asserts" so that an adventurous debugger 2183 // could try this in the product build provided some subset of 2184 // the conditions were met, provided they were interested in the 2185 // results and knew that the computation below wouldn't interfere 2186 // with other concurrent computations mutating the structures 2187 // being read or written. 2188 assert(SafepointSynchronize::is_at_safepoint(), 2189 "Else mutations in object graph will make answer suspect"); 2190 assert(have_cms_token(), "Should hold cms token"); 2191 assert(haveFreelistLocks(), "must hold free list locks"); 2192 assert_lock_strong(bitMapLock()); 2193 2194 // Clear the marking bit map array before starting, but, just 2195 // for kicks, first report if the given address is already marked 2196 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2197 _markBitMap.isMarked(addr) ? "" : " not"); 2198 2199 if (verify_after_remark()) { 2200 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2201 bool result = verification_mark_bm()->isMarked(addr); 2202 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2203 result ? "IS" : "is NOT"); 2204 return result; 2205 } else { 2206 tty->print_cr("Could not compute result"); 2207 return false; 2208 } 2209 } 2210 #endif 2211 2212 void 2213 CMSCollector::print_on_error(outputStream* st) { 2214 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2215 if (collector != NULL) { 2216 CMSBitMap* bitmap = &collector->_markBitMap; 2217 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2218 bitmap->print_on_error(st, " Bits: "); 2219 2220 st->cr(); 2221 2222 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2223 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2224 mut_bitmap->print_on_error(st, " Bits: "); 2225 } 2226 } 2227 2228 //////////////////////////////////////////////////////// 2229 // CMS Verification Support 2230 //////////////////////////////////////////////////////// 2231 // Following the remark phase, the following invariant 2232 // should hold -- each object in the CMS heap which is 2233 // marked in markBitMap() should be marked in the verification_mark_bm(). 2234 2235 class VerifyMarkedClosure: public BitMapClosure { 2236 CMSBitMap* _marks; 2237 bool _failed; 2238 2239 public: 2240 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2241 2242 bool do_bit(size_t offset) { 2243 HeapWord* addr = _marks->offsetToHeapWord(offset); 2244 if (!_marks->isMarked(addr)) { 2245 Log(gc, verify) log; 2246 ResourceMark rm; 2247 oop(addr)->print_on(log.error_stream()); 2248 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2249 _failed = true; 2250 } 2251 return true; 2252 } 2253 2254 bool failed() { return _failed; } 2255 }; 2256 2257 bool CMSCollector::verify_after_remark() { 2258 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2259 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2260 static bool init = false; 2261 2262 assert(SafepointSynchronize::is_at_safepoint(), 2263 "Else mutations in object graph will make answer suspect"); 2264 assert(have_cms_token(), 2265 "Else there may be mutual interference in use of " 2266 " verification data structures"); 2267 assert(_collectorState > Marking && _collectorState <= Sweeping, 2268 "Else marking info checked here may be obsolete"); 2269 assert(haveFreelistLocks(), "must hold free list locks"); 2270 assert_lock_strong(bitMapLock()); 2271 2272 2273 // Allocate marking bit map if not already allocated 2274 if (!init) { // first time 2275 if (!verification_mark_bm()->allocate(_span)) { 2276 return false; 2277 } 2278 init = true; 2279 } 2280 2281 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2282 2283 // Turn off refs discovery -- so we will be tracing through refs. 2284 // This is as intended, because by this time 2285 // GC must already have cleared any refs that need to be cleared, 2286 // and traced those that need to be marked; moreover, 2287 // the marking done here is not going to interfere in any 2288 // way with the marking information used by GC. 2289 NoRefDiscovery no_discovery(ref_processor()); 2290 2291 #if defined(COMPILER2) || INCLUDE_JVMCI 2292 DerivedPointerTableDeactivate dpt_deact; 2293 #endif 2294 2295 // Clear any marks from a previous round 2296 verification_mark_bm()->clear_all(); 2297 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2298 verify_work_stacks_empty(); 2299 2300 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2301 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2302 // Update the saved marks which may affect the root scans. 2303 gch->save_marks(); 2304 2305 if (CMSRemarkVerifyVariant == 1) { 2306 // In this first variant of verification, we complete 2307 // all marking, then check if the new marks-vector is 2308 // a subset of the CMS marks-vector. 2309 verify_after_remark_work_1(); 2310 } else { 2311 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2312 // In this second variant of verification, we flag an error 2313 // (i.e. an object reachable in the new marks-vector not reachable 2314 // in the CMS marks-vector) immediately, also indicating the 2315 // identify of an object (A) that references the unmarked object (B) -- 2316 // presumably, a mutation to A failed to be picked up by preclean/remark? 2317 verify_after_remark_work_2(); 2318 } 2319 2320 return true; 2321 } 2322 2323 void CMSCollector::verify_after_remark_work_1() { 2324 ResourceMark rm; 2325 HandleMark hm; 2326 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2327 2328 // Get a clear set of claim bits for the roots processing to work with. 2329 ClassLoaderDataGraph::clear_claimed_marks(); 2330 2331 // Mark from roots one level into CMS 2332 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2333 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2334 2335 { 2336 StrongRootsScope srs(1); 2337 2338 gch->gen_process_roots(&srs, 2339 GenCollectedHeap::OldGen, 2340 true, // young gen as roots 2341 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2342 should_unload_classes(), 2343 ¬Older, 2344 NULL, 2345 NULL); 2346 } 2347 2348 // Now mark from the roots 2349 MarkFromRootsClosure markFromRootsClosure(this, _span, 2350 verification_mark_bm(), verification_mark_stack(), 2351 false /* don't yield */, true /* verifying */); 2352 assert(_restart_addr == NULL, "Expected pre-condition"); 2353 verification_mark_bm()->iterate(&markFromRootsClosure); 2354 while (_restart_addr != NULL) { 2355 // Deal with stack overflow: by restarting at the indicated 2356 // address. 2357 HeapWord* ra = _restart_addr; 2358 markFromRootsClosure.reset(ra); 2359 _restart_addr = NULL; 2360 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2361 } 2362 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2363 verify_work_stacks_empty(); 2364 2365 // Marking completed -- now verify that each bit marked in 2366 // verification_mark_bm() is also marked in markBitMap(); flag all 2367 // errors by printing corresponding objects. 2368 VerifyMarkedClosure vcl(markBitMap()); 2369 verification_mark_bm()->iterate(&vcl); 2370 if (vcl.failed()) { 2371 Log(gc, verify) log; 2372 log.error("Failed marking verification after remark"); 2373 ResourceMark rm; 2374 gch->print_on(log.error_stream()); 2375 fatal("CMS: failed marking verification after remark"); 2376 } 2377 } 2378 2379 class VerifyKlassOopsKlassClosure : public KlassClosure { 2380 class VerifyKlassOopsClosure : public OopClosure { 2381 CMSBitMap* _bitmap; 2382 public: 2383 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2384 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2385 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2386 } _oop_closure; 2387 public: 2388 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2389 void do_klass(Klass* k) { 2390 k->oops_do(&_oop_closure); 2391 } 2392 }; 2393 2394 void CMSCollector::verify_after_remark_work_2() { 2395 ResourceMark rm; 2396 HandleMark hm; 2397 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2398 2399 // Get a clear set of claim bits for the roots processing to work with. 2400 ClassLoaderDataGraph::clear_claimed_marks(); 2401 2402 // Mark from roots one level into CMS 2403 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2404 markBitMap()); 2405 CLDToOopClosure cld_closure(¬Older, true); 2406 2407 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2408 2409 { 2410 StrongRootsScope srs(1); 2411 2412 gch->gen_process_roots(&srs, 2413 GenCollectedHeap::OldGen, 2414 true, // young gen as roots 2415 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2416 should_unload_classes(), 2417 ¬Older, 2418 NULL, 2419 &cld_closure); 2420 } 2421 2422 // Now mark from the roots 2423 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2424 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2425 assert(_restart_addr == NULL, "Expected pre-condition"); 2426 verification_mark_bm()->iterate(&markFromRootsClosure); 2427 while (_restart_addr != NULL) { 2428 // Deal with stack overflow: by restarting at the indicated 2429 // address. 2430 HeapWord* ra = _restart_addr; 2431 markFromRootsClosure.reset(ra); 2432 _restart_addr = NULL; 2433 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2434 } 2435 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2436 verify_work_stacks_empty(); 2437 2438 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2439 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2440 2441 // Marking completed -- now verify that each bit marked in 2442 // verification_mark_bm() is also marked in markBitMap(); flag all 2443 // errors by printing corresponding objects. 2444 VerifyMarkedClosure vcl(markBitMap()); 2445 verification_mark_bm()->iterate(&vcl); 2446 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2447 } 2448 2449 void ConcurrentMarkSweepGeneration::save_marks() { 2450 // delegate to CMS space 2451 cmsSpace()->save_marks(); 2452 } 2453 2454 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2455 return cmsSpace()->no_allocs_since_save_marks(); 2456 } 2457 2458 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2459 \ 2460 void ConcurrentMarkSweepGeneration:: \ 2461 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2462 cl->set_generation(this); \ 2463 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2464 cl->reset_generation(); \ 2465 save_marks(); \ 2466 } 2467 2468 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2469 2470 void 2471 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2472 if (freelistLock()->owned_by_self()) { 2473 Generation::oop_iterate(cl); 2474 } else { 2475 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2476 Generation::oop_iterate(cl); 2477 } 2478 } 2479 2480 void 2481 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2482 if (freelistLock()->owned_by_self()) { 2483 Generation::object_iterate(cl); 2484 } else { 2485 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2486 Generation::object_iterate(cl); 2487 } 2488 } 2489 2490 void 2491 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2492 if (freelistLock()->owned_by_self()) { 2493 Generation::safe_object_iterate(cl); 2494 } else { 2495 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2496 Generation::safe_object_iterate(cl); 2497 } 2498 } 2499 2500 void 2501 ConcurrentMarkSweepGeneration::post_compact() { 2502 } 2503 2504 void 2505 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2506 // Fix the linear allocation blocks to look like free blocks. 2507 2508 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2509 // are not called when the heap is verified during universe initialization and 2510 // at vm shutdown. 2511 if (freelistLock()->owned_by_self()) { 2512 cmsSpace()->prepare_for_verify(); 2513 } else { 2514 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2515 cmsSpace()->prepare_for_verify(); 2516 } 2517 } 2518 2519 void 2520 ConcurrentMarkSweepGeneration::verify() { 2521 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2522 // are not called when the heap is verified during universe initialization and 2523 // at vm shutdown. 2524 if (freelistLock()->owned_by_self()) { 2525 cmsSpace()->verify(); 2526 } else { 2527 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2528 cmsSpace()->verify(); 2529 } 2530 } 2531 2532 void CMSCollector::verify() { 2533 _cmsGen->verify(); 2534 } 2535 2536 #ifndef PRODUCT 2537 bool CMSCollector::overflow_list_is_empty() const { 2538 assert(_num_par_pushes >= 0, "Inconsistency"); 2539 if (_overflow_list == NULL) { 2540 assert(_num_par_pushes == 0, "Inconsistency"); 2541 } 2542 return _overflow_list == NULL; 2543 } 2544 2545 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2546 // merely consolidate assertion checks that appear to occur together frequently. 2547 void CMSCollector::verify_work_stacks_empty() const { 2548 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2549 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2550 } 2551 2552 void CMSCollector::verify_overflow_empty() const { 2553 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2554 assert(no_preserved_marks(), "No preserved marks"); 2555 } 2556 #endif // PRODUCT 2557 2558 // Decide if we want to enable class unloading as part of the 2559 // ensuing concurrent GC cycle. We will collect and 2560 // unload classes if it's the case that: 2561 // (1) an explicit gc request has been made and the flag 2562 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2563 // (2) (a) class unloading is enabled at the command line, and 2564 // (b) old gen is getting really full 2565 // NOTE: Provided there is no change in the state of the heap between 2566 // calls to this method, it should have idempotent results. Moreover, 2567 // its results should be monotonically increasing (i.e. going from 0 to 1, 2568 // but not 1 to 0) between successive calls between which the heap was 2569 // not collected. For the implementation below, it must thus rely on 2570 // the property that concurrent_cycles_since_last_unload() 2571 // will not decrease unless a collection cycle happened and that 2572 // _cmsGen->is_too_full() are 2573 // themselves also monotonic in that sense. See check_monotonicity() 2574 // below. 2575 void CMSCollector::update_should_unload_classes() { 2576 _should_unload_classes = false; 2577 // Condition 1 above 2578 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2579 _should_unload_classes = true; 2580 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2581 // Disjuncts 2.b.(i,ii,iii) above 2582 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2583 CMSClassUnloadingMaxInterval) 2584 || _cmsGen->is_too_full(); 2585 } 2586 } 2587 2588 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2589 bool res = should_concurrent_collect(); 2590 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2591 return res; 2592 } 2593 2594 void CMSCollector::setup_cms_unloading_and_verification_state() { 2595 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2596 || VerifyBeforeExit; 2597 const int rso = GenCollectedHeap::SO_AllCodeCache; 2598 2599 // We set the proper root for this CMS cycle here. 2600 if (should_unload_classes()) { // Should unload classes this cycle 2601 remove_root_scanning_option(rso); // Shrink the root set appropriately 2602 set_verifying(should_verify); // Set verification state for this cycle 2603 return; // Nothing else needs to be done at this time 2604 } 2605 2606 // Not unloading classes this cycle 2607 assert(!should_unload_classes(), "Inconsistency!"); 2608 2609 // If we are not unloading classes then add SO_AllCodeCache to root 2610 // scanning options. 2611 add_root_scanning_option(rso); 2612 2613 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2614 set_verifying(true); 2615 } else if (verifying() && !should_verify) { 2616 // We were verifying, but some verification flags got disabled. 2617 set_verifying(false); 2618 // Exclude symbols, strings and code cache elements from root scanning to 2619 // reduce IM and RM pauses. 2620 remove_root_scanning_option(rso); 2621 } 2622 } 2623 2624 2625 #ifndef PRODUCT 2626 HeapWord* CMSCollector::block_start(const void* p) const { 2627 const HeapWord* addr = (HeapWord*)p; 2628 if (_span.contains(p)) { 2629 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2630 return _cmsGen->cmsSpace()->block_start(p); 2631 } 2632 } 2633 return NULL; 2634 } 2635 #endif 2636 2637 HeapWord* 2638 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2639 bool tlab, 2640 bool parallel) { 2641 CMSSynchronousYieldRequest yr; 2642 assert(!tlab, "Can't deal with TLAB allocation"); 2643 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2644 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2645 if (GCExpandToAllocateDelayMillis > 0) { 2646 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2647 } 2648 return have_lock_and_allocate(word_size, tlab); 2649 } 2650 2651 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2652 size_t bytes, 2653 size_t expand_bytes, 2654 CMSExpansionCause::Cause cause) 2655 { 2656 2657 bool success = expand(bytes, expand_bytes); 2658 2659 // remember why we expanded; this information is used 2660 // by shouldConcurrentCollect() when making decisions on whether to start 2661 // a new CMS cycle. 2662 if (success) { 2663 set_expansion_cause(cause); 2664 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2665 } 2666 } 2667 2668 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2669 HeapWord* res = NULL; 2670 MutexLocker x(ParGCRareEvent_lock); 2671 while (true) { 2672 // Expansion by some other thread might make alloc OK now: 2673 res = ps->lab.alloc(word_sz); 2674 if (res != NULL) return res; 2675 // If there's not enough expansion space available, give up. 2676 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2677 return NULL; 2678 } 2679 // Otherwise, we try expansion. 2680 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2681 // Now go around the loop and try alloc again; 2682 // A competing par_promote might beat us to the expansion space, 2683 // so we may go around the loop again if promotion fails again. 2684 if (GCExpandToAllocateDelayMillis > 0) { 2685 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2686 } 2687 } 2688 } 2689 2690 2691 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2692 PromotionInfo* promo) { 2693 MutexLocker x(ParGCRareEvent_lock); 2694 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2695 while (true) { 2696 // Expansion by some other thread might make alloc OK now: 2697 if (promo->ensure_spooling_space()) { 2698 assert(promo->has_spooling_space(), 2699 "Post-condition of successful ensure_spooling_space()"); 2700 return true; 2701 } 2702 // If there's not enough expansion space available, give up. 2703 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2704 return false; 2705 } 2706 // Otherwise, we try expansion. 2707 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2708 // Now go around the loop and try alloc again; 2709 // A competing allocation might beat us to the expansion space, 2710 // so we may go around the loop again if allocation fails again. 2711 if (GCExpandToAllocateDelayMillis > 0) { 2712 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2713 } 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2718 // Only shrink if a compaction was done so that all the free space 2719 // in the generation is in a contiguous block at the end. 2720 if (did_compact()) { 2721 CardGeneration::shrink(bytes); 2722 } 2723 } 2724 2725 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2726 assert_locked_or_safepoint(Heap_lock); 2727 } 2728 2729 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2730 assert_locked_or_safepoint(Heap_lock); 2731 assert_lock_strong(freelistLock()); 2732 log_trace(gc)("Shrinking of CMS not yet implemented"); 2733 return; 2734 } 2735 2736 2737 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2738 // phases. 2739 class CMSPhaseAccounting: public StackObj { 2740 public: 2741 CMSPhaseAccounting(CMSCollector *collector, 2742 const char *title); 2743 ~CMSPhaseAccounting(); 2744 2745 private: 2746 CMSCollector *_collector; 2747 const char *_title; 2748 GCTraceConcTime(Info, gc) _trace_time; 2749 2750 public: 2751 // Not MT-safe; so do not pass around these StackObj's 2752 // where they may be accessed by other threads. 2753 double wallclock_millis() { 2754 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2755 } 2756 }; 2757 2758 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2759 const char *title) : 2760 _collector(collector), _title(title), _trace_time(title) { 2761 2762 _collector->resetYields(); 2763 _collector->resetTimer(); 2764 _collector->startTimer(); 2765 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2766 } 2767 2768 CMSPhaseAccounting::~CMSPhaseAccounting() { 2769 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2770 _collector->stopTimer(); 2771 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2772 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2773 } 2774 2775 // CMS work 2776 2777 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2778 class CMSParMarkTask : public AbstractGangTask { 2779 protected: 2780 CMSCollector* _collector; 2781 uint _n_workers; 2782 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2783 AbstractGangTask(name), 2784 _collector(collector), 2785 _n_workers(n_workers) {} 2786 // Work method in support of parallel rescan ... of young gen spaces 2787 void do_young_space_rescan(OopsInGenClosure* cl, 2788 ContiguousSpace* space, 2789 HeapWord** chunk_array, size_t chunk_top); 2790 void work_on_young_gen_roots(OopsInGenClosure* cl); 2791 }; 2792 2793 // Parallel initial mark task 2794 class CMSParInitialMarkTask: public CMSParMarkTask { 2795 StrongRootsScope* _strong_roots_scope; 2796 public: 2797 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2798 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2799 _strong_roots_scope(strong_roots_scope) {} 2800 void work(uint worker_id); 2801 }; 2802 2803 // Checkpoint the roots into this generation from outside 2804 // this generation. [Note this initial checkpoint need only 2805 // be approximate -- we'll do a catch up phase subsequently.] 2806 void CMSCollector::checkpointRootsInitial() { 2807 assert(_collectorState == InitialMarking, "Wrong collector state"); 2808 check_correct_thread_executing(); 2809 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2810 2811 save_heap_summary(); 2812 report_heap_summary(GCWhen::BeforeGC); 2813 2814 ReferenceProcessor* rp = ref_processor(); 2815 assert(_restart_addr == NULL, "Control point invariant"); 2816 { 2817 // acquire locks for subsequent manipulations 2818 MutexLockerEx x(bitMapLock(), 2819 Mutex::_no_safepoint_check_flag); 2820 checkpointRootsInitialWork(); 2821 // enable ("weak") refs discovery 2822 rp->enable_discovery(); 2823 _collectorState = Marking; 2824 } 2825 } 2826 2827 void CMSCollector::checkpointRootsInitialWork() { 2828 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2829 assert(_collectorState == InitialMarking, "just checking"); 2830 2831 // Already have locks. 2832 assert_lock_strong(bitMapLock()); 2833 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2834 2835 // Setup the verification and class unloading state for this 2836 // CMS collection cycle. 2837 setup_cms_unloading_and_verification_state(); 2838 2839 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2840 2841 // Reset all the PLAB chunk arrays if necessary. 2842 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2843 reset_survivor_plab_arrays(); 2844 } 2845 2846 ResourceMark rm; 2847 HandleMark hm; 2848 2849 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2850 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2851 2852 verify_work_stacks_empty(); 2853 verify_overflow_empty(); 2854 2855 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2856 // Update the saved marks which may affect the root scans. 2857 gch->save_marks(); 2858 2859 // weak reference processing has not started yet. 2860 ref_processor()->set_enqueuing_is_done(false); 2861 2862 // Need to remember all newly created CLDs, 2863 // so that we can guarantee that the remark finds them. 2864 ClassLoaderDataGraph::remember_new_clds(true); 2865 2866 // Whenever a CLD is found, it will be claimed before proceeding to mark 2867 // the klasses. The claimed marks need to be cleared before marking starts. 2868 ClassLoaderDataGraph::clear_claimed_marks(); 2869 2870 print_eden_and_survivor_chunk_arrays(); 2871 2872 { 2873 #if defined(COMPILER2) || INCLUDE_JVMCI 2874 DerivedPointerTableDeactivate dpt_deact; 2875 #endif 2876 if (CMSParallelInitialMarkEnabled) { 2877 // The parallel version. 2878 WorkGang* workers = gch->workers(); 2879 assert(workers != NULL, "Need parallel worker threads."); 2880 uint n_workers = workers->active_workers(); 2881 2882 StrongRootsScope srs(n_workers); 2883 2884 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2885 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2886 if (n_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 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2895 2896 StrongRootsScope srs(1); 2897 2898 gch->gen_process_roots(&srs, 2899 GenCollectedHeap::OldGen, 2900 true, // young gen as roots 2901 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2902 should_unload_classes(), 2903 ¬Older, 2904 NULL, 2905 &cld_closure); 2906 } 2907 } 2908 2909 // Clear mod-union table; it will be dirtied in the prologue of 2910 // CMS generation per each young generation collection. 2911 2912 assert(_modUnionTable.isAllClear(), 2913 "Was cleared in most recent final checkpoint phase" 2914 " or no bits are set in the gc_prologue before the start of the next " 2915 "subsequent marking phase."); 2916 2917 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2918 2919 // Save the end of the used_region of the constituent generations 2920 // to be used to limit the extent of sweep in each generation. 2921 save_sweep_limits(); 2922 verify_overflow_empty(); 2923 } 2924 2925 bool CMSCollector::markFromRoots() { 2926 // we might be tempted to assert that: 2927 // assert(!SafepointSynchronize::is_at_safepoint(), 2928 // "inconsistent argument?"); 2929 // However that wouldn't be right, because it's possible that 2930 // a safepoint is indeed in progress as a young generation 2931 // stop-the-world GC happens even as we mark in this generation. 2932 assert(_collectorState == Marking, "inconsistent state?"); 2933 check_correct_thread_executing(); 2934 verify_overflow_empty(); 2935 2936 // Weak ref discovery note: We may be discovering weak 2937 // refs in this generation concurrent (but interleaved) with 2938 // weak ref discovery by the young generation collector. 2939 2940 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2941 GCTraceCPUTime tcpu; 2942 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2943 bool res = markFromRootsWork(); 2944 if (res) { 2945 _collectorState = Precleaning; 2946 } else { // We failed and a foreground collection wants to take over 2947 assert(_foregroundGCIsActive, "internal state inconsistency"); 2948 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2949 log_debug(gc)("bailing out to foreground collection"); 2950 } 2951 verify_overflow_empty(); 2952 return res; 2953 } 2954 2955 bool CMSCollector::markFromRootsWork() { 2956 // iterate over marked bits in bit map, doing a full scan and mark 2957 // from these roots using the following algorithm: 2958 // . if oop is to the right of the current scan pointer, 2959 // mark corresponding bit (we'll process it later) 2960 // . else (oop is to left of current scan pointer) 2961 // push oop on marking stack 2962 // . drain the marking stack 2963 2964 // Note that when we do a marking step we need to hold the 2965 // bit map lock -- recall that direct allocation (by mutators) 2966 // and promotion (by the young generation collector) is also 2967 // marking the bit map. [the so-called allocate live policy.] 2968 // Because the implementation of bit map marking is not 2969 // robust wrt simultaneous marking of bits in the same word, 2970 // we need to make sure that there is no such interference 2971 // between concurrent such updates. 2972 2973 // already have locks 2974 assert_lock_strong(bitMapLock()); 2975 2976 verify_work_stacks_empty(); 2977 verify_overflow_empty(); 2978 bool result = false; 2979 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2980 result = do_marking_mt(); 2981 } else { 2982 result = do_marking_st(); 2983 } 2984 return result; 2985 } 2986 2987 // Forward decl 2988 class CMSConcMarkingTask; 2989 2990 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2991 CMSCollector* _collector; 2992 CMSConcMarkingTask* _task; 2993 public: 2994 virtual void yield(); 2995 2996 // "n_threads" is the number of threads to be terminated. 2997 // "queue_set" is a set of work queues of other threads. 