1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsHeap.hpp" 33 #include "gc/cms/cmsOopClosures.inline.hpp" 34 #include "gc/cms/compactibleFreeListSpace.hpp" 35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc/cms/concurrentMarkSweepThread.hpp" 37 #include "gc/cms/parNewGeneration.hpp" 38 #include "gc/cms/vmCMSOperations.hpp" 39 #include "gc/serial/genMarkSweep.hpp" 40 #include "gc/serial/tenuredGeneration.hpp" 41 #include "gc/shared/adaptiveSizePolicy.hpp" 42 #include "gc/shared/cardGeneration.inline.hpp" 43 #include "gc/shared/cardTableRS.hpp" 44 #include "gc/shared/collectedHeap.inline.hpp" 45 #include "gc/shared/collectorCounters.hpp" 46 #include "gc/shared/collectorPolicy.hpp" 47 #include "gc/shared/gcLocker.hpp" 48 #include "gc/shared/gcPolicyCounters.hpp" 49 #include "gc/shared/gcTimer.hpp" 50 #include "gc/shared/gcTrace.hpp" 51 #include "gc/shared/gcTraceTime.inline.hpp" 52 #include "gc/shared/genCollectedHeap.hpp" 53 #include "gc/shared/genOopClosures.inline.hpp" 54 #include "gc/shared/isGCActiveMark.hpp" 55 #include "gc/shared/referencePolicy.hpp" 56 #include "gc/shared/strongRootsScope.hpp" 57 #include "gc/shared/taskqueue.inline.hpp" 58 #include "gc/shared/weakProcessor.hpp" 59 #include "logging/log.hpp" 60 #include "logging/logStream.hpp" 61 #include "memory/allocation.hpp" 62 #include "memory/binaryTreeDictionary.inline.hpp" 63 #include "memory/iterator.inline.hpp" 64 #include "memory/padded.hpp" 65 #include "memory/resourceArea.hpp" 66 #include "oops/access.inline.hpp" 67 #include "oops/oop.inline.hpp" 68 #include "prims/jvmtiExport.hpp" 69 #include "runtime/atomic.hpp" 70 #include "runtime/globals_extension.hpp" 71 #include "runtime/handles.inline.hpp" 72 #include "runtime/java.hpp" 73 #include "runtime/orderAccess.inline.hpp" 74 #include "runtime/timer.hpp" 75 #include "runtime/vmThread.hpp" 76 #include "services/memoryService.hpp" 77 #include "services/runtimeService.hpp" 78 #include "utilities/align.hpp" 79 #include "utilities/stack.inline.hpp" 80 81 // statics 82 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 83 bool CMSCollector::_full_gc_requested = false; 84 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 85 86 ////////////////////////////////////////////////////////////////// 87 // In support of CMS/VM thread synchronization 88 ////////////////////////////////////////////////////////////////// 89 // We split use of the CGC_lock into 2 "levels". 90 // The low-level locking is of the usual CGC_lock monitor. We introduce 91 // a higher level "token" (hereafter "CMS token") built on top of the 92 // low level monitor (hereafter "CGC lock"). 93 // The token-passing protocol gives priority to the VM thread. The 94 // CMS-lock doesn't provide any fairness guarantees, but clients 95 // should ensure that it is only held for very short, bounded 96 // durations. 97 // 98 // When either of the CMS thread or the VM thread is involved in 99 // collection operations during which it does not want the other 100 // thread to interfere, it obtains the CMS token. 101 // 102 // If either thread tries to get the token while the other has 103 // it, that thread waits. However, if the VM thread and CMS thread 104 // both want the token, then the VM thread gets priority while the 105 // CMS thread waits. This ensures, for instance, that the "concurrent" 106 // phases of the CMS thread's work do not block out the VM thread 107 // for long periods of time as the CMS thread continues to hog 108 // the token. (See bug 4616232). 109 // 110 // The baton-passing functions are, however, controlled by the 111 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 112 // and here the low-level CMS lock, not the high level token, 113 // ensures mutual exclusion. 114 // 115 // Two important conditions that we have to satisfy: 116 // 1. if a thread does a low-level wait on the CMS lock, then it 117 // relinquishes the CMS token if it were holding that token 118 // when it acquired the low-level CMS lock. 119 // 2. any low-level notifications on the low-level lock 120 // should only be sent when a thread has relinquished the token. 121 // 122 // In the absence of either property, we'd have potential deadlock. 123 // 124 // We protect each of the CMS (concurrent and sequential) phases 125 // with the CMS _token_, not the CMS _lock_. 126 // 127 // The only code protected by CMS lock is the token acquisition code 128 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 129 // baton-passing code. 130 // 131 // Unfortunately, i couldn't come up with a good abstraction to factor and 132 // hide the naked CGC_lock manipulation in the baton-passing code 133 // further below. That's something we should try to do. Also, the proof 134 // of correctness of this 2-level locking scheme is far from obvious, 135 // and potentially quite slippery. We have an uneasy suspicion, for instance, 136 // that there may be a theoretical possibility of delay/starvation in the 137 // low-level lock/wait/notify scheme used for the baton-passing because of 138 // potential interference with the priority scheme embodied in the 139 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 140 // invocation further below and marked with "XXX 20011219YSR". 141 // Indeed, as we note elsewhere, this may become yet more slippery 142 // in the presence of multiple CMS and/or multiple VM threads. XXX 143 144 class CMSTokenSync: public StackObj { 145 private: 146 bool _is_cms_thread; 147 public: 148 CMSTokenSync(bool is_cms_thread): 149 _is_cms_thread(is_cms_thread) { 150 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 151 "Incorrect argument to constructor"); 152 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 153 } 154 155 ~CMSTokenSync() { 156 assert(_is_cms_thread ? 157 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 158 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 159 "Incorrect state"); 160 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 161 } 162 }; 163 164 // Convenience class that does a CMSTokenSync, and then acquires 165 // upto three locks. 166 class CMSTokenSyncWithLocks: public CMSTokenSync { 167 private: 168 // Note: locks are acquired in textual declaration order 169 // and released in the opposite order 170 MutexLockerEx _locker1, _locker2, _locker3; 171 public: 172 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 173 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 174 CMSTokenSync(is_cms_thread), 175 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 176 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 177 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 178 { } 179 }; 180 181 182 ////////////////////////////////////////////////////////////////// 183 // Concurrent Mark-Sweep Generation ///////////////////////////// 184 ////////////////////////////////////////////////////////////////// 185 186 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 187 188 // This struct contains per-thread things necessary to support parallel 189 // young-gen collection. 190 class CMSParGCThreadState: public CHeapObj<mtGC> { 191 public: 192 CompactibleFreeListSpaceLAB lab; 193 PromotionInfo promo; 194 195 // Constructor. 196 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 197 promo.setSpace(cfls); 198 } 199 }; 200 201 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 202 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 203 CardGeneration(rs, initial_byte_size, ct), 204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 205 _did_compact(false) 206 { 207 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 208 HeapWord* end = (HeapWord*) _virtual_space.high(); 209 210 _direct_allocated_words = 0; 211 NOT_PRODUCT( 212 _numObjectsPromoted = 0; 213 _numWordsPromoted = 0; 214 _numObjectsAllocated = 0; 215 _numWordsAllocated = 0; 216 ) 217 218 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 219 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 220 _cmsSpace->_old_gen = this; 221 222 _gc_stats = new CMSGCStats(); 223 224 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 225 // offsets match. The ability to tell free chunks from objects 226 // depends on this property. 227 debug_only( 228 FreeChunk* junk = NULL; 229 assert(UseCompressedClassPointers || 230 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 231 "Offset of FreeChunk::_prev within FreeChunk must match" 232 " that of OopDesc::_klass within OopDesc"); 233 ) 234 235 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 236 for (uint i = 0; i < ParallelGCThreads; i++) { 237 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 238 } 239 240 _incremental_collection_failed = false; 241 // The "dilatation_factor" is the expansion that can occur on 242 // account of the fact that the minimum object size in the CMS 243 // generation may be larger than that in, say, a contiguous young 244 // generation. 245 // Ideally, in the calculation below, we'd compute the dilatation 246 // factor as: MinChunkSize/(promoting_gen's min object size) 247 // Since we do not have such a general query interface for the 248 // promoting generation, we'll instead just use the minimum 249 // object size (which today is a header's worth of space); 250 // note that all arithmetic is in units of HeapWords. 251 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 252 assert(_dilatation_factor >= 1.0, "from previous assert"); 253 } 254 255 256 // The field "_initiating_occupancy" represents the occupancy percentage 257 // at which we trigger a new collection cycle. Unless explicitly specified 258 // via CMSInitiatingOccupancyFraction (argument "io" below), it 259 // is calculated by: 260 // 261 // Let "f" be MinHeapFreeRatio in 262 // 263 // _initiating_occupancy = 100-f + 264 // f * (CMSTriggerRatio/100) 265 // where CMSTriggerRatio is the argument "tr" below. 266 // 267 // That is, if we assume the heap is at its desired maximum occupancy at the 268 // end of a collection, we let CMSTriggerRatio of the (purported) free 269 // space be allocated before initiating a new collection cycle. 270 // 271 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 272 assert(io <= 100 && tr <= 100, "Check the arguments"); 273 if (io >= 0) { 274 _initiating_occupancy = (double)io / 100.0; 275 } else { 276 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 277 (double)(tr * MinHeapFreeRatio) / 100.0) 278 / 100.0; 279 } 280 } 281 282 void ConcurrentMarkSweepGeneration::ref_processor_init() { 283 assert(collector() != NULL, "no collector"); 284 collector()->ref_processor_init(); 285 } 286 287 void CMSCollector::ref_processor_init() { 288 if (_ref_processor == NULL) { 289 // Allocate and initialize a reference processor 290 _ref_processor = 291 new ReferenceProcessor(_span, // span 292 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 293 ParallelGCThreads, // mt processing degree 294 _cmsGen->refs_discovery_is_mt(), // mt discovery 295 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 296 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 297 &_is_alive_closure); // closure for liveness info 298 // Initialize the _ref_processor field of CMSGen 299 _cmsGen->set_ref_processor(_ref_processor); 300 301 } 302 } 303 304 AdaptiveSizePolicy* CMSCollector::size_policy() { 305 return CMSHeap::heap()->size_policy(); 306 } 307 308 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 309 310 const char* gen_name = "old"; 311 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 312 // Generation Counters - generation 1, 1 subspace 313 _gen_counters = new GenerationCounters(gen_name, 1, 1, 314 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 315 316 _space_counters = new GSpaceCounters(gen_name, 0, 317 _virtual_space.reserved_size(), 318 this, _gen_counters); 319 } 320 321 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 322 _cms_gen(cms_gen) 323 { 324 assert(alpha <= 100, "bad value"); 325 _saved_alpha = alpha; 326 327 // Initialize the alphas to the bootstrap value of 100. 328 _gc0_alpha = _cms_alpha = 100; 329 330 _cms_begin_time.update(); 331 _cms_end_time.update(); 332 333 _gc0_duration = 0.0; 334 _gc0_period = 0.0; 335 _gc0_promoted = 0; 336 337 _cms_duration = 0.0; 338 _cms_period = 0.0; 339 _cms_allocated = 0; 340 341 _cms_used_at_gc0_begin = 0; 342 _cms_used_at_gc0_end = 0; 343 _allow_duty_cycle_reduction = false; 344 _valid_bits = 0; 345 } 346 347 double CMSStats::cms_free_adjustment_factor(size_t free) const { 348 // TBD: CR 6909490 349 return 1.0; 350 } 351 352 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 353 } 354 355 // If promotion failure handling is on use 356 // the padded average size of the promotion for each 357 // young generation collection. 358 double CMSStats::time_until_cms_gen_full() const { 359 size_t cms_free = _cms_gen->cmsSpace()->free(); 360 CMSHeap* heap = CMSHeap::heap(); 361 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 362 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 363 if (cms_free > expected_promotion) { 364 // Start a cms collection if there isn't enough space to promote 365 // for the next young collection. Use the padded average as 366 // a safety factor. 367 cms_free -= expected_promotion; 368 369 // Adjust by the safety factor. 370 double cms_free_dbl = (double)cms_free; 371 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 372 // Apply a further correction factor which tries to adjust 373 // for recent occurance of concurrent mode failures. 374 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 375 cms_free_dbl = cms_free_dbl * cms_adjustment; 376 377 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 378 cms_free, expected_promotion); 379 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 380 // Add 1 in case the consumption rate goes to zero. 381 return cms_free_dbl / (cms_consumption_rate() + 1.0); 382 } 383 return 0.0; 384 } 385 386 // Compare the duration of the cms collection to the 387 // time remaining before the cms generation is empty. 388 // Note that the time from the start of the cms collection 389 // to the start of the cms sweep (less than the total 390 // duration of the cms collection) can be used. This 391 // has been tried and some applications experienced 392 // promotion failures early in execution. This was 393 // possibly because the averages were not accurate 394 // enough at the beginning. 395 double CMSStats::time_until_cms_start() const { 396 // We add "gc0_period" to the "work" calculation 397 // below because this query is done (mostly) at the 398 // end of a scavenge, so we need to conservatively 399 // account for that much possible delay 400 // in the query so as to avoid concurrent mode failures 401 // due to starting the collection just a wee bit too 402 // late. 403 double work = cms_duration() + gc0_period(); 404 double deadline = time_until_cms_gen_full(); 405 // If a concurrent mode failure occurred recently, we want to be 406 // more conservative and halve our expected time_until_cms_gen_full() 407 if (work > deadline) { 408 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 409 cms_duration(), gc0_period(), time_until_cms_gen_full()); 410 return 0.0; 411 } 412 return work - deadline; 413 } 414 415 #ifndef PRODUCT 416 void CMSStats::print_on(outputStream *st) const { 417 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 418 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 419 gc0_duration(), gc0_period(), gc0_promoted()); 420 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 421 cms_duration(), cms_period(), cms_allocated()); 422 st->print(",cms_since_beg=%g,cms_since_end=%g", 423 cms_time_since_begin(), cms_time_since_end()); 424 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 425 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 426 427 if (valid()) { 428 st->print(",promo_rate=%g,cms_alloc_rate=%g", 429 promotion_rate(), cms_allocation_rate()); 430 st->print(",cms_consumption_rate=%g,time_until_full=%g", 431 cms_consumption_rate(), time_until_cms_gen_full()); 432 } 433 st->cr(); 434 } 435 #endif // #ifndef PRODUCT 436 437 CMSCollector::CollectorState CMSCollector::_collectorState = 438 CMSCollector::Idling; 439 bool CMSCollector::_foregroundGCIsActive = false; 440 bool CMSCollector::_foregroundGCShouldWait = false; 441 442 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 443 CardTableRS* ct, 444 ConcurrentMarkSweepPolicy* cp): 445 _cmsGen(cmsGen), 446 _ct(ct), 447 _ref_processor(NULL), // will be set later 448 _conc_workers(NULL), // may be set later 449 _abort_preclean(false), 450 _start_sampling(false), 451 _between_prologue_and_epilogue(false), 452 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 453 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 454 -1 /* lock-free */, "No_lock" /* dummy */), 455 _modUnionClosurePar(&_modUnionTable), 456 // Adjust my span to cover old (cms) gen 457 _span(cmsGen->reserved()), 458 // Construct the is_alive_closure with _span & markBitMap 459 _is_alive_closure(_span, &_markBitMap), 460 _restart_addr(NULL), 461 _overflow_list(NULL), 462 _stats(cmsGen), 463 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 464 //verify that this lock should be acquired with safepoint check. 465 Monitor::_safepoint_check_sometimes)), 466 _eden_chunk_array(NULL), // may be set in ctor body 467 _eden_chunk_capacity(0), // -- ditto -- 468 _eden_chunk_index(0), // -- ditto -- 469 _survivor_plab_array(NULL), // -- ditto -- 470 _survivor_chunk_array(NULL), // -- ditto -- 471 _survivor_chunk_capacity(0), // -- ditto -- 472 _survivor_chunk_index(0), // -- ditto -- 473 _ser_pmc_preclean_ovflw(0), 474 _ser_kac_preclean_ovflw(0), 475 _ser_pmc_remark_ovflw(0), 476 _par_pmc_remark_ovflw(0), 477 _ser_kac_ovflw(0), 478 _par_kac_ovflw(0), 479 #ifndef PRODUCT 480 _num_par_pushes(0), 481 #endif 482 _collection_count_start(0), 483 _verifying(false), 484 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 485 _completed_initialization(false), 486 _collector_policy(cp), 487 _should_unload_classes(CMSClassUnloadingEnabled), 488 _concurrent_cycles_since_last_unload(0), 489 _roots_scanning_options(GenCollectedHeap::SO_None), 490 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 491 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 492 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 493 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 494 _cms_start_registered(false) 495 { 496 // Now expand the span and allocate the collection support structures 497 // (MUT, marking bit map etc.) to cover both generations subject to 498 // collection. 499 500 // For use by dirty card to oop closures. 501 _cmsGen->cmsSpace()->set_collector(this); 502 503 // Allocate MUT and marking bit map 504 { 505 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 506 if (!_markBitMap.allocate(_span)) { 507 log_warning(gc)("Failed to allocate CMS Bit Map"); 508 return; 509 } 510 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 511 } 512 { 513 _modUnionTable.allocate(_span); 514 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 515 } 516 517 if (!_markStack.allocate(MarkStackSize)) { 518 log_warning(gc)("Failed to allocate CMS Marking Stack"); 519 return; 520 } 521 522 // Support for multi-threaded concurrent phases 523 if (CMSConcurrentMTEnabled) { 524 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 525 // just for now 526 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 527 } 528 if (ConcGCThreads > 1) { 529 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 530 ConcGCThreads, true); 531 if (_conc_workers == NULL) { 532 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 533 CMSConcurrentMTEnabled = false; 534 } else { 535 _conc_workers->initialize_workers(); 536 } 537 } else { 538 CMSConcurrentMTEnabled = false; 539 } 540 } 541 if (!CMSConcurrentMTEnabled) { 542 ConcGCThreads = 0; 543 } else { 544 // Turn off CMSCleanOnEnter optimization temporarily for 545 // the MT case where it's not fixed yet; see 6178663. 546 CMSCleanOnEnter = false; 547 } 548 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 549 "Inconsistency"); 550 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 551 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 552 553 // Parallel task queues; these are shared for the 554 // concurrent and stop-world phases of CMS, but 555 // are not shared with parallel scavenge (ParNew). 556 { 557 uint i; 558 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 559 560 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 561 || ParallelRefProcEnabled) 562 && num_queues > 0) { 563 _task_queues = new OopTaskQueueSet(num_queues); 564 if (_task_queues == NULL) { 565 log_warning(gc)("task_queues allocation failure."); 566 return; 567 } 568 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 569 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 570 for (i = 0; i < num_queues; i++) { 571 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 572 if (q == NULL) { 573 log_warning(gc)("work_queue allocation failure."); 574 return; 575 } 576 _task_queues->register_queue(i, q); 577 } 578 for (i = 0; i < num_queues; i++) { 579 _task_queues->queue(i)->initialize(); 580 _hash_seed[i] = 17; // copied from ParNew 581 } 582 } 583 } 584 585 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 586 587 // Clip CMSBootstrapOccupancy between 0 and 100. 588 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 589 590 // Now tell CMS generations the identity of their collector 591 ConcurrentMarkSweepGeneration::set_collector(this); 592 593 // Create & start a CMS thread for this CMS collector 594 _cmsThread = ConcurrentMarkSweepThread::start(this); 595 assert(cmsThread() != NULL, "CMS Thread should have been created"); 596 assert(cmsThread()->collector() == this, 597 "CMS Thread should refer to this gen"); 598 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 599 600 // Support for parallelizing young gen rescan 601 CMSHeap* heap = CMSHeap::heap(); 602 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 603 _young_gen = (ParNewGeneration*)heap->young_gen(); 604 if (heap->supports_inline_contig_alloc()) { 605 _top_addr = heap->top_addr(); 606 _end_addr = heap->end_addr(); 607 assert(_young_gen != NULL, "no _young_gen"); 608 _eden_chunk_index = 0; 609 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 610 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 611 } 612 613 // Support for parallelizing survivor space rescan 614 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 615 const size_t max_plab_samples = 616 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 617 618 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 619 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 620 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 621 _survivor_chunk_capacity = max_plab_samples; 622 for (uint i = 0; i < ParallelGCThreads; i++) { 623 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 624 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 625 assert(cur->end() == 0, "Should be 0"); 626 assert(cur->array() == vec, "Should be vec"); 627 assert(cur->capacity() == max_plab_samples, "Error"); 628 } 629 } 630 631 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 632 _gc_counters = new CollectorCounters("CMS", 1); 633 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 634 _completed_initialization = true; 635 _inter_sweep_timer.start(); // start of time 636 } 637 638 const char* ConcurrentMarkSweepGeneration::name() const { 639 return "concurrent mark-sweep generation"; 640 } 641 void ConcurrentMarkSweepGeneration::update_counters() { 642 if (UsePerfData) { 643 _space_counters->update_all(); 644 _gen_counters->update_all(); 645 } 646 } 647 648 // this is an optimized version of update_counters(). it takes the 649 // used value as a parameter rather than computing it. 650 // 651 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 652 if (UsePerfData) { 653 _space_counters->update_used(used); 654 _space_counters->update_capacity(); 655 _gen_counters->update_all(); 656 } 657 } 658 659 void ConcurrentMarkSweepGeneration::print() const { 660 Generation::print(); 661 cmsSpace()->print(); 662 } 663 664 #ifndef PRODUCT 665 void ConcurrentMarkSweepGeneration::print_statistics() { 666 cmsSpace()->printFLCensus(0); 667 } 668 #endif 669 670 size_t 671 ConcurrentMarkSweepGeneration::contiguous_available() const { 672 // dld proposes an improvement in precision here. If the committed 673 // part of the space ends in a free block we should add that to 674 // uncommitted size in the calculation below. Will make this 675 // change later, staying with the approximation below for the 676 // time being. -- ysr. 677 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 678 } 679 680 size_t 681 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 682 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 683 } 684 685 size_t ConcurrentMarkSweepGeneration::max_available() const { 686 return free() + _virtual_space.uncommitted_size(); 687 } 688 689 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 690 size_t available = max_available(); 691 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 692 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 693 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 694 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 695 return res; 696 } 697 698 // At a promotion failure dump information on block layout in heap 699 // (cms old generation). 700 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 701 Log(gc, promotion) log; 702 if (log.is_trace()) { 703 LogStream ls(log.trace()); 704 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 705 } 706 } 707 708 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 709 // Clear the promotion information. These pointers can be adjusted 710 // along with all the other pointers into the heap but 711 // compaction is expected to be a rare event with 712 // a heap using cms so don't do it without seeing the need. 713 for (uint i = 0; i < ParallelGCThreads; i++) { 714 _par_gc_thread_states[i]->promo.reset(); 715 } 716 } 717 718 void ConcurrentMarkSweepGeneration::compute_new_size() { 719 assert_locked_or_safepoint(Heap_lock); 720 721 // If incremental collection failed, we just want to expand 722 // to the limit. 723 if (incremental_collection_failed()) { 724 clear_incremental_collection_failed(); 725 grow_to_reserved(); 726 return; 727 } 728 729 // The heap has been compacted but not reset yet. 730 // Any metric such as free() or used() will be incorrect. 731 732 CardGeneration::compute_new_size(); 733 734 // Reset again after a possible resizing 735 if (did_compact()) { 736 cmsSpace()->reset_after_compaction(); 737 } 738 } 739 740 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 741 assert_locked_or_safepoint(Heap_lock); 742 743 // If incremental collection failed, we just want to expand 744 // to the limit. 745 if (incremental_collection_failed()) { 746 clear_incremental_collection_failed(); 747 grow_to_reserved(); 748 return; 749 } 750 751 double free_percentage = ((double) free()) / capacity(); 752 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 753 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 754 755 // compute expansion delta needed for reaching desired free percentage 756 if (free_percentage < desired_free_percentage) { 757 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 758 assert(desired_capacity >= capacity(), "invalid expansion size"); 759 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 760 Log(gc) log; 761 if (log.is_trace()) { 762 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 763 log.trace("From compute_new_size: "); 764 log.trace(" Free fraction %f", free_percentage); 765 log.trace(" Desired free fraction %f", desired_free_percentage); 766 log.trace(" Maximum free fraction %f", maximum_free_percentage); 767 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 768 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 769 CMSHeap* heap = CMSHeap::heap(); 770 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation"); 771 size_t young_size = heap->young_gen()->capacity(); 772 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 773 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 774 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 775 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 776 } 777 // safe if expansion fails 778 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 779 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 780 } else { 781 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 782 assert(desired_capacity <= capacity(), "invalid expansion size"); 783 size_t shrink_bytes = capacity() - desired_capacity; 784 // Don't shrink unless the delta is greater than the minimum shrink we want 785 if (shrink_bytes >= MinHeapDeltaBytes) { 786 shrink_free_list_by(shrink_bytes); 787 } 788 } 789 } 790 791 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 792 return cmsSpace()->freelistLock(); 793 } 794 795 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 796 CMSSynchronousYieldRequest yr; 797 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 798 return have_lock_and_allocate(size, tlab); 799 } 800 801 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 802 bool tlab /* ignored */) { 803 assert_lock_strong(freelistLock()); 804 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 805 HeapWord* res = cmsSpace()->allocate(adjustedSize); 806 // Allocate the object live (grey) if the background collector has 807 // started marking. This is necessary because the marker may 808 // have passed this address and consequently this object will 809 // not otherwise be greyed and would be incorrectly swept up. 810 // Note that if this object contains references, the writing 811 // of those references will dirty the card containing this object 812 // allowing the object to be blackened (and its references scanned) 813 // either during a preclean phase or at the final checkpoint. 814 if (res != NULL) { 815 // We may block here with an uninitialized object with 816 // its mark-bit or P-bits not yet set. Such objects need 817 // to be safely navigable by block_start(). 818 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 819 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 820 collector()->direct_allocated(res, adjustedSize); 821 _direct_allocated_words += adjustedSize; 822 // allocation counters 823 NOT_PRODUCT( 824 _numObjectsAllocated++; 825 _numWordsAllocated += (int)adjustedSize; 826 ) 827 } 828 return res; 829 } 830 831 // In the case of direct allocation by mutators in a generation that 832 // is being concurrently collected, the object must be allocated 833 // live (grey) if the background collector has started marking. 834 // This is necessary because the marker may 835 // have passed this address and consequently this object will 836 // not otherwise be greyed and would be incorrectly swept up. 837 // Note that if this object contains references, the writing 838 // of those references will dirty the card containing this object 839 // allowing the object to be blackened (and its references scanned) 840 // either during a preclean phase or at the final checkpoint. 841 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 842 assert(_markBitMap.covers(start, size), "Out of bounds"); 843 if (_collectorState >= Marking) { 844 MutexLockerEx y(_markBitMap.lock(), 845 Mutex::_no_safepoint_check_flag); 846 // [see comments preceding SweepClosure::do_blk() below for details] 847 // 848 // Can the P-bits be deleted now? JJJ 849 // 850 // 1. need to mark the object as live so it isn't collected 851 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 852 // 3. need to mark the end of the object so marking, precleaning or sweeping 853 // can skip over uninitialized or unparsable objects. An allocated 854 // object is considered uninitialized for our purposes as long as 855 // its klass word is NULL. All old gen objects are parsable 856 // as soon as they are initialized.) 857 _markBitMap.mark(start); // object is live 858 _markBitMap.mark(start + 1); // object is potentially uninitialized? 859 _markBitMap.mark(start + size - 1); 860 // mark end of object 861 } 862 // check that oop looks uninitialized 863 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 864 } 865 866 void CMSCollector::promoted(bool par, HeapWord* start, 867 bool is_obj_array, size_t obj_size) { 868 assert(_markBitMap.covers(start), "Out of bounds"); 869 // See comment in direct_allocated() about when objects should 870 // be allocated live. 871 if (_collectorState >= Marking) { 872 // we already hold the marking bit map lock, taken in 873 // the prologue 874 if (par) { 875 _markBitMap.par_mark(start); 876 } else { 877 _markBitMap.mark(start); 878 } 879 // We don't need to mark the object as uninitialized (as 880 // in direct_allocated above) because this is being done with the 881 // world stopped and the object will be initialized by the 882 // time the marking, precleaning or sweeping get to look at it. 883 // But see the code for copying objects into the CMS generation, 884 // where we need to ensure that concurrent readers of the 885 // block offset table are able to safely navigate a block that 886 // is in flux from being free to being allocated (and in 887 // transition while being copied into) and subsequently 888 // becoming a bona-fide object when the copy/promotion is complete. 889 assert(SafepointSynchronize::is_at_safepoint(), 890 "expect promotion only at safepoints"); 891 892 if (_collectorState < Sweeping) { 893 // Mark the appropriate cards in the modUnionTable, so that 894 // this object gets scanned before the sweep. If this is 895 // not done, CMS generation references in the object might 896 // not get marked. 897 // For the case of arrays, which are otherwise precisely 898 // marked, we need to dirty the entire array, not just its head. 899 if (is_obj_array) { 900 // The [par_]mark_range() method expects mr.end() below to 901 // be aligned to the granularity of a bit's representation 902 // in the heap. In the case of the MUT below, that's a 903 // card size. 904 MemRegion mr(start, 905 align_up(start + obj_size, 906 CardTable::card_size /* bytes */)); 907 if (par) { 908 _modUnionTable.par_mark_range(mr); 909 } else { 910 _modUnionTable.mark_range(mr); 911 } 912 } else { // not an obj array; we can just mark the head 913 if (par) { 914 _modUnionTable.par_mark(start); 915 } else { 916 _modUnionTable.mark(start); 917 } 918 } 919 } 920 } 921 } 922 923 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 924 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 925 // allocate, copy and if necessary update promoinfo -- 926 // delegate to underlying space. 927 assert_lock_strong(freelistLock()); 928 929 #ifndef PRODUCT 930 if (CMSHeap::heap()->promotion_should_fail()) { 931 return NULL; 932 } 933 #endif // #ifndef PRODUCT 934 935 oop res = _cmsSpace->promote(obj, obj_size); 936 if (res == NULL) { 937 // expand and retry 938 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 939 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 940 // Since this is the old generation, we don't try to promote 941 // into a more senior generation. 942 res = _cmsSpace->promote(obj, obj_size); 943 } 944 if (res != NULL) { 945 // See comment in allocate() about when objects should 946 // be allocated live. 947 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 948 collector()->promoted(false, // Not parallel 949 (HeapWord*)res, obj->is_objArray(), obj_size); 950 // promotion counters 951 NOT_PRODUCT( 952 _numObjectsPromoted++; 953 _numWordsPromoted += 954 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 955 ) 956 } 957 return res; 958 } 959 960 961 // IMPORTANT: Notes on object size recognition in CMS. 962 // --------------------------------------------------- 963 // A block of storage in the CMS generation is always in 964 // one of three states. A free block (FREE), an allocated 965 // object (OBJECT) whose size() method reports the correct size, 966 // and an intermediate state (TRANSIENT) in which its size cannot 967 // be accurately determined. 968 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 969 // ----------------------------------------------------- 970 // FREE: klass_word & 1 == 1; mark_word holds block size 971 // 972 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 973 // obj->size() computes correct size 974 // 975 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 976 // 977 // STATE IDENTIFICATION: (64 bit+COOPS) 978 // ------------------------------------ 979 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 980 // 981 // OBJECT: klass_word installed; klass_word != 0; 982 // obj->size() computes correct size 983 // 984 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 985 // 986 // 987 // STATE TRANSITION DIAGRAM 988 // 989 // mut / parnew mut / parnew 990 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 991 // ^ | 992 // |------------------------ DEAD <------------------------------------| 993 // sweep mut 994 // 995 // While a block is in TRANSIENT state its size cannot be determined 996 // so readers will either need to come back later or stall until 997 // the size can be determined. Note that for the case of direct 998 // allocation, P-bits, when available, may be used to determine the 999 // size of an object that may not yet have been initialized. 1000 1001 // Things to support parallel young-gen collection. 1002 oop 1003 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1004 oop old, markOop m, 1005 size_t word_sz) { 1006 #ifndef PRODUCT 1007 if (CMSHeap::heap()->promotion_should_fail()) { 1008 return NULL; 1009 } 1010 #endif // #ifndef PRODUCT 1011 1012 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1013 PromotionInfo* promoInfo = &ps->promo; 1014 // if we are tracking promotions, then first ensure space for 1015 // promotion (including spooling space for saving header if necessary). 1016 // then allocate and copy, then track promoted info if needed. 1017 // When tracking (see PromotionInfo::track()), the mark word may 1018 // be displaced and in this case restoration of the mark word 1019 // occurs in the (oop_since_save_marks_)iterate phase. 1020 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1021 // Out of space for allocating spooling buffers; 1022 // try expanding and allocating spooling buffers. 1023 if (!expand_and_ensure_spooling_space(promoInfo)) { 1024 return NULL; 1025 } 1026 } 1027 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1028 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1029 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1030 if (obj_ptr == NULL) { 1031 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1032 if (obj_ptr == NULL) { 1033 return NULL; 1034 } 1035 } 1036 oop obj = oop(obj_ptr); 1037 OrderAccess::storestore(); 1038 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1039 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1040 // IMPORTANT: See note on object initialization for CMS above. 1041 // Otherwise, copy the object. Here we must be careful to insert the 1042 // klass pointer last, since this marks the block as an allocated object. 1043 // Except with compressed oops it's the mark word. 1044 HeapWord* old_ptr = (HeapWord*)old; 1045 // Restore the mark word copied above. 