1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.inline.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->print(" "); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 warning("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 warning("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 warning("GC/CMS: _conc_workers allocation failure: " 531 "forcing -CMSConcurrentMTEnabled"); 532 CMSConcurrentMTEnabled = false; 533 } else { 534 _conc_workers->initialize_workers(); 535 } 536 } else { 537 CMSConcurrentMTEnabled = false; 538 } 539 } 540 if (!CMSConcurrentMTEnabled) { 541 ConcGCThreads = 0; 542 } else { 543 // Turn off CMSCleanOnEnter optimization temporarily for 544 // the MT case where it's not fixed yet; see 6178663. 545 CMSCleanOnEnter = false; 546 } 547 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 548 "Inconsistency"); 549 550 // Parallel task queues; these are shared for the 551 // concurrent and stop-world phases of CMS, but 552 // are not shared with parallel scavenge (ParNew). 553 { 554 uint i; 555 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 556 557 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 558 || ParallelRefProcEnabled) 559 && num_queues > 0) { 560 _task_queues = new OopTaskQueueSet(num_queues); 561 if (_task_queues == NULL) { 562 warning("task_queues allocation failure."); 563 return; 564 } 565 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 566 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 567 for (i = 0; i < num_queues; i++) { 568 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 569 if (q == NULL) { 570 warning("work_queue allocation failure."); 571 return; 572 } 573 _task_queues->register_queue(i, q); 574 } 575 for (i = 0; i < num_queues; i++) { 576 _task_queues->queue(i)->initialize(); 577 _hash_seed[i] = 17; // copied from ParNew 578 } 579 } 580 } 581 582 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 583 584 // Clip CMSBootstrapOccupancy between 0 and 100. 585 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 586 587 // Now tell CMS generations the identity of their collector 588 ConcurrentMarkSweepGeneration::set_collector(this); 589 590 // Create & start a CMS thread for this CMS collector 591 _cmsThread = ConcurrentMarkSweepThread::start(this); 592 assert(cmsThread() != NULL, "CMS Thread should have been created"); 593 assert(cmsThread()->collector() == this, 594 "CMS Thread should refer to this gen"); 595 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 596 597 // Support for parallelizing young gen rescan 598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 599 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 600 _young_gen = (ParNewGeneration*)gch->young_gen(); 601 if (gch->supports_inline_contig_alloc()) { 602 _top_addr = gch->top_addr(); 603 _end_addr = gch->end_addr(); 604 assert(_young_gen != NULL, "no _young_gen"); 605 _eden_chunk_index = 0; 606 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 607 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 608 } 609 610 // Support for parallelizing survivor space rescan 611 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 612 const size_t max_plab_samples = 613 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 614 615 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 616 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 617 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 618 _survivor_chunk_capacity = max_plab_samples; 619 for (uint i = 0; i < ParallelGCThreads; i++) { 620 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 621 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 622 assert(cur->end() == 0, "Should be 0"); 623 assert(cur->array() == vec, "Should be vec"); 624 assert(cur->capacity() == max_plab_samples, "Error"); 625 } 626 } 627 628 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 629 _gc_counters = new CollectorCounters("CMS", 1); 630 _completed_initialization = true; 631 _inter_sweep_timer.start(); // start of time 632 } 633 634 const char* ConcurrentMarkSweepGeneration::name() const { 635 return "concurrent mark-sweep generation"; 636 } 637 void ConcurrentMarkSweepGeneration::update_counters() { 638 if (UsePerfData) { 639 _space_counters->update_all(); 640 _gen_counters->update_all(); 641 } 642 } 643 644 // this is an optimized version of update_counters(). it takes the 645 // used value as a parameter rather than computing it. 646 // 647 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 648 if (UsePerfData) { 649 _space_counters->update_used(used); 650 _space_counters->update_capacity(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 void ConcurrentMarkSweepGeneration::print() const { 656 Generation::print(); 657 cmsSpace()->print(); 658 } 659 660 #ifndef PRODUCT 661 void ConcurrentMarkSweepGeneration::print_statistics() { 662 cmsSpace()->printFLCensus(0); 663 } 664 #endif 665 666 size_t 667 ConcurrentMarkSweepGeneration::contiguous_available() const { 668 // dld proposes an improvement in precision here. If the committed 669 // part of the space ends in a free block we should add that to 670 // uncommitted size in the calculation below. Will make this 671 // change later, staying with the approximation below for the 672 // time being. -- ysr. 673 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 674 } 675 676 size_t 677 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 678 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 679 } 680 681 size_t ConcurrentMarkSweepGeneration::max_available() const { 682 return free() + _virtual_space.uncommitted_size(); 683 } 684 685 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 686 size_t available = max_available(); 687 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 688 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 689 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 690 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 691 return res; 692 } 693 694 // At a promotion failure dump information on block layout in heap 695 // (cms old generation). 696 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 697 LogHandle(gc, promotion) log; 698 if (log.is_trace()) { 699 ResourceMark rm; 700 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 701 } 702 } 703 704 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 705 // Clear the promotion information. These pointers can be adjusted 706 // along with all the other pointers into the heap but 707 // compaction is expected to be a rare event with 708 // a heap using cms so don't do it without seeing the need. 709 for (uint i = 0; i < ParallelGCThreads; i++) { 710 _par_gc_thread_states[i]->promo.reset(); 711 } 712 } 713 714 void ConcurrentMarkSweepGeneration::compute_new_size() { 715 assert_locked_or_safepoint(Heap_lock); 716 717 // If incremental collection failed, we just want to expand 718 // to the limit. 719 if (incremental_collection_failed()) { 720 clear_incremental_collection_failed(); 721 grow_to_reserved(); 722 return; 723 } 724 725 // The heap has been compacted but not reset yet. 726 // Any metric such as free() or used() will be incorrect. 727 728 CardGeneration::compute_new_size(); 729 730 // Reset again after a possible resizing 731 if (did_compact()) { 732 cmsSpace()->reset_after_compaction(); 733 } 734 } 735 736 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 737 assert_locked_or_safepoint(Heap_lock); 738 739 // If incremental collection failed, we just want to expand 740 // to the limit. 741 if (incremental_collection_failed()) { 742 clear_incremental_collection_failed(); 743 grow_to_reserved(); 744 return; 745 } 746 747 double free_percentage = ((double) free()) / capacity(); 748 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 749 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 750 751 // compute expansion delta needed for reaching desired free percentage 752 if (free_percentage < desired_free_percentage) { 753 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 754 assert(desired_capacity >= capacity(), "invalid expansion size"); 755 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 756 LogHandle(gc) log; 757 if (log.is_trace()) { 758 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 759 log.trace("From compute_new_size: "); 760 log.trace(" Free fraction %f", free_percentage); 761 log.trace(" Desired free fraction %f", desired_free_percentage); 762 log.trace(" Maximum free fraction %f", maximum_free_percentage); 763 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 764 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 765 GenCollectedHeap* gch = GenCollectedHeap::heap(); 766 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 767 size_t young_size = gch->young_gen()->capacity(); 768 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 769 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 770 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 771 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 772 } 773 // safe if expansion fails 774 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 775 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 776 } else { 777 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 778 assert(desired_capacity <= capacity(), "invalid expansion size"); 779 size_t shrink_bytes = capacity() - desired_capacity; 780 // Don't shrink unless the delta is greater than the minimum shrink we want 781 if (shrink_bytes >= MinHeapDeltaBytes) { 782 shrink_free_list_by(shrink_bytes); 783 } 784 } 785 } 786 787 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 788 return cmsSpace()->freelistLock(); 789 } 790 791 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 792 CMSSynchronousYieldRequest yr; 793 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 794 return have_lock_and_allocate(size, tlab); 795 } 796 797 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 798 bool tlab /* ignored */) { 799 assert_lock_strong(freelistLock()); 800 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 801 HeapWord* res = cmsSpace()->allocate(adjustedSize); 802 // Allocate the object live (grey) if the background collector has 803 // started marking. This is necessary because the marker may 804 // have passed this address and consequently this object will 805 // not otherwise be greyed and would be incorrectly swept up. 806 // Note that if this object contains references, the writing 807 // of those references will dirty the card containing this object 808 // allowing the object to be blackened (and its references scanned) 809 // either during a preclean phase or at the final checkpoint. 810 if (res != NULL) { 811 // We may block here with an uninitialized object with 812 // its mark-bit or P-bits not yet set. Such objects need 813 // to be safely navigable by block_start(). 814 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 815 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 816 collector()->direct_allocated(res, adjustedSize); 817 _direct_allocated_words += adjustedSize; 818 // allocation counters 819 NOT_PRODUCT( 820 _numObjectsAllocated++; 821 _numWordsAllocated += (int)adjustedSize; 822 ) 823 } 824 return res; 825 } 826 827 // In the case of direct allocation by mutators in a generation that 828 // is being concurrently collected, the object must be allocated 829 // live (grey) if the background collector has started marking. 830 // This is necessary because the marker may 831 // have passed this address and consequently this object will 832 // not otherwise be greyed and would be incorrectly swept up. 833 // Note that if this object contains references, the writing 834 // of those references will dirty the card containing this object 835 // allowing the object to be blackened (and its references scanned) 836 // either during a preclean phase or at the final checkpoint. 837 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 838 assert(_markBitMap.covers(start, size), "Out of bounds"); 839 if (_collectorState >= Marking) { 840 MutexLockerEx y(_markBitMap.lock(), 841 Mutex::_no_safepoint_check_flag); 842 // [see comments preceding SweepClosure::do_blk() below for details] 843 // 844 // Can the P-bits be deleted now? JJJ 845 // 846 // 1. need to mark the object as live so it isn't collected 847 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 848 // 3. need to mark the end of the object so marking, precleaning or sweeping 849 // can skip over uninitialized or unparsable objects. An allocated 850 // object is considered uninitialized for our purposes as long as 851 // its klass word is NULL. All old gen objects are parsable 852 // as soon as they are initialized.) 853 _markBitMap.mark(start); // object is live 854 _markBitMap.mark(start + 1); // object is potentially uninitialized? 855 _markBitMap.mark(start + size - 1); 856 // mark end of object 857 } 858 // check that oop looks uninitialized 859 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 860 } 861 862 void CMSCollector::promoted(bool par, HeapWord* start, 863 bool is_obj_array, size_t obj_size) { 864 assert(_markBitMap.covers(start), "Out of bounds"); 865 // See comment in direct_allocated() about when objects should 866 // be allocated live. 867 if (_collectorState >= Marking) { 868 // we already hold the marking bit map lock, taken in 869 // the prologue 870 if (par) { 871 _markBitMap.par_mark(start); 872 } else { 873 _markBitMap.mark(start); 874 } 875 // We don't need to mark the object as uninitialized (as 876 // in direct_allocated above) because this is being done with the 877 // world stopped and the object will be initialized by the 878 // time the marking, precleaning or sweeping get to look at it. 879 // But see the code for copying objects into the CMS generation, 880 // where we need to ensure that concurrent readers of the 881 // block offset table are able to safely navigate a block that 882 // is in flux from being free to being allocated (and in 883 // transition while being copied into) and subsequently 884 // becoming a bona-fide object when the copy/promotion is complete. 885 assert(SafepointSynchronize::is_at_safepoint(), 886 "expect promotion only at safepoints"); 887 888 if (_collectorState < Sweeping) { 889 // Mark the appropriate cards in the modUnionTable, so that 890 // this object gets scanned before the sweep. If this is 891 // not done, CMS generation references in the object might 892 // not get marked. 893 // For the case of arrays, which are otherwise precisely 894 // marked, we need to dirty the entire array, not just its head. 895 if (is_obj_array) { 896 // The [par_]mark_range() method expects mr.end() below to 897 // be aligned to the granularity of a bit's representation 898 // in the heap. In the case of the MUT below, that's a 899 // card size. 900 MemRegion mr(start, 901 (HeapWord*)round_to((intptr_t)(start + obj_size), 902 CardTableModRefBS::card_size /* bytes */)); 903 if (par) { 904 _modUnionTable.par_mark_range(mr); 905 } else { 906 _modUnionTable.mark_range(mr); 907 } 908 } else { // not an obj array; we can just mark the head 909 if (par) { 910 _modUnionTable.par_mark(start); 911 } else { 912 _modUnionTable.mark(start); 913 } 914 } 915 } 916 } 917 } 918 919 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 920 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 921 // allocate, copy and if necessary update promoinfo -- 922 // delegate to underlying space. 923 assert_lock_strong(freelistLock()); 924 925 #ifndef PRODUCT 926 if (GenCollectedHeap::heap()->promotion_should_fail()) { 927 return NULL; 928 } 929 #endif // #ifndef PRODUCT 930 931 oop res = _cmsSpace->promote(obj, obj_size); 932 if (res == NULL) { 933 // expand and retry 934 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 935 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 936 // Since this is the old generation, we don't try to promote 937 // into a more senior generation. 938 res = _cmsSpace->promote(obj, obj_size); 939 } 940 if (res != NULL) { 941 // See comment in allocate() about when objects should 942 // be allocated live. 943 assert(obj->is_oop(), "Will dereference klass pointer below"); 944 collector()->promoted(false, // Not parallel 945 (HeapWord*)res, obj->is_objArray(), obj_size); 946 // promotion counters 947 NOT_PRODUCT( 948 _numObjectsPromoted++; 949 _numWordsPromoted += 950 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 951 ) 952 } 953 return res; 954 } 955 956 957 // IMPORTANT: Notes on object size recognition in CMS. 958 // --------------------------------------------------- 959 // A block of storage in the CMS generation is always in 960 // one of three states. A free block (FREE), an allocated 961 // object (OBJECT) whose size() method reports the correct size, 962 // and an intermediate state (TRANSIENT) in which its size cannot 963 // be accurately determined. 964 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 965 // ----------------------------------------------------- 966 // FREE: klass_word & 1 == 1; mark_word holds block size 967 // 968 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 969 // obj->size() computes correct size 970 // 971 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 972 // 973 // STATE IDENTIFICATION: (64 bit+COOPS) 974 // ------------------------------------ 975 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 976 // 977 // OBJECT: klass_word installed; klass_word != 0; 978 // obj->size() computes correct size 979 // 980 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 981 // 982 // 983 // STATE TRANSITION DIAGRAM 984 // 985 // mut / parnew mut / parnew 986 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 987 // ^ | 988 // |------------------------ DEAD <------------------------------------| 989 // sweep mut 990 // 991 // While a block is in TRANSIENT state its size cannot be determined 992 // so readers will either need to come back later or stall until 993 // the size can be determined. Note that for the case of direct 994 // allocation, P-bits, when available, may be used to determine the 995 // size of an object that may not yet have been initialized. 996 997 // Things to support parallel young-gen collection. 998 oop 999 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1000 oop old, markOop m, 1001 size_t word_sz) { 1002 #ifndef PRODUCT 1003 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1004 return NULL; 1005 } 1006 #endif // #ifndef PRODUCT 1007 1008 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1009 PromotionInfo* promoInfo = &ps->promo; 1010 // if we are tracking promotions, then first ensure space for 1011 // promotion (including spooling space for saving header if necessary). 1012 // then allocate and copy, then track promoted info if needed. 1013 // When tracking (see PromotionInfo::track()), the mark word may 1014 // be displaced and in this case restoration of the mark word 1015 // occurs in the (oop_since_save_marks_)iterate phase. 1016 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1017 // Out of space for allocating spooling buffers; 1018 // try expanding and allocating spooling buffers. 1019 if (!expand_and_ensure_spooling_space(promoInfo)) { 1020 return NULL; 1021 } 1022 } 1023 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1024 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1025 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1026 if (obj_ptr == NULL) { 1027 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1028 if (obj_ptr == NULL) { 1029 return NULL; 1030 } 1031 } 1032 oop obj = oop(obj_ptr); 1033 OrderAccess::storestore(); 1034 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1036 // IMPORTANT: See note on object initialization for CMS above. 1037 // Otherwise, copy the object. Here we must be careful to insert the 1038 // klass pointer last, since this marks the block as an allocated object. 1039 // Except with compressed oops it's the mark word. 1040 HeapWord* old_ptr = (HeapWord*)old; 1041 // Restore the mark word copied above. 1042 obj->set_mark(m); 1043 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1044 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1045 OrderAccess::storestore(); 1046 1047 if (UseCompressedClassPointers) { 1048 // Copy gap missed by (aligned) header size calculation below 1049 obj->set_klass_gap(old->klass_gap()); 1050 } 1051 if (word_sz > (size_t)oopDesc::header_size()) { 1052 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1053 obj_ptr + oopDesc::header_size(), 1054 word_sz - oopDesc::header_size()); 1055 } 1056 1057 // Now we can track the promoted object, if necessary. We take care 1058 // to delay the transition from uninitialized to full object 1059 // (i.e., insertion of klass pointer) until after, so that it 1060 // atomically becomes a promoted object. 1061 if (promoInfo->tracking()) { 1062 promoInfo->track((PromotedObject*)obj, old->klass()); 1063 } 1064 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1065 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1066 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1067 1068 // Finally, install the klass pointer (this should be volatile). 1069 OrderAccess::storestore(); 1070 obj->set_klass(old->klass()); 1071 // We should now be able to calculate the right size for this object 1072 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1073 1074 collector()->promoted(true, // parallel 1075 obj_ptr, old->is_objArray(), word_sz); 1076 1077 NOT_PRODUCT( 1078 Atomic::inc_ptr(&_numObjectsPromoted); 1079 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1080 ) 1081 1082 return obj; 1083 } 1084 1085 void 1086 ConcurrentMarkSweepGeneration:: 1087 par_promote_alloc_done(int thread_num) { 1088 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1089 ps->lab.retire(thread_num); 1090 } 1091 1092 void 1093 ConcurrentMarkSweepGeneration:: 1094 par_oop_since_save_marks_iterate_done(int thread_num) { 1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1096 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1097 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1098 } 1099 1100 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1101 size_t size, 1102 bool tlab) 1103 { 1104 // We allow a STW collection only if a full 1105 // collection was requested. 1106 return full || should_allocate(size, tlab); // FIX ME !!! 1107 // This and promotion failure handling are connected at the 1108 // hip and should be fixed by untying them. 1109 } 1110 1111 bool CMSCollector::shouldConcurrentCollect() { 1112 if (_full_gc_requested) { 1113 log_trace(gc)("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1114 return true; 1115 } 1116 1117 FreelistLocker x(this); 1118 // ------------------------------------------------------------------ 1119 // Print out lots of information which affects the initiation of 1120 // a collection. 1121 LogHandle(gc) log; 1122 if (log.is_trace() && stats().valid()) { 1123 log.trace("CMSCollector shouldConcurrentCollect: "); 1124 ResourceMark rm; 1125 stats().print_on(log.debug_stream()); 1126 log.trace("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1127 log.trace("free=" SIZE_FORMAT, _cmsGen->free()); 1128 log.trace("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1129 log.trace("promotion_rate=%g", stats().promotion_rate()); 1130 log.trace("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1131 log.trace("occupancy=%3.7f", _cmsGen->occupancy()); 1132 log.trace("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1133 log.trace("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1134 log.trace("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1135 log.trace("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1136 } 1137 // ------------------------------------------------------------------ 1138 1139 // If the estimated time to complete a cms collection (cms_duration()) 1140 // is less than the estimated time remaining until the cms generation 1141 // is full, start a collection. 1142 if (!UseCMSInitiatingOccupancyOnly) { 1143 if (stats().valid()) { 1144 if (stats().time_until_cms_start() == 0.0) { 1145 return true; 1146 } 1147 } else { 1148 // We want to conservatively collect somewhat early in order 1149 // to try and "bootstrap" our CMS/promotion statistics; 1150 // this branch will not fire after the first successful CMS 1151 // collection because the stats should then be valid. 1152 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1153 log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1154 _cmsGen->occupancy(), _bootstrap_occupancy); 1155 return true; 1156 } 1157 } 1158 } 1159 1160 // Otherwise, we start a collection cycle if 1161 // old gen want a collection cycle started. Each may use 1162 // an appropriate criterion for making this decision. 1163 // XXX We need to make sure that the gen expansion 1164 // criterion dovetails well with this. XXX NEED TO FIX THIS 1165 if (_cmsGen->should_concurrent_collect()) { 1166 log_trace(gc)("CMS old gen initiated"); 1167 return true; 1168 } 1169 1170 // We start a collection if we believe an incremental collection may fail; 1171 // this is not likely to be productive in practice because it's probably too 1172 // late anyway. 1173 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1174 assert(gch->collector_policy()->is_generation_policy(), 1175 "You may want to check the correctness of the following"); 1176 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1177 log_trace(gc)("CMSCollector: collect because incremental collection will fail "); 1178 return true; 1179 } 1180 1181 if (MetaspaceGC::should_concurrent_collect()) { 1182 log_trace(gc)("CMSCollector: collect for metadata allocation "); 1183 return true; 1184 } 1185 1186 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1187 if (CMSTriggerInterval >= 0) { 1188 if (CMSTriggerInterval == 0) { 1189 // Trigger always 1190 return true; 1191 } 1192 1193 // Check the CMS time since begin (we do not check the stats validity 1194 // as we want to be able to trigger the first CMS cycle as well) 1195 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1196 if (stats().valid()) { 1197 log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1198 stats().cms_time_since_begin()); 1199 } else { 1200 log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)"); 1201 } 1202 return true; 1203 } 1204 } 1205 1206 return false; 1207 } 1208 1209 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1210 1211 // Clear _expansion_cause fields of constituent generations 1212 void CMSCollector::clear_expansion_cause() { 1213 _cmsGen->clear_expansion_cause(); 1214 } 1215 1216 // We should be conservative in starting a collection cycle. To 1217 // start too eagerly runs the risk of collecting too often in the 1218 // extreme. To collect too rarely falls back on full collections, 1219 // which works, even if not optimum in terms of concurrent work. 1220 // As a work around for too eagerly collecting, use the flag 1221 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1222 // giving the user an easily understandable way of controlling the 1223 // collections. 1224 // We want to start a new collection cycle if any of the following 1225 // conditions hold: 1226 // . our current occupancy exceeds the configured initiating occupancy 1227 // for this generation, or 1228 // . we recently needed to expand this space and have not, since that 1229 // expansion, done a collection of this generation, or 1230 // . the underlying space believes that it may be a good idea to initiate 1231 // a concurrent collection (this may be based on criteria such as the 1232 // following: the space uses linear allocation and linear allocation is 1233 // going to fail, or there is believed to be excessive fragmentation in 1234 // the generation, etc... or ... 1235 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1236 // the case of the old generation; see CR 6543076): 1237 // we may be approaching a point at which allocation requests may fail because 1238 // we will be out of sufficient free space given allocation rate estimates.] 1239 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1240 1241 assert_lock_strong(freelistLock()); 1242 if (occupancy() > initiating_occupancy()) { 1243 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1244 short_name(), occupancy(), initiating_occupancy()); 1245 return true; 1246 } 1247 if (UseCMSInitiatingOccupancyOnly) { 1248 return false; 1249 } 1250 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1251 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1252 return true; 1253 } 1254 return false; 1255 } 1256 1257 void ConcurrentMarkSweepGeneration::collect(bool full, 1258 bool clear_all_soft_refs, 1259 size_t size, 1260 bool tlab) 1261 { 1262 collector()->collect(full, clear_all_soft_refs, size, tlab); 1263 } 1264 1265 void CMSCollector::collect(bool full, 1266 bool clear_all_soft_refs, 1267 size_t size, 1268 bool tlab) 1269 { 1270 // The following "if" branch is present for defensive reasons. 1271 // In the current uses of this interface, it can be replaced with: 1272 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1273 // But I am not placing that assert here to allow future 1274 // generality in invoking this interface. 1275 if (GCLocker::is_active()) { 1276 // A consistency test for GCLocker 1277 assert(GCLocker::needs_gc(), "Should have been set already"); 1278 // Skip this foreground collection, instead 1279 // expanding the heap if necessary. 1280 // Need the free list locks for the call to free() in compute_new_size() 1281 compute_new_size(); 1282 return; 1283 } 1284 acquire_control_and_collect(full, clear_all_soft_refs); 1285 } 1286 1287 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1288 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1289 unsigned int gc_count = gch->total_full_collections(); 1290 if (gc_count == full_gc_count) { 1291 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1292 _full_gc_requested = true; 1293 _full_gc_cause = cause; 1294 CGC_lock->notify(); // nudge CMS thread 1295 } else { 1296 assert(gc_count > full_gc_count, "Error: causal loop"); 1297 } 1298 } 1299 1300 bool CMSCollector::is_external_interruption() { 1301 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1302 return GCCause::is_user_requested_gc(cause) || 1303 GCCause::is_serviceability_requested_gc(cause); 1304 } 1305 1306 void CMSCollector::report_concurrent_mode_interruption() { 1307 if (is_external_interruption()) { 1308 log_debug(gc)("Concurrent mode interrupted"); 1309 } else { 1310 log_debug(gc)("Concurrent mode failure"); 1311 _gc_tracer_cm->report_concurrent_mode_failure(); 1312 } 1313 } 1314 1315 1316 // The foreground and background collectors need to coordinate in order 1317 // to make sure that they do not mutually interfere with CMS collections. 1318 // When a background collection is active, 1319 // the foreground collector may need to take over (preempt) and 1320 // synchronously complete an ongoing collection. Depending on the 1321 // frequency of the background collections and the heap usage 1322 // of the application, this preemption can be seldom or frequent. 1323 // There are only certain 1324 // points in the background collection that the "collection-baton" 1325 // can be passed to the foreground collector. 1326 // 1327 // The foreground collector will wait for the baton before 1328 // starting any part of the collection. The foreground collector 1329 // will only wait at one location. 1330 // 1331 // The background collector will yield the baton before starting a new 1332 // phase of the collection (e.g., before initial marking, marking from roots, 1333 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1334 // of the loop which switches the phases. The background collector does some 1335 // of the phases (initial mark, final re-mark) with the world stopped. 1336 // Because of locking involved in stopping the world, 1337 // the foreground collector should not block waiting for the background 1338 // collector when it is doing a stop-the-world phase. The background 1339 // collector will yield the baton at an additional point just before 1340 // it enters a stop-the-world phase. Once the world is stopped, the 1341 // background collector checks the phase of the collection. If the 1342 // phase has not changed, it proceeds with the collection. If the 1343 // phase has changed, it skips that phase of the collection. See 1344 // the comments on the use of the Heap_lock in collect_in_background(). 1345 // 1346 // Variable used in baton passing. 1347 // _foregroundGCIsActive - Set to true by the foreground collector when 1348 // it wants the baton. The foreground clears it when it has finished 1349 // the collection. 1350 // _foregroundGCShouldWait - Set to true by the background collector 1351 // when it is running. The foreground collector waits while 1352 // _foregroundGCShouldWait is true. 1353 // CGC_lock - monitor used to protect access to the above variables 1354 // and to notify the foreground and background collectors. 1355 // _collectorState - current state of the CMS collection. 1356 // 1357 // The foreground collector 1358 // acquires the CGC_lock 1359 // sets _foregroundGCIsActive 1360 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1361 // various locks acquired in preparation for the collection 1362 // are released so as not to block the background collector 1363 // that is in the midst of a collection 1364 // proceeds with the collection 1365 // clears _foregroundGCIsActive 1366 // returns 1367 // 1368 // The background collector in a loop iterating on the phases of the 1369 // collection 1370 // acquires the CGC_lock 1371 // sets _foregroundGCShouldWait 1372 // if _foregroundGCIsActive is set 1373 // clears _foregroundGCShouldWait, notifies _CGC_lock 1374 // waits on _CGC_lock for _foregroundGCIsActive to become false 1375 // and exits the loop. 1376 // otherwise 1377 // proceed with that phase of the collection 1378 // if the phase is a stop-the-world phase, 1379 // yield the baton once more just before enqueueing 1380 // the stop-world CMS operation (executed by the VM thread). 1381 // returns after all phases of the collection are done 1382 // 1383 1384 void CMSCollector::acquire_control_and_collect(bool full, 1385 bool clear_all_soft_refs) { 1386 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1387 assert(!Thread::current()->is_ConcurrentGC_thread(), 1388 "shouldn't try to acquire control from self!"); 1389 1390 // Start the protocol for acquiring control of the 1391 // collection from the background collector (aka CMS thread). 1392 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1393 "VM thread should have CMS token"); 1394 // Remember the possibly interrupted state of an ongoing 1395 // concurrent collection 1396 CollectorState first_state = _collectorState; 1397 1398 // Signal to a possibly ongoing concurrent collection that 1399 // we want to do a foreground collection. 1400 _foregroundGCIsActive = true; 1401 1402 // release locks and wait for a notify from the background collector 1403 // releasing the locks in only necessary for phases which 1404 // do yields to improve the granularity of the collection. 1405 assert_lock_strong(bitMapLock()); 1406 // We need to lock the Free list lock for the space that we are 1407 // currently collecting. 1408 assert(haveFreelistLocks(), "Must be holding free list locks"); 1409 bitMapLock()->unlock(); 1410 releaseFreelistLocks(); 1411 { 1412 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1413 if (_foregroundGCShouldWait) { 1414 // We are going to be waiting for action for the CMS thread; 1415 // it had better not be gone (for instance at shutdown)! 1416 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1417 "CMS thread must be running"); 1418 // Wait here until the background collector gives us the go-ahead 1419 ConcurrentMarkSweepThread::clear_CMS_flag( 1420 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1421 // Get a possibly blocked CMS thread going: 1422 // Note that we set _foregroundGCIsActive true above, 1423 // without protection of the CGC_lock. 1424 CGC_lock->notify(); 1425 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1426 "Possible deadlock"); 1427 while (_foregroundGCShouldWait) { 1428 // wait for notification 1429 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1430 // Possibility of delay/starvation here, since CMS token does 1431 // not know to give priority to VM thread? Actually, i think 1432 // there wouldn't be any delay/starvation, but the proof of 1433 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1434 } 1435 ConcurrentMarkSweepThread::set_CMS_flag( 1436 ConcurrentMarkSweepThread::CMS_vm_has_token); 1437 } 1438 } 1439 // The CMS_token is already held. Get back the other locks. 