2998 // "collector" is the CMS collector associated with this task terminator. 2999 // "yield" indicates whether we need the gang as a whole to yield. 3000 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3001 ParallelTaskTerminator(n_threads, queue_set), 3002 _collector(collector) { } 3003 3004 void set_task(CMSConcMarkingTask* task) { 3005 _task = task; 3006 } 3007 }; 3008 3009 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3010 CMSConcMarkingTask* _task; 3011 public: 3012 bool should_exit_termination(); 3013 void set_task(CMSConcMarkingTask* task) { 3014 _task = task; 3015 } 3016 }; 3017 3018 // MT Concurrent Marking Task 3019 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3020 CMSCollector* _collector; 3021 uint _n_workers; // requested/desired # workers 3022 bool _result; 3023 CompactibleFreeListSpace* _cms_space; 3024 char _pad_front[64]; // padding to ... 3025 HeapWord* _global_finger; // ... avoid sharing cache line 3026 char _pad_back[64]; 3027 HeapWord* _restart_addr; 3028 3029 // Exposed here for yielding support 3030 Mutex* const _bit_map_lock; 3031 3032 // The per thread work queues, available here for stealing 3033 OopTaskQueueSet* _task_queues; 3034 3035 // Termination (and yielding) support 3036 CMSConcMarkingTerminator _term; 3037 CMSConcMarkingTerminatorTerminator _term_term; 3038 3039 public: 3040 CMSConcMarkingTask(CMSCollector* collector, 3041 CompactibleFreeListSpace* cms_space, 3042 YieldingFlexibleWorkGang* workers, 3043 OopTaskQueueSet* task_queues): 3044 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3045 _collector(collector), 3046 _cms_space(cms_space), 3047 _n_workers(0), _result(true), 3048 _task_queues(task_queues), 3049 _term(_n_workers, task_queues, _collector), 3050 _bit_map_lock(collector->bitMapLock()) 3051 { 3052 _requested_size = _n_workers; 3053 _term.set_task(this); 3054 _term_term.set_task(this); 3055 _restart_addr = _global_finger = _cms_space->bottom(); 3056 } 3057 3058 3059 OopTaskQueueSet* task_queues() { return _task_queues; } 3060 3061 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3062 3063 HeapWord** global_finger_addr() { return &_global_finger; } 3064 3065 CMSConcMarkingTerminator* terminator() { return &_term; } 3066 3067 virtual void set_for_termination(uint active_workers) { 3068 terminator()->reset_for_reuse(active_workers); 3069 } 3070 3071 void work(uint worker_id); 3072 bool should_yield() { 3073 return ConcurrentMarkSweepThread::should_yield() 3074 && !_collector->foregroundGCIsActive(); 3075 } 3076 3077 virtual void coordinator_yield(); // stuff done by coordinator 3078 bool result() { return _result; } 3079 3080 void reset(HeapWord* ra) { 3081 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3082 _restart_addr = _global_finger = ra; 3083 _term.reset_for_reuse(); 3084 } 3085 3086 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3087 OopTaskQueue* work_q); 3088 3089 private: 3090 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3091 void do_work_steal(int i); 3092 void bump_global_finger(HeapWord* f); 3093 }; 3094 3095 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3096 assert(_task != NULL, "Error"); 3097 return _task->yielding(); 3098 // Note that we do not need the disjunct || _task->should_yield() above 3099 // because we want terminating threads to yield only if the task 3100 // is already in the midst of yielding, which happens only after at least one 3101 // thread has yielded. 3102 } 3103 3104 void CMSConcMarkingTerminator::yield() { 3105 if (_task->should_yield()) { 3106 _task->yield(); 3107 } else { 3108 ParallelTaskTerminator::yield(); 3109 } 3110 } 3111 3112 //////////////////////////////////////////////////////////////// 3113 // Concurrent Marking Algorithm Sketch 3114 //////////////////////////////////////////////////////////////// 3115 // Until all tasks exhausted (both spaces): 3116 // -- claim next available chunk 3117 // -- bump global finger via CAS 3118 // -- find first object that starts in this chunk 3119 // and start scanning bitmap from that position 3120 // -- scan marked objects for oops 3121 // -- CAS-mark target, and if successful: 3122 // . if target oop is above global finger (volatile read) 3123 // nothing to do 3124 // . if target oop is in chunk and above local finger 3125 // then nothing to do 3126 // . else push on work-queue 3127 // -- Deal with possible overflow issues: 3128 // . local work-queue overflow causes stuff to be pushed on 3129 // global (common) overflow queue 3130 // . always first empty local work queue 3131 // . then get a batch of oops from global work queue if any 3132 // . then do work stealing 3133 // -- When all tasks claimed (both spaces) 3134 // and local work queue empty, 3135 // then in a loop do: 3136 // . check global overflow stack; steal a batch of oops and trace 3137 // . try to steal from other threads oif GOS is empty 3138 // . if neither is available, offer termination 3139 // -- Terminate and return result 3140 // 3141 void CMSConcMarkingTask::work(uint worker_id) { 3142 elapsedTimer _timer; 3143 ResourceMark rm; 3144 HandleMark hm; 3145 3146 DEBUG_ONLY(_collector->verify_overflow_empty();) 3147 3148 // Before we begin work, our work queue should be empty 3149 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3150 // Scan the bitmap covering _cms_space, tracing through grey objects. 3151 _timer.start(); 3152 do_scan_and_mark(worker_id, _cms_space); 3153 _timer.stop(); 3154 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3155 3156 // ... do work stealing 3157 _timer.reset(); 3158 _timer.start(); 3159 do_work_steal(worker_id); 3160 _timer.stop(); 3161 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3162 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3163 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3164 // Note that under the current task protocol, the 3165 // following assertion is true even of the spaces 3166 // expanded since the completion of the concurrent 3167 // marking. XXX This will likely change under a strict 3168 // ABORT semantics. 3169 // After perm removal the comparison was changed to 3170 // greater than or equal to from strictly greater than. 3171 // Before perm removal the highest address sweep would 3172 // have been at the end of perm gen but now is at the 3173 // end of the tenured gen. 3174 assert(_global_finger >= _cms_space->end(), 3175 "All tasks have been completed"); 3176 DEBUG_ONLY(_collector->verify_overflow_empty();) 3177 } 3178 3179 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3180 HeapWord* read = _global_finger; 3181 HeapWord* cur = read; 3182 while (f > read) { 3183 cur = read; 3184 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3185 if (cur == read) { 3186 // our cas succeeded 3187 assert(_global_finger >= f, "protocol consistency"); 3188 break; 3189 } 3190 } 3191 } 3192 3193 // This is really inefficient, and should be redone by 3194 // using (not yet available) block-read and -write interfaces to the 3195 // stack and the work_queue. XXX FIX ME !!! 3196 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3197 OopTaskQueue* work_q) { 3198 // Fast lock-free check 3199 if (ovflw_stk->length() == 0) { 3200 return false; 3201 } 3202 assert(work_q->size() == 0, "Shouldn't steal"); 3203 MutexLockerEx ml(ovflw_stk->par_lock(), 3204 Mutex::_no_safepoint_check_flag); 3205 // Grab up to 1/4 the size of the work queue 3206 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3207 (size_t)ParGCDesiredObjsFromOverflowList); 3208 num = MIN2(num, ovflw_stk->length()); 3209 for (int i = (int) num; i > 0; i--) { 3210 oop cur = ovflw_stk->pop(); 3211 assert(cur != NULL, "Counted wrong?"); 3212 work_q->push(cur); 3213 } 3214 return num > 0; 3215 } 3216 3217 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3218 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3219 int n_tasks = pst->n_tasks(); 3220 // We allow that there may be no tasks to do here because 3221 // we are restarting after a stack overflow. 3222 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3223 uint nth_task = 0; 3224 3225 HeapWord* aligned_start = sp->bottom(); 3226 if (sp->used_region().contains(_restart_addr)) { 3227 // Align down to a card boundary for the start of 0th task 3228 // for this space. 3229 aligned_start = 3230 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3231 CardTableModRefBS::card_size); 3232 } 3233 3234 size_t chunk_size = sp->marking_task_size(); 3235 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3236 // Having claimed the nth task in this space, 3237 // compute the chunk that it corresponds to: 3238 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3239 aligned_start + (nth_task+1)*chunk_size); 3240 // Try and bump the global finger via a CAS; 3241 // note that we need to do the global finger bump 3242 // _before_ taking the intersection below, because 3243 // the task corresponding to that region will be 3244 // deemed done even if the used_region() expands 3245 // because of allocation -- as it almost certainly will 3246 // during start-up while the threads yield in the 3247 // closure below. 3248 HeapWord* finger = span.end(); 3249 bump_global_finger(finger); // atomically 3250 // There are null tasks here corresponding to chunks 3251 // beyond the "top" address of the space. 3252 span = span.intersection(sp->used_region()); 3253 if (!span.is_empty()) { // Non-null task 3254 HeapWord* prev_obj; 3255 assert(!span.contains(_restart_addr) || nth_task == 0, 3256 "Inconsistency"); 3257 if (nth_task == 0) { 3258 // For the 0th task, we'll not need to compute a block_start. 3259 if (span.contains(_restart_addr)) { 3260 // In the case of a restart because of stack overflow, 3261 // we might additionally skip a chunk prefix. 3262 prev_obj = _restart_addr; 3263 } else { 3264 prev_obj = span.start(); 3265 } 3266 } else { 3267 // We want to skip the first object because 3268 // the protocol is to scan any object in its entirety 3269 // that _starts_ in this span; a fortiori, any 3270 // object starting in an earlier span is scanned 3271 // as part of an earlier claimed task. 3272 // Below we use the "careful" version of block_start 3273 // so we do not try to navigate uninitialized objects. 3274 prev_obj = sp->block_start_careful(span.start()); 3275 // Below we use a variant of block_size that uses the 3276 // Printezis bits to avoid waiting for allocated 3277 // objects to become initialized/parsable. 3278 while (prev_obj < span.start()) { 3279 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3280 if (sz > 0) { 3281 prev_obj += sz; 3282 } else { 3283 // In this case we may end up doing a bit of redundant 3284 // scanning, but that appears unavoidable, short of 3285 // locking the free list locks; see bug 6324141. 3286 break; 3287 } 3288 } 3289 } 3290 if (prev_obj < span.end()) { 3291 MemRegion my_span = MemRegion(prev_obj, span.end()); 3292 // Do the marking work within a non-empty span -- 3293 // the last argument to the constructor indicates whether the 3294 // iteration should be incremental with periodic yields. 3295 ParMarkFromRootsClosure cl(this, _collector, my_span, 3296 &_collector->_markBitMap, 3297 work_queue(i), 3298 &_collector->_markStack); 3299 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3300 } // else nothing to do for this task 3301 } // else nothing to do for this task 3302 } 3303 // We'd be tempted to assert here that since there are no 3304 // more tasks left to claim in this space, the global_finger 3305 // must exceed space->top() and a fortiori space->end(). However, 3306 // that would not quite be correct because the bumping of 3307 // global_finger occurs strictly after the claiming of a task, 3308 // so by the time we reach here the global finger may not yet 3309 // have been bumped up by the thread that claimed the last 3310 // task. 3311 pst->all_tasks_completed(); 3312 } 3313 3314 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3315 private: 3316 CMSCollector* _collector; 3317 CMSConcMarkingTask* _task; 3318 MemRegion _span; 3319 CMSBitMap* _bit_map; 3320 CMSMarkStack* _overflow_stack; 3321 OopTaskQueue* _work_queue; 3322 protected: 3323 DO_OOP_WORK_DEFN 3324 public: 3325 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3326 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3327 MetadataAwareOopClosure(collector->ref_processor()), 3328 _collector(collector), 3329 _task(task), 3330 _span(collector->_span), 3331 _work_queue(work_queue), 3332 _bit_map(bit_map), 3333 _overflow_stack(overflow_stack) 3334 { } 3335 virtual void do_oop(oop* p); 3336 virtual void do_oop(narrowOop* p); 3337 3338 void trim_queue(size_t max); 3339 void handle_stack_overflow(HeapWord* lost); 3340 void do_yield_check() { 3341 if (_task->should_yield()) { 3342 _task->yield(); 3343 } 3344 } 3345 }; 3346 3347 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3348 3349 // Grey object scanning during work stealing phase -- 3350 // the salient assumption here is that any references 3351 // that are in these stolen objects being scanned must 3352 // already have been initialized (else they would not have 3353 // been published), so we do not need to check for 3354 // uninitialized objects before pushing here. 3355 void ParConcMarkingClosure::do_oop(oop obj) { 3356 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3357 HeapWord* addr = (HeapWord*)obj; 3358 // Check if oop points into the CMS generation 3359 // and is not marked 3360 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3361 // a white object ... 3362 // If we manage to "claim" the object, by being the 3363 // first thread to mark it, then we push it on our 3364 // marking stack 3365 if (_bit_map->par_mark(addr)) { // ... now grey 3366 // push on work queue (grey set) 3367 bool simulate_overflow = false; 3368 NOT_PRODUCT( 3369 if (CMSMarkStackOverflowALot && 3370 _collector->simulate_overflow()) { 3371 // simulate a stack overflow 3372 simulate_overflow = true; 3373 } 3374 ) 3375 if (simulate_overflow || 3376 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3377 // stack overflow 3378 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3379 // We cannot assert that the overflow stack is full because 3380 // it may have been emptied since. 3381 assert(simulate_overflow || 3382 _work_queue->size() == _work_queue->max_elems(), 3383 "Else push should have succeeded"); 3384 handle_stack_overflow(addr); 3385 } 3386 } // Else, some other thread got there first 3387 do_yield_check(); 3388 } 3389 } 3390 3391 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3392 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3393 3394 void ParConcMarkingClosure::trim_queue(size_t max) { 3395 while (_work_queue->size() > max) { 3396 oop new_oop; 3397 if (_work_queue->pop_local(new_oop)) { 3398 assert(new_oop->is_oop(), "Should be an oop"); 3399 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3400 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3401 new_oop->oop_iterate(this); // do_oop() above 3402 do_yield_check(); 3403 } 3404 } 3405 } 3406 3407 // Upon stack overflow, we discard (part of) the stack, 3408 // remembering the least address amongst those discarded 3409 // in CMSCollector's _restart_address. 3410 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3411 // We need to do this under a mutex to prevent other 3412 // workers from interfering with the work done below. 3413 MutexLockerEx ml(_overflow_stack->par_lock(), 3414 Mutex::_no_safepoint_check_flag); 3415 // Remember the least grey address discarded 3416 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3417 _collector->lower_restart_addr(ra); 3418 _overflow_stack->reset(); // discard stack contents 3419 _overflow_stack->expand(); // expand the stack if possible 3420 } 3421 3422 3423 void CMSConcMarkingTask::do_work_steal(int i) { 3424 OopTaskQueue* work_q = work_queue(i); 3425 oop obj_to_scan; 3426 CMSBitMap* bm = &(_collector->_markBitMap); 3427 CMSMarkStack* ovflw = &(_collector->_markStack); 3428 int* seed = _collector->hash_seed(i); 3429 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3430 while (true) { 3431 cl.trim_queue(0); 3432 assert(work_q->size() == 0, "Should have been emptied above"); 3433 if (get_work_from_overflow_stack(ovflw, work_q)) { 3434 // Can't assert below because the work obtained from the 3435 // overflow stack may already have been stolen from us. 3436 // assert(work_q->size() > 0, "Work from overflow stack"); 3437 continue; 3438 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3439 assert(obj_to_scan->is_oop(), "Should be an oop"); 3440 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3441 obj_to_scan->oop_iterate(&cl); 3442 } else if (terminator()->offer_termination(&_term_term)) { 3443 assert(work_q->size() == 0, "Impossible!"); 3444 break; 3445 } else if (yielding() || should_yield()) { 3446 yield(); 3447 } 3448 } 3449 } 3450 3451 // This is run by the CMS (coordinator) thread. 3452 void CMSConcMarkingTask::coordinator_yield() { 3453 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3454 "CMS thread should hold CMS token"); 3455 // First give up the locks, then yield, then re-lock 3456 // We should probably use a constructor/destructor idiom to 3457 // do this unlock/lock or modify the MutexUnlocker class to 3458 // serve our purpose. XXX 3459 assert_lock_strong(_bit_map_lock); 3460 _bit_map_lock->unlock(); 3461 ConcurrentMarkSweepThread::desynchronize(true); 3462 _collector->stopTimer(); 3463 _collector->incrementYields(); 3464 3465 // It is possible for whichever thread initiated the yield request 3466 // not to get a chance to wake up and take the bitmap lock between 3467 // this thread releasing it and reacquiring it. So, while the 3468 // should_yield() flag is on, let's sleep for a bit to give the 3469 // other thread a chance to wake up. The limit imposed on the number 3470 // of iterations is defensive, to avoid any unforseen circumstances 3471 // putting us into an infinite loop. Since it's always been this 3472 // (coordinator_yield()) method that was observed to cause the 3473 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3474 // which is by default non-zero. For the other seven methods that 3475 // also perform the yield operation, as are using a different 3476 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3477 // can enable the sleeping for those methods too, if necessary. 3478 // See 6442774. 3479 // 3480 // We really need to reconsider the synchronization between the GC 3481 // thread and the yield-requesting threads in the future and we 3482 // should really use wait/notify, which is the recommended 3483 // way of doing this type of interaction. Additionally, we should 3484 // consolidate the eight methods that do the yield operation and they 3485 // are almost identical into one for better maintainability and 3486 // readability. See 6445193. 3487 // 3488 // Tony 2006.06.29 3489 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3490 ConcurrentMarkSweepThread::should_yield() && 3491 !CMSCollector::foregroundGCIsActive(); ++i) { 3492 os::sleep(Thread::current(), 1, false); 3493 } 3494 3495 ConcurrentMarkSweepThread::synchronize(true); 3496 _bit_map_lock->lock_without_safepoint_check(); 3497 _collector->startTimer(); 3498 } 3499 3500 bool CMSCollector::do_marking_mt() { 3501 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3502 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3503 conc_workers()->active_workers(), 3504 Threads::number_of_non_daemon_threads()); 3505 conc_workers()->set_active_workers(num_workers); 3506 3507 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3508 3509 CMSConcMarkingTask tsk(this, 3510 cms_space, 3511 conc_workers(), 3512 task_queues()); 3513 3514 // Since the actual number of workers we get may be different 3515 // from the number we requested above, do we need to do anything different 3516 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3517 // class?? XXX 3518 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3519 3520 // Refs discovery is already non-atomic. 3521 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3522 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3523 conc_workers()->start_task(&tsk); 3524 while (tsk.yielded()) { 3525 tsk.coordinator_yield(); 3526 conc_workers()->continue_task(&tsk); 3527 } 3528 // If the task was aborted, _restart_addr will be non-NULL 3529 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3530 while (_restart_addr != NULL) { 3531 // XXX For now we do not make use of ABORTED state and have not 3532 // yet implemented the right abort semantics (even in the original 3533 // single-threaded CMS case). That needs some more investigation 3534 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3535 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3536 // If _restart_addr is non-NULL, a marking stack overflow 3537 // occurred; we need to do a fresh marking iteration from the 3538 // indicated restart address. 3539 if (_foregroundGCIsActive) { 3540 // We may be running into repeated stack overflows, having 3541 // reached the limit of the stack size, while making very 3542 // slow forward progress. It may be best to bail out and 3543 // let the foreground collector do its job. 3544 // Clear _restart_addr, so that foreground GC 3545 // works from scratch. This avoids the headache of 3546 // a "rescan" which would otherwise be needed because 3547 // of the dirty mod union table & card table. 3548 _restart_addr = NULL; 3549 return false; 3550 } 3551 // Adjust the task to restart from _restart_addr 3552 tsk.reset(_restart_addr); 3553 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3554 _restart_addr); 3555 _restart_addr = NULL; 3556 // Get the workers going again 3557 conc_workers()->start_task(&tsk); 3558 while (tsk.yielded()) { 3559 tsk.coordinator_yield(); 3560 conc_workers()->continue_task(&tsk); 3561 } 3562 } 3563 assert(tsk.completed(), "Inconsistency"); 3564 assert(tsk.result() == true, "Inconsistency"); 3565 return true; 3566 } 3567 3568 bool CMSCollector::do_marking_st() { 3569 ResourceMark rm; 3570 HandleMark hm; 3571 3572 // Temporarily make refs discovery single threaded (non-MT) 3573 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3574 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3575 &_markStack, CMSYield); 3576 // the last argument to iterate indicates whether the iteration 3577 // should be incremental with periodic yields. 3578 _markBitMap.iterate(&markFromRootsClosure); 3579 // If _restart_addr is non-NULL, a marking stack overflow 3580 // occurred; we need to do a fresh iteration from the 3581 // indicated restart address. 3582 while (_restart_addr != NULL) { 3583 if (_foregroundGCIsActive) { 3584 // We may be running into repeated stack overflows, having 3585 // reached the limit of the stack size, while making very 3586 // slow forward progress. It may be best to bail out and 3587 // let the foreground collector do its job. 3588 // Clear _restart_addr, so that foreground GC 3589 // works from scratch. This avoids the headache of 3590 // a "rescan" which would otherwise be needed because 3591 // of the dirty mod union table & card table. 3592 _restart_addr = NULL; 3593 return false; // indicating failure to complete marking 3594 } 3595 // Deal with stack overflow: 3596 // we restart marking from _restart_addr 3597 HeapWord* ra = _restart_addr; 3598 markFromRootsClosure.reset(ra); 3599 _restart_addr = NULL; 3600 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3601 } 3602 return true; 3603 } 3604 3605 void CMSCollector::preclean() { 3606 check_correct_thread_executing(); 3607 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3608 verify_work_stacks_empty(); 3609 verify_overflow_empty(); 3610 _abort_preclean = false; 3611 if (CMSPrecleaningEnabled) { 3612 if (!CMSEdenChunksRecordAlways) { 3613 _eden_chunk_index = 0; 3614 } 3615 size_t used = get_eden_used(); 3616 size_t capacity = get_eden_capacity(); 3617 // Don't start sampling unless we will get sufficiently 3618 // many samples. 3619 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3620 * CMSScheduleRemarkEdenPenetration)) { 3621 _start_sampling = true; 3622 } else { 3623 _start_sampling = false; 3624 } 3625 GCTraceCPUTime tcpu; 3626 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3627 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3628 } 3629 CMSTokenSync x(true); // is cms thread 3630 if (CMSPrecleaningEnabled) { 3631 sample_eden(); 3632 _collectorState = AbortablePreclean; 3633 } else { 3634 _collectorState = FinalMarking; 3635 } 3636 verify_work_stacks_empty(); 3637 verify_overflow_empty(); 3638 } 3639 3640 // Try and schedule the remark such that young gen 3641 // occupancy is CMSScheduleRemarkEdenPenetration %. 3642 void CMSCollector::abortable_preclean() { 3643 check_correct_thread_executing(); 3644 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3645 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3646 3647 // If Eden's current occupancy is below this threshold, 3648 // immediately schedule the remark; else preclean 3649 // past the next scavenge in an effort to 3650 // schedule the pause as described above. By choosing 3651 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3652 // we will never do an actual abortable preclean cycle. 3653 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3654 GCTraceCPUTime tcpu; 3655 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3656 // We need more smarts in the abortable preclean 3657 // loop below to deal with cases where allocation 3658 // in young gen is very very slow, and our precleaning 3659 // is running a losing race against a horde of 3660 // mutators intent on flooding us with CMS updates 3661 // (dirty cards). 3662 // One, admittedly dumb, strategy is to give up 3663 // after a certain number of abortable precleaning loops 3664 // or after a certain maximum time. We want to make 3665 // this smarter in the next iteration. 3666 // XXX FIX ME!!! YSR 3667 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3668 while (!(should_abort_preclean() || 3669 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3670 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3671 cumworkdone += workdone; 3672 loops++; 3673 // Voluntarily terminate abortable preclean phase if we have 3674 // been at it for too long. 3675 if ((CMSMaxAbortablePrecleanLoops != 0) && 3676 loops >= CMSMaxAbortablePrecleanLoops) { 3677 log_debug(gc)(" CMS: abort preclean due to loops "); 3678 break; 3679 } 3680 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3681 log_debug(gc)(" CMS: abort preclean due to time "); 3682 break; 3683 } 3684 // If we are doing little work each iteration, we should 3685 // take a short break. 3686 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3687 // Sleep for some time, waiting for work to accumulate 3688 stopTimer(); 3689 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3690 startTimer(); 3691 waited++; 3692 } 3693 } 3694 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3695 loops, waited, cumworkdone); 3696 } 3697 CMSTokenSync x(true); // is cms thread 3698 if (_collectorState != Idling) { 3699 assert(_collectorState == AbortablePreclean, 3700 "Spontaneous state transition?"); 3701 _collectorState = FinalMarking; 3702 } // Else, a foreground collection completed this CMS cycle. 3703 return; 3704 } 3705 3706 // Respond to an Eden sampling opportunity 3707 void CMSCollector::sample_eden() { 3708 // Make sure a young gc cannot sneak in between our 3709 // reading and recording of a sample. 3710 assert(Thread::current()->is_ConcurrentGC_thread(), 3711 "Only the cms thread may collect Eden samples"); 3712 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3713 "Should collect samples while holding CMS token"); 3714 if (!_start_sampling) { 3715 return; 3716 } 3717 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3718 // is populated by the young generation. 3719 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3720 if (_eden_chunk_index < _eden_chunk_capacity) { 3721 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3722 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3723 "Unexpected state of Eden"); 3724 // We'd like to check that what we just sampled is an oop-start address; 3725 // however, we cannot do that here since the object may not yet have been 3726 // initialized. So we'll instead do the check when we _use_ this sample 3727 // later. 