1046 obj->set_mark(m); 1047 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1048 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1049 OrderAccess::storestore(); 1050 1051 if (UseCompressedClassPointers) { 1052 // Copy gap missed by (aligned) header size calculation below 1053 obj->set_klass_gap(old->klass_gap()); 1054 } 1055 if (word_sz > (size_t)oopDesc::header_size()) { 1056 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1057 obj_ptr + oopDesc::header_size(), 1058 word_sz - oopDesc::header_size()); 1059 } 1060 1061 // Now we can track the promoted object, if necessary. We take care 1062 // to delay the transition from uninitialized to full object 1063 // (i.e., insertion of klass pointer) until after, so that it 1064 // atomically becomes a promoted object. 1065 if (promoInfo->tracking()) { 1066 promoInfo->track((PromotedObject*)obj, old->klass()); 1067 } 1068 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1069 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1070 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1071 1072 // Finally, install the klass pointer (this should be volatile). 1073 OrderAccess::storestore(); 1074 obj->set_klass(old->klass()); 1075 // We should now be able to calculate the right size for this object 1076 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1077 1078 collector()->promoted(true, // parallel 1079 obj_ptr, old->is_objArray(), word_sz); 1080 1081 NOT_PRODUCT( 1082 Atomic::inc(&_numObjectsPromoted); 1083 Atomic::add(alloc_sz, &_numWordsPromoted); 1084 ) 1085 1086 return obj; 1087 } 1088 1089 void 1090 ConcurrentMarkSweepGeneration:: 1091 par_promote_alloc_done(int thread_num) { 1092 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1093 ps->lab.retire(thread_num); 1094 } 1095 1096 void 1097 ConcurrentMarkSweepGeneration:: 1098 par_oop_since_save_marks_iterate_done(int thread_num) { 1099 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1100 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1101 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1102 1103 // Because card-scanning has been completed, subsequent phases 1104 // (e.g., reference processing) will not need to recognize which 1105 // objects have been promoted during this GC. So, we can now disable 1106 // promotion tracking. 1107 ps->promo.stopTrackingPromotions(); 1108 } 1109 1110 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1111 size_t size, 1112 bool tlab) 1113 { 1114 // We allow a STW collection only if a full 1115 // collection was requested. 1116 return full || should_allocate(size, tlab); // FIX ME !!! 1117 // This and promotion failure handling are connected at the 1118 // hip and should be fixed by untying them. 1119 } 1120 1121 bool CMSCollector::shouldConcurrentCollect() { 1122 LogTarget(Trace, gc) log; 1123 1124 if (_full_gc_requested) { 1125 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1126 return true; 1127 } 1128 1129 FreelistLocker x(this); 1130 // ------------------------------------------------------------------ 1131 // Print out lots of information which affects the initiation of 1132 // a collection. 1133 if (log.is_enabled() && stats().valid()) { 1134 log.print("CMSCollector shouldConcurrentCollect: "); 1135 1136 LogStream out(log); 1137 stats().print_on(&out); 1138 1139 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1140 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1141 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1142 log.print("promotion_rate=%g", stats().promotion_rate()); 1143 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1144 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1145 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1146 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1147 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1148 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1149 } 1150 // ------------------------------------------------------------------ 1151 1152 // If the estimated time to complete a cms collection (cms_duration()) 1153 // is less than the estimated time remaining until the cms generation 1154 // is full, start a collection. 1155 if (!UseCMSInitiatingOccupancyOnly) { 1156 if (stats().valid()) { 1157 if (stats().time_until_cms_start() == 0.0) { 1158 return true; 1159 } 1160 } else { 1161 // We want to conservatively collect somewhat early in order 1162 // to try and "bootstrap" our CMS/promotion statistics; 1163 // this branch will not fire after the first successful CMS 1164 // collection because the stats should then be valid. 1165 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1166 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1167 _cmsGen->occupancy(), _bootstrap_occupancy); 1168 return true; 1169 } 1170 } 1171 } 1172 1173 // Otherwise, we start a collection cycle if 1174 // old gen want a collection cycle started. Each may use 1175 // an appropriate criterion for making this decision. 1176 // XXX We need to make sure that the gen expansion 1177 // criterion dovetails well with this. XXX NEED TO FIX THIS 1178 if (_cmsGen->should_concurrent_collect()) { 1179 log.print("CMS old gen initiated"); 1180 return true; 1181 } 1182 1183 // We start a collection if we believe an incremental collection may fail; 1184 // this is not likely to be productive in practice because it's probably too 1185 // late anyway. 1186 CMSHeap* heap = CMSHeap::heap(); 1187 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1188 log.print("CMSCollector: collect because incremental collection will fail "); 1189 return true; 1190 } 1191 1192 if (MetaspaceGC::should_concurrent_collect()) { 1193 log.print("CMSCollector: collect for metadata allocation "); 1194 return true; 1195 } 1196 1197 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1198 if (CMSTriggerInterval >= 0) { 1199 if (CMSTriggerInterval == 0) { 1200 // Trigger always 1201 return true; 1202 } 1203 1204 // Check the CMS time since begin (we do not check the stats validity 1205 // as we want to be able to trigger the first CMS cycle as well) 1206 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1207 if (stats().valid()) { 1208 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1209 stats().cms_time_since_begin()); 1210 } else { 1211 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1212 } 1213 return true; 1214 } 1215 } 1216 1217 return false; 1218 } 1219 1220 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1221 1222 // Clear _expansion_cause fields of constituent generations 1223 void CMSCollector::clear_expansion_cause() { 1224 _cmsGen->clear_expansion_cause(); 1225 } 1226 1227 // We should be conservative in starting a collection cycle. To 1228 // start too eagerly runs the risk of collecting too often in the 1229 // extreme. To collect too rarely falls back on full collections, 1230 // which works, even if not optimum in terms of concurrent work. 1231 // As a work around for too eagerly collecting, use the flag 1232 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1233 // giving the user an easily understandable way of controlling the 1234 // collections. 1235 // We want to start a new collection cycle if any of the following 1236 // conditions hold: 1237 // . our current occupancy exceeds the configured initiating occupancy 1238 // for this generation, or 1239 // . we recently needed to expand this space and have not, since that 1240 // expansion, done a collection of this generation, or 1241 // . the underlying space believes that it may be a good idea to initiate 1242 // a concurrent collection (this may be based on criteria such as the 1243 // following: the space uses linear allocation and linear allocation is 1244 // going to fail, or there is believed to be excessive fragmentation in 1245 // the generation, etc... or ... 1246 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1247 // the case of the old generation; see CR 6543076): 1248 // we may be approaching a point at which allocation requests may fail because 1249 // we will be out of sufficient free space given allocation rate estimates.] 1250 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1251 1252 assert_lock_strong(freelistLock()); 1253 if (occupancy() > initiating_occupancy()) { 1254 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1255 short_name(), occupancy(), initiating_occupancy()); 1256 return true; 1257 } 1258 if (UseCMSInitiatingOccupancyOnly) { 1259 return false; 1260 } 1261 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1262 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1263 return true; 1264 } 1265 return false; 1266 } 1267 1268 void ConcurrentMarkSweepGeneration::collect(bool full, 1269 bool clear_all_soft_refs, 1270 size_t size, 1271 bool tlab) 1272 { 1273 collector()->collect(full, clear_all_soft_refs, size, tlab); 1274 } 1275 1276 void CMSCollector::collect(bool full, 1277 bool clear_all_soft_refs, 1278 size_t size, 1279 bool tlab) 1280 { 1281 // The following "if" branch is present for defensive reasons. 1282 // In the current uses of this interface, it can be replaced with: 1283 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1284 // But I am not placing that assert here to allow future 1285 // generality in invoking this interface. 1286 if (GCLocker::is_active()) { 1287 // A consistency test for GCLocker 1288 assert(GCLocker::needs_gc(), "Should have been set already"); 1289 // Skip this foreground collection, instead 1290 // expanding the heap if necessary. 1291 // Need the free list locks for the call to free() in compute_new_size() 1292 compute_new_size(); 1293 return; 1294 } 1295 acquire_control_and_collect(full, clear_all_soft_refs); 1296 } 1297 1298 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1299 CMSHeap* heap = CMSHeap::heap(); 1300 unsigned int gc_count = heap->total_full_collections(); 1301 if (gc_count == full_gc_count) { 1302 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1303 _full_gc_requested = true; 1304 _full_gc_cause = cause; 1305 CGC_lock->notify(); // nudge CMS thread 1306 } else { 1307 assert(gc_count > full_gc_count, "Error: causal loop"); 1308 } 1309 } 1310 1311 bool CMSCollector::is_external_interruption() { 1312 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1313 return GCCause::is_user_requested_gc(cause) || 1314 GCCause::is_serviceability_requested_gc(cause); 1315 } 1316 1317 void CMSCollector::report_concurrent_mode_interruption() { 1318 if (is_external_interruption()) { 1319 log_debug(gc)("Concurrent mode interrupted"); 1320 } else { 1321 log_debug(gc)("Concurrent mode failure"); 1322 _gc_tracer_cm->report_concurrent_mode_failure(); 1323 } 1324 } 1325 1326 1327 // The foreground and background collectors need to coordinate in order 1328 // to make sure that they do not mutually interfere with CMS collections. 1329 // When a background collection is active, 1330 // the foreground collector may need to take over (preempt) and 1331 // synchronously complete an ongoing collection. Depending on the 1332 // frequency of the background collections and the heap usage 1333 // of the application, this preemption can be seldom or frequent. 1334 // There are only certain 1335 // points in the background collection that the "collection-baton" 1336 // can be passed to the foreground collector. 1337 // 1338 // The foreground collector will wait for the baton before 1339 // starting any part of the collection. The foreground collector 1340 // will only wait at one location. 1341 // 1342 // The background collector will yield the baton before starting a new 1343 // phase of the collection (e.g., before initial marking, marking from roots, 1344 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1345 // of the loop which switches the phases. The background collector does some 1346 // of the phases (initial mark, final re-mark) with the world stopped. 1347 // Because of locking involved in stopping the world, 1348 // the foreground collector should not block waiting for the background 1349 // collector when it is doing a stop-the-world phase. The background 1350 // collector will yield the baton at an additional point just before 1351 // it enters a stop-the-world phase. Once the world is stopped, the 1352 // background collector checks the phase of the collection. If the 1353 // phase has not changed, it proceeds with the collection. If the 1354 // phase has changed, it skips that phase of the collection. See 1355 // the comments on the use of the Heap_lock in collect_in_background(). 1356 // 1357 // Variable used in baton passing. 1358 // _foregroundGCIsActive - Set to true by the foreground collector when 1359 // it wants the baton. The foreground clears it when it has finished 1360 // the collection. 1361 // _foregroundGCShouldWait - Set to true by the background collector 1362 // when it is running. The foreground collector waits while 1363 // _foregroundGCShouldWait is true. 1364 // CGC_lock - monitor used to protect access to the above variables 1365 // and to notify the foreground and background collectors. 1366 // _collectorState - current state of the CMS collection. 1367 // 1368 // The foreground collector 1369 // acquires the CGC_lock 1370 // sets _foregroundGCIsActive 1371 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1372 // various locks acquired in preparation for the collection 1373 // are released so as not to block the background collector 1374 // that is in the midst of a collection 1375 // proceeds with the collection 1376 // clears _foregroundGCIsActive 1377 // returns 1378 // 1379 // The background collector in a loop iterating on the phases of the 1380 // collection 1381 // acquires the CGC_lock 1382 // sets _foregroundGCShouldWait 1383 // if _foregroundGCIsActive is set 1384 // clears _foregroundGCShouldWait, notifies _CGC_lock 1385 // waits on _CGC_lock for _foregroundGCIsActive to become false 1386 // and exits the loop. 1387 // otherwise 1388 // proceed with that phase of the collection 1389 // if the phase is a stop-the-world phase, 1390 // yield the baton once more just before enqueueing 1391 // the stop-world CMS operation (executed by the VM thread). 1392 // returns after all phases of the collection are done 1393 // 1394 1395 void CMSCollector::acquire_control_and_collect(bool full, 1396 bool clear_all_soft_refs) { 1397 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1398 assert(!Thread::current()->is_ConcurrentGC_thread(), 1399 "shouldn't try to acquire control from self!"); 1400 1401 // Start the protocol for acquiring control of the 1402 // collection from the background collector (aka CMS thread). 1403 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1404 "VM thread should have CMS token"); 1405 // Remember the possibly interrupted state of an ongoing 1406 // concurrent collection 1407 CollectorState first_state = _collectorState; 1408 1409 // Signal to a possibly ongoing concurrent collection that 1410 // we want to do a foreground collection. 1411 _foregroundGCIsActive = true; 1412 1413 // release locks and wait for a notify from the background collector 1414 // releasing the locks in only necessary for phases which 1415 // do yields to improve the granularity of the collection. 1416 assert_lock_strong(bitMapLock()); 1417 // We need to lock the Free list lock for the space that we are 1418 // currently collecting. 1419 assert(haveFreelistLocks(), "Must be holding free list locks"); 1420 bitMapLock()->unlock(); 1421 releaseFreelistLocks(); 1422 { 1423 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1424 if (_foregroundGCShouldWait) { 1425 // We are going to be waiting for action for the CMS thread; 1426 // it had better not be gone (for instance at shutdown)! 1427 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1428 "CMS thread must be running"); 1429 // Wait here until the background collector gives us the go-ahead 1430 ConcurrentMarkSweepThread::clear_CMS_flag( 1431 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1432 // Get a possibly blocked CMS thread going: 1433 // Note that we set _foregroundGCIsActive true above, 1434 // without protection of the CGC_lock. 1435 CGC_lock->notify(); 1436 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1437 "Possible deadlock"); 1438 while (_foregroundGCShouldWait) { 1439 // wait for notification 1440 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1441 // Possibility of delay/starvation here, since CMS token does 1442 // not know to give priority to VM thread? Actually, i think 1443 // there wouldn't be any delay/starvation, but the proof of 1444 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1445 } 1446 ConcurrentMarkSweepThread::set_CMS_flag( 1447 ConcurrentMarkSweepThread::CMS_vm_has_token); 1448 } 1449 } 1450 // The CMS_token is already held. Get back the other locks. 1451 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1452 "VM thread should have CMS token"); 1453 getFreelistLocks(); 1454 bitMapLock()->lock_without_safepoint_check(); 1455 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1456 p2i(Thread::current()), first_state); 1457 log_debug(gc, state)(" gets control with state %d", _collectorState); 1458 1459 // Inform cms gen if this was due to partial collection failing. 1460 // The CMS gen may use this fact to determine its expansion policy. 1461 CMSHeap* heap = CMSHeap::heap(); 1462 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1463 assert(!_cmsGen->incremental_collection_failed(), 1464 "Should have been noticed, reacted to and cleared"); 1465 _cmsGen->set_incremental_collection_failed(); 1466 } 1467 1468 if (first_state > Idling) { 1469 report_concurrent_mode_interruption(); 1470 } 1471 1472 set_did_compact(true); 1473 1474 // If the collection is being acquired from the background 1475 // collector, there may be references on the discovered 1476 // references lists. Abandon those references, since some 1477 // of them may have become unreachable after concurrent 1478 // discovery; the STW compacting collector will redo discovery 1479 // more precisely, without being subject to floating garbage. 1480 // Leaving otherwise unreachable references in the discovered 1481 // lists would require special handling. 1482 ref_processor()->disable_discovery(); 1483 ref_processor()->abandon_partial_discovery(); 1484 ref_processor()->verify_no_references_recorded(); 1485 1486 if (first_state > Idling) { 1487 save_heap_summary(); 1488 } 1489 1490 do_compaction_work(clear_all_soft_refs); 1491 1492 // Has the GC time limit been exceeded? 1493 size_t max_eden_size = _young_gen->max_eden_size(); 1494 GCCause::Cause gc_cause = heap->gc_cause(); 1495 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1496 _young_gen->eden()->used(), 1497 _cmsGen->max_capacity(), 1498 max_eden_size, 1499 full, 1500 gc_cause, 1501 heap->soft_ref_policy()); 1502 1503 // Reset the expansion cause, now that we just completed 1504 // a collection cycle. 1505 clear_expansion_cause(); 1506 _foregroundGCIsActive = false; 1507 return; 1508 } 1509 1510 // Resize the tenured generation 1511 // after obtaining the free list locks for the 1512 // two generations. 1513 void CMSCollector::compute_new_size() { 1514 assert_locked_or_safepoint(Heap_lock); 1515 FreelistLocker z(this); 1516 MetaspaceGC::compute_new_size(); 1517 _cmsGen->compute_new_size_free_list(); 1518 } 1519 1520 // A work method used by the foreground collector to do 1521 // a mark-sweep-compact. 1522 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1523 CMSHeap* heap = CMSHeap::heap(); 1524 1525 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1526 gc_timer->register_gc_start(); 1527 1528 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1529 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1530 1531 heap->pre_full_gc_dump(gc_timer); 1532 1533 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1534 1535 // Temporarily widen the span of the weak reference processing to 1536 // the entire heap. 1537 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1538 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1539 // Temporarily, clear the "is_alive_non_header" field of the 1540 // reference processor. 1541 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1542 // Temporarily make reference _processing_ single threaded (non-MT). 1543 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1544 // Temporarily make refs discovery atomic 1545 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1546 // Temporarily make reference _discovery_ single threaded (non-MT) 1547 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1548 1549 ref_processor()->set_enqueuing_is_done(false); 1550 ref_processor()->enable_discovery(); 1551 ref_processor()->setup_policy(clear_all_soft_refs); 1552 // If an asynchronous collection finishes, the _modUnionTable is 1553 // all clear. If we are assuming the collection from an asynchronous 1554 // collection, clear the _modUnionTable. 1555 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1556 "_modUnionTable should be clear if the baton was not passed"); 1557 _modUnionTable.clear_all(); 1558 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1559 "mod union for klasses should be clear if the baton was passed"); 1560 _ct->cld_rem_set()->clear_mod_union(); 1561 1562 1563 // We must adjust the allocation statistics being maintained 1564 // in the free list space. We do so by reading and clearing 1565 // the sweep timer and updating the block flux rate estimates below. 1566 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1567 if (_inter_sweep_timer.is_active()) { 1568 _inter_sweep_timer.stop(); 1569 // Note that we do not use this sample to update the _inter_sweep_estimate. 1570 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1571 _inter_sweep_estimate.padded_average(), 1572 _intra_sweep_estimate.padded_average()); 1573 } 1574 1575 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1576 #ifdef ASSERT 1577 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1578 size_t free_size = cms_space->free(); 1579 assert(free_size == 1580 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1581 * HeapWordSize, 1582 "All the free space should be compacted into one chunk at top"); 1583 assert(cms_space->dictionary()->total_chunk_size( 1584 debug_only(cms_space->freelistLock())) == 0 || 1585 cms_space->totalSizeInIndexedFreeLists() == 0, 1586 "All the free space should be in a single chunk"); 1587 size_t num = cms_space->totalCount(); 1588 assert((free_size == 0 && num == 0) || 1589 (free_size > 0 && (num == 1 || num == 2)), 1590 "There should be at most 2 free chunks after compaction"); 1591 #endif // ASSERT 1592 _collectorState = Resetting; 1593 assert(_restart_addr == NULL, 1594 "Should have been NULL'd before baton was passed"); 1595 reset_stw(); 1596 _cmsGen->reset_after_compaction(); 1597 _concurrent_cycles_since_last_unload = 0; 1598 1599 // Clear any data recorded in the PLAB chunk arrays. 1600 if (_survivor_plab_array != NULL) { 1601 reset_survivor_plab_arrays(); 1602 } 1603 1604 // Adjust the per-size allocation stats for the next epoch. 1605 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1606 // Restart the "inter sweep timer" for the next epoch. 1607 _inter_sweep_timer.reset(); 1608 _inter_sweep_timer.start(); 1609 1610 // No longer a need to do a concurrent collection for Metaspace. 1611 MetaspaceGC::set_should_concurrent_collect(false); 1612 1613 heap->post_full_gc_dump(gc_timer); 1614 1615 gc_timer->register_gc_end(); 1616 1617 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1618 1619 // For a mark-sweep-compact, compute_new_size() will be called 1620 // in the heap's do_collection() method. 1621 } 1622 1623 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1624 Log(gc, heap) log; 1625 if (!log.is_trace()) { 1626 return; 1627 } 1628 1629 ContiguousSpace* eden_space = _young_gen->eden(); 1630 ContiguousSpace* from_space = _young_gen->from(); 1631 ContiguousSpace* to_space = _young_gen->to(); 1632 // Eden 1633 if (_eden_chunk_array != NULL) { 1634 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1635 p2i(eden_space->bottom()), p2i(eden_space->top()), 1636 p2i(eden_space->end()), eden_space->capacity()); 1637 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1638 _eden_chunk_index, _eden_chunk_capacity); 1639 for (size_t i = 0; i < _eden_chunk_index; i++) { 1640 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1641 } 1642 } 1643 // Survivor 1644 if (_survivor_chunk_array != NULL) { 1645 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1646 p2i(from_space->bottom()), p2i(from_space->top()), 1647 p2i(from_space->end()), from_space->capacity()); 1648 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1649 _survivor_chunk_index, _survivor_chunk_capacity); 1650 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1651 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1652 } 1653 } 1654 } 1655 1656 void CMSCollector::getFreelistLocks() const { 1657 // Get locks for all free lists in all generations that this 1658 // collector is responsible for 1659 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1660 } 1661 1662 void CMSCollector::releaseFreelistLocks() const { 1663 // Release locks for all free lists in all generations that this 1664 // collector is responsible for 1665 _cmsGen->freelistLock()->unlock(); 1666 } 1667 1668 bool CMSCollector::haveFreelistLocks() const { 1669 // Check locks for all free lists in all generations that this 1670 // collector is responsible for 1671 assert_lock_strong(_cmsGen->freelistLock()); 1672 PRODUCT_ONLY(ShouldNotReachHere()); 1673 return true; 1674 } 1675 1676 // A utility class that is used by the CMS collector to 1677 // temporarily "release" the foreground collector from its 1678 // usual obligation to wait for the background collector to 1679 // complete an ongoing phase before proceeding. 1680 class ReleaseForegroundGC: public StackObj { 1681 private: 1682 CMSCollector* _c; 1683 public: 1684 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1685 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1686 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1687 // allow a potentially blocked foreground collector to proceed 1688 _c->_foregroundGCShouldWait = false; 1689 if (_c->_foregroundGCIsActive) { 1690 CGC_lock->notify(); 1691 } 1692 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1693 "Possible deadlock"); 1694 } 1695 1696 ~ReleaseForegroundGC() { 1697 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1698 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1699 _c->_foregroundGCShouldWait = true; 1700 } 1701 }; 1702 1703 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1704 assert(Thread::current()->is_ConcurrentGC_thread(), 1705 "A CMS asynchronous collection is only allowed on a CMS thread."); 1706 1707 CMSHeap* heap = CMSHeap::heap(); 1708 { 1709 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1710 MutexLockerEx hl(Heap_lock, safepoint_check); 1711 FreelistLocker fll(this); 1712 MutexLockerEx x(CGC_lock, safepoint_check); 1713 if (_foregroundGCIsActive) { 1714 // The foreground collector is. Skip this 1715 // background collection. 1716 assert(!_foregroundGCShouldWait, "Should be clear"); 1717 return; 1718 } else { 1719 assert(_collectorState == Idling, "Should be idling before start."); 1720 _collectorState = InitialMarking; 1721 register_gc_start(cause); 1722 // Reset the expansion cause, now that we are about to begin 1723 // a new cycle. 1724 clear_expansion_cause(); 1725 1726 // Clear the MetaspaceGC flag since a concurrent collection 1727 // is starting but also clear it after the collection. 1728 MetaspaceGC::set_should_concurrent_collect(false); 1729 } 1730 // Decide if we want to enable class unloading as part of the 1731 // ensuing concurrent GC cycle. 1732 update_should_unload_classes(); 1733 _full_gc_requested = false; // acks all outstanding full gc requests 1734 _full_gc_cause = GCCause::_no_gc; 1735 // Signal that we are about to start a collection 1736 heap->increment_total_full_collections(); // ... starting a collection cycle 1737 _collection_count_start = heap->total_full_collections(); 1738 } 1739 1740 size_t prev_used = _cmsGen->used(); 1741 1742 // The change of the collection state is normally done at this level; 1743 // the exceptions are phases that are executed while the world is 1744 // stopped. For those phases the change of state is done while the 1745 // world is stopped. For baton passing purposes this allows the 1746 // background collector to finish the phase and change state atomically. 1747 // The foreground collector cannot wait on a phase that is done 1748 // while the world is stopped because the foreground collector already 1749 // has the world stopped and would deadlock. 1750 while (_collectorState != Idling) { 1751 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1752 p2i(Thread::current()), _collectorState); 1753 // The foreground collector 1754 // holds the Heap_lock throughout its collection. 1755 // holds the CMS token (but not the lock) 1756 // except while it is waiting for the background collector to yield. 1757 // 1758 // The foreground collector should be blocked (not for long) 1759 // if the background collector is about to start a phase 1760 // executed with world stopped. If the background 1761 // collector has already started such a phase, the 1762 // foreground collector is blocked waiting for the 1763 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1764 // are executed in the VM thread. 1765 // 1766 // The locking order is 1767 // PendingListLock (PLL) -- if applicable (FinalMarking) 1768 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1769 // CMS token (claimed in 1770 // stop_world_and_do() --> 1771 // safepoint_synchronize() --> 1772 // CMSThread::synchronize()) 1773 1774 { 1775 // Check if the FG collector wants us to yield. 1776 CMSTokenSync x(true); // is cms thread 1777 if (waitForForegroundGC()) { 1778 // We yielded to a foreground GC, nothing more to be 1779 // done this round. 1780 assert(_foregroundGCShouldWait == false, "We set it to false in " 1781 "waitForForegroundGC()"); 1782 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1783 p2i(Thread::current()), _collectorState); 1784 return; 1785 } else { 1786 // The background collector can run but check to see if the 1787 // foreground collector has done a collection while the 1788 // background collector was waiting to get the CGC_lock 1789 // above. If yes, break so that _foregroundGCShouldWait 1790 // is cleared before returning. 1791 if (_collectorState == Idling) { 1792 break; 1793 } 1794 } 1795 } 1796 1797 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1798 "should be waiting"); 1799 1800 switch (_collectorState) { 1801 case InitialMarking: 1802 { 1803 ReleaseForegroundGC x(this); 1804 stats().record_cms_begin(); 1805 VM_CMS_Initial_Mark initial_mark_op(this); 1806 VMThread::execute(&initial_mark_op); 1807 } 1808 // The collector state may be any legal state at this point 1809 // since the background collector may have yielded to the 1810 // foreground collector. 1811 break; 1812 case Marking: 1813 // initial marking in checkpointRootsInitialWork has been completed 1814 if (markFromRoots()) { // we were successful 1815 assert(_collectorState == Precleaning, "Collector state should " 1816 "have changed"); 1817 } else { 1818 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1819 } 1820 break; 1821 case Precleaning: 1822 // marking from roots in markFromRoots has been completed 1823 preclean(); 1824 assert(_collectorState == AbortablePreclean || 1825 _collectorState == FinalMarking, 1826 "Collector state should have changed"); 1827 break; 1828 case AbortablePreclean: 1829 abortable_preclean(); 1830 assert(_collectorState == FinalMarking, "Collector state should " 1831 "have changed"); 1832 break; 1833 case FinalMarking: 1834 { 1835 ReleaseForegroundGC x(this); 1836 1837 VM_CMS_Final_Remark final_remark_op(this); 1838 VMThread::execute(&final_remark_op); 1839 } 1840 assert(_foregroundGCShouldWait, "block post-condition"); 1841 break; 1842 case Sweeping: 1843 // final marking in checkpointRootsFinal has been completed 1844 sweep(); 1845 assert(_collectorState == Resizing, "Collector state change " 1846 "to Resizing must be done under the free_list_lock"); 1847 1848 case Resizing: { 1849 // Sweeping has been completed... 1850 // At this point the background collection has completed. 1851 // Don't move the call to compute_new_size() down 1852 // into code that might be executed if the background 1853 // collection was preempted. 1854 { 1855 ReleaseForegroundGC x(this); // unblock FG collection 1856 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1857 CMSTokenSync z(true); // not strictly needed. 1858 if (_collectorState == Resizing) { 1859 compute_new_size(); 1860 save_heap_summary(); 1861 _collectorState = Resetting; 1862 } else { 1863 assert(_collectorState == Idling, "The state should only change" 1864 " because the foreground collector has finished the collection"); 1865 } 1866 } 1867 break; 1868 } 1869 case Resetting: 1870 // CMS heap resizing has been completed 1871 reset_concurrent(); 1872 assert(_collectorState == Idling, "Collector state should " 1873 "have changed"); 1874 1875 MetaspaceGC::set_should_concurrent_collect(false); 1876 1877 stats().record_cms_end(); 1878 // Don't move the concurrent_phases_end() and compute_new_size() 1879 // calls to here because a preempted background collection 1880 // has it's state set to "Resetting". 1881 break; 1882 case Idling: 1883 default: 1884 ShouldNotReachHere(); 1885 break; 1886 } 1887 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1888 p2i(Thread::current()), _collectorState); 1889 assert(_foregroundGCShouldWait, "block post-condition"); 1890 } 1891 1892 // Should this be in gc_epilogue? 1893 heap->counters()->update_counters(); 1894 1895 { 1896 // Clear _foregroundGCShouldWait and, in the event that the 1897 // foreground collector is waiting, notify it, before 1898 // returning. 1899 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1900 _foregroundGCShouldWait = false; 1901 if (_foregroundGCIsActive) { 1902 CGC_lock->notify(); 1903 } 1904 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1905 "Possible deadlock"); 1906 } 1907 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1908 p2i(Thread::current()), _collectorState); 1909 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1910 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1911 } 1912 1913 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1914 _cms_start_registered = true; 1915 _gc_timer_cm->register_gc_start(); 1916 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1917 } 1918 1919 void CMSCollector::register_gc_end() { 1920 if (_cms_start_registered) { 1921 report_heap_summary(GCWhen::AfterGC); 1922 1923 _gc_timer_cm->register_gc_end(); 1924 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1925 _cms_start_registered = false; 1926 } 1927 } 1928 1929 void CMSCollector::save_heap_summary() { 1930 CMSHeap* heap = CMSHeap::heap(); 1931 _last_heap_summary = heap->create_heap_summary(); 1932 _last_metaspace_summary = heap->create_metaspace_summary(); 1933 } 1934 1935 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1936 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1937 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1938 } 1939 1940 bool CMSCollector::waitForForegroundGC() { 1941 bool res = false; 1942 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1943 "CMS thread should have CMS token"); 1944 // Block the foreground collector until the 1945 // background collectors decides whether to 1946 // yield. 1947 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1948 _foregroundGCShouldWait = true; 1949 if (_foregroundGCIsActive) { 1950 // The background collector yields to the 1951 // foreground collector and returns a value 1952 // indicating that it has yielded. The foreground 1953 // collector can proceed. 1954 res = true; 1955 _foregroundGCShouldWait = false; 1956 ConcurrentMarkSweepThread::clear_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_has_token); 1958 ConcurrentMarkSweepThread::set_CMS_flag( 1959 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1960 // Get a possibly blocked foreground thread going 1961 CGC_lock->notify(); 1962 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1963 p2i(Thread::current()), _collectorState); 1964 while (_foregroundGCIsActive) { 1965 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1966 } 1967 ConcurrentMarkSweepThread::set_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_has_token); 1969 ConcurrentMarkSweepThread::clear_CMS_flag( 1970 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1971 } 1972 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1973 p2i(Thread::current()), _collectorState); 1974 return res; 1975 } 1976 1977 // Because of the need to lock the free lists and other structures in 1978 // the collector, common to all the generations that the collector is 1979 // collecting, we need the gc_prologues of individual CMS generations 1980 // delegate to their collector. It may have been simpler had the 1981 // current infrastructure allowed one to call a prologue on a 1982 // collector. In the absence of that we have the generation's 1983 // prologue delegate to the collector, which delegates back 1984 // some "local" work to a worker method in the individual generations 1985 // that it's responsible for collecting, while itself doing any 1986 // work common to all generations it's responsible for. A similar 1987 // comment applies to the gc_epilogue()'s. 1988 // The role of the variable _between_prologue_and_epilogue is to 1989 // enforce the invocation protocol. 1990 void CMSCollector::gc_prologue(bool full) { 1991 // Call gc_prologue_work() for the CMSGen 1992 // we are responsible for. 1993 1994 // The following locking discipline assumes that we are only called 1995 // when the world is stopped. 1996 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1997 1998 // The CMSCollector prologue must call the gc_prologues for the 1999 // "generations" that it's responsible 2000 // for. 2001 2002 assert( Thread::current()->is_VM_thread() 2003 || ( CMSScavengeBeforeRemark 2004 && Thread::current()->is_ConcurrentGC_thread()), 2005 "Incorrect thread type for prologue execution"); 2006 2007 if (_between_prologue_and_epilogue) { 2008 // We have already been invoked; this is a gc_prologue delegation 2009 // from yet another CMS generation that we are responsible for, just 2010 // ignore it since all relevant work has already been done. 2011 return; 2012 } 2013 2014 // set a bit saying prologue has been called; cleared in epilogue 2015 _between_prologue_and_epilogue = true; 2016 // Claim locks for common data structures, then call gc_prologue_work() 2017 // for each CMSGen. 2018 2019 getFreelistLocks(); // gets free list locks on constituent spaces 2020 bitMapLock()->lock_without_safepoint_check(); 2021 2022 // Should call gc_prologue_work() for all cms gens we are responsible for 2023 bool duringMarking = _collectorState >= Marking 2024 && _collectorState < Sweeping; 2025 2026 // The young collections clear the modified oops state, which tells if 2027 // there are any modified oops in the class. The remark phase also needs 2028 // that information. Tell the young collection to save the union of all 2029 // modified klasses. 