1440 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1441 "VM thread should have CMS token"); 1442 getFreelistLocks(); 1443 bitMapLock()->lock_without_safepoint_check(); 1444 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1445 p2i(Thread::current()), first_state); 1446 log_debug(gc, state)(" gets control with state %d", _collectorState); 1447 1448 // Inform cms gen if this was due to partial collection failing. 1449 // The CMS gen may use this fact to determine its expansion policy. 1450 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1451 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1452 assert(!_cmsGen->incremental_collection_failed(), 1453 "Should have been noticed, reacted to and cleared"); 1454 _cmsGen->set_incremental_collection_failed(); 1455 } 1456 1457 if (first_state > Idling) { 1458 report_concurrent_mode_interruption(); 1459 } 1460 1461 set_did_compact(true); 1462 1463 // If the collection is being acquired from the background 1464 // collector, there may be references on the discovered 1465 // references lists. Abandon those references, since some 1466 // of them may have become unreachable after concurrent 1467 // discovery; the STW compacting collector will redo discovery 1468 // more precisely, without being subject to floating garbage. 1469 // Leaving otherwise unreachable references in the discovered 1470 // lists would require special handling. 1471 ref_processor()->disable_discovery(); 1472 ref_processor()->abandon_partial_discovery(); 1473 ref_processor()->verify_no_references_recorded(); 1474 1475 if (first_state > Idling) { 1476 save_heap_summary(); 1477 } 1478 1479 do_compaction_work(clear_all_soft_refs); 1480 1481 // Has the GC time limit been exceeded? 1482 size_t max_eden_size = _young_gen->max_eden_size(); 1483 GCCause::Cause gc_cause = gch->gc_cause(); 1484 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1485 _young_gen->eden()->used(), 1486 _cmsGen->max_capacity(), 1487 max_eden_size, 1488 full, 1489 gc_cause, 1490 gch->collector_policy()); 1491 1492 // Reset the expansion cause, now that we just completed 1493 // a collection cycle. 1494 clear_expansion_cause(); 1495 _foregroundGCIsActive = false; 1496 return; 1497 } 1498 1499 // Resize the tenured generation 1500 // after obtaining the free list locks for the 1501 // two generations. 1502 void CMSCollector::compute_new_size() { 1503 assert_locked_or_safepoint(Heap_lock); 1504 FreelistLocker z(this); 1505 MetaspaceGC::compute_new_size(); 1506 _cmsGen->compute_new_size_free_list(); 1507 } 1508 1509 // A work method used by the foreground collector to do 1510 // a mark-sweep-compact. 1511 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1512 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1513 1514 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1515 gc_timer->register_gc_start(); 1516 1517 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1518 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1519 1520 gch->pre_full_gc_dump(gc_timer); 1521 1522 GCTraceTime(Trace, gc) t("CMS:MSC"); 1523 1524 // Temporarily widen the span of the weak reference processing to 1525 // the entire heap. 1526 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1527 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1528 // Temporarily, clear the "is_alive_non_header" field of the 1529 // reference processor. 1530 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1531 // Temporarily make reference _processing_ single threaded (non-MT). 1532 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1533 // Temporarily make refs discovery atomic 1534 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1535 // Temporarily make reference _discovery_ single threaded (non-MT) 1536 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1537 1538 ref_processor()->set_enqueuing_is_done(false); 1539 ref_processor()->enable_discovery(); 1540 ref_processor()->setup_policy(clear_all_soft_refs); 1541 // If an asynchronous collection finishes, the _modUnionTable is 1542 // all clear. If we are assuming the collection from an asynchronous 1543 // collection, clear the _modUnionTable. 1544 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1545 "_modUnionTable should be clear if the baton was not passed"); 1546 _modUnionTable.clear_all(); 1547 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1548 "mod union for klasses should be clear if the baton was passed"); 1549 _ct->klass_rem_set()->clear_mod_union(); 1550 1551 // We must adjust the allocation statistics being maintained 1552 // in the free list space. We do so by reading and clearing 1553 // the sweep timer and updating the block flux rate estimates below. 1554 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1555 if (_inter_sweep_timer.is_active()) { 1556 _inter_sweep_timer.stop(); 1557 // Note that we do not use this sample to update the _inter_sweep_estimate. 1558 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1559 _inter_sweep_estimate.padded_average(), 1560 _intra_sweep_estimate.padded_average()); 1561 } 1562 1563 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1564 #ifdef ASSERT 1565 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1566 size_t free_size = cms_space->free(); 1567 assert(free_size == 1568 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1569 * HeapWordSize, 1570 "All the free space should be compacted into one chunk at top"); 1571 assert(cms_space->dictionary()->total_chunk_size( 1572 debug_only(cms_space->freelistLock())) == 0 || 1573 cms_space->totalSizeInIndexedFreeLists() == 0, 1574 "All the free space should be in a single chunk"); 1575 size_t num = cms_space->totalCount(); 1576 assert((free_size == 0 && num == 0) || 1577 (free_size > 0 && (num == 1 || num == 2)), 1578 "There should be at most 2 free chunks after compaction"); 1579 #endif // ASSERT 1580 _collectorState = Resetting; 1581 assert(_restart_addr == NULL, 1582 "Should have been NULL'd before baton was passed"); 1583 reset_stw(); 1584 _cmsGen->reset_after_compaction(); 1585 _concurrent_cycles_since_last_unload = 0; 1586 1587 // Clear any data recorded in the PLAB chunk arrays. 1588 if (_survivor_plab_array != NULL) { 1589 reset_survivor_plab_arrays(); 1590 } 1591 1592 // Adjust the per-size allocation stats for the next epoch. 1593 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1594 // Restart the "inter sweep timer" for the next epoch. 1595 _inter_sweep_timer.reset(); 1596 _inter_sweep_timer.start(); 1597 1598 gch->post_full_gc_dump(gc_timer); 1599 1600 gc_timer->register_gc_end(); 1601 1602 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1603 1604 // For a mark-sweep-compact, compute_new_size() will be called 1605 // in the heap's do_collection() method. 1606 } 1607 1608 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1609 LogHandle(gc, heap) log; 1610 if (!log.is_trace()) { 1611 return; 1612 } 1613 1614 ContiguousSpace* eden_space = _young_gen->eden(); 1615 ContiguousSpace* from_space = _young_gen->from(); 1616 ContiguousSpace* to_space = _young_gen->to(); 1617 // Eden 1618 if (_eden_chunk_array != NULL) { 1619 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1620 p2i(eden_space->bottom()), p2i(eden_space->top()), 1621 p2i(eden_space->end()), eden_space->capacity()); 1622 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1623 _eden_chunk_index, _eden_chunk_capacity); 1624 for (size_t i = 0; i < _eden_chunk_index; i++) { 1625 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1626 } 1627 } 1628 // Survivor 1629 if (_survivor_chunk_array != NULL) { 1630 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1631 p2i(from_space->bottom()), p2i(from_space->top()), 1632 p2i(from_space->end()), from_space->capacity()); 1633 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1634 _survivor_chunk_index, _survivor_chunk_capacity); 1635 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1636 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1637 } 1638 } 1639 } 1640 1641 void CMSCollector::getFreelistLocks() const { 1642 // Get locks for all free lists in all generations that this 1643 // collector is responsible for 1644 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1645 } 1646 1647 void CMSCollector::releaseFreelistLocks() const { 1648 // Release locks for all free lists in all generations that this 1649 // collector is responsible for 1650 _cmsGen->freelistLock()->unlock(); 1651 } 1652 1653 bool CMSCollector::haveFreelistLocks() const { 1654 // Check locks for all free lists in all generations that this 1655 // collector is responsible for 1656 assert_lock_strong(_cmsGen->freelistLock()); 1657 PRODUCT_ONLY(ShouldNotReachHere()); 1658 return true; 1659 } 1660 1661 // A utility class that is used by the CMS collector to 1662 // temporarily "release" the foreground collector from its 1663 // usual obligation to wait for the background collector to 1664 // complete an ongoing phase before proceeding. 1665 class ReleaseForegroundGC: public StackObj { 1666 private: 1667 CMSCollector* _c; 1668 public: 1669 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1670 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1671 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1672 // allow a potentially blocked foreground collector to proceed 1673 _c->_foregroundGCShouldWait = false; 1674 if (_c->_foregroundGCIsActive) { 1675 CGC_lock->notify(); 1676 } 1677 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1678 "Possible deadlock"); 1679 } 1680 1681 ~ReleaseForegroundGC() { 1682 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1683 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1684 _c->_foregroundGCShouldWait = true; 1685 } 1686 }; 1687 1688 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1689 assert(Thread::current()->is_ConcurrentGC_thread(), 1690 "A CMS asynchronous collection is only allowed on a CMS thread."); 1691 1692 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1693 { 1694 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1695 MutexLockerEx hl(Heap_lock, safepoint_check); 1696 FreelistLocker fll(this); 1697 MutexLockerEx x(CGC_lock, safepoint_check); 1698 if (_foregroundGCIsActive) { 1699 // The foreground collector is. Skip this 1700 // background collection. 1701 assert(!_foregroundGCShouldWait, "Should be clear"); 1702 return; 1703 } else { 1704 assert(_collectorState == Idling, "Should be idling before start."); 1705 _collectorState = InitialMarking; 1706 register_gc_start(cause); 1707 // Reset the expansion cause, now that we are about to begin 1708 // a new cycle. 1709 clear_expansion_cause(); 1710 1711 // Clear the MetaspaceGC flag since a concurrent collection 1712 // is starting but also clear it after the collection. 1713 MetaspaceGC::set_should_concurrent_collect(false); 1714 } 1715 // Decide if we want to enable class unloading as part of the 1716 // ensuing concurrent GC cycle. 1717 update_should_unload_classes(); 1718 _full_gc_requested = false; // acks all outstanding full gc requests 1719 _full_gc_cause = GCCause::_no_gc; 1720 // Signal that we are about to start a collection 1721 gch->increment_total_full_collections(); // ... starting a collection cycle 1722 _collection_count_start = gch->total_full_collections(); 1723 } 1724 1725 size_t prev_used = _cmsGen->used(); 1726 1727 // The change of the collection state is normally done at this level; 1728 // the exceptions are phases that are executed while the world is 1729 // stopped. For those phases the change of state is done while the 1730 // world is stopped. For baton passing purposes this allows the 1731 // background collector to finish the phase and change state atomically. 1732 // The foreground collector cannot wait on a phase that is done 1733 // while the world is stopped because the foreground collector already 1734 // has the world stopped and would deadlock. 1735 while (_collectorState != Idling) { 1736 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1737 p2i(Thread::current()), _collectorState); 1738 // The foreground collector 1739 // holds the Heap_lock throughout its collection. 1740 // holds the CMS token (but not the lock) 1741 // except while it is waiting for the background collector to yield. 1742 // 1743 // The foreground collector should be blocked (not for long) 1744 // if the background collector is about to start a phase 1745 // executed with world stopped. If the background 1746 // collector has already started such a phase, the 1747 // foreground collector is blocked waiting for the 1748 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1749 // are executed in the VM thread. 1750 // 1751 // The locking order is 1752 // PendingListLock (PLL) -- if applicable (FinalMarking) 1753 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1754 // CMS token (claimed in 1755 // stop_world_and_do() --> 1756 // safepoint_synchronize() --> 1757 // CMSThread::synchronize()) 1758 1759 { 1760 // Check if the FG collector wants us to yield. 1761 CMSTokenSync x(true); // is cms thread 1762 if (waitForForegroundGC()) { 1763 // We yielded to a foreground GC, nothing more to be 1764 // done this round. 1765 assert(_foregroundGCShouldWait == false, "We set it to false in " 1766 "waitForForegroundGC()"); 1767 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1768 p2i(Thread::current()), _collectorState); 1769 return; 1770 } else { 1771 // The background collector can run but check to see if the 1772 // foreground collector has done a collection while the 1773 // background collector was waiting to get the CGC_lock 1774 // above. If yes, break so that _foregroundGCShouldWait 1775 // is cleared before returning. 1776 if (_collectorState == Idling) { 1777 break; 1778 } 1779 } 1780 } 1781 1782 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1783 "should be waiting"); 1784 1785 switch (_collectorState) { 1786 case InitialMarking: 1787 { 1788 ReleaseForegroundGC x(this); 1789 stats().record_cms_begin(); 1790 VM_CMS_Initial_Mark initial_mark_op(this); 1791 VMThread::execute(&initial_mark_op); 1792 } 1793 // The collector state may be any legal state at this point 1794 // since the background collector may have yielded to the 1795 // foreground collector. 1796 break; 1797 case Marking: 1798 // initial marking in checkpointRootsInitialWork has been completed 1799 if (markFromRoots()) { // we were successful 1800 assert(_collectorState == Precleaning, "Collector state should " 1801 "have changed"); 1802 } else { 1803 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1804 } 1805 break; 1806 case Precleaning: 1807 // marking from roots in markFromRoots has been completed 1808 preclean(); 1809 assert(_collectorState == AbortablePreclean || 1810 _collectorState == FinalMarking, 1811 "Collector state should have changed"); 1812 break; 1813 case AbortablePreclean: 1814 abortable_preclean(); 1815 assert(_collectorState == FinalMarking, "Collector state should " 1816 "have changed"); 1817 break; 1818 case FinalMarking: 1819 { 1820 ReleaseForegroundGC x(this); 1821 1822 VM_CMS_Final_Remark final_remark_op(this); 1823 VMThread::execute(&final_remark_op); 1824 } 1825 assert(_foregroundGCShouldWait, "block post-condition"); 1826 break; 1827 case Sweeping: 1828 // final marking in checkpointRootsFinal has been completed 1829 sweep(); 1830 assert(_collectorState == Resizing, "Collector state change " 1831 "to Resizing must be done under the free_list_lock"); 1832 1833 case Resizing: { 1834 // Sweeping has been completed... 1835 // At this point the background collection has completed. 1836 // Don't move the call to compute_new_size() down 1837 // into code that might be executed if the background 1838 // collection was preempted. 1839 { 1840 ReleaseForegroundGC x(this); // unblock FG collection 1841 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1842 CMSTokenSync z(true); // not strictly needed. 1843 if (_collectorState == Resizing) { 1844 compute_new_size(); 1845 save_heap_summary(); 1846 _collectorState = Resetting; 1847 } else { 1848 assert(_collectorState == Idling, "The state should only change" 1849 " because the foreground collector has finished the collection"); 1850 } 1851 } 1852 break; 1853 } 1854 case Resetting: 1855 // CMS heap resizing has been completed 1856 reset_concurrent(); 1857 assert(_collectorState == Idling, "Collector state should " 1858 "have changed"); 1859 1860 MetaspaceGC::set_should_concurrent_collect(false); 1861 1862 stats().record_cms_end(); 1863 // Don't move the concurrent_phases_end() and compute_new_size() 1864 // calls to here because a preempted background collection 1865 // has it's state set to "Resetting". 1866 break; 1867 case Idling: 1868 default: 1869 ShouldNotReachHere(); 1870 break; 1871 } 1872 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1873 p2i(Thread::current()), _collectorState); 1874 assert(_foregroundGCShouldWait, "block post-condition"); 1875 } 1876 1877 // Should this be in gc_epilogue? 1878 collector_policy()->counters()->update_counters(); 1879 1880 { 1881 // Clear _foregroundGCShouldWait and, in the event that the 1882 // foreground collector is waiting, notify it, before 1883 // returning. 1884 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1885 _foregroundGCShouldWait = false; 1886 if (_foregroundGCIsActive) { 1887 CGC_lock->notify(); 1888 } 1889 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1890 "Possible deadlock"); 1891 } 1892 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1893 p2i(Thread::current()), _collectorState); 1894 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1895 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1896 } 1897 1898 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1899 _cms_start_registered = true; 1900 _gc_timer_cm->register_gc_start(); 1901 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1902 } 1903 1904 void CMSCollector::register_gc_end() { 1905 if (_cms_start_registered) { 1906 report_heap_summary(GCWhen::AfterGC); 1907 1908 _gc_timer_cm->register_gc_end(); 1909 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1910 _cms_start_registered = false; 1911 } 1912 } 1913 1914 void CMSCollector::save_heap_summary() { 1915 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1916 _last_heap_summary = gch->create_heap_summary(); 1917 _last_metaspace_summary = gch->create_metaspace_summary(); 1918 } 1919 1920 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1921 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1922 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1923 } 1924 1925 bool CMSCollector::waitForForegroundGC() { 1926 bool res = false; 1927 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1928 "CMS thread should have CMS token"); 1929 // Block the foreground collector until the 1930 // background collectors decides whether to 1931 // yield. 1932 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1933 _foregroundGCShouldWait = true; 1934 if (_foregroundGCIsActive) { 1935 // The background collector yields to the 1936 // foreground collector and returns a value 1937 // indicating that it has yielded. The foreground 1938 // collector can proceed. 1939 res = true; 1940 _foregroundGCShouldWait = false; 1941 ConcurrentMarkSweepThread::clear_CMS_flag( 1942 ConcurrentMarkSweepThread::CMS_cms_has_token); 1943 ConcurrentMarkSweepThread::set_CMS_flag( 1944 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1945 // Get a possibly blocked foreground thread going 1946 CGC_lock->notify(); 1947 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1948 p2i(Thread::current()), _collectorState); 1949 while (_foregroundGCIsActive) { 1950 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1951 } 1952 ConcurrentMarkSweepThread::set_CMS_flag( 1953 ConcurrentMarkSweepThread::CMS_cms_has_token); 1954 ConcurrentMarkSweepThread::clear_CMS_flag( 1955 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1956 } 1957 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1958 p2i(Thread::current()), _collectorState); 1959 return res; 1960 } 1961 1962 // Because of the need to lock the free lists and other structures in 1963 // the collector, common to all the generations that the collector is 1964 // collecting, we need the gc_prologues of individual CMS generations 1965 // delegate to their collector. It may have been simpler had the 1966 // current infrastructure allowed one to call a prologue on a 1967 // collector. In the absence of that we have the generation's 1968 // prologue delegate to the collector, which delegates back 1969 // some "local" work to a worker method in the individual generations 1970 // that it's responsible for collecting, while itself doing any 1971 // work common to all generations it's responsible for. A similar 1972 // comment applies to the gc_epilogue()'s. 1973 // The role of the variable _between_prologue_and_epilogue is to 1974 // enforce the invocation protocol. 1975 void CMSCollector::gc_prologue(bool full) { 1976 // Call gc_prologue_work() for the CMSGen 1977 // we are responsible for. 1978 1979 // The following locking discipline assumes that we are only called 1980 // when the world is stopped. 1981 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1982 1983 // The CMSCollector prologue must call the gc_prologues for the 1984 // "generations" that it's responsible 1985 // for. 1986 1987 assert( Thread::current()->is_VM_thread() 1988 || ( CMSScavengeBeforeRemark 1989 && Thread::current()->is_ConcurrentGC_thread()), 1990 "Incorrect thread type for prologue execution"); 1991 1992 if (_between_prologue_and_epilogue) { 1993 // We have already been invoked; this is a gc_prologue delegation 1994 // from yet another CMS generation that we are responsible for, just 1995 // ignore it since all relevant work has already been done. 1996 return; 1997 } 1998 1999 // set a bit saying prologue has been called; cleared in epilogue 2000 _between_prologue_and_epilogue = true; 2001 // Claim locks for common data structures, then call gc_prologue_work() 2002 // for each CMSGen. 2003 2004 getFreelistLocks(); // gets free list locks on constituent spaces 2005 bitMapLock()->lock_without_safepoint_check(); 2006 2007 // Should call gc_prologue_work() for all cms gens we are responsible for 2008 bool duringMarking = _collectorState >= Marking 2009 && _collectorState < Sweeping; 2010 2011 // The young collections clear the modified oops state, which tells if 2012 // there are any modified oops in the class. The remark phase also needs 2013 // that information. Tell the young collection to save the union of all 2014 // modified klasses. 2015 if (duringMarking) { 2016 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2017 } 2018 2019 bool registerClosure = duringMarking; 2020 2021 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2022 2023 if (!full) { 2024 stats().record_gc0_begin(); 2025 } 2026 } 2027 2028 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2029 2030 _capacity_at_prologue = capacity(); 2031 _used_at_prologue = used(); 2032 2033 // Delegate to CMScollector which knows how to coordinate between 2034 // this and any other CMS generations that it is responsible for 2035 // collecting. 2036 collector()->gc_prologue(full); 2037 } 2038 2039 // This is a "private" interface for use by this generation's CMSCollector. 2040 // Not to be called directly by any other entity (for instance, 2041 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2042 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2043 bool registerClosure, ModUnionClosure* modUnionClosure) { 2044 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2045 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2046 "Should be NULL"); 2047 if (registerClosure) { 2048 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2049 } 2050 cmsSpace()->gc_prologue(); 2051 // Clear stat counters 2052 NOT_PRODUCT( 2053 assert(_numObjectsPromoted == 0, "check"); 2054 assert(_numWordsPromoted == 0, "check"); 2055 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2056 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2057 _numObjectsAllocated = 0; 2058 _numWordsAllocated = 0; 2059 ) 2060 } 2061 2062 void CMSCollector::gc_epilogue(bool full) { 2063 // The following locking discipline assumes that we are only called 2064 // when the world is stopped. 2065 assert(SafepointSynchronize::is_at_safepoint(), 2066 "world is stopped assumption"); 2067 2068 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2069 // if linear allocation blocks need to be appropriately marked to allow the 2070 // the blocks to be parsable. We also check here whether we need to nudge the 2071 // CMS collector thread to start a new cycle (if it's not already active). 2072 assert( Thread::current()->is_VM_thread() 2073 || ( CMSScavengeBeforeRemark 2074 && Thread::current()->is_ConcurrentGC_thread()), 2075 "Incorrect thread type for epilogue execution"); 2076 2077 if (!_between_prologue_and_epilogue) { 2078 // We have already been invoked; this is a gc_epilogue delegation 2079 // from yet another CMS generation that we are responsible for, just 2080 // ignore it since all relevant work has already been done. 2081 return; 2082 } 2083 assert(haveFreelistLocks(), "must have freelist locks"); 2084 assert_lock_strong(bitMapLock()); 2085 2086 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2087 2088 _cmsGen->gc_epilogue_work(full); 2089 2090 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2091 // in case sampling was not already enabled, enable it 2092 _start_sampling = true; 2093 } 2094 // reset _eden_chunk_array so sampling starts afresh 2095 _eden_chunk_index = 0; 2096 2097 size_t cms_used = _cmsGen->cmsSpace()->used(); 2098 2099 // update performance counters - this uses a special version of 2100 // update_counters() that allows the utilization to be passed as a 2101 // parameter, avoiding multiple calls to used(). 2102 // 2103 _cmsGen->update_counters(cms_used); 2104 2105 bitMapLock()->unlock(); 2106 releaseFreelistLocks(); 2107 2108 if (!CleanChunkPoolAsync) { 2109 Chunk::clean_chunk_pool(); 2110 } 2111 2112 set_did_compact(false); 2113 _between_prologue_and_epilogue = false; // ready for next cycle 2114 } 2115 2116 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2117 collector()->gc_epilogue(full); 2118 2119 // Also reset promotion tracking in par gc thread states. 2120 for (uint i = 0; i < ParallelGCThreads; i++) { 2121 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2122 } 2123 } 2124 2125 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2126 assert(!incremental_collection_failed(), "Should have been cleared"); 2127 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2128 cmsSpace()->gc_epilogue(); 2129 // Print stat counters 2130 NOT_PRODUCT( 2131 assert(_numObjectsAllocated == 0, "check"); 2132 assert(_numWordsAllocated == 0, "check"); 2133 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2134 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2135 _numObjectsPromoted = 0; 2136 _numWordsPromoted = 0; 2137 ) 2138 2139 // Call down the chain in contiguous_available needs the freelistLock 2140 // so print this out before releasing the freeListLock. 2141 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2142 } 2143 2144 #ifndef PRODUCT 2145 bool CMSCollector::have_cms_token() { 2146 Thread* thr = Thread::current(); 2147 if (thr->is_VM_thread()) { 2148 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2149 } else if (thr->is_ConcurrentGC_thread()) { 2150 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2151 } else if (thr->is_GC_task_thread()) { 2152 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2153 ParGCRareEvent_lock->owned_by_self(); 2154 } 2155 return false; 2156 } 2157 2158 // Check reachability of the given heap address in CMS generation, 2159 // treating all other generations as roots. 2160 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2161 // We could "guarantee" below, rather than assert, but I'll 2162 // leave these as "asserts" so that an adventurous debugger 2163 // could try this in the product build provided some subset of 2164 // the conditions were met, provided they were interested in the 2165 // results and knew that the computation below wouldn't interfere 2166 // with other concurrent computations mutating the structures 2167 // being read or written. 2168 assert(SafepointSynchronize::is_at_safepoint(), 2169 "Else mutations in object graph will make answer suspect"); 2170 assert(have_cms_token(), "Should hold cms token"); 2171 assert(haveFreelistLocks(), "must hold free list locks"); 2172 assert_lock_strong(bitMapLock()); 2173 2174 // Clear the marking bit map array before starting, but, just 2175 // for kicks, first report if the given address is already marked 2176 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2177 _markBitMap.isMarked(addr) ? "" : " not"); 2178 2179 if (verify_after_remark()) { 2180 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2181 bool result = verification_mark_bm()->isMarked(addr); 2182 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2183 result ? "IS" : "is NOT"); 2184 return result; 2185 } else { 2186 tty->print_cr("Could not compute result"); 2187 return false; 2188 } 2189 } 2190 #endif 2191 2192 void 2193 CMSCollector::print_on_error(outputStream* st) { 2194 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2195 if (collector != NULL) { 2196 CMSBitMap* bitmap = &collector->_markBitMap; 2197 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2198 bitmap->print_on_error(st, " Bits: "); 2199 2200 st->cr(); 2201 2202 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2203 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2204 mut_bitmap->print_on_error(st, " Bits: "); 2205 } 2206 } 2207 2208 //////////////////////////////////////////////////////// 2209 // CMS Verification Support 2210 //////////////////////////////////////////////////////// 2211 // Following the remark phase, the following invariant 2212 // should hold -- each object in the CMS heap which is 2213 // marked in markBitMap() should be marked in the verification_mark_bm(). 2214 2215 class VerifyMarkedClosure: public BitMapClosure { 2216 CMSBitMap* _marks; 2217 bool _failed; 2218 2219 public: 2220 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2221 2222 bool do_bit(size_t offset) { 2223 HeapWord* addr = _marks->offsetToHeapWord(offset); 2224 if (!_marks->isMarked(addr)) { 2225 LogHandle(gc, verify) log; 2226 ResourceMark rm; 2227 oop(addr)->print_on(log.error_stream()); 2228 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2229 _failed = true; 2230 } 2231 return true; 2232 } 2233 2234 bool failed() { return _failed; } 2235 }; 2236 2237 bool CMSCollector::verify_after_remark() { 2238 GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking."); 2239 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2240 static bool init = false; 2241 2242 assert(SafepointSynchronize::is_at_safepoint(), 2243 "Else mutations in object graph will make answer suspect"); 2244 assert(have_cms_token(), 2245 "Else there may be mutual interference in use of " 2246 " verification data structures"); 2247 assert(_collectorState > Marking && _collectorState <= Sweeping, 2248 "Else marking info checked here may be obsolete"); 2249 assert(haveFreelistLocks(), "must hold free list locks"); 2250 assert_lock_strong(bitMapLock()); 2251 2252 2253 // Allocate marking bit map if not already allocated 2254 if (!init) { // first time 2255 if (!verification_mark_bm()->allocate(_span)) { 2256 return false; 2257 } 2258 init = true; 2259 } 2260 2261 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2262 2263 // Turn off refs discovery -- so we will be tracing through refs. 2264 // This is as intended, because by this time 2265 // GC must already have cleared any refs that need to be cleared, 2266 // and traced those that need to be marked; moreover, 2267 // the marking done here is not going to interfere in any 2268 // way with the marking information used by GC. 2269 NoRefDiscovery no_discovery(ref_processor()); 2270 2271 #if defined(COMPILER2) || INCLUDE_JVMCI 2272 DerivedPointerTableDeactivate dpt_deact; 2273 #endif 2274 2275 // Clear any marks from a previous round 2276 verification_mark_bm()->clear_all(); 2277 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2278 verify_work_stacks_empty(); 2279 2280 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2281 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2282 // Update the saved marks which may affect the root scans. 2283 gch->save_marks(); 2284 2285 if (CMSRemarkVerifyVariant == 1) { 2286 // In this first variant of verification, we complete 2287 // all marking, then check if the new marks-vector is 2288 // a subset of the CMS marks-vector. 2289 verify_after_remark_work_1(); 2290 } else { 2291 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2292 // In this second variant of verification, we flag an error 2293 // (i.e. an object reachable in the new marks-vector not reachable 2294 // in the CMS marks-vector) immediately, also indicating the 2295 // identify of an object (A) that references the unmarked object (B) -- 2296 // presumably, a mutation to A failed to be picked up by preclean/remark? 2297 verify_after_remark_work_2(); 2298 } 2299 2300 return true; 2301 } 2302 2303 void CMSCollector::verify_after_remark_work_1() { 2304 ResourceMark rm; 2305 HandleMark hm; 2306 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2307 2308 // Get a clear set of claim bits for the roots processing to work with. 2309 ClassLoaderDataGraph::clear_claimed_marks(); 2310 2311 // Mark from roots one level into CMS 2312 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2313 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2314 2315 { 2316 StrongRootsScope srs(1); 2317 2318 gch->gen_process_roots(&srs, 2319 GenCollectedHeap::OldGen, 2320 true, // young gen as roots 2321 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2322 should_unload_classes(), 2323 ¬Older, 2324 NULL, 2325 NULL); 2326 } 2327 2328 // Now mark from the roots 2329 MarkFromRootsClosure markFromRootsClosure(this, _span, 2330 verification_mark_bm(), verification_mark_stack(), 2331 false /* don't yield */, true /* verifying */); 2332 assert(_restart_addr == NULL, "Expected pre-condition"); 2333 verification_mark_bm()->iterate(&markFromRootsClosure); 2334 while (_restart_addr != NULL) { 2335 // Deal with stack overflow: by restarting at the indicated 2336 // address. 2337 HeapWord* ra = _restart_addr; 2338 markFromRootsClosure.reset(ra); 2339 _restart_addr = NULL; 2340 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2341 } 2342 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2343 verify_work_stacks_empty(); 2344 2345 // Marking completed -- now verify that each bit marked in 2346 // verification_mark_bm() is also marked in markBitMap(); flag all 2347 // errors by printing corresponding objects. 2348 VerifyMarkedClosure vcl(markBitMap()); 2349 verification_mark_bm()->iterate(&vcl); 2350 if (vcl.failed()) { 2351 LogHandle(gc, verify) log; 2352 log.error("Failed marking verification after remark"); 2353 ResourceMark rm; 2354 gch->print_on(log.error_stream()); 2355 fatal("CMS: failed marking verification after remark"); 2356 } 2357 } 2358 2359 class VerifyKlassOopsKlassClosure : public KlassClosure { 2360 class VerifyKlassOopsClosure : public OopClosure { 2361 CMSBitMap* _bitmap; 2362 public: 2363 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2364 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2365 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2366 } _oop_closure; 2367 public: 2368 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2369 void do_klass(Klass* k) { 2370 k->oops_do(&_oop_closure); 2371 } 2372 }; 2373 2374 void CMSCollector::verify_after_remark_work_2() { 2375 ResourceMark rm; 2376 HandleMark hm; 2377 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2378 2379 // Get a clear set of claim bits for the roots processing to work with. 2380 ClassLoaderDataGraph::clear_claimed_marks(); 2381 2382 // Mark from roots one level into CMS 2383 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2384 markBitMap()); 2385 CLDToOopClosure cld_closure(¬Older, true); 2386 2387 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2388 2389 { 2390 StrongRootsScope srs(1); 2391 2392 gch->gen_process_roots(&srs, 2393 GenCollectedHeap::OldGen, 2394 true, // young gen as roots 2395 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2396 should_unload_classes(), 2397 ¬Older, 2398 NULL, 2399 &cld_closure); 2400 } 2401 2402 // Now mark from the roots 2403 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2404 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2405 assert(_restart_addr == NULL, "Expected pre-condition"); 2406 verification_mark_bm()->iterate(&markFromRootsClosure); 2407 while (_restart_addr != NULL) { 2408 // Deal with stack overflow: by restarting at the indicated 2409 // address. 