3728 if (_eden_chunk_index == 0 || 3729 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3730 _eden_chunk_array[_eden_chunk_index-1]) 3731 >= CMSSamplingGrain)) { 3732 _eden_chunk_index++; // commit sample 3733 } 3734 } 3735 } 3736 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3737 size_t used = get_eden_used(); 3738 size_t capacity = get_eden_capacity(); 3739 assert(used <= capacity, "Unexpected state of Eden"); 3740 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3741 _abort_preclean = true; 3742 } 3743 } 3744 } 3745 3746 3747 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3748 assert(_collectorState == Precleaning || 3749 _collectorState == AbortablePreclean, "incorrect state"); 3750 ResourceMark rm; 3751 HandleMark hm; 3752 3753 // Precleaning is currently not MT but the reference processor 3754 // may be set for MT. Disable it temporarily here. 3755 ReferenceProcessor* rp = ref_processor(); 3756 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3757 3758 // Do one pass of scrubbing the discovered reference lists 3759 // to remove any reference objects with strongly-reachable 3760 // referents. 3761 if (clean_refs) { 3762 CMSPrecleanRefsYieldClosure yield_cl(this); 3763 assert(rp->span().equals(_span), "Spans should be equal"); 3764 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3765 &_markStack, true /* preclean */); 3766 CMSDrainMarkingStackClosure complete_trace(this, 3767 _span, &_markBitMap, &_markStack, 3768 &keep_alive, true /* preclean */); 3769 3770 // We don't want this step to interfere with a young 3771 // collection because we don't want to take CPU 3772 // or memory bandwidth away from the young GC threads 3773 // (which may be as many as there are CPUs). 3774 // Note that we don't need to protect ourselves from 3775 // interference with mutators because they can't 3776 // manipulate the discovered reference lists nor affect 3777 // the computed reachability of the referents, the 3778 // only properties manipulated by the precleaning 3779 // of these reference lists. 3780 stopTimer(); 3781 CMSTokenSyncWithLocks x(true /* is cms thread */, 3782 bitMapLock()); 3783 startTimer(); 3784 sample_eden(); 3785 3786 // The following will yield to allow foreground 3787 // collection to proceed promptly. XXX YSR: 3788 // The code in this method may need further 3789 // tweaking for better performance and some restructuring 3790 // for cleaner interfaces. 3791 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3792 rp->preclean_discovered_references( 3793 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3794 gc_timer); 3795 } 3796 3797 if (clean_survivor) { // preclean the active survivor space(s) 3798 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3799 &_markBitMap, &_modUnionTable, 3800 &_markStack, true /* precleaning phase */); 3801 stopTimer(); 3802 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3803 bitMapLock()); 3804 startTimer(); 3805 unsigned int before_count = 3806 GenCollectedHeap::heap()->total_collections(); 3807 SurvivorSpacePrecleanClosure 3808 sss_cl(this, _span, &_markBitMap, &_markStack, 3809 &pam_cl, before_count, CMSYield); 3810 _young_gen->from()->object_iterate_careful(&sss_cl); 3811 _young_gen->to()->object_iterate_careful(&sss_cl); 3812 } 3813 MarkRefsIntoAndScanClosure 3814 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3815 &_markStack, this, CMSYield, 3816 true /* precleaning phase */); 3817 // CAUTION: The following closure has persistent state that may need to 3818 // be reset upon a decrease in the sequence of addresses it 3819 // processes. 3820 ScanMarkedObjectsAgainCarefullyClosure 3821 smoac_cl(this, _span, 3822 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3823 3824 // Preclean dirty cards in ModUnionTable and CardTable using 3825 // appropriate convergence criterion; 3826 // repeat CMSPrecleanIter times unless we find that 3827 // we are losing. 3828 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3829 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3830 "Bad convergence multiplier"); 3831 assert(CMSPrecleanThreshold >= 100, 3832 "Unreasonably low CMSPrecleanThreshold"); 3833 3834 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3835 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3836 numIter < CMSPrecleanIter; 3837 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3838 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3839 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3840 // Either there are very few dirty cards, so re-mark 3841 // pause will be small anyway, or our pre-cleaning isn't 3842 // that much faster than the rate at which cards are being 3843 // dirtied, so we might as well stop and re-mark since 3844 // precleaning won't improve our re-mark time by much. 3845 if (curNumCards <= CMSPrecleanThreshold || 3846 (numIter > 0 && 3847 (curNumCards * CMSPrecleanDenominator > 3848 lastNumCards * CMSPrecleanNumerator))) { 3849 numIter++; 3850 cumNumCards += curNumCards; 3851 break; 3852 } 3853 } 3854 3855 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3856 3857 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3858 cumNumCards += curNumCards; 3859 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3860 curNumCards, cumNumCards, numIter); 3861 return cumNumCards; // as a measure of useful work done 3862 } 3863 3864 // PRECLEANING NOTES: 3865 // Precleaning involves: 3866 // . reading the bits of the modUnionTable and clearing the set bits. 3867 // . For the cards corresponding to the set bits, we scan the 3868 // objects on those cards. This means we need the free_list_lock 3869 // so that we can safely iterate over the CMS space when scanning 3870 // for oops. 3871 // . When we scan the objects, we'll be both reading and setting 3872 // marks in the marking bit map, so we'll need the marking bit map. 3873 // . For protecting _collector_state transitions, we take the CGC_lock. 3874 // Note that any races in the reading of of card table entries by the 3875 // CMS thread on the one hand and the clearing of those entries by the 3876 // VM thread or the setting of those entries by the mutator threads on the 3877 // other are quite benign. However, for efficiency it makes sense to keep 3878 // the VM thread from racing with the CMS thread while the latter is 3879 // dirty card info to the modUnionTable. We therefore also use the 3880 // CGC_lock to protect the reading of the card table and the mod union 3881 // table by the CM thread. 3882 // . We run concurrently with mutator updates, so scanning 3883 // needs to be done carefully -- we should not try to scan 3884 // potentially uninitialized objects. 3885 // 3886 // Locking strategy: While holding the CGC_lock, we scan over and 3887 // reset a maximal dirty range of the mod union / card tables, then lock 3888 // the free_list_lock and bitmap lock to do a full marking, then 3889 // release these locks; and repeat the cycle. This allows for a 3890 // certain amount of fairness in the sharing of these locks between 3891 // the CMS collector on the one hand, and the VM thread and the 3892 // mutators on the other. 3893 3894 // NOTE: preclean_mod_union_table() and preclean_card_table() 3895 // further below are largely identical; if you need to modify 3896 // one of these methods, please check the other method too. 3897 3898 size_t CMSCollector::preclean_mod_union_table( 3899 ConcurrentMarkSweepGeneration* old_gen, 3900 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3901 verify_work_stacks_empty(); 3902 verify_overflow_empty(); 3903 3904 // strategy: starting with the first card, accumulate contiguous 3905 // ranges of dirty cards; clear these cards, then scan the region 3906 // covered by these cards. 3907 3908 // Since all of the MUT is committed ahead, we can just use 3909 // that, in case the generations expand while we are precleaning. 3910 // It might also be fine to just use the committed part of the 3911 // generation, but we might potentially miss cards when the 3912 // generation is rapidly expanding while we are in the midst 3913 // of precleaning. 3914 HeapWord* startAddr = old_gen->reserved().start(); 3915 HeapWord* endAddr = old_gen->reserved().end(); 3916 3917 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3918 3919 size_t numDirtyCards, cumNumDirtyCards; 3920 HeapWord *nextAddr, *lastAddr; 3921 for (cumNumDirtyCards = numDirtyCards = 0, 3922 nextAddr = lastAddr = startAddr; 3923 nextAddr < endAddr; 3924 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3925 3926 ResourceMark rm; 3927 HandleMark hm; 3928 3929 MemRegion dirtyRegion; 3930 { 3931 stopTimer(); 3932 // Potential yield point 3933 CMSTokenSync ts(true); 3934 startTimer(); 3935 sample_eden(); 3936 // Get dirty region starting at nextOffset (inclusive), 3937 // simultaneously clearing it. 3938 dirtyRegion = 3939 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3940 assert(dirtyRegion.start() >= nextAddr, 3941 "returned region inconsistent?"); 3942 } 3943 // Remember where the next search should begin. 3944 // The returned region (if non-empty) is a right open interval, 3945 // so lastOffset is obtained from the right end of that 3946 // interval. 3947 lastAddr = dirtyRegion.end(); 3948 // Should do something more transparent and less hacky XXX 3949 numDirtyCards = 3950 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3951 3952 // We'll scan the cards in the dirty region (with periodic 3953 // yields for foreground GC as needed). 3954 if (!dirtyRegion.is_empty()) { 3955 assert(numDirtyCards > 0, "consistency check"); 3956 HeapWord* stop_point = NULL; 3957 stopTimer(); 3958 // Potential yield point 3959 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3960 bitMapLock()); 3961 startTimer(); 3962 { 3963 verify_work_stacks_empty(); 3964 verify_overflow_empty(); 3965 sample_eden(); 3966 stop_point = 3967 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3968 } 3969 if (stop_point != NULL) { 3970 // The careful iteration stopped early either because it found an 3971 // uninitialized object, or because we were in the midst of an 3972 // "abortable preclean", which should now be aborted. Redirty 3973 // the bits corresponding to the partially-scanned or unscanned 3974 // cards. We'll either restart at the next block boundary or 3975 // abort the preclean. 3976 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3977 "Should only be AbortablePreclean."); 3978 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3979 if (should_abort_preclean()) { 3980 break; // out of preclean loop 3981 } else { 3982 // Compute the next address at which preclean should pick up; 3983 // might need bitMapLock in order to read P-bits. 3984 lastAddr = next_card_start_after_block(stop_point); 3985 } 3986 } 3987 } else { 3988 assert(lastAddr == endAddr, "consistency check"); 3989 assert(numDirtyCards == 0, "consistency check"); 3990 break; 3991 } 3992 } 3993 verify_work_stacks_empty(); 3994 verify_overflow_empty(); 3995 return cumNumDirtyCards; 3996 } 3997 3998 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3999 // below are largely identical; if you need to modify 4000 // one of these methods, please check the other method too. 4001 4002 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4003 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4004 // strategy: it's similar to precleamModUnionTable above, in that 4005 // we accumulate contiguous ranges of dirty cards, mark these cards 4006 // precleaned, then scan the region covered by these cards. 4007 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4008 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4009 4010 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4011 4012 size_t numDirtyCards, cumNumDirtyCards; 4013 HeapWord *lastAddr, *nextAddr; 4014 4015 for (cumNumDirtyCards = numDirtyCards = 0, 4016 nextAddr = lastAddr = startAddr; 4017 nextAddr < endAddr; 4018 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4019 4020 ResourceMark rm; 4021 HandleMark hm; 4022 4023 MemRegion dirtyRegion; 4024 { 4025 // See comments in "Precleaning notes" above on why we 4026 // do this locking. XXX Could the locking overheads be 4027 // too high when dirty cards are sparse? [I don't think so.] 4028 stopTimer(); 4029 CMSTokenSync x(true); // is cms thread 4030 startTimer(); 4031 sample_eden(); 4032 // Get and clear dirty region from card table 4033 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4034 MemRegion(nextAddr, endAddr), 4035 true, 4036 CardTableModRefBS::precleaned_card_val()); 4037 4038 assert(dirtyRegion.start() >= nextAddr, 4039 "returned region inconsistent?"); 4040 } 4041 lastAddr = dirtyRegion.end(); 4042 numDirtyCards = 4043 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4044 4045 if (!dirtyRegion.is_empty()) { 4046 stopTimer(); 4047 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4048 startTimer(); 4049 sample_eden(); 4050 verify_work_stacks_empty(); 4051 verify_overflow_empty(); 4052 HeapWord* stop_point = 4053 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4054 if (stop_point != NULL) { 4055 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4056 "Should only be AbortablePreclean."); 4057 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4058 if (should_abort_preclean()) { 4059 break; // out of preclean loop 4060 } else { 4061 // Compute the next address at which preclean should pick up. 4062 lastAddr = next_card_start_after_block(stop_point); 4063 } 4064 } 4065 } else { 4066 break; 4067 } 4068 } 4069 verify_work_stacks_empty(); 4070 verify_overflow_empty(); 4071 return cumNumDirtyCards; 4072 } 4073 4074 class PrecleanKlassClosure : public KlassClosure { 4075 KlassToOopClosure _cm_klass_closure; 4076 public: 4077 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4078 void do_klass(Klass* k) { 4079 if (k->has_accumulated_modified_oops()) { 4080 k->clear_accumulated_modified_oops(); 4081 4082 _cm_klass_closure.do_klass(k); 4083 } 4084 } 4085 }; 4086 4087 // The freelist lock is needed to prevent asserts, is it really needed? 4088 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4089 4090 cl->set_freelistLock(freelistLock); 4091 4092 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4093 4094 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4095 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4096 PrecleanKlassClosure preclean_klass_closure(cl); 4097 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4098 4099 verify_work_stacks_empty(); 4100 verify_overflow_empty(); 4101 } 4102 4103 void CMSCollector::checkpointRootsFinal() { 4104 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4105 check_correct_thread_executing(); 4106 // world is stopped at this checkpoint 4107 assert(SafepointSynchronize::is_at_safepoint(), 4108 "world should be stopped"); 4109 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4110 4111 verify_work_stacks_empty(); 4112 verify_overflow_empty(); 4113 4114 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4115 _young_gen->used() / K, _young_gen->capacity() / K); 4116 { 4117 if (CMSScavengeBeforeRemark) { 4118 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4119 // Temporarily set flag to false, GCH->do_collection will 4120 // expect it to be false and set to true 4121 FlagSetting fl(gch->_is_gc_active, false); 4122 4123 gch->do_collection(true, // full (i.e. force, see below) 4124 false, // !clear_all_soft_refs 4125 0, // size 4126 false, // is_tlab 4127 GenCollectedHeap::YoungGen // type 4128 ); 4129 } 4130 FreelistLocker x(this); 4131 MutexLockerEx y(bitMapLock(), 4132 Mutex::_no_safepoint_check_flag); 4133 checkpointRootsFinalWork(); 4134 } 4135 verify_work_stacks_empty(); 4136 verify_overflow_empty(); 4137 } 4138 4139 void CMSCollector::checkpointRootsFinalWork() { 4140 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4141 4142 assert(haveFreelistLocks(), "must have free list locks"); 4143 assert_lock_strong(bitMapLock()); 4144 4145 ResourceMark rm; 4146 HandleMark hm; 4147 4148 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4149 4150 if (should_unload_classes()) { 4151 CodeCache::gc_prologue(); 4152 } 4153 assert(haveFreelistLocks(), "must have free list locks"); 4154 assert_lock_strong(bitMapLock()); 4155 4156 // We might assume that we need not fill TLAB's when 4157 // CMSScavengeBeforeRemark is set, because we may have just done 4158 // a scavenge which would have filled all TLAB's -- and besides 4159 // Eden would be empty. This however may not always be the case -- 4160 // for instance although we asked for a scavenge, it may not have 4161 // happened because of a JNI critical section. We probably need 4162 // a policy for deciding whether we can in that case wait until 4163 // the critical section releases and then do the remark following 4164 // the scavenge, and skip it here. In the absence of that policy, 4165 // or of an indication of whether the scavenge did indeed occur, 4166 // we cannot rely on TLAB's having been filled and must do 4167 // so here just in case a scavenge did not happen. 4168 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4169 // Update the saved marks which may affect the root scans. 4170 gch->save_marks(); 4171 4172 print_eden_and_survivor_chunk_arrays(); 4173 4174 { 4175 #if defined(COMPILER2) || INCLUDE_JVMCI 4176 DerivedPointerTableDeactivate dpt_deact; 4177 #endif 4178 4179 // Note on the role of the mod union table: 4180 // Since the marker in "markFromRoots" marks concurrently with 4181 // mutators, it is possible for some reachable objects not to have been 4182 // scanned. For instance, an only reference to an object A was 4183 // placed in object B after the marker scanned B. Unless B is rescanned, 4184 // A would be collected. Such updates to references in marked objects 4185 // are detected via the mod union table which is the set of all cards 4186 // dirtied since the first checkpoint in this GC cycle and prior to 4187 // the most recent young generation GC, minus those cleaned up by the 4188 // concurrent precleaning. 4189 if (CMSParallelRemarkEnabled) { 4190 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4191 do_remark_parallel(); 4192 } else { 4193 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4194 do_remark_non_parallel(); 4195 } 4196 } 4197 verify_work_stacks_empty(); 4198 verify_overflow_empty(); 4199 4200 { 4201 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4202 refProcessingWork(); 4203 } 4204 verify_work_stacks_empty(); 4205 verify_overflow_empty(); 4206 4207 if (should_unload_classes()) { 4208 CodeCache::gc_epilogue(); 4209 } 4210 JvmtiExport::gc_epilogue(); 4211 4212 // If we encountered any (marking stack / work queue) overflow 4213 // events during the current CMS cycle, take appropriate 4214 // remedial measures, where possible, so as to try and avoid 4215 // recurrence of that condition. 4216 assert(_markStack.isEmpty(), "No grey objects"); 4217 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4218 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4219 if (ser_ovflw > 0) { 4220 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4221 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4222 _markStack.expand(); 4223 _ser_pmc_remark_ovflw = 0; 4224 _ser_pmc_preclean_ovflw = 0; 4225 _ser_kac_preclean_ovflw = 0; 4226 _ser_kac_ovflw = 0; 4227 } 4228 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4229 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4230 _par_pmc_remark_ovflw, _par_kac_ovflw); 4231 _par_pmc_remark_ovflw = 0; 4232 _par_kac_ovflw = 0; 4233 } 4234 if (_markStack._hit_limit > 0) { 4235 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4236 _markStack._hit_limit); 4237 } 4238 if (_markStack._failed_double > 0) { 4239 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4240 _markStack._failed_double, _markStack.capacity()); 4241 } 4242 _markStack._hit_limit = 0; 4243 _markStack._failed_double = 0; 4244 4245 if ((VerifyAfterGC || VerifyDuringGC) && 4246 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4247 verify_after_remark(); 4248 } 4249 4250 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4251 4252 // Change under the freelistLocks. 4253 _collectorState = Sweeping; 4254 // Call isAllClear() under bitMapLock 4255 assert(_modUnionTable.isAllClear(), 4256 "Should be clear by end of the final marking"); 4257 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4258 "Should be clear by end of the final marking"); 4259 } 4260 4261 void CMSParInitialMarkTask::work(uint worker_id) { 4262 elapsedTimer _timer; 4263 ResourceMark rm; 4264 HandleMark hm; 4265 4266 // ---------- scan from roots -------------- 4267 _timer.start(); 4268 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4269 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4270 4271 // ---------- young gen roots -------------- 4272 { 4273 work_on_young_gen_roots(&par_mri_cl); 4274 _timer.stop(); 4275 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4276 } 4277 4278 // ---------- remaining roots -------------- 4279 _timer.reset(); 4280 _timer.start(); 4281 4282 CLDToOopClosure cld_closure(&par_mri_cl, true); 4283 4284 gch->gen_process_roots(_strong_roots_scope, 4285 GenCollectedHeap::OldGen, 4286 false, // yg was scanned above 4287 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4288 _collector->should_unload_classes(), 4289 &par_mri_cl, 4290 NULL, 4291 &cld_closure); 4292 assert(_collector->should_unload_classes() 4293 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4294 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4295 _timer.stop(); 4296 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4297 } 4298 4299 // Parallel remark task 4300 class CMSParRemarkTask: public CMSParMarkTask { 4301 CompactibleFreeListSpace* _cms_space; 4302 4303 // The per-thread work queues, available here for stealing. 4304 OopTaskQueueSet* _task_queues; 4305 ParallelTaskTerminator _term; 4306 StrongRootsScope* _strong_roots_scope; 4307 4308 public: 4309 // A value of 0 passed to n_workers will cause the number of 4310 // workers to be taken from the active workers in the work gang. 4311 CMSParRemarkTask(CMSCollector* collector, 4312 CompactibleFreeListSpace* cms_space, 4313 uint n_workers, WorkGang* workers, 4314 OopTaskQueueSet* task_queues, 4315 StrongRootsScope* strong_roots_scope): 4316 CMSParMarkTask("Rescan roots and grey objects in parallel", 4317 collector, n_workers), 4318 _cms_space(cms_space), 4319 _task_queues(task_queues), 4320 _term(n_workers, task_queues), 4321 _strong_roots_scope(strong_roots_scope) { } 4322 4323 OopTaskQueueSet* task_queues() { return _task_queues; } 4324 4325 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4326 4327 ParallelTaskTerminator* terminator() { return &_term; } 4328 uint n_workers() { return _n_workers; } 4329 4330 void work(uint worker_id); 4331 4332 private: 4333 // ... of dirty cards in old space 4334 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4335 ParMarkRefsIntoAndScanClosure* cl); 4336 4337 // ... work stealing for the above 4338 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4339 }; 4340 4341 class RemarkKlassClosure : public KlassClosure { 4342 KlassToOopClosure _cm_klass_closure; 4343 public: 4344 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4345 void do_klass(Klass* k) { 4346 // Check if we have modified any oops in the Klass during the concurrent marking. 4347 if (k->has_accumulated_modified_oops()) { 4348 k->clear_accumulated_modified_oops(); 4349 4350 // We could have transfered the current modified marks to the accumulated marks, 4351 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4352 } else if (k->has_modified_oops()) { 4353 // Don't clear anything, this info is needed by the next young collection. 4354 } else { 4355 // No modified oops in the Klass. 4356 return; 4357 } 4358 4359 // The klass has modified fields, need to scan the klass. 4360 _cm_klass_closure.do_klass(k); 4361 } 4362 }; 4363 4364 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4365 ParNewGeneration* young_gen = _collector->_young_gen; 4366 ContiguousSpace* eden_space = young_gen->eden(); 4367 ContiguousSpace* from_space = young_gen->from(); 4368 ContiguousSpace* to_space = young_gen->to(); 4369 4370 HeapWord** eca = _collector->_eden_chunk_array; 4371 size_t ect = _collector->_eden_chunk_index; 4372 HeapWord** sca = _collector->_survivor_chunk_array; 4373 size_t sct = _collector->_survivor_chunk_index; 4374 4375 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4376 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4377 4378 do_young_space_rescan(cl, to_space, NULL, 0); 4379 do_young_space_rescan(cl, from_space, sca, sct); 4380 do_young_space_rescan(cl, eden_space, eca, ect); 4381 } 4382 4383 // work_queue(i) is passed to the closure 4384 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4385 // also is passed to do_dirty_card_rescan_tasks() and to 4386 // do_work_steal() to select the i-th task_queue. 4387 4388 void CMSParRemarkTask::work(uint worker_id) { 4389 elapsedTimer _timer; 4390 ResourceMark rm; 4391 HandleMark hm; 4392 4393 // ---------- rescan from roots -------------- 4394 _timer.start(); 4395 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4396 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4397 _collector->_span, _collector->ref_processor(), 4398 &(_collector->_markBitMap), 4399 work_queue(worker_id)); 4400 4401 // Rescan young gen roots first since these are likely 4402 // coarsely partitioned and may, on that account, constitute 4403 // the critical path; thus, it's best to start off that 4404 // work first. 4405 // ---------- young gen roots -------------- 4406 { 4407 work_on_young_gen_roots(&par_mrias_cl); 4408 _timer.stop(); 4409 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4410 } 4411 4412 // ---------- remaining roots -------------- 4413 _timer.reset(); 4414 _timer.start(); 4415 gch->gen_process_roots(_strong_roots_scope, 4416 GenCollectedHeap::OldGen, 4417 false, // yg was scanned above 4418 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4419 _collector->should_unload_classes(), 4420 &par_mrias_cl, 4421 NULL, 4422 NULL); // The dirty klasses will be handled below 4423 4424 assert(_collector->should_unload_classes() 4425 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4426 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4427 _timer.stop(); 4428 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4429 4430 // ---------- unhandled CLD scanning ---------- 4431 if (worker_id == 0) { // Single threaded at the moment. 4432 _timer.reset(); 4433 _timer.start(); 4434 4435 // Scan all new class loader data objects and new dependencies that were 4436 // introduced during concurrent marking. 4437 ResourceMark rm; 4438 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4439 for (int i = 0; i < array->length(); i++) { 4440 par_mrias_cl.do_cld_nv(array->at(i)); 4441 } 4442 4443 // We don't need to keep track of new CLDs anymore. 4444 ClassLoaderDataGraph::remember_new_clds(false); 4445 4446 _timer.stop(); 4447 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4448 } 4449 4450 // ---------- dirty klass scanning ---------- 4451 if (worker_id == 0) { // Single threaded at the moment. 4452 _timer.reset(); 4453 _timer.start(); 4454 4455 // Scan all classes that was dirtied during the concurrent marking phase. 4456 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4457 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4458 4459 _timer.stop(); 4460 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4461 } 4462 4463 // We might have added oops to ClassLoaderData::_handles during the 4464 // concurrent marking phase. These oops point to newly allocated objects 4465 // that are guaranteed to be kept alive. Either by the direct allocation 4466 // code, or when the young collector processes the roots. Hence, 4467 // we don't have to revisit the _handles block during the remark phase. 4468 4469 // ---------- rescan dirty cards ------------ 4470 _timer.reset(); 4471 _timer.start(); 4472 4473 // Do the rescan tasks for each of the two spaces 4474 // (cms_space) in turn. 4475 // "worker_id" is passed to select the task_queue for "worker_id" 4476 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4477 _timer.stop(); 4478 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4479 4480 // ---------- steal work from other threads ... 4481 // ---------- ... and drain overflow list. 4482 _timer.reset(); 4483 _timer.start(); 4484 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4485 _timer.stop(); 4486 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4487 } 4488 4489 void 4490 CMSParMarkTask::do_young_space_rescan( 4491 OopsInGenClosure* cl, ContiguousSpace* space, 4492 HeapWord** chunk_array, size_t chunk_top) { 4493 // Until all tasks completed: 4494 // . claim an unclaimed task 4495 // . compute region boundaries corresponding to task claimed 4496 // using chunk_array 4497 // . par_oop_iterate(cl) over that region 4498 4499 ResourceMark rm; 4500 HandleMark hm; 4501 4502 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4503 4504 uint nth_task = 0; 4505 uint n_tasks = pst->n_tasks(); 4506 4507 if (n_tasks > 0) { 4508 assert(pst->valid(), "Uninitialized use?"); 4509 HeapWord *start, *end; 4510 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4511 // We claimed task # nth_task; compute its boundaries. 