2030 if (duringMarking) { 2031 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2032 } 2033 2034 bool registerClosure = duringMarking; 2035 2036 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2037 2038 if (!full) { 2039 stats().record_gc0_begin(); 2040 } 2041 } 2042 2043 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2044 2045 _capacity_at_prologue = capacity(); 2046 _used_at_prologue = used(); 2047 2048 // We enable promotion tracking so that card-scanning can recognize 2049 // which objects have been promoted during this GC and skip them. 2050 for (uint i = 0; i < ParallelGCThreads; i++) { 2051 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2052 } 2053 2054 // Delegate to CMScollector which knows how to coordinate between 2055 // this and any other CMS generations that it is responsible for 2056 // collecting. 2057 collector()->gc_prologue(full); 2058 } 2059 2060 // This is a "private" interface for use by this generation's CMSCollector. 2061 // Not to be called directly by any other entity (for instance, 2062 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2063 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2064 bool registerClosure, ModUnionClosure* modUnionClosure) { 2065 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2066 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2067 "Should be NULL"); 2068 if (registerClosure) { 2069 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2070 } 2071 cmsSpace()->gc_prologue(); 2072 // Clear stat counters 2073 NOT_PRODUCT( 2074 assert(_numObjectsPromoted == 0, "check"); 2075 assert(_numWordsPromoted == 0, "check"); 2076 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2077 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2078 _numObjectsAllocated = 0; 2079 _numWordsAllocated = 0; 2080 ) 2081 } 2082 2083 void CMSCollector::gc_epilogue(bool full) { 2084 // The following locking discipline assumes that we are only called 2085 // when the world is stopped. 2086 assert(SafepointSynchronize::is_at_safepoint(), 2087 "world is stopped assumption"); 2088 2089 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2090 // if linear allocation blocks need to be appropriately marked to allow the 2091 // the blocks to be parsable. We also check here whether we need to nudge the 2092 // CMS collector thread to start a new cycle (if it's not already active). 2093 assert( Thread::current()->is_VM_thread() 2094 || ( CMSScavengeBeforeRemark 2095 && Thread::current()->is_ConcurrentGC_thread()), 2096 "Incorrect thread type for epilogue execution"); 2097 2098 if (!_between_prologue_and_epilogue) { 2099 // We have already been invoked; this is a gc_epilogue delegation 2100 // from yet another CMS generation that we are responsible for, just 2101 // ignore it since all relevant work has already been done. 2102 return; 2103 } 2104 assert(haveFreelistLocks(), "must have freelist locks"); 2105 assert_lock_strong(bitMapLock()); 2106 2107 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2108 2109 _cmsGen->gc_epilogue_work(full); 2110 2111 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2112 // in case sampling was not already enabled, enable it 2113 _start_sampling = true; 2114 } 2115 // reset _eden_chunk_array so sampling starts afresh 2116 _eden_chunk_index = 0; 2117 2118 size_t cms_used = _cmsGen->cmsSpace()->used(); 2119 2120 // update performance counters - this uses a special version of 2121 // update_counters() that allows the utilization to be passed as a 2122 // parameter, avoiding multiple calls to used(). 2123 // 2124 _cmsGen->update_counters(cms_used); 2125 2126 bitMapLock()->unlock(); 2127 releaseFreelistLocks(); 2128 2129 if (!CleanChunkPoolAsync) { 2130 Chunk::clean_chunk_pool(); 2131 } 2132 2133 set_did_compact(false); 2134 _between_prologue_and_epilogue = false; // ready for next cycle 2135 } 2136 2137 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2138 collector()->gc_epilogue(full); 2139 2140 // When using ParNew, promotion tracking should have already been 2141 // disabled. However, the prologue (which enables promotion 2142 // tracking) and epilogue are called irrespective of the type of 2143 // GC. So they will also be called before and after Full GCs, during 2144 // which promotion tracking will not be explicitly disabled. So, 2145 // it's safer to also disable it here too (to be symmetric with 2146 // enabling it in the prologue). 2147 for (uint i = 0; i < ParallelGCThreads; i++) { 2148 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2149 } 2150 } 2151 2152 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2153 assert(!incremental_collection_failed(), "Should have been cleared"); 2154 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2155 cmsSpace()->gc_epilogue(); 2156 // Print stat counters 2157 NOT_PRODUCT( 2158 assert(_numObjectsAllocated == 0, "check"); 2159 assert(_numWordsAllocated == 0, "check"); 2160 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2161 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2162 _numObjectsPromoted = 0; 2163 _numWordsPromoted = 0; 2164 ) 2165 2166 // Call down the chain in contiguous_available needs the freelistLock 2167 // so print this out before releasing the freeListLock. 2168 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2169 } 2170 2171 #ifndef PRODUCT 2172 bool CMSCollector::have_cms_token() { 2173 Thread* thr = Thread::current(); 2174 if (thr->is_VM_thread()) { 2175 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2176 } else if (thr->is_ConcurrentGC_thread()) { 2177 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2178 } else if (thr->is_GC_task_thread()) { 2179 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2180 ParGCRareEvent_lock->owned_by_self(); 2181 } 2182 return false; 2183 } 2184 2185 // Check reachability of the given heap address in CMS generation, 2186 // treating all other generations as roots. 2187 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2188 // We could "guarantee" below, rather than assert, but I'll 2189 // leave these as "asserts" so that an adventurous debugger 2190 // could try this in the product build provided some subset of 2191 // the conditions were met, provided they were interested in the 2192 // results and knew that the computation below wouldn't interfere 2193 // with other concurrent computations mutating the structures 2194 // being read or written. 2195 assert(SafepointSynchronize::is_at_safepoint(), 2196 "Else mutations in object graph will make answer suspect"); 2197 assert(have_cms_token(), "Should hold cms token"); 2198 assert(haveFreelistLocks(), "must hold free list locks"); 2199 assert_lock_strong(bitMapLock()); 2200 2201 // Clear the marking bit map array before starting, but, just 2202 // for kicks, first report if the given address is already marked 2203 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2204 _markBitMap.isMarked(addr) ? "" : " not"); 2205 2206 if (verify_after_remark()) { 2207 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2208 bool result = verification_mark_bm()->isMarked(addr); 2209 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2210 result ? "IS" : "is NOT"); 2211 return result; 2212 } else { 2213 tty->print_cr("Could not compute result"); 2214 return false; 2215 } 2216 } 2217 #endif 2218 2219 void 2220 CMSCollector::print_on_error(outputStream* st) { 2221 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2222 if (collector != NULL) { 2223 CMSBitMap* bitmap = &collector->_markBitMap; 2224 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2225 bitmap->print_on_error(st, " Bits: "); 2226 2227 st->cr(); 2228 2229 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2230 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2231 mut_bitmap->print_on_error(st, " Bits: "); 2232 } 2233 } 2234 2235 //////////////////////////////////////////////////////// 2236 // CMS Verification Support 2237 //////////////////////////////////////////////////////// 2238 // Following the remark phase, the following invariant 2239 // should hold -- each object in the CMS heap which is 2240 // marked in markBitMap() should be marked in the verification_mark_bm(). 2241 2242 class VerifyMarkedClosure: public BitMapClosure { 2243 CMSBitMap* _marks; 2244 bool _failed; 2245 2246 public: 2247 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2248 2249 bool do_bit(size_t offset) { 2250 HeapWord* addr = _marks->offsetToHeapWord(offset); 2251 if (!_marks->isMarked(addr)) { 2252 Log(gc, verify) log; 2253 ResourceMark rm; 2254 LogStream ls(log.error()); 2255 oop(addr)->print_on(&ls); 2256 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2257 _failed = true; 2258 } 2259 return true; 2260 } 2261 2262 bool failed() { return _failed; } 2263 }; 2264 2265 bool CMSCollector::verify_after_remark() { 2266 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2267 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2268 static bool init = false; 2269 2270 assert(SafepointSynchronize::is_at_safepoint(), 2271 "Else mutations in object graph will make answer suspect"); 2272 assert(have_cms_token(), 2273 "Else there may be mutual interference in use of " 2274 " verification data structures"); 2275 assert(_collectorState > Marking && _collectorState <= Sweeping, 2276 "Else marking info checked here may be obsolete"); 2277 assert(haveFreelistLocks(), "must hold free list locks"); 2278 assert_lock_strong(bitMapLock()); 2279 2280 2281 // Allocate marking bit map if not already allocated 2282 if (!init) { // first time 2283 if (!verification_mark_bm()->allocate(_span)) { 2284 return false; 2285 } 2286 init = true; 2287 } 2288 2289 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2290 2291 // Turn off refs discovery -- so we will be tracing through refs. 2292 // This is as intended, because by this time 2293 // GC must already have cleared any refs that need to be cleared, 2294 // and traced those that need to be marked; moreover, 2295 // the marking done here is not going to interfere in any 2296 // way with the marking information used by GC. 2297 NoRefDiscovery no_discovery(ref_processor()); 2298 2299 #if COMPILER2_OR_JVMCI 2300 DerivedPointerTableDeactivate dpt_deact; 2301 #endif 2302 2303 // Clear any marks from a previous round 2304 verification_mark_bm()->clear_all(); 2305 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2306 verify_work_stacks_empty(); 2307 2308 CMSHeap* heap = CMSHeap::heap(); 2309 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2310 // Update the saved marks which may affect the root scans. 2311 heap->save_marks(); 2312 2313 if (CMSRemarkVerifyVariant == 1) { 2314 // In this first variant of verification, we complete 2315 // all marking, then check if the new marks-vector is 2316 // a subset of the CMS marks-vector. 2317 verify_after_remark_work_1(); 2318 } else { 2319 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2320 // In this second variant of verification, we flag an error 2321 // (i.e. an object reachable in the new marks-vector not reachable 2322 // in the CMS marks-vector) immediately, also indicating the 2323 // identify of an object (A) that references the unmarked object (B) -- 2324 // presumably, a mutation to A failed to be picked up by preclean/remark? 2325 verify_after_remark_work_2(); 2326 } 2327 2328 return true; 2329 } 2330 2331 void CMSCollector::verify_after_remark_work_1() { 2332 ResourceMark rm; 2333 HandleMark hm; 2334 CMSHeap* heap = CMSHeap::heap(); 2335 2336 // Get a clear set of claim bits for the roots processing to work with. 2337 ClassLoaderDataGraph::clear_claimed_marks(); 2338 2339 // Mark from roots one level into CMS 2340 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2341 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2342 2343 { 2344 StrongRootsScope srs(1); 2345 2346 heap->cms_process_roots(&srs, 2347 true, // young gen as roots 2348 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2349 should_unload_classes(), 2350 ¬Older, 2351 NULL); 2352 } 2353 2354 // Now mark from the roots 2355 MarkFromRootsClosure markFromRootsClosure(this, _span, 2356 verification_mark_bm(), verification_mark_stack(), 2357 false /* don't yield */, true /* verifying */); 2358 assert(_restart_addr == NULL, "Expected pre-condition"); 2359 verification_mark_bm()->iterate(&markFromRootsClosure); 2360 while (_restart_addr != NULL) { 2361 // Deal with stack overflow: by restarting at the indicated 2362 // address. 2363 HeapWord* ra = _restart_addr; 2364 markFromRootsClosure.reset(ra); 2365 _restart_addr = NULL; 2366 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2367 } 2368 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2369 verify_work_stacks_empty(); 2370 2371 // Marking completed -- now verify that each bit marked in 2372 // verification_mark_bm() is also marked in markBitMap(); flag all 2373 // errors by printing corresponding objects. 2374 VerifyMarkedClosure vcl(markBitMap()); 2375 verification_mark_bm()->iterate(&vcl); 2376 if (vcl.failed()) { 2377 Log(gc, verify) log; 2378 log.error("Failed marking verification after remark"); 2379 ResourceMark rm; 2380 LogStream ls(log.error()); 2381 heap->print_on(&ls); 2382 fatal("CMS: failed marking verification after remark"); 2383 } 2384 } 2385 2386 class VerifyCLDOopsCLDClosure : public CLDClosure { 2387 class VerifyCLDOopsClosure : public OopClosure { 2388 CMSBitMap* _bitmap; 2389 public: 2390 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2391 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2392 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2393 } _oop_closure; 2394 public: 2395 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2396 void do_cld(ClassLoaderData* cld) { 2397 cld->oops_do(&_oop_closure, false, false); 2398 } 2399 }; 2400 2401 void CMSCollector::verify_after_remark_work_2() { 2402 ResourceMark rm; 2403 HandleMark hm; 2404 CMSHeap* heap = CMSHeap::heap(); 2405 2406 // Get a clear set of claim bits for the roots processing to work with. 2407 ClassLoaderDataGraph::clear_claimed_marks(); 2408 2409 // Mark from roots one level into CMS 2410 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2411 markBitMap()); 2412 CLDToOopClosure cld_closure(¬Older, true); 2413 2414 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2415 2416 { 2417 StrongRootsScope srs(1); 2418 2419 heap->cms_process_roots(&srs, 2420 true, // young gen as roots 2421 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2422 should_unload_classes(), 2423 ¬Older, 2424 &cld_closure); 2425 } 2426 2427 // Now mark from the roots 2428 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2429 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2430 assert(_restart_addr == NULL, "Expected pre-condition"); 2431 verification_mark_bm()->iterate(&markFromRootsClosure); 2432 while (_restart_addr != NULL) { 2433 // Deal with stack overflow: by restarting at the indicated 2434 // address. 2435 HeapWord* ra = _restart_addr; 2436 markFromRootsClosure.reset(ra); 2437 _restart_addr = NULL; 2438 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2439 } 2440 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2441 verify_work_stacks_empty(); 2442 2443 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2444 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2445 2446 // Marking completed -- now verify that each bit marked in 2447 // verification_mark_bm() is also marked in markBitMap(); flag all 2448 // errors by printing corresponding objects. 2449 VerifyMarkedClosure vcl(markBitMap()); 2450 verification_mark_bm()->iterate(&vcl); 2451 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2452 } 2453 2454 void ConcurrentMarkSweepGeneration::save_marks() { 2455 // delegate to CMS space 2456 cmsSpace()->save_marks(); 2457 } 2458 2459 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2460 return cmsSpace()->no_allocs_since_save_marks(); 2461 } 2462 2463 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2464 \ 2465 void ConcurrentMarkSweepGeneration:: \ 2466 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2467 cl->set_generation(this); \ 2468 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2469 cl->reset_generation(); \ 2470 save_marks(); \ 2471 } 2472 2473 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2474 2475 void 2476 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2477 if (freelistLock()->owned_by_self()) { 2478 Generation::oop_iterate(cl); 2479 } else { 2480 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2481 Generation::oop_iterate(cl); 2482 } 2483 } 2484 2485 void 2486 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2487 if (freelistLock()->owned_by_self()) { 2488 Generation::object_iterate(cl); 2489 } else { 2490 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2491 Generation::object_iterate(cl); 2492 } 2493 } 2494 2495 void 2496 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2497 if (freelistLock()->owned_by_self()) { 2498 Generation::safe_object_iterate(cl); 2499 } else { 2500 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2501 Generation::safe_object_iterate(cl); 2502 } 2503 } 2504 2505 void 2506 ConcurrentMarkSweepGeneration::post_compact() { 2507 } 2508 2509 void 2510 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2511 // Fix the linear allocation blocks to look like free blocks. 2512 2513 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2514 // are not called when the heap is verified during universe initialization and 2515 // at vm shutdown. 2516 if (freelistLock()->owned_by_self()) { 2517 cmsSpace()->prepare_for_verify(); 2518 } else { 2519 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2520 cmsSpace()->prepare_for_verify(); 2521 } 2522 } 2523 2524 void 2525 ConcurrentMarkSweepGeneration::verify() { 2526 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2527 // are not called when the heap is verified during universe initialization and 2528 // at vm shutdown. 2529 if (freelistLock()->owned_by_self()) { 2530 cmsSpace()->verify(); 2531 } else { 2532 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2533 cmsSpace()->verify(); 2534 } 2535 } 2536 2537 void CMSCollector::verify() { 2538 _cmsGen->verify(); 2539 } 2540 2541 #ifndef PRODUCT 2542 bool CMSCollector::overflow_list_is_empty() const { 2543 assert(_num_par_pushes >= 0, "Inconsistency"); 2544 if (_overflow_list == NULL) { 2545 assert(_num_par_pushes == 0, "Inconsistency"); 2546 } 2547 return _overflow_list == NULL; 2548 } 2549 2550 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2551 // merely consolidate assertion checks that appear to occur together frequently. 2552 void CMSCollector::verify_work_stacks_empty() const { 2553 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2554 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2555 } 2556 2557 void CMSCollector::verify_overflow_empty() const { 2558 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2559 assert(no_preserved_marks(), "No preserved marks"); 2560 } 2561 #endif // PRODUCT 2562 2563 // Decide if we want to enable class unloading as part of the 2564 // ensuing concurrent GC cycle. We will collect and 2565 // unload classes if it's the case that: 2566 // (a) class unloading is enabled at the command line, and 2567 // (b) old gen is getting really full 2568 // NOTE: Provided there is no change in the state of the heap between 2569 // calls to this method, it should have idempotent results. Moreover, 2570 // its results should be monotonically increasing (i.e. going from 0 to 1, 2571 // but not 1 to 0) between successive calls between which the heap was 2572 // not collected. For the implementation below, it must thus rely on 2573 // the property that concurrent_cycles_since_last_unload() 2574 // will not decrease unless a collection cycle happened and that 2575 // _cmsGen->is_too_full() are 2576 // themselves also monotonic in that sense. See check_monotonicity() 2577 // below. 2578 void CMSCollector::update_should_unload_classes() { 2579 _should_unload_classes = false; 2580 if (CMSClassUnloadingEnabled) { 2581 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2582 CMSClassUnloadingMaxInterval) 2583 || _cmsGen->is_too_full(); 2584 } 2585 } 2586 2587 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2588 bool res = should_concurrent_collect(); 2589 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2590 return res; 2591 } 2592 2593 void CMSCollector::setup_cms_unloading_and_verification_state() { 2594 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2595 || VerifyBeforeExit; 2596 const int rso = GenCollectedHeap::SO_AllCodeCache; 2597 2598 // We set the proper root for this CMS cycle here. 2599 if (should_unload_classes()) { // Should unload classes this cycle 2600 remove_root_scanning_option(rso); // Shrink the root set appropriately 2601 set_verifying(should_verify); // Set verification state for this cycle 2602 return; // Nothing else needs to be done at this time 2603 } 2604 2605 // Not unloading classes this cycle 2606 assert(!should_unload_classes(), "Inconsistency!"); 2607 2608 // If we are not unloading classes then add SO_AllCodeCache to root 2609 // scanning options. 2610 add_root_scanning_option(rso); 2611 2612 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2613 set_verifying(true); 2614 } else if (verifying() && !should_verify) { 2615 // We were verifying, but some verification flags got disabled. 2616 set_verifying(false); 2617 // Exclude symbols, strings and code cache elements from root scanning to 2618 // reduce IM and RM pauses. 2619 remove_root_scanning_option(rso); 2620 } 2621 } 2622 2623 2624 #ifndef PRODUCT 2625 HeapWord* CMSCollector::block_start(const void* p) const { 2626 const HeapWord* addr = (HeapWord*)p; 2627 if (_span.contains(p)) { 2628 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2629 return _cmsGen->cmsSpace()->block_start(p); 2630 } 2631 } 2632 return NULL; 2633 } 2634 #endif 2635 2636 HeapWord* 2637 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2638 bool tlab, 2639 bool parallel) { 2640 CMSSynchronousYieldRequest yr; 2641 assert(!tlab, "Can't deal with TLAB allocation"); 2642 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2643 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2644 if (GCExpandToAllocateDelayMillis > 0) { 2645 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2646 } 2647 return have_lock_and_allocate(word_size, tlab); 2648 } 2649 2650 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2651 size_t bytes, 2652 size_t expand_bytes, 2653 CMSExpansionCause::Cause cause) 2654 { 2655 2656 bool success = expand(bytes, expand_bytes); 2657 2658 // remember why we expanded; this information is used 2659 // by shouldConcurrentCollect() when making decisions on whether to start 2660 // a new CMS cycle. 2661 if (success) { 2662 set_expansion_cause(cause); 2663 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2664 } 2665 } 2666 2667 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2668 HeapWord* res = NULL; 2669 MutexLocker x(ParGCRareEvent_lock); 2670 while (true) { 2671 // Expansion by some other thread might make alloc OK now: 2672 res = ps->lab.alloc(word_sz); 2673 if (res != NULL) return res; 2674 // If there's not enough expansion space available, give up. 2675 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2676 return NULL; 2677 } 2678 // Otherwise, we try expansion. 2679 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2680 // Now go around the loop and try alloc again; 2681 // A competing par_promote might beat us to the expansion space, 2682 // so we may go around the loop again if promotion fails again. 2683 if (GCExpandToAllocateDelayMillis > 0) { 2684 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2685 } 2686 } 2687 } 2688 2689 2690 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2691 PromotionInfo* promo) { 2692 MutexLocker x(ParGCRareEvent_lock); 2693 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2694 while (true) { 2695 // Expansion by some other thread might make alloc OK now: 2696 if (promo->ensure_spooling_space()) { 2697 assert(promo->has_spooling_space(), 2698 "Post-condition of successful ensure_spooling_space()"); 2699 return true; 2700 } 2701 // If there's not enough expansion space available, give up. 2702 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2703 return false; 2704 } 2705 // Otherwise, we try expansion. 2706 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2707 // Now go around the loop and try alloc again; 2708 // A competing allocation might beat us to the expansion space, 2709 // so we may go around the loop again if allocation fails again. 2710 if (GCExpandToAllocateDelayMillis > 0) { 2711 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2712 } 2713 } 2714 } 2715 2716 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2717 // Only shrink if a compaction was done so that all the free space 2718 // in the generation is in a contiguous block at the end. 2719 if (did_compact()) { 2720 CardGeneration::shrink(bytes); 2721 } 2722 } 2723 2724 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2725 assert_locked_or_safepoint(Heap_lock); 2726 } 2727 2728 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2729 assert_locked_or_safepoint(Heap_lock); 2730 assert_lock_strong(freelistLock()); 2731 log_trace(gc)("Shrinking of CMS not yet implemented"); 2732 return; 2733 } 2734 2735 2736 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2737 // phases. 2738 class CMSPhaseAccounting: public StackObj { 2739 public: 2740 CMSPhaseAccounting(CMSCollector *collector, 2741 const char *title); 2742 ~CMSPhaseAccounting(); 2743 2744 private: 2745 CMSCollector *_collector; 2746 const char *_title; 2747 GCTraceConcTime(Info, gc) _trace_time; 2748 2749 public: 2750 // Not MT-safe; so do not pass around these StackObj's 2751 // where they may be accessed by other threads. 2752 double wallclock_millis() { 2753 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2754 } 2755 }; 2756 2757 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2758 const char *title) : 2759 _collector(collector), _title(title), _trace_time(title) { 2760 2761 _collector->resetYields(); 2762 _collector->resetTimer(); 2763 _collector->startTimer(); 2764 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2765 } 2766 2767 CMSPhaseAccounting::~CMSPhaseAccounting() { 2768 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2769 _collector->stopTimer(); 2770 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2771 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2772 } 2773 2774 // CMS work 2775 2776 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2777 class CMSParMarkTask : public AbstractGangTask { 2778 protected: 2779 CMSCollector* _collector; 2780 uint _n_workers; 2781 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2782 AbstractGangTask(name), 2783 _collector(collector), 2784 _n_workers(n_workers) {} 2785 // Work method in support of parallel rescan ... of young gen spaces 2786 void do_young_space_rescan(OopsInGenClosure* cl, 2787 ContiguousSpace* space, 2788 HeapWord** chunk_array, size_t chunk_top); 2789 void work_on_young_gen_roots(OopsInGenClosure* cl); 2790 }; 2791 2792 // Parallel initial mark task 2793 class CMSParInitialMarkTask: public CMSParMarkTask { 2794 StrongRootsScope* _strong_roots_scope; 2795 public: 2796 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2797 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2798 _strong_roots_scope(strong_roots_scope) {} 2799 void work(uint worker_id); 2800 }; 2801 2802 // Checkpoint the roots into this generation from outside 2803 // this generation. [Note this initial checkpoint need only 2804 // be approximate -- we'll do a catch up phase subsequently.] 2805 void CMSCollector::checkpointRootsInitial() { 2806 assert(_collectorState == InitialMarking, "Wrong collector state"); 2807 check_correct_thread_executing(); 2808 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2809 2810 save_heap_summary(); 2811 report_heap_summary(GCWhen::BeforeGC); 2812 2813 ReferenceProcessor* rp = ref_processor(); 2814 assert(_restart_addr == NULL, "Control point invariant"); 2815 { 2816 // acquire locks for subsequent manipulations 2817 MutexLockerEx x(bitMapLock(), 2818 Mutex::_no_safepoint_check_flag); 2819 checkpointRootsInitialWork(); 2820 // enable ("weak") refs discovery 2821 rp->enable_discovery(); 2822 _collectorState = Marking; 2823 } 2824 } 2825 2826 void CMSCollector::checkpointRootsInitialWork() { 2827 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2828 assert(_collectorState == InitialMarking, "just checking"); 2829 2830 // Already have locks. 2831 assert_lock_strong(bitMapLock()); 2832 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2833 2834 // Setup the verification and class unloading state for this 2835 // CMS collection cycle. 2836 setup_cms_unloading_and_verification_state(); 2837 2838 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2839 2840 // Reset all the PLAB chunk arrays if necessary. 2841 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2842 reset_survivor_plab_arrays(); 2843 } 2844 2845 ResourceMark rm; 2846 HandleMark hm; 2847 2848 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2849 CMSHeap* heap = CMSHeap::heap(); 2850 2851 verify_work_stacks_empty(); 2852 verify_overflow_empty(); 2853 2854 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2855 // Update the saved marks which may affect the root scans. 2856 heap->save_marks(); 2857 2858 // weak reference processing has not started yet. 2859 ref_processor()->set_enqueuing_is_done(false); 2860 2861 // Need to remember all newly created CLDs, 2862 // so that we can guarantee that the remark finds them. 2863 ClassLoaderDataGraph::remember_new_clds(true); 2864 2865 // Whenever a CLD is found, it will be claimed before proceeding to mark 2866 // the klasses. The claimed marks need to be cleared before marking starts. 2867 ClassLoaderDataGraph::clear_claimed_marks(); 2868 2869 print_eden_and_survivor_chunk_arrays(); 2870 2871 { 2872 #if COMPILER2_OR_JVMCI 2873 DerivedPointerTableDeactivate dpt_deact; 2874 #endif 2875 if (CMSParallelInitialMarkEnabled) { 2876 // The parallel version. 2877 WorkGang* workers = heap->workers(); 2878 assert(workers != NULL, "Need parallel worker threads."); 2879 uint n_workers = workers->active_workers(); 2880 2881 StrongRootsScope srs(n_workers); 2882 2883 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2884 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2885 // If the total workers is greater than 1, then multiple workers 2886 // may be used at some time and the initialization has been set 2887 // such that the single threaded path cannot be used. 2888 if (workers->total_workers() > 1) { 2889 workers->run_task(&tsk); 2890 } else { 2891 tsk.work(0); 2892 } 2893 } else { 2894 // The serial version. 2895 CLDToOopClosure cld_closure(¬Older, true); 2896 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2897 2898 StrongRootsScope srs(1); 2899 2900 heap->cms_process_roots(&srs, 2901 true, // young gen as roots 2902 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2903 should_unload_classes(), 2904 ¬Older, 2905 &cld_closure); 2906 } 2907 } 2908 2909 // Clear mod-union table; it will be dirtied in the prologue of 2910 // CMS generation per each young generation collection. 2911 2912 assert(_modUnionTable.isAllClear(), 2913 "Was cleared in most recent final checkpoint phase" 2914 " or no bits are set in the gc_prologue before the start of the next " 2915 "subsequent marking phase."); 2916 2917 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2918 2919 // Save the end of the used_region of the constituent generations 2920 // to be used to limit the extent of sweep in each generation. 2921 save_sweep_limits(); 2922 verify_overflow_empty(); 2923 } 2924 2925 bool CMSCollector::markFromRoots() { 2926 // we might be tempted to assert that: 2927 // assert(!SafepointSynchronize::is_at_safepoint(), 2928 // "inconsistent argument?"); 2929 // However that wouldn't be right, because it's possible that 2930 // a safepoint is indeed in progress as a young generation 2931 // stop-the-world GC happens even as we mark in this generation. 2932 assert(_collectorState == Marking, "inconsistent state?"); 2933 check_correct_thread_executing(); 2934 verify_overflow_empty(); 2935 2936 // Weak ref discovery note: We may be discovering weak 2937 // refs in this generation concurrent (but interleaved) with 2938 // weak ref discovery by the young generation collector. 2939 2940 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2941 GCTraceCPUTime tcpu; 2942 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2943 bool res = markFromRootsWork(); 2944 if (res) { 2945 _collectorState = Precleaning; 2946 } else { // We failed and a foreground collection wants to take over 2947 assert(_foregroundGCIsActive, "internal state inconsistency"); 2948 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2949 log_debug(gc)("bailing out to foreground collection"); 2950 } 2951 verify_overflow_empty(); 2952 return res; 2953 } 2954 2955 bool CMSCollector::markFromRootsWork() { 2956 // iterate over marked bits in bit map, doing a full scan and mark 2957 // from these roots using the following algorithm: 2958 // . if oop is to the right of the current scan pointer, 2959 // mark corresponding bit (we'll process it later) 2960 // . else (oop is to left of current scan pointer) 2961 // push oop on marking stack 2962 // . drain the marking stack 2963 2964 // Note that when we do a marking step we need to hold the 2965 // bit map lock -- recall that direct allocation (by mutators) 2966 // and promotion (by the young generation collector) is also 2967 // marking the bit map. [the so-called allocate live policy.] 2968 // Because the implementation of bit map marking is not 2969 // robust wrt simultaneous marking of bits in the same word, 2970 // we need to make sure that there is no such interference 2971 // between concurrent such updates. 2972 2973 // already have locks 2974 assert_lock_strong(bitMapLock()); 2975 2976 verify_work_stacks_empty(); 2977 verify_overflow_empty(); 2978 bool result = false; 2979 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2980 result = do_marking_mt(); 2981 } else { 2982 result = do_marking_st(); 2983 } 2984 return result; 2985 } 2986 2987 // Forward decl 2988 class CMSConcMarkingTask; 2989 2990 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2991 CMSCollector* _collector; 2992 CMSConcMarkingTask* _task; 2993 public: 2994 virtual void yield(); 2995 2996 // "n_threads" is the number of threads to be terminated. 2997 // "queue_set" is a set of work queues of other threads. 2998 // "collector" is the CMS collector associated with this task terminator. 2999 // "yield" indicates whether we need the gang as a whole to yield. 3000 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3001 ParallelTaskTerminator(n_threads, queue_set), 3002 _collector(collector) { } 3003 3004 void set_task(CMSConcMarkingTask* task) { 3005 _task = task; 3006 } 3007 }; 3008 3009 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3010 CMSConcMarkingTask* _task; 3011 public: 3012 bool should_exit_termination(); 3013 void set_task(CMSConcMarkingTask* task) { 3014 _task = task; 3015 } 3016 }; 3017 3018 // MT Concurrent Marking Task 3019 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3020 CMSCollector* _collector; 3021 uint _n_workers; // requested/desired # workers 3022 bool _result; 3023 CompactibleFreeListSpace* _cms_space; 3024 char _pad_front[64]; // padding to ... 3025 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3026 char _pad_back[64]; 3027 HeapWord* _restart_addr; 3028 3029 // Exposed here for yielding support 3030 Mutex* const _bit_map_lock; 3031 3032 // The per thread work queues, available here for stealing 3033 OopTaskQueueSet* _task_queues; 3034 3035 // Termination (and yielding) support 3036 CMSConcMarkingTerminator _term; 3037 CMSConcMarkingTerminatorTerminator _term_term; 3038 3039 public: 3040 CMSConcMarkingTask(CMSCollector* collector, 3041 CompactibleFreeListSpace* cms_space, 3042 YieldingFlexibleWorkGang* workers, 3043 OopTaskQueueSet* task_queues): 3044 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3045 _collector(collector), 3046 _cms_space(cms_space), 3047 _n_workers(0), _result(true), 3048 _task_queues(task_queues), 3049 _term(_n_workers, task_queues, _collector), 3050 _bit_map_lock(collector->bitMapLock()) 3051 { 3052 _requested_size = _n_workers; 3053 _term.set_task(this); 3054 _term_term.set_task(this); 3055 _restart_addr = _global_finger = _cms_space->bottom(); 3056 } 3057 3058 3059 OopTaskQueueSet* task_queues() { return _task_queues; } 3060 3061 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3062 3063 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3064 3065 CMSConcMarkingTerminator* terminator() { return &_term; } 3066 3067 virtual void set_for_termination(uint active_workers) { 3068 terminator()->reset_for_reuse(active_workers); 3069 } 3070 3071 void work(uint worker_id); 3072 bool should_yield() { 3073 return ConcurrentMarkSweepThread::should_yield() 3074 && !_collector->foregroundGCIsActive(); 3075 } 3076 3077 virtual void coordinator_yield(); // stuff done by coordinator 3078 bool result() { return _result; } 3079 3080 void reset(HeapWord* ra) { 3081 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3082 _restart_addr = _global_finger = ra; 3083 _term.reset_for_reuse(); 3084 } 3085 3086 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3087 OopTaskQueue* work_q); 3088 3089 private: 3090 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3091 void do_work_steal(int i); 3092 void bump_global_finger(HeapWord* f); 3093 }; 3094 3095 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3096 assert(_task != NULL, "Error"); 3097 return _task->yielding(); 3098 // Note that we do not need the disjunct || _task->should_yield() above 3099 // because we want terminating threads to yield only if the task 3100 // is already in the midst of yielding, which happens only after at least one 3101 // thread has yielded. 3102 } 3103 3104 void CMSConcMarkingTerminator::yield() { 3105 if (_task->should_yield()) { 3106 _task->yield(); 3107 } else { 3108 ParallelTaskTerminator::yield(); 3109 } 3110 } 3111 3112 //////////////////////////////////////////////////////////////// 3113 // Concurrent Marking Algorithm Sketch 3114 //////////////////////////////////////////////////////////////// 3115 // Until all tasks exhausted (both spaces): 3116 // -- claim next available chunk 3117 // -- bump global finger via CAS 3118 // -- find first object that starts in this chunk 3119 // and start scanning bitmap from that position 3120 // -- scan marked objects for oops 3121 // -- CAS-mark target, and if successful: 3122 // . if target oop is above global finger (volatile read) 3123 // nothing to do 3124 // . if target oop is in chunk and above local finger 3125 // then nothing to do 3126 // . else push on work-queue 3127 // -- Deal with possible overflow issues: 3128 // . local work-queue overflow causes stuff to be pushed on 3129 // global (common) overflow queue 3130 // . always first empty local work queue 3131 // . then get a batch of oops from global work queue if any 3132 // . then do work stealing 3133 // -- When all tasks claimed (both spaces) 3134 // and local work queue empty, 3135 // then in a loop do: 3136 // . check global overflow stack; steal a batch of oops and trace 3137 // . try to steal from other threads oif GOS is empty 3138 // . if neither is available, offer termination 3139 // -- Terminate and return result 3140 // 3141 void CMSConcMarkingTask::work(uint worker_id) { 3142 elapsedTimer _timer; 3143 ResourceMark rm; 3144 HandleMark hm; 3145 3146 DEBUG_ONLY(_collector->verify_overflow_empty();) 3147 3148 // Before we begin work, our work queue should be empty 3149 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3150 // Scan the bitmap covering _cms_space, tracing through grey objects. 