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