4512 if (chunk_top == 0) { // no samples were taken 4513 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4514 start = space->bottom(); 4515 end = space->top(); 4516 } else if (nth_task == 0) { 4517 start = space->bottom(); 4518 end = chunk_array[nth_task]; 4519 } else if (nth_task < (uint)chunk_top) { 4520 assert(nth_task >= 1, "Control point invariant"); 4521 start = chunk_array[nth_task - 1]; 4522 end = chunk_array[nth_task]; 4523 } else { 4524 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4525 start = chunk_array[chunk_top - 1]; 4526 end = space->top(); 4527 } 4528 MemRegion mr(start, end); 4529 // Verify that mr is in space 4530 assert(mr.is_empty() || space->used_region().contains(mr), 4531 "Should be in space"); 4532 // Verify that "start" is an object boundary 4533 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4534 "Should be an oop"); 4535 space->par_oop_iterate(mr, cl); 4536 } 4537 pst->all_tasks_completed(); 4538 } 4539 } 4540 4541 void 4542 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4543 CompactibleFreeListSpace* sp, int i, 4544 ParMarkRefsIntoAndScanClosure* cl) { 4545 // Until all tasks completed: 4546 // . claim an unclaimed task 4547 // . compute region boundaries corresponding to task claimed 4548 // . transfer dirty bits ct->mut for that region 4549 // . apply rescanclosure to dirty mut bits for that region 4550 4551 ResourceMark rm; 4552 HandleMark hm; 4553 4554 OopTaskQueue* work_q = work_queue(i); 4555 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4556 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4557 // CAUTION: This closure has state that persists across calls to 4558 // the work method dirty_range_iterate_clear() in that it has 4559 // embedded in it a (subtype of) UpwardsObjectClosure. The 4560 // use of that state in the embedded UpwardsObjectClosure instance 4561 // assumes that the cards are always iterated (even if in parallel 4562 // by several threads) in monotonically increasing order per each 4563 // thread. This is true of the implementation below which picks 4564 // card ranges (chunks) in monotonically increasing order globally 4565 // and, a-fortiori, in monotonically increasing order per thread 4566 // (the latter order being a subsequence of the former). 4567 // If the work code below is ever reorganized into a more chaotic 4568 // work-partitioning form than the current "sequential tasks" 4569 // paradigm, the use of that persistent state will have to be 4570 // revisited and modified appropriately. See also related 4571 // bug 4756801 work on which should examine this code to make 4572 // sure that the changes there do not run counter to the 4573 // assumptions made here and necessary for correctness and 4574 // efficiency. Note also that this code might yield inefficient 4575 // behavior in the case of very large objects that span one or 4576 // more work chunks. Such objects would potentially be scanned 4577 // several times redundantly. Work on 4756801 should try and 4578 // address that performance anomaly if at all possible. XXX 4579 MemRegion full_span = _collector->_span; 4580 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4581 MarkFromDirtyCardsClosure 4582 greyRescanClosure(_collector, full_span, // entire span of interest 4583 sp, bm, work_q, cl); 4584 4585 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4586 assert(pst->valid(), "Uninitialized use?"); 4587 uint nth_task = 0; 4588 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4589 MemRegion span = sp->used_region(); 4590 HeapWord* start_addr = span.start(); 4591 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4592 alignment); 4593 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4594 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4595 start_addr, "Check alignment"); 4596 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4597 chunk_size, "Check alignment"); 4598 4599 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4600 // Having claimed the nth_task, compute corresponding mem-region, 4601 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4602 // The alignment restriction ensures that we do not need any 4603 // synchronization with other gang-workers while setting or 4604 // clearing bits in thus chunk of the MUT. 4605 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4606 start_addr + (nth_task+1)*chunk_size); 4607 // The last chunk's end might be way beyond end of the 4608 // used region. In that case pull back appropriately. 4609 if (this_span.end() > end_addr) { 4610 this_span.set_end(end_addr); 4611 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4612 } 4613 // Iterate over the dirty cards covering this chunk, marking them 4614 // precleaned, and setting the corresponding bits in the mod union 4615 // table. Since we have been careful to partition at Card and MUT-word 4616 // boundaries no synchronization is needed between parallel threads. 4617 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4618 &modUnionClosure); 4619 4620 // Having transferred these marks into the modUnionTable, 4621 // rescan the marked objects on the dirty cards in the modUnionTable. 4622 // Even if this is at a synchronous collection, the initial marking 4623 // may have been done during an asynchronous collection so there 4624 // may be dirty bits in the mod-union table. 4625 _collector->_modUnionTable.dirty_range_iterate_clear( 4626 this_span, &greyRescanClosure); 4627 _collector->_modUnionTable.verifyNoOneBitsInRange( 4628 this_span.start(), 4629 this_span.end()); 4630 } 4631 pst->all_tasks_completed(); // declare that i am done 4632 } 4633 4634 // . see if we can share work_queues with ParNew? XXX 4635 void 4636 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4637 int* seed) { 4638 OopTaskQueue* work_q = work_queue(i); 4639 NOT_PRODUCT(int num_steals = 0;) 4640 oop obj_to_scan; 4641 CMSBitMap* bm = &(_collector->_markBitMap); 4642 4643 while (true) { 4644 // Completely finish any left over work from (an) earlier round(s) 4645 cl->trim_queue(0); 4646 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4647 (size_t)ParGCDesiredObjsFromOverflowList); 4648 // Now check if there's any work in the overflow list 4649 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4650 // only affects the number of attempts made to get work from the 4651 // overflow list and does not affect the number of workers. Just 4652 // pass ParallelGCThreads so this behavior is unchanged. 4653 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4654 work_q, 4655 ParallelGCThreads)) { 4656 // found something in global overflow list; 4657 // not yet ready to go stealing work from others. 4658 // We'd like to assert(work_q->size() != 0, ...) 4659 // because we just took work from the overflow list, 4660 // but of course we can't since all of that could have 4661 // been already stolen from us. 4662 // "He giveth and He taketh away." 4663 continue; 4664 } 4665 // Verify that we have no work before we resort to stealing 4666 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4667 // Try to steal from other queues that have work 4668 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4669 NOT_PRODUCT(num_steals++;) 4670 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4671 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4672 // Do scanning work 4673 obj_to_scan->oop_iterate(cl); 4674 // Loop around, finish this work, and try to steal some more 4675 } else if (terminator()->offer_termination()) { 4676 break; // nirvana from the infinite cycle 4677 } 4678 } 4679 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4680 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4681 "Else our work is not yet done"); 4682 } 4683 4684 // Record object boundaries in _eden_chunk_array by sampling the eden 4685 // top in the slow-path eden object allocation code path and record 4686 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4687 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4688 // sampling in sample_eden() that activates during the part of the 4689 // preclean phase. 4690 void CMSCollector::sample_eden_chunk() { 4691 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4692 if (_eden_chunk_lock->try_lock()) { 4693 // Record a sample. This is the critical section. The contents 4694 // of the _eden_chunk_array have to be non-decreasing in the 4695 // address order. 4696 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4697 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4698 "Unexpected state of Eden"); 4699 if (_eden_chunk_index == 0 || 4700 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4701 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4702 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4703 _eden_chunk_index++; // commit sample 4704 } 4705 _eden_chunk_lock->unlock(); 4706 } 4707 } 4708 } 4709 4710 // Return a thread-local PLAB recording array, as appropriate. 4711 void* CMSCollector::get_data_recorder(int thr_num) { 4712 if (_survivor_plab_array != NULL && 4713 (CMSPLABRecordAlways || 4714 (_collectorState > Marking && _collectorState < FinalMarking))) { 4715 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4716 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4717 ca->reset(); // clear it so that fresh data is recorded 4718 return (void*) ca; 4719 } else { 4720 return NULL; 4721 } 4722 } 4723 4724 // Reset all the thread-local PLAB recording arrays 4725 void CMSCollector::reset_survivor_plab_arrays() { 4726 for (uint i = 0; i < ParallelGCThreads; i++) { 4727 _survivor_plab_array[i].reset(); 4728 } 4729 } 4730 4731 // Merge the per-thread plab arrays into the global survivor chunk 4732 // array which will provide the partitioning of the survivor space 4733 // for CMS initial scan and rescan. 4734 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4735 int no_of_gc_threads) { 4736 assert(_survivor_plab_array != NULL, "Error"); 4737 assert(_survivor_chunk_array != NULL, "Error"); 4738 assert(_collectorState == FinalMarking || 4739 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4740 for (int j = 0; j < no_of_gc_threads; j++) { 4741 _cursor[j] = 0; 4742 } 4743 HeapWord* top = surv->top(); 4744 size_t i; 4745 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4746 HeapWord* min_val = top; // Higher than any PLAB address 4747 uint min_tid = 0; // position of min_val this round 4748 for (int j = 0; j < no_of_gc_threads; j++) { 4749 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4750 if (_cursor[j] == cur_sca->end()) { 4751 continue; 4752 } 4753 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4754 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4755 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4756 if (cur_val < min_val) { 4757 min_tid = j; 4758 min_val = cur_val; 4759 } else { 4760 assert(cur_val < top, "All recorded addresses should be less"); 4761 } 4762 } 4763 // At this point min_val and min_tid are respectively 4764 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4765 // and the thread (j) that witnesses that address. 4766 // We record this address in the _survivor_chunk_array[i] 4767 // and increment _cursor[min_tid] prior to the next round i. 4768 if (min_val == top) { 4769 break; 4770 } 4771 _survivor_chunk_array[i] = min_val; 4772 _cursor[min_tid]++; 4773 } 4774 // We are all done; record the size of the _survivor_chunk_array 4775 _survivor_chunk_index = i; // exclusive: [0, i) 4776 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4777 // Verify that we used up all the recorded entries 4778 #ifdef ASSERT 4779 size_t total = 0; 4780 for (int j = 0; j < no_of_gc_threads; j++) { 4781 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4782 total += _cursor[j]; 4783 } 4784 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4785 // Check that the merged array is in sorted order 4786 if (total > 0) { 4787 for (size_t i = 0; i < total - 1; i++) { 4788 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4789 i, p2i(_survivor_chunk_array[i])); 4790 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4791 "Not sorted"); 4792 } 4793 } 4794 #endif // ASSERT 4795 } 4796 4797 // Set up the space's par_seq_tasks structure for work claiming 4798 // for parallel initial scan and rescan of young gen. 4799 // See ParRescanTask where this is currently used. 4800 void 4801 CMSCollector:: 4802 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4803 assert(n_threads > 0, "Unexpected n_threads argument"); 4804 4805 // Eden space 4806 if (!_young_gen->eden()->is_empty()) { 4807 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4808 assert(!pst->valid(), "Clobbering existing data?"); 4809 // Each valid entry in [0, _eden_chunk_index) represents a task. 4810 size_t n_tasks = _eden_chunk_index + 1; 4811 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4812 // Sets the condition for completion of the subtask (how many threads 4813 // need to finish in order to be done). 4814 pst->set_n_threads(n_threads); 4815 pst->set_n_tasks((int)n_tasks); 4816 } 4817 4818 // Merge the survivor plab arrays into _survivor_chunk_array 4819 if (_survivor_plab_array != NULL) { 4820 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4821 } else { 4822 assert(_survivor_chunk_index == 0, "Error"); 4823 } 4824 4825 // To space 4826 { 4827 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4828 assert(!pst->valid(), "Clobbering existing data?"); 4829 // Sets the condition for completion of the subtask (how many threads 4830 // need to finish in order to be done). 4831 pst->set_n_threads(n_threads); 4832 pst->set_n_tasks(1); 4833 assert(pst->valid(), "Error"); 4834 } 4835 4836 // From space 4837 { 4838 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4839 assert(!pst->valid(), "Clobbering existing data?"); 4840 size_t n_tasks = _survivor_chunk_index + 1; 4841 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4842 // Sets the condition for completion of the subtask (how many threads 4843 // need to finish in order to be done). 4844 pst->set_n_threads(n_threads); 4845 pst->set_n_tasks((int)n_tasks); 4846 assert(pst->valid(), "Error"); 4847 } 4848 } 4849 4850 // Parallel version of remark 4851 void CMSCollector::do_remark_parallel() { 4852 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4853 WorkGang* workers = gch->workers(); 4854 assert(workers != NULL, "Need parallel worker threads."); 4855 // Choose to use the number of GC workers most recently set 4856 // into "active_workers". 4857 uint n_workers = workers->active_workers(); 4858 4859 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4860 4861 StrongRootsScope srs(n_workers); 4862 4863 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4864 4865 // We won't be iterating over the cards in the card table updating 4866 // the younger_gen cards, so we shouldn't call the following else 4867 // the verification code as well as subsequent younger_refs_iterate 4868 // code would get confused. XXX 4869 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4870 4871 // The young gen rescan work will not be done as part of 4872 // process_roots (which currently doesn't know how to 4873 // parallelize such a scan), but rather will be broken up into 4874 // a set of parallel tasks (via the sampling that the [abortable] 4875 // preclean phase did of eden, plus the [two] tasks of 4876 // scanning the [two] survivor spaces. Further fine-grain 4877 // parallelization of the scanning of the survivor spaces 4878 // themselves, and of precleaning of the young gen itself 4879 // is deferred to the future. 4880 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4881 4882 // The dirty card rescan work is broken up into a "sequence" 4883 // of parallel tasks (per constituent space) that are dynamically 4884 // claimed by the parallel threads. 4885 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4886 4887 // It turns out that even when we're using 1 thread, doing the work in a 4888 // separate thread causes wide variance in run times. We can't help this 4889 // in the multi-threaded case, but we special-case n=1 here to get 4890 // repeatable measurements of the 1-thread overhead of the parallel code. 4891 if (n_workers > 1) { 4892 // Make refs discovery MT-safe, if it isn't already: it may not 4893 // necessarily be so, since it's possible that we are doing 4894 // ST marking. 4895 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4896 workers->run_task(&tsk); 4897 } else { 4898 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4899 tsk.work(0); 4900 } 4901 4902 // restore, single-threaded for now, any preserved marks 4903 // as a result of work_q overflow 4904 restore_preserved_marks_if_any(); 4905 } 4906 4907 // Non-parallel version of remark 4908 void CMSCollector::do_remark_non_parallel() { 4909 ResourceMark rm; 4910 HandleMark hm; 4911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4912 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4913 4914 MarkRefsIntoAndScanClosure 4915 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4916 &_markStack, this, 4917 false /* should_yield */, false /* not precleaning */); 4918 MarkFromDirtyCardsClosure 4919 markFromDirtyCardsClosure(this, _span, 4920 NULL, // space is set further below 4921 &_markBitMap, &_markStack, &mrias_cl); 4922 { 4923 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4924 // Iterate over the dirty cards, setting the corresponding bits in the 4925 // mod union table. 4926 { 4927 ModUnionClosure modUnionClosure(&_modUnionTable); 4928 _ct->ct_bs()->dirty_card_iterate( 4929 _cmsGen->used_region(), 4930 &modUnionClosure); 4931 } 4932 // Having transferred these marks into the modUnionTable, we just need 4933 // to rescan the marked objects on the dirty cards in the modUnionTable. 4934 // The initial marking may have been done during an asynchronous 4935 // collection so there may be dirty bits in the mod-union table. 4936 const int alignment = 4937 CardTableModRefBS::card_size * BitsPerWord; 4938 { 4939 // ... First handle dirty cards in CMS gen 4940 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4941 MemRegion ur = _cmsGen->used_region(); 4942 HeapWord* lb = ur.start(); 4943 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4944 MemRegion cms_span(lb, ub); 4945 _modUnionTable.dirty_range_iterate_clear(cms_span, 4946 &markFromDirtyCardsClosure); 4947 verify_work_stacks_empty(); 4948 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4949 } 4950 } 4951 if (VerifyDuringGC && 4952 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4953 HandleMark hm; // Discard invalid handles created during verification 4954 Universe::verify(); 4955 } 4956 { 4957 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4958 4959 verify_work_stacks_empty(); 4960 4961 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4962 StrongRootsScope srs(1); 4963 4964 gch->gen_process_roots(&srs, 4965 GenCollectedHeap::OldGen, 4966 true, // young gen as roots 4967 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4968 should_unload_classes(), 4969 &mrias_cl, 4970 NULL, 4971 NULL); // The dirty klasses will be handled below 4972 4973 assert(should_unload_classes() 4974 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4975 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4976 } 4977 4978 { 4979 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4980 4981 verify_work_stacks_empty(); 4982 4983 // Scan all class loader data objects that might have been introduced 4984 // during concurrent marking. 4985 ResourceMark rm; 4986 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4987 for (int i = 0; i < array->length(); i++) { 4988 mrias_cl.do_cld_nv(array->at(i)); 4989 } 4990 4991 // We don't need to keep track of new CLDs anymore. 4992 ClassLoaderDataGraph::remember_new_clds(false); 4993 4994 verify_work_stacks_empty(); 4995 } 4996 4997 { 4998 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm); 4999 5000 verify_work_stacks_empty(); 5001 5002 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5003 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5004 5005 verify_work_stacks_empty(); 5006 } 5007 5008 // We might have added oops to ClassLoaderData::_handles during the 5009 // concurrent marking phase. These oops point to newly allocated objects 5010 // that are guaranteed to be kept alive. Either by the direct allocation 5011 // code, or when the young collector processes the roots. Hence, 5012 // we don't have to revisit the _handles block during the remark phase. 5013 5014 verify_work_stacks_empty(); 5015 // Restore evacuated mark words, if any, used for overflow list links 5016 restore_preserved_marks_if_any(); 5017 5018 verify_overflow_empty(); 5019 } 5020 5021 //////////////////////////////////////////////////////// 5022 // Parallel Reference Processing Task Proxy Class 5023 //////////////////////////////////////////////////////// 5024 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5025 OopTaskQueueSet* _queues; 5026 ParallelTaskTerminator _terminator; 5027 public: 5028 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5029 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5030 ParallelTaskTerminator* terminator() { return &_terminator; } 5031 OopTaskQueueSet* queues() { return _queues; } 5032 }; 5033 5034 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5035 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5036 CMSCollector* _collector; 5037 CMSBitMap* _mark_bit_map; 5038 const MemRegion _span; 5039 ProcessTask& _task; 5040 5041 public: 5042 CMSRefProcTaskProxy(ProcessTask& task, 5043 CMSCollector* collector, 5044 const MemRegion& span, 5045 CMSBitMap* mark_bit_map, 5046 AbstractWorkGang* workers, 5047 OopTaskQueueSet* task_queues): 5048 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5049 task_queues, 5050 workers->active_workers()), 5051 _task(task), 5052 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5053 { 5054 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5055 "Inconsistency in _span"); 5056 } 5057 5058 OopTaskQueueSet* task_queues() { return queues(); } 5059 5060 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5061 5062 void do_work_steal(int i, 5063 CMSParDrainMarkingStackClosure* drain, 5064 CMSParKeepAliveClosure* keep_alive, 5065 int* seed); 5066 5067 virtual void work(uint worker_id); 5068 }; 5069 5070 void CMSRefProcTaskProxy::work(uint worker_id) { 5071 ResourceMark rm; 5072 HandleMark hm; 5073 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5074 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5075 _mark_bit_map, 5076 work_queue(worker_id)); 5077 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5078 _mark_bit_map, 5079 work_queue(worker_id)); 5080 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5081 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5082 if (_task.marks_oops_alive()) { 5083 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5084 _collector->hash_seed(worker_id)); 5085 } 5086 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5087 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5088 } 5089 5090 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5091 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5092 EnqueueTask& _task; 5093 5094 public: 5095 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5096 : AbstractGangTask("Enqueue reference objects in parallel"), 5097 _task(task) 5098 { } 5099 5100 virtual void work(uint worker_id) 5101 { 5102 _task.work(worker_id); 5103 } 5104 }; 5105 5106 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5107 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5108 _span(span), 5109 _bit_map(bit_map), 5110 _work_queue(work_queue), 5111 _mark_and_push(collector, span, bit_map, work_queue), 5112 _low_water_mark(MIN2((work_queue->max_elems()/4), 5113 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5114 { } 5115 5116 // . see if we can share work_queues with ParNew? XXX 5117 void CMSRefProcTaskProxy::do_work_steal(int i, 5118 CMSParDrainMarkingStackClosure* drain, 5119 CMSParKeepAliveClosure* keep_alive, 5120 int* seed) { 5121 OopTaskQueue* work_q = work_queue(i); 5122 NOT_PRODUCT(int num_steals = 0;) 5123 oop obj_to_scan; 5124 5125 while (true) { 5126 // Completely finish any left over work from (an) earlier round(s) 5127 drain->trim_queue(0); 5128 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5129 (size_t)ParGCDesiredObjsFromOverflowList); 5130 // Now check if there's any work in the overflow list 5131 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5132 // only affects the number of attempts made to get work from the 5133 // overflow list and does not affect the number of workers. Just 5134 // pass ParallelGCThreads so this behavior is unchanged. 5135 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5136 work_q, 5137 ParallelGCThreads)) { 5138 // Found something in global overflow list; 5139 // not yet ready to go stealing work from others. 5140 // We'd like to assert(work_q->size() != 0, ...) 5141 // because we just took work from the overflow list, 5142 // but of course we can't, since all of that might have 5143 // been already stolen from us. 5144 continue; 5145 } 5146 // Verify that we have no work before we resort to stealing 5147 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5148 // Try to steal from other queues that have work 5149 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5150 NOT_PRODUCT(num_steals++;) 5151 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5152 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5153 // Do scanning work 5154 obj_to_scan->oop_iterate(keep_alive); 5155 // Loop around, finish this work, and try to steal some more 5156 } else if (terminator()->offer_termination()) { 5157 break; // nirvana from the infinite cycle 5158 } 5159 } 5160 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5161 } 5162 5163 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5164 { 5165 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5166 WorkGang* workers = gch->workers(); 5167 assert(workers != NULL, "Need parallel worker threads."); 5168 CMSRefProcTaskProxy rp_task(task, &_collector, 5169 _collector.ref_processor()->span(), 5170 _collector.markBitMap(), 5171 workers, _collector.task_queues()); 5172 workers->run_task(&rp_task); 5173 } 5174 5175 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5176 { 5177 5178 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5179 WorkGang* workers = gch->workers(); 5180 assert(workers != NULL, "Need parallel worker threads."); 5181 CMSRefEnqueueTaskProxy enq_task(task); 5182 workers->run_task(&enq_task); 5183 } 5184 5185 void CMSCollector::refProcessingWork() { 5186 ResourceMark rm; 5187 HandleMark hm; 5188 5189 ReferenceProcessor* rp = ref_processor(); 5190 assert(rp->span().equals(_span), "Spans should be equal"); 5191 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5192 // Process weak references. 5193 rp->setup_policy(false); 5194 verify_work_stacks_empty(); 5195 5196 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5197 &_markStack, false /* !preclean */); 5198 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5199 _span, &_markBitMap, &_markStack, 5200 &cmsKeepAliveClosure, false /* !preclean */); 5201 { 5202 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5203 5204 ReferenceProcessorStats stats; 5205 if (rp->processing_is_mt()) { 5206 // Set the degree of MT here. If the discovery is done MT, there 5207 // may have been a different number of threads doing the discovery 5208 // and a different number of discovered lists may have Ref objects. 5209 // That is OK as long as the Reference lists are balanced (see 5210 // balance_all_queues() and balance_queues()). 5211 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5212 uint active_workers = ParallelGCThreads; 5213 WorkGang* workers = gch->workers(); 5214 if (workers != NULL) { 5215 active_workers = workers->active_workers(); 5216 // The expectation is that active_workers will have already 5217 // been set to a reasonable value. If it has not been set, 5218 // investigate. 5219 assert(active_workers > 0, "Should have been set during scavenge"); 5220 } 5221 rp->set_active_mt_degree(active_workers); 5222 CMSRefProcTaskExecutor task_executor(*this); 5223 stats = rp->process_discovered_references(&_is_alive_closure, 5224 &cmsKeepAliveClosure, 5225 &cmsDrainMarkingStackClosure, 5226 &task_executor, 5227 _gc_timer_cm); 5228 } else { 5229 stats = rp->process_discovered_references(&_is_alive_closure, 5230 &cmsKeepAliveClosure, 5231 &cmsDrainMarkingStackClosure, 5232 NULL, 5233 _gc_timer_cm); 5234 } 5235 _gc_tracer_cm->report_gc_reference_stats(stats); 5236 5237 } 5238 5239 // This is the point where the entire marking should have completed. 