3151 _timer.start(); 3152 do_scan_and_mark(worker_id, _cms_space); 3153 _timer.stop(); 3154 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3155 3156 // ... do work stealing 3157 _timer.reset(); 3158 _timer.start(); 3159 do_work_steal(worker_id); 3160 _timer.stop(); 3161 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3162 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3163 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3164 // Note that under the current task protocol, the 3165 // following assertion is true even of the spaces 3166 // expanded since the completion of the concurrent 3167 // marking. XXX This will likely change under a strict 3168 // ABORT semantics. 3169 // After perm removal the comparison was changed to 3170 // greater than or equal to from strictly greater than. 3171 // Before perm removal the highest address sweep would 3172 // have been at the end of perm gen but now is at the 3173 // end of the tenured gen. 3174 assert(_global_finger >= _cms_space->end(), 3175 "All tasks have been completed"); 3176 DEBUG_ONLY(_collector->verify_overflow_empty();) 3177 } 3178 3179 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3180 HeapWord* read = _global_finger; 3181 HeapWord* cur = read; 3182 while (f > read) { 3183 cur = read; 3184 read = Atomic::cmpxchg(f, &_global_finger, cur); 3185 if (cur == read) { 3186 // our cas succeeded 3187 assert(_global_finger >= f, "protocol consistency"); 3188 break; 3189 } 3190 } 3191 } 3192 3193 // This is really inefficient, and should be redone by 3194 // using (not yet available) block-read and -write interfaces to the 3195 // stack and the work_queue. XXX FIX ME !!! 3196 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3197 OopTaskQueue* work_q) { 3198 // Fast lock-free check 3199 if (ovflw_stk->length() == 0) { 3200 return false; 3201 } 3202 assert(work_q->size() == 0, "Shouldn't steal"); 3203 MutexLockerEx ml(ovflw_stk->par_lock(), 3204 Mutex::_no_safepoint_check_flag); 3205 // Grab up to 1/4 the size of the work queue 3206 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3207 (size_t)ParGCDesiredObjsFromOverflowList); 3208 num = MIN2(num, ovflw_stk->length()); 3209 for (int i = (int) num; i > 0; i--) { 3210 oop cur = ovflw_stk->pop(); 3211 assert(cur != NULL, "Counted wrong?"); 3212 work_q->push(cur); 3213 } 3214 return num > 0; 3215 } 3216 3217 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3218 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3219 int n_tasks = pst->n_tasks(); 3220 // We allow that there may be no tasks to do here because 3221 // we are restarting after a stack overflow. 3222 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3223 uint nth_task = 0; 3224 3225 HeapWord* aligned_start = sp->bottom(); 3226 if (sp->used_region().contains(_restart_addr)) { 3227 // Align down to a card boundary for the start of 0th task 3228 // for this space. 3229 aligned_start = align_down(_restart_addr, CardTable::card_size); 3230 } 3231 3232 size_t chunk_size = sp->marking_task_size(); 3233 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3234 // Having claimed the nth task in this space, 3235 // compute the chunk that it corresponds to: 3236 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3237 aligned_start + (nth_task+1)*chunk_size); 3238 // Try and bump the global finger via a CAS; 3239 // note that we need to do the global finger bump 3240 // _before_ taking the intersection below, because 3241 // the task corresponding to that region will be 3242 // deemed done even if the used_region() expands 3243 // because of allocation -- as it almost certainly will 3244 // during start-up while the threads yield in the 3245 // closure below. 3246 HeapWord* finger = span.end(); 3247 bump_global_finger(finger); // atomically 3248 // There are null tasks here corresponding to chunks 3249 // beyond the "top" address of the space. 3250 span = span.intersection(sp->used_region()); 3251 if (!span.is_empty()) { // Non-null task 3252 HeapWord* prev_obj; 3253 assert(!span.contains(_restart_addr) || nth_task == 0, 3254 "Inconsistency"); 3255 if (nth_task == 0) { 3256 // For the 0th task, we'll not need to compute a block_start. 3257 if (span.contains(_restart_addr)) { 3258 // In the case of a restart because of stack overflow, 3259 // we might additionally skip a chunk prefix. 3260 prev_obj = _restart_addr; 3261 } else { 3262 prev_obj = span.start(); 3263 } 3264 } else { 3265 // We want to skip the first object because 3266 // the protocol is to scan any object in its entirety 3267 // that _starts_ in this span; a fortiori, any 3268 // object starting in an earlier span is scanned 3269 // as part of an earlier claimed task. 3270 // Below we use the "careful" version of block_start 3271 // so we do not try to navigate uninitialized objects. 3272 prev_obj = sp->block_start_careful(span.start()); 3273 // Below we use a variant of block_size that uses the 3274 // Printezis bits to avoid waiting for allocated 3275 // objects to become initialized/parsable. 3276 while (prev_obj < span.start()) { 3277 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3278 if (sz > 0) { 3279 prev_obj += sz; 3280 } else { 3281 // In this case we may end up doing a bit of redundant 3282 // scanning, but that appears unavoidable, short of 3283 // locking the free list locks; see bug 6324141. 3284 break; 3285 } 3286 } 3287 } 3288 if (prev_obj < span.end()) { 3289 MemRegion my_span = MemRegion(prev_obj, span.end()); 3290 // Do the marking work within a non-empty span -- 3291 // the last argument to the constructor indicates whether the 3292 // iteration should be incremental with periodic yields. 3293 ParMarkFromRootsClosure cl(this, _collector, my_span, 3294 &_collector->_markBitMap, 3295 work_queue(i), 3296 &_collector->_markStack); 3297 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3298 } // else nothing to do for this task 3299 } // else nothing to do for this task 3300 } 3301 // We'd be tempted to assert here that since there are no 3302 // more tasks left to claim in this space, the global_finger 3303 // must exceed space->top() and a fortiori space->end(). However, 3304 // that would not quite be correct because the bumping of 3305 // global_finger occurs strictly after the claiming of a task, 3306 // so by the time we reach here the global finger may not yet 3307 // have been bumped up by the thread that claimed the last 3308 // task. 3309 pst->all_tasks_completed(); 3310 } 3311 3312 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3313 private: 3314 CMSCollector* _collector; 3315 CMSConcMarkingTask* _task; 3316 MemRegion _span; 3317 CMSBitMap* _bit_map; 3318 CMSMarkStack* _overflow_stack; 3319 OopTaskQueue* _work_queue; 3320 protected: 3321 DO_OOP_WORK_DEFN 3322 public: 3323 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3324 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3325 MetadataAwareOopClosure(collector->ref_processor()), 3326 _collector(collector), 3327 _task(task), 3328 _span(collector->_span), 3329 _work_queue(work_queue), 3330 _bit_map(bit_map), 3331 _overflow_stack(overflow_stack) 3332 { } 3333 virtual void do_oop(oop* p); 3334 virtual void do_oop(narrowOop* p); 3335 3336 void trim_queue(size_t max); 3337 void handle_stack_overflow(HeapWord* lost); 3338 void do_yield_check() { 3339 if (_task->should_yield()) { 3340 _task->yield(); 3341 } 3342 } 3343 }; 3344 3345 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3346 3347 // Grey object scanning during work stealing phase -- 3348 // the salient assumption here is that any references 3349 // that are in these stolen objects being scanned must 3350 // already have been initialized (else they would not have 3351 // been published), so we do not need to check for 3352 // uninitialized objects before pushing here. 3353 void ParConcMarkingClosure::do_oop(oop obj) { 3354 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3355 HeapWord* addr = (HeapWord*)obj; 3356 // Check if oop points into the CMS generation 3357 // and is not marked 3358 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3359 // a white object ... 3360 // If we manage to "claim" the object, by being the 3361 // first thread to mark it, then we push it on our 3362 // marking stack 3363 if (_bit_map->par_mark(addr)) { // ... now grey 3364 // push on work queue (grey set) 3365 bool simulate_overflow = false; 3366 NOT_PRODUCT( 3367 if (CMSMarkStackOverflowALot && 3368 _collector->simulate_overflow()) { 3369 // simulate a stack overflow 3370 simulate_overflow = true; 3371 } 3372 ) 3373 if (simulate_overflow || 3374 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3375 // stack overflow 3376 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3377 // We cannot assert that the overflow stack is full because 3378 // it may have been emptied since. 3379 assert(simulate_overflow || 3380 _work_queue->size() == _work_queue->max_elems(), 3381 "Else push should have succeeded"); 3382 handle_stack_overflow(addr); 3383 } 3384 } // Else, some other thread got there first 3385 do_yield_check(); 3386 } 3387 } 3388 3389 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3390 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3391 3392 void ParConcMarkingClosure::trim_queue(size_t max) { 3393 while (_work_queue->size() > max) { 3394 oop new_oop; 3395 if (_work_queue->pop_local(new_oop)) { 3396 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3397 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3398 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3399 new_oop->oop_iterate(this); // do_oop() above 3400 do_yield_check(); 3401 } 3402 } 3403 } 3404 3405 // Upon stack overflow, we discard (part of) the stack, 3406 // remembering the least address amongst those discarded 3407 // in CMSCollector's _restart_address. 3408 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3409 // We need to do this under a mutex to prevent other 3410 // workers from interfering with the work done below. 3411 MutexLockerEx ml(_overflow_stack->par_lock(), 3412 Mutex::_no_safepoint_check_flag); 3413 // Remember the least grey address discarded 3414 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3415 _collector->lower_restart_addr(ra); 3416 _overflow_stack->reset(); // discard stack contents 3417 _overflow_stack->expand(); // expand the stack if possible 3418 } 3419 3420 3421 void CMSConcMarkingTask::do_work_steal(int i) { 3422 OopTaskQueue* work_q = work_queue(i); 3423 oop obj_to_scan; 3424 CMSBitMap* bm = &(_collector->_markBitMap); 3425 CMSMarkStack* ovflw = &(_collector->_markStack); 3426 int* seed = _collector->hash_seed(i); 3427 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3428 while (true) { 3429 cl.trim_queue(0); 3430 assert(work_q->size() == 0, "Should have been emptied above"); 3431 if (get_work_from_overflow_stack(ovflw, work_q)) { 3432 // Can't assert below because the work obtained from the 3433 // overflow stack may already have been stolen from us. 3434 // assert(work_q->size() > 0, "Work from overflow stack"); 3435 continue; 3436 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3437 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3438 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3439 obj_to_scan->oop_iterate(&cl); 3440 } else if (terminator()->offer_termination(&_term_term)) { 3441 assert(work_q->size() == 0, "Impossible!"); 3442 break; 3443 } else if (yielding() || should_yield()) { 3444 yield(); 3445 } 3446 } 3447 } 3448 3449 // This is run by the CMS (coordinator) thread. 3450 void CMSConcMarkingTask::coordinator_yield() { 3451 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3452 "CMS thread should hold CMS token"); 3453 // First give up the locks, then yield, then re-lock 3454 // We should probably use a constructor/destructor idiom to 3455 // do this unlock/lock or modify the MutexUnlocker class to 3456 // serve our purpose. XXX 3457 assert_lock_strong(_bit_map_lock); 3458 _bit_map_lock->unlock(); 3459 ConcurrentMarkSweepThread::desynchronize(true); 3460 _collector->stopTimer(); 3461 _collector->incrementYields(); 3462 3463 // It is possible for whichever thread initiated the yield request 3464 // not to get a chance to wake up and take the bitmap lock between 3465 // this thread releasing it and reacquiring it. So, while the 3466 // should_yield() flag is on, let's sleep for a bit to give the 3467 // other thread a chance to wake up. The limit imposed on the number 3468 // of iterations is defensive, to avoid any unforseen circumstances 3469 // putting us into an infinite loop. Since it's always been this 3470 // (coordinator_yield()) method that was observed to cause the 3471 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3472 // which is by default non-zero. For the other seven methods that 3473 // also perform the yield operation, as are using a different 3474 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3475 // can enable the sleeping for those methods too, if necessary. 3476 // See 6442774. 3477 // 3478 // We really need to reconsider the synchronization between the GC 3479 // thread and the yield-requesting threads in the future and we 3480 // should really use wait/notify, which is the recommended 3481 // way of doing this type of interaction. Additionally, we should 3482 // consolidate the eight methods that do the yield operation and they 3483 // are almost identical into one for better maintainability and 3484 // readability. See 6445193. 3485 // 3486 // Tony 2006.06.29 3487 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3488 ConcurrentMarkSweepThread::should_yield() && 3489 !CMSCollector::foregroundGCIsActive(); ++i) { 3490 os::sleep(Thread::current(), 1, false); 3491 } 3492 3493 ConcurrentMarkSweepThread::synchronize(true); 3494 _bit_map_lock->lock_without_safepoint_check(); 3495 _collector->startTimer(); 3496 } 3497 3498 bool CMSCollector::do_marking_mt() { 3499 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3500 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3501 conc_workers()->active_workers(), 3502 Threads::number_of_non_daemon_threads()); 3503 num_workers = conc_workers()->update_active_workers(num_workers); 3504 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3505 3506 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3507 3508 CMSConcMarkingTask tsk(this, 3509 cms_space, 3510 conc_workers(), 3511 task_queues()); 3512 3513 // Since the actual number of workers we get may be different 3514 // from the number we requested above, do we need to do anything different 3515 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3516 // class?? XXX 3517 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3518 3519 // Refs discovery is already non-atomic. 3520 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3521 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3522 conc_workers()->start_task(&tsk); 3523 while (tsk.yielded()) { 3524 tsk.coordinator_yield(); 3525 conc_workers()->continue_task(&tsk); 3526 } 3527 // If the task was aborted, _restart_addr will be non-NULL 3528 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3529 while (_restart_addr != NULL) { 3530 // XXX For now we do not make use of ABORTED state and have not 3531 // yet implemented the right abort semantics (even in the original 3532 // single-threaded CMS case). That needs some more investigation 3533 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3534 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3535 // If _restart_addr is non-NULL, a marking stack overflow 3536 // occurred; we need to do a fresh marking iteration from the 3537 // indicated restart address. 3538 if (_foregroundGCIsActive) { 3539 // We may be running into repeated stack overflows, having 3540 // reached the limit of the stack size, while making very 3541 // slow forward progress. It may be best to bail out and 3542 // let the foreground collector do its job. 3543 // Clear _restart_addr, so that foreground GC 3544 // works from scratch. This avoids the headache of 3545 // a "rescan" which would otherwise be needed because 3546 // of the dirty mod union table & card table. 3547 _restart_addr = NULL; 3548 return false; 3549 } 3550 // Adjust the task to restart from _restart_addr 3551 tsk.reset(_restart_addr); 3552 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3553 _restart_addr); 3554 _restart_addr = NULL; 3555 // Get the workers going again 3556 conc_workers()->start_task(&tsk); 3557 while (tsk.yielded()) { 3558 tsk.coordinator_yield(); 3559 conc_workers()->continue_task(&tsk); 3560 } 3561 } 3562 assert(tsk.completed(), "Inconsistency"); 3563 assert(tsk.result() == true, "Inconsistency"); 3564 return true; 3565 } 3566 3567 bool CMSCollector::do_marking_st() { 3568 ResourceMark rm; 3569 HandleMark hm; 3570 3571 // Temporarily make refs discovery single threaded (non-MT) 3572 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3573 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3574 &_markStack, CMSYield); 3575 // the last argument to iterate indicates whether the iteration 3576 // should be incremental with periodic yields. 3577 _markBitMap.iterate(&markFromRootsClosure); 3578 // If _restart_addr is non-NULL, a marking stack overflow 3579 // occurred; we need to do a fresh iteration from the 3580 // indicated restart address. 3581 while (_restart_addr != NULL) { 3582 if (_foregroundGCIsActive) { 3583 // We may be running into repeated stack overflows, having 3584 // reached the limit of the stack size, while making very 3585 // slow forward progress. It may be best to bail out and 3586 // let the foreground collector do its job. 3587 // Clear _restart_addr, so that foreground GC 3588 // works from scratch. This avoids the headache of 3589 // a "rescan" which would otherwise be needed because 3590 // of the dirty mod union table & card table. 3591 _restart_addr = NULL; 3592 return false; // indicating failure to complete marking 3593 } 3594 // Deal with stack overflow: 3595 // we restart marking from _restart_addr 3596 HeapWord* ra = _restart_addr; 3597 markFromRootsClosure.reset(ra); 3598 _restart_addr = NULL; 3599 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3600 } 3601 return true; 3602 } 3603 3604 void CMSCollector::preclean() { 3605 check_correct_thread_executing(); 3606 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3607 verify_work_stacks_empty(); 3608 verify_overflow_empty(); 3609 _abort_preclean = false; 3610 if (CMSPrecleaningEnabled) { 3611 if (!CMSEdenChunksRecordAlways) { 3612 _eden_chunk_index = 0; 3613 } 3614 size_t used = get_eden_used(); 3615 size_t capacity = get_eden_capacity(); 3616 // Don't start sampling unless we will get sufficiently 3617 // many samples. 3618 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3619 * CMSScheduleRemarkEdenPenetration)) { 3620 _start_sampling = true; 3621 } else { 3622 _start_sampling = false; 3623 } 3624 GCTraceCPUTime tcpu; 3625 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3626 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3627 } 3628 CMSTokenSync x(true); // is cms thread 3629 if (CMSPrecleaningEnabled) { 3630 sample_eden(); 3631 _collectorState = AbortablePreclean; 3632 } else { 3633 _collectorState = FinalMarking; 3634 } 3635 verify_work_stacks_empty(); 3636 verify_overflow_empty(); 3637 } 3638 3639 // Try and schedule the remark such that young gen 3640 // occupancy is CMSScheduleRemarkEdenPenetration %. 3641 void CMSCollector::abortable_preclean() { 3642 check_correct_thread_executing(); 3643 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3644 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3645 3646 // If Eden's current occupancy is below this threshold, 3647 // immediately schedule the remark; else preclean 3648 // past the next scavenge in an effort to 3649 // schedule the pause as described above. By choosing 3650 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3651 // we will never do an actual abortable preclean cycle. 3652 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3653 GCTraceCPUTime tcpu; 3654 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3655 // We need more smarts in the abortable preclean 3656 // loop below to deal with cases where allocation 3657 // in young gen is very very slow, and our precleaning 3658 // is running a losing race against a horde of 3659 // mutators intent on flooding us with CMS updates 3660 // (dirty cards). 3661 // One, admittedly dumb, strategy is to give up 3662 // after a certain number of abortable precleaning loops 3663 // or after a certain maximum time. We want to make 3664 // this smarter in the next iteration. 3665 // XXX FIX ME!!! YSR 3666 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3667 while (!(should_abort_preclean() || 3668 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3669 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3670 cumworkdone += workdone; 3671 loops++; 3672 // Voluntarily terminate abortable preclean phase if we have 3673 // been at it for too long. 3674 if ((CMSMaxAbortablePrecleanLoops != 0) && 3675 loops >= CMSMaxAbortablePrecleanLoops) { 3676 log_debug(gc)(" CMS: abort preclean due to loops "); 3677 break; 3678 } 3679 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3680 log_debug(gc)(" CMS: abort preclean due to time "); 3681 break; 3682 } 3683 // If we are doing little work each iteration, we should 3684 // take a short break. 3685 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3686 // Sleep for some time, waiting for work to accumulate 3687 stopTimer(); 3688 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3689 startTimer(); 3690 waited++; 3691 } 3692 } 3693 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3694 loops, waited, cumworkdone); 3695 } 3696 CMSTokenSync x(true); // is cms thread 3697 if (_collectorState != Idling) { 3698 assert(_collectorState == AbortablePreclean, 3699 "Spontaneous state transition?"); 3700 _collectorState = FinalMarking; 3701 } // Else, a foreground collection completed this CMS cycle. 3702 return; 3703 } 3704 3705 // Respond to an Eden sampling opportunity 3706 void CMSCollector::sample_eden() { 3707 // Make sure a young gc cannot sneak in between our 3708 // reading and recording of a sample. 3709 assert(Thread::current()->is_ConcurrentGC_thread(), 3710 "Only the cms thread may collect Eden samples"); 3711 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3712 "Should collect samples while holding CMS token"); 3713 if (!_start_sampling) { 3714 return; 3715 } 3716 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3717 // is populated by the young generation. 3718 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3719 if (_eden_chunk_index < _eden_chunk_capacity) { 3720 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3721 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3722 "Unexpected state of Eden"); 3723 // We'd like to check that what we just sampled is an oop-start address; 3724 // however, we cannot do that here since the object may not yet have been 3725 // initialized. So we'll instead do the check when we _use_ this sample 3726 // later. 3727 if (_eden_chunk_index == 0 || 3728 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3729 _eden_chunk_array[_eden_chunk_index-1]) 3730 >= CMSSamplingGrain)) { 3731 _eden_chunk_index++; // commit sample 3732 } 3733 } 3734 } 3735 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3736 size_t used = get_eden_used(); 3737 size_t capacity = get_eden_capacity(); 3738 assert(used <= capacity, "Unexpected state of Eden"); 3739 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3740 _abort_preclean = true; 3741 } 3742 } 3743 } 3744 3745 3746 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3747 assert(_collectorState == Precleaning || 3748 _collectorState == AbortablePreclean, "incorrect state"); 3749 ResourceMark rm; 3750 HandleMark hm; 3751 3752 // Precleaning is currently not MT but the reference processor 3753 // may be set for MT. Disable it temporarily here. 3754 ReferenceProcessor* rp = ref_processor(); 3755 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3756 3757 // Do one pass of scrubbing the discovered reference lists 3758 // to remove any reference objects with strongly-reachable 3759 // referents. 3760 if (clean_refs) { 3761 CMSPrecleanRefsYieldClosure yield_cl(this); 3762 assert(rp->span().equals(_span), "Spans should be equal"); 3763 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3764 &_markStack, true /* preclean */); 3765 CMSDrainMarkingStackClosure complete_trace(this, 3766 _span, &_markBitMap, &_markStack, 3767 &keep_alive, true /* preclean */); 3768 3769 // We don't want this step to interfere with a young 3770 // collection because we don't want to take CPU 3771 // or memory bandwidth away from the young GC threads 3772 // (which may be as many as there are CPUs). 3773 // Note that we don't need to protect ourselves from 3774 // interference with mutators because they can't 3775 // manipulate the discovered reference lists nor affect 3776 // the computed reachability of the referents, the 3777 // only properties manipulated by the precleaning 3778 // of these reference lists. 3779 stopTimer(); 3780 CMSTokenSyncWithLocks x(true /* is cms thread */, 3781 bitMapLock()); 3782 startTimer(); 3783 sample_eden(); 3784 3785 // The following will yield to allow foreground 3786 // collection to proceed promptly. XXX YSR: 3787 // The code in this method may need further 3788 // tweaking for better performance and some restructuring 3789 // for cleaner interfaces. 3790 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3791 rp->preclean_discovered_references( 3792 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3793 gc_timer); 3794 } 3795 3796 if (clean_survivor) { // preclean the active survivor space(s) 3797 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3798 &_markBitMap, &_modUnionTable, 3799 &_markStack, true /* precleaning phase */); 3800 stopTimer(); 3801 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3802 bitMapLock()); 3803 startTimer(); 3804 unsigned int before_count = 3805 CMSHeap::heap()->total_collections(); 3806 SurvivorSpacePrecleanClosure 3807 sss_cl(this, _span, &_markBitMap, &_markStack, 3808 &pam_cl, before_count, CMSYield); 3809 _young_gen->from()->object_iterate_careful(&sss_cl); 3810 _young_gen->to()->object_iterate_careful(&sss_cl); 3811 } 3812 MarkRefsIntoAndScanClosure 3813 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3814 &_markStack, this, CMSYield, 3815 true /* precleaning phase */); 3816 // CAUTION: The following closure has persistent state that may need to 3817 // be reset upon a decrease in the sequence of addresses it 3818 // processes. 3819 ScanMarkedObjectsAgainCarefullyClosure 3820 smoac_cl(this, _span, 3821 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3822 3823 // Preclean dirty cards in ModUnionTable and CardTable using 3824 // appropriate convergence criterion; 3825 // repeat CMSPrecleanIter times unless we find that 3826 // we are losing. 3827 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3828 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3829 "Bad convergence multiplier"); 3830 assert(CMSPrecleanThreshold >= 100, 3831 "Unreasonably low CMSPrecleanThreshold"); 3832 3833 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3834 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3835 numIter < CMSPrecleanIter; 3836 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3837 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3838 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3839 // Either there are very few dirty cards, so re-mark 3840 // pause will be small anyway, or our pre-cleaning isn't 3841 // that much faster than the rate at which cards are being 3842 // dirtied, so we might as well stop and re-mark since 3843 // precleaning won't improve our re-mark time by much. 3844 if (curNumCards <= CMSPrecleanThreshold || 3845 (numIter > 0 && 3846 (curNumCards * CMSPrecleanDenominator > 3847 lastNumCards * CMSPrecleanNumerator))) { 3848 numIter++; 3849 cumNumCards += curNumCards; 3850 break; 3851 } 3852 } 3853 3854 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3855 3856 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3857 cumNumCards += curNumCards; 3858 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3859 curNumCards, cumNumCards, numIter); 3860 return cumNumCards; // as a measure of useful work done 3861 } 3862 3863 // PRECLEANING NOTES: 3864 // Precleaning involves: 3865 // . reading the bits of the modUnionTable and clearing the set bits. 3866 // . For the cards corresponding to the set bits, we scan the 3867 // objects on those cards. This means we need the free_list_lock 3868 // so that we can safely iterate over the CMS space when scanning 3869 // for oops. 3870 // . When we scan the objects, we'll be both reading and setting 3871 // marks in the marking bit map, so we'll need the marking bit map. 3872 // . For protecting _collector_state transitions, we take the CGC_lock. 3873 // Note that any races in the reading of of card table entries by the 3874 // CMS thread on the one hand and the clearing of those entries by the 3875 // VM thread or the setting of those entries by the mutator threads on the 3876 // other are quite benign. However, for efficiency it makes sense to keep 3877 // the VM thread from racing with the CMS thread while the latter is 3878 // dirty card info to the modUnionTable. We therefore also use the 3879 // CGC_lock to protect the reading of the card table and the mod union 3880 // table by the CM thread. 3881 // . We run concurrently with mutator updates, so scanning 3882 // needs to be done carefully -- we should not try to scan 3883 // potentially uninitialized objects. 3884 // 3885 // Locking strategy: While holding the CGC_lock, we scan over and 3886 // reset a maximal dirty range of the mod union / card tables, then lock 3887 // the free_list_lock and bitmap lock to do a full marking, then 3888 // release these locks; and repeat the cycle. This allows for a 3889 // certain amount of fairness in the sharing of these locks between 3890 // the CMS collector on the one hand, and the VM thread and the 3891 // mutators on the other. 3892 3893 // NOTE: preclean_mod_union_table() and preclean_card_table() 3894 // further below are largely identical; if you need to modify 3895 // one of these methods, please check the other method too. 3896 3897 size_t CMSCollector::preclean_mod_union_table( 3898 ConcurrentMarkSweepGeneration* old_gen, 3899 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3900 verify_work_stacks_empty(); 3901 verify_overflow_empty(); 3902 3903 // strategy: starting with the first card, accumulate contiguous 3904 // ranges of dirty cards; clear these cards, then scan the region 3905 // covered by these cards. 3906 3907 // Since all of the MUT is committed ahead, we can just use 3908 // that, in case the generations expand while we are precleaning. 3909 // It might also be fine to just use the committed part of the 3910 // generation, but we might potentially miss cards when the 3911 // generation is rapidly expanding while we are in the midst 3912 // of precleaning. 3913 HeapWord* startAddr = old_gen->reserved().start(); 3914 HeapWord* endAddr = old_gen->reserved().end(); 3915 3916 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3917 3918 size_t numDirtyCards, cumNumDirtyCards; 3919 HeapWord *nextAddr, *lastAddr; 3920 for (cumNumDirtyCards = numDirtyCards = 0, 3921 nextAddr = lastAddr = startAddr; 3922 nextAddr < endAddr; 3923 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3924 3925 ResourceMark rm; 3926 HandleMark hm; 3927 3928 MemRegion dirtyRegion; 3929 { 3930 stopTimer(); 3931 // Potential yield point 3932 CMSTokenSync ts(true); 3933 startTimer(); 3934 sample_eden(); 3935 // Get dirty region starting at nextOffset (inclusive), 3936 // simultaneously clearing it. 3937 dirtyRegion = 3938 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3939 assert(dirtyRegion.start() >= nextAddr, 3940 "returned region inconsistent?"); 3941 } 3942 // Remember where the next search should begin. 3943 // The returned region (if non-empty) is a right open interval, 3944 // so lastOffset is obtained from the right end of that 3945 // interval. 3946 lastAddr = dirtyRegion.end(); 3947 // Should do something more transparent and less hacky XXX 3948 numDirtyCards = 3949 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3950 3951 // We'll scan the cards in the dirty region (with periodic 3952 // yields for foreground GC as needed). 3953 if (!dirtyRegion.is_empty()) { 3954 assert(numDirtyCards > 0, "consistency check"); 3955 HeapWord* stop_point = NULL; 3956 stopTimer(); 3957 // Potential yield point 3958 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3959 bitMapLock()); 3960 startTimer(); 3961 { 3962 verify_work_stacks_empty(); 3963 verify_overflow_empty(); 3964 sample_eden(); 3965 stop_point = 3966 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3967 } 3968 if (stop_point != NULL) { 3969 // The careful iteration stopped early either because it found an 3970 // uninitialized object, or because we were in the midst of an 3971 // "abortable preclean", which should now be aborted. Redirty 3972 // the bits corresponding to the partially-scanned or unscanned 3973 // cards. We'll either restart at the next block boundary or 3974 // abort the preclean. 3975 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3976 "Should only be AbortablePreclean."); 3977 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3978 if (should_abort_preclean()) { 3979 break; // out of preclean loop 3980 } else { 3981 // Compute the next address at which preclean should pick up; 3982 // might need bitMapLock in order to read P-bits. 3983 lastAddr = next_card_start_after_block(stop_point); 3984 } 3985 } 3986 } else { 3987 assert(lastAddr == endAddr, "consistency check"); 3988 assert(numDirtyCards == 0, "consistency check"); 3989 break; 3990 } 3991 } 3992 verify_work_stacks_empty(); 3993 verify_overflow_empty(); 3994 return cumNumDirtyCards; 3995 } 3996 3997 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3998 // below are largely identical; if you need to modify 3999 // one of these methods, please check the other method too. 4000 4001 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4002 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4003 // strategy: it's similar to precleamModUnionTable above, in that 4004 // we accumulate contiguous ranges of dirty cards, mark these cards 4005 // precleaned, then scan the region covered by these cards. 4006 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4007 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4008 4009 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4010 4011 size_t numDirtyCards, cumNumDirtyCards; 4012 HeapWord *lastAddr, *nextAddr; 4013 4014 for (cumNumDirtyCards = numDirtyCards = 0, 4015 nextAddr = lastAddr = startAddr; 4016 nextAddr < endAddr; 4017 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4018 4019 ResourceMark rm; 4020 HandleMark hm; 4021 4022 MemRegion dirtyRegion; 4023 { 4024 // See comments in "Precleaning notes" above on why we 4025 // do this locking. XXX Could the locking overheads be 4026 // too high when dirty cards are sparse? [I don't think so.] 4027 stopTimer(); 4028 CMSTokenSync x(true); // is cms thread 4029 startTimer(); 4030 sample_eden(); 4031 // Get and clear dirty region from card table 4032 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4033 true, 4034 CardTable::precleaned_card_val()); 4035 4036 assert(dirtyRegion.start() >= nextAddr, 4037 "returned region inconsistent?"); 4038 } 4039 lastAddr = dirtyRegion.end(); 4040 numDirtyCards = 4041 dirtyRegion.word_size()/CardTable::card_size_in_words; 4042 4043 if (!dirtyRegion.is_empty()) { 4044 stopTimer(); 4045 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4046 startTimer(); 4047 sample_eden(); 4048 verify_work_stacks_empty(); 4049 verify_overflow_empty(); 4050 HeapWord* stop_point = 4051 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4052 if (stop_point != NULL) { 4053 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4054 "Should only be AbortablePreclean."); 4055 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4056 if (should_abort_preclean()) { 4057 break; // out of preclean loop 4058 } else { 4059 // Compute the next address at which preclean should pick up. 4060 lastAddr = next_card_start_after_block(stop_point); 4061 } 4062 } 4063 } else { 4064 break; 4065 } 4066 } 4067 verify_work_stacks_empty(); 4068 verify_overflow_empty(); 4069 return cumNumDirtyCards; 4070 } 4071 4072 class PrecleanCLDClosure : public CLDClosure { 4073 MetadataAwareOopsInGenClosure* _cm_closure; 4074 public: 4075 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4076 void do_cld(ClassLoaderData* cld) { 4077 if (cld->has_accumulated_modified_oops()) { 4078 cld->clear_accumulated_modified_oops(); 4079 4080 _cm_closure->do_cld(cld); 4081 } 4082 } 4083 }; 4084 4085 // The freelist lock is needed to prevent asserts, is it really needed? 4086 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4087 4088 cl->set_freelistLock(freelistLock); 4089 4090 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4091 4092 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4093 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4094 PrecleanCLDClosure preclean_closure(cl); 4095 ClassLoaderDataGraph::cld_do(&preclean_closure); 4096 4097 verify_work_stacks_empty(); 4098 verify_overflow_empty(); 4099 } 4100 4101 void CMSCollector::checkpointRootsFinal() { 4102 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4103 check_correct_thread_executing(); 4104 // world is stopped at this checkpoint 4105 assert(SafepointSynchronize::is_at_safepoint(), 4106 "world should be stopped"); 4107 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4108 4109 verify_work_stacks_empty(); 4110 verify_overflow_empty(); 4111 4112 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4113 _young_gen->used() / K, _young_gen->capacity() / K); 4114 { 4115 if (CMSScavengeBeforeRemark) { 4116 CMSHeap* heap = CMSHeap::heap(); 4117 // Temporarily set flag to false, GCH->do_collection will 4118 // expect it to be false and set to true 4119 FlagSetting fl(heap->_is_gc_active, false); 4120 4121 heap->do_collection(true, // full (i.e. force, see below) 4122 false, // !clear_all_soft_refs 4123 0, // size 4124 false, // is_tlab 4125 GenCollectedHeap::YoungGen // type 4126 ); 4127 } 4128 FreelistLocker x(this); 4129 MutexLockerEx y(bitMapLock(), 4130 Mutex::_no_safepoint_check_flag); 4131 checkpointRootsFinalWork(); 4132 } 4133 verify_work_stacks_empty(); 4134 verify_overflow_empty(); 4135 } 4136 4137 void CMSCollector::checkpointRootsFinalWork() { 4138 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4139 4140 assert(haveFreelistLocks(), "must have free list locks"); 4141 assert_lock_strong(bitMapLock()); 4142 4143 ResourceMark rm; 4144 HandleMark hm; 4145 4146 CMSHeap* heap = CMSHeap::heap(); 4147 4148 if (should_unload_classes()) { 4149 CodeCache::gc_prologue(); 4150 } 4151 assert(haveFreelistLocks(), "must have free list locks"); 4152 assert_lock_strong(bitMapLock()); 4153 4154 // We might assume that we need not fill TLAB's when 4155 // CMSScavengeBeforeRemark is set, because we may have just done 4156 // a scavenge which would have filled all TLAB's -- and besides 4157 // Eden would be empty. This however may not always be the case -- 4158 // for instance although we asked for a scavenge, it may not have 4159 // happened because of a JNI critical section. We probably need 4160 // a policy for deciding whether we can in that case wait until 4161 // the critical section releases and then do the remark following 4162 // the scavenge, and skip it here. In the absence of that policy, 4163 // or of an indication of whether the scavenge did indeed occur, 4164 // we cannot rely on TLAB's having been filled and must do 4165 // so here just in case a scavenge did not happen. 