2410 HeapWord* ra = _restart_addr; 2411 markFromRootsClosure.reset(ra); 2412 _restart_addr = NULL; 2413 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2414 } 2415 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2416 verify_work_stacks_empty(); 2417 2418 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2419 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2420 2421 // Marking completed -- now verify that each bit marked in 2422 // verification_mark_bm() is also marked in markBitMap(); flag all 2423 // errors by printing corresponding objects. 2424 VerifyMarkedClosure vcl(markBitMap()); 2425 verification_mark_bm()->iterate(&vcl); 2426 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2427 } 2428 2429 void ConcurrentMarkSweepGeneration::save_marks() { 2430 // delegate to CMS space 2431 cmsSpace()->save_marks(); 2432 for (uint i = 0; i < ParallelGCThreads; i++) { 2433 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2434 } 2435 } 2436 2437 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2438 return cmsSpace()->no_allocs_since_save_marks(); 2439 } 2440 2441 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2442 \ 2443 void ConcurrentMarkSweepGeneration:: \ 2444 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2445 cl->set_generation(this); \ 2446 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2447 cl->reset_generation(); \ 2448 save_marks(); \ 2449 } 2450 2451 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2452 2453 void 2454 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2455 if (freelistLock()->owned_by_self()) { 2456 Generation::oop_iterate(cl); 2457 } else { 2458 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2459 Generation::oop_iterate(cl); 2460 } 2461 } 2462 2463 void 2464 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2465 if (freelistLock()->owned_by_self()) { 2466 Generation::object_iterate(cl); 2467 } else { 2468 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2469 Generation::object_iterate(cl); 2470 } 2471 } 2472 2473 void 2474 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2475 if (freelistLock()->owned_by_self()) { 2476 Generation::safe_object_iterate(cl); 2477 } else { 2478 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2479 Generation::safe_object_iterate(cl); 2480 } 2481 } 2482 2483 void 2484 ConcurrentMarkSweepGeneration::post_compact() { 2485 } 2486 2487 void 2488 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2489 // Fix the linear allocation blocks to look like free blocks. 2490 2491 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2492 // are not called when the heap is verified during universe initialization and 2493 // at vm shutdown. 2494 if (freelistLock()->owned_by_self()) { 2495 cmsSpace()->prepare_for_verify(); 2496 } else { 2497 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2498 cmsSpace()->prepare_for_verify(); 2499 } 2500 } 2501 2502 void 2503 ConcurrentMarkSweepGeneration::verify() { 2504 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2505 // are not called when the heap is verified during universe initialization and 2506 // at vm shutdown. 2507 if (freelistLock()->owned_by_self()) { 2508 cmsSpace()->verify(); 2509 } else { 2510 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2511 cmsSpace()->verify(); 2512 } 2513 } 2514 2515 void CMSCollector::verify() { 2516 _cmsGen->verify(); 2517 } 2518 2519 #ifndef PRODUCT 2520 bool CMSCollector::overflow_list_is_empty() const { 2521 assert(_num_par_pushes >= 0, "Inconsistency"); 2522 if (_overflow_list == NULL) { 2523 assert(_num_par_pushes == 0, "Inconsistency"); 2524 } 2525 return _overflow_list == NULL; 2526 } 2527 2528 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2529 // merely consolidate assertion checks that appear to occur together frequently. 2530 void CMSCollector::verify_work_stacks_empty() const { 2531 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2532 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2533 } 2534 2535 void CMSCollector::verify_overflow_empty() const { 2536 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2537 assert(no_preserved_marks(), "No preserved marks"); 2538 } 2539 #endif // PRODUCT 2540 2541 // Decide if we want to enable class unloading as part of the 2542 // ensuing concurrent GC cycle. We will collect and 2543 // unload classes if it's the case that: 2544 // (1) an explicit gc request has been made and the flag 2545 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2546 // (2) (a) class unloading is enabled at the command line, and 2547 // (b) old gen is getting really full 2548 // NOTE: Provided there is no change in the state of the heap between 2549 // calls to this method, it should have idempotent results. Moreover, 2550 // its results should be monotonically increasing (i.e. going from 0 to 1, 2551 // but not 1 to 0) between successive calls between which the heap was 2552 // not collected. For the implementation below, it must thus rely on 2553 // the property that concurrent_cycles_since_last_unload() 2554 // will not decrease unless a collection cycle happened and that 2555 // _cmsGen->is_too_full() are 2556 // themselves also monotonic in that sense. See check_monotonicity() 2557 // below. 2558 void CMSCollector::update_should_unload_classes() { 2559 _should_unload_classes = false; 2560 // Condition 1 above 2561 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2562 _should_unload_classes = true; 2563 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2564 // Disjuncts 2.b.(i,ii,iii) above 2565 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2566 CMSClassUnloadingMaxInterval) 2567 || _cmsGen->is_too_full(); 2568 } 2569 } 2570 2571 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2572 bool res = should_concurrent_collect(); 2573 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2574 return res; 2575 } 2576 2577 void CMSCollector::setup_cms_unloading_and_verification_state() { 2578 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2579 || VerifyBeforeExit; 2580 const int rso = GenCollectedHeap::SO_AllCodeCache; 2581 2582 // We set the proper root for this CMS cycle here. 2583 if (should_unload_classes()) { // Should unload classes this cycle 2584 remove_root_scanning_option(rso); // Shrink the root set appropriately 2585 set_verifying(should_verify); // Set verification state for this cycle 2586 return; // Nothing else needs to be done at this time 2587 } 2588 2589 // Not unloading classes this cycle 2590 assert(!should_unload_classes(), "Inconsistency!"); 2591 2592 // If we are not unloading classes then add SO_AllCodeCache to root 2593 // scanning options. 2594 add_root_scanning_option(rso); 2595 2596 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2597 set_verifying(true); 2598 } else if (verifying() && !should_verify) { 2599 // We were verifying, but some verification flags got disabled. 2600 set_verifying(false); 2601 // Exclude symbols, strings and code cache elements from root scanning to 2602 // reduce IM and RM pauses. 2603 remove_root_scanning_option(rso); 2604 } 2605 } 2606 2607 2608 #ifndef PRODUCT 2609 HeapWord* CMSCollector::block_start(const void* p) const { 2610 const HeapWord* addr = (HeapWord*)p; 2611 if (_span.contains(p)) { 2612 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2613 return _cmsGen->cmsSpace()->block_start(p); 2614 } 2615 } 2616 return NULL; 2617 } 2618 #endif 2619 2620 HeapWord* 2621 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2622 bool tlab, 2623 bool parallel) { 2624 CMSSynchronousYieldRequest yr; 2625 assert(!tlab, "Can't deal with TLAB allocation"); 2626 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2627 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2628 if (GCExpandToAllocateDelayMillis > 0) { 2629 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2630 } 2631 return have_lock_and_allocate(word_size, tlab); 2632 } 2633 2634 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2635 size_t bytes, 2636 size_t expand_bytes, 2637 CMSExpansionCause::Cause cause) 2638 { 2639 2640 bool success = expand(bytes, expand_bytes); 2641 2642 // remember why we expanded; this information is used 2643 // by shouldConcurrentCollect() when making decisions on whether to start 2644 // a new CMS cycle. 2645 if (success) { 2646 set_expansion_cause(cause); 2647 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2648 } 2649 } 2650 2651 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2652 HeapWord* res = NULL; 2653 MutexLocker x(ParGCRareEvent_lock); 2654 while (true) { 2655 // Expansion by some other thread might make alloc OK now: 2656 res = ps->lab.alloc(word_sz); 2657 if (res != NULL) return res; 2658 // If there's not enough expansion space available, give up. 2659 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2660 return NULL; 2661 } 2662 // Otherwise, we try expansion. 2663 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2664 // Now go around the loop and try alloc again; 2665 // A competing par_promote might beat us to the expansion space, 2666 // so we may go around the loop again if promotion fails again. 2667 if (GCExpandToAllocateDelayMillis > 0) { 2668 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2669 } 2670 } 2671 } 2672 2673 2674 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2675 PromotionInfo* promo) { 2676 MutexLocker x(ParGCRareEvent_lock); 2677 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2678 while (true) { 2679 // Expansion by some other thread might make alloc OK now: 2680 if (promo->ensure_spooling_space()) { 2681 assert(promo->has_spooling_space(), 2682 "Post-condition of successful ensure_spooling_space()"); 2683 return true; 2684 } 2685 // If there's not enough expansion space available, give up. 2686 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2687 return false; 2688 } 2689 // Otherwise, we try expansion. 2690 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2691 // Now go around the loop and try alloc again; 2692 // A competing allocation might beat us to the expansion space, 2693 // so we may go around the loop again if allocation fails again. 2694 if (GCExpandToAllocateDelayMillis > 0) { 2695 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2696 } 2697 } 2698 } 2699 2700 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2701 // Only shrink if a compaction was done so that all the free space 2702 // in the generation is in a contiguous block at the end. 2703 if (did_compact()) { 2704 CardGeneration::shrink(bytes); 2705 } 2706 } 2707 2708 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2709 assert_locked_or_safepoint(Heap_lock); 2710 } 2711 2712 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2713 assert_locked_or_safepoint(Heap_lock); 2714 assert_lock_strong(freelistLock()); 2715 log_trace(gc)("Shrinking of CMS not yet implemented"); 2716 return; 2717 } 2718 2719 2720 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2721 // phases. 2722 class CMSPhaseAccounting: public StackObj { 2723 public: 2724 CMSPhaseAccounting(CMSCollector *collector, 2725 const char *title); 2726 ~CMSPhaseAccounting(); 2727 2728 private: 2729 CMSCollector *_collector; 2730 const char *_title; 2731 GCTraceConcTime(Info, gc) _trace_time; 2732 2733 public: 2734 // Not MT-safe; so do not pass around these StackObj's 2735 // where they may be accessed by other threads. 2736 double wallclock_millis() { 2737 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2738 } 2739 }; 2740 2741 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2742 const char *title) : 2743 _collector(collector), _title(title), _trace_time(title) { 2744 2745 _collector->resetYields(); 2746 _collector->resetTimer(); 2747 _collector->startTimer(); 2748 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2749 } 2750 2751 CMSPhaseAccounting::~CMSPhaseAccounting() { 2752 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2753 _collector->stopTimer(); 2754 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2755 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2756 } 2757 2758 // CMS work 2759 2760 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2761 class CMSParMarkTask : public AbstractGangTask { 2762 protected: 2763 CMSCollector* _collector; 2764 uint _n_workers; 2765 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2766 AbstractGangTask(name), 2767 _collector(collector), 2768 _n_workers(n_workers) {} 2769 // Work method in support of parallel rescan ... of young gen spaces 2770 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2771 ContiguousSpace* space, 2772 HeapWord** chunk_array, size_t chunk_top); 2773 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2774 }; 2775 2776 // Parallel initial mark task 2777 class CMSParInitialMarkTask: public CMSParMarkTask { 2778 StrongRootsScope* _strong_roots_scope; 2779 public: 2780 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2781 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2782 _strong_roots_scope(strong_roots_scope) {} 2783 void work(uint worker_id); 2784 }; 2785 2786 // Checkpoint the roots into this generation from outside 2787 // this generation. [Note this initial checkpoint need only 2788 // be approximate -- we'll do a catch up phase subsequently.] 2789 void CMSCollector::checkpointRootsInitial() { 2790 assert(_collectorState == InitialMarking, "Wrong collector state"); 2791 check_correct_thread_executing(); 2792 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2793 2794 save_heap_summary(); 2795 report_heap_summary(GCWhen::BeforeGC); 2796 2797 ReferenceProcessor* rp = ref_processor(); 2798 assert(_restart_addr == NULL, "Control point invariant"); 2799 { 2800 // acquire locks for subsequent manipulations 2801 MutexLockerEx x(bitMapLock(), 2802 Mutex::_no_safepoint_check_flag); 2803 checkpointRootsInitialWork(); 2804 // enable ("weak") refs discovery 2805 rp->enable_discovery(); 2806 _collectorState = Marking; 2807 } 2808 } 2809 2810 void CMSCollector::checkpointRootsInitialWork() { 2811 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2812 assert(_collectorState == InitialMarking, "just checking"); 2813 2814 // Already have locks. 2815 assert_lock_strong(bitMapLock()); 2816 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2817 2818 // Setup the verification and class unloading state for this 2819 // CMS collection cycle. 2820 setup_cms_unloading_and_verification_state(); 2821 2822 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2823 2824 // Reset all the PLAB chunk arrays if necessary. 2825 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2826 reset_survivor_plab_arrays(); 2827 } 2828 2829 ResourceMark rm; 2830 HandleMark hm; 2831 2832 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2833 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2834 2835 verify_work_stacks_empty(); 2836 verify_overflow_empty(); 2837 2838 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2839 // Update the saved marks which may affect the root scans. 2840 gch->save_marks(); 2841 2842 // weak reference processing has not started yet. 2843 ref_processor()->set_enqueuing_is_done(false); 2844 2845 // Need to remember all newly created CLDs, 2846 // so that we can guarantee that the remark finds them. 2847 ClassLoaderDataGraph::remember_new_clds(true); 2848 2849 // Whenever a CLD is found, it will be claimed before proceeding to mark 2850 // the klasses. The claimed marks need to be cleared before marking starts. 2851 ClassLoaderDataGraph::clear_claimed_marks(); 2852 2853 print_eden_and_survivor_chunk_arrays(); 2854 2855 { 2856 #if defined(COMPILER2) || INCLUDE_JVMCI 2857 DerivedPointerTableDeactivate dpt_deact; 2858 #endif 2859 if (CMSParallelInitialMarkEnabled) { 2860 // The parallel version. 2861 WorkGang* workers = gch->workers(); 2862 assert(workers != NULL, "Need parallel worker threads."); 2863 uint n_workers = workers->active_workers(); 2864 2865 StrongRootsScope srs(n_workers); 2866 2867 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2868 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2869 if (n_workers > 1) { 2870 workers->run_task(&tsk); 2871 } else { 2872 tsk.work(0); 2873 } 2874 } else { 2875 // The serial version. 2876 CLDToOopClosure cld_closure(¬Older, true); 2877 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2878 2879 StrongRootsScope srs(1); 2880 2881 gch->gen_process_roots(&srs, 2882 GenCollectedHeap::OldGen, 2883 true, // young gen as roots 2884 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2885 should_unload_classes(), 2886 ¬Older, 2887 NULL, 2888 &cld_closure); 2889 } 2890 } 2891 2892 // Clear mod-union table; it will be dirtied in the prologue of 2893 // CMS generation per each young generation collection. 2894 2895 assert(_modUnionTable.isAllClear(), 2896 "Was cleared in most recent final checkpoint phase" 2897 " or no bits are set in the gc_prologue before the start of the next " 2898 "subsequent marking phase."); 2899 2900 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2901 2902 // Save the end of the used_region of the constituent generations 2903 // to be used to limit the extent of sweep in each generation. 2904 save_sweep_limits(); 2905 verify_overflow_empty(); 2906 } 2907 2908 bool CMSCollector::markFromRoots() { 2909 // we might be tempted to assert that: 2910 // assert(!SafepointSynchronize::is_at_safepoint(), 2911 // "inconsistent argument?"); 2912 // However that wouldn't be right, because it's possible that 2913 // a safepoint is indeed in progress as a young generation 2914 // stop-the-world GC happens even as we mark in this generation. 2915 assert(_collectorState == Marking, "inconsistent state?"); 2916 check_correct_thread_executing(); 2917 verify_overflow_empty(); 2918 2919 // Weak ref discovery note: We may be discovering weak 2920 // refs in this generation concurrent (but interleaved) with 2921 // weak ref discovery by the young generation collector. 2922 2923 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2924 GCTraceCPUTime tcpu; 2925 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2926 bool res = markFromRootsWork(); 2927 if (res) { 2928 _collectorState = Precleaning; 2929 } else { // We failed and a foreground collection wants to take over 2930 assert(_foregroundGCIsActive, "internal state inconsistency"); 2931 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2932 log_debug(gc)("bailing out to foreground collection"); 2933 } 2934 verify_overflow_empty(); 2935 return res; 2936 } 2937 2938 bool CMSCollector::markFromRootsWork() { 2939 // iterate over marked bits in bit map, doing a full scan and mark 2940 // from these roots using the following algorithm: 2941 // . if oop is to the right of the current scan pointer, 2942 // mark corresponding bit (we'll process it later) 2943 // . else (oop is to left of current scan pointer) 2944 // push oop on marking stack 2945 // . drain the marking stack 2946 2947 // Note that when we do a marking step we need to hold the 2948 // bit map lock -- recall that direct allocation (by mutators) 2949 // and promotion (by the young generation collector) is also 2950 // marking the bit map. [the so-called allocate live policy.] 2951 // Because the implementation of bit map marking is not 2952 // robust wrt simultaneous marking of bits in the same word, 2953 // we need to make sure that there is no such interference 2954 // between concurrent such updates. 2955 2956 // already have locks 2957 assert_lock_strong(bitMapLock()); 2958 2959 verify_work_stacks_empty(); 2960 verify_overflow_empty(); 2961 bool result = false; 2962 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2963 result = do_marking_mt(); 2964 } else { 2965 result = do_marking_st(); 2966 } 2967 return result; 2968 } 2969 2970 // Forward decl 2971 class CMSConcMarkingTask; 2972 2973 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2974 CMSCollector* _collector; 2975 CMSConcMarkingTask* _task; 2976 public: 2977 virtual void yield(); 2978 2979 // "n_threads" is the number of threads to be terminated. 2980 // "queue_set" is a set of work queues of other threads. 2981 // "collector" is the CMS collector associated with this task terminator. 2982 // "yield" indicates whether we need the gang as a whole to yield. 2983 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2984 ParallelTaskTerminator(n_threads, queue_set), 2985 _collector(collector) { } 2986 2987 void set_task(CMSConcMarkingTask* task) { 2988 _task = task; 2989 } 2990 }; 2991 2992 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2993 CMSConcMarkingTask* _task; 2994 public: 2995 bool should_exit_termination(); 2996 void set_task(CMSConcMarkingTask* task) { 2997 _task = task; 2998 } 2999 }; 3000 3001 // MT Concurrent Marking Task 3002 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3003 CMSCollector* _collector; 3004 uint _n_workers; // requested/desired # workers 3005 bool _result; 3006 CompactibleFreeListSpace* _cms_space; 3007 char _pad_front[64]; // padding to ... 3008 HeapWord* _global_finger; // ... avoid sharing cache line 3009 char _pad_back[64]; 3010 HeapWord* _restart_addr; 3011 3012 // Exposed here for yielding support 3013 Mutex* const _bit_map_lock; 3014 3015 // The per thread work queues, available here for stealing 3016 OopTaskQueueSet* _task_queues; 3017 3018 // Termination (and yielding) support 3019 CMSConcMarkingTerminator _term; 3020 CMSConcMarkingTerminatorTerminator _term_term; 3021 3022 public: 3023 CMSConcMarkingTask(CMSCollector* collector, 3024 CompactibleFreeListSpace* cms_space, 3025 YieldingFlexibleWorkGang* workers, 3026 OopTaskQueueSet* task_queues): 3027 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3028 _collector(collector), 3029 _cms_space(cms_space), 3030 _n_workers(0), _result(true), 3031 _task_queues(task_queues), 3032 _term(_n_workers, task_queues, _collector), 3033 _bit_map_lock(collector->bitMapLock()) 3034 { 3035 _requested_size = _n_workers; 3036 _term.set_task(this); 3037 _term_term.set_task(this); 3038 _restart_addr = _global_finger = _cms_space->bottom(); 3039 } 3040 3041 3042 OopTaskQueueSet* task_queues() { return _task_queues; } 3043 3044 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3045 3046 HeapWord** global_finger_addr() { return &_global_finger; } 3047 3048 CMSConcMarkingTerminator* terminator() { return &_term; } 3049 3050 virtual void set_for_termination(uint active_workers) { 3051 terminator()->reset_for_reuse(active_workers); 3052 } 3053 3054 void work(uint worker_id); 3055 bool should_yield() { 3056 return ConcurrentMarkSweepThread::should_yield() 3057 && !_collector->foregroundGCIsActive(); 3058 } 3059 3060 virtual void coordinator_yield(); // stuff done by coordinator 3061 bool result() { return _result; } 3062 3063 void reset(HeapWord* ra) { 3064 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3065 _restart_addr = _global_finger = ra; 3066 _term.reset_for_reuse(); 3067 } 3068 3069 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3070 OopTaskQueue* work_q); 3071 3072 private: 3073 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3074 void do_work_steal(int i); 3075 void bump_global_finger(HeapWord* f); 3076 }; 3077 3078 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3079 assert(_task != NULL, "Error"); 3080 return _task->yielding(); 3081 // Note that we do not need the disjunct || _task->should_yield() above 3082 // because we want terminating threads to yield only if the task 3083 // is already in the midst of yielding, which happens only after at least one 3084 // thread has yielded. 3085 } 3086 3087 void CMSConcMarkingTerminator::yield() { 3088 if (_task->should_yield()) { 3089 _task->yield(); 3090 } else { 3091 ParallelTaskTerminator::yield(); 3092 } 3093 } 3094 3095 //////////////////////////////////////////////////////////////// 3096 // Concurrent Marking Algorithm Sketch 3097 //////////////////////////////////////////////////////////////// 3098 // Until all tasks exhausted (both spaces): 3099 // -- claim next available chunk 3100 // -- bump global finger via CAS 3101 // -- find first object that starts in this chunk 3102 // and start scanning bitmap from that position 3103 // -- scan marked objects for oops 3104 // -- CAS-mark target, and if successful: 3105 // . if target oop is above global finger (volatile read) 3106 // nothing to do 3107 // . if target oop is in chunk and above local finger 3108 // then nothing to do 3109 // . else push on work-queue 3110 // -- Deal with possible overflow issues: 3111 // . local work-queue overflow causes stuff to be pushed on 3112 // global (common) overflow queue 3113 // . always first empty local work queue 3114 // . then get a batch of oops from global work queue if any 3115 // . then do work stealing 3116 // -- When all tasks claimed (both spaces) 3117 // and local work queue empty, 3118 // then in a loop do: 3119 // . check global overflow stack; steal a batch of oops and trace 3120 // . try to steal from other threads oif GOS is empty 3121 // . if neither is available, offer termination 3122 // -- Terminate and return result 3123 // 3124 void CMSConcMarkingTask::work(uint worker_id) { 3125 elapsedTimer _timer; 3126 ResourceMark rm; 3127 HandleMark hm; 3128 3129 DEBUG_ONLY(_collector->verify_overflow_empty();) 3130 3131 // Before we begin work, our work queue should be empty 3132 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3133 // Scan the bitmap covering _cms_space, tracing through grey objects. 3134 _timer.start(); 3135 do_scan_and_mark(worker_id, _cms_space); 3136 _timer.stop(); 3137 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3138 3139 // ... do work stealing 3140 _timer.reset(); 3141 _timer.start(); 3142 do_work_steal(worker_id); 3143 _timer.stop(); 3144 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3145 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3146 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3147 // Note that under the current task protocol, the 3148 // following assertion is true even of the spaces 3149 // expanded since the completion of the concurrent 3150 // marking. XXX This will likely change under a strict 3151 // ABORT semantics. 3152 // After perm removal the comparison was changed to 3153 // greater than or equal to from strictly greater than. 3154 // Before perm removal the highest address sweep would 3155 // have been at the end of perm gen but now is at the 3156 // end of the tenured gen. 3157 assert(_global_finger >= _cms_space->end(), 3158 "All tasks have been completed"); 3159 DEBUG_ONLY(_collector->verify_overflow_empty();) 3160 } 3161 3162 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3163 HeapWord* read = _global_finger; 3164 HeapWord* cur = read; 3165 while (f > read) { 3166 cur = read; 3167 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3168 if (cur == read) { 3169 // our cas succeeded 3170 assert(_global_finger >= f, "protocol consistency"); 3171 break; 3172 } 3173 } 3174 } 3175 3176 // This is really inefficient, and should be redone by 3177 // using (not yet available) block-read and -write interfaces to the 3178 // stack and the work_queue. XXX FIX ME !!! 3179 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3180 OopTaskQueue* work_q) { 3181 // Fast lock-free check 3182 if (ovflw_stk->length() == 0) { 3183 return false; 3184 } 3185 assert(work_q->size() == 0, "Shouldn't steal"); 3186 MutexLockerEx ml(ovflw_stk->par_lock(), 3187 Mutex::_no_safepoint_check_flag); 3188 // Grab up to 1/4 the size of the work queue 3189 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3190 (size_t)ParGCDesiredObjsFromOverflowList); 3191 num = MIN2(num, ovflw_stk->length()); 3192 for (int i = (int) num; i > 0; i--) { 3193 oop cur = ovflw_stk->pop(); 3194 assert(cur != NULL, "Counted wrong?"); 3195 work_q->push(cur); 3196 } 3197 return num > 0; 3198 } 3199 3200 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3201 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3202 int n_tasks = pst->n_tasks(); 3203 // We allow that there may be no tasks to do here because 3204 // we are restarting after a stack overflow. 3205 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3206 uint nth_task = 0; 3207 3208 HeapWord* aligned_start = sp->bottom(); 3209 if (sp->used_region().contains(_restart_addr)) { 3210 // Align down to a card boundary for the start of 0th task 3211 // for this space. 3212 aligned_start = 3213 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3214 CardTableModRefBS::card_size); 3215 } 3216 3217 size_t chunk_size = sp->marking_task_size(); 3218 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3219 // Having claimed the nth task in this space, 3220 // compute the chunk that it corresponds to: 3221 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3222 aligned_start + (nth_task+1)*chunk_size); 3223 // Try and bump the global finger via a CAS; 3224 // note that we need to do the global finger bump 3225 // _before_ taking the intersection below, because 3226 // the task corresponding to that region will be 3227 // deemed done even if the used_region() expands 3228 // because of allocation -- as it almost certainly will 3229 // during start-up while the threads yield in the 3230 // closure below. 3231 HeapWord* finger = span.end(); 3232 bump_global_finger(finger); // atomically 3233 // There are null tasks here corresponding to chunks 3234 // beyond the "top" address of the space. 3235 span = span.intersection(sp->used_region()); 3236 if (!span.is_empty()) { // Non-null task 3237 HeapWord* prev_obj; 3238 assert(!span.contains(_restart_addr) || nth_task == 0, 3239 "Inconsistency"); 3240 if (nth_task == 0) { 3241 // For the 0th task, we'll not need to compute a block_start. 3242 if (span.contains(_restart_addr)) { 3243 // In the case of a restart because of stack overflow, 3244 // we might additionally skip a chunk prefix. 3245 prev_obj = _restart_addr; 3246 } else { 3247 prev_obj = span.start(); 3248 } 3249 } else { 3250 // We want to skip the first object because 3251 // the protocol is to scan any object in its entirety 3252 // that _starts_ in this span; a fortiori, any 3253 // object starting in an earlier span is scanned 3254 // as part of an earlier claimed task. 3255 // Below we use the "careful" version of block_start 3256 // so we do not try to navigate uninitialized objects. 3257 prev_obj = sp->block_start_careful(span.start()); 3258 // Below we use a variant of block_size that uses the 3259 // Printezis bits to avoid waiting for allocated 3260 // objects to become initialized/parsable. 3261 while (prev_obj < span.start()) { 3262 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3263 if (sz > 0) { 3264 prev_obj += sz; 3265 } else { 3266 // In this case we may end up doing a bit of redundant 3267 // scanning, but that appears unavoidable, short of 3268 // locking the free list locks; see bug 6324141. 3269 break; 3270 } 3271 } 3272 } 3273 if (prev_obj < span.end()) { 3274 MemRegion my_span = MemRegion(prev_obj, span.end()); 3275 // Do the marking work within a non-empty span -- 3276 // the last argument to the constructor indicates whether the 3277 // iteration should be incremental with periodic yields. 3278 ParMarkFromRootsClosure cl(this, _collector, my_span, 3279 &_collector->_markBitMap, 3280 work_queue(i), 3281 &_collector->_markStack); 3282 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3283 } // else nothing to do for this task 3284 } // else nothing to do for this task 3285 } 3286 // We'd be tempted to assert here that since there are no 3287 // more tasks left to claim in this space, the global_finger 3288 // must exceed space->top() and a fortiori space->end(). However, 3289 // that would not quite be correct because the bumping of 3290 // global_finger occurs strictly after the claiming of a task, 3291 // so by the time we reach here the global finger may not yet 3292 // have been bumped up by the thread that claimed the last 3293 // task. 3294 pst->all_tasks_completed(); 3295 } 3296 3297 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3298 private: 3299 CMSCollector* _collector; 3300 CMSConcMarkingTask* _task; 3301 MemRegion _span; 3302 CMSBitMap* _bit_map; 3303 CMSMarkStack* _overflow_stack; 3304 OopTaskQueue* _work_queue; 3305 protected: 3306 DO_OOP_WORK_DEFN 3307 public: 3308 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3309 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3310 MetadataAwareOopClosure(collector->ref_processor()), 3311 _collector(collector), 3312 _task(task), 3313 _span(collector->_span), 3314 _work_queue(work_queue), 3315 _bit_map(bit_map), 3316 _overflow_stack(overflow_stack) 3317 { } 3318 virtual void do_oop(oop* p); 3319 virtual void do_oop(narrowOop* p); 3320 3321 void trim_queue(size_t max); 3322 void handle_stack_overflow(HeapWord* lost); 3323 void do_yield_check() { 3324 if (_task->should_yield()) { 3325 _task->yield(); 3326 } 3327 } 3328 }; 3329 3330 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3331 3332 // Grey object scanning during work stealing phase -- 3333 // the salient assumption here is that any references 3334 // that are in these stolen objects being scanned must 3335 // already have been initialized (else they would not have 3336 // been published), so we do not need to check for 3337 // uninitialized objects before pushing here. 3338 void ParConcMarkingClosure::do_oop(oop obj) { 3339 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3340 HeapWord* addr = (HeapWord*)obj; 3341 // Check if oop points into the CMS generation 3342 // and is not marked 3343 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3344 // a white object ... 3345 // If we manage to "claim" the object, by being the 3346 // first thread to mark it, then we push it on our 3347 // marking stack 3348 if (_bit_map->par_mark(addr)) { // ... now grey 3349 // push on work queue (grey set) 3350 bool simulate_overflow = false; 3351 NOT_PRODUCT( 3352 if (CMSMarkStackOverflowALot && 3353 _collector->simulate_overflow()) { 3354 // simulate a stack overflow 3355 simulate_overflow = true; 3356 } 3357 ) 3358 if (simulate_overflow || 3359 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3360 // stack overflow 3361 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3362 // We cannot assert that the overflow stack is full because 3363 // it may have been emptied since. 3364 assert(simulate_overflow || 3365 _work_queue->size() == _work_queue->max_elems(), 3366 "Else push should have succeeded"); 3367 handle_stack_overflow(addr); 3368 } 3369 } // Else, some other thread got there first 3370 do_yield_check(); 3371 } 3372 } 3373 3374 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3375 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3376 3377 void ParConcMarkingClosure::trim_queue(size_t max) { 3378 while (_work_queue->size() > max) { 3379 oop new_oop; 3380 if (_work_queue->pop_local(new_oop)) { 3381 assert(new_oop->is_oop(), "Should be an oop"); 3382 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3383 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3384 new_oop->oop_iterate(this); // do_oop() above 3385 do_yield_check(); 3386 } 3387 } 3388 } 3389 3390 // Upon stack overflow, we discard (part of) the stack, 3391 // remembering the least address amongst those discarded 3392 // in CMSCollector's _restart_address. 3393 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3394 // We need to do this under a mutex to prevent other 3395 // workers from interfering with the work done below. 3396 MutexLockerEx ml(_overflow_stack->par_lock(), 3397 Mutex::_no_safepoint_check_flag); 3398 // Remember the least grey address discarded 3399 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3400 _collector->lower_restart_addr(ra); 3401 _overflow_stack->reset(); // discard stack contents 3402 _overflow_stack->expand(); // expand the stack if possible 3403 } 3404 3405 3406 void CMSConcMarkingTask::do_work_steal(int i) { 3407 OopTaskQueue* work_q = work_queue(i); 3408 oop obj_to_scan; 3409 CMSBitMap* bm = &(_collector->_markBitMap); 3410 CMSMarkStack* ovflw = &(_collector->_markStack); 3411 int* seed = _collector->hash_seed(i); 3412 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3413 while (true) { 3414 cl.trim_queue(0); 3415 assert(work_q->size() == 0, "Should have been emptied above"); 3416 if (get_work_from_overflow_stack(ovflw, work_q)) { 3417 // Can't assert below because the work obtained from the 3418 // overflow stack may already have been stolen from us. 3419 // assert(work_q->size() > 0, "Work from overflow stack"); 3420 continue; 3421 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3422 assert(obj_to_scan->is_oop(), "Should be an oop"); 3423 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3424 obj_to_scan->oop_iterate(&cl); 3425 } else if (terminator()->offer_termination(&_term_term)) { 3426 assert(work_q->size() == 0, "Impossible!"); 3427 break; 3428 } else if (yielding() || should_yield()) { 3429 yield(); 3430 } 3431 } 3432 } 3433 3434 // This is run by the CMS (coordinator) thread. 