5240 verify_work_stacks_empty(); 5241 5242 if (should_unload_classes()) { 5243 { 5244 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5245 5246 // Unload classes and purge the SystemDictionary. 5247 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5248 5249 // Unload nmethods. 5250 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5251 5252 // Prune dead klasses from subklass/sibling/implementor lists. 5253 Klass::clean_weak_klass_links(&_is_alive_closure); 5254 } 5255 5256 { 5257 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5258 // Clean up unreferenced symbols in symbol table. 5259 SymbolTable::unlink(); 5260 } 5261 5262 { 5263 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5264 // Delete entries for dead interned strings. 5265 StringTable::unlink(&_is_alive_closure); 5266 } 5267 } 5268 5269 5270 // Restore any preserved marks as a result of mark stack or 5271 // work queue overflow 5272 restore_preserved_marks_if_any(); // done single-threaded for now 5273 5274 rp->set_enqueuing_is_done(true); 5275 if (rp->processing_is_mt()) { 5276 rp->balance_all_queues(); 5277 CMSRefProcTaskExecutor task_executor(*this); 5278 rp->enqueue_discovered_references(&task_executor); 5279 } else { 5280 rp->enqueue_discovered_references(NULL); 5281 } 5282 rp->verify_no_references_recorded(); 5283 assert(!rp->discovery_enabled(), "should have been disabled"); 5284 } 5285 5286 #ifndef PRODUCT 5287 void CMSCollector::check_correct_thread_executing() { 5288 Thread* t = Thread::current(); 5289 // Only the VM thread or the CMS thread should be here. 5290 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5291 "Unexpected thread type"); 5292 // If this is the vm thread, the foreground process 5293 // should not be waiting. Note that _foregroundGCIsActive is 5294 // true while the foreground collector is waiting. 5295 if (_foregroundGCShouldWait) { 5296 // We cannot be the VM thread 5297 assert(t->is_ConcurrentGC_thread(), 5298 "Should be CMS thread"); 5299 } else { 5300 // We can be the CMS thread only if we are in a stop-world 5301 // phase of CMS collection. 5302 if (t->is_ConcurrentGC_thread()) { 5303 assert(_collectorState == InitialMarking || 5304 _collectorState == FinalMarking, 5305 "Should be a stop-world phase"); 5306 // The CMS thread should be holding the CMS_token. 5307 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5308 "Potential interference with concurrently " 5309 "executing VM thread"); 5310 } 5311 } 5312 } 5313 #endif 5314 5315 void CMSCollector::sweep() { 5316 assert(_collectorState == Sweeping, "just checking"); 5317 check_correct_thread_executing(); 5318 verify_work_stacks_empty(); 5319 verify_overflow_empty(); 5320 increment_sweep_count(); 5321 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5322 5323 _inter_sweep_timer.stop(); 5324 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5325 5326 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5327 _intra_sweep_timer.reset(); 5328 _intra_sweep_timer.start(); 5329 { 5330 GCTraceCPUTime tcpu; 5331 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5332 // First sweep the old gen 5333 { 5334 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5335 bitMapLock()); 5336 sweepWork(_cmsGen); 5337 } 5338 5339 // Update Universe::_heap_*_at_gc figures. 5340 // We need all the free list locks to make the abstract state 5341 // transition from Sweeping to Resetting. See detailed note 5342 // further below. 5343 { 5344 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5345 // Update heap occupancy information which is used as 5346 // input to soft ref clearing policy at the next gc. 5347 Universe::update_heap_info_at_gc(); 5348 _collectorState = Resizing; 5349 } 5350 } 5351 verify_work_stacks_empty(); 5352 verify_overflow_empty(); 5353 5354 if (should_unload_classes()) { 5355 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5356 // requires that the virtual spaces are stable and not deleted. 5357 ClassLoaderDataGraph::set_should_purge(true); 5358 } 5359 5360 _intra_sweep_timer.stop(); 5361 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5362 5363 _inter_sweep_timer.reset(); 5364 _inter_sweep_timer.start(); 5365 5366 // We need to use a monotonically non-decreasing time in ms 5367 // or we will see time-warp warnings and os::javaTimeMillis() 5368 // does not guarantee monotonicity. 5369 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5370 update_time_of_last_gc(now); 5371 5372 // NOTE on abstract state transitions: 5373 // Mutators allocate-live and/or mark the mod-union table dirty 5374 // based on the state of the collection. The former is done in 5375 // the interval [Marking, Sweeping] and the latter in the interval 5376 // [Marking, Sweeping). Thus the transitions into the Marking state 5377 // and out of the Sweeping state must be synchronously visible 5378 // globally to the mutators. 5379 // The transition into the Marking state happens with the world 5380 // stopped so the mutators will globally see it. Sweeping is 5381 // done asynchronously by the background collector so the transition 5382 // from the Sweeping state to the Resizing state must be done 5383 // under the freelistLock (as is the check for whether to 5384 // allocate-live and whether to dirty the mod-union table). 5385 assert(_collectorState == Resizing, "Change of collector state to" 5386 " Resizing must be done under the freelistLocks (plural)"); 5387 5388 // Now that sweeping has been completed, we clear 5389 // the incremental_collection_failed flag, 5390 // thus inviting a younger gen collection to promote into 5391 // this generation. If such a promotion may still fail, 5392 // the flag will be set again when a young collection is 5393 // attempted. 5394 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5395 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5396 gch->update_full_collections_completed(_collection_count_start); 5397 } 5398 5399 // FIX ME!!! Looks like this belongs in CFLSpace, with 5400 // CMSGen merely delegating to it. 5401 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5402 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5403 HeapWord* minAddr = _cmsSpace->bottom(); 5404 HeapWord* largestAddr = 5405 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5406 if (largestAddr == NULL) { 5407 // The dictionary appears to be empty. In this case 5408 // try to coalesce at the end of the heap. 5409 largestAddr = _cmsSpace->end(); 5410 } 5411 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5412 size_t nearLargestOffset = 5413 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5414 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5415 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5416 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5417 } 5418 5419 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5420 return addr >= _cmsSpace->nearLargestChunk(); 5421 } 5422 5423 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5424 return _cmsSpace->find_chunk_at_end(); 5425 } 5426 5427 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5428 bool full) { 5429 // If the young generation has been collected, gather any statistics 5430 // that are of interest at this point. 5431 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5432 if (!full && current_is_young) { 5433 // Gather statistics on the young generation collection. 5434 collector()->stats().record_gc0_end(used()); 5435 } 5436 } 5437 5438 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5439 // We iterate over the space(s) underlying this generation, 5440 // checking the mark bit map to see if the bits corresponding 5441 // to specific blocks are marked or not. Blocks that are 5442 // marked are live and are not swept up. All remaining blocks 5443 // are swept up, with coalescing on-the-fly as we sweep up 5444 // contiguous free and/or garbage blocks: 5445 // We need to ensure that the sweeper synchronizes with allocators 5446 // and stop-the-world collectors. In particular, the following 5447 // locks are used: 5448 // . CMS token: if this is held, a stop the world collection cannot occur 5449 // . freelistLock: if this is held no allocation can occur from this 5450 // generation by another thread 5451 // . bitMapLock: if this is held, no other thread can access or update 5452 // 5453 5454 // Note that we need to hold the freelistLock if we use 5455 // block iterate below; else the iterator might go awry if 5456 // a mutator (or promotion) causes block contents to change 5457 // (for instance if the allocator divvies up a block). 5458 // If we hold the free list lock, for all practical purposes 5459 // young generation GC's can't occur (they'll usually need to 5460 // promote), so we might as well prevent all young generation 5461 // GC's while we do a sweeping step. For the same reason, we might 5462 // as well take the bit map lock for the entire duration 5463 5464 // check that we hold the requisite locks 5465 assert(have_cms_token(), "Should hold cms token"); 5466 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5467 assert_lock_strong(old_gen->freelistLock()); 5468 assert_lock_strong(bitMapLock()); 5469 5470 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5471 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5472 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5473 _inter_sweep_estimate.padded_average(), 5474 _intra_sweep_estimate.padded_average()); 5475 old_gen->setNearLargestChunk(); 5476 5477 { 5478 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5479 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5480 // We need to free-up/coalesce garbage/blocks from a 5481 // co-terminal free run. This is done in the SweepClosure 5482 // destructor; so, do not remove this scope, else the 5483 // end-of-sweep-census below will be off by a little bit. 5484 } 5485 old_gen->cmsSpace()->sweep_completed(); 5486 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5487 if (should_unload_classes()) { // unloaded classes this cycle, 5488 _concurrent_cycles_since_last_unload = 0; // ... reset count 5489 } else { // did not unload classes, 5490 _concurrent_cycles_since_last_unload++; // ... increment count 5491 } 5492 } 5493 5494 // Reset CMS data structures (for now just the marking bit map) 5495 // preparatory for the next cycle. 5496 void CMSCollector::reset_concurrent() { 5497 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5498 5499 // If the state is not "Resetting", the foreground thread 5500 // has done a collection and the resetting. 5501 if (_collectorState != Resetting) { 5502 assert(_collectorState == Idling, "The state should only change" 5503 " because the foreground collector has finished the collection"); 5504 return; 5505 } 5506 5507 { 5508 // Clear the mark bitmap (no grey objects to start with) 5509 // for the next cycle. 5510 GCTraceCPUTime tcpu; 5511 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5512 5513 HeapWord* curAddr = _markBitMap.startWord(); 5514 while (curAddr < _markBitMap.endWord()) { 5515 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5516 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5517 _markBitMap.clear_large_range(chunk); 5518 if (ConcurrentMarkSweepThread::should_yield() && 5519 !foregroundGCIsActive() && 5520 CMSYield) { 5521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5522 "CMS thread should hold CMS token"); 5523 assert_lock_strong(bitMapLock()); 5524 bitMapLock()->unlock(); 5525 ConcurrentMarkSweepThread::desynchronize(true); 5526 stopTimer(); 5527 incrementYields(); 5528 5529 // See the comment in coordinator_yield() 5530 for (unsigned i = 0; i < CMSYieldSleepCount && 5531 ConcurrentMarkSweepThread::should_yield() && 5532 !CMSCollector::foregroundGCIsActive(); ++i) { 5533 os::sleep(Thread::current(), 1, false); 5534 } 5535 5536 ConcurrentMarkSweepThread::synchronize(true); 5537 bitMapLock()->lock_without_safepoint_check(); 5538 startTimer(); 5539 } 5540 curAddr = chunk.end(); 5541 } 5542 // A successful mostly concurrent collection has been done. 5543 // Because only the full (i.e., concurrent mode failure) collections 5544 // are being measured for gc overhead limits, clean the "near" flag 5545 // and count. 5546 size_policy()->reset_gc_overhead_limit_count(); 5547 _collectorState = Idling; 5548 } 5549 5550 register_gc_end(); 5551 } 5552 5553 // Same as above but for STW paths 5554 void CMSCollector::reset_stw() { 5555 // already have the lock 5556 assert(_collectorState == Resetting, "just checking"); 5557 assert_lock_strong(bitMapLock()); 5558 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5559 _markBitMap.clear_all(); 5560 _collectorState = Idling; 5561 register_gc_end(); 5562 } 5563 5564 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5565 GCTraceCPUTime tcpu; 5566 TraceCollectorStats tcs(counters()); 5567 5568 switch (op) { 5569 case CMS_op_checkpointRootsInitial: { 5570 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5571 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5572 checkpointRootsInitial(); 5573 break; 5574 } 5575 case CMS_op_checkpointRootsFinal: { 5576 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5577 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5578 checkpointRootsFinal(); 5579 break; 5580 } 5581 default: 5582 fatal("No such CMS_op"); 5583 } 5584 } 5585 5586 #ifndef PRODUCT 5587 size_t const CMSCollector::skip_header_HeapWords() { 5588 return FreeChunk::header_size(); 5589 } 5590 5591 // Try and collect here conditions that should hold when 5592 // CMS thread is exiting. The idea is that the foreground GC 5593 // thread should not be blocked if it wants to terminate 5594 // the CMS thread and yet continue to run the VM for a while 5595 // after that. 5596 void CMSCollector::verify_ok_to_terminate() const { 5597 assert(Thread::current()->is_ConcurrentGC_thread(), 5598 "should be called by CMS thread"); 5599 assert(!_foregroundGCShouldWait, "should be false"); 5600 // We could check here that all the various low-level locks 5601 // are not held by the CMS thread, but that is overkill; see 5602 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5603 // is checked. 5604 } 5605 #endif 5606 5607 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5608 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5609 "missing Printezis mark?"); 5610 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5611 size_t size = pointer_delta(nextOneAddr + 1, addr); 5612 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5613 "alignment problem"); 5614 assert(size >= 3, "Necessary for Printezis marks to work"); 5615 return size; 5616 } 5617 5618 // A variant of the above (block_size_using_printezis_bits()) except 5619 // that we return 0 if the P-bits are not yet set. 5620 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5621 if (_markBitMap.isMarked(addr + 1)) { 5622 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5623 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5624 size_t size = pointer_delta(nextOneAddr + 1, addr); 5625 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5626 "alignment problem"); 5627 assert(size >= 3, "Necessary for Printezis marks to work"); 5628 return size; 5629 } 5630 return 0; 5631 } 5632 5633 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5634 size_t sz = 0; 5635 oop p = (oop)addr; 5636 if (p->klass_or_null() != NULL) { 5637 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5638 } else { 5639 sz = block_size_using_printezis_bits(addr); 5640 } 5641 assert(sz > 0, "size must be nonzero"); 5642 HeapWord* next_block = addr + sz; 5643 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5644 CardTableModRefBS::card_size); 5645 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5646 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5647 "must be different cards"); 5648 return next_card; 5649 } 5650 5651 5652 // CMS Bit Map Wrapper ///////////////////////////////////////// 5653 5654 // Construct a CMS bit map infrastructure, but don't create the 5655 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5656 // further below. 5657 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5658 _bm(), 5659 _shifter(shifter), 5660 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5661 Monitor::_safepoint_check_sometimes) : NULL) 5662 { 5663 _bmStartWord = 0; 5664 _bmWordSize = 0; 5665 } 5666 5667 bool CMSBitMap::allocate(MemRegion mr) { 5668 _bmStartWord = mr.start(); 5669 _bmWordSize = mr.word_size(); 5670 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5671 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5672 if (!brs.is_reserved()) { 5673 log_warning(gc)("CMS bit map allocation failure"); 5674 return false; 5675 } 5676 // For now we'll just commit all of the bit map up front. 5677 // Later on we'll try to be more parsimonious with swap. 5678 if (!_virtual_space.initialize(brs, brs.size())) { 5679 log_warning(gc)("CMS bit map backing store failure"); 5680 return false; 5681 } 5682 assert(_virtual_space.committed_size() == brs.size(), 5683 "didn't reserve backing store for all of CMS bit map?"); 5684 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5685 _bmWordSize, "inconsistency in bit map sizing"); 5686 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5687 5688 // bm.clear(); // can we rely on getting zero'd memory? verify below 5689 assert(isAllClear(), 5690 "Expected zero'd memory from ReservedSpace constructor"); 5691 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5692 "consistency check"); 5693 return true; 5694 } 5695 5696 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5697 HeapWord *next_addr, *end_addr, *last_addr; 5698 assert_locked(); 5699 assert(covers(mr), "out-of-range error"); 5700 // XXX assert that start and end are appropriately aligned 5701 for (next_addr = mr.start(), end_addr = mr.end(); 5702 next_addr < end_addr; next_addr = last_addr) { 5703 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5704 last_addr = dirty_region.end(); 5705 if (!dirty_region.is_empty()) { 5706 cl->do_MemRegion(dirty_region); 5707 } else { 5708 assert(last_addr == end_addr, "program logic"); 5709 return; 5710 } 5711 } 5712 } 5713 5714 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5715 _bm.print_on_error(st, prefix); 5716 } 5717 5718 #ifndef PRODUCT 5719 void CMSBitMap::assert_locked() const { 5720 CMSLockVerifier::assert_locked(lock()); 5721 } 5722 5723 bool CMSBitMap::covers(MemRegion mr) const { 5724 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5725 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5726 "size inconsistency"); 5727 return (mr.start() >= _bmStartWord) && 5728 (mr.end() <= endWord()); 5729 } 5730 5731 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5732 return (start >= _bmStartWord && (start + size) <= endWord()); 5733 } 5734 5735 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5736 // verify that there are no 1 bits in the interval [left, right) 5737 FalseBitMapClosure falseBitMapClosure; 5738 iterate(&falseBitMapClosure, left, right); 5739 } 5740 5741 void CMSBitMap::region_invariant(MemRegion mr) 5742 { 5743 assert_locked(); 5744 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5745 assert(!mr.is_empty(), "unexpected empty region"); 5746 assert(covers(mr), "mr should be covered by bit map"); 5747 // convert address range into offset range 5748 size_t start_ofs = heapWordToOffset(mr.start()); 5749 // Make sure that end() is appropriately aligned 5750 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5751 (1 << (_shifter+LogHeapWordSize))), 5752 "Misaligned mr.end()"); 5753 size_t end_ofs = heapWordToOffset(mr.end()); 5754 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5755 } 5756 5757 #endif 5758 5759 bool CMSMarkStack::allocate(size_t size) { 5760 // allocate a stack of the requisite depth 5761 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5762 size * sizeof(oop))); 5763 if (!rs.is_reserved()) { 5764 log_warning(gc)("CMSMarkStack allocation failure"); 5765 return false; 5766 } 5767 if (!_virtual_space.initialize(rs, rs.size())) { 5768 log_warning(gc)("CMSMarkStack backing store failure"); 5769 return false; 5770 } 5771 assert(_virtual_space.committed_size() == rs.size(), 5772 "didn't reserve backing store for all of CMS stack?"); 5773 _base = (oop*)(_virtual_space.low()); 5774 _index = 0; 5775 _capacity = size; 5776 NOT_PRODUCT(_max_depth = 0); 5777 return true; 5778 } 5779 5780 // XXX FIX ME !!! In the MT case we come in here holding a 5781 // leaf lock. For printing we need to take a further lock 5782 // which has lower rank. We need to recalibrate the two 5783 // lock-ranks involved in order to be able to print the 5784 // messages below. (Or defer the printing to the caller. 5785 // For now we take the expedient path of just disabling the 5786 // messages for the problematic case.) 5787 void CMSMarkStack::expand() { 5788 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5789 if (_capacity == MarkStackSizeMax) { 5790 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5791 // We print a warning message only once per CMS cycle. 5792 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5793 } 5794 return; 5795 } 5796 // Double capacity if possible 5797 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5798 // Do not give up existing stack until we have managed to 5799 // get the double capacity that we desired. 5800 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5801 new_capacity * sizeof(oop))); 5802 if (rs.is_reserved()) { 5803 // Release the backing store associated with old stack 5804 _virtual_space.release(); 5805 // Reinitialize virtual space for new stack 5806 if (!_virtual_space.initialize(rs, rs.size())) { 5807 fatal("Not enough swap for expanded marking stack"); 5808 } 5809 _base = (oop*)(_virtual_space.low()); 5810 _index = 0; 5811 _capacity = new_capacity; 5812 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5813 // Failed to double capacity, continue; 5814 // we print a detail message only once per CMS cycle. 5815 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5816 _capacity / K, new_capacity / K); 5817 } 5818 } 5819 5820 5821 // Closures 5822 // XXX: there seems to be a lot of code duplication here; 5823 // should refactor and consolidate common code. 5824 5825 // This closure is used to mark refs into the CMS generation in 5826 // the CMS bit map. Called at the first checkpoint. This closure 5827 // assumes that we do not need to re-mark dirty cards; if the CMS 5828 // generation on which this is used is not an oldest 5829 // generation then this will lose younger_gen cards! 5830 5831 MarkRefsIntoClosure::MarkRefsIntoClosure( 5832 MemRegion span, CMSBitMap* bitMap): 5833 _span(span), 5834 _bitMap(bitMap) 5835 { 5836 assert(ref_processor() == NULL, "deliberately left NULL"); 5837 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5838 } 5839 5840 void MarkRefsIntoClosure::do_oop(oop obj) { 5841 // if p points into _span, then mark corresponding bit in _markBitMap 5842 assert(obj->is_oop(), "expected an oop"); 5843 HeapWord* addr = (HeapWord*)obj; 5844 if (_span.contains(addr)) { 5845 // this should be made more efficient 5846 _bitMap->mark(addr); 5847 } 5848 } 5849 5850 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5851 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5852 5853 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5854 MemRegion span, CMSBitMap* bitMap): 5855 _span(span), 5856 _bitMap(bitMap) 5857 { 5858 assert(ref_processor() == NULL, "deliberately left NULL"); 5859 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5860 } 5861 5862 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5863 // if p points into _span, then mark corresponding bit in _markBitMap 5864 assert(obj->is_oop(), "expected an oop"); 5865 HeapWord* addr = (HeapWord*)obj; 5866 if (_span.contains(addr)) { 5867 // this should be made more efficient 5868 _bitMap->par_mark(addr); 5869 } 5870 } 5871 5872 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5873 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5874 5875 // A variant of the above, used for CMS marking verification. 5876 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5877 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5878 _span(span), 5879 _verification_bm(verification_bm), 5880 _cms_bm(cms_bm) 5881 { 5882 assert(ref_processor() == NULL, "deliberately left NULL"); 5883 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5884 } 5885 5886 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5887 // if p points into _span, then mark corresponding bit in _markBitMap 5888 assert(obj->is_oop(), "expected an oop"); 5889 HeapWord* addr = (HeapWord*)obj; 5890 if (_span.contains(addr)) { 5891 _verification_bm->mark(addr); 5892 if (!_cms_bm->isMarked(addr)) { 5893 Log(gc, verify) log; 5894 ResourceMark rm; 5895 oop(addr)->print_on(log.error_stream()); 5896 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5897 fatal("... aborting"); 5898 } 5899 } 5900 } 5901 5902 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5903 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5904 5905 ////////////////////////////////////////////////// 5906 // MarkRefsIntoAndScanClosure 5907 ////////////////////////////////////////////////// 5908 5909 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5910 ReferenceProcessor* rp, 5911 CMSBitMap* bit_map, 5912 CMSBitMap* mod_union_table, 5913 CMSMarkStack* mark_stack, 5914 CMSCollector* collector, 5915 bool should_yield, 5916 bool concurrent_precleaning): 5917 _collector(collector), 5918 _span(span), 5919 _bit_map(bit_map), 5920 _mark_stack(mark_stack), 5921 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5922 mark_stack, concurrent_precleaning), 5923 _yield(should_yield), 5924 _concurrent_precleaning(concurrent_precleaning), 5925 _freelistLock(NULL) 5926 { 5927 // FIXME: Should initialize in base class constructor. 5928 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5929 set_ref_processor_internal(rp); 5930 } 5931 5932 // This closure is used to mark refs into the CMS generation at the 5933 // second (final) checkpoint, and to scan and transitively follow 5934 // the unmarked oops. It is also used during the concurrent precleaning 5935 // phase while scanning objects on dirty cards in the CMS generation. 5936 // The marks are made in the marking bit map and the marking stack is 5937 // used for keeping the (newly) grey objects during the scan. 5938 // The parallel version (Par_...) appears further below. 5939 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5940 if (obj != NULL) { 5941 assert(obj->is_oop(), "expected an oop"); 5942 HeapWord* addr = (HeapWord*)obj; 5943 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5944 assert(_collector->overflow_list_is_empty(), 5945 "overflow list should be empty"); 5946 if (_span.contains(addr) && 5947 !_bit_map->isMarked(addr)) { 5948 // mark bit map (object is now grey) 5949 _bit_map->mark(addr); 5950 // push on marking stack (stack should be empty), and drain the 5951 // stack by applying this closure to the oops in the oops popped 5952 // from the stack (i.e. blacken the grey objects) 5953 bool res = _mark_stack->push(obj); 5954 assert(res, "Should have space to push on empty stack"); 5955 do { 5956 oop new_oop = _mark_stack->pop(); 5957 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5958 assert(_bit_map->isMarked((HeapWord*)new_oop), 5959 "only grey objects on this stack"); 5960 // iterate over the oops in this oop, marking and pushing 5961 // the ones in CMS heap (i.e. in _span). 5962 new_oop->oop_iterate(&_pushAndMarkClosure); 5963 // check if it's time to yield 5964 do_yield_check(); 5965 } while (!_mark_stack->isEmpty() || 5966 (!