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