3435 void CMSConcMarkingTask::coordinator_yield() { 3436 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3437 "CMS thread should hold CMS token"); 3438 // First give up the locks, then yield, then re-lock 3439 // We should probably use a constructor/destructor idiom to 3440 // do this unlock/lock or modify the MutexUnlocker class to 3441 // serve our purpose. XXX 3442 assert_lock_strong(_bit_map_lock); 3443 _bit_map_lock->unlock(); 3444 ConcurrentMarkSweepThread::desynchronize(true); 3445 _collector->stopTimer(); 3446 _collector->incrementYields(); 3447 3448 // It is possible for whichever thread initiated the yield request 3449 // not to get a chance to wake up and take the bitmap lock between 3450 // this thread releasing it and reacquiring it. So, while the 3451 // should_yield() flag is on, let's sleep for a bit to give the 3452 // other thread a chance to wake up. The limit imposed on the number 3453 // of iterations is defensive, to avoid any unforseen circumstances 3454 // putting us into an infinite loop. Since it's always been this 3455 // (coordinator_yield()) method that was observed to cause the 3456 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3457 // which is by default non-zero. For the other seven methods that 3458 // also perform the yield operation, as are using a different 3459 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3460 // can enable the sleeping for those methods too, if necessary. 3461 // See 6442774. 3462 // 3463 // We really need to reconsider the synchronization between the GC 3464 // thread and the yield-requesting threads in the future and we 3465 // should really use wait/notify, which is the recommended 3466 // way of doing this type of interaction. Additionally, we should 3467 // consolidate the eight methods that do the yield operation and they 3468 // are almost identical into one for better maintainability and 3469 // readability. See 6445193. 3470 // 3471 // Tony 2006.06.29 3472 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3473 ConcurrentMarkSweepThread::should_yield() && 3474 !CMSCollector::foregroundGCIsActive(); ++i) { 3475 os::sleep(Thread::current(), 1, false); 3476 } 3477 3478 ConcurrentMarkSweepThread::synchronize(true); 3479 _bit_map_lock->lock_without_safepoint_check(); 3480 _collector->startTimer(); 3481 } 3482 3483 bool CMSCollector::do_marking_mt() { 3484 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3485 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3486 conc_workers()->active_workers(), 3487 Threads::number_of_non_daemon_threads()); 3488 conc_workers()->set_active_workers(num_workers); 3489 3490 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3491 3492 CMSConcMarkingTask tsk(this, 3493 cms_space, 3494 conc_workers(), 3495 task_queues()); 3496 3497 // Since the actual number of workers we get may be different 3498 // from the number we requested above, do we need to do anything different 3499 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3500 // class?? XXX 3501 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3502 3503 // Refs discovery is already non-atomic. 3504 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3505 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3506 conc_workers()->start_task(&tsk); 3507 while (tsk.yielded()) { 3508 tsk.coordinator_yield(); 3509 conc_workers()->continue_task(&tsk); 3510 } 3511 // If the task was aborted, _restart_addr will be non-NULL 3512 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3513 while (_restart_addr != NULL) { 3514 // XXX For now we do not make use of ABORTED state and have not 3515 // yet implemented the right abort semantics (even in the original 3516 // single-threaded CMS case). That needs some more investigation 3517 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3518 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3519 // If _restart_addr is non-NULL, a marking stack overflow 3520 // occurred; we need to do a fresh marking iteration from the 3521 // indicated restart address. 3522 if (_foregroundGCIsActive) { 3523 // We may be running into repeated stack overflows, having 3524 // reached the limit of the stack size, while making very 3525 // slow forward progress. It may be best to bail out and 3526 // let the foreground collector do its job. 3527 // Clear _restart_addr, so that foreground GC 3528 // works from scratch. This avoids the headache of 3529 // a "rescan" which would otherwise be needed because 3530 // of the dirty mod union table & card table. 3531 _restart_addr = NULL; 3532 return false; 3533 } 3534 // Adjust the task to restart from _restart_addr 3535 tsk.reset(_restart_addr); 3536 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3537 _restart_addr); 3538 _restart_addr = NULL; 3539 // Get the workers going again 3540 conc_workers()->start_task(&tsk); 3541 while (tsk.yielded()) { 3542 tsk.coordinator_yield(); 3543 conc_workers()->continue_task(&tsk); 3544 } 3545 } 3546 assert(tsk.completed(), "Inconsistency"); 3547 assert(tsk.result() == true, "Inconsistency"); 3548 return true; 3549 } 3550 3551 bool CMSCollector::do_marking_st() { 3552 ResourceMark rm; 3553 HandleMark hm; 3554 3555 // Temporarily make refs discovery single threaded (non-MT) 3556 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3557 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3558 &_markStack, CMSYield); 3559 // the last argument to iterate indicates whether the iteration 3560 // should be incremental with periodic yields. 3561 _markBitMap.iterate(&markFromRootsClosure); 3562 // If _restart_addr is non-NULL, a marking stack overflow 3563 // occurred; we need to do a fresh iteration from the 3564 // indicated restart address. 3565 while (_restart_addr != NULL) { 3566 if (_foregroundGCIsActive) { 3567 // We may be running into repeated stack overflows, having 3568 // reached the limit of the stack size, while making very 3569 // slow forward progress. It may be best to bail out and 3570 // let the foreground collector do its job. 3571 // Clear _restart_addr, so that foreground GC 3572 // works from scratch. This avoids the headache of 3573 // a "rescan" which would otherwise be needed because 3574 // of the dirty mod union table & card table. 3575 _restart_addr = NULL; 3576 return false; // indicating failure to complete marking 3577 } 3578 // Deal with stack overflow: 3579 // we restart marking from _restart_addr 3580 HeapWord* ra = _restart_addr; 3581 markFromRootsClosure.reset(ra); 3582 _restart_addr = NULL; 3583 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3584 } 3585 return true; 3586 } 3587 3588 void CMSCollector::preclean() { 3589 check_correct_thread_executing(); 3590 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3591 verify_work_stacks_empty(); 3592 verify_overflow_empty(); 3593 _abort_preclean = false; 3594 if (CMSPrecleaningEnabled) { 3595 if (!CMSEdenChunksRecordAlways) { 3596 _eden_chunk_index = 0; 3597 } 3598 size_t used = get_eden_used(); 3599 size_t capacity = get_eden_capacity(); 3600 // Don't start sampling unless we will get sufficiently 3601 // many samples. 3602 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3603 * CMSScheduleRemarkEdenPenetration)) { 3604 _start_sampling = true; 3605 } else { 3606 _start_sampling = false; 3607 } 3608 GCTraceCPUTime tcpu; 3609 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3610 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3611 } 3612 CMSTokenSync x(true); // is cms thread 3613 if (CMSPrecleaningEnabled) { 3614 sample_eden(); 3615 _collectorState = AbortablePreclean; 3616 } else { 3617 _collectorState = FinalMarking; 3618 } 3619 verify_work_stacks_empty(); 3620 verify_overflow_empty(); 3621 } 3622 3623 // Try and schedule the remark such that young gen 3624 // occupancy is CMSScheduleRemarkEdenPenetration %. 3625 void CMSCollector::abortable_preclean() { 3626 check_correct_thread_executing(); 3627 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3628 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3629 3630 // If Eden's current occupancy is below this threshold, 3631 // immediately schedule the remark; else preclean 3632 // past the next scavenge in an effort to 3633 // schedule the pause as described above. By choosing 3634 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3635 // we will never do an actual abortable preclean cycle. 3636 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3637 GCTraceCPUTime tcpu; 3638 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3639 // We need more smarts in the abortable preclean 3640 // loop below to deal with cases where allocation 3641 // in young gen is very very slow, and our precleaning 3642 // is running a losing race against a horde of 3643 // mutators intent on flooding us with CMS updates 3644 // (dirty cards). 3645 // One, admittedly dumb, strategy is to give up 3646 // after a certain number of abortable precleaning loops 3647 // or after a certain maximum time. We want to make 3648 // this smarter in the next iteration. 3649 // XXX FIX ME!!! YSR 3650 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3651 while (!(should_abort_preclean() || 3652 ConcurrentMarkSweepThread::should_terminate())) { 3653 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3654 cumworkdone += workdone; 3655 loops++; 3656 // Voluntarily terminate abortable preclean phase if we have 3657 // been at it for too long. 3658 if ((CMSMaxAbortablePrecleanLoops != 0) && 3659 loops >= CMSMaxAbortablePrecleanLoops) { 3660 log_debug(gc)(" CMS: abort preclean due to loops "); 3661 break; 3662 } 3663 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3664 log_debug(gc)(" CMS: abort preclean due to time "); 3665 break; 3666 } 3667 // If we are doing little work each iteration, we should 3668 // take a short break. 3669 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3670 // Sleep for some time, waiting for work to accumulate 3671 stopTimer(); 3672 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3673 startTimer(); 3674 waited++; 3675 } 3676 } 3677 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3678 loops, waited, cumworkdone); 3679 } 3680 CMSTokenSync x(true); // is cms thread 3681 if (_collectorState != Idling) { 3682 assert(_collectorState == AbortablePreclean, 3683 "Spontaneous state transition?"); 3684 _collectorState = FinalMarking; 3685 } // Else, a foreground collection completed this CMS cycle. 3686 return; 3687 } 3688 3689 // Respond to an Eden sampling opportunity 3690 void CMSCollector::sample_eden() { 3691 // Make sure a young gc cannot sneak in between our 3692 // reading and recording of a sample. 3693 assert(Thread::current()->is_ConcurrentGC_thread(), 3694 "Only the cms thread may collect Eden samples"); 3695 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3696 "Should collect samples while holding CMS token"); 3697 if (!_start_sampling) { 3698 return; 3699 } 3700 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3701 // is populated by the young generation. 3702 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3703 if (_eden_chunk_index < _eden_chunk_capacity) { 3704 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3705 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3706 "Unexpected state of Eden"); 3707 // We'd like to check that what we just sampled is an oop-start address; 3708 // however, we cannot do that here since the object may not yet have been 3709 // initialized. So we'll instead do the check when we _use_ this sample 3710 // later. 3711 if (_eden_chunk_index == 0 || 3712 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3713 _eden_chunk_array[_eden_chunk_index-1]) 3714 >= CMSSamplingGrain)) { 3715 _eden_chunk_index++; // commit sample 3716 } 3717 } 3718 } 3719 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3720 size_t used = get_eden_used(); 3721 size_t capacity = get_eden_capacity(); 3722 assert(used <= capacity, "Unexpected state of Eden"); 3723 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3724 _abort_preclean = true; 3725 } 3726 } 3727 } 3728 3729 3730 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3731 assert(_collectorState == Precleaning || 3732 _collectorState == AbortablePreclean, "incorrect state"); 3733 ResourceMark rm; 3734 HandleMark hm; 3735 3736 // Precleaning is currently not MT but the reference processor 3737 // may be set for MT. Disable it temporarily here. 3738 ReferenceProcessor* rp = ref_processor(); 3739 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3740 3741 // Do one pass of scrubbing the discovered reference lists 3742 // to remove any reference objects with strongly-reachable 3743 // referents. 3744 if (clean_refs) { 3745 CMSPrecleanRefsYieldClosure yield_cl(this); 3746 assert(rp->span().equals(_span), "Spans should be equal"); 3747 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3748 &_markStack, true /* preclean */); 3749 CMSDrainMarkingStackClosure complete_trace(this, 3750 _span, &_markBitMap, &_markStack, 3751 &keep_alive, true /* preclean */); 3752 3753 // We don't want this step to interfere with a young 3754 // collection because we don't want to take CPU 3755 // or memory bandwidth away from the young GC threads 3756 // (which may be as many as there are CPUs). 3757 // Note that we don't need to protect ourselves from 3758 // interference with mutators because they can't 3759 // manipulate the discovered reference lists nor affect 3760 // the computed reachability of the referents, the 3761 // only properties manipulated by the precleaning 3762 // of these reference lists. 3763 stopTimer(); 3764 CMSTokenSyncWithLocks x(true /* is cms thread */, 3765 bitMapLock()); 3766 startTimer(); 3767 sample_eden(); 3768 3769 // The following will yield to allow foreground 3770 // collection to proceed promptly. XXX YSR: 3771 // The code in this method may need further 3772 // tweaking for better performance and some restructuring 3773 // for cleaner interfaces. 3774 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3775 rp->preclean_discovered_references( 3776 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3777 gc_timer); 3778 } 3779 3780 if (clean_survivor) { // preclean the active survivor space(s) 3781 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3782 &_markBitMap, &_modUnionTable, 3783 &_markStack, true /* precleaning phase */); 3784 stopTimer(); 3785 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3786 bitMapLock()); 3787 startTimer(); 3788 unsigned int before_count = 3789 GenCollectedHeap::heap()->total_collections(); 3790 SurvivorSpacePrecleanClosure 3791 sss_cl(this, _span, &_markBitMap, &_markStack, 3792 &pam_cl, before_count, CMSYield); 3793 _young_gen->from()->object_iterate_careful(&sss_cl); 3794 _young_gen->to()->object_iterate_careful(&sss_cl); 3795 } 3796 MarkRefsIntoAndScanClosure 3797 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3798 &_markStack, this, CMSYield, 3799 true /* precleaning phase */); 3800 // CAUTION: The following closure has persistent state that may need to 3801 // be reset upon a decrease in the sequence of addresses it 3802 // processes. 3803 ScanMarkedObjectsAgainCarefullyClosure 3804 smoac_cl(this, _span, 3805 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3806 3807 // Preclean dirty cards in ModUnionTable and CardTable using 3808 // appropriate convergence criterion; 3809 // repeat CMSPrecleanIter times unless we find that 3810 // we are losing. 3811 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3812 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3813 "Bad convergence multiplier"); 3814 assert(CMSPrecleanThreshold >= 100, 3815 "Unreasonably low CMSPrecleanThreshold"); 3816 3817 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3818 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3819 numIter < CMSPrecleanIter; 3820 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3821 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3822 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3823 // Either there are very few dirty cards, so re-mark 3824 // pause will be small anyway, or our pre-cleaning isn't 3825 // that much faster than the rate at which cards are being 3826 // dirtied, so we might as well stop and re-mark since 3827 // precleaning won't improve our re-mark time by much. 3828 if (curNumCards <= CMSPrecleanThreshold || 3829 (numIter > 0 && 3830 (curNumCards * CMSPrecleanDenominator > 3831 lastNumCards * CMSPrecleanNumerator))) { 3832 numIter++; 3833 cumNumCards += curNumCards; 3834 break; 3835 } 3836 } 3837 3838 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3839 3840 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3841 cumNumCards += curNumCards; 3842 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3843 curNumCards, cumNumCards, numIter); 3844 return cumNumCards; // as a measure of useful work done 3845 } 3846 3847 // PRECLEANING NOTES: 3848 // Precleaning involves: 3849 // . reading the bits of the modUnionTable and clearing the set bits. 3850 // . For the cards corresponding to the set bits, we scan the 3851 // objects on those cards. This means we need the free_list_lock 3852 // so that we can safely iterate over the CMS space when scanning 3853 // for oops. 3854 // . When we scan the objects, we'll be both reading and setting 3855 // marks in the marking bit map, so we'll need the marking bit map. 3856 // . For protecting _collector_state transitions, we take the CGC_lock. 3857 // Note that any races in the reading of of card table entries by the 3858 // CMS thread on the one hand and the clearing of those entries by the 3859 // VM thread or the setting of those entries by the mutator threads on the 3860 // other are quite benign. However, for efficiency it makes sense to keep 3861 // the VM thread from racing with the CMS thread while the latter is 3862 // dirty card info to the modUnionTable. We therefore also use the 3863 // CGC_lock to protect the reading of the card table and the mod union 3864 // table by the CM thread. 3865 // . We run concurrently with mutator updates, so scanning 3866 // needs to be done carefully -- we should not try to scan 3867 // potentially uninitialized objects. 3868 // 3869 // Locking strategy: While holding the CGC_lock, we scan over and 3870 // reset a maximal dirty range of the mod union / card tables, then lock 3871 // the free_list_lock and bitmap lock to do a full marking, then 3872 // release these locks; and repeat the cycle. This allows for a 3873 // certain amount of fairness in the sharing of these locks between 3874 // the CMS collector on the one hand, and the VM thread and the 3875 // mutators on the other. 3876 3877 // NOTE: preclean_mod_union_table() and preclean_card_table() 3878 // further below are largely identical; if you need to modify 3879 // one of these methods, please check the other method too. 3880 3881 size_t CMSCollector::preclean_mod_union_table( 3882 ConcurrentMarkSweepGeneration* old_gen, 3883 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3884 verify_work_stacks_empty(); 3885 verify_overflow_empty(); 3886 3887 // strategy: starting with the first card, accumulate contiguous 3888 // ranges of dirty cards; clear these cards, then scan the region 3889 // covered by these cards. 3890 3891 // Since all of the MUT is committed ahead, we can just use 3892 // that, in case the generations expand while we are precleaning. 3893 // It might also be fine to just use the committed part of the 3894 // generation, but we might potentially miss cards when the 3895 // generation is rapidly expanding while we are in the midst 3896 // of precleaning. 3897 HeapWord* startAddr = old_gen->reserved().start(); 3898 HeapWord* endAddr = old_gen->reserved().end(); 3899 3900 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3901 3902 size_t numDirtyCards, cumNumDirtyCards; 3903 HeapWord *nextAddr, *lastAddr; 3904 for (cumNumDirtyCards = numDirtyCards = 0, 3905 nextAddr = lastAddr = startAddr; 3906 nextAddr < endAddr; 3907 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3908 3909 ResourceMark rm; 3910 HandleMark hm; 3911 3912 MemRegion dirtyRegion; 3913 { 3914 stopTimer(); 3915 // Potential yield point 3916 CMSTokenSync ts(true); 3917 startTimer(); 3918 sample_eden(); 3919 // Get dirty region starting at nextOffset (inclusive), 3920 // simultaneously clearing it. 3921 dirtyRegion = 3922 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3923 assert(dirtyRegion.start() >= nextAddr, 3924 "returned region inconsistent?"); 3925 } 3926 // Remember where the next search should begin. 3927 // The returned region (if non-empty) is a right open interval, 3928 // so lastOffset is obtained from the right end of that 3929 // interval. 3930 lastAddr = dirtyRegion.end(); 3931 // Should do something more transparent and less hacky XXX 3932 numDirtyCards = 3933 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3934 3935 // We'll scan the cards in the dirty region (with periodic 3936 // yields for foreground GC as needed). 3937 if (!dirtyRegion.is_empty()) { 3938 assert(numDirtyCards > 0, "consistency check"); 3939 HeapWord* stop_point = NULL; 3940 stopTimer(); 3941 // Potential yield point 3942 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3943 bitMapLock()); 3944 startTimer(); 3945 { 3946 verify_work_stacks_empty(); 3947 verify_overflow_empty(); 3948 sample_eden(); 3949 stop_point = 3950 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3951 } 3952 if (stop_point != NULL) { 3953 // The careful iteration stopped early either because it found an 3954 // uninitialized object, or because we were in the midst of an 3955 // "abortable preclean", which should now be aborted. Redirty 3956 // the bits corresponding to the partially-scanned or unscanned 3957 // cards. We'll either restart at the next block boundary or 3958 // abort the preclean. 3959 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3960 "Should only be AbortablePreclean."); 3961 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3962 if (should_abort_preclean()) { 3963 break; // out of preclean loop 3964 } else { 3965 // Compute the next address at which preclean should pick up; 3966 // might need bitMapLock in order to read P-bits. 3967 lastAddr = next_card_start_after_block(stop_point); 3968 } 3969 } 3970 } else { 3971 assert(lastAddr == endAddr, "consistency check"); 3972 assert(numDirtyCards == 0, "consistency check"); 3973 break; 3974 } 3975 } 3976 verify_work_stacks_empty(); 3977 verify_overflow_empty(); 3978 return cumNumDirtyCards; 3979 } 3980 3981 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3982 // below are largely identical; if you need to modify 3983 // one of these methods, please check the other method too. 3984 3985 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3986 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3987 // strategy: it's similar to precleamModUnionTable above, in that 3988 // we accumulate contiguous ranges of dirty cards, mark these cards 3989 // precleaned, then scan the region covered by these cards. 3990 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3991 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3992 3993 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3994 3995 size_t numDirtyCards, cumNumDirtyCards; 3996 HeapWord *lastAddr, *nextAddr; 3997 3998 for (cumNumDirtyCards = numDirtyCards = 0, 3999 nextAddr = lastAddr = startAddr; 4000 nextAddr < endAddr; 4001 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4002 4003 ResourceMark rm; 4004 HandleMark hm; 4005 4006 MemRegion dirtyRegion; 4007 { 4008 // See comments in "Precleaning notes" above on why we 4009 // do this locking. XXX Could the locking overheads be 4010 // too high when dirty cards are sparse? [I don't think so.] 4011 stopTimer(); 4012 CMSTokenSync x(true); // is cms thread 4013 startTimer(); 4014 sample_eden(); 4015 // Get and clear dirty region from card table 4016 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4017 MemRegion(nextAddr, endAddr), 4018 true, 4019 CardTableModRefBS::precleaned_card_val()); 4020 4021 assert(dirtyRegion.start() >= nextAddr, 4022 "returned region inconsistent?"); 4023 } 4024 lastAddr = dirtyRegion.end(); 4025 numDirtyCards = 4026 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4027 4028 if (!dirtyRegion.is_empty()) { 4029 stopTimer(); 4030 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4031 startTimer(); 4032 sample_eden(); 4033 verify_work_stacks_empty(); 4034 verify_overflow_empty(); 4035 HeapWord* stop_point = 4036 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4037 if (stop_point != NULL) { 4038 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4039 "Should only be AbortablePreclean."); 4040 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4041 if (should_abort_preclean()) { 4042 break; // out of preclean loop 4043 } else { 4044 // Compute the next address at which preclean should pick up. 4045 lastAddr = next_card_start_after_block(stop_point); 4046 } 4047 } 4048 } else { 4049 break; 4050 } 4051 } 4052 verify_work_stacks_empty(); 4053 verify_overflow_empty(); 4054 return cumNumDirtyCards; 4055 } 4056 4057 class PrecleanKlassClosure : public KlassClosure { 4058 KlassToOopClosure _cm_klass_closure; 4059 public: 4060 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4061 void do_klass(Klass* k) { 4062 if (k->has_accumulated_modified_oops()) { 4063 k->clear_accumulated_modified_oops(); 4064 4065 _cm_klass_closure.do_klass(k); 4066 } 4067 } 4068 }; 4069 4070 // The freelist lock is needed to prevent asserts, is it really needed? 4071 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4072 4073 cl->set_freelistLock(freelistLock); 4074 4075 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4076 4077 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4078 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4079 PrecleanKlassClosure preclean_klass_closure(cl); 4080 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4081 4082 verify_work_stacks_empty(); 4083 verify_overflow_empty(); 4084 } 4085 4086 void CMSCollector::checkpointRootsFinal() { 4087 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4088 check_correct_thread_executing(); 4089 // world is stopped at this checkpoint 4090 assert(SafepointSynchronize::is_at_safepoint(), 4091 "world should be stopped"); 4092 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4093 4094 verify_work_stacks_empty(); 4095 verify_overflow_empty(); 4096 4097 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4098 _young_gen->used() / K, _young_gen->capacity() / K); 4099 { 4100 if (CMSScavengeBeforeRemark) { 4101 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4102 // Temporarily set flag to false, GCH->do_collection will 4103 // expect it to be false and set to true 4104 FlagSetting fl(gch->_is_gc_active, false); 4105 4106 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4107 4108 gch->do_collection(true, // full (i.e. force, see below) 4109 false, // !clear_all_soft_refs 4110 0, // size 4111 false, // is_tlab 4112 GenCollectedHeap::YoungGen // type 4113 ); 4114 } 4115 FreelistLocker x(this); 4116 MutexLockerEx y(bitMapLock(), 4117 Mutex::_no_safepoint_check_flag); 4118 checkpointRootsFinalWork(); 4119 } 4120 verify_work_stacks_empty(); 4121 verify_overflow_empty(); 4122 } 4123 4124 void CMSCollector::checkpointRootsFinalWork() { 4125 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4126 4127 assert(haveFreelistLocks(), "must have free list locks"); 4128 assert_lock_strong(bitMapLock()); 4129 4130 ResourceMark rm; 4131 HandleMark hm; 4132 4133 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4134 4135 if (should_unload_classes()) { 4136 CodeCache::gc_prologue(); 4137 } 4138 assert(haveFreelistLocks(), "must have free list locks"); 4139 assert_lock_strong(bitMapLock()); 4140 4141 // We might assume that we need not fill TLAB's when 4142 // CMSScavengeBeforeRemark is set, because we may have just done 4143 // a scavenge which would have filled all TLAB's -- and besides 4144 // Eden would be empty. This however may not always be the case -- 4145 // for instance although we asked for a scavenge, it may not have 4146 // happened because of a JNI critical section. We probably need 4147 // a policy for deciding whether we can in that case wait until 4148 // the critical section releases and then do the remark following 4149 // the scavenge, and skip it here. In the absence of that policy, 4150 // or of an indication of whether the scavenge did indeed occur, 4151 // we cannot rely on TLAB's having been filled and must do 4152 // so here just in case a scavenge did not happen. 4153 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4154 // Update the saved marks which may affect the root scans. 4155 gch->save_marks(); 4156 4157 print_eden_and_survivor_chunk_arrays(); 4158 4159 { 4160 #if defined(COMPILER2) || INCLUDE_JVMCI 4161 DerivedPointerTableDeactivate dpt_deact; 4162 #endif 4163 4164 // Note on the role of the mod union table: 4165 // Since the marker in "markFromRoots" marks concurrently with 4166 // mutators, it is possible for some reachable objects not to have been 4167 // scanned. For instance, an only reference to an object A was 4168 // placed in object B after the marker scanned B. Unless B is rescanned, 4169 // A would be collected. Such updates to references in marked objects 4170 // are detected via the mod union table which is the set of all cards 4171 // dirtied since the first checkpoint in this GC cycle and prior to 4172 // the most recent young generation GC, minus those cleaned up by the 4173 // concurrent precleaning. 4174 if (CMSParallelRemarkEnabled) { 4175 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4176 do_remark_parallel(); 4177 } else { 4178 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4179 do_remark_non_parallel(); 4180 } 4181 } 4182 verify_work_stacks_empty(); 4183 verify_overflow_empty(); 4184 4185 { 4186 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4187 refProcessingWork(); 4188 } 4189 verify_work_stacks_empty(); 4190 verify_overflow_empty(); 4191 4192 if (should_unload_classes()) { 4193 CodeCache::gc_epilogue(); 4194 } 4195 JvmtiExport::gc_epilogue(); 4196 4197 // If we encountered any (marking stack / work queue) overflow 4198 // events during the current CMS cycle, take appropriate 4199 // remedial measures, where possible, so as to try and avoid 4200 // recurrence of that condition. 4201 assert(_markStack.isEmpty(), "No grey objects"); 4202 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4203 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4204 if (ser_ovflw > 0) { 4205 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4206 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4207 _markStack.expand(); 4208 _ser_pmc_remark_ovflw = 0; 4209 _ser_pmc_preclean_ovflw = 0; 4210 _ser_kac_preclean_ovflw = 0; 4211 _ser_kac_ovflw = 0; 4212 } 4213 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4214 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4215 _par_pmc_remark_ovflw, _par_kac_ovflw); 4216 _par_pmc_remark_ovflw = 0; 4217 _par_kac_ovflw = 0; 4218 } 4219 if (_markStack._hit_limit > 0) { 4220 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4221 _markStack._hit_limit); 4222 } 4223 if (_markStack._failed_double > 0) { 4224 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4225 _markStack._failed_double, _markStack.capacity()); 4226 } 4227 _markStack._hit_limit = 0; 4228 _markStack._failed_double = 0; 4229 4230 if ((VerifyAfterGC || VerifyDuringGC) && 4231 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4232 verify_after_remark(); 4233 } 4234 4235 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4236 4237 // Change under the freelistLocks. 4238 _collectorState = Sweeping; 4239 // Call isAllClear() under bitMapLock 4240 assert(_modUnionTable.isAllClear(), 4241 "Should be clear by end of the final marking"); 4242 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4243 "Should be clear by end of the final marking"); 4244 } 4245 4246 void CMSParInitialMarkTask::work(uint worker_id) { 4247 elapsedTimer _timer; 4248 ResourceMark rm; 4249 HandleMark hm; 4250 4251 // ---------- scan from roots -------------- 4252 _timer.start(); 4253 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4254 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4255 4256 // ---------- young gen roots -------------- 4257 { 4258 work_on_young_gen_roots(worker_id, &par_mri_cl); 4259 _timer.stop(); 4260 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4261 } 4262 4263 // ---------- remaining roots -------------- 4264 _timer.reset(); 4265 _timer.start(); 4266 4267 CLDToOopClosure cld_closure(&par_mri_cl, true); 4268 4269 gch->gen_process_roots(_strong_roots_scope, 4270 GenCollectedHeap::OldGen, 4271 false, // yg was scanned above 4272 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4273 _collector->should_unload_classes(), 4274 &par_mri_cl, 4275 NULL, 4276 &cld_closure); 4277 assert(_collector->should_unload_classes() 4278 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4279 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4280 _timer.stop(); 4281 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4282 } 4283 4284 // Parallel remark task 4285 class CMSParRemarkTask: public CMSParMarkTask { 4286 CompactibleFreeListSpace* _cms_space; 4287 4288 // The per-thread work queues, available here for stealing. 4289 OopTaskQueueSet* _task_queues; 4290 ParallelTaskTerminator _term; 4291 StrongRootsScope* _strong_roots_scope; 4292 4293 public: 4294 // A value of 0 passed to n_workers will cause the number of 4295 // workers to be taken from the active workers in the work gang. 4296 CMSParRemarkTask(CMSCollector* collector, 4297 CompactibleFreeListSpace* cms_space, 4298 uint n_workers, WorkGang* workers, 4299 OopTaskQueueSet* task_queues, 4300 StrongRootsScope* strong_roots_scope): 4301 CMSParMarkTask("Rescan roots and grey objects in parallel", 4302 collector, n_workers), 4303 _cms_space(cms_space), 4304 _task_queues(task_queues), 4305 _term(n_workers, task_queues), 4306 _strong_roots_scope(strong_roots_scope) { } 4307 4308 OopTaskQueueSet* task_queues() { return _task_queues; } 4309 4310 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4311 4312 ParallelTaskTerminator* terminator() { return &_term; } 4313 uint n_workers() { return _n_workers; } 4314 4315 void work(uint worker_id); 4316 4317 private: 4318 // ... of dirty cards in old space 4319 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4320 ParMarkRefsIntoAndScanClosure* cl); 4321 4322 // ... work stealing for the above 4323 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4324 }; 4325 4326 class RemarkKlassClosure : public KlassClosure { 4327 KlassToOopClosure _cm_klass_closure; 4328 public: 4329 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4330 void do_klass(Klass* k) { 4331 // Check if we have modified any oops in the Klass during the concurrent marking. 4332 if (k->has_accumulated_modified_oops()) { 4333 k->clear_accumulated_modified_oops(); 4334 4335 // We could have transfered the current modified marks to the accumulated marks, 4336 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4337 } else if (k->has_modified_oops()) { 4338 // Don't clear anything, this info is needed by the next young collection. 4339 } else { 4340 // No modified oops in the Klass. 4341 return; 4342 } 4343 4344 // The klass has modified fields, need to scan the klass. 4345 _cm_klass_closure.do_klass(k); 4346 } 4347 }; 4348 4349 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4350 ParNewGeneration* young_gen = _collector->_young_gen; 4351 ContiguousSpace* eden_space = young_gen->eden(); 4352 ContiguousSpace* from_space = young_gen->from(); 4353 ContiguousSpace* to_space = young_gen->to(); 4354 4355 HeapWord** eca = _collector->_eden_chunk_array; 4356 size_t ect = _collector->_eden_chunk_index; 4357 HeapWord** sca = _collector->_survivor_chunk_array; 4358 size_t sct = _collector->_survivor_chunk_index; 4359 4360 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4361 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4362 4363 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4364 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4365 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4366 } 4367 4368 // work_queue(i) is passed to the closure 4369 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4370 // also is passed to do_dirty_card_rescan_tasks() and to 4371 // do_work_steal() to select the i-th task_queue. 4372 4373 void CMSParRemarkTask::work(uint worker_id) { 4374 elapsedTimer _timer; 4375 ResourceMark rm; 4376 HandleMark hm; 4377 4378 // ---------- rescan from roots -------------- 4379 _timer.start(); 4380 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4381 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4382 _collector->_span, _collector->ref_processor(), 4383 &(_collector->_markBitMap), 4384 work_queue(worker_id)); 4385 4386 // Rescan young gen roots first since these are likely 4387 // coarsely partitioned and may, on that account, constitute 4388 // the critical path; thus, it's best to start off that 4389 // work first. 4390 // ---------- young gen roots -------------- 4391 { 4392 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4393 _timer.stop(); 4394 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4395 } 4396 4397 // ---------- remaining roots -------------- 4398 _timer.reset(); 4399 _timer.start(); 4400 gch->gen_process_roots(_strong_roots_scope, 4401 GenCollectedHeap::OldGen, 4402 false, // yg was scanned above 4403 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4404 _collector->should_unload_classes(), 4405 &par_mrias_cl, 4406 NULL, 4407 NULL); // The dirty klasses will be handled below 4408 4409 assert(_collector->should_unload_classes() 4410 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4411 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4412 _timer.stop(); 4413 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4414 4415 // ---------- unhandled CLD scanning ---------- 4416 if (worker_id == 0) { // Single threaded at the moment. 4417 _timer.reset(); 4418 _timer.start(); 4419 4420 // Scan all new class loader data objects and new dependencies that were 4421 // introduced during concurrent marking. 4422 ResourceMark rm; 4423 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4424 for (int i = 0; i < array->length(); i++) { 4425 par_mrias_cl.do_cld_nv(array->at(i)); 4426 } 4427 4428 // We don't need to keep track of new CLDs anymore. 4429 ClassLoaderDataGraph::remember_new_clds(false); 4430 4431 _timer.stop(); 4432 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4433 } 4434 4435 // ---------- dirty klass scanning ---------- 4436 if (worker_id == 0) { // Single threaded at the moment. 4437 _timer.reset(); 4438 _timer.start(); 4439 4440 // Scan all classes that was dirtied during the concurrent marking phase. 4441 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4442 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4443 4444 _timer.stop(); 4445 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4446 } 4447 4448 // We might have added oops to ClassLoaderData::_handles during the 4449 // concurrent marking phase. These oops point to newly allocated objects 4450 // that are guaranteed to be kept alive. Either by the direct allocation 4451 // code, or when the young collector processes the roots. Hence, 4452 // we don't have to revisit the _handles block during the remark phase. 4453 4454 // ---------- rescan dirty cards ------------ 4455 _timer.reset(); 4456 _timer.start(); 4457 4458 // Do the rescan tasks for each of the two spaces 4459 // (cms_space) in turn. 4460 // "worker_id" is passed to select the task_queue for "worker_id" 4461 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4462 _timer.stop(); 4463 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4464 4465 // ---------- steal work from other threads ... 4466 // ---------- ... and drain overflow list. 4467 _timer.reset(); 4468 _timer.start(); 4469 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4470 _timer.stop(); 4471 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4472 } 4473 4474 // Note that parameter "i" is not used. 4475 void 4476 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4477 OopsInGenClosure* cl, ContiguousSpace* space, 4478 HeapWord** chunk_array, size_t chunk_top) { 4479 // Until all tasks completed: 4480 // . claim an unclaimed task 4481 // . compute region boundaries corresponding to task claimed 4482 // using chunk_array 4483 // . par_oop_iterate(cl) over that region 4484 4485 ResourceMark rm; 4486 HandleMark hm; 4487 4488 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4489 4490 uint nth_task = 0; 4491 uint n_tasks = pst->n_tasks(); 4492 4493 if (n_tasks > 0) { 4494 assert(pst->valid(), "Uninitialized use?"); 4495 HeapWord *start, *end; 4496 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4497 // We claimed task # nth_task; compute its boundaries. 4498 if (chunk_top == 0) { // no samples were taken 4499 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4500 start = space->bottom(); 4501 end = space->top(); 4502 } else if (nth_task == 0) { 4503 start = space->bottom(); 4504 end = chunk_array[nth_task]; 4505 } else if (nth_task < (uint)chunk_top) { 4506 assert(nth_task >= 1, "Control point invariant"); 4507 start = chunk_array[nth_task - 1]; 4508 end = chunk_array[nth_task]; 4509 } else { 4510 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4511 start = chunk_array[chunk_top - 1]; 4512 end = space->top(); 4513 } 4514 MemRegion mr(start, end); 4515 // Verify that mr is in space 4516 assert(mr.is_empty() || space->used_region().contains(mr), 4517 "Should be in space"); 4518 // Verify that "start" is an object boundary 4519 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4520 "Should be an oop"); 4521 space->par_oop_iterate(mr, cl); 4522 } 4523 pst->all_tasks_completed(); 4524 } 4525 } 4526 4527 void 4528 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4529 CompactibleFreeListSpace* sp, int i, 4530 ParMarkRefsIntoAndScanClosure* cl) { 4531 // Until all tasks completed: 4532 // . claim an unclaimed task 4533 // . compute region boundaries corresponding to task claimed 4534 // . transfer dirty bits ct->mut for that region 4535 // . apply rescanclosure to dirty mut bits for that region 4536 4537 ResourceMark rm; 4538 HandleMark hm; 4539 4540 OopTaskQueue* work_q = work_queue(i); 4541 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4542 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4543 // CAUTION: This closure has state that persists across calls to 4544 // the work method dirty_range_iterate_clear() in that it has 4545 // embedded in it a (subtype of) UpwardsObjectClosure. The 4546 // use of that state in the embedded UpwardsObjectClosure instance 4547 // assumes that the cards are always iterated (even if in parallel 4548 // by several threads) in monotonically increasing order per each 4549 // thread. This is true of the implementation below which picks 4550 // card ranges (chunks) in monotonically increasing order globally 4551 // and, a-fortiori, in monotonically increasing order per thread 4552 // (the latter order being a subsequence of the former). 4553 // If the work code below is ever reorganized into a more chaotic 4554 // work-partitioning form than the current "sequential tasks" 4555 // paradigm, the use of that persistent state will have to be 4556 // revisited and modified appropriately. See also related 4557 // bug 4756801 work on which should examine this code to make 4558 // sure that the changes there do not run counter to the 4559 // assumptions made here and necessary for correctness and 4560 // efficiency. Note also that this code might yield inefficient 4561 // behavior in the case of very large objects that span one or 4562 // more work chunks. Such objects would potentially be scanned 4563 // several times redundantly. Work on 4756801 should try and 4564 // address that performance anomaly if at all possible. XXX 4565 MemRegion full_span = _collector->_span; 4566 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4567 MarkFromDirtyCardsClosure 4568 greyRescanClosure(_collector, full_span, // entire span of interest 4569 sp, bm, work_q, cl); 4570 4571 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4572 assert(pst->valid(), "Uninitialized use?"); 4573 uint nth_task = 0; 4574 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4575 MemRegion span = sp->used_region(); 4576 HeapWord* start_addr = span.start(); 4577 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4578 alignment); 4579 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4580 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4581 start_addr, "Check alignment"); 4582 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4583 chunk_size, "Check alignment"); 4584 4585 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4586 // Having claimed the nth_task, compute corresponding mem-region, 4587 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4588 // The alignment restriction ensures that we do not need any 4589 // synchronization with other gang-workers while setting or 4590 // clearing bits in thus chunk of the MUT. 4591 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4592 start_addr + (nth_task+1)*chunk_size); 4593 // The last chunk's end might be way beyond end of the 4594 // used region. In that case pull back appropriately. 4595 if (this_span.end() > end_addr) { 4596 this_span.set_end(end_addr); 4597 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4598 } 4599 // Iterate over the dirty cards covering this chunk, marking them 4600 // precleaned, and setting the corresponding bits in the mod union 4601 // table. Since we have been careful to partition at Card and MUT-word 4602 // boundaries no synchronization is needed between parallel threads. 4603 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4604 &modUnionClosure); 4605 4606 // Having transferred these marks into the modUnionTable, 4607 // rescan the marked objects on the dirty cards in the modUnionTable. 4608 // Even if this is at a synchronous collection, the initial marking 4609 // may have been done during an asynchronous collection so there 4610 // may be dirty bits in the mod-union table. 4611 _collector->_modUnionTable.dirty_range_iterate_clear( 4612 this_span, &greyRescanClosure); 4613 _collector->_modUnionTable.verifyNoOneBitsInRange( 4614 this_span.start(), 4615 this_span.end()); 4616 } 4617 pst->all_tasks_completed(); // declare that i am done 4618 } 4619 4620 // . see if we can share work_queues with ParNew? XXX 4621 void 4622 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4623 int* seed) { 4624 OopTaskQueue* work_q = work_queue(i); 4625 NOT_PRODUCT(int num_steals = 0;) 4626 oop obj_to_scan; 4627 CMSBitMap* bm = &(_collector->_markBitMap); 4628 4629 while (true) { 4630 // Completely finish any left over work from (an) earlier round(s) 4631 cl->trim_queue(0); 4632 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4633 (size_t)ParGCDesiredObjsFromOverflowList); 4634 // Now check if there's any work in the overflow list 4635 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4636 // only affects the number of attempts made to get work from the 4637 // overflow list and does not affect the number of workers. Just 4638 // pass ParallelGCThreads so this behavior is unchanged. 4639 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4640 work_q, 4641 ParallelGCThreads)) { 4642 // found something in global overflow list; 4643 // not yet ready to go stealing work from others. 4644 // We'd like to assert(work_q->size() != 0, ...) 4645 // because we just took work from the overflow list, 4646 // but of course we can't since all of that could have 4647 // been already stolen from us. 4648 // "He giveth and He taketh away." 4649 continue; 4650 } 4651 // Verify that we have no work before we resort to stealing 4652 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4653 // Try to steal from other queues that have work 4654 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4655 NOT_PRODUCT(num_steals++;) 4656 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4657 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4658 // Do scanning work 4659 obj_to_scan->oop_iterate(cl); 4660 // Loop around, finish this work, and try to steal some more 4661 } else if (terminator()->offer_termination()) { 4662 break; // nirvana from the infinite cycle 4663 } 4664 } 4665 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4666 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4667 "Else our work is not yet done"); 4668 } 4669 4670 // Record object boundaries in _eden_chunk_array by sampling the eden 4671 // top in the slow-path eden object allocation code path and record 4672 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4673 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4674 // sampling in sample_eden() that activates during the part of the 4675 // preclean phase. 4676 void CMSCollector::sample_eden_chunk() { 4677 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4678 if (_eden_chunk_lock->try_lock()) { 4679 // Record a sample. This is the critical section. The contents 4680 // of the _eden_chunk_array have to be non-decreasing in the 4681 // address order. 4682 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4683 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4684 "Unexpected state of Eden"); 4685 if (_eden_chunk_index == 0 || 4686 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4687 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4688 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4689 _eden_chunk_index++; // commit sample 4690 } 4691 _eden_chunk_lock->unlock(); 4692 } 4693 } 4694 } 4695 4696 // Return a thread-local PLAB recording array, as appropriate. 4697 void* CMSCollector::get_data_recorder(int thr_num) { 4698 if (_survivor_plab_array != NULL && 4699 (CMSPLABRecordAlways || 4700 (_collectorState > Marking && _collectorState < FinalMarking))) { 4701 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4702 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4703 ca->reset(); // clear it so that fresh data is recorded 4704 return (void*) ca; 4705 } else { 4706 return NULL; 4707 } 4708 } 4709 4710 // Reset all the thread-local PLAB recording arrays 4711 void CMSCollector::reset_survivor_plab_arrays() { 4712 for (uint i = 0; i < ParallelGCThreads; i++) { 4713 _survivor_plab_array[i].reset(); 4714 } 4715 } 4716 4717 // Merge the per-thread plab arrays into the global survivor chunk 4718 // array which will provide the partitioning of the survivor space 4719 // for CMS initial scan and rescan. 4720 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4721 int no_of_gc_threads) { 4722 assert(_survivor_plab_array != NULL, "Error"); 4723 assert(_survivor_chunk_array != NULL, "Error"); 4724 assert(_collectorState == FinalMarking || 4725 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4726 for (int j = 0; j < no_of_gc_threads; j++) { 4727 _cursor[j] = 0; 4728 } 4729 HeapWord* top = surv->top(); 4730 size_t i; 4731 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4732 HeapWord* min_val = top; // Higher than any PLAB address 4733 uint min_tid = 0; // position of min_val this round 4734 for (int j = 0; j < no_of_gc_threads; j++) { 4735 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4736 if (_cursor[j] == cur_sca->end()) { 4737 continue; 4738 } 4739 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4740 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4741 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4742 if (cur_val < min_val) { 4743 min_tid = j; 4744 min_val = cur_val; 4745 } else { 4746 assert(cur_val < top, "All recorded addresses should be less"); 4747 } 4748 } 4749 // At this point min_val and min_tid are respectively 4750 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4751 // and the thread (j) that witnesses that address. 4752 // We record this address in the _survivor_chunk_array[i] 4753 // and increment _cursor[min_tid] prior to the next round i. 4754 if (min_val == top) { 4755 break; 4756 } 4757 _survivor_chunk_array[i] = min_val; 4758 _cursor[min_tid]++; 4759 } 4760 // We are all done; record the size of the _survivor_chunk_array 4761 _survivor_chunk_index = i; // exclusive: [0, i) 4762 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4763 // Verify that we used up all the recorded entries 4764 #ifdef ASSERT 4765 size_t total = 0; 4766 for (int j = 0; j < no_of_gc_threads; j++) { 4767 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4768 total += _cursor[j]; 4769 } 4770 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4771 // Check that the merged array is in sorted order 4772 if (total > 0) { 4773 for (size_t i = 0; i < total - 1; i++) { 4774 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4775 i, p2i(_survivor_chunk_array[i])); 4776 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4777 "Not sorted"); 4778 } 4779 } 4780 #endif // ASSERT 4781 } 4782 4783 // Set up the space's par_seq_tasks structure for work claiming 4784 // for parallel initial scan and rescan of young gen. 4785 // See ParRescanTask where this is currently used. 4786 void 4787 CMSCollector:: 4788 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4789 assert(n_threads > 0, "Unexpected n_threads argument"); 4790 4791 // Eden space 4792 if (!_young_gen->eden()->is_empty()) { 4793 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4794 assert(!pst->valid(), "Clobbering existing data?"); 4795 // Each valid entry in [0, _eden_chunk_index) represents a task. 4796 size_t n_tasks = _eden_chunk_index + 1; 4797 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4798 // Sets the condition for completion of the subtask (how many threads 4799 // need to finish in order to be done). 4800 pst->set_n_threads(n_threads); 4801 pst->set_n_tasks((int)n_tasks); 4802 } 4803 4804 // Merge the survivor plab arrays into _survivor_chunk_array 4805 if (_survivor_plab_array != NULL) { 4806 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4807 } else { 4808 assert(_survivor_chunk_index == 0, "Error"); 4809 } 4810 4811 // To space 4812 { 4813 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4814 assert(!pst->valid(), "Clobbering existing data?"); 4815 // Sets the condition for completion of the subtask (how many threads 4816 // need to finish in order to be done). 4817 pst->set_n_threads(n_threads); 4818 pst->set_n_tasks(1); 4819 assert(pst->valid(), "Error"); 4820 } 4821 4822 // From space 4823 { 4824 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4825 assert(!pst->valid(), "Clobbering existing data?"); 4826 size_t n_tasks = _survivor_chunk_index + 1; 4827 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4828 // Sets the condition for completion of the subtask (how many threads 4829 // need to finish in order to be done). 4830 pst->set_n_threads(n_threads); 4831 pst->set_n_tasks((int)n_tasks); 4832 assert(pst->valid(), "Error"); 4833 } 4834 } 4835 4836 // Parallel version of remark 4837 void CMSCollector::do_remark_parallel() { 4838 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4839 WorkGang* workers = gch->workers(); 4840 assert(workers != NULL, "Need parallel worker threads."); 4841 // Choose to use the number of GC workers most recently set 4842 // into "active_workers". 4843 uint n_workers = workers->active_workers(); 4844 4845 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4846 4847 StrongRootsScope srs(n_workers); 4848 4849 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4850 4851 // We won't be iterating over the cards in the card table updating 4852 // the younger_gen cards, so we shouldn't call the following else 4853 // the verification code as well as subsequent younger_refs_iterate 4854 // code would get confused. XXX 4855 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4856 4857 // The young gen rescan work will not be done as part of 4858 // process_roots (which currently doesn't know how to 4859 // parallelize such a scan), but rather will be broken up into 4860 // a set of parallel tasks (via the sampling that the [abortable] 4861 // preclean phase did of eden, plus the [two] tasks of 4862 // scanning the [two] survivor spaces. Further fine-grain 4863 // parallelization of the scanning of the survivor spaces 4864 // themselves, and of precleaning of the young gen itself 4865 // is deferred to the future. 4866 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4867 4868 // The dirty card rescan work is broken up into a "sequence" 4869 // of parallel tasks (per constituent space) that are dynamically 4870 // claimed by the parallel threads. 4871 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4872 4873 // It turns out that even when we're using 1 thread, doing the work in a 4874 // separate thread causes wide variance in run times. We can't help this 4875 // in the multi-threaded case, but we special-case n=1 here to get 4876 // repeatable measurements of the 1-thread overhead of the parallel code. 4877 if (n_workers > 1) { 4878 // Make refs discovery MT-safe, if it isn't already: it may not 4879 // necessarily be so, since it's possible that we are doing 4880 // ST marking. 4881 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4882 workers->run_task(&tsk); 4883 } else { 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4885 tsk.work(0); 4886 } 4887 4888 // restore, single-threaded for now, any preserved marks 4889 // as a result of work_q overflow 4890 restore_preserved_marks_if_any(); 4891 } 4892 4893 // Non-parallel version of remark 4894 void CMSCollector::do_remark_non_parallel() { 4895 ResourceMark rm; 4896 HandleMark hm; 4897 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4898 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4899 4900 MarkRefsIntoAndScanClosure 4901 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4902 &_markStack, this, 4903 false /* should_yield */, false /* not precleaning */); 4904 MarkFromDirtyCardsClosure 4905 markFromDirtyCardsClosure(this, _span, 4906 NULL, // space is set further below 4907 &_markBitMap, &_markStack, &mrias_cl); 4908 { 4909 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4910 // Iterate over the dirty cards, setting the corresponding bits in the 4911 // mod union table. 4912 { 4913 ModUnionClosure modUnionClosure(&_modUnionTable); 4914 _ct->ct_bs()->dirty_card_iterate( 4915 _cmsGen->used_region(), 4916 &modUnionClosure); 4917 } 4918 // Having transferred these marks into the modUnionTable, we just need 4919 // to rescan the marked objects on the dirty cards in the modUnionTable. 4920 // The initial marking may have been done during an asynchronous 4921 // collection so there may be dirty bits in the mod-union table. 4922 const int alignment = 4923 CardTableModRefBS::card_size * BitsPerWord; 4924 { 4925 // ... First handle dirty cards in CMS gen 4926 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4927 MemRegion ur = _cmsGen->used_region(); 4928 HeapWord* lb = ur.start(); 4929 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4930 MemRegion cms_span(lb, ub); 4931 _modUnionTable.dirty_range_iterate_clear(cms_span, 4932 &markFromDirtyCardsClosure); 4933 verify_work_stacks_empty(); 4934 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4935 } 4936 } 4937 if (VerifyDuringGC && 4938 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4939 HandleMark hm; // Discard invalid handles created during verification 4940 Universe::verify(); 4941 } 4942 { 4943 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4944 4945 verify_work_stacks_empty(); 4946 4947 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4948 StrongRootsScope srs(1); 4949 4950 gch->gen_process_roots(&srs, 4951 GenCollectedHeap::OldGen, 4952 true, // young gen as roots 4953 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4954 should_unload_classes(), 4955 &mrias_cl, 4956 NULL, 4957 NULL); // The dirty klasses will be handled below 4958 4959 assert(should_unload_classes() 4960 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4961 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4962 } 4963 4964 { 4965 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4966 4967 verify_work_stacks_empty(); 4968 4969 // Scan all class loader data objects that might have been introduced 4970 // during concurrent marking. 4971 ResourceMark rm; 4972 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4973 for (int i = 0; i < array->length(); i++) { 4974 mrias_cl.do_cld_nv(array->at(i)); 4975 } 4976 4977 // We don't need to keep track of new CLDs anymore. 4978 ClassLoaderDataGraph::remember_new_clds(false); 4979 4980 verify_work_stacks_empty(); 4981 } 4982 4983 { 4984 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4985 4986 verify_work_stacks_empty(); 4987 4988 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4989 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4990 4991 verify_work_stacks_empty(); 4992 } 4993 4994 // We might have added oops to ClassLoaderData::_handles during the 4995 // concurrent marking phase. These oops point to newly allocated objects 4996 // that are guaranteed to be kept alive. Either by the direct allocation 4997 // code, or when the young collector processes the roots. Hence, 4998 // we don't have to revisit the _handles block during the remark phase. 4999 5000 verify_work_stacks_empty(); 5001 // Restore evacuated mark words, if any, used for overflow list links 5002 restore_preserved_marks_if_any(); 5003 5004 verify_overflow_empty(); 5005 } 5006 5007 //////////////////////////////////////////////////////// 5008 // Parallel Reference Processing Task Proxy Class 5009 //////////////////////////////////////////////////////// 5010 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5011 OopTaskQueueSet* _queues; 5012 ParallelTaskTerminator _terminator; 5013 public: 5014 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5015 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5016 ParallelTaskTerminator* terminator() { return &_terminator; } 5017 OopTaskQueueSet* queues() { return _queues; } 5018 }; 5019 5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5021 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5022 CMSCollector* _collector; 5023 CMSBitMap* _mark_bit_map; 5024 const MemRegion _span; 5025 ProcessTask& _task; 5026 5027 public: 5028 CMSRefProcTaskProxy(ProcessTask& task, 5029 CMSCollector* collector, 5030 const MemRegion& span, 5031 CMSBitMap* mark_bit_map, 5032 AbstractWorkGang* workers, 5033 OopTaskQueueSet* task_queues): 5034 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5035 task_queues, 5036 workers->active_workers()), 5037 _task(task), 5038 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5039 { 5040 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5041 "Inconsistency in _span"); 5042 } 5043 5044 OopTaskQueueSet* task_queues() { return queues(); } 5045 5046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5047 5048 void do_work_steal(int i, 5049 CMSParDrainMarkingStackClosure* drain, 5050 CMSParKeepAliveClosure* keep_alive, 5051 int* seed); 5052 5053 virtual void work(uint worker_id); 5054 }; 5055 5056 void CMSRefProcTaskProxy::work(uint worker_id) { 5057 ResourceMark rm; 5058 HandleMark hm; 5059 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5060 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5061 _mark_bit_map, 5062 work_queue(worker_id)); 5063 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5064 _mark_bit_map, 5065 work_queue(worker_id)); 5066 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5067 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5068 if (_task.marks_oops_alive()) { 5069 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5070 _collector->hash_seed(worker_id)); 5071 } 5072 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5073 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5074 } 5075 5076 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5077 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5078 EnqueueTask& _task; 5079 5080 public: 5081 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5082 : AbstractGangTask("Enqueue reference objects in parallel"), 5083 _task(task) 5084 { } 5085 5086 virtual void work(uint worker_id) 5087 { 5088 _task.work(worker_id); 5089 } 5090 }; 5091 5092 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5093 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5094 _span(span), 5095 _bit_map(bit_map), 5096 _work_queue(work_queue), 5097 _mark_and_push(collector, span, bit_map, work_queue), 5098 _low_water_mark(MIN2((work_queue->max_elems()/4), 5099 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5100 { } 5101 5102 // . see if we can share work_queues with ParNew? XXX 5103 void CMSRefProcTaskProxy::do_work_steal(int i, 5104 CMSParDrainMarkingStackClosure* drain, 5105 CMSParKeepAliveClosure* keep_alive, 5106 int* seed) { 5107 OopTaskQueue* work_q = work_queue(i); 5108 NOT_PRODUCT(int num_steals = 0;) 5109 oop obj_to_scan; 5110 5111 while (true) { 5112 // Completely finish any left over work from (an) earlier round(s) 5113 drain->trim_queue(0); 5114 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5115 (size_t)ParGCDesiredObjsFromOverflowList); 5116 // Now check if there's any work in the overflow list 5117 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5118 // only affects the number of attempts made to get work from the 5119 // overflow list and does not affect the number of workers. Just 5120 // pass ParallelGCThreads so this behavior is unchanged. 5121 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5122 work_q, 5123 ParallelGCThreads)) { 5124 // Found something in global overflow list; 5125 // not yet ready to go stealing work from others. 5126 // We'd like to assert(work_q->size() != 0, ...) 5127 // because we just took work from the overflow list, 5128 // but of course we can't, since all of that might have 5129 // been already stolen from us. 5130 continue; 5131 } 5132 // Verify that we have no work before we resort to stealing 5133 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5134 // Try to steal from other queues that have work 5135 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5136 NOT_PRODUCT(num_steals++;) 5137 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5138 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5139 // Do scanning work 5140 obj_to_scan->oop_iterate(keep_alive); 5141 // Loop around, finish this work, and try to steal some more 5142 } else if (terminator()->offer_termination()) { 5143 break; // nirvana from the infinite cycle 5144 } 5145 } 5146 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5147 } 5148 5149 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5150 { 5151 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5152 WorkGang* workers = gch->workers(); 5153 assert(workers != NULL, "Need parallel worker threads."); 5154 CMSRefProcTaskProxy rp_task(task, &_collector, 5155 _collector.ref_processor()->span(), 5156 _collector.markBitMap(), 5157 workers, _collector.task_queues()); 5158 workers->run_task(&rp_task); 5159 } 5160 5161 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5162 { 5163 5164 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5165 WorkGang* workers = gch->workers(); 5166 assert(workers != NULL, "Need parallel worker threads."); 5167 CMSRefEnqueueTaskProxy enq_task(task); 5168 workers->run_task(&enq_task); 5169 } 5170 5171 void CMSCollector::refProcessingWork() { 5172 ResourceMark rm; 5173 HandleMark hm; 5174 5175 ReferenceProcessor* rp = ref_processor(); 5176 assert(rp->span().equals(_span), "Spans should be equal"); 5177 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5178 // Process weak references. 5179 rp->setup_policy(false); 5180 verify_work_stacks_empty(); 5181 5182 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5183 &_markStack, false /* !preclean */); 5184 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5185 _span, &_markBitMap, &_markStack, 5186 &cmsKeepAliveClosure, false /* !preclean */); 5187 { 5188 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5189 5190 ReferenceProcessorStats stats; 5191 if (rp->processing_is_mt()) { 5192 // Set the degree of MT here. If the discovery is done MT, there 5193 // may have been a different number of threads doing the discovery 5194 // and a different number of discovered lists may have Ref objects. 5195 // That is OK as long as the Reference lists are balanced (see 5196 // balance_all_queues() and balance_queues()). 5197 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5198 uint active_workers = ParallelGCThreads; 5199 WorkGang* workers = gch->workers(); 5200 if (workers != NULL) { 5201 active_workers = workers->active_workers(); 5202 // The expectation is that active_workers will have already 5203 // been set to a reasonable value. If it has not been set, 5204 // investigate. 5205 assert(active_workers > 0, "Should have been set during scavenge"); 5206 } 5207 rp->set_active_mt_degree(active_workers); 5208 CMSRefProcTaskExecutor task_executor(*this); 5209 stats = rp->process_discovered_references(&_is_alive_closure, 5210 &cmsKeepAliveClosure, 5211 &cmsDrainMarkingStackClosure, 5212 &task_executor, 5213 _gc_timer_cm); 5214 } else { 5215 stats = rp->process_discovered_references(&_is_alive_closure, 5216 &cmsKeepAliveClosure, 5217 &cmsDrainMarkingStackClosure, 5218 NULL, 5219 _gc_timer_cm); 5220 } 5221 _gc_tracer_cm->report_gc_reference_stats(stats); 5222 5223 } 5224 5225 // This is the point where the entire marking should have completed. 5226 verify_work_stacks_empty(); 5227 5228 if (should_unload_classes()) { 5229 { 5230 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5231 5232 // Unload classes and purge the SystemDictionary. 5233 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5234 5235 // Unload nmethods. 5236 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5237 5238 // Prune dead klasses from subklass/sibling/implementor lists. 5239 Klass::clean_weak_klass_links(&_is_alive_closure); 5240 } 5241 5242 { 5243 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5244 // Clean up unreferenced symbols in symbol table. 5245 SymbolTable::unlink(); 5246 } 5247 5248 { 5249 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5250 // Delete entries for dead interned strings. 5251 StringTable::unlink(&_is_alive_closure); 5252 } 5253 } 5254 5255 5256 // Restore any preserved marks as a result of mark stack or 5257 // work queue overflow 5258 restore_preserved_marks_if_any(); // done single-threaded for now 5259 5260 rp->set_enqueuing_is_done(true); 5261 if (rp->processing_is_mt()) { 5262 rp->balance_all_queues(); 5263 CMSRefProcTaskExecutor task_executor(*this); 5264 rp->enqueue_discovered_references(&task_executor); 5265 } else { 5266 rp->enqueue_discovered_references(NULL); 5267 } 5268 rp->verify_no_references_recorded(); 5269 assert(!rp->discovery_enabled(), "should have been disabled"); 5270 } 5271 5272 #ifndef PRODUCT 5273 void CMSCollector::check_correct_thread_executing() { 5274 Thread* t = Thread::current(); 5275 // Only the VM thread or the CMS thread should be here. 5276 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5277 "Unexpected thread type"); 5278 // If this is the vm thread, the foreground process 5279 // should not be waiting. Note that _foregroundGCIsActive is 5280 // true while the foreground collector is waiting. 5281 if (_foregroundGCShouldWait) { 5282 // We cannot be the VM thread 5283 assert(t->is_ConcurrentGC_thread(), 5284 "Should be CMS thread"); 5285 } else { 5286 // We can be the CMS thread only if we are in a stop-world 5287 // phase of CMS collection. 5288 if (t->is_ConcurrentGC_thread()) { 5289 assert(_collectorState == InitialMarking || 5290 _collectorState == FinalMarking, 5291 "Should be a stop-world phase"); 5292 // The CMS thread should be holding the CMS_token. 5293 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5294 "Potential interference with concurrently " 5295 "executing VM thread"); 5296 } 5297 } 5298 } 5299 #endif 5300 5301 void CMSCollector::sweep() { 5302 assert(_collectorState == Sweeping, "just checking"); 5303 check_correct_thread_executing(); 5304 verify_work_stacks_empty(); 5305 verify_overflow_empty(); 5306 increment_sweep_count(); 5307 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5308 5309 _inter_sweep_timer.stop(); 5310 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5311 5312 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5313 _intra_sweep_timer.reset(); 5314 _intra_sweep_timer.start(); 5315 { 5316 GCTraceCPUTime tcpu; 5317 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5318 // First sweep the old gen 5319 { 5320 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5321 bitMapLock()); 5322 sweepWork(_cmsGen); 5323 } 5324 5325 // Update Universe::_heap_*_at_gc figures. 5326 // We need all the free list locks to make the abstract state 5327 // transition from Sweeping to Resetting. See detailed note 5328 // further below. 5329 { 5330 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5331 // Update heap occupancy information which is used as 5332 // input to soft ref clearing policy at the next gc. 5333 Universe::update_heap_info_at_gc(); 5334 _collectorState = Resizing; 5335 } 5336 } 5337 verify_work_stacks_empty(); 5338 verify_overflow_empty(); 5339 5340 if (should_unload_classes()) { 5341 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5342 // requires that the virtual spaces are stable and not deleted. 5343 ClassLoaderDataGraph::set_should_purge(true); 5344 } 5345 5346 _intra_sweep_timer.stop(); 5347 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5348 5349 _inter_sweep_timer.reset(); 5350 _inter_sweep_timer.start(); 5351 5352 // We need to use a monotonically non-decreasing time in ms 5353 // or we will see time-warp warnings and os::javaTimeMillis() 5354 // does not guarantee monotonicity. 5355 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5356 update_time_of_last_gc(now); 5357 5358 // NOTE on abstract state transitions: 5359 // Mutators allocate-live and/or mark the mod-union table dirty 5360 // based on the state of the collection. The former is done in 5361 // the interval [Marking, Sweeping] and the latter in the interval 5362 // [Marking, Sweeping). Thus the transitions into the Marking state 5363 // and out of the Sweeping state must be synchronously visible 5364 // globally to the mutators. 5365 // The transition into the Marking state happens with the world 5366 // stopped so the mutators will globally see it. Sweeping is 5367 // done asynchronously by the background collector so the transition 5368 // from the Sweeping state to the Resizing state must be done 5369 // under the freelistLock (as is the check for whether to 5370 // allocate-live and whether to dirty the mod-union table). 5371 assert(_collectorState == Resizing, "Change of collector state to" 5372 " Resizing must be done under the freelistLocks (plural)"); 5373 5374 // Now that sweeping has been completed, we clear 5375 // the incremental_collection_failed flag, 5376 // thus inviting a younger gen collection to promote into 5377 // this generation. If such a promotion may still fail, 5378 // the flag will be set again when a young collection is 5379 // attempted. 5380 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5381 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5382 gch->update_full_collections_completed(_collection_count_start); 5383 } 5384 5385 // FIX ME!!! Looks like this belongs in CFLSpace, with 5386 // CMSGen merely delegating to it. 5387 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5388 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5389 HeapWord* minAddr = _cmsSpace->bottom(); 5390 HeapWord* largestAddr = 5391 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5392 if (largestAddr == NULL) { 5393 // The dictionary appears to be empty. In this case 5394 // try to coalesce at the end of the heap. 5395 largestAddr = _cmsSpace->end(); 5396 } 5397 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5398 size_t nearLargestOffset = 5399 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5400 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5401 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5402 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5403 } 5404 5405 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5406 return addr >= _cmsSpace->nearLargestChunk(); 5407 } 5408 5409 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5410 return _cmsSpace->find_chunk_at_end(); 5411 } 5412 5413 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5414 bool full) { 5415 // If the young generation has been collected, gather any statistics 5416 // that are of interest at this point. 5417 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5418 if (!full && current_is_young) { 5419 // Gather statistics on the young generation collection. 5420 collector()->stats().record_gc0_end(used()); 5421 } 5422 } 5423 5424 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5425 // We iterate over the space(s) underlying this generation, 5426 // checking the mark bit map to see if the bits corresponding 5427 // to specific blocks are marked or not. Blocks that are 5428 // marked are live and are not swept up. All remaining blocks 5429 // are swept up, with coalescing on-the-fly as we sweep up 5430 // contiguous free and/or garbage blocks: 5431 // We need to ensure that the sweeper synchronizes with allocators 5432 // and stop-the-world collectors. In particular, the following 5433 // locks are used: 5434 // . CMS token: if this is held, a stop the world collection cannot occur 5435 // . freelistLock: if this is held no allocation can occur from this 5436 // generation by another thread 5437 // . bitMapLock: if this is held, no other thread can access or update 5438 // 5439 5440 // Note that we need to hold the freelistLock if we use 5441 // block iterate below; else the iterator might go awry if 5442 // a mutator (or promotion) causes block contents to change 5443 // (for instance if the allocator divvies up a block). 5444 // If we hold the free list lock, for all practical purposes 5445 // young generation GC's can't occur (they'll usually need to 5446 // promote), so we might as well prevent all young generation 5447 // GC's while we do a sweeping step. For the same reason, we might 5448 // as well take the bit map lock for the entire duration 5449 5450 // check that we hold the requisite locks 5451 assert(have_cms_token(), "Should hold cms token"); 5452 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5453 assert_lock_strong(old_gen->freelistLock()); 5454 assert_lock_strong(bitMapLock()); 5455 5456 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5457 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5458 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5459 _inter_sweep_estimate.padded_average(), 5460 _intra_sweep_estimate.padded_average()); 5461 old_gen->setNearLargestChunk(); 5462 5463 { 5464 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5465 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5466 // We need to free-up/coalesce garbage/blocks from a 5467 // co-terminal free run. This is done in the SweepClosure 5468 // destructor; so, do not remove this scope, else the 5469 // end-of-sweep-census below will be off by a little bit. 5470 } 5471 old_gen->cmsSpace()->sweep_completed(); 5472 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5473 if (should_unload_classes()) { // unloaded classes this cycle, 5474 _concurrent_cycles_since_last_unload = 0; // ... reset count 5475 } else { // did not unload classes, 5476 _concurrent_cycles_since_last_unload++; // ... increment count 5477 } 5478 } 5479 5480 // Reset CMS data structures (for now just the marking bit map) 5481 // preparatory for the next cycle. 5482 void CMSCollector::reset_concurrent() { 5483 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5484 5485 // If the state is not "Resetting", the foreground thread 5486 // has done a collection and the resetting. 5487 if (_collectorState != Resetting) { 5488 assert(_collectorState == Idling, "The state should only change" 5489 " because the foreground collector has finished the collection"); 5490 return; 5491 } 5492 5493 { 5494 // Clear the mark bitmap (no grey objects to start with) 5495 // for the next cycle. 5496 GCTraceCPUTime tcpu; 5497 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5498 5499 HeapWord* curAddr = _markBitMap.startWord(); 5500 while (curAddr < _markBitMap.endWord()) { 5501 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5502 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5503 _markBitMap.clear_large_range(chunk); 5504 if (ConcurrentMarkSweepThread::should_yield() && 5505 !foregroundGCIsActive() && 5506 CMSYield) { 5507 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5508 "CMS thread should hold CMS token"); 5509 assert_lock_strong(bitMapLock()); 5510 bitMapLock()->unlock(); 5511 ConcurrentMarkSweepThread::desynchronize(true); 5512 stopTimer(); 5513 incrementYields(); 5514 5515 // See the comment in coordinator_yield() 5516 for (unsigned i = 0; i < CMSYieldSleepCount && 5517 ConcurrentMarkSweepThread::should_yield() && 5518 !CMSCollector::foregroundGCIsActive(); ++i) { 5519 os::sleep(Thread::current(), 1, false); 5520 } 5521 5522 ConcurrentMarkSweepThread::synchronize(true); 5523 bitMapLock()->lock_without_safepoint_check(); 5524 startTimer(); 5525 } 5526 curAddr = chunk.end(); 5527 } 5528 // A successful mostly concurrent collection has been done. 5529 // Because only the full (i.e., concurrent mode failure) collections 5530 // are being measured for gc overhead limits, clean the "near" flag 5531 // and count. 5532 size_policy()->reset_gc_overhead_limit_count(); 5533 _collectorState = Idling; 5534 } 5535 5536 register_gc_end(); 5537 } 5538 5539 // Same as above but for STW paths 5540 void CMSCollector::reset_stw() { 5541 // already have the lock 5542 assert(_collectorState == Resetting, "just checking"); 5543 assert_lock_strong(bitMapLock()); 5544 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5545 _markBitMap.clear_all(); 5546 _collectorState = Idling; 5547 register_gc_end(); 5548 } 5549 5550 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5551 GCTraceCPUTime tcpu; 5552 TraceCollectorStats tcs(counters()); 5553 5554 switch (op) { 5555 case CMS_op_checkpointRootsInitial: { 5556 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5557 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5558 checkpointRootsInitial(); 5559 break; 5560 } 5561 case CMS_op_checkpointRootsFinal: { 5562 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5563 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5564 checkpointRootsFinal(); 5565 break; 5566 } 5567 default: 5568 fatal("No such CMS_op"); 5569 } 5570 } 5571 5572 #ifndef PRODUCT 5573 size_t const CMSCollector::skip_header_HeapWords() { 5574 return FreeChunk::header_size(); 5575 } 5576 5577 // Try and collect here conditions that should hold when 5578 // CMS thread is exiting. The idea is that the foreground GC 5579 // thread should not be blocked if it wants to terminate 5580 // the CMS thread and yet continue to run the VM for a while 5581 // after that. 5582 void CMSCollector::verify_ok_to_terminate() const { 5583 assert(Thread::current()->is_ConcurrentGC_thread(), 5584 "should be called by CMS thread"); 5585 assert(!_foregroundGCShouldWait, "should be false"); 5586 // We could check here that all the various low-level locks 5587 // are not held by the CMS thread, but that is overkill; see 5588 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5589 // is checked. 5590 } 5591 #endif 5592 5593 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5594 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5595 "missing Printezis mark?"); 5596 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5597 size_t size = pointer_delta(nextOneAddr + 1, addr); 5598 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5599 "alignment problem"); 5600 assert(size >= 3, "Necessary for Printezis marks to work"); 5601 return size; 5602 } 5603 5604 // A variant of the above (block_size_using_printezis_bits()) except 5605 // that we return 0 if the P-bits are not yet set. 5606 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5607 if (_markBitMap.isMarked(addr + 1)) { 5608 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5609 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5610 size_t size = pointer_delta(nextOneAddr + 1, addr); 5611 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5612 "alignment problem"); 5613 assert(size >= 3, "Necessary for Printezis marks to work"); 5614 return size; 5615 } 5616 return 0; 5617 } 5618 5619 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5620 size_t sz = 0; 5621 oop p = (oop)addr; 5622 if (p->klass_or_null() != NULL) { 5623 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5624 } else { 5625 sz = block_size_using_printezis_bits(addr); 5626 } 5627 assert(sz > 0, "size must be nonzero"); 5628 HeapWord* next_block = addr + sz; 5629 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5630 CardTableModRefBS::card_size); 5631 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5632 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5633 "must be different cards"); 5634 return next_card; 5635 } 5636 5637 5638 // CMS Bit Map Wrapper ///////////////////////////////////////// 5639 5640 // Construct a CMS bit map infrastructure, but don't create the 5641 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5642 // further below. 5643 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5644 _bm(), 5645 _shifter(shifter), 5646 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5647 Monitor::_safepoint_check_sometimes) : NULL) 5648 { 5649 _bmStartWord = 0; 5650 _bmWordSize = 0; 5651 } 5652 5653 bool CMSBitMap::allocate(MemRegion mr) { 5654 _bmStartWord = mr.start(); 5655 _bmWordSize = mr.word_size(); 5656 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5657 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5658 if (!brs.is_reserved()) { 5659 warning("CMS bit map allocation failure"); 5660 return false; 5661 } 5662 // For now we'll just commit all of the bit map up front. 5663 // Later on we'll try to be more parsimonious with swap. 5664 if (!_virtual_space.initialize(brs, brs.size())) { 5665 warning("CMS bit map backing store failure"); 5666 return false; 5667 } 5668 assert(_virtual_space.committed_size() == brs.size(), 5669 "didn't reserve backing store for all of CMS bit map?"); 5670 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5671 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5672 _bmWordSize, "inconsistency in bit map sizing"); 5673 _bm.set_size(_bmWordSize >> _shifter); 5674 5675 // bm.clear(); // can we rely on getting zero'd memory? verify below 5676 assert(isAllClear(), 5677 "Expected zero'd memory from ReservedSpace constructor"); 5678 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5679 "consistency check"); 5680 return true; 5681 } 5682 5683 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5684 HeapWord *next_addr, *end_addr, *last_addr; 5685 assert_locked(); 5686 assert(covers(mr), "out-of-range error"); 5687 // XXX assert that start and end are appropriately aligned 5688 for (next_addr = mr.start(), end_addr = mr.end(); 5689 next_addr < end_addr; next_addr = last_addr) { 5690 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5691 last_addr = dirty_region.end(); 5692 if (!dirty_region.is_empty()) { 5693 cl->do_MemRegion(dirty_region); 5694 } else { 5695 assert(last_addr == end_addr, "program logic"); 5696 return; 5697 } 5698 } 5699 } 5700 5701 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5702 _bm.print_on_error(st, prefix); 5703 } 5704 5705 #ifndef PRODUCT 5706 void CMSBitMap::assert_locked() const { 5707 CMSLockVerifier::assert_locked(lock()); 5708 } 5709 5710 bool CMSBitMap::covers(MemRegion mr) const { 5711 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5712 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5713 "size inconsistency"); 5714 return (mr.start() >= _bmStartWord) && 5715 (mr.end() <= endWord()); 5716 } 5717 5718 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5719 return (start >= _bmStartWord && (start + size) <= endWord()); 5720 } 5721 5722 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5723 // verify that there are no 1 bits in the interval [left, right) 5724 FalseBitMapClosure falseBitMapClosure; 5725 iterate(&falseBitMapClosure, left, right); 5726 } 5727 5728 void CMSBitMap::region_invariant(MemRegion mr) 5729 { 5730 assert_locked(); 5731 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5732 assert(!mr.is_empty(), "unexpected empty region"); 5733 assert(covers(mr), "mr should be covered by bit map"); 5734 // convert address range into offset range 5735 size_t start_ofs = heapWordToOffset(mr.start()); 5736 // Make sure that end() is appropriately aligned 5737 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5738 (1 << (_shifter+LogHeapWordSize))), 5739 "Misaligned mr.end()"); 5740 size_t end_ofs = heapWordToOffset(mr.end()); 5741 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5742 } 5743 5744 #endif 5745 5746 bool CMSMarkStack::allocate(size_t size) { 5747 // allocate a stack of the requisite depth 5748 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5749 size * sizeof(oop))); 5750 if (!rs.is_reserved()) { 5751 warning("CMSMarkStack allocation failure"); 5752 return false; 5753 } 5754 if (!_virtual_space.initialize(rs, rs.size())) { 5755 warning("CMSMarkStack backing store failure"); 5756 return false; 5757 } 5758 assert(_virtual_space.committed_size() == rs.size(), 5759 "didn't reserve backing store for all of CMS stack?"); 5760 _base = (oop*)(_virtual_space.low()); 5761 _index = 0; 5762 _capacity = size; 5763 NOT_PRODUCT(_max_depth = 0); 5764 return true; 5765 } 5766 5767 // XXX FIX ME !!! In the MT case we come in here holding a 5768 // leaf lock. For printing we need to take a further lock 5769 // which has lower rank. We need to recalibrate the two 5770 // lock-ranks involved in order to be able to print the 5771 // messages below. (Or defer the printing to the caller. 5772 // For now we take the expedient path of just disabling the 5773 // messages for the problematic case.) 5774 void CMSMarkStack::expand() { 5775 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5776 if (_capacity == MarkStackSizeMax) { 5777 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5778 // We print a warning message only once per CMS cycle. 5779 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5780 } 5781 return; 5782 } 5783 // Double capacity if possible 5784 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5785 // Do not give up existing stack until we have managed to 5786 // get the double capacity that we desired. 5787 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5788 new_capacity * sizeof(oop))); 5789 if (rs.is_reserved()) { 5790 // Release the backing store associated with old stack 5791 _virtual_space.release(); 5792 // Reinitialize virtual space for new stack 5793 if (!_virtual_space.initialize(rs, rs.size())) { 5794 fatal("Not enough swap for expanded marking stack"); 5795 } 5796 _base = (oop*)(_virtual_space.low()); 5797 _index = 0; 5798 _capacity = new_capacity; 5799 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5800 // Failed to double capacity, continue; 5801 // we print a detail message only once per CMS cycle. 5802 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5803 _capacity / K, new_capacity / K); 5804 } 5805 } 5806 5807 5808 // Closures 5809 // XXX: there seems to be a lot of code duplication here; 5810 // should refactor and consolidate common code. 5811 5812 // This closure is used to mark refs into the CMS generation in 5813 // the CMS bit map. Called at the first checkpoint. This closure 5814 // assumes that we do not need to re-mark dirty cards; if the CMS 5815 // generation on which this is used is not an oldest 5816 // generation then this will lose younger_gen cards! 5817 5818 MarkRefsIntoClosure::MarkRefsIntoClosure( 5819 MemRegion span, CMSBitMap* bitMap): 5820 _span(span), 5821 _bitMap(bitMap) 5822 { 5823 assert(ref_processor() == NULL, "deliberately left NULL"); 5824 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5825 } 5826 5827 void MarkRefsIntoClosure::do_oop(oop obj) { 5828 // if p points into _span, then mark corresponding bit in _markBitMap 5829 assert(obj->is_oop(), "expected an oop"); 5830 HeapWord* addr = (HeapWord*)obj; 5831 if (_span.contains(addr)) { 5832 // this should be made more efficient 5833 _bitMap->mark(addr); 5834 } 5835 } 5836 5837 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5838 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5839 5840 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5841 MemRegion span, CMSBitMap* bitMap): 5842 _span(span), 5843 _bitMap(bitMap) 5844 { 5845 assert(ref_processor() == NULL, "deliberately left NULL"); 5846 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5847 } 5848 5849 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5850 // if p points into _span, then mark corresponding bit in _markBitMap 5851 assert(obj->is_oop(), "expected an oop"); 5852 HeapWord* addr = (HeapWord*)obj; 5853 if (_span.contains(addr)) { 5854 // this should be made more efficient 5855 _bitMap->par_mark(addr); 5856 } 5857 } 5858 5859 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5860 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5861 5862 // A variant of the above, used for CMS marking verification. 5863 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5864 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5865 _span(span), 5866 _verification_bm(verification_bm), 5867 _cms_bm(cms_bm) 5868 { 5869 assert(ref_processor() == NULL, "deliberately left NULL"); 5870 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5871 } 5872 5873 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5874 // if p points into _span, then mark corresponding bit in _markBitMap 5875 assert(obj->is_oop(), "expected an oop"); 5876 HeapWord* addr = (HeapWord*)obj; 5877 if (_span.contains(addr)) { 5878 _verification_bm->mark(addr); 5879 if (!_cms_bm->isMarked(addr)) { 5880 LogHandle(gc, verify) log; 5881 ResourceMark rm; 5882 oop(addr)->print_on(log.error_stream()); 5883 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5884 fatal("... aborting"); 5885 } 5886 } 5887 } 5888 5889 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5890 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5891 5892 ////////////////////////////////////////////////// 5893 // MarkRefsIntoAndScanClosure 5894 ////////////////////////////////////////////////// 5895 5896 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5897 ReferenceProcessor* rp, 5898 CMSBitMap* bit_map, 5899 CMSBitMap* mod_union_table, 5900 CMSMarkStack* mark_stack, 5901 CMSCollector* collector, 5902 bool should_yield, 5903 bool concurrent_precleaning): 5904 _collector(collector), 5905 _span(span), 5906 _bit_map(bit_map), 5907 _mark_stack(mark_stack), 5908 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5909 mark_stack, concurrent_precleaning), 5910 _yield(should_yield), 5911 _concurrent_precleaning(concurrent_precleaning), 5912 _freelistLock(NULL) 5913 { 5914 // FIXME: Should initialize in base class constructor. 5915 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5916 set_ref_processor_internal(rp); 5917 } 5918 5919 // This closure is used to mark refs into the CMS generation at the 5920 // second (final) checkpoint, and to scan and transitively follow 5921 // the unmarked oops. It is also used during the concurrent precleaning 5922 // phase while scanning objects on dirty cards in the CMS generation. 5923 // The marks are made in the marking bit map and the marking stack is 5924 // used for keeping the (newly) grey objects during the scan. 5925 // The parallel version (Par_...) appears further below. 5926 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5927 if (obj != NULL) { 5928 assert(obj->is_oop(), "expected an oop"); 5929 HeapWord* addr = (HeapWord*)obj; 5930 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5931 assert(_collector->overflow_list_is_empty(), 5932 "overflow list should be empty"); 5933 if (_span.contains(addr) && 5934 !_bit_map->isMarked(addr)) { 5935 // mark bit map (object is now grey) 5936 _bit_map->mark(addr); 5937 // push on marking stack (stack should be empty), and drain the 5938 // stack by applying this closure to the oops in the oops popped 5939 // from the stack (i.e. blacken the grey objects) 5940 bool res = _mark_stack->push(obj); 5941 assert(res, "Should have space to push on empty stack"); 5942 do { 5943 oop new_oop = _mark_stack->pop(); 5944 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5945 assert(_bit_map->isMarked((HeapWord*)new_oop), 5946 "only grey objects on this stack"); 5947 // iterate over the oops in this oop, marking and pushing 5948 // the ones in CMS heap (i.e. in _span). 5949 new_oop->oop_iterate(&_pushAndMarkClosure); 5950 // check if it's time to yield 5951 do_yield_check(); 5952 } while (!_mark_stack->isEmpty() || 5953 (!_concurrent_precleaning && take_from_overflow_list())); 5954 // if marking stack is empty, and we are not doing this 5955 // during precleaning, then check the overflow list 5956 } 5957 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5958 assert(_collector->overflow_list_is_empty(), 5959 "overflow list was drained above"); 5960 5961 assert(_collector->no_preserved_marks(), 5962 "All preserved marks should have been restored above"); 5963 } 5964 } 5965 5966 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5967 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5968 5969 void MarkRefsIntoAndScanClosure::do_yield_work() { 5970 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5971 "CMS thread should hold CMS token"); 5972 assert_lock_strong(_freelistLock); 5973 assert_lock_strong(_bit_map->lock()); 5974 // relinquish the free_list_lock and bitMaplock() 5975 _bit_map->lock()->unlock(); 5976 _freelistLock->unlock(); 5977 ConcurrentMarkSweepThread::desynchronize(true); 5978 _collector->stopTimer(); 5979 _collector->incrementYields(); 5980 5981 // See the comment in coordinator_yield() 5982 for (unsigned i = 0; 5983 i < CMSYieldSleepCount && 5984 ConcurrentMarkSweepThread::should_yield() && 5985 !CMSCollector::foregroundGCIsActive(); 5986 ++i) { 5987 os::sleep(Thread::current(), 1, false); 5988 } 5989 5990 ConcurrentMarkSweepThread::synchronize(true); 5991 _freelistLock->lock_without_safepoint_check(); 5992 _bit_map->lock()->lock_without_safepoint_check(); 5993 _collector->startTimer(); 5994 } 5995 5996 /////////////////////////////////////////////////////////// 5997 // ParMarkRefsIntoAndScanClosure: a parallel version of 5998 // MarkRefsIntoAndScanClosure 5999 /////////////////////////////////////////////////////////// 6000 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6001 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6002 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6003 _span(span), 6004 _bit_map(bit_map), 6005 _work_queue(work_queue), 6006 _low_water_mark(MIN2((work_queue->max_elems()/4), 6007 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6008 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6009 { 6010 // FIXME: Should initialize in base class constructor. 6011 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6012 set_ref_processor_internal(rp); 6013 } 6014 6015 // This closure is used to mark refs into the CMS generation at the 6016 // second (final) checkpoint, and to scan and transitively follow 6017 // the unmarked oops. The marks are made in the marking bit map and 6018 // the work_queue is used for keeping the (newly) grey objects during 6019 // the scan phase whence they are also available for stealing by parallel 6020 // threads. Since the marking bit map is shared, updates are 6021 // synchronized (via CAS). 6022 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6023 if (obj != NULL) { 6024 // Ignore mark word because this could be an already marked oop 6025 // that may be chained at the end of the overflow list. 6026 assert(obj->is_oop(true), "expected an oop"); 6027 HeapWord* addr = (HeapWord*)obj; 6028 if (_span.contains(addr) && 6029 !_bit_map->isMarked(addr)) { 6030 // mark bit map (object will become grey): 6031 // It is possible for several threads to be 6032 // trying to "claim" this object concurrently; 6033 // the unique thread that succeeds in marking the 6034 // object first will do the subsequent push on 6035 // to the work queue (or overflow list). 6036 if (_bit_map->par_mark(addr)) { 6037 // push on work_queue (which may not be empty), and trim the 6038 // queue to an appropriate length by applying this closure to 6039 // the oops in the oops popped from the stack (i.e. blacken the 6040 // grey objects) 6041 bool res = _work_queue->push(obj); 6042 assert(res, "Low water mark should be less than capacity?"); 6043 trim_queue(_low_water_mark); 6044 } // Else, another thread claimed the object 6045 } 6046 } 6047 } 6048 6049 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6050 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6051 6052 // This closure is used to rescan the marked objects on the dirty cards 6053 // in the mod union table and the card table proper. 6054 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6055 oop p, MemRegion mr) { 6056 6057 size_t size = 0; 6058 HeapWord* addr = (HeapWord*)p; 6059 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6060 assert(_span.contains(addr), "we are scanning the CMS generation"); 6061 // check if it's time to yield 6062 if (do_yield_check()) { 6063 // We yielded for some foreground stop-world work, 6064 // and we have been asked to abort this ongoing preclean cycle. 6065 return 0; 6066 } 6067 if (_bitMap->isMarked(addr)) { 6068 // it's marked; is it potentially uninitialized? 6069 if (p->klass_or_null() != NULL) { 6070 // an initialized object; ignore mark word in verification below 6071 // since we are running concurrent with mutators 6072 assert(p->is_oop(true), "should be an oop"); 6073 if (p->is_objArray()) { 6074 // objArrays are precisely marked; restrict scanning 6075 // to dirty cards only. 6076 size = CompactibleFreeListSpace::adjustObjectSize( 6077 p->oop_iterate_size(_scanningClosure, mr)); 6078 } else { 6079 // A non-array may have been imprecisely marked; we need 6080 // to scan object in its entirety. 6081 size = CompactibleFreeListSpace::adjustObjectSize( 6082 p->oop_iterate_size(_scanningClosure)); 6083 } 6084 #ifdef ASSERT 6085 size_t direct_size = 6086 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6087 assert(size == direct_size, "Inconsistency in size"); 6088 assert(size >= 3, "Necessary for Printezis marks to work"); 6089 if (!_bitMap->isMarked(addr+1)) { 6090 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6091 } else { 6092 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6093 assert(_bitMap->isMarked(addr+size-1), 6094 "inconsistent Printezis mark"); 6095 } 6096 #endif // ASSERT 6097 } else { 6098 // An uninitialized object. 6099 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6100 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6101 size = pointer_delta(nextOneAddr + 1, addr); 6102 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6103 "alignment problem"); 6104 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6105 // will dirty the card when the klass pointer is installed in the 6106 // object (signaling the completion of initialization). 6107 } 6108 } else { 6109 // Either a not yet marked object or an uninitialized object 6110 if (p->klass_or_null() == NULL) { 6111 // An uninitialized object, skip to the next card, since 6112 // we may not be able to read its P-bits yet. 6113 assert(size == 0, "Initial value"); 6114 } else { 6115 // An object not (yet) reached by marking: we merely need to 6116 // compute its size so as to go look at the next block. 6117 assert(p->is_oop(true), "should be an oop"); 6118 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6119 } 6120 } 6121 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6122 return size; 6123 } 6124 6125 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6126 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6127 "CMS thread should hold CMS token"); 6128 assert_lock_strong(_freelistLock); 6129 assert_lock_strong(_bitMap->lock()); 6130 // relinquish the free_list_lock and bitMaplock() 6131 _bitMap->lock()->unlock(); 6132 _freelistLock->unlock(); 6133 ConcurrentMarkSweepThread::desynchronize(true); 6134 _collector->stopTimer(); 6135 _collector->incrementYields(); 6136 6137 // See the comment in coordinator_yield() 6138 for (unsigned i = 0; i < CMSYieldSleepCount && 6139 ConcurrentMarkSweepThread::should_yield() && 6140 !CMSCollector::foregroundGCIsActive(); ++i) { 6141 os::sleep(Thread::current(), 1, false); 6142 } 6143 6144 ConcurrentMarkSweepThread::synchronize(true); 6145 _freelistLock->lock_without_safepoint_check(); 6146 _bitMap->lock()->lock_without_safepoint_check(); 6147 _collector->startTimer(); 6148 } 6149 6150 6151 ////////////////////////////////////////////////////////////////// 6152 // SurvivorSpacePrecleanClosure 6153 ////////////////////////////////////////////////////////////////// 6154 // This (single-threaded) closure is used to preclean the oops in 6155 // the survivor spaces. 6156 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6157 6158 HeapWord* addr = (HeapWord*)p; 6159 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6160 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6161 assert(p->klass_or_null() != NULL, "object should be initialized"); 6162 // an initialized object; ignore mark word in verification below 6163 // since we are running concurrent with mutators 6164 assert(p->is_oop(true), "should be an oop"); 6165 // Note that we do not yield while we iterate over 6166 // the interior oops of p, pushing the relevant ones 6167 // on our marking stack. 6168 size_t size = p->oop_iterate_size(_scanning_closure); 6169 do_yield_check(); 6170 // Observe that below, we do not abandon the preclean 6171 // phase as soon as we should; rather we empty the 6172 // marking stack before returning. This is to satisfy 6173 // some existing assertions. In general, it may be a 6174 // good idea to abort immediately and complete the marking 6175 // from the grey objects at a later time. 6176 while (!_mark_stack->isEmpty()) { 6177 oop new_oop = _mark_stack->pop(); 6178 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6179 assert(_bit_map->isMarked((HeapWord*)new_oop), 6180 "only grey objects on this stack"); 6181 // iterate over the oops in this oop, marking and pushing 6182 // the ones in CMS heap (i.e. in _span). 6183 new_oop->oop_iterate(_scanning_closure); 6184 // check if it's time to yield 6185 do_yield_check(); 6186 } 6187 unsigned int after_count = 6188 GenCollectedHeap::heap()->total_collections(); 6189 bool abort = (_before_count != after_count) || 6190 _collector->should_abort_preclean(); 6191 return abort ? 0 : size; 6192 } 6193 6194 void SurvivorSpacePrecleanClosure::do_yield_work() { 6195 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6196 "CMS thread should hold CMS token"); 6197 assert_lock_strong(_bit_map->lock()); 6198 // Relinquish the bit map lock 6199 _bit_map->lock()->unlock(); 6200 ConcurrentMarkSweepThread::desynchronize(true); 6201 _collector->stopTimer(); 6202 _collector->incrementYields(); 6203 6204 // See the comment in coordinator_yield() 6205 for (unsigned i = 0; i < CMSYieldSleepCount && 6206 ConcurrentMarkSweepThread::should_yield() && 6207 !CMSCollector::foregroundGCIsActive(); ++i) { 6208 os::sleep(Thread::current(), 1, false); 6209 } 6210 6211 ConcurrentMarkSweepThread::synchronize(true); 6212 _bit_map->lock()->lock_without_safepoint_check(); 6213 _collector->startTimer(); 6214 } 6215 6216 // This closure is used to rescan the marked objects on the dirty cards 6217 // in the mod union table and the card table proper. In the parallel 6218 // case, although the bitMap is shared, we do a single read so the 6219 // isMarked() query is "safe". 6220 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6221 // Ignore mark word because we are running concurrent with mutators 6222 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6223 HeapWord* addr = (HeapWord*)p; 6224 assert(_span.contains(addr), "we are scanning the CMS generation"); 6225 bool is_obj_array = false; 6226 #ifdef ASSERT 6227 if (!_parallel) { 6228 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6229 assert(_collector->overflow_list_is_empty(), 6230 "overflow list should be empty"); 6231 6232 } 6233 #endif // ASSERT 6234 if (_bit_map->isMarked(addr)) { 6235 // Obj arrays are precisely marked, non-arrays are not; 6236 // so we scan objArrays precisely and non-arrays in their 6237 // entirety. 6238 if (p->is_objArray()) { 6239 is_obj_array = true; 6240 if (_parallel) { 6241 p->oop_iterate(_par_scan_closure, mr); 6242 } else { 6243 p->oop_iterate(_scan_closure, mr); 6244 } 6245 } else { 6246 if (_parallel) { 6247 p->oop_iterate(_par_scan_closure); 6248 } else { 6249 p->oop_iterate(_scan_closure); 6250 } 6251 } 6252 } 6253 #ifdef ASSERT 6254 if (!_parallel) { 6255 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6256 assert(_collector->overflow_list_is_empty(), 6257 "overflow list should be empty"); 6258 6259 } 6260 #endif // ASSERT 6261 return is_obj_array; 6262 } 6263 6264 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6265 MemRegion span, 6266 CMSBitMap* bitMap, CMSMarkStack* markStack, 6267 bool should_yield, bool verifying): 6268 _collector(collector), 6269 _span(span), 6270 _bitMap(bitMap), 6271 _mut(&collector->_modUnionTable), 6272 _markStack(markStack), 6273 _yield(should_yield), 6274 _skipBits(0) 6275 { 6276 assert(_markStack->isEmpty(), "stack should be empty"); 6277 _finger = _bitMap->startWord(); 6278 _threshold = _finger; 6279 assert(_collector->_restart_addr == NULL, "Sanity check"); 6280 assert(_span.contains(_finger), "Out of bounds _finger?"); 6281 DEBUG_ONLY(_verifying = verifying;) 6282 } 6283 6284 void MarkFromRootsClosure::reset(HeapWord* addr) { 6285 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6286 assert(_span.contains(addr), "Out of bounds _finger?"); 6287 _finger = addr; 6288 _threshold = (HeapWord*)round_to( 6289 (intptr_t)_finger, CardTableModRefBS::card_size); 6290 } 6291 6292 // Should revisit to see if this should be restructured for 6293 // greater efficiency. 6294 bool MarkFromRootsClosure::do_bit(size_t offset) { 6295 if (_skipBits > 0) { 6296 _skipBits--; 6297 return true; 6298 } 6299 // convert offset into a HeapWord* 6300 HeapWord* addr = _bitMap->startWord() + offset; 6301 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6302 "address out of range"); 6303 assert(_bitMap->isMarked(addr), "tautology"); 6304 if (_bitMap->isMarked(addr+1)) { 6305 // this is an allocated but not yet initialized object 6306 assert(_skipBits == 0, "tautology"); 6307 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6308 oop p = oop(addr); 6309 if (p->klass_or_null() == NULL) { 6310 DEBUG_ONLY(if (!_verifying) {) 6311 // We re-dirty the cards on which this object lies and increase 6312 // the _threshold so that we'll come back to scan this object 6313 // during the preclean or remark phase. (CMSCleanOnEnter) 6314 if (CMSCleanOnEnter) { 6315 size_t sz = _collector->block_size_using_printezis_bits(addr); 6316 HeapWord* end_card_addr = (HeapWord*)round_to( 6317 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6318 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6319 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6320 // Bump _threshold to end_card_addr; note that 6321 // _threshold cannot possibly exceed end_card_addr, anyhow. 6322 // This prevents future clearing of the card as the scan proceeds 6323 // to the right. 6324 assert(_threshold <= end_card_addr, 6325 "Because we are just scanning into this object"); 6326 if (_threshold < end_card_addr) { 6327 _threshold = end_card_addr; 6328 } 6329 if (p->klass_or_null() != NULL) { 6330 // Redirty the range of cards... 6331 _mut->mark_range(redirty_range); 6332 } // ...else the setting of klass will dirty the card anyway. 6333 } 6334 DEBUG_ONLY(}) 6335 return true; 6336 } 6337 } 6338 scanOopsInOop(addr); 6339 return true; 6340 } 6341 6342 // We take a break if we've been at this for a while, 6343 // so as to avoid monopolizing the locks involved. 6344 void MarkFromRootsClosure::do_yield_work() { 6345 // First give up the locks, then yield, then re-lock 6346 // We should probably use a constructor/destructor idiom to 6347 // do this unlock/lock or modify the MutexUnlocker class to 6348 // serve our purpose. XXX 6349 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6350 "CMS thread should hold CMS token"); 6351 assert_lock_strong(_bitMap->lock()); 6352 _bitMap->lock()->unlock(); 6353 ConcurrentMarkSweepThread::desynchronize(true); 6354 _collector->stopTimer(); 6355 _collector->incrementYields(); 6356 6357 // See the comment in coordinator_yield() 6358 for (unsigned i = 0; i < CMSYieldSleepCount && 6359 ConcurrentMarkSweepThread::should_yield() && 6360 !CMSCollector::foregroundGCIsActive(); ++i) { 6361 os::sleep(Thread::current(), 1, false); 6362 } 6363 6364 ConcurrentMarkSweepThread::synchronize(true); 6365 _bitMap->lock()->lock_without_safepoint_check(); 6366 _collector->startTimer(); 6367 } 6368 6369 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6370 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6371 assert(_markStack->isEmpty(), 6372 "should drain stack to limit stack usage"); 6373 // convert ptr to an oop preparatory to scanning 6374 oop obj = oop(ptr); 6375 // Ignore mark word in verification below, since we 6376 // may be running concurrent with mutators. 6377 assert(obj->is_oop(true), "should be an oop"); 6378 assert(_finger <= ptr, "_finger runneth ahead"); 6379 // advance the finger to right end of this object 6380 _finger = ptr + obj->size(); 6381 assert(_finger > ptr, "we just incremented it above"); 6382 // On large heaps, it may take us some time to get through 6383 // the marking phase. During 6384 // this time it's possible that a lot of mutations have 6385 // accumulated in the card table and the mod union table -- 6386 // these mutation records are redundant until we have 6387 // actually traced into the corresponding card. 6388 // Here, we check whether advancing the finger would make 6389 // us cross into a new card, and if so clear corresponding 6390 // cards in the MUT (preclean them in the card-table in the 6391 // future). 6392 6393 DEBUG_ONLY(if (!_verifying) {) 6394 // The clean-on-enter optimization is disabled by default, 6395 // until we fix 6178663. 6396 if (CMSCleanOnEnter && (_finger > _threshold)) { 6397 // [_threshold, _finger) represents the interval 6398 // of cards to be cleared in MUT (or precleaned in card table). 6399 // The set of cards to be cleared is all those that overlap 6400 // with the interval [_threshold, _finger); note that 6401 // _threshold is always kept card-aligned but _finger isn't 6402 // always card-aligned. 6403 HeapWord* old_threshold = _threshold; 6404 assert(old_threshold == (HeapWord*)round_to( 6405 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6406 "_threshold should always be card-aligned"); 6407 _threshold = (HeapWord*)round_to( 6408 (intptr_t)_finger, CardTableModRefBS::card_size); 6409 MemRegion mr(old_threshold, _threshold); 6410 assert(!mr.is_empty(), "Control point invariant"); 6411 assert(_span.contains(mr), "Should clear within span"); 6412 _mut->clear_range(mr); 6413 } 6414 DEBUG_ONLY(}) 6415 // Note: the finger doesn't advance while we drain 6416 // the stack below. 6417 PushOrMarkClosure pushOrMarkClosure(_collector, 6418 _span, _bitMap, _markStack, 6419 _finger, this); 6420 bool res = _markStack->push(obj); 6421 assert(res, "Empty non-zero size stack should have space for single push"); 6422 while (!_markStack->isEmpty()) { 6423 oop new_oop = _markStack->pop(); 6424 // Skip verifying header mark word below because we are 6425 // running concurrent with mutators. 6426 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6427 // now scan this oop's oops 6428 new_oop->oop_iterate(&pushOrMarkClosure); 6429 do_yield_check(); 6430 } 6431 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6432 } 6433 6434 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6435 CMSCollector* collector, MemRegion span, 6436 CMSBitMap* bit_map, 6437 OopTaskQueue* work_queue, 6438 CMSMarkStack* overflow_stack): 6439 _collector(collector), 6440 _whole_span(collector->_span), 6441 _span(span), 6442 _bit_map(bit_map), 6443 _mut(&collector->_modUnionTable), 6444 _work_queue(work_queue), 6445 _overflow_stack(overflow_stack), 6446 _skip_bits(0), 6447 _task(task) 6448 { 6449 assert(_work_queue->size() == 0, "work_queue should be empty"); 6450 _finger = span.start(); 6451 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6452 assert(_span.contains(_finger), "Out of bounds _finger?"); 6453 } 6454 6455 // Should revisit to see if this should be restructured for 6456 // greater efficiency. 6457 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6458 if (_skip_bits > 0) { 6459 _skip_bits--; 6460 return true; 6461 } 6462 // convert offset into a HeapWord* 6463 HeapWord* addr = _bit_map->startWord() + offset; 6464 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6465 "address out of range"); 6466 assert(_bit_map->isMarked(addr), "tautology"); 6467 if (_bit_map->isMarked(addr+1)) { 6468 // this is an allocated object that might not yet be initialized 6469 assert(_skip_bits == 0, "tautology"); 6470 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6471 oop p = oop(addr); 6472 if (p->klass_or_null() == NULL) { 6473 // in the case of Clean-on-Enter optimization, redirty card 6474 // and avoid clearing card by increasing the threshold. 6475 return true; 6476 } 6477 } 6478 scan_oops_in_oop(addr); 6479 return true; 6480 } 6481 6482 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6483 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6484 // Should we assert that our work queue is empty or 6485 // below some drain limit? 6486 assert(_work_queue->size() == 0, 6487 "should drain stack to limit stack usage"); 6488 // convert ptr to an oop preparatory to scanning 6489 oop obj = oop(ptr); 6490 // Ignore mark word in verification below, since we 6491 // may be running concurrent with mutators. 6492 assert(obj->is_oop(true), "should be an oop"); 6493 assert(_finger <= ptr, "_finger runneth ahead"); 6494 // advance the finger to right end of this object 6495 _finger = ptr + obj->size(); 6496 assert(_finger > ptr, "we just incremented it above"); 6497 // On large heaps, it may take us some time to get through 6498 // the marking phase. During 6499 // this time it's possible that a lot of mutations have 6500 // accumulated in the card table and the mod union table -- 6501 // these mutation records are redundant until we have 6502 // actually traced into the corresponding card. 6503 // Here, we check whether advancing the finger would make 6504 // us cross into a new card, and if so clear corresponding 6505 // cards in the MUT (preclean them in the card-table in the 6506 // future). 6507 6508 // The clean-on-enter optimization is disabled by default, 6509 // until we fix 6178663. 6510 if (CMSCleanOnEnter && (_finger > _threshold)) { 6511 // [_threshold, _finger) represents the interval 6512 // of cards to be cleared in MUT (or precleaned in card table). 6513 // The set of cards to be cleared is all those that overlap 6514 // with the interval [_threshold, _finger); note that 6515 // _threshold is always kept card-aligned but _finger isn't 6516 // always card-aligned. 6517 HeapWord* old_threshold = _threshold; 6518 assert(old_threshold == (HeapWord*)round_to( 6519 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6520 "_threshold should always be card-aligned"); 6521 _threshold = (HeapWord*)round_to( 6522 (intptr_t)_finger, CardTableModRefBS::card_size); 6523 MemRegion mr(old_threshold, _threshold); 6524 assert(!mr.is_empty(), "Control point invariant"); 6525 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6526 _mut->clear_range(mr); 6527 } 6528 6529 // Note: the local finger doesn't advance while we drain 6530 // the stack below, but the global finger sure can and will. 6531 HeapWord** gfa = _task->global_finger_addr(); 6532 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6533 _span, _bit_map, 6534 _work_queue, 6535 _overflow_stack, 6536 _finger, 6537 gfa, this); 6538 bool res = _work_queue->push(obj); // overflow could occur here 6539 assert(res, "Will hold once we use workqueues"); 6540 while (true) { 6541 oop new_oop; 6542 if (!_work_queue->pop_local(new_oop)) { 6543 // We emptied our work_queue; check if there's stuff that can 6544 // be gotten from the overflow stack. 6545 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6546 _overflow_stack, _work_queue)) { 6547 do_yield_check(); 6548 continue; 6549 } else { // done 6550 break; 6551 } 6552 } 6553 // Skip verifying header mark word below because we are 6554 // running concurrent with mutators. 6555 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6556 // now scan this oop's oops 6557 new_oop->oop_iterate(&pushOrMarkClosure); 6558 do_yield_check(); 6559 } 6560 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6561 } 6562 6563 // Yield in response to a request from VM Thread or 6564 // from mutators. 6565 void ParMarkFromRootsClosure::do_yield_work() { 6566 assert(_task != NULL, "sanity"); 6567 _task->yield(); 6568 } 6569 6570 // A variant of the above used for verifying CMS marking work. 6571 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6572 MemRegion span, 6573 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6574 CMSMarkStack* mark_stack): 6575 _collector(collector), 6576 _span(span), 6577 _verification_bm(verification_bm), 6578 _cms_bm(cms_bm), 6579 _mark_stack(mark_stack), 6580 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6581 mark_stack) 6582 { 6583 assert(_mark_stack->isEmpty(), "stack should be empty"); 6584 _finger = _verification_bm->startWord(); 6585 assert(_collector->_restart_addr == NULL, "Sanity check"); 6586 assert(_span.contains(_finger), "Out of bounds _finger?"); 6587 } 6588 6589 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6590 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6591 assert(_span.contains(addr), "Out of bounds _finger?"); 6592 _finger = addr; 6593 } 6594 6595 // Should revisit to see if this should be restructured for 6596 // greater efficiency. 6597 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6598 // convert offset into a HeapWord* 6599 HeapWord* addr = _verification_bm->startWord() + offset; 6600 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6601 "address out of range"); 6602 assert(_verification_bm->isMarked(addr), "tautology"); 6603 assert(_cms_bm->isMarked(addr), "tautology"); 6604 6605 assert(_mark_stack->isEmpty(), 6606 "should drain stack to limit stack usage"); 6607 // convert addr to an oop preparatory to scanning 6608 oop obj = oop(addr); 6609 assert(obj->is_oop(), "should be an oop"); 6610 assert(_finger <= addr, "_finger runneth ahead"); 6611 // advance the finger to right end of this object 6612 _finger = addr + obj->size(); 6613 assert(_finger > addr, "we just incremented it above"); 6614 // Note: the finger doesn't advance while we drain 6615 // the stack below. 6616 bool res = _mark_stack->push(obj); 6617 assert(res, "Empty non-zero size stack should have space for single push"); 6618 while (!_mark_stack->isEmpty()) { 6619 oop new_oop = _mark_stack->pop(); 6620 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6621 // now scan this oop's oops 6622 new_oop->oop_iterate(&_pam_verify_closure); 6623 } 6624 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6625 return true; 6626 } 6627 6628 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6629 CMSCollector* collector, MemRegion span, 6630 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6631 CMSMarkStack* mark_stack): 6632 MetadataAwareOopClosure(collector->ref_processor()), 6633 _collector(collector), 6634 _span(span), 6635 _verification_bm(verification_bm), 6636 _cms_bm(cms_bm), 6637 _mark_stack(mark_stack) 6638 { } 6639 6640 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6641 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6642 6643 // Upon stack overflow, we discard (part of) the stack, 6644 // remembering the least address amongst those discarded 6645 // in CMSCollector's _restart_address. 6646 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6647 // Remember the least grey address discarded 6648 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6649 _collector->lower_restart_addr(ra); 6650 _mark_stack->reset(); // discard stack contents 6651 _mark_stack->expand(); // expand the stack if possible 6652 } 6653 6654 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6655 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6656 HeapWord* addr = (HeapWord*)obj; 6657 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6658 // Oop lies in _span and isn't yet grey or black 6659 _verification_bm->mark(addr); // now grey 6660 if (!_cms_bm->isMarked(addr)) { 6661 LogHandle(gc, verify) log; 6662 ResourceMark rm; 6663 oop(addr)->print_on(log.error_stream()); 6664 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6665 fatal("... aborting"); 6666 } 6667 6668 if (!_mark_stack->push(obj)) { // stack overflow 6669 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6670 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6671 handle_stack_overflow(addr); 6672 } 6673 // anything including and to the right of _finger 6674 // will be scanned as we iterate over the remainder of the 6675 // bit map 6676 } 6677 } 6678 6679 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6680 MemRegion span, 6681 CMSBitMap* bitMap, CMSMarkStack* markStack, 6682 HeapWord* finger, MarkFromRootsClosure* parent) : 6683 MetadataAwareOopClosure(collector->ref_processor()), 6684 _collector(collector), 6685 _span(span), 6686 _bitMap(bitMap), 6687 _markStack(markStack), 6688 _finger(finger), 6689 _parent(parent) 6690 { } 6691 6692 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6693 MemRegion span, 6694 CMSBitMap* bit_map, 6695 OopTaskQueue* work_queue, 6696 CMSMarkStack* overflow_stack, 6697 HeapWord* finger, 6698 HeapWord** global_finger_addr, 6699 ParMarkFromRootsClosure* parent) : 6700 MetadataAwareOopClosure(collector->ref_processor()), 6701 _collector(collector), 6702 _whole_span(collector->_span), 6703 _span(span), 6704 _bit_map(bit_map), 6705 _work_queue(work_queue), 6706 _overflow_stack(overflow_stack), 6707 _finger(finger), 6708 _global_finger_addr(global_finger_addr), 6709 _parent(parent) 6710 { } 6711 6712 // Assumes thread-safe access by callers, who are 6713 // responsible for mutual exclusion. 6714 void CMSCollector::lower_restart_addr(HeapWord* low) { 6715 assert(_span.contains(low), "Out of bounds addr"); 6716 if (_restart_addr == NULL) { 6717 _restart_addr = low; 6718 } else { 6719 _restart_addr = MIN2(_restart_addr, low); 6720 } 6721 } 6722 6723 // Upon stack overflow, we discard (part of) the stack, 6724 // remembering the least address amongst those discarded 6725 // in CMSCollector's _restart_address. 6726 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6727 // Remember the least grey address discarded 6728 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6729 _collector->lower_restart_addr(ra); 6730 _markStack->reset(); // discard stack contents 6731 _markStack->expand(); // expand the stack if possible 6732 } 6733 6734 // Upon stack overflow, we discard (part of) the stack, 6735 // remembering the least address amongst those discarded 6736 // in CMSCollector's _restart_address. 6737 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6738 // We need to do this under a mutex to prevent other 6739 // workers from interfering with the work done below. 6740 MutexLockerEx ml(_overflow_stack->par_lock(), 6741 Mutex::_no_safepoint_check_flag); 6742 // Remember the least grey address discarded 6743 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6744 _collector->lower_restart_addr(ra); 6745 _overflow_stack->reset(); // discard stack contents 6746 _overflow_stack->expand(); // expand the stack if possible 6747 } 6748 6749 void PushOrMarkClosure::do_oop(oop obj) { 6750 // Ignore mark word because we are running concurrent with mutators. 6751 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6752 HeapWord* addr = (HeapWord*)obj; 6753 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6754 // Oop lies in _span and isn't yet grey or black 6755 _bitMap->mark(addr); // now grey 6756 if (addr < _finger) { 6757 // the bit map iteration has already either passed, or 6758 // sampled, this bit in the bit map; we'll need to 6759 // use the marking stack to scan this oop's oops. 6760 bool simulate_overflow = false; 6761 NOT_PRODUCT( 6762 if (CMSMarkStackOverflowALot && 6763 _collector->simulate_overflow()) { 6764 // simulate a stack overflow 6765 simulate_overflow = true; 6766 } 6767 ) 6768 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6769 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6770 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6771 handle_stack_overflow(addr); 6772 } 6773 } 6774 // anything including and to the right of _finger 6775 // will be scanned as we iterate over the remainder of the 6776 // bit map 6777 do_yield_check(); 6778 } 6779 } 6780 6781 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6782 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6783 6784 void ParPushOrMarkClosure::do_oop(oop obj) { 6785 // Ignore mark word because we are running concurrent with mutators. 6786 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6787 HeapWord* addr = (HeapWord*)obj; 6788 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6789 // Oop lies in _span and isn't yet grey or black 6790 // We read the global_finger (volatile read) strictly after marking oop 6791 bool res = _bit_map->par_mark(addr); // now grey 6792 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6793 // Should we push this marked oop on our stack? 6794 // -- if someone else marked it, nothing to do 6795 // -- if target oop is above global finger nothing to do 6796 // -- if target oop is in chunk and above local finger 6797 // then nothing to do 6798 // -- else push on work queue 6799 if ( !res // someone else marked it, they will deal with it 6800 || (addr >= *gfa) // will be scanned in a later task 6801 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6802 return; 6803 } 6804 // the bit map iteration has already either passed, or 6805 // sampled, this bit in the bit map; we'll need to 6806 // use the marking stack to scan this oop's oops. 6807 bool simulate_overflow = false; 6808 NOT_PRODUCT( 6809 if (CMSMarkStackOverflowALot && 6810 _collector->simulate_overflow()) { 6811 // simulate a stack overflow 6812 simulate_overflow = true; 6813 } 6814 ) 6815 if (simulate_overflow || 6816 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6817 // stack overflow 6818 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6819 // We cannot assert that the overflow stack is full because 6820 // it may have been emptied since. 6821 assert(simulate_overflow || 6822 _work_queue->size() == _work_queue->max_elems(), 6823 "Else push should have succeeded"); 6824 handle_stack_overflow(addr); 6825 } 6826 do_yield_check(); 6827 } 6828 } 6829 6830 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6831 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6832 6833 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6834 MemRegion span, 6835 ReferenceProcessor* rp, 6836 CMSBitMap* bit_map, 6837 CMSBitMap* mod_union_table, 6838 CMSMarkStack* mark_stack, 6839 bool concurrent_precleaning): 6840 MetadataAwareOopClosure(rp), 6841 _collector(collector), 6842 _span(span), 6843 _bit_map(bit_map), 6844 _mod_union_table(mod_union_table), 6845 _mark_stack(mark_stack), 6846 _concurrent_precleaning(concurrent_precleaning) 6847 { 6848 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6849 } 6850 6851 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6852 // the non-parallel version (the parallel version appears further below.) 6853 void PushAndMarkClosure::do_oop(oop obj) { 6854 // Ignore mark word verification. If during concurrent precleaning, 6855 // the object monitor may be locked. If during the checkpoint 6856 // phases, the object may already have been reached by a different 6857 // path and may be at the end of the global overflow list (so 6858 // the mark word may be NULL). 6859 assert(obj->is_oop_or_null(true /* ignore mark word */), 6860 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6861 HeapWord* addr = (HeapWord*)obj; 6862 // Check if oop points into the CMS generation 6863 // and is not marked 6864 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6865 // a white object ... 6866 _bit_map->mark(addr); // ... now grey 6867 // push on the marking stack (grey set) 6868 bool simulate_overflow = false; 6869 NOT_PRODUCT( 6870 if (CMSMarkStackOverflowALot && 6871 _collector->simulate_overflow()) { 6872 // simulate a stack overflow 6873 simulate_overflow = true; 6874 } 6875 ) 6876 if (simulate_overflow || !_mark_stack->push(obj)) { 6877 if (_concurrent_precleaning) { 6878 // During precleaning we can just dirty the appropriate card(s) 6879 // in the mod union table, thus ensuring that the object remains 6880 // in the grey set and continue. In the case of object arrays 6881 // we need to dirty all of the cards that the object spans, 6882 // since the rescan of object arrays will be limited to the 6883 // dirty cards. 6884 // Note that no one can be interfering with us in this action 6885 // of dirtying the mod union table, so no locking or atomics 6886 // are required. 6887 if (obj->is_objArray()) { 6888 size_t sz = obj->size(); 6889 HeapWord* end_card_addr = (HeapWord*)round_to( 6890 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6891 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6892 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6893 _mod_union_table->mark_range(redirty_range); 6894 } else { 6895 _mod_union_table->mark(addr); 6896 } 6897 _collector->_ser_pmc_preclean_ovflw++; 6898 } else { 6899 // During the remark phase, we need to remember this oop 6900 // in the overflow list. 6901 _collector->push_on_overflow_list(obj); 6902 _collector->_ser_pmc_remark_ovflw++; 6903 } 6904 } 6905 } 6906 } 6907 6908 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6909 MemRegion span, 6910 ReferenceProcessor* rp, 6911 CMSBitMap* bit_map, 6912 OopTaskQueue* work_queue): 6913 MetadataAwareOopClosure(rp), 6914 _collector(collector), 6915 _span(span), 6916 _bit_map(bit_map), 6917 _work_queue(work_queue) 6918 { 6919 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6920 } 6921 6922 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6923 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6924 6925 // Grey object rescan during second checkpoint phase -- 6926 // the parallel version. 6927 void ParPushAndMarkClosure::do_oop(oop obj) { 6928 // In the assert below, we ignore the mark word because 6929 // this oop may point to an already visited object that is 6930 // on the overflow stack (in which case the mark word has 6931 // been hijacked for chaining into the overflow stack -- 6932 // if this is the last object in the overflow stack then 6933 // its mark word will be NULL). Because this object may 6934 // have been subsequently popped off the global overflow 6935 // stack, and the mark word possibly restored to the prototypical 6936 // value, by the time we get to examined this failing assert in 6937 // the debugger, is_oop_or_null(false) may subsequently start 6938 // to hold. 6939 assert(obj->is_oop_or_null(true), 6940 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6941 HeapWord* addr = (HeapWord*)obj; 6942 // Check if oop points into the CMS generation 6943 // and is not marked 6944 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6945 // a white object ... 6946 // If we manage to "claim" the object, by being the 6947 // first thread to mark it, then we push it on our 6948 // marking stack 6949 if (_bit_map->par_mark(addr)) { // ... now grey 6950 // push on work queue (grey set) 6951 bool simulate_overflow = false; 6952 NOT_PRODUCT( 6953 if (CMSMarkStackOverflowALot && 6954 _collector->par_simulate_overflow()) { 6955 // simulate a stack overflow 6956 simulate_overflow = true; 6957 } 6958 ) 6959 if (simulate_overflow || !_work_queue->push(obj)) { 6960 _collector->par_push_on_overflow_list(obj); 6961 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6962 } 6963 } // Else, some other thread got there first 6964 } 6965 } 6966 6967 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6968 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6969 6970 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6971 Mutex* bml = _collector->bitMapLock(); 6972 assert_lock_strong(bml); 6973 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6974 "CMS thread should hold CMS token"); 6975 6976 bml->unlock(); 6977 ConcurrentMarkSweepThread::desynchronize(true); 6978 6979 _collector->stopTimer(); 6980 _collector->incrementYields(); 6981 6982 // See the comment in coordinator_yield() 6983 for (unsigned i = 0; i < CMSYieldSleepCount && 6984 ConcurrentMarkSweepThread::should_yield() && 6985 !CMSCollector::foregroundGCIsActive(); ++i) { 6986 os::sleep(Thread::current(), 1, false); 6987 } 6988 6989 ConcurrentMarkSweepThread::synchronize(true); 6990 bml->lock(); 6991 6992 _collector->startTimer(); 6993 } 6994 6995 bool CMSPrecleanRefsYieldClosure::should_return() { 6996 if (ConcurrentMarkSweepThread::should_yield()) { 6997 do_yield_work(); 6998 } 6999 return _collector->foregroundGCIsActive(); 7000 } 7001 7002 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7003 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7004 "mr should be aligned to start at a card boundary"); 7005 // We'd like to assert: 7006 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7007 // "mr should be a range of cards"); 7008 // However, that would be too strong in one case -- the last 7009 // partition ends at _unallocated_block which, in general, can be 7010 // an arbitrary boundary, not necessarily card aligned. 7011 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7012 _space->object_iterate_mem(mr, &_scan_cl); 7013 } 7014 7015 SweepClosure::SweepClosure(CMSCollector* collector, 7016 ConcurrentMarkSweepGeneration* g, 7017 CMSBitMap* bitMap, bool should_yield) : 7018 _collector(collector), 7019 _g(g), 7020 _sp(g->cmsSpace()), 7021 _limit(_sp->sweep_limit()), 7022 _freelistLock(_sp->freelistLock()), 7023 _bitMap(bitMap), 7024 _yield(should_yield), 7025 _inFreeRange(false), // No free range at beginning of sweep 7026 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7027 _lastFreeRangeCoalesced(false), 7028 _freeFinger(g->used_region().start()) 7029 { 7030 NOT_PRODUCT( 7031 _numObjectsFreed = 0; 7032 _numWordsFreed = 0; 7033 _numObjectsLive = 0; 7034 _numWordsLive = 0; 7035 _numObjectsAlreadyFree = 0; 7036 _numWordsAlreadyFree = 0; 7037 _last_fc = NULL; 7038 7039 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7040 _sp->dictionary()->initialize_dict_returned_bytes(); 7041 ) 7042 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7043 "sweep _limit out of bounds"); 7044 log_develop_trace(gc, sweep)("===================="); 7045 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7046 } 7047 7048 void SweepClosure::print_on(outputStream* st) const { 7049 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7050 p2i(_sp->bottom()), p2i(_sp->end())); 7051 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7052 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7053 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7054 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7055 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7056 } 7057 7058 #ifndef PRODUCT 7059 // Assertion checking only: no useful work in product mode -- 7060 // however, if any of the flags below become product flags, 7061 // you may need to review this code to see if it needs to be 7062 // enabled in product mode. 7063 SweepClosure::~SweepClosure() { 7064 assert_lock_strong(_freelistLock); 7065 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7066 "sweep _limit out of bounds"); 7067 if (inFreeRange()) { 7068 LogHandle(gc, sweep) log; 7069 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7070 ResourceMark rm; 7071 print_on(log.error_stream()); 7072 ShouldNotReachHere(); 7073 } 7074 7075 if (log_is_enabled(Debug, gc, sweep)) { 7076 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7077 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7078 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7079 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7080 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7081 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7082 } 7083 7084 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7085 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7086 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7087 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7088 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7089 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7090 } 7091 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7092 log_develop_trace(gc, sweep)("================"); 7093 } 7094 #endif // PRODUCT 7095 7096 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7097 bool freeRangeInFreeLists) { 7098 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7099 p2i(freeFinger), freeRangeInFreeLists); 7100 assert(!inFreeRange(), "Trampling existing free range"); 7101 set_inFreeRange(true); 7102 set_lastFreeRangeCoalesced(false); 7103 7104 set_freeFinger(freeFinger); 7105 set_freeRangeInFreeLists(freeRangeInFreeLists); 7106 if (CMSTestInFreeList) { 7107 if (freeRangeInFreeLists) { 7108 FreeChunk* fc = (FreeChunk*) freeFinger; 7109 assert(fc->is_free(), "A chunk on the free list should be free."); 7110 assert(fc->size() > 0, "Free range should have a size"); 7111 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7112 } 7113 } 7114 } 7115 7116 // Note that the sweeper runs concurrently with mutators. Thus, 7117 // it is possible for direct allocation in this generation to happen 7118 // in the middle of the sweep. Note that the sweeper also coalesces 7119 // contiguous free blocks. Thus, unless the sweeper and the allocator 7120 // synchronize appropriately freshly allocated blocks may get swept up. 7121 // This is accomplished by the sweeper locking the free lists while 7122 // it is sweeping. Thus blocks that are determined to be free are 7123 // indeed free. There is however one additional complication: 7124 // blocks that have been allocated since the final checkpoint and 7125 // mark, will not have been marked and so would be treated as 7126 // unreachable and swept up. To prevent this, the allocator marks 7127 // the bit map when allocating during the sweep phase. This leads, 7128 // however, to a further complication -- objects may have been allocated 7129 // but not yet initialized -- in the sense that the header isn't yet 7130 // installed. The sweeper can not then determine the size of the block 7131 // in order to skip over it. To deal with this case, we use a technique 7132 // (due to Printezis) to encode such uninitialized block sizes in the 7133 // bit map. Since the bit map uses a bit per every HeapWord, but the 7134 // CMS generation has a minimum object size of 3 HeapWords, it follows 7135 // that "normal marks" won't be adjacent in the bit map (there will 7136 // always be at least two 0 bits between successive 1 bits). We make use 7137 // of these "unused" bits to represent uninitialized blocks -- the bit 7138 // corresponding to the start of the uninitialized object and the next 7139 // bit are both set. Finally, a 1 bit marks the end of the object that 7140 // started with the two consecutive 1 bits to indicate its potentially 7141 // uninitialized state. 7142 7143 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7144 FreeChunk* fc = (FreeChunk*)addr; 7145 size_t res; 7146 7147 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7148 // than "addr == _limit" because although _limit was a block boundary when 7149 // we started the sweep, it may no longer be one because heap expansion 7150 // may have caused us to coalesce the block ending at the address _limit 7151 // with a newly expanded chunk (this happens when _limit was set to the 7152 // previous _end of the space), so we may have stepped past _limit: 7153 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7154 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7155 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7156 "sweep _limit out of bounds"); 7157 assert(addr < _sp->end(), "addr out of bounds"); 7158 // Flush any free range we might be holding as a single 7159 // coalesced chunk to the appropriate free list. 7160 if (inFreeRange()) { 7161 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7162 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7163 flush_cur_free_chunk(freeFinger(), 7164 pointer_delta(addr, freeFinger())); 7165 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7166 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7167 lastFreeRangeCoalesced() ? 1 : 0); 7168 } 7169 7170 // help the iterator loop finish 7171 return pointer_delta(_sp->end(), addr); 7172 } 7173 7174 assert(addr < _limit, "sweep invariant"); 7175 // check if we should yield 7176 do_yield_check(addr); 7177 if (fc->is_free()) { 7178 // Chunk that is already free 7179 res = fc->size(); 7180 do_already_free_chunk(fc); 7181 debug_only(_sp->verifyFreeLists()); 7182 // If we flush the chunk at hand in lookahead_and_flush() 7183 // and it's coalesced with a preceding chunk, then the 7184 // process of "mangling" the payload of the coalesced block 7185 // will cause erasure of the size information from the 7186 // (erstwhile) header of all the coalesced blocks but the 7187 // first, so the first disjunct in the assert will not hold 7188 // in that specific case (in which case the second disjunct 7189 // will hold). 7190 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7191 "Otherwise the size info doesn't change at this step"); 7192 NOT_PRODUCT( 7193 _numObjectsAlreadyFree++; 7194 _numWordsAlreadyFree += res; 7195 ) 7196 NOT_PRODUCT(_last_fc = fc;) 7197 } else if (!_bitMap->isMarked(addr)) { 7198 // Chunk is fresh garbage 7199 res = do_garbage_chunk(fc); 7200 debug_only(_sp->verifyFreeLists()); 7201 NOT_PRODUCT( 7202 _numObjectsFreed++; 7203 _numWordsFreed += res; 7204 ) 7205 } else { 7206 // Chunk that is alive. 7207 res = do_live_chunk(fc); 7208 debug_only(_sp->verifyFreeLists()); 7209 NOT_PRODUCT( 7210 _numObjectsLive++; 7211 _numWordsLive += res; 7212 ) 7213 } 7214 return res; 7215 } 7216 7217 // For the smart allocation, record following 7218 // split deaths - a free chunk is removed from its free list because 7219 // it is being split into two or more chunks. 7220 // split birth - a free chunk is being added to its free list because 7221 // a larger free chunk has been split and resulted in this free chunk. 7222 // coal death - a free chunk is being removed from its free list because 7223 // it is being coalesced into a large free chunk. 7224 // coal birth - a free chunk is being added to its free list because 7225 // it was created when two or more free chunks where coalesced into 7226 // this free chunk. 7227 // 7228 // These statistics are used to determine the desired number of free 7229 // chunks of a given size. The desired number is chosen to be relative 7230 // to the end of a CMS sweep. The desired number at the end of a sweep 7231 // is the 7232 // count-at-end-of-previous-sweep (an amount that was enough) 7233 // - count-at-beginning-of-current-sweep (the excess) 7234 // + split-births (gains in this size during interval) 7235 // - split-deaths (demands on this size during interval) 7236 // where the interval is from the end of one sweep to the end of the 7237 // next. 7238 // 7239 // When sweeping the sweeper maintains an accumulated chunk which is 7240 // the chunk that is made up of chunks that have been coalesced. That 7241 // will be termed the left-hand chunk. A new chunk of garbage that 7242 // is being considered for coalescing will be referred to as the 7243 // right-hand chunk. 7244 // 7245 // When making a decision on whether to coalesce a right-hand chunk with 7246 // the current left-hand chunk, the current count vs. the desired count 7247 // of the left-hand chunk is considered. Also if the right-hand chunk 7248 // is near the large chunk at the end of the heap (see 7249 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7250 // left-hand chunk is coalesced. 7251 // 7252 // When making a decision about whether to split a chunk, the desired count 7253 // vs. the current count of the candidate to be split is also considered. 7254 // If the candidate is underpopulated (currently fewer chunks than desired) 7255 // a chunk of an overpopulated (currently more chunks than desired) size may 7256 // be chosen. The "hint" associated with a free list, if non-null, points 7257 // to a free list which may be overpopulated. 7258 // 7259 7260 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7261 const size_t size = fc->size(); 7262 // Chunks that cannot be coalesced are not in the 7263 // free lists. 7264 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7265 assert(_sp->verify_chunk_in_free_list(fc), 7266 "free chunk should be in free lists"); 7267 } 7268 // a chunk that is already free, should not have been 7269 // marked in the bit map 7270 HeapWord* const addr = (HeapWord*) fc; 7271 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7272 // Verify that the bit map has no bits marked between 7273 // addr and purported end of this block. 7274 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7275 7276 // Some chunks cannot be coalesced under any circumstances. 7277 // See the definition of cantCoalesce(). 7278 if (!fc->cantCoalesce()) { 7279 // This chunk can potentially be coalesced. 7280 // All the work is done in 7281 do_post_free_or_garbage_chunk(fc, size); 7282 // Note that if the chunk is not coalescable (the else arm 7283 // below), we unconditionally flush, without needing to do 7284 // a "lookahead," as we do below. 7285 if (inFreeRange()) lookahead_and_flush(fc, size); 7286 } else { 7287 // Code path common to both original and adaptive free lists. 7288 7289 // cant coalesce with previous block; this should be treated 7290 // as the end of a free run if any 7291 if (inFreeRange()) { 7292 // we kicked some butt; time to pick up the garbage 7293 assert(freeFinger() < addr, "freeFinger points too high"); 7294 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7295 } 7296 // else, nothing to do, just continue 7297 } 7298 } 7299 7300 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7301 // This is a chunk of garbage. It is not in any free list. 7302 // Add it to a free list or let it possibly be coalesced into 7303 // a larger chunk. 7304 HeapWord* const addr = (HeapWord*) fc; 7305 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7306 7307 // Verify that the bit map has no bits marked between 7308 // addr and purported end of just dead object. 7309 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7310 do_post_free_or_garbage_chunk(fc, size); 7311 7312 assert(_limit >= addr + size, 7313 "A freshly garbage chunk can't possibly straddle over _limit"); 7314 if (inFreeRange()) lookahead_and_flush(fc, size); 7315 return size; 7316 } 7317 7318 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7319 HeapWord* addr = (HeapWord*) fc; 7320 // The sweeper has just found a live object. Return any accumulated 7321 // left hand chunk to the free lists. 7322 if (inFreeRange()) { 7323 assert(freeFinger() < addr, "freeFinger points too high"); 7324 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7325 } 7326 7327 // This object is live: we'd normally expect this to be 7328 // an oop, and like to assert the following: 7329 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7330 // However, as we commented above, this may be an object whose 7331 // header hasn't yet been initialized. 7332 size_t size; 7333 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7334 if (_bitMap->isMarked(addr + 1)) { 7335 // Determine the size from the bit map, rather than trying to 7336 // compute it from the object header. 7337 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7338 size = pointer_delta(nextOneAddr + 1, addr); 7339 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7340 "alignment problem"); 7341 7342 #ifdef ASSERT 7343 if (oop(addr)->klass_or_null() != NULL) { 7344 // Ignore mark word because we are running concurrent with mutators 7345 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7346 assert(size == 7347 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7348 "P-mark and computed size do not agree"); 7349 } 7350 #endif 7351 7352 } else { 7353 // This should be an initialized object that's alive. 7354 assert(oop(addr)->klass_or_null() != NULL, 7355 "Should be an initialized object"); 7356 // Ignore mark word because we are running concurrent with mutators 7357 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7358 // Verify that the bit map has no bits marked between 7359 // addr and purported end of this block. 7360 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7361 assert(size >= 3, "Necessary for Printezis marks to work"); 7362 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7363 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7364 } 7365 return size; 7366 } 7367 7368 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7369 size_t chunkSize) { 7370 // do_post_free_or_garbage_chunk() should only be called in the case 7371 // of the adaptive free list allocator. 7372 const bool fcInFreeLists = fc->is_free(); 7373 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7374 if (CMSTestInFreeList && fcInFreeLists) { 7375 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7376 } 7377 7378 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7379 7380 HeapWord* const fc_addr = (HeapWord*) fc; 7381 7382 bool coalesce = false; 7383 const size_t left = pointer_delta(fc_addr, freeFinger()); 7384 const size_t right = chunkSize; 7385 switch (FLSCoalescePolicy) { 7386 // numeric value forms a coalition aggressiveness metric 7387 case 0: { // never coalesce 7388 coalesce = false; 7389 break; 7390 } 7391 case 1: { // coalesce if left & right chunks on overpopulated lists 7392 coalesce = _sp->coalOverPopulated(left) && 7393 _sp->coalOverPopulated(right); 7394 break; 7395 } 7396 case 2: { // coalesce if left chunk on overpopulated list (default) 7397 coalesce = _sp->coalOverPopulated(left); 7398 break; 7399 } 7400 case 3: { // coalesce if left OR right chunk on overpopulated list 7401 coalesce = _sp->coalOverPopulated(left) || 7402 _sp->coalOverPopulated(right); 7403 break; 7404 } 7405 case 4: { // always coalesce 7406 coalesce = true; 7407 break; 7408 } 7409 default: 7410 ShouldNotReachHere(); 7411 } 7412 7413 // Should the current free range be coalesced? 7414 // If the chunk is in a free range and either we decided to coalesce above 7415 // or the chunk is near the large block at the end of the heap 7416 // (isNearLargestChunk() returns true), then coalesce this chunk. 7417 const bool doCoalesce = inFreeRange() 7418 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7419 if (doCoalesce) { 7420 // Coalesce the current free range on the left with the new 7421 // chunk on the right. If either is on a free list, 7422 // it must be removed from the list and stashed in the closure. 7423 if (freeRangeInFreeLists()) { 7424 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7425 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7426 "Size of free range is inconsistent with chunk size."); 7427 if (CMSTestInFreeList) { 7428 assert(_sp->verify_chunk_in_free_list(ffc), 7429 "Chunk is not in free lists"); 7430 } 7431 _sp->coalDeath(ffc->size()); 7432 _sp->removeFreeChunkFromFreeLists(ffc); 7433 set_freeRangeInFreeLists(false); 7434 } 7435 if (fcInFreeLists) { 7436 _sp->coalDeath(chunkSize); 7437 assert(fc->size() == chunkSize, 7438 "The chunk has the wrong size or is not in the free lists"); 7439 _sp->removeFreeChunkFromFreeLists(fc); 7440 } 7441 set_lastFreeRangeCoalesced(true); 7442 print_free_block_coalesced(fc); 7443 } else { // not in a free range and/or should not coalesce 7444 // Return the current free range and start a new one. 7445 if (inFreeRange()) { 7446 // In a free range but cannot coalesce with the right hand chunk. 7447 // Put the current free range into the free lists. 7448 flush_cur_free_chunk(freeFinger(), 7449 pointer_delta(fc_addr, freeFinger())); 7450 } 7451 // Set up for new free range. Pass along whether the right hand 7452 // chunk is in the free lists. 7453 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7454 } 7455 } 7456 7457 // Lookahead flush: 7458 // If we are tracking a free range, and this is the last chunk that 7459 // we'll look at because its end crosses past _limit, we'll preemptively 7460 // flush it along with any free range we may be holding on to. Note that 7461 // this can be the case only for an already free or freshly garbage 7462 // chunk. If this block is an object, it can never straddle 7463 // over _limit. The "straddling" occurs when _limit is set at 7464 // the previous end of the space when this cycle started, and 7465 // a subsequent heap expansion caused the previously co-terminal 7466 // free block to be coalesced with the newly expanded portion, 7467 // thus rendering _limit a non-block-boundary making it dangerous 7468 // for the sweeper to step over and examine. 7469 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7470 assert(inFreeRange(), "Should only be called if currently in a free range."); 7471 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7472 assert(_sp->used_region().contains(eob - 1), 7473 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7474 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7475 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7476 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7477 if (eob >= _limit) { 7478 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7479 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7480 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7481 "[" PTR_FORMAT "," PTR_FORMAT ")", 7482 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7483 // Return the storage we are tracking back into the free lists. 7484 log_develop_trace(gc, sweep)("Flushing ... "); 7485 assert(freeFinger() < eob, "Error"); 7486 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7487 } 7488 } 7489 7490 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7491 assert(inFreeRange(), "Should only be called if currently in a free range."); 7492 assert(size > 0, 7493 "A zero sized chunk cannot be added to the free lists."); 7494 if (!freeRangeInFreeLists()) { 7495 if (CMSTestInFreeList) { 7496 FreeChunk* fc = (FreeChunk*) chunk; 7497 fc->set_size(size); 7498 assert(!_sp->verify_chunk_in_free_list(fc), 7499 "chunk should not be in free lists yet"); 7500 } 7501 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7502 // A new free range is going to be starting. The current 7503 // free range has not been added to the free lists yet or 7504 // was removed so add it back. 7505 // If the current free range was coalesced, then the death 7506 // of the free range was recorded. Record a birth now. 7507 if (lastFreeRangeCoalesced()) { 7508 _sp->coalBirth(size); 7509 } 7510 _sp->addChunkAndRepairOffsetTable(chunk, size, 7511 lastFreeRangeCoalesced()); 7512 } else { 7513 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7514 } 7515 set_inFreeRange(false); 7516 set_freeRangeInFreeLists(false); 7517 } 7518 7519 // We take a break if we've been at this for a while, 7520 // so as to avoid monopolizing the locks involved. 7521 void SweepClosure::do_yield_work(HeapWord* addr) { 7522 // Return current free chunk being used for coalescing (if any) 7523 // to the appropriate freelist. After yielding, the next 7524 // free block encountered will start a coalescing range of 7525 // free blocks. If the next free block is adjacent to the 7526 // chunk just flushed, they will need to wait for the next 7527 // sweep to be coalesced. 7528 if (inFreeRange()) { 7529 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7530 } 7531 7532 // First give up the locks, then yield, then re-lock. 7533 // We should probably use a constructor/destructor idiom to 7534 // do this unlock/lock or modify the MutexUnlocker class to 7535 // serve our purpose. XXX 7536 assert_lock_strong(_bitMap->lock()); 7537 assert_lock_strong(_freelistLock); 7538 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7539 "CMS thread should hold CMS token"); 7540 _bitMap->lock()->unlock(); 7541 _freelistLock->unlock(); 7542 ConcurrentMarkSweepThread::desynchronize(true); 7543 _collector->stopTimer(); 7544 _collector->incrementYields(); 7545 7546 // See the comment in coordinator_yield() 7547 for (unsigned i = 0; i < CMSYieldSleepCount && 7548 ConcurrentMarkSweepThread::should_yield() && 7549 !CMSCollector::foregroundGCIsActive(); ++i) { 7550 os::sleep(Thread::current(), 1, false); 7551 } 7552 7553 ConcurrentMarkSweepThread::synchronize(true); 7554 _freelistLock->lock(); 7555 _bitMap->lock()->lock_without_safepoint_check(); 7556 _collector->startTimer(); 7557 } 7558 7559 #ifndef PRODUCT 7560 // This is actually very useful in a product build if it can 7561 // be called from the debugger. Compile it into the product 7562 // as needed. 7563 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7564 return debug_cms_space->verify_chunk_in_free_list(fc); 7565 } 7566 #endif 7567 7568 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7569 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7570 p2i(fc), fc->size()); 7571 } 7572 7573 // CMSIsAliveClosure 7574 bool CMSIsAliveClosure::do_object_b(oop obj) { 7575 HeapWord* addr = (HeapWord*)obj; 7576 return addr != NULL && 7577 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7578 } 7579 7580 7581 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7582 MemRegion span, 7583 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7584 bool cpc): 7585 _collector(collector), 7586 _span(span), 7587 _bit_map(bit_map), 7588 _mark_stack(mark_stack), 7589 _concurrent_precleaning(cpc) { 7590 assert(!_span.is_empty(), "Empty span could spell trouble"); 7591 } 7592 7593 7594 // CMSKeepAliveClosure: the serial version 7595 void CMSKeepAliveClosure::do_oop(oop obj) { 7596 HeapWord* addr = (HeapWord*)obj; 7597 if (_span.contains(addr) && 7598 !_bit_map->isMarked(addr)) { 7599 _bit_map->mark(addr); 7600 bool simulate_overflow = false; 7601 NOT_PRODUCT( 7602 if (CMSMarkStackOverflowALot && 7603 _collector->simulate_overflow()) { 7604 // simulate a stack overflow 7605 simulate_overflow = true; 7606 } 7607 ) 7608 if (simulate_overflow || !_mark_stack->push(obj)) { 7609 if (_concurrent_precleaning) { 7610 // We dirty the overflown object and let the remark 7611 // phase deal with it. 7612 assert(_collector->overflow_list_is_empty(), "Error"); 7613 // In the case of object arrays, we need to dirty all of 7614 // the cards that the object spans. No locking or atomics 7615 // are needed since no one else can be mutating the mod union 7616 // table. 7617 if (obj->is_objArray()) { 7618 size_t sz = obj->size(); 7619 HeapWord* end_card_addr = 7620 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7621 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7622 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7623 _collector->_modUnionTable.mark_range(redirty_range); 7624 } else { 7625 _collector->_modUnionTable.mark(addr); 7626 } 7627 _collector->_ser_kac_preclean_ovflw++; 7628 } else { 7629 _collector->push_on_overflow_list(obj); 7630 _collector->_ser_kac_ovflw++; 7631 } 7632 } 7633 } 7634 } 7635 7636 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7637 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7638 7639 // CMSParKeepAliveClosure: a parallel version of the above. 7640 // The work queues are private to each closure (thread), 7641 // but (may be) available for stealing by other threads. 7642 void CMSParKeepAliveClosure::do_oop(oop obj) { 7643 HeapWord* addr = (HeapWord*)obj; 7644 if (_span.contains(addr) && 7645 !_bit_map->isMarked(addr)) { 7646 // In general, during recursive tracing, several threads 7647 // may be concurrently getting here; the first one to 7648 // "tag" it, claims it. 7649 if (_bit_map->par_mark(addr)) { 7650 bool res = _work_queue->push(obj); 7651 assert(res, "Low water mark should be much less than capacity"); 7652 // Do a recursive trim in the hope that this will keep 7653 // stack usage lower, but leave some oops for potential stealers 7654 trim_queue(_low_water_mark); 7655 } // Else, another thread got there first 7656 } 7657 } 7658 7659 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7660 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7661 7662 void CMSParKeepAliveClosure::trim_queue(uint max) { 7663 while (_work_queue->size() > max) { 7664 oop new_oop; 7665 if (_work_queue->pop_local(new_oop)) { 7666 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7667 assert(_bit_map->isMarked((HeapWord*)new_oop), 7668 "no white objects on this stack!"); 7669 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7670 // iterate over the oops in this oop, marking and pushing 7671 // the ones in CMS heap (i.e. in _span). 7672 new_oop->oop_iterate(&_mark_and_push); 7673 } 7674 } 7675 } 7676 7677 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7678 CMSCollector* collector, 7679 MemRegion span, CMSBitMap* bit_map, 7680 OopTaskQueue* work_queue): 7681 _collector(collector), 7682 _span(span), 7683 _bit_map(bit_map), 7684 _work_queue(work_queue) { } 7685 7686 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7687 HeapWord* addr = (HeapWord*)obj; 7688 if (_span.contains(addr) && 7689 !_bit_map->isMarked(addr)) { 7690 if (_bit_map->par_mark(addr)) { 7691 bool simulate_overflow = false; 7692 NOT_PRODUCT( 7693 if (CMSMarkStackOverflowALot && 7694 _collector->par_simulate_overflow()) { 7695 // simulate a stack overflow 7696 simulate_overflow = true; 7697 } 7698 ) 7699 if (simulate_overflow || !_work_queue->push(obj)) { 7700 _collector->par_push_on_overflow_list(obj); 7701 _collector->_par_kac_ovflw++; 7702 } 7703 } // Else another thread got there already 7704 } 7705 } 7706 7707 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7708 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7709 7710 ////////////////////////////////////////////////////////////////// 7711 // CMSExpansionCause ///////////////////////////// 7712 ////////////////////////////////////////////////////////////////// 7713 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7714 switch (cause) { 7715 case _no_expansion: 7716 return "No expansion"; 7717 case _satisfy_free_ratio: 7718 return "Free ratio"; 7719 case _satisfy_promotion: 7720 return "Satisfy promotion"; 7721 case _satisfy_allocation: 7722 return "allocation"; 7723 case _allocate_par_lab: 7724 return "Par LAB"; 7725 case _allocate_par_spooling_space: 7726 return "Par Spooling Space"; 7727 case _adaptive_size_policy: 7728 return "Ergonomics"; 7729 default: 7730 return "unknown"; 7731 } 7732 } 7733 7734 void CMSDrainMarkingStackClosure::do_void() { 7735 // the max number to take from overflow list at a time 7736 const size_t num = _mark_stack->capacity()/4; 7737 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7738 "Overflow list should be NULL during concurrent phases"); 7739 while (!_mark_stack->isEmpty() || 7740 // if stack is empty, check the overflow list 7741 _collector->take_from_overflow_list(num, _mark_stack)) { 7742 oop obj = _mark_stack->pop(); 7743 HeapWord* addr = (HeapWord*)obj; 7744 assert(_span.contains(addr), "Should be within span"); 7745 assert(_bit_map->isMarked(addr), "Should be marked"); 7746 assert(obj->is_oop(), "Should be an oop"); 7747 obj->oop_iterate(_keep_alive); 7748 } 7749 } 7750 7751 void CMSParDrainMarkingStackClosure::do_void() { 7752 // drain queue 7753 trim_queue(0); 7754 } 7755 7756 // Trim our work_queue so its length is below max at return 7757 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7758 while (_work_queue->size() > max) { 7759 oop new_oop; 7760 if (_work_queue->pop_local(new_oop)) { 7761 assert(new_oop->is_oop(), "Expected an oop"); 7762 assert(_bit_map->isMarked((HeapWord*)new_oop), 7763 "no white objects on this stack!"); 7764 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7765 // iterate over the oops in this oop, marking and pushing 7766 // the ones in CMS heap (i.e. in _span). 7767 new_oop->oop_iterate(&_mark_and_push); 7768 } 7769 } 7770 } 7771 7772 //////////////////////////////////////////////////////////////////// 7773 // Support for Marking Stack Overflow list handling and related code 7774 //////////////////////////////////////////////////////////////////// 7775 // Much of the following code is similar in shape and spirit to the 7776 // code used in ParNewGC. We should try and share that code 7777 // as much as possible in the future. 7778 7779 #ifndef PRODUCT 7780 // Debugging support for CMSStackOverflowALot 7781 7782 // It's OK to call this multi-threaded; the worst thing 7783 // that can happen is that we'll get a bunch of closely 7784 // spaced simulated overflows, but that's OK, in fact 7785 // probably good as it would exercise the overflow code 7786 // under contention. 7787 bool CMSCollector::simulate_overflow() { 7788 if (_overflow_counter-- <= 0) { // just being defensive 7789 _overflow_counter = CMSMarkStackOverflowInterval; 7790 return true; 7791 } else { 7792 return false; 7793 } 7794 } 7795 7796 bool CMSCollector::par_simulate_overflow() { 7797 return simulate_overflow(); 7798 } 7799 #endif 7800 7801 // Single-threaded 7802 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7803 assert(stack->isEmpty(), "Expected precondition"); 7804 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7805 size_t i = num; 7806 oop cur = _overflow_list; 7807 const markOop proto = markOopDesc::prototype(); 7808 NOT_PRODUCT(ssize_t n = 0;) 7809 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7810 next = oop(cur->mark()); 7811 cur->set_mark(proto); // until proven otherwise 7812 assert(cur->is_oop(), "Should be an oop"); 7813 bool res = stack->push(cur); 7814 assert(res, "Bit off more than can chew?"); 7815 NOT_PRODUCT(n++;) 7816 } 7817 _overflow_list = cur; 7818 #ifndef PRODUCT 7819 assert(_num_par_pushes >= n, "Too many pops?"); 7820 _num_par_pushes -=n; 7821 #endif 7822 return !stack->isEmpty(); 7823 } 7824 7825 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7826 // (MT-safe) Get a prefix of at most "num" from the list. 7827 // The overflow list is chained through the mark word of 7828 // each object in the list. We fetch the entire list, 7829 // break off a prefix of the right size and return the 7830 // remainder. If other threads try to take objects from 7831 // the overflow list at that time, they will wait for 7832 // some time to see if data becomes available. If (and 7833 // only if) another thread places one or more object(s) 7834 // on the global list before we have returned the suffix 7835 // to the global list, we will walk down our local list 7836 // to find its end and append the global list to 7837 // our suffix before returning it. This suffix walk can 7838 // prove to be expensive (quadratic in the amount of traffic) 7839 // when there are many objects in the overflow list and 7840 // there is much producer-consumer contention on the list. 7841 // *NOTE*: The overflow list manipulation code here and 7842 // in ParNewGeneration:: are very similar in shape, 7843 // except that in the ParNew case we use the old (from/eden) 7844 // copy of the object to thread the list via its klass word. 7845 // Because of the common code, if you make any changes in 7846 // the code below, please check the ParNew version to see if 7847 // similar changes might be needed. 7848 // CR 6797058 has been filed to consolidate the common code. 7849 bool CMSCollector::par_take_from_overflow_list(size_t num, 7850 OopTaskQueue* work_q, 7851 int no_of_gc_threads) { 7852 assert(work_q->size() == 0, "First empty local work queue"); 7853 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7854 if (_overflow_list == NULL) { 7855 return false; 7856 } 7857 // Grab the entire list; we'll put back a suffix 7858 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7859 Thread* tid = Thread::current(); 7860 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7861 // set to ParallelGCThreads. 7862 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7863 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7864 // If the list is busy, we spin for a short while, 7865 // sleeping between attempts to get the list. 7866 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7867 os::sleep(tid, sleep_time_millis, false); 7868 if (_overflow_list == NULL) { 7869 // Nothing left to take 7870 return false; 7871 } else if (_overflow_list != BUSY) { 7872 // Try and grab the prefix 7873 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7874 } 7875 } 7876 // If the list was found to be empty, or we spun long 7877 // enough, we give up and return empty-handed. If we leave 7878 // the list in the BUSY state below, it must be the case that 7879 // some other thread holds the overflow list and will set it 7880 // to a non-BUSY state in the future. 7881 if (prefix == NULL || prefix == BUSY) { 7882 // Nothing to take or waited long enough 7883 if (prefix == NULL) { 7884 // Write back the NULL in case we overwrote it with BUSY above 7885 // and it is still the same value. 7886 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7887 } 7888 return false; 7889 } 7890 assert(prefix != NULL && prefix != BUSY, "Error"); 7891 size_t i = num; 7892 oop cur = prefix; 7893 // Walk down the first "num" objects, unless we reach the end. 7894 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7895 if (cur->mark() == NULL) { 7896 // We have "num" or fewer elements in the list, so there 7897 // is nothing to return to the global list. 7898 // Write back the NULL in lieu of the BUSY we wrote 7899 // above, if it is still the same value. 7900 if (_overflow_list == BUSY) { 7901 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7902 } 7903 } else { 7904 // Chop off the suffix and return it to the global list. 7905 assert(cur->mark() != BUSY, "Error"); 7906 oop suffix_head = cur->mark(); // suffix will be put back on global list 7907 cur->set_mark(NULL); // break off suffix 7908 // It's possible that the list is still in the empty(busy) state 7909 // we left it in a short while ago; in that case we may be 7910 // able to place back the suffix without incurring the cost 7911 // of a walk down the list. 7912 oop observed_overflow_list = _overflow_list; 7913 oop cur_overflow_list = observed_overflow_list; 7914 bool attached = false; 7915 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7916 observed_overflow_list = 7917 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7918 if (cur_overflow_list == observed_overflow_list) { 7919 attached = true; 7920 break; 7921 } else cur_overflow_list = observed_overflow_list; 7922 } 7923 if (!attached) { 7924 // Too bad, someone else sneaked in (at least) an element; we'll need 7925 // to do a splice. Find tail of suffix so we can prepend suffix to global 7926 // list. 7927 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7928 oop suffix_tail = cur; 7929 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7930 "Tautology"); 7931 observed_overflow_list = _overflow_list; 7932 do { 7933 cur_overflow_list = observed_overflow_list; 7934 if (cur_overflow_list != BUSY) { 7935 // Do the splice ... 7936 suffix_tail->set_mark(markOop(cur_overflow_list)); 7937 } else { // cur_overflow_list == BUSY 7938 suffix_tail->set_mark(NULL); 7939 } 7940 // ... and try to place spliced list back on overflow_list ... 7941 observed_overflow_list = 7942 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7943 } while (cur_overflow_list != observed_overflow_list); 7944 // ... until we have succeeded in doing so. 7945 } 7946 } 7947 7948 // Push the prefix elements on work_q 7949 assert(prefix != NULL, "control point invariant"); 7950 const markOop proto = markOopDesc::prototype(); 7951 oop next; 7952 NOT_PRODUCT(ssize_t n = 0;) 7953 for (cur = prefix; cur != NULL; cur = next) { 7954 next = oop(cur->mark()); 7955 cur->set_mark(proto); // until proven otherwise 7956 assert(cur->is_oop(), "Should be an oop"); 7957 bool res = work_q->push(cur); 7958 assert(res, "Bit off more than we can chew?"); 7959 NOT_PRODUCT(n++;) 7960 } 7961 #ifndef PRODUCT 7962 assert(_num_par_pushes >= n, "Too many pops?"); 7963 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7964 #endif 7965 return true; 7966 } 7967 7968 // Single-threaded 7969 void CMSCollector::push_on_overflow_list(oop p) { 7970 NOT_PRODUCT(_num_par_pushes++;) 7971 assert(p->is_oop(), "Not an oop"); 7972 preserve_mark_if_necessary(p); 7973 p->set_mark((markOop)_overflow_list); 7974 _overflow_list = p; 7975 } 7976 7977 // Multi-threaded; use CAS to prepend to overflow list 7978 void CMSCollector::par_push_on_overflow_list(oop p) { 7979 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7980 assert(p->is_oop(), "Not an oop"); 7981 par_preserve_mark_if_necessary(p); 7982 oop observed_overflow_list = _overflow_list; 7983 oop cur_overflow_list; 7984 do { 7985 cur_overflow_list = observed_overflow_list; 7986 if (cur_overflow_list != BUSY) { 7987 p->set_mark(markOop(cur_overflow_list)); 7988 } else { 7989 p->set_mark(NULL); 7990 } 7991 observed_overflow_list = 7992 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7993 } while (cur_overflow_list != observed_overflow_list); 7994 } 7995 #undef BUSY 7996 7997 // Single threaded 7998 // General Note on GrowableArray: pushes may silently fail 7999 // because we are (temporarily) out of C-heap for expanding 8000 // the stack. The problem is quite ubiquitous and affects 8001 // a lot of code in the JVM. The prudent thing for GrowableArray 8002 // to do (for now) is to exit with an error. However, that may 8003 // be too draconian in some cases because the caller may be 8004 // able to recover without much harm. For such cases, we 8005 // should probably introduce a "soft_push" method which returns 8006 // an indication of success or failure with the assumption that 8007 // the caller may be able to recover from a failure; code in 8008 // the VM can then be changed, incrementally, to deal with such 8009 // failures where possible, thus, incrementally hardening the VM 8010 // in such low resource situations. 8011 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8012 _preserved_oop_stack.push(p); 8013 _preserved_mark_stack.push(m); 8014 assert(m == p->mark(), "Mark word changed"); 8015 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8016 "bijection"); 8017 } 8018 8019 // Single threaded 8020 void CMSCollector::preserve_mark_if_necessary(oop p) { 8021 markOop m = p->mark(); 8022 if (m->must_be_preserved(p)) { 8023 preserve_mark_work(p, m); 8024 } 8025 } 8026 8027 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8028 markOop m = p->mark(); 8029 if (m->must_be_preserved(p)) { 8030 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8031 // Even though we read the mark word without holding 8032 // the lock, we are assured that it will not change 8033 // because we "own" this oop, so no other thread can 8034 // be trying to push it on the overflow list; see 8035 // the assertion in preserve_mark_work() that checks 8036 // that m == p->mark(). 8037 preserve_mark_work(p, m); 8038 } 8039 } 8040 8041 // We should be able to do this multi-threaded, 8042 // a chunk of stack being a task (this is 8043 // correct because each oop only ever appears 8044 // once in the overflow list. However, it's 8045 // not very easy to completely overlap this with 8046 // other operations, so will generally not be done 8047 // until all work's been completed. Because we 8048 // expect the preserved oop stack (set) to be small, 8049 // it's probably fine to do this single-threaded. 8050 // We can explore cleverer concurrent/overlapped/parallel 8051 // processing of preserved marks if we feel the 8052 // need for this in the future. Stack overflow should 8053 // be so rare in practice and, when it happens, its 8054 // effect on performance so great that this will 8055 // likely just be in the noise anyway. 8056 void CMSCollector::restore_preserved_marks_if_any() { 8057 assert(SafepointSynchronize::is_at_safepoint(), 8058 "world should be stopped"); 8059 assert(Thread::current()->is_ConcurrentGC_thread() || 8060 Thread::current()->is_VM_thread(), 8061 "should be single-threaded"); 8062 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8063 "bijection"); 8064 8065 while (!_preserved_oop_stack.is_empty()) { 8066 oop p = _preserved_oop_stack.pop(); 8067 assert(p->is_oop(), "Should be an oop"); 8068 assert(_span.contains(p), "oop should be in _span"); 8069 assert(p->mark() == markOopDesc::prototype(), 8070 "Set when taken from overflow list"); 8071 markOop m = _preserved_mark_stack.pop(); 8072 p->set_mark(m); 8073 } 8074 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8075 "stacks were cleared above"); 8076 } 8077 8078 #ifndef PRODUCT 8079 bool CMSCollector::no_preserved_marks() const { 8080 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8081 } 8082 #endif 8083 8084 // Transfer some number of overflown objects to usual marking 8085 // stack. Return true if some objects were transferred. 8086 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8087 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8088 (size_t)ParGCDesiredObjsFromOverflowList); 8089 8090 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8091 assert(_collector->overflow_list_is_empty() || res, 8092 "If list is not empty, we should have taken something"); 8093 assert(!res || !_mark_stack->isEmpty(), 8094 "If we took something, it should now be on our stack"); 8095 return res; 8096 } 8097 8098 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8099 size_t res = _sp->block_size_no_stall(addr, _collector); 8100 if (_sp->block_is_obj(addr)) { 8101 if (_live_bit_map->isMarked(addr)) { 8102 // It can't have been dead in a previous cycle 8103 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8104 } else { 8105 _dead_bit_map->mark(addr); // mark the dead object 8106 } 8107 } 8108 // Could be 0, if the block size could not be computed without stalling. 8109 return res; 8110 } 8111 8112 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8113 8114 switch (phase) { 8115 case CMSCollector::InitialMarking: 8116 initialize(true /* fullGC */ , 8117 cause /* cause of the GC */, 8118 true /* recordGCBeginTime */, 8119 true /* recordPreGCUsage */, 8120 false /* recordPeakUsage */, 8121 false /* recordPostGCusage */, 8122 true /* recordAccumulatedGCTime */, 8123 false /* recordGCEndTime */, 8124 false /* countCollection */ ); 8125 break; 8126 8127 case CMSCollector::FinalMarking: 8128 initialize(true /* fullGC */ , 8129 cause /* cause of the GC */, 8130 false /* recordGCBeginTime */, 8131 false /* recordPreGCUsage */, 8132 false /* recordPeakUsage */, 8133 false /* recordPostGCusage */, 8134 true /* recordAccumulatedGCTime */, 8135 false /* recordGCEndTime */, 8136 false /* countCollection */ ); 8137 break; 8138 8139 case CMSCollector::Sweeping: 8140 initialize(true /* fullGC */ , 8141 cause /* cause of the GC */, 8142 false /* recordGCBeginTime */, 8143 false /* recordPreGCUsage */, 8144 true /* recordPeakUsage */, 8145 true /* recordPostGCusage */, 8146 false /* recordAccumulatedGCTime */, 8147 true /* recordGCEndTime */, 8148 true /* countCollection */ ); 8149 break; 8150 8151 default: 8152 ShouldNotReachHere(); 8153 } 8154 }