_concurrent_precleaning && take_from_overflow_list())); 5967 // if marking stack is empty, and we are not doing this 5968 // during precleaning, then check the overflow list 5969 } 5970 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5971 assert(_collector->overflow_list_is_empty(), 5972 "overflow list was drained above"); 5973 5974 assert(_collector->no_preserved_marks(), 5975 "All preserved marks should have been restored above"); 5976 } 5977 } 5978 5979 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5980 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5981 5982 void MarkRefsIntoAndScanClosure::do_yield_work() { 5983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5984 "CMS thread should hold CMS token"); 5985 assert_lock_strong(_freelistLock); 5986 assert_lock_strong(_bit_map->lock()); 5987 // relinquish the free_list_lock and bitMaplock() 5988 _bit_map->lock()->unlock(); 5989 _freelistLock->unlock(); 5990 ConcurrentMarkSweepThread::desynchronize(true); 5991 _collector->stopTimer(); 5992 _collector->incrementYields(); 5993 5994 // See the comment in coordinator_yield() 5995 for (unsigned i = 0; 5996 i < CMSYieldSleepCount && 5997 ConcurrentMarkSweepThread::should_yield() && 5998 !CMSCollector::foregroundGCIsActive(); 5999 ++i) { 6000 os::sleep(Thread::current(), 1, false); 6001 } 6002 6003 ConcurrentMarkSweepThread::synchronize(true); 6004 _freelistLock->lock_without_safepoint_check(); 6005 _bit_map->lock()->lock_without_safepoint_check(); 6006 _collector->startTimer(); 6007 } 6008 6009 /////////////////////////////////////////////////////////// 6010 // ParMarkRefsIntoAndScanClosure: a parallel version of 6011 // MarkRefsIntoAndScanClosure 6012 /////////////////////////////////////////////////////////// 6013 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6014 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6015 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6016 _span(span), 6017 _bit_map(bit_map), 6018 _work_queue(work_queue), 6019 _low_water_mark(MIN2((work_queue->max_elems()/4), 6020 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6021 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6022 { 6023 // FIXME: Should initialize in base class constructor. 6024 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6025 set_ref_processor_internal(rp); 6026 } 6027 6028 // This closure is used to mark refs into the CMS generation at the 6029 // second (final) checkpoint, and to scan and transitively follow 6030 // the unmarked oops. The marks are made in the marking bit map and 6031 // the work_queue is used for keeping the (newly) grey objects during 6032 // the scan phase whence they are also available for stealing by parallel 6033 // threads. Since the marking bit map is shared, updates are 6034 // synchronized (via CAS). 6035 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6036 if (obj != NULL) { 6037 // Ignore mark word because this could be an already marked oop 6038 // that may be chained at the end of the overflow list. 6039 assert(obj->is_oop(true), "expected an oop"); 6040 HeapWord* addr = (HeapWord*)obj; 6041 if (_span.contains(addr) && 6042 !_bit_map->isMarked(addr)) { 6043 // mark bit map (object will become grey): 6044 // It is possible for several threads to be 6045 // trying to "claim" this object concurrently; 6046 // the unique thread that succeeds in marking the 6047 // object first will do the subsequent push on 6048 // to the work queue (or overflow list). 6049 if (_bit_map->par_mark(addr)) { 6050 // push on work_queue (which may not be empty), and trim the 6051 // queue to an appropriate length by applying this closure to 6052 // the oops in the oops popped from the stack (i.e. blacken the 6053 // grey objects) 6054 bool res = _work_queue->push(obj); 6055 assert(res, "Low water mark should be less than capacity?"); 6056 trim_queue(_low_water_mark); 6057 } // Else, another thread claimed the object 6058 } 6059 } 6060 } 6061 6062 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6063 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6064 6065 // This closure is used to rescan the marked objects on the dirty cards 6066 // in the mod union table and the card table proper. 6067 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6068 oop p, MemRegion mr) { 6069 6070 size_t size = 0; 6071 HeapWord* addr = (HeapWord*)p; 6072 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6073 assert(_span.contains(addr), "we are scanning the CMS generation"); 6074 // check if it's time to yield 6075 if (do_yield_check()) { 6076 // We yielded for some foreground stop-world work, 6077 // and we have been asked to abort this ongoing preclean cycle. 6078 return 0; 6079 } 6080 if (_bitMap->isMarked(addr)) { 6081 // it's marked; is it potentially uninitialized? 6082 if (p->klass_or_null() != NULL) { 6083 // an initialized object; ignore mark word in verification below 6084 // since we are running concurrent with mutators 6085 assert(p->is_oop(true), "should be an oop"); 6086 if (p->is_objArray()) { 6087 // objArrays are precisely marked; restrict scanning 6088 // to dirty cards only. 6089 size = CompactibleFreeListSpace::adjustObjectSize( 6090 p->oop_iterate_size(_scanningClosure, mr)); 6091 } else { 6092 // A non-array may have been imprecisely marked; we need 6093 // to scan object in its entirety. 6094 size = CompactibleFreeListSpace::adjustObjectSize( 6095 p->oop_iterate_size(_scanningClosure)); 6096 } 6097 #ifdef ASSERT 6098 size_t direct_size = 6099 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6100 assert(size == direct_size, "Inconsistency in size"); 6101 assert(size >= 3, "Necessary for Printezis marks to work"); 6102 if (!_bitMap->isMarked(addr+1)) { 6103 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6104 } else { 6105 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6106 assert(_bitMap->isMarked(addr+size-1), 6107 "inconsistent Printezis mark"); 6108 } 6109 #endif // ASSERT 6110 } else { 6111 // An uninitialized object. 6112 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6113 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6114 size = pointer_delta(nextOneAddr + 1, addr); 6115 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6116 "alignment problem"); 6117 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6118 // will dirty the card when the klass pointer is installed in the 6119 // object (signaling the completion of initialization). 6120 } 6121 } else { 6122 // Either a not yet marked object or an uninitialized object 6123 if (p->klass_or_null() == NULL) { 6124 // An uninitialized object, skip to the next card, since 6125 // we may not be able to read its P-bits yet. 6126 assert(size == 0, "Initial value"); 6127 } else { 6128 // An object not (yet) reached by marking: we merely need to 6129 // compute its size so as to go look at the next block. 6130 assert(p->is_oop(true), "should be an oop"); 6131 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6132 } 6133 } 6134 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6135 return size; 6136 } 6137 6138 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6139 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6140 "CMS thread should hold CMS token"); 6141 assert_lock_strong(_freelistLock); 6142 assert_lock_strong(_bitMap->lock()); 6143 // relinquish the free_list_lock and bitMaplock() 6144 _bitMap->lock()->unlock(); 6145 _freelistLock->unlock(); 6146 ConcurrentMarkSweepThread::desynchronize(true); 6147 _collector->stopTimer(); 6148 _collector->incrementYields(); 6149 6150 // See the comment in coordinator_yield() 6151 for (unsigned i = 0; i < CMSYieldSleepCount && 6152 ConcurrentMarkSweepThread::should_yield() && 6153 !CMSCollector::foregroundGCIsActive(); ++i) { 6154 os::sleep(Thread::current(), 1, false); 6155 } 6156 6157 ConcurrentMarkSweepThread::synchronize(true); 6158 _freelistLock->lock_without_safepoint_check(); 6159 _bitMap->lock()->lock_without_safepoint_check(); 6160 _collector->startTimer(); 6161 } 6162 6163 6164 ////////////////////////////////////////////////////////////////// 6165 // SurvivorSpacePrecleanClosure 6166 ////////////////////////////////////////////////////////////////// 6167 // This (single-threaded) closure is used to preclean the oops in 6168 // the survivor spaces. 6169 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6170 6171 HeapWord* addr = (HeapWord*)p; 6172 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6173 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6174 assert(p->klass_or_null() != NULL, "object should be initialized"); 6175 // an initialized object; ignore mark word in verification below 6176 // since we are running concurrent with mutators 6177 assert(p->is_oop(true), "should be an oop"); 6178 // Note that we do not yield while we iterate over 6179 // the interior oops of p, pushing the relevant ones 6180 // on our marking stack. 6181 size_t size = p->oop_iterate_size(_scanning_closure); 6182 do_yield_check(); 6183 // Observe that below, we do not abandon the preclean 6184 // phase as soon as we should; rather we empty the 6185 // marking stack before returning. This is to satisfy 6186 // some existing assertions. In general, it may be a 6187 // good idea to abort immediately and complete the marking 6188 // from the grey objects at a later time. 6189 while (!_mark_stack->isEmpty()) { 6190 oop new_oop = _mark_stack->pop(); 6191 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6192 assert(_bit_map->isMarked((HeapWord*)new_oop), 6193 "only grey objects on this stack"); 6194 // iterate over the oops in this oop, marking and pushing 6195 // the ones in CMS heap (i.e. in _span). 6196 new_oop->oop_iterate(_scanning_closure); 6197 // check if it's time to yield 6198 do_yield_check(); 6199 } 6200 unsigned int after_count = 6201 GenCollectedHeap::heap()->total_collections(); 6202 bool abort = (_before_count != after_count) || 6203 _collector->should_abort_preclean(); 6204 return abort ? 0 : size; 6205 } 6206 6207 void SurvivorSpacePrecleanClosure::do_yield_work() { 6208 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6209 "CMS thread should hold CMS token"); 6210 assert_lock_strong(_bit_map->lock()); 6211 // Relinquish the bit map lock 6212 _bit_map->lock()->unlock(); 6213 ConcurrentMarkSweepThread::desynchronize(true); 6214 _collector->stopTimer(); 6215 _collector->incrementYields(); 6216 6217 // See the comment in coordinator_yield() 6218 for (unsigned i = 0; i < CMSYieldSleepCount && 6219 ConcurrentMarkSweepThread::should_yield() && 6220 !CMSCollector::foregroundGCIsActive(); ++i) { 6221 os::sleep(Thread::current(), 1, false); 6222 } 6223 6224 ConcurrentMarkSweepThread::synchronize(true); 6225 _bit_map->lock()->lock_without_safepoint_check(); 6226 _collector->startTimer(); 6227 } 6228 6229 // This closure is used to rescan the marked objects on the dirty cards 6230 // in the mod union table and the card table proper. In the parallel 6231 // case, although the bitMap is shared, we do a single read so the 6232 // isMarked() query is "safe". 6233 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6234 // Ignore mark word because we are running concurrent with mutators 6235 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6236 HeapWord* addr = (HeapWord*)p; 6237 assert(_span.contains(addr), "we are scanning the CMS generation"); 6238 bool is_obj_array = false; 6239 #ifdef ASSERT 6240 if (!_parallel) { 6241 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6242 assert(_collector->overflow_list_is_empty(), 6243 "overflow list should be empty"); 6244 6245 } 6246 #endif // ASSERT 6247 if (_bit_map->isMarked(addr)) { 6248 // Obj arrays are precisely marked, non-arrays are not; 6249 // so we scan objArrays precisely and non-arrays in their 6250 // entirety. 6251 if (p->is_objArray()) { 6252 is_obj_array = true; 6253 if (_parallel) { 6254 p->oop_iterate(_par_scan_closure, mr); 6255 } else { 6256 p->oop_iterate(_scan_closure, mr); 6257 } 6258 } else { 6259 if (_parallel) { 6260 p->oop_iterate(_par_scan_closure); 6261 } else { 6262 p->oop_iterate(_scan_closure); 6263 } 6264 } 6265 } 6266 #ifdef ASSERT 6267 if (!_parallel) { 6268 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6269 assert(_collector->overflow_list_is_empty(), 6270 "overflow list should be empty"); 6271 6272 } 6273 #endif // ASSERT 6274 return is_obj_array; 6275 } 6276 6277 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6278 MemRegion span, 6279 CMSBitMap* bitMap, CMSMarkStack* markStack, 6280 bool should_yield, bool verifying): 6281 _collector(collector), 6282 _span(span), 6283 _bitMap(bitMap), 6284 _mut(&collector->_modUnionTable), 6285 _markStack(markStack), 6286 _yield(should_yield), 6287 _skipBits(0) 6288 { 6289 assert(_markStack->isEmpty(), "stack should be empty"); 6290 _finger = _bitMap->startWord(); 6291 _threshold = _finger; 6292 assert(_collector->_restart_addr == NULL, "Sanity check"); 6293 assert(_span.contains(_finger), "Out of bounds _finger?"); 6294 DEBUG_ONLY(_verifying = verifying;) 6295 } 6296 6297 void MarkFromRootsClosure::reset(HeapWord* addr) { 6298 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6299 assert(_span.contains(addr), "Out of bounds _finger?"); 6300 _finger = addr; 6301 _threshold = (HeapWord*)round_to( 6302 (intptr_t)_finger, CardTableModRefBS::card_size); 6303 } 6304 6305 // Should revisit to see if this should be restructured for 6306 // greater efficiency. 6307 bool MarkFromRootsClosure::do_bit(size_t offset) { 6308 if (_skipBits > 0) { 6309 _skipBits--; 6310 return true; 6311 } 6312 // convert offset into a HeapWord* 6313 HeapWord* addr = _bitMap->startWord() + offset; 6314 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6315 "address out of range"); 6316 assert(_bitMap->isMarked(addr), "tautology"); 6317 if (_bitMap->isMarked(addr+1)) { 6318 // this is an allocated but not yet initialized object 6319 assert(_skipBits == 0, "tautology"); 6320 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6321 oop p = oop(addr); 6322 if (p->klass_or_null() == NULL) { 6323 DEBUG_ONLY(if (!_verifying) {) 6324 // We re-dirty the cards on which this object lies and increase 6325 // the _threshold so that we'll come back to scan this object 6326 // during the preclean or remark phase. (CMSCleanOnEnter) 6327 if (CMSCleanOnEnter) { 6328 size_t sz = _collector->block_size_using_printezis_bits(addr); 6329 HeapWord* end_card_addr = (HeapWord*)round_to( 6330 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6331 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6332 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6333 // Bump _threshold to end_card_addr; note that 6334 // _threshold cannot possibly exceed end_card_addr, anyhow. 6335 // This prevents future clearing of the card as the scan proceeds 6336 // to the right. 6337 assert(_threshold <= end_card_addr, 6338 "Because we are just scanning into this object"); 6339 if (_threshold < end_card_addr) { 6340 _threshold = end_card_addr; 6341 } 6342 if (p->klass_or_null() != NULL) { 6343 // Redirty the range of cards... 6344 _mut->mark_range(redirty_range); 6345 } // ...else the setting of klass will dirty the card anyway. 6346 } 6347 DEBUG_ONLY(}) 6348 return true; 6349 } 6350 } 6351 scanOopsInOop(addr); 6352 return true; 6353 } 6354 6355 // We take a break if we've been at this for a while, 6356 // so as to avoid monopolizing the locks involved. 6357 void MarkFromRootsClosure::do_yield_work() { 6358 // First give up the locks, then yield, then re-lock 6359 // We should probably use a constructor/destructor idiom to 6360 // do this unlock/lock or modify the MutexUnlocker class to 6361 // serve our purpose. XXX 6362 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6363 "CMS thread should hold CMS token"); 6364 assert_lock_strong(_bitMap->lock()); 6365 _bitMap->lock()->unlock(); 6366 ConcurrentMarkSweepThread::desynchronize(true); 6367 _collector->stopTimer(); 6368 _collector->incrementYields(); 6369 6370 // See the comment in coordinator_yield() 6371 for (unsigned i = 0; i < CMSYieldSleepCount && 6372 ConcurrentMarkSweepThread::should_yield() && 6373 !CMSCollector::foregroundGCIsActive(); ++i) { 6374 os::sleep(Thread::current(), 1, false); 6375 } 6376 6377 ConcurrentMarkSweepThread::synchronize(true); 6378 _bitMap->lock()->lock_without_safepoint_check(); 6379 _collector->startTimer(); 6380 } 6381 6382 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6383 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6384 assert(_markStack->isEmpty(), 6385 "should drain stack to limit stack usage"); 6386 // convert ptr to an oop preparatory to scanning 6387 oop obj = oop(ptr); 6388 // Ignore mark word in verification below, since we 6389 // may be running concurrent with mutators. 6390 assert(obj->is_oop(true), "should be an oop"); 6391 assert(_finger <= ptr, "_finger runneth ahead"); 6392 // advance the finger to right end of this object 6393 _finger = ptr + obj->size(); 6394 assert(_finger > ptr, "we just incremented it above"); 6395 // On large heaps, it may take us some time to get through 6396 // the marking phase. During 6397 // this time it's possible that a lot of mutations have 6398 // accumulated in the card table and the mod union table -- 6399 // these mutation records are redundant until we have 6400 // actually traced into the corresponding card. 6401 // Here, we check whether advancing the finger would make 6402 // us cross into a new card, and if so clear corresponding 6403 // cards in the MUT (preclean them in the card-table in the 6404 // future). 6405 6406 DEBUG_ONLY(if (!_verifying) {) 6407 // The clean-on-enter optimization is disabled by default, 6408 // until we fix 6178663. 6409 if (CMSCleanOnEnter && (_finger > _threshold)) { 6410 // [_threshold, _finger) represents the interval 6411 // of cards to be cleared in MUT (or precleaned in card table). 6412 // The set of cards to be cleared is all those that overlap 6413 // with the interval [_threshold, _finger); note that 6414 // _threshold is always kept card-aligned but _finger isn't 6415 // always card-aligned. 6416 HeapWord* old_threshold = _threshold; 6417 assert(old_threshold == (HeapWord*)round_to( 6418 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6419 "_threshold should always be card-aligned"); 6420 _threshold = (HeapWord*)round_to( 6421 (intptr_t)_finger, CardTableModRefBS::card_size); 6422 MemRegion mr(old_threshold, _threshold); 6423 assert(!mr.is_empty(), "Control point invariant"); 6424 assert(_span.contains(mr), "Should clear within span"); 6425 _mut->clear_range(mr); 6426 } 6427 DEBUG_ONLY(}) 6428 // Note: the finger doesn't advance while we drain 6429 // the stack below. 6430 PushOrMarkClosure pushOrMarkClosure(_collector, 6431 _span, _bitMap, _markStack, 6432 _finger, this); 6433 bool res = _markStack->push(obj); 6434 assert(res, "Empty non-zero size stack should have space for single push"); 6435 while (!_markStack->isEmpty()) { 6436 oop new_oop = _markStack->pop(); 6437 // Skip verifying header mark word below because we are 6438 // running concurrent with mutators. 6439 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6440 // now scan this oop's oops 6441 new_oop->oop_iterate(&pushOrMarkClosure); 6442 do_yield_check(); 6443 } 6444 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6445 } 6446 6447 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6448 CMSCollector* collector, MemRegion span, 6449 CMSBitMap* bit_map, 6450 OopTaskQueue* work_queue, 6451 CMSMarkStack* overflow_stack): 6452 _collector(collector), 6453 _whole_span(collector->_span), 6454 _span(span), 6455 _bit_map(bit_map), 6456 _mut(&collector->_modUnionTable), 6457 _work_queue(work_queue), 6458 _overflow_stack(overflow_stack), 6459 _skip_bits(0), 6460 _task(task) 6461 { 6462 assert(_work_queue->size() == 0, "work_queue should be empty"); 6463 _finger = span.start(); 6464 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6465 assert(_span.contains(_finger), "Out of bounds _finger?"); 6466 } 6467 6468 // Should revisit to see if this should be restructured for 6469 // greater efficiency. 6470 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6471 if (_skip_bits > 0) { 6472 _skip_bits--; 6473 return true; 6474 } 6475 // convert offset into a HeapWord* 6476 HeapWord* addr = _bit_map->startWord() + offset; 6477 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6478 "address out of range"); 6479 assert(_bit_map->isMarked(addr), "tautology"); 6480 if (_bit_map->isMarked(addr+1)) { 6481 // this is an allocated object that might not yet be initialized 6482 assert(_skip_bits == 0, "tautology"); 6483 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6484 oop p = oop(addr); 6485 if (p->klass_or_null() == NULL) { 6486 // in the case of Clean-on-Enter optimization, redirty card 6487 // and avoid clearing card by increasing the threshold. 6488 return true; 6489 } 6490 } 6491 scan_oops_in_oop(addr); 6492 return true; 6493 } 6494 6495 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6496 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6497 // Should we assert that our work queue is empty or 6498 // below some drain limit? 6499 assert(_work_queue->size() == 0, 6500 "should drain stack to limit stack usage"); 6501 // convert ptr to an oop preparatory to scanning 6502 oop obj = oop(ptr); 6503 // Ignore mark word in verification below, since we 6504 // may be running concurrent with mutators. 6505 assert(obj->is_oop(true), "should be an oop"); 6506 assert(_finger <= ptr, "_finger runneth ahead"); 6507 // advance the finger to right end of this object 6508 _finger = ptr + obj->size(); 6509 assert(_finger > ptr, "we just incremented it above"); 6510 // On large heaps, it may take us some time to get through 6511 // the marking phase. During 6512 // this time it's possible that a lot of mutations have 6513 // accumulated in the card table and the mod union table -- 6514 // these mutation records are redundant until we have 6515 // actually traced into the corresponding card. 6516 // Here, we check whether advancing the finger would make 6517 // us cross into a new card, and if so clear corresponding 6518 // cards in the MUT (preclean them in the card-table in the 6519 // future). 6520 6521 // The clean-on-enter optimization is disabled by default, 6522 // until we fix 6178663. 6523 if (CMSCleanOnEnter && (_finger > _threshold)) { 6524 // [_threshold, _finger) represents the interval 6525 // of cards to be cleared in MUT (or precleaned in card table). 6526 // The set of cards to be cleared is all those that overlap 6527 // with the interval [_threshold, _finger); note that 6528 // _threshold is always kept card-aligned but _finger isn't 6529 // always card-aligned. 6530 HeapWord* old_threshold = _threshold; 6531 assert(old_threshold == (HeapWord*)round_to( 6532 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6533 "_threshold should always be card-aligned"); 6534 _threshold = (HeapWord*)round_to( 6535 (intptr_t)_finger, CardTableModRefBS::card_size); 6536 MemRegion mr(old_threshold, _threshold); 6537 assert(!mr.is_empty(), "Control point invariant"); 6538 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6539 _mut->clear_range(mr); 6540 } 6541 6542 // Note: the local finger doesn't advance while we drain 6543 // the stack below, but the global finger sure can and will. 6544 HeapWord** gfa = _task->global_finger_addr(); 6545 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6546 _span, _bit_map, 6547 _work_queue, 6548 _overflow_stack, 6549 _finger, 6550 gfa, this); 6551 bool res = _work_queue->push(obj); // overflow could occur here 6552 assert(res, "Will hold once we use workqueues"); 6553 while (true) { 6554 oop new_oop; 6555 if (!_work_queue->pop_local(new_oop)) { 6556 // We emptied our work_queue; check if there's stuff that can 6557 // be gotten from the overflow stack. 6558 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6559 _overflow_stack, _work_queue)) { 6560 do_yield_check(); 6561 continue; 6562 } else { // done 6563 break; 6564 } 6565 } 6566 // Skip verifying header mark word below because we are 6567 // running concurrent with mutators. 6568 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6569 // now scan this oop's oops 6570 new_oop->oop_iterate(&pushOrMarkClosure); 6571 do_yield_check(); 6572 } 6573 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6574 } 6575 6576 // Yield in response to a request from VM Thread or 6577 // from mutators. 6578 void ParMarkFromRootsClosure::do_yield_work() { 6579 assert(_task != NULL, "sanity"); 6580 _task->yield(); 6581 } 6582 6583 // A variant of the above used for verifying CMS marking work. 6584 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6585 MemRegion span, 6586 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6587 CMSMarkStack* mark_stack): 6588 _collector(collector), 6589 _span(span), 6590 _verification_bm(verification_bm), 6591 _cms_bm(cms_bm), 6592 _mark_stack(mark_stack), 6593 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6594 mark_stack) 6595 { 6596 assert(_mark_stack->isEmpty(), "stack should be empty"); 6597 _finger = _verification_bm->startWord(); 6598 assert(_collector->_restart_addr == NULL, "Sanity check"); 6599 assert(_span.contains(_finger), "Out of bounds _finger?"); 6600 } 6601 6602 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6603 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6604 assert(_span.contains(addr), "Out of bounds _finger?"); 6605 _finger = addr; 6606 } 6607 6608 // Should revisit to see if this should be restructured for 6609 // greater efficiency. 6610 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6611 // convert offset into a HeapWord* 6612 HeapWord* addr = _verification_bm->startWord() + offset; 6613 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6614 "address out of range"); 6615 assert(_verification_bm->isMarked(addr), "tautology"); 6616 assert(_cms_bm->isMarked(addr), "tautology"); 6617 6618 assert(_mark_stack->isEmpty(), 6619 "should drain stack to limit stack usage"); 6620 // convert addr to an oop preparatory to scanning 6621 oop obj = oop(addr); 6622 assert(obj->is_oop(), "should be an oop"); 6623 assert(_finger <= addr, "_finger runneth ahead"); 6624 // advance the finger to right end of this object 6625 _finger = addr + obj->size(); 6626 assert(_finger > addr, "we just incremented it above"); 6627 // Note: the finger doesn't advance while we drain 6628 // the stack below. 6629 bool res = _mark_stack->push(obj); 6630 assert(res, "Empty non-zero size stack should have space for single push"); 6631 while (!_mark_stack->isEmpty()) { 6632 oop new_oop = _mark_stack->pop(); 6633 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6634 // now scan this oop's oops 6635 new_oop->oop_iterate(&_pam_verify_closure); 6636 } 6637 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6638 return true; 6639 } 6640 6641 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6642 CMSCollector* collector, MemRegion span, 6643 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6644 CMSMarkStack* mark_stack): 6645 MetadataAwareOopClosure(collector->ref_processor()), 6646 _collector(collector), 6647 _span(span), 6648 _verification_bm(verification_bm), 6649 _cms_bm(cms_bm), 6650 _mark_stack(mark_stack) 6651 { } 6652 6653 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6654 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6655 6656 // Upon stack overflow, we discard (part of) the stack, 6657 // remembering the least address amongst those discarded 6658 // in CMSCollector's _restart_address. 6659 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6660 // Remember the least grey address discarded 6661 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6662 _collector->lower_restart_addr(ra); 6663 _mark_stack->reset(); // discard stack contents 6664 _mark_stack->expand(); // expand the stack if possible 6665 } 6666 6667 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6668 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6669 HeapWord* addr = (HeapWord*)obj; 6670 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6671 // Oop lies in _span and isn't yet grey or black 6672 _verification_bm->mark(addr); // now grey 6673 if (!_cms_bm->isMarked(addr)) { 6674 Log(gc, verify) log; 6675 ResourceMark rm; 6676 oop(addr)->print_on(log.error_stream()); 6677 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6678 fatal("... aborting"); 6679 } 6680 6681 if (!_mark_stack->push(obj)) { // stack overflow 6682 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6683 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6684 handle_stack_overflow(addr); 6685 } 6686 // anything including and to the right of _finger 6687 // will be scanned as we iterate over the remainder of the 6688 // bit map 6689 } 6690 } 6691 6692 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6693 MemRegion span, 6694 CMSBitMap* bitMap, CMSMarkStack* markStack, 6695 HeapWord* finger, MarkFromRootsClosure* parent) : 6696 MetadataAwareOopClosure(collector->ref_processor()), 6697 _collector(collector), 6698 _span(span), 6699 _bitMap(bitMap), 6700 _markStack(markStack), 6701 _finger(finger), 6702 _parent(parent) 6703 { } 6704 6705 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6706 MemRegion span, 6707 CMSBitMap* bit_map, 6708 OopTaskQueue* work_queue, 6709 CMSMarkStack* overflow_stack, 6710 HeapWord* finger, 6711 HeapWord** global_finger_addr, 6712 ParMarkFromRootsClosure* parent) : 6713 MetadataAwareOopClosure(collector->ref_processor()), 6714 _collector(collector), 6715 _whole_span(collector->_span), 6716 _span(span), 6717 _bit_map(bit_map), 6718 _work_queue(work_queue), 6719 _overflow_stack(overflow_stack), 6720 _finger(finger), 6721 _global_finger_addr(global_finger_addr), 6722 _parent(parent) 6723 { } 6724 6725 // Assumes thread-safe access by callers, who are 6726 // responsible for mutual exclusion. 6727 void CMSCollector::lower_restart_addr(HeapWord* low) { 6728 assert(_span.contains(low), "Out of bounds addr"); 6729 if (_restart_addr == NULL) { 6730 _restart_addr = low; 6731 } else { 6732 _restart_addr = MIN2(_restart_addr, low); 6733 } 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 PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6740 // Remember the least grey address discarded 6741 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6742 _collector->lower_restart_addr(ra); 6743 _markStack->reset(); // discard stack contents 6744 _markStack->expand(); // expand the stack if possible 6745 } 6746 6747 // Upon stack overflow, we discard (part of) the stack, 6748 // remembering the least address amongst those discarded 6749 // in CMSCollector's _restart_address. 6750 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6751 // We need to do this under a mutex to prevent other 6752 // workers from interfering with the work done below. 6753 MutexLockerEx ml(_overflow_stack->par_lock(), 6754 Mutex::_no_safepoint_check_flag); 6755 // Remember the least grey address discarded 6756 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6757 _collector->lower_restart_addr(ra); 6758 _overflow_stack->reset(); // discard stack contents 6759 _overflow_stack->expand(); // expand the stack if possible 6760 } 6761 6762 void PushOrMarkClosure::do_oop(oop obj) { 6763 // Ignore mark word because we are running concurrent with mutators. 6764 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6765 HeapWord* addr = (HeapWord*)obj; 6766 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6767 // Oop lies in _span and isn't yet grey or black 6768 _bitMap->mark(addr); // now grey 6769 if (addr < _finger) { 6770 // the bit map iteration has already either passed, or 6771 // sampled, this bit in the bit map; we'll need to 6772 // use the marking stack to scan this oop's oops. 6773 bool simulate_overflow = false; 6774 NOT_PRODUCT( 6775 if (CMSMarkStackOverflowALot && 6776 _collector->simulate_overflow()) { 6777 // simulate a stack overflow 6778 simulate_overflow = true; 6779 } 6780 ) 6781 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6782 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6783 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6784 handle_stack_overflow(addr); 6785 } 6786 } 6787 // anything including and to the right of _finger 6788 // will be scanned as we iterate over the remainder of the 6789 // bit map 6790 do_yield_check(); 6791 } 6792 } 6793 6794 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6795 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6796 6797 void ParPushOrMarkClosure::do_oop(oop obj) { 6798 // Ignore mark word because we are running concurrent with mutators. 6799 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6800 HeapWord* addr = (HeapWord*)obj; 6801 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6802 // Oop lies in _span and isn't yet grey or black 6803 // We read the global_finger (volatile read) strictly after marking oop 6804 bool res = _bit_map->par_mark(addr); // now grey 6805 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6806 // Should we push this marked oop on our stack? 6807 // -- if someone else marked it, nothing to do 6808 // -- if target oop is above global finger nothing to do 6809 // -- if target oop is in chunk and above local finger 6810 // then nothing to do 6811 // -- else push on work queue 6812 if ( !res // someone else marked it, they will deal with it 6813 || (addr >= *gfa) // will be scanned in a later task 6814 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6815 return; 6816 } 6817 // the bit map iteration has already either passed, or 6818 // sampled, this bit in the bit map; we'll need to 6819 // use the marking stack to scan this oop's oops. 6820 bool simulate_overflow = false; 6821 NOT_PRODUCT( 6822 if (CMSMarkStackOverflowALot && 6823 _collector->simulate_overflow()) { 6824 // simulate a stack overflow 6825 simulate_overflow = true; 6826 } 6827 ) 6828 if (simulate_overflow || 6829 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6830 // stack overflow 6831 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6832 // We cannot assert that the overflow stack is full because 6833 // it may have been emptied since. 6834 assert(simulate_overflow || 6835 _work_queue->size() == _work_queue->max_elems(), 6836 "Else push should have succeeded"); 6837 handle_stack_overflow(addr); 6838 } 6839 do_yield_check(); 6840 } 6841 } 6842 6843 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6844 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6845 6846 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6847 MemRegion span, 6848 ReferenceProcessor* rp, 6849 CMSBitMap* bit_map, 6850 CMSBitMap* mod_union_table, 6851 CMSMarkStack* mark_stack, 6852 bool concurrent_precleaning): 6853 MetadataAwareOopClosure(rp), 6854 _collector(collector), 6855 _span(span), 6856 _bit_map(bit_map), 6857 _mod_union_table(mod_union_table), 6858 _mark_stack(mark_stack), 6859 _concurrent_precleaning(concurrent_precleaning) 6860 { 6861 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6862 } 6863 6864 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6865 // the non-parallel version (the parallel version appears further below.) 6866 void PushAndMarkClosure::do_oop(oop obj) { 6867 // Ignore mark word verification. If during concurrent precleaning, 6868 // the object monitor may be locked. If during the checkpoint 6869 // phases, the object may already have been reached by a different 6870 // path and may be at the end of the global overflow list (so 6871 // the mark word may be NULL). 6872 assert(obj->is_oop_or_null(true /* ignore mark word */), 6873 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6874 HeapWord* addr = (HeapWord*)obj; 6875 // Check if oop points into the CMS generation 6876 // and is not marked 6877 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6878 // a white object ... 6879 _bit_map->mark(addr); // ... now grey 6880 // push on the marking stack (grey set) 6881 bool simulate_overflow = false; 6882 NOT_PRODUCT( 6883 if (CMSMarkStackOverflowALot && 6884 _collector->simulate_overflow()) { 6885 // simulate a stack overflow 6886 simulate_overflow = true; 6887 } 6888 ) 6889 if (simulate_overflow || !_mark_stack->push(obj)) { 6890 if (_concurrent_precleaning) { 6891 // During precleaning we can just dirty the appropriate card(s) 6892 // in the mod union table, thus ensuring that the object remains 6893 // in the grey set and continue. In the case of object arrays 6894 // we need to dirty all of the cards that the object spans, 6895 // since the rescan of object arrays will be limited to the 6896 // dirty cards. 6897 // Note that no one can be interfering with us in this action 6898 // of dirtying the mod union table, so no locking or atomics 6899 // are required. 6900 if (obj->is_objArray()) { 6901 size_t sz = obj->size(); 6902 HeapWord* end_card_addr = (HeapWord*)round_to( 6903 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6904 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6905 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6906 _mod_union_table->mark_range(redirty_range); 6907 } else { 6908 _mod_union_table->mark(addr); 6909 } 6910 _collector->_ser_pmc_preclean_ovflw++; 6911 } else { 6912 // During the remark phase, we need to remember this oop 6913 // in the overflow list. 6914 _collector->push_on_overflow_list(obj); 6915 _collector->_ser_pmc_remark_ovflw++; 6916 } 6917 } 6918 } 6919 } 6920 6921 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6922 MemRegion span, 6923 ReferenceProcessor* rp, 6924 CMSBitMap* bit_map, 6925 OopTaskQueue* work_queue): 6926 MetadataAwareOopClosure(rp), 6927 _collector(collector), 6928 _span(span), 6929 _bit_map(bit_map), 6930 _work_queue(work_queue) 6931 { 6932 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6933 } 6934 6935 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6936 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6937 6938 // Grey object rescan during second checkpoint phase -- 6939 // the parallel version. 6940 void ParPushAndMarkClosure::do_oop(oop obj) { 6941 // In the assert below, we ignore the mark word because 6942 // this oop may point to an already visited object that is 6943 // on the overflow stack (in which case the mark word has 6944 // been hijacked for chaining into the overflow stack -- 6945 // if this is the last object in the overflow stack then 6946 // its mark word will be NULL). Because this object may 6947 // have been subsequently popped off the global overflow 6948 // stack, and the mark word possibly restored to the prototypical 6949 // value, by the time we get to examined this failing assert in 6950 // the debugger, is_oop_or_null(false) may subsequently start 6951 // to hold. 6952 assert(obj->is_oop_or_null(true), 6953 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6954 HeapWord* addr = (HeapWord*)obj; 6955 // Check if oop points into the CMS generation 6956 // and is not marked 6957 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6958 // a white object ... 6959 // If we manage to "claim" the object, by being the 6960 // first thread to mark it, then we push it on our 6961 // marking stack 6962 if (_bit_map->par_mark(addr)) { // ... now grey 6963 // push on work queue (grey set) 6964 bool simulate_overflow = false; 6965 NOT_PRODUCT( 6966 if (CMSMarkStackOverflowALot && 6967 _collector->par_simulate_overflow()) { 6968 // simulate a stack overflow 6969 simulate_overflow = true; 6970 } 6971 ) 6972 if (simulate_overflow || !_work_queue->push(obj)) { 6973 _collector->par_push_on_overflow_list(obj); 6974 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6975 } 6976 } // Else, some other thread got there first 6977 } 6978 } 6979 6980 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6981 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6982 6983 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6984 Mutex* bml = _collector->bitMapLock(); 6985 assert_lock_strong(bml); 6986 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6987 "CMS thread should hold CMS token"); 6988 6989 bml->unlock(); 6990 ConcurrentMarkSweepThread::desynchronize(true); 6991 6992 _collector->stopTimer(); 6993 _collector->incrementYields(); 6994 6995 // See the comment in coordinator_yield() 6996 for (unsigned i = 0; i < CMSYieldSleepCount && 6997 ConcurrentMarkSweepThread::should_yield() && 6998 !CMSCollector::foregroundGCIsActive(); ++i) { 6999 os::sleep(Thread::current(), 1, false); 7000 } 7001 7002 ConcurrentMarkSweepThread::synchronize(true); 7003 bml->lock(); 7004 7005 _collector->startTimer(); 7006 } 7007 7008 bool CMSPrecleanRefsYieldClosure::should_return() { 7009 if (ConcurrentMarkSweepThread::should_yield()) { 7010 do_yield_work(); 7011 } 7012 return _collector->foregroundGCIsActive(); 7013 } 7014 7015 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7016 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7017 "mr should be aligned to start at a card boundary"); 7018 // We'd like to assert: 7019 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7020 // "mr should be a range of cards"); 7021 // However, that would be too strong in one case -- the last 7022 // partition ends at _unallocated_block which, in general, can be 7023 // an arbitrary boundary, not necessarily card aligned. 7024 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7025 _space->object_iterate_mem(mr, &_scan_cl); 7026 } 7027 7028 SweepClosure::SweepClosure(CMSCollector* collector, 7029 ConcurrentMarkSweepGeneration* g, 7030 CMSBitMap* bitMap, bool should_yield) : 7031 _collector(collector), 7032 _g(g), 7033 _sp(g->cmsSpace()), 7034 _limit(_sp->sweep_limit()), 7035 _freelistLock(_sp->freelistLock()), 7036 _bitMap(bitMap), 7037 _yield(should_yield), 7038 _inFreeRange(false), // No free range at beginning of sweep 7039 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7040 _lastFreeRangeCoalesced(false), 7041 _freeFinger(g->used_region().start()) 7042 { 7043 NOT_PRODUCT( 7044 _numObjectsFreed = 0; 7045 _numWordsFreed = 0; 7046 _numObjectsLive = 0; 7047 _numWordsLive = 0; 7048 _numObjectsAlreadyFree = 0; 7049 _numWordsAlreadyFree = 0; 7050 _last_fc = NULL; 7051 7052 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7053 _sp->dictionary()->initialize_dict_returned_bytes(); 7054 ) 7055 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7056 "sweep _limit out of bounds"); 7057 log_develop_trace(gc, sweep)("===================="); 7058 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7059 } 7060 7061 void SweepClosure::print_on(outputStream* st) const { 7062 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7063 p2i(_sp->bottom()), p2i(_sp->end())); 7064 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7065 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7066 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7067 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7068 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7069 } 7070 7071 #ifndef PRODUCT 7072 // Assertion checking only: no useful work in product mode -- 7073 // however, if any of the flags below become product flags, 7074 // you may need to review this code to see if it needs to be 7075 // enabled in product mode. 7076 SweepClosure::~SweepClosure() { 7077 assert_lock_strong(_freelistLock); 7078 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7079 "sweep _limit out of bounds"); 7080 if (inFreeRange()) { 7081 Log(gc, sweep) log; 7082 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7083 ResourceMark rm; 7084 print_on(log.error_stream()); 7085 ShouldNotReachHere(); 7086 } 7087 7088 if (log_is_enabled(Debug, gc, sweep)) { 7089 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7090 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7091 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7092 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7093 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7094 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7095 } 7096 7097 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7098 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7099 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7100 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7101 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7102 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7103 } 7104 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7105 log_develop_trace(gc, sweep)("================"); 7106 } 7107 #endif // PRODUCT 7108 7109 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7110 bool freeRangeInFreeLists) { 7111 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7112 p2i(freeFinger), freeRangeInFreeLists); 7113 assert(!inFreeRange(), "Trampling existing free range"); 7114 set_inFreeRange(true); 7115 set_lastFreeRangeCoalesced(false); 7116 7117 set_freeFinger(freeFinger); 7118 set_freeRangeInFreeLists(freeRangeInFreeLists); 7119 if (CMSTestInFreeList) { 7120 if (freeRangeInFreeLists) { 7121 FreeChunk* fc = (FreeChunk*) freeFinger; 7122 assert(fc->is_free(), "A chunk on the free list should be free."); 7123 assert(fc->size() > 0, "Free range should have a size"); 7124 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7125 } 7126 } 7127 } 7128 7129 // Note that the sweeper runs concurrently with mutators. Thus, 7130 // it is possible for direct allocation in this generation to happen 7131 // in the middle of the sweep. Note that the sweeper also coalesces 7132 // contiguous free blocks. Thus, unless the sweeper and the allocator 7133 // synchronize appropriately freshly allocated blocks may get swept up. 7134 // This is accomplished by the sweeper locking the free lists while 7135 // it is sweeping. Thus blocks that are determined to be free are 7136 // indeed free. There is however one additional complication: 7137 // blocks that have been allocated since the final checkpoint and 7138 // mark, will not have been marked and so would be treated as 7139 // unreachable and swept up. To prevent this, the allocator marks 7140 // the bit map when allocating during the sweep phase. This leads, 7141 // however, to a further complication -- objects may have been allocated 7142 // but not yet initialized -- in the sense that the header isn't yet 7143 // installed. The sweeper can not then determine the size of the block 7144 // in order to skip over it. To deal with this case, we use a technique 7145 // (due to Printezis) to encode such uninitialized block sizes in the 7146 // bit map. Since the bit map uses a bit per every HeapWord, but the 7147 // CMS generation has a minimum object size of 3 HeapWords, it follows 7148 // that "normal marks" won't be adjacent in the bit map (there will 7149 // always be at least two 0 bits between successive 1 bits). We make use 7150 // of these "unused" bits to represent uninitialized blocks -- the bit 7151 // corresponding to the start of the uninitialized object and the next 7152 // bit are both set. Finally, a 1 bit marks the end of the object that 7153 // started with the two consecutive 1 bits to indicate its potentially 7154 // uninitialized state. 7155 7156 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7157 FreeChunk* fc = (FreeChunk*)addr; 7158 size_t res; 7159 7160 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7161 // than "addr == _limit" because although _limit was a block boundary when 7162 // we started the sweep, it may no longer be one because heap expansion 7163 // may have caused us to coalesce the block ending at the address _limit 7164 // with a newly expanded chunk (this happens when _limit was set to the 7165 // previous _end of the space), so we may have stepped past _limit: 7166 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7167 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7168 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7169 "sweep _limit out of bounds"); 7170 assert(addr < _sp->end(), "addr out of bounds"); 7171 // Flush any free range we might be holding as a single 7172 // coalesced chunk to the appropriate free list. 7173 if (inFreeRange()) { 7174 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7175 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7176 flush_cur_free_chunk(freeFinger(), 7177 pointer_delta(addr, freeFinger())); 7178 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7179 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7180 lastFreeRangeCoalesced() ? 1 : 0); 7181 } 7182 7183 // help the iterator loop finish 7184 return pointer_delta(_sp->end(), addr); 7185 } 7186 7187 assert(addr < _limit, "sweep invariant"); 7188 // check if we should yield 7189 do_yield_check(addr); 7190 if (fc->is_free()) { 7191 // Chunk that is already free 7192 res = fc->size(); 7193 do_already_free_chunk(fc); 7194 debug_only(_sp->verifyFreeLists()); 7195 // If we flush the chunk at hand in lookahead_and_flush() 7196 // and it's coalesced with a preceding chunk, then the 7197 // process of "mangling" the payload of the coalesced block 7198 // will cause erasure of the size information from the 7199 // (erstwhile) header of all the coalesced blocks but the 7200 // first, so the first disjunct in the assert will not hold 7201 // in that specific case (in which case the second disjunct 7202 // will hold). 7203 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7204 "Otherwise the size info doesn't change at this step"); 7205 NOT_PRODUCT( 7206 _numObjectsAlreadyFree++; 7207 _numWordsAlreadyFree += res; 7208 ) 7209 NOT_PRODUCT(_last_fc = fc;) 7210 } else if (!_bitMap->isMarked(addr)) { 7211 // Chunk is fresh garbage 7212 res = do_garbage_chunk(fc); 7213 debug_only(_sp->verifyFreeLists()); 7214 NOT_PRODUCT( 7215 _numObjectsFreed++; 7216 _numWordsFreed += res; 7217 ) 7218 } else { 7219 // Chunk that is alive. 7220 res = do_live_chunk(fc); 7221 debug_only(_sp->verifyFreeLists()); 7222 NOT_PRODUCT( 7223 _numObjectsLive++; 7224 _numWordsLive += res; 7225 ) 7226 } 7227 return res; 7228 } 7229 7230 // For the smart allocation, record following 7231 // split deaths - a free chunk is removed from its free list because 7232 // it is being split into two or more chunks. 7233 // split birth - a free chunk is being added to its free list because 7234 // a larger free chunk has been split and resulted in this free chunk. 7235 // coal death - a free chunk is being removed from its free list because 7236 // it is being coalesced into a large free chunk. 7237 // coal birth - a free chunk is being added to its free list because 7238 // it was created when two or more free chunks where coalesced into 7239 // this free chunk. 7240 // 7241 // These statistics are used to determine the desired number of free 7242 // chunks of a given size. The desired number is chosen to be relative 7243 // to the end of a CMS sweep. The desired number at the end of a sweep 7244 // is the 7245 // count-at-end-of-previous-sweep (an amount that was enough) 7246 // - count-at-beginning-of-current-sweep (the excess) 7247 // + split-births (gains in this size during interval) 7248 // - split-deaths (demands on this size during interval) 7249 // where the interval is from the end of one sweep to the end of the 7250 // next. 7251 // 7252 // When sweeping the sweeper maintains an accumulated chunk which is 7253 // the chunk that is made up of chunks that have been coalesced. That 7254 // will be termed the left-hand chunk. A new chunk of garbage that 7255 // is being considered for coalescing will be referred to as the 7256 // right-hand chunk. 7257 // 7258 // When making a decision on whether to coalesce a right-hand chunk with 7259 // the current left-hand chunk, the current count vs. the desired count 7260 // of the left-hand chunk is considered. Also if the right-hand chunk 7261 // is near the large chunk at the end of the heap (see 7262 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7263 // left-hand chunk is coalesced. 7264 // 7265 // When making a decision about whether to split a chunk, the desired count 7266 // vs. the current count of the candidate to be split is also considered. 7267 // If the candidate is underpopulated (currently fewer chunks than desired) 7268 // a chunk of an overpopulated (currently more chunks than desired) size may 7269 // be chosen. The "hint" associated with a free list, if non-null, points 7270 // to a free list which may be overpopulated. 7271 // 7272 7273 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7274 const size_t size = fc->size(); 7275 // Chunks that cannot be coalesced are not in the 7276 // free lists. 7277 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7278 assert(_sp->verify_chunk_in_free_list(fc), 7279 "free chunk should be in free lists"); 7280 } 7281 // a chunk that is already free, should not have been 7282 // marked in the bit map 7283 HeapWord* const addr = (HeapWord*) fc; 7284 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7285 // Verify that the bit map has no bits marked between 7286 // addr and purported end of this block. 7287 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7288 7289 // Some chunks cannot be coalesced under any circumstances. 7290 // See the definition of cantCoalesce(). 7291 if (!fc->cantCoalesce()) { 7292 // This chunk can potentially be coalesced. 7293 // All the work is done in 7294 do_post_free_or_garbage_chunk(fc, size); 7295 // Note that if the chunk is not coalescable (the else arm 7296 // below), we unconditionally flush, without needing to do 7297 // a "lookahead," as we do below. 7298 if (inFreeRange()) lookahead_and_flush(fc, size); 7299 } else { 7300 // Code path common to both original and adaptive free lists. 7301 7302 // cant coalesce with previous block; this should be treated 7303 // as the end of a free run if any 7304 if (inFreeRange()) { 7305 // we kicked some butt; time to pick up the garbage 7306 assert(freeFinger() < addr, "freeFinger points too high"); 7307 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7308 } 7309 // else, nothing to do, just continue 7310 } 7311 } 7312 7313 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7314 // This is a chunk of garbage. It is not in any free list. 7315 // Add it to a free list or let it possibly be coalesced into 7316 // a larger chunk. 7317 HeapWord* const addr = (HeapWord*) fc; 7318 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7319 7320 // Verify that the bit map has no bits marked between 7321 // addr and purported end of just dead object. 7322 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7323 do_post_free_or_garbage_chunk(fc, size); 7324 7325 assert(_limit >= addr + size, 7326 "A freshly garbage chunk can't possibly straddle over _limit"); 7327 if (inFreeRange()) lookahead_and_flush(fc, size); 7328 return size; 7329 } 7330 7331 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7332 HeapWord* addr = (HeapWord*) fc; 7333 // The sweeper has just found a live object. Return any accumulated 7334 // left hand chunk to the free lists. 7335 if (inFreeRange()) { 7336 assert(freeFinger() < addr, "freeFinger points too high"); 7337 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7338 } 7339 7340 // This object is live: we'd normally expect this to be 7341 // an oop, and like to assert the following: 7342 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7343 // However, as we commented above, this may be an object whose 7344 // header hasn't yet been initialized. 7345 size_t size; 7346 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7347 if (_bitMap->isMarked(addr + 1)) { 7348 // Determine the size from the bit map, rather than trying to 7349 // compute it from the object header. 7350 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7351 size = pointer_delta(nextOneAddr + 1, addr); 7352 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7353 "alignment problem"); 7354 7355 #ifdef ASSERT 7356 if (oop(addr)->klass_or_null() != NULL) { 7357 // Ignore mark word because we are running concurrent with mutators 7358 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7359 assert(size == 7360 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7361 "P-mark and computed size do not agree"); 7362 } 7363 #endif 7364 7365 } else { 7366 // This should be an initialized object that's alive. 7367 assert(oop(addr)->klass_or_null() != NULL, 7368 "Should be an initialized object"); 7369 // Ignore mark word because we are running concurrent with mutators 7370 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7371 // Verify that the bit map has no bits marked between 7372 // addr and purported end of this block. 7373 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7374 assert(size >= 3, "Necessary for Printezis marks to work"); 7375 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7376 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7377 } 7378 return size; 7379 } 7380 7381 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7382 size_t chunkSize) { 7383 // do_post_free_or_garbage_chunk() should only be called in the case 7384 // of the adaptive free list allocator. 7385 const bool fcInFreeLists = fc->is_free(); 7386 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7387 if (CMSTestInFreeList && fcInFreeLists) { 7388 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7389 } 7390 7391 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7392 7393 HeapWord* const fc_addr = (HeapWord*) fc; 7394 7395 bool coalesce = false; 7396 const size_t left = pointer_delta(fc_addr, freeFinger()); 7397 const size_t right = chunkSize; 7398 switch (FLSCoalescePolicy) { 7399 // numeric value forms a coalition aggressiveness metric 7400 case 0: { // never coalesce 7401 coalesce = false; 7402 break; 7403 } 7404 case 1: { // coalesce if left & right chunks on overpopulated lists 7405 coalesce = _sp->coalOverPopulated(left) && 7406 _sp->coalOverPopulated(right); 7407 break; 7408 } 7409 case 2: { // coalesce if left chunk on overpopulated list (default) 7410 coalesce = _sp->coalOverPopulated(left); 7411 break; 7412 } 7413 case 3: { // coalesce if left OR right chunk on overpopulated list 7414 coalesce = _sp->coalOverPopulated(left) || 7415 _sp->coalOverPopulated(right); 7416 break; 7417 } 7418 case 4: { // always coalesce 7419 coalesce = true; 7420 break; 7421 } 7422 default: 7423 ShouldNotReachHere(); 7424 } 7425 7426 // Should the current free range be coalesced? 7427 // If the chunk is in a free range and either we decided to coalesce above 7428 // or the chunk is near the large block at the end of the heap 7429 // (isNearLargestChunk() returns true), then coalesce this chunk. 7430 const bool doCoalesce = inFreeRange() 7431 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7432 if (doCoalesce) { 7433 // Coalesce the current free range on the left with the new 7434 // chunk on the right. If either is on a free list, 7435 // it must be removed from the list and stashed in the closure. 7436 if (freeRangeInFreeLists()) { 7437 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7438 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7439 "Size of free range is inconsistent with chunk size."); 7440 if (CMSTestInFreeList) { 7441 assert(_sp->verify_chunk_in_free_list(ffc), 7442 "Chunk is not in free lists"); 7443 } 7444 _sp->coalDeath(ffc->size()); 7445 _sp->removeFreeChunkFromFreeLists(ffc); 7446 set_freeRangeInFreeLists(false); 7447 } 7448 if (fcInFreeLists) { 7449 _sp->coalDeath(chunkSize); 7450 assert(fc->size() == chunkSize, 7451 "The chunk has the wrong size or is not in the free lists"); 7452 _sp->removeFreeChunkFromFreeLists(fc); 7453 } 7454 set_lastFreeRangeCoalesced(true); 7455 print_free_block_coalesced(fc); 7456 } else { // not in a free range and/or should not coalesce 7457 // Return the current free range and start a new one. 7458 if (inFreeRange()) { 7459 // In a free range but cannot coalesce with the right hand chunk. 7460 // Put the current free range into the free lists. 7461 flush_cur_free_chunk(freeFinger(), 7462 pointer_delta(fc_addr, freeFinger())); 7463 } 7464 // Set up for new free range. Pass along whether the right hand 7465 // chunk is in the free lists. 7466 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7467 } 7468 } 7469 7470 // Lookahead flush: 7471 // If we are tracking a free range, and this is the last chunk that 7472 // we'll look at because its end crosses past _limit, we'll preemptively 7473 // flush it along with any free range we may be holding on to. Note that 7474 // this can be the case only for an already free or freshly garbage 7475 // chunk. If this block is an object, it can never straddle 7476 // over _limit. The "straddling" occurs when _limit is set at 7477 // the previous end of the space when this cycle started, and 7478 // a subsequent heap expansion caused the previously co-terminal 7479 // free block to be coalesced with the newly expanded portion, 7480 // thus rendering _limit a non-block-boundary making it dangerous 7481 // for the sweeper to step over and examine. 7482 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7483 assert(inFreeRange(), "Should only be called if currently in a free range."); 7484 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7485 assert(_sp->used_region().contains(eob - 1), 7486 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7487 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7488 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7489 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7490 if (eob >= _limit) { 7491 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7492 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7493 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7494 "[" PTR_FORMAT "," PTR_FORMAT ")", 7495 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7496 // Return the storage we are tracking back into the free lists. 7497 log_develop_trace(gc, sweep)("Flushing ... "); 7498 assert(freeFinger() < eob, "Error"); 7499 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7500 } 7501 } 7502 7503 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7504 assert(inFreeRange(), "Should only be called if currently in a free range."); 7505 assert(size > 0, 7506 "A zero sized chunk cannot be added to the free lists."); 7507 if (!freeRangeInFreeLists()) { 7508 if (CMSTestInFreeList) { 7509 FreeChunk* fc = (FreeChunk*) chunk; 7510 fc->set_size(size); 7511 assert(!_sp->verify_chunk_in_free_list(fc), 7512 "chunk should not be in free lists yet"); 7513 } 7514 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7515 // A new free range is going to be starting. The current 7516 // free range has not been added to the free lists yet or 7517 // was removed so add it back. 7518 // If the current free range was coalesced, then the death 7519 // of the free range was recorded. Record a birth now. 7520 if (lastFreeRangeCoalesced()) { 7521 _sp->coalBirth(size); 7522 } 7523 _sp->addChunkAndRepairOffsetTable(chunk, size, 7524 lastFreeRangeCoalesced()); 7525 } else { 7526 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7527 } 7528 set_inFreeRange(false); 7529 set_freeRangeInFreeLists(false); 7530 } 7531 7532 // We take a break if we've been at this for a while, 7533 // so as to avoid monopolizing the locks involved. 7534 void SweepClosure::do_yield_work(HeapWord* addr) { 7535 // Return current free chunk being used for coalescing (if any) 7536 // to the appropriate freelist. After yielding, the next 7537 // free block encountered will start a coalescing range of 7538 // free blocks. If the next free block is adjacent to the 7539 // chunk just flushed, they will need to wait for the next 7540 // sweep to be coalesced. 7541 if (inFreeRange()) { 7542 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7543 } 7544 7545 // First give up the locks, then yield, then re-lock. 7546 // We should probably use a constructor/destructor idiom to 7547 // do this unlock/lock or modify the MutexUnlocker class to 7548 // serve our purpose. XXX 7549 assert_lock_strong(_bitMap->lock()); 7550 assert_lock_strong(_freelistLock); 7551 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7552 "CMS thread should hold CMS token"); 7553 _bitMap->lock()->unlock(); 7554 _freelistLock->unlock(); 7555 ConcurrentMarkSweepThread::desynchronize(true); 7556 _collector->stopTimer(); 7557 _collector->incrementYields(); 7558 7559 // See the comment in coordinator_yield() 7560 for (unsigned i = 0; i < CMSYieldSleepCount && 7561 ConcurrentMarkSweepThread::should_yield() && 7562 !CMSCollector::foregroundGCIsActive(); ++i) { 7563 os::sleep(Thread::current(), 1, false); 7564 } 7565 7566 ConcurrentMarkSweepThread::synchronize(true); 7567 _freelistLock->lock(); 7568 _bitMap->lock()->lock_without_safepoint_check(); 7569 _collector->startTimer(); 7570 } 7571 7572 #ifndef PRODUCT 7573 // This is actually very useful in a product build if it can 7574 // be called from the debugger. Compile it into the product 7575 // as needed. 7576 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7577 return debug_cms_space->verify_chunk_in_free_list(fc); 7578 } 7579 #endif 7580 7581 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7582 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7583 p2i(fc), fc->size()); 7584 } 7585 7586 // CMSIsAliveClosure 7587 bool CMSIsAliveClosure::do_object_b(oop obj) { 7588 HeapWord* addr = (HeapWord*)obj; 7589 return addr != NULL && 7590 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7591 } 7592 7593 7594 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7595 MemRegion span, 7596 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7597 bool cpc): 7598 _collector(collector), 7599 _span(span), 7600 _bit_map(bit_map), 7601 _mark_stack(mark_stack), 7602 _concurrent_precleaning(cpc) { 7603 assert(!_span.is_empty(), "Empty span could spell trouble"); 7604 } 7605 7606 7607 // CMSKeepAliveClosure: the serial version 7608 void CMSKeepAliveClosure::do_oop(oop obj) { 7609 HeapWord* addr = (HeapWord*)obj; 7610 if (_span.contains(addr) && 7611 !_bit_map->isMarked(addr)) { 7612 _bit_map->mark(addr); 7613 bool simulate_overflow = false; 7614 NOT_PRODUCT( 7615 if (CMSMarkStackOverflowALot && 7616 _collector->simulate_overflow()) { 7617 // simulate a stack overflow 7618 simulate_overflow = true; 7619 } 7620 ) 7621 if (simulate_overflow || !_mark_stack->push(obj)) { 7622 if (_concurrent_precleaning) { 7623 // We dirty the overflown object and let the remark 7624 // phase deal with it. 7625 assert(_collector->overflow_list_is_empty(), "Error"); 7626 // In the case of object arrays, we need to dirty all of 7627 // the cards that the object spans. No locking or atomics 7628 // are needed since no one else can be mutating the mod union 7629 // table. 7630 if (obj->is_objArray()) { 7631 size_t sz = obj->size(); 7632 HeapWord* end_card_addr = 7633 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7634 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7635 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7636 _collector->_modUnionTable.mark_range(redirty_range); 7637 } else { 7638 _collector->_modUnionTable.mark(addr); 7639 } 7640 _collector->_ser_kac_preclean_ovflw++; 7641 } else { 7642 _collector->push_on_overflow_list(obj); 7643 _collector->_ser_kac_ovflw++; 7644 } 7645 } 7646 } 7647 } 7648 7649 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7650 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7651 7652 // CMSParKeepAliveClosure: a parallel version of the above. 7653 // The work queues are private to each closure (thread), 7654 // but (may be) available for stealing by other threads. 7655 void CMSParKeepAliveClosure::do_oop(oop obj) { 7656 HeapWord* addr = (HeapWord*)obj; 7657 if (_span.contains(addr) && 7658 !_bit_map->isMarked(addr)) { 7659 // In general, during recursive tracing, several threads 7660 // may be concurrently getting here; the first one to 7661 // "tag" it, claims it. 7662 if (_bit_map->par_mark(addr)) { 7663 bool res = _work_queue->push(obj); 7664 assert(res, "Low water mark should be much less than capacity"); 7665 // Do a recursive trim in the hope that this will keep 7666 // stack usage lower, but leave some oops for potential stealers 7667 trim_queue(_low_water_mark); 7668 } // Else, another thread got there first 7669 } 7670 } 7671 7672 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7673 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7674 7675 void CMSParKeepAliveClosure::trim_queue(uint max) { 7676 while (_work_queue->size() > max) { 7677 oop new_oop; 7678 if (_work_queue->pop_local(new_oop)) { 7679 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7680 assert(_bit_map->isMarked((HeapWord*)new_oop), 7681 "no white objects on this stack!"); 7682 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7683 // iterate over the oops in this oop, marking and pushing 7684 // the ones in CMS heap (i.e. in _span). 7685 new_oop->oop_iterate(&_mark_and_push); 7686 } 7687 } 7688 } 7689 7690 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7691 CMSCollector* collector, 7692 MemRegion span, CMSBitMap* bit_map, 7693 OopTaskQueue* work_queue): 7694 _collector(collector), 7695 _span(span), 7696 _bit_map(bit_map), 7697 _work_queue(work_queue) { } 7698 7699 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7700 HeapWord* addr = (HeapWord*)obj; 7701 if (_span.contains(addr) && 7702 !_bit_map->isMarked(addr)) { 7703 if (_bit_map->par_mark(addr)) { 7704 bool simulate_overflow = false; 7705 NOT_PRODUCT( 7706 if (CMSMarkStackOverflowALot && 7707 _collector->par_simulate_overflow()) { 7708 // simulate a stack overflow 7709 simulate_overflow = true; 7710 } 7711 ) 7712 if (simulate_overflow || !_work_queue->push(obj)) { 7713 _collector->par_push_on_overflow_list(obj); 7714 _collector->_par_kac_ovflw++; 7715 } 7716 } // Else another thread got there already 7717 } 7718 } 7719 7720 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7721 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7722 7723 ////////////////////////////////////////////////////////////////// 7724 // CMSExpansionCause ///////////////////////////// 7725 ////////////////////////////////////////////////////////////////// 7726 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7727 switch (cause) { 7728 case _no_expansion: 7729 return "No expansion"; 7730 case _satisfy_free_ratio: 7731 return "Free ratio"; 7732 case _satisfy_promotion: 7733 return "Satisfy promotion"; 7734 case _satisfy_allocation: 7735 return "allocation"; 7736 case _allocate_par_lab: 7737 return "Par LAB"; 7738 case _allocate_par_spooling_space: 7739 return "Par Spooling Space"; 7740 case _adaptive_size_policy: 7741 return "Ergonomics"; 7742 default: 7743 return "unknown"; 7744 } 7745 } 7746 7747 void CMSDrainMarkingStackClosure::do_void() { 7748 // the max number to take from overflow list at a time 7749 const size_t num = _mark_stack->capacity()/4; 7750 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7751 "Overflow list should be NULL during concurrent phases"); 7752 while (!_mark_stack->isEmpty() || 7753 // if stack is empty, check the overflow list 7754 _collector->take_from_overflow_list(num, _mark_stack)) { 7755 oop obj = _mark_stack->pop(); 7756 HeapWord* addr = (HeapWord*)obj; 7757 assert(_span.contains(addr), "Should be within span"); 7758 assert(_bit_map->isMarked(addr), "Should be marked"); 7759 assert(obj->is_oop(), "Should be an oop"); 7760 obj->oop_iterate(_keep_alive); 7761 } 7762 } 7763 7764 void CMSParDrainMarkingStackClosure::do_void() { 7765 // drain queue 7766 trim_queue(0); 7767 } 7768 7769 // Trim our work_queue so its length is below max at return 7770 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7771 while (_work_queue->size() > max) { 7772 oop new_oop; 7773 if (_work_queue->pop_local(new_oop)) { 7774 assert(new_oop->is_oop(), "Expected an oop"); 7775 assert(_bit_map->isMarked((HeapWord*)new_oop), 7776 "no white objects on this stack!"); 7777 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7778 // iterate over the oops in this oop, marking and pushing 7779 // the ones in CMS heap (i.e. in _span). 7780 new_oop->oop_iterate(&_mark_and_push); 7781 } 7782 } 7783 } 7784 7785 //////////////////////////////////////////////////////////////////// 7786 // Support for Marking Stack Overflow list handling and related code 7787 //////////////////////////////////////////////////////////////////// 7788 // Much of the following code is similar in shape and spirit to the 7789 // code used in ParNewGC. We should try and share that code 7790 // as much as possible in the future. 7791 7792 #ifndef PRODUCT 7793 // Debugging support for CMSStackOverflowALot 7794 7795 // It's OK to call this multi-threaded; the worst thing 7796 // that can happen is that we'll get a bunch of closely 7797 // spaced simulated overflows, but that's OK, in fact 7798 // probably good as it would exercise the overflow code 7799 // under contention. 7800 bool CMSCollector::simulate_overflow() { 7801 if (_overflow_counter-- <= 0) { // just being defensive 7802 _overflow_counter = CMSMarkStackOverflowInterval; 7803 return true; 7804 } else { 7805 return false; 7806 } 7807 } 7808 7809 bool CMSCollector::par_simulate_overflow() { 7810 return simulate_overflow(); 7811 } 7812 #endif 7813 7814 // Single-threaded 7815 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7816 assert(stack->isEmpty(), "Expected precondition"); 7817 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7818 size_t i = num; 7819 oop cur = _overflow_list; 7820 const markOop proto = markOopDesc::prototype(); 7821 NOT_PRODUCT(ssize_t n = 0;) 7822 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7823 next = oop(cur->mark()); 7824 cur->set_mark(proto); // until proven otherwise 7825 assert(cur->is_oop(), "Should be an oop"); 7826 bool res = stack->push(cur); 7827 assert(res, "Bit off more than can chew?"); 7828 NOT_PRODUCT(n++;) 7829 } 7830 _overflow_list = cur; 7831 #ifndef PRODUCT 7832 assert(_num_par_pushes >= n, "Too many pops?"); 7833 _num_par_pushes -=n; 7834 #endif 7835 return !stack->isEmpty(); 7836 } 7837 7838 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7839 // (MT-safe) Get a prefix of at most "num" from the list. 7840 // The overflow list is chained through the mark word of 7841 // each object in the list. We fetch the entire list, 7842 // break off a prefix of the right size and return the 7843 // remainder. If other threads try to take objects from 7844 // the overflow list at that time, they will wait for 7845 // some time to see if data becomes available. If (and 7846 // only if) another thread places one or more object(s) 7847 // on the global list before we have returned the suffix 7848 // to the global list, we will walk down our local list 7849 // to find its end and append the global list to 7850 // our suffix before returning it. This suffix walk can 7851 // prove to be expensive (quadratic in the amount of traffic) 7852 // when there are many objects in the overflow list and 7853 // there is much producer-consumer contention on the list. 7854 // *NOTE*: The overflow list manipulation code here and 7855 // in ParNewGeneration:: are very similar in shape, 7856 // except that in the ParNew case we use the old (from/eden) 7857 // copy of the object to thread the list via its klass word. 7858 // Because of the common code, if you make any changes in 7859 // the code below, please check the ParNew version to see if 7860 // similar changes might be needed. 7861 // CR 6797058 has been filed to consolidate the common code. 7862 bool CMSCollector::par_take_from_overflow_list(size_t num, 7863 OopTaskQueue* work_q, 7864 int no_of_gc_threads) { 7865 assert(work_q->size() == 0, "First empty local work queue"); 7866 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7867 if (_overflow_list == NULL) { 7868 return false; 7869 } 7870 // Grab the entire list; we'll put back a suffix 7871 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7872 Thread* tid = Thread::current(); 7873 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7874 // set to ParallelGCThreads. 7875 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7876 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7877 // If the list is busy, we spin for a short while, 7878 // sleeping between attempts to get the list. 7879 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7880 os::sleep(tid, sleep_time_millis, false); 7881 if (_overflow_list == NULL) { 7882 // Nothing left to take 7883 return false; 7884 } else if (_overflow_list != BUSY) { 7885 // Try and grab the prefix 7886 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7887 } 7888 } 7889 // If the list was found to be empty, or we spun long 7890 // enough, we give up and return empty-handed. If we leave 7891 // the list in the BUSY state below, it must be the case that 7892 // some other thread holds the overflow list and will set it 7893 // to a non-BUSY state in the future. 7894 if (prefix == NULL || prefix == BUSY) { 7895 // Nothing to take or waited long enough 7896 if (prefix == NULL) { 7897 // Write back the NULL in case we overwrote it with BUSY above 7898 // and it is still the same value. 7899 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7900 } 7901 return false; 7902 } 7903 assert(prefix != NULL && prefix != BUSY, "Error"); 7904 size_t i = num; 7905 oop cur = prefix; 7906 // Walk down the first "num" objects, unless we reach the end. 7907 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7908 if (cur->mark() == NULL) { 7909 // We have "num" or fewer elements in the list, so there 7910 // is nothing to return to the global list. 7911 // Write back the NULL in lieu of the BUSY we wrote 7912 // above, if it is still the same value. 7913 if (_overflow_list == BUSY) { 7914 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7915 } 7916 } else { 7917 // Chop off the suffix and return it to the global list. 7918 assert(cur->mark() != BUSY, "Error"); 7919 oop suffix_head = cur->mark(); // suffix will be put back on global list 7920 cur->set_mark(NULL); // break off suffix 7921 // It's possible that the list is still in the empty(busy) state 7922 // we left it in a short while ago; in that case we may be 7923 // able to place back the suffix without incurring the cost 7924 // of a walk down the list. 7925 oop observed_overflow_list = _overflow_list; 7926 oop cur_overflow_list = observed_overflow_list; 7927 bool attached = false; 7928 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7929 observed_overflow_list = 7930 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7931 if (cur_overflow_list == observed_overflow_list) { 7932 attached = true; 7933 break; 7934 } else cur_overflow_list = observed_overflow_list; 7935 } 7936 if (!attached) { 7937 // Too bad, someone else sneaked in (at least) an element; we'll need 7938 // to do a splice. Find tail of suffix so we can prepend suffix to global 7939 // list. 7940 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7941 oop suffix_tail = cur; 7942 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7943 "Tautology"); 7944 observed_overflow_list = _overflow_list; 7945 do { 7946 cur_overflow_list = observed_overflow_list; 7947 if (cur_overflow_list != BUSY) { 7948 // Do the splice ... 7949 suffix_tail->set_mark(markOop(cur_overflow_list)); 7950 } else { // cur_overflow_list == BUSY 7951 suffix_tail->set_mark(NULL); 7952 } 7953 // ... and try to place spliced list back on overflow_list ... 7954 observed_overflow_list = 7955 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7956 } while (cur_overflow_list != observed_overflow_list); 7957 // ... until we have succeeded in doing so. 7958 } 7959 } 7960 7961 // Push the prefix elements on work_q 7962 assert(prefix != NULL, "control point invariant"); 7963 const markOop proto = markOopDesc::prototype(); 7964 oop next; 7965 NOT_PRODUCT(ssize_t n = 0;) 7966 for (cur = prefix; cur != NULL; cur = next) { 7967 next = oop(cur->mark()); 7968 cur->set_mark(proto); // until proven otherwise 7969 assert(cur->is_oop(), "Should be an oop"); 7970 bool res = work_q->push(cur); 7971 assert(res, "Bit off more than we can chew?"); 7972 NOT_PRODUCT(n++;) 7973 } 7974 #ifndef PRODUCT 7975 assert(_num_par_pushes >= n, "Too many pops?"); 7976 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7977 #endif 7978 return true; 7979 } 7980 7981 // Single-threaded 7982 void CMSCollector::push_on_overflow_list(oop p) { 7983 NOT_PRODUCT(_num_par_pushes++;) 7984 assert(p->is_oop(), "Not an oop"); 7985 preserve_mark_if_necessary(p); 7986 p->set_mark((markOop)_overflow_list); 7987 _overflow_list = p; 7988 } 7989 7990 // Multi-threaded; use CAS to prepend to overflow list 7991 void CMSCollector::par_push_on_overflow_list(oop p) { 7992 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7993 assert(p->is_oop(), "Not an oop"); 7994 par_preserve_mark_if_necessary(p); 7995 oop observed_overflow_list = _overflow_list; 7996 oop cur_overflow_list; 7997 do { 7998 cur_overflow_list = observed_overflow_list; 7999 if (cur_overflow_list != BUSY) { 8000 p->set_mark(markOop(cur_overflow_list)); 8001 } else { 8002 p->set_mark(NULL); 8003 } 8004 observed_overflow_list = 8005 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8006 } while (cur_overflow_list != observed_overflow_list); 8007 } 8008 #undef BUSY 8009 8010 // Single threaded 8011 // General Note on GrowableArray: pushes may silently fail 8012 // because we are (temporarily) out of C-heap for expanding 8013 // the stack. The problem is quite ubiquitous and affects 8014 // a lot of code in the JVM. The prudent thing for GrowableArray 8015 // to do (for now) is to exit with an error. However, that may 8016 // be too draconian in some cases because the caller may be 8017 // able to recover without much harm. For such cases, we 8018 // should probably introduce a "soft_push" method which returns 8019 // an indication of success or failure with the assumption that 8020 // the caller may be able to recover from a failure; code in 8021 // the VM can then be changed, incrementally, to deal with such 8022 // failures where possible, thus, incrementally hardening the VM 8023 // in such low resource situations. 8024 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8025 _preserved_oop_stack.push(p); 8026 _preserved_mark_stack.push(m); 8027 assert(m == p->mark(), "Mark word changed"); 8028 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8029 "bijection"); 8030 } 8031 8032 // Single threaded 8033 void CMSCollector::preserve_mark_if_necessary(oop p) { 8034 markOop m = p->mark(); 8035 if (m->must_be_preserved(p)) { 8036 preserve_mark_work(p, m); 8037 } 8038 } 8039 8040 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8041 markOop m = p->mark(); 8042 if (m->must_be_preserved(p)) { 8043 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8044 // Even though we read the mark word without holding 8045 // the lock, we are assured that it will not change 8046 // because we "own" this oop, so no other thread can 8047 // be trying to push it on the overflow list; see 8048 // the assertion in preserve_mark_work() that checks 8049 // that m == p->mark(). 8050 preserve_mark_work(p, m); 8051 } 8052 } 8053 8054 // We should be able to do this multi-threaded, 8055 // a chunk of stack being a task (this is 8056 // correct because each oop only ever appears 8057 // once in the overflow list. However, it's 8058 // not very easy to completely overlap this with 8059 // other operations, so will generally not be done 8060 // until all work's been completed. Because we 8061 // expect the preserved oop stack (set) to be small, 8062 // it's probably fine to do this single-threaded. 8063 // We can explore cleverer concurrent/overlapped/parallel 8064 // processing of preserved marks if we feel the 8065 // need for this in the future. Stack overflow should 8066 // be so rare in practice and, when it happens, its 8067 // effect on performance so great that this will 8068 // likely just be in the noise anyway. 8069 void CMSCollector::restore_preserved_marks_if_any() { 8070 assert(SafepointSynchronize::is_at_safepoint(), 8071 "world should be stopped"); 8072 assert(Thread::current()->is_ConcurrentGC_thread() || 8073 Thread::current()->is_VM_thread(), 8074 "should be single-threaded"); 8075 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8076 "bijection"); 8077 8078 while (!_preserved_oop_stack.is_empty()) { 8079 oop p = _preserved_oop_stack.pop(); 8080 assert(p->is_oop(), "Should be an oop"); 8081 assert(_span.contains(p), "oop should be in _span"); 8082 assert(p->mark() == markOopDesc::prototype(), 8083 "Set when taken from overflow list"); 8084 markOop m = _preserved_mark_stack.pop(); 8085 p->set_mark(m); 8086 } 8087 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8088 "stacks were cleared above"); 8089 } 8090 8091 #ifndef PRODUCT 8092 bool CMSCollector::no_preserved_marks() const { 8093 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8094 } 8095 #endif 8096 8097 // Transfer some number of overflown objects to usual marking 8098 // stack. Return true if some objects were transferred. 8099 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8100 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8101 (size_t)ParGCDesiredObjsFromOverflowList); 8102 8103 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8104 assert(_collector->overflow_list_is_empty() || res, 8105 "If list is not empty, we should have taken something"); 8106 assert(!res || !_mark_stack->isEmpty(), 8107 "If we took something, it should now be on our stack"); 8108 return res; 8109 } 8110 8111 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8112 size_t res = _sp->block_size_no_stall(addr, _collector); 8113 if (_sp->block_is_obj(addr)) { 8114 if (_live_bit_map->isMarked(addr)) { 8115 // It can't have been dead in a previous cycle 8116 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8117 } else { 8118 _dead_bit_map->mark(addr); // mark the dead object 8119 } 8120 } 8121 // Could be 0, if the block size could not be computed without stalling. 8122 return res; 8123 } 8124 8125 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8126 8127 switch (phase) { 8128 case CMSCollector::InitialMarking: 8129 initialize(true /* fullGC */ , 8130 cause /* cause of the GC */, 8131 true /* recordGCBeginTime */, 8132 true /* recordPreGCUsage */, 8133 false /* recordPeakUsage */, 8134 false /* recordPostGCusage */, 8135 true /* recordAccumulatedGCTime */, 8136 false /* recordGCEndTime */, 8137 false /* countCollection */ ); 8138 break; 8139 8140 case CMSCollector::FinalMarking: 8141 initialize(true /* fullGC */ , 8142 cause /* cause of the GC */, 8143 false /* recordGCBeginTime */, 8144 false /* recordPreGCUsage */, 8145 false /* recordPeakUsage */, 8146 false /* recordPostGCusage */, 8147 true /* recordAccumulatedGCTime */, 8148 false /* recordGCEndTime */, 8149 false /* countCollection */ ); 8150 break; 8151 8152 case CMSCollector::Sweeping: 8153 initialize(true /* fullGC */ , 8154 cause /* cause of the GC */, 8155 false /* recordGCBeginTime */, 8156 false /* recordPreGCUsage */, 8157 true /* recordPeakUsage */, 8158 true /* recordPostGCusage */, 8159 false /* recordAccumulatedGCTime */, 8160 true /* recordGCEndTime */, 8161 true /* countCollection */ ); 8162 break; 8163 8164 default: 8165 ShouldNotReachHere(); 8166 } 8167 }