4166 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4167 // Update the saved marks which may affect the root scans. 4168 heap->save_marks(); 4169 4170 print_eden_and_survivor_chunk_arrays(); 4171 4172 { 4173 #if COMPILER2_OR_JVMCI 4174 DerivedPointerTableDeactivate dpt_deact; 4175 #endif 4176 4177 // Note on the role of the mod union table: 4178 // Since the marker in "markFromRoots" marks concurrently with 4179 // mutators, it is possible for some reachable objects not to have been 4180 // scanned. For instance, an only reference to an object A was 4181 // placed in object B after the marker scanned B. Unless B is rescanned, 4182 // A would be collected. Such updates to references in marked objects 4183 // are detected via the mod union table which is the set of all cards 4184 // dirtied since the first checkpoint in this GC cycle and prior to 4185 // the most recent young generation GC, minus those cleaned up by the 4186 // concurrent precleaning. 4187 if (CMSParallelRemarkEnabled) { 4188 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4189 do_remark_parallel(); 4190 } else { 4191 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4192 do_remark_non_parallel(); 4193 } 4194 } 4195 verify_work_stacks_empty(); 4196 verify_overflow_empty(); 4197 4198 { 4199 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4200 refProcessingWork(); 4201 } 4202 verify_work_stacks_empty(); 4203 verify_overflow_empty(); 4204 4205 if (should_unload_classes()) { 4206 CodeCache::gc_epilogue(); 4207 } 4208 JvmtiExport::gc_epilogue(); 4209 4210 // If we encountered any (marking stack / work queue) overflow 4211 // events during the current CMS cycle, take appropriate 4212 // remedial measures, where possible, so as to try and avoid 4213 // recurrence of that condition. 4214 assert(_markStack.isEmpty(), "No grey objects"); 4215 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4216 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4217 if (ser_ovflw > 0) { 4218 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4219 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4220 _markStack.expand(); 4221 _ser_pmc_remark_ovflw = 0; 4222 _ser_pmc_preclean_ovflw = 0; 4223 _ser_kac_preclean_ovflw = 0; 4224 _ser_kac_ovflw = 0; 4225 } 4226 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4227 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4228 _par_pmc_remark_ovflw, _par_kac_ovflw); 4229 _par_pmc_remark_ovflw = 0; 4230 _par_kac_ovflw = 0; 4231 } 4232 if (_markStack._hit_limit > 0) { 4233 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4234 _markStack._hit_limit); 4235 } 4236 if (_markStack._failed_double > 0) { 4237 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4238 _markStack._failed_double, _markStack.capacity()); 4239 } 4240 _markStack._hit_limit = 0; 4241 _markStack._failed_double = 0; 4242 4243 if ((VerifyAfterGC || VerifyDuringGC) && 4244 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4245 verify_after_remark(); 4246 } 4247 4248 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4249 4250 // Change under the freelistLocks. 4251 _collectorState = Sweeping; 4252 // Call isAllClear() under bitMapLock 4253 assert(_modUnionTable.isAllClear(), 4254 "Should be clear by end of the final marking"); 4255 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4256 "Should be clear by end of the final marking"); 4257 } 4258 4259 void CMSParInitialMarkTask::work(uint worker_id) { 4260 elapsedTimer _timer; 4261 ResourceMark rm; 4262 HandleMark hm; 4263 4264 // ---------- scan from roots -------------- 4265 _timer.start(); 4266 CMSHeap* heap = CMSHeap::heap(); 4267 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4268 4269 // ---------- young gen roots -------------- 4270 { 4271 work_on_young_gen_roots(&par_mri_cl); 4272 _timer.stop(); 4273 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4274 } 4275 4276 // ---------- remaining roots -------------- 4277 _timer.reset(); 4278 _timer.start(); 4279 4280 CLDToOopClosure cld_closure(&par_mri_cl, true); 4281 4282 heap->cms_process_roots(_strong_roots_scope, 4283 false, // yg was scanned above 4284 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4285 _collector->should_unload_classes(), 4286 &par_mri_cl, 4287 &cld_closure); 4288 assert(_collector->should_unload_classes() 4289 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4290 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4291 _timer.stop(); 4292 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4293 } 4294 4295 // Parallel remark task 4296 class CMSParRemarkTask: public CMSParMarkTask { 4297 CompactibleFreeListSpace* _cms_space; 4298 4299 // The per-thread work queues, available here for stealing. 4300 OopTaskQueueSet* _task_queues; 4301 ParallelTaskTerminator _term; 4302 StrongRootsScope* _strong_roots_scope; 4303 4304 public: 4305 // A value of 0 passed to n_workers will cause the number of 4306 // workers to be taken from the active workers in the work gang. 4307 CMSParRemarkTask(CMSCollector* collector, 4308 CompactibleFreeListSpace* cms_space, 4309 uint n_workers, WorkGang* workers, 4310 OopTaskQueueSet* task_queues, 4311 StrongRootsScope* strong_roots_scope): 4312 CMSParMarkTask("Rescan roots and grey objects in parallel", 4313 collector, n_workers), 4314 _cms_space(cms_space), 4315 _task_queues(task_queues), 4316 _term(n_workers, task_queues), 4317 _strong_roots_scope(strong_roots_scope) { } 4318 4319 OopTaskQueueSet* task_queues() { return _task_queues; } 4320 4321 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4322 4323 ParallelTaskTerminator* terminator() { return &_term; } 4324 uint n_workers() { return _n_workers; } 4325 4326 void work(uint worker_id); 4327 4328 private: 4329 // ... of dirty cards in old space 4330 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4331 ParMarkRefsIntoAndScanClosure* cl); 4332 4333 // ... work stealing for the above 4334 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4335 }; 4336 4337 class RemarkCLDClosure : public CLDClosure { 4338 CLDToOopClosure _cm_closure; 4339 public: 4340 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4341 void do_cld(ClassLoaderData* cld) { 4342 // Check if we have modified any oops in the CLD during the concurrent marking. 4343 if (cld->has_accumulated_modified_oops()) { 4344 cld->clear_accumulated_modified_oops(); 4345 4346 // We could have transfered the current modified marks to the accumulated marks, 4347 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4348 } else if (cld->has_modified_oops()) { 4349 // Don't clear anything, this info is needed by the next young collection. 4350 } else { 4351 // No modified oops in the ClassLoaderData. 4352 return; 4353 } 4354 4355 // The klass has modified fields, need to scan the klass. 4356 _cm_closure.do_cld(cld); 4357 } 4358 }; 4359 4360 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4361 ParNewGeneration* young_gen = _collector->_young_gen; 4362 ContiguousSpace* eden_space = young_gen->eden(); 4363 ContiguousSpace* from_space = young_gen->from(); 4364 ContiguousSpace* to_space = young_gen->to(); 4365 4366 HeapWord** eca = _collector->_eden_chunk_array; 4367 size_t ect = _collector->_eden_chunk_index; 4368 HeapWord** sca = _collector->_survivor_chunk_array; 4369 size_t sct = _collector->_survivor_chunk_index; 4370 4371 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4372 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4373 4374 do_young_space_rescan(cl, to_space, NULL, 0); 4375 do_young_space_rescan(cl, from_space, sca, sct); 4376 do_young_space_rescan(cl, eden_space, eca, ect); 4377 } 4378 4379 // work_queue(i) is passed to the closure 4380 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4381 // also is passed to do_dirty_card_rescan_tasks() and to 4382 // do_work_steal() to select the i-th task_queue. 4383 4384 void CMSParRemarkTask::work(uint worker_id) { 4385 elapsedTimer _timer; 4386 ResourceMark rm; 4387 HandleMark hm; 4388 4389 // ---------- rescan from roots -------------- 4390 _timer.start(); 4391 CMSHeap* heap = CMSHeap::heap(); 4392 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4393 _collector->_span, _collector->ref_processor(), 4394 &(_collector->_markBitMap), 4395 work_queue(worker_id)); 4396 4397 // Rescan young gen roots first since these are likely 4398 // coarsely partitioned and may, on that account, constitute 4399 // the critical path; thus, it's best to start off that 4400 // work first. 4401 // ---------- young gen roots -------------- 4402 { 4403 work_on_young_gen_roots(&par_mrias_cl); 4404 _timer.stop(); 4405 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4406 } 4407 4408 // ---------- remaining roots -------------- 4409 _timer.reset(); 4410 _timer.start(); 4411 heap->cms_process_roots(_strong_roots_scope, 4412 false, // yg was scanned above 4413 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4414 _collector->should_unload_classes(), 4415 &par_mrias_cl, 4416 NULL); // The dirty klasses will be handled below 4417 4418 assert(_collector->should_unload_classes() 4419 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4420 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4421 _timer.stop(); 4422 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4423 4424 // ---------- unhandled CLD scanning ---------- 4425 if (worker_id == 0) { // Single threaded at the moment. 4426 _timer.reset(); 4427 _timer.start(); 4428 4429 // Scan all new class loader data objects and new dependencies that were 4430 // introduced during concurrent marking. 4431 ResourceMark rm; 4432 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4433 for (int i = 0; i < array->length(); i++) { 4434 par_mrias_cl.do_cld_nv(array->at(i)); 4435 } 4436 4437 // We don't need to keep track of new CLDs anymore. 4438 ClassLoaderDataGraph::remember_new_clds(false); 4439 4440 _timer.stop(); 4441 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4442 } 4443 4444 // We might have added oops to ClassLoaderData::_handles during the 4445 // concurrent marking phase. These oops do not always point to newly allocated objects 4446 // that are guaranteed to be kept alive. Hence, 4447 // we do have to revisit the _handles block during the remark phase. 4448 4449 // ---------- dirty CLD scanning ---------- 4450 if (worker_id == 0) { // Single threaded at the moment. 4451 _timer.reset(); 4452 _timer.start(); 4453 4454 // Scan all classes that was dirtied during the concurrent marking phase. 4455 RemarkCLDClosure remark_closure(&par_mrias_cl); 4456 ClassLoaderDataGraph::cld_do(&remark_closure); 4457 4458 _timer.stop(); 4459 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4460 } 4461 4462 4463 // ---------- rescan dirty cards ------------ 4464 _timer.reset(); 4465 _timer.start(); 4466 4467 // Do the rescan tasks for each of the two spaces 4468 // (cms_space) in turn. 4469 // "worker_id" is passed to select the task_queue for "worker_id" 4470 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4471 _timer.stop(); 4472 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4473 4474 // ---------- steal work from other threads ... 4475 // ---------- ... and drain overflow list. 4476 _timer.reset(); 4477 _timer.start(); 4478 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4479 _timer.stop(); 4480 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4481 } 4482 4483 void 4484 CMSParMarkTask::do_young_space_rescan( 4485 OopsInGenClosure* cl, ContiguousSpace* space, 4486 HeapWord** chunk_array, size_t chunk_top) { 4487 // Until all tasks completed: 4488 // . claim an unclaimed task 4489 // . compute region boundaries corresponding to task claimed 4490 // using chunk_array 4491 // . par_oop_iterate(cl) over that region 4492 4493 ResourceMark rm; 4494 HandleMark hm; 4495 4496 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4497 4498 uint nth_task = 0; 4499 uint n_tasks = pst->n_tasks(); 4500 4501 if (n_tasks > 0) { 4502 assert(pst->valid(), "Uninitialized use?"); 4503 HeapWord *start, *end; 4504 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4505 // We claimed task # nth_task; compute its boundaries. 4506 if (chunk_top == 0) { // no samples were taken 4507 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4508 start = space->bottom(); 4509 end = space->top(); 4510 } else if (nth_task == 0) { 4511 start = space->bottom(); 4512 end = chunk_array[nth_task]; 4513 } else if (nth_task < (uint)chunk_top) { 4514 assert(nth_task >= 1, "Control point invariant"); 4515 start = chunk_array[nth_task - 1]; 4516 end = chunk_array[nth_task]; 4517 } else { 4518 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4519 start = chunk_array[chunk_top - 1]; 4520 end = space->top(); 4521 } 4522 MemRegion mr(start, end); 4523 // Verify that mr is in space 4524 assert(mr.is_empty() || space->used_region().contains(mr), 4525 "Should be in space"); 4526 // Verify that "start" is an object boundary 4527 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4528 "Should be an oop"); 4529 space->par_oop_iterate(mr, cl); 4530 } 4531 pst->all_tasks_completed(); 4532 } 4533 } 4534 4535 void 4536 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4537 CompactibleFreeListSpace* sp, int i, 4538 ParMarkRefsIntoAndScanClosure* cl) { 4539 // Until all tasks completed: 4540 // . claim an unclaimed task 4541 // . compute region boundaries corresponding to task claimed 4542 // . transfer dirty bits ct->mut for that region 4543 // . apply rescanclosure to dirty mut bits for that region 4544 4545 ResourceMark rm; 4546 HandleMark hm; 4547 4548 OopTaskQueue* work_q = work_queue(i); 4549 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4550 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4551 // CAUTION: This closure has state that persists across calls to 4552 // the work method dirty_range_iterate_clear() in that it has 4553 // embedded in it a (subtype of) UpwardsObjectClosure. The 4554 // use of that state in the embedded UpwardsObjectClosure instance 4555 // assumes that the cards are always iterated (even if in parallel 4556 // by several threads) in monotonically increasing order per each 4557 // thread. This is true of the implementation below which picks 4558 // card ranges (chunks) in monotonically increasing order globally 4559 // and, a-fortiori, in monotonically increasing order per thread 4560 // (the latter order being a subsequence of the former). 4561 // If the work code below is ever reorganized into a more chaotic 4562 // work-partitioning form than the current "sequential tasks" 4563 // paradigm, the use of that persistent state will have to be 4564 // revisited and modified appropriately. See also related 4565 // bug 4756801 work on which should examine this code to make 4566 // sure that the changes there do not run counter to the 4567 // assumptions made here and necessary for correctness and 4568 // efficiency. Note also that this code might yield inefficient 4569 // behavior in the case of very large objects that span one or 4570 // more work chunks. Such objects would potentially be scanned 4571 // several times redundantly. Work on 4756801 should try and 4572 // address that performance anomaly if at all possible. XXX 4573 MemRegion full_span = _collector->_span; 4574 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4575 MarkFromDirtyCardsClosure 4576 greyRescanClosure(_collector, full_span, // entire span of interest 4577 sp, bm, work_q, cl); 4578 4579 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4580 assert(pst->valid(), "Uninitialized use?"); 4581 uint nth_task = 0; 4582 const int alignment = CardTable::card_size * BitsPerWord; 4583 MemRegion span = sp->used_region(); 4584 HeapWord* start_addr = span.start(); 4585 HeapWord* end_addr = align_up(span.end(), alignment); 4586 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4587 assert(is_aligned(start_addr, alignment), "Check alignment"); 4588 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4589 4590 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4591 // Having claimed the nth_task, compute corresponding mem-region, 4592 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4593 // The alignment restriction ensures that we do not need any 4594 // synchronization with other gang-workers while setting or 4595 // clearing bits in thus chunk of the MUT. 4596 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4597 start_addr + (nth_task+1)*chunk_size); 4598 // The last chunk's end might be way beyond end of the 4599 // used region. In that case pull back appropriately. 4600 if (this_span.end() > end_addr) { 4601 this_span.set_end(end_addr); 4602 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4603 } 4604 // Iterate over the dirty cards covering this chunk, marking them 4605 // precleaned, and setting the corresponding bits in the mod union 4606 // table. Since we have been careful to partition at Card and MUT-word 4607 // boundaries no synchronization is needed between parallel threads. 4608 _collector->_ct->dirty_card_iterate(this_span, 4609 &modUnionClosure); 4610 4611 // Having transferred these marks into the modUnionTable, 4612 // rescan the marked objects on the dirty cards in the modUnionTable. 4613 // Even if this is at a synchronous collection, the initial marking 4614 // may have been done during an asynchronous collection so there 4615 // may be dirty bits in the mod-union table. 4616 _collector->_modUnionTable.dirty_range_iterate_clear( 4617 this_span, &greyRescanClosure); 4618 _collector->_modUnionTable.verifyNoOneBitsInRange( 4619 this_span.start(), 4620 this_span.end()); 4621 } 4622 pst->all_tasks_completed(); // declare that i am done 4623 } 4624 4625 // . see if we can share work_queues with ParNew? XXX 4626 void 4627 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4628 int* seed) { 4629 OopTaskQueue* work_q = work_queue(i); 4630 NOT_PRODUCT(int num_steals = 0;) 4631 oop obj_to_scan; 4632 CMSBitMap* bm = &(_collector->_markBitMap); 4633 4634 while (true) { 4635 // Completely finish any left over work from (an) earlier round(s) 4636 cl->trim_queue(0); 4637 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4638 (size_t)ParGCDesiredObjsFromOverflowList); 4639 // Now check if there's any work in the overflow list 4640 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4641 // only affects the number of attempts made to get work from the 4642 // overflow list and does not affect the number of workers. Just 4643 // pass ParallelGCThreads so this behavior is unchanged. 4644 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4645 work_q, 4646 ParallelGCThreads)) { 4647 // found something in global overflow list; 4648 // not yet ready to go stealing work from others. 4649 // We'd like to assert(work_q->size() != 0, ...) 4650 // because we just took work from the overflow list, 4651 // but of course we can't since all of that could have 4652 // been already stolen from us. 4653 // "He giveth and He taketh away." 4654 continue; 4655 } 4656 // Verify that we have no work before we resort to stealing 4657 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4658 // Try to steal from other queues that have work 4659 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4660 NOT_PRODUCT(num_steals++;) 4661 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4662 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4663 // Do scanning work 4664 obj_to_scan->oop_iterate(cl); 4665 // Loop around, finish this work, and try to steal some more 4666 } else if (terminator()->offer_termination()) { 4667 break; // nirvana from the infinite cycle 4668 } 4669 } 4670 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4671 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4672 "Else our work is not yet done"); 4673 } 4674 4675 // Record object boundaries in _eden_chunk_array by sampling the eden 4676 // top in the slow-path eden object allocation code path and record 4677 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4678 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4679 // sampling in sample_eden() that activates during the part of the 4680 // preclean phase. 4681 void CMSCollector::sample_eden_chunk() { 4682 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4683 if (_eden_chunk_lock->try_lock()) { 4684 // Record a sample. This is the critical section. The contents 4685 // of the _eden_chunk_array have to be non-decreasing in the 4686 // address order. 4687 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4688 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4689 "Unexpected state of Eden"); 4690 if (_eden_chunk_index == 0 || 4691 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4692 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4693 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4694 _eden_chunk_index++; // commit sample 4695 } 4696 _eden_chunk_lock->unlock(); 4697 } 4698 } 4699 } 4700 4701 // Return a thread-local PLAB recording array, as appropriate. 4702 void* CMSCollector::get_data_recorder(int thr_num) { 4703 if (_survivor_plab_array != NULL && 4704 (CMSPLABRecordAlways || 4705 (_collectorState > Marking && _collectorState < FinalMarking))) { 4706 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4707 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4708 ca->reset(); // clear it so that fresh data is recorded 4709 return (void*) ca; 4710 } else { 4711 return NULL; 4712 } 4713 } 4714 4715 // Reset all the thread-local PLAB recording arrays 4716 void CMSCollector::reset_survivor_plab_arrays() { 4717 for (uint i = 0; i < ParallelGCThreads; i++) { 4718 _survivor_plab_array[i].reset(); 4719 } 4720 } 4721 4722 // Merge the per-thread plab arrays into the global survivor chunk 4723 // array which will provide the partitioning of the survivor space 4724 // for CMS initial scan and rescan. 4725 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4726 int no_of_gc_threads) { 4727 assert(_survivor_plab_array != NULL, "Error"); 4728 assert(_survivor_chunk_array != NULL, "Error"); 4729 assert(_collectorState == FinalMarking || 4730 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4731 for (int j = 0; j < no_of_gc_threads; j++) { 4732 _cursor[j] = 0; 4733 } 4734 HeapWord* top = surv->top(); 4735 size_t i; 4736 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4737 HeapWord* min_val = top; // Higher than any PLAB address 4738 uint min_tid = 0; // position of min_val this round 4739 for (int j = 0; j < no_of_gc_threads; j++) { 4740 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4741 if (_cursor[j] == cur_sca->end()) { 4742 continue; 4743 } 4744 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4745 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4746 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4747 if (cur_val < min_val) { 4748 min_tid = j; 4749 min_val = cur_val; 4750 } else { 4751 assert(cur_val < top, "All recorded addresses should be less"); 4752 } 4753 } 4754 // At this point min_val and min_tid are respectively 4755 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4756 // and the thread (j) that witnesses that address. 4757 // We record this address in the _survivor_chunk_array[i] 4758 // and increment _cursor[min_tid] prior to the next round i. 4759 if (min_val == top) { 4760 break; 4761 } 4762 _survivor_chunk_array[i] = min_val; 4763 _cursor[min_tid]++; 4764 } 4765 // We are all done; record the size of the _survivor_chunk_array 4766 _survivor_chunk_index = i; // exclusive: [0, i) 4767 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4768 // Verify that we used up all the recorded entries 4769 #ifdef ASSERT 4770 size_t total = 0; 4771 for (int j = 0; j < no_of_gc_threads; j++) { 4772 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4773 total += _cursor[j]; 4774 } 4775 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4776 // Check that the merged array is in sorted order 4777 if (total > 0) { 4778 for (size_t i = 0; i < total - 1; i++) { 4779 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4780 i, p2i(_survivor_chunk_array[i])); 4781 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4782 "Not sorted"); 4783 } 4784 } 4785 #endif // ASSERT 4786 } 4787 4788 // Set up the space's par_seq_tasks structure for work claiming 4789 // for parallel initial scan and rescan of young gen. 4790 // See ParRescanTask where this is currently used. 4791 void 4792 CMSCollector:: 4793 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4794 assert(n_threads > 0, "Unexpected n_threads argument"); 4795 4796 // Eden space 4797 if (!_young_gen->eden()->is_empty()) { 4798 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4799 assert(!pst->valid(), "Clobbering existing data?"); 4800 // Each valid entry in [0, _eden_chunk_index) represents a task. 4801 size_t n_tasks = _eden_chunk_index + 1; 4802 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4803 // Sets the condition for completion of the subtask (how many threads 4804 // need to finish in order to be done). 4805 pst->set_n_threads(n_threads); 4806 pst->set_n_tasks((int)n_tasks); 4807 } 4808 4809 // Merge the survivor plab arrays into _survivor_chunk_array 4810 if (_survivor_plab_array != NULL) { 4811 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4812 } else { 4813 assert(_survivor_chunk_index == 0, "Error"); 4814 } 4815 4816 // To space 4817 { 4818 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4819 assert(!pst->valid(), "Clobbering existing data?"); 4820 // Sets the condition for completion of the subtask (how many threads 4821 // need to finish in order to be done). 4822 pst->set_n_threads(n_threads); 4823 pst->set_n_tasks(1); 4824 assert(pst->valid(), "Error"); 4825 } 4826 4827 // From space 4828 { 4829 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4830 assert(!pst->valid(), "Clobbering existing data?"); 4831 size_t n_tasks = _survivor_chunk_index + 1; 4832 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4833 // Sets the condition for completion of the subtask (how many threads 4834 // need to finish in order to be done). 4835 pst->set_n_threads(n_threads); 4836 pst->set_n_tasks((int)n_tasks); 4837 assert(pst->valid(), "Error"); 4838 } 4839 } 4840 4841 // Parallel version of remark 4842 void CMSCollector::do_remark_parallel() { 4843 CMSHeap* heap = CMSHeap::heap(); 4844 WorkGang* workers = heap->workers(); 4845 assert(workers != NULL, "Need parallel worker threads."); 4846 // Choose to use the number of GC workers most recently set 4847 // into "active_workers". 4848 uint n_workers = workers->active_workers(); 4849 4850 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4851 4852 StrongRootsScope srs(n_workers); 4853 4854 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4855 4856 // We won't be iterating over the cards in the card table updating 4857 // the younger_gen cards, so we shouldn't call the following else 4858 // the verification code as well as subsequent younger_refs_iterate 4859 // code would get confused. XXX 4860 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4861 4862 // The young gen rescan work will not be done as part of 4863 // process_roots (which currently doesn't know how to 4864 // parallelize such a scan), but rather will be broken up into 4865 // a set of parallel tasks (via the sampling that the [abortable] 4866 // preclean phase did of eden, plus the [two] tasks of 4867 // scanning the [two] survivor spaces. Further fine-grain 4868 // parallelization of the scanning of the survivor spaces 4869 // themselves, and of precleaning of the young gen itself 4870 // is deferred to the future. 4871 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4872 4873 // The dirty card rescan work is broken up into a "sequence" 4874 // of parallel tasks (per constituent space) that are dynamically 4875 // claimed by the parallel threads. 4876 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4877 4878 // It turns out that even when we're using 1 thread, doing the work in a 4879 // separate thread causes wide variance in run times. We can't help this 4880 // in the multi-threaded case, but we special-case n=1 here to get 4881 // repeatable measurements of the 1-thread overhead of the parallel code. 4882 if (n_workers > 1) { 4883 // Make refs discovery MT-safe, if it isn't already: it may not 4884 // necessarily be so, since it's possible that we are doing 4885 // ST marking. 4886 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4887 workers->run_task(&tsk); 4888 } else { 4889 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4890 tsk.work(0); 4891 } 4892 4893 // restore, single-threaded for now, any preserved marks 4894 // as a result of work_q overflow 4895 restore_preserved_marks_if_any(); 4896 } 4897 4898 // Non-parallel version of remark 4899 void CMSCollector::do_remark_non_parallel() { 4900 ResourceMark rm; 4901 HandleMark hm; 4902 CMSHeap* heap = CMSHeap::heap(); 4903 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4904 4905 MarkRefsIntoAndScanClosure 4906 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4907 &_markStack, this, 4908 false /* should_yield */, false /* not precleaning */); 4909 MarkFromDirtyCardsClosure 4910 markFromDirtyCardsClosure(this, _span, 4911 NULL, // space is set further below 4912 &_markBitMap, &_markStack, &mrias_cl); 4913 { 4914 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4915 // Iterate over the dirty cards, setting the corresponding bits in the 4916 // mod union table. 4917 { 4918 ModUnionClosure modUnionClosure(&_modUnionTable); 4919 _ct->dirty_card_iterate(_cmsGen->used_region(), 4920 &modUnionClosure); 4921 } 4922 // Having transferred these marks into the modUnionTable, we just need 4923 // to rescan the marked objects on the dirty cards in the modUnionTable. 4924 // The initial marking may have been done during an asynchronous 4925 // collection so there may be dirty bits in the mod-union table. 4926 const int alignment = CardTable::card_size * BitsPerWord; 4927 { 4928 // ... First handle dirty cards in CMS gen 4929 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4930 MemRegion ur = _cmsGen->used_region(); 4931 HeapWord* lb = ur.start(); 4932 HeapWord* ub = align_up(ur.end(), alignment); 4933 MemRegion cms_span(lb, ub); 4934 _modUnionTable.dirty_range_iterate_clear(cms_span, 4935 &markFromDirtyCardsClosure); 4936 verify_work_stacks_empty(); 4937 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4938 } 4939 } 4940 if (VerifyDuringGC && 4941 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4942 HandleMark hm; // Discard invalid handles created during verification 4943 Universe::verify(); 4944 } 4945 { 4946 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4947 4948 verify_work_stacks_empty(); 4949 4950 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4951 StrongRootsScope srs(1); 4952 4953 heap->cms_process_roots(&srs, 4954 true, // young gen as roots 4955 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4956 should_unload_classes(), 4957 &mrias_cl, 4958 NULL); // The dirty klasses will be handled below 4959 4960 assert(should_unload_classes() 4961 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4962 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4963 } 4964 4965 { 4966 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4967 4968 verify_work_stacks_empty(); 4969 4970 // Scan all class loader data objects that might have been introduced 4971 // during concurrent marking. 4972 ResourceMark rm; 4973 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4974 for (int i = 0; i < array->length(); i++) { 4975 mrias_cl.do_cld_nv(array->at(i)); 4976 } 4977 4978 // We don't need to keep track of new CLDs anymore. 4979 ClassLoaderDataGraph::remember_new_clds(false); 4980 4981 verify_work_stacks_empty(); 4982 } 4983 4984 // We might have added oops to ClassLoaderData::_handles during the 4985 // concurrent marking phase. These oops do not point to newly allocated objects 4986 // that are guaranteed to be kept alive. Hence, 4987 // we do have to revisit the _handles block during the remark phase. 4988 { 4989 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4990 4991 verify_work_stacks_empty(); 4992 4993 RemarkCLDClosure remark_closure(&mrias_cl); 4994 ClassLoaderDataGraph::cld_do(&remark_closure); 4995 4996 verify_work_stacks_empty(); 4997 } 4998 4999 verify_work_stacks_empty(); 5000 // Restore evacuated mark words, if any, used for overflow list links 5001 restore_preserved_marks_if_any(); 5002 5003 verify_overflow_empty(); 5004 } 5005 5006 //////////////////////////////////////////////////////// 5007 // Parallel Reference Processing Task Proxy Class 5008 //////////////////////////////////////////////////////// 5009 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5010 OopTaskQueueSet* _queues; 5011 ParallelTaskTerminator _terminator; 5012 public: 5013 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5014 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5015 ParallelTaskTerminator* terminator() { return &_terminator; } 5016 OopTaskQueueSet* queues() { return _queues; } 5017 }; 5018 5019 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5020 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5021 CMSCollector* _collector; 5022 CMSBitMap* _mark_bit_map; 5023 const MemRegion _span; 5024 ProcessTask& _task; 5025 5026 public: 5027 CMSRefProcTaskProxy(ProcessTask& task, 5028 CMSCollector* collector, 5029 const MemRegion& span, 5030 CMSBitMap* mark_bit_map, 5031 AbstractWorkGang* workers, 5032 OopTaskQueueSet* task_queues): 5033 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5034 task_queues, 5035 workers->active_workers()), 5036 _task(task), 5037 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5038 { 5039 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5040 "Inconsistency in _span"); 5041 } 5042 5043 OopTaskQueueSet* task_queues() { return queues(); } 5044 5045 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5046 5047 void do_work_steal(int i, 5048 CMSParDrainMarkingStackClosure* drain, 5049 CMSParKeepAliveClosure* keep_alive, 5050 int* seed); 5051 5052 virtual void work(uint worker_id); 5053 }; 5054 5055 void CMSRefProcTaskProxy::work(uint worker_id) { 5056 ResourceMark rm; 5057 HandleMark hm; 5058 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5059 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5060 _mark_bit_map, 5061 work_queue(worker_id)); 5062 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5063 _mark_bit_map, 5064 work_queue(worker_id)); 5065 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5066 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5067 if (_task.marks_oops_alive()) { 5068 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5069 _collector->hash_seed(worker_id)); 5070 } 5071 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5072 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5073 } 5074 5075 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5076 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5077 EnqueueTask& _task; 5078 5079 public: 5080 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5081 : AbstractGangTask("Enqueue reference objects in parallel"), 5082 _task(task) 5083 { } 5084 5085 virtual void work(uint worker_id) 5086 { 5087 _task.work(worker_id); 5088 } 5089 }; 5090 5091 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5092 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5093 _span(span), 5094 _bit_map(bit_map), 5095 _work_queue(work_queue), 5096 _mark_and_push(collector, span, bit_map, work_queue), 5097 _low_water_mark(MIN2((work_queue->max_elems()/4), 5098 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5099 { } 5100 5101 // . see if we can share work_queues with ParNew? XXX 5102 void CMSRefProcTaskProxy::do_work_steal(int i, 5103 CMSParDrainMarkingStackClosure* drain, 5104 CMSParKeepAliveClosure* keep_alive, 5105 int* seed) { 5106 OopTaskQueue* work_q = work_queue(i); 5107 NOT_PRODUCT(int num_steals = 0;) 5108 oop obj_to_scan; 5109 5110 while (true) { 5111 // Completely finish any left over work from (an) earlier round(s) 5112 drain->trim_queue(0); 5113 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5114 (size_t)ParGCDesiredObjsFromOverflowList); 5115 // Now check if there's any work in the overflow list 5116 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5117 // only affects the number of attempts made to get work from the 5118 // overflow list and does not affect the number of workers. Just 5119 // pass ParallelGCThreads so this behavior is unchanged. 5120 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5121 work_q, 5122 ParallelGCThreads)) { 5123 // Found something in global overflow list; 5124 // not yet ready to go stealing work from others. 5125 // We'd like to assert(work_q->size() != 0, ...) 5126 // because we just took work from the overflow list, 5127 // but of course we can't, since all of that might have 5128 // been already stolen from us. 5129 continue; 5130 } 5131 // Verify that we have no work before we resort to stealing 5132 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5133 // Try to steal from other queues that have work 5134 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5135 NOT_PRODUCT(num_steals++;) 5136 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5137 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5138 // Do scanning work 5139 obj_to_scan->oop_iterate(keep_alive); 5140 // Loop around, finish this work, and try to steal some more 5141 } else if (terminator()->offer_termination()) { 5142 break; // nirvana from the infinite cycle 5143 } 5144 } 5145 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5146 } 5147 5148 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5149 { 5150 CMSHeap* heap = CMSHeap::heap(); 5151 WorkGang* workers = heap->workers(); 5152 assert(workers != NULL, "Need parallel worker threads."); 5153 CMSRefProcTaskProxy rp_task(task, &_collector, 5154 _collector.ref_processor()->span(), 5155 _collector.markBitMap(), 5156 workers, _collector.task_queues()); 5157 workers->run_task(&rp_task); 5158 } 5159 5160 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5161 { 5162 5163 CMSHeap* heap = CMSHeap::heap(); 5164 WorkGang* workers = heap->workers(); 5165 assert(workers != NULL, "Need parallel worker threads."); 5166 CMSRefEnqueueTaskProxy enq_task(task); 5167 workers->run_task(&enq_task); 5168 } 5169 5170 void CMSCollector::refProcessingWork() { 5171 ResourceMark rm; 5172 HandleMark hm; 5173 5174 ReferenceProcessor* rp = ref_processor(); 5175 assert(rp->span().equals(_span), "Spans should be equal"); 5176 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5177 // Process weak references. 5178 rp->setup_policy(false); 5179 verify_work_stacks_empty(); 5180 5181 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q()); 5182 { 5183 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5184 5185 // Setup keep_alive and complete closures. 5186 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5187 &_markStack, false /* !preclean */); 5188 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5189 _span, &_markBitMap, &_markStack, 5190 &cmsKeepAliveClosure, false /* !preclean */); 5191 5192 ReferenceProcessorStats stats; 5193 if (rp->processing_is_mt()) { 5194 // Set the degree of MT here. If the discovery is done MT, there 5195 // may have been a different number of threads doing the discovery 5196 // and a different number of discovered lists may have Ref objects. 5197 // That is OK as long as the Reference lists are balanced (see 5198 // balance_all_queues() and balance_queues()). 5199 CMSHeap* heap = CMSHeap::heap(); 5200 uint active_workers = ParallelGCThreads; 5201 WorkGang* workers = heap->workers(); 5202 if (workers != NULL) { 5203 active_workers = workers->active_workers(); 5204 // The expectation is that active_workers will have already 5205 // been set to a reasonable value. If it has not been set, 5206 // investigate. 5207 assert(active_workers > 0, "Should have been set during scavenge"); 5208 } 5209 rp->set_active_mt_degree(active_workers); 5210 CMSRefProcTaskExecutor task_executor(*this); 5211 stats = rp->process_discovered_references(&_is_alive_closure, 5212 &cmsKeepAliveClosure, 5213 &cmsDrainMarkingStackClosure, 5214 &task_executor, 5215 &pt); 5216 } else { 5217 stats = rp->process_discovered_references(&_is_alive_closure, 5218 &cmsKeepAliveClosure, 5219 &cmsDrainMarkingStackClosure, 5220 NULL, 5221 &pt); 5222 } 5223 _gc_tracer_cm->report_gc_reference_stats(stats); 5224 pt.print_all_references(); 5225 } 5226 5227 // This is the point where the entire marking should have completed. 5228 verify_work_stacks_empty(); 5229 5230 { 5231 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5232 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5233 } 5234 5235 if (should_unload_classes()) { 5236 { 5237 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5238 5239 // Unload classes and purge the SystemDictionary. 5240 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5241 5242 // Unload nmethods. 5243 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5244 5245 // Prune dead klasses from subklass/sibling/implementor lists. 5246 Klass::clean_weak_klass_links(&_is_alive_closure); 5247 } 5248 5249 { 5250 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5251 // Clean up unreferenced symbols in symbol table. 5252 SymbolTable::unlink(); 5253 } 5254 5255 { 5256 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5257 // Delete entries for dead interned strings. 5258 StringTable::unlink(&_is_alive_closure); 5259 } 5260 } 5261 5262 // Restore any preserved marks as a result of mark stack or 5263 // work queue overflow 5264 restore_preserved_marks_if_any(); // done single-threaded for now 5265 5266 rp->set_enqueuing_is_done(true); 5267 if (rp->processing_is_mt()) { 5268 rp->balance_all_queues(); 5269 CMSRefProcTaskExecutor task_executor(*this); 5270 rp->enqueue_discovered_references(&task_executor, &pt); 5271 } else { 5272 rp->enqueue_discovered_references(NULL, &pt); 5273 } 5274 rp->verify_no_references_recorded(); 5275 pt.print_enqueue_phase(); 5276 assert(!rp->discovery_enabled(), "should have been disabled"); 5277 } 5278 5279 #ifndef PRODUCT 5280 void CMSCollector::check_correct_thread_executing() { 5281 Thread* t = Thread::current(); 5282 // Only the VM thread or the CMS thread should be here. 5283 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5284 "Unexpected thread type"); 5285 // If this is the vm thread, the foreground process 5286 // should not be waiting. Note that _foregroundGCIsActive is 5287 // true while the foreground collector is waiting. 5288 if (_foregroundGCShouldWait) { 5289 // We cannot be the VM thread 5290 assert(t->is_ConcurrentGC_thread(), 5291 "Should be CMS thread"); 5292 } else { 5293 // We can be the CMS thread only if we are in a stop-world 5294 // phase of CMS collection. 5295 if (t->is_ConcurrentGC_thread()) { 5296 assert(_collectorState == InitialMarking || 5297 _collectorState == FinalMarking, 5298 "Should be a stop-world phase"); 5299 // The CMS thread should be holding the CMS_token. 5300 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5301 "Potential interference with concurrently " 5302 "executing VM thread"); 5303 } 5304 } 5305 } 5306 #endif 5307 5308 void CMSCollector::sweep() { 5309 assert(_collectorState == Sweeping, "just checking"); 5310 check_correct_thread_executing(); 5311 verify_work_stacks_empty(); 5312 verify_overflow_empty(); 5313 increment_sweep_count(); 5314 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5315 5316 _inter_sweep_timer.stop(); 5317 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5318 5319 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5320 _intra_sweep_timer.reset(); 5321 _intra_sweep_timer.start(); 5322 { 5323 GCTraceCPUTime tcpu; 5324 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5325 // First sweep the old gen 5326 { 5327 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5328 bitMapLock()); 5329 sweepWork(_cmsGen); 5330 } 5331 5332 // Update Universe::_heap_*_at_gc figures. 5333 // We need all the free list locks to make the abstract state 5334 // transition from Sweeping to Resetting. See detailed note 5335 // further below. 5336 { 5337 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5338 // Update heap occupancy information which is used as 5339 // input to soft ref clearing policy at the next gc. 5340 Universe::update_heap_info_at_gc(); 5341 _collectorState = Resizing; 5342 } 5343 } 5344 verify_work_stacks_empty(); 5345 verify_overflow_empty(); 5346 5347 if (should_unload_classes()) { 5348 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5349 // requires that the virtual spaces are stable and not deleted. 5350 ClassLoaderDataGraph::set_should_purge(true); 5351 } 5352 5353 _intra_sweep_timer.stop(); 5354 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5355 5356 _inter_sweep_timer.reset(); 5357 _inter_sweep_timer.start(); 5358 5359 // We need to use a monotonically non-decreasing time in ms 5360 // or we will see time-warp warnings and os::javaTimeMillis() 5361 // does not guarantee monotonicity. 5362 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5363 update_time_of_last_gc(now); 5364 5365 // NOTE on abstract state transitions: 5366 // Mutators allocate-live and/or mark the mod-union table dirty 5367 // based on the state of the collection. The former is done in 5368 // the interval [Marking, Sweeping] and the latter in the interval 5369 // [Marking, Sweeping). Thus the transitions into the Marking state 5370 // and out of the Sweeping state must be synchronously visible 5371 // globally to the mutators. 5372 // The transition into the Marking state happens with the world 5373 // stopped so the mutators will globally see it. Sweeping is 5374 // done asynchronously by the background collector so the transition 5375 // from the Sweeping state to the Resizing state must be done 5376 // under the freelistLock (as is the check for whether to 5377 // allocate-live and whether to dirty the mod-union table). 5378 assert(_collectorState == Resizing, "Change of collector state to" 5379 " Resizing must be done under the freelistLocks (plural)"); 5380 5381 // Now that sweeping has been completed, we clear 5382 // the incremental_collection_failed flag, 5383 // thus inviting a younger gen collection to promote into 5384 // this generation. If such a promotion may still fail, 5385 // the flag will be set again when a young collection is 5386 // attempted. 5387 CMSHeap* heap = CMSHeap::heap(); 5388 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5389 heap->update_full_collections_completed(_collection_count_start); 5390 } 5391 5392 // FIX ME!!! Looks like this belongs in CFLSpace, with 5393 // CMSGen merely delegating to it. 5394 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5395 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5396 HeapWord* minAddr = _cmsSpace->bottom(); 5397 HeapWord* largestAddr = 5398 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5399 if (largestAddr == NULL) { 5400 // The dictionary appears to be empty. In this case 5401 // try to coalesce at the end of the heap. 5402 largestAddr = _cmsSpace->end(); 5403 } 5404 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5405 size_t nearLargestOffset = 5406 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5407 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5408 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5409 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5410 } 5411 5412 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5413 return addr >= _cmsSpace->nearLargestChunk(); 5414 } 5415 5416 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5417 return _cmsSpace->find_chunk_at_end(); 5418 } 5419 5420 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5421 bool full) { 5422 // If the young generation has been collected, gather any statistics 5423 // that are of interest at this point. 5424 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5425 if (!full && current_is_young) { 5426 // Gather statistics on the young generation collection. 5427 collector()->stats().record_gc0_end(used()); 5428 } 5429 } 5430 5431 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5432 // We iterate over the space(s) underlying this generation, 5433 // checking the mark bit map to see if the bits corresponding 5434 // to specific blocks are marked or not. Blocks that are 5435 // marked are live and are not swept up. All remaining blocks 5436 // are swept up, with coalescing on-the-fly as we sweep up 5437 // contiguous free and/or garbage blocks: 5438 // We need to ensure that the sweeper synchronizes with allocators 5439 // and stop-the-world collectors. In particular, the following 5440 // locks are used: 5441 // . CMS token: if this is held, a stop the world collection cannot occur 5442 // . freelistLock: if this is held no allocation can occur from this 5443 // generation by another thread 5444 // . bitMapLock: if this is held, no other thread can access or update 5445 // 5446 5447 // Note that we need to hold the freelistLock if we use 5448 // block iterate below; else the iterator might go awry if 5449 // a mutator (or promotion) causes block contents to change 5450 // (for instance if the allocator divvies up a block). 5451 // If we hold the free list lock, for all practical purposes 5452 // young generation GC's can't occur (they'll usually need to 5453 // promote), so we might as well prevent all young generation 5454 // GC's while we do a sweeping step. For the same reason, we might 5455 // as well take the bit map lock for the entire duration 5456 5457 // check that we hold the requisite locks 5458 assert(have_cms_token(), "Should hold cms token"); 5459 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5460 assert_lock_strong(old_gen->freelistLock()); 5461 assert_lock_strong(bitMapLock()); 5462 5463 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5464 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5465 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5466 _inter_sweep_estimate.padded_average(), 5467 _intra_sweep_estimate.padded_average()); 5468 old_gen->setNearLargestChunk(); 5469 5470 { 5471 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5472 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5473 // We need to free-up/coalesce garbage/blocks from a 5474 // co-terminal free run. This is done in the SweepClosure 5475 // destructor; so, do not remove this scope, else the 5476 // end-of-sweep-census below will be off by a little bit. 5477 } 5478 old_gen->cmsSpace()->sweep_completed(); 5479 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5480 if (should_unload_classes()) { // unloaded classes this cycle, 5481 _concurrent_cycles_since_last_unload = 0; // ... reset count 5482 } else { // did not unload classes, 5483 _concurrent_cycles_since_last_unload++; // ... increment count 5484 } 5485 } 5486 5487 // Reset CMS data structures (for now just the marking bit map) 5488 // preparatory for the next cycle. 5489 void CMSCollector::reset_concurrent() { 5490 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5491 5492 // If the state is not "Resetting", the foreground thread 5493 // has done a collection and the resetting. 5494 if (_collectorState != Resetting) { 5495 assert(_collectorState == Idling, "The state should only change" 5496 " because the foreground collector has finished the collection"); 5497 return; 5498 } 5499 5500 { 5501 // Clear the mark bitmap (no grey objects to start with) 5502 // for the next cycle. 5503 GCTraceCPUTime tcpu; 5504 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5505 5506 HeapWord* curAddr = _markBitMap.startWord(); 5507 while (curAddr < _markBitMap.endWord()) { 5508 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5509 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5510 _markBitMap.clear_large_range(chunk); 5511 if (ConcurrentMarkSweepThread::should_yield() && 5512 !foregroundGCIsActive() && 5513 CMSYield) { 5514 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5515 "CMS thread should hold CMS token"); 5516 assert_lock_strong(bitMapLock()); 5517 bitMapLock()->unlock(); 5518 ConcurrentMarkSweepThread::desynchronize(true); 5519 stopTimer(); 5520 incrementYields(); 5521 5522 // See the comment in coordinator_yield() 5523 for (unsigned i = 0; i < CMSYieldSleepCount && 5524 ConcurrentMarkSweepThread::should_yield() && 5525 !CMSCollector::foregroundGCIsActive(); ++i) { 5526 os::sleep(Thread::current(), 1, false); 5527 } 5528 5529 ConcurrentMarkSweepThread::synchronize(true); 5530 bitMapLock()->lock_without_safepoint_check(); 5531 startTimer(); 5532 } 5533 curAddr = chunk.end(); 5534 } 5535 // A successful mostly concurrent collection has been done. 5536 // Because only the full (i.e., concurrent mode failure) collections 5537 // are being measured for gc overhead limits, clean the "near" flag 5538 // and count. 5539 size_policy()->reset_gc_overhead_limit_count(); 5540 _collectorState = Idling; 5541 } 5542 5543 register_gc_end(); 5544 } 5545 5546 // Same as above but for STW paths 5547 void CMSCollector::reset_stw() { 5548 // already have the lock 5549 assert(_collectorState == Resetting, "just checking"); 5550 assert_lock_strong(bitMapLock()); 5551 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5552 _markBitMap.clear_all(); 5553 _collectorState = Idling; 5554 register_gc_end(); 5555 } 5556 5557 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5558 GCTraceCPUTime tcpu; 5559 TraceCollectorStats tcs_cgc(cgc_counters()); 5560 5561 switch (op) { 5562 case CMS_op_checkpointRootsInitial: { 5563 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5564 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5565 checkpointRootsInitial(); 5566 break; 5567 } 5568 case CMS_op_checkpointRootsFinal: { 5569 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5570 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5571 checkpointRootsFinal(); 5572 break; 5573 } 5574 default: 5575 fatal("No such CMS_op"); 5576 } 5577 } 5578 5579 #ifndef PRODUCT 5580 size_t const CMSCollector::skip_header_HeapWords() { 5581 return FreeChunk::header_size(); 5582 } 5583 5584 // Try and collect here conditions that should hold when 5585 // CMS thread is exiting. The idea is that the foreground GC 5586 // thread should not be blocked if it wants to terminate 5587 // the CMS thread and yet continue to run the VM for a while 5588 // after that. 5589 void CMSCollector::verify_ok_to_terminate() const { 5590 assert(Thread::current()->is_ConcurrentGC_thread(), 5591 "should be called by CMS thread"); 5592 assert(!_foregroundGCShouldWait, "should be false"); 5593 // We could check here that all the various low-level locks 5594 // are not held by the CMS thread, but that is overkill; see 5595 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5596 // is checked. 5597 } 5598 #endif 5599 5600 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5601 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5602 "missing Printezis mark?"); 5603 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5604 size_t size = pointer_delta(nextOneAddr + 1, addr); 5605 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5606 "alignment problem"); 5607 assert(size >= 3, "Necessary for Printezis marks to work"); 5608 return size; 5609 } 5610 5611 // A variant of the above (block_size_using_printezis_bits()) except 5612 // that we return 0 if the P-bits are not yet set. 5613 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5614 if (_markBitMap.isMarked(addr + 1)) { 5615 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5616 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5617 size_t size = pointer_delta(nextOneAddr + 1, addr); 5618 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5619 "alignment problem"); 5620 assert(size >= 3, "Necessary for Printezis marks to work"); 5621 return size; 5622 } 5623 return 0; 5624 } 5625 5626 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5627 size_t sz = 0; 5628 oop p = (oop)addr; 5629 if (p->klass_or_null_acquire() != NULL) { 5630 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5631 } else { 5632 sz = block_size_using_printezis_bits(addr); 5633 } 5634 assert(sz > 0, "size must be nonzero"); 5635 HeapWord* next_block = addr + sz; 5636 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5637 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5638 align_down((uintptr_t)next_card, CardTable::card_size), 5639 "must be different cards"); 5640 return next_card; 5641 } 5642 5643 5644 // CMS Bit Map Wrapper ///////////////////////////////////////// 5645 5646 // Construct a CMS bit map infrastructure, but don't create the 5647 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5648 // further below. 5649 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5650 _bm(), 5651 _shifter(shifter), 5652 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5653 Monitor::_safepoint_check_sometimes) : NULL) 5654 { 5655 _bmStartWord = 0; 5656 _bmWordSize = 0; 5657 } 5658 5659 bool CMSBitMap::allocate(MemRegion mr) { 5660 _bmStartWord = mr.start(); 5661 _bmWordSize = mr.word_size(); 5662 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5663 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5664 if (!brs.is_reserved()) { 5665 log_warning(gc)("CMS bit map allocation failure"); 5666 return false; 5667 } 5668 // For now we'll just commit all of the bit map up front. 5669 // Later on we'll try to be more parsimonious with swap. 5670 if (!_virtual_space.initialize(brs, brs.size())) { 5671 log_warning(gc)("CMS bit map backing store failure"); 5672 return false; 5673 } 5674 assert(_virtual_space.committed_size() == brs.size(), 5675 "didn't reserve backing store for all of CMS bit map?"); 5676 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5677 _bmWordSize, "inconsistency in bit map sizing"); 5678 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5679 5680 // bm.clear(); // can we rely on getting zero'd memory? verify below 5681 assert(isAllClear(), 5682 "Expected zero'd memory from ReservedSpace constructor"); 5683 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5684 "consistency check"); 5685 return true; 5686 } 5687 5688 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5689 HeapWord *next_addr, *end_addr, *last_addr; 5690 assert_locked(); 5691 assert(covers(mr), "out-of-range error"); 5692 // XXX assert that start and end are appropriately aligned 5693 for (next_addr = mr.start(), end_addr = mr.end(); 5694 next_addr < end_addr; next_addr = last_addr) { 5695 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5696 last_addr = dirty_region.end(); 5697 if (!dirty_region.is_empty()) { 5698 cl->do_MemRegion(dirty_region); 5699 } else { 5700 assert(last_addr == end_addr, "program logic"); 5701 return; 5702 } 5703 } 5704 } 5705 5706 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5707 _bm.print_on_error(st, prefix); 5708 } 5709 5710 #ifndef PRODUCT 5711 void CMSBitMap::assert_locked() const { 5712 CMSLockVerifier::assert_locked(lock()); 5713 } 5714 5715 bool CMSBitMap::covers(MemRegion mr) const { 5716 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5717 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5718 "size inconsistency"); 5719 return (mr.start() >= _bmStartWord) && 5720 (mr.end() <= endWord()); 5721 } 5722 5723 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5724 return (start >= _bmStartWord && (start + size) <= endWord()); 5725 } 5726 5727 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5728 // verify that there are no 1 bits in the interval [left, right) 5729 FalseBitMapClosure falseBitMapClosure; 5730 iterate(&falseBitMapClosure, left, right); 5731 } 5732 5733 void CMSBitMap::region_invariant(MemRegion mr) 5734 { 5735 assert_locked(); 5736 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5737 assert(!mr.is_empty(), "unexpected empty region"); 5738 assert(covers(mr), "mr should be covered by bit map"); 5739 // convert address range into offset range 5740 size_t start_ofs = heapWordToOffset(mr.start()); 5741 // Make sure that end() is appropriately aligned 5742 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5743 "Misaligned mr.end()"); 5744 size_t end_ofs = heapWordToOffset(mr.end()); 5745 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5746 } 5747 5748 #endif 5749 5750 bool CMSMarkStack::allocate(size_t size) { 5751 // allocate a stack of the requisite depth 5752 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5753 size * sizeof(oop))); 5754 if (!rs.is_reserved()) { 5755 log_warning(gc)("CMSMarkStack allocation failure"); 5756 return false; 5757 } 5758 if (!_virtual_space.initialize(rs, rs.size())) { 5759 log_warning(gc)("CMSMarkStack backing store failure"); 5760 return false; 5761 } 5762 assert(_virtual_space.committed_size() == rs.size(), 5763 "didn't reserve backing store for all of CMS stack?"); 5764 _base = (oop*)(_virtual_space.low()); 5765 _index = 0; 5766 _capacity = size; 5767 NOT_PRODUCT(_max_depth = 0); 5768 return true; 5769 } 5770 5771 // XXX FIX ME !!! In the MT case we come in here holding a 5772 // leaf lock. For printing we need to take a further lock 5773 // which has lower rank. We need to recalibrate the two 5774 // lock-ranks involved in order to be able to print the 5775 // messages below. (Or defer the printing to the caller. 5776 // For now we take the expedient path of just disabling the 5777 // messages for the problematic case.) 5778 void CMSMarkStack::expand() { 5779 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5780 if (_capacity == MarkStackSizeMax) { 5781 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5782 // We print a warning message only once per CMS cycle. 5783 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5784 } 5785 return; 5786 } 5787 // Double capacity if possible 5788 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5789 // Do not give up existing stack until we have managed to 5790 // get the double capacity that we desired. 5791 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5792 new_capacity * sizeof(oop))); 5793 if (rs.is_reserved()) { 5794 // Release the backing store associated with old stack 5795 _virtual_space.release(); 5796 // Reinitialize virtual space for new stack 5797 if (!_virtual_space.initialize(rs, rs.size())) { 5798 fatal("Not enough swap for expanded marking stack"); 5799 } 5800 _base = (oop*)(_virtual_space.low()); 5801 _index = 0; 5802 _capacity = new_capacity; 5803 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5804 // Failed to double capacity, continue; 5805 // we print a detail message only once per CMS cycle. 5806 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5807 _capacity / K, new_capacity / K); 5808 } 5809 } 5810 5811 5812 // Closures 5813 // XXX: there seems to be a lot of code duplication here; 5814 // should refactor and consolidate common code. 5815 5816 // This closure is used to mark refs into the CMS generation in 5817 // the CMS bit map. Called at the first checkpoint. This closure 5818 // assumes that we do not need to re-mark dirty cards; if the CMS 5819 // generation on which this is used is not an oldest 5820 // generation then this will lose younger_gen cards! 5821 5822 MarkRefsIntoClosure::MarkRefsIntoClosure( 5823 MemRegion span, CMSBitMap* bitMap): 5824 _span(span), 5825 _bitMap(bitMap) 5826 { 5827 assert(ref_processor() == NULL, "deliberately left NULL"); 5828 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5829 } 5830 5831 void MarkRefsIntoClosure::do_oop(oop obj) { 5832 // if p points into _span, then mark corresponding bit in _markBitMap 5833 assert(oopDesc::is_oop(obj), "expected an oop"); 5834 HeapWord* addr = (HeapWord*)obj; 5835 if (_span.contains(addr)) { 5836 // this should be made more efficient 5837 _bitMap->mark(addr); 5838 } 5839 } 5840 5841 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5842 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5843 5844 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5845 MemRegion span, CMSBitMap* bitMap): 5846 _span(span), 5847 _bitMap(bitMap) 5848 { 5849 assert(ref_processor() == NULL, "deliberately left NULL"); 5850 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5851 } 5852 5853 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5854 // if p points into _span, then mark corresponding bit in _markBitMap 5855 assert(oopDesc::is_oop(obj), "expected an oop"); 5856 HeapWord* addr = (HeapWord*)obj; 5857 if (_span.contains(addr)) { 5858 // this should be made more efficient 5859 _bitMap->par_mark(addr); 5860 } 5861 } 5862 5863 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5864 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5865 5866 // A variant of the above, used for CMS marking verification. 5867 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5868 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5869 _span(span), 5870 _verification_bm(verification_bm), 5871 _cms_bm(cms_bm) 5872 { 5873 assert(ref_processor() == NULL, "deliberately left NULL"); 5874 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5875 } 5876 5877 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5878 // if p points into _span, then mark corresponding bit in _markBitMap 5879 assert(oopDesc::is_oop(obj), "expected an oop"); 5880 HeapWord* addr = (HeapWord*)obj; 5881 if (_span.contains(addr)) { 5882 _verification_bm->mark(addr); 5883 if (!_cms_bm->isMarked(addr)) { 5884 Log(gc, verify) log; 5885 ResourceMark rm; 5886 LogStream ls(log.error()); 5887 oop(addr)->print_on(&ls); 5888 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5889 fatal("... aborting"); 5890 } 5891 } 5892 } 5893 5894 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5895 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5896 5897 ////////////////////////////////////////////////// 5898 // MarkRefsIntoAndScanClosure 5899 ////////////////////////////////////////////////// 5900 5901 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5902 ReferenceProcessor* rp, 5903 CMSBitMap* bit_map, 5904 CMSBitMap* mod_union_table, 5905 CMSMarkStack* mark_stack, 5906 CMSCollector* collector, 5907 bool should_yield, 5908 bool concurrent_precleaning): 5909 _collector(collector), 5910 _span(span), 5911 _bit_map(bit_map), 5912 _mark_stack(mark_stack), 5913 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5914 mark_stack, concurrent_precleaning), 5915 _yield(should_yield), 5916 _concurrent_precleaning(concurrent_precleaning), 5917 _freelistLock(NULL) 5918 { 5919 // FIXME: Should initialize in base class constructor. 5920 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5921 set_ref_processor_internal(rp); 5922 } 5923 5924 // This closure is used to mark refs into the CMS generation at the 5925 // second (final) checkpoint, and to scan and transitively follow 5926 // the unmarked oops. It is also used during the concurrent precleaning 5927 // phase while scanning objects on dirty cards in the CMS generation. 5928 // The marks are made in the marking bit map and the marking stack is 5929 // used for keeping the (newly) grey objects during the scan. 5930 // The parallel version (Par_...) appears further below. 5931 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5932 if (obj != NULL) { 5933 assert(oopDesc::is_oop(obj), "expected an oop"); 5934 HeapWord* addr = (HeapWord*)obj; 5935 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5936 assert(_collector->overflow_list_is_empty(), 5937 "overflow list should be empty"); 5938 if (_span.contains(addr) && 5939 !_bit_map->isMarked(addr)) { 5940 // mark bit map (object is now grey) 5941 _bit_map->mark(addr); 5942 // push on marking stack (stack should be empty), and drain the 5943 // stack by applying this closure to the oops in the oops popped 5944 // from the stack (i.e. blacken the grey objects) 5945 bool res = _mark_stack->push(obj); 5946 assert(res, "Should have space to push on empty stack"); 5947 do { 5948 oop new_oop = _mark_stack->pop(); 5949 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5950 assert(_bit_map->isMarked((HeapWord*)new_oop), 5951 "only grey objects on this stack"); 5952 // iterate over the oops in this oop, marking and pushing 5953 // the ones in CMS heap (i.e. in _span). 5954 new_oop->oop_iterate(&_pushAndMarkClosure); 5955 // check if it's time to yield 5956 do_yield_check(); 5957 } while (!_mark_stack->isEmpty() || 5958 (!_concurrent_precleaning && take_from_overflow_list())); 5959 // if marking stack is empty, and we are not doing this 5960 // during precleaning, then check the overflow list 5961 } 5962 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5963 assert(_collector->overflow_list_is_empty(), 5964 "overflow list was drained above"); 5965 5966 assert(_collector->no_preserved_marks(), 5967 "All preserved marks should have been restored above"); 5968 } 5969 } 5970 5971 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5972 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5973 5974 void MarkRefsIntoAndScanClosure::do_yield_work() { 5975 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5976 "CMS thread should hold CMS token"); 5977 assert_lock_strong(_freelistLock); 5978 assert_lock_strong(_bit_map->lock()); 5979 // relinquish the free_list_lock and bitMaplock() 5980 _bit_map->lock()->unlock(); 5981 _freelistLock->unlock(); 5982 ConcurrentMarkSweepThread::desynchronize(true); 5983 _collector->stopTimer(); 5984 _collector->incrementYields(); 5985 5986 // See the comment in coordinator_yield() 5987 for (unsigned i = 0; 5988 i < CMSYieldSleepCount && 5989 ConcurrentMarkSweepThread::should_yield() && 5990 !CMSCollector::foregroundGCIsActive(); 5991 ++i) { 5992 os::sleep(Thread::current(), 1, false); 5993 } 5994 5995 ConcurrentMarkSweepThread::synchronize(true); 5996 _freelistLock->lock_without_safepoint_check(); 5997 _bit_map->lock()->lock_without_safepoint_check(); 5998 _collector->startTimer(); 5999 } 6000 6001 /////////////////////////////////////////////////////////// 6002 // ParMarkRefsIntoAndScanClosure: a parallel version of 6003 // MarkRefsIntoAndScanClosure 6004 /////////////////////////////////////////////////////////// 6005 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6006 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6007 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6008 _span(span), 6009 _bit_map(bit_map), 6010 _work_queue(work_queue), 6011 _low_water_mark(MIN2((work_queue->max_elems()/4), 6012 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6013 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6014 { 6015 // FIXME: Should initialize in base class constructor. 6016 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6017 set_ref_processor_internal(rp); 6018 } 6019 6020 // This closure is used to mark refs into the CMS generation at the 6021 // second (final) checkpoint, and to scan and transitively follow 6022 // the unmarked oops. The marks are made in the marking bit map and 6023 // the work_queue is used for keeping the (newly) grey objects during 6024 // the scan phase whence they are also available for stealing by parallel 6025 // threads. Since the marking bit map is shared, updates are 6026 // synchronized (via CAS). 6027 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6028 if (obj != NULL) { 6029 // Ignore mark word because this could be an already marked oop 6030 // that may be chained at the end of the overflow list. 6031 assert(oopDesc::is_oop(obj, true), "expected an oop"); 6032 HeapWord* addr = (HeapWord*)obj; 6033 if (_span.contains(addr) && 6034 !_bit_map->isMarked(addr)) { 6035 // mark bit map (object will become grey): 6036 // It is possible for several threads to be 6037 // trying to "claim" this object concurrently; 6038 // the unique thread that succeeds in marking the 6039 // object first will do the subsequent push on 6040 // to the work queue (or overflow list). 6041 if (_bit_map->par_mark(addr)) { 6042 // push on work_queue (which may not be empty), and trim the 6043 // queue to an appropriate length by applying this closure to 6044 // the oops in the oops popped from the stack (i.e. blacken the 6045 // grey objects) 6046 bool res = _work_queue->push(obj); 6047 assert(res, "Low water mark should be less than capacity?"); 6048 trim_queue(_low_water_mark); 6049 } // Else, another thread claimed the object 6050 } 6051 } 6052 } 6053 6054 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6055 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6056 6057 // This closure is used to rescan the marked objects on the dirty cards 6058 // in the mod union table and the card table proper. 6059 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6060 oop p, MemRegion mr) { 6061 6062 size_t size = 0; 6063 HeapWord* addr = (HeapWord*)p; 6064 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6065 assert(_span.contains(addr), "we are scanning the CMS generation"); 6066 // check if it's time to yield 6067 if (do_yield_check()) { 6068 // We yielded for some foreground stop-world work, 6069 // and we have been asked to abort this ongoing preclean cycle. 6070 return 0; 6071 } 6072 if (_bitMap->isMarked(addr)) { 6073 // it's marked; is it potentially uninitialized? 6074 if (p->klass_or_null_acquire() != NULL) { 6075 // an initialized object; ignore mark word in verification below 6076 // since we are running concurrent with mutators 6077 assert(oopDesc::is_oop(p, true), "should be an oop"); 6078 if (p->is_objArray()) { 6079 // objArrays are precisely marked; restrict scanning 6080 // to dirty cards only. 6081 size = CompactibleFreeListSpace::adjustObjectSize( 6082 p->oop_iterate_size(_scanningClosure, mr)); 6083 } else { 6084 // A non-array may have been imprecisely marked; we need 6085 // to scan object in its entirety. 6086 size = CompactibleFreeListSpace::adjustObjectSize( 6087 p->oop_iterate_size(_scanningClosure)); 6088 } 6089 #ifdef ASSERT 6090 size_t direct_size = 6091 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6092 assert(size == direct_size, "Inconsistency in size"); 6093 assert(size >= 3, "Necessary for Printezis marks to work"); 6094 HeapWord* start_pbit = addr + 1; 6095 HeapWord* end_pbit = addr + size - 1; 6096 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6097 "inconsistent Printezis mark"); 6098 // Verify inner mark bits (between Printezis bits) are clear, 6099 // but don't repeat if there are multiple dirty regions for 6100 // the same object, to avoid potential O(N^2) performance. 6101 if (addr != _last_scanned_object) { 6102 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6103 _last_scanned_object = addr; 6104 } 6105 #endif // ASSERT 6106 } else { 6107 // An uninitialized object. 6108 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6109 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6110 size = pointer_delta(nextOneAddr + 1, addr); 6111 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6112 "alignment problem"); 6113 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6114 // will dirty the card when the klass pointer is installed in the 6115 // object (signaling the completion of initialization). 6116 } 6117 } else { 6118 // Either a not yet marked object or an uninitialized object 6119 if (p->klass_or_null_acquire() == NULL) { 6120 // An uninitialized object, skip to the next card, since 6121 // we may not be able to read its P-bits yet. 6122 assert(size == 0, "Initial value"); 6123 } else { 6124 // An object not (yet) reached by marking: we merely need to 6125 // compute its size so as to go look at the next block. 6126 assert(oopDesc::is_oop(p, true), "should be an oop"); 6127 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6128 } 6129 } 6130 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6131 return size; 6132 } 6133 6134 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6135 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6136 "CMS thread should hold CMS token"); 6137 assert_lock_strong(_freelistLock); 6138 assert_lock_strong(_bitMap->lock()); 6139 // relinquish the free_list_lock and bitMaplock() 6140 _bitMap->lock()->unlock(); 6141 _freelistLock->unlock(); 6142 ConcurrentMarkSweepThread::desynchronize(true); 6143 _collector->stopTimer(); 6144 _collector->incrementYields(); 6145 6146 // See the comment in coordinator_yield() 6147 for (unsigned i = 0; i < CMSYieldSleepCount && 6148 ConcurrentMarkSweepThread::should_yield() && 6149 !CMSCollector::foregroundGCIsActive(); ++i) { 6150 os::sleep(Thread::current(), 1, false); 6151 } 6152 6153 ConcurrentMarkSweepThread::synchronize(true); 6154 _freelistLock->lock_without_safepoint_check(); 6155 _bitMap->lock()->lock_without_safepoint_check(); 6156 _collector->startTimer(); 6157 } 6158 6159 6160 ////////////////////////////////////////////////////////////////// 6161 // SurvivorSpacePrecleanClosure 6162 ////////////////////////////////////////////////////////////////// 6163 // This (single-threaded) closure is used to preclean the oops in 6164 // the survivor spaces. 6165 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6166 6167 HeapWord* addr = (HeapWord*)p; 6168 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6169 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6170 assert(p->klass_or_null() != NULL, "object should be initialized"); 6171 // an initialized object; ignore mark word in verification below 6172 // since we are running concurrent with mutators 6173 assert(oopDesc::is_oop(p, true), "should be an oop"); 6174 // Note that we do not yield while we iterate over 6175 // the interior oops of p, pushing the relevant ones 6176 // on our marking stack. 6177 size_t size = p->oop_iterate_size(_scanning_closure); 6178 do_yield_check(); 6179 // Observe that below, we do not abandon the preclean 6180 // phase as soon as we should; rather we empty the 6181 // marking stack before returning. This is to satisfy 6182 // some existing assertions. In general, it may be a 6183 // good idea to abort immediately and complete the marking 6184 // from the grey objects at a later time. 6185 while (!_mark_stack->isEmpty()) { 6186 oop new_oop = _mark_stack->pop(); 6187 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6188 assert(_bit_map->isMarked((HeapWord*)new_oop), 6189 "only grey objects on this stack"); 6190 // iterate over the oops in this oop, marking and pushing 6191 // the ones in CMS heap (i.e. in _span). 6192 new_oop->oop_iterate(_scanning_closure); 6193 // check if it's time to yield 6194 do_yield_check(); 6195 } 6196 unsigned int after_count = 6197 CMSHeap::heap()->total_collections(); 6198 bool abort = (_before_count != after_count) || 6199 _collector->should_abort_preclean(); 6200 return abort ? 0 : size; 6201 } 6202 6203 void SurvivorSpacePrecleanClosure::do_yield_work() { 6204 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6205 "CMS thread should hold CMS token"); 6206 assert_lock_strong(_bit_map->lock()); 6207 // Relinquish the bit map lock 6208 _bit_map->lock()->unlock(); 6209 ConcurrentMarkSweepThread::desynchronize(true); 6210 _collector->stopTimer(); 6211 _collector->incrementYields(); 6212 6213 // See the comment in coordinator_yield() 6214 for (unsigned i = 0; i < CMSYieldSleepCount && 6215 ConcurrentMarkSweepThread::should_yield() && 6216 !CMSCollector::foregroundGCIsActive(); ++i) { 6217 os::sleep(Thread::current(), 1, false); 6218 } 6219 6220 ConcurrentMarkSweepThread::synchronize(true); 6221 _bit_map->lock()->lock_without_safepoint_check(); 6222 _collector->startTimer(); 6223 } 6224 6225 // This closure is used to rescan the marked objects on the dirty cards 6226 // in the mod union table and the card table proper. In the parallel 6227 // case, although the bitMap is shared, we do a single read so the 6228 // isMarked() query is "safe". 6229 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6230 // Ignore mark word because we are running concurrent with mutators 6231 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6232 HeapWord* addr = (HeapWord*)p; 6233 assert(_span.contains(addr), "we are scanning the CMS generation"); 6234 bool is_obj_array = false; 6235 #ifdef ASSERT 6236 if (!_parallel) { 6237 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6238 assert(_collector->overflow_list_is_empty(), 6239 "overflow list should be empty"); 6240 6241 } 6242 #endif // ASSERT 6243 if (_bit_map->isMarked(addr)) { 6244 // Obj arrays are precisely marked, non-arrays are not; 6245 // so we scan objArrays precisely and non-arrays in their 6246 // entirety. 6247 if (p->is_objArray()) { 6248 is_obj_array = true; 6249 if (_parallel) { 6250 p->oop_iterate(_par_scan_closure, mr); 6251 } else { 6252 p->oop_iterate(_scan_closure, mr); 6253 } 6254 } else { 6255 if (_parallel) { 6256 p->oop_iterate(_par_scan_closure); 6257 } else { 6258 p->oop_iterate(_scan_closure); 6259 } 6260 } 6261 } 6262 #ifdef ASSERT 6263 if (!_parallel) { 6264 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6265 assert(_collector->overflow_list_is_empty(), 6266 "overflow list should be empty"); 6267 6268 } 6269 #endif // ASSERT 6270 return is_obj_array; 6271 } 6272 6273 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6274 MemRegion span, 6275 CMSBitMap* bitMap, CMSMarkStack* markStack, 6276 bool should_yield, bool verifying): 6277 _collector(collector), 6278 _span(span), 6279 _bitMap(bitMap), 6280 _mut(&collector->_modUnionTable), 6281 _markStack(markStack), 6282 _yield(should_yield), 6283 _skipBits(0) 6284 { 6285 assert(_markStack->isEmpty(), "stack should be empty"); 6286 _finger = _bitMap->startWord(); 6287 _threshold = _finger; 6288 assert(_collector->_restart_addr == NULL, "Sanity check"); 6289 assert(_span.contains(_finger), "Out of bounds _finger?"); 6290 DEBUG_ONLY(_verifying = verifying;) 6291 } 6292 6293 void MarkFromRootsClosure::reset(HeapWord* addr) { 6294 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6295 assert(_span.contains(addr), "Out of bounds _finger?"); 6296 _finger = addr; 6297 _threshold = align_up(_finger, CardTable::card_size); 6298 } 6299 6300 // Should revisit to see if this should be restructured for 6301 // greater efficiency. 6302 bool MarkFromRootsClosure::do_bit(size_t offset) { 6303 if (_skipBits > 0) { 6304 _skipBits--; 6305 return true; 6306 } 6307 // convert offset into a HeapWord* 6308 HeapWord* addr = _bitMap->startWord() + offset; 6309 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6310 "address out of range"); 6311 assert(_bitMap->isMarked(addr), "tautology"); 6312 if (_bitMap->isMarked(addr+1)) { 6313 // this is an allocated but not yet initialized object 6314 assert(_skipBits == 0, "tautology"); 6315 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6316 oop p = oop(addr); 6317 if (p->klass_or_null_acquire() == NULL) { 6318 DEBUG_ONLY(if (!_verifying) {) 6319 // We re-dirty the cards on which this object lies and increase 6320 // the _threshold so that we'll come back to scan this object 6321 // during the preclean or remark phase. (CMSCleanOnEnter) 6322 if (CMSCleanOnEnter) { 6323 size_t sz = _collector->block_size_using_printezis_bits(addr); 6324 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6325 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6326 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6327 // Bump _threshold to end_card_addr; note that 6328 // _threshold cannot possibly exceed end_card_addr, anyhow. 6329 // This prevents future clearing of the card as the scan proceeds 6330 // to the right. 6331 assert(_threshold <= end_card_addr, 6332 "Because we are just scanning into this object"); 6333 if (_threshold < end_card_addr) { 6334 _threshold = end_card_addr; 6335 } 6336 if (p->klass_or_null_acquire() != NULL) { 6337 // Redirty the range of cards... 6338 _mut->mark_range(redirty_range); 6339 } // ...else the setting of klass will dirty the card anyway. 6340 } 6341 DEBUG_ONLY(}) 6342 return true; 6343 } 6344 } 6345 scanOopsInOop(addr); 6346 return true; 6347 } 6348 6349 // We take a break if we've been at this for a while, 6350 // so as to avoid monopolizing the locks involved. 6351 void MarkFromRootsClosure::do_yield_work() { 6352 // First give up the locks, then yield, then re-lock 6353 // We should probably use a constructor/destructor idiom to 6354 // do this unlock/lock or modify the MutexUnlocker class to 6355 // serve our purpose. XXX 6356 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6357 "CMS thread should hold CMS token"); 6358 assert_lock_strong(_bitMap->lock()); 6359 _bitMap->lock()->unlock(); 6360 ConcurrentMarkSweepThread::desynchronize(true); 6361 _collector->stopTimer(); 6362 _collector->incrementYields(); 6363 6364 // See the comment in coordinator_yield() 6365 for (unsigned i = 0; i < CMSYieldSleepCount && 6366 ConcurrentMarkSweepThread::should_yield() && 6367 !CMSCollector::foregroundGCIsActive(); ++i) { 6368 os::sleep(Thread::current(), 1, false); 6369 } 6370 6371 ConcurrentMarkSweepThread::synchronize(true); 6372 _bitMap->lock()->lock_without_safepoint_check(); 6373 _collector->startTimer(); 6374 } 6375 6376 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6377 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6378 assert(_markStack->isEmpty(), 6379 "should drain stack to limit stack usage"); 6380 // convert ptr to an oop preparatory to scanning 6381 oop obj = oop(ptr); 6382 // Ignore mark word in verification below, since we 6383 // may be running concurrent with mutators. 6384 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6385 assert(_finger <= ptr, "_finger runneth ahead"); 6386 // advance the finger to right end of this object 6387 _finger = ptr + obj->size(); 6388 assert(_finger > ptr, "we just incremented it above"); 6389 // On large heaps, it may take us some time to get through 6390 // the marking phase. During 6391 // this time it's possible that a lot of mutations have 6392 // accumulated in the card table and the mod union table -- 6393 // these mutation records are redundant until we have 6394 // actually traced into the corresponding card. 6395 // Here, we check whether advancing the finger would make 6396 // us cross into a new card, and if so clear corresponding 6397 // cards in the MUT (preclean them in the card-table in the 6398 // future). 6399 6400 DEBUG_ONLY(if (!_verifying) {) 6401 // The clean-on-enter optimization is disabled by default, 6402 // until we fix 6178663. 6403 if (CMSCleanOnEnter && (_finger > _threshold)) { 6404 // [_threshold, _finger) represents the interval 6405 // of cards to be cleared in MUT (or precleaned in card table). 6406 // The set of cards to be cleared is all those that overlap 6407 // with the interval [_threshold, _finger); note that 6408 // _threshold is always kept card-aligned but _finger isn't 6409 // always card-aligned. 6410 HeapWord* old_threshold = _threshold; 6411 assert(is_aligned(old_threshold, CardTable::card_size), 6412 "_threshold should always be card-aligned"); 6413 _threshold = align_up(_finger, CardTable::card_size); 6414 MemRegion mr(old_threshold, _threshold); 6415 assert(!mr.is_empty(), "Control point invariant"); 6416 assert(_span.contains(mr), "Should clear within span"); 6417 _mut->clear_range(mr); 6418 } 6419 DEBUG_ONLY(}) 6420 // Note: the finger doesn't advance while we drain 6421 // the stack below. 6422 PushOrMarkClosure pushOrMarkClosure(_collector, 6423 _span, _bitMap, _markStack, 6424 _finger, this); 6425 bool res = _markStack->push(obj); 6426 assert(res, "Empty non-zero size stack should have space for single push"); 6427 while (!_markStack->isEmpty()) { 6428 oop new_oop = _markStack->pop(); 6429 // Skip verifying header mark word below because we are 6430 // running concurrent with mutators. 6431 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6432 // now scan this oop's oops 6433 new_oop->oop_iterate(&pushOrMarkClosure); 6434 do_yield_check(); 6435 } 6436 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6437 } 6438 6439 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6440 CMSCollector* collector, MemRegion span, 6441 CMSBitMap* bit_map, 6442 OopTaskQueue* work_queue, 6443 CMSMarkStack* overflow_stack): 6444 _collector(collector), 6445 _whole_span(collector->_span), 6446 _span(span), 6447 _bit_map(bit_map), 6448 _mut(&collector->_modUnionTable), 6449 _work_queue(work_queue), 6450 _overflow_stack(overflow_stack), 6451 _skip_bits(0), 6452 _task(task) 6453 { 6454 assert(_work_queue->size() == 0, "work_queue should be empty"); 6455 _finger = span.start(); 6456 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6457 assert(_span.contains(_finger), "Out of bounds _finger?"); 6458 } 6459 6460 // Should revisit to see if this should be restructured for 6461 // greater efficiency. 6462 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6463 if (_skip_bits > 0) { 6464 _skip_bits--; 6465 return true; 6466 } 6467 // convert offset into a HeapWord* 6468 HeapWord* addr = _bit_map->startWord() + offset; 6469 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6470 "address out of range"); 6471 assert(_bit_map->isMarked(addr), "tautology"); 6472 if (_bit_map->isMarked(addr+1)) { 6473 // this is an allocated object that might not yet be initialized 6474 assert(_skip_bits == 0, "tautology"); 6475 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6476 oop p = oop(addr); 6477 if (p->klass_or_null_acquire() == NULL) { 6478 // in the case of Clean-on-Enter optimization, redirty card 6479 // and avoid clearing card by increasing the threshold. 6480 return true; 6481 } 6482 } 6483 scan_oops_in_oop(addr); 6484 return true; 6485 } 6486 6487 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6488 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6489 // Should we assert that our work queue is empty or 6490 // below some drain limit? 6491 assert(_work_queue->size() == 0, 6492 "should drain stack to limit stack usage"); 6493 // convert ptr to an oop preparatory to scanning 6494 oop obj = oop(ptr); 6495 // Ignore mark word in verification below, since we 6496 // may be running concurrent with mutators. 6497 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6498 assert(_finger <= ptr, "_finger runneth ahead"); 6499 // advance the finger to right end of this object 6500 _finger = ptr + obj->size(); 6501 assert(_finger > ptr, "we just incremented it above"); 6502 // On large heaps, it may take us some time to get through 6503 // the marking phase. During 6504 // this time it's possible that a lot of mutations have 6505 // accumulated in the card table and the mod union table -- 6506 // these mutation records are redundant until we have 6507 // actually traced into the corresponding card. 6508 // Here, we check whether advancing the finger would make 6509 // us cross into a new card, and if so clear corresponding 6510 // cards in the MUT (preclean them in the card-table in the 6511 // future). 6512 6513 // The clean-on-enter optimization is disabled by default, 6514 // until we fix 6178663. 6515 if (CMSCleanOnEnter && (_finger > _threshold)) { 6516 // [_threshold, _finger) represents the interval 6517 // of cards to be cleared in MUT (or precleaned in card table). 6518 // The set of cards to be cleared is all those that overlap 6519 // with the interval [_threshold, _finger); note that 6520 // _threshold is always kept card-aligned but _finger isn't 6521 // always card-aligned. 6522 HeapWord* old_threshold = _threshold; 6523 assert(is_aligned(old_threshold, CardTable::card_size), 6524 "_threshold should always be card-aligned"); 6525 _threshold = align_up(_finger, CardTable::card_size); 6526 MemRegion mr(old_threshold, _threshold); 6527 assert(!mr.is_empty(), "Control point invariant"); 6528 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6529 _mut->clear_range(mr); 6530 } 6531 6532 // Note: the local finger doesn't advance while we drain 6533 // the stack below, but the global finger sure can and will. 6534 HeapWord* volatile* gfa = _task->global_finger_addr(); 6535 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6536 _span, _bit_map, 6537 _work_queue, 6538 _overflow_stack, 6539 _finger, 6540 gfa, this); 6541 bool res = _work_queue->push(obj); // overflow could occur here 6542 assert(res, "Will hold once we use workqueues"); 6543 while (true) { 6544 oop new_oop; 6545 if (!_work_queue->pop_local(new_oop)) { 6546 // We emptied our work_queue; check if there's stuff that can 6547 // be gotten from the overflow stack. 6548 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6549 _overflow_stack, _work_queue)) { 6550 do_yield_check(); 6551 continue; 6552 } else { // done 6553 break; 6554 } 6555 } 6556 // Skip verifying header mark word below because we are 6557 // running concurrent with mutators. 6558 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6559 // now scan this oop's oops 6560 new_oop->oop_iterate(&pushOrMarkClosure); 6561 do_yield_check(); 6562 } 6563 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6564 } 6565 6566 // Yield in response to a request from VM Thread or 6567 // from mutators. 6568 void ParMarkFromRootsClosure::do_yield_work() { 6569 assert(_task != NULL, "sanity"); 6570 _task->yield(); 6571 } 6572 6573 // A variant of the above used for verifying CMS marking work. 6574 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6575 MemRegion span, 6576 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6577 CMSMarkStack* mark_stack): 6578 _collector(collector), 6579 _span(span), 6580 _verification_bm(verification_bm), 6581 _cms_bm(cms_bm), 6582 _mark_stack(mark_stack), 6583 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6584 mark_stack) 6585 { 6586 assert(_mark_stack->isEmpty(), "stack should be empty"); 6587 _finger = _verification_bm->startWord(); 6588 assert(_collector->_restart_addr == NULL, "Sanity check"); 6589 assert(_span.contains(_finger), "Out of bounds _finger?"); 6590 } 6591 6592 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6593 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6594 assert(_span.contains(addr), "Out of bounds _finger?"); 6595 _finger = addr; 6596 } 6597 6598 // Should revisit to see if this should be restructured for 6599 // greater efficiency. 6600 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6601 // convert offset into a HeapWord* 6602 HeapWord* addr = _verification_bm->startWord() + offset; 6603 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6604 "address out of range"); 6605 assert(_verification_bm->isMarked(addr), "tautology"); 6606 assert(_cms_bm->isMarked(addr), "tautology"); 6607 6608 assert(_mark_stack->isEmpty(), 6609 "should drain stack to limit stack usage"); 6610 // convert addr to an oop preparatory to scanning 6611 oop obj = oop(addr); 6612 assert(oopDesc::is_oop(obj), "should be an oop"); 6613 assert(_finger <= addr, "_finger runneth ahead"); 6614 // advance the finger to right end of this object 6615 _finger = addr + obj->size(); 6616 assert(_finger > addr, "we just incremented it above"); 6617 // Note: the finger doesn't advance while we drain 6618 // the stack below. 6619 bool res = _mark_stack->push(obj); 6620 assert(res, "Empty non-zero size stack should have space for single push"); 6621 while (!_mark_stack->isEmpty()) { 6622 oop new_oop = _mark_stack->pop(); 6623 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6624 // now scan this oop's oops 6625 new_oop->oop_iterate(&_pam_verify_closure); 6626 } 6627 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6628 return true; 6629 } 6630 6631 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6632 CMSCollector* collector, MemRegion span, 6633 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6634 CMSMarkStack* mark_stack): 6635 MetadataAwareOopClosure(collector->ref_processor()), 6636 _collector(collector), 6637 _span(span), 6638 _verification_bm(verification_bm), 6639 _cms_bm(cms_bm), 6640 _mark_stack(mark_stack) 6641 { } 6642 6643 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6644 oop obj = RawAccess<>::oop_load(p); 6645 do_oop(obj); 6646 } 6647 6648 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6649 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6650 6651 // Upon stack overflow, we discard (part of) the stack, 6652 // remembering the least address amongst those discarded 6653 // in CMSCollector's _restart_address. 6654 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6655 // Remember the least grey address discarded 6656 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6657 _collector->lower_restart_addr(ra); 6658 _mark_stack->reset(); // discard stack contents 6659 _mark_stack->expand(); // expand the stack if possible 6660 } 6661 6662 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6663 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6664 HeapWord* addr = (HeapWord*)obj; 6665 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6666 // Oop lies in _span and isn't yet grey or black 6667 _verification_bm->mark(addr); // now grey 6668 if (!_cms_bm->isMarked(addr)) { 6669 Log(gc, verify) log; 6670 ResourceMark rm; 6671 LogStream ls(log.error()); 6672 oop(addr)->print_on(&ls); 6673 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6674 fatal("... aborting"); 6675 } 6676 6677 if (!_mark_stack->push(obj)) { // stack overflow 6678 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6679 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6680 handle_stack_overflow(addr); 6681 } 6682 // anything including and to the right of _finger 6683 // will be scanned as we iterate over the remainder of the 6684 // bit map 6685 } 6686 } 6687 6688 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6689 MemRegion span, 6690 CMSBitMap* bitMap, CMSMarkStack* markStack, 6691 HeapWord* finger, MarkFromRootsClosure* parent) : 6692 MetadataAwareOopClosure(collector->ref_processor()), 6693 _collector(collector), 6694 _span(span), 6695 _bitMap(bitMap), 6696 _markStack(markStack), 6697 _finger(finger), 6698 _parent(parent) 6699 { } 6700 6701 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6702 MemRegion span, 6703 CMSBitMap* bit_map, 6704 OopTaskQueue* work_queue, 6705 CMSMarkStack* overflow_stack, 6706 HeapWord* finger, 6707 HeapWord* volatile* global_finger_addr, 6708 ParMarkFromRootsClosure* parent) : 6709 MetadataAwareOopClosure(collector->ref_processor()), 6710 _collector(collector), 6711 _whole_span(collector->_span), 6712 _span(span), 6713 _bit_map(bit_map), 6714 _work_queue(work_queue), 6715 _overflow_stack(overflow_stack), 6716 _finger(finger), 6717 _global_finger_addr(global_finger_addr), 6718 _parent(parent) 6719 { } 6720 6721 // Assumes thread-safe access by callers, who are 6722 // responsible for mutual exclusion. 6723 void CMSCollector::lower_restart_addr(HeapWord* low) { 6724 assert(_span.contains(low), "Out of bounds addr"); 6725 if (_restart_addr == NULL) { 6726 _restart_addr = low; 6727 } else { 6728 _restart_addr = MIN2(_restart_addr, low); 6729 } 6730 } 6731 6732 // Upon stack overflow, we discard (part of) the stack, 6733 // remembering the least address amongst those discarded 6734 // in CMSCollector's _restart_address. 6735 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6736 // Remember the least grey address discarded 6737 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6738 _collector->lower_restart_addr(ra); 6739 _markStack->reset(); // discard stack contents 6740 _markStack->expand(); // expand the stack if possible 6741 } 6742 6743 // Upon stack overflow, we discard (part of) the stack, 6744 // remembering the least address amongst those discarded 6745 // in CMSCollector's _restart_address. 6746 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6747 // We need to do this under a mutex to prevent other 6748 // workers from interfering with the work done below. 6749 MutexLockerEx ml(_overflow_stack->par_lock(), 6750 Mutex::_no_safepoint_check_flag); 6751 // Remember the least grey address discarded 6752 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6753 _collector->lower_restart_addr(ra); 6754 _overflow_stack->reset(); // discard stack contents 6755 _overflow_stack->expand(); // expand the stack if possible 6756 } 6757 6758 void PushOrMarkClosure::do_oop(oop obj) { 6759 // Ignore mark word because we are running concurrent with mutators. 6760 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6761 HeapWord* addr = (HeapWord*)obj; 6762 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6763 // Oop lies in _span and isn't yet grey or black 6764 _bitMap->mark(addr); // now grey 6765 if (addr < _finger) { 6766 // the bit map iteration has already either passed, or 6767 // sampled, this bit in the bit map; we'll need to 6768 // use the marking stack to scan this oop's oops. 6769 bool simulate_overflow = false; 6770 NOT_PRODUCT( 6771 if (CMSMarkStackOverflowALot && 6772 _collector->simulate_overflow()) { 6773 // simulate a stack overflow 6774 simulate_overflow = true; 6775 } 6776 ) 6777 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6778 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6779 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6780 handle_stack_overflow(addr); 6781 } 6782 } 6783 // anything including and to the right of _finger 6784 // will be scanned as we iterate over the remainder of the 6785 // bit map 6786 do_yield_check(); 6787 } 6788 } 6789 6790 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6791 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6792 6793 void ParPushOrMarkClosure::do_oop(oop obj) { 6794 // Ignore mark word because we are running concurrent with mutators. 6795 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6796 HeapWord* addr = (HeapWord*)obj; 6797 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6798 // Oop lies in _span and isn't yet grey or black 6799 // We read the global_finger (volatile read) strictly after marking oop 6800 bool res = _bit_map->par_mark(addr); // now grey 6801 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6802 // Should we push this marked oop on our stack? 6803 // -- if someone else marked it, nothing to do 6804 // -- if target oop is above global finger nothing to do 6805 // -- if target oop is in chunk and above local finger 6806 // then nothing to do 6807 // -- else push on work queue 6808 if ( !res // someone else marked it, they will deal with it 6809 || (addr >= *gfa) // will be scanned in a later task 6810 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6811 return; 6812 } 6813 // the bit map iteration has already either passed, or 6814 // sampled, this bit in the bit map; we'll need to 6815 // use the marking stack to scan this oop's oops. 6816 bool simulate_overflow = false; 6817 NOT_PRODUCT( 6818 if (CMSMarkStackOverflowALot && 6819 _collector->simulate_overflow()) { 6820 // simulate a stack overflow 6821 simulate_overflow = true; 6822 } 6823 ) 6824 if (simulate_overflow || 6825 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6826 // stack overflow 6827 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6828 // We cannot assert that the overflow stack is full because 6829 // it may have been emptied since. 6830 assert(simulate_overflow || 6831 _work_queue->size() == _work_queue->max_elems(), 6832 "Else push should have succeeded"); 6833 handle_stack_overflow(addr); 6834 } 6835 do_yield_check(); 6836 } 6837 } 6838 6839 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6840 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6841 6842 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6843 MemRegion span, 6844 ReferenceProcessor* rp, 6845 CMSBitMap* bit_map, 6846 CMSBitMap* mod_union_table, 6847 CMSMarkStack* mark_stack, 6848 bool concurrent_precleaning): 6849 MetadataAwareOopClosure(rp), 6850 _collector(collector), 6851 _span(span), 6852 _bit_map(bit_map), 6853 _mod_union_table(mod_union_table), 6854 _mark_stack(mark_stack), 6855 _concurrent_precleaning(concurrent_precleaning) 6856 { 6857 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6858 } 6859 6860 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6861 // the non-parallel version (the parallel version appears further below.) 6862 void PushAndMarkClosure::do_oop(oop obj) { 6863 // Ignore mark word verification. If during concurrent precleaning, 6864 // the object monitor may be locked. If during the checkpoint 6865 // phases, the object may already have been reached by a different 6866 // path and may be at the end of the global overflow list (so 6867 // the mark word may be NULL). 6868 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6869 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6870 HeapWord* addr = (HeapWord*)obj; 6871 // Check if oop points into the CMS generation 6872 // and is not marked 6873 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6874 // a white object ... 6875 _bit_map->mark(addr); // ... now grey 6876 // push on the marking stack (grey set) 6877 bool simulate_overflow = false; 6878 NOT_PRODUCT( 6879 if (CMSMarkStackOverflowALot && 6880 _collector->simulate_overflow()) { 6881 // simulate a stack overflow 6882 simulate_overflow = true; 6883 } 6884 ) 6885 if (simulate_overflow || !_mark_stack->push(obj)) { 6886 if (_concurrent_precleaning) { 6887 // During precleaning we can just dirty the appropriate card(s) 6888 // in the mod union table, thus ensuring that the object remains 6889 // in the grey set and continue. In the case of object arrays 6890 // we need to dirty all of the cards that the object spans, 6891 // since the rescan of object arrays will be limited to the 6892 // dirty cards. 6893 // Note that no one can be interfering with us in this action 6894 // of dirtying the mod union table, so no locking or atomics 6895 // are required. 6896 if (obj->is_objArray()) { 6897 size_t sz = obj->size(); 6898 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6899 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6900 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6901 _mod_union_table->mark_range(redirty_range); 6902 } else { 6903 _mod_union_table->mark(addr); 6904 } 6905 _collector->_ser_pmc_preclean_ovflw++; 6906 } else { 6907 // During the remark phase, we need to remember this oop 6908 // in the overflow list. 6909 _collector->push_on_overflow_list(obj); 6910 _collector->_ser_pmc_remark_ovflw++; 6911 } 6912 } 6913 } 6914 } 6915 6916 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6917 MemRegion span, 6918 ReferenceProcessor* rp, 6919 CMSBitMap* bit_map, 6920 OopTaskQueue* work_queue): 6921 MetadataAwareOopClosure(rp), 6922 _collector(collector), 6923 _span(span), 6924 _bit_map(bit_map), 6925 _work_queue(work_queue) 6926 { 6927 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6928 } 6929 6930 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6931 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6932 6933 // Grey object rescan during second checkpoint phase -- 6934 // the parallel version. 6935 void ParPushAndMarkClosure::do_oop(oop obj) { 6936 // In the assert below, we ignore the mark word because 6937 // this oop may point to an already visited object that is 6938 // on the overflow stack (in which case the mark word has 6939 // been hijacked for chaining into the overflow stack -- 6940 // if this is the last object in the overflow stack then 6941 // its mark word will be NULL). Because this object may 6942 // have been subsequently popped off the global overflow 6943 // stack, and the mark word possibly restored to the prototypical 6944 // value, by the time we get to examined this failing assert in 6945 // the debugger, is_oop_or_null(false) may subsequently start 6946 // to hold. 6947 assert(oopDesc::is_oop_or_null(obj, true), 6948 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6949 HeapWord* addr = (HeapWord*)obj; 6950 // Check if oop points into the CMS generation 6951 // and is not marked 6952 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6953 // a white object ... 6954 // If we manage to "claim" the object, by being the 6955 // first thread to mark it, then we push it on our 6956 // marking stack 6957 if (_bit_map->par_mark(addr)) { // ... now grey 6958 // push on work queue (grey set) 6959 bool simulate_overflow = false; 6960 NOT_PRODUCT( 6961 if (CMSMarkStackOverflowALot && 6962 _collector->par_simulate_overflow()) { 6963 // simulate a stack overflow 6964 simulate_overflow = true; 6965 } 6966 ) 6967 if (simulate_overflow || !_work_queue->push(obj)) { 6968 _collector->par_push_on_overflow_list(obj); 6969 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6970 } 6971 } // Else, some other thread got there first 6972 } 6973 } 6974 6975 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6976 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6977 6978 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6979 Mutex* bml = _collector->bitMapLock(); 6980 assert_lock_strong(bml); 6981 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6982 "CMS thread should hold CMS token"); 6983 6984 bml->unlock(); 6985 ConcurrentMarkSweepThread::desynchronize(true); 6986 6987 _collector->stopTimer(); 6988 _collector->incrementYields(); 6989 6990 // See the comment in coordinator_yield() 6991 for (unsigned i = 0; i < CMSYieldSleepCount && 6992 ConcurrentMarkSweepThread::should_yield() && 6993 !CMSCollector::foregroundGCIsActive(); ++i) { 6994 os::sleep(Thread::current(), 1, false); 6995 } 6996 6997 ConcurrentMarkSweepThread::synchronize(true); 6998 bml->lock(); 6999 7000 _collector->startTimer(); 7001 } 7002 7003 bool CMSPrecleanRefsYieldClosure::should_return() { 7004 if (ConcurrentMarkSweepThread::should_yield()) { 7005 do_yield_work(); 7006 } 7007 return _collector->foregroundGCIsActive(); 7008 } 7009 7010 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7011 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7012 "mr should be aligned to start at a card boundary"); 7013 // We'd like to assert: 7014 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7015 // "mr should be a range of cards"); 7016 // However, that would be too strong in one case -- the last 7017 // partition ends at _unallocated_block which, in general, can be 7018 // an arbitrary boundary, not necessarily card aligned. 7019 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7020 _space->object_iterate_mem(mr, &_scan_cl); 7021 } 7022 7023 SweepClosure::SweepClosure(CMSCollector* collector, 7024 ConcurrentMarkSweepGeneration* g, 7025 CMSBitMap* bitMap, bool should_yield) : 7026 _collector(collector), 7027 _g(g), 7028 _sp(g->cmsSpace()), 7029 _limit(_sp->sweep_limit()), 7030 _freelistLock(_sp->freelistLock()), 7031 _bitMap(bitMap), 7032 _yield(should_yield), 7033 _inFreeRange(false), // No free range at beginning of sweep 7034 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7035 _lastFreeRangeCoalesced(false), 7036 _freeFinger(g->used_region().start()) 7037 { 7038 NOT_PRODUCT( 7039 _numObjectsFreed = 0; 7040 _numWordsFreed = 0; 7041 _numObjectsLive = 0; 7042 _numWordsLive = 0; 7043 _numObjectsAlreadyFree = 0; 7044 _numWordsAlreadyFree = 0; 7045 _last_fc = NULL; 7046 7047 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7048 _sp->dictionary()->initialize_dict_returned_bytes(); 7049 ) 7050 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7051 "sweep _limit out of bounds"); 7052 log_develop_trace(gc, sweep)("===================="); 7053 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7054 } 7055 7056 void SweepClosure::print_on(outputStream* st) const { 7057 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7058 p2i(_sp->bottom()), p2i(_sp->end())); 7059 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7060 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7061 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7062 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7063 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7064 } 7065 7066 #ifndef PRODUCT 7067 // Assertion checking only: no useful work in product mode -- 7068 // however, if any of the flags below become product flags, 7069 // you may need to review this code to see if it needs to be 7070 // enabled in product mode. 7071 SweepClosure::~SweepClosure() { 7072 assert_lock_strong(_freelistLock); 7073 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7074 "sweep _limit out of bounds"); 7075 if (inFreeRange()) { 7076 Log(gc, sweep) log; 7077 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7078 ResourceMark rm; 7079 LogStream ls(log.error()); 7080 print_on(&ls); 7081 ShouldNotReachHere(); 7082 } 7083 7084 if (log_is_enabled(Debug, gc, sweep)) { 7085 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7086 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7087 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7088 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7089 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7090 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7091 } 7092 7093 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7094 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7095 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7096 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7097 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7098 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7099 } 7100 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7101 log_develop_trace(gc, sweep)("================"); 7102 } 7103 #endif // PRODUCT 7104 7105 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7106 bool freeRangeInFreeLists) { 7107 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7108 p2i(freeFinger), freeRangeInFreeLists); 7109 assert(!inFreeRange(), "Trampling existing free range"); 7110 set_inFreeRange(true); 7111 set_lastFreeRangeCoalesced(false); 7112 7113 set_freeFinger(freeFinger); 7114 set_freeRangeInFreeLists(freeRangeInFreeLists); 7115 if (CMSTestInFreeList) { 7116 if (freeRangeInFreeLists) { 7117 FreeChunk* fc = (FreeChunk*) freeFinger; 7118 assert(fc->is_free(), "A chunk on the free list should be free."); 7119 assert(fc->size() > 0, "Free range should have a size"); 7120 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7121 } 7122 } 7123 } 7124 7125 // Note that the sweeper runs concurrently with mutators. Thus, 7126 // it is possible for direct allocation in this generation to happen 7127 // in the middle of the sweep. Note that the sweeper also coalesces 7128 // contiguous free blocks. Thus, unless the sweeper and the allocator 7129 // synchronize appropriately freshly allocated blocks may get swept up. 7130 // This is accomplished by the sweeper locking the free lists while 7131 // it is sweeping. Thus blocks that are determined to be free are 7132 // indeed free. There is however one additional complication: 7133 // blocks that have been allocated since the final checkpoint and 7134 // mark, will not have been marked and so would be treated as 7135 // unreachable and swept up. To prevent this, the allocator marks 7136 // the bit map when allocating during the sweep phase. This leads, 7137 // however, to a further complication -- objects may have been allocated 7138 // but not yet initialized -- in the sense that the header isn't yet 7139 // installed. The sweeper can not then determine the size of the block 7140 // in order to skip over it. To deal with this case, we use a technique 7141 // (due to Printezis) to encode such uninitialized block sizes in the 7142 // bit map. Since the bit map uses a bit per every HeapWord, but the 7143 // CMS generation has a minimum object size of 3 HeapWords, it follows 7144 // that "normal marks" won't be adjacent in the bit map (there will 7145 // always be at least two 0 bits between successive 1 bits). We make use 7146 // of these "unused" bits to represent uninitialized blocks -- the bit 7147 // corresponding to the start of the uninitialized object and the next 7148 // bit are both set. Finally, a 1 bit marks the end of the object that 7149 // started with the two consecutive 1 bits to indicate its potentially 7150 // uninitialized state. 7151 7152 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7153 FreeChunk* fc = (FreeChunk*)addr; 7154 size_t res; 7155 7156 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7157 // than "addr == _limit" because although _limit was a block boundary when 7158 // we started the sweep, it may no longer be one because heap expansion 7159 // may have caused us to coalesce the block ending at the address _limit 7160 // with a newly expanded chunk (this happens when _limit was set to the 7161 // previous _end of the space), so we may have stepped past _limit: 7162 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7163 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7164 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7165 "sweep _limit out of bounds"); 7166 assert(addr < _sp->end(), "addr out of bounds"); 7167 // Flush any free range we might be holding as a single 7168 // coalesced chunk to the appropriate free list. 7169 if (inFreeRange()) { 7170 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7171 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7172 flush_cur_free_chunk(freeFinger(), 7173 pointer_delta(addr, freeFinger())); 7174 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7175 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7176 lastFreeRangeCoalesced() ? 1 : 0); 7177 } 7178 7179 // help the iterator loop finish 7180 return pointer_delta(_sp->end(), addr); 7181 } 7182 7183 assert(addr < _limit, "sweep invariant"); 7184 // check if we should yield 7185 do_yield_check(addr); 7186 if (fc->is_free()) { 7187 // Chunk that is already free 7188 res = fc->size(); 7189 do_already_free_chunk(fc); 7190 debug_only(_sp->verifyFreeLists()); 7191 // If we flush the chunk at hand in lookahead_and_flush() 7192 // and it's coalesced with a preceding chunk, then the 7193 // process of "mangling" the payload of the coalesced block 7194 // will cause erasure of the size information from the 7195 // (erstwhile) header of all the coalesced blocks but the 7196 // first, so the first disjunct in the assert will not hold 7197 // in that specific case (in which case the second disjunct 7198 // will hold). 7199 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7200 "Otherwise the size info doesn't change at this step"); 7201 NOT_PRODUCT( 7202 _numObjectsAlreadyFree++; 7203 _numWordsAlreadyFree += res; 7204 ) 7205 NOT_PRODUCT(_last_fc = fc;) 7206 } else if (!_bitMap->isMarked(addr)) { 7207 // Chunk is fresh garbage 7208 res = do_garbage_chunk(fc); 7209 debug_only(_sp->verifyFreeLists()); 7210 NOT_PRODUCT( 7211 _numObjectsFreed++; 7212 _numWordsFreed += res; 7213 ) 7214 } else { 7215 // Chunk that is alive. 7216 res = do_live_chunk(fc); 7217 debug_only(_sp->verifyFreeLists()); 7218 NOT_PRODUCT( 7219 _numObjectsLive++; 7220 _numWordsLive += res; 7221 ) 7222 } 7223 return res; 7224 } 7225 7226 // For the smart allocation, record following 7227 // split deaths - a free chunk is removed from its free list because 7228 // it is being split into two or more chunks. 7229 // split birth - a free chunk is being added to its free list because 7230 // a larger free chunk has been split and resulted in this free chunk. 7231 // coal death - a free chunk is being removed from its free list because 7232 // it is being coalesced into a large free chunk. 7233 // coal birth - a free chunk is being added to its free list because 7234 // it was created when two or more free chunks where coalesced into 7235 // this free chunk. 7236 // 7237 // These statistics are used to determine the desired number of free 7238 // chunks of a given size. The desired number is chosen to be relative 7239 // to the end of a CMS sweep. The desired number at the end of a sweep 7240 // is the 7241 // count-at-end-of-previous-sweep (an amount that was enough) 7242 // - count-at-beginning-of-current-sweep (the excess) 7243 // + split-births (gains in this size during interval) 7244 // - split-deaths (demands on this size during interval) 7245 // where the interval is from the end of one sweep to the end of the 7246 // next. 7247 // 7248 // When sweeping the sweeper maintains an accumulated chunk which is 7249 // the chunk that is made up of chunks that have been coalesced. That 7250 // will be termed the left-hand chunk. A new chunk of garbage that 7251 // is being considered for coalescing will be referred to as the 7252 // right-hand chunk. 7253 // 7254 // When making a decision on whether to coalesce a right-hand chunk with 7255 // the current left-hand chunk, the current count vs. the desired count 7256 // of the left-hand chunk is considered. Also if the right-hand chunk 7257 // is near the large chunk at the end of the heap (see 7258 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7259 // left-hand chunk is coalesced. 7260 // 7261 // When making a decision about whether to split a chunk, the desired count 7262 // vs. the current count of the candidate to be split is also considered. 7263 // If the candidate is underpopulated (currently fewer chunks than desired) 7264 // a chunk of an overpopulated (currently more chunks than desired) size may 7265 // be chosen. The "hint" associated with a free list, if non-null, points 7266 // to a free list which may be overpopulated. 7267 // 7268 7269 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7270 const size_t size = fc->size(); 7271 // Chunks that cannot be coalesced are not in the 7272 // free lists. 7273 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7274 assert(_sp->verify_chunk_in_free_list(fc), 7275 "free chunk should be in free lists"); 7276 } 7277 // a chunk that is already free, should not have been 7278 // marked in the bit map 7279 HeapWord* const addr = (HeapWord*) fc; 7280 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7281 // Verify that the bit map has no bits marked between 7282 // addr and purported end of this block. 7283 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7284 7285 // Some chunks cannot be coalesced under any circumstances. 7286 // See the definition of cantCoalesce(). 7287 if (!fc->cantCoalesce()) { 7288 // This chunk can potentially be coalesced. 7289 // All the work is done in 7290 do_post_free_or_garbage_chunk(fc, size); 7291 // Note that if the chunk is not coalescable (the else arm 7292 // below), we unconditionally flush, without needing to do 7293 // a "lookahead," as we do below. 7294 if (inFreeRange()) lookahead_and_flush(fc, size); 7295 } else { 7296 // Code path common to both original and adaptive free lists. 7297 7298 // cant coalesce with previous block; this should be treated 7299 // as the end of a free run if any 7300 if (inFreeRange()) { 7301 // we kicked some butt; time to pick up the garbage 7302 assert(freeFinger() < addr, "freeFinger points too high"); 7303 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7304 } 7305 // else, nothing to do, just continue 7306 } 7307 } 7308 7309 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7310 // This is a chunk of garbage. It is not in any free list. 7311 // Add it to a free list or let it possibly be coalesced into 7312 // a larger chunk. 7313 HeapWord* const addr = (HeapWord*) fc; 7314 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7315 7316 // Verify that the bit map has no bits marked between 7317 // addr and purported end of just dead object. 7318 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7319 do_post_free_or_garbage_chunk(fc, size); 7320 7321 assert(_limit >= addr + size, 7322 "A freshly garbage chunk can't possibly straddle over _limit"); 7323 if (inFreeRange()) lookahead_and_flush(fc, size); 7324 return size; 7325 } 7326 7327 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7328 HeapWord* addr = (HeapWord*) fc; 7329 // The sweeper has just found a live object. Return any accumulated 7330 // left hand chunk to the free lists. 7331 if (inFreeRange()) { 7332 assert(freeFinger() < addr, "freeFinger points too high"); 7333 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7334 } 7335 7336 // This object is live: we'd normally expect this to be 7337 // an oop, and like to assert the following: 7338 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7339 // However, as we commented above, this may be an object whose 7340 // header hasn't yet been initialized. 7341 size_t size; 7342 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7343 if (_bitMap->isMarked(addr + 1)) { 7344 // Determine the size from the bit map, rather than trying to 7345 // compute it from the object header. 7346 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7347 size = pointer_delta(nextOneAddr + 1, addr); 7348 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7349 "alignment problem"); 7350 7351 #ifdef ASSERT 7352 if (oop(addr)->klass_or_null_acquire() != NULL) { 7353 // Ignore mark word because we are running concurrent with mutators 7354 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7355 assert(size == 7356 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7357 "P-mark and computed size do not agree"); 7358 } 7359 #endif 7360 7361 } else { 7362 // This should be an initialized object that's alive. 7363 assert(oop(addr)->klass_or_null_acquire() != NULL, 7364 "Should be an initialized object"); 7365 // Ignore mark word because we are running concurrent with mutators 7366 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7367 // Verify that the bit map has no bits marked between 7368 // addr and purported end of this block. 7369 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7370 assert(size >= 3, "Necessary for Printezis marks to work"); 7371 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7372 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7373 } 7374 return size; 7375 } 7376 7377 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7378 size_t chunkSize) { 7379 // do_post_free_or_garbage_chunk() should only be called in the case 7380 // of the adaptive free list allocator. 7381 const bool fcInFreeLists = fc->is_free(); 7382 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7383 if (CMSTestInFreeList && fcInFreeLists) { 7384 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7385 } 7386 7387 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7388 7389 HeapWord* const fc_addr = (HeapWord*) fc; 7390 7391 bool coalesce = false; 7392 const size_t left = pointer_delta(fc_addr, freeFinger()); 7393 const size_t right = chunkSize; 7394 switch (FLSCoalescePolicy) { 7395 // numeric value forms a coalition aggressiveness metric 7396 case 0: { // never coalesce 7397 coalesce = false; 7398 break; 7399 } 7400 case 1: { // coalesce if left & right chunks on overpopulated lists 7401 coalesce = _sp->coalOverPopulated(left) && 7402 _sp->coalOverPopulated(right); 7403 break; 7404 } 7405 case 2: { // coalesce if left chunk on overpopulated list (default) 7406 coalesce = _sp->coalOverPopulated(left); 7407 break; 7408 } 7409 case 3: { // coalesce if left OR right chunk on overpopulated list 7410 coalesce = _sp->coalOverPopulated(left) || 7411 _sp->coalOverPopulated(right); 7412 break; 7413 } 7414 case 4: { // always coalesce 7415 coalesce = true; 7416 break; 7417 } 7418 default: 7419 ShouldNotReachHere(); 7420 } 7421 7422 // Should the current free range be coalesced? 7423 // If the chunk is in a free range and either we decided to coalesce above 7424 // or the chunk is near the large block at the end of the heap 7425 // (isNearLargestChunk() returns true), then coalesce this chunk. 7426 const bool doCoalesce = inFreeRange() 7427 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7428 if (doCoalesce) { 7429 // Coalesce the current free range on the left with the new 7430 // chunk on the right. If either is on a free list, 7431 // it must be removed from the list and stashed in the closure. 7432 if (freeRangeInFreeLists()) { 7433 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7434 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7435 "Size of free range is inconsistent with chunk size."); 7436 if (CMSTestInFreeList) { 7437 assert(_sp->verify_chunk_in_free_list(ffc), 7438 "Chunk is not in free lists"); 7439 } 7440 _sp->coalDeath(ffc->size()); 7441 _sp->removeFreeChunkFromFreeLists(ffc); 7442 set_freeRangeInFreeLists(false); 7443 } 7444 if (fcInFreeLists) { 7445 _sp->coalDeath(chunkSize); 7446 assert(fc->size() == chunkSize, 7447 "The chunk has the wrong size or is not in the free lists"); 7448 _sp->removeFreeChunkFromFreeLists(fc); 7449 } 7450 set_lastFreeRangeCoalesced(true); 7451 print_free_block_coalesced(fc); 7452 } else { // not in a free range and/or should not coalesce 7453 // Return the current free range and start a new one. 7454 if (inFreeRange()) { 7455 // In a free range but cannot coalesce with the right hand chunk. 7456 // Put the current free range into the free lists. 7457 flush_cur_free_chunk(freeFinger(), 7458 pointer_delta(fc_addr, freeFinger())); 7459 } 7460 // Set up for new free range. Pass along whether the right hand 7461 // chunk is in the free lists. 7462 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7463 } 7464 } 7465 7466 // Lookahead flush: 7467 // If we are tracking a free range, and this is the last chunk that 7468 // we'll look at because its end crosses past _limit, we'll preemptively 7469 // flush it along with any free range we may be holding on to. Note that 7470 // this can be the case only for an already free or freshly garbage 7471 // chunk. If this block is an object, it can never straddle 7472 // over _limit. The "straddling" occurs when _limit is set at 7473 // the previous end of the space when this cycle started, and 7474 // a subsequent heap expansion caused the previously co-terminal 7475 // free block to be coalesced with the newly expanded portion, 7476 // thus rendering _limit a non-block-boundary making it dangerous 7477 // for the sweeper to step over and examine. 7478 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7479 assert(inFreeRange(), "Should only be called if currently in a free range."); 7480 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7481 assert(_sp->used_region().contains(eob - 1), 7482 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7483 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7484 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7485 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7486 if (eob >= _limit) { 7487 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7488 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7489 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7490 "[" PTR_FORMAT "," PTR_FORMAT ")", 7491 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7492 // Return the storage we are tracking back into the free lists. 7493 log_develop_trace(gc, sweep)("Flushing ... "); 7494 assert(freeFinger() < eob, "Error"); 7495 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7496 } 7497 } 7498 7499 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7500 assert(inFreeRange(), "Should only be called if currently in a free range."); 7501 assert(size > 0, 7502 "A zero sized chunk cannot be added to the free lists."); 7503 if (!freeRangeInFreeLists()) { 7504 if (CMSTestInFreeList) { 7505 FreeChunk* fc = (FreeChunk*) chunk; 7506 fc->set_size(size); 7507 assert(!_sp->verify_chunk_in_free_list(fc), 7508 "chunk should not be in free lists yet"); 7509 } 7510 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7511 // A new free range is going to be starting. The current 7512 // free range has not been added to the free lists yet or 7513 // was removed so add it back. 7514 // If the current free range was coalesced, then the death 7515 // of the free range was recorded. Record a birth now. 7516 if (lastFreeRangeCoalesced()) { 7517 _sp->coalBirth(size); 7518 } 7519 _sp->addChunkAndRepairOffsetTable(chunk, size, 7520 lastFreeRangeCoalesced()); 7521 } else { 7522 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7523 } 7524 set_inFreeRange(false); 7525 set_freeRangeInFreeLists(false); 7526 } 7527 7528 // We take a break if we've been at this for a while, 7529 // so as to avoid monopolizing the locks involved. 7530 void SweepClosure::do_yield_work(HeapWord* addr) { 7531 // Return current free chunk being used for coalescing (if any) 7532 // to the appropriate freelist. After yielding, the next 7533 // free block encountered will start a coalescing range of 7534 // free blocks. If the next free block is adjacent to the 7535 // chunk just flushed, they will need to wait for the next 7536 // sweep to be coalesced. 7537 if (inFreeRange()) { 7538 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7539 } 7540 7541 // First give up the locks, then yield, then re-lock. 7542 // We should probably use a constructor/destructor idiom to 7543 // do this unlock/lock or modify the MutexUnlocker class to 7544 // serve our purpose. XXX 7545 assert_lock_strong(_bitMap->lock()); 7546 assert_lock_strong(_freelistLock); 7547 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7548 "CMS thread should hold CMS token"); 7549 _bitMap->lock()->unlock(); 7550 _freelistLock->unlock(); 7551 ConcurrentMarkSweepThread::desynchronize(true); 7552 _collector->stopTimer(); 7553 _collector->incrementYields(); 7554 7555 // See the comment in coordinator_yield() 7556 for (unsigned i = 0; i < CMSYieldSleepCount && 7557 ConcurrentMarkSweepThread::should_yield() && 7558 !CMSCollector::foregroundGCIsActive(); ++i) { 7559 os::sleep(Thread::current(), 1, false); 7560 } 7561 7562 ConcurrentMarkSweepThread::synchronize(true); 7563 _freelistLock->lock(); 7564 _bitMap->lock()->lock_without_safepoint_check(); 7565 _collector->startTimer(); 7566 } 7567 7568 #ifndef PRODUCT 7569 // This is actually very useful in a product build if it can 7570 // be called from the debugger. Compile it into the product 7571 // as needed. 7572 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7573 return debug_cms_space->verify_chunk_in_free_list(fc); 7574 } 7575 #endif 7576 7577 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7578 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7579 p2i(fc), fc->size()); 7580 } 7581 7582 // CMSIsAliveClosure 7583 bool CMSIsAliveClosure::do_object_b(oop obj) { 7584 HeapWord* addr = (HeapWord*)obj; 7585 return addr != NULL && 7586 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7587 } 7588 7589 7590 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7591 MemRegion span, 7592 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7593 bool cpc): 7594 _collector(collector), 7595 _span(span), 7596 _bit_map(bit_map), 7597 _mark_stack(mark_stack), 7598 _concurrent_precleaning(cpc) { 7599 assert(!_span.is_empty(), "Empty span could spell trouble"); 7600 } 7601 7602 7603 // CMSKeepAliveClosure: the serial version 7604 void CMSKeepAliveClosure::do_oop(oop obj) { 7605 HeapWord* addr = (HeapWord*)obj; 7606 if (_span.contains(addr) && 7607 !_bit_map->isMarked(addr)) { 7608 _bit_map->mark(addr); 7609 bool simulate_overflow = false; 7610 NOT_PRODUCT( 7611 if (CMSMarkStackOverflowALot && 7612 _collector->simulate_overflow()) { 7613 // simulate a stack overflow 7614 simulate_overflow = true; 7615 } 7616 ) 7617 if (simulate_overflow || !_mark_stack->push(obj)) { 7618 if (_concurrent_precleaning) { 7619 // We dirty the overflown object and let the remark 7620 // phase deal with it. 7621 assert(_collector->overflow_list_is_empty(), "Error"); 7622 // In the case of object arrays, we need to dirty all of 7623 // the cards that the object spans. No locking or atomics 7624 // are needed since no one else can be mutating the mod union 7625 // table. 7626 if (obj->is_objArray()) { 7627 size_t sz = obj->size(); 7628 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7629 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7630 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7631 _collector->_modUnionTable.mark_range(redirty_range); 7632 } else { 7633 _collector->_modUnionTable.mark(addr); 7634 } 7635 _collector->_ser_kac_preclean_ovflw++; 7636 } else { 7637 _collector->push_on_overflow_list(obj); 7638 _collector->_ser_kac_ovflw++; 7639 } 7640 } 7641 } 7642 } 7643 7644 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7645 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7646 7647 // CMSParKeepAliveClosure: a parallel version of the above. 7648 // The work queues are private to each closure (thread), 7649 // but (may be) available for stealing by other threads. 7650 void CMSParKeepAliveClosure::do_oop(oop obj) { 7651 HeapWord* addr = (HeapWord*)obj; 7652 if (_span.contains(addr) && 7653 !_bit_map->isMarked(addr)) { 7654 // In general, during recursive tracing, several threads 7655 // may be concurrently getting here; the first one to 7656 // "tag" it, claims it. 7657 if (_bit_map->par_mark(addr)) { 7658 bool res = _work_queue->push(obj); 7659 assert(res, "Low water mark should be much less than capacity"); 7660 // Do a recursive trim in the hope that this will keep 7661 // stack usage lower, but leave some oops for potential stealers 7662 trim_queue(_low_water_mark); 7663 } // Else, another thread got there first 7664 } 7665 } 7666 7667 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7668 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7669 7670 void CMSParKeepAliveClosure::trim_queue(uint max) { 7671 while (_work_queue->size() > max) { 7672 oop new_oop; 7673 if (_work_queue->pop_local(new_oop)) { 7674 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7675 assert(_bit_map->isMarked((HeapWord*)new_oop), 7676 "no white objects on this stack!"); 7677 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7678 // iterate over the oops in this oop, marking and pushing 7679 // the ones in CMS heap (i.e. in _span). 7680 new_oop->oop_iterate(&_mark_and_push); 7681 } 7682 } 7683 } 7684 7685 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7686 CMSCollector* collector, 7687 MemRegion span, CMSBitMap* bit_map, 7688 OopTaskQueue* work_queue): 7689 _collector(collector), 7690 _span(span), 7691 _bit_map(bit_map), 7692 _work_queue(work_queue) { } 7693 7694 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7695 HeapWord* addr = (HeapWord*)obj; 7696 if (_span.contains(addr) && 7697 !_bit_map->isMarked(addr)) { 7698 if (_bit_map->par_mark(addr)) { 7699 bool simulate_overflow = false; 7700 NOT_PRODUCT( 7701 if (CMSMarkStackOverflowALot && 7702 _collector->par_simulate_overflow()) { 7703 // simulate a stack overflow 7704 simulate_overflow = true; 7705 } 7706 ) 7707 if (simulate_overflow || !_work_queue->push(obj)) { 7708 _collector->par_push_on_overflow_list(obj); 7709 _collector->_par_kac_ovflw++; 7710 } 7711 } // Else another thread got there already 7712 } 7713 } 7714 7715 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7716 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7717 7718 ////////////////////////////////////////////////////////////////// 7719 // CMSExpansionCause ///////////////////////////// 7720 ////////////////////////////////////////////////////////////////// 7721 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7722 switch (cause) { 7723 case _no_expansion: 7724 return "No expansion"; 7725 case _satisfy_free_ratio: 7726 return "Free ratio"; 7727 case _satisfy_promotion: 7728 return "Satisfy promotion"; 7729 case _satisfy_allocation: 7730 return "allocation"; 7731 case _allocate_par_lab: 7732 return "Par LAB"; 7733 case _allocate_par_spooling_space: 7734 return "Par Spooling Space"; 7735 case _adaptive_size_policy: 7736 return "Ergonomics"; 7737 default: 7738 return "unknown"; 7739 } 7740 } 7741 7742 void CMSDrainMarkingStackClosure::do_void() { 7743 // the max number to take from overflow list at a time 7744 const size_t num = _mark_stack->capacity()/4; 7745 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7746 "Overflow list should be NULL during concurrent phases"); 7747 while (!_mark_stack->isEmpty() || 7748 // if stack is empty, check the overflow list 7749 _collector->take_from_overflow_list(num, _mark_stack)) { 7750 oop obj = _mark_stack->pop(); 7751 HeapWord* addr = (HeapWord*)obj; 7752 assert(_span.contains(addr), "Should be within span"); 7753 assert(_bit_map->isMarked(addr), "Should be marked"); 7754 assert(oopDesc::is_oop(obj), "Should be an oop"); 7755 obj->oop_iterate(_keep_alive); 7756 } 7757 } 7758 7759 void CMSParDrainMarkingStackClosure::do_void() { 7760 // drain queue 7761 trim_queue(0); 7762 } 7763 7764 // Trim our work_queue so its length is below max at return 7765 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7766 while (_work_queue->size() > max) { 7767 oop new_oop; 7768 if (_work_queue->pop_local(new_oop)) { 7769 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7770 assert(_bit_map->isMarked((HeapWord*)new_oop), 7771 "no white objects on this stack!"); 7772 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7773 // iterate over the oops in this oop, marking and pushing 7774 // the ones in CMS heap (i.e. in _span). 7775 new_oop->oop_iterate(&_mark_and_push); 7776 } 7777 } 7778 } 7779 7780 //////////////////////////////////////////////////////////////////// 7781 // Support for Marking Stack Overflow list handling and related code 7782 //////////////////////////////////////////////////////////////////// 7783 // Much of the following code is similar in shape and spirit to the 7784 // code used in ParNewGC. We should try and share that code 7785 // as much as possible in the future. 7786 7787 #ifndef PRODUCT 7788 // Debugging support for CMSStackOverflowALot 7789 7790 // It's OK to call this multi-threaded; the worst thing 7791 // that can happen is that we'll get a bunch of closely 7792 // spaced simulated overflows, but that's OK, in fact 7793 // probably good as it would exercise the overflow code 7794 // under contention. 7795 bool CMSCollector::simulate_overflow() { 7796 if (_overflow_counter-- <= 0) { // just being defensive 7797 _overflow_counter = CMSMarkStackOverflowInterval; 7798 return true; 7799 } else { 7800 return false; 7801 } 7802 } 7803 7804 bool CMSCollector::par_simulate_overflow() { 7805 return simulate_overflow(); 7806 } 7807 #endif 7808 7809 // Single-threaded 7810 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7811 assert(stack->isEmpty(), "Expected precondition"); 7812 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7813 size_t i = num; 7814 oop cur = _overflow_list; 7815 const markOop proto = markOopDesc::prototype(); 7816 NOT_PRODUCT(ssize_t n = 0;) 7817 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7818 next = oop(cur->mark()); 7819 cur->set_mark(proto); // until proven otherwise 7820 assert(oopDesc::is_oop(cur), "Should be an oop"); 7821 bool res = stack->push(cur); 7822 assert(res, "Bit off more than can chew?"); 7823 NOT_PRODUCT(n++;) 7824 } 7825 _overflow_list = cur; 7826 #ifndef PRODUCT 7827 assert(_num_par_pushes >= n, "Too many pops?"); 7828 _num_par_pushes -=n; 7829 #endif 7830 return !stack->isEmpty(); 7831 } 7832 7833 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7834 // (MT-safe) Get a prefix of at most "num" from the list. 7835 // The overflow list is chained through the mark word of 7836 // each object in the list. We fetch the entire list, 7837 // break off a prefix of the right size and return the 7838 // remainder. If other threads try to take objects from 7839 // the overflow list at that time, they will wait for 7840 // some time to see if data becomes available. If (and 7841 // only if) another thread places one or more object(s) 7842 // on the global list before we have returned the suffix 7843 // to the global list, we will walk down our local list 7844 // to find its end and append the global list to 7845 // our suffix before returning it. This suffix walk can 7846 // prove to be expensive (quadratic in the amount of traffic) 7847 // when there are many objects in the overflow list and 7848 // there is much producer-consumer contention on the list. 7849 // *NOTE*: The overflow list manipulation code here and 7850 // in ParNewGeneration:: are very similar in shape, 7851 // except that in the ParNew case we use the old (from/eden) 7852 // copy of the object to thread the list via its klass word. 7853 // Because of the common code, if you make any changes in 7854 // the code below, please check the ParNew version to see if 7855 // similar changes might be needed. 7856 // CR 6797058 has been filed to consolidate the common code. 7857 bool CMSCollector::par_take_from_overflow_list(size_t num, 7858 OopTaskQueue* work_q, 7859 int no_of_gc_threads) { 7860 assert(work_q->size() == 0, "First empty local work queue"); 7861 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7862 if (_overflow_list == NULL) { 7863 return false; 7864 } 7865 // Grab the entire list; we'll put back a suffix 7866 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7867 Thread* tid = Thread::current(); 7868 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7869 // set to ParallelGCThreads. 7870 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7871 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7872 // If the list is busy, we spin for a short while, 7873 // sleeping between attempts to get the list. 7874 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7875 os::sleep(tid, sleep_time_millis, false); 7876 if (_overflow_list == NULL) { 7877 // Nothing left to take 7878 return false; 7879 } else if (_overflow_list != BUSY) { 7880 // Try and grab the prefix 7881 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7882 } 7883 } 7884 // If the list was found to be empty, or we spun long 7885 // enough, we give up and return empty-handed. If we leave 7886 // the list in the BUSY state below, it must be the case that 7887 // some other thread holds the overflow list and will set it 7888 // to a non-BUSY state in the future. 7889 if (prefix == NULL || prefix == BUSY) { 7890 // Nothing to take or waited long enough 7891 if (prefix == NULL) { 7892 // Write back the NULL in case we overwrote it with BUSY above 7893 // and it is still the same value. 7894 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7895 } 7896 return false; 7897 } 7898 assert(prefix != NULL && prefix != BUSY, "Error"); 7899 size_t i = num; 7900 oop cur = prefix; 7901 // Walk down the first "num" objects, unless we reach the end. 7902 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7903 if (cur->mark() == NULL) { 7904 // We have "num" or fewer elements in the list, so there 7905 // is nothing to return to the global list. 7906 // Write back the NULL in lieu of the BUSY we wrote 7907 // above, if it is still the same value. 7908 if (_overflow_list == BUSY) { 7909 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7910 } 7911 } else { 7912 // Chop off the suffix and return it to the global list. 7913 assert(cur->mark() != BUSY, "Error"); 7914 oop suffix_head = cur->mark(); // suffix will be put back on global list 7915 cur->set_mark(NULL); // break off suffix 7916 // It's possible that the list is still in the empty(busy) state 7917 // we left it in a short while ago; in that case we may be 7918 // able to place back the suffix without incurring the cost 7919 // of a walk down the list. 7920 oop observed_overflow_list = _overflow_list; 7921 oop cur_overflow_list = observed_overflow_list; 7922 bool attached = false; 7923 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7924 observed_overflow_list = 7925 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7926 if (cur_overflow_list == observed_overflow_list) { 7927 attached = true; 7928 break; 7929 } else cur_overflow_list = observed_overflow_list; 7930 } 7931 if (!attached) { 7932 // Too bad, someone else sneaked in (at least) an element; we'll need 7933 // to do a splice. Find tail of suffix so we can prepend suffix to global 7934 // list. 7935 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7936 oop suffix_tail = cur; 7937 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7938 "Tautology"); 7939 observed_overflow_list = _overflow_list; 7940 do { 7941 cur_overflow_list = observed_overflow_list; 7942 if (cur_overflow_list != BUSY) { 7943 // Do the splice ... 7944 suffix_tail->set_mark(markOop(cur_overflow_list)); 7945 } else { // cur_overflow_list == BUSY 7946 suffix_tail->set_mark(NULL); 7947 } 7948 // ... and try to place spliced list back on overflow_list ... 7949 observed_overflow_list = 7950 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7951 } while (cur_overflow_list != observed_overflow_list); 7952 // ... until we have succeeded in doing so. 7953 } 7954 } 7955 7956 // Push the prefix elements on work_q 7957 assert(prefix != NULL, "control point invariant"); 7958 const markOop proto = markOopDesc::prototype(); 7959 oop next; 7960 NOT_PRODUCT(ssize_t n = 0;) 7961 for (cur = prefix; cur != NULL; cur = next) { 7962 next = oop(cur->mark()); 7963 cur->set_mark(proto); // until proven otherwise 7964 assert(oopDesc::is_oop(cur), "Should be an oop"); 7965 bool res = work_q->push(cur); 7966 assert(res, "Bit off more than we can chew?"); 7967 NOT_PRODUCT(n++;) 7968 } 7969 #ifndef PRODUCT 7970 assert(_num_par_pushes >= n, "Too many pops?"); 7971 Atomic::sub(n, &_num_par_pushes); 7972 #endif 7973 return true; 7974 } 7975 7976 // Single-threaded 7977 void CMSCollector::push_on_overflow_list(oop p) { 7978 NOT_PRODUCT(_num_par_pushes++;) 7979 assert(oopDesc::is_oop(p), "Not an oop"); 7980 preserve_mark_if_necessary(p); 7981 p->set_mark((markOop)_overflow_list); 7982 _overflow_list = p; 7983 } 7984 7985 // Multi-threaded; use CAS to prepend to overflow list 7986 void CMSCollector::par_push_on_overflow_list(oop p) { 7987 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7988 assert(oopDesc::is_oop(p), "Not an oop"); 7989 par_preserve_mark_if_necessary(p); 7990 oop observed_overflow_list = _overflow_list; 7991 oop cur_overflow_list; 7992 do { 7993 cur_overflow_list = observed_overflow_list; 7994 if (cur_overflow_list != BUSY) { 7995 p->set_mark(markOop(cur_overflow_list)); 7996 } else { 7997 p->set_mark(NULL); 7998 } 7999 observed_overflow_list = 8000 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 8001 } while (cur_overflow_list != observed_overflow_list); 8002 } 8003 #undef BUSY 8004 8005 // Single threaded 8006 // General Note on GrowableArray: pushes may silently fail 8007 // because we are (temporarily) out of C-heap for expanding 8008 // the stack. The problem is quite ubiquitous and affects 8009 // a lot of code in the JVM. The prudent thing for GrowableArray 8010 // to do (for now) is to exit with an error. However, that may 8011 // be too draconian in some cases because the caller may be 8012 // able to recover without much harm. For such cases, we 8013 // should probably introduce a "soft_push" method which returns 8014 // an indication of success or failure with the assumption that 8015 // the caller may be able to recover from a failure; code in 8016 // the VM can then be changed, incrementally, to deal with such 8017 // failures where possible, thus, incrementally hardening the VM 8018 // in such low resource situations. 8019 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8020 _preserved_oop_stack.push(p); 8021 _preserved_mark_stack.push(m); 8022 assert(m == p->mark(), "Mark word changed"); 8023 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8024 "bijection"); 8025 } 8026 8027 // Single threaded 8028 void CMSCollector::preserve_mark_if_necessary(oop p) { 8029 markOop m = p->mark(); 8030 if (m->must_be_preserved(p)) { 8031 preserve_mark_work(p, m); 8032 } 8033 } 8034 8035 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8036 markOop m = p->mark(); 8037 if (m->must_be_preserved(p)) { 8038 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8039 // Even though we read the mark word without holding 8040 // the lock, we are assured that it will not change 8041 // because we "own" this oop, so no other thread can 8042 // be trying to push it on the overflow list; see 8043 // the assertion in preserve_mark_work() that checks 8044 // that m == p->mark(). 8045 preserve_mark_work(p, m); 8046 } 8047 } 8048 8049 // We should be able to do this multi-threaded, 8050 // a chunk of stack being a task (this is 8051 // correct because each oop only ever appears 8052 // once in the overflow list. However, it's 8053 // not very easy to completely overlap this with 8054 // other operations, so will generally not be done 8055 // until all work's been completed. Because we 8056 // expect the preserved oop stack (set) to be small, 8057 // it's probably fine to do this single-threaded. 8058 // We can explore cleverer concurrent/overlapped/parallel 8059 // processing of preserved marks if we feel the 8060 // need for this in the future. Stack overflow should 8061 // be so rare in practice and, when it happens, its 8062 // effect on performance so great that this will 8063 // likely just be in the noise anyway. 8064 void CMSCollector::restore_preserved_marks_if_any() { 8065 assert(SafepointSynchronize::is_at_safepoint(), 8066 "world should be stopped"); 8067 assert(Thread::current()->is_ConcurrentGC_thread() || 8068 Thread::current()->is_VM_thread(), 8069 "should be single-threaded"); 8070 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8071 "bijection"); 8072 8073 while (!_preserved_oop_stack.is_empty()) { 8074 oop p = _preserved_oop_stack.pop(); 8075 assert(oopDesc::is_oop(p), "Should be an oop"); 8076 assert(_span.contains(p), "oop should be in _span"); 8077 assert(p->mark() == markOopDesc::prototype(), 8078 "Set when taken from overflow list"); 8079 markOop m = _preserved_mark_stack.pop(); 8080 p->set_mark(m); 8081 } 8082 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8083 "stacks were cleared above"); 8084 } 8085 8086 #ifndef PRODUCT 8087 bool CMSCollector::no_preserved_marks() const { 8088 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8089 } 8090 #endif 8091 8092 // Transfer some number of overflown objects to usual marking 8093 // stack. Return true if some objects were transferred. 8094 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8095 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8096 (size_t)ParGCDesiredObjsFromOverflowList); 8097 8098 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8099 assert(_collector->overflow_list_is_empty() || res, 8100 "If list is not empty, we should have taken something"); 8101 assert(!res || !_mark_stack->isEmpty(), 8102 "If we took something, it should now be on our stack"); 8103 return res; 8104 } 8105 8106 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8107 size_t res = _sp->block_size_no_stall(addr, _collector); 8108 if (_sp->block_is_obj(addr)) { 8109 if (_live_bit_map->isMarked(addr)) { 8110 // It can't have been dead in a previous cycle 8111 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8112 } else { 8113 _dead_bit_map->mark(addr); // mark the dead object 8114 } 8115 } 8116 // Could be 0, if the block size could not be computed without stalling. 8117 return res; 8118 } 8119 8120 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8121 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8122 switch (phase) { 8123 case CMSCollector::InitialMarking: 8124 initialize(manager /* GC manager */ , 8125 cause /* cause of the GC */, 8126 true /* recordGCBeginTime */, 8127 true /* recordPreGCUsage */, 8128 false /* recordPeakUsage */, 8129 false /* recordPostGCusage */, 8130 true /* recordAccumulatedGCTime */, 8131 false /* recordGCEndTime */, 8132 false /* countCollection */ ); 8133 break; 8134 8135 case CMSCollector::FinalMarking: 8136 initialize(manager /* GC manager */ , 8137 cause /* cause of the GC */, 8138 false /* recordGCBeginTime */, 8139 false /* recordPreGCUsage */, 8140 false /* recordPeakUsage */, 8141 false /* recordPostGCusage */, 8142 true /* recordAccumulatedGCTime */, 8143 false /* recordGCEndTime */, 8144 false /* countCollection */ ); 8145 break; 8146 8147 case CMSCollector::Sweeping: 8148 initialize(manager /* GC manager */ , 8149 cause /* cause of the GC */, 8150 false /* recordGCBeginTime */, 8151 false /* recordPreGCUsage */, 8152 true /* recordPeakUsage */, 8153 true /* recordPostGCusage */, 8154 false /* recordAccumulatedGCTime */, 8155 true /* recordGCEndTime */, 8156 true /* countCollection */ ); 8157 break; 8158 8159 default: 8160 ShouldNotReachHere